1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com 3 * Copyright (c) 2016 Facebook 4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io 5 */ 6 #include <uapi/linux/btf.h> 7 #include <linux/bpf-cgroup.h> 8 #include <linux/kernel.h> 9 #include <linux/types.h> 10 #include <linux/slab.h> 11 #include <linux/bpf.h> 12 #include <linux/btf.h> 13 #include <linux/bpf_verifier.h> 14 #include <linux/filter.h> 15 #include <net/netlink.h> 16 #include <linux/file.h> 17 #include <linux/vmalloc.h> 18 #include <linux/stringify.h> 19 #include <linux/bsearch.h> 20 #include <linux/sort.h> 21 #include <linux/perf_event.h> 22 #include <linux/ctype.h> 23 #include <linux/error-injection.h> 24 #include <linux/bpf_lsm.h> 25 #include <linux/btf_ids.h> 26 #include <linux/poison.h> 27 #include <linux/module.h> 28 29 #include "disasm.h" 30 31 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = { 32 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \ 33 [_id] = & _name ## _verifier_ops, 34 #define BPF_MAP_TYPE(_id, _ops) 35 #define BPF_LINK_TYPE(_id, _name) 36 #include <linux/bpf_types.h> 37 #undef BPF_PROG_TYPE 38 #undef BPF_MAP_TYPE 39 #undef BPF_LINK_TYPE 40 }; 41 42 /* bpf_check() is a static code analyzer that walks eBPF program 43 * instruction by instruction and updates register/stack state. 44 * All paths of conditional branches are analyzed until 'bpf_exit' insn. 45 * 46 * The first pass is depth-first-search to check that the program is a DAG. 47 * It rejects the following programs: 48 * - larger than BPF_MAXINSNS insns 49 * - if loop is present (detected via back-edge) 50 * - unreachable insns exist (shouldn't be a forest. program = one function) 51 * - out of bounds or malformed jumps 52 * The second pass is all possible path descent from the 1st insn. 53 * Since it's analyzing all paths through the program, the length of the 54 * analysis is limited to 64k insn, which may be hit even if total number of 55 * insn is less then 4K, but there are too many branches that change stack/regs. 56 * Number of 'branches to be analyzed' is limited to 1k 57 * 58 * On entry to each instruction, each register has a type, and the instruction 59 * changes the types of the registers depending on instruction semantics. 60 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is 61 * copied to R1. 62 * 63 * All registers are 64-bit. 64 * R0 - return register 65 * R1-R5 argument passing registers 66 * R6-R9 callee saved registers 67 * R10 - frame pointer read-only 68 * 69 * At the start of BPF program the register R1 contains a pointer to bpf_context 70 * and has type PTR_TO_CTX. 71 * 72 * Verifier tracks arithmetic operations on pointers in case: 73 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10), 74 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20), 75 * 1st insn copies R10 (which has FRAME_PTR) type into R1 76 * and 2nd arithmetic instruction is pattern matched to recognize 77 * that it wants to construct a pointer to some element within stack. 78 * So after 2nd insn, the register R1 has type PTR_TO_STACK 79 * (and -20 constant is saved for further stack bounds checking). 80 * Meaning that this reg is a pointer to stack plus known immediate constant. 81 * 82 * Most of the time the registers have SCALAR_VALUE type, which 83 * means the register has some value, but it's not a valid pointer. 84 * (like pointer plus pointer becomes SCALAR_VALUE type) 85 * 86 * When verifier sees load or store instructions the type of base register 87 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are 88 * four pointer types recognized by check_mem_access() function. 89 * 90 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value' 91 * and the range of [ptr, ptr + map's value_size) is accessible. 92 * 93 * registers used to pass values to function calls are checked against 94 * function argument constraints. 95 * 96 * ARG_PTR_TO_MAP_KEY is one of such argument constraints. 97 * It means that the register type passed to this function must be 98 * PTR_TO_STACK and it will be used inside the function as 99 * 'pointer to map element key' 100 * 101 * For example the argument constraints for bpf_map_lookup_elem(): 102 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL, 103 * .arg1_type = ARG_CONST_MAP_PTR, 104 * .arg2_type = ARG_PTR_TO_MAP_KEY, 105 * 106 * ret_type says that this function returns 'pointer to map elem value or null' 107 * function expects 1st argument to be a const pointer to 'struct bpf_map' and 108 * 2nd argument should be a pointer to stack, which will be used inside 109 * the helper function as a pointer to map element key. 110 * 111 * On the kernel side the helper function looks like: 112 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 113 * { 114 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1; 115 * void *key = (void *) (unsigned long) r2; 116 * void *value; 117 * 118 * here kernel can access 'key' and 'map' pointers safely, knowing that 119 * [key, key + map->key_size) bytes are valid and were initialized on 120 * the stack of eBPF program. 121 * } 122 * 123 * Corresponding eBPF program may look like: 124 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR 125 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK 126 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP 127 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem), 128 * here verifier looks at prototype of map_lookup_elem() and sees: 129 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok, 130 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes 131 * 132 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far, 133 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits 134 * and were initialized prior to this call. 135 * If it's ok, then verifier allows this BPF_CALL insn and looks at 136 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets 137 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function 138 * returns either pointer to map value or NULL. 139 * 140 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off' 141 * insn, the register holding that pointer in the true branch changes state to 142 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false 143 * branch. See check_cond_jmp_op(). 144 * 145 * After the call R0 is set to return type of the function and registers R1-R5 146 * are set to NOT_INIT to indicate that they are no longer readable. 147 * 148 * The following reference types represent a potential reference to a kernel 149 * resource which, after first being allocated, must be checked and freed by 150 * the BPF program: 151 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET 152 * 153 * When the verifier sees a helper call return a reference type, it allocates a 154 * pointer id for the reference and stores it in the current function state. 155 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into 156 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type 157 * passes through a NULL-check conditional. For the branch wherein the state is 158 * changed to CONST_IMM, the verifier releases the reference. 159 * 160 * For each helper function that allocates a reference, such as 161 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as 162 * bpf_sk_release(). When a reference type passes into the release function, 163 * the verifier also releases the reference. If any unchecked or unreleased 164 * reference remains at the end of the program, the verifier rejects it. 165 */ 166 167 /* verifier_state + insn_idx are pushed to stack when branch is encountered */ 168 struct bpf_verifier_stack_elem { 169 /* verifer state is 'st' 170 * before processing instruction 'insn_idx' 171 * and after processing instruction 'prev_insn_idx' 172 */ 173 struct bpf_verifier_state st; 174 int insn_idx; 175 int prev_insn_idx; 176 struct bpf_verifier_stack_elem *next; 177 /* length of verifier log at the time this state was pushed on stack */ 178 u32 log_pos; 179 }; 180 181 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192 182 #define BPF_COMPLEXITY_LIMIT_STATES 64 183 184 #define BPF_MAP_KEY_POISON (1ULL << 63) 185 #define BPF_MAP_KEY_SEEN (1ULL << 62) 186 187 #define BPF_MAP_PTR_UNPRIV 1UL 188 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \ 189 POISON_POINTER_DELTA)) 190 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV)) 191 192 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx); 193 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id); 194 static void invalidate_non_owning_refs(struct bpf_verifier_env *env); 195 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env); 196 static int ref_set_non_owning(struct bpf_verifier_env *env, 197 struct bpf_reg_state *reg); 198 static void specialize_kfunc(struct bpf_verifier_env *env, 199 u32 func_id, u16 offset, unsigned long *addr); 200 static bool is_trusted_reg(const struct bpf_reg_state *reg); 201 202 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux) 203 { 204 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON; 205 } 206 207 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux) 208 { 209 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV; 210 } 211 212 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux, 213 const struct bpf_map *map, bool unpriv) 214 { 215 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV); 216 unpriv |= bpf_map_ptr_unpriv(aux); 217 aux->map_ptr_state = (unsigned long)map | 218 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL); 219 } 220 221 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux) 222 { 223 return aux->map_key_state & BPF_MAP_KEY_POISON; 224 } 225 226 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux) 227 { 228 return !(aux->map_key_state & BPF_MAP_KEY_SEEN); 229 } 230 231 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux) 232 { 233 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON); 234 } 235 236 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state) 237 { 238 bool poisoned = bpf_map_key_poisoned(aux); 239 240 aux->map_key_state = state | BPF_MAP_KEY_SEEN | 241 (poisoned ? BPF_MAP_KEY_POISON : 0ULL); 242 } 243 244 static bool bpf_helper_call(const struct bpf_insn *insn) 245 { 246 return insn->code == (BPF_JMP | BPF_CALL) && 247 insn->src_reg == 0; 248 } 249 250 static bool bpf_pseudo_call(const struct bpf_insn *insn) 251 { 252 return insn->code == (BPF_JMP | BPF_CALL) && 253 insn->src_reg == BPF_PSEUDO_CALL; 254 } 255 256 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn) 257 { 258 return insn->code == (BPF_JMP | BPF_CALL) && 259 insn->src_reg == BPF_PSEUDO_KFUNC_CALL; 260 } 261 262 struct bpf_call_arg_meta { 263 struct bpf_map *map_ptr; 264 bool raw_mode; 265 bool pkt_access; 266 u8 release_regno; 267 int regno; 268 int access_size; 269 int mem_size; 270 u64 msize_max_value; 271 int ref_obj_id; 272 int dynptr_id; 273 int map_uid; 274 int func_id; 275 struct btf *btf; 276 u32 btf_id; 277 struct btf *ret_btf; 278 u32 ret_btf_id; 279 u32 subprogno; 280 struct btf_field *kptr_field; 281 }; 282 283 struct bpf_kfunc_call_arg_meta { 284 /* In parameters */ 285 struct btf *btf; 286 u32 func_id; 287 u32 kfunc_flags; 288 const struct btf_type *func_proto; 289 const char *func_name; 290 /* Out parameters */ 291 u32 ref_obj_id; 292 u8 release_regno; 293 bool r0_rdonly; 294 u32 ret_btf_id; 295 u64 r0_size; 296 u32 subprogno; 297 struct { 298 u64 value; 299 bool found; 300 } arg_constant; 301 302 /* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling, 303 * generally to pass info about user-defined local kptr types to later 304 * verification logic 305 * bpf_obj_drop 306 * Record the local kptr type to be drop'd 307 * bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type) 308 * Record the local kptr type to be refcount_incr'd and use 309 * arg_owning_ref to determine whether refcount_acquire should be 310 * fallible 311 */ 312 struct btf *arg_btf; 313 u32 arg_btf_id; 314 bool arg_owning_ref; 315 316 struct { 317 struct btf_field *field; 318 } arg_list_head; 319 struct { 320 struct btf_field *field; 321 } arg_rbtree_root; 322 struct { 323 enum bpf_dynptr_type type; 324 u32 id; 325 u32 ref_obj_id; 326 } initialized_dynptr; 327 struct { 328 u8 spi; 329 u8 frameno; 330 } iter; 331 u64 mem_size; 332 }; 333 334 struct btf *btf_vmlinux; 335 336 static DEFINE_MUTEX(bpf_verifier_lock); 337 338 static const struct bpf_line_info * 339 find_linfo(const struct bpf_verifier_env *env, u32 insn_off) 340 { 341 const struct bpf_line_info *linfo; 342 const struct bpf_prog *prog; 343 u32 i, nr_linfo; 344 345 prog = env->prog; 346 nr_linfo = prog->aux->nr_linfo; 347 348 if (!nr_linfo || insn_off >= prog->len) 349 return NULL; 350 351 linfo = prog->aux->linfo; 352 for (i = 1; i < nr_linfo; i++) 353 if (insn_off < linfo[i].insn_off) 354 break; 355 356 return &linfo[i - 1]; 357 } 358 359 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...) 360 { 361 struct bpf_verifier_env *env = private_data; 362 va_list args; 363 364 if (!bpf_verifier_log_needed(&env->log)) 365 return; 366 367 va_start(args, fmt); 368 bpf_verifier_vlog(&env->log, fmt, args); 369 va_end(args); 370 } 371 372 static const char *ltrim(const char *s) 373 { 374 while (isspace(*s)) 375 s++; 376 377 return s; 378 } 379 380 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env, 381 u32 insn_off, 382 const char *prefix_fmt, ...) 383 { 384 const struct bpf_line_info *linfo; 385 386 if (!bpf_verifier_log_needed(&env->log)) 387 return; 388 389 linfo = find_linfo(env, insn_off); 390 if (!linfo || linfo == env->prev_linfo) 391 return; 392 393 if (prefix_fmt) { 394 va_list args; 395 396 va_start(args, prefix_fmt); 397 bpf_verifier_vlog(&env->log, prefix_fmt, args); 398 va_end(args); 399 } 400 401 verbose(env, "%s\n", 402 ltrim(btf_name_by_offset(env->prog->aux->btf, 403 linfo->line_off))); 404 405 env->prev_linfo = linfo; 406 } 407 408 static void verbose_invalid_scalar(struct bpf_verifier_env *env, 409 struct bpf_reg_state *reg, 410 struct tnum *range, const char *ctx, 411 const char *reg_name) 412 { 413 char tn_buf[48]; 414 415 verbose(env, "At %s the register %s ", ctx, reg_name); 416 if (!tnum_is_unknown(reg->var_off)) { 417 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 418 verbose(env, "has value %s", tn_buf); 419 } else { 420 verbose(env, "has unknown scalar value"); 421 } 422 tnum_strn(tn_buf, sizeof(tn_buf), *range); 423 verbose(env, " should have been in %s\n", tn_buf); 424 } 425 426 static bool type_is_pkt_pointer(enum bpf_reg_type type) 427 { 428 type = base_type(type); 429 return type == PTR_TO_PACKET || 430 type == PTR_TO_PACKET_META; 431 } 432 433 static bool type_is_sk_pointer(enum bpf_reg_type type) 434 { 435 return type == PTR_TO_SOCKET || 436 type == PTR_TO_SOCK_COMMON || 437 type == PTR_TO_TCP_SOCK || 438 type == PTR_TO_XDP_SOCK; 439 } 440 441 static bool type_may_be_null(u32 type) 442 { 443 return type & PTR_MAYBE_NULL; 444 } 445 446 static bool reg_not_null(const struct bpf_reg_state *reg) 447 { 448 enum bpf_reg_type type; 449 450 type = reg->type; 451 if (type_may_be_null(type)) 452 return false; 453 454 type = base_type(type); 455 return type == PTR_TO_SOCKET || 456 type == PTR_TO_TCP_SOCK || 457 type == PTR_TO_MAP_VALUE || 458 type == PTR_TO_MAP_KEY || 459 type == PTR_TO_SOCK_COMMON || 460 (type == PTR_TO_BTF_ID && is_trusted_reg(reg)) || 461 type == PTR_TO_MEM; 462 } 463 464 static bool type_is_ptr_alloc_obj(u32 type) 465 { 466 return base_type(type) == PTR_TO_BTF_ID && type_flag(type) & MEM_ALLOC; 467 } 468 469 static bool type_is_non_owning_ref(u32 type) 470 { 471 return type_is_ptr_alloc_obj(type) && type_flag(type) & NON_OWN_REF; 472 } 473 474 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg) 475 { 476 struct btf_record *rec = NULL; 477 struct btf_struct_meta *meta; 478 479 if (reg->type == PTR_TO_MAP_VALUE) { 480 rec = reg->map_ptr->record; 481 } else if (type_is_ptr_alloc_obj(reg->type)) { 482 meta = btf_find_struct_meta(reg->btf, reg->btf_id); 483 if (meta) 484 rec = meta->record; 485 } 486 return rec; 487 } 488 489 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog) 490 { 491 struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux; 492 493 return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL; 494 } 495 496 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg) 497 { 498 return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK); 499 } 500 501 static bool type_is_rdonly_mem(u32 type) 502 { 503 return type & MEM_RDONLY; 504 } 505 506 static bool is_acquire_function(enum bpf_func_id func_id, 507 const struct bpf_map *map) 508 { 509 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC; 510 511 if (func_id == BPF_FUNC_sk_lookup_tcp || 512 func_id == BPF_FUNC_sk_lookup_udp || 513 func_id == BPF_FUNC_skc_lookup_tcp || 514 func_id == BPF_FUNC_ringbuf_reserve || 515 func_id == BPF_FUNC_kptr_xchg) 516 return true; 517 518 if (func_id == BPF_FUNC_map_lookup_elem && 519 (map_type == BPF_MAP_TYPE_SOCKMAP || 520 map_type == BPF_MAP_TYPE_SOCKHASH)) 521 return true; 522 523 return false; 524 } 525 526 static bool is_ptr_cast_function(enum bpf_func_id func_id) 527 { 528 return func_id == BPF_FUNC_tcp_sock || 529 func_id == BPF_FUNC_sk_fullsock || 530 func_id == BPF_FUNC_skc_to_tcp_sock || 531 func_id == BPF_FUNC_skc_to_tcp6_sock || 532 func_id == BPF_FUNC_skc_to_udp6_sock || 533 func_id == BPF_FUNC_skc_to_mptcp_sock || 534 func_id == BPF_FUNC_skc_to_tcp_timewait_sock || 535 func_id == BPF_FUNC_skc_to_tcp_request_sock; 536 } 537 538 static bool is_dynptr_ref_function(enum bpf_func_id func_id) 539 { 540 return func_id == BPF_FUNC_dynptr_data; 541 } 542 543 static bool is_callback_calling_kfunc(u32 btf_id); 544 545 static bool is_callback_calling_function(enum bpf_func_id func_id) 546 { 547 return func_id == BPF_FUNC_for_each_map_elem || 548 func_id == BPF_FUNC_timer_set_callback || 549 func_id == BPF_FUNC_find_vma || 550 func_id == BPF_FUNC_loop || 551 func_id == BPF_FUNC_user_ringbuf_drain; 552 } 553 554 static bool is_async_callback_calling_function(enum bpf_func_id func_id) 555 { 556 return func_id == BPF_FUNC_timer_set_callback; 557 } 558 559 static bool is_storage_get_function(enum bpf_func_id func_id) 560 { 561 return func_id == BPF_FUNC_sk_storage_get || 562 func_id == BPF_FUNC_inode_storage_get || 563 func_id == BPF_FUNC_task_storage_get || 564 func_id == BPF_FUNC_cgrp_storage_get; 565 } 566 567 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id, 568 const struct bpf_map *map) 569 { 570 int ref_obj_uses = 0; 571 572 if (is_ptr_cast_function(func_id)) 573 ref_obj_uses++; 574 if (is_acquire_function(func_id, map)) 575 ref_obj_uses++; 576 if (is_dynptr_ref_function(func_id)) 577 ref_obj_uses++; 578 579 return ref_obj_uses > 1; 580 } 581 582 static bool is_cmpxchg_insn(const struct bpf_insn *insn) 583 { 584 return BPF_CLASS(insn->code) == BPF_STX && 585 BPF_MODE(insn->code) == BPF_ATOMIC && 586 insn->imm == BPF_CMPXCHG; 587 } 588 589 /* string representation of 'enum bpf_reg_type' 590 * 591 * Note that reg_type_str() can not appear more than once in a single verbose() 592 * statement. 593 */ 594 static const char *reg_type_str(struct bpf_verifier_env *env, 595 enum bpf_reg_type type) 596 { 597 char postfix[16] = {0}, prefix[64] = {0}; 598 static const char * const str[] = { 599 [NOT_INIT] = "?", 600 [SCALAR_VALUE] = "scalar", 601 [PTR_TO_CTX] = "ctx", 602 [CONST_PTR_TO_MAP] = "map_ptr", 603 [PTR_TO_MAP_VALUE] = "map_value", 604 [PTR_TO_STACK] = "fp", 605 [PTR_TO_PACKET] = "pkt", 606 [PTR_TO_PACKET_META] = "pkt_meta", 607 [PTR_TO_PACKET_END] = "pkt_end", 608 [PTR_TO_FLOW_KEYS] = "flow_keys", 609 [PTR_TO_SOCKET] = "sock", 610 [PTR_TO_SOCK_COMMON] = "sock_common", 611 [PTR_TO_TCP_SOCK] = "tcp_sock", 612 [PTR_TO_TP_BUFFER] = "tp_buffer", 613 [PTR_TO_XDP_SOCK] = "xdp_sock", 614 [PTR_TO_BTF_ID] = "ptr_", 615 [PTR_TO_MEM] = "mem", 616 [PTR_TO_BUF] = "buf", 617 [PTR_TO_FUNC] = "func", 618 [PTR_TO_MAP_KEY] = "map_key", 619 [CONST_PTR_TO_DYNPTR] = "dynptr_ptr", 620 }; 621 622 if (type & PTR_MAYBE_NULL) { 623 if (base_type(type) == PTR_TO_BTF_ID) 624 strncpy(postfix, "or_null_", 16); 625 else 626 strncpy(postfix, "_or_null", 16); 627 } 628 629 snprintf(prefix, sizeof(prefix), "%s%s%s%s%s%s%s", 630 type & MEM_RDONLY ? "rdonly_" : "", 631 type & MEM_RINGBUF ? "ringbuf_" : "", 632 type & MEM_USER ? "user_" : "", 633 type & MEM_PERCPU ? "percpu_" : "", 634 type & MEM_RCU ? "rcu_" : "", 635 type & PTR_UNTRUSTED ? "untrusted_" : "", 636 type & PTR_TRUSTED ? "trusted_" : "" 637 ); 638 639 snprintf(env->tmp_str_buf, TMP_STR_BUF_LEN, "%s%s%s", 640 prefix, str[base_type(type)], postfix); 641 return env->tmp_str_buf; 642 } 643 644 static char slot_type_char[] = { 645 [STACK_INVALID] = '?', 646 [STACK_SPILL] = 'r', 647 [STACK_MISC] = 'm', 648 [STACK_ZERO] = '0', 649 [STACK_DYNPTR] = 'd', 650 [STACK_ITER] = 'i', 651 }; 652 653 static void print_liveness(struct bpf_verifier_env *env, 654 enum bpf_reg_liveness live) 655 { 656 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE)) 657 verbose(env, "_"); 658 if (live & REG_LIVE_READ) 659 verbose(env, "r"); 660 if (live & REG_LIVE_WRITTEN) 661 verbose(env, "w"); 662 if (live & REG_LIVE_DONE) 663 verbose(env, "D"); 664 } 665 666 static int __get_spi(s32 off) 667 { 668 return (-off - 1) / BPF_REG_SIZE; 669 } 670 671 static struct bpf_func_state *func(struct bpf_verifier_env *env, 672 const struct bpf_reg_state *reg) 673 { 674 struct bpf_verifier_state *cur = env->cur_state; 675 676 return cur->frame[reg->frameno]; 677 } 678 679 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots) 680 { 681 int allocated_slots = state->allocated_stack / BPF_REG_SIZE; 682 683 /* We need to check that slots between [spi - nr_slots + 1, spi] are 684 * within [0, allocated_stack). 685 * 686 * Please note that the spi grows downwards. For example, a dynptr 687 * takes the size of two stack slots; the first slot will be at 688 * spi and the second slot will be at spi - 1. 689 */ 690 return spi - nr_slots + 1 >= 0 && spi < allocated_slots; 691 } 692 693 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 694 const char *obj_kind, int nr_slots) 695 { 696 int off, spi; 697 698 if (!tnum_is_const(reg->var_off)) { 699 verbose(env, "%s has to be at a constant offset\n", obj_kind); 700 return -EINVAL; 701 } 702 703 off = reg->off + reg->var_off.value; 704 if (off % BPF_REG_SIZE) { 705 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off); 706 return -EINVAL; 707 } 708 709 spi = __get_spi(off); 710 if (spi + 1 < nr_slots) { 711 verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off); 712 return -EINVAL; 713 } 714 715 if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots)) 716 return -ERANGE; 717 return spi; 718 } 719 720 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 721 { 722 return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS); 723 } 724 725 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots) 726 { 727 return stack_slot_obj_get_spi(env, reg, "iter", nr_slots); 728 } 729 730 static const char *btf_type_name(const struct btf *btf, u32 id) 731 { 732 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off); 733 } 734 735 static const char *dynptr_type_str(enum bpf_dynptr_type type) 736 { 737 switch (type) { 738 case BPF_DYNPTR_TYPE_LOCAL: 739 return "local"; 740 case BPF_DYNPTR_TYPE_RINGBUF: 741 return "ringbuf"; 742 case BPF_DYNPTR_TYPE_SKB: 743 return "skb"; 744 case BPF_DYNPTR_TYPE_XDP: 745 return "xdp"; 746 case BPF_DYNPTR_TYPE_INVALID: 747 return "<invalid>"; 748 default: 749 WARN_ONCE(1, "unknown dynptr type %d\n", type); 750 return "<unknown>"; 751 } 752 } 753 754 static const char *iter_type_str(const struct btf *btf, u32 btf_id) 755 { 756 if (!btf || btf_id == 0) 757 return "<invalid>"; 758 759 /* we already validated that type is valid and has conforming name */ 760 return btf_type_name(btf, btf_id) + sizeof(ITER_PREFIX) - 1; 761 } 762 763 static const char *iter_state_str(enum bpf_iter_state state) 764 { 765 switch (state) { 766 case BPF_ITER_STATE_ACTIVE: 767 return "active"; 768 case BPF_ITER_STATE_DRAINED: 769 return "drained"; 770 case BPF_ITER_STATE_INVALID: 771 return "<invalid>"; 772 default: 773 WARN_ONCE(1, "unknown iter state %d\n", state); 774 return "<unknown>"; 775 } 776 } 777 778 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno) 779 { 780 env->scratched_regs |= 1U << regno; 781 } 782 783 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi) 784 { 785 env->scratched_stack_slots |= 1ULL << spi; 786 } 787 788 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno) 789 { 790 return (env->scratched_regs >> regno) & 1; 791 } 792 793 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno) 794 { 795 return (env->scratched_stack_slots >> regno) & 1; 796 } 797 798 static bool verifier_state_scratched(const struct bpf_verifier_env *env) 799 { 800 return env->scratched_regs || env->scratched_stack_slots; 801 } 802 803 static void mark_verifier_state_clean(struct bpf_verifier_env *env) 804 { 805 env->scratched_regs = 0U; 806 env->scratched_stack_slots = 0ULL; 807 } 808 809 /* Used for printing the entire verifier state. */ 810 static void mark_verifier_state_scratched(struct bpf_verifier_env *env) 811 { 812 env->scratched_regs = ~0U; 813 env->scratched_stack_slots = ~0ULL; 814 } 815 816 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type) 817 { 818 switch (arg_type & DYNPTR_TYPE_FLAG_MASK) { 819 case DYNPTR_TYPE_LOCAL: 820 return BPF_DYNPTR_TYPE_LOCAL; 821 case DYNPTR_TYPE_RINGBUF: 822 return BPF_DYNPTR_TYPE_RINGBUF; 823 case DYNPTR_TYPE_SKB: 824 return BPF_DYNPTR_TYPE_SKB; 825 case DYNPTR_TYPE_XDP: 826 return BPF_DYNPTR_TYPE_XDP; 827 default: 828 return BPF_DYNPTR_TYPE_INVALID; 829 } 830 } 831 832 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type) 833 { 834 switch (type) { 835 case BPF_DYNPTR_TYPE_LOCAL: 836 return DYNPTR_TYPE_LOCAL; 837 case BPF_DYNPTR_TYPE_RINGBUF: 838 return DYNPTR_TYPE_RINGBUF; 839 case BPF_DYNPTR_TYPE_SKB: 840 return DYNPTR_TYPE_SKB; 841 case BPF_DYNPTR_TYPE_XDP: 842 return DYNPTR_TYPE_XDP; 843 default: 844 return 0; 845 } 846 } 847 848 static bool dynptr_type_refcounted(enum bpf_dynptr_type type) 849 { 850 return type == BPF_DYNPTR_TYPE_RINGBUF; 851 } 852 853 static void __mark_dynptr_reg(struct bpf_reg_state *reg, 854 enum bpf_dynptr_type type, 855 bool first_slot, int dynptr_id); 856 857 static void __mark_reg_not_init(const struct bpf_verifier_env *env, 858 struct bpf_reg_state *reg); 859 860 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env, 861 struct bpf_reg_state *sreg1, 862 struct bpf_reg_state *sreg2, 863 enum bpf_dynptr_type type) 864 { 865 int id = ++env->id_gen; 866 867 __mark_dynptr_reg(sreg1, type, true, id); 868 __mark_dynptr_reg(sreg2, type, false, id); 869 } 870 871 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env, 872 struct bpf_reg_state *reg, 873 enum bpf_dynptr_type type) 874 { 875 __mark_dynptr_reg(reg, type, true, ++env->id_gen); 876 } 877 878 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env, 879 struct bpf_func_state *state, int spi); 880 881 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 882 enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id) 883 { 884 struct bpf_func_state *state = func(env, reg); 885 enum bpf_dynptr_type type; 886 int spi, i, err; 887 888 spi = dynptr_get_spi(env, reg); 889 if (spi < 0) 890 return spi; 891 892 /* We cannot assume both spi and spi - 1 belong to the same dynptr, 893 * hence we need to call destroy_if_dynptr_stack_slot twice for both, 894 * to ensure that for the following example: 895 * [d1][d1][d2][d2] 896 * spi 3 2 1 0 897 * So marking spi = 2 should lead to destruction of both d1 and d2. In 898 * case they do belong to same dynptr, second call won't see slot_type 899 * as STACK_DYNPTR and will simply skip destruction. 900 */ 901 err = destroy_if_dynptr_stack_slot(env, state, spi); 902 if (err) 903 return err; 904 err = destroy_if_dynptr_stack_slot(env, state, spi - 1); 905 if (err) 906 return err; 907 908 for (i = 0; i < BPF_REG_SIZE; i++) { 909 state->stack[spi].slot_type[i] = STACK_DYNPTR; 910 state->stack[spi - 1].slot_type[i] = STACK_DYNPTR; 911 } 912 913 type = arg_to_dynptr_type(arg_type); 914 if (type == BPF_DYNPTR_TYPE_INVALID) 915 return -EINVAL; 916 917 mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr, 918 &state->stack[spi - 1].spilled_ptr, type); 919 920 if (dynptr_type_refcounted(type)) { 921 /* The id is used to track proper releasing */ 922 int id; 923 924 if (clone_ref_obj_id) 925 id = clone_ref_obj_id; 926 else 927 id = acquire_reference_state(env, insn_idx); 928 929 if (id < 0) 930 return id; 931 932 state->stack[spi].spilled_ptr.ref_obj_id = id; 933 state->stack[spi - 1].spilled_ptr.ref_obj_id = id; 934 } 935 936 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; 937 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN; 938 939 return 0; 940 } 941 942 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi) 943 { 944 int i; 945 946 for (i = 0; i < BPF_REG_SIZE; i++) { 947 state->stack[spi].slot_type[i] = STACK_INVALID; 948 state->stack[spi - 1].slot_type[i] = STACK_INVALID; 949 } 950 951 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr); 952 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr); 953 954 /* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot? 955 * 956 * While we don't allow reading STACK_INVALID, it is still possible to 957 * do <8 byte writes marking some but not all slots as STACK_MISC. Then, 958 * helpers or insns can do partial read of that part without failing, 959 * but check_stack_range_initialized, check_stack_read_var_off, and 960 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of 961 * the slot conservatively. Hence we need to prevent those liveness 962 * marking walks. 963 * 964 * This was not a problem before because STACK_INVALID is only set by 965 * default (where the default reg state has its reg->parent as NULL), or 966 * in clean_live_states after REG_LIVE_DONE (at which point 967 * mark_reg_read won't walk reg->parent chain), but not randomly during 968 * verifier state exploration (like we did above). Hence, for our case 969 * parentage chain will still be live (i.e. reg->parent may be 970 * non-NULL), while earlier reg->parent was NULL, so we need 971 * REG_LIVE_WRITTEN to screen off read marker propagation when it is 972 * done later on reads or by mark_dynptr_read as well to unnecessary 973 * mark registers in verifier state. 974 */ 975 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; 976 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN; 977 } 978 979 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 980 { 981 struct bpf_func_state *state = func(env, reg); 982 int spi, ref_obj_id, i; 983 984 spi = dynptr_get_spi(env, reg); 985 if (spi < 0) 986 return spi; 987 988 if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) { 989 invalidate_dynptr(env, state, spi); 990 return 0; 991 } 992 993 ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id; 994 995 /* If the dynptr has a ref_obj_id, then we need to invalidate 996 * two things: 997 * 998 * 1) Any dynptrs with a matching ref_obj_id (clones) 999 * 2) Any slices derived from this dynptr. 1000 */ 1001 1002 /* Invalidate any slices associated with this dynptr */ 1003 WARN_ON_ONCE(release_reference(env, ref_obj_id)); 1004 1005 /* Invalidate any dynptr clones */ 1006 for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) { 1007 if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id) 1008 continue; 1009 1010 /* it should always be the case that if the ref obj id 1011 * matches then the stack slot also belongs to a 1012 * dynptr 1013 */ 1014 if (state->stack[i].slot_type[0] != STACK_DYNPTR) { 1015 verbose(env, "verifier internal error: misconfigured ref_obj_id\n"); 1016 return -EFAULT; 1017 } 1018 if (state->stack[i].spilled_ptr.dynptr.first_slot) 1019 invalidate_dynptr(env, state, i); 1020 } 1021 1022 return 0; 1023 } 1024 1025 static void __mark_reg_unknown(const struct bpf_verifier_env *env, 1026 struct bpf_reg_state *reg); 1027 1028 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg) 1029 { 1030 if (!env->allow_ptr_leaks) 1031 __mark_reg_not_init(env, reg); 1032 else 1033 __mark_reg_unknown(env, reg); 1034 } 1035 1036 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env, 1037 struct bpf_func_state *state, int spi) 1038 { 1039 struct bpf_func_state *fstate; 1040 struct bpf_reg_state *dreg; 1041 int i, dynptr_id; 1042 1043 /* We always ensure that STACK_DYNPTR is never set partially, 1044 * hence just checking for slot_type[0] is enough. This is 1045 * different for STACK_SPILL, where it may be only set for 1046 * 1 byte, so code has to use is_spilled_reg. 1047 */ 1048 if (state->stack[spi].slot_type[0] != STACK_DYNPTR) 1049 return 0; 1050 1051 /* Reposition spi to first slot */ 1052 if (!state->stack[spi].spilled_ptr.dynptr.first_slot) 1053 spi = spi + 1; 1054 1055 if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) { 1056 verbose(env, "cannot overwrite referenced dynptr\n"); 1057 return -EINVAL; 1058 } 1059 1060 mark_stack_slot_scratched(env, spi); 1061 mark_stack_slot_scratched(env, spi - 1); 1062 1063 /* Writing partially to one dynptr stack slot destroys both. */ 1064 for (i = 0; i < BPF_REG_SIZE; i++) { 1065 state->stack[spi].slot_type[i] = STACK_INVALID; 1066 state->stack[spi - 1].slot_type[i] = STACK_INVALID; 1067 } 1068 1069 dynptr_id = state->stack[spi].spilled_ptr.id; 1070 /* Invalidate any slices associated with this dynptr */ 1071 bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({ 1072 /* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */ 1073 if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM) 1074 continue; 1075 if (dreg->dynptr_id == dynptr_id) 1076 mark_reg_invalid(env, dreg); 1077 })); 1078 1079 /* Do not release reference state, we are destroying dynptr on stack, 1080 * not using some helper to release it. Just reset register. 1081 */ 1082 __mark_reg_not_init(env, &state->stack[spi].spilled_ptr); 1083 __mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr); 1084 1085 /* Same reason as unmark_stack_slots_dynptr above */ 1086 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; 1087 state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN; 1088 1089 return 0; 1090 } 1091 1092 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 1093 { 1094 int spi; 1095 1096 if (reg->type == CONST_PTR_TO_DYNPTR) 1097 return false; 1098 1099 spi = dynptr_get_spi(env, reg); 1100 1101 /* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an 1102 * error because this just means the stack state hasn't been updated yet. 1103 * We will do check_mem_access to check and update stack bounds later. 1104 */ 1105 if (spi < 0 && spi != -ERANGE) 1106 return false; 1107 1108 /* We don't need to check if the stack slots are marked by previous 1109 * dynptr initializations because we allow overwriting existing unreferenced 1110 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls 1111 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are 1112 * touching are completely destructed before we reinitialize them for a new 1113 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early 1114 * instead of delaying it until the end where the user will get "Unreleased 1115 * reference" error. 1116 */ 1117 return true; 1118 } 1119 1120 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 1121 { 1122 struct bpf_func_state *state = func(env, reg); 1123 int i, spi; 1124 1125 /* This already represents first slot of initialized bpf_dynptr. 1126 * 1127 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to 1128 * check_func_arg_reg_off's logic, so we don't need to check its 1129 * offset and alignment. 1130 */ 1131 if (reg->type == CONST_PTR_TO_DYNPTR) 1132 return true; 1133 1134 spi = dynptr_get_spi(env, reg); 1135 if (spi < 0) 1136 return false; 1137 if (!state->stack[spi].spilled_ptr.dynptr.first_slot) 1138 return false; 1139 1140 for (i = 0; i < BPF_REG_SIZE; i++) { 1141 if (state->stack[spi].slot_type[i] != STACK_DYNPTR || 1142 state->stack[spi - 1].slot_type[i] != STACK_DYNPTR) 1143 return false; 1144 } 1145 1146 return true; 1147 } 1148 1149 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 1150 enum bpf_arg_type arg_type) 1151 { 1152 struct bpf_func_state *state = func(env, reg); 1153 enum bpf_dynptr_type dynptr_type; 1154 int spi; 1155 1156 /* ARG_PTR_TO_DYNPTR takes any type of dynptr */ 1157 if (arg_type == ARG_PTR_TO_DYNPTR) 1158 return true; 1159 1160 dynptr_type = arg_to_dynptr_type(arg_type); 1161 if (reg->type == CONST_PTR_TO_DYNPTR) { 1162 return reg->dynptr.type == dynptr_type; 1163 } else { 1164 spi = dynptr_get_spi(env, reg); 1165 if (spi < 0) 1166 return false; 1167 return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type; 1168 } 1169 } 1170 1171 static void __mark_reg_known_zero(struct bpf_reg_state *reg); 1172 1173 static int mark_stack_slots_iter(struct bpf_verifier_env *env, 1174 struct bpf_reg_state *reg, int insn_idx, 1175 struct btf *btf, u32 btf_id, int nr_slots) 1176 { 1177 struct bpf_func_state *state = func(env, reg); 1178 int spi, i, j, id; 1179 1180 spi = iter_get_spi(env, reg, nr_slots); 1181 if (spi < 0) 1182 return spi; 1183 1184 id = acquire_reference_state(env, insn_idx); 1185 if (id < 0) 1186 return id; 1187 1188 for (i = 0; i < nr_slots; i++) { 1189 struct bpf_stack_state *slot = &state->stack[spi - i]; 1190 struct bpf_reg_state *st = &slot->spilled_ptr; 1191 1192 __mark_reg_known_zero(st); 1193 st->type = PTR_TO_STACK; /* we don't have dedicated reg type */ 1194 st->live |= REG_LIVE_WRITTEN; 1195 st->ref_obj_id = i == 0 ? id : 0; 1196 st->iter.btf = btf; 1197 st->iter.btf_id = btf_id; 1198 st->iter.state = BPF_ITER_STATE_ACTIVE; 1199 st->iter.depth = 0; 1200 1201 for (j = 0; j < BPF_REG_SIZE; j++) 1202 slot->slot_type[j] = STACK_ITER; 1203 1204 mark_stack_slot_scratched(env, spi - i); 1205 } 1206 1207 return 0; 1208 } 1209 1210 static int unmark_stack_slots_iter(struct bpf_verifier_env *env, 1211 struct bpf_reg_state *reg, int nr_slots) 1212 { 1213 struct bpf_func_state *state = func(env, reg); 1214 int spi, i, j; 1215 1216 spi = iter_get_spi(env, reg, nr_slots); 1217 if (spi < 0) 1218 return spi; 1219 1220 for (i = 0; i < nr_slots; i++) { 1221 struct bpf_stack_state *slot = &state->stack[spi - i]; 1222 struct bpf_reg_state *st = &slot->spilled_ptr; 1223 1224 if (i == 0) 1225 WARN_ON_ONCE(release_reference(env, st->ref_obj_id)); 1226 1227 __mark_reg_not_init(env, st); 1228 1229 /* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */ 1230 st->live |= REG_LIVE_WRITTEN; 1231 1232 for (j = 0; j < BPF_REG_SIZE; j++) 1233 slot->slot_type[j] = STACK_INVALID; 1234 1235 mark_stack_slot_scratched(env, spi - i); 1236 } 1237 1238 return 0; 1239 } 1240 1241 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env, 1242 struct bpf_reg_state *reg, int nr_slots) 1243 { 1244 struct bpf_func_state *state = func(env, reg); 1245 int spi, i, j; 1246 1247 /* For -ERANGE (i.e. spi not falling into allocated stack slots), we 1248 * will do check_mem_access to check and update stack bounds later, so 1249 * return true for that case. 1250 */ 1251 spi = iter_get_spi(env, reg, nr_slots); 1252 if (spi == -ERANGE) 1253 return true; 1254 if (spi < 0) 1255 return false; 1256 1257 for (i = 0; i < nr_slots; i++) { 1258 struct bpf_stack_state *slot = &state->stack[spi - i]; 1259 1260 for (j = 0; j < BPF_REG_SIZE; j++) 1261 if (slot->slot_type[j] == STACK_ITER) 1262 return false; 1263 } 1264 1265 return true; 1266 } 1267 1268 static bool is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 1269 struct btf *btf, u32 btf_id, int nr_slots) 1270 { 1271 struct bpf_func_state *state = func(env, reg); 1272 int spi, i, j; 1273 1274 spi = iter_get_spi(env, reg, nr_slots); 1275 if (spi < 0) 1276 return false; 1277 1278 for (i = 0; i < nr_slots; i++) { 1279 struct bpf_stack_state *slot = &state->stack[spi - i]; 1280 struct bpf_reg_state *st = &slot->spilled_ptr; 1281 1282 /* only main (first) slot has ref_obj_id set */ 1283 if (i == 0 && !st->ref_obj_id) 1284 return false; 1285 if (i != 0 && st->ref_obj_id) 1286 return false; 1287 if (st->iter.btf != btf || st->iter.btf_id != btf_id) 1288 return false; 1289 1290 for (j = 0; j < BPF_REG_SIZE; j++) 1291 if (slot->slot_type[j] != STACK_ITER) 1292 return false; 1293 } 1294 1295 return true; 1296 } 1297 1298 /* Check if given stack slot is "special": 1299 * - spilled register state (STACK_SPILL); 1300 * - dynptr state (STACK_DYNPTR); 1301 * - iter state (STACK_ITER). 1302 */ 1303 static bool is_stack_slot_special(const struct bpf_stack_state *stack) 1304 { 1305 enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1]; 1306 1307 switch (type) { 1308 case STACK_SPILL: 1309 case STACK_DYNPTR: 1310 case STACK_ITER: 1311 return true; 1312 case STACK_INVALID: 1313 case STACK_MISC: 1314 case STACK_ZERO: 1315 return false; 1316 default: 1317 WARN_ONCE(1, "unknown stack slot type %d\n", type); 1318 return true; 1319 } 1320 } 1321 1322 /* The reg state of a pointer or a bounded scalar was saved when 1323 * it was spilled to the stack. 1324 */ 1325 static bool is_spilled_reg(const struct bpf_stack_state *stack) 1326 { 1327 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL; 1328 } 1329 1330 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack) 1331 { 1332 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL && 1333 stack->spilled_ptr.type == SCALAR_VALUE; 1334 } 1335 1336 static void scrub_spilled_slot(u8 *stype) 1337 { 1338 if (*stype != STACK_INVALID) 1339 *stype = STACK_MISC; 1340 } 1341 1342 static void print_verifier_state(struct bpf_verifier_env *env, 1343 const struct bpf_func_state *state, 1344 bool print_all) 1345 { 1346 const struct bpf_reg_state *reg; 1347 enum bpf_reg_type t; 1348 int i; 1349 1350 if (state->frameno) 1351 verbose(env, " frame%d:", state->frameno); 1352 for (i = 0; i < MAX_BPF_REG; i++) { 1353 reg = &state->regs[i]; 1354 t = reg->type; 1355 if (t == NOT_INIT) 1356 continue; 1357 if (!print_all && !reg_scratched(env, i)) 1358 continue; 1359 verbose(env, " R%d", i); 1360 print_liveness(env, reg->live); 1361 verbose(env, "="); 1362 if (t == SCALAR_VALUE && reg->precise) 1363 verbose(env, "P"); 1364 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) && 1365 tnum_is_const(reg->var_off)) { 1366 /* reg->off should be 0 for SCALAR_VALUE */ 1367 verbose(env, "%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t)); 1368 verbose(env, "%lld", reg->var_off.value + reg->off); 1369 } else { 1370 const char *sep = ""; 1371 1372 verbose(env, "%s", reg_type_str(env, t)); 1373 if (base_type(t) == PTR_TO_BTF_ID) 1374 verbose(env, "%s", btf_type_name(reg->btf, reg->btf_id)); 1375 verbose(env, "("); 1376 /* 1377 * _a stands for append, was shortened to avoid multiline statements below. 1378 * This macro is used to output a comma separated list of attributes. 1379 */ 1380 #define verbose_a(fmt, ...) ({ verbose(env, "%s" fmt, sep, __VA_ARGS__); sep = ","; }) 1381 1382 if (reg->id) 1383 verbose_a("id=%d", reg->id); 1384 if (reg->ref_obj_id) 1385 verbose_a("ref_obj_id=%d", reg->ref_obj_id); 1386 if (type_is_non_owning_ref(reg->type)) 1387 verbose_a("%s", "non_own_ref"); 1388 if (t != SCALAR_VALUE) 1389 verbose_a("off=%d", reg->off); 1390 if (type_is_pkt_pointer(t)) 1391 verbose_a("r=%d", reg->range); 1392 else if (base_type(t) == CONST_PTR_TO_MAP || 1393 base_type(t) == PTR_TO_MAP_KEY || 1394 base_type(t) == PTR_TO_MAP_VALUE) 1395 verbose_a("ks=%d,vs=%d", 1396 reg->map_ptr->key_size, 1397 reg->map_ptr->value_size); 1398 if (tnum_is_const(reg->var_off)) { 1399 /* Typically an immediate SCALAR_VALUE, but 1400 * could be a pointer whose offset is too big 1401 * for reg->off 1402 */ 1403 verbose_a("imm=%llx", reg->var_off.value); 1404 } else { 1405 if (reg->smin_value != reg->umin_value && 1406 reg->smin_value != S64_MIN) 1407 verbose_a("smin=%lld", (long long)reg->smin_value); 1408 if (reg->smax_value != reg->umax_value && 1409 reg->smax_value != S64_MAX) 1410 verbose_a("smax=%lld", (long long)reg->smax_value); 1411 if (reg->umin_value != 0) 1412 verbose_a("umin=%llu", (unsigned long long)reg->umin_value); 1413 if (reg->umax_value != U64_MAX) 1414 verbose_a("umax=%llu", (unsigned long long)reg->umax_value); 1415 if (!tnum_is_unknown(reg->var_off)) { 1416 char tn_buf[48]; 1417 1418 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 1419 verbose_a("var_off=%s", tn_buf); 1420 } 1421 if (reg->s32_min_value != reg->smin_value && 1422 reg->s32_min_value != S32_MIN) 1423 verbose_a("s32_min=%d", (int)(reg->s32_min_value)); 1424 if (reg->s32_max_value != reg->smax_value && 1425 reg->s32_max_value != S32_MAX) 1426 verbose_a("s32_max=%d", (int)(reg->s32_max_value)); 1427 if (reg->u32_min_value != reg->umin_value && 1428 reg->u32_min_value != U32_MIN) 1429 verbose_a("u32_min=%d", (int)(reg->u32_min_value)); 1430 if (reg->u32_max_value != reg->umax_value && 1431 reg->u32_max_value != U32_MAX) 1432 verbose_a("u32_max=%d", (int)(reg->u32_max_value)); 1433 } 1434 #undef verbose_a 1435 1436 verbose(env, ")"); 1437 } 1438 } 1439 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { 1440 char types_buf[BPF_REG_SIZE + 1]; 1441 bool valid = false; 1442 int j; 1443 1444 for (j = 0; j < BPF_REG_SIZE; j++) { 1445 if (state->stack[i].slot_type[j] != STACK_INVALID) 1446 valid = true; 1447 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]]; 1448 } 1449 types_buf[BPF_REG_SIZE] = 0; 1450 if (!valid) 1451 continue; 1452 if (!print_all && !stack_slot_scratched(env, i)) 1453 continue; 1454 switch (state->stack[i].slot_type[BPF_REG_SIZE - 1]) { 1455 case STACK_SPILL: 1456 reg = &state->stack[i].spilled_ptr; 1457 t = reg->type; 1458 1459 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE); 1460 print_liveness(env, reg->live); 1461 verbose(env, "=%s", t == SCALAR_VALUE ? "" : reg_type_str(env, t)); 1462 if (t == SCALAR_VALUE && reg->precise) 1463 verbose(env, "P"); 1464 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off)) 1465 verbose(env, "%lld", reg->var_off.value + reg->off); 1466 break; 1467 case STACK_DYNPTR: 1468 i += BPF_DYNPTR_NR_SLOTS - 1; 1469 reg = &state->stack[i].spilled_ptr; 1470 1471 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE); 1472 print_liveness(env, reg->live); 1473 verbose(env, "=dynptr_%s", dynptr_type_str(reg->dynptr.type)); 1474 if (reg->ref_obj_id) 1475 verbose(env, "(ref_id=%d)", reg->ref_obj_id); 1476 break; 1477 case STACK_ITER: 1478 /* only main slot has ref_obj_id set; skip others */ 1479 reg = &state->stack[i].spilled_ptr; 1480 if (!reg->ref_obj_id) 1481 continue; 1482 1483 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE); 1484 print_liveness(env, reg->live); 1485 verbose(env, "=iter_%s(ref_id=%d,state=%s,depth=%u)", 1486 iter_type_str(reg->iter.btf, reg->iter.btf_id), 1487 reg->ref_obj_id, iter_state_str(reg->iter.state), 1488 reg->iter.depth); 1489 break; 1490 case STACK_MISC: 1491 case STACK_ZERO: 1492 default: 1493 reg = &state->stack[i].spilled_ptr; 1494 1495 for (j = 0; j < BPF_REG_SIZE; j++) 1496 types_buf[j] = slot_type_char[state->stack[i].slot_type[j]]; 1497 types_buf[BPF_REG_SIZE] = 0; 1498 1499 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE); 1500 print_liveness(env, reg->live); 1501 verbose(env, "=%s", types_buf); 1502 break; 1503 } 1504 } 1505 if (state->acquired_refs && state->refs[0].id) { 1506 verbose(env, " refs=%d", state->refs[0].id); 1507 for (i = 1; i < state->acquired_refs; i++) 1508 if (state->refs[i].id) 1509 verbose(env, ",%d", state->refs[i].id); 1510 } 1511 if (state->in_callback_fn) 1512 verbose(env, " cb"); 1513 if (state->in_async_callback_fn) 1514 verbose(env, " async_cb"); 1515 verbose(env, "\n"); 1516 mark_verifier_state_clean(env); 1517 } 1518 1519 static inline u32 vlog_alignment(u32 pos) 1520 { 1521 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT), 1522 BPF_LOG_MIN_ALIGNMENT) - pos - 1; 1523 } 1524 1525 static void print_insn_state(struct bpf_verifier_env *env, 1526 const struct bpf_func_state *state) 1527 { 1528 if (env->prev_log_pos && env->prev_log_pos == env->log.end_pos) { 1529 /* remove new line character */ 1530 bpf_vlog_reset(&env->log, env->prev_log_pos - 1); 1531 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_pos), ' '); 1532 } else { 1533 verbose(env, "%d:", env->insn_idx); 1534 } 1535 print_verifier_state(env, state, false); 1536 } 1537 1538 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too 1539 * small to hold src. This is different from krealloc since we don't want to preserve 1540 * the contents of dst. 1541 * 1542 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could 1543 * not be allocated. 1544 */ 1545 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags) 1546 { 1547 size_t alloc_bytes; 1548 void *orig = dst; 1549 size_t bytes; 1550 1551 if (ZERO_OR_NULL_PTR(src)) 1552 goto out; 1553 1554 if (unlikely(check_mul_overflow(n, size, &bytes))) 1555 return NULL; 1556 1557 alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes)); 1558 dst = krealloc(orig, alloc_bytes, flags); 1559 if (!dst) { 1560 kfree(orig); 1561 return NULL; 1562 } 1563 1564 memcpy(dst, src, bytes); 1565 out: 1566 return dst ? dst : ZERO_SIZE_PTR; 1567 } 1568 1569 /* resize an array from old_n items to new_n items. the array is reallocated if it's too 1570 * small to hold new_n items. new items are zeroed out if the array grows. 1571 * 1572 * Contrary to krealloc_array, does not free arr if new_n is zero. 1573 */ 1574 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size) 1575 { 1576 size_t alloc_size; 1577 void *new_arr; 1578 1579 if (!new_n || old_n == new_n) 1580 goto out; 1581 1582 alloc_size = kmalloc_size_roundup(size_mul(new_n, size)); 1583 new_arr = krealloc(arr, alloc_size, GFP_KERNEL); 1584 if (!new_arr) { 1585 kfree(arr); 1586 return NULL; 1587 } 1588 arr = new_arr; 1589 1590 if (new_n > old_n) 1591 memset(arr + old_n * size, 0, (new_n - old_n) * size); 1592 1593 out: 1594 return arr ? arr : ZERO_SIZE_PTR; 1595 } 1596 1597 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src) 1598 { 1599 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs, 1600 sizeof(struct bpf_reference_state), GFP_KERNEL); 1601 if (!dst->refs) 1602 return -ENOMEM; 1603 1604 dst->acquired_refs = src->acquired_refs; 1605 return 0; 1606 } 1607 1608 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src) 1609 { 1610 size_t n = src->allocated_stack / BPF_REG_SIZE; 1611 1612 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state), 1613 GFP_KERNEL); 1614 if (!dst->stack) 1615 return -ENOMEM; 1616 1617 dst->allocated_stack = src->allocated_stack; 1618 return 0; 1619 } 1620 1621 static int resize_reference_state(struct bpf_func_state *state, size_t n) 1622 { 1623 state->refs = realloc_array(state->refs, state->acquired_refs, n, 1624 sizeof(struct bpf_reference_state)); 1625 if (!state->refs) 1626 return -ENOMEM; 1627 1628 state->acquired_refs = n; 1629 return 0; 1630 } 1631 1632 static int grow_stack_state(struct bpf_func_state *state, int size) 1633 { 1634 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE; 1635 1636 if (old_n >= n) 1637 return 0; 1638 1639 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state)); 1640 if (!state->stack) 1641 return -ENOMEM; 1642 1643 state->allocated_stack = size; 1644 return 0; 1645 } 1646 1647 /* Acquire a pointer id from the env and update the state->refs to include 1648 * this new pointer reference. 1649 * On success, returns a valid pointer id to associate with the register 1650 * On failure, returns a negative errno. 1651 */ 1652 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx) 1653 { 1654 struct bpf_func_state *state = cur_func(env); 1655 int new_ofs = state->acquired_refs; 1656 int id, err; 1657 1658 err = resize_reference_state(state, state->acquired_refs + 1); 1659 if (err) 1660 return err; 1661 id = ++env->id_gen; 1662 state->refs[new_ofs].id = id; 1663 state->refs[new_ofs].insn_idx = insn_idx; 1664 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0; 1665 1666 return id; 1667 } 1668 1669 /* release function corresponding to acquire_reference_state(). Idempotent. */ 1670 static int release_reference_state(struct bpf_func_state *state, int ptr_id) 1671 { 1672 int i, last_idx; 1673 1674 last_idx = state->acquired_refs - 1; 1675 for (i = 0; i < state->acquired_refs; i++) { 1676 if (state->refs[i].id == ptr_id) { 1677 /* Cannot release caller references in callbacks */ 1678 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno) 1679 return -EINVAL; 1680 if (last_idx && i != last_idx) 1681 memcpy(&state->refs[i], &state->refs[last_idx], 1682 sizeof(*state->refs)); 1683 memset(&state->refs[last_idx], 0, sizeof(*state->refs)); 1684 state->acquired_refs--; 1685 return 0; 1686 } 1687 } 1688 return -EINVAL; 1689 } 1690 1691 static void free_func_state(struct bpf_func_state *state) 1692 { 1693 if (!state) 1694 return; 1695 kfree(state->refs); 1696 kfree(state->stack); 1697 kfree(state); 1698 } 1699 1700 static void clear_jmp_history(struct bpf_verifier_state *state) 1701 { 1702 kfree(state->jmp_history); 1703 state->jmp_history = NULL; 1704 state->jmp_history_cnt = 0; 1705 } 1706 1707 static void free_verifier_state(struct bpf_verifier_state *state, 1708 bool free_self) 1709 { 1710 int i; 1711 1712 for (i = 0; i <= state->curframe; i++) { 1713 free_func_state(state->frame[i]); 1714 state->frame[i] = NULL; 1715 } 1716 clear_jmp_history(state); 1717 if (free_self) 1718 kfree(state); 1719 } 1720 1721 /* copy verifier state from src to dst growing dst stack space 1722 * when necessary to accommodate larger src stack 1723 */ 1724 static int copy_func_state(struct bpf_func_state *dst, 1725 const struct bpf_func_state *src) 1726 { 1727 int err; 1728 1729 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs)); 1730 err = copy_reference_state(dst, src); 1731 if (err) 1732 return err; 1733 return copy_stack_state(dst, src); 1734 } 1735 1736 static int copy_verifier_state(struct bpf_verifier_state *dst_state, 1737 const struct bpf_verifier_state *src) 1738 { 1739 struct bpf_func_state *dst; 1740 int i, err; 1741 1742 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history, 1743 src->jmp_history_cnt, sizeof(struct bpf_idx_pair), 1744 GFP_USER); 1745 if (!dst_state->jmp_history) 1746 return -ENOMEM; 1747 dst_state->jmp_history_cnt = src->jmp_history_cnt; 1748 1749 /* if dst has more stack frames then src frame, free them */ 1750 for (i = src->curframe + 1; i <= dst_state->curframe; i++) { 1751 free_func_state(dst_state->frame[i]); 1752 dst_state->frame[i] = NULL; 1753 } 1754 dst_state->speculative = src->speculative; 1755 dst_state->active_rcu_lock = src->active_rcu_lock; 1756 dst_state->curframe = src->curframe; 1757 dst_state->active_lock.ptr = src->active_lock.ptr; 1758 dst_state->active_lock.id = src->active_lock.id; 1759 dst_state->branches = src->branches; 1760 dst_state->parent = src->parent; 1761 dst_state->first_insn_idx = src->first_insn_idx; 1762 dst_state->last_insn_idx = src->last_insn_idx; 1763 for (i = 0; i <= src->curframe; i++) { 1764 dst = dst_state->frame[i]; 1765 if (!dst) { 1766 dst = kzalloc(sizeof(*dst), GFP_KERNEL); 1767 if (!dst) 1768 return -ENOMEM; 1769 dst_state->frame[i] = dst; 1770 } 1771 err = copy_func_state(dst, src->frame[i]); 1772 if (err) 1773 return err; 1774 } 1775 return 0; 1776 } 1777 1778 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st) 1779 { 1780 while (st) { 1781 u32 br = --st->branches; 1782 1783 /* WARN_ON(br > 1) technically makes sense here, 1784 * but see comment in push_stack(), hence: 1785 */ 1786 WARN_ONCE((int)br < 0, 1787 "BUG update_branch_counts:branches_to_explore=%d\n", 1788 br); 1789 if (br) 1790 break; 1791 st = st->parent; 1792 } 1793 } 1794 1795 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx, 1796 int *insn_idx, bool pop_log) 1797 { 1798 struct bpf_verifier_state *cur = env->cur_state; 1799 struct bpf_verifier_stack_elem *elem, *head = env->head; 1800 int err; 1801 1802 if (env->head == NULL) 1803 return -ENOENT; 1804 1805 if (cur) { 1806 err = copy_verifier_state(cur, &head->st); 1807 if (err) 1808 return err; 1809 } 1810 if (pop_log) 1811 bpf_vlog_reset(&env->log, head->log_pos); 1812 if (insn_idx) 1813 *insn_idx = head->insn_idx; 1814 if (prev_insn_idx) 1815 *prev_insn_idx = head->prev_insn_idx; 1816 elem = head->next; 1817 free_verifier_state(&head->st, false); 1818 kfree(head); 1819 env->head = elem; 1820 env->stack_size--; 1821 return 0; 1822 } 1823 1824 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env, 1825 int insn_idx, int prev_insn_idx, 1826 bool speculative) 1827 { 1828 struct bpf_verifier_state *cur = env->cur_state; 1829 struct bpf_verifier_stack_elem *elem; 1830 int err; 1831 1832 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL); 1833 if (!elem) 1834 goto err; 1835 1836 elem->insn_idx = insn_idx; 1837 elem->prev_insn_idx = prev_insn_idx; 1838 elem->next = env->head; 1839 elem->log_pos = env->log.end_pos; 1840 env->head = elem; 1841 env->stack_size++; 1842 err = copy_verifier_state(&elem->st, cur); 1843 if (err) 1844 goto err; 1845 elem->st.speculative |= speculative; 1846 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) { 1847 verbose(env, "The sequence of %d jumps is too complex.\n", 1848 env->stack_size); 1849 goto err; 1850 } 1851 if (elem->st.parent) { 1852 ++elem->st.parent->branches; 1853 /* WARN_ON(branches > 2) technically makes sense here, 1854 * but 1855 * 1. speculative states will bump 'branches' for non-branch 1856 * instructions 1857 * 2. is_state_visited() heuristics may decide not to create 1858 * a new state for a sequence of branches and all such current 1859 * and cloned states will be pointing to a single parent state 1860 * which might have large 'branches' count. 1861 */ 1862 } 1863 return &elem->st; 1864 err: 1865 free_verifier_state(env->cur_state, true); 1866 env->cur_state = NULL; 1867 /* pop all elements and return */ 1868 while (!pop_stack(env, NULL, NULL, false)); 1869 return NULL; 1870 } 1871 1872 #define CALLER_SAVED_REGS 6 1873 static const int caller_saved[CALLER_SAVED_REGS] = { 1874 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5 1875 }; 1876 1877 /* This helper doesn't clear reg->id */ 1878 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm) 1879 { 1880 reg->var_off = tnum_const(imm); 1881 reg->smin_value = (s64)imm; 1882 reg->smax_value = (s64)imm; 1883 reg->umin_value = imm; 1884 reg->umax_value = imm; 1885 1886 reg->s32_min_value = (s32)imm; 1887 reg->s32_max_value = (s32)imm; 1888 reg->u32_min_value = (u32)imm; 1889 reg->u32_max_value = (u32)imm; 1890 } 1891 1892 /* Mark the unknown part of a register (variable offset or scalar value) as 1893 * known to have the value @imm. 1894 */ 1895 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm) 1896 { 1897 /* Clear off and union(map_ptr, range) */ 1898 memset(((u8 *)reg) + sizeof(reg->type), 0, 1899 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type)); 1900 reg->id = 0; 1901 reg->ref_obj_id = 0; 1902 ___mark_reg_known(reg, imm); 1903 } 1904 1905 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm) 1906 { 1907 reg->var_off = tnum_const_subreg(reg->var_off, imm); 1908 reg->s32_min_value = (s32)imm; 1909 reg->s32_max_value = (s32)imm; 1910 reg->u32_min_value = (u32)imm; 1911 reg->u32_max_value = (u32)imm; 1912 } 1913 1914 /* Mark the 'variable offset' part of a register as zero. This should be 1915 * used only on registers holding a pointer type. 1916 */ 1917 static void __mark_reg_known_zero(struct bpf_reg_state *reg) 1918 { 1919 __mark_reg_known(reg, 0); 1920 } 1921 1922 static void __mark_reg_const_zero(struct bpf_reg_state *reg) 1923 { 1924 __mark_reg_known(reg, 0); 1925 reg->type = SCALAR_VALUE; 1926 } 1927 1928 static void mark_reg_known_zero(struct bpf_verifier_env *env, 1929 struct bpf_reg_state *regs, u32 regno) 1930 { 1931 if (WARN_ON(regno >= MAX_BPF_REG)) { 1932 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno); 1933 /* Something bad happened, let's kill all regs */ 1934 for (regno = 0; regno < MAX_BPF_REG; regno++) 1935 __mark_reg_not_init(env, regs + regno); 1936 return; 1937 } 1938 __mark_reg_known_zero(regs + regno); 1939 } 1940 1941 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type, 1942 bool first_slot, int dynptr_id) 1943 { 1944 /* reg->type has no meaning for STACK_DYNPTR, but when we set reg for 1945 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply 1946 * set it unconditionally as it is ignored for STACK_DYNPTR anyway. 1947 */ 1948 __mark_reg_known_zero(reg); 1949 reg->type = CONST_PTR_TO_DYNPTR; 1950 /* Give each dynptr a unique id to uniquely associate slices to it. */ 1951 reg->id = dynptr_id; 1952 reg->dynptr.type = type; 1953 reg->dynptr.first_slot = first_slot; 1954 } 1955 1956 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg) 1957 { 1958 if (base_type(reg->type) == PTR_TO_MAP_VALUE) { 1959 const struct bpf_map *map = reg->map_ptr; 1960 1961 if (map->inner_map_meta) { 1962 reg->type = CONST_PTR_TO_MAP; 1963 reg->map_ptr = map->inner_map_meta; 1964 /* transfer reg's id which is unique for every map_lookup_elem 1965 * as UID of the inner map. 1966 */ 1967 if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER)) 1968 reg->map_uid = reg->id; 1969 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) { 1970 reg->type = PTR_TO_XDP_SOCK; 1971 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP || 1972 map->map_type == BPF_MAP_TYPE_SOCKHASH) { 1973 reg->type = PTR_TO_SOCKET; 1974 } else { 1975 reg->type = PTR_TO_MAP_VALUE; 1976 } 1977 return; 1978 } 1979 1980 reg->type &= ~PTR_MAYBE_NULL; 1981 } 1982 1983 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno, 1984 struct btf_field_graph_root *ds_head) 1985 { 1986 __mark_reg_known_zero(®s[regno]); 1987 regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC; 1988 regs[regno].btf = ds_head->btf; 1989 regs[regno].btf_id = ds_head->value_btf_id; 1990 regs[regno].off = ds_head->node_offset; 1991 } 1992 1993 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg) 1994 { 1995 return type_is_pkt_pointer(reg->type); 1996 } 1997 1998 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg) 1999 { 2000 return reg_is_pkt_pointer(reg) || 2001 reg->type == PTR_TO_PACKET_END; 2002 } 2003 2004 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg) 2005 { 2006 return base_type(reg->type) == PTR_TO_MEM && 2007 (reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP); 2008 } 2009 2010 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */ 2011 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg, 2012 enum bpf_reg_type which) 2013 { 2014 /* The register can already have a range from prior markings. 2015 * This is fine as long as it hasn't been advanced from its 2016 * origin. 2017 */ 2018 return reg->type == which && 2019 reg->id == 0 && 2020 reg->off == 0 && 2021 tnum_equals_const(reg->var_off, 0); 2022 } 2023 2024 /* Reset the min/max bounds of a register */ 2025 static void __mark_reg_unbounded(struct bpf_reg_state *reg) 2026 { 2027 reg->smin_value = S64_MIN; 2028 reg->smax_value = S64_MAX; 2029 reg->umin_value = 0; 2030 reg->umax_value = U64_MAX; 2031 2032 reg->s32_min_value = S32_MIN; 2033 reg->s32_max_value = S32_MAX; 2034 reg->u32_min_value = 0; 2035 reg->u32_max_value = U32_MAX; 2036 } 2037 2038 static void __mark_reg64_unbounded(struct bpf_reg_state *reg) 2039 { 2040 reg->smin_value = S64_MIN; 2041 reg->smax_value = S64_MAX; 2042 reg->umin_value = 0; 2043 reg->umax_value = U64_MAX; 2044 } 2045 2046 static void __mark_reg32_unbounded(struct bpf_reg_state *reg) 2047 { 2048 reg->s32_min_value = S32_MIN; 2049 reg->s32_max_value = S32_MAX; 2050 reg->u32_min_value = 0; 2051 reg->u32_max_value = U32_MAX; 2052 } 2053 2054 static void __update_reg32_bounds(struct bpf_reg_state *reg) 2055 { 2056 struct tnum var32_off = tnum_subreg(reg->var_off); 2057 2058 /* min signed is max(sign bit) | min(other bits) */ 2059 reg->s32_min_value = max_t(s32, reg->s32_min_value, 2060 var32_off.value | (var32_off.mask & S32_MIN)); 2061 /* max signed is min(sign bit) | max(other bits) */ 2062 reg->s32_max_value = min_t(s32, reg->s32_max_value, 2063 var32_off.value | (var32_off.mask & S32_MAX)); 2064 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value); 2065 reg->u32_max_value = min(reg->u32_max_value, 2066 (u32)(var32_off.value | var32_off.mask)); 2067 } 2068 2069 static void __update_reg64_bounds(struct bpf_reg_state *reg) 2070 { 2071 /* min signed is max(sign bit) | min(other bits) */ 2072 reg->smin_value = max_t(s64, reg->smin_value, 2073 reg->var_off.value | (reg->var_off.mask & S64_MIN)); 2074 /* max signed is min(sign bit) | max(other bits) */ 2075 reg->smax_value = min_t(s64, reg->smax_value, 2076 reg->var_off.value | (reg->var_off.mask & S64_MAX)); 2077 reg->umin_value = max(reg->umin_value, reg->var_off.value); 2078 reg->umax_value = min(reg->umax_value, 2079 reg->var_off.value | reg->var_off.mask); 2080 } 2081 2082 static void __update_reg_bounds(struct bpf_reg_state *reg) 2083 { 2084 __update_reg32_bounds(reg); 2085 __update_reg64_bounds(reg); 2086 } 2087 2088 /* Uses signed min/max values to inform unsigned, and vice-versa */ 2089 static void __reg32_deduce_bounds(struct bpf_reg_state *reg) 2090 { 2091 /* Learn sign from signed bounds. 2092 * If we cannot cross the sign boundary, then signed and unsigned bounds 2093 * are the same, so combine. This works even in the negative case, e.g. 2094 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff. 2095 */ 2096 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) { 2097 reg->s32_min_value = reg->u32_min_value = 2098 max_t(u32, reg->s32_min_value, reg->u32_min_value); 2099 reg->s32_max_value = reg->u32_max_value = 2100 min_t(u32, reg->s32_max_value, reg->u32_max_value); 2101 return; 2102 } 2103 /* Learn sign from unsigned bounds. Signed bounds cross the sign 2104 * boundary, so we must be careful. 2105 */ 2106 if ((s32)reg->u32_max_value >= 0) { 2107 /* Positive. We can't learn anything from the smin, but smax 2108 * is positive, hence safe. 2109 */ 2110 reg->s32_min_value = reg->u32_min_value; 2111 reg->s32_max_value = reg->u32_max_value = 2112 min_t(u32, reg->s32_max_value, reg->u32_max_value); 2113 } else if ((s32)reg->u32_min_value < 0) { 2114 /* Negative. We can't learn anything from the smax, but smin 2115 * is negative, hence safe. 2116 */ 2117 reg->s32_min_value = reg->u32_min_value = 2118 max_t(u32, reg->s32_min_value, reg->u32_min_value); 2119 reg->s32_max_value = reg->u32_max_value; 2120 } 2121 } 2122 2123 static void __reg64_deduce_bounds(struct bpf_reg_state *reg) 2124 { 2125 /* Learn sign from signed bounds. 2126 * If we cannot cross the sign boundary, then signed and unsigned bounds 2127 * are the same, so combine. This works even in the negative case, e.g. 2128 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff. 2129 */ 2130 if (reg->smin_value >= 0 || reg->smax_value < 0) { 2131 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value, 2132 reg->umin_value); 2133 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value, 2134 reg->umax_value); 2135 return; 2136 } 2137 /* Learn sign from unsigned bounds. Signed bounds cross the sign 2138 * boundary, so we must be careful. 2139 */ 2140 if ((s64)reg->umax_value >= 0) { 2141 /* Positive. We can't learn anything from the smin, but smax 2142 * is positive, hence safe. 2143 */ 2144 reg->smin_value = reg->umin_value; 2145 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value, 2146 reg->umax_value); 2147 } else if ((s64)reg->umin_value < 0) { 2148 /* Negative. We can't learn anything from the smax, but smin 2149 * is negative, hence safe. 2150 */ 2151 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value, 2152 reg->umin_value); 2153 reg->smax_value = reg->umax_value; 2154 } 2155 } 2156 2157 static void __reg_deduce_bounds(struct bpf_reg_state *reg) 2158 { 2159 __reg32_deduce_bounds(reg); 2160 __reg64_deduce_bounds(reg); 2161 } 2162 2163 /* Attempts to improve var_off based on unsigned min/max information */ 2164 static void __reg_bound_offset(struct bpf_reg_state *reg) 2165 { 2166 struct tnum var64_off = tnum_intersect(reg->var_off, 2167 tnum_range(reg->umin_value, 2168 reg->umax_value)); 2169 struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off), 2170 tnum_range(reg->u32_min_value, 2171 reg->u32_max_value)); 2172 2173 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off); 2174 } 2175 2176 static void reg_bounds_sync(struct bpf_reg_state *reg) 2177 { 2178 /* We might have learned new bounds from the var_off. */ 2179 __update_reg_bounds(reg); 2180 /* We might have learned something about the sign bit. */ 2181 __reg_deduce_bounds(reg); 2182 /* We might have learned some bits from the bounds. */ 2183 __reg_bound_offset(reg); 2184 /* Intersecting with the old var_off might have improved our bounds 2185 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc), 2186 * then new var_off is (0; 0x7f...fc) which improves our umax. 2187 */ 2188 __update_reg_bounds(reg); 2189 } 2190 2191 static bool __reg32_bound_s64(s32 a) 2192 { 2193 return a >= 0 && a <= S32_MAX; 2194 } 2195 2196 static void __reg_assign_32_into_64(struct bpf_reg_state *reg) 2197 { 2198 reg->umin_value = reg->u32_min_value; 2199 reg->umax_value = reg->u32_max_value; 2200 2201 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must 2202 * be positive otherwise set to worse case bounds and refine later 2203 * from tnum. 2204 */ 2205 if (__reg32_bound_s64(reg->s32_min_value) && 2206 __reg32_bound_s64(reg->s32_max_value)) { 2207 reg->smin_value = reg->s32_min_value; 2208 reg->smax_value = reg->s32_max_value; 2209 } else { 2210 reg->smin_value = 0; 2211 reg->smax_value = U32_MAX; 2212 } 2213 } 2214 2215 static void __reg_combine_32_into_64(struct bpf_reg_state *reg) 2216 { 2217 /* special case when 64-bit register has upper 32-bit register 2218 * zeroed. Typically happens after zext or <<32, >>32 sequence 2219 * allowing us to use 32-bit bounds directly, 2220 */ 2221 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) { 2222 __reg_assign_32_into_64(reg); 2223 } else { 2224 /* Otherwise the best we can do is push lower 32bit known and 2225 * unknown bits into register (var_off set from jmp logic) 2226 * then learn as much as possible from the 64-bit tnum 2227 * known and unknown bits. The previous smin/smax bounds are 2228 * invalid here because of jmp32 compare so mark them unknown 2229 * so they do not impact tnum bounds calculation. 2230 */ 2231 __mark_reg64_unbounded(reg); 2232 } 2233 reg_bounds_sync(reg); 2234 } 2235 2236 static bool __reg64_bound_s32(s64 a) 2237 { 2238 return a >= S32_MIN && a <= S32_MAX; 2239 } 2240 2241 static bool __reg64_bound_u32(u64 a) 2242 { 2243 return a >= U32_MIN && a <= U32_MAX; 2244 } 2245 2246 static void __reg_combine_64_into_32(struct bpf_reg_state *reg) 2247 { 2248 __mark_reg32_unbounded(reg); 2249 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) { 2250 reg->s32_min_value = (s32)reg->smin_value; 2251 reg->s32_max_value = (s32)reg->smax_value; 2252 } 2253 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) { 2254 reg->u32_min_value = (u32)reg->umin_value; 2255 reg->u32_max_value = (u32)reg->umax_value; 2256 } 2257 reg_bounds_sync(reg); 2258 } 2259 2260 /* Mark a register as having a completely unknown (scalar) value. */ 2261 static void __mark_reg_unknown(const struct bpf_verifier_env *env, 2262 struct bpf_reg_state *reg) 2263 { 2264 /* 2265 * Clear type, off, and union(map_ptr, range) and 2266 * padding between 'type' and union 2267 */ 2268 memset(reg, 0, offsetof(struct bpf_reg_state, var_off)); 2269 reg->type = SCALAR_VALUE; 2270 reg->id = 0; 2271 reg->ref_obj_id = 0; 2272 reg->var_off = tnum_unknown; 2273 reg->frameno = 0; 2274 reg->precise = !env->bpf_capable; 2275 __mark_reg_unbounded(reg); 2276 } 2277 2278 static void mark_reg_unknown(struct bpf_verifier_env *env, 2279 struct bpf_reg_state *regs, u32 regno) 2280 { 2281 if (WARN_ON(regno >= MAX_BPF_REG)) { 2282 verbose(env, "mark_reg_unknown(regs, %u)\n", regno); 2283 /* Something bad happened, let's kill all regs except FP */ 2284 for (regno = 0; regno < BPF_REG_FP; regno++) 2285 __mark_reg_not_init(env, regs + regno); 2286 return; 2287 } 2288 __mark_reg_unknown(env, regs + regno); 2289 } 2290 2291 static void __mark_reg_not_init(const struct bpf_verifier_env *env, 2292 struct bpf_reg_state *reg) 2293 { 2294 __mark_reg_unknown(env, reg); 2295 reg->type = NOT_INIT; 2296 } 2297 2298 static void mark_reg_not_init(struct bpf_verifier_env *env, 2299 struct bpf_reg_state *regs, u32 regno) 2300 { 2301 if (WARN_ON(regno >= MAX_BPF_REG)) { 2302 verbose(env, "mark_reg_not_init(regs, %u)\n", regno); 2303 /* Something bad happened, let's kill all regs except FP */ 2304 for (regno = 0; regno < BPF_REG_FP; regno++) 2305 __mark_reg_not_init(env, regs + regno); 2306 return; 2307 } 2308 __mark_reg_not_init(env, regs + regno); 2309 } 2310 2311 static void mark_btf_ld_reg(struct bpf_verifier_env *env, 2312 struct bpf_reg_state *regs, u32 regno, 2313 enum bpf_reg_type reg_type, 2314 struct btf *btf, u32 btf_id, 2315 enum bpf_type_flag flag) 2316 { 2317 if (reg_type == SCALAR_VALUE) { 2318 mark_reg_unknown(env, regs, regno); 2319 return; 2320 } 2321 mark_reg_known_zero(env, regs, regno); 2322 regs[regno].type = PTR_TO_BTF_ID | flag; 2323 regs[regno].btf = btf; 2324 regs[regno].btf_id = btf_id; 2325 } 2326 2327 #define DEF_NOT_SUBREG (0) 2328 static void init_reg_state(struct bpf_verifier_env *env, 2329 struct bpf_func_state *state) 2330 { 2331 struct bpf_reg_state *regs = state->regs; 2332 int i; 2333 2334 for (i = 0; i < MAX_BPF_REG; i++) { 2335 mark_reg_not_init(env, regs, i); 2336 regs[i].live = REG_LIVE_NONE; 2337 regs[i].parent = NULL; 2338 regs[i].subreg_def = DEF_NOT_SUBREG; 2339 } 2340 2341 /* frame pointer */ 2342 regs[BPF_REG_FP].type = PTR_TO_STACK; 2343 mark_reg_known_zero(env, regs, BPF_REG_FP); 2344 regs[BPF_REG_FP].frameno = state->frameno; 2345 } 2346 2347 #define BPF_MAIN_FUNC (-1) 2348 static void init_func_state(struct bpf_verifier_env *env, 2349 struct bpf_func_state *state, 2350 int callsite, int frameno, int subprogno) 2351 { 2352 state->callsite = callsite; 2353 state->frameno = frameno; 2354 state->subprogno = subprogno; 2355 state->callback_ret_range = tnum_range(0, 0); 2356 init_reg_state(env, state); 2357 mark_verifier_state_scratched(env); 2358 } 2359 2360 /* Similar to push_stack(), but for async callbacks */ 2361 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env, 2362 int insn_idx, int prev_insn_idx, 2363 int subprog) 2364 { 2365 struct bpf_verifier_stack_elem *elem; 2366 struct bpf_func_state *frame; 2367 2368 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL); 2369 if (!elem) 2370 goto err; 2371 2372 elem->insn_idx = insn_idx; 2373 elem->prev_insn_idx = prev_insn_idx; 2374 elem->next = env->head; 2375 elem->log_pos = env->log.end_pos; 2376 env->head = elem; 2377 env->stack_size++; 2378 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) { 2379 verbose(env, 2380 "The sequence of %d jumps is too complex for async cb.\n", 2381 env->stack_size); 2382 goto err; 2383 } 2384 /* Unlike push_stack() do not copy_verifier_state(). 2385 * The caller state doesn't matter. 2386 * This is async callback. It starts in a fresh stack. 2387 * Initialize it similar to do_check_common(). 2388 */ 2389 elem->st.branches = 1; 2390 frame = kzalloc(sizeof(*frame), GFP_KERNEL); 2391 if (!frame) 2392 goto err; 2393 init_func_state(env, frame, 2394 BPF_MAIN_FUNC /* callsite */, 2395 0 /* frameno within this callchain */, 2396 subprog /* subprog number within this prog */); 2397 elem->st.frame[0] = frame; 2398 return &elem->st; 2399 err: 2400 free_verifier_state(env->cur_state, true); 2401 env->cur_state = NULL; 2402 /* pop all elements and return */ 2403 while (!pop_stack(env, NULL, NULL, false)); 2404 return NULL; 2405 } 2406 2407 2408 enum reg_arg_type { 2409 SRC_OP, /* register is used as source operand */ 2410 DST_OP, /* register is used as destination operand */ 2411 DST_OP_NO_MARK /* same as above, check only, don't mark */ 2412 }; 2413 2414 static int cmp_subprogs(const void *a, const void *b) 2415 { 2416 return ((struct bpf_subprog_info *)a)->start - 2417 ((struct bpf_subprog_info *)b)->start; 2418 } 2419 2420 static int find_subprog(struct bpf_verifier_env *env, int off) 2421 { 2422 struct bpf_subprog_info *p; 2423 2424 p = bsearch(&off, env->subprog_info, env->subprog_cnt, 2425 sizeof(env->subprog_info[0]), cmp_subprogs); 2426 if (!p) 2427 return -ENOENT; 2428 return p - env->subprog_info; 2429 2430 } 2431 2432 static int add_subprog(struct bpf_verifier_env *env, int off) 2433 { 2434 int insn_cnt = env->prog->len; 2435 int ret; 2436 2437 if (off >= insn_cnt || off < 0) { 2438 verbose(env, "call to invalid destination\n"); 2439 return -EINVAL; 2440 } 2441 ret = find_subprog(env, off); 2442 if (ret >= 0) 2443 return ret; 2444 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) { 2445 verbose(env, "too many subprograms\n"); 2446 return -E2BIG; 2447 } 2448 /* determine subprog starts. The end is one before the next starts */ 2449 env->subprog_info[env->subprog_cnt++].start = off; 2450 sort(env->subprog_info, env->subprog_cnt, 2451 sizeof(env->subprog_info[0]), cmp_subprogs, NULL); 2452 return env->subprog_cnt - 1; 2453 } 2454 2455 #define MAX_KFUNC_DESCS 256 2456 #define MAX_KFUNC_BTFS 256 2457 2458 struct bpf_kfunc_desc { 2459 struct btf_func_model func_model; 2460 u32 func_id; 2461 s32 imm; 2462 u16 offset; 2463 unsigned long addr; 2464 }; 2465 2466 struct bpf_kfunc_btf { 2467 struct btf *btf; 2468 struct module *module; 2469 u16 offset; 2470 }; 2471 2472 struct bpf_kfunc_desc_tab { 2473 /* Sorted by func_id (BTF ID) and offset (fd_array offset) during 2474 * verification. JITs do lookups by bpf_insn, where func_id may not be 2475 * available, therefore at the end of verification do_misc_fixups() 2476 * sorts this by imm and offset. 2477 */ 2478 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS]; 2479 u32 nr_descs; 2480 }; 2481 2482 struct bpf_kfunc_btf_tab { 2483 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS]; 2484 u32 nr_descs; 2485 }; 2486 2487 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b) 2488 { 2489 const struct bpf_kfunc_desc *d0 = a; 2490 const struct bpf_kfunc_desc *d1 = b; 2491 2492 /* func_id is not greater than BTF_MAX_TYPE */ 2493 return d0->func_id - d1->func_id ?: d0->offset - d1->offset; 2494 } 2495 2496 static int kfunc_btf_cmp_by_off(const void *a, const void *b) 2497 { 2498 const struct bpf_kfunc_btf *d0 = a; 2499 const struct bpf_kfunc_btf *d1 = b; 2500 2501 return d0->offset - d1->offset; 2502 } 2503 2504 static const struct bpf_kfunc_desc * 2505 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset) 2506 { 2507 struct bpf_kfunc_desc desc = { 2508 .func_id = func_id, 2509 .offset = offset, 2510 }; 2511 struct bpf_kfunc_desc_tab *tab; 2512 2513 tab = prog->aux->kfunc_tab; 2514 return bsearch(&desc, tab->descs, tab->nr_descs, 2515 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off); 2516 } 2517 2518 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id, 2519 u16 btf_fd_idx, u8 **func_addr) 2520 { 2521 const struct bpf_kfunc_desc *desc; 2522 2523 desc = find_kfunc_desc(prog, func_id, btf_fd_idx); 2524 if (!desc) 2525 return -EFAULT; 2526 2527 *func_addr = (u8 *)desc->addr; 2528 return 0; 2529 } 2530 2531 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env, 2532 s16 offset) 2533 { 2534 struct bpf_kfunc_btf kf_btf = { .offset = offset }; 2535 struct bpf_kfunc_btf_tab *tab; 2536 struct bpf_kfunc_btf *b; 2537 struct module *mod; 2538 struct btf *btf; 2539 int btf_fd; 2540 2541 tab = env->prog->aux->kfunc_btf_tab; 2542 b = bsearch(&kf_btf, tab->descs, tab->nr_descs, 2543 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off); 2544 if (!b) { 2545 if (tab->nr_descs == MAX_KFUNC_BTFS) { 2546 verbose(env, "too many different module BTFs\n"); 2547 return ERR_PTR(-E2BIG); 2548 } 2549 2550 if (bpfptr_is_null(env->fd_array)) { 2551 verbose(env, "kfunc offset > 0 without fd_array is invalid\n"); 2552 return ERR_PTR(-EPROTO); 2553 } 2554 2555 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array, 2556 offset * sizeof(btf_fd), 2557 sizeof(btf_fd))) 2558 return ERR_PTR(-EFAULT); 2559 2560 btf = btf_get_by_fd(btf_fd); 2561 if (IS_ERR(btf)) { 2562 verbose(env, "invalid module BTF fd specified\n"); 2563 return btf; 2564 } 2565 2566 if (!btf_is_module(btf)) { 2567 verbose(env, "BTF fd for kfunc is not a module BTF\n"); 2568 btf_put(btf); 2569 return ERR_PTR(-EINVAL); 2570 } 2571 2572 mod = btf_try_get_module(btf); 2573 if (!mod) { 2574 btf_put(btf); 2575 return ERR_PTR(-ENXIO); 2576 } 2577 2578 b = &tab->descs[tab->nr_descs++]; 2579 b->btf = btf; 2580 b->module = mod; 2581 b->offset = offset; 2582 2583 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]), 2584 kfunc_btf_cmp_by_off, NULL); 2585 } 2586 return b->btf; 2587 } 2588 2589 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab) 2590 { 2591 if (!tab) 2592 return; 2593 2594 while (tab->nr_descs--) { 2595 module_put(tab->descs[tab->nr_descs].module); 2596 btf_put(tab->descs[tab->nr_descs].btf); 2597 } 2598 kfree(tab); 2599 } 2600 2601 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset) 2602 { 2603 if (offset) { 2604 if (offset < 0) { 2605 /* In the future, this can be allowed to increase limit 2606 * of fd index into fd_array, interpreted as u16. 2607 */ 2608 verbose(env, "negative offset disallowed for kernel module function call\n"); 2609 return ERR_PTR(-EINVAL); 2610 } 2611 2612 return __find_kfunc_desc_btf(env, offset); 2613 } 2614 return btf_vmlinux ?: ERR_PTR(-ENOENT); 2615 } 2616 2617 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset) 2618 { 2619 const struct btf_type *func, *func_proto; 2620 struct bpf_kfunc_btf_tab *btf_tab; 2621 struct bpf_kfunc_desc_tab *tab; 2622 struct bpf_prog_aux *prog_aux; 2623 struct bpf_kfunc_desc *desc; 2624 const char *func_name; 2625 struct btf *desc_btf; 2626 unsigned long call_imm; 2627 unsigned long addr; 2628 int err; 2629 2630 prog_aux = env->prog->aux; 2631 tab = prog_aux->kfunc_tab; 2632 btf_tab = prog_aux->kfunc_btf_tab; 2633 if (!tab) { 2634 if (!btf_vmlinux) { 2635 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n"); 2636 return -ENOTSUPP; 2637 } 2638 2639 if (!env->prog->jit_requested) { 2640 verbose(env, "JIT is required for calling kernel function\n"); 2641 return -ENOTSUPP; 2642 } 2643 2644 if (!bpf_jit_supports_kfunc_call()) { 2645 verbose(env, "JIT does not support calling kernel function\n"); 2646 return -ENOTSUPP; 2647 } 2648 2649 if (!env->prog->gpl_compatible) { 2650 verbose(env, "cannot call kernel function from non-GPL compatible program\n"); 2651 return -EINVAL; 2652 } 2653 2654 tab = kzalloc(sizeof(*tab), GFP_KERNEL); 2655 if (!tab) 2656 return -ENOMEM; 2657 prog_aux->kfunc_tab = tab; 2658 } 2659 2660 /* func_id == 0 is always invalid, but instead of returning an error, be 2661 * conservative and wait until the code elimination pass before returning 2662 * error, so that invalid calls that get pruned out can be in BPF programs 2663 * loaded from userspace. It is also required that offset be untouched 2664 * for such calls. 2665 */ 2666 if (!func_id && !offset) 2667 return 0; 2668 2669 if (!btf_tab && offset) { 2670 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL); 2671 if (!btf_tab) 2672 return -ENOMEM; 2673 prog_aux->kfunc_btf_tab = btf_tab; 2674 } 2675 2676 desc_btf = find_kfunc_desc_btf(env, offset); 2677 if (IS_ERR(desc_btf)) { 2678 verbose(env, "failed to find BTF for kernel function\n"); 2679 return PTR_ERR(desc_btf); 2680 } 2681 2682 if (find_kfunc_desc(env->prog, func_id, offset)) 2683 return 0; 2684 2685 if (tab->nr_descs == MAX_KFUNC_DESCS) { 2686 verbose(env, "too many different kernel function calls\n"); 2687 return -E2BIG; 2688 } 2689 2690 func = btf_type_by_id(desc_btf, func_id); 2691 if (!func || !btf_type_is_func(func)) { 2692 verbose(env, "kernel btf_id %u is not a function\n", 2693 func_id); 2694 return -EINVAL; 2695 } 2696 func_proto = btf_type_by_id(desc_btf, func->type); 2697 if (!func_proto || !btf_type_is_func_proto(func_proto)) { 2698 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n", 2699 func_id); 2700 return -EINVAL; 2701 } 2702 2703 func_name = btf_name_by_offset(desc_btf, func->name_off); 2704 addr = kallsyms_lookup_name(func_name); 2705 if (!addr) { 2706 verbose(env, "cannot find address for kernel function %s\n", 2707 func_name); 2708 return -EINVAL; 2709 } 2710 specialize_kfunc(env, func_id, offset, &addr); 2711 2712 if (bpf_jit_supports_far_kfunc_call()) { 2713 call_imm = func_id; 2714 } else { 2715 call_imm = BPF_CALL_IMM(addr); 2716 /* Check whether the relative offset overflows desc->imm */ 2717 if ((unsigned long)(s32)call_imm != call_imm) { 2718 verbose(env, "address of kernel function %s is out of range\n", 2719 func_name); 2720 return -EINVAL; 2721 } 2722 } 2723 2724 if (bpf_dev_bound_kfunc_id(func_id)) { 2725 err = bpf_dev_bound_kfunc_check(&env->log, prog_aux); 2726 if (err) 2727 return err; 2728 } 2729 2730 desc = &tab->descs[tab->nr_descs++]; 2731 desc->func_id = func_id; 2732 desc->imm = call_imm; 2733 desc->offset = offset; 2734 desc->addr = addr; 2735 err = btf_distill_func_proto(&env->log, desc_btf, 2736 func_proto, func_name, 2737 &desc->func_model); 2738 if (!err) 2739 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]), 2740 kfunc_desc_cmp_by_id_off, NULL); 2741 return err; 2742 } 2743 2744 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b) 2745 { 2746 const struct bpf_kfunc_desc *d0 = a; 2747 const struct bpf_kfunc_desc *d1 = b; 2748 2749 if (d0->imm != d1->imm) 2750 return d0->imm < d1->imm ? -1 : 1; 2751 if (d0->offset != d1->offset) 2752 return d0->offset < d1->offset ? -1 : 1; 2753 return 0; 2754 } 2755 2756 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog) 2757 { 2758 struct bpf_kfunc_desc_tab *tab; 2759 2760 tab = prog->aux->kfunc_tab; 2761 if (!tab) 2762 return; 2763 2764 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]), 2765 kfunc_desc_cmp_by_imm_off, NULL); 2766 } 2767 2768 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog) 2769 { 2770 return !!prog->aux->kfunc_tab; 2771 } 2772 2773 const struct btf_func_model * 2774 bpf_jit_find_kfunc_model(const struct bpf_prog *prog, 2775 const struct bpf_insn *insn) 2776 { 2777 const struct bpf_kfunc_desc desc = { 2778 .imm = insn->imm, 2779 .offset = insn->off, 2780 }; 2781 const struct bpf_kfunc_desc *res; 2782 struct bpf_kfunc_desc_tab *tab; 2783 2784 tab = prog->aux->kfunc_tab; 2785 res = bsearch(&desc, tab->descs, tab->nr_descs, 2786 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off); 2787 2788 return res ? &res->func_model : NULL; 2789 } 2790 2791 static int add_subprog_and_kfunc(struct bpf_verifier_env *env) 2792 { 2793 struct bpf_subprog_info *subprog = env->subprog_info; 2794 struct bpf_insn *insn = env->prog->insnsi; 2795 int i, ret, insn_cnt = env->prog->len; 2796 2797 /* Add entry function. */ 2798 ret = add_subprog(env, 0); 2799 if (ret) 2800 return ret; 2801 2802 for (i = 0; i < insn_cnt; i++, insn++) { 2803 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) && 2804 !bpf_pseudo_kfunc_call(insn)) 2805 continue; 2806 2807 if (!env->bpf_capable) { 2808 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n"); 2809 return -EPERM; 2810 } 2811 2812 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn)) 2813 ret = add_subprog(env, i + insn->imm + 1); 2814 else 2815 ret = add_kfunc_call(env, insn->imm, insn->off); 2816 2817 if (ret < 0) 2818 return ret; 2819 } 2820 2821 /* Add a fake 'exit' subprog which could simplify subprog iteration 2822 * logic. 'subprog_cnt' should not be increased. 2823 */ 2824 subprog[env->subprog_cnt].start = insn_cnt; 2825 2826 if (env->log.level & BPF_LOG_LEVEL2) 2827 for (i = 0; i < env->subprog_cnt; i++) 2828 verbose(env, "func#%d @%d\n", i, subprog[i].start); 2829 2830 return 0; 2831 } 2832 2833 static int check_subprogs(struct bpf_verifier_env *env) 2834 { 2835 int i, subprog_start, subprog_end, off, cur_subprog = 0; 2836 struct bpf_subprog_info *subprog = env->subprog_info; 2837 struct bpf_insn *insn = env->prog->insnsi; 2838 int insn_cnt = env->prog->len; 2839 2840 /* now check that all jumps are within the same subprog */ 2841 subprog_start = subprog[cur_subprog].start; 2842 subprog_end = subprog[cur_subprog + 1].start; 2843 for (i = 0; i < insn_cnt; i++) { 2844 u8 code = insn[i].code; 2845 2846 if (code == (BPF_JMP | BPF_CALL) && 2847 insn[i].src_reg == 0 && 2848 insn[i].imm == BPF_FUNC_tail_call) 2849 subprog[cur_subprog].has_tail_call = true; 2850 if (BPF_CLASS(code) == BPF_LD && 2851 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND)) 2852 subprog[cur_subprog].has_ld_abs = true; 2853 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32) 2854 goto next; 2855 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL) 2856 goto next; 2857 off = i + insn[i].off + 1; 2858 if (off < subprog_start || off >= subprog_end) { 2859 verbose(env, "jump out of range from insn %d to %d\n", i, off); 2860 return -EINVAL; 2861 } 2862 next: 2863 if (i == subprog_end - 1) { 2864 /* to avoid fall-through from one subprog into another 2865 * the last insn of the subprog should be either exit 2866 * or unconditional jump back 2867 */ 2868 if (code != (BPF_JMP | BPF_EXIT) && 2869 code != (BPF_JMP | BPF_JA)) { 2870 verbose(env, "last insn is not an exit or jmp\n"); 2871 return -EINVAL; 2872 } 2873 subprog_start = subprog_end; 2874 cur_subprog++; 2875 if (cur_subprog < env->subprog_cnt) 2876 subprog_end = subprog[cur_subprog + 1].start; 2877 } 2878 } 2879 return 0; 2880 } 2881 2882 /* Parentage chain of this register (or stack slot) should take care of all 2883 * issues like callee-saved registers, stack slot allocation time, etc. 2884 */ 2885 static int mark_reg_read(struct bpf_verifier_env *env, 2886 const struct bpf_reg_state *state, 2887 struct bpf_reg_state *parent, u8 flag) 2888 { 2889 bool writes = parent == state->parent; /* Observe write marks */ 2890 int cnt = 0; 2891 2892 while (parent) { 2893 /* if read wasn't screened by an earlier write ... */ 2894 if (writes && state->live & REG_LIVE_WRITTEN) 2895 break; 2896 if (parent->live & REG_LIVE_DONE) { 2897 verbose(env, "verifier BUG type %s var_off %lld off %d\n", 2898 reg_type_str(env, parent->type), 2899 parent->var_off.value, parent->off); 2900 return -EFAULT; 2901 } 2902 /* The first condition is more likely to be true than the 2903 * second, checked it first. 2904 */ 2905 if ((parent->live & REG_LIVE_READ) == flag || 2906 parent->live & REG_LIVE_READ64) 2907 /* The parentage chain never changes and 2908 * this parent was already marked as LIVE_READ. 2909 * There is no need to keep walking the chain again and 2910 * keep re-marking all parents as LIVE_READ. 2911 * This case happens when the same register is read 2912 * multiple times without writes into it in-between. 2913 * Also, if parent has the stronger REG_LIVE_READ64 set, 2914 * then no need to set the weak REG_LIVE_READ32. 2915 */ 2916 break; 2917 /* ... then we depend on parent's value */ 2918 parent->live |= flag; 2919 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */ 2920 if (flag == REG_LIVE_READ64) 2921 parent->live &= ~REG_LIVE_READ32; 2922 state = parent; 2923 parent = state->parent; 2924 writes = true; 2925 cnt++; 2926 } 2927 2928 if (env->longest_mark_read_walk < cnt) 2929 env->longest_mark_read_walk = cnt; 2930 return 0; 2931 } 2932 2933 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 2934 { 2935 struct bpf_func_state *state = func(env, reg); 2936 int spi, ret; 2937 2938 /* For CONST_PTR_TO_DYNPTR, it must have already been done by 2939 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in 2940 * check_kfunc_call. 2941 */ 2942 if (reg->type == CONST_PTR_TO_DYNPTR) 2943 return 0; 2944 spi = dynptr_get_spi(env, reg); 2945 if (spi < 0) 2946 return spi; 2947 /* Caller ensures dynptr is valid and initialized, which means spi is in 2948 * bounds and spi is the first dynptr slot. Simply mark stack slot as 2949 * read. 2950 */ 2951 ret = mark_reg_read(env, &state->stack[spi].spilled_ptr, 2952 state->stack[spi].spilled_ptr.parent, REG_LIVE_READ64); 2953 if (ret) 2954 return ret; 2955 return mark_reg_read(env, &state->stack[spi - 1].spilled_ptr, 2956 state->stack[spi - 1].spilled_ptr.parent, REG_LIVE_READ64); 2957 } 2958 2959 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 2960 int spi, int nr_slots) 2961 { 2962 struct bpf_func_state *state = func(env, reg); 2963 int err, i; 2964 2965 for (i = 0; i < nr_slots; i++) { 2966 struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr; 2967 2968 err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64); 2969 if (err) 2970 return err; 2971 2972 mark_stack_slot_scratched(env, spi - i); 2973 } 2974 2975 return 0; 2976 } 2977 2978 /* This function is supposed to be used by the following 32-bit optimization 2979 * code only. It returns TRUE if the source or destination register operates 2980 * on 64-bit, otherwise return FALSE. 2981 */ 2982 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn, 2983 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t) 2984 { 2985 u8 code, class, op; 2986 2987 code = insn->code; 2988 class = BPF_CLASS(code); 2989 op = BPF_OP(code); 2990 if (class == BPF_JMP) { 2991 /* BPF_EXIT for "main" will reach here. Return TRUE 2992 * conservatively. 2993 */ 2994 if (op == BPF_EXIT) 2995 return true; 2996 if (op == BPF_CALL) { 2997 /* BPF to BPF call will reach here because of marking 2998 * caller saved clobber with DST_OP_NO_MARK for which we 2999 * don't care the register def because they are anyway 3000 * marked as NOT_INIT already. 3001 */ 3002 if (insn->src_reg == BPF_PSEUDO_CALL) 3003 return false; 3004 /* Helper call will reach here because of arg type 3005 * check, conservatively return TRUE. 3006 */ 3007 if (t == SRC_OP) 3008 return true; 3009 3010 return false; 3011 } 3012 } 3013 3014 if (class == BPF_ALU64 || class == BPF_JMP || 3015 /* BPF_END always use BPF_ALU class. */ 3016 (class == BPF_ALU && op == BPF_END && insn->imm == 64)) 3017 return true; 3018 3019 if (class == BPF_ALU || class == BPF_JMP32) 3020 return false; 3021 3022 if (class == BPF_LDX) { 3023 if (t != SRC_OP) 3024 return BPF_SIZE(code) == BPF_DW; 3025 /* LDX source must be ptr. */ 3026 return true; 3027 } 3028 3029 if (class == BPF_STX) { 3030 /* BPF_STX (including atomic variants) has multiple source 3031 * operands, one of which is a ptr. Check whether the caller is 3032 * asking about it. 3033 */ 3034 if (t == SRC_OP && reg->type != SCALAR_VALUE) 3035 return true; 3036 return BPF_SIZE(code) == BPF_DW; 3037 } 3038 3039 if (class == BPF_LD) { 3040 u8 mode = BPF_MODE(code); 3041 3042 /* LD_IMM64 */ 3043 if (mode == BPF_IMM) 3044 return true; 3045 3046 /* Both LD_IND and LD_ABS return 32-bit data. */ 3047 if (t != SRC_OP) 3048 return false; 3049 3050 /* Implicit ctx ptr. */ 3051 if (regno == BPF_REG_6) 3052 return true; 3053 3054 /* Explicit source could be any width. */ 3055 return true; 3056 } 3057 3058 if (class == BPF_ST) 3059 /* The only source register for BPF_ST is a ptr. */ 3060 return true; 3061 3062 /* Conservatively return true at default. */ 3063 return true; 3064 } 3065 3066 /* Return the regno defined by the insn, or -1. */ 3067 static int insn_def_regno(const struct bpf_insn *insn) 3068 { 3069 switch (BPF_CLASS(insn->code)) { 3070 case BPF_JMP: 3071 case BPF_JMP32: 3072 case BPF_ST: 3073 return -1; 3074 case BPF_STX: 3075 if (BPF_MODE(insn->code) == BPF_ATOMIC && 3076 (insn->imm & BPF_FETCH)) { 3077 if (insn->imm == BPF_CMPXCHG) 3078 return BPF_REG_0; 3079 else 3080 return insn->src_reg; 3081 } else { 3082 return -1; 3083 } 3084 default: 3085 return insn->dst_reg; 3086 } 3087 } 3088 3089 /* Return TRUE if INSN has defined any 32-bit value explicitly. */ 3090 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn) 3091 { 3092 int dst_reg = insn_def_regno(insn); 3093 3094 if (dst_reg == -1) 3095 return false; 3096 3097 return !is_reg64(env, insn, dst_reg, NULL, DST_OP); 3098 } 3099 3100 static void mark_insn_zext(struct bpf_verifier_env *env, 3101 struct bpf_reg_state *reg) 3102 { 3103 s32 def_idx = reg->subreg_def; 3104 3105 if (def_idx == DEF_NOT_SUBREG) 3106 return; 3107 3108 env->insn_aux_data[def_idx - 1].zext_dst = true; 3109 /* The dst will be zero extended, so won't be sub-register anymore. */ 3110 reg->subreg_def = DEF_NOT_SUBREG; 3111 } 3112 3113 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno, 3114 enum reg_arg_type t) 3115 { 3116 struct bpf_verifier_state *vstate = env->cur_state; 3117 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 3118 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx; 3119 struct bpf_reg_state *reg, *regs = state->regs; 3120 bool rw64; 3121 3122 if (regno >= MAX_BPF_REG) { 3123 verbose(env, "R%d is invalid\n", regno); 3124 return -EINVAL; 3125 } 3126 3127 mark_reg_scratched(env, regno); 3128 3129 reg = ®s[regno]; 3130 rw64 = is_reg64(env, insn, regno, reg, t); 3131 if (t == SRC_OP) { 3132 /* check whether register used as source operand can be read */ 3133 if (reg->type == NOT_INIT) { 3134 verbose(env, "R%d !read_ok\n", regno); 3135 return -EACCES; 3136 } 3137 /* We don't need to worry about FP liveness because it's read-only */ 3138 if (regno == BPF_REG_FP) 3139 return 0; 3140 3141 if (rw64) 3142 mark_insn_zext(env, reg); 3143 3144 return mark_reg_read(env, reg, reg->parent, 3145 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32); 3146 } else { 3147 /* check whether register used as dest operand can be written to */ 3148 if (regno == BPF_REG_FP) { 3149 verbose(env, "frame pointer is read only\n"); 3150 return -EACCES; 3151 } 3152 reg->live |= REG_LIVE_WRITTEN; 3153 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1; 3154 if (t == DST_OP) 3155 mark_reg_unknown(env, regs, regno); 3156 } 3157 return 0; 3158 } 3159 3160 static void mark_jmp_point(struct bpf_verifier_env *env, int idx) 3161 { 3162 env->insn_aux_data[idx].jmp_point = true; 3163 } 3164 3165 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx) 3166 { 3167 return env->insn_aux_data[insn_idx].jmp_point; 3168 } 3169 3170 /* for any branch, call, exit record the history of jmps in the given state */ 3171 static int push_jmp_history(struct bpf_verifier_env *env, 3172 struct bpf_verifier_state *cur) 3173 { 3174 u32 cnt = cur->jmp_history_cnt; 3175 struct bpf_idx_pair *p; 3176 size_t alloc_size; 3177 3178 if (!is_jmp_point(env, env->insn_idx)) 3179 return 0; 3180 3181 cnt++; 3182 alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p))); 3183 p = krealloc(cur->jmp_history, alloc_size, GFP_USER); 3184 if (!p) 3185 return -ENOMEM; 3186 p[cnt - 1].idx = env->insn_idx; 3187 p[cnt - 1].prev_idx = env->prev_insn_idx; 3188 cur->jmp_history = p; 3189 cur->jmp_history_cnt = cnt; 3190 return 0; 3191 } 3192 3193 /* Backtrack one insn at a time. If idx is not at the top of recorded 3194 * history then previous instruction came from straight line execution. 3195 */ 3196 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i, 3197 u32 *history) 3198 { 3199 u32 cnt = *history; 3200 3201 if (cnt && st->jmp_history[cnt - 1].idx == i) { 3202 i = st->jmp_history[cnt - 1].prev_idx; 3203 (*history)--; 3204 } else { 3205 i--; 3206 } 3207 return i; 3208 } 3209 3210 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn) 3211 { 3212 const struct btf_type *func; 3213 struct btf *desc_btf; 3214 3215 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL) 3216 return NULL; 3217 3218 desc_btf = find_kfunc_desc_btf(data, insn->off); 3219 if (IS_ERR(desc_btf)) 3220 return "<error>"; 3221 3222 func = btf_type_by_id(desc_btf, insn->imm); 3223 return btf_name_by_offset(desc_btf, func->name_off); 3224 } 3225 3226 static inline void bt_init(struct backtrack_state *bt, u32 frame) 3227 { 3228 bt->frame = frame; 3229 } 3230 3231 static inline void bt_reset(struct backtrack_state *bt) 3232 { 3233 struct bpf_verifier_env *env = bt->env; 3234 3235 memset(bt, 0, sizeof(*bt)); 3236 bt->env = env; 3237 } 3238 3239 static inline u32 bt_empty(struct backtrack_state *bt) 3240 { 3241 u64 mask = 0; 3242 int i; 3243 3244 for (i = 0; i <= bt->frame; i++) 3245 mask |= bt->reg_masks[i] | bt->stack_masks[i]; 3246 3247 return mask == 0; 3248 } 3249 3250 static inline int bt_subprog_enter(struct backtrack_state *bt) 3251 { 3252 if (bt->frame == MAX_CALL_FRAMES - 1) { 3253 verbose(bt->env, "BUG subprog enter from frame %d\n", bt->frame); 3254 WARN_ONCE(1, "verifier backtracking bug"); 3255 return -EFAULT; 3256 } 3257 bt->frame++; 3258 return 0; 3259 } 3260 3261 static inline int bt_subprog_exit(struct backtrack_state *bt) 3262 { 3263 if (bt->frame == 0) { 3264 verbose(bt->env, "BUG subprog exit from frame 0\n"); 3265 WARN_ONCE(1, "verifier backtracking bug"); 3266 return -EFAULT; 3267 } 3268 bt->frame--; 3269 return 0; 3270 } 3271 3272 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg) 3273 { 3274 bt->reg_masks[frame] |= 1 << reg; 3275 } 3276 3277 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg) 3278 { 3279 bt->reg_masks[frame] &= ~(1 << reg); 3280 } 3281 3282 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg) 3283 { 3284 bt_set_frame_reg(bt, bt->frame, reg); 3285 } 3286 3287 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg) 3288 { 3289 bt_clear_frame_reg(bt, bt->frame, reg); 3290 } 3291 3292 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot) 3293 { 3294 bt->stack_masks[frame] |= 1ull << slot; 3295 } 3296 3297 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot) 3298 { 3299 bt->stack_masks[frame] &= ~(1ull << slot); 3300 } 3301 3302 static inline void bt_set_slot(struct backtrack_state *bt, u32 slot) 3303 { 3304 bt_set_frame_slot(bt, bt->frame, slot); 3305 } 3306 3307 static inline void bt_clear_slot(struct backtrack_state *bt, u32 slot) 3308 { 3309 bt_clear_frame_slot(bt, bt->frame, slot); 3310 } 3311 3312 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame) 3313 { 3314 return bt->reg_masks[frame]; 3315 } 3316 3317 static inline u32 bt_reg_mask(struct backtrack_state *bt) 3318 { 3319 return bt->reg_masks[bt->frame]; 3320 } 3321 3322 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame) 3323 { 3324 return bt->stack_masks[frame]; 3325 } 3326 3327 static inline u64 bt_stack_mask(struct backtrack_state *bt) 3328 { 3329 return bt->stack_masks[bt->frame]; 3330 } 3331 3332 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg) 3333 { 3334 return bt->reg_masks[bt->frame] & (1 << reg); 3335 } 3336 3337 static inline bool bt_is_slot_set(struct backtrack_state *bt, u32 slot) 3338 { 3339 return bt->stack_masks[bt->frame] & (1ull << slot); 3340 } 3341 3342 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */ 3343 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask) 3344 { 3345 DECLARE_BITMAP(mask, 64); 3346 bool first = true; 3347 int i, n; 3348 3349 buf[0] = '\0'; 3350 3351 bitmap_from_u64(mask, reg_mask); 3352 for_each_set_bit(i, mask, 32) { 3353 n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i); 3354 first = false; 3355 buf += n; 3356 buf_sz -= n; 3357 if (buf_sz < 0) 3358 break; 3359 } 3360 } 3361 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */ 3362 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask) 3363 { 3364 DECLARE_BITMAP(mask, 64); 3365 bool first = true; 3366 int i, n; 3367 3368 buf[0] = '\0'; 3369 3370 bitmap_from_u64(mask, stack_mask); 3371 for_each_set_bit(i, mask, 64) { 3372 n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8); 3373 first = false; 3374 buf += n; 3375 buf_sz -= n; 3376 if (buf_sz < 0) 3377 break; 3378 } 3379 } 3380 3381 /* For given verifier state backtrack_insn() is called from the last insn to 3382 * the first insn. Its purpose is to compute a bitmask of registers and 3383 * stack slots that needs precision in the parent verifier state. 3384 * 3385 * @idx is an index of the instruction we are currently processing; 3386 * @subseq_idx is an index of the subsequent instruction that: 3387 * - *would be* executed next, if jump history is viewed in forward order; 3388 * - *was* processed previously during backtracking. 3389 */ 3390 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx, 3391 struct backtrack_state *bt) 3392 { 3393 const struct bpf_insn_cbs cbs = { 3394 .cb_call = disasm_kfunc_name, 3395 .cb_print = verbose, 3396 .private_data = env, 3397 }; 3398 struct bpf_insn *insn = env->prog->insnsi + idx; 3399 u8 class = BPF_CLASS(insn->code); 3400 u8 opcode = BPF_OP(insn->code); 3401 u8 mode = BPF_MODE(insn->code); 3402 u32 dreg = insn->dst_reg; 3403 u32 sreg = insn->src_reg; 3404 u32 spi, i; 3405 3406 if (insn->code == 0) 3407 return 0; 3408 if (env->log.level & BPF_LOG_LEVEL2) { 3409 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt)); 3410 verbose(env, "mark_precise: frame%d: regs=%s ", 3411 bt->frame, env->tmp_str_buf); 3412 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt)); 3413 verbose(env, "stack=%s before ", env->tmp_str_buf); 3414 verbose(env, "%d: ", idx); 3415 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks); 3416 } 3417 3418 if (class == BPF_ALU || class == BPF_ALU64) { 3419 if (!bt_is_reg_set(bt, dreg)) 3420 return 0; 3421 if (opcode == BPF_MOV) { 3422 if (BPF_SRC(insn->code) == BPF_X) { 3423 /* dreg = sreg 3424 * dreg needs precision after this insn 3425 * sreg needs precision before this insn 3426 */ 3427 bt_clear_reg(bt, dreg); 3428 bt_set_reg(bt, sreg); 3429 } else { 3430 /* dreg = K 3431 * dreg needs precision after this insn. 3432 * Corresponding register is already marked 3433 * as precise=true in this verifier state. 3434 * No further markings in parent are necessary 3435 */ 3436 bt_clear_reg(bt, dreg); 3437 } 3438 } else { 3439 if (BPF_SRC(insn->code) == BPF_X) { 3440 /* dreg += sreg 3441 * both dreg and sreg need precision 3442 * before this insn 3443 */ 3444 bt_set_reg(bt, sreg); 3445 } /* else dreg += K 3446 * dreg still needs precision before this insn 3447 */ 3448 } 3449 } else if (class == BPF_LDX) { 3450 if (!bt_is_reg_set(bt, dreg)) 3451 return 0; 3452 bt_clear_reg(bt, dreg); 3453 3454 /* scalars can only be spilled into stack w/o losing precision. 3455 * Load from any other memory can be zero extended. 3456 * The desire to keep that precision is already indicated 3457 * by 'precise' mark in corresponding register of this state. 3458 * No further tracking necessary. 3459 */ 3460 if (insn->src_reg != BPF_REG_FP) 3461 return 0; 3462 3463 /* dreg = *(u64 *)[fp - off] was a fill from the stack. 3464 * that [fp - off] slot contains scalar that needs to be 3465 * tracked with precision 3466 */ 3467 spi = (-insn->off - 1) / BPF_REG_SIZE; 3468 if (spi >= 64) { 3469 verbose(env, "BUG spi %d\n", spi); 3470 WARN_ONCE(1, "verifier backtracking bug"); 3471 return -EFAULT; 3472 } 3473 bt_set_slot(bt, spi); 3474 } else if (class == BPF_STX || class == BPF_ST) { 3475 if (bt_is_reg_set(bt, dreg)) 3476 /* stx & st shouldn't be using _scalar_ dst_reg 3477 * to access memory. It means backtracking 3478 * encountered a case of pointer subtraction. 3479 */ 3480 return -ENOTSUPP; 3481 /* scalars can only be spilled into stack */ 3482 if (insn->dst_reg != BPF_REG_FP) 3483 return 0; 3484 spi = (-insn->off - 1) / BPF_REG_SIZE; 3485 if (spi >= 64) { 3486 verbose(env, "BUG spi %d\n", spi); 3487 WARN_ONCE(1, "verifier backtracking bug"); 3488 return -EFAULT; 3489 } 3490 if (!bt_is_slot_set(bt, spi)) 3491 return 0; 3492 bt_clear_slot(bt, spi); 3493 if (class == BPF_STX) 3494 bt_set_reg(bt, sreg); 3495 } else if (class == BPF_JMP || class == BPF_JMP32) { 3496 if (bpf_pseudo_call(insn)) { 3497 int subprog_insn_idx, subprog; 3498 3499 subprog_insn_idx = idx + insn->imm + 1; 3500 subprog = find_subprog(env, subprog_insn_idx); 3501 if (subprog < 0) 3502 return -EFAULT; 3503 3504 if (subprog_is_global(env, subprog)) { 3505 /* check that jump history doesn't have any 3506 * extra instructions from subprog; the next 3507 * instruction after call to global subprog 3508 * should be literally next instruction in 3509 * caller program 3510 */ 3511 WARN_ONCE(idx + 1 != subseq_idx, "verifier backtracking bug"); 3512 /* r1-r5 are invalidated after subprog call, 3513 * so for global func call it shouldn't be set 3514 * anymore 3515 */ 3516 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) { 3517 verbose(env, "BUG regs %x\n", bt_reg_mask(bt)); 3518 WARN_ONCE(1, "verifier backtracking bug"); 3519 return -EFAULT; 3520 } 3521 /* global subprog always sets R0 */ 3522 bt_clear_reg(bt, BPF_REG_0); 3523 return 0; 3524 } else { 3525 /* static subprog call instruction, which 3526 * means that we are exiting current subprog, 3527 * so only r1-r5 could be still requested as 3528 * precise, r0 and r6-r10 or any stack slot in 3529 * the current frame should be zero by now 3530 */ 3531 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) { 3532 verbose(env, "BUG regs %x\n", bt_reg_mask(bt)); 3533 WARN_ONCE(1, "verifier backtracking bug"); 3534 return -EFAULT; 3535 } 3536 /* we don't track register spills perfectly, 3537 * so fallback to force-precise instead of failing */ 3538 if (bt_stack_mask(bt) != 0) 3539 return -ENOTSUPP; 3540 /* propagate r1-r5 to the caller */ 3541 for (i = BPF_REG_1; i <= BPF_REG_5; i++) { 3542 if (bt_is_reg_set(bt, i)) { 3543 bt_clear_reg(bt, i); 3544 bt_set_frame_reg(bt, bt->frame - 1, i); 3545 } 3546 } 3547 if (bt_subprog_exit(bt)) 3548 return -EFAULT; 3549 return 0; 3550 } 3551 } else if ((bpf_helper_call(insn) && 3552 is_callback_calling_function(insn->imm) && 3553 !is_async_callback_calling_function(insn->imm)) || 3554 (bpf_pseudo_kfunc_call(insn) && is_callback_calling_kfunc(insn->imm))) { 3555 /* callback-calling helper or kfunc call, which means 3556 * we are exiting from subprog, but unlike the subprog 3557 * call handling above, we shouldn't propagate 3558 * precision of r1-r5 (if any requested), as they are 3559 * not actually arguments passed directly to callback 3560 * subprogs 3561 */ 3562 if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) { 3563 verbose(env, "BUG regs %x\n", bt_reg_mask(bt)); 3564 WARN_ONCE(1, "verifier backtracking bug"); 3565 return -EFAULT; 3566 } 3567 if (bt_stack_mask(bt) != 0) 3568 return -ENOTSUPP; 3569 /* clear r1-r5 in callback subprog's mask */ 3570 for (i = BPF_REG_1; i <= BPF_REG_5; i++) 3571 bt_clear_reg(bt, i); 3572 if (bt_subprog_exit(bt)) 3573 return -EFAULT; 3574 return 0; 3575 } else if (opcode == BPF_CALL) { 3576 /* kfunc with imm==0 is invalid and fixup_kfunc_call will 3577 * catch this error later. Make backtracking conservative 3578 * with ENOTSUPP. 3579 */ 3580 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0) 3581 return -ENOTSUPP; 3582 /* regular helper call sets R0 */ 3583 bt_clear_reg(bt, BPF_REG_0); 3584 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) { 3585 /* if backtracing was looking for registers R1-R5 3586 * they should have been found already. 3587 */ 3588 verbose(env, "BUG regs %x\n", bt_reg_mask(bt)); 3589 WARN_ONCE(1, "verifier backtracking bug"); 3590 return -EFAULT; 3591 } 3592 } else if (opcode == BPF_EXIT) { 3593 bool r0_precise; 3594 3595 if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) { 3596 /* if backtracing was looking for registers R1-R5 3597 * they should have been found already. 3598 */ 3599 verbose(env, "BUG regs %x\n", bt_reg_mask(bt)); 3600 WARN_ONCE(1, "verifier backtracking bug"); 3601 return -EFAULT; 3602 } 3603 3604 /* BPF_EXIT in subprog or callback always returns 3605 * right after the call instruction, so by checking 3606 * whether the instruction at subseq_idx-1 is subprog 3607 * call or not we can distinguish actual exit from 3608 * *subprog* from exit from *callback*. In the former 3609 * case, we need to propagate r0 precision, if 3610 * necessary. In the former we never do that. 3611 */ 3612 r0_precise = subseq_idx - 1 >= 0 && 3613 bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) && 3614 bt_is_reg_set(bt, BPF_REG_0); 3615 3616 bt_clear_reg(bt, BPF_REG_0); 3617 if (bt_subprog_enter(bt)) 3618 return -EFAULT; 3619 3620 if (r0_precise) 3621 bt_set_reg(bt, BPF_REG_0); 3622 /* r6-r9 and stack slots will stay set in caller frame 3623 * bitmasks until we return back from callee(s) 3624 */ 3625 return 0; 3626 } else if (BPF_SRC(insn->code) == BPF_X) { 3627 if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg)) 3628 return 0; 3629 /* dreg <cond> sreg 3630 * Both dreg and sreg need precision before 3631 * this insn. If only sreg was marked precise 3632 * before it would be equally necessary to 3633 * propagate it to dreg. 3634 */ 3635 bt_set_reg(bt, dreg); 3636 bt_set_reg(bt, sreg); 3637 /* else dreg <cond> K 3638 * Only dreg still needs precision before 3639 * this insn, so for the K-based conditional 3640 * there is nothing new to be marked. 3641 */ 3642 } 3643 } else if (class == BPF_LD) { 3644 if (!bt_is_reg_set(bt, dreg)) 3645 return 0; 3646 bt_clear_reg(bt, dreg); 3647 /* It's ld_imm64 or ld_abs or ld_ind. 3648 * For ld_imm64 no further tracking of precision 3649 * into parent is necessary 3650 */ 3651 if (mode == BPF_IND || mode == BPF_ABS) 3652 /* to be analyzed */ 3653 return -ENOTSUPP; 3654 } 3655 return 0; 3656 } 3657 3658 /* the scalar precision tracking algorithm: 3659 * . at the start all registers have precise=false. 3660 * . scalar ranges are tracked as normal through alu and jmp insns. 3661 * . once precise value of the scalar register is used in: 3662 * . ptr + scalar alu 3663 * . if (scalar cond K|scalar) 3664 * . helper_call(.., scalar, ...) where ARG_CONST is expected 3665 * backtrack through the verifier states and mark all registers and 3666 * stack slots with spilled constants that these scalar regisers 3667 * should be precise. 3668 * . during state pruning two registers (or spilled stack slots) 3669 * are equivalent if both are not precise. 3670 * 3671 * Note the verifier cannot simply walk register parentage chain, 3672 * since many different registers and stack slots could have been 3673 * used to compute single precise scalar. 3674 * 3675 * The approach of starting with precise=true for all registers and then 3676 * backtrack to mark a register as not precise when the verifier detects 3677 * that program doesn't care about specific value (e.g., when helper 3678 * takes register as ARG_ANYTHING parameter) is not safe. 3679 * 3680 * It's ok to walk single parentage chain of the verifier states. 3681 * It's possible that this backtracking will go all the way till 1st insn. 3682 * All other branches will be explored for needing precision later. 3683 * 3684 * The backtracking needs to deal with cases like: 3685 * R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0) 3686 * r9 -= r8 3687 * r5 = r9 3688 * if r5 > 0x79f goto pc+7 3689 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff)) 3690 * r5 += 1 3691 * ... 3692 * call bpf_perf_event_output#25 3693 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO 3694 * 3695 * and this case: 3696 * r6 = 1 3697 * call foo // uses callee's r6 inside to compute r0 3698 * r0 += r6 3699 * if r0 == 0 goto 3700 * 3701 * to track above reg_mask/stack_mask needs to be independent for each frame. 3702 * 3703 * Also if parent's curframe > frame where backtracking started, 3704 * the verifier need to mark registers in both frames, otherwise callees 3705 * may incorrectly prune callers. This is similar to 3706 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences") 3707 * 3708 * For now backtracking falls back into conservative marking. 3709 */ 3710 static void mark_all_scalars_precise(struct bpf_verifier_env *env, 3711 struct bpf_verifier_state *st) 3712 { 3713 struct bpf_func_state *func; 3714 struct bpf_reg_state *reg; 3715 int i, j; 3716 3717 if (env->log.level & BPF_LOG_LEVEL2) { 3718 verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n", 3719 st->curframe); 3720 } 3721 3722 /* big hammer: mark all scalars precise in this path. 3723 * pop_stack may still get !precise scalars. 3724 * We also skip current state and go straight to first parent state, 3725 * because precision markings in current non-checkpointed state are 3726 * not needed. See why in the comment in __mark_chain_precision below. 3727 */ 3728 for (st = st->parent; st; st = st->parent) { 3729 for (i = 0; i <= st->curframe; i++) { 3730 func = st->frame[i]; 3731 for (j = 0; j < BPF_REG_FP; j++) { 3732 reg = &func->regs[j]; 3733 if (reg->type != SCALAR_VALUE || reg->precise) 3734 continue; 3735 reg->precise = true; 3736 if (env->log.level & BPF_LOG_LEVEL2) { 3737 verbose(env, "force_precise: frame%d: forcing r%d to be precise\n", 3738 i, j); 3739 } 3740 } 3741 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) { 3742 if (!is_spilled_reg(&func->stack[j])) 3743 continue; 3744 reg = &func->stack[j].spilled_ptr; 3745 if (reg->type != SCALAR_VALUE || reg->precise) 3746 continue; 3747 reg->precise = true; 3748 if (env->log.level & BPF_LOG_LEVEL2) { 3749 verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n", 3750 i, -(j + 1) * 8); 3751 } 3752 } 3753 } 3754 } 3755 } 3756 3757 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st) 3758 { 3759 struct bpf_func_state *func; 3760 struct bpf_reg_state *reg; 3761 int i, j; 3762 3763 for (i = 0; i <= st->curframe; i++) { 3764 func = st->frame[i]; 3765 for (j = 0; j < BPF_REG_FP; j++) { 3766 reg = &func->regs[j]; 3767 if (reg->type != SCALAR_VALUE) 3768 continue; 3769 reg->precise = false; 3770 } 3771 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) { 3772 if (!is_spilled_reg(&func->stack[j])) 3773 continue; 3774 reg = &func->stack[j].spilled_ptr; 3775 if (reg->type != SCALAR_VALUE) 3776 continue; 3777 reg->precise = false; 3778 } 3779 } 3780 } 3781 3782 static bool idset_contains(struct bpf_idset *s, u32 id) 3783 { 3784 u32 i; 3785 3786 for (i = 0; i < s->count; ++i) 3787 if (s->ids[i] == id) 3788 return true; 3789 3790 return false; 3791 } 3792 3793 static int idset_push(struct bpf_idset *s, u32 id) 3794 { 3795 if (WARN_ON_ONCE(s->count >= ARRAY_SIZE(s->ids))) 3796 return -EFAULT; 3797 s->ids[s->count++] = id; 3798 return 0; 3799 } 3800 3801 static void idset_reset(struct bpf_idset *s) 3802 { 3803 s->count = 0; 3804 } 3805 3806 /* Collect a set of IDs for all registers currently marked as precise in env->bt. 3807 * Mark all registers with these IDs as precise. 3808 */ 3809 static int mark_precise_scalar_ids(struct bpf_verifier_env *env, struct bpf_verifier_state *st) 3810 { 3811 struct bpf_idset *precise_ids = &env->idset_scratch; 3812 struct backtrack_state *bt = &env->bt; 3813 struct bpf_func_state *func; 3814 struct bpf_reg_state *reg; 3815 DECLARE_BITMAP(mask, 64); 3816 int i, fr; 3817 3818 idset_reset(precise_ids); 3819 3820 for (fr = bt->frame; fr >= 0; fr--) { 3821 func = st->frame[fr]; 3822 3823 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr)); 3824 for_each_set_bit(i, mask, 32) { 3825 reg = &func->regs[i]; 3826 if (!reg->id || reg->type != SCALAR_VALUE) 3827 continue; 3828 if (idset_push(precise_ids, reg->id)) 3829 return -EFAULT; 3830 } 3831 3832 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr)); 3833 for_each_set_bit(i, mask, 64) { 3834 if (i >= func->allocated_stack / BPF_REG_SIZE) 3835 break; 3836 if (!is_spilled_scalar_reg(&func->stack[i])) 3837 continue; 3838 reg = &func->stack[i].spilled_ptr; 3839 if (!reg->id) 3840 continue; 3841 if (idset_push(precise_ids, reg->id)) 3842 return -EFAULT; 3843 } 3844 } 3845 3846 for (fr = 0; fr <= st->curframe; ++fr) { 3847 func = st->frame[fr]; 3848 3849 for (i = BPF_REG_0; i < BPF_REG_10; ++i) { 3850 reg = &func->regs[i]; 3851 if (!reg->id) 3852 continue; 3853 if (!idset_contains(precise_ids, reg->id)) 3854 continue; 3855 bt_set_frame_reg(bt, fr, i); 3856 } 3857 for (i = 0; i < func->allocated_stack / BPF_REG_SIZE; ++i) { 3858 if (!is_spilled_scalar_reg(&func->stack[i])) 3859 continue; 3860 reg = &func->stack[i].spilled_ptr; 3861 if (!reg->id) 3862 continue; 3863 if (!idset_contains(precise_ids, reg->id)) 3864 continue; 3865 bt_set_frame_slot(bt, fr, i); 3866 } 3867 } 3868 3869 return 0; 3870 } 3871 3872 /* 3873 * __mark_chain_precision() backtracks BPF program instruction sequence and 3874 * chain of verifier states making sure that register *regno* (if regno >= 0) 3875 * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked 3876 * SCALARS, as well as any other registers and slots that contribute to 3877 * a tracked state of given registers/stack slots, depending on specific BPF 3878 * assembly instructions (see backtrack_insns() for exact instruction handling 3879 * logic). This backtracking relies on recorded jmp_history and is able to 3880 * traverse entire chain of parent states. This process ends only when all the 3881 * necessary registers/slots and their transitive dependencies are marked as 3882 * precise. 3883 * 3884 * One important and subtle aspect is that precise marks *do not matter* in 3885 * the currently verified state (current state). It is important to understand 3886 * why this is the case. 3887 * 3888 * First, note that current state is the state that is not yet "checkpointed", 3889 * i.e., it is not yet put into env->explored_states, and it has no children 3890 * states as well. It's ephemeral, and can end up either a) being discarded if 3891 * compatible explored state is found at some point or BPF_EXIT instruction is 3892 * reached or b) checkpointed and put into env->explored_states, branching out 3893 * into one or more children states. 3894 * 3895 * In the former case, precise markings in current state are completely 3896 * ignored by state comparison code (see regsafe() for details). Only 3897 * checkpointed ("old") state precise markings are important, and if old 3898 * state's register/slot is precise, regsafe() assumes current state's 3899 * register/slot as precise and checks value ranges exactly and precisely. If 3900 * states turn out to be compatible, current state's necessary precise 3901 * markings and any required parent states' precise markings are enforced 3902 * after the fact with propagate_precision() logic, after the fact. But it's 3903 * important to realize that in this case, even after marking current state 3904 * registers/slots as precise, we immediately discard current state. So what 3905 * actually matters is any of the precise markings propagated into current 3906 * state's parent states, which are always checkpointed (due to b) case above). 3907 * As such, for scenario a) it doesn't matter if current state has precise 3908 * markings set or not. 3909 * 3910 * Now, for the scenario b), checkpointing and forking into child(ren) 3911 * state(s). Note that before current state gets to checkpointing step, any 3912 * processed instruction always assumes precise SCALAR register/slot 3913 * knowledge: if precise value or range is useful to prune jump branch, BPF 3914 * verifier takes this opportunity enthusiastically. Similarly, when 3915 * register's value is used to calculate offset or memory address, exact 3916 * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to 3917 * what we mentioned above about state comparison ignoring precise markings 3918 * during state comparison, BPF verifier ignores and also assumes precise 3919 * markings *at will* during instruction verification process. But as verifier 3920 * assumes precision, it also propagates any precision dependencies across 3921 * parent states, which are not yet finalized, so can be further restricted 3922 * based on new knowledge gained from restrictions enforced by their children 3923 * states. This is so that once those parent states are finalized, i.e., when 3924 * they have no more active children state, state comparison logic in 3925 * is_state_visited() would enforce strict and precise SCALAR ranges, if 3926 * required for correctness. 3927 * 3928 * To build a bit more intuition, note also that once a state is checkpointed, 3929 * the path we took to get to that state is not important. This is crucial 3930 * property for state pruning. When state is checkpointed and finalized at 3931 * some instruction index, it can be correctly and safely used to "short 3932 * circuit" any *compatible* state that reaches exactly the same instruction 3933 * index. I.e., if we jumped to that instruction from a completely different 3934 * code path than original finalized state was derived from, it doesn't 3935 * matter, current state can be discarded because from that instruction 3936 * forward having a compatible state will ensure we will safely reach the 3937 * exit. States describe preconditions for further exploration, but completely 3938 * forget the history of how we got here. 3939 * 3940 * This also means that even if we needed precise SCALAR range to get to 3941 * finalized state, but from that point forward *that same* SCALAR register is 3942 * never used in a precise context (i.e., it's precise value is not needed for 3943 * correctness), it's correct and safe to mark such register as "imprecise" 3944 * (i.e., precise marking set to false). This is what we rely on when we do 3945 * not set precise marking in current state. If no child state requires 3946 * precision for any given SCALAR register, it's safe to dictate that it can 3947 * be imprecise. If any child state does require this register to be precise, 3948 * we'll mark it precise later retroactively during precise markings 3949 * propagation from child state to parent states. 3950 * 3951 * Skipping precise marking setting in current state is a mild version of 3952 * relying on the above observation. But we can utilize this property even 3953 * more aggressively by proactively forgetting any precise marking in the 3954 * current state (which we inherited from the parent state), right before we 3955 * checkpoint it and branch off into new child state. This is done by 3956 * mark_all_scalars_imprecise() to hopefully get more permissive and generic 3957 * finalized states which help in short circuiting more future states. 3958 */ 3959 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno) 3960 { 3961 struct backtrack_state *bt = &env->bt; 3962 struct bpf_verifier_state *st = env->cur_state; 3963 int first_idx = st->first_insn_idx; 3964 int last_idx = env->insn_idx; 3965 int subseq_idx = -1; 3966 struct bpf_func_state *func; 3967 struct bpf_reg_state *reg; 3968 bool skip_first = true; 3969 int i, fr, err; 3970 3971 if (!env->bpf_capable) 3972 return 0; 3973 3974 /* set frame number from which we are starting to backtrack */ 3975 bt_init(bt, env->cur_state->curframe); 3976 3977 /* Do sanity checks against current state of register and/or stack 3978 * slot, but don't set precise flag in current state, as precision 3979 * tracking in the current state is unnecessary. 3980 */ 3981 func = st->frame[bt->frame]; 3982 if (regno >= 0) { 3983 reg = &func->regs[regno]; 3984 if (reg->type != SCALAR_VALUE) { 3985 WARN_ONCE(1, "backtracing misuse"); 3986 return -EFAULT; 3987 } 3988 bt_set_reg(bt, regno); 3989 } 3990 3991 if (bt_empty(bt)) 3992 return 0; 3993 3994 for (;;) { 3995 DECLARE_BITMAP(mask, 64); 3996 u32 history = st->jmp_history_cnt; 3997 3998 if (env->log.level & BPF_LOG_LEVEL2) { 3999 verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n", 4000 bt->frame, last_idx, first_idx, subseq_idx); 4001 } 4002 4003 /* If some register with scalar ID is marked as precise, 4004 * make sure that all registers sharing this ID are also precise. 4005 * This is needed to estimate effect of find_equal_scalars(). 4006 * Do this at the last instruction of each state, 4007 * bpf_reg_state::id fields are valid for these instructions. 4008 * 4009 * Allows to track precision in situation like below: 4010 * 4011 * r2 = unknown value 4012 * ... 4013 * --- state #0 --- 4014 * ... 4015 * r1 = r2 // r1 and r2 now share the same ID 4016 * ... 4017 * --- state #1 {r1.id = A, r2.id = A} --- 4018 * ... 4019 * if (r2 > 10) goto exit; // find_equal_scalars() assigns range to r1 4020 * ... 4021 * --- state #2 {r1.id = A, r2.id = A} --- 4022 * r3 = r10 4023 * r3 += r1 // need to mark both r1 and r2 4024 */ 4025 if (mark_precise_scalar_ids(env, st)) 4026 return -EFAULT; 4027 4028 if (last_idx < 0) { 4029 /* we are at the entry into subprog, which 4030 * is expected for global funcs, but only if 4031 * requested precise registers are R1-R5 4032 * (which are global func's input arguments) 4033 */ 4034 if (st->curframe == 0 && 4035 st->frame[0]->subprogno > 0 && 4036 st->frame[0]->callsite == BPF_MAIN_FUNC && 4037 bt_stack_mask(bt) == 0 && 4038 (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) { 4039 bitmap_from_u64(mask, bt_reg_mask(bt)); 4040 for_each_set_bit(i, mask, 32) { 4041 reg = &st->frame[0]->regs[i]; 4042 if (reg->type != SCALAR_VALUE) { 4043 bt_clear_reg(bt, i); 4044 continue; 4045 } 4046 reg->precise = true; 4047 } 4048 return 0; 4049 } 4050 4051 verbose(env, "BUG backtracking func entry subprog %d reg_mask %x stack_mask %llx\n", 4052 st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt)); 4053 WARN_ONCE(1, "verifier backtracking bug"); 4054 return -EFAULT; 4055 } 4056 4057 for (i = last_idx;;) { 4058 if (skip_first) { 4059 err = 0; 4060 skip_first = false; 4061 } else { 4062 err = backtrack_insn(env, i, subseq_idx, bt); 4063 } 4064 if (err == -ENOTSUPP) { 4065 mark_all_scalars_precise(env, env->cur_state); 4066 bt_reset(bt); 4067 return 0; 4068 } else if (err) { 4069 return err; 4070 } 4071 if (bt_empty(bt)) 4072 /* Found assignment(s) into tracked register in this state. 4073 * Since this state is already marked, just return. 4074 * Nothing to be tracked further in the parent state. 4075 */ 4076 return 0; 4077 if (i == first_idx) 4078 break; 4079 subseq_idx = i; 4080 i = get_prev_insn_idx(st, i, &history); 4081 if (i >= env->prog->len) { 4082 /* This can happen if backtracking reached insn 0 4083 * and there are still reg_mask or stack_mask 4084 * to backtrack. 4085 * It means the backtracking missed the spot where 4086 * particular register was initialized with a constant. 4087 */ 4088 verbose(env, "BUG backtracking idx %d\n", i); 4089 WARN_ONCE(1, "verifier backtracking bug"); 4090 return -EFAULT; 4091 } 4092 } 4093 st = st->parent; 4094 if (!st) 4095 break; 4096 4097 for (fr = bt->frame; fr >= 0; fr--) { 4098 func = st->frame[fr]; 4099 bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr)); 4100 for_each_set_bit(i, mask, 32) { 4101 reg = &func->regs[i]; 4102 if (reg->type != SCALAR_VALUE) { 4103 bt_clear_frame_reg(bt, fr, i); 4104 continue; 4105 } 4106 if (reg->precise) 4107 bt_clear_frame_reg(bt, fr, i); 4108 else 4109 reg->precise = true; 4110 } 4111 4112 bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr)); 4113 for_each_set_bit(i, mask, 64) { 4114 if (i >= func->allocated_stack / BPF_REG_SIZE) { 4115 /* the sequence of instructions: 4116 * 2: (bf) r3 = r10 4117 * 3: (7b) *(u64 *)(r3 -8) = r0 4118 * 4: (79) r4 = *(u64 *)(r10 -8) 4119 * doesn't contain jmps. It's backtracked 4120 * as a single block. 4121 * During backtracking insn 3 is not recognized as 4122 * stack access, so at the end of backtracking 4123 * stack slot fp-8 is still marked in stack_mask. 4124 * However the parent state may not have accessed 4125 * fp-8 and it's "unallocated" stack space. 4126 * In such case fallback to conservative. 4127 */ 4128 mark_all_scalars_precise(env, env->cur_state); 4129 bt_reset(bt); 4130 return 0; 4131 } 4132 4133 if (!is_spilled_scalar_reg(&func->stack[i])) { 4134 bt_clear_frame_slot(bt, fr, i); 4135 continue; 4136 } 4137 reg = &func->stack[i].spilled_ptr; 4138 if (reg->precise) 4139 bt_clear_frame_slot(bt, fr, i); 4140 else 4141 reg->precise = true; 4142 } 4143 if (env->log.level & BPF_LOG_LEVEL2) { 4144 fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, 4145 bt_frame_reg_mask(bt, fr)); 4146 verbose(env, "mark_precise: frame%d: parent state regs=%s ", 4147 fr, env->tmp_str_buf); 4148 fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, 4149 bt_frame_stack_mask(bt, fr)); 4150 verbose(env, "stack=%s: ", env->tmp_str_buf); 4151 print_verifier_state(env, func, true); 4152 } 4153 } 4154 4155 if (bt_empty(bt)) 4156 return 0; 4157 4158 subseq_idx = first_idx; 4159 last_idx = st->last_insn_idx; 4160 first_idx = st->first_insn_idx; 4161 } 4162 4163 /* if we still have requested precise regs or slots, we missed 4164 * something (e.g., stack access through non-r10 register), so 4165 * fallback to marking all precise 4166 */ 4167 if (!bt_empty(bt)) { 4168 mark_all_scalars_precise(env, env->cur_state); 4169 bt_reset(bt); 4170 } 4171 4172 return 0; 4173 } 4174 4175 int mark_chain_precision(struct bpf_verifier_env *env, int regno) 4176 { 4177 return __mark_chain_precision(env, regno); 4178 } 4179 4180 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to 4181 * desired reg and stack masks across all relevant frames 4182 */ 4183 static int mark_chain_precision_batch(struct bpf_verifier_env *env) 4184 { 4185 return __mark_chain_precision(env, -1); 4186 } 4187 4188 static bool is_spillable_regtype(enum bpf_reg_type type) 4189 { 4190 switch (base_type(type)) { 4191 case PTR_TO_MAP_VALUE: 4192 case PTR_TO_STACK: 4193 case PTR_TO_CTX: 4194 case PTR_TO_PACKET: 4195 case PTR_TO_PACKET_META: 4196 case PTR_TO_PACKET_END: 4197 case PTR_TO_FLOW_KEYS: 4198 case CONST_PTR_TO_MAP: 4199 case PTR_TO_SOCKET: 4200 case PTR_TO_SOCK_COMMON: 4201 case PTR_TO_TCP_SOCK: 4202 case PTR_TO_XDP_SOCK: 4203 case PTR_TO_BTF_ID: 4204 case PTR_TO_BUF: 4205 case PTR_TO_MEM: 4206 case PTR_TO_FUNC: 4207 case PTR_TO_MAP_KEY: 4208 return true; 4209 default: 4210 return false; 4211 } 4212 } 4213 4214 /* Does this register contain a constant zero? */ 4215 static bool register_is_null(struct bpf_reg_state *reg) 4216 { 4217 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0); 4218 } 4219 4220 static bool register_is_const(struct bpf_reg_state *reg) 4221 { 4222 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off); 4223 } 4224 4225 static bool __is_scalar_unbounded(struct bpf_reg_state *reg) 4226 { 4227 return tnum_is_unknown(reg->var_off) && 4228 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX && 4229 reg->umin_value == 0 && reg->umax_value == U64_MAX && 4230 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX && 4231 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX; 4232 } 4233 4234 static bool register_is_bounded(struct bpf_reg_state *reg) 4235 { 4236 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg); 4237 } 4238 4239 static bool __is_pointer_value(bool allow_ptr_leaks, 4240 const struct bpf_reg_state *reg) 4241 { 4242 if (allow_ptr_leaks) 4243 return false; 4244 4245 return reg->type != SCALAR_VALUE; 4246 } 4247 4248 /* Copy src state preserving dst->parent and dst->live fields */ 4249 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src) 4250 { 4251 struct bpf_reg_state *parent = dst->parent; 4252 enum bpf_reg_liveness live = dst->live; 4253 4254 *dst = *src; 4255 dst->parent = parent; 4256 dst->live = live; 4257 } 4258 4259 static void save_register_state(struct bpf_func_state *state, 4260 int spi, struct bpf_reg_state *reg, 4261 int size) 4262 { 4263 int i; 4264 4265 copy_register_state(&state->stack[spi].spilled_ptr, reg); 4266 if (size == BPF_REG_SIZE) 4267 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; 4268 4269 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--) 4270 state->stack[spi].slot_type[i - 1] = STACK_SPILL; 4271 4272 /* size < 8 bytes spill */ 4273 for (; i; i--) 4274 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]); 4275 } 4276 4277 static bool is_bpf_st_mem(struct bpf_insn *insn) 4278 { 4279 return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM; 4280 } 4281 4282 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers, 4283 * stack boundary and alignment are checked in check_mem_access() 4284 */ 4285 static int check_stack_write_fixed_off(struct bpf_verifier_env *env, 4286 /* stack frame we're writing to */ 4287 struct bpf_func_state *state, 4288 int off, int size, int value_regno, 4289 int insn_idx) 4290 { 4291 struct bpf_func_state *cur; /* state of the current function */ 4292 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err; 4293 struct bpf_insn *insn = &env->prog->insnsi[insn_idx]; 4294 struct bpf_reg_state *reg = NULL; 4295 u32 dst_reg = insn->dst_reg; 4296 4297 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE)); 4298 if (err) 4299 return err; 4300 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0, 4301 * so it's aligned access and [off, off + size) are within stack limits 4302 */ 4303 if (!env->allow_ptr_leaks && 4304 state->stack[spi].slot_type[0] == STACK_SPILL && 4305 size != BPF_REG_SIZE) { 4306 verbose(env, "attempt to corrupt spilled pointer on stack\n"); 4307 return -EACCES; 4308 } 4309 4310 cur = env->cur_state->frame[env->cur_state->curframe]; 4311 if (value_regno >= 0) 4312 reg = &cur->regs[value_regno]; 4313 if (!env->bypass_spec_v4) { 4314 bool sanitize = reg && is_spillable_regtype(reg->type); 4315 4316 for (i = 0; i < size; i++) { 4317 u8 type = state->stack[spi].slot_type[i]; 4318 4319 if (type != STACK_MISC && type != STACK_ZERO) { 4320 sanitize = true; 4321 break; 4322 } 4323 } 4324 4325 if (sanitize) 4326 env->insn_aux_data[insn_idx].sanitize_stack_spill = true; 4327 } 4328 4329 err = destroy_if_dynptr_stack_slot(env, state, spi); 4330 if (err) 4331 return err; 4332 4333 mark_stack_slot_scratched(env, spi); 4334 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) && 4335 !register_is_null(reg) && env->bpf_capable) { 4336 if (dst_reg != BPF_REG_FP) { 4337 /* The backtracking logic can only recognize explicit 4338 * stack slot address like [fp - 8]. Other spill of 4339 * scalar via different register has to be conservative. 4340 * Backtrack from here and mark all registers as precise 4341 * that contributed into 'reg' being a constant. 4342 */ 4343 err = mark_chain_precision(env, value_regno); 4344 if (err) 4345 return err; 4346 } 4347 save_register_state(state, spi, reg, size); 4348 /* Break the relation on a narrowing spill. */ 4349 if (fls64(reg->umax_value) > BITS_PER_BYTE * size) 4350 state->stack[spi].spilled_ptr.id = 0; 4351 } else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) && 4352 insn->imm != 0 && env->bpf_capable) { 4353 struct bpf_reg_state fake_reg = {}; 4354 4355 __mark_reg_known(&fake_reg, (u32)insn->imm); 4356 fake_reg.type = SCALAR_VALUE; 4357 save_register_state(state, spi, &fake_reg, size); 4358 } else if (reg && is_spillable_regtype(reg->type)) { 4359 /* register containing pointer is being spilled into stack */ 4360 if (size != BPF_REG_SIZE) { 4361 verbose_linfo(env, insn_idx, "; "); 4362 verbose(env, "invalid size of register spill\n"); 4363 return -EACCES; 4364 } 4365 if (state != cur && reg->type == PTR_TO_STACK) { 4366 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n"); 4367 return -EINVAL; 4368 } 4369 save_register_state(state, spi, reg, size); 4370 } else { 4371 u8 type = STACK_MISC; 4372 4373 /* regular write of data into stack destroys any spilled ptr */ 4374 state->stack[spi].spilled_ptr.type = NOT_INIT; 4375 /* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */ 4376 if (is_stack_slot_special(&state->stack[spi])) 4377 for (i = 0; i < BPF_REG_SIZE; i++) 4378 scrub_spilled_slot(&state->stack[spi].slot_type[i]); 4379 4380 /* only mark the slot as written if all 8 bytes were written 4381 * otherwise read propagation may incorrectly stop too soon 4382 * when stack slots are partially written. 4383 * This heuristic means that read propagation will be 4384 * conservative, since it will add reg_live_read marks 4385 * to stack slots all the way to first state when programs 4386 * writes+reads less than 8 bytes 4387 */ 4388 if (size == BPF_REG_SIZE) 4389 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN; 4390 4391 /* when we zero initialize stack slots mark them as such */ 4392 if ((reg && register_is_null(reg)) || 4393 (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) { 4394 /* backtracking doesn't work for STACK_ZERO yet. */ 4395 err = mark_chain_precision(env, value_regno); 4396 if (err) 4397 return err; 4398 type = STACK_ZERO; 4399 } 4400 4401 /* Mark slots affected by this stack write. */ 4402 for (i = 0; i < size; i++) 4403 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] = 4404 type; 4405 } 4406 return 0; 4407 } 4408 4409 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is 4410 * known to contain a variable offset. 4411 * This function checks whether the write is permitted and conservatively 4412 * tracks the effects of the write, considering that each stack slot in the 4413 * dynamic range is potentially written to. 4414 * 4415 * 'off' includes 'regno->off'. 4416 * 'value_regno' can be -1, meaning that an unknown value is being written to 4417 * the stack. 4418 * 4419 * Spilled pointers in range are not marked as written because we don't know 4420 * what's going to be actually written. This means that read propagation for 4421 * future reads cannot be terminated by this write. 4422 * 4423 * For privileged programs, uninitialized stack slots are considered 4424 * initialized by this write (even though we don't know exactly what offsets 4425 * are going to be written to). The idea is that we don't want the verifier to 4426 * reject future reads that access slots written to through variable offsets. 4427 */ 4428 static int check_stack_write_var_off(struct bpf_verifier_env *env, 4429 /* func where register points to */ 4430 struct bpf_func_state *state, 4431 int ptr_regno, int off, int size, 4432 int value_regno, int insn_idx) 4433 { 4434 struct bpf_func_state *cur; /* state of the current function */ 4435 int min_off, max_off; 4436 int i, err; 4437 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL; 4438 struct bpf_insn *insn = &env->prog->insnsi[insn_idx]; 4439 bool writing_zero = false; 4440 /* set if the fact that we're writing a zero is used to let any 4441 * stack slots remain STACK_ZERO 4442 */ 4443 bool zero_used = false; 4444 4445 cur = env->cur_state->frame[env->cur_state->curframe]; 4446 ptr_reg = &cur->regs[ptr_regno]; 4447 min_off = ptr_reg->smin_value + off; 4448 max_off = ptr_reg->smax_value + off + size; 4449 if (value_regno >= 0) 4450 value_reg = &cur->regs[value_regno]; 4451 if ((value_reg && register_is_null(value_reg)) || 4452 (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0)) 4453 writing_zero = true; 4454 4455 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE)); 4456 if (err) 4457 return err; 4458 4459 for (i = min_off; i < max_off; i++) { 4460 int spi; 4461 4462 spi = __get_spi(i); 4463 err = destroy_if_dynptr_stack_slot(env, state, spi); 4464 if (err) 4465 return err; 4466 } 4467 4468 /* Variable offset writes destroy any spilled pointers in range. */ 4469 for (i = min_off; i < max_off; i++) { 4470 u8 new_type, *stype; 4471 int slot, spi; 4472 4473 slot = -i - 1; 4474 spi = slot / BPF_REG_SIZE; 4475 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE]; 4476 mark_stack_slot_scratched(env, spi); 4477 4478 if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) { 4479 /* Reject the write if range we may write to has not 4480 * been initialized beforehand. If we didn't reject 4481 * here, the ptr status would be erased below (even 4482 * though not all slots are actually overwritten), 4483 * possibly opening the door to leaks. 4484 * 4485 * We do however catch STACK_INVALID case below, and 4486 * only allow reading possibly uninitialized memory 4487 * later for CAP_PERFMON, as the write may not happen to 4488 * that slot. 4489 */ 4490 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d", 4491 insn_idx, i); 4492 return -EINVAL; 4493 } 4494 4495 /* Erase all spilled pointers. */ 4496 state->stack[spi].spilled_ptr.type = NOT_INIT; 4497 4498 /* Update the slot type. */ 4499 new_type = STACK_MISC; 4500 if (writing_zero && *stype == STACK_ZERO) { 4501 new_type = STACK_ZERO; 4502 zero_used = true; 4503 } 4504 /* If the slot is STACK_INVALID, we check whether it's OK to 4505 * pretend that it will be initialized by this write. The slot 4506 * might not actually be written to, and so if we mark it as 4507 * initialized future reads might leak uninitialized memory. 4508 * For privileged programs, we will accept such reads to slots 4509 * that may or may not be written because, if we're reject 4510 * them, the error would be too confusing. 4511 */ 4512 if (*stype == STACK_INVALID && !env->allow_uninit_stack) { 4513 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d", 4514 insn_idx, i); 4515 return -EINVAL; 4516 } 4517 *stype = new_type; 4518 } 4519 if (zero_used) { 4520 /* backtracking doesn't work for STACK_ZERO yet. */ 4521 err = mark_chain_precision(env, value_regno); 4522 if (err) 4523 return err; 4524 } 4525 return 0; 4526 } 4527 4528 /* When register 'dst_regno' is assigned some values from stack[min_off, 4529 * max_off), we set the register's type according to the types of the 4530 * respective stack slots. If all the stack values are known to be zeros, then 4531 * so is the destination reg. Otherwise, the register is considered to be 4532 * SCALAR. This function does not deal with register filling; the caller must 4533 * ensure that all spilled registers in the stack range have been marked as 4534 * read. 4535 */ 4536 static void mark_reg_stack_read(struct bpf_verifier_env *env, 4537 /* func where src register points to */ 4538 struct bpf_func_state *ptr_state, 4539 int min_off, int max_off, int dst_regno) 4540 { 4541 struct bpf_verifier_state *vstate = env->cur_state; 4542 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 4543 int i, slot, spi; 4544 u8 *stype; 4545 int zeros = 0; 4546 4547 for (i = min_off; i < max_off; i++) { 4548 slot = -i - 1; 4549 spi = slot / BPF_REG_SIZE; 4550 mark_stack_slot_scratched(env, spi); 4551 stype = ptr_state->stack[spi].slot_type; 4552 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO) 4553 break; 4554 zeros++; 4555 } 4556 if (zeros == max_off - min_off) { 4557 /* any access_size read into register is zero extended, 4558 * so the whole register == const_zero 4559 */ 4560 __mark_reg_const_zero(&state->regs[dst_regno]); 4561 /* backtracking doesn't support STACK_ZERO yet, 4562 * so mark it precise here, so that later 4563 * backtracking can stop here. 4564 * Backtracking may not need this if this register 4565 * doesn't participate in pointer adjustment. 4566 * Forward propagation of precise flag is not 4567 * necessary either. This mark is only to stop 4568 * backtracking. Any register that contributed 4569 * to const 0 was marked precise before spill. 4570 */ 4571 state->regs[dst_regno].precise = true; 4572 } else { 4573 /* have read misc data from the stack */ 4574 mark_reg_unknown(env, state->regs, dst_regno); 4575 } 4576 state->regs[dst_regno].live |= REG_LIVE_WRITTEN; 4577 } 4578 4579 /* Read the stack at 'off' and put the results into the register indicated by 4580 * 'dst_regno'. It handles reg filling if the addressed stack slot is a 4581 * spilled reg. 4582 * 4583 * 'dst_regno' can be -1, meaning that the read value is not going to a 4584 * register. 4585 * 4586 * The access is assumed to be within the current stack bounds. 4587 */ 4588 static int check_stack_read_fixed_off(struct bpf_verifier_env *env, 4589 /* func where src register points to */ 4590 struct bpf_func_state *reg_state, 4591 int off, int size, int dst_regno) 4592 { 4593 struct bpf_verifier_state *vstate = env->cur_state; 4594 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 4595 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE; 4596 struct bpf_reg_state *reg; 4597 u8 *stype, type; 4598 4599 stype = reg_state->stack[spi].slot_type; 4600 reg = ®_state->stack[spi].spilled_ptr; 4601 4602 mark_stack_slot_scratched(env, spi); 4603 4604 if (is_spilled_reg(®_state->stack[spi])) { 4605 u8 spill_size = 1; 4606 4607 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--) 4608 spill_size++; 4609 4610 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) { 4611 if (reg->type != SCALAR_VALUE) { 4612 verbose_linfo(env, env->insn_idx, "; "); 4613 verbose(env, "invalid size of register fill\n"); 4614 return -EACCES; 4615 } 4616 4617 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64); 4618 if (dst_regno < 0) 4619 return 0; 4620 4621 if (!(off % BPF_REG_SIZE) && size == spill_size) { 4622 /* The earlier check_reg_arg() has decided the 4623 * subreg_def for this insn. Save it first. 4624 */ 4625 s32 subreg_def = state->regs[dst_regno].subreg_def; 4626 4627 copy_register_state(&state->regs[dst_regno], reg); 4628 state->regs[dst_regno].subreg_def = subreg_def; 4629 } else { 4630 for (i = 0; i < size; i++) { 4631 type = stype[(slot - i) % BPF_REG_SIZE]; 4632 if (type == STACK_SPILL) 4633 continue; 4634 if (type == STACK_MISC) 4635 continue; 4636 if (type == STACK_INVALID && env->allow_uninit_stack) 4637 continue; 4638 verbose(env, "invalid read from stack off %d+%d size %d\n", 4639 off, i, size); 4640 return -EACCES; 4641 } 4642 mark_reg_unknown(env, state->regs, dst_regno); 4643 } 4644 state->regs[dst_regno].live |= REG_LIVE_WRITTEN; 4645 return 0; 4646 } 4647 4648 if (dst_regno >= 0) { 4649 /* restore register state from stack */ 4650 copy_register_state(&state->regs[dst_regno], reg); 4651 /* mark reg as written since spilled pointer state likely 4652 * has its liveness marks cleared by is_state_visited() 4653 * which resets stack/reg liveness for state transitions 4654 */ 4655 state->regs[dst_regno].live |= REG_LIVE_WRITTEN; 4656 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) { 4657 /* If dst_regno==-1, the caller is asking us whether 4658 * it is acceptable to use this value as a SCALAR_VALUE 4659 * (e.g. for XADD). 4660 * We must not allow unprivileged callers to do that 4661 * with spilled pointers. 4662 */ 4663 verbose(env, "leaking pointer from stack off %d\n", 4664 off); 4665 return -EACCES; 4666 } 4667 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64); 4668 } else { 4669 for (i = 0; i < size; i++) { 4670 type = stype[(slot - i) % BPF_REG_SIZE]; 4671 if (type == STACK_MISC) 4672 continue; 4673 if (type == STACK_ZERO) 4674 continue; 4675 if (type == STACK_INVALID && env->allow_uninit_stack) 4676 continue; 4677 verbose(env, "invalid read from stack off %d+%d size %d\n", 4678 off, i, size); 4679 return -EACCES; 4680 } 4681 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64); 4682 if (dst_regno >= 0) 4683 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno); 4684 } 4685 return 0; 4686 } 4687 4688 enum bpf_access_src { 4689 ACCESS_DIRECT = 1, /* the access is performed by an instruction */ 4690 ACCESS_HELPER = 2, /* the access is performed by a helper */ 4691 }; 4692 4693 static int check_stack_range_initialized(struct bpf_verifier_env *env, 4694 int regno, int off, int access_size, 4695 bool zero_size_allowed, 4696 enum bpf_access_src type, 4697 struct bpf_call_arg_meta *meta); 4698 4699 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno) 4700 { 4701 return cur_regs(env) + regno; 4702 } 4703 4704 /* Read the stack at 'ptr_regno + off' and put the result into the register 4705 * 'dst_regno'. 4706 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'), 4707 * but not its variable offset. 4708 * 'size' is assumed to be <= reg size and the access is assumed to be aligned. 4709 * 4710 * As opposed to check_stack_read_fixed_off, this function doesn't deal with 4711 * filling registers (i.e. reads of spilled register cannot be detected when 4712 * the offset is not fixed). We conservatively mark 'dst_regno' as containing 4713 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable 4714 * offset; for a fixed offset check_stack_read_fixed_off should be used 4715 * instead. 4716 */ 4717 static int check_stack_read_var_off(struct bpf_verifier_env *env, 4718 int ptr_regno, int off, int size, int dst_regno) 4719 { 4720 /* The state of the source register. */ 4721 struct bpf_reg_state *reg = reg_state(env, ptr_regno); 4722 struct bpf_func_state *ptr_state = func(env, reg); 4723 int err; 4724 int min_off, max_off; 4725 4726 /* Note that we pass a NULL meta, so raw access will not be permitted. 4727 */ 4728 err = check_stack_range_initialized(env, ptr_regno, off, size, 4729 false, ACCESS_DIRECT, NULL); 4730 if (err) 4731 return err; 4732 4733 min_off = reg->smin_value + off; 4734 max_off = reg->smax_value + off; 4735 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno); 4736 return 0; 4737 } 4738 4739 /* check_stack_read dispatches to check_stack_read_fixed_off or 4740 * check_stack_read_var_off. 4741 * 4742 * The caller must ensure that the offset falls within the allocated stack 4743 * bounds. 4744 * 4745 * 'dst_regno' is a register which will receive the value from the stack. It 4746 * can be -1, meaning that the read value is not going to a register. 4747 */ 4748 static int check_stack_read(struct bpf_verifier_env *env, 4749 int ptr_regno, int off, int size, 4750 int dst_regno) 4751 { 4752 struct bpf_reg_state *reg = reg_state(env, ptr_regno); 4753 struct bpf_func_state *state = func(env, reg); 4754 int err; 4755 /* Some accesses are only permitted with a static offset. */ 4756 bool var_off = !tnum_is_const(reg->var_off); 4757 4758 /* The offset is required to be static when reads don't go to a 4759 * register, in order to not leak pointers (see 4760 * check_stack_read_fixed_off). 4761 */ 4762 if (dst_regno < 0 && var_off) { 4763 char tn_buf[48]; 4764 4765 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 4766 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n", 4767 tn_buf, off, size); 4768 return -EACCES; 4769 } 4770 /* Variable offset is prohibited for unprivileged mode for simplicity 4771 * since it requires corresponding support in Spectre masking for stack 4772 * ALU. See also retrieve_ptr_limit(). The check in 4773 * check_stack_access_for_ptr_arithmetic() called by 4774 * adjust_ptr_min_max_vals() prevents users from creating stack pointers 4775 * with variable offsets, therefore no check is required here. Further, 4776 * just checking it here would be insufficient as speculative stack 4777 * writes could still lead to unsafe speculative behaviour. 4778 */ 4779 if (!var_off) { 4780 off += reg->var_off.value; 4781 err = check_stack_read_fixed_off(env, state, off, size, 4782 dst_regno); 4783 } else { 4784 /* Variable offset stack reads need more conservative handling 4785 * than fixed offset ones. Note that dst_regno >= 0 on this 4786 * branch. 4787 */ 4788 err = check_stack_read_var_off(env, ptr_regno, off, size, 4789 dst_regno); 4790 } 4791 return err; 4792 } 4793 4794 4795 /* check_stack_write dispatches to check_stack_write_fixed_off or 4796 * check_stack_write_var_off. 4797 * 4798 * 'ptr_regno' is the register used as a pointer into the stack. 4799 * 'off' includes 'ptr_regno->off', but not its variable offset (if any). 4800 * 'value_regno' is the register whose value we're writing to the stack. It can 4801 * be -1, meaning that we're not writing from a register. 4802 * 4803 * The caller must ensure that the offset falls within the maximum stack size. 4804 */ 4805 static int check_stack_write(struct bpf_verifier_env *env, 4806 int ptr_regno, int off, int size, 4807 int value_regno, int insn_idx) 4808 { 4809 struct bpf_reg_state *reg = reg_state(env, ptr_regno); 4810 struct bpf_func_state *state = func(env, reg); 4811 int err; 4812 4813 if (tnum_is_const(reg->var_off)) { 4814 off += reg->var_off.value; 4815 err = check_stack_write_fixed_off(env, state, off, size, 4816 value_regno, insn_idx); 4817 } else { 4818 /* Variable offset stack reads need more conservative handling 4819 * than fixed offset ones. 4820 */ 4821 err = check_stack_write_var_off(env, state, 4822 ptr_regno, off, size, 4823 value_regno, insn_idx); 4824 } 4825 return err; 4826 } 4827 4828 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno, 4829 int off, int size, enum bpf_access_type type) 4830 { 4831 struct bpf_reg_state *regs = cur_regs(env); 4832 struct bpf_map *map = regs[regno].map_ptr; 4833 u32 cap = bpf_map_flags_to_cap(map); 4834 4835 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) { 4836 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n", 4837 map->value_size, off, size); 4838 return -EACCES; 4839 } 4840 4841 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) { 4842 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n", 4843 map->value_size, off, size); 4844 return -EACCES; 4845 } 4846 4847 return 0; 4848 } 4849 4850 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */ 4851 static int __check_mem_access(struct bpf_verifier_env *env, int regno, 4852 int off, int size, u32 mem_size, 4853 bool zero_size_allowed) 4854 { 4855 bool size_ok = size > 0 || (size == 0 && zero_size_allowed); 4856 struct bpf_reg_state *reg; 4857 4858 if (off >= 0 && size_ok && (u64)off + size <= mem_size) 4859 return 0; 4860 4861 reg = &cur_regs(env)[regno]; 4862 switch (reg->type) { 4863 case PTR_TO_MAP_KEY: 4864 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n", 4865 mem_size, off, size); 4866 break; 4867 case PTR_TO_MAP_VALUE: 4868 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n", 4869 mem_size, off, size); 4870 break; 4871 case PTR_TO_PACKET: 4872 case PTR_TO_PACKET_META: 4873 case PTR_TO_PACKET_END: 4874 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n", 4875 off, size, regno, reg->id, off, mem_size); 4876 break; 4877 case PTR_TO_MEM: 4878 default: 4879 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n", 4880 mem_size, off, size); 4881 } 4882 4883 return -EACCES; 4884 } 4885 4886 /* check read/write into a memory region with possible variable offset */ 4887 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno, 4888 int off, int size, u32 mem_size, 4889 bool zero_size_allowed) 4890 { 4891 struct bpf_verifier_state *vstate = env->cur_state; 4892 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 4893 struct bpf_reg_state *reg = &state->regs[regno]; 4894 int err; 4895 4896 /* We may have adjusted the register pointing to memory region, so we 4897 * need to try adding each of min_value and max_value to off 4898 * to make sure our theoretical access will be safe. 4899 * 4900 * The minimum value is only important with signed 4901 * comparisons where we can't assume the floor of a 4902 * value is 0. If we are using signed variables for our 4903 * index'es we need to make sure that whatever we use 4904 * will have a set floor within our range. 4905 */ 4906 if (reg->smin_value < 0 && 4907 (reg->smin_value == S64_MIN || 4908 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) || 4909 reg->smin_value + off < 0)) { 4910 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n", 4911 regno); 4912 return -EACCES; 4913 } 4914 err = __check_mem_access(env, regno, reg->smin_value + off, size, 4915 mem_size, zero_size_allowed); 4916 if (err) { 4917 verbose(env, "R%d min value is outside of the allowed memory range\n", 4918 regno); 4919 return err; 4920 } 4921 4922 /* If we haven't set a max value then we need to bail since we can't be 4923 * sure we won't do bad things. 4924 * If reg->umax_value + off could overflow, treat that as unbounded too. 4925 */ 4926 if (reg->umax_value >= BPF_MAX_VAR_OFF) { 4927 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n", 4928 regno); 4929 return -EACCES; 4930 } 4931 err = __check_mem_access(env, regno, reg->umax_value + off, size, 4932 mem_size, zero_size_allowed); 4933 if (err) { 4934 verbose(env, "R%d max value is outside of the allowed memory range\n", 4935 regno); 4936 return err; 4937 } 4938 4939 return 0; 4940 } 4941 4942 static int __check_ptr_off_reg(struct bpf_verifier_env *env, 4943 const struct bpf_reg_state *reg, int regno, 4944 bool fixed_off_ok) 4945 { 4946 /* Access to this pointer-typed register or passing it to a helper 4947 * is only allowed in its original, unmodified form. 4948 */ 4949 4950 if (reg->off < 0) { 4951 verbose(env, "negative offset %s ptr R%d off=%d disallowed\n", 4952 reg_type_str(env, reg->type), regno, reg->off); 4953 return -EACCES; 4954 } 4955 4956 if (!fixed_off_ok && reg->off) { 4957 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n", 4958 reg_type_str(env, reg->type), regno, reg->off); 4959 return -EACCES; 4960 } 4961 4962 if (!tnum_is_const(reg->var_off) || reg->var_off.value) { 4963 char tn_buf[48]; 4964 4965 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 4966 verbose(env, "variable %s access var_off=%s disallowed\n", 4967 reg_type_str(env, reg->type), tn_buf); 4968 return -EACCES; 4969 } 4970 4971 return 0; 4972 } 4973 4974 int check_ptr_off_reg(struct bpf_verifier_env *env, 4975 const struct bpf_reg_state *reg, int regno) 4976 { 4977 return __check_ptr_off_reg(env, reg, regno, false); 4978 } 4979 4980 static int map_kptr_match_type(struct bpf_verifier_env *env, 4981 struct btf_field *kptr_field, 4982 struct bpf_reg_state *reg, u32 regno) 4983 { 4984 const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id); 4985 int perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU; 4986 const char *reg_name = ""; 4987 4988 /* Only unreferenced case accepts untrusted pointers */ 4989 if (kptr_field->type == BPF_KPTR_UNREF) 4990 perm_flags |= PTR_UNTRUSTED; 4991 4992 if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags)) 4993 goto bad_type; 4994 4995 if (!btf_is_kernel(reg->btf)) { 4996 verbose(env, "R%d must point to kernel BTF\n", regno); 4997 return -EINVAL; 4998 } 4999 /* We need to verify reg->type and reg->btf, before accessing reg->btf */ 5000 reg_name = btf_type_name(reg->btf, reg->btf_id); 5001 5002 /* For ref_ptr case, release function check should ensure we get one 5003 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the 5004 * normal store of unreferenced kptr, we must ensure var_off is zero. 5005 * Since ref_ptr cannot be accessed directly by BPF insns, checks for 5006 * reg->off and reg->ref_obj_id are not needed here. 5007 */ 5008 if (__check_ptr_off_reg(env, reg, regno, true)) 5009 return -EACCES; 5010 5011 /* A full type match is needed, as BTF can be vmlinux or module BTF, and 5012 * we also need to take into account the reg->off. 5013 * 5014 * We want to support cases like: 5015 * 5016 * struct foo { 5017 * struct bar br; 5018 * struct baz bz; 5019 * }; 5020 * 5021 * struct foo *v; 5022 * v = func(); // PTR_TO_BTF_ID 5023 * val->foo = v; // reg->off is zero, btf and btf_id match type 5024 * val->bar = &v->br; // reg->off is still zero, but we need to retry with 5025 * // first member type of struct after comparison fails 5026 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked 5027 * // to match type 5028 * 5029 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off 5030 * is zero. We must also ensure that btf_struct_ids_match does not walk 5031 * the struct to match type against first member of struct, i.e. reject 5032 * second case from above. Hence, when type is BPF_KPTR_REF, we set 5033 * strict mode to true for type match. 5034 */ 5035 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off, 5036 kptr_field->kptr.btf, kptr_field->kptr.btf_id, 5037 kptr_field->type == BPF_KPTR_REF)) 5038 goto bad_type; 5039 return 0; 5040 bad_type: 5041 verbose(env, "invalid kptr access, R%d type=%s%s ", regno, 5042 reg_type_str(env, reg->type), reg_name); 5043 verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name); 5044 if (kptr_field->type == BPF_KPTR_UNREF) 5045 verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED), 5046 targ_name); 5047 else 5048 verbose(env, "\n"); 5049 return -EINVAL; 5050 } 5051 5052 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock() 5053 * can dereference RCU protected pointers and result is PTR_TRUSTED. 5054 */ 5055 static bool in_rcu_cs(struct bpf_verifier_env *env) 5056 { 5057 return env->cur_state->active_rcu_lock || !env->prog->aux->sleepable; 5058 } 5059 5060 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */ 5061 BTF_SET_START(rcu_protected_types) 5062 BTF_ID(struct, prog_test_ref_kfunc) 5063 BTF_ID(struct, cgroup) 5064 BTF_ID(struct, bpf_cpumask) 5065 BTF_ID(struct, task_struct) 5066 BTF_SET_END(rcu_protected_types) 5067 5068 static bool rcu_protected_object(const struct btf *btf, u32 btf_id) 5069 { 5070 if (!btf_is_kernel(btf)) 5071 return false; 5072 return btf_id_set_contains(&rcu_protected_types, btf_id); 5073 } 5074 5075 static bool rcu_safe_kptr(const struct btf_field *field) 5076 { 5077 const struct btf_field_kptr *kptr = &field->kptr; 5078 5079 return field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id); 5080 } 5081 5082 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno, 5083 int value_regno, int insn_idx, 5084 struct btf_field *kptr_field) 5085 { 5086 struct bpf_insn *insn = &env->prog->insnsi[insn_idx]; 5087 int class = BPF_CLASS(insn->code); 5088 struct bpf_reg_state *val_reg; 5089 5090 /* Things we already checked for in check_map_access and caller: 5091 * - Reject cases where variable offset may touch kptr 5092 * - size of access (must be BPF_DW) 5093 * - tnum_is_const(reg->var_off) 5094 * - kptr_field->offset == off + reg->var_off.value 5095 */ 5096 /* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */ 5097 if (BPF_MODE(insn->code) != BPF_MEM) { 5098 verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n"); 5099 return -EACCES; 5100 } 5101 5102 /* We only allow loading referenced kptr, since it will be marked as 5103 * untrusted, similar to unreferenced kptr. 5104 */ 5105 if (class != BPF_LDX && kptr_field->type == BPF_KPTR_REF) { 5106 verbose(env, "store to referenced kptr disallowed\n"); 5107 return -EACCES; 5108 } 5109 5110 if (class == BPF_LDX) { 5111 val_reg = reg_state(env, value_regno); 5112 /* We can simply mark the value_regno receiving the pointer 5113 * value from map as PTR_TO_BTF_ID, with the correct type. 5114 */ 5115 mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID, kptr_field->kptr.btf, 5116 kptr_field->kptr.btf_id, 5117 rcu_safe_kptr(kptr_field) && in_rcu_cs(env) ? 5118 PTR_MAYBE_NULL | MEM_RCU : 5119 PTR_MAYBE_NULL | PTR_UNTRUSTED); 5120 /* For mark_ptr_or_null_reg */ 5121 val_reg->id = ++env->id_gen; 5122 } else if (class == BPF_STX) { 5123 val_reg = reg_state(env, value_regno); 5124 if (!register_is_null(val_reg) && 5125 map_kptr_match_type(env, kptr_field, val_reg, value_regno)) 5126 return -EACCES; 5127 } else if (class == BPF_ST) { 5128 if (insn->imm) { 5129 verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n", 5130 kptr_field->offset); 5131 return -EACCES; 5132 } 5133 } else { 5134 verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n"); 5135 return -EACCES; 5136 } 5137 return 0; 5138 } 5139 5140 /* check read/write into a map element with possible variable offset */ 5141 static int check_map_access(struct bpf_verifier_env *env, u32 regno, 5142 int off, int size, bool zero_size_allowed, 5143 enum bpf_access_src src) 5144 { 5145 struct bpf_verifier_state *vstate = env->cur_state; 5146 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 5147 struct bpf_reg_state *reg = &state->regs[regno]; 5148 struct bpf_map *map = reg->map_ptr; 5149 struct btf_record *rec; 5150 int err, i; 5151 5152 err = check_mem_region_access(env, regno, off, size, map->value_size, 5153 zero_size_allowed); 5154 if (err) 5155 return err; 5156 5157 if (IS_ERR_OR_NULL(map->record)) 5158 return 0; 5159 rec = map->record; 5160 for (i = 0; i < rec->cnt; i++) { 5161 struct btf_field *field = &rec->fields[i]; 5162 u32 p = field->offset; 5163 5164 /* If any part of a field can be touched by load/store, reject 5165 * this program. To check that [x1, x2) overlaps with [y1, y2), 5166 * it is sufficient to check x1 < y2 && y1 < x2. 5167 */ 5168 if (reg->smin_value + off < p + btf_field_type_size(field->type) && 5169 p < reg->umax_value + off + size) { 5170 switch (field->type) { 5171 case BPF_KPTR_UNREF: 5172 case BPF_KPTR_REF: 5173 if (src != ACCESS_DIRECT) { 5174 verbose(env, "kptr cannot be accessed indirectly by helper\n"); 5175 return -EACCES; 5176 } 5177 if (!tnum_is_const(reg->var_off)) { 5178 verbose(env, "kptr access cannot have variable offset\n"); 5179 return -EACCES; 5180 } 5181 if (p != off + reg->var_off.value) { 5182 verbose(env, "kptr access misaligned expected=%u off=%llu\n", 5183 p, off + reg->var_off.value); 5184 return -EACCES; 5185 } 5186 if (size != bpf_size_to_bytes(BPF_DW)) { 5187 verbose(env, "kptr access size must be BPF_DW\n"); 5188 return -EACCES; 5189 } 5190 break; 5191 default: 5192 verbose(env, "%s cannot be accessed directly by load/store\n", 5193 btf_field_type_name(field->type)); 5194 return -EACCES; 5195 } 5196 } 5197 } 5198 return 0; 5199 } 5200 5201 #define MAX_PACKET_OFF 0xffff 5202 5203 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env, 5204 const struct bpf_call_arg_meta *meta, 5205 enum bpf_access_type t) 5206 { 5207 enum bpf_prog_type prog_type = resolve_prog_type(env->prog); 5208 5209 switch (prog_type) { 5210 /* Program types only with direct read access go here! */ 5211 case BPF_PROG_TYPE_LWT_IN: 5212 case BPF_PROG_TYPE_LWT_OUT: 5213 case BPF_PROG_TYPE_LWT_SEG6LOCAL: 5214 case BPF_PROG_TYPE_SK_REUSEPORT: 5215 case BPF_PROG_TYPE_FLOW_DISSECTOR: 5216 case BPF_PROG_TYPE_CGROUP_SKB: 5217 if (t == BPF_WRITE) 5218 return false; 5219 fallthrough; 5220 5221 /* Program types with direct read + write access go here! */ 5222 case BPF_PROG_TYPE_SCHED_CLS: 5223 case BPF_PROG_TYPE_SCHED_ACT: 5224 case BPF_PROG_TYPE_XDP: 5225 case BPF_PROG_TYPE_LWT_XMIT: 5226 case BPF_PROG_TYPE_SK_SKB: 5227 case BPF_PROG_TYPE_SK_MSG: 5228 if (meta) 5229 return meta->pkt_access; 5230 5231 env->seen_direct_write = true; 5232 return true; 5233 5234 case BPF_PROG_TYPE_CGROUP_SOCKOPT: 5235 if (t == BPF_WRITE) 5236 env->seen_direct_write = true; 5237 5238 return true; 5239 5240 default: 5241 return false; 5242 } 5243 } 5244 5245 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off, 5246 int size, bool zero_size_allowed) 5247 { 5248 struct bpf_reg_state *regs = cur_regs(env); 5249 struct bpf_reg_state *reg = ®s[regno]; 5250 int err; 5251 5252 /* We may have added a variable offset to the packet pointer; but any 5253 * reg->range we have comes after that. We are only checking the fixed 5254 * offset. 5255 */ 5256 5257 /* We don't allow negative numbers, because we aren't tracking enough 5258 * detail to prove they're safe. 5259 */ 5260 if (reg->smin_value < 0) { 5261 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n", 5262 regno); 5263 return -EACCES; 5264 } 5265 5266 err = reg->range < 0 ? -EINVAL : 5267 __check_mem_access(env, regno, off, size, reg->range, 5268 zero_size_allowed); 5269 if (err) { 5270 verbose(env, "R%d offset is outside of the packet\n", regno); 5271 return err; 5272 } 5273 5274 /* __check_mem_access has made sure "off + size - 1" is within u16. 5275 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff, 5276 * otherwise find_good_pkt_pointers would have refused to set range info 5277 * that __check_mem_access would have rejected this pkt access. 5278 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32. 5279 */ 5280 env->prog->aux->max_pkt_offset = 5281 max_t(u32, env->prog->aux->max_pkt_offset, 5282 off + reg->umax_value + size - 1); 5283 5284 return err; 5285 } 5286 5287 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */ 5288 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size, 5289 enum bpf_access_type t, enum bpf_reg_type *reg_type, 5290 struct btf **btf, u32 *btf_id) 5291 { 5292 struct bpf_insn_access_aux info = { 5293 .reg_type = *reg_type, 5294 .log = &env->log, 5295 }; 5296 5297 if (env->ops->is_valid_access && 5298 env->ops->is_valid_access(off, size, t, env->prog, &info)) { 5299 /* A non zero info.ctx_field_size indicates that this field is a 5300 * candidate for later verifier transformation to load the whole 5301 * field and then apply a mask when accessed with a narrower 5302 * access than actual ctx access size. A zero info.ctx_field_size 5303 * will only allow for whole field access and rejects any other 5304 * type of narrower access. 5305 */ 5306 *reg_type = info.reg_type; 5307 5308 if (base_type(*reg_type) == PTR_TO_BTF_ID) { 5309 *btf = info.btf; 5310 *btf_id = info.btf_id; 5311 } else { 5312 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size; 5313 } 5314 /* remember the offset of last byte accessed in ctx */ 5315 if (env->prog->aux->max_ctx_offset < off + size) 5316 env->prog->aux->max_ctx_offset = off + size; 5317 return 0; 5318 } 5319 5320 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size); 5321 return -EACCES; 5322 } 5323 5324 static int check_flow_keys_access(struct bpf_verifier_env *env, int off, 5325 int size) 5326 { 5327 if (size < 0 || off < 0 || 5328 (u64)off + size > sizeof(struct bpf_flow_keys)) { 5329 verbose(env, "invalid access to flow keys off=%d size=%d\n", 5330 off, size); 5331 return -EACCES; 5332 } 5333 return 0; 5334 } 5335 5336 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx, 5337 u32 regno, int off, int size, 5338 enum bpf_access_type t) 5339 { 5340 struct bpf_reg_state *regs = cur_regs(env); 5341 struct bpf_reg_state *reg = ®s[regno]; 5342 struct bpf_insn_access_aux info = {}; 5343 bool valid; 5344 5345 if (reg->smin_value < 0) { 5346 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n", 5347 regno); 5348 return -EACCES; 5349 } 5350 5351 switch (reg->type) { 5352 case PTR_TO_SOCK_COMMON: 5353 valid = bpf_sock_common_is_valid_access(off, size, t, &info); 5354 break; 5355 case PTR_TO_SOCKET: 5356 valid = bpf_sock_is_valid_access(off, size, t, &info); 5357 break; 5358 case PTR_TO_TCP_SOCK: 5359 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info); 5360 break; 5361 case PTR_TO_XDP_SOCK: 5362 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info); 5363 break; 5364 default: 5365 valid = false; 5366 } 5367 5368 5369 if (valid) { 5370 env->insn_aux_data[insn_idx].ctx_field_size = 5371 info.ctx_field_size; 5372 return 0; 5373 } 5374 5375 verbose(env, "R%d invalid %s access off=%d size=%d\n", 5376 regno, reg_type_str(env, reg->type), off, size); 5377 5378 return -EACCES; 5379 } 5380 5381 static bool is_pointer_value(struct bpf_verifier_env *env, int regno) 5382 { 5383 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno)); 5384 } 5385 5386 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno) 5387 { 5388 const struct bpf_reg_state *reg = reg_state(env, regno); 5389 5390 return reg->type == PTR_TO_CTX; 5391 } 5392 5393 static bool is_sk_reg(struct bpf_verifier_env *env, int regno) 5394 { 5395 const struct bpf_reg_state *reg = reg_state(env, regno); 5396 5397 return type_is_sk_pointer(reg->type); 5398 } 5399 5400 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno) 5401 { 5402 const struct bpf_reg_state *reg = reg_state(env, regno); 5403 5404 return type_is_pkt_pointer(reg->type); 5405 } 5406 5407 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno) 5408 { 5409 const struct bpf_reg_state *reg = reg_state(env, regno); 5410 5411 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */ 5412 return reg->type == PTR_TO_FLOW_KEYS; 5413 } 5414 5415 static bool is_trusted_reg(const struct bpf_reg_state *reg) 5416 { 5417 /* A referenced register is always trusted. */ 5418 if (reg->ref_obj_id) 5419 return true; 5420 5421 /* If a register is not referenced, it is trusted if it has the 5422 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the 5423 * other type modifiers may be safe, but we elect to take an opt-in 5424 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are 5425 * not. 5426 * 5427 * Eventually, we should make PTR_TRUSTED the single source of truth 5428 * for whether a register is trusted. 5429 */ 5430 return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS && 5431 !bpf_type_has_unsafe_modifiers(reg->type); 5432 } 5433 5434 static bool is_rcu_reg(const struct bpf_reg_state *reg) 5435 { 5436 return reg->type & MEM_RCU; 5437 } 5438 5439 static void clear_trusted_flags(enum bpf_type_flag *flag) 5440 { 5441 *flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU); 5442 } 5443 5444 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env, 5445 const struct bpf_reg_state *reg, 5446 int off, int size, bool strict) 5447 { 5448 struct tnum reg_off; 5449 int ip_align; 5450 5451 /* Byte size accesses are always allowed. */ 5452 if (!strict || size == 1) 5453 return 0; 5454 5455 /* For platforms that do not have a Kconfig enabling 5456 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of 5457 * NET_IP_ALIGN is universally set to '2'. And on platforms 5458 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get 5459 * to this code only in strict mode where we want to emulate 5460 * the NET_IP_ALIGN==2 checking. Therefore use an 5461 * unconditional IP align value of '2'. 5462 */ 5463 ip_align = 2; 5464 5465 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off)); 5466 if (!tnum_is_aligned(reg_off, size)) { 5467 char tn_buf[48]; 5468 5469 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 5470 verbose(env, 5471 "misaligned packet access off %d+%s+%d+%d size %d\n", 5472 ip_align, tn_buf, reg->off, off, size); 5473 return -EACCES; 5474 } 5475 5476 return 0; 5477 } 5478 5479 static int check_generic_ptr_alignment(struct bpf_verifier_env *env, 5480 const struct bpf_reg_state *reg, 5481 const char *pointer_desc, 5482 int off, int size, bool strict) 5483 { 5484 struct tnum reg_off; 5485 5486 /* Byte size accesses are always allowed. */ 5487 if (!strict || size == 1) 5488 return 0; 5489 5490 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off)); 5491 if (!tnum_is_aligned(reg_off, size)) { 5492 char tn_buf[48]; 5493 5494 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 5495 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n", 5496 pointer_desc, tn_buf, reg->off, off, size); 5497 return -EACCES; 5498 } 5499 5500 return 0; 5501 } 5502 5503 static int check_ptr_alignment(struct bpf_verifier_env *env, 5504 const struct bpf_reg_state *reg, int off, 5505 int size, bool strict_alignment_once) 5506 { 5507 bool strict = env->strict_alignment || strict_alignment_once; 5508 const char *pointer_desc = ""; 5509 5510 switch (reg->type) { 5511 case PTR_TO_PACKET: 5512 case PTR_TO_PACKET_META: 5513 /* Special case, because of NET_IP_ALIGN. Given metadata sits 5514 * right in front, treat it the very same way. 5515 */ 5516 return check_pkt_ptr_alignment(env, reg, off, size, strict); 5517 case PTR_TO_FLOW_KEYS: 5518 pointer_desc = "flow keys "; 5519 break; 5520 case PTR_TO_MAP_KEY: 5521 pointer_desc = "key "; 5522 break; 5523 case PTR_TO_MAP_VALUE: 5524 pointer_desc = "value "; 5525 break; 5526 case PTR_TO_CTX: 5527 pointer_desc = "context "; 5528 break; 5529 case PTR_TO_STACK: 5530 pointer_desc = "stack "; 5531 /* The stack spill tracking logic in check_stack_write_fixed_off() 5532 * and check_stack_read_fixed_off() relies on stack accesses being 5533 * aligned. 5534 */ 5535 strict = true; 5536 break; 5537 case PTR_TO_SOCKET: 5538 pointer_desc = "sock "; 5539 break; 5540 case PTR_TO_SOCK_COMMON: 5541 pointer_desc = "sock_common "; 5542 break; 5543 case PTR_TO_TCP_SOCK: 5544 pointer_desc = "tcp_sock "; 5545 break; 5546 case PTR_TO_XDP_SOCK: 5547 pointer_desc = "xdp_sock "; 5548 break; 5549 default: 5550 break; 5551 } 5552 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size, 5553 strict); 5554 } 5555 5556 static int update_stack_depth(struct bpf_verifier_env *env, 5557 const struct bpf_func_state *func, 5558 int off) 5559 { 5560 u16 stack = env->subprog_info[func->subprogno].stack_depth; 5561 5562 if (stack >= -off) 5563 return 0; 5564 5565 /* update known max for given subprogram */ 5566 env->subprog_info[func->subprogno].stack_depth = -off; 5567 return 0; 5568 } 5569 5570 /* starting from main bpf function walk all instructions of the function 5571 * and recursively walk all callees that given function can call. 5572 * Ignore jump and exit insns. 5573 * Since recursion is prevented by check_cfg() this algorithm 5574 * only needs a local stack of MAX_CALL_FRAMES to remember callsites 5575 */ 5576 static int check_max_stack_depth(struct bpf_verifier_env *env) 5577 { 5578 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end; 5579 struct bpf_subprog_info *subprog = env->subprog_info; 5580 struct bpf_insn *insn = env->prog->insnsi; 5581 bool tail_call_reachable = false; 5582 int ret_insn[MAX_CALL_FRAMES]; 5583 int ret_prog[MAX_CALL_FRAMES]; 5584 int j; 5585 5586 process_func: 5587 /* protect against potential stack overflow that might happen when 5588 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack 5589 * depth for such case down to 256 so that the worst case scenario 5590 * would result in 8k stack size (32 which is tailcall limit * 256 = 5591 * 8k). 5592 * 5593 * To get the idea what might happen, see an example: 5594 * func1 -> sub rsp, 128 5595 * subfunc1 -> sub rsp, 256 5596 * tailcall1 -> add rsp, 256 5597 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320) 5598 * subfunc2 -> sub rsp, 64 5599 * subfunc22 -> sub rsp, 128 5600 * tailcall2 -> add rsp, 128 5601 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416) 5602 * 5603 * tailcall will unwind the current stack frame but it will not get rid 5604 * of caller's stack as shown on the example above. 5605 */ 5606 if (idx && subprog[idx].has_tail_call && depth >= 256) { 5607 verbose(env, 5608 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n", 5609 depth); 5610 return -EACCES; 5611 } 5612 /* round up to 32-bytes, since this is granularity 5613 * of interpreter stack size 5614 */ 5615 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32); 5616 if (depth > MAX_BPF_STACK) { 5617 verbose(env, "combined stack size of %d calls is %d. Too large\n", 5618 frame + 1, depth); 5619 return -EACCES; 5620 } 5621 continue_func: 5622 subprog_end = subprog[idx + 1].start; 5623 for (; i < subprog_end; i++) { 5624 int next_insn; 5625 5626 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i)) 5627 continue; 5628 /* remember insn and function to return to */ 5629 ret_insn[frame] = i + 1; 5630 ret_prog[frame] = idx; 5631 5632 /* find the callee */ 5633 next_insn = i + insn[i].imm + 1; 5634 idx = find_subprog(env, next_insn); 5635 if (idx < 0) { 5636 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n", 5637 next_insn); 5638 return -EFAULT; 5639 } 5640 if (subprog[idx].is_async_cb) { 5641 if (subprog[idx].has_tail_call) { 5642 verbose(env, "verifier bug. subprog has tail_call and async cb\n"); 5643 return -EFAULT; 5644 } 5645 /* async callbacks don't increase bpf prog stack size unless called directly */ 5646 if (!bpf_pseudo_call(insn + i)) 5647 continue; 5648 } 5649 i = next_insn; 5650 5651 if (subprog[idx].has_tail_call) 5652 tail_call_reachable = true; 5653 5654 frame++; 5655 if (frame >= MAX_CALL_FRAMES) { 5656 verbose(env, "the call stack of %d frames is too deep !\n", 5657 frame); 5658 return -E2BIG; 5659 } 5660 goto process_func; 5661 } 5662 /* if tail call got detected across bpf2bpf calls then mark each of the 5663 * currently present subprog frames as tail call reachable subprogs; 5664 * this info will be utilized by JIT so that we will be preserving the 5665 * tail call counter throughout bpf2bpf calls combined with tailcalls 5666 */ 5667 if (tail_call_reachable) 5668 for (j = 0; j < frame; j++) 5669 subprog[ret_prog[j]].tail_call_reachable = true; 5670 if (subprog[0].tail_call_reachable) 5671 env->prog->aux->tail_call_reachable = true; 5672 5673 /* end of for() loop means the last insn of the 'subprog' 5674 * was reached. Doesn't matter whether it was JA or EXIT 5675 */ 5676 if (frame == 0) 5677 return 0; 5678 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32); 5679 frame--; 5680 i = ret_insn[frame]; 5681 idx = ret_prog[frame]; 5682 goto continue_func; 5683 } 5684 5685 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 5686 static int get_callee_stack_depth(struct bpf_verifier_env *env, 5687 const struct bpf_insn *insn, int idx) 5688 { 5689 int start = idx + insn->imm + 1, subprog; 5690 5691 subprog = find_subprog(env, start); 5692 if (subprog < 0) { 5693 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n", 5694 start); 5695 return -EFAULT; 5696 } 5697 return env->subprog_info[subprog].stack_depth; 5698 } 5699 #endif 5700 5701 static int __check_buffer_access(struct bpf_verifier_env *env, 5702 const char *buf_info, 5703 const struct bpf_reg_state *reg, 5704 int regno, int off, int size) 5705 { 5706 if (off < 0) { 5707 verbose(env, 5708 "R%d invalid %s buffer access: off=%d, size=%d\n", 5709 regno, buf_info, off, size); 5710 return -EACCES; 5711 } 5712 if (!tnum_is_const(reg->var_off) || reg->var_off.value) { 5713 char tn_buf[48]; 5714 5715 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 5716 verbose(env, 5717 "R%d invalid variable buffer offset: off=%d, var_off=%s\n", 5718 regno, off, tn_buf); 5719 return -EACCES; 5720 } 5721 5722 return 0; 5723 } 5724 5725 static int check_tp_buffer_access(struct bpf_verifier_env *env, 5726 const struct bpf_reg_state *reg, 5727 int regno, int off, int size) 5728 { 5729 int err; 5730 5731 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size); 5732 if (err) 5733 return err; 5734 5735 if (off + size > env->prog->aux->max_tp_access) 5736 env->prog->aux->max_tp_access = off + size; 5737 5738 return 0; 5739 } 5740 5741 static int check_buffer_access(struct bpf_verifier_env *env, 5742 const struct bpf_reg_state *reg, 5743 int regno, int off, int size, 5744 bool zero_size_allowed, 5745 u32 *max_access) 5746 { 5747 const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr"; 5748 int err; 5749 5750 err = __check_buffer_access(env, buf_info, reg, regno, off, size); 5751 if (err) 5752 return err; 5753 5754 if (off + size > *max_access) 5755 *max_access = off + size; 5756 5757 return 0; 5758 } 5759 5760 /* BPF architecture zero extends alu32 ops into 64-bit registesr */ 5761 static void zext_32_to_64(struct bpf_reg_state *reg) 5762 { 5763 reg->var_off = tnum_subreg(reg->var_off); 5764 __reg_assign_32_into_64(reg); 5765 } 5766 5767 /* truncate register to smaller size (in bytes) 5768 * must be called with size < BPF_REG_SIZE 5769 */ 5770 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size) 5771 { 5772 u64 mask; 5773 5774 /* clear high bits in bit representation */ 5775 reg->var_off = tnum_cast(reg->var_off, size); 5776 5777 /* fix arithmetic bounds */ 5778 mask = ((u64)1 << (size * 8)) - 1; 5779 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) { 5780 reg->umin_value &= mask; 5781 reg->umax_value &= mask; 5782 } else { 5783 reg->umin_value = 0; 5784 reg->umax_value = mask; 5785 } 5786 reg->smin_value = reg->umin_value; 5787 reg->smax_value = reg->umax_value; 5788 5789 /* If size is smaller than 32bit register the 32bit register 5790 * values are also truncated so we push 64-bit bounds into 5791 * 32-bit bounds. Above were truncated < 32-bits already. 5792 */ 5793 if (size >= 4) 5794 return; 5795 __reg_combine_64_into_32(reg); 5796 } 5797 5798 static bool bpf_map_is_rdonly(const struct bpf_map *map) 5799 { 5800 /* A map is considered read-only if the following condition are true: 5801 * 5802 * 1) BPF program side cannot change any of the map content. The 5803 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map 5804 * and was set at map creation time. 5805 * 2) The map value(s) have been initialized from user space by a 5806 * loader and then "frozen", such that no new map update/delete 5807 * operations from syscall side are possible for the rest of 5808 * the map's lifetime from that point onwards. 5809 * 3) Any parallel/pending map update/delete operations from syscall 5810 * side have been completed. Only after that point, it's safe to 5811 * assume that map value(s) are immutable. 5812 */ 5813 return (map->map_flags & BPF_F_RDONLY_PROG) && 5814 READ_ONCE(map->frozen) && 5815 !bpf_map_write_active(map); 5816 } 5817 5818 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val) 5819 { 5820 void *ptr; 5821 u64 addr; 5822 int err; 5823 5824 err = map->ops->map_direct_value_addr(map, &addr, off); 5825 if (err) 5826 return err; 5827 ptr = (void *)(long)addr + off; 5828 5829 switch (size) { 5830 case sizeof(u8): 5831 *val = (u64)*(u8 *)ptr; 5832 break; 5833 case sizeof(u16): 5834 *val = (u64)*(u16 *)ptr; 5835 break; 5836 case sizeof(u32): 5837 *val = (u64)*(u32 *)ptr; 5838 break; 5839 case sizeof(u64): 5840 *val = *(u64 *)ptr; 5841 break; 5842 default: 5843 return -EINVAL; 5844 } 5845 return 0; 5846 } 5847 5848 #define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu) 5849 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null) 5850 #define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted) 5851 5852 /* 5853 * Allow list few fields as RCU trusted or full trusted. 5854 * This logic doesn't allow mix tagging and will be removed once GCC supports 5855 * btf_type_tag. 5856 */ 5857 5858 /* RCU trusted: these fields are trusted in RCU CS and never NULL */ 5859 BTF_TYPE_SAFE_RCU(struct task_struct) { 5860 const cpumask_t *cpus_ptr; 5861 struct css_set __rcu *cgroups; 5862 struct task_struct __rcu *real_parent; 5863 struct task_struct *group_leader; 5864 }; 5865 5866 BTF_TYPE_SAFE_RCU(struct cgroup) { 5867 /* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */ 5868 struct kernfs_node *kn; 5869 }; 5870 5871 BTF_TYPE_SAFE_RCU(struct css_set) { 5872 struct cgroup *dfl_cgrp; 5873 }; 5874 5875 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */ 5876 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) { 5877 struct file __rcu *exe_file; 5878 }; 5879 5880 /* skb->sk, req->sk are not RCU protected, but we mark them as such 5881 * because bpf prog accessible sockets are SOCK_RCU_FREE. 5882 */ 5883 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) { 5884 struct sock *sk; 5885 }; 5886 5887 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) { 5888 struct sock *sk; 5889 }; 5890 5891 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */ 5892 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) { 5893 struct seq_file *seq; 5894 }; 5895 5896 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) { 5897 struct bpf_iter_meta *meta; 5898 struct task_struct *task; 5899 }; 5900 5901 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) { 5902 struct file *file; 5903 }; 5904 5905 BTF_TYPE_SAFE_TRUSTED(struct file) { 5906 struct inode *f_inode; 5907 }; 5908 5909 BTF_TYPE_SAFE_TRUSTED(struct dentry) { 5910 /* no negative dentry-s in places where bpf can see it */ 5911 struct inode *d_inode; 5912 }; 5913 5914 BTF_TYPE_SAFE_TRUSTED(struct socket) { 5915 struct sock *sk; 5916 }; 5917 5918 static bool type_is_rcu(struct bpf_verifier_env *env, 5919 struct bpf_reg_state *reg, 5920 const char *field_name, u32 btf_id) 5921 { 5922 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct)); 5923 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup)); 5924 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set)); 5925 5926 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu"); 5927 } 5928 5929 static bool type_is_rcu_or_null(struct bpf_verifier_env *env, 5930 struct bpf_reg_state *reg, 5931 const char *field_name, u32 btf_id) 5932 { 5933 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct)); 5934 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff)); 5935 BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock)); 5936 5937 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null"); 5938 } 5939 5940 static bool type_is_trusted(struct bpf_verifier_env *env, 5941 struct bpf_reg_state *reg, 5942 const char *field_name, u32 btf_id) 5943 { 5944 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta)); 5945 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task)); 5946 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm)); 5947 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file)); 5948 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct dentry)); 5949 BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct socket)); 5950 5951 return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted"); 5952 } 5953 5954 static int check_ptr_to_btf_access(struct bpf_verifier_env *env, 5955 struct bpf_reg_state *regs, 5956 int regno, int off, int size, 5957 enum bpf_access_type atype, 5958 int value_regno) 5959 { 5960 struct bpf_reg_state *reg = regs + regno; 5961 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id); 5962 const char *tname = btf_name_by_offset(reg->btf, t->name_off); 5963 const char *field_name = NULL; 5964 enum bpf_type_flag flag = 0; 5965 u32 btf_id = 0; 5966 int ret; 5967 5968 if (!env->allow_ptr_leaks) { 5969 verbose(env, 5970 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n", 5971 tname); 5972 return -EPERM; 5973 } 5974 if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) { 5975 verbose(env, 5976 "Cannot access kernel 'struct %s' from non-GPL compatible program\n", 5977 tname); 5978 return -EINVAL; 5979 } 5980 if (off < 0) { 5981 verbose(env, 5982 "R%d is ptr_%s invalid negative access: off=%d\n", 5983 regno, tname, off); 5984 return -EACCES; 5985 } 5986 if (!tnum_is_const(reg->var_off) || reg->var_off.value) { 5987 char tn_buf[48]; 5988 5989 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 5990 verbose(env, 5991 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n", 5992 regno, tname, off, tn_buf); 5993 return -EACCES; 5994 } 5995 5996 if (reg->type & MEM_USER) { 5997 verbose(env, 5998 "R%d is ptr_%s access user memory: off=%d\n", 5999 regno, tname, off); 6000 return -EACCES; 6001 } 6002 6003 if (reg->type & MEM_PERCPU) { 6004 verbose(env, 6005 "R%d is ptr_%s access percpu memory: off=%d\n", 6006 regno, tname, off); 6007 return -EACCES; 6008 } 6009 6010 if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) { 6011 if (!btf_is_kernel(reg->btf)) { 6012 verbose(env, "verifier internal error: reg->btf must be kernel btf\n"); 6013 return -EFAULT; 6014 } 6015 ret = env->ops->btf_struct_access(&env->log, reg, off, size); 6016 } else { 6017 /* Writes are permitted with default btf_struct_access for 6018 * program allocated objects (which always have ref_obj_id > 0), 6019 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC. 6020 */ 6021 if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) { 6022 verbose(env, "only read is supported\n"); 6023 return -EACCES; 6024 } 6025 6026 if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) && 6027 !reg->ref_obj_id) { 6028 verbose(env, "verifier internal error: ref_obj_id for allocated object must be non-zero\n"); 6029 return -EFAULT; 6030 } 6031 6032 ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name); 6033 } 6034 6035 if (ret < 0) 6036 return ret; 6037 6038 if (ret != PTR_TO_BTF_ID) { 6039 /* just mark; */ 6040 6041 } else if (type_flag(reg->type) & PTR_UNTRUSTED) { 6042 /* If this is an untrusted pointer, all pointers formed by walking it 6043 * also inherit the untrusted flag. 6044 */ 6045 flag = PTR_UNTRUSTED; 6046 6047 } else if (is_trusted_reg(reg) || is_rcu_reg(reg)) { 6048 /* By default any pointer obtained from walking a trusted pointer is no 6049 * longer trusted, unless the field being accessed has explicitly been 6050 * marked as inheriting its parent's state of trust (either full or RCU). 6051 * For example: 6052 * 'cgroups' pointer is untrusted if task->cgroups dereference 6053 * happened in a sleepable program outside of bpf_rcu_read_lock() 6054 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU). 6055 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED. 6056 * 6057 * A regular RCU-protected pointer with __rcu tag can also be deemed 6058 * trusted if we are in an RCU CS. Such pointer can be NULL. 6059 */ 6060 if (type_is_trusted(env, reg, field_name, btf_id)) { 6061 flag |= PTR_TRUSTED; 6062 } else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) { 6063 if (type_is_rcu(env, reg, field_name, btf_id)) { 6064 /* ignore __rcu tag and mark it MEM_RCU */ 6065 flag |= MEM_RCU; 6066 } else if (flag & MEM_RCU || 6067 type_is_rcu_or_null(env, reg, field_name, btf_id)) { 6068 /* __rcu tagged pointers can be NULL */ 6069 flag |= MEM_RCU | PTR_MAYBE_NULL; 6070 } else if (flag & (MEM_PERCPU | MEM_USER)) { 6071 /* keep as-is */ 6072 } else { 6073 /* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */ 6074 clear_trusted_flags(&flag); 6075 } 6076 } else { 6077 /* 6078 * If not in RCU CS or MEM_RCU pointer can be NULL then 6079 * aggressively mark as untrusted otherwise such 6080 * pointers will be plain PTR_TO_BTF_ID without flags 6081 * and will be allowed to be passed into helpers for 6082 * compat reasons. 6083 */ 6084 flag = PTR_UNTRUSTED; 6085 } 6086 } else { 6087 /* Old compat. Deprecated */ 6088 clear_trusted_flags(&flag); 6089 } 6090 6091 if (atype == BPF_READ && value_regno >= 0) 6092 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag); 6093 6094 return 0; 6095 } 6096 6097 static int check_ptr_to_map_access(struct bpf_verifier_env *env, 6098 struct bpf_reg_state *regs, 6099 int regno, int off, int size, 6100 enum bpf_access_type atype, 6101 int value_regno) 6102 { 6103 struct bpf_reg_state *reg = regs + regno; 6104 struct bpf_map *map = reg->map_ptr; 6105 struct bpf_reg_state map_reg; 6106 enum bpf_type_flag flag = 0; 6107 const struct btf_type *t; 6108 const char *tname; 6109 u32 btf_id; 6110 int ret; 6111 6112 if (!btf_vmlinux) { 6113 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n"); 6114 return -ENOTSUPP; 6115 } 6116 6117 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) { 6118 verbose(env, "map_ptr access not supported for map type %d\n", 6119 map->map_type); 6120 return -ENOTSUPP; 6121 } 6122 6123 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id); 6124 tname = btf_name_by_offset(btf_vmlinux, t->name_off); 6125 6126 if (!env->allow_ptr_leaks) { 6127 verbose(env, 6128 "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n", 6129 tname); 6130 return -EPERM; 6131 } 6132 6133 if (off < 0) { 6134 verbose(env, "R%d is %s invalid negative access: off=%d\n", 6135 regno, tname, off); 6136 return -EACCES; 6137 } 6138 6139 if (atype != BPF_READ) { 6140 verbose(env, "only read from %s is supported\n", tname); 6141 return -EACCES; 6142 } 6143 6144 /* Simulate access to a PTR_TO_BTF_ID */ 6145 memset(&map_reg, 0, sizeof(map_reg)); 6146 mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID, btf_vmlinux, *map->ops->map_btf_id, 0); 6147 ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL); 6148 if (ret < 0) 6149 return ret; 6150 6151 if (value_regno >= 0) 6152 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag); 6153 6154 return 0; 6155 } 6156 6157 /* Check that the stack access at the given offset is within bounds. The 6158 * maximum valid offset is -1. 6159 * 6160 * The minimum valid offset is -MAX_BPF_STACK for writes, and 6161 * -state->allocated_stack for reads. 6162 */ 6163 static int check_stack_slot_within_bounds(int off, 6164 struct bpf_func_state *state, 6165 enum bpf_access_type t) 6166 { 6167 int min_valid_off; 6168 6169 if (t == BPF_WRITE) 6170 min_valid_off = -MAX_BPF_STACK; 6171 else 6172 min_valid_off = -state->allocated_stack; 6173 6174 if (off < min_valid_off || off > -1) 6175 return -EACCES; 6176 return 0; 6177 } 6178 6179 /* Check that the stack access at 'regno + off' falls within the maximum stack 6180 * bounds. 6181 * 6182 * 'off' includes `regno->offset`, but not its dynamic part (if any). 6183 */ 6184 static int check_stack_access_within_bounds( 6185 struct bpf_verifier_env *env, 6186 int regno, int off, int access_size, 6187 enum bpf_access_src src, enum bpf_access_type type) 6188 { 6189 struct bpf_reg_state *regs = cur_regs(env); 6190 struct bpf_reg_state *reg = regs + regno; 6191 struct bpf_func_state *state = func(env, reg); 6192 int min_off, max_off; 6193 int err; 6194 char *err_extra; 6195 6196 if (src == ACCESS_HELPER) 6197 /* We don't know if helpers are reading or writing (or both). */ 6198 err_extra = " indirect access to"; 6199 else if (type == BPF_READ) 6200 err_extra = " read from"; 6201 else 6202 err_extra = " write to"; 6203 6204 if (tnum_is_const(reg->var_off)) { 6205 min_off = reg->var_off.value + off; 6206 if (access_size > 0) 6207 max_off = min_off + access_size - 1; 6208 else 6209 max_off = min_off; 6210 } else { 6211 if (reg->smax_value >= BPF_MAX_VAR_OFF || 6212 reg->smin_value <= -BPF_MAX_VAR_OFF) { 6213 verbose(env, "invalid unbounded variable-offset%s stack R%d\n", 6214 err_extra, regno); 6215 return -EACCES; 6216 } 6217 min_off = reg->smin_value + off; 6218 if (access_size > 0) 6219 max_off = reg->smax_value + off + access_size - 1; 6220 else 6221 max_off = min_off; 6222 } 6223 6224 err = check_stack_slot_within_bounds(min_off, state, type); 6225 if (!err) 6226 err = check_stack_slot_within_bounds(max_off, state, type); 6227 6228 if (err) { 6229 if (tnum_is_const(reg->var_off)) { 6230 verbose(env, "invalid%s stack R%d off=%d size=%d\n", 6231 err_extra, regno, off, access_size); 6232 } else { 6233 char tn_buf[48]; 6234 6235 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 6236 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n", 6237 err_extra, regno, tn_buf, access_size); 6238 } 6239 } 6240 return err; 6241 } 6242 6243 /* check whether memory at (regno + off) is accessible for t = (read | write) 6244 * if t==write, value_regno is a register which value is stored into memory 6245 * if t==read, value_regno is a register which will receive the value from memory 6246 * if t==write && value_regno==-1, some unknown value is stored into memory 6247 * if t==read && value_regno==-1, don't care what we read from memory 6248 */ 6249 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno, 6250 int off, int bpf_size, enum bpf_access_type t, 6251 int value_regno, bool strict_alignment_once) 6252 { 6253 struct bpf_reg_state *regs = cur_regs(env); 6254 struct bpf_reg_state *reg = regs + regno; 6255 struct bpf_func_state *state; 6256 int size, err = 0; 6257 6258 size = bpf_size_to_bytes(bpf_size); 6259 if (size < 0) 6260 return size; 6261 6262 /* alignment checks will add in reg->off themselves */ 6263 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once); 6264 if (err) 6265 return err; 6266 6267 /* for access checks, reg->off is just part of off */ 6268 off += reg->off; 6269 6270 if (reg->type == PTR_TO_MAP_KEY) { 6271 if (t == BPF_WRITE) { 6272 verbose(env, "write to change key R%d not allowed\n", regno); 6273 return -EACCES; 6274 } 6275 6276 err = check_mem_region_access(env, regno, off, size, 6277 reg->map_ptr->key_size, false); 6278 if (err) 6279 return err; 6280 if (value_regno >= 0) 6281 mark_reg_unknown(env, regs, value_regno); 6282 } else if (reg->type == PTR_TO_MAP_VALUE) { 6283 struct btf_field *kptr_field = NULL; 6284 6285 if (t == BPF_WRITE && value_regno >= 0 && 6286 is_pointer_value(env, value_regno)) { 6287 verbose(env, "R%d leaks addr into map\n", value_regno); 6288 return -EACCES; 6289 } 6290 err = check_map_access_type(env, regno, off, size, t); 6291 if (err) 6292 return err; 6293 err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT); 6294 if (err) 6295 return err; 6296 if (tnum_is_const(reg->var_off)) 6297 kptr_field = btf_record_find(reg->map_ptr->record, 6298 off + reg->var_off.value, BPF_KPTR); 6299 if (kptr_field) { 6300 err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field); 6301 } else if (t == BPF_READ && value_regno >= 0) { 6302 struct bpf_map *map = reg->map_ptr; 6303 6304 /* if map is read-only, track its contents as scalars */ 6305 if (tnum_is_const(reg->var_off) && 6306 bpf_map_is_rdonly(map) && 6307 map->ops->map_direct_value_addr) { 6308 int map_off = off + reg->var_off.value; 6309 u64 val = 0; 6310 6311 err = bpf_map_direct_read(map, map_off, size, 6312 &val); 6313 if (err) 6314 return err; 6315 6316 regs[value_regno].type = SCALAR_VALUE; 6317 __mark_reg_known(®s[value_regno], val); 6318 } else { 6319 mark_reg_unknown(env, regs, value_regno); 6320 } 6321 } 6322 } else if (base_type(reg->type) == PTR_TO_MEM) { 6323 bool rdonly_mem = type_is_rdonly_mem(reg->type); 6324 6325 if (type_may_be_null(reg->type)) { 6326 verbose(env, "R%d invalid mem access '%s'\n", regno, 6327 reg_type_str(env, reg->type)); 6328 return -EACCES; 6329 } 6330 6331 if (t == BPF_WRITE && rdonly_mem) { 6332 verbose(env, "R%d cannot write into %s\n", 6333 regno, reg_type_str(env, reg->type)); 6334 return -EACCES; 6335 } 6336 6337 if (t == BPF_WRITE && value_regno >= 0 && 6338 is_pointer_value(env, value_regno)) { 6339 verbose(env, "R%d leaks addr into mem\n", value_regno); 6340 return -EACCES; 6341 } 6342 6343 err = check_mem_region_access(env, regno, off, size, 6344 reg->mem_size, false); 6345 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem)) 6346 mark_reg_unknown(env, regs, value_regno); 6347 } else if (reg->type == PTR_TO_CTX) { 6348 enum bpf_reg_type reg_type = SCALAR_VALUE; 6349 struct btf *btf = NULL; 6350 u32 btf_id = 0; 6351 6352 if (t == BPF_WRITE && value_regno >= 0 && 6353 is_pointer_value(env, value_regno)) { 6354 verbose(env, "R%d leaks addr into ctx\n", value_regno); 6355 return -EACCES; 6356 } 6357 6358 err = check_ptr_off_reg(env, reg, regno); 6359 if (err < 0) 6360 return err; 6361 6362 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, 6363 &btf_id); 6364 if (err) 6365 verbose_linfo(env, insn_idx, "; "); 6366 if (!err && t == BPF_READ && value_regno >= 0) { 6367 /* ctx access returns either a scalar, or a 6368 * PTR_TO_PACKET[_META,_END]. In the latter 6369 * case, we know the offset is zero. 6370 */ 6371 if (reg_type == SCALAR_VALUE) { 6372 mark_reg_unknown(env, regs, value_regno); 6373 } else { 6374 mark_reg_known_zero(env, regs, 6375 value_regno); 6376 if (type_may_be_null(reg_type)) 6377 regs[value_regno].id = ++env->id_gen; 6378 /* A load of ctx field could have different 6379 * actual load size with the one encoded in the 6380 * insn. When the dst is PTR, it is for sure not 6381 * a sub-register. 6382 */ 6383 regs[value_regno].subreg_def = DEF_NOT_SUBREG; 6384 if (base_type(reg_type) == PTR_TO_BTF_ID) { 6385 regs[value_regno].btf = btf; 6386 regs[value_regno].btf_id = btf_id; 6387 } 6388 } 6389 regs[value_regno].type = reg_type; 6390 } 6391 6392 } else if (reg->type == PTR_TO_STACK) { 6393 /* Basic bounds checks. */ 6394 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t); 6395 if (err) 6396 return err; 6397 6398 state = func(env, reg); 6399 err = update_stack_depth(env, state, off); 6400 if (err) 6401 return err; 6402 6403 if (t == BPF_READ) 6404 err = check_stack_read(env, regno, off, size, 6405 value_regno); 6406 else 6407 err = check_stack_write(env, regno, off, size, 6408 value_regno, insn_idx); 6409 } else if (reg_is_pkt_pointer(reg)) { 6410 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) { 6411 verbose(env, "cannot write into packet\n"); 6412 return -EACCES; 6413 } 6414 if (t == BPF_WRITE && value_regno >= 0 && 6415 is_pointer_value(env, value_regno)) { 6416 verbose(env, "R%d leaks addr into packet\n", 6417 value_regno); 6418 return -EACCES; 6419 } 6420 err = check_packet_access(env, regno, off, size, false); 6421 if (!err && t == BPF_READ && value_regno >= 0) 6422 mark_reg_unknown(env, regs, value_regno); 6423 } else if (reg->type == PTR_TO_FLOW_KEYS) { 6424 if (t == BPF_WRITE && value_regno >= 0 && 6425 is_pointer_value(env, value_regno)) { 6426 verbose(env, "R%d leaks addr into flow keys\n", 6427 value_regno); 6428 return -EACCES; 6429 } 6430 6431 err = check_flow_keys_access(env, off, size); 6432 if (!err && t == BPF_READ && value_regno >= 0) 6433 mark_reg_unknown(env, regs, value_regno); 6434 } else if (type_is_sk_pointer(reg->type)) { 6435 if (t == BPF_WRITE) { 6436 verbose(env, "R%d cannot write into %s\n", 6437 regno, reg_type_str(env, reg->type)); 6438 return -EACCES; 6439 } 6440 err = check_sock_access(env, insn_idx, regno, off, size, t); 6441 if (!err && value_regno >= 0) 6442 mark_reg_unknown(env, regs, value_regno); 6443 } else if (reg->type == PTR_TO_TP_BUFFER) { 6444 err = check_tp_buffer_access(env, reg, regno, off, size); 6445 if (!err && t == BPF_READ && value_regno >= 0) 6446 mark_reg_unknown(env, regs, value_regno); 6447 } else if (base_type(reg->type) == PTR_TO_BTF_ID && 6448 !type_may_be_null(reg->type)) { 6449 err = check_ptr_to_btf_access(env, regs, regno, off, size, t, 6450 value_regno); 6451 } else if (reg->type == CONST_PTR_TO_MAP) { 6452 err = check_ptr_to_map_access(env, regs, regno, off, size, t, 6453 value_regno); 6454 } else if (base_type(reg->type) == PTR_TO_BUF) { 6455 bool rdonly_mem = type_is_rdonly_mem(reg->type); 6456 u32 *max_access; 6457 6458 if (rdonly_mem) { 6459 if (t == BPF_WRITE) { 6460 verbose(env, "R%d cannot write into %s\n", 6461 regno, reg_type_str(env, reg->type)); 6462 return -EACCES; 6463 } 6464 max_access = &env->prog->aux->max_rdonly_access; 6465 } else { 6466 max_access = &env->prog->aux->max_rdwr_access; 6467 } 6468 6469 err = check_buffer_access(env, reg, regno, off, size, false, 6470 max_access); 6471 6472 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ)) 6473 mark_reg_unknown(env, regs, value_regno); 6474 } else { 6475 verbose(env, "R%d invalid mem access '%s'\n", regno, 6476 reg_type_str(env, reg->type)); 6477 return -EACCES; 6478 } 6479 6480 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ && 6481 regs[value_regno].type == SCALAR_VALUE) { 6482 /* b/h/w load zero-extends, mark upper bits as known 0 */ 6483 coerce_reg_to_size(®s[value_regno], size); 6484 } 6485 return err; 6486 } 6487 6488 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn) 6489 { 6490 int load_reg; 6491 int err; 6492 6493 switch (insn->imm) { 6494 case BPF_ADD: 6495 case BPF_ADD | BPF_FETCH: 6496 case BPF_AND: 6497 case BPF_AND | BPF_FETCH: 6498 case BPF_OR: 6499 case BPF_OR | BPF_FETCH: 6500 case BPF_XOR: 6501 case BPF_XOR | BPF_FETCH: 6502 case BPF_XCHG: 6503 case BPF_CMPXCHG: 6504 break; 6505 default: 6506 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm); 6507 return -EINVAL; 6508 } 6509 6510 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) { 6511 verbose(env, "invalid atomic operand size\n"); 6512 return -EINVAL; 6513 } 6514 6515 /* check src1 operand */ 6516 err = check_reg_arg(env, insn->src_reg, SRC_OP); 6517 if (err) 6518 return err; 6519 6520 /* check src2 operand */ 6521 err = check_reg_arg(env, insn->dst_reg, SRC_OP); 6522 if (err) 6523 return err; 6524 6525 if (insn->imm == BPF_CMPXCHG) { 6526 /* Check comparison of R0 with memory location */ 6527 const u32 aux_reg = BPF_REG_0; 6528 6529 err = check_reg_arg(env, aux_reg, SRC_OP); 6530 if (err) 6531 return err; 6532 6533 if (is_pointer_value(env, aux_reg)) { 6534 verbose(env, "R%d leaks addr into mem\n", aux_reg); 6535 return -EACCES; 6536 } 6537 } 6538 6539 if (is_pointer_value(env, insn->src_reg)) { 6540 verbose(env, "R%d leaks addr into mem\n", insn->src_reg); 6541 return -EACCES; 6542 } 6543 6544 if (is_ctx_reg(env, insn->dst_reg) || 6545 is_pkt_reg(env, insn->dst_reg) || 6546 is_flow_key_reg(env, insn->dst_reg) || 6547 is_sk_reg(env, insn->dst_reg)) { 6548 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n", 6549 insn->dst_reg, 6550 reg_type_str(env, reg_state(env, insn->dst_reg)->type)); 6551 return -EACCES; 6552 } 6553 6554 if (insn->imm & BPF_FETCH) { 6555 if (insn->imm == BPF_CMPXCHG) 6556 load_reg = BPF_REG_0; 6557 else 6558 load_reg = insn->src_reg; 6559 6560 /* check and record load of old value */ 6561 err = check_reg_arg(env, load_reg, DST_OP); 6562 if (err) 6563 return err; 6564 } else { 6565 /* This instruction accesses a memory location but doesn't 6566 * actually load it into a register. 6567 */ 6568 load_reg = -1; 6569 } 6570 6571 /* Check whether we can read the memory, with second call for fetch 6572 * case to simulate the register fill. 6573 */ 6574 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off, 6575 BPF_SIZE(insn->code), BPF_READ, -1, true); 6576 if (!err && load_reg >= 0) 6577 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off, 6578 BPF_SIZE(insn->code), BPF_READ, load_reg, 6579 true); 6580 if (err) 6581 return err; 6582 6583 /* Check whether we can write into the same memory. */ 6584 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off, 6585 BPF_SIZE(insn->code), BPF_WRITE, -1, true); 6586 if (err) 6587 return err; 6588 6589 return 0; 6590 } 6591 6592 /* When register 'regno' is used to read the stack (either directly or through 6593 * a helper function) make sure that it's within stack boundary and, depending 6594 * on the access type, that all elements of the stack are initialized. 6595 * 6596 * 'off' includes 'regno->off', but not its dynamic part (if any). 6597 * 6598 * All registers that have been spilled on the stack in the slots within the 6599 * read offsets are marked as read. 6600 */ 6601 static int check_stack_range_initialized( 6602 struct bpf_verifier_env *env, int regno, int off, 6603 int access_size, bool zero_size_allowed, 6604 enum bpf_access_src type, struct bpf_call_arg_meta *meta) 6605 { 6606 struct bpf_reg_state *reg = reg_state(env, regno); 6607 struct bpf_func_state *state = func(env, reg); 6608 int err, min_off, max_off, i, j, slot, spi; 6609 char *err_extra = type == ACCESS_HELPER ? " indirect" : ""; 6610 enum bpf_access_type bounds_check_type; 6611 /* Some accesses can write anything into the stack, others are 6612 * read-only. 6613 */ 6614 bool clobber = false; 6615 6616 if (access_size == 0 && !zero_size_allowed) { 6617 verbose(env, "invalid zero-sized read\n"); 6618 return -EACCES; 6619 } 6620 6621 if (type == ACCESS_HELPER) { 6622 /* The bounds checks for writes are more permissive than for 6623 * reads. However, if raw_mode is not set, we'll do extra 6624 * checks below. 6625 */ 6626 bounds_check_type = BPF_WRITE; 6627 clobber = true; 6628 } else { 6629 bounds_check_type = BPF_READ; 6630 } 6631 err = check_stack_access_within_bounds(env, regno, off, access_size, 6632 type, bounds_check_type); 6633 if (err) 6634 return err; 6635 6636 6637 if (tnum_is_const(reg->var_off)) { 6638 min_off = max_off = reg->var_off.value + off; 6639 } else { 6640 /* Variable offset is prohibited for unprivileged mode for 6641 * simplicity since it requires corresponding support in 6642 * Spectre masking for stack ALU. 6643 * See also retrieve_ptr_limit(). 6644 */ 6645 if (!env->bypass_spec_v1) { 6646 char tn_buf[48]; 6647 6648 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 6649 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n", 6650 regno, err_extra, tn_buf); 6651 return -EACCES; 6652 } 6653 /* Only initialized buffer on stack is allowed to be accessed 6654 * with variable offset. With uninitialized buffer it's hard to 6655 * guarantee that whole memory is marked as initialized on 6656 * helper return since specific bounds are unknown what may 6657 * cause uninitialized stack leaking. 6658 */ 6659 if (meta && meta->raw_mode) 6660 meta = NULL; 6661 6662 min_off = reg->smin_value + off; 6663 max_off = reg->smax_value + off; 6664 } 6665 6666 if (meta && meta->raw_mode) { 6667 /* Ensure we won't be overwriting dynptrs when simulating byte 6668 * by byte access in check_helper_call using meta.access_size. 6669 * This would be a problem if we have a helper in the future 6670 * which takes: 6671 * 6672 * helper(uninit_mem, len, dynptr) 6673 * 6674 * Now, uninint_mem may overlap with dynptr pointer. Hence, it 6675 * may end up writing to dynptr itself when touching memory from 6676 * arg 1. This can be relaxed on a case by case basis for known 6677 * safe cases, but reject due to the possibilitiy of aliasing by 6678 * default. 6679 */ 6680 for (i = min_off; i < max_off + access_size; i++) { 6681 int stack_off = -i - 1; 6682 6683 spi = __get_spi(i); 6684 /* raw_mode may write past allocated_stack */ 6685 if (state->allocated_stack <= stack_off) 6686 continue; 6687 if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) { 6688 verbose(env, "potential write to dynptr at off=%d disallowed\n", i); 6689 return -EACCES; 6690 } 6691 } 6692 meta->access_size = access_size; 6693 meta->regno = regno; 6694 return 0; 6695 } 6696 6697 for (i = min_off; i < max_off + access_size; i++) { 6698 u8 *stype; 6699 6700 slot = -i - 1; 6701 spi = slot / BPF_REG_SIZE; 6702 if (state->allocated_stack <= slot) 6703 goto err; 6704 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE]; 6705 if (*stype == STACK_MISC) 6706 goto mark; 6707 if ((*stype == STACK_ZERO) || 6708 (*stype == STACK_INVALID && env->allow_uninit_stack)) { 6709 if (clobber) { 6710 /* helper can write anything into the stack */ 6711 *stype = STACK_MISC; 6712 } 6713 goto mark; 6714 } 6715 6716 if (is_spilled_reg(&state->stack[spi]) && 6717 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE || 6718 env->allow_ptr_leaks)) { 6719 if (clobber) { 6720 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr); 6721 for (j = 0; j < BPF_REG_SIZE; j++) 6722 scrub_spilled_slot(&state->stack[spi].slot_type[j]); 6723 } 6724 goto mark; 6725 } 6726 6727 err: 6728 if (tnum_is_const(reg->var_off)) { 6729 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n", 6730 err_extra, regno, min_off, i - min_off, access_size); 6731 } else { 6732 char tn_buf[48]; 6733 6734 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 6735 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n", 6736 err_extra, regno, tn_buf, i - min_off, access_size); 6737 } 6738 return -EACCES; 6739 mark: 6740 /* reading any byte out of 8-byte 'spill_slot' will cause 6741 * the whole slot to be marked as 'read' 6742 */ 6743 mark_reg_read(env, &state->stack[spi].spilled_ptr, 6744 state->stack[spi].spilled_ptr.parent, 6745 REG_LIVE_READ64); 6746 /* We do not set REG_LIVE_WRITTEN for stack slot, as we can not 6747 * be sure that whether stack slot is written to or not. Hence, 6748 * we must still conservatively propagate reads upwards even if 6749 * helper may write to the entire memory range. 6750 */ 6751 } 6752 return update_stack_depth(env, state, min_off); 6753 } 6754 6755 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno, 6756 int access_size, bool zero_size_allowed, 6757 struct bpf_call_arg_meta *meta) 6758 { 6759 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 6760 u32 *max_access; 6761 6762 switch (base_type(reg->type)) { 6763 case PTR_TO_PACKET: 6764 case PTR_TO_PACKET_META: 6765 return check_packet_access(env, regno, reg->off, access_size, 6766 zero_size_allowed); 6767 case PTR_TO_MAP_KEY: 6768 if (meta && meta->raw_mode) { 6769 verbose(env, "R%d cannot write into %s\n", regno, 6770 reg_type_str(env, reg->type)); 6771 return -EACCES; 6772 } 6773 return check_mem_region_access(env, regno, reg->off, access_size, 6774 reg->map_ptr->key_size, false); 6775 case PTR_TO_MAP_VALUE: 6776 if (check_map_access_type(env, regno, reg->off, access_size, 6777 meta && meta->raw_mode ? BPF_WRITE : 6778 BPF_READ)) 6779 return -EACCES; 6780 return check_map_access(env, regno, reg->off, access_size, 6781 zero_size_allowed, ACCESS_HELPER); 6782 case PTR_TO_MEM: 6783 if (type_is_rdonly_mem(reg->type)) { 6784 if (meta && meta->raw_mode) { 6785 verbose(env, "R%d cannot write into %s\n", regno, 6786 reg_type_str(env, reg->type)); 6787 return -EACCES; 6788 } 6789 } 6790 return check_mem_region_access(env, regno, reg->off, 6791 access_size, reg->mem_size, 6792 zero_size_allowed); 6793 case PTR_TO_BUF: 6794 if (type_is_rdonly_mem(reg->type)) { 6795 if (meta && meta->raw_mode) { 6796 verbose(env, "R%d cannot write into %s\n", regno, 6797 reg_type_str(env, reg->type)); 6798 return -EACCES; 6799 } 6800 6801 max_access = &env->prog->aux->max_rdonly_access; 6802 } else { 6803 max_access = &env->prog->aux->max_rdwr_access; 6804 } 6805 return check_buffer_access(env, reg, regno, reg->off, 6806 access_size, zero_size_allowed, 6807 max_access); 6808 case PTR_TO_STACK: 6809 return check_stack_range_initialized( 6810 env, 6811 regno, reg->off, access_size, 6812 zero_size_allowed, ACCESS_HELPER, meta); 6813 case PTR_TO_BTF_ID: 6814 return check_ptr_to_btf_access(env, regs, regno, reg->off, 6815 access_size, BPF_READ, -1); 6816 case PTR_TO_CTX: 6817 /* in case the function doesn't know how to access the context, 6818 * (because we are in a program of type SYSCALL for example), we 6819 * can not statically check its size. 6820 * Dynamically check it now. 6821 */ 6822 if (!env->ops->convert_ctx_access) { 6823 enum bpf_access_type atype = meta && meta->raw_mode ? BPF_WRITE : BPF_READ; 6824 int offset = access_size - 1; 6825 6826 /* Allow zero-byte read from PTR_TO_CTX */ 6827 if (access_size == 0) 6828 return zero_size_allowed ? 0 : -EACCES; 6829 6830 return check_mem_access(env, env->insn_idx, regno, offset, BPF_B, 6831 atype, -1, false); 6832 } 6833 6834 fallthrough; 6835 default: /* scalar_value or invalid ptr */ 6836 /* Allow zero-byte read from NULL, regardless of pointer type */ 6837 if (zero_size_allowed && access_size == 0 && 6838 register_is_null(reg)) 6839 return 0; 6840 6841 verbose(env, "R%d type=%s ", regno, 6842 reg_type_str(env, reg->type)); 6843 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK)); 6844 return -EACCES; 6845 } 6846 } 6847 6848 static int check_mem_size_reg(struct bpf_verifier_env *env, 6849 struct bpf_reg_state *reg, u32 regno, 6850 bool zero_size_allowed, 6851 struct bpf_call_arg_meta *meta) 6852 { 6853 int err; 6854 6855 /* This is used to refine r0 return value bounds for helpers 6856 * that enforce this value as an upper bound on return values. 6857 * See do_refine_retval_range() for helpers that can refine 6858 * the return value. C type of helper is u32 so we pull register 6859 * bound from umax_value however, if negative verifier errors 6860 * out. Only upper bounds can be learned because retval is an 6861 * int type and negative retvals are allowed. 6862 */ 6863 meta->msize_max_value = reg->umax_value; 6864 6865 /* The register is SCALAR_VALUE; the access check 6866 * happens using its boundaries. 6867 */ 6868 if (!tnum_is_const(reg->var_off)) 6869 /* For unprivileged variable accesses, disable raw 6870 * mode so that the program is required to 6871 * initialize all the memory that the helper could 6872 * just partially fill up. 6873 */ 6874 meta = NULL; 6875 6876 if (reg->smin_value < 0) { 6877 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n", 6878 regno); 6879 return -EACCES; 6880 } 6881 6882 if (reg->umin_value == 0) { 6883 err = check_helper_mem_access(env, regno - 1, 0, 6884 zero_size_allowed, 6885 meta); 6886 if (err) 6887 return err; 6888 } 6889 6890 if (reg->umax_value >= BPF_MAX_VAR_SIZ) { 6891 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n", 6892 regno); 6893 return -EACCES; 6894 } 6895 err = check_helper_mem_access(env, regno - 1, 6896 reg->umax_value, 6897 zero_size_allowed, meta); 6898 if (!err) 6899 err = mark_chain_precision(env, regno); 6900 return err; 6901 } 6902 6903 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 6904 u32 regno, u32 mem_size) 6905 { 6906 bool may_be_null = type_may_be_null(reg->type); 6907 struct bpf_reg_state saved_reg; 6908 struct bpf_call_arg_meta meta; 6909 int err; 6910 6911 if (register_is_null(reg)) 6912 return 0; 6913 6914 memset(&meta, 0, sizeof(meta)); 6915 /* Assuming that the register contains a value check if the memory 6916 * access is safe. Temporarily save and restore the register's state as 6917 * the conversion shouldn't be visible to a caller. 6918 */ 6919 if (may_be_null) { 6920 saved_reg = *reg; 6921 mark_ptr_not_null_reg(reg); 6922 } 6923 6924 err = check_helper_mem_access(env, regno, mem_size, true, &meta); 6925 /* Check access for BPF_WRITE */ 6926 meta.raw_mode = true; 6927 err = err ?: check_helper_mem_access(env, regno, mem_size, true, &meta); 6928 6929 if (may_be_null) 6930 *reg = saved_reg; 6931 6932 return err; 6933 } 6934 6935 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg, 6936 u32 regno) 6937 { 6938 struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1]; 6939 bool may_be_null = type_may_be_null(mem_reg->type); 6940 struct bpf_reg_state saved_reg; 6941 struct bpf_call_arg_meta meta; 6942 int err; 6943 6944 WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5); 6945 6946 memset(&meta, 0, sizeof(meta)); 6947 6948 if (may_be_null) { 6949 saved_reg = *mem_reg; 6950 mark_ptr_not_null_reg(mem_reg); 6951 } 6952 6953 err = check_mem_size_reg(env, reg, regno, true, &meta); 6954 /* Check access for BPF_WRITE */ 6955 meta.raw_mode = true; 6956 err = err ?: check_mem_size_reg(env, reg, regno, true, &meta); 6957 6958 if (may_be_null) 6959 *mem_reg = saved_reg; 6960 return err; 6961 } 6962 6963 /* Implementation details: 6964 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL. 6965 * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL. 6966 * Two bpf_map_lookups (even with the same key) will have different reg->id. 6967 * Two separate bpf_obj_new will also have different reg->id. 6968 * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier 6969 * clears reg->id after value_or_null->value transition, since the verifier only 6970 * cares about the range of access to valid map value pointer and doesn't care 6971 * about actual address of the map element. 6972 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps 6973 * reg->id > 0 after value_or_null->value transition. By doing so 6974 * two bpf_map_lookups will be considered two different pointers that 6975 * point to different bpf_spin_locks. Likewise for pointers to allocated objects 6976 * returned from bpf_obj_new. 6977 * The verifier allows taking only one bpf_spin_lock at a time to avoid 6978 * dead-locks. 6979 * Since only one bpf_spin_lock is allowed the checks are simpler than 6980 * reg_is_refcounted() logic. The verifier needs to remember only 6981 * one spin_lock instead of array of acquired_refs. 6982 * cur_state->active_lock remembers which map value element or allocated 6983 * object got locked and clears it after bpf_spin_unlock. 6984 */ 6985 static int process_spin_lock(struct bpf_verifier_env *env, int regno, 6986 bool is_lock) 6987 { 6988 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 6989 struct bpf_verifier_state *cur = env->cur_state; 6990 bool is_const = tnum_is_const(reg->var_off); 6991 u64 val = reg->var_off.value; 6992 struct bpf_map *map = NULL; 6993 struct btf *btf = NULL; 6994 struct btf_record *rec; 6995 6996 if (!is_const) { 6997 verbose(env, 6998 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n", 6999 regno); 7000 return -EINVAL; 7001 } 7002 if (reg->type == PTR_TO_MAP_VALUE) { 7003 map = reg->map_ptr; 7004 if (!map->btf) { 7005 verbose(env, 7006 "map '%s' has to have BTF in order to use bpf_spin_lock\n", 7007 map->name); 7008 return -EINVAL; 7009 } 7010 } else { 7011 btf = reg->btf; 7012 } 7013 7014 rec = reg_btf_record(reg); 7015 if (!btf_record_has_field(rec, BPF_SPIN_LOCK)) { 7016 verbose(env, "%s '%s' has no valid bpf_spin_lock\n", map ? "map" : "local", 7017 map ? map->name : "kptr"); 7018 return -EINVAL; 7019 } 7020 if (rec->spin_lock_off != val + reg->off) { 7021 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock' that is at %d\n", 7022 val + reg->off, rec->spin_lock_off); 7023 return -EINVAL; 7024 } 7025 if (is_lock) { 7026 if (cur->active_lock.ptr) { 7027 verbose(env, 7028 "Locking two bpf_spin_locks are not allowed\n"); 7029 return -EINVAL; 7030 } 7031 if (map) 7032 cur->active_lock.ptr = map; 7033 else 7034 cur->active_lock.ptr = btf; 7035 cur->active_lock.id = reg->id; 7036 } else { 7037 void *ptr; 7038 7039 if (map) 7040 ptr = map; 7041 else 7042 ptr = btf; 7043 7044 if (!cur->active_lock.ptr) { 7045 verbose(env, "bpf_spin_unlock without taking a lock\n"); 7046 return -EINVAL; 7047 } 7048 if (cur->active_lock.ptr != ptr || 7049 cur->active_lock.id != reg->id) { 7050 verbose(env, "bpf_spin_unlock of different lock\n"); 7051 return -EINVAL; 7052 } 7053 7054 invalidate_non_owning_refs(env); 7055 7056 cur->active_lock.ptr = NULL; 7057 cur->active_lock.id = 0; 7058 } 7059 return 0; 7060 } 7061 7062 static int process_timer_func(struct bpf_verifier_env *env, int regno, 7063 struct bpf_call_arg_meta *meta) 7064 { 7065 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 7066 bool is_const = tnum_is_const(reg->var_off); 7067 struct bpf_map *map = reg->map_ptr; 7068 u64 val = reg->var_off.value; 7069 7070 if (!is_const) { 7071 verbose(env, 7072 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n", 7073 regno); 7074 return -EINVAL; 7075 } 7076 if (!map->btf) { 7077 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n", 7078 map->name); 7079 return -EINVAL; 7080 } 7081 if (!btf_record_has_field(map->record, BPF_TIMER)) { 7082 verbose(env, "map '%s' has no valid bpf_timer\n", map->name); 7083 return -EINVAL; 7084 } 7085 if (map->record->timer_off != val + reg->off) { 7086 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n", 7087 val + reg->off, map->record->timer_off); 7088 return -EINVAL; 7089 } 7090 if (meta->map_ptr) { 7091 verbose(env, "verifier bug. Two map pointers in a timer helper\n"); 7092 return -EFAULT; 7093 } 7094 meta->map_uid = reg->map_uid; 7095 meta->map_ptr = map; 7096 return 0; 7097 } 7098 7099 static int process_kptr_func(struct bpf_verifier_env *env, int regno, 7100 struct bpf_call_arg_meta *meta) 7101 { 7102 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 7103 struct bpf_map *map_ptr = reg->map_ptr; 7104 struct btf_field *kptr_field; 7105 u32 kptr_off; 7106 7107 if (!tnum_is_const(reg->var_off)) { 7108 verbose(env, 7109 "R%d doesn't have constant offset. kptr has to be at the constant offset\n", 7110 regno); 7111 return -EINVAL; 7112 } 7113 if (!map_ptr->btf) { 7114 verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n", 7115 map_ptr->name); 7116 return -EINVAL; 7117 } 7118 if (!btf_record_has_field(map_ptr->record, BPF_KPTR)) { 7119 verbose(env, "map '%s' has no valid kptr\n", map_ptr->name); 7120 return -EINVAL; 7121 } 7122 7123 meta->map_ptr = map_ptr; 7124 kptr_off = reg->off + reg->var_off.value; 7125 kptr_field = btf_record_find(map_ptr->record, kptr_off, BPF_KPTR); 7126 if (!kptr_field) { 7127 verbose(env, "off=%d doesn't point to kptr\n", kptr_off); 7128 return -EACCES; 7129 } 7130 if (kptr_field->type != BPF_KPTR_REF) { 7131 verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off); 7132 return -EACCES; 7133 } 7134 meta->kptr_field = kptr_field; 7135 return 0; 7136 } 7137 7138 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK 7139 * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR. 7140 * 7141 * In both cases we deal with the first 8 bytes, but need to mark the next 8 7142 * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of 7143 * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object. 7144 * 7145 * Mutability of bpf_dynptr is at two levels, one is at the level of struct 7146 * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct 7147 * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can 7148 * mutate the view of the dynptr and also possibly destroy it. In the latter 7149 * case, it cannot mutate the bpf_dynptr itself but it can still mutate the 7150 * memory that dynptr points to. 7151 * 7152 * The verifier will keep track both levels of mutation (bpf_dynptr's in 7153 * reg->type and the memory's in reg->dynptr.type), but there is no support for 7154 * readonly dynptr view yet, hence only the first case is tracked and checked. 7155 * 7156 * This is consistent with how C applies the const modifier to a struct object, 7157 * where the pointer itself inside bpf_dynptr becomes const but not what it 7158 * points to. 7159 * 7160 * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument 7161 * type, and declare it as 'const struct bpf_dynptr *' in their prototype. 7162 */ 7163 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx, 7164 enum bpf_arg_type arg_type, int clone_ref_obj_id) 7165 { 7166 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 7167 int err; 7168 7169 /* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an 7170 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*): 7171 */ 7172 if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) { 7173 verbose(env, "verifier internal error: misconfigured dynptr helper type flags\n"); 7174 return -EFAULT; 7175 } 7176 7177 /* MEM_UNINIT - Points to memory that is an appropriate candidate for 7178 * constructing a mutable bpf_dynptr object. 7179 * 7180 * Currently, this is only possible with PTR_TO_STACK 7181 * pointing to a region of at least 16 bytes which doesn't 7182 * contain an existing bpf_dynptr. 7183 * 7184 * MEM_RDONLY - Points to a initialized bpf_dynptr that will not be 7185 * mutated or destroyed. However, the memory it points to 7186 * may be mutated. 7187 * 7188 * None - Points to a initialized dynptr that can be mutated and 7189 * destroyed, including mutation of the memory it points 7190 * to. 7191 */ 7192 if (arg_type & MEM_UNINIT) { 7193 int i; 7194 7195 if (!is_dynptr_reg_valid_uninit(env, reg)) { 7196 verbose(env, "Dynptr has to be an uninitialized dynptr\n"); 7197 return -EINVAL; 7198 } 7199 7200 /* we write BPF_DW bits (8 bytes) at a time */ 7201 for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) { 7202 err = check_mem_access(env, insn_idx, regno, 7203 i, BPF_DW, BPF_WRITE, -1, false); 7204 if (err) 7205 return err; 7206 } 7207 7208 err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id); 7209 } else /* MEM_RDONLY and None case from above */ { 7210 /* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */ 7211 if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) { 7212 verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n"); 7213 return -EINVAL; 7214 } 7215 7216 if (!is_dynptr_reg_valid_init(env, reg)) { 7217 verbose(env, 7218 "Expected an initialized dynptr as arg #%d\n", 7219 regno); 7220 return -EINVAL; 7221 } 7222 7223 /* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */ 7224 if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) { 7225 verbose(env, 7226 "Expected a dynptr of type %s as arg #%d\n", 7227 dynptr_type_str(arg_to_dynptr_type(arg_type)), regno); 7228 return -EINVAL; 7229 } 7230 7231 err = mark_dynptr_read(env, reg); 7232 } 7233 return err; 7234 } 7235 7236 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi) 7237 { 7238 struct bpf_func_state *state = func(env, reg); 7239 7240 return state->stack[spi].spilled_ptr.ref_obj_id; 7241 } 7242 7243 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta) 7244 { 7245 return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY); 7246 } 7247 7248 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta) 7249 { 7250 return meta->kfunc_flags & KF_ITER_NEW; 7251 } 7252 7253 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta) 7254 { 7255 return meta->kfunc_flags & KF_ITER_NEXT; 7256 } 7257 7258 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta) 7259 { 7260 return meta->kfunc_flags & KF_ITER_DESTROY; 7261 } 7262 7263 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg) 7264 { 7265 /* btf_check_iter_kfuncs() guarantees that first argument of any iter 7266 * kfunc is iter state pointer 7267 */ 7268 return arg == 0 && is_iter_kfunc(meta); 7269 } 7270 7271 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx, 7272 struct bpf_kfunc_call_arg_meta *meta) 7273 { 7274 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 7275 const struct btf_type *t; 7276 const struct btf_param *arg; 7277 int spi, err, i, nr_slots; 7278 u32 btf_id; 7279 7280 /* btf_check_iter_kfuncs() ensures we don't need to validate anything here */ 7281 arg = &btf_params(meta->func_proto)[0]; 7282 t = btf_type_skip_modifiers(meta->btf, arg->type, NULL); /* PTR */ 7283 t = btf_type_skip_modifiers(meta->btf, t->type, &btf_id); /* STRUCT */ 7284 nr_slots = t->size / BPF_REG_SIZE; 7285 7286 if (is_iter_new_kfunc(meta)) { 7287 /* bpf_iter_<type>_new() expects pointer to uninit iter state */ 7288 if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) { 7289 verbose(env, "expected uninitialized iter_%s as arg #%d\n", 7290 iter_type_str(meta->btf, btf_id), regno); 7291 return -EINVAL; 7292 } 7293 7294 for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) { 7295 err = check_mem_access(env, insn_idx, regno, 7296 i, BPF_DW, BPF_WRITE, -1, false); 7297 if (err) 7298 return err; 7299 } 7300 7301 err = mark_stack_slots_iter(env, reg, insn_idx, meta->btf, btf_id, nr_slots); 7302 if (err) 7303 return err; 7304 } else { 7305 /* iter_next() or iter_destroy() expect initialized iter state*/ 7306 if (!is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots)) { 7307 verbose(env, "expected an initialized iter_%s as arg #%d\n", 7308 iter_type_str(meta->btf, btf_id), regno); 7309 return -EINVAL; 7310 } 7311 7312 spi = iter_get_spi(env, reg, nr_slots); 7313 if (spi < 0) 7314 return spi; 7315 7316 err = mark_iter_read(env, reg, spi, nr_slots); 7317 if (err) 7318 return err; 7319 7320 /* remember meta->iter info for process_iter_next_call() */ 7321 meta->iter.spi = spi; 7322 meta->iter.frameno = reg->frameno; 7323 meta->ref_obj_id = iter_ref_obj_id(env, reg, spi); 7324 7325 if (is_iter_destroy_kfunc(meta)) { 7326 err = unmark_stack_slots_iter(env, reg, nr_slots); 7327 if (err) 7328 return err; 7329 } 7330 } 7331 7332 return 0; 7333 } 7334 7335 /* process_iter_next_call() is called when verifier gets to iterator's next 7336 * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer 7337 * to it as just "iter_next()" in comments below. 7338 * 7339 * BPF verifier relies on a crucial contract for any iter_next() 7340 * implementation: it should *eventually* return NULL, and once that happens 7341 * it should keep returning NULL. That is, once iterator exhausts elements to 7342 * iterate, it should never reset or spuriously return new elements. 7343 * 7344 * With the assumption of such contract, process_iter_next_call() simulates 7345 * a fork in the verifier state to validate loop logic correctness and safety 7346 * without having to simulate infinite amount of iterations. 7347 * 7348 * In current state, we first assume that iter_next() returned NULL and 7349 * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such 7350 * conditions we should not form an infinite loop and should eventually reach 7351 * exit. 7352 * 7353 * Besides that, we also fork current state and enqueue it for later 7354 * verification. In a forked state we keep iterator state as ACTIVE 7355 * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We 7356 * also bump iteration depth to prevent erroneous infinite loop detection 7357 * later on (see iter_active_depths_differ() comment for details). In this 7358 * state we assume that we'll eventually loop back to another iter_next() 7359 * calls (it could be in exactly same location or in some other instruction, 7360 * it doesn't matter, we don't make any unnecessary assumptions about this, 7361 * everything revolves around iterator state in a stack slot, not which 7362 * instruction is calling iter_next()). When that happens, we either will come 7363 * to iter_next() with equivalent state and can conclude that next iteration 7364 * will proceed in exactly the same way as we just verified, so it's safe to 7365 * assume that loop converges. If not, we'll go on another iteration 7366 * simulation with a different input state, until all possible starting states 7367 * are validated or we reach maximum number of instructions limit. 7368 * 7369 * This way, we will either exhaustively discover all possible input states 7370 * that iterator loop can start with and eventually will converge, or we'll 7371 * effectively regress into bounded loop simulation logic and either reach 7372 * maximum number of instructions if loop is not provably convergent, or there 7373 * is some statically known limit on number of iterations (e.g., if there is 7374 * an explicit `if n > 100 then break;` statement somewhere in the loop). 7375 * 7376 * One very subtle but very important aspect is that we *always* simulate NULL 7377 * condition first (as the current state) before we simulate non-NULL case. 7378 * This has to do with intricacies of scalar precision tracking. By simulating 7379 * "exit condition" of iter_next() returning NULL first, we make sure all the 7380 * relevant precision marks *that will be set **after** we exit iterator loop* 7381 * are propagated backwards to common parent state of NULL and non-NULL 7382 * branches. Thanks to that, state equivalence checks done later in forked 7383 * state, when reaching iter_next() for ACTIVE iterator, can assume that 7384 * precision marks are finalized and won't change. Because simulating another 7385 * ACTIVE iterator iteration won't change them (because given same input 7386 * states we'll end up with exactly same output states which we are currently 7387 * comparing; and verification after the loop already propagated back what 7388 * needs to be **additionally** tracked as precise). It's subtle, grok 7389 * precision tracking for more intuitive understanding. 7390 */ 7391 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx, 7392 struct bpf_kfunc_call_arg_meta *meta) 7393 { 7394 struct bpf_verifier_state *cur_st = env->cur_state, *queued_st; 7395 struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr; 7396 struct bpf_reg_state *cur_iter, *queued_iter; 7397 int iter_frameno = meta->iter.frameno; 7398 int iter_spi = meta->iter.spi; 7399 7400 BTF_TYPE_EMIT(struct bpf_iter); 7401 7402 cur_iter = &env->cur_state->frame[iter_frameno]->stack[iter_spi].spilled_ptr; 7403 7404 if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE && 7405 cur_iter->iter.state != BPF_ITER_STATE_DRAINED) { 7406 verbose(env, "verifier internal error: unexpected iterator state %d (%s)\n", 7407 cur_iter->iter.state, iter_state_str(cur_iter->iter.state)); 7408 return -EFAULT; 7409 } 7410 7411 if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) { 7412 /* branch out active iter state */ 7413 queued_st = push_stack(env, insn_idx + 1, insn_idx, false); 7414 if (!queued_st) 7415 return -ENOMEM; 7416 7417 queued_iter = &queued_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr; 7418 queued_iter->iter.state = BPF_ITER_STATE_ACTIVE; 7419 queued_iter->iter.depth++; 7420 7421 queued_fr = queued_st->frame[queued_st->curframe]; 7422 mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]); 7423 } 7424 7425 /* switch to DRAINED state, but keep the depth unchanged */ 7426 /* mark current iter state as drained and assume returned NULL */ 7427 cur_iter->iter.state = BPF_ITER_STATE_DRAINED; 7428 __mark_reg_const_zero(&cur_fr->regs[BPF_REG_0]); 7429 7430 return 0; 7431 } 7432 7433 static bool arg_type_is_mem_size(enum bpf_arg_type type) 7434 { 7435 return type == ARG_CONST_SIZE || 7436 type == ARG_CONST_SIZE_OR_ZERO; 7437 } 7438 7439 static bool arg_type_is_release(enum bpf_arg_type type) 7440 { 7441 return type & OBJ_RELEASE; 7442 } 7443 7444 static bool arg_type_is_dynptr(enum bpf_arg_type type) 7445 { 7446 return base_type(type) == ARG_PTR_TO_DYNPTR; 7447 } 7448 7449 static int int_ptr_type_to_size(enum bpf_arg_type type) 7450 { 7451 if (type == ARG_PTR_TO_INT) 7452 return sizeof(u32); 7453 else if (type == ARG_PTR_TO_LONG) 7454 return sizeof(u64); 7455 7456 return -EINVAL; 7457 } 7458 7459 static int resolve_map_arg_type(struct bpf_verifier_env *env, 7460 const struct bpf_call_arg_meta *meta, 7461 enum bpf_arg_type *arg_type) 7462 { 7463 if (!meta->map_ptr) { 7464 /* kernel subsystem misconfigured verifier */ 7465 verbose(env, "invalid map_ptr to access map->type\n"); 7466 return -EACCES; 7467 } 7468 7469 switch (meta->map_ptr->map_type) { 7470 case BPF_MAP_TYPE_SOCKMAP: 7471 case BPF_MAP_TYPE_SOCKHASH: 7472 if (*arg_type == ARG_PTR_TO_MAP_VALUE) { 7473 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON; 7474 } else { 7475 verbose(env, "invalid arg_type for sockmap/sockhash\n"); 7476 return -EINVAL; 7477 } 7478 break; 7479 case BPF_MAP_TYPE_BLOOM_FILTER: 7480 if (meta->func_id == BPF_FUNC_map_peek_elem) 7481 *arg_type = ARG_PTR_TO_MAP_VALUE; 7482 break; 7483 default: 7484 break; 7485 } 7486 return 0; 7487 } 7488 7489 struct bpf_reg_types { 7490 const enum bpf_reg_type types[10]; 7491 u32 *btf_id; 7492 }; 7493 7494 static const struct bpf_reg_types sock_types = { 7495 .types = { 7496 PTR_TO_SOCK_COMMON, 7497 PTR_TO_SOCKET, 7498 PTR_TO_TCP_SOCK, 7499 PTR_TO_XDP_SOCK, 7500 }, 7501 }; 7502 7503 #ifdef CONFIG_NET 7504 static const struct bpf_reg_types btf_id_sock_common_types = { 7505 .types = { 7506 PTR_TO_SOCK_COMMON, 7507 PTR_TO_SOCKET, 7508 PTR_TO_TCP_SOCK, 7509 PTR_TO_XDP_SOCK, 7510 PTR_TO_BTF_ID, 7511 PTR_TO_BTF_ID | PTR_TRUSTED, 7512 }, 7513 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON], 7514 }; 7515 #endif 7516 7517 static const struct bpf_reg_types mem_types = { 7518 .types = { 7519 PTR_TO_STACK, 7520 PTR_TO_PACKET, 7521 PTR_TO_PACKET_META, 7522 PTR_TO_MAP_KEY, 7523 PTR_TO_MAP_VALUE, 7524 PTR_TO_MEM, 7525 PTR_TO_MEM | MEM_RINGBUF, 7526 PTR_TO_BUF, 7527 PTR_TO_BTF_ID | PTR_TRUSTED, 7528 }, 7529 }; 7530 7531 static const struct bpf_reg_types int_ptr_types = { 7532 .types = { 7533 PTR_TO_STACK, 7534 PTR_TO_PACKET, 7535 PTR_TO_PACKET_META, 7536 PTR_TO_MAP_KEY, 7537 PTR_TO_MAP_VALUE, 7538 }, 7539 }; 7540 7541 static const struct bpf_reg_types spin_lock_types = { 7542 .types = { 7543 PTR_TO_MAP_VALUE, 7544 PTR_TO_BTF_ID | MEM_ALLOC, 7545 } 7546 }; 7547 7548 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } }; 7549 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } }; 7550 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } }; 7551 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } }; 7552 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } }; 7553 static const struct bpf_reg_types btf_ptr_types = { 7554 .types = { 7555 PTR_TO_BTF_ID, 7556 PTR_TO_BTF_ID | PTR_TRUSTED, 7557 PTR_TO_BTF_ID | MEM_RCU, 7558 }, 7559 }; 7560 static const struct bpf_reg_types percpu_btf_ptr_types = { 7561 .types = { 7562 PTR_TO_BTF_ID | MEM_PERCPU, 7563 PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED, 7564 } 7565 }; 7566 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } }; 7567 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } }; 7568 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } }; 7569 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } }; 7570 static const struct bpf_reg_types kptr_types = { .types = { PTR_TO_MAP_VALUE } }; 7571 static const struct bpf_reg_types dynptr_types = { 7572 .types = { 7573 PTR_TO_STACK, 7574 CONST_PTR_TO_DYNPTR, 7575 } 7576 }; 7577 7578 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = { 7579 [ARG_PTR_TO_MAP_KEY] = &mem_types, 7580 [ARG_PTR_TO_MAP_VALUE] = &mem_types, 7581 [ARG_CONST_SIZE] = &scalar_types, 7582 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types, 7583 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types, 7584 [ARG_CONST_MAP_PTR] = &const_map_ptr_types, 7585 [ARG_PTR_TO_CTX] = &context_types, 7586 [ARG_PTR_TO_SOCK_COMMON] = &sock_types, 7587 #ifdef CONFIG_NET 7588 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types, 7589 #endif 7590 [ARG_PTR_TO_SOCKET] = &fullsock_types, 7591 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types, 7592 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types, 7593 [ARG_PTR_TO_MEM] = &mem_types, 7594 [ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types, 7595 [ARG_PTR_TO_INT] = &int_ptr_types, 7596 [ARG_PTR_TO_LONG] = &int_ptr_types, 7597 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types, 7598 [ARG_PTR_TO_FUNC] = &func_ptr_types, 7599 [ARG_PTR_TO_STACK] = &stack_ptr_types, 7600 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types, 7601 [ARG_PTR_TO_TIMER] = &timer_types, 7602 [ARG_PTR_TO_KPTR] = &kptr_types, 7603 [ARG_PTR_TO_DYNPTR] = &dynptr_types, 7604 }; 7605 7606 static int check_reg_type(struct bpf_verifier_env *env, u32 regno, 7607 enum bpf_arg_type arg_type, 7608 const u32 *arg_btf_id, 7609 struct bpf_call_arg_meta *meta) 7610 { 7611 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 7612 enum bpf_reg_type expected, type = reg->type; 7613 const struct bpf_reg_types *compatible; 7614 int i, j; 7615 7616 compatible = compatible_reg_types[base_type(arg_type)]; 7617 if (!compatible) { 7618 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type); 7619 return -EFAULT; 7620 } 7621 7622 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY, 7623 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY 7624 * 7625 * Same for MAYBE_NULL: 7626 * 7627 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL, 7628 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL 7629 * 7630 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type. 7631 * 7632 * Therefore we fold these flags depending on the arg_type before comparison. 7633 */ 7634 if (arg_type & MEM_RDONLY) 7635 type &= ~MEM_RDONLY; 7636 if (arg_type & PTR_MAYBE_NULL) 7637 type &= ~PTR_MAYBE_NULL; 7638 if (base_type(arg_type) == ARG_PTR_TO_MEM) 7639 type &= ~DYNPTR_TYPE_FLAG_MASK; 7640 7641 if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type)) 7642 type &= ~MEM_ALLOC; 7643 7644 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) { 7645 expected = compatible->types[i]; 7646 if (expected == NOT_INIT) 7647 break; 7648 7649 if (type == expected) 7650 goto found; 7651 } 7652 7653 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type)); 7654 for (j = 0; j + 1 < i; j++) 7655 verbose(env, "%s, ", reg_type_str(env, compatible->types[j])); 7656 verbose(env, "%s\n", reg_type_str(env, compatible->types[j])); 7657 return -EACCES; 7658 7659 found: 7660 if (base_type(reg->type) != PTR_TO_BTF_ID) 7661 return 0; 7662 7663 if (compatible == &mem_types) { 7664 if (!(arg_type & MEM_RDONLY)) { 7665 verbose(env, 7666 "%s() may write into memory pointed by R%d type=%s\n", 7667 func_id_name(meta->func_id), 7668 regno, reg_type_str(env, reg->type)); 7669 return -EACCES; 7670 } 7671 return 0; 7672 } 7673 7674 switch ((int)reg->type) { 7675 case PTR_TO_BTF_ID: 7676 case PTR_TO_BTF_ID | PTR_TRUSTED: 7677 case PTR_TO_BTF_ID | MEM_RCU: 7678 case PTR_TO_BTF_ID | PTR_MAYBE_NULL: 7679 case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU: 7680 { 7681 /* For bpf_sk_release, it needs to match against first member 7682 * 'struct sock_common', hence make an exception for it. This 7683 * allows bpf_sk_release to work for multiple socket types. 7684 */ 7685 bool strict_type_match = arg_type_is_release(arg_type) && 7686 meta->func_id != BPF_FUNC_sk_release; 7687 7688 if (type_may_be_null(reg->type) && 7689 (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) { 7690 verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno); 7691 return -EACCES; 7692 } 7693 7694 if (!arg_btf_id) { 7695 if (!compatible->btf_id) { 7696 verbose(env, "verifier internal error: missing arg compatible BTF ID\n"); 7697 return -EFAULT; 7698 } 7699 arg_btf_id = compatible->btf_id; 7700 } 7701 7702 if (meta->func_id == BPF_FUNC_kptr_xchg) { 7703 if (map_kptr_match_type(env, meta->kptr_field, reg, regno)) 7704 return -EACCES; 7705 } else { 7706 if (arg_btf_id == BPF_PTR_POISON) { 7707 verbose(env, "verifier internal error:"); 7708 verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n", 7709 regno); 7710 return -EACCES; 7711 } 7712 7713 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off, 7714 btf_vmlinux, *arg_btf_id, 7715 strict_type_match)) { 7716 verbose(env, "R%d is of type %s but %s is expected\n", 7717 regno, btf_type_name(reg->btf, reg->btf_id), 7718 btf_type_name(btf_vmlinux, *arg_btf_id)); 7719 return -EACCES; 7720 } 7721 } 7722 break; 7723 } 7724 case PTR_TO_BTF_ID | MEM_ALLOC: 7725 if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock && 7726 meta->func_id != BPF_FUNC_kptr_xchg) { 7727 verbose(env, "verifier internal error: unimplemented handling of MEM_ALLOC\n"); 7728 return -EFAULT; 7729 } 7730 /* Handled by helper specific checks */ 7731 break; 7732 case PTR_TO_BTF_ID | MEM_PERCPU: 7733 case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED: 7734 /* Handled by helper specific checks */ 7735 break; 7736 default: 7737 verbose(env, "verifier internal error: invalid PTR_TO_BTF_ID register for type match\n"); 7738 return -EFAULT; 7739 } 7740 return 0; 7741 } 7742 7743 static struct btf_field * 7744 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields) 7745 { 7746 struct btf_field *field; 7747 struct btf_record *rec; 7748 7749 rec = reg_btf_record(reg); 7750 if (!rec) 7751 return NULL; 7752 7753 field = btf_record_find(rec, off, fields); 7754 if (!field) 7755 return NULL; 7756 7757 return field; 7758 } 7759 7760 int check_func_arg_reg_off(struct bpf_verifier_env *env, 7761 const struct bpf_reg_state *reg, int regno, 7762 enum bpf_arg_type arg_type) 7763 { 7764 u32 type = reg->type; 7765 7766 /* When referenced register is passed to release function, its fixed 7767 * offset must be 0. 7768 * 7769 * We will check arg_type_is_release reg has ref_obj_id when storing 7770 * meta->release_regno. 7771 */ 7772 if (arg_type_is_release(arg_type)) { 7773 /* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it 7774 * may not directly point to the object being released, but to 7775 * dynptr pointing to such object, which might be at some offset 7776 * on the stack. In that case, we simply to fallback to the 7777 * default handling. 7778 */ 7779 if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK) 7780 return 0; 7781 7782 if ((type_is_ptr_alloc_obj(type) || type_is_non_owning_ref(type)) && reg->off) { 7783 if (reg_find_field_offset(reg, reg->off, BPF_GRAPH_NODE_OR_ROOT)) 7784 return __check_ptr_off_reg(env, reg, regno, true); 7785 7786 verbose(env, "R%d must have zero offset when passed to release func\n", 7787 regno); 7788 verbose(env, "No graph node or root found at R%d type:%s off:%d\n", regno, 7789 btf_type_name(reg->btf, reg->btf_id), reg->off); 7790 return -EINVAL; 7791 } 7792 7793 /* Doing check_ptr_off_reg check for the offset will catch this 7794 * because fixed_off_ok is false, but checking here allows us 7795 * to give the user a better error message. 7796 */ 7797 if (reg->off) { 7798 verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n", 7799 regno); 7800 return -EINVAL; 7801 } 7802 return __check_ptr_off_reg(env, reg, regno, false); 7803 } 7804 7805 switch (type) { 7806 /* Pointer types where both fixed and variable offset is explicitly allowed: */ 7807 case PTR_TO_STACK: 7808 case PTR_TO_PACKET: 7809 case PTR_TO_PACKET_META: 7810 case PTR_TO_MAP_KEY: 7811 case PTR_TO_MAP_VALUE: 7812 case PTR_TO_MEM: 7813 case PTR_TO_MEM | MEM_RDONLY: 7814 case PTR_TO_MEM | MEM_RINGBUF: 7815 case PTR_TO_BUF: 7816 case PTR_TO_BUF | MEM_RDONLY: 7817 case SCALAR_VALUE: 7818 return 0; 7819 /* All the rest must be rejected, except PTR_TO_BTF_ID which allows 7820 * fixed offset. 7821 */ 7822 case PTR_TO_BTF_ID: 7823 case PTR_TO_BTF_ID | MEM_ALLOC: 7824 case PTR_TO_BTF_ID | PTR_TRUSTED: 7825 case PTR_TO_BTF_ID | MEM_RCU: 7826 case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF: 7827 /* When referenced PTR_TO_BTF_ID is passed to release function, 7828 * its fixed offset must be 0. In the other cases, fixed offset 7829 * can be non-zero. This was already checked above. So pass 7830 * fixed_off_ok as true to allow fixed offset for all other 7831 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we 7832 * still need to do checks instead of returning. 7833 */ 7834 return __check_ptr_off_reg(env, reg, regno, true); 7835 default: 7836 return __check_ptr_off_reg(env, reg, regno, false); 7837 } 7838 } 7839 7840 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env, 7841 const struct bpf_func_proto *fn, 7842 struct bpf_reg_state *regs) 7843 { 7844 struct bpf_reg_state *state = NULL; 7845 int i; 7846 7847 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) 7848 if (arg_type_is_dynptr(fn->arg_type[i])) { 7849 if (state) { 7850 verbose(env, "verifier internal error: multiple dynptr args\n"); 7851 return NULL; 7852 } 7853 state = ®s[BPF_REG_1 + i]; 7854 } 7855 7856 if (!state) 7857 verbose(env, "verifier internal error: no dynptr arg found\n"); 7858 7859 return state; 7860 } 7861 7862 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 7863 { 7864 struct bpf_func_state *state = func(env, reg); 7865 int spi; 7866 7867 if (reg->type == CONST_PTR_TO_DYNPTR) 7868 return reg->id; 7869 spi = dynptr_get_spi(env, reg); 7870 if (spi < 0) 7871 return spi; 7872 return state->stack[spi].spilled_ptr.id; 7873 } 7874 7875 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 7876 { 7877 struct bpf_func_state *state = func(env, reg); 7878 int spi; 7879 7880 if (reg->type == CONST_PTR_TO_DYNPTR) 7881 return reg->ref_obj_id; 7882 spi = dynptr_get_spi(env, reg); 7883 if (spi < 0) 7884 return spi; 7885 return state->stack[spi].spilled_ptr.ref_obj_id; 7886 } 7887 7888 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env, 7889 struct bpf_reg_state *reg) 7890 { 7891 struct bpf_func_state *state = func(env, reg); 7892 int spi; 7893 7894 if (reg->type == CONST_PTR_TO_DYNPTR) 7895 return reg->dynptr.type; 7896 7897 spi = __get_spi(reg->off); 7898 if (spi < 0) { 7899 verbose(env, "verifier internal error: invalid spi when querying dynptr type\n"); 7900 return BPF_DYNPTR_TYPE_INVALID; 7901 } 7902 7903 return state->stack[spi].spilled_ptr.dynptr.type; 7904 } 7905 7906 static int check_func_arg(struct bpf_verifier_env *env, u32 arg, 7907 struct bpf_call_arg_meta *meta, 7908 const struct bpf_func_proto *fn, 7909 int insn_idx) 7910 { 7911 u32 regno = BPF_REG_1 + arg; 7912 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno]; 7913 enum bpf_arg_type arg_type = fn->arg_type[arg]; 7914 enum bpf_reg_type type = reg->type; 7915 u32 *arg_btf_id = NULL; 7916 int err = 0; 7917 7918 if (arg_type == ARG_DONTCARE) 7919 return 0; 7920 7921 err = check_reg_arg(env, regno, SRC_OP); 7922 if (err) 7923 return err; 7924 7925 if (arg_type == ARG_ANYTHING) { 7926 if (is_pointer_value(env, regno)) { 7927 verbose(env, "R%d leaks addr into helper function\n", 7928 regno); 7929 return -EACCES; 7930 } 7931 return 0; 7932 } 7933 7934 if (type_is_pkt_pointer(type) && 7935 !may_access_direct_pkt_data(env, meta, BPF_READ)) { 7936 verbose(env, "helper access to the packet is not allowed\n"); 7937 return -EACCES; 7938 } 7939 7940 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) { 7941 err = resolve_map_arg_type(env, meta, &arg_type); 7942 if (err) 7943 return err; 7944 } 7945 7946 if (register_is_null(reg) && type_may_be_null(arg_type)) 7947 /* A NULL register has a SCALAR_VALUE type, so skip 7948 * type checking. 7949 */ 7950 goto skip_type_check; 7951 7952 /* arg_btf_id and arg_size are in a union. */ 7953 if (base_type(arg_type) == ARG_PTR_TO_BTF_ID || 7954 base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK) 7955 arg_btf_id = fn->arg_btf_id[arg]; 7956 7957 err = check_reg_type(env, regno, arg_type, arg_btf_id, meta); 7958 if (err) 7959 return err; 7960 7961 err = check_func_arg_reg_off(env, reg, regno, arg_type); 7962 if (err) 7963 return err; 7964 7965 skip_type_check: 7966 if (arg_type_is_release(arg_type)) { 7967 if (arg_type_is_dynptr(arg_type)) { 7968 struct bpf_func_state *state = func(env, reg); 7969 int spi; 7970 7971 /* Only dynptr created on stack can be released, thus 7972 * the get_spi and stack state checks for spilled_ptr 7973 * should only be done before process_dynptr_func for 7974 * PTR_TO_STACK. 7975 */ 7976 if (reg->type == PTR_TO_STACK) { 7977 spi = dynptr_get_spi(env, reg); 7978 if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) { 7979 verbose(env, "arg %d is an unacquired reference\n", regno); 7980 return -EINVAL; 7981 } 7982 } else { 7983 verbose(env, "cannot release unowned const bpf_dynptr\n"); 7984 return -EINVAL; 7985 } 7986 } else if (!reg->ref_obj_id && !register_is_null(reg)) { 7987 verbose(env, "R%d must be referenced when passed to release function\n", 7988 regno); 7989 return -EINVAL; 7990 } 7991 if (meta->release_regno) { 7992 verbose(env, "verifier internal error: more than one release argument\n"); 7993 return -EFAULT; 7994 } 7995 meta->release_regno = regno; 7996 } 7997 7998 if (reg->ref_obj_id) { 7999 if (meta->ref_obj_id) { 8000 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n", 8001 regno, reg->ref_obj_id, 8002 meta->ref_obj_id); 8003 return -EFAULT; 8004 } 8005 meta->ref_obj_id = reg->ref_obj_id; 8006 } 8007 8008 switch (base_type(arg_type)) { 8009 case ARG_CONST_MAP_PTR: 8010 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */ 8011 if (meta->map_ptr) { 8012 /* Use map_uid (which is unique id of inner map) to reject: 8013 * inner_map1 = bpf_map_lookup_elem(outer_map, key1) 8014 * inner_map2 = bpf_map_lookup_elem(outer_map, key2) 8015 * if (inner_map1 && inner_map2) { 8016 * timer = bpf_map_lookup_elem(inner_map1); 8017 * if (timer) 8018 * // mismatch would have been allowed 8019 * bpf_timer_init(timer, inner_map2); 8020 * } 8021 * 8022 * Comparing map_ptr is enough to distinguish normal and outer maps. 8023 */ 8024 if (meta->map_ptr != reg->map_ptr || 8025 meta->map_uid != reg->map_uid) { 8026 verbose(env, 8027 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n", 8028 meta->map_uid, reg->map_uid); 8029 return -EINVAL; 8030 } 8031 } 8032 meta->map_ptr = reg->map_ptr; 8033 meta->map_uid = reg->map_uid; 8034 break; 8035 case ARG_PTR_TO_MAP_KEY: 8036 /* bpf_map_xxx(..., map_ptr, ..., key) call: 8037 * check that [key, key + map->key_size) are within 8038 * stack limits and initialized 8039 */ 8040 if (!meta->map_ptr) { 8041 /* in function declaration map_ptr must come before 8042 * map_key, so that it's verified and known before 8043 * we have to check map_key here. Otherwise it means 8044 * that kernel subsystem misconfigured verifier 8045 */ 8046 verbose(env, "invalid map_ptr to access map->key\n"); 8047 return -EACCES; 8048 } 8049 err = check_helper_mem_access(env, regno, 8050 meta->map_ptr->key_size, false, 8051 NULL); 8052 break; 8053 case ARG_PTR_TO_MAP_VALUE: 8054 if (type_may_be_null(arg_type) && register_is_null(reg)) 8055 return 0; 8056 8057 /* bpf_map_xxx(..., map_ptr, ..., value) call: 8058 * check [value, value + map->value_size) validity 8059 */ 8060 if (!meta->map_ptr) { 8061 /* kernel subsystem misconfigured verifier */ 8062 verbose(env, "invalid map_ptr to access map->value\n"); 8063 return -EACCES; 8064 } 8065 meta->raw_mode = arg_type & MEM_UNINIT; 8066 err = check_helper_mem_access(env, regno, 8067 meta->map_ptr->value_size, false, 8068 meta); 8069 break; 8070 case ARG_PTR_TO_PERCPU_BTF_ID: 8071 if (!reg->btf_id) { 8072 verbose(env, "Helper has invalid btf_id in R%d\n", regno); 8073 return -EACCES; 8074 } 8075 meta->ret_btf = reg->btf; 8076 meta->ret_btf_id = reg->btf_id; 8077 break; 8078 case ARG_PTR_TO_SPIN_LOCK: 8079 if (in_rbtree_lock_required_cb(env)) { 8080 verbose(env, "can't spin_{lock,unlock} in rbtree cb\n"); 8081 return -EACCES; 8082 } 8083 if (meta->func_id == BPF_FUNC_spin_lock) { 8084 err = process_spin_lock(env, regno, true); 8085 if (err) 8086 return err; 8087 } else if (meta->func_id == BPF_FUNC_spin_unlock) { 8088 err = process_spin_lock(env, regno, false); 8089 if (err) 8090 return err; 8091 } else { 8092 verbose(env, "verifier internal error\n"); 8093 return -EFAULT; 8094 } 8095 break; 8096 case ARG_PTR_TO_TIMER: 8097 err = process_timer_func(env, regno, meta); 8098 if (err) 8099 return err; 8100 break; 8101 case ARG_PTR_TO_FUNC: 8102 meta->subprogno = reg->subprogno; 8103 break; 8104 case ARG_PTR_TO_MEM: 8105 /* The access to this pointer is only checked when we hit the 8106 * next is_mem_size argument below. 8107 */ 8108 meta->raw_mode = arg_type & MEM_UNINIT; 8109 if (arg_type & MEM_FIXED_SIZE) { 8110 err = check_helper_mem_access(env, regno, 8111 fn->arg_size[arg], false, 8112 meta); 8113 } 8114 break; 8115 case ARG_CONST_SIZE: 8116 err = check_mem_size_reg(env, reg, regno, false, meta); 8117 break; 8118 case ARG_CONST_SIZE_OR_ZERO: 8119 err = check_mem_size_reg(env, reg, regno, true, meta); 8120 break; 8121 case ARG_PTR_TO_DYNPTR: 8122 err = process_dynptr_func(env, regno, insn_idx, arg_type, 0); 8123 if (err) 8124 return err; 8125 break; 8126 case ARG_CONST_ALLOC_SIZE_OR_ZERO: 8127 if (!tnum_is_const(reg->var_off)) { 8128 verbose(env, "R%d is not a known constant'\n", 8129 regno); 8130 return -EACCES; 8131 } 8132 meta->mem_size = reg->var_off.value; 8133 err = mark_chain_precision(env, regno); 8134 if (err) 8135 return err; 8136 break; 8137 case ARG_PTR_TO_INT: 8138 case ARG_PTR_TO_LONG: 8139 { 8140 int size = int_ptr_type_to_size(arg_type); 8141 8142 err = check_helper_mem_access(env, regno, size, false, meta); 8143 if (err) 8144 return err; 8145 err = check_ptr_alignment(env, reg, 0, size, true); 8146 break; 8147 } 8148 case ARG_PTR_TO_CONST_STR: 8149 { 8150 struct bpf_map *map = reg->map_ptr; 8151 int map_off; 8152 u64 map_addr; 8153 char *str_ptr; 8154 8155 if (!bpf_map_is_rdonly(map)) { 8156 verbose(env, "R%d does not point to a readonly map'\n", regno); 8157 return -EACCES; 8158 } 8159 8160 if (!tnum_is_const(reg->var_off)) { 8161 verbose(env, "R%d is not a constant address'\n", regno); 8162 return -EACCES; 8163 } 8164 8165 if (!map->ops->map_direct_value_addr) { 8166 verbose(env, "no direct value access support for this map type\n"); 8167 return -EACCES; 8168 } 8169 8170 err = check_map_access(env, regno, reg->off, 8171 map->value_size - reg->off, false, 8172 ACCESS_HELPER); 8173 if (err) 8174 return err; 8175 8176 map_off = reg->off + reg->var_off.value; 8177 err = map->ops->map_direct_value_addr(map, &map_addr, map_off); 8178 if (err) { 8179 verbose(env, "direct value access on string failed\n"); 8180 return err; 8181 } 8182 8183 str_ptr = (char *)(long)(map_addr); 8184 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) { 8185 verbose(env, "string is not zero-terminated\n"); 8186 return -EINVAL; 8187 } 8188 break; 8189 } 8190 case ARG_PTR_TO_KPTR: 8191 err = process_kptr_func(env, regno, meta); 8192 if (err) 8193 return err; 8194 break; 8195 } 8196 8197 return err; 8198 } 8199 8200 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id) 8201 { 8202 enum bpf_attach_type eatype = env->prog->expected_attach_type; 8203 enum bpf_prog_type type = resolve_prog_type(env->prog); 8204 8205 if (func_id != BPF_FUNC_map_update_elem) 8206 return false; 8207 8208 /* It's not possible to get access to a locked struct sock in these 8209 * contexts, so updating is safe. 8210 */ 8211 switch (type) { 8212 case BPF_PROG_TYPE_TRACING: 8213 if (eatype == BPF_TRACE_ITER) 8214 return true; 8215 break; 8216 case BPF_PROG_TYPE_SOCKET_FILTER: 8217 case BPF_PROG_TYPE_SCHED_CLS: 8218 case BPF_PROG_TYPE_SCHED_ACT: 8219 case BPF_PROG_TYPE_XDP: 8220 case BPF_PROG_TYPE_SK_REUSEPORT: 8221 case BPF_PROG_TYPE_FLOW_DISSECTOR: 8222 case BPF_PROG_TYPE_SK_LOOKUP: 8223 return true; 8224 default: 8225 break; 8226 } 8227 8228 verbose(env, "cannot update sockmap in this context\n"); 8229 return false; 8230 } 8231 8232 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env) 8233 { 8234 return env->prog->jit_requested && 8235 bpf_jit_supports_subprog_tailcalls(); 8236 } 8237 8238 static int check_map_func_compatibility(struct bpf_verifier_env *env, 8239 struct bpf_map *map, int func_id) 8240 { 8241 if (!map) 8242 return 0; 8243 8244 /* We need a two way check, first is from map perspective ... */ 8245 switch (map->map_type) { 8246 case BPF_MAP_TYPE_PROG_ARRAY: 8247 if (func_id != BPF_FUNC_tail_call) 8248 goto error; 8249 break; 8250 case BPF_MAP_TYPE_PERF_EVENT_ARRAY: 8251 if (func_id != BPF_FUNC_perf_event_read && 8252 func_id != BPF_FUNC_perf_event_output && 8253 func_id != BPF_FUNC_skb_output && 8254 func_id != BPF_FUNC_perf_event_read_value && 8255 func_id != BPF_FUNC_xdp_output) 8256 goto error; 8257 break; 8258 case BPF_MAP_TYPE_RINGBUF: 8259 if (func_id != BPF_FUNC_ringbuf_output && 8260 func_id != BPF_FUNC_ringbuf_reserve && 8261 func_id != BPF_FUNC_ringbuf_query && 8262 func_id != BPF_FUNC_ringbuf_reserve_dynptr && 8263 func_id != BPF_FUNC_ringbuf_submit_dynptr && 8264 func_id != BPF_FUNC_ringbuf_discard_dynptr) 8265 goto error; 8266 break; 8267 case BPF_MAP_TYPE_USER_RINGBUF: 8268 if (func_id != BPF_FUNC_user_ringbuf_drain) 8269 goto error; 8270 break; 8271 case BPF_MAP_TYPE_STACK_TRACE: 8272 if (func_id != BPF_FUNC_get_stackid) 8273 goto error; 8274 break; 8275 case BPF_MAP_TYPE_CGROUP_ARRAY: 8276 if (func_id != BPF_FUNC_skb_under_cgroup && 8277 func_id != BPF_FUNC_current_task_under_cgroup) 8278 goto error; 8279 break; 8280 case BPF_MAP_TYPE_CGROUP_STORAGE: 8281 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE: 8282 if (func_id != BPF_FUNC_get_local_storage) 8283 goto error; 8284 break; 8285 case BPF_MAP_TYPE_DEVMAP: 8286 case BPF_MAP_TYPE_DEVMAP_HASH: 8287 if (func_id != BPF_FUNC_redirect_map && 8288 func_id != BPF_FUNC_map_lookup_elem) 8289 goto error; 8290 break; 8291 /* Restrict bpf side of cpumap and xskmap, open when use-cases 8292 * appear. 8293 */ 8294 case BPF_MAP_TYPE_CPUMAP: 8295 if (func_id != BPF_FUNC_redirect_map) 8296 goto error; 8297 break; 8298 case BPF_MAP_TYPE_XSKMAP: 8299 if (func_id != BPF_FUNC_redirect_map && 8300 func_id != BPF_FUNC_map_lookup_elem) 8301 goto error; 8302 break; 8303 case BPF_MAP_TYPE_ARRAY_OF_MAPS: 8304 case BPF_MAP_TYPE_HASH_OF_MAPS: 8305 if (func_id != BPF_FUNC_map_lookup_elem) 8306 goto error; 8307 break; 8308 case BPF_MAP_TYPE_SOCKMAP: 8309 if (func_id != BPF_FUNC_sk_redirect_map && 8310 func_id != BPF_FUNC_sock_map_update && 8311 func_id != BPF_FUNC_map_delete_elem && 8312 func_id != BPF_FUNC_msg_redirect_map && 8313 func_id != BPF_FUNC_sk_select_reuseport && 8314 func_id != BPF_FUNC_map_lookup_elem && 8315 !may_update_sockmap(env, func_id)) 8316 goto error; 8317 break; 8318 case BPF_MAP_TYPE_SOCKHASH: 8319 if (func_id != BPF_FUNC_sk_redirect_hash && 8320 func_id != BPF_FUNC_sock_hash_update && 8321 func_id != BPF_FUNC_map_delete_elem && 8322 func_id != BPF_FUNC_msg_redirect_hash && 8323 func_id != BPF_FUNC_sk_select_reuseport && 8324 func_id != BPF_FUNC_map_lookup_elem && 8325 !may_update_sockmap(env, func_id)) 8326 goto error; 8327 break; 8328 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY: 8329 if (func_id != BPF_FUNC_sk_select_reuseport) 8330 goto error; 8331 break; 8332 case BPF_MAP_TYPE_QUEUE: 8333 case BPF_MAP_TYPE_STACK: 8334 if (func_id != BPF_FUNC_map_peek_elem && 8335 func_id != BPF_FUNC_map_pop_elem && 8336 func_id != BPF_FUNC_map_push_elem) 8337 goto error; 8338 break; 8339 case BPF_MAP_TYPE_SK_STORAGE: 8340 if (func_id != BPF_FUNC_sk_storage_get && 8341 func_id != BPF_FUNC_sk_storage_delete && 8342 func_id != BPF_FUNC_kptr_xchg) 8343 goto error; 8344 break; 8345 case BPF_MAP_TYPE_INODE_STORAGE: 8346 if (func_id != BPF_FUNC_inode_storage_get && 8347 func_id != BPF_FUNC_inode_storage_delete && 8348 func_id != BPF_FUNC_kptr_xchg) 8349 goto error; 8350 break; 8351 case BPF_MAP_TYPE_TASK_STORAGE: 8352 if (func_id != BPF_FUNC_task_storage_get && 8353 func_id != BPF_FUNC_task_storage_delete && 8354 func_id != BPF_FUNC_kptr_xchg) 8355 goto error; 8356 break; 8357 case BPF_MAP_TYPE_CGRP_STORAGE: 8358 if (func_id != BPF_FUNC_cgrp_storage_get && 8359 func_id != BPF_FUNC_cgrp_storage_delete && 8360 func_id != BPF_FUNC_kptr_xchg) 8361 goto error; 8362 break; 8363 case BPF_MAP_TYPE_BLOOM_FILTER: 8364 if (func_id != BPF_FUNC_map_peek_elem && 8365 func_id != BPF_FUNC_map_push_elem) 8366 goto error; 8367 break; 8368 default: 8369 break; 8370 } 8371 8372 /* ... and second from the function itself. */ 8373 switch (func_id) { 8374 case BPF_FUNC_tail_call: 8375 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY) 8376 goto error; 8377 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) { 8378 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n"); 8379 return -EINVAL; 8380 } 8381 break; 8382 case BPF_FUNC_perf_event_read: 8383 case BPF_FUNC_perf_event_output: 8384 case BPF_FUNC_perf_event_read_value: 8385 case BPF_FUNC_skb_output: 8386 case BPF_FUNC_xdp_output: 8387 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY) 8388 goto error; 8389 break; 8390 case BPF_FUNC_ringbuf_output: 8391 case BPF_FUNC_ringbuf_reserve: 8392 case BPF_FUNC_ringbuf_query: 8393 case BPF_FUNC_ringbuf_reserve_dynptr: 8394 case BPF_FUNC_ringbuf_submit_dynptr: 8395 case BPF_FUNC_ringbuf_discard_dynptr: 8396 if (map->map_type != BPF_MAP_TYPE_RINGBUF) 8397 goto error; 8398 break; 8399 case BPF_FUNC_user_ringbuf_drain: 8400 if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF) 8401 goto error; 8402 break; 8403 case BPF_FUNC_get_stackid: 8404 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE) 8405 goto error; 8406 break; 8407 case BPF_FUNC_current_task_under_cgroup: 8408 case BPF_FUNC_skb_under_cgroup: 8409 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY) 8410 goto error; 8411 break; 8412 case BPF_FUNC_redirect_map: 8413 if (map->map_type != BPF_MAP_TYPE_DEVMAP && 8414 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH && 8415 map->map_type != BPF_MAP_TYPE_CPUMAP && 8416 map->map_type != BPF_MAP_TYPE_XSKMAP) 8417 goto error; 8418 break; 8419 case BPF_FUNC_sk_redirect_map: 8420 case BPF_FUNC_msg_redirect_map: 8421 case BPF_FUNC_sock_map_update: 8422 if (map->map_type != BPF_MAP_TYPE_SOCKMAP) 8423 goto error; 8424 break; 8425 case BPF_FUNC_sk_redirect_hash: 8426 case BPF_FUNC_msg_redirect_hash: 8427 case BPF_FUNC_sock_hash_update: 8428 if (map->map_type != BPF_MAP_TYPE_SOCKHASH) 8429 goto error; 8430 break; 8431 case BPF_FUNC_get_local_storage: 8432 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE && 8433 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE) 8434 goto error; 8435 break; 8436 case BPF_FUNC_sk_select_reuseport: 8437 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY && 8438 map->map_type != BPF_MAP_TYPE_SOCKMAP && 8439 map->map_type != BPF_MAP_TYPE_SOCKHASH) 8440 goto error; 8441 break; 8442 case BPF_FUNC_map_pop_elem: 8443 if (map->map_type != BPF_MAP_TYPE_QUEUE && 8444 map->map_type != BPF_MAP_TYPE_STACK) 8445 goto error; 8446 break; 8447 case BPF_FUNC_map_peek_elem: 8448 case BPF_FUNC_map_push_elem: 8449 if (map->map_type != BPF_MAP_TYPE_QUEUE && 8450 map->map_type != BPF_MAP_TYPE_STACK && 8451 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER) 8452 goto error; 8453 break; 8454 case BPF_FUNC_map_lookup_percpu_elem: 8455 if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY && 8456 map->map_type != BPF_MAP_TYPE_PERCPU_HASH && 8457 map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH) 8458 goto error; 8459 break; 8460 case BPF_FUNC_sk_storage_get: 8461 case BPF_FUNC_sk_storage_delete: 8462 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE) 8463 goto error; 8464 break; 8465 case BPF_FUNC_inode_storage_get: 8466 case BPF_FUNC_inode_storage_delete: 8467 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE) 8468 goto error; 8469 break; 8470 case BPF_FUNC_task_storage_get: 8471 case BPF_FUNC_task_storage_delete: 8472 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE) 8473 goto error; 8474 break; 8475 case BPF_FUNC_cgrp_storage_get: 8476 case BPF_FUNC_cgrp_storage_delete: 8477 if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE) 8478 goto error; 8479 break; 8480 default: 8481 break; 8482 } 8483 8484 return 0; 8485 error: 8486 verbose(env, "cannot pass map_type %d into func %s#%d\n", 8487 map->map_type, func_id_name(func_id), func_id); 8488 return -EINVAL; 8489 } 8490 8491 static bool check_raw_mode_ok(const struct bpf_func_proto *fn) 8492 { 8493 int count = 0; 8494 8495 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM) 8496 count++; 8497 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM) 8498 count++; 8499 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM) 8500 count++; 8501 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM) 8502 count++; 8503 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM) 8504 count++; 8505 8506 /* We only support one arg being in raw mode at the moment, 8507 * which is sufficient for the helper functions we have 8508 * right now. 8509 */ 8510 return count <= 1; 8511 } 8512 8513 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg) 8514 { 8515 bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE; 8516 bool has_size = fn->arg_size[arg] != 0; 8517 bool is_next_size = false; 8518 8519 if (arg + 1 < ARRAY_SIZE(fn->arg_type)) 8520 is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]); 8521 8522 if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM) 8523 return is_next_size; 8524 8525 return has_size == is_next_size || is_next_size == is_fixed; 8526 } 8527 8528 static bool check_arg_pair_ok(const struct bpf_func_proto *fn) 8529 { 8530 /* bpf_xxx(..., buf, len) call will access 'len' 8531 * bytes from memory 'buf'. Both arg types need 8532 * to be paired, so make sure there's no buggy 8533 * helper function specification. 8534 */ 8535 if (arg_type_is_mem_size(fn->arg1_type) || 8536 check_args_pair_invalid(fn, 0) || 8537 check_args_pair_invalid(fn, 1) || 8538 check_args_pair_invalid(fn, 2) || 8539 check_args_pair_invalid(fn, 3) || 8540 check_args_pair_invalid(fn, 4)) 8541 return false; 8542 8543 return true; 8544 } 8545 8546 static bool check_btf_id_ok(const struct bpf_func_proto *fn) 8547 { 8548 int i; 8549 8550 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) { 8551 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID) 8552 return !!fn->arg_btf_id[i]; 8553 if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK) 8554 return fn->arg_btf_id[i] == BPF_PTR_POISON; 8555 if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] && 8556 /* arg_btf_id and arg_size are in a union. */ 8557 (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM || 8558 !(fn->arg_type[i] & MEM_FIXED_SIZE))) 8559 return false; 8560 } 8561 8562 return true; 8563 } 8564 8565 static int check_func_proto(const struct bpf_func_proto *fn, int func_id) 8566 { 8567 return check_raw_mode_ok(fn) && 8568 check_arg_pair_ok(fn) && 8569 check_btf_id_ok(fn) ? 0 : -EINVAL; 8570 } 8571 8572 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END] 8573 * are now invalid, so turn them into unknown SCALAR_VALUE. 8574 * 8575 * This also applies to dynptr slices belonging to skb and xdp dynptrs, 8576 * since these slices point to packet data. 8577 */ 8578 static void clear_all_pkt_pointers(struct bpf_verifier_env *env) 8579 { 8580 struct bpf_func_state *state; 8581 struct bpf_reg_state *reg; 8582 8583 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({ 8584 if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg)) 8585 mark_reg_invalid(env, reg); 8586 })); 8587 } 8588 8589 enum { 8590 AT_PKT_END = -1, 8591 BEYOND_PKT_END = -2, 8592 }; 8593 8594 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open) 8595 { 8596 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 8597 struct bpf_reg_state *reg = &state->regs[regn]; 8598 8599 if (reg->type != PTR_TO_PACKET) 8600 /* PTR_TO_PACKET_META is not supported yet */ 8601 return; 8602 8603 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end. 8604 * How far beyond pkt_end it goes is unknown. 8605 * if (!range_open) it's the case of pkt >= pkt_end 8606 * if (range_open) it's the case of pkt > pkt_end 8607 * hence this pointer is at least 1 byte bigger than pkt_end 8608 */ 8609 if (range_open) 8610 reg->range = BEYOND_PKT_END; 8611 else 8612 reg->range = AT_PKT_END; 8613 } 8614 8615 /* The pointer with the specified id has released its reference to kernel 8616 * resources. Identify all copies of the same pointer and clear the reference. 8617 */ 8618 static int release_reference(struct bpf_verifier_env *env, 8619 int ref_obj_id) 8620 { 8621 struct bpf_func_state *state; 8622 struct bpf_reg_state *reg; 8623 int err; 8624 8625 err = release_reference_state(cur_func(env), ref_obj_id); 8626 if (err) 8627 return err; 8628 8629 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({ 8630 if (reg->ref_obj_id == ref_obj_id) 8631 mark_reg_invalid(env, reg); 8632 })); 8633 8634 return 0; 8635 } 8636 8637 static void invalidate_non_owning_refs(struct bpf_verifier_env *env) 8638 { 8639 struct bpf_func_state *unused; 8640 struct bpf_reg_state *reg; 8641 8642 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({ 8643 if (type_is_non_owning_ref(reg->type)) 8644 mark_reg_invalid(env, reg); 8645 })); 8646 } 8647 8648 static void clear_caller_saved_regs(struct bpf_verifier_env *env, 8649 struct bpf_reg_state *regs) 8650 { 8651 int i; 8652 8653 /* after the call registers r0 - r5 were scratched */ 8654 for (i = 0; i < CALLER_SAVED_REGS; i++) { 8655 mark_reg_not_init(env, regs, caller_saved[i]); 8656 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK); 8657 } 8658 } 8659 8660 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env, 8661 struct bpf_func_state *caller, 8662 struct bpf_func_state *callee, 8663 int insn_idx); 8664 8665 static int set_callee_state(struct bpf_verifier_env *env, 8666 struct bpf_func_state *caller, 8667 struct bpf_func_state *callee, int insn_idx); 8668 8669 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn, 8670 int *insn_idx, int subprog, 8671 set_callee_state_fn set_callee_state_cb) 8672 { 8673 struct bpf_verifier_state *state = env->cur_state; 8674 struct bpf_func_state *caller, *callee; 8675 int err; 8676 8677 if (state->curframe + 1 >= MAX_CALL_FRAMES) { 8678 verbose(env, "the call stack of %d frames is too deep\n", 8679 state->curframe + 2); 8680 return -E2BIG; 8681 } 8682 8683 caller = state->frame[state->curframe]; 8684 if (state->frame[state->curframe + 1]) { 8685 verbose(env, "verifier bug. Frame %d already allocated\n", 8686 state->curframe + 1); 8687 return -EFAULT; 8688 } 8689 8690 err = btf_check_subprog_call(env, subprog, caller->regs); 8691 if (err == -EFAULT) 8692 return err; 8693 if (subprog_is_global(env, subprog)) { 8694 if (err) { 8695 verbose(env, "Caller passes invalid args into func#%d\n", 8696 subprog); 8697 return err; 8698 } else { 8699 if (env->log.level & BPF_LOG_LEVEL) 8700 verbose(env, 8701 "Func#%d is global and valid. Skipping.\n", 8702 subprog); 8703 clear_caller_saved_regs(env, caller->regs); 8704 8705 /* All global functions return a 64-bit SCALAR_VALUE */ 8706 mark_reg_unknown(env, caller->regs, BPF_REG_0); 8707 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG; 8708 8709 /* continue with next insn after call */ 8710 return 0; 8711 } 8712 } 8713 8714 /* set_callee_state is used for direct subprog calls, but we are 8715 * interested in validating only BPF helpers that can call subprogs as 8716 * callbacks 8717 */ 8718 if (set_callee_state_cb != set_callee_state) { 8719 if (bpf_pseudo_kfunc_call(insn) && 8720 !is_callback_calling_kfunc(insn->imm)) { 8721 verbose(env, "verifier bug: kfunc %s#%d not marked as callback-calling\n", 8722 func_id_name(insn->imm), insn->imm); 8723 return -EFAULT; 8724 } else if (!bpf_pseudo_kfunc_call(insn) && 8725 !is_callback_calling_function(insn->imm)) { /* helper */ 8726 verbose(env, "verifier bug: helper %s#%d not marked as callback-calling\n", 8727 func_id_name(insn->imm), insn->imm); 8728 return -EFAULT; 8729 } 8730 } 8731 8732 if (insn->code == (BPF_JMP | BPF_CALL) && 8733 insn->src_reg == 0 && 8734 insn->imm == BPF_FUNC_timer_set_callback) { 8735 struct bpf_verifier_state *async_cb; 8736 8737 /* there is no real recursion here. timer callbacks are async */ 8738 env->subprog_info[subprog].is_async_cb = true; 8739 async_cb = push_async_cb(env, env->subprog_info[subprog].start, 8740 *insn_idx, subprog); 8741 if (!async_cb) 8742 return -EFAULT; 8743 callee = async_cb->frame[0]; 8744 callee->async_entry_cnt = caller->async_entry_cnt + 1; 8745 8746 /* Convert bpf_timer_set_callback() args into timer callback args */ 8747 err = set_callee_state_cb(env, caller, callee, *insn_idx); 8748 if (err) 8749 return err; 8750 8751 clear_caller_saved_regs(env, caller->regs); 8752 mark_reg_unknown(env, caller->regs, BPF_REG_0); 8753 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG; 8754 /* continue with next insn after call */ 8755 return 0; 8756 } 8757 8758 callee = kzalloc(sizeof(*callee), GFP_KERNEL); 8759 if (!callee) 8760 return -ENOMEM; 8761 state->frame[state->curframe + 1] = callee; 8762 8763 /* callee cannot access r0, r6 - r9 for reading and has to write 8764 * into its own stack before reading from it. 8765 * callee can read/write into caller's stack 8766 */ 8767 init_func_state(env, callee, 8768 /* remember the callsite, it will be used by bpf_exit */ 8769 *insn_idx /* callsite */, 8770 state->curframe + 1 /* frameno within this callchain */, 8771 subprog /* subprog number within this prog */); 8772 8773 /* Transfer references to the callee */ 8774 err = copy_reference_state(callee, caller); 8775 if (err) 8776 goto err_out; 8777 8778 err = set_callee_state_cb(env, caller, callee, *insn_idx); 8779 if (err) 8780 goto err_out; 8781 8782 clear_caller_saved_regs(env, caller->regs); 8783 8784 /* only increment it after check_reg_arg() finished */ 8785 state->curframe++; 8786 8787 /* and go analyze first insn of the callee */ 8788 *insn_idx = env->subprog_info[subprog].start - 1; 8789 8790 if (env->log.level & BPF_LOG_LEVEL) { 8791 verbose(env, "caller:\n"); 8792 print_verifier_state(env, caller, true); 8793 verbose(env, "callee:\n"); 8794 print_verifier_state(env, callee, true); 8795 } 8796 return 0; 8797 8798 err_out: 8799 free_func_state(callee); 8800 state->frame[state->curframe + 1] = NULL; 8801 return err; 8802 } 8803 8804 int map_set_for_each_callback_args(struct bpf_verifier_env *env, 8805 struct bpf_func_state *caller, 8806 struct bpf_func_state *callee) 8807 { 8808 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn, 8809 * void *callback_ctx, u64 flags); 8810 * callback_fn(struct bpf_map *map, void *key, void *value, 8811 * void *callback_ctx); 8812 */ 8813 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1]; 8814 8815 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY; 8816 __mark_reg_known_zero(&callee->regs[BPF_REG_2]); 8817 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr; 8818 8819 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE; 8820 __mark_reg_known_zero(&callee->regs[BPF_REG_3]); 8821 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr; 8822 8823 /* pointer to stack or null */ 8824 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3]; 8825 8826 /* unused */ 8827 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]); 8828 return 0; 8829 } 8830 8831 static int set_callee_state(struct bpf_verifier_env *env, 8832 struct bpf_func_state *caller, 8833 struct bpf_func_state *callee, int insn_idx) 8834 { 8835 int i; 8836 8837 /* copy r1 - r5 args that callee can access. The copy includes parent 8838 * pointers, which connects us up to the liveness chain 8839 */ 8840 for (i = BPF_REG_1; i <= BPF_REG_5; i++) 8841 callee->regs[i] = caller->regs[i]; 8842 return 0; 8843 } 8844 8845 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn, 8846 int *insn_idx) 8847 { 8848 int subprog, target_insn; 8849 8850 target_insn = *insn_idx + insn->imm + 1; 8851 subprog = find_subprog(env, target_insn); 8852 if (subprog < 0) { 8853 verbose(env, "verifier bug. No program starts at insn %d\n", 8854 target_insn); 8855 return -EFAULT; 8856 } 8857 8858 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state); 8859 } 8860 8861 static int set_map_elem_callback_state(struct bpf_verifier_env *env, 8862 struct bpf_func_state *caller, 8863 struct bpf_func_state *callee, 8864 int insn_idx) 8865 { 8866 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx]; 8867 struct bpf_map *map; 8868 int err; 8869 8870 if (bpf_map_ptr_poisoned(insn_aux)) { 8871 verbose(env, "tail_call abusing map_ptr\n"); 8872 return -EINVAL; 8873 } 8874 8875 map = BPF_MAP_PTR(insn_aux->map_ptr_state); 8876 if (!map->ops->map_set_for_each_callback_args || 8877 !map->ops->map_for_each_callback) { 8878 verbose(env, "callback function not allowed for map\n"); 8879 return -ENOTSUPP; 8880 } 8881 8882 err = map->ops->map_set_for_each_callback_args(env, caller, callee); 8883 if (err) 8884 return err; 8885 8886 callee->in_callback_fn = true; 8887 callee->callback_ret_range = tnum_range(0, 1); 8888 return 0; 8889 } 8890 8891 static int set_loop_callback_state(struct bpf_verifier_env *env, 8892 struct bpf_func_state *caller, 8893 struct bpf_func_state *callee, 8894 int insn_idx) 8895 { 8896 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx, 8897 * u64 flags); 8898 * callback_fn(u32 index, void *callback_ctx); 8899 */ 8900 callee->regs[BPF_REG_1].type = SCALAR_VALUE; 8901 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3]; 8902 8903 /* unused */ 8904 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]); 8905 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]); 8906 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]); 8907 8908 callee->in_callback_fn = true; 8909 callee->callback_ret_range = tnum_range(0, 1); 8910 return 0; 8911 } 8912 8913 static int set_timer_callback_state(struct bpf_verifier_env *env, 8914 struct bpf_func_state *caller, 8915 struct bpf_func_state *callee, 8916 int insn_idx) 8917 { 8918 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr; 8919 8920 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn); 8921 * callback_fn(struct bpf_map *map, void *key, void *value); 8922 */ 8923 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP; 8924 __mark_reg_known_zero(&callee->regs[BPF_REG_1]); 8925 callee->regs[BPF_REG_1].map_ptr = map_ptr; 8926 8927 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY; 8928 __mark_reg_known_zero(&callee->regs[BPF_REG_2]); 8929 callee->regs[BPF_REG_2].map_ptr = map_ptr; 8930 8931 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE; 8932 __mark_reg_known_zero(&callee->regs[BPF_REG_3]); 8933 callee->regs[BPF_REG_3].map_ptr = map_ptr; 8934 8935 /* unused */ 8936 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]); 8937 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]); 8938 callee->in_async_callback_fn = true; 8939 callee->callback_ret_range = tnum_range(0, 1); 8940 return 0; 8941 } 8942 8943 static int set_find_vma_callback_state(struct bpf_verifier_env *env, 8944 struct bpf_func_state *caller, 8945 struct bpf_func_state *callee, 8946 int insn_idx) 8947 { 8948 /* bpf_find_vma(struct task_struct *task, u64 addr, 8949 * void *callback_fn, void *callback_ctx, u64 flags) 8950 * (callback_fn)(struct task_struct *task, 8951 * struct vm_area_struct *vma, void *callback_ctx); 8952 */ 8953 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1]; 8954 8955 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID; 8956 __mark_reg_known_zero(&callee->regs[BPF_REG_2]); 8957 callee->regs[BPF_REG_2].btf = btf_vmlinux; 8958 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA], 8959 8960 /* pointer to stack or null */ 8961 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4]; 8962 8963 /* unused */ 8964 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]); 8965 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]); 8966 callee->in_callback_fn = true; 8967 callee->callback_ret_range = tnum_range(0, 1); 8968 return 0; 8969 } 8970 8971 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env, 8972 struct bpf_func_state *caller, 8973 struct bpf_func_state *callee, 8974 int insn_idx) 8975 { 8976 /* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void 8977 * callback_ctx, u64 flags); 8978 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx); 8979 */ 8980 __mark_reg_not_init(env, &callee->regs[BPF_REG_0]); 8981 mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL); 8982 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3]; 8983 8984 /* unused */ 8985 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]); 8986 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]); 8987 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]); 8988 8989 callee->in_callback_fn = true; 8990 callee->callback_ret_range = tnum_range(0, 1); 8991 return 0; 8992 } 8993 8994 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env, 8995 struct bpf_func_state *caller, 8996 struct bpf_func_state *callee, 8997 int insn_idx) 8998 { 8999 /* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node, 9000 * bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b)); 9001 * 9002 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset 9003 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd 9004 * by this point, so look at 'root' 9005 */ 9006 struct btf_field *field; 9007 9008 field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off, 9009 BPF_RB_ROOT); 9010 if (!field || !field->graph_root.value_btf_id) 9011 return -EFAULT; 9012 9013 mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root); 9014 ref_set_non_owning(env, &callee->regs[BPF_REG_1]); 9015 mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root); 9016 ref_set_non_owning(env, &callee->regs[BPF_REG_2]); 9017 9018 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]); 9019 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]); 9020 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]); 9021 callee->in_callback_fn = true; 9022 callee->callback_ret_range = tnum_range(0, 1); 9023 return 0; 9024 } 9025 9026 static bool is_rbtree_lock_required_kfunc(u32 btf_id); 9027 9028 /* Are we currently verifying the callback for a rbtree helper that must 9029 * be called with lock held? If so, no need to complain about unreleased 9030 * lock 9031 */ 9032 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env) 9033 { 9034 struct bpf_verifier_state *state = env->cur_state; 9035 struct bpf_insn *insn = env->prog->insnsi; 9036 struct bpf_func_state *callee; 9037 int kfunc_btf_id; 9038 9039 if (!state->curframe) 9040 return false; 9041 9042 callee = state->frame[state->curframe]; 9043 9044 if (!callee->in_callback_fn) 9045 return false; 9046 9047 kfunc_btf_id = insn[callee->callsite].imm; 9048 return is_rbtree_lock_required_kfunc(kfunc_btf_id); 9049 } 9050 9051 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx) 9052 { 9053 struct bpf_verifier_state *state = env->cur_state; 9054 struct bpf_func_state *caller, *callee; 9055 struct bpf_reg_state *r0; 9056 int err; 9057 9058 callee = state->frame[state->curframe]; 9059 r0 = &callee->regs[BPF_REG_0]; 9060 if (r0->type == PTR_TO_STACK) { 9061 /* technically it's ok to return caller's stack pointer 9062 * (or caller's caller's pointer) back to the caller, 9063 * since these pointers are valid. Only current stack 9064 * pointer will be invalid as soon as function exits, 9065 * but let's be conservative 9066 */ 9067 verbose(env, "cannot return stack pointer to the caller\n"); 9068 return -EINVAL; 9069 } 9070 9071 caller = state->frame[state->curframe - 1]; 9072 if (callee->in_callback_fn) { 9073 /* enforce R0 return value range [0, 1]. */ 9074 struct tnum range = callee->callback_ret_range; 9075 9076 if (r0->type != SCALAR_VALUE) { 9077 verbose(env, "R0 not a scalar value\n"); 9078 return -EACCES; 9079 } 9080 if (!tnum_in(range, r0->var_off)) { 9081 verbose_invalid_scalar(env, r0, &range, "callback return", "R0"); 9082 return -EINVAL; 9083 } 9084 } else { 9085 /* return to the caller whatever r0 had in the callee */ 9086 caller->regs[BPF_REG_0] = *r0; 9087 } 9088 9089 /* callback_fn frame should have released its own additions to parent's 9090 * reference state at this point, or check_reference_leak would 9091 * complain, hence it must be the same as the caller. There is no need 9092 * to copy it back. 9093 */ 9094 if (!callee->in_callback_fn) { 9095 /* Transfer references to the caller */ 9096 err = copy_reference_state(caller, callee); 9097 if (err) 9098 return err; 9099 } 9100 9101 *insn_idx = callee->callsite + 1; 9102 if (env->log.level & BPF_LOG_LEVEL) { 9103 verbose(env, "returning from callee:\n"); 9104 print_verifier_state(env, callee, true); 9105 verbose(env, "to caller at %d:\n", *insn_idx); 9106 print_verifier_state(env, caller, true); 9107 } 9108 /* clear everything in the callee */ 9109 free_func_state(callee); 9110 state->frame[state->curframe--] = NULL; 9111 return 0; 9112 } 9113 9114 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type, 9115 int func_id, 9116 struct bpf_call_arg_meta *meta) 9117 { 9118 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0]; 9119 9120 if (ret_type != RET_INTEGER || 9121 (func_id != BPF_FUNC_get_stack && 9122 func_id != BPF_FUNC_get_task_stack && 9123 func_id != BPF_FUNC_probe_read_str && 9124 func_id != BPF_FUNC_probe_read_kernel_str && 9125 func_id != BPF_FUNC_probe_read_user_str)) 9126 return; 9127 9128 ret_reg->smax_value = meta->msize_max_value; 9129 ret_reg->s32_max_value = meta->msize_max_value; 9130 ret_reg->smin_value = -MAX_ERRNO; 9131 ret_reg->s32_min_value = -MAX_ERRNO; 9132 reg_bounds_sync(ret_reg); 9133 } 9134 9135 static int 9136 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta, 9137 int func_id, int insn_idx) 9138 { 9139 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx]; 9140 struct bpf_map *map = meta->map_ptr; 9141 9142 if (func_id != BPF_FUNC_tail_call && 9143 func_id != BPF_FUNC_map_lookup_elem && 9144 func_id != BPF_FUNC_map_update_elem && 9145 func_id != BPF_FUNC_map_delete_elem && 9146 func_id != BPF_FUNC_map_push_elem && 9147 func_id != BPF_FUNC_map_pop_elem && 9148 func_id != BPF_FUNC_map_peek_elem && 9149 func_id != BPF_FUNC_for_each_map_elem && 9150 func_id != BPF_FUNC_redirect_map && 9151 func_id != BPF_FUNC_map_lookup_percpu_elem) 9152 return 0; 9153 9154 if (map == NULL) { 9155 verbose(env, "kernel subsystem misconfigured verifier\n"); 9156 return -EINVAL; 9157 } 9158 9159 /* In case of read-only, some additional restrictions 9160 * need to be applied in order to prevent altering the 9161 * state of the map from program side. 9162 */ 9163 if ((map->map_flags & BPF_F_RDONLY_PROG) && 9164 (func_id == BPF_FUNC_map_delete_elem || 9165 func_id == BPF_FUNC_map_update_elem || 9166 func_id == BPF_FUNC_map_push_elem || 9167 func_id == BPF_FUNC_map_pop_elem)) { 9168 verbose(env, "write into map forbidden\n"); 9169 return -EACCES; 9170 } 9171 9172 if (!BPF_MAP_PTR(aux->map_ptr_state)) 9173 bpf_map_ptr_store(aux, meta->map_ptr, 9174 !meta->map_ptr->bypass_spec_v1); 9175 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr) 9176 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON, 9177 !meta->map_ptr->bypass_spec_v1); 9178 return 0; 9179 } 9180 9181 static int 9182 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta, 9183 int func_id, int insn_idx) 9184 { 9185 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx]; 9186 struct bpf_reg_state *regs = cur_regs(env), *reg; 9187 struct bpf_map *map = meta->map_ptr; 9188 u64 val, max; 9189 int err; 9190 9191 if (func_id != BPF_FUNC_tail_call) 9192 return 0; 9193 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) { 9194 verbose(env, "kernel subsystem misconfigured verifier\n"); 9195 return -EINVAL; 9196 } 9197 9198 reg = ®s[BPF_REG_3]; 9199 val = reg->var_off.value; 9200 max = map->max_entries; 9201 9202 if (!(register_is_const(reg) && val < max)) { 9203 bpf_map_key_store(aux, BPF_MAP_KEY_POISON); 9204 return 0; 9205 } 9206 9207 err = mark_chain_precision(env, BPF_REG_3); 9208 if (err) 9209 return err; 9210 if (bpf_map_key_unseen(aux)) 9211 bpf_map_key_store(aux, val); 9212 else if (!bpf_map_key_poisoned(aux) && 9213 bpf_map_key_immediate(aux) != val) 9214 bpf_map_key_store(aux, BPF_MAP_KEY_POISON); 9215 return 0; 9216 } 9217 9218 static int check_reference_leak(struct bpf_verifier_env *env) 9219 { 9220 struct bpf_func_state *state = cur_func(env); 9221 bool refs_lingering = false; 9222 int i; 9223 9224 if (state->frameno && !state->in_callback_fn) 9225 return 0; 9226 9227 for (i = 0; i < state->acquired_refs; i++) { 9228 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno) 9229 continue; 9230 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n", 9231 state->refs[i].id, state->refs[i].insn_idx); 9232 refs_lingering = true; 9233 } 9234 return refs_lingering ? -EINVAL : 0; 9235 } 9236 9237 static int check_bpf_snprintf_call(struct bpf_verifier_env *env, 9238 struct bpf_reg_state *regs) 9239 { 9240 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3]; 9241 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5]; 9242 struct bpf_map *fmt_map = fmt_reg->map_ptr; 9243 struct bpf_bprintf_data data = {}; 9244 int err, fmt_map_off, num_args; 9245 u64 fmt_addr; 9246 char *fmt; 9247 9248 /* data must be an array of u64 */ 9249 if (data_len_reg->var_off.value % 8) 9250 return -EINVAL; 9251 num_args = data_len_reg->var_off.value / 8; 9252 9253 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const 9254 * and map_direct_value_addr is set. 9255 */ 9256 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value; 9257 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr, 9258 fmt_map_off); 9259 if (err) { 9260 verbose(env, "verifier bug\n"); 9261 return -EFAULT; 9262 } 9263 fmt = (char *)(long)fmt_addr + fmt_map_off; 9264 9265 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we 9266 * can focus on validating the format specifiers. 9267 */ 9268 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data); 9269 if (err < 0) 9270 verbose(env, "Invalid format string\n"); 9271 9272 return err; 9273 } 9274 9275 static int check_get_func_ip(struct bpf_verifier_env *env) 9276 { 9277 enum bpf_prog_type type = resolve_prog_type(env->prog); 9278 int func_id = BPF_FUNC_get_func_ip; 9279 9280 if (type == BPF_PROG_TYPE_TRACING) { 9281 if (!bpf_prog_has_trampoline(env->prog)) { 9282 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n", 9283 func_id_name(func_id), func_id); 9284 return -ENOTSUPP; 9285 } 9286 return 0; 9287 } else if (type == BPF_PROG_TYPE_KPROBE) { 9288 return 0; 9289 } 9290 9291 verbose(env, "func %s#%d not supported for program type %d\n", 9292 func_id_name(func_id), func_id, type); 9293 return -ENOTSUPP; 9294 } 9295 9296 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env) 9297 { 9298 return &env->insn_aux_data[env->insn_idx]; 9299 } 9300 9301 static bool loop_flag_is_zero(struct bpf_verifier_env *env) 9302 { 9303 struct bpf_reg_state *regs = cur_regs(env); 9304 struct bpf_reg_state *reg = ®s[BPF_REG_4]; 9305 bool reg_is_null = register_is_null(reg); 9306 9307 if (reg_is_null) 9308 mark_chain_precision(env, BPF_REG_4); 9309 9310 return reg_is_null; 9311 } 9312 9313 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno) 9314 { 9315 struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state; 9316 9317 if (!state->initialized) { 9318 state->initialized = 1; 9319 state->fit_for_inline = loop_flag_is_zero(env); 9320 state->callback_subprogno = subprogno; 9321 return; 9322 } 9323 9324 if (!state->fit_for_inline) 9325 return; 9326 9327 state->fit_for_inline = (loop_flag_is_zero(env) && 9328 state->callback_subprogno == subprogno); 9329 } 9330 9331 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn, 9332 int *insn_idx_p) 9333 { 9334 enum bpf_prog_type prog_type = resolve_prog_type(env->prog); 9335 const struct bpf_func_proto *fn = NULL; 9336 enum bpf_return_type ret_type; 9337 enum bpf_type_flag ret_flag; 9338 struct bpf_reg_state *regs; 9339 struct bpf_call_arg_meta meta; 9340 int insn_idx = *insn_idx_p; 9341 bool changes_data; 9342 int i, err, func_id; 9343 9344 /* find function prototype */ 9345 func_id = insn->imm; 9346 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) { 9347 verbose(env, "invalid func %s#%d\n", func_id_name(func_id), 9348 func_id); 9349 return -EINVAL; 9350 } 9351 9352 if (env->ops->get_func_proto) 9353 fn = env->ops->get_func_proto(func_id, env->prog); 9354 if (!fn) { 9355 verbose(env, "unknown func %s#%d\n", func_id_name(func_id), 9356 func_id); 9357 return -EINVAL; 9358 } 9359 9360 /* eBPF programs must be GPL compatible to use GPL-ed functions */ 9361 if (!env->prog->gpl_compatible && fn->gpl_only) { 9362 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n"); 9363 return -EINVAL; 9364 } 9365 9366 if (fn->allowed && !fn->allowed(env->prog)) { 9367 verbose(env, "helper call is not allowed in probe\n"); 9368 return -EINVAL; 9369 } 9370 9371 if (!env->prog->aux->sleepable && fn->might_sleep) { 9372 verbose(env, "helper call might sleep in a non-sleepable prog\n"); 9373 return -EINVAL; 9374 } 9375 9376 /* With LD_ABS/IND some JITs save/restore skb from r1. */ 9377 changes_data = bpf_helper_changes_pkt_data(fn->func); 9378 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) { 9379 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n", 9380 func_id_name(func_id), func_id); 9381 return -EINVAL; 9382 } 9383 9384 memset(&meta, 0, sizeof(meta)); 9385 meta.pkt_access = fn->pkt_access; 9386 9387 err = check_func_proto(fn, func_id); 9388 if (err) { 9389 verbose(env, "kernel subsystem misconfigured func %s#%d\n", 9390 func_id_name(func_id), func_id); 9391 return err; 9392 } 9393 9394 if (env->cur_state->active_rcu_lock) { 9395 if (fn->might_sleep) { 9396 verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n", 9397 func_id_name(func_id), func_id); 9398 return -EINVAL; 9399 } 9400 9401 if (env->prog->aux->sleepable && is_storage_get_function(func_id)) 9402 env->insn_aux_data[insn_idx].storage_get_func_atomic = true; 9403 } 9404 9405 meta.func_id = func_id; 9406 /* check args */ 9407 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) { 9408 err = check_func_arg(env, i, &meta, fn, insn_idx); 9409 if (err) 9410 return err; 9411 } 9412 9413 err = record_func_map(env, &meta, func_id, insn_idx); 9414 if (err) 9415 return err; 9416 9417 err = record_func_key(env, &meta, func_id, insn_idx); 9418 if (err) 9419 return err; 9420 9421 /* Mark slots with STACK_MISC in case of raw mode, stack offset 9422 * is inferred from register state. 9423 */ 9424 for (i = 0; i < meta.access_size; i++) { 9425 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B, 9426 BPF_WRITE, -1, false); 9427 if (err) 9428 return err; 9429 } 9430 9431 regs = cur_regs(env); 9432 9433 if (meta.release_regno) { 9434 err = -EINVAL; 9435 /* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot 9436 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr 9437 * is safe to do directly. 9438 */ 9439 if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) { 9440 if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) { 9441 verbose(env, "verifier internal error: CONST_PTR_TO_DYNPTR cannot be released\n"); 9442 return -EFAULT; 9443 } 9444 err = unmark_stack_slots_dynptr(env, ®s[meta.release_regno]); 9445 } else if (meta.ref_obj_id) { 9446 err = release_reference(env, meta.ref_obj_id); 9447 } else if (register_is_null(®s[meta.release_regno])) { 9448 /* meta.ref_obj_id can only be 0 if register that is meant to be 9449 * released is NULL, which must be > R0. 9450 */ 9451 err = 0; 9452 } 9453 if (err) { 9454 verbose(env, "func %s#%d reference has not been acquired before\n", 9455 func_id_name(func_id), func_id); 9456 return err; 9457 } 9458 } 9459 9460 switch (func_id) { 9461 case BPF_FUNC_tail_call: 9462 err = check_reference_leak(env); 9463 if (err) { 9464 verbose(env, "tail_call would lead to reference leak\n"); 9465 return err; 9466 } 9467 break; 9468 case BPF_FUNC_get_local_storage: 9469 /* check that flags argument in get_local_storage(map, flags) is 0, 9470 * this is required because get_local_storage() can't return an error. 9471 */ 9472 if (!register_is_null(®s[BPF_REG_2])) { 9473 verbose(env, "get_local_storage() doesn't support non-zero flags\n"); 9474 return -EINVAL; 9475 } 9476 break; 9477 case BPF_FUNC_for_each_map_elem: 9478 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno, 9479 set_map_elem_callback_state); 9480 break; 9481 case BPF_FUNC_timer_set_callback: 9482 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno, 9483 set_timer_callback_state); 9484 break; 9485 case BPF_FUNC_find_vma: 9486 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno, 9487 set_find_vma_callback_state); 9488 break; 9489 case BPF_FUNC_snprintf: 9490 err = check_bpf_snprintf_call(env, regs); 9491 break; 9492 case BPF_FUNC_loop: 9493 update_loop_inline_state(env, meta.subprogno); 9494 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno, 9495 set_loop_callback_state); 9496 break; 9497 case BPF_FUNC_dynptr_from_mem: 9498 if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) { 9499 verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n", 9500 reg_type_str(env, regs[BPF_REG_1].type)); 9501 return -EACCES; 9502 } 9503 break; 9504 case BPF_FUNC_set_retval: 9505 if (prog_type == BPF_PROG_TYPE_LSM && 9506 env->prog->expected_attach_type == BPF_LSM_CGROUP) { 9507 if (!env->prog->aux->attach_func_proto->type) { 9508 /* Make sure programs that attach to void 9509 * hooks don't try to modify return value. 9510 */ 9511 verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n"); 9512 return -EINVAL; 9513 } 9514 } 9515 break; 9516 case BPF_FUNC_dynptr_data: 9517 { 9518 struct bpf_reg_state *reg; 9519 int id, ref_obj_id; 9520 9521 reg = get_dynptr_arg_reg(env, fn, regs); 9522 if (!reg) 9523 return -EFAULT; 9524 9525 9526 if (meta.dynptr_id) { 9527 verbose(env, "verifier internal error: meta.dynptr_id already set\n"); 9528 return -EFAULT; 9529 } 9530 if (meta.ref_obj_id) { 9531 verbose(env, "verifier internal error: meta.ref_obj_id already set\n"); 9532 return -EFAULT; 9533 } 9534 9535 id = dynptr_id(env, reg); 9536 if (id < 0) { 9537 verbose(env, "verifier internal error: failed to obtain dynptr id\n"); 9538 return id; 9539 } 9540 9541 ref_obj_id = dynptr_ref_obj_id(env, reg); 9542 if (ref_obj_id < 0) { 9543 verbose(env, "verifier internal error: failed to obtain dynptr ref_obj_id\n"); 9544 return ref_obj_id; 9545 } 9546 9547 meta.dynptr_id = id; 9548 meta.ref_obj_id = ref_obj_id; 9549 9550 break; 9551 } 9552 case BPF_FUNC_dynptr_write: 9553 { 9554 enum bpf_dynptr_type dynptr_type; 9555 struct bpf_reg_state *reg; 9556 9557 reg = get_dynptr_arg_reg(env, fn, regs); 9558 if (!reg) 9559 return -EFAULT; 9560 9561 dynptr_type = dynptr_get_type(env, reg); 9562 if (dynptr_type == BPF_DYNPTR_TYPE_INVALID) 9563 return -EFAULT; 9564 9565 if (dynptr_type == BPF_DYNPTR_TYPE_SKB) 9566 /* this will trigger clear_all_pkt_pointers(), which will 9567 * invalidate all dynptr slices associated with the skb 9568 */ 9569 changes_data = true; 9570 9571 break; 9572 } 9573 case BPF_FUNC_user_ringbuf_drain: 9574 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno, 9575 set_user_ringbuf_callback_state); 9576 break; 9577 } 9578 9579 if (err) 9580 return err; 9581 9582 /* reset caller saved regs */ 9583 for (i = 0; i < CALLER_SAVED_REGS; i++) { 9584 mark_reg_not_init(env, regs, caller_saved[i]); 9585 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK); 9586 } 9587 9588 /* helper call returns 64-bit value. */ 9589 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG; 9590 9591 /* update return register (already marked as written above) */ 9592 ret_type = fn->ret_type; 9593 ret_flag = type_flag(ret_type); 9594 9595 switch (base_type(ret_type)) { 9596 case RET_INTEGER: 9597 /* sets type to SCALAR_VALUE */ 9598 mark_reg_unknown(env, regs, BPF_REG_0); 9599 break; 9600 case RET_VOID: 9601 regs[BPF_REG_0].type = NOT_INIT; 9602 break; 9603 case RET_PTR_TO_MAP_VALUE: 9604 /* There is no offset yet applied, variable or fixed */ 9605 mark_reg_known_zero(env, regs, BPF_REG_0); 9606 /* remember map_ptr, so that check_map_access() 9607 * can check 'value_size' boundary of memory access 9608 * to map element returned from bpf_map_lookup_elem() 9609 */ 9610 if (meta.map_ptr == NULL) { 9611 verbose(env, 9612 "kernel subsystem misconfigured verifier\n"); 9613 return -EINVAL; 9614 } 9615 regs[BPF_REG_0].map_ptr = meta.map_ptr; 9616 regs[BPF_REG_0].map_uid = meta.map_uid; 9617 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag; 9618 if (!type_may_be_null(ret_type) && 9619 btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK)) { 9620 regs[BPF_REG_0].id = ++env->id_gen; 9621 } 9622 break; 9623 case RET_PTR_TO_SOCKET: 9624 mark_reg_known_zero(env, regs, BPF_REG_0); 9625 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag; 9626 break; 9627 case RET_PTR_TO_SOCK_COMMON: 9628 mark_reg_known_zero(env, regs, BPF_REG_0); 9629 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag; 9630 break; 9631 case RET_PTR_TO_TCP_SOCK: 9632 mark_reg_known_zero(env, regs, BPF_REG_0); 9633 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag; 9634 break; 9635 case RET_PTR_TO_MEM: 9636 mark_reg_known_zero(env, regs, BPF_REG_0); 9637 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag; 9638 regs[BPF_REG_0].mem_size = meta.mem_size; 9639 break; 9640 case RET_PTR_TO_MEM_OR_BTF_ID: 9641 { 9642 const struct btf_type *t; 9643 9644 mark_reg_known_zero(env, regs, BPF_REG_0); 9645 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL); 9646 if (!btf_type_is_struct(t)) { 9647 u32 tsize; 9648 const struct btf_type *ret; 9649 const char *tname; 9650 9651 /* resolve the type size of ksym. */ 9652 ret = btf_resolve_size(meta.ret_btf, t, &tsize); 9653 if (IS_ERR(ret)) { 9654 tname = btf_name_by_offset(meta.ret_btf, t->name_off); 9655 verbose(env, "unable to resolve the size of type '%s': %ld\n", 9656 tname, PTR_ERR(ret)); 9657 return -EINVAL; 9658 } 9659 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag; 9660 regs[BPF_REG_0].mem_size = tsize; 9661 } else { 9662 /* MEM_RDONLY may be carried from ret_flag, but it 9663 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise 9664 * it will confuse the check of PTR_TO_BTF_ID in 9665 * check_mem_access(). 9666 */ 9667 ret_flag &= ~MEM_RDONLY; 9668 9669 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag; 9670 regs[BPF_REG_0].btf = meta.ret_btf; 9671 regs[BPF_REG_0].btf_id = meta.ret_btf_id; 9672 } 9673 break; 9674 } 9675 case RET_PTR_TO_BTF_ID: 9676 { 9677 struct btf *ret_btf; 9678 int ret_btf_id; 9679 9680 mark_reg_known_zero(env, regs, BPF_REG_0); 9681 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag; 9682 if (func_id == BPF_FUNC_kptr_xchg) { 9683 ret_btf = meta.kptr_field->kptr.btf; 9684 ret_btf_id = meta.kptr_field->kptr.btf_id; 9685 if (!btf_is_kernel(ret_btf)) 9686 regs[BPF_REG_0].type |= MEM_ALLOC; 9687 } else { 9688 if (fn->ret_btf_id == BPF_PTR_POISON) { 9689 verbose(env, "verifier internal error:"); 9690 verbose(env, "func %s has non-overwritten BPF_PTR_POISON return type\n", 9691 func_id_name(func_id)); 9692 return -EINVAL; 9693 } 9694 ret_btf = btf_vmlinux; 9695 ret_btf_id = *fn->ret_btf_id; 9696 } 9697 if (ret_btf_id == 0) { 9698 verbose(env, "invalid return type %u of func %s#%d\n", 9699 base_type(ret_type), func_id_name(func_id), 9700 func_id); 9701 return -EINVAL; 9702 } 9703 regs[BPF_REG_0].btf = ret_btf; 9704 regs[BPF_REG_0].btf_id = ret_btf_id; 9705 break; 9706 } 9707 default: 9708 verbose(env, "unknown return type %u of func %s#%d\n", 9709 base_type(ret_type), func_id_name(func_id), func_id); 9710 return -EINVAL; 9711 } 9712 9713 if (type_may_be_null(regs[BPF_REG_0].type)) 9714 regs[BPF_REG_0].id = ++env->id_gen; 9715 9716 if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) { 9717 verbose(env, "verifier internal error: func %s#%d sets ref_obj_id more than once\n", 9718 func_id_name(func_id), func_id); 9719 return -EFAULT; 9720 } 9721 9722 if (is_dynptr_ref_function(func_id)) 9723 regs[BPF_REG_0].dynptr_id = meta.dynptr_id; 9724 9725 if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) { 9726 /* For release_reference() */ 9727 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id; 9728 } else if (is_acquire_function(func_id, meta.map_ptr)) { 9729 int id = acquire_reference_state(env, insn_idx); 9730 9731 if (id < 0) 9732 return id; 9733 /* For mark_ptr_or_null_reg() */ 9734 regs[BPF_REG_0].id = id; 9735 /* For release_reference() */ 9736 regs[BPF_REG_0].ref_obj_id = id; 9737 } 9738 9739 do_refine_retval_range(regs, fn->ret_type, func_id, &meta); 9740 9741 err = check_map_func_compatibility(env, meta.map_ptr, func_id); 9742 if (err) 9743 return err; 9744 9745 if ((func_id == BPF_FUNC_get_stack || 9746 func_id == BPF_FUNC_get_task_stack) && 9747 !env->prog->has_callchain_buf) { 9748 const char *err_str; 9749 9750 #ifdef CONFIG_PERF_EVENTS 9751 err = get_callchain_buffers(sysctl_perf_event_max_stack); 9752 err_str = "cannot get callchain buffer for func %s#%d\n"; 9753 #else 9754 err = -ENOTSUPP; 9755 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n"; 9756 #endif 9757 if (err) { 9758 verbose(env, err_str, func_id_name(func_id), func_id); 9759 return err; 9760 } 9761 9762 env->prog->has_callchain_buf = true; 9763 } 9764 9765 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack) 9766 env->prog->call_get_stack = true; 9767 9768 if (func_id == BPF_FUNC_get_func_ip) { 9769 if (check_get_func_ip(env)) 9770 return -ENOTSUPP; 9771 env->prog->call_get_func_ip = true; 9772 } 9773 9774 if (changes_data) 9775 clear_all_pkt_pointers(env); 9776 return 0; 9777 } 9778 9779 /* mark_btf_func_reg_size() is used when the reg size is determined by 9780 * the BTF func_proto's return value size and argument. 9781 */ 9782 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno, 9783 size_t reg_size) 9784 { 9785 struct bpf_reg_state *reg = &cur_regs(env)[regno]; 9786 9787 if (regno == BPF_REG_0) { 9788 /* Function return value */ 9789 reg->live |= REG_LIVE_WRITTEN; 9790 reg->subreg_def = reg_size == sizeof(u64) ? 9791 DEF_NOT_SUBREG : env->insn_idx + 1; 9792 } else { 9793 /* Function argument */ 9794 if (reg_size == sizeof(u64)) { 9795 mark_insn_zext(env, reg); 9796 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64); 9797 } else { 9798 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32); 9799 } 9800 } 9801 } 9802 9803 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta) 9804 { 9805 return meta->kfunc_flags & KF_ACQUIRE; 9806 } 9807 9808 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta) 9809 { 9810 return meta->kfunc_flags & KF_RELEASE; 9811 } 9812 9813 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta) 9814 { 9815 return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta); 9816 } 9817 9818 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta) 9819 { 9820 return meta->kfunc_flags & KF_SLEEPABLE; 9821 } 9822 9823 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta) 9824 { 9825 return meta->kfunc_flags & KF_DESTRUCTIVE; 9826 } 9827 9828 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta) 9829 { 9830 return meta->kfunc_flags & KF_RCU; 9831 } 9832 9833 static bool __kfunc_param_match_suffix(const struct btf *btf, 9834 const struct btf_param *arg, 9835 const char *suffix) 9836 { 9837 int suffix_len = strlen(suffix), len; 9838 const char *param_name; 9839 9840 /* In the future, this can be ported to use BTF tagging */ 9841 param_name = btf_name_by_offset(btf, arg->name_off); 9842 if (str_is_empty(param_name)) 9843 return false; 9844 len = strlen(param_name); 9845 if (len < suffix_len) 9846 return false; 9847 param_name += len - suffix_len; 9848 return !strncmp(param_name, suffix, suffix_len); 9849 } 9850 9851 static bool is_kfunc_arg_mem_size(const struct btf *btf, 9852 const struct btf_param *arg, 9853 const struct bpf_reg_state *reg) 9854 { 9855 const struct btf_type *t; 9856 9857 t = btf_type_skip_modifiers(btf, arg->type, NULL); 9858 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE) 9859 return false; 9860 9861 return __kfunc_param_match_suffix(btf, arg, "__sz"); 9862 } 9863 9864 static bool is_kfunc_arg_const_mem_size(const struct btf *btf, 9865 const struct btf_param *arg, 9866 const struct bpf_reg_state *reg) 9867 { 9868 const struct btf_type *t; 9869 9870 t = btf_type_skip_modifiers(btf, arg->type, NULL); 9871 if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE) 9872 return false; 9873 9874 return __kfunc_param_match_suffix(btf, arg, "__szk"); 9875 } 9876 9877 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg) 9878 { 9879 return __kfunc_param_match_suffix(btf, arg, "__opt"); 9880 } 9881 9882 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg) 9883 { 9884 return __kfunc_param_match_suffix(btf, arg, "__k"); 9885 } 9886 9887 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg) 9888 { 9889 return __kfunc_param_match_suffix(btf, arg, "__ign"); 9890 } 9891 9892 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg) 9893 { 9894 return __kfunc_param_match_suffix(btf, arg, "__alloc"); 9895 } 9896 9897 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg) 9898 { 9899 return __kfunc_param_match_suffix(btf, arg, "__uninit"); 9900 } 9901 9902 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg) 9903 { 9904 return __kfunc_param_match_suffix(btf, arg, "__refcounted_kptr"); 9905 } 9906 9907 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf, 9908 const struct btf_param *arg, 9909 const char *name) 9910 { 9911 int len, target_len = strlen(name); 9912 const char *param_name; 9913 9914 param_name = btf_name_by_offset(btf, arg->name_off); 9915 if (str_is_empty(param_name)) 9916 return false; 9917 len = strlen(param_name); 9918 if (len != target_len) 9919 return false; 9920 if (strcmp(param_name, name)) 9921 return false; 9922 9923 return true; 9924 } 9925 9926 enum { 9927 KF_ARG_DYNPTR_ID, 9928 KF_ARG_LIST_HEAD_ID, 9929 KF_ARG_LIST_NODE_ID, 9930 KF_ARG_RB_ROOT_ID, 9931 KF_ARG_RB_NODE_ID, 9932 }; 9933 9934 BTF_ID_LIST(kf_arg_btf_ids) 9935 BTF_ID(struct, bpf_dynptr_kern) 9936 BTF_ID(struct, bpf_list_head) 9937 BTF_ID(struct, bpf_list_node) 9938 BTF_ID(struct, bpf_rb_root) 9939 BTF_ID(struct, bpf_rb_node) 9940 9941 static bool __is_kfunc_ptr_arg_type(const struct btf *btf, 9942 const struct btf_param *arg, int type) 9943 { 9944 const struct btf_type *t; 9945 u32 res_id; 9946 9947 t = btf_type_skip_modifiers(btf, arg->type, NULL); 9948 if (!t) 9949 return false; 9950 if (!btf_type_is_ptr(t)) 9951 return false; 9952 t = btf_type_skip_modifiers(btf, t->type, &res_id); 9953 if (!t) 9954 return false; 9955 return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]); 9956 } 9957 9958 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg) 9959 { 9960 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID); 9961 } 9962 9963 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg) 9964 { 9965 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID); 9966 } 9967 9968 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg) 9969 { 9970 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID); 9971 } 9972 9973 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg) 9974 { 9975 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID); 9976 } 9977 9978 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg) 9979 { 9980 return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID); 9981 } 9982 9983 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf, 9984 const struct btf_param *arg) 9985 { 9986 const struct btf_type *t; 9987 9988 t = btf_type_resolve_func_ptr(btf, arg->type, NULL); 9989 if (!t) 9990 return false; 9991 9992 return true; 9993 } 9994 9995 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */ 9996 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env, 9997 const struct btf *btf, 9998 const struct btf_type *t, int rec) 9999 { 10000 const struct btf_type *member_type; 10001 const struct btf_member *member; 10002 u32 i; 10003 10004 if (!btf_type_is_struct(t)) 10005 return false; 10006 10007 for_each_member(i, t, member) { 10008 const struct btf_array *array; 10009 10010 member_type = btf_type_skip_modifiers(btf, member->type, NULL); 10011 if (btf_type_is_struct(member_type)) { 10012 if (rec >= 3) { 10013 verbose(env, "max struct nesting depth exceeded\n"); 10014 return false; 10015 } 10016 if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1)) 10017 return false; 10018 continue; 10019 } 10020 if (btf_type_is_array(member_type)) { 10021 array = btf_array(member_type); 10022 if (!array->nelems) 10023 return false; 10024 member_type = btf_type_skip_modifiers(btf, array->type, NULL); 10025 if (!btf_type_is_scalar(member_type)) 10026 return false; 10027 continue; 10028 } 10029 if (!btf_type_is_scalar(member_type)) 10030 return false; 10031 } 10032 return true; 10033 } 10034 10035 10036 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = { 10037 #ifdef CONFIG_NET 10038 [PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK], 10039 [PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON], 10040 [PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP], 10041 #endif 10042 }; 10043 10044 enum kfunc_ptr_arg_type { 10045 KF_ARG_PTR_TO_CTX, 10046 KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */ 10047 KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */ 10048 KF_ARG_PTR_TO_DYNPTR, 10049 KF_ARG_PTR_TO_ITER, 10050 KF_ARG_PTR_TO_LIST_HEAD, 10051 KF_ARG_PTR_TO_LIST_NODE, 10052 KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */ 10053 KF_ARG_PTR_TO_MEM, 10054 KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */ 10055 KF_ARG_PTR_TO_CALLBACK, 10056 KF_ARG_PTR_TO_RB_ROOT, 10057 KF_ARG_PTR_TO_RB_NODE, 10058 }; 10059 10060 enum special_kfunc_type { 10061 KF_bpf_obj_new_impl, 10062 KF_bpf_obj_drop_impl, 10063 KF_bpf_refcount_acquire_impl, 10064 KF_bpf_list_push_front_impl, 10065 KF_bpf_list_push_back_impl, 10066 KF_bpf_list_pop_front, 10067 KF_bpf_list_pop_back, 10068 KF_bpf_cast_to_kern_ctx, 10069 KF_bpf_rdonly_cast, 10070 KF_bpf_rcu_read_lock, 10071 KF_bpf_rcu_read_unlock, 10072 KF_bpf_rbtree_remove, 10073 KF_bpf_rbtree_add_impl, 10074 KF_bpf_rbtree_first, 10075 KF_bpf_dynptr_from_skb, 10076 KF_bpf_dynptr_from_xdp, 10077 KF_bpf_dynptr_slice, 10078 KF_bpf_dynptr_slice_rdwr, 10079 KF_bpf_dynptr_clone, 10080 }; 10081 10082 BTF_SET_START(special_kfunc_set) 10083 BTF_ID(func, bpf_obj_new_impl) 10084 BTF_ID(func, bpf_obj_drop_impl) 10085 BTF_ID(func, bpf_refcount_acquire_impl) 10086 BTF_ID(func, bpf_list_push_front_impl) 10087 BTF_ID(func, bpf_list_push_back_impl) 10088 BTF_ID(func, bpf_list_pop_front) 10089 BTF_ID(func, bpf_list_pop_back) 10090 BTF_ID(func, bpf_cast_to_kern_ctx) 10091 BTF_ID(func, bpf_rdonly_cast) 10092 BTF_ID(func, bpf_rbtree_remove) 10093 BTF_ID(func, bpf_rbtree_add_impl) 10094 BTF_ID(func, bpf_rbtree_first) 10095 BTF_ID(func, bpf_dynptr_from_skb) 10096 BTF_ID(func, bpf_dynptr_from_xdp) 10097 BTF_ID(func, bpf_dynptr_slice) 10098 BTF_ID(func, bpf_dynptr_slice_rdwr) 10099 BTF_ID(func, bpf_dynptr_clone) 10100 BTF_SET_END(special_kfunc_set) 10101 10102 BTF_ID_LIST(special_kfunc_list) 10103 BTF_ID(func, bpf_obj_new_impl) 10104 BTF_ID(func, bpf_obj_drop_impl) 10105 BTF_ID(func, bpf_refcount_acquire_impl) 10106 BTF_ID(func, bpf_list_push_front_impl) 10107 BTF_ID(func, bpf_list_push_back_impl) 10108 BTF_ID(func, bpf_list_pop_front) 10109 BTF_ID(func, bpf_list_pop_back) 10110 BTF_ID(func, bpf_cast_to_kern_ctx) 10111 BTF_ID(func, bpf_rdonly_cast) 10112 BTF_ID(func, bpf_rcu_read_lock) 10113 BTF_ID(func, bpf_rcu_read_unlock) 10114 BTF_ID(func, bpf_rbtree_remove) 10115 BTF_ID(func, bpf_rbtree_add_impl) 10116 BTF_ID(func, bpf_rbtree_first) 10117 BTF_ID(func, bpf_dynptr_from_skb) 10118 BTF_ID(func, bpf_dynptr_from_xdp) 10119 BTF_ID(func, bpf_dynptr_slice) 10120 BTF_ID(func, bpf_dynptr_slice_rdwr) 10121 BTF_ID(func, bpf_dynptr_clone) 10122 10123 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta) 10124 { 10125 if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] && 10126 meta->arg_owning_ref) { 10127 return false; 10128 } 10129 10130 return meta->kfunc_flags & KF_RET_NULL; 10131 } 10132 10133 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta) 10134 { 10135 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock]; 10136 } 10137 10138 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta) 10139 { 10140 return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock]; 10141 } 10142 10143 static enum kfunc_ptr_arg_type 10144 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env, 10145 struct bpf_kfunc_call_arg_meta *meta, 10146 const struct btf_type *t, const struct btf_type *ref_t, 10147 const char *ref_tname, const struct btf_param *args, 10148 int argno, int nargs) 10149 { 10150 u32 regno = argno + 1; 10151 struct bpf_reg_state *regs = cur_regs(env); 10152 struct bpf_reg_state *reg = ®s[regno]; 10153 bool arg_mem_size = false; 10154 10155 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) 10156 return KF_ARG_PTR_TO_CTX; 10157 10158 /* In this function, we verify the kfunc's BTF as per the argument type, 10159 * leaving the rest of the verification with respect to the register 10160 * type to our caller. When a set of conditions hold in the BTF type of 10161 * arguments, we resolve it to a known kfunc_ptr_arg_type. 10162 */ 10163 if (btf_get_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno)) 10164 return KF_ARG_PTR_TO_CTX; 10165 10166 if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno])) 10167 return KF_ARG_PTR_TO_ALLOC_BTF_ID; 10168 10169 if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno])) 10170 return KF_ARG_PTR_TO_REFCOUNTED_KPTR; 10171 10172 if (is_kfunc_arg_dynptr(meta->btf, &args[argno])) 10173 return KF_ARG_PTR_TO_DYNPTR; 10174 10175 if (is_kfunc_arg_iter(meta, argno)) 10176 return KF_ARG_PTR_TO_ITER; 10177 10178 if (is_kfunc_arg_list_head(meta->btf, &args[argno])) 10179 return KF_ARG_PTR_TO_LIST_HEAD; 10180 10181 if (is_kfunc_arg_list_node(meta->btf, &args[argno])) 10182 return KF_ARG_PTR_TO_LIST_NODE; 10183 10184 if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno])) 10185 return KF_ARG_PTR_TO_RB_ROOT; 10186 10187 if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno])) 10188 return KF_ARG_PTR_TO_RB_NODE; 10189 10190 if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) { 10191 if (!btf_type_is_struct(ref_t)) { 10192 verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n", 10193 meta->func_name, argno, btf_type_str(ref_t), ref_tname); 10194 return -EINVAL; 10195 } 10196 return KF_ARG_PTR_TO_BTF_ID; 10197 } 10198 10199 if (is_kfunc_arg_callback(env, meta->btf, &args[argno])) 10200 return KF_ARG_PTR_TO_CALLBACK; 10201 10202 10203 if (argno + 1 < nargs && 10204 (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]) || 10205 is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], ®s[regno + 1]))) 10206 arg_mem_size = true; 10207 10208 /* This is the catch all argument type of register types supported by 10209 * check_helper_mem_access. However, we only allow when argument type is 10210 * pointer to scalar, or struct composed (recursively) of scalars. When 10211 * arg_mem_size is true, the pointer can be void *. 10212 */ 10213 if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) && 10214 (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) { 10215 verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n", 10216 argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : ""); 10217 return -EINVAL; 10218 } 10219 return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM; 10220 } 10221 10222 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env, 10223 struct bpf_reg_state *reg, 10224 const struct btf_type *ref_t, 10225 const char *ref_tname, u32 ref_id, 10226 struct bpf_kfunc_call_arg_meta *meta, 10227 int argno) 10228 { 10229 const struct btf_type *reg_ref_t; 10230 bool strict_type_match = false; 10231 const struct btf *reg_btf; 10232 const char *reg_ref_tname; 10233 u32 reg_ref_id; 10234 10235 if (base_type(reg->type) == PTR_TO_BTF_ID) { 10236 reg_btf = reg->btf; 10237 reg_ref_id = reg->btf_id; 10238 } else { 10239 reg_btf = btf_vmlinux; 10240 reg_ref_id = *reg2btf_ids[base_type(reg->type)]; 10241 } 10242 10243 /* Enforce strict type matching for calls to kfuncs that are acquiring 10244 * or releasing a reference, or are no-cast aliases. We do _not_ 10245 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default, 10246 * as we want to enable BPF programs to pass types that are bitwise 10247 * equivalent without forcing them to explicitly cast with something 10248 * like bpf_cast_to_kern_ctx(). 10249 * 10250 * For example, say we had a type like the following: 10251 * 10252 * struct bpf_cpumask { 10253 * cpumask_t cpumask; 10254 * refcount_t usage; 10255 * }; 10256 * 10257 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed 10258 * to a struct cpumask, so it would be safe to pass a struct 10259 * bpf_cpumask * to a kfunc expecting a struct cpumask *. 10260 * 10261 * The philosophy here is similar to how we allow scalars of different 10262 * types to be passed to kfuncs as long as the size is the same. The 10263 * only difference here is that we're simply allowing 10264 * btf_struct_ids_match() to walk the struct at the 0th offset, and 10265 * resolve types. 10266 */ 10267 if (is_kfunc_acquire(meta) || 10268 (is_kfunc_release(meta) && reg->ref_obj_id) || 10269 btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id)) 10270 strict_type_match = true; 10271 10272 WARN_ON_ONCE(is_kfunc_trusted_args(meta) && reg->off); 10273 10274 reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, ®_ref_id); 10275 reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off); 10276 if (!btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match)) { 10277 verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n", 10278 meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1, 10279 btf_type_str(reg_ref_t), reg_ref_tname); 10280 return -EINVAL; 10281 } 10282 return 0; 10283 } 10284 10285 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 10286 { 10287 struct bpf_verifier_state *state = env->cur_state; 10288 10289 if (!state->active_lock.ptr) { 10290 verbose(env, "verifier internal error: ref_set_non_owning w/o active lock\n"); 10291 return -EFAULT; 10292 } 10293 10294 if (type_flag(reg->type) & NON_OWN_REF) { 10295 verbose(env, "verifier internal error: NON_OWN_REF already set\n"); 10296 return -EFAULT; 10297 } 10298 10299 reg->type |= NON_OWN_REF; 10300 return 0; 10301 } 10302 10303 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id) 10304 { 10305 struct bpf_func_state *state, *unused; 10306 struct bpf_reg_state *reg; 10307 int i; 10308 10309 state = cur_func(env); 10310 10311 if (!ref_obj_id) { 10312 verbose(env, "verifier internal error: ref_obj_id is zero for " 10313 "owning -> non-owning conversion\n"); 10314 return -EFAULT; 10315 } 10316 10317 for (i = 0; i < state->acquired_refs; i++) { 10318 if (state->refs[i].id != ref_obj_id) 10319 continue; 10320 10321 /* Clear ref_obj_id here so release_reference doesn't clobber 10322 * the whole reg 10323 */ 10324 bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({ 10325 if (reg->ref_obj_id == ref_obj_id) { 10326 reg->ref_obj_id = 0; 10327 ref_set_non_owning(env, reg); 10328 } 10329 })); 10330 return 0; 10331 } 10332 10333 verbose(env, "verifier internal error: ref state missing for ref_obj_id\n"); 10334 return -EFAULT; 10335 } 10336 10337 /* Implementation details: 10338 * 10339 * Each register points to some region of memory, which we define as an 10340 * allocation. Each allocation may embed a bpf_spin_lock which protects any 10341 * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same 10342 * allocation. The lock and the data it protects are colocated in the same 10343 * memory region. 10344 * 10345 * Hence, everytime a register holds a pointer value pointing to such 10346 * allocation, the verifier preserves a unique reg->id for it. 10347 * 10348 * The verifier remembers the lock 'ptr' and the lock 'id' whenever 10349 * bpf_spin_lock is called. 10350 * 10351 * To enable this, lock state in the verifier captures two values: 10352 * active_lock.ptr = Register's type specific pointer 10353 * active_lock.id = A unique ID for each register pointer value 10354 * 10355 * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two 10356 * supported register types. 10357 * 10358 * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of 10359 * allocated objects is the reg->btf pointer. 10360 * 10361 * The active_lock.id is non-unique for maps supporting direct_value_addr, as we 10362 * can establish the provenance of the map value statically for each distinct 10363 * lookup into such maps. They always contain a single map value hence unique 10364 * IDs for each pseudo load pessimizes the algorithm and rejects valid programs. 10365 * 10366 * So, in case of global variables, they use array maps with max_entries = 1, 10367 * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point 10368 * into the same map value as max_entries is 1, as described above). 10369 * 10370 * In case of inner map lookups, the inner map pointer has same map_ptr as the 10371 * outer map pointer (in verifier context), but each lookup into an inner map 10372 * assigns a fresh reg->id to the lookup, so while lookups into distinct inner 10373 * maps from the same outer map share the same map_ptr as active_lock.ptr, they 10374 * will get different reg->id assigned to each lookup, hence different 10375 * active_lock.id. 10376 * 10377 * In case of allocated objects, active_lock.ptr is the reg->btf, and the 10378 * reg->id is a unique ID preserved after the NULL pointer check on the pointer 10379 * returned from bpf_obj_new. Each allocation receives a new reg->id. 10380 */ 10381 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg) 10382 { 10383 void *ptr; 10384 u32 id; 10385 10386 switch ((int)reg->type) { 10387 case PTR_TO_MAP_VALUE: 10388 ptr = reg->map_ptr; 10389 break; 10390 case PTR_TO_BTF_ID | MEM_ALLOC: 10391 ptr = reg->btf; 10392 break; 10393 default: 10394 verbose(env, "verifier internal error: unknown reg type for lock check\n"); 10395 return -EFAULT; 10396 } 10397 id = reg->id; 10398 10399 if (!env->cur_state->active_lock.ptr) 10400 return -EINVAL; 10401 if (env->cur_state->active_lock.ptr != ptr || 10402 env->cur_state->active_lock.id != id) { 10403 verbose(env, "held lock and object are not in the same allocation\n"); 10404 return -EINVAL; 10405 } 10406 return 0; 10407 } 10408 10409 static bool is_bpf_list_api_kfunc(u32 btf_id) 10410 { 10411 return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] || 10412 btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] || 10413 btf_id == special_kfunc_list[KF_bpf_list_pop_front] || 10414 btf_id == special_kfunc_list[KF_bpf_list_pop_back]; 10415 } 10416 10417 static bool is_bpf_rbtree_api_kfunc(u32 btf_id) 10418 { 10419 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] || 10420 btf_id == special_kfunc_list[KF_bpf_rbtree_remove] || 10421 btf_id == special_kfunc_list[KF_bpf_rbtree_first]; 10422 } 10423 10424 static bool is_bpf_graph_api_kfunc(u32 btf_id) 10425 { 10426 return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) || 10427 btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]; 10428 } 10429 10430 static bool is_callback_calling_kfunc(u32 btf_id) 10431 { 10432 return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]; 10433 } 10434 10435 static bool is_rbtree_lock_required_kfunc(u32 btf_id) 10436 { 10437 return is_bpf_rbtree_api_kfunc(btf_id); 10438 } 10439 10440 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env, 10441 enum btf_field_type head_field_type, 10442 u32 kfunc_btf_id) 10443 { 10444 bool ret; 10445 10446 switch (head_field_type) { 10447 case BPF_LIST_HEAD: 10448 ret = is_bpf_list_api_kfunc(kfunc_btf_id); 10449 break; 10450 case BPF_RB_ROOT: 10451 ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id); 10452 break; 10453 default: 10454 verbose(env, "verifier internal error: unexpected graph root argument type %s\n", 10455 btf_field_type_name(head_field_type)); 10456 return false; 10457 } 10458 10459 if (!ret) 10460 verbose(env, "verifier internal error: %s head arg for unknown kfunc\n", 10461 btf_field_type_name(head_field_type)); 10462 return ret; 10463 } 10464 10465 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env, 10466 enum btf_field_type node_field_type, 10467 u32 kfunc_btf_id) 10468 { 10469 bool ret; 10470 10471 switch (node_field_type) { 10472 case BPF_LIST_NODE: 10473 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] || 10474 kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]); 10475 break; 10476 case BPF_RB_NODE: 10477 ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] || 10478 kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl]); 10479 break; 10480 default: 10481 verbose(env, "verifier internal error: unexpected graph node argument type %s\n", 10482 btf_field_type_name(node_field_type)); 10483 return false; 10484 } 10485 10486 if (!ret) 10487 verbose(env, "verifier internal error: %s node arg for unknown kfunc\n", 10488 btf_field_type_name(node_field_type)); 10489 return ret; 10490 } 10491 10492 static int 10493 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env, 10494 struct bpf_reg_state *reg, u32 regno, 10495 struct bpf_kfunc_call_arg_meta *meta, 10496 enum btf_field_type head_field_type, 10497 struct btf_field **head_field) 10498 { 10499 const char *head_type_name; 10500 struct btf_field *field; 10501 struct btf_record *rec; 10502 u32 head_off; 10503 10504 if (meta->btf != btf_vmlinux) { 10505 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n"); 10506 return -EFAULT; 10507 } 10508 10509 if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id)) 10510 return -EFAULT; 10511 10512 head_type_name = btf_field_type_name(head_field_type); 10513 if (!tnum_is_const(reg->var_off)) { 10514 verbose(env, 10515 "R%d doesn't have constant offset. %s has to be at the constant offset\n", 10516 regno, head_type_name); 10517 return -EINVAL; 10518 } 10519 10520 rec = reg_btf_record(reg); 10521 head_off = reg->off + reg->var_off.value; 10522 field = btf_record_find(rec, head_off, head_field_type); 10523 if (!field) { 10524 verbose(env, "%s not found at offset=%u\n", head_type_name, head_off); 10525 return -EINVAL; 10526 } 10527 10528 /* All functions require bpf_list_head to be protected using a bpf_spin_lock */ 10529 if (check_reg_allocation_locked(env, reg)) { 10530 verbose(env, "bpf_spin_lock at off=%d must be held for %s\n", 10531 rec->spin_lock_off, head_type_name); 10532 return -EINVAL; 10533 } 10534 10535 if (*head_field) { 10536 verbose(env, "verifier internal error: repeating %s arg\n", head_type_name); 10537 return -EFAULT; 10538 } 10539 *head_field = field; 10540 return 0; 10541 } 10542 10543 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env, 10544 struct bpf_reg_state *reg, u32 regno, 10545 struct bpf_kfunc_call_arg_meta *meta) 10546 { 10547 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD, 10548 &meta->arg_list_head.field); 10549 } 10550 10551 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env, 10552 struct bpf_reg_state *reg, u32 regno, 10553 struct bpf_kfunc_call_arg_meta *meta) 10554 { 10555 return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT, 10556 &meta->arg_rbtree_root.field); 10557 } 10558 10559 static int 10560 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env, 10561 struct bpf_reg_state *reg, u32 regno, 10562 struct bpf_kfunc_call_arg_meta *meta, 10563 enum btf_field_type head_field_type, 10564 enum btf_field_type node_field_type, 10565 struct btf_field **node_field) 10566 { 10567 const char *node_type_name; 10568 const struct btf_type *et, *t; 10569 struct btf_field *field; 10570 u32 node_off; 10571 10572 if (meta->btf != btf_vmlinux) { 10573 verbose(env, "verifier internal error: unexpected btf mismatch in kfunc call\n"); 10574 return -EFAULT; 10575 } 10576 10577 if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id)) 10578 return -EFAULT; 10579 10580 node_type_name = btf_field_type_name(node_field_type); 10581 if (!tnum_is_const(reg->var_off)) { 10582 verbose(env, 10583 "R%d doesn't have constant offset. %s has to be at the constant offset\n", 10584 regno, node_type_name); 10585 return -EINVAL; 10586 } 10587 10588 node_off = reg->off + reg->var_off.value; 10589 field = reg_find_field_offset(reg, node_off, node_field_type); 10590 if (!field || field->offset != node_off) { 10591 verbose(env, "%s not found at offset=%u\n", node_type_name, node_off); 10592 return -EINVAL; 10593 } 10594 10595 field = *node_field; 10596 10597 et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id); 10598 t = btf_type_by_id(reg->btf, reg->btf_id); 10599 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf, 10600 field->graph_root.value_btf_id, true)) { 10601 verbose(env, "operation on %s expects arg#1 %s at offset=%d " 10602 "in struct %s, but arg is at offset=%d in struct %s\n", 10603 btf_field_type_name(head_field_type), 10604 btf_field_type_name(node_field_type), 10605 field->graph_root.node_offset, 10606 btf_name_by_offset(field->graph_root.btf, et->name_off), 10607 node_off, btf_name_by_offset(reg->btf, t->name_off)); 10608 return -EINVAL; 10609 } 10610 meta->arg_btf = reg->btf; 10611 meta->arg_btf_id = reg->btf_id; 10612 10613 if (node_off != field->graph_root.node_offset) { 10614 verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n", 10615 node_off, btf_field_type_name(node_field_type), 10616 field->graph_root.node_offset, 10617 btf_name_by_offset(field->graph_root.btf, et->name_off)); 10618 return -EINVAL; 10619 } 10620 10621 return 0; 10622 } 10623 10624 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env, 10625 struct bpf_reg_state *reg, u32 regno, 10626 struct bpf_kfunc_call_arg_meta *meta) 10627 { 10628 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta, 10629 BPF_LIST_HEAD, BPF_LIST_NODE, 10630 &meta->arg_list_head.field); 10631 } 10632 10633 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env, 10634 struct bpf_reg_state *reg, u32 regno, 10635 struct bpf_kfunc_call_arg_meta *meta) 10636 { 10637 return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta, 10638 BPF_RB_ROOT, BPF_RB_NODE, 10639 &meta->arg_rbtree_root.field); 10640 } 10641 10642 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta, 10643 int insn_idx) 10644 { 10645 const char *func_name = meta->func_name, *ref_tname; 10646 const struct btf *btf = meta->btf; 10647 const struct btf_param *args; 10648 struct btf_record *rec; 10649 u32 i, nargs; 10650 int ret; 10651 10652 args = (const struct btf_param *)(meta->func_proto + 1); 10653 nargs = btf_type_vlen(meta->func_proto); 10654 if (nargs > MAX_BPF_FUNC_REG_ARGS) { 10655 verbose(env, "Function %s has %d > %d args\n", func_name, nargs, 10656 MAX_BPF_FUNC_REG_ARGS); 10657 return -EINVAL; 10658 } 10659 10660 /* Check that BTF function arguments match actual types that the 10661 * verifier sees. 10662 */ 10663 for (i = 0; i < nargs; i++) { 10664 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[i + 1]; 10665 const struct btf_type *t, *ref_t, *resolve_ret; 10666 enum bpf_arg_type arg_type = ARG_DONTCARE; 10667 u32 regno = i + 1, ref_id, type_size; 10668 bool is_ret_buf_sz = false; 10669 int kf_arg_type; 10670 10671 t = btf_type_skip_modifiers(btf, args[i].type, NULL); 10672 10673 if (is_kfunc_arg_ignore(btf, &args[i])) 10674 continue; 10675 10676 if (btf_type_is_scalar(t)) { 10677 if (reg->type != SCALAR_VALUE) { 10678 verbose(env, "R%d is not a scalar\n", regno); 10679 return -EINVAL; 10680 } 10681 10682 if (is_kfunc_arg_constant(meta->btf, &args[i])) { 10683 if (meta->arg_constant.found) { 10684 verbose(env, "verifier internal error: only one constant argument permitted\n"); 10685 return -EFAULT; 10686 } 10687 if (!tnum_is_const(reg->var_off)) { 10688 verbose(env, "R%d must be a known constant\n", regno); 10689 return -EINVAL; 10690 } 10691 ret = mark_chain_precision(env, regno); 10692 if (ret < 0) 10693 return ret; 10694 meta->arg_constant.found = true; 10695 meta->arg_constant.value = reg->var_off.value; 10696 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) { 10697 meta->r0_rdonly = true; 10698 is_ret_buf_sz = true; 10699 } else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) { 10700 is_ret_buf_sz = true; 10701 } 10702 10703 if (is_ret_buf_sz) { 10704 if (meta->r0_size) { 10705 verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc"); 10706 return -EINVAL; 10707 } 10708 10709 if (!tnum_is_const(reg->var_off)) { 10710 verbose(env, "R%d is not a const\n", regno); 10711 return -EINVAL; 10712 } 10713 10714 meta->r0_size = reg->var_off.value; 10715 ret = mark_chain_precision(env, regno); 10716 if (ret) 10717 return ret; 10718 } 10719 continue; 10720 } 10721 10722 if (!btf_type_is_ptr(t)) { 10723 verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t)); 10724 return -EINVAL; 10725 } 10726 10727 if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) && 10728 (register_is_null(reg) || type_may_be_null(reg->type))) { 10729 verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i); 10730 return -EACCES; 10731 } 10732 10733 if (reg->ref_obj_id) { 10734 if (is_kfunc_release(meta) && meta->ref_obj_id) { 10735 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n", 10736 regno, reg->ref_obj_id, 10737 meta->ref_obj_id); 10738 return -EFAULT; 10739 } 10740 meta->ref_obj_id = reg->ref_obj_id; 10741 if (is_kfunc_release(meta)) 10742 meta->release_regno = regno; 10743 } 10744 10745 ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id); 10746 ref_tname = btf_name_by_offset(btf, ref_t->name_off); 10747 10748 kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs); 10749 if (kf_arg_type < 0) 10750 return kf_arg_type; 10751 10752 switch (kf_arg_type) { 10753 case KF_ARG_PTR_TO_ALLOC_BTF_ID: 10754 case KF_ARG_PTR_TO_BTF_ID: 10755 if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta)) 10756 break; 10757 10758 if (!is_trusted_reg(reg)) { 10759 if (!is_kfunc_rcu(meta)) { 10760 verbose(env, "R%d must be referenced or trusted\n", regno); 10761 return -EINVAL; 10762 } 10763 if (!is_rcu_reg(reg)) { 10764 verbose(env, "R%d must be a rcu pointer\n", regno); 10765 return -EINVAL; 10766 } 10767 } 10768 10769 fallthrough; 10770 case KF_ARG_PTR_TO_CTX: 10771 /* Trusted arguments have the same offset checks as release arguments */ 10772 arg_type |= OBJ_RELEASE; 10773 break; 10774 case KF_ARG_PTR_TO_DYNPTR: 10775 case KF_ARG_PTR_TO_ITER: 10776 case KF_ARG_PTR_TO_LIST_HEAD: 10777 case KF_ARG_PTR_TO_LIST_NODE: 10778 case KF_ARG_PTR_TO_RB_ROOT: 10779 case KF_ARG_PTR_TO_RB_NODE: 10780 case KF_ARG_PTR_TO_MEM: 10781 case KF_ARG_PTR_TO_MEM_SIZE: 10782 case KF_ARG_PTR_TO_CALLBACK: 10783 case KF_ARG_PTR_TO_REFCOUNTED_KPTR: 10784 /* Trusted by default */ 10785 break; 10786 default: 10787 WARN_ON_ONCE(1); 10788 return -EFAULT; 10789 } 10790 10791 if (is_kfunc_release(meta) && reg->ref_obj_id) 10792 arg_type |= OBJ_RELEASE; 10793 ret = check_func_arg_reg_off(env, reg, regno, arg_type); 10794 if (ret < 0) 10795 return ret; 10796 10797 switch (kf_arg_type) { 10798 case KF_ARG_PTR_TO_CTX: 10799 if (reg->type != PTR_TO_CTX) { 10800 verbose(env, "arg#%d expected pointer to ctx, but got %s\n", i, btf_type_str(t)); 10801 return -EINVAL; 10802 } 10803 10804 if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) { 10805 ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog)); 10806 if (ret < 0) 10807 return -EINVAL; 10808 meta->ret_btf_id = ret; 10809 } 10810 break; 10811 case KF_ARG_PTR_TO_ALLOC_BTF_ID: 10812 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) { 10813 verbose(env, "arg#%d expected pointer to allocated object\n", i); 10814 return -EINVAL; 10815 } 10816 if (!reg->ref_obj_id) { 10817 verbose(env, "allocated object must be referenced\n"); 10818 return -EINVAL; 10819 } 10820 if (meta->btf == btf_vmlinux && 10821 meta->func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) { 10822 meta->arg_btf = reg->btf; 10823 meta->arg_btf_id = reg->btf_id; 10824 } 10825 break; 10826 case KF_ARG_PTR_TO_DYNPTR: 10827 { 10828 enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR; 10829 int clone_ref_obj_id = 0; 10830 10831 if (reg->type != PTR_TO_STACK && 10832 reg->type != CONST_PTR_TO_DYNPTR) { 10833 verbose(env, "arg#%d expected pointer to stack or dynptr_ptr\n", i); 10834 return -EINVAL; 10835 } 10836 10837 if (reg->type == CONST_PTR_TO_DYNPTR) 10838 dynptr_arg_type |= MEM_RDONLY; 10839 10840 if (is_kfunc_arg_uninit(btf, &args[i])) 10841 dynptr_arg_type |= MEM_UNINIT; 10842 10843 if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) { 10844 dynptr_arg_type |= DYNPTR_TYPE_SKB; 10845 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) { 10846 dynptr_arg_type |= DYNPTR_TYPE_XDP; 10847 } else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] && 10848 (dynptr_arg_type & MEM_UNINIT)) { 10849 enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type; 10850 10851 if (parent_type == BPF_DYNPTR_TYPE_INVALID) { 10852 verbose(env, "verifier internal error: no dynptr type for parent of clone\n"); 10853 return -EFAULT; 10854 } 10855 10856 dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type); 10857 clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id; 10858 if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) { 10859 verbose(env, "verifier internal error: missing ref obj id for parent of clone\n"); 10860 return -EFAULT; 10861 } 10862 } 10863 10864 ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id); 10865 if (ret < 0) 10866 return ret; 10867 10868 if (!(dynptr_arg_type & MEM_UNINIT)) { 10869 int id = dynptr_id(env, reg); 10870 10871 if (id < 0) { 10872 verbose(env, "verifier internal error: failed to obtain dynptr id\n"); 10873 return id; 10874 } 10875 meta->initialized_dynptr.id = id; 10876 meta->initialized_dynptr.type = dynptr_get_type(env, reg); 10877 meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg); 10878 } 10879 10880 break; 10881 } 10882 case KF_ARG_PTR_TO_ITER: 10883 ret = process_iter_arg(env, regno, insn_idx, meta); 10884 if (ret < 0) 10885 return ret; 10886 break; 10887 case KF_ARG_PTR_TO_LIST_HEAD: 10888 if (reg->type != PTR_TO_MAP_VALUE && 10889 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) { 10890 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i); 10891 return -EINVAL; 10892 } 10893 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) { 10894 verbose(env, "allocated object must be referenced\n"); 10895 return -EINVAL; 10896 } 10897 ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta); 10898 if (ret < 0) 10899 return ret; 10900 break; 10901 case KF_ARG_PTR_TO_RB_ROOT: 10902 if (reg->type != PTR_TO_MAP_VALUE && 10903 reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) { 10904 verbose(env, "arg#%d expected pointer to map value or allocated object\n", i); 10905 return -EINVAL; 10906 } 10907 if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) { 10908 verbose(env, "allocated object must be referenced\n"); 10909 return -EINVAL; 10910 } 10911 ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta); 10912 if (ret < 0) 10913 return ret; 10914 break; 10915 case KF_ARG_PTR_TO_LIST_NODE: 10916 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) { 10917 verbose(env, "arg#%d expected pointer to allocated object\n", i); 10918 return -EINVAL; 10919 } 10920 if (!reg->ref_obj_id) { 10921 verbose(env, "allocated object must be referenced\n"); 10922 return -EINVAL; 10923 } 10924 ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta); 10925 if (ret < 0) 10926 return ret; 10927 break; 10928 case KF_ARG_PTR_TO_RB_NODE: 10929 if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_remove]) { 10930 if (!type_is_non_owning_ref(reg->type) || reg->ref_obj_id) { 10931 verbose(env, "rbtree_remove node input must be non-owning ref\n"); 10932 return -EINVAL; 10933 } 10934 if (in_rbtree_lock_required_cb(env)) { 10935 verbose(env, "rbtree_remove not allowed in rbtree cb\n"); 10936 return -EINVAL; 10937 } 10938 } else { 10939 if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) { 10940 verbose(env, "arg#%d expected pointer to allocated object\n", i); 10941 return -EINVAL; 10942 } 10943 if (!reg->ref_obj_id) { 10944 verbose(env, "allocated object must be referenced\n"); 10945 return -EINVAL; 10946 } 10947 } 10948 10949 ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta); 10950 if (ret < 0) 10951 return ret; 10952 break; 10953 case KF_ARG_PTR_TO_BTF_ID: 10954 /* Only base_type is checked, further checks are done here */ 10955 if ((base_type(reg->type) != PTR_TO_BTF_ID || 10956 (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) && 10957 !reg2btf_ids[base_type(reg->type)]) { 10958 verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type)); 10959 verbose(env, "expected %s or socket\n", 10960 reg_type_str(env, base_type(reg->type) | 10961 (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS))); 10962 return -EINVAL; 10963 } 10964 ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i); 10965 if (ret < 0) 10966 return ret; 10967 break; 10968 case KF_ARG_PTR_TO_MEM: 10969 resolve_ret = btf_resolve_size(btf, ref_t, &type_size); 10970 if (IS_ERR(resolve_ret)) { 10971 verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n", 10972 i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret)); 10973 return -EINVAL; 10974 } 10975 ret = check_mem_reg(env, reg, regno, type_size); 10976 if (ret < 0) 10977 return ret; 10978 break; 10979 case KF_ARG_PTR_TO_MEM_SIZE: 10980 { 10981 struct bpf_reg_state *buff_reg = ®s[regno]; 10982 const struct btf_param *buff_arg = &args[i]; 10983 struct bpf_reg_state *size_reg = ®s[regno + 1]; 10984 const struct btf_param *size_arg = &args[i + 1]; 10985 10986 if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) { 10987 ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1); 10988 if (ret < 0) { 10989 verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1); 10990 return ret; 10991 } 10992 } 10993 10994 if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) { 10995 if (meta->arg_constant.found) { 10996 verbose(env, "verifier internal error: only one constant argument permitted\n"); 10997 return -EFAULT; 10998 } 10999 if (!tnum_is_const(size_reg->var_off)) { 11000 verbose(env, "R%d must be a known constant\n", regno + 1); 11001 return -EINVAL; 11002 } 11003 meta->arg_constant.found = true; 11004 meta->arg_constant.value = size_reg->var_off.value; 11005 } 11006 11007 /* Skip next '__sz' or '__szk' argument */ 11008 i++; 11009 break; 11010 } 11011 case KF_ARG_PTR_TO_CALLBACK: 11012 meta->subprogno = reg->subprogno; 11013 break; 11014 case KF_ARG_PTR_TO_REFCOUNTED_KPTR: 11015 if (!type_is_ptr_alloc_obj(reg->type)) { 11016 verbose(env, "arg#%d is neither owning or non-owning ref\n", i); 11017 return -EINVAL; 11018 } 11019 if (!type_is_non_owning_ref(reg->type)) 11020 meta->arg_owning_ref = true; 11021 11022 rec = reg_btf_record(reg); 11023 if (!rec) { 11024 verbose(env, "verifier internal error: Couldn't find btf_record\n"); 11025 return -EFAULT; 11026 } 11027 11028 if (rec->refcount_off < 0) { 11029 verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i); 11030 return -EINVAL; 11031 } 11032 if (rec->refcount_off >= 0) { 11033 verbose(env, "bpf_refcount_acquire calls are disabled for now\n"); 11034 return -EINVAL; 11035 } 11036 meta->arg_btf = reg->btf; 11037 meta->arg_btf_id = reg->btf_id; 11038 break; 11039 } 11040 } 11041 11042 if (is_kfunc_release(meta) && !meta->release_regno) { 11043 verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n", 11044 func_name); 11045 return -EINVAL; 11046 } 11047 11048 return 0; 11049 } 11050 11051 static int fetch_kfunc_meta(struct bpf_verifier_env *env, 11052 struct bpf_insn *insn, 11053 struct bpf_kfunc_call_arg_meta *meta, 11054 const char **kfunc_name) 11055 { 11056 const struct btf_type *func, *func_proto; 11057 u32 func_id, *kfunc_flags; 11058 const char *func_name; 11059 struct btf *desc_btf; 11060 11061 if (kfunc_name) 11062 *kfunc_name = NULL; 11063 11064 if (!insn->imm) 11065 return -EINVAL; 11066 11067 desc_btf = find_kfunc_desc_btf(env, insn->off); 11068 if (IS_ERR(desc_btf)) 11069 return PTR_ERR(desc_btf); 11070 11071 func_id = insn->imm; 11072 func = btf_type_by_id(desc_btf, func_id); 11073 func_name = btf_name_by_offset(desc_btf, func->name_off); 11074 if (kfunc_name) 11075 *kfunc_name = func_name; 11076 func_proto = btf_type_by_id(desc_btf, func->type); 11077 11078 kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog); 11079 if (!kfunc_flags) { 11080 return -EACCES; 11081 } 11082 11083 memset(meta, 0, sizeof(*meta)); 11084 meta->btf = desc_btf; 11085 meta->func_id = func_id; 11086 meta->kfunc_flags = *kfunc_flags; 11087 meta->func_proto = func_proto; 11088 meta->func_name = func_name; 11089 11090 return 0; 11091 } 11092 11093 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, 11094 int *insn_idx_p) 11095 { 11096 const struct btf_type *t, *ptr_type; 11097 u32 i, nargs, ptr_type_id, release_ref_obj_id; 11098 struct bpf_reg_state *regs = cur_regs(env); 11099 const char *func_name, *ptr_type_name; 11100 bool sleepable, rcu_lock, rcu_unlock; 11101 struct bpf_kfunc_call_arg_meta meta; 11102 struct bpf_insn_aux_data *insn_aux; 11103 int err, insn_idx = *insn_idx_p; 11104 const struct btf_param *args; 11105 const struct btf_type *ret_t; 11106 struct btf *desc_btf; 11107 11108 /* skip for now, but return error when we find this in fixup_kfunc_call */ 11109 if (!insn->imm) 11110 return 0; 11111 11112 err = fetch_kfunc_meta(env, insn, &meta, &func_name); 11113 if (err == -EACCES && func_name) 11114 verbose(env, "calling kernel function %s is not allowed\n", func_name); 11115 if (err) 11116 return err; 11117 desc_btf = meta.btf; 11118 insn_aux = &env->insn_aux_data[insn_idx]; 11119 11120 insn_aux->is_iter_next = is_iter_next_kfunc(&meta); 11121 11122 if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) { 11123 verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n"); 11124 return -EACCES; 11125 } 11126 11127 sleepable = is_kfunc_sleepable(&meta); 11128 if (sleepable && !env->prog->aux->sleepable) { 11129 verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name); 11130 return -EACCES; 11131 } 11132 11133 rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta); 11134 rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta); 11135 11136 if (env->cur_state->active_rcu_lock) { 11137 struct bpf_func_state *state; 11138 struct bpf_reg_state *reg; 11139 11140 if (rcu_lock) { 11141 verbose(env, "nested rcu read lock (kernel function %s)\n", func_name); 11142 return -EINVAL; 11143 } else if (rcu_unlock) { 11144 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({ 11145 if (reg->type & MEM_RCU) { 11146 reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL); 11147 reg->type |= PTR_UNTRUSTED; 11148 } 11149 })); 11150 env->cur_state->active_rcu_lock = false; 11151 } else if (sleepable) { 11152 verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name); 11153 return -EACCES; 11154 } 11155 } else if (rcu_lock) { 11156 env->cur_state->active_rcu_lock = true; 11157 } else if (rcu_unlock) { 11158 verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name); 11159 return -EINVAL; 11160 } 11161 11162 /* Check the arguments */ 11163 err = check_kfunc_args(env, &meta, insn_idx); 11164 if (err < 0) 11165 return err; 11166 /* In case of release function, we get register number of refcounted 11167 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now. 11168 */ 11169 if (meta.release_regno) { 11170 err = release_reference(env, regs[meta.release_regno].ref_obj_id); 11171 if (err) { 11172 verbose(env, "kfunc %s#%d reference has not been acquired before\n", 11173 func_name, meta.func_id); 11174 return err; 11175 } 11176 } 11177 11178 if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] || 11179 meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] || 11180 meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) { 11181 release_ref_obj_id = regs[BPF_REG_2].ref_obj_id; 11182 insn_aux->insert_off = regs[BPF_REG_2].off; 11183 insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id); 11184 err = ref_convert_owning_non_owning(env, release_ref_obj_id); 11185 if (err) { 11186 verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n", 11187 func_name, meta.func_id); 11188 return err; 11189 } 11190 11191 err = release_reference(env, release_ref_obj_id); 11192 if (err) { 11193 verbose(env, "kfunc %s#%d reference has not been acquired before\n", 11194 func_name, meta.func_id); 11195 return err; 11196 } 11197 } 11198 11199 if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) { 11200 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno, 11201 set_rbtree_add_callback_state); 11202 if (err) { 11203 verbose(env, "kfunc %s#%d failed callback verification\n", 11204 func_name, meta.func_id); 11205 return err; 11206 } 11207 } 11208 11209 for (i = 0; i < CALLER_SAVED_REGS; i++) 11210 mark_reg_not_init(env, regs, caller_saved[i]); 11211 11212 /* Check return type */ 11213 t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL); 11214 11215 if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) { 11216 /* Only exception is bpf_obj_new_impl */ 11217 if (meta.btf != btf_vmlinux || 11218 (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] && 11219 meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) { 11220 verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n"); 11221 return -EINVAL; 11222 } 11223 } 11224 11225 if (btf_type_is_scalar(t)) { 11226 mark_reg_unknown(env, regs, BPF_REG_0); 11227 mark_btf_func_reg_size(env, BPF_REG_0, t->size); 11228 } else if (btf_type_is_ptr(t)) { 11229 ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id); 11230 11231 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) { 11232 if (meta.func_id == special_kfunc_list[KF_bpf_obj_new_impl]) { 11233 struct btf *ret_btf; 11234 u32 ret_btf_id; 11235 11236 if (unlikely(!bpf_global_ma_set)) 11237 return -ENOMEM; 11238 11239 if (((u64)(u32)meta.arg_constant.value) != meta.arg_constant.value) { 11240 verbose(env, "local type ID argument must be in range [0, U32_MAX]\n"); 11241 return -EINVAL; 11242 } 11243 11244 ret_btf = env->prog->aux->btf; 11245 ret_btf_id = meta.arg_constant.value; 11246 11247 /* This may be NULL due to user not supplying a BTF */ 11248 if (!ret_btf) { 11249 verbose(env, "bpf_obj_new requires prog BTF\n"); 11250 return -EINVAL; 11251 } 11252 11253 ret_t = btf_type_by_id(ret_btf, ret_btf_id); 11254 if (!ret_t || !__btf_type_is_struct(ret_t)) { 11255 verbose(env, "bpf_obj_new type ID argument must be of a struct\n"); 11256 return -EINVAL; 11257 } 11258 11259 mark_reg_known_zero(env, regs, BPF_REG_0); 11260 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC; 11261 regs[BPF_REG_0].btf = ret_btf; 11262 regs[BPF_REG_0].btf_id = ret_btf_id; 11263 11264 insn_aux->obj_new_size = ret_t->size; 11265 insn_aux->kptr_struct_meta = 11266 btf_find_struct_meta(ret_btf, ret_btf_id); 11267 } else if (meta.func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) { 11268 mark_reg_known_zero(env, regs, BPF_REG_0); 11269 regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC; 11270 regs[BPF_REG_0].btf = meta.arg_btf; 11271 regs[BPF_REG_0].btf_id = meta.arg_btf_id; 11272 11273 insn_aux->kptr_struct_meta = 11274 btf_find_struct_meta(meta.arg_btf, 11275 meta.arg_btf_id); 11276 } else if (meta.func_id == special_kfunc_list[KF_bpf_list_pop_front] || 11277 meta.func_id == special_kfunc_list[KF_bpf_list_pop_back]) { 11278 struct btf_field *field = meta.arg_list_head.field; 11279 11280 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root); 11281 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_remove] || 11282 meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) { 11283 struct btf_field *field = meta.arg_rbtree_root.field; 11284 11285 mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root); 11286 } else if (meta.func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) { 11287 mark_reg_known_zero(env, regs, BPF_REG_0); 11288 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED; 11289 regs[BPF_REG_0].btf = desc_btf; 11290 regs[BPF_REG_0].btf_id = meta.ret_btf_id; 11291 } else if (meta.func_id == special_kfunc_list[KF_bpf_rdonly_cast]) { 11292 ret_t = btf_type_by_id(desc_btf, meta.arg_constant.value); 11293 if (!ret_t || !btf_type_is_struct(ret_t)) { 11294 verbose(env, 11295 "kfunc bpf_rdonly_cast type ID argument must be of a struct\n"); 11296 return -EINVAL; 11297 } 11298 11299 mark_reg_known_zero(env, regs, BPF_REG_0); 11300 regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED; 11301 regs[BPF_REG_0].btf = desc_btf; 11302 regs[BPF_REG_0].btf_id = meta.arg_constant.value; 11303 } else if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice] || 11304 meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) { 11305 enum bpf_type_flag type_flag = get_dynptr_type_flag(meta.initialized_dynptr.type); 11306 11307 mark_reg_known_zero(env, regs, BPF_REG_0); 11308 11309 if (!meta.arg_constant.found) { 11310 verbose(env, "verifier internal error: bpf_dynptr_slice(_rdwr) no constant size\n"); 11311 return -EFAULT; 11312 } 11313 11314 regs[BPF_REG_0].mem_size = meta.arg_constant.value; 11315 11316 /* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */ 11317 regs[BPF_REG_0].type = PTR_TO_MEM | type_flag; 11318 11319 if (meta.func_id == special_kfunc_list[KF_bpf_dynptr_slice]) { 11320 regs[BPF_REG_0].type |= MEM_RDONLY; 11321 } else { 11322 /* this will set env->seen_direct_write to true */ 11323 if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) { 11324 verbose(env, "the prog does not allow writes to packet data\n"); 11325 return -EINVAL; 11326 } 11327 } 11328 11329 if (!meta.initialized_dynptr.id) { 11330 verbose(env, "verifier internal error: no dynptr id\n"); 11331 return -EFAULT; 11332 } 11333 regs[BPF_REG_0].dynptr_id = meta.initialized_dynptr.id; 11334 11335 /* we don't need to set BPF_REG_0's ref obj id 11336 * because packet slices are not refcounted (see 11337 * dynptr_type_refcounted) 11338 */ 11339 } else { 11340 verbose(env, "kernel function %s unhandled dynamic return type\n", 11341 meta.func_name); 11342 return -EFAULT; 11343 } 11344 } else if (!__btf_type_is_struct(ptr_type)) { 11345 if (!meta.r0_size) { 11346 __u32 sz; 11347 11348 if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) { 11349 meta.r0_size = sz; 11350 meta.r0_rdonly = true; 11351 } 11352 } 11353 if (!meta.r0_size) { 11354 ptr_type_name = btf_name_by_offset(desc_btf, 11355 ptr_type->name_off); 11356 verbose(env, 11357 "kernel function %s returns pointer type %s %s is not supported\n", 11358 func_name, 11359 btf_type_str(ptr_type), 11360 ptr_type_name); 11361 return -EINVAL; 11362 } 11363 11364 mark_reg_known_zero(env, regs, BPF_REG_0); 11365 regs[BPF_REG_0].type = PTR_TO_MEM; 11366 regs[BPF_REG_0].mem_size = meta.r0_size; 11367 11368 if (meta.r0_rdonly) 11369 regs[BPF_REG_0].type |= MEM_RDONLY; 11370 11371 /* Ensures we don't access the memory after a release_reference() */ 11372 if (meta.ref_obj_id) 11373 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id; 11374 } else { 11375 mark_reg_known_zero(env, regs, BPF_REG_0); 11376 regs[BPF_REG_0].btf = desc_btf; 11377 regs[BPF_REG_0].type = PTR_TO_BTF_ID; 11378 regs[BPF_REG_0].btf_id = ptr_type_id; 11379 } 11380 11381 if (is_kfunc_ret_null(&meta)) { 11382 regs[BPF_REG_0].type |= PTR_MAYBE_NULL; 11383 /* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */ 11384 regs[BPF_REG_0].id = ++env->id_gen; 11385 } 11386 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *)); 11387 if (is_kfunc_acquire(&meta)) { 11388 int id = acquire_reference_state(env, insn_idx); 11389 11390 if (id < 0) 11391 return id; 11392 if (is_kfunc_ret_null(&meta)) 11393 regs[BPF_REG_0].id = id; 11394 regs[BPF_REG_0].ref_obj_id = id; 11395 } else if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_first]) { 11396 ref_set_non_owning(env, ®s[BPF_REG_0]); 11397 } 11398 11399 if (reg_may_point_to_spin_lock(®s[BPF_REG_0]) && !regs[BPF_REG_0].id) 11400 regs[BPF_REG_0].id = ++env->id_gen; 11401 } else if (btf_type_is_void(t)) { 11402 if (meta.btf == btf_vmlinux && btf_id_set_contains(&special_kfunc_set, meta.func_id)) { 11403 if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl]) { 11404 insn_aux->kptr_struct_meta = 11405 btf_find_struct_meta(meta.arg_btf, 11406 meta.arg_btf_id); 11407 } 11408 } 11409 } 11410 11411 nargs = btf_type_vlen(meta.func_proto); 11412 args = (const struct btf_param *)(meta.func_proto + 1); 11413 for (i = 0; i < nargs; i++) { 11414 u32 regno = i + 1; 11415 11416 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL); 11417 if (btf_type_is_ptr(t)) 11418 mark_btf_func_reg_size(env, regno, sizeof(void *)); 11419 else 11420 /* scalar. ensured by btf_check_kfunc_arg_match() */ 11421 mark_btf_func_reg_size(env, regno, t->size); 11422 } 11423 11424 if (is_iter_next_kfunc(&meta)) { 11425 err = process_iter_next_call(env, insn_idx, &meta); 11426 if (err) 11427 return err; 11428 } 11429 11430 return 0; 11431 } 11432 11433 static bool signed_add_overflows(s64 a, s64 b) 11434 { 11435 /* Do the add in u64, where overflow is well-defined */ 11436 s64 res = (s64)((u64)a + (u64)b); 11437 11438 if (b < 0) 11439 return res > a; 11440 return res < a; 11441 } 11442 11443 static bool signed_add32_overflows(s32 a, s32 b) 11444 { 11445 /* Do the add in u32, where overflow is well-defined */ 11446 s32 res = (s32)((u32)a + (u32)b); 11447 11448 if (b < 0) 11449 return res > a; 11450 return res < a; 11451 } 11452 11453 static bool signed_sub_overflows(s64 a, s64 b) 11454 { 11455 /* Do the sub in u64, where overflow is well-defined */ 11456 s64 res = (s64)((u64)a - (u64)b); 11457 11458 if (b < 0) 11459 return res < a; 11460 return res > a; 11461 } 11462 11463 static bool signed_sub32_overflows(s32 a, s32 b) 11464 { 11465 /* Do the sub in u32, where overflow is well-defined */ 11466 s32 res = (s32)((u32)a - (u32)b); 11467 11468 if (b < 0) 11469 return res < a; 11470 return res > a; 11471 } 11472 11473 static bool check_reg_sane_offset(struct bpf_verifier_env *env, 11474 const struct bpf_reg_state *reg, 11475 enum bpf_reg_type type) 11476 { 11477 bool known = tnum_is_const(reg->var_off); 11478 s64 val = reg->var_off.value; 11479 s64 smin = reg->smin_value; 11480 11481 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) { 11482 verbose(env, "math between %s pointer and %lld is not allowed\n", 11483 reg_type_str(env, type), val); 11484 return false; 11485 } 11486 11487 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) { 11488 verbose(env, "%s pointer offset %d is not allowed\n", 11489 reg_type_str(env, type), reg->off); 11490 return false; 11491 } 11492 11493 if (smin == S64_MIN) { 11494 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n", 11495 reg_type_str(env, type)); 11496 return false; 11497 } 11498 11499 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) { 11500 verbose(env, "value %lld makes %s pointer be out of bounds\n", 11501 smin, reg_type_str(env, type)); 11502 return false; 11503 } 11504 11505 return true; 11506 } 11507 11508 enum { 11509 REASON_BOUNDS = -1, 11510 REASON_TYPE = -2, 11511 REASON_PATHS = -3, 11512 REASON_LIMIT = -4, 11513 REASON_STACK = -5, 11514 }; 11515 11516 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg, 11517 u32 *alu_limit, bool mask_to_left) 11518 { 11519 u32 max = 0, ptr_limit = 0; 11520 11521 switch (ptr_reg->type) { 11522 case PTR_TO_STACK: 11523 /* Offset 0 is out-of-bounds, but acceptable start for the 11524 * left direction, see BPF_REG_FP. Also, unknown scalar 11525 * offset where we would need to deal with min/max bounds is 11526 * currently prohibited for unprivileged. 11527 */ 11528 max = MAX_BPF_STACK + mask_to_left; 11529 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off); 11530 break; 11531 case PTR_TO_MAP_VALUE: 11532 max = ptr_reg->map_ptr->value_size; 11533 ptr_limit = (mask_to_left ? 11534 ptr_reg->smin_value : 11535 ptr_reg->umax_value) + ptr_reg->off; 11536 break; 11537 default: 11538 return REASON_TYPE; 11539 } 11540 11541 if (ptr_limit >= max) 11542 return REASON_LIMIT; 11543 *alu_limit = ptr_limit; 11544 return 0; 11545 } 11546 11547 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env, 11548 const struct bpf_insn *insn) 11549 { 11550 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K; 11551 } 11552 11553 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux, 11554 u32 alu_state, u32 alu_limit) 11555 { 11556 /* If we arrived here from different branches with different 11557 * state or limits to sanitize, then this won't work. 11558 */ 11559 if (aux->alu_state && 11560 (aux->alu_state != alu_state || 11561 aux->alu_limit != alu_limit)) 11562 return REASON_PATHS; 11563 11564 /* Corresponding fixup done in do_misc_fixups(). */ 11565 aux->alu_state = alu_state; 11566 aux->alu_limit = alu_limit; 11567 return 0; 11568 } 11569 11570 static int sanitize_val_alu(struct bpf_verifier_env *env, 11571 struct bpf_insn *insn) 11572 { 11573 struct bpf_insn_aux_data *aux = cur_aux(env); 11574 11575 if (can_skip_alu_sanitation(env, insn)) 11576 return 0; 11577 11578 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0); 11579 } 11580 11581 static bool sanitize_needed(u8 opcode) 11582 { 11583 return opcode == BPF_ADD || opcode == BPF_SUB; 11584 } 11585 11586 struct bpf_sanitize_info { 11587 struct bpf_insn_aux_data aux; 11588 bool mask_to_left; 11589 }; 11590 11591 static struct bpf_verifier_state * 11592 sanitize_speculative_path(struct bpf_verifier_env *env, 11593 const struct bpf_insn *insn, 11594 u32 next_idx, u32 curr_idx) 11595 { 11596 struct bpf_verifier_state *branch; 11597 struct bpf_reg_state *regs; 11598 11599 branch = push_stack(env, next_idx, curr_idx, true); 11600 if (branch && insn) { 11601 regs = branch->frame[branch->curframe]->regs; 11602 if (BPF_SRC(insn->code) == BPF_K) { 11603 mark_reg_unknown(env, regs, insn->dst_reg); 11604 } else if (BPF_SRC(insn->code) == BPF_X) { 11605 mark_reg_unknown(env, regs, insn->dst_reg); 11606 mark_reg_unknown(env, regs, insn->src_reg); 11607 } 11608 } 11609 return branch; 11610 } 11611 11612 static int sanitize_ptr_alu(struct bpf_verifier_env *env, 11613 struct bpf_insn *insn, 11614 const struct bpf_reg_state *ptr_reg, 11615 const struct bpf_reg_state *off_reg, 11616 struct bpf_reg_state *dst_reg, 11617 struct bpf_sanitize_info *info, 11618 const bool commit_window) 11619 { 11620 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux; 11621 struct bpf_verifier_state *vstate = env->cur_state; 11622 bool off_is_imm = tnum_is_const(off_reg->var_off); 11623 bool off_is_neg = off_reg->smin_value < 0; 11624 bool ptr_is_dst_reg = ptr_reg == dst_reg; 11625 u8 opcode = BPF_OP(insn->code); 11626 u32 alu_state, alu_limit; 11627 struct bpf_reg_state tmp; 11628 bool ret; 11629 int err; 11630 11631 if (can_skip_alu_sanitation(env, insn)) 11632 return 0; 11633 11634 /* We already marked aux for masking from non-speculative 11635 * paths, thus we got here in the first place. We only care 11636 * to explore bad access from here. 11637 */ 11638 if (vstate->speculative) 11639 goto do_sim; 11640 11641 if (!commit_window) { 11642 if (!tnum_is_const(off_reg->var_off) && 11643 (off_reg->smin_value < 0) != (off_reg->smax_value < 0)) 11644 return REASON_BOUNDS; 11645 11646 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) || 11647 (opcode == BPF_SUB && !off_is_neg); 11648 } 11649 11650 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left); 11651 if (err < 0) 11652 return err; 11653 11654 if (commit_window) { 11655 /* In commit phase we narrow the masking window based on 11656 * the observed pointer move after the simulated operation. 11657 */ 11658 alu_state = info->aux.alu_state; 11659 alu_limit = abs(info->aux.alu_limit - alu_limit); 11660 } else { 11661 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0; 11662 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0; 11663 alu_state |= ptr_is_dst_reg ? 11664 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST; 11665 11666 /* Limit pruning on unknown scalars to enable deep search for 11667 * potential masking differences from other program paths. 11668 */ 11669 if (!off_is_imm) 11670 env->explore_alu_limits = true; 11671 } 11672 11673 err = update_alu_sanitation_state(aux, alu_state, alu_limit); 11674 if (err < 0) 11675 return err; 11676 do_sim: 11677 /* If we're in commit phase, we're done here given we already 11678 * pushed the truncated dst_reg into the speculative verification 11679 * stack. 11680 * 11681 * Also, when register is a known constant, we rewrite register-based 11682 * operation to immediate-based, and thus do not need masking (and as 11683 * a consequence, do not need to simulate the zero-truncation either). 11684 */ 11685 if (commit_window || off_is_imm) 11686 return 0; 11687 11688 /* Simulate and find potential out-of-bounds access under 11689 * speculative execution from truncation as a result of 11690 * masking when off was not within expected range. If off 11691 * sits in dst, then we temporarily need to move ptr there 11692 * to simulate dst (== 0) +/-= ptr. Needed, for example, 11693 * for cases where we use K-based arithmetic in one direction 11694 * and truncated reg-based in the other in order to explore 11695 * bad access. 11696 */ 11697 if (!ptr_is_dst_reg) { 11698 tmp = *dst_reg; 11699 copy_register_state(dst_reg, ptr_reg); 11700 } 11701 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1, 11702 env->insn_idx); 11703 if (!ptr_is_dst_reg && ret) 11704 *dst_reg = tmp; 11705 return !ret ? REASON_STACK : 0; 11706 } 11707 11708 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env) 11709 { 11710 struct bpf_verifier_state *vstate = env->cur_state; 11711 11712 /* If we simulate paths under speculation, we don't update the 11713 * insn as 'seen' such that when we verify unreachable paths in 11714 * the non-speculative domain, sanitize_dead_code() can still 11715 * rewrite/sanitize them. 11716 */ 11717 if (!vstate->speculative) 11718 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt; 11719 } 11720 11721 static int sanitize_err(struct bpf_verifier_env *env, 11722 const struct bpf_insn *insn, int reason, 11723 const struct bpf_reg_state *off_reg, 11724 const struct bpf_reg_state *dst_reg) 11725 { 11726 static const char *err = "pointer arithmetic with it prohibited for !root"; 11727 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub"; 11728 u32 dst = insn->dst_reg, src = insn->src_reg; 11729 11730 switch (reason) { 11731 case REASON_BOUNDS: 11732 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n", 11733 off_reg == dst_reg ? dst : src, err); 11734 break; 11735 case REASON_TYPE: 11736 verbose(env, "R%d has pointer with unsupported alu operation, %s\n", 11737 off_reg == dst_reg ? src : dst, err); 11738 break; 11739 case REASON_PATHS: 11740 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n", 11741 dst, op, err); 11742 break; 11743 case REASON_LIMIT: 11744 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n", 11745 dst, op, err); 11746 break; 11747 case REASON_STACK: 11748 verbose(env, "R%d could not be pushed for speculative verification, %s\n", 11749 dst, err); 11750 break; 11751 default: 11752 verbose(env, "verifier internal error: unknown reason (%d)\n", 11753 reason); 11754 break; 11755 } 11756 11757 return -EACCES; 11758 } 11759 11760 /* check that stack access falls within stack limits and that 'reg' doesn't 11761 * have a variable offset. 11762 * 11763 * Variable offset is prohibited for unprivileged mode for simplicity since it 11764 * requires corresponding support in Spectre masking for stack ALU. See also 11765 * retrieve_ptr_limit(). 11766 * 11767 * 11768 * 'off' includes 'reg->off'. 11769 */ 11770 static int check_stack_access_for_ptr_arithmetic( 11771 struct bpf_verifier_env *env, 11772 int regno, 11773 const struct bpf_reg_state *reg, 11774 int off) 11775 { 11776 if (!tnum_is_const(reg->var_off)) { 11777 char tn_buf[48]; 11778 11779 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off); 11780 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n", 11781 regno, tn_buf, off); 11782 return -EACCES; 11783 } 11784 11785 if (off >= 0 || off < -MAX_BPF_STACK) { 11786 verbose(env, "R%d stack pointer arithmetic goes out of range, " 11787 "prohibited for !root; off=%d\n", regno, off); 11788 return -EACCES; 11789 } 11790 11791 return 0; 11792 } 11793 11794 static int sanitize_check_bounds(struct bpf_verifier_env *env, 11795 const struct bpf_insn *insn, 11796 const struct bpf_reg_state *dst_reg) 11797 { 11798 u32 dst = insn->dst_reg; 11799 11800 /* For unprivileged we require that resulting offset must be in bounds 11801 * in order to be able to sanitize access later on. 11802 */ 11803 if (env->bypass_spec_v1) 11804 return 0; 11805 11806 switch (dst_reg->type) { 11807 case PTR_TO_STACK: 11808 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg, 11809 dst_reg->off + dst_reg->var_off.value)) 11810 return -EACCES; 11811 break; 11812 case PTR_TO_MAP_VALUE: 11813 if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) { 11814 verbose(env, "R%d pointer arithmetic of map value goes out of range, " 11815 "prohibited for !root\n", dst); 11816 return -EACCES; 11817 } 11818 break; 11819 default: 11820 break; 11821 } 11822 11823 return 0; 11824 } 11825 11826 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off. 11827 * Caller should also handle BPF_MOV case separately. 11828 * If we return -EACCES, caller may want to try again treating pointer as a 11829 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks. 11830 */ 11831 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env, 11832 struct bpf_insn *insn, 11833 const struct bpf_reg_state *ptr_reg, 11834 const struct bpf_reg_state *off_reg) 11835 { 11836 struct bpf_verifier_state *vstate = env->cur_state; 11837 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 11838 struct bpf_reg_state *regs = state->regs, *dst_reg; 11839 bool known = tnum_is_const(off_reg->var_off); 11840 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value, 11841 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value; 11842 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value, 11843 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value; 11844 struct bpf_sanitize_info info = {}; 11845 u8 opcode = BPF_OP(insn->code); 11846 u32 dst = insn->dst_reg; 11847 int ret; 11848 11849 dst_reg = ®s[dst]; 11850 11851 if ((known && (smin_val != smax_val || umin_val != umax_val)) || 11852 smin_val > smax_val || umin_val > umax_val) { 11853 /* Taint dst register if offset had invalid bounds derived from 11854 * e.g. dead branches. 11855 */ 11856 __mark_reg_unknown(env, dst_reg); 11857 return 0; 11858 } 11859 11860 if (BPF_CLASS(insn->code) != BPF_ALU64) { 11861 /* 32-bit ALU ops on pointers produce (meaningless) scalars */ 11862 if (opcode == BPF_SUB && env->allow_ptr_leaks) { 11863 __mark_reg_unknown(env, dst_reg); 11864 return 0; 11865 } 11866 11867 verbose(env, 11868 "R%d 32-bit pointer arithmetic prohibited\n", 11869 dst); 11870 return -EACCES; 11871 } 11872 11873 if (ptr_reg->type & PTR_MAYBE_NULL) { 11874 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n", 11875 dst, reg_type_str(env, ptr_reg->type)); 11876 return -EACCES; 11877 } 11878 11879 switch (base_type(ptr_reg->type)) { 11880 case CONST_PTR_TO_MAP: 11881 /* smin_val represents the known value */ 11882 if (known && smin_val == 0 && opcode == BPF_ADD) 11883 break; 11884 fallthrough; 11885 case PTR_TO_PACKET_END: 11886 case PTR_TO_SOCKET: 11887 case PTR_TO_SOCK_COMMON: 11888 case PTR_TO_TCP_SOCK: 11889 case PTR_TO_XDP_SOCK: 11890 verbose(env, "R%d pointer arithmetic on %s prohibited\n", 11891 dst, reg_type_str(env, ptr_reg->type)); 11892 return -EACCES; 11893 default: 11894 break; 11895 } 11896 11897 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id. 11898 * The id may be overwritten later if we create a new variable offset. 11899 */ 11900 dst_reg->type = ptr_reg->type; 11901 dst_reg->id = ptr_reg->id; 11902 11903 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) || 11904 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type)) 11905 return -EINVAL; 11906 11907 /* pointer types do not carry 32-bit bounds at the moment. */ 11908 __mark_reg32_unbounded(dst_reg); 11909 11910 if (sanitize_needed(opcode)) { 11911 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg, 11912 &info, false); 11913 if (ret < 0) 11914 return sanitize_err(env, insn, ret, off_reg, dst_reg); 11915 } 11916 11917 switch (opcode) { 11918 case BPF_ADD: 11919 /* We can take a fixed offset as long as it doesn't overflow 11920 * the s32 'off' field 11921 */ 11922 if (known && (ptr_reg->off + smin_val == 11923 (s64)(s32)(ptr_reg->off + smin_val))) { 11924 /* pointer += K. Accumulate it into fixed offset */ 11925 dst_reg->smin_value = smin_ptr; 11926 dst_reg->smax_value = smax_ptr; 11927 dst_reg->umin_value = umin_ptr; 11928 dst_reg->umax_value = umax_ptr; 11929 dst_reg->var_off = ptr_reg->var_off; 11930 dst_reg->off = ptr_reg->off + smin_val; 11931 dst_reg->raw = ptr_reg->raw; 11932 break; 11933 } 11934 /* A new variable offset is created. Note that off_reg->off 11935 * == 0, since it's a scalar. 11936 * dst_reg gets the pointer type and since some positive 11937 * integer value was added to the pointer, give it a new 'id' 11938 * if it's a PTR_TO_PACKET. 11939 * this creates a new 'base' pointer, off_reg (variable) gets 11940 * added into the variable offset, and we copy the fixed offset 11941 * from ptr_reg. 11942 */ 11943 if (signed_add_overflows(smin_ptr, smin_val) || 11944 signed_add_overflows(smax_ptr, smax_val)) { 11945 dst_reg->smin_value = S64_MIN; 11946 dst_reg->smax_value = S64_MAX; 11947 } else { 11948 dst_reg->smin_value = smin_ptr + smin_val; 11949 dst_reg->smax_value = smax_ptr + smax_val; 11950 } 11951 if (umin_ptr + umin_val < umin_ptr || 11952 umax_ptr + umax_val < umax_ptr) { 11953 dst_reg->umin_value = 0; 11954 dst_reg->umax_value = U64_MAX; 11955 } else { 11956 dst_reg->umin_value = umin_ptr + umin_val; 11957 dst_reg->umax_value = umax_ptr + umax_val; 11958 } 11959 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off); 11960 dst_reg->off = ptr_reg->off; 11961 dst_reg->raw = ptr_reg->raw; 11962 if (reg_is_pkt_pointer(ptr_reg)) { 11963 dst_reg->id = ++env->id_gen; 11964 /* something was added to pkt_ptr, set range to zero */ 11965 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw)); 11966 } 11967 break; 11968 case BPF_SUB: 11969 if (dst_reg == off_reg) { 11970 /* scalar -= pointer. Creates an unknown scalar */ 11971 verbose(env, "R%d tried to subtract pointer from scalar\n", 11972 dst); 11973 return -EACCES; 11974 } 11975 /* We don't allow subtraction from FP, because (according to 11976 * test_verifier.c test "invalid fp arithmetic", JITs might not 11977 * be able to deal with it. 11978 */ 11979 if (ptr_reg->type == PTR_TO_STACK) { 11980 verbose(env, "R%d subtraction from stack pointer prohibited\n", 11981 dst); 11982 return -EACCES; 11983 } 11984 if (known && (ptr_reg->off - smin_val == 11985 (s64)(s32)(ptr_reg->off - smin_val))) { 11986 /* pointer -= K. Subtract it from fixed offset */ 11987 dst_reg->smin_value = smin_ptr; 11988 dst_reg->smax_value = smax_ptr; 11989 dst_reg->umin_value = umin_ptr; 11990 dst_reg->umax_value = umax_ptr; 11991 dst_reg->var_off = ptr_reg->var_off; 11992 dst_reg->id = ptr_reg->id; 11993 dst_reg->off = ptr_reg->off - smin_val; 11994 dst_reg->raw = ptr_reg->raw; 11995 break; 11996 } 11997 /* A new variable offset is created. If the subtrahend is known 11998 * nonnegative, then any reg->range we had before is still good. 11999 */ 12000 if (signed_sub_overflows(smin_ptr, smax_val) || 12001 signed_sub_overflows(smax_ptr, smin_val)) { 12002 /* Overflow possible, we know nothing */ 12003 dst_reg->smin_value = S64_MIN; 12004 dst_reg->smax_value = S64_MAX; 12005 } else { 12006 dst_reg->smin_value = smin_ptr - smax_val; 12007 dst_reg->smax_value = smax_ptr - smin_val; 12008 } 12009 if (umin_ptr < umax_val) { 12010 /* Overflow possible, we know nothing */ 12011 dst_reg->umin_value = 0; 12012 dst_reg->umax_value = U64_MAX; 12013 } else { 12014 /* Cannot overflow (as long as bounds are consistent) */ 12015 dst_reg->umin_value = umin_ptr - umax_val; 12016 dst_reg->umax_value = umax_ptr - umin_val; 12017 } 12018 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off); 12019 dst_reg->off = ptr_reg->off; 12020 dst_reg->raw = ptr_reg->raw; 12021 if (reg_is_pkt_pointer(ptr_reg)) { 12022 dst_reg->id = ++env->id_gen; 12023 /* something was added to pkt_ptr, set range to zero */ 12024 if (smin_val < 0) 12025 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw)); 12026 } 12027 break; 12028 case BPF_AND: 12029 case BPF_OR: 12030 case BPF_XOR: 12031 /* bitwise ops on pointers are troublesome, prohibit. */ 12032 verbose(env, "R%d bitwise operator %s on pointer prohibited\n", 12033 dst, bpf_alu_string[opcode >> 4]); 12034 return -EACCES; 12035 default: 12036 /* other operators (e.g. MUL,LSH) produce non-pointer results */ 12037 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n", 12038 dst, bpf_alu_string[opcode >> 4]); 12039 return -EACCES; 12040 } 12041 12042 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type)) 12043 return -EINVAL; 12044 reg_bounds_sync(dst_reg); 12045 if (sanitize_check_bounds(env, insn, dst_reg) < 0) 12046 return -EACCES; 12047 if (sanitize_needed(opcode)) { 12048 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg, 12049 &info, true); 12050 if (ret < 0) 12051 return sanitize_err(env, insn, ret, off_reg, dst_reg); 12052 } 12053 12054 return 0; 12055 } 12056 12057 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg, 12058 struct bpf_reg_state *src_reg) 12059 { 12060 s32 smin_val = src_reg->s32_min_value; 12061 s32 smax_val = src_reg->s32_max_value; 12062 u32 umin_val = src_reg->u32_min_value; 12063 u32 umax_val = src_reg->u32_max_value; 12064 12065 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) || 12066 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) { 12067 dst_reg->s32_min_value = S32_MIN; 12068 dst_reg->s32_max_value = S32_MAX; 12069 } else { 12070 dst_reg->s32_min_value += smin_val; 12071 dst_reg->s32_max_value += smax_val; 12072 } 12073 if (dst_reg->u32_min_value + umin_val < umin_val || 12074 dst_reg->u32_max_value + umax_val < umax_val) { 12075 dst_reg->u32_min_value = 0; 12076 dst_reg->u32_max_value = U32_MAX; 12077 } else { 12078 dst_reg->u32_min_value += umin_val; 12079 dst_reg->u32_max_value += umax_val; 12080 } 12081 } 12082 12083 static void scalar_min_max_add(struct bpf_reg_state *dst_reg, 12084 struct bpf_reg_state *src_reg) 12085 { 12086 s64 smin_val = src_reg->smin_value; 12087 s64 smax_val = src_reg->smax_value; 12088 u64 umin_val = src_reg->umin_value; 12089 u64 umax_val = src_reg->umax_value; 12090 12091 if (signed_add_overflows(dst_reg->smin_value, smin_val) || 12092 signed_add_overflows(dst_reg->smax_value, smax_val)) { 12093 dst_reg->smin_value = S64_MIN; 12094 dst_reg->smax_value = S64_MAX; 12095 } else { 12096 dst_reg->smin_value += smin_val; 12097 dst_reg->smax_value += smax_val; 12098 } 12099 if (dst_reg->umin_value + umin_val < umin_val || 12100 dst_reg->umax_value + umax_val < umax_val) { 12101 dst_reg->umin_value = 0; 12102 dst_reg->umax_value = U64_MAX; 12103 } else { 12104 dst_reg->umin_value += umin_val; 12105 dst_reg->umax_value += umax_val; 12106 } 12107 } 12108 12109 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg, 12110 struct bpf_reg_state *src_reg) 12111 { 12112 s32 smin_val = src_reg->s32_min_value; 12113 s32 smax_val = src_reg->s32_max_value; 12114 u32 umin_val = src_reg->u32_min_value; 12115 u32 umax_val = src_reg->u32_max_value; 12116 12117 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) || 12118 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) { 12119 /* Overflow possible, we know nothing */ 12120 dst_reg->s32_min_value = S32_MIN; 12121 dst_reg->s32_max_value = S32_MAX; 12122 } else { 12123 dst_reg->s32_min_value -= smax_val; 12124 dst_reg->s32_max_value -= smin_val; 12125 } 12126 if (dst_reg->u32_min_value < umax_val) { 12127 /* Overflow possible, we know nothing */ 12128 dst_reg->u32_min_value = 0; 12129 dst_reg->u32_max_value = U32_MAX; 12130 } else { 12131 /* Cannot overflow (as long as bounds are consistent) */ 12132 dst_reg->u32_min_value -= umax_val; 12133 dst_reg->u32_max_value -= umin_val; 12134 } 12135 } 12136 12137 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg, 12138 struct bpf_reg_state *src_reg) 12139 { 12140 s64 smin_val = src_reg->smin_value; 12141 s64 smax_val = src_reg->smax_value; 12142 u64 umin_val = src_reg->umin_value; 12143 u64 umax_val = src_reg->umax_value; 12144 12145 if (signed_sub_overflows(dst_reg->smin_value, smax_val) || 12146 signed_sub_overflows(dst_reg->smax_value, smin_val)) { 12147 /* Overflow possible, we know nothing */ 12148 dst_reg->smin_value = S64_MIN; 12149 dst_reg->smax_value = S64_MAX; 12150 } else { 12151 dst_reg->smin_value -= smax_val; 12152 dst_reg->smax_value -= smin_val; 12153 } 12154 if (dst_reg->umin_value < umax_val) { 12155 /* Overflow possible, we know nothing */ 12156 dst_reg->umin_value = 0; 12157 dst_reg->umax_value = U64_MAX; 12158 } else { 12159 /* Cannot overflow (as long as bounds are consistent) */ 12160 dst_reg->umin_value -= umax_val; 12161 dst_reg->umax_value -= umin_val; 12162 } 12163 } 12164 12165 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg, 12166 struct bpf_reg_state *src_reg) 12167 { 12168 s32 smin_val = src_reg->s32_min_value; 12169 u32 umin_val = src_reg->u32_min_value; 12170 u32 umax_val = src_reg->u32_max_value; 12171 12172 if (smin_val < 0 || dst_reg->s32_min_value < 0) { 12173 /* Ain't nobody got time to multiply that sign */ 12174 __mark_reg32_unbounded(dst_reg); 12175 return; 12176 } 12177 /* Both values are positive, so we can work with unsigned and 12178 * copy the result to signed (unless it exceeds S32_MAX). 12179 */ 12180 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) { 12181 /* Potential overflow, we know nothing */ 12182 __mark_reg32_unbounded(dst_reg); 12183 return; 12184 } 12185 dst_reg->u32_min_value *= umin_val; 12186 dst_reg->u32_max_value *= umax_val; 12187 if (dst_reg->u32_max_value > S32_MAX) { 12188 /* Overflow possible, we know nothing */ 12189 dst_reg->s32_min_value = S32_MIN; 12190 dst_reg->s32_max_value = S32_MAX; 12191 } else { 12192 dst_reg->s32_min_value = dst_reg->u32_min_value; 12193 dst_reg->s32_max_value = dst_reg->u32_max_value; 12194 } 12195 } 12196 12197 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg, 12198 struct bpf_reg_state *src_reg) 12199 { 12200 s64 smin_val = src_reg->smin_value; 12201 u64 umin_val = src_reg->umin_value; 12202 u64 umax_val = src_reg->umax_value; 12203 12204 if (smin_val < 0 || dst_reg->smin_value < 0) { 12205 /* Ain't nobody got time to multiply that sign */ 12206 __mark_reg64_unbounded(dst_reg); 12207 return; 12208 } 12209 /* Both values are positive, so we can work with unsigned and 12210 * copy the result to signed (unless it exceeds S64_MAX). 12211 */ 12212 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) { 12213 /* Potential overflow, we know nothing */ 12214 __mark_reg64_unbounded(dst_reg); 12215 return; 12216 } 12217 dst_reg->umin_value *= umin_val; 12218 dst_reg->umax_value *= umax_val; 12219 if (dst_reg->umax_value > S64_MAX) { 12220 /* Overflow possible, we know nothing */ 12221 dst_reg->smin_value = S64_MIN; 12222 dst_reg->smax_value = S64_MAX; 12223 } else { 12224 dst_reg->smin_value = dst_reg->umin_value; 12225 dst_reg->smax_value = dst_reg->umax_value; 12226 } 12227 } 12228 12229 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg, 12230 struct bpf_reg_state *src_reg) 12231 { 12232 bool src_known = tnum_subreg_is_const(src_reg->var_off); 12233 bool dst_known = tnum_subreg_is_const(dst_reg->var_off); 12234 struct tnum var32_off = tnum_subreg(dst_reg->var_off); 12235 s32 smin_val = src_reg->s32_min_value; 12236 u32 umax_val = src_reg->u32_max_value; 12237 12238 if (src_known && dst_known) { 12239 __mark_reg32_known(dst_reg, var32_off.value); 12240 return; 12241 } 12242 12243 /* We get our minimum from the var_off, since that's inherently 12244 * bitwise. Our maximum is the minimum of the operands' maxima. 12245 */ 12246 dst_reg->u32_min_value = var32_off.value; 12247 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val); 12248 if (dst_reg->s32_min_value < 0 || smin_val < 0) { 12249 /* Lose signed bounds when ANDing negative numbers, 12250 * ain't nobody got time for that. 12251 */ 12252 dst_reg->s32_min_value = S32_MIN; 12253 dst_reg->s32_max_value = S32_MAX; 12254 } else { 12255 /* ANDing two positives gives a positive, so safe to 12256 * cast result into s64. 12257 */ 12258 dst_reg->s32_min_value = dst_reg->u32_min_value; 12259 dst_reg->s32_max_value = dst_reg->u32_max_value; 12260 } 12261 } 12262 12263 static void scalar_min_max_and(struct bpf_reg_state *dst_reg, 12264 struct bpf_reg_state *src_reg) 12265 { 12266 bool src_known = tnum_is_const(src_reg->var_off); 12267 bool dst_known = tnum_is_const(dst_reg->var_off); 12268 s64 smin_val = src_reg->smin_value; 12269 u64 umax_val = src_reg->umax_value; 12270 12271 if (src_known && dst_known) { 12272 __mark_reg_known(dst_reg, dst_reg->var_off.value); 12273 return; 12274 } 12275 12276 /* We get our minimum from the var_off, since that's inherently 12277 * bitwise. Our maximum is the minimum of the operands' maxima. 12278 */ 12279 dst_reg->umin_value = dst_reg->var_off.value; 12280 dst_reg->umax_value = min(dst_reg->umax_value, umax_val); 12281 if (dst_reg->smin_value < 0 || smin_val < 0) { 12282 /* Lose signed bounds when ANDing negative numbers, 12283 * ain't nobody got time for that. 12284 */ 12285 dst_reg->smin_value = S64_MIN; 12286 dst_reg->smax_value = S64_MAX; 12287 } else { 12288 /* ANDing two positives gives a positive, so safe to 12289 * cast result into s64. 12290 */ 12291 dst_reg->smin_value = dst_reg->umin_value; 12292 dst_reg->smax_value = dst_reg->umax_value; 12293 } 12294 /* We may learn something more from the var_off */ 12295 __update_reg_bounds(dst_reg); 12296 } 12297 12298 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg, 12299 struct bpf_reg_state *src_reg) 12300 { 12301 bool src_known = tnum_subreg_is_const(src_reg->var_off); 12302 bool dst_known = tnum_subreg_is_const(dst_reg->var_off); 12303 struct tnum var32_off = tnum_subreg(dst_reg->var_off); 12304 s32 smin_val = src_reg->s32_min_value; 12305 u32 umin_val = src_reg->u32_min_value; 12306 12307 if (src_known && dst_known) { 12308 __mark_reg32_known(dst_reg, var32_off.value); 12309 return; 12310 } 12311 12312 /* We get our maximum from the var_off, and our minimum is the 12313 * maximum of the operands' minima 12314 */ 12315 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val); 12316 dst_reg->u32_max_value = var32_off.value | var32_off.mask; 12317 if (dst_reg->s32_min_value < 0 || smin_val < 0) { 12318 /* Lose signed bounds when ORing negative numbers, 12319 * ain't nobody got time for that. 12320 */ 12321 dst_reg->s32_min_value = S32_MIN; 12322 dst_reg->s32_max_value = S32_MAX; 12323 } else { 12324 /* ORing two positives gives a positive, so safe to 12325 * cast result into s64. 12326 */ 12327 dst_reg->s32_min_value = dst_reg->u32_min_value; 12328 dst_reg->s32_max_value = dst_reg->u32_max_value; 12329 } 12330 } 12331 12332 static void scalar_min_max_or(struct bpf_reg_state *dst_reg, 12333 struct bpf_reg_state *src_reg) 12334 { 12335 bool src_known = tnum_is_const(src_reg->var_off); 12336 bool dst_known = tnum_is_const(dst_reg->var_off); 12337 s64 smin_val = src_reg->smin_value; 12338 u64 umin_val = src_reg->umin_value; 12339 12340 if (src_known && dst_known) { 12341 __mark_reg_known(dst_reg, dst_reg->var_off.value); 12342 return; 12343 } 12344 12345 /* We get our maximum from the var_off, and our minimum is the 12346 * maximum of the operands' minima 12347 */ 12348 dst_reg->umin_value = max(dst_reg->umin_value, umin_val); 12349 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask; 12350 if (dst_reg->smin_value < 0 || smin_val < 0) { 12351 /* Lose signed bounds when ORing negative numbers, 12352 * ain't nobody got time for that. 12353 */ 12354 dst_reg->smin_value = S64_MIN; 12355 dst_reg->smax_value = S64_MAX; 12356 } else { 12357 /* ORing two positives gives a positive, so safe to 12358 * cast result into s64. 12359 */ 12360 dst_reg->smin_value = dst_reg->umin_value; 12361 dst_reg->smax_value = dst_reg->umax_value; 12362 } 12363 /* We may learn something more from the var_off */ 12364 __update_reg_bounds(dst_reg); 12365 } 12366 12367 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg, 12368 struct bpf_reg_state *src_reg) 12369 { 12370 bool src_known = tnum_subreg_is_const(src_reg->var_off); 12371 bool dst_known = tnum_subreg_is_const(dst_reg->var_off); 12372 struct tnum var32_off = tnum_subreg(dst_reg->var_off); 12373 s32 smin_val = src_reg->s32_min_value; 12374 12375 if (src_known && dst_known) { 12376 __mark_reg32_known(dst_reg, var32_off.value); 12377 return; 12378 } 12379 12380 /* We get both minimum and maximum from the var32_off. */ 12381 dst_reg->u32_min_value = var32_off.value; 12382 dst_reg->u32_max_value = var32_off.value | var32_off.mask; 12383 12384 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) { 12385 /* XORing two positive sign numbers gives a positive, 12386 * so safe to cast u32 result into s32. 12387 */ 12388 dst_reg->s32_min_value = dst_reg->u32_min_value; 12389 dst_reg->s32_max_value = dst_reg->u32_max_value; 12390 } else { 12391 dst_reg->s32_min_value = S32_MIN; 12392 dst_reg->s32_max_value = S32_MAX; 12393 } 12394 } 12395 12396 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg, 12397 struct bpf_reg_state *src_reg) 12398 { 12399 bool src_known = tnum_is_const(src_reg->var_off); 12400 bool dst_known = tnum_is_const(dst_reg->var_off); 12401 s64 smin_val = src_reg->smin_value; 12402 12403 if (src_known && dst_known) { 12404 /* dst_reg->var_off.value has been updated earlier */ 12405 __mark_reg_known(dst_reg, dst_reg->var_off.value); 12406 return; 12407 } 12408 12409 /* We get both minimum and maximum from the var_off. */ 12410 dst_reg->umin_value = dst_reg->var_off.value; 12411 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask; 12412 12413 if (dst_reg->smin_value >= 0 && smin_val >= 0) { 12414 /* XORing two positive sign numbers gives a positive, 12415 * so safe to cast u64 result into s64. 12416 */ 12417 dst_reg->smin_value = dst_reg->umin_value; 12418 dst_reg->smax_value = dst_reg->umax_value; 12419 } else { 12420 dst_reg->smin_value = S64_MIN; 12421 dst_reg->smax_value = S64_MAX; 12422 } 12423 12424 __update_reg_bounds(dst_reg); 12425 } 12426 12427 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg, 12428 u64 umin_val, u64 umax_val) 12429 { 12430 /* We lose all sign bit information (except what we can pick 12431 * up from var_off) 12432 */ 12433 dst_reg->s32_min_value = S32_MIN; 12434 dst_reg->s32_max_value = S32_MAX; 12435 /* If we might shift our top bit out, then we know nothing */ 12436 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) { 12437 dst_reg->u32_min_value = 0; 12438 dst_reg->u32_max_value = U32_MAX; 12439 } else { 12440 dst_reg->u32_min_value <<= umin_val; 12441 dst_reg->u32_max_value <<= umax_val; 12442 } 12443 } 12444 12445 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg, 12446 struct bpf_reg_state *src_reg) 12447 { 12448 u32 umax_val = src_reg->u32_max_value; 12449 u32 umin_val = src_reg->u32_min_value; 12450 /* u32 alu operation will zext upper bits */ 12451 struct tnum subreg = tnum_subreg(dst_reg->var_off); 12452 12453 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val); 12454 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val)); 12455 /* Not required but being careful mark reg64 bounds as unknown so 12456 * that we are forced to pick them up from tnum and zext later and 12457 * if some path skips this step we are still safe. 12458 */ 12459 __mark_reg64_unbounded(dst_reg); 12460 __update_reg32_bounds(dst_reg); 12461 } 12462 12463 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg, 12464 u64 umin_val, u64 umax_val) 12465 { 12466 /* Special case <<32 because it is a common compiler pattern to sign 12467 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are 12468 * positive we know this shift will also be positive so we can track 12469 * bounds correctly. Otherwise we lose all sign bit information except 12470 * what we can pick up from var_off. Perhaps we can generalize this 12471 * later to shifts of any length. 12472 */ 12473 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0) 12474 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32; 12475 else 12476 dst_reg->smax_value = S64_MAX; 12477 12478 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0) 12479 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32; 12480 else 12481 dst_reg->smin_value = S64_MIN; 12482 12483 /* If we might shift our top bit out, then we know nothing */ 12484 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) { 12485 dst_reg->umin_value = 0; 12486 dst_reg->umax_value = U64_MAX; 12487 } else { 12488 dst_reg->umin_value <<= umin_val; 12489 dst_reg->umax_value <<= umax_val; 12490 } 12491 } 12492 12493 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg, 12494 struct bpf_reg_state *src_reg) 12495 { 12496 u64 umax_val = src_reg->umax_value; 12497 u64 umin_val = src_reg->umin_value; 12498 12499 /* scalar64 calc uses 32bit unshifted bounds so must be called first */ 12500 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val); 12501 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val); 12502 12503 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val); 12504 /* We may learn something more from the var_off */ 12505 __update_reg_bounds(dst_reg); 12506 } 12507 12508 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg, 12509 struct bpf_reg_state *src_reg) 12510 { 12511 struct tnum subreg = tnum_subreg(dst_reg->var_off); 12512 u32 umax_val = src_reg->u32_max_value; 12513 u32 umin_val = src_reg->u32_min_value; 12514 12515 /* BPF_RSH is an unsigned shift. If the value in dst_reg might 12516 * be negative, then either: 12517 * 1) src_reg might be zero, so the sign bit of the result is 12518 * unknown, so we lose our signed bounds 12519 * 2) it's known negative, thus the unsigned bounds capture the 12520 * signed bounds 12521 * 3) the signed bounds cross zero, so they tell us nothing 12522 * about the result 12523 * If the value in dst_reg is known nonnegative, then again the 12524 * unsigned bounds capture the signed bounds. 12525 * Thus, in all cases it suffices to blow away our signed bounds 12526 * and rely on inferring new ones from the unsigned bounds and 12527 * var_off of the result. 12528 */ 12529 dst_reg->s32_min_value = S32_MIN; 12530 dst_reg->s32_max_value = S32_MAX; 12531 12532 dst_reg->var_off = tnum_rshift(subreg, umin_val); 12533 dst_reg->u32_min_value >>= umax_val; 12534 dst_reg->u32_max_value >>= umin_val; 12535 12536 __mark_reg64_unbounded(dst_reg); 12537 __update_reg32_bounds(dst_reg); 12538 } 12539 12540 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg, 12541 struct bpf_reg_state *src_reg) 12542 { 12543 u64 umax_val = src_reg->umax_value; 12544 u64 umin_val = src_reg->umin_value; 12545 12546 /* BPF_RSH is an unsigned shift. If the value in dst_reg might 12547 * be negative, then either: 12548 * 1) src_reg might be zero, so the sign bit of the result is 12549 * unknown, so we lose our signed bounds 12550 * 2) it's known negative, thus the unsigned bounds capture the 12551 * signed bounds 12552 * 3) the signed bounds cross zero, so they tell us nothing 12553 * about the result 12554 * If the value in dst_reg is known nonnegative, then again the 12555 * unsigned bounds capture the signed bounds. 12556 * Thus, in all cases it suffices to blow away our signed bounds 12557 * and rely on inferring new ones from the unsigned bounds and 12558 * var_off of the result. 12559 */ 12560 dst_reg->smin_value = S64_MIN; 12561 dst_reg->smax_value = S64_MAX; 12562 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val); 12563 dst_reg->umin_value >>= umax_val; 12564 dst_reg->umax_value >>= umin_val; 12565 12566 /* Its not easy to operate on alu32 bounds here because it depends 12567 * on bits being shifted in. Take easy way out and mark unbounded 12568 * so we can recalculate later from tnum. 12569 */ 12570 __mark_reg32_unbounded(dst_reg); 12571 __update_reg_bounds(dst_reg); 12572 } 12573 12574 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg, 12575 struct bpf_reg_state *src_reg) 12576 { 12577 u64 umin_val = src_reg->u32_min_value; 12578 12579 /* Upon reaching here, src_known is true and 12580 * umax_val is equal to umin_val. 12581 */ 12582 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val); 12583 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val); 12584 12585 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32); 12586 12587 /* blow away the dst_reg umin_value/umax_value and rely on 12588 * dst_reg var_off to refine the result. 12589 */ 12590 dst_reg->u32_min_value = 0; 12591 dst_reg->u32_max_value = U32_MAX; 12592 12593 __mark_reg64_unbounded(dst_reg); 12594 __update_reg32_bounds(dst_reg); 12595 } 12596 12597 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg, 12598 struct bpf_reg_state *src_reg) 12599 { 12600 u64 umin_val = src_reg->umin_value; 12601 12602 /* Upon reaching here, src_known is true and umax_val is equal 12603 * to umin_val. 12604 */ 12605 dst_reg->smin_value >>= umin_val; 12606 dst_reg->smax_value >>= umin_val; 12607 12608 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64); 12609 12610 /* blow away the dst_reg umin_value/umax_value and rely on 12611 * dst_reg var_off to refine the result. 12612 */ 12613 dst_reg->umin_value = 0; 12614 dst_reg->umax_value = U64_MAX; 12615 12616 /* Its not easy to operate on alu32 bounds here because it depends 12617 * on bits being shifted in from upper 32-bits. Take easy way out 12618 * and mark unbounded so we can recalculate later from tnum. 12619 */ 12620 __mark_reg32_unbounded(dst_reg); 12621 __update_reg_bounds(dst_reg); 12622 } 12623 12624 /* WARNING: This function does calculations on 64-bit values, but the actual 12625 * execution may occur on 32-bit values. Therefore, things like bitshifts 12626 * need extra checks in the 32-bit case. 12627 */ 12628 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env, 12629 struct bpf_insn *insn, 12630 struct bpf_reg_state *dst_reg, 12631 struct bpf_reg_state src_reg) 12632 { 12633 struct bpf_reg_state *regs = cur_regs(env); 12634 u8 opcode = BPF_OP(insn->code); 12635 bool src_known; 12636 s64 smin_val, smax_val; 12637 u64 umin_val, umax_val; 12638 s32 s32_min_val, s32_max_val; 12639 u32 u32_min_val, u32_max_val; 12640 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32; 12641 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64); 12642 int ret; 12643 12644 smin_val = src_reg.smin_value; 12645 smax_val = src_reg.smax_value; 12646 umin_val = src_reg.umin_value; 12647 umax_val = src_reg.umax_value; 12648 12649 s32_min_val = src_reg.s32_min_value; 12650 s32_max_val = src_reg.s32_max_value; 12651 u32_min_val = src_reg.u32_min_value; 12652 u32_max_val = src_reg.u32_max_value; 12653 12654 if (alu32) { 12655 src_known = tnum_subreg_is_const(src_reg.var_off); 12656 if ((src_known && 12657 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) || 12658 s32_min_val > s32_max_val || u32_min_val > u32_max_val) { 12659 /* Taint dst register if offset had invalid bounds 12660 * derived from e.g. dead branches. 12661 */ 12662 __mark_reg_unknown(env, dst_reg); 12663 return 0; 12664 } 12665 } else { 12666 src_known = tnum_is_const(src_reg.var_off); 12667 if ((src_known && 12668 (smin_val != smax_val || umin_val != umax_val)) || 12669 smin_val > smax_val || umin_val > umax_val) { 12670 /* Taint dst register if offset had invalid bounds 12671 * derived from e.g. dead branches. 12672 */ 12673 __mark_reg_unknown(env, dst_reg); 12674 return 0; 12675 } 12676 } 12677 12678 if (!src_known && 12679 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) { 12680 __mark_reg_unknown(env, dst_reg); 12681 return 0; 12682 } 12683 12684 if (sanitize_needed(opcode)) { 12685 ret = sanitize_val_alu(env, insn); 12686 if (ret < 0) 12687 return sanitize_err(env, insn, ret, NULL, NULL); 12688 } 12689 12690 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops. 12691 * There are two classes of instructions: The first class we track both 12692 * alu32 and alu64 sign/unsigned bounds independently this provides the 12693 * greatest amount of precision when alu operations are mixed with jmp32 12694 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD, 12695 * and BPF_OR. This is possible because these ops have fairly easy to 12696 * understand and calculate behavior in both 32-bit and 64-bit alu ops. 12697 * See alu32 verifier tests for examples. The second class of 12698 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy 12699 * with regards to tracking sign/unsigned bounds because the bits may 12700 * cross subreg boundaries in the alu64 case. When this happens we mark 12701 * the reg unbounded in the subreg bound space and use the resulting 12702 * tnum to calculate an approximation of the sign/unsigned bounds. 12703 */ 12704 switch (opcode) { 12705 case BPF_ADD: 12706 scalar32_min_max_add(dst_reg, &src_reg); 12707 scalar_min_max_add(dst_reg, &src_reg); 12708 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off); 12709 break; 12710 case BPF_SUB: 12711 scalar32_min_max_sub(dst_reg, &src_reg); 12712 scalar_min_max_sub(dst_reg, &src_reg); 12713 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off); 12714 break; 12715 case BPF_MUL: 12716 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off); 12717 scalar32_min_max_mul(dst_reg, &src_reg); 12718 scalar_min_max_mul(dst_reg, &src_reg); 12719 break; 12720 case BPF_AND: 12721 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off); 12722 scalar32_min_max_and(dst_reg, &src_reg); 12723 scalar_min_max_and(dst_reg, &src_reg); 12724 break; 12725 case BPF_OR: 12726 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off); 12727 scalar32_min_max_or(dst_reg, &src_reg); 12728 scalar_min_max_or(dst_reg, &src_reg); 12729 break; 12730 case BPF_XOR: 12731 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off); 12732 scalar32_min_max_xor(dst_reg, &src_reg); 12733 scalar_min_max_xor(dst_reg, &src_reg); 12734 break; 12735 case BPF_LSH: 12736 if (umax_val >= insn_bitness) { 12737 /* Shifts greater than 31 or 63 are undefined. 12738 * This includes shifts by a negative number. 12739 */ 12740 mark_reg_unknown(env, regs, insn->dst_reg); 12741 break; 12742 } 12743 if (alu32) 12744 scalar32_min_max_lsh(dst_reg, &src_reg); 12745 else 12746 scalar_min_max_lsh(dst_reg, &src_reg); 12747 break; 12748 case BPF_RSH: 12749 if (umax_val >= insn_bitness) { 12750 /* Shifts greater than 31 or 63 are undefined. 12751 * This includes shifts by a negative number. 12752 */ 12753 mark_reg_unknown(env, regs, insn->dst_reg); 12754 break; 12755 } 12756 if (alu32) 12757 scalar32_min_max_rsh(dst_reg, &src_reg); 12758 else 12759 scalar_min_max_rsh(dst_reg, &src_reg); 12760 break; 12761 case BPF_ARSH: 12762 if (umax_val >= insn_bitness) { 12763 /* Shifts greater than 31 or 63 are undefined. 12764 * This includes shifts by a negative number. 12765 */ 12766 mark_reg_unknown(env, regs, insn->dst_reg); 12767 break; 12768 } 12769 if (alu32) 12770 scalar32_min_max_arsh(dst_reg, &src_reg); 12771 else 12772 scalar_min_max_arsh(dst_reg, &src_reg); 12773 break; 12774 default: 12775 mark_reg_unknown(env, regs, insn->dst_reg); 12776 break; 12777 } 12778 12779 /* ALU32 ops are zero extended into 64bit register */ 12780 if (alu32) 12781 zext_32_to_64(dst_reg); 12782 reg_bounds_sync(dst_reg); 12783 return 0; 12784 } 12785 12786 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max 12787 * and var_off. 12788 */ 12789 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env, 12790 struct bpf_insn *insn) 12791 { 12792 struct bpf_verifier_state *vstate = env->cur_state; 12793 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 12794 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg; 12795 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0}; 12796 u8 opcode = BPF_OP(insn->code); 12797 int err; 12798 12799 dst_reg = ®s[insn->dst_reg]; 12800 src_reg = NULL; 12801 if (dst_reg->type != SCALAR_VALUE) 12802 ptr_reg = dst_reg; 12803 else 12804 /* Make sure ID is cleared otherwise dst_reg min/max could be 12805 * incorrectly propagated into other registers by find_equal_scalars() 12806 */ 12807 dst_reg->id = 0; 12808 if (BPF_SRC(insn->code) == BPF_X) { 12809 src_reg = ®s[insn->src_reg]; 12810 if (src_reg->type != SCALAR_VALUE) { 12811 if (dst_reg->type != SCALAR_VALUE) { 12812 /* Combining two pointers by any ALU op yields 12813 * an arbitrary scalar. Disallow all math except 12814 * pointer subtraction 12815 */ 12816 if (opcode == BPF_SUB && env->allow_ptr_leaks) { 12817 mark_reg_unknown(env, regs, insn->dst_reg); 12818 return 0; 12819 } 12820 verbose(env, "R%d pointer %s pointer prohibited\n", 12821 insn->dst_reg, 12822 bpf_alu_string[opcode >> 4]); 12823 return -EACCES; 12824 } else { 12825 /* scalar += pointer 12826 * This is legal, but we have to reverse our 12827 * src/dest handling in computing the range 12828 */ 12829 err = mark_chain_precision(env, insn->dst_reg); 12830 if (err) 12831 return err; 12832 return adjust_ptr_min_max_vals(env, insn, 12833 src_reg, dst_reg); 12834 } 12835 } else if (ptr_reg) { 12836 /* pointer += scalar */ 12837 err = mark_chain_precision(env, insn->src_reg); 12838 if (err) 12839 return err; 12840 return adjust_ptr_min_max_vals(env, insn, 12841 dst_reg, src_reg); 12842 } else if (dst_reg->precise) { 12843 /* if dst_reg is precise, src_reg should be precise as well */ 12844 err = mark_chain_precision(env, insn->src_reg); 12845 if (err) 12846 return err; 12847 } 12848 } else { 12849 /* Pretend the src is a reg with a known value, since we only 12850 * need to be able to read from this state. 12851 */ 12852 off_reg.type = SCALAR_VALUE; 12853 __mark_reg_known(&off_reg, insn->imm); 12854 src_reg = &off_reg; 12855 if (ptr_reg) /* pointer += K */ 12856 return adjust_ptr_min_max_vals(env, insn, 12857 ptr_reg, src_reg); 12858 } 12859 12860 /* Got here implies adding two SCALAR_VALUEs */ 12861 if (WARN_ON_ONCE(ptr_reg)) { 12862 print_verifier_state(env, state, true); 12863 verbose(env, "verifier internal error: unexpected ptr_reg\n"); 12864 return -EINVAL; 12865 } 12866 if (WARN_ON(!src_reg)) { 12867 print_verifier_state(env, state, true); 12868 verbose(env, "verifier internal error: no src_reg\n"); 12869 return -EINVAL; 12870 } 12871 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg); 12872 } 12873 12874 /* check validity of 32-bit and 64-bit arithmetic operations */ 12875 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn) 12876 { 12877 struct bpf_reg_state *regs = cur_regs(env); 12878 u8 opcode = BPF_OP(insn->code); 12879 int err; 12880 12881 if (opcode == BPF_END || opcode == BPF_NEG) { 12882 if (opcode == BPF_NEG) { 12883 if (BPF_SRC(insn->code) != BPF_K || 12884 insn->src_reg != BPF_REG_0 || 12885 insn->off != 0 || insn->imm != 0) { 12886 verbose(env, "BPF_NEG uses reserved fields\n"); 12887 return -EINVAL; 12888 } 12889 } else { 12890 if (insn->src_reg != BPF_REG_0 || insn->off != 0 || 12891 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) || 12892 BPF_CLASS(insn->code) == BPF_ALU64) { 12893 verbose(env, "BPF_END uses reserved fields\n"); 12894 return -EINVAL; 12895 } 12896 } 12897 12898 /* check src operand */ 12899 err = check_reg_arg(env, insn->dst_reg, SRC_OP); 12900 if (err) 12901 return err; 12902 12903 if (is_pointer_value(env, insn->dst_reg)) { 12904 verbose(env, "R%d pointer arithmetic prohibited\n", 12905 insn->dst_reg); 12906 return -EACCES; 12907 } 12908 12909 /* check dest operand */ 12910 err = check_reg_arg(env, insn->dst_reg, DST_OP); 12911 if (err) 12912 return err; 12913 12914 } else if (opcode == BPF_MOV) { 12915 12916 if (BPF_SRC(insn->code) == BPF_X) { 12917 if (insn->imm != 0 || insn->off != 0) { 12918 verbose(env, "BPF_MOV uses reserved fields\n"); 12919 return -EINVAL; 12920 } 12921 12922 /* check src operand */ 12923 err = check_reg_arg(env, insn->src_reg, SRC_OP); 12924 if (err) 12925 return err; 12926 } else { 12927 if (insn->src_reg != BPF_REG_0 || insn->off != 0) { 12928 verbose(env, "BPF_MOV uses reserved fields\n"); 12929 return -EINVAL; 12930 } 12931 } 12932 12933 /* check dest operand, mark as required later */ 12934 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK); 12935 if (err) 12936 return err; 12937 12938 if (BPF_SRC(insn->code) == BPF_X) { 12939 struct bpf_reg_state *src_reg = regs + insn->src_reg; 12940 struct bpf_reg_state *dst_reg = regs + insn->dst_reg; 12941 bool need_id = src_reg->type == SCALAR_VALUE && !src_reg->id && 12942 !tnum_is_const(src_reg->var_off); 12943 12944 if (BPF_CLASS(insn->code) == BPF_ALU64) { 12945 /* case: R1 = R2 12946 * copy register state to dest reg 12947 */ 12948 if (need_id) 12949 /* Assign src and dst registers the same ID 12950 * that will be used by find_equal_scalars() 12951 * to propagate min/max range. 12952 */ 12953 src_reg->id = ++env->id_gen; 12954 copy_register_state(dst_reg, src_reg); 12955 dst_reg->live |= REG_LIVE_WRITTEN; 12956 dst_reg->subreg_def = DEF_NOT_SUBREG; 12957 } else { 12958 /* R1 = (u32) R2 */ 12959 if (is_pointer_value(env, insn->src_reg)) { 12960 verbose(env, 12961 "R%d partial copy of pointer\n", 12962 insn->src_reg); 12963 return -EACCES; 12964 } else if (src_reg->type == SCALAR_VALUE) { 12965 bool is_src_reg_u32 = src_reg->umax_value <= U32_MAX; 12966 12967 if (is_src_reg_u32 && need_id) 12968 src_reg->id = ++env->id_gen; 12969 copy_register_state(dst_reg, src_reg); 12970 /* Make sure ID is cleared if src_reg is not in u32 range otherwise 12971 * dst_reg min/max could be incorrectly 12972 * propagated into src_reg by find_equal_scalars() 12973 */ 12974 if (!is_src_reg_u32) 12975 dst_reg->id = 0; 12976 dst_reg->live |= REG_LIVE_WRITTEN; 12977 dst_reg->subreg_def = env->insn_idx + 1; 12978 } else { 12979 mark_reg_unknown(env, regs, 12980 insn->dst_reg); 12981 } 12982 zext_32_to_64(dst_reg); 12983 reg_bounds_sync(dst_reg); 12984 } 12985 } else { 12986 /* case: R = imm 12987 * remember the value we stored into this reg 12988 */ 12989 /* clear any state __mark_reg_known doesn't set */ 12990 mark_reg_unknown(env, regs, insn->dst_reg); 12991 regs[insn->dst_reg].type = SCALAR_VALUE; 12992 if (BPF_CLASS(insn->code) == BPF_ALU64) { 12993 __mark_reg_known(regs + insn->dst_reg, 12994 insn->imm); 12995 } else { 12996 __mark_reg_known(regs + insn->dst_reg, 12997 (u32)insn->imm); 12998 } 12999 } 13000 13001 } else if (opcode > BPF_END) { 13002 verbose(env, "invalid BPF_ALU opcode %x\n", opcode); 13003 return -EINVAL; 13004 13005 } else { /* all other ALU ops: and, sub, xor, add, ... */ 13006 13007 if (BPF_SRC(insn->code) == BPF_X) { 13008 if (insn->imm != 0 || insn->off != 0) { 13009 verbose(env, "BPF_ALU uses reserved fields\n"); 13010 return -EINVAL; 13011 } 13012 /* check src1 operand */ 13013 err = check_reg_arg(env, insn->src_reg, SRC_OP); 13014 if (err) 13015 return err; 13016 } else { 13017 if (insn->src_reg != BPF_REG_0 || insn->off != 0) { 13018 verbose(env, "BPF_ALU uses reserved fields\n"); 13019 return -EINVAL; 13020 } 13021 } 13022 13023 /* check src2 operand */ 13024 err = check_reg_arg(env, insn->dst_reg, SRC_OP); 13025 if (err) 13026 return err; 13027 13028 if ((opcode == BPF_MOD || opcode == BPF_DIV) && 13029 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) { 13030 verbose(env, "div by zero\n"); 13031 return -EINVAL; 13032 } 13033 13034 if ((opcode == BPF_LSH || opcode == BPF_RSH || 13035 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) { 13036 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32; 13037 13038 if (insn->imm < 0 || insn->imm >= size) { 13039 verbose(env, "invalid shift %d\n", insn->imm); 13040 return -EINVAL; 13041 } 13042 } 13043 13044 /* check dest operand */ 13045 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK); 13046 if (err) 13047 return err; 13048 13049 return adjust_reg_min_max_vals(env, insn); 13050 } 13051 13052 return 0; 13053 } 13054 13055 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate, 13056 struct bpf_reg_state *dst_reg, 13057 enum bpf_reg_type type, 13058 bool range_right_open) 13059 { 13060 struct bpf_func_state *state; 13061 struct bpf_reg_state *reg; 13062 int new_range; 13063 13064 if (dst_reg->off < 0 || 13065 (dst_reg->off == 0 && range_right_open)) 13066 /* This doesn't give us any range */ 13067 return; 13068 13069 if (dst_reg->umax_value > MAX_PACKET_OFF || 13070 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF) 13071 /* Risk of overflow. For instance, ptr + (1<<63) may be less 13072 * than pkt_end, but that's because it's also less than pkt. 13073 */ 13074 return; 13075 13076 new_range = dst_reg->off; 13077 if (range_right_open) 13078 new_range++; 13079 13080 /* Examples for register markings: 13081 * 13082 * pkt_data in dst register: 13083 * 13084 * r2 = r3; 13085 * r2 += 8; 13086 * if (r2 > pkt_end) goto <handle exception> 13087 * <access okay> 13088 * 13089 * r2 = r3; 13090 * r2 += 8; 13091 * if (r2 < pkt_end) goto <access okay> 13092 * <handle exception> 13093 * 13094 * Where: 13095 * r2 == dst_reg, pkt_end == src_reg 13096 * r2=pkt(id=n,off=8,r=0) 13097 * r3=pkt(id=n,off=0,r=0) 13098 * 13099 * pkt_data in src register: 13100 * 13101 * r2 = r3; 13102 * r2 += 8; 13103 * if (pkt_end >= r2) goto <access okay> 13104 * <handle exception> 13105 * 13106 * r2 = r3; 13107 * r2 += 8; 13108 * if (pkt_end <= r2) goto <handle exception> 13109 * <access okay> 13110 * 13111 * Where: 13112 * pkt_end == dst_reg, r2 == src_reg 13113 * r2=pkt(id=n,off=8,r=0) 13114 * r3=pkt(id=n,off=0,r=0) 13115 * 13116 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8) 13117 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8) 13118 * and [r3, r3 + 8-1) respectively is safe to access depending on 13119 * the check. 13120 */ 13121 13122 /* If our ids match, then we must have the same max_value. And we 13123 * don't care about the other reg's fixed offset, since if it's too big 13124 * the range won't allow anything. 13125 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16. 13126 */ 13127 bpf_for_each_reg_in_vstate(vstate, state, reg, ({ 13128 if (reg->type == type && reg->id == dst_reg->id) 13129 /* keep the maximum range already checked */ 13130 reg->range = max(reg->range, new_range); 13131 })); 13132 } 13133 13134 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode) 13135 { 13136 struct tnum subreg = tnum_subreg(reg->var_off); 13137 s32 sval = (s32)val; 13138 13139 switch (opcode) { 13140 case BPF_JEQ: 13141 if (tnum_is_const(subreg)) 13142 return !!tnum_equals_const(subreg, val); 13143 else if (val < reg->u32_min_value || val > reg->u32_max_value) 13144 return 0; 13145 break; 13146 case BPF_JNE: 13147 if (tnum_is_const(subreg)) 13148 return !tnum_equals_const(subreg, val); 13149 else if (val < reg->u32_min_value || val > reg->u32_max_value) 13150 return 1; 13151 break; 13152 case BPF_JSET: 13153 if ((~subreg.mask & subreg.value) & val) 13154 return 1; 13155 if (!((subreg.mask | subreg.value) & val)) 13156 return 0; 13157 break; 13158 case BPF_JGT: 13159 if (reg->u32_min_value > val) 13160 return 1; 13161 else if (reg->u32_max_value <= val) 13162 return 0; 13163 break; 13164 case BPF_JSGT: 13165 if (reg->s32_min_value > sval) 13166 return 1; 13167 else if (reg->s32_max_value <= sval) 13168 return 0; 13169 break; 13170 case BPF_JLT: 13171 if (reg->u32_max_value < val) 13172 return 1; 13173 else if (reg->u32_min_value >= val) 13174 return 0; 13175 break; 13176 case BPF_JSLT: 13177 if (reg->s32_max_value < sval) 13178 return 1; 13179 else if (reg->s32_min_value >= sval) 13180 return 0; 13181 break; 13182 case BPF_JGE: 13183 if (reg->u32_min_value >= val) 13184 return 1; 13185 else if (reg->u32_max_value < val) 13186 return 0; 13187 break; 13188 case BPF_JSGE: 13189 if (reg->s32_min_value >= sval) 13190 return 1; 13191 else if (reg->s32_max_value < sval) 13192 return 0; 13193 break; 13194 case BPF_JLE: 13195 if (reg->u32_max_value <= val) 13196 return 1; 13197 else if (reg->u32_min_value > val) 13198 return 0; 13199 break; 13200 case BPF_JSLE: 13201 if (reg->s32_max_value <= sval) 13202 return 1; 13203 else if (reg->s32_min_value > sval) 13204 return 0; 13205 break; 13206 } 13207 13208 return -1; 13209 } 13210 13211 13212 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode) 13213 { 13214 s64 sval = (s64)val; 13215 13216 switch (opcode) { 13217 case BPF_JEQ: 13218 if (tnum_is_const(reg->var_off)) 13219 return !!tnum_equals_const(reg->var_off, val); 13220 else if (val < reg->umin_value || val > reg->umax_value) 13221 return 0; 13222 break; 13223 case BPF_JNE: 13224 if (tnum_is_const(reg->var_off)) 13225 return !tnum_equals_const(reg->var_off, val); 13226 else if (val < reg->umin_value || val > reg->umax_value) 13227 return 1; 13228 break; 13229 case BPF_JSET: 13230 if ((~reg->var_off.mask & reg->var_off.value) & val) 13231 return 1; 13232 if (!((reg->var_off.mask | reg->var_off.value) & val)) 13233 return 0; 13234 break; 13235 case BPF_JGT: 13236 if (reg->umin_value > val) 13237 return 1; 13238 else if (reg->umax_value <= val) 13239 return 0; 13240 break; 13241 case BPF_JSGT: 13242 if (reg->smin_value > sval) 13243 return 1; 13244 else if (reg->smax_value <= sval) 13245 return 0; 13246 break; 13247 case BPF_JLT: 13248 if (reg->umax_value < val) 13249 return 1; 13250 else if (reg->umin_value >= val) 13251 return 0; 13252 break; 13253 case BPF_JSLT: 13254 if (reg->smax_value < sval) 13255 return 1; 13256 else if (reg->smin_value >= sval) 13257 return 0; 13258 break; 13259 case BPF_JGE: 13260 if (reg->umin_value >= val) 13261 return 1; 13262 else if (reg->umax_value < val) 13263 return 0; 13264 break; 13265 case BPF_JSGE: 13266 if (reg->smin_value >= sval) 13267 return 1; 13268 else if (reg->smax_value < sval) 13269 return 0; 13270 break; 13271 case BPF_JLE: 13272 if (reg->umax_value <= val) 13273 return 1; 13274 else if (reg->umin_value > val) 13275 return 0; 13276 break; 13277 case BPF_JSLE: 13278 if (reg->smax_value <= sval) 13279 return 1; 13280 else if (reg->smin_value > sval) 13281 return 0; 13282 break; 13283 } 13284 13285 return -1; 13286 } 13287 13288 /* compute branch direction of the expression "if (reg opcode val) goto target;" 13289 * and return: 13290 * 1 - branch will be taken and "goto target" will be executed 13291 * 0 - branch will not be taken and fall-through to next insn 13292 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value 13293 * range [0,10] 13294 */ 13295 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode, 13296 bool is_jmp32) 13297 { 13298 if (__is_pointer_value(false, reg)) { 13299 if (!reg_not_null(reg)) 13300 return -1; 13301 13302 /* If pointer is valid tests against zero will fail so we can 13303 * use this to direct branch taken. 13304 */ 13305 if (val != 0) 13306 return -1; 13307 13308 switch (opcode) { 13309 case BPF_JEQ: 13310 return 0; 13311 case BPF_JNE: 13312 return 1; 13313 default: 13314 return -1; 13315 } 13316 } 13317 13318 if (is_jmp32) 13319 return is_branch32_taken(reg, val, opcode); 13320 return is_branch64_taken(reg, val, opcode); 13321 } 13322 13323 static int flip_opcode(u32 opcode) 13324 { 13325 /* How can we transform "a <op> b" into "b <op> a"? */ 13326 static const u8 opcode_flip[16] = { 13327 /* these stay the same */ 13328 [BPF_JEQ >> 4] = BPF_JEQ, 13329 [BPF_JNE >> 4] = BPF_JNE, 13330 [BPF_JSET >> 4] = BPF_JSET, 13331 /* these swap "lesser" and "greater" (L and G in the opcodes) */ 13332 [BPF_JGE >> 4] = BPF_JLE, 13333 [BPF_JGT >> 4] = BPF_JLT, 13334 [BPF_JLE >> 4] = BPF_JGE, 13335 [BPF_JLT >> 4] = BPF_JGT, 13336 [BPF_JSGE >> 4] = BPF_JSLE, 13337 [BPF_JSGT >> 4] = BPF_JSLT, 13338 [BPF_JSLE >> 4] = BPF_JSGE, 13339 [BPF_JSLT >> 4] = BPF_JSGT 13340 }; 13341 return opcode_flip[opcode >> 4]; 13342 } 13343 13344 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg, 13345 struct bpf_reg_state *src_reg, 13346 u8 opcode) 13347 { 13348 struct bpf_reg_state *pkt; 13349 13350 if (src_reg->type == PTR_TO_PACKET_END) { 13351 pkt = dst_reg; 13352 } else if (dst_reg->type == PTR_TO_PACKET_END) { 13353 pkt = src_reg; 13354 opcode = flip_opcode(opcode); 13355 } else { 13356 return -1; 13357 } 13358 13359 if (pkt->range >= 0) 13360 return -1; 13361 13362 switch (opcode) { 13363 case BPF_JLE: 13364 /* pkt <= pkt_end */ 13365 fallthrough; 13366 case BPF_JGT: 13367 /* pkt > pkt_end */ 13368 if (pkt->range == BEYOND_PKT_END) 13369 /* pkt has at last one extra byte beyond pkt_end */ 13370 return opcode == BPF_JGT; 13371 break; 13372 case BPF_JLT: 13373 /* pkt < pkt_end */ 13374 fallthrough; 13375 case BPF_JGE: 13376 /* pkt >= pkt_end */ 13377 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END) 13378 return opcode == BPF_JGE; 13379 break; 13380 } 13381 return -1; 13382 } 13383 13384 /* Adjusts the register min/max values in the case that the dst_reg is the 13385 * variable register that we are working on, and src_reg is a constant or we're 13386 * simply doing a BPF_K check. 13387 * In JEQ/JNE cases we also adjust the var_off values. 13388 */ 13389 static void reg_set_min_max(struct bpf_reg_state *true_reg, 13390 struct bpf_reg_state *false_reg, 13391 u64 val, u32 val32, 13392 u8 opcode, bool is_jmp32) 13393 { 13394 struct tnum false_32off = tnum_subreg(false_reg->var_off); 13395 struct tnum false_64off = false_reg->var_off; 13396 struct tnum true_32off = tnum_subreg(true_reg->var_off); 13397 struct tnum true_64off = true_reg->var_off; 13398 s64 sval = (s64)val; 13399 s32 sval32 = (s32)val32; 13400 13401 /* If the dst_reg is a pointer, we can't learn anything about its 13402 * variable offset from the compare (unless src_reg were a pointer into 13403 * the same object, but we don't bother with that. 13404 * Since false_reg and true_reg have the same type by construction, we 13405 * only need to check one of them for pointerness. 13406 */ 13407 if (__is_pointer_value(false, false_reg)) 13408 return; 13409 13410 switch (opcode) { 13411 /* JEQ/JNE comparison doesn't change the register equivalence. 13412 * 13413 * r1 = r2; 13414 * if (r1 == 42) goto label; 13415 * ... 13416 * label: // here both r1 and r2 are known to be 42. 13417 * 13418 * Hence when marking register as known preserve it's ID. 13419 */ 13420 case BPF_JEQ: 13421 if (is_jmp32) { 13422 __mark_reg32_known(true_reg, val32); 13423 true_32off = tnum_subreg(true_reg->var_off); 13424 } else { 13425 ___mark_reg_known(true_reg, val); 13426 true_64off = true_reg->var_off; 13427 } 13428 break; 13429 case BPF_JNE: 13430 if (is_jmp32) { 13431 __mark_reg32_known(false_reg, val32); 13432 false_32off = tnum_subreg(false_reg->var_off); 13433 } else { 13434 ___mark_reg_known(false_reg, val); 13435 false_64off = false_reg->var_off; 13436 } 13437 break; 13438 case BPF_JSET: 13439 if (is_jmp32) { 13440 false_32off = tnum_and(false_32off, tnum_const(~val32)); 13441 if (is_power_of_2(val32)) 13442 true_32off = tnum_or(true_32off, 13443 tnum_const(val32)); 13444 } else { 13445 false_64off = tnum_and(false_64off, tnum_const(~val)); 13446 if (is_power_of_2(val)) 13447 true_64off = tnum_or(true_64off, 13448 tnum_const(val)); 13449 } 13450 break; 13451 case BPF_JGE: 13452 case BPF_JGT: 13453 { 13454 if (is_jmp32) { 13455 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1; 13456 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32; 13457 13458 false_reg->u32_max_value = min(false_reg->u32_max_value, 13459 false_umax); 13460 true_reg->u32_min_value = max(true_reg->u32_min_value, 13461 true_umin); 13462 } else { 13463 u64 false_umax = opcode == BPF_JGT ? val : val - 1; 13464 u64 true_umin = opcode == BPF_JGT ? val + 1 : val; 13465 13466 false_reg->umax_value = min(false_reg->umax_value, false_umax); 13467 true_reg->umin_value = max(true_reg->umin_value, true_umin); 13468 } 13469 break; 13470 } 13471 case BPF_JSGE: 13472 case BPF_JSGT: 13473 { 13474 if (is_jmp32) { 13475 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1; 13476 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32; 13477 13478 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax); 13479 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin); 13480 } else { 13481 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1; 13482 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval; 13483 13484 false_reg->smax_value = min(false_reg->smax_value, false_smax); 13485 true_reg->smin_value = max(true_reg->smin_value, true_smin); 13486 } 13487 break; 13488 } 13489 case BPF_JLE: 13490 case BPF_JLT: 13491 { 13492 if (is_jmp32) { 13493 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1; 13494 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32; 13495 13496 false_reg->u32_min_value = max(false_reg->u32_min_value, 13497 false_umin); 13498 true_reg->u32_max_value = min(true_reg->u32_max_value, 13499 true_umax); 13500 } else { 13501 u64 false_umin = opcode == BPF_JLT ? val : val + 1; 13502 u64 true_umax = opcode == BPF_JLT ? val - 1 : val; 13503 13504 false_reg->umin_value = max(false_reg->umin_value, false_umin); 13505 true_reg->umax_value = min(true_reg->umax_value, true_umax); 13506 } 13507 break; 13508 } 13509 case BPF_JSLE: 13510 case BPF_JSLT: 13511 { 13512 if (is_jmp32) { 13513 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1; 13514 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32; 13515 13516 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin); 13517 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax); 13518 } else { 13519 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1; 13520 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval; 13521 13522 false_reg->smin_value = max(false_reg->smin_value, false_smin); 13523 true_reg->smax_value = min(true_reg->smax_value, true_smax); 13524 } 13525 break; 13526 } 13527 default: 13528 return; 13529 } 13530 13531 if (is_jmp32) { 13532 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off), 13533 tnum_subreg(false_32off)); 13534 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off), 13535 tnum_subreg(true_32off)); 13536 __reg_combine_32_into_64(false_reg); 13537 __reg_combine_32_into_64(true_reg); 13538 } else { 13539 false_reg->var_off = false_64off; 13540 true_reg->var_off = true_64off; 13541 __reg_combine_64_into_32(false_reg); 13542 __reg_combine_64_into_32(true_reg); 13543 } 13544 } 13545 13546 /* Same as above, but for the case that dst_reg holds a constant and src_reg is 13547 * the variable reg. 13548 */ 13549 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg, 13550 struct bpf_reg_state *false_reg, 13551 u64 val, u32 val32, 13552 u8 opcode, bool is_jmp32) 13553 { 13554 opcode = flip_opcode(opcode); 13555 /* This uses zero as "not present in table"; luckily the zero opcode, 13556 * BPF_JA, can't get here. 13557 */ 13558 if (opcode) 13559 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32); 13560 } 13561 13562 /* Regs are known to be equal, so intersect their min/max/var_off */ 13563 static void __reg_combine_min_max(struct bpf_reg_state *src_reg, 13564 struct bpf_reg_state *dst_reg) 13565 { 13566 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value, 13567 dst_reg->umin_value); 13568 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value, 13569 dst_reg->umax_value); 13570 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value, 13571 dst_reg->smin_value); 13572 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value, 13573 dst_reg->smax_value); 13574 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off, 13575 dst_reg->var_off); 13576 reg_bounds_sync(src_reg); 13577 reg_bounds_sync(dst_reg); 13578 } 13579 13580 static void reg_combine_min_max(struct bpf_reg_state *true_src, 13581 struct bpf_reg_state *true_dst, 13582 struct bpf_reg_state *false_src, 13583 struct bpf_reg_state *false_dst, 13584 u8 opcode) 13585 { 13586 switch (opcode) { 13587 case BPF_JEQ: 13588 __reg_combine_min_max(true_src, true_dst); 13589 break; 13590 case BPF_JNE: 13591 __reg_combine_min_max(false_src, false_dst); 13592 break; 13593 } 13594 } 13595 13596 static void mark_ptr_or_null_reg(struct bpf_func_state *state, 13597 struct bpf_reg_state *reg, u32 id, 13598 bool is_null) 13599 { 13600 if (type_may_be_null(reg->type) && reg->id == id && 13601 (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) { 13602 /* Old offset (both fixed and variable parts) should have been 13603 * known-zero, because we don't allow pointer arithmetic on 13604 * pointers that might be NULL. If we see this happening, don't 13605 * convert the register. 13606 * 13607 * But in some cases, some helpers that return local kptrs 13608 * advance offset for the returned pointer. In those cases, it 13609 * is fine to expect to see reg->off. 13610 */ 13611 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0))) 13612 return; 13613 if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) && 13614 WARN_ON_ONCE(reg->off)) 13615 return; 13616 13617 if (is_null) { 13618 reg->type = SCALAR_VALUE; 13619 /* We don't need id and ref_obj_id from this point 13620 * onwards anymore, thus we should better reset it, 13621 * so that state pruning has chances to take effect. 13622 */ 13623 reg->id = 0; 13624 reg->ref_obj_id = 0; 13625 13626 return; 13627 } 13628 13629 mark_ptr_not_null_reg(reg); 13630 13631 if (!reg_may_point_to_spin_lock(reg)) { 13632 /* For not-NULL ptr, reg->ref_obj_id will be reset 13633 * in release_reference(). 13634 * 13635 * reg->id is still used by spin_lock ptr. Other 13636 * than spin_lock ptr type, reg->id can be reset. 13637 */ 13638 reg->id = 0; 13639 } 13640 } 13641 } 13642 13643 /* The logic is similar to find_good_pkt_pointers(), both could eventually 13644 * be folded together at some point. 13645 */ 13646 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno, 13647 bool is_null) 13648 { 13649 struct bpf_func_state *state = vstate->frame[vstate->curframe]; 13650 struct bpf_reg_state *regs = state->regs, *reg; 13651 u32 ref_obj_id = regs[regno].ref_obj_id; 13652 u32 id = regs[regno].id; 13653 13654 if (ref_obj_id && ref_obj_id == id && is_null) 13655 /* regs[regno] is in the " == NULL" branch. 13656 * No one could have freed the reference state before 13657 * doing the NULL check. 13658 */ 13659 WARN_ON_ONCE(release_reference_state(state, id)); 13660 13661 bpf_for_each_reg_in_vstate(vstate, state, reg, ({ 13662 mark_ptr_or_null_reg(state, reg, id, is_null); 13663 })); 13664 } 13665 13666 static bool try_match_pkt_pointers(const struct bpf_insn *insn, 13667 struct bpf_reg_state *dst_reg, 13668 struct bpf_reg_state *src_reg, 13669 struct bpf_verifier_state *this_branch, 13670 struct bpf_verifier_state *other_branch) 13671 { 13672 if (BPF_SRC(insn->code) != BPF_X) 13673 return false; 13674 13675 /* Pointers are always 64-bit. */ 13676 if (BPF_CLASS(insn->code) == BPF_JMP32) 13677 return false; 13678 13679 switch (BPF_OP(insn->code)) { 13680 case BPF_JGT: 13681 if ((dst_reg->type == PTR_TO_PACKET && 13682 src_reg->type == PTR_TO_PACKET_END) || 13683 (dst_reg->type == PTR_TO_PACKET_META && 13684 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { 13685 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */ 13686 find_good_pkt_pointers(this_branch, dst_reg, 13687 dst_reg->type, false); 13688 mark_pkt_end(other_branch, insn->dst_reg, true); 13689 } else if ((dst_reg->type == PTR_TO_PACKET_END && 13690 src_reg->type == PTR_TO_PACKET) || 13691 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && 13692 src_reg->type == PTR_TO_PACKET_META)) { 13693 /* pkt_end > pkt_data', pkt_data > pkt_meta' */ 13694 find_good_pkt_pointers(other_branch, src_reg, 13695 src_reg->type, true); 13696 mark_pkt_end(this_branch, insn->src_reg, false); 13697 } else { 13698 return false; 13699 } 13700 break; 13701 case BPF_JLT: 13702 if ((dst_reg->type == PTR_TO_PACKET && 13703 src_reg->type == PTR_TO_PACKET_END) || 13704 (dst_reg->type == PTR_TO_PACKET_META && 13705 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { 13706 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */ 13707 find_good_pkt_pointers(other_branch, dst_reg, 13708 dst_reg->type, true); 13709 mark_pkt_end(this_branch, insn->dst_reg, false); 13710 } else if ((dst_reg->type == PTR_TO_PACKET_END && 13711 src_reg->type == PTR_TO_PACKET) || 13712 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && 13713 src_reg->type == PTR_TO_PACKET_META)) { 13714 /* pkt_end < pkt_data', pkt_data > pkt_meta' */ 13715 find_good_pkt_pointers(this_branch, src_reg, 13716 src_reg->type, false); 13717 mark_pkt_end(other_branch, insn->src_reg, true); 13718 } else { 13719 return false; 13720 } 13721 break; 13722 case BPF_JGE: 13723 if ((dst_reg->type == PTR_TO_PACKET && 13724 src_reg->type == PTR_TO_PACKET_END) || 13725 (dst_reg->type == PTR_TO_PACKET_META && 13726 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { 13727 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */ 13728 find_good_pkt_pointers(this_branch, dst_reg, 13729 dst_reg->type, true); 13730 mark_pkt_end(other_branch, insn->dst_reg, false); 13731 } else if ((dst_reg->type == PTR_TO_PACKET_END && 13732 src_reg->type == PTR_TO_PACKET) || 13733 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && 13734 src_reg->type == PTR_TO_PACKET_META)) { 13735 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */ 13736 find_good_pkt_pointers(other_branch, src_reg, 13737 src_reg->type, false); 13738 mark_pkt_end(this_branch, insn->src_reg, true); 13739 } else { 13740 return false; 13741 } 13742 break; 13743 case BPF_JLE: 13744 if ((dst_reg->type == PTR_TO_PACKET && 13745 src_reg->type == PTR_TO_PACKET_END) || 13746 (dst_reg->type == PTR_TO_PACKET_META && 13747 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) { 13748 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */ 13749 find_good_pkt_pointers(other_branch, dst_reg, 13750 dst_reg->type, false); 13751 mark_pkt_end(this_branch, insn->dst_reg, true); 13752 } else if ((dst_reg->type == PTR_TO_PACKET_END && 13753 src_reg->type == PTR_TO_PACKET) || 13754 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) && 13755 src_reg->type == PTR_TO_PACKET_META)) { 13756 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */ 13757 find_good_pkt_pointers(this_branch, src_reg, 13758 src_reg->type, true); 13759 mark_pkt_end(other_branch, insn->src_reg, false); 13760 } else { 13761 return false; 13762 } 13763 break; 13764 default: 13765 return false; 13766 } 13767 13768 return true; 13769 } 13770 13771 static void find_equal_scalars(struct bpf_verifier_state *vstate, 13772 struct bpf_reg_state *known_reg) 13773 { 13774 struct bpf_func_state *state; 13775 struct bpf_reg_state *reg; 13776 13777 bpf_for_each_reg_in_vstate(vstate, state, reg, ({ 13778 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id) 13779 copy_register_state(reg, known_reg); 13780 })); 13781 } 13782 13783 static int check_cond_jmp_op(struct bpf_verifier_env *env, 13784 struct bpf_insn *insn, int *insn_idx) 13785 { 13786 struct bpf_verifier_state *this_branch = env->cur_state; 13787 struct bpf_verifier_state *other_branch; 13788 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs; 13789 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL; 13790 struct bpf_reg_state *eq_branch_regs; 13791 u8 opcode = BPF_OP(insn->code); 13792 bool is_jmp32; 13793 int pred = -1; 13794 int err; 13795 13796 /* Only conditional jumps are expected to reach here. */ 13797 if (opcode == BPF_JA || opcode > BPF_JSLE) { 13798 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode); 13799 return -EINVAL; 13800 } 13801 13802 if (BPF_SRC(insn->code) == BPF_X) { 13803 if (insn->imm != 0) { 13804 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n"); 13805 return -EINVAL; 13806 } 13807 13808 /* check src1 operand */ 13809 err = check_reg_arg(env, insn->src_reg, SRC_OP); 13810 if (err) 13811 return err; 13812 13813 if (is_pointer_value(env, insn->src_reg)) { 13814 verbose(env, "R%d pointer comparison prohibited\n", 13815 insn->src_reg); 13816 return -EACCES; 13817 } 13818 src_reg = ®s[insn->src_reg]; 13819 } else { 13820 if (insn->src_reg != BPF_REG_0) { 13821 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n"); 13822 return -EINVAL; 13823 } 13824 } 13825 13826 /* check src2 operand */ 13827 err = check_reg_arg(env, insn->dst_reg, SRC_OP); 13828 if (err) 13829 return err; 13830 13831 dst_reg = ®s[insn->dst_reg]; 13832 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32; 13833 13834 if (BPF_SRC(insn->code) == BPF_K) { 13835 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32); 13836 } else if (src_reg->type == SCALAR_VALUE && 13837 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) { 13838 pred = is_branch_taken(dst_reg, 13839 tnum_subreg(src_reg->var_off).value, 13840 opcode, 13841 is_jmp32); 13842 } else if (src_reg->type == SCALAR_VALUE && 13843 !is_jmp32 && tnum_is_const(src_reg->var_off)) { 13844 pred = is_branch_taken(dst_reg, 13845 src_reg->var_off.value, 13846 opcode, 13847 is_jmp32); 13848 } else if (dst_reg->type == SCALAR_VALUE && 13849 is_jmp32 && tnum_is_const(tnum_subreg(dst_reg->var_off))) { 13850 pred = is_branch_taken(src_reg, 13851 tnum_subreg(dst_reg->var_off).value, 13852 flip_opcode(opcode), 13853 is_jmp32); 13854 } else if (dst_reg->type == SCALAR_VALUE && 13855 !is_jmp32 && tnum_is_const(dst_reg->var_off)) { 13856 pred = is_branch_taken(src_reg, 13857 dst_reg->var_off.value, 13858 flip_opcode(opcode), 13859 is_jmp32); 13860 } else if (reg_is_pkt_pointer_any(dst_reg) && 13861 reg_is_pkt_pointer_any(src_reg) && 13862 !is_jmp32) { 13863 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode); 13864 } 13865 13866 if (pred >= 0) { 13867 /* If we get here with a dst_reg pointer type it is because 13868 * above is_branch_taken() special cased the 0 comparison. 13869 */ 13870 if (!__is_pointer_value(false, dst_reg)) 13871 err = mark_chain_precision(env, insn->dst_reg); 13872 if (BPF_SRC(insn->code) == BPF_X && !err && 13873 !__is_pointer_value(false, src_reg)) 13874 err = mark_chain_precision(env, insn->src_reg); 13875 if (err) 13876 return err; 13877 } 13878 13879 if (pred == 1) { 13880 /* Only follow the goto, ignore fall-through. If needed, push 13881 * the fall-through branch for simulation under speculative 13882 * execution. 13883 */ 13884 if (!env->bypass_spec_v1 && 13885 !sanitize_speculative_path(env, insn, *insn_idx + 1, 13886 *insn_idx)) 13887 return -EFAULT; 13888 *insn_idx += insn->off; 13889 return 0; 13890 } else if (pred == 0) { 13891 /* Only follow the fall-through branch, since that's where the 13892 * program will go. If needed, push the goto branch for 13893 * simulation under speculative execution. 13894 */ 13895 if (!env->bypass_spec_v1 && 13896 !sanitize_speculative_path(env, insn, 13897 *insn_idx + insn->off + 1, 13898 *insn_idx)) 13899 return -EFAULT; 13900 return 0; 13901 } 13902 13903 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx, 13904 false); 13905 if (!other_branch) 13906 return -EFAULT; 13907 other_branch_regs = other_branch->frame[other_branch->curframe]->regs; 13908 13909 /* detect if we are comparing against a constant value so we can adjust 13910 * our min/max values for our dst register. 13911 * this is only legit if both are scalars (or pointers to the same 13912 * object, I suppose, see the PTR_MAYBE_NULL related if block below), 13913 * because otherwise the different base pointers mean the offsets aren't 13914 * comparable. 13915 */ 13916 if (BPF_SRC(insn->code) == BPF_X) { 13917 struct bpf_reg_state *src_reg = ®s[insn->src_reg]; 13918 13919 if (dst_reg->type == SCALAR_VALUE && 13920 src_reg->type == SCALAR_VALUE) { 13921 if (tnum_is_const(src_reg->var_off) || 13922 (is_jmp32 && 13923 tnum_is_const(tnum_subreg(src_reg->var_off)))) 13924 reg_set_min_max(&other_branch_regs[insn->dst_reg], 13925 dst_reg, 13926 src_reg->var_off.value, 13927 tnum_subreg(src_reg->var_off).value, 13928 opcode, is_jmp32); 13929 else if (tnum_is_const(dst_reg->var_off) || 13930 (is_jmp32 && 13931 tnum_is_const(tnum_subreg(dst_reg->var_off)))) 13932 reg_set_min_max_inv(&other_branch_regs[insn->src_reg], 13933 src_reg, 13934 dst_reg->var_off.value, 13935 tnum_subreg(dst_reg->var_off).value, 13936 opcode, is_jmp32); 13937 else if (!is_jmp32 && 13938 (opcode == BPF_JEQ || opcode == BPF_JNE)) 13939 /* Comparing for equality, we can combine knowledge */ 13940 reg_combine_min_max(&other_branch_regs[insn->src_reg], 13941 &other_branch_regs[insn->dst_reg], 13942 src_reg, dst_reg, opcode); 13943 if (src_reg->id && 13944 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) { 13945 find_equal_scalars(this_branch, src_reg); 13946 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]); 13947 } 13948 13949 } 13950 } else if (dst_reg->type == SCALAR_VALUE) { 13951 reg_set_min_max(&other_branch_regs[insn->dst_reg], 13952 dst_reg, insn->imm, (u32)insn->imm, 13953 opcode, is_jmp32); 13954 } 13955 13956 if (dst_reg->type == SCALAR_VALUE && dst_reg->id && 13957 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) { 13958 find_equal_scalars(this_branch, dst_reg); 13959 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]); 13960 } 13961 13962 /* if one pointer register is compared to another pointer 13963 * register check if PTR_MAYBE_NULL could be lifted. 13964 * E.g. register A - maybe null 13965 * register B - not null 13966 * for JNE A, B, ... - A is not null in the false branch; 13967 * for JEQ A, B, ... - A is not null in the true branch. 13968 * 13969 * Since PTR_TO_BTF_ID points to a kernel struct that does 13970 * not need to be null checked by the BPF program, i.e., 13971 * could be null even without PTR_MAYBE_NULL marking, so 13972 * only propagate nullness when neither reg is that type. 13973 */ 13974 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X && 13975 __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) && 13976 type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) && 13977 base_type(src_reg->type) != PTR_TO_BTF_ID && 13978 base_type(dst_reg->type) != PTR_TO_BTF_ID) { 13979 eq_branch_regs = NULL; 13980 switch (opcode) { 13981 case BPF_JEQ: 13982 eq_branch_regs = other_branch_regs; 13983 break; 13984 case BPF_JNE: 13985 eq_branch_regs = regs; 13986 break; 13987 default: 13988 /* do nothing */ 13989 break; 13990 } 13991 if (eq_branch_regs) { 13992 if (type_may_be_null(src_reg->type)) 13993 mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]); 13994 else 13995 mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]); 13996 } 13997 } 13998 13999 /* detect if R == 0 where R is returned from bpf_map_lookup_elem(). 14000 * NOTE: these optimizations below are related with pointer comparison 14001 * which will never be JMP32. 14002 */ 14003 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K && 14004 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) && 14005 type_may_be_null(dst_reg->type)) { 14006 /* Mark all identical registers in each branch as either 14007 * safe or unknown depending R == 0 or R != 0 conditional. 14008 */ 14009 mark_ptr_or_null_regs(this_branch, insn->dst_reg, 14010 opcode == BPF_JNE); 14011 mark_ptr_or_null_regs(other_branch, insn->dst_reg, 14012 opcode == BPF_JEQ); 14013 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg], 14014 this_branch, other_branch) && 14015 is_pointer_value(env, insn->dst_reg)) { 14016 verbose(env, "R%d pointer comparison prohibited\n", 14017 insn->dst_reg); 14018 return -EACCES; 14019 } 14020 if (env->log.level & BPF_LOG_LEVEL) 14021 print_insn_state(env, this_branch->frame[this_branch->curframe]); 14022 return 0; 14023 } 14024 14025 /* verify BPF_LD_IMM64 instruction */ 14026 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn) 14027 { 14028 struct bpf_insn_aux_data *aux = cur_aux(env); 14029 struct bpf_reg_state *regs = cur_regs(env); 14030 struct bpf_reg_state *dst_reg; 14031 struct bpf_map *map; 14032 int err; 14033 14034 if (BPF_SIZE(insn->code) != BPF_DW) { 14035 verbose(env, "invalid BPF_LD_IMM insn\n"); 14036 return -EINVAL; 14037 } 14038 if (insn->off != 0) { 14039 verbose(env, "BPF_LD_IMM64 uses reserved fields\n"); 14040 return -EINVAL; 14041 } 14042 14043 err = check_reg_arg(env, insn->dst_reg, DST_OP); 14044 if (err) 14045 return err; 14046 14047 dst_reg = ®s[insn->dst_reg]; 14048 if (insn->src_reg == 0) { 14049 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm; 14050 14051 dst_reg->type = SCALAR_VALUE; 14052 __mark_reg_known(®s[insn->dst_reg], imm); 14053 return 0; 14054 } 14055 14056 /* All special src_reg cases are listed below. From this point onwards 14057 * we either succeed and assign a corresponding dst_reg->type after 14058 * zeroing the offset, or fail and reject the program. 14059 */ 14060 mark_reg_known_zero(env, regs, insn->dst_reg); 14061 14062 if (insn->src_reg == BPF_PSEUDO_BTF_ID) { 14063 dst_reg->type = aux->btf_var.reg_type; 14064 switch (base_type(dst_reg->type)) { 14065 case PTR_TO_MEM: 14066 dst_reg->mem_size = aux->btf_var.mem_size; 14067 break; 14068 case PTR_TO_BTF_ID: 14069 dst_reg->btf = aux->btf_var.btf; 14070 dst_reg->btf_id = aux->btf_var.btf_id; 14071 break; 14072 default: 14073 verbose(env, "bpf verifier is misconfigured\n"); 14074 return -EFAULT; 14075 } 14076 return 0; 14077 } 14078 14079 if (insn->src_reg == BPF_PSEUDO_FUNC) { 14080 struct bpf_prog_aux *aux = env->prog->aux; 14081 u32 subprogno = find_subprog(env, 14082 env->insn_idx + insn->imm + 1); 14083 14084 if (!aux->func_info) { 14085 verbose(env, "missing btf func_info\n"); 14086 return -EINVAL; 14087 } 14088 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) { 14089 verbose(env, "callback function not static\n"); 14090 return -EINVAL; 14091 } 14092 14093 dst_reg->type = PTR_TO_FUNC; 14094 dst_reg->subprogno = subprogno; 14095 return 0; 14096 } 14097 14098 map = env->used_maps[aux->map_index]; 14099 dst_reg->map_ptr = map; 14100 14101 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE || 14102 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) { 14103 dst_reg->type = PTR_TO_MAP_VALUE; 14104 dst_reg->off = aux->map_off; 14105 WARN_ON_ONCE(map->max_entries != 1); 14106 /* We want reg->id to be same (0) as map_value is not distinct */ 14107 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD || 14108 insn->src_reg == BPF_PSEUDO_MAP_IDX) { 14109 dst_reg->type = CONST_PTR_TO_MAP; 14110 } else { 14111 verbose(env, "bpf verifier is misconfigured\n"); 14112 return -EINVAL; 14113 } 14114 14115 return 0; 14116 } 14117 14118 static bool may_access_skb(enum bpf_prog_type type) 14119 { 14120 switch (type) { 14121 case BPF_PROG_TYPE_SOCKET_FILTER: 14122 case BPF_PROG_TYPE_SCHED_CLS: 14123 case BPF_PROG_TYPE_SCHED_ACT: 14124 return true; 14125 default: 14126 return false; 14127 } 14128 } 14129 14130 /* verify safety of LD_ABS|LD_IND instructions: 14131 * - they can only appear in the programs where ctx == skb 14132 * - since they are wrappers of function calls, they scratch R1-R5 registers, 14133 * preserve R6-R9, and store return value into R0 14134 * 14135 * Implicit input: 14136 * ctx == skb == R6 == CTX 14137 * 14138 * Explicit input: 14139 * SRC == any register 14140 * IMM == 32-bit immediate 14141 * 14142 * Output: 14143 * R0 - 8/16/32-bit skb data converted to cpu endianness 14144 */ 14145 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn) 14146 { 14147 struct bpf_reg_state *regs = cur_regs(env); 14148 static const int ctx_reg = BPF_REG_6; 14149 u8 mode = BPF_MODE(insn->code); 14150 int i, err; 14151 14152 if (!may_access_skb(resolve_prog_type(env->prog))) { 14153 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n"); 14154 return -EINVAL; 14155 } 14156 14157 if (!env->ops->gen_ld_abs) { 14158 verbose(env, "bpf verifier is misconfigured\n"); 14159 return -EINVAL; 14160 } 14161 14162 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 || 14163 BPF_SIZE(insn->code) == BPF_DW || 14164 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) { 14165 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n"); 14166 return -EINVAL; 14167 } 14168 14169 /* check whether implicit source operand (register R6) is readable */ 14170 err = check_reg_arg(env, ctx_reg, SRC_OP); 14171 if (err) 14172 return err; 14173 14174 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as 14175 * gen_ld_abs() may terminate the program at runtime, leading to 14176 * reference leak. 14177 */ 14178 err = check_reference_leak(env); 14179 if (err) { 14180 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n"); 14181 return err; 14182 } 14183 14184 if (env->cur_state->active_lock.ptr) { 14185 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n"); 14186 return -EINVAL; 14187 } 14188 14189 if (env->cur_state->active_rcu_lock) { 14190 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_rcu_read_lock-ed region\n"); 14191 return -EINVAL; 14192 } 14193 14194 if (regs[ctx_reg].type != PTR_TO_CTX) { 14195 verbose(env, 14196 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n"); 14197 return -EINVAL; 14198 } 14199 14200 if (mode == BPF_IND) { 14201 /* check explicit source operand */ 14202 err = check_reg_arg(env, insn->src_reg, SRC_OP); 14203 if (err) 14204 return err; 14205 } 14206 14207 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg); 14208 if (err < 0) 14209 return err; 14210 14211 /* reset caller saved regs to unreadable */ 14212 for (i = 0; i < CALLER_SAVED_REGS; i++) { 14213 mark_reg_not_init(env, regs, caller_saved[i]); 14214 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK); 14215 } 14216 14217 /* mark destination R0 register as readable, since it contains 14218 * the value fetched from the packet. 14219 * Already marked as written above. 14220 */ 14221 mark_reg_unknown(env, regs, BPF_REG_0); 14222 /* ld_abs load up to 32-bit skb data. */ 14223 regs[BPF_REG_0].subreg_def = env->insn_idx + 1; 14224 return 0; 14225 } 14226 14227 static int check_return_code(struct bpf_verifier_env *env) 14228 { 14229 struct tnum enforce_attach_type_range = tnum_unknown; 14230 const struct bpf_prog *prog = env->prog; 14231 struct bpf_reg_state *reg; 14232 struct tnum range = tnum_range(0, 1); 14233 enum bpf_prog_type prog_type = resolve_prog_type(env->prog); 14234 int err; 14235 struct bpf_func_state *frame = env->cur_state->frame[0]; 14236 const bool is_subprog = frame->subprogno; 14237 14238 /* LSM and struct_ops func-ptr's return type could be "void" */ 14239 if (!is_subprog) { 14240 switch (prog_type) { 14241 case BPF_PROG_TYPE_LSM: 14242 if (prog->expected_attach_type == BPF_LSM_CGROUP) 14243 /* See below, can be 0 or 0-1 depending on hook. */ 14244 break; 14245 fallthrough; 14246 case BPF_PROG_TYPE_STRUCT_OPS: 14247 if (!prog->aux->attach_func_proto->type) 14248 return 0; 14249 break; 14250 default: 14251 break; 14252 } 14253 } 14254 14255 /* eBPF calling convention is such that R0 is used 14256 * to return the value from eBPF program. 14257 * Make sure that it's readable at this time 14258 * of bpf_exit, which means that program wrote 14259 * something into it earlier 14260 */ 14261 err = check_reg_arg(env, BPF_REG_0, SRC_OP); 14262 if (err) 14263 return err; 14264 14265 if (is_pointer_value(env, BPF_REG_0)) { 14266 verbose(env, "R0 leaks addr as return value\n"); 14267 return -EACCES; 14268 } 14269 14270 reg = cur_regs(env) + BPF_REG_0; 14271 14272 if (frame->in_async_callback_fn) { 14273 /* enforce return zero from async callbacks like timer */ 14274 if (reg->type != SCALAR_VALUE) { 14275 verbose(env, "In async callback the register R0 is not a known value (%s)\n", 14276 reg_type_str(env, reg->type)); 14277 return -EINVAL; 14278 } 14279 14280 if (!tnum_in(tnum_const(0), reg->var_off)) { 14281 verbose_invalid_scalar(env, reg, &range, "async callback", "R0"); 14282 return -EINVAL; 14283 } 14284 return 0; 14285 } 14286 14287 if (is_subprog) { 14288 if (reg->type != SCALAR_VALUE) { 14289 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n", 14290 reg_type_str(env, reg->type)); 14291 return -EINVAL; 14292 } 14293 return 0; 14294 } 14295 14296 switch (prog_type) { 14297 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR: 14298 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG || 14299 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG || 14300 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME || 14301 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME || 14302 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME || 14303 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME) 14304 range = tnum_range(1, 1); 14305 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND || 14306 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND) 14307 range = tnum_range(0, 3); 14308 break; 14309 case BPF_PROG_TYPE_CGROUP_SKB: 14310 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) { 14311 range = tnum_range(0, 3); 14312 enforce_attach_type_range = tnum_range(2, 3); 14313 } 14314 break; 14315 case BPF_PROG_TYPE_CGROUP_SOCK: 14316 case BPF_PROG_TYPE_SOCK_OPS: 14317 case BPF_PROG_TYPE_CGROUP_DEVICE: 14318 case BPF_PROG_TYPE_CGROUP_SYSCTL: 14319 case BPF_PROG_TYPE_CGROUP_SOCKOPT: 14320 break; 14321 case BPF_PROG_TYPE_RAW_TRACEPOINT: 14322 if (!env->prog->aux->attach_btf_id) 14323 return 0; 14324 range = tnum_const(0); 14325 break; 14326 case BPF_PROG_TYPE_TRACING: 14327 switch (env->prog->expected_attach_type) { 14328 case BPF_TRACE_FENTRY: 14329 case BPF_TRACE_FEXIT: 14330 range = tnum_const(0); 14331 break; 14332 case BPF_TRACE_RAW_TP: 14333 case BPF_MODIFY_RETURN: 14334 return 0; 14335 case BPF_TRACE_ITER: 14336 break; 14337 default: 14338 return -ENOTSUPP; 14339 } 14340 break; 14341 case BPF_PROG_TYPE_SK_LOOKUP: 14342 range = tnum_range(SK_DROP, SK_PASS); 14343 break; 14344 14345 case BPF_PROG_TYPE_LSM: 14346 if (env->prog->expected_attach_type != BPF_LSM_CGROUP) { 14347 /* Regular BPF_PROG_TYPE_LSM programs can return 14348 * any value. 14349 */ 14350 return 0; 14351 } 14352 if (!env->prog->aux->attach_func_proto->type) { 14353 /* Make sure programs that attach to void 14354 * hooks don't try to modify return value. 14355 */ 14356 range = tnum_range(1, 1); 14357 } 14358 break; 14359 14360 case BPF_PROG_TYPE_NETFILTER: 14361 range = tnum_range(NF_DROP, NF_ACCEPT); 14362 break; 14363 case BPF_PROG_TYPE_EXT: 14364 /* freplace program can return anything as its return value 14365 * depends on the to-be-replaced kernel func or bpf program. 14366 */ 14367 default: 14368 return 0; 14369 } 14370 14371 if (reg->type != SCALAR_VALUE) { 14372 verbose(env, "At program exit the register R0 is not a known value (%s)\n", 14373 reg_type_str(env, reg->type)); 14374 return -EINVAL; 14375 } 14376 14377 if (!tnum_in(range, reg->var_off)) { 14378 verbose_invalid_scalar(env, reg, &range, "program exit", "R0"); 14379 if (prog->expected_attach_type == BPF_LSM_CGROUP && 14380 prog_type == BPF_PROG_TYPE_LSM && 14381 !prog->aux->attach_func_proto->type) 14382 verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n"); 14383 return -EINVAL; 14384 } 14385 14386 if (!tnum_is_unknown(enforce_attach_type_range) && 14387 tnum_in(enforce_attach_type_range, reg->var_off)) 14388 env->prog->enforce_expected_attach_type = 1; 14389 return 0; 14390 } 14391 14392 /* non-recursive DFS pseudo code 14393 * 1 procedure DFS-iterative(G,v): 14394 * 2 label v as discovered 14395 * 3 let S be a stack 14396 * 4 S.push(v) 14397 * 5 while S is not empty 14398 * 6 t <- S.peek() 14399 * 7 if t is what we're looking for: 14400 * 8 return t 14401 * 9 for all edges e in G.adjacentEdges(t) do 14402 * 10 if edge e is already labelled 14403 * 11 continue with the next edge 14404 * 12 w <- G.adjacentVertex(t,e) 14405 * 13 if vertex w is not discovered and not explored 14406 * 14 label e as tree-edge 14407 * 15 label w as discovered 14408 * 16 S.push(w) 14409 * 17 continue at 5 14410 * 18 else if vertex w is discovered 14411 * 19 label e as back-edge 14412 * 20 else 14413 * 21 // vertex w is explored 14414 * 22 label e as forward- or cross-edge 14415 * 23 label t as explored 14416 * 24 S.pop() 14417 * 14418 * convention: 14419 * 0x10 - discovered 14420 * 0x11 - discovered and fall-through edge labelled 14421 * 0x12 - discovered and fall-through and branch edges labelled 14422 * 0x20 - explored 14423 */ 14424 14425 enum { 14426 DISCOVERED = 0x10, 14427 EXPLORED = 0x20, 14428 FALLTHROUGH = 1, 14429 BRANCH = 2, 14430 }; 14431 14432 static u32 state_htab_size(struct bpf_verifier_env *env) 14433 { 14434 return env->prog->len; 14435 } 14436 14437 static struct bpf_verifier_state_list **explored_state( 14438 struct bpf_verifier_env *env, 14439 int idx) 14440 { 14441 struct bpf_verifier_state *cur = env->cur_state; 14442 struct bpf_func_state *state = cur->frame[cur->curframe]; 14443 14444 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)]; 14445 } 14446 14447 static void mark_prune_point(struct bpf_verifier_env *env, int idx) 14448 { 14449 env->insn_aux_data[idx].prune_point = true; 14450 } 14451 14452 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx) 14453 { 14454 return env->insn_aux_data[insn_idx].prune_point; 14455 } 14456 14457 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx) 14458 { 14459 env->insn_aux_data[idx].force_checkpoint = true; 14460 } 14461 14462 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx) 14463 { 14464 return env->insn_aux_data[insn_idx].force_checkpoint; 14465 } 14466 14467 14468 enum { 14469 DONE_EXPLORING = 0, 14470 KEEP_EXPLORING = 1, 14471 }; 14472 14473 /* t, w, e - match pseudo-code above: 14474 * t - index of current instruction 14475 * w - next instruction 14476 * e - edge 14477 */ 14478 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env, 14479 bool loop_ok) 14480 { 14481 int *insn_stack = env->cfg.insn_stack; 14482 int *insn_state = env->cfg.insn_state; 14483 14484 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH)) 14485 return DONE_EXPLORING; 14486 14487 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH)) 14488 return DONE_EXPLORING; 14489 14490 if (w < 0 || w >= env->prog->len) { 14491 verbose_linfo(env, t, "%d: ", t); 14492 verbose(env, "jump out of range from insn %d to %d\n", t, w); 14493 return -EINVAL; 14494 } 14495 14496 if (e == BRANCH) { 14497 /* mark branch target for state pruning */ 14498 mark_prune_point(env, w); 14499 mark_jmp_point(env, w); 14500 } 14501 14502 if (insn_state[w] == 0) { 14503 /* tree-edge */ 14504 insn_state[t] = DISCOVERED | e; 14505 insn_state[w] = DISCOVERED; 14506 if (env->cfg.cur_stack >= env->prog->len) 14507 return -E2BIG; 14508 insn_stack[env->cfg.cur_stack++] = w; 14509 return KEEP_EXPLORING; 14510 } else if ((insn_state[w] & 0xF0) == DISCOVERED) { 14511 if (loop_ok && env->bpf_capable) 14512 return DONE_EXPLORING; 14513 verbose_linfo(env, t, "%d: ", t); 14514 verbose_linfo(env, w, "%d: ", w); 14515 verbose(env, "back-edge from insn %d to %d\n", t, w); 14516 return -EINVAL; 14517 } else if (insn_state[w] == EXPLORED) { 14518 /* forward- or cross-edge */ 14519 insn_state[t] = DISCOVERED | e; 14520 } else { 14521 verbose(env, "insn state internal bug\n"); 14522 return -EFAULT; 14523 } 14524 return DONE_EXPLORING; 14525 } 14526 14527 static int visit_func_call_insn(int t, struct bpf_insn *insns, 14528 struct bpf_verifier_env *env, 14529 bool visit_callee) 14530 { 14531 int ret; 14532 14533 ret = push_insn(t, t + 1, FALLTHROUGH, env, false); 14534 if (ret) 14535 return ret; 14536 14537 mark_prune_point(env, t + 1); 14538 /* when we exit from subprog, we need to record non-linear history */ 14539 mark_jmp_point(env, t + 1); 14540 14541 if (visit_callee) { 14542 mark_prune_point(env, t); 14543 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env, 14544 /* It's ok to allow recursion from CFG point of 14545 * view. __check_func_call() will do the actual 14546 * check. 14547 */ 14548 bpf_pseudo_func(insns + t)); 14549 } 14550 return ret; 14551 } 14552 14553 /* Visits the instruction at index t and returns one of the following: 14554 * < 0 - an error occurred 14555 * DONE_EXPLORING - the instruction was fully explored 14556 * KEEP_EXPLORING - there is still work to be done before it is fully explored 14557 */ 14558 static int visit_insn(int t, struct bpf_verifier_env *env) 14559 { 14560 struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t]; 14561 int ret; 14562 14563 if (bpf_pseudo_func(insn)) 14564 return visit_func_call_insn(t, insns, env, true); 14565 14566 /* All non-branch instructions have a single fall-through edge. */ 14567 if (BPF_CLASS(insn->code) != BPF_JMP && 14568 BPF_CLASS(insn->code) != BPF_JMP32) 14569 return push_insn(t, t + 1, FALLTHROUGH, env, false); 14570 14571 switch (BPF_OP(insn->code)) { 14572 case BPF_EXIT: 14573 return DONE_EXPLORING; 14574 14575 case BPF_CALL: 14576 if (insn->src_reg == 0 && insn->imm == BPF_FUNC_timer_set_callback) 14577 /* Mark this call insn as a prune point to trigger 14578 * is_state_visited() check before call itself is 14579 * processed by __check_func_call(). Otherwise new 14580 * async state will be pushed for further exploration. 14581 */ 14582 mark_prune_point(env, t); 14583 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) { 14584 struct bpf_kfunc_call_arg_meta meta; 14585 14586 ret = fetch_kfunc_meta(env, insn, &meta, NULL); 14587 if (ret == 0 && is_iter_next_kfunc(&meta)) { 14588 mark_prune_point(env, t); 14589 /* Checking and saving state checkpoints at iter_next() call 14590 * is crucial for fast convergence of open-coded iterator loop 14591 * logic, so we need to force it. If we don't do that, 14592 * is_state_visited() might skip saving a checkpoint, causing 14593 * unnecessarily long sequence of not checkpointed 14594 * instructions and jumps, leading to exhaustion of jump 14595 * history buffer, and potentially other undesired outcomes. 14596 * It is expected that with correct open-coded iterators 14597 * convergence will happen quickly, so we don't run a risk of 14598 * exhausting memory. 14599 */ 14600 mark_force_checkpoint(env, t); 14601 } 14602 } 14603 return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL); 14604 14605 case BPF_JA: 14606 if (BPF_SRC(insn->code) != BPF_K) 14607 return -EINVAL; 14608 14609 /* unconditional jump with single edge */ 14610 ret = push_insn(t, t + insn->off + 1, FALLTHROUGH, env, 14611 true); 14612 if (ret) 14613 return ret; 14614 14615 mark_prune_point(env, t + insn->off + 1); 14616 mark_jmp_point(env, t + insn->off + 1); 14617 14618 return ret; 14619 14620 default: 14621 /* conditional jump with two edges */ 14622 mark_prune_point(env, t); 14623 14624 ret = push_insn(t, t + 1, FALLTHROUGH, env, true); 14625 if (ret) 14626 return ret; 14627 14628 return push_insn(t, t + insn->off + 1, BRANCH, env, true); 14629 } 14630 } 14631 14632 /* non-recursive depth-first-search to detect loops in BPF program 14633 * loop == back-edge in directed graph 14634 */ 14635 static int check_cfg(struct bpf_verifier_env *env) 14636 { 14637 int insn_cnt = env->prog->len; 14638 int *insn_stack, *insn_state; 14639 int ret = 0; 14640 int i; 14641 14642 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL); 14643 if (!insn_state) 14644 return -ENOMEM; 14645 14646 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL); 14647 if (!insn_stack) { 14648 kvfree(insn_state); 14649 return -ENOMEM; 14650 } 14651 14652 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */ 14653 insn_stack[0] = 0; /* 0 is the first instruction */ 14654 env->cfg.cur_stack = 1; 14655 14656 while (env->cfg.cur_stack > 0) { 14657 int t = insn_stack[env->cfg.cur_stack - 1]; 14658 14659 ret = visit_insn(t, env); 14660 switch (ret) { 14661 case DONE_EXPLORING: 14662 insn_state[t] = EXPLORED; 14663 env->cfg.cur_stack--; 14664 break; 14665 case KEEP_EXPLORING: 14666 break; 14667 default: 14668 if (ret > 0) { 14669 verbose(env, "visit_insn internal bug\n"); 14670 ret = -EFAULT; 14671 } 14672 goto err_free; 14673 } 14674 } 14675 14676 if (env->cfg.cur_stack < 0) { 14677 verbose(env, "pop stack internal bug\n"); 14678 ret = -EFAULT; 14679 goto err_free; 14680 } 14681 14682 for (i = 0; i < insn_cnt; i++) { 14683 if (insn_state[i] != EXPLORED) { 14684 verbose(env, "unreachable insn %d\n", i); 14685 ret = -EINVAL; 14686 goto err_free; 14687 } 14688 } 14689 ret = 0; /* cfg looks good */ 14690 14691 err_free: 14692 kvfree(insn_state); 14693 kvfree(insn_stack); 14694 env->cfg.insn_state = env->cfg.insn_stack = NULL; 14695 return ret; 14696 } 14697 14698 static int check_abnormal_return(struct bpf_verifier_env *env) 14699 { 14700 int i; 14701 14702 for (i = 1; i < env->subprog_cnt; i++) { 14703 if (env->subprog_info[i].has_ld_abs) { 14704 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n"); 14705 return -EINVAL; 14706 } 14707 if (env->subprog_info[i].has_tail_call) { 14708 verbose(env, "tail_call is not allowed in subprogs without BTF\n"); 14709 return -EINVAL; 14710 } 14711 } 14712 return 0; 14713 } 14714 14715 /* The minimum supported BTF func info size */ 14716 #define MIN_BPF_FUNCINFO_SIZE 8 14717 #define MAX_FUNCINFO_REC_SIZE 252 14718 14719 static int check_btf_func(struct bpf_verifier_env *env, 14720 const union bpf_attr *attr, 14721 bpfptr_t uattr) 14722 { 14723 const struct btf_type *type, *func_proto, *ret_type; 14724 u32 i, nfuncs, urec_size, min_size; 14725 u32 krec_size = sizeof(struct bpf_func_info); 14726 struct bpf_func_info *krecord; 14727 struct bpf_func_info_aux *info_aux = NULL; 14728 struct bpf_prog *prog; 14729 const struct btf *btf; 14730 bpfptr_t urecord; 14731 u32 prev_offset = 0; 14732 bool scalar_return; 14733 int ret = -ENOMEM; 14734 14735 nfuncs = attr->func_info_cnt; 14736 if (!nfuncs) { 14737 if (check_abnormal_return(env)) 14738 return -EINVAL; 14739 return 0; 14740 } 14741 14742 if (nfuncs != env->subprog_cnt) { 14743 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n"); 14744 return -EINVAL; 14745 } 14746 14747 urec_size = attr->func_info_rec_size; 14748 if (urec_size < MIN_BPF_FUNCINFO_SIZE || 14749 urec_size > MAX_FUNCINFO_REC_SIZE || 14750 urec_size % sizeof(u32)) { 14751 verbose(env, "invalid func info rec size %u\n", urec_size); 14752 return -EINVAL; 14753 } 14754 14755 prog = env->prog; 14756 btf = prog->aux->btf; 14757 14758 urecord = make_bpfptr(attr->func_info, uattr.is_kernel); 14759 min_size = min_t(u32, krec_size, urec_size); 14760 14761 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN); 14762 if (!krecord) 14763 return -ENOMEM; 14764 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN); 14765 if (!info_aux) 14766 goto err_free; 14767 14768 for (i = 0; i < nfuncs; i++) { 14769 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size); 14770 if (ret) { 14771 if (ret == -E2BIG) { 14772 verbose(env, "nonzero tailing record in func info"); 14773 /* set the size kernel expects so loader can zero 14774 * out the rest of the record. 14775 */ 14776 if (copy_to_bpfptr_offset(uattr, 14777 offsetof(union bpf_attr, func_info_rec_size), 14778 &min_size, sizeof(min_size))) 14779 ret = -EFAULT; 14780 } 14781 goto err_free; 14782 } 14783 14784 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) { 14785 ret = -EFAULT; 14786 goto err_free; 14787 } 14788 14789 /* check insn_off */ 14790 ret = -EINVAL; 14791 if (i == 0) { 14792 if (krecord[i].insn_off) { 14793 verbose(env, 14794 "nonzero insn_off %u for the first func info record", 14795 krecord[i].insn_off); 14796 goto err_free; 14797 } 14798 } else if (krecord[i].insn_off <= prev_offset) { 14799 verbose(env, 14800 "same or smaller insn offset (%u) than previous func info record (%u)", 14801 krecord[i].insn_off, prev_offset); 14802 goto err_free; 14803 } 14804 14805 if (env->subprog_info[i].start != krecord[i].insn_off) { 14806 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n"); 14807 goto err_free; 14808 } 14809 14810 /* check type_id */ 14811 type = btf_type_by_id(btf, krecord[i].type_id); 14812 if (!type || !btf_type_is_func(type)) { 14813 verbose(env, "invalid type id %d in func info", 14814 krecord[i].type_id); 14815 goto err_free; 14816 } 14817 info_aux[i].linkage = BTF_INFO_VLEN(type->info); 14818 14819 func_proto = btf_type_by_id(btf, type->type); 14820 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto))) 14821 /* btf_func_check() already verified it during BTF load */ 14822 goto err_free; 14823 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL); 14824 scalar_return = 14825 btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type); 14826 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) { 14827 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n"); 14828 goto err_free; 14829 } 14830 if (i && !scalar_return && env->subprog_info[i].has_tail_call) { 14831 verbose(env, "tail_call is only allowed in functions that return 'int'.\n"); 14832 goto err_free; 14833 } 14834 14835 prev_offset = krecord[i].insn_off; 14836 bpfptr_add(&urecord, urec_size); 14837 } 14838 14839 prog->aux->func_info = krecord; 14840 prog->aux->func_info_cnt = nfuncs; 14841 prog->aux->func_info_aux = info_aux; 14842 return 0; 14843 14844 err_free: 14845 kvfree(krecord); 14846 kfree(info_aux); 14847 return ret; 14848 } 14849 14850 static void adjust_btf_func(struct bpf_verifier_env *env) 14851 { 14852 struct bpf_prog_aux *aux = env->prog->aux; 14853 int i; 14854 14855 if (!aux->func_info) 14856 return; 14857 14858 for (i = 0; i < env->subprog_cnt; i++) 14859 aux->func_info[i].insn_off = env->subprog_info[i].start; 14860 } 14861 14862 #define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col) 14863 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE 14864 14865 static int check_btf_line(struct bpf_verifier_env *env, 14866 const union bpf_attr *attr, 14867 bpfptr_t uattr) 14868 { 14869 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0; 14870 struct bpf_subprog_info *sub; 14871 struct bpf_line_info *linfo; 14872 struct bpf_prog *prog; 14873 const struct btf *btf; 14874 bpfptr_t ulinfo; 14875 int err; 14876 14877 nr_linfo = attr->line_info_cnt; 14878 if (!nr_linfo) 14879 return 0; 14880 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info)) 14881 return -EINVAL; 14882 14883 rec_size = attr->line_info_rec_size; 14884 if (rec_size < MIN_BPF_LINEINFO_SIZE || 14885 rec_size > MAX_LINEINFO_REC_SIZE || 14886 rec_size & (sizeof(u32) - 1)) 14887 return -EINVAL; 14888 14889 /* Need to zero it in case the userspace may 14890 * pass in a smaller bpf_line_info object. 14891 */ 14892 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info), 14893 GFP_KERNEL | __GFP_NOWARN); 14894 if (!linfo) 14895 return -ENOMEM; 14896 14897 prog = env->prog; 14898 btf = prog->aux->btf; 14899 14900 s = 0; 14901 sub = env->subprog_info; 14902 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel); 14903 expected_size = sizeof(struct bpf_line_info); 14904 ncopy = min_t(u32, expected_size, rec_size); 14905 for (i = 0; i < nr_linfo; i++) { 14906 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size); 14907 if (err) { 14908 if (err == -E2BIG) { 14909 verbose(env, "nonzero tailing record in line_info"); 14910 if (copy_to_bpfptr_offset(uattr, 14911 offsetof(union bpf_attr, line_info_rec_size), 14912 &expected_size, sizeof(expected_size))) 14913 err = -EFAULT; 14914 } 14915 goto err_free; 14916 } 14917 14918 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) { 14919 err = -EFAULT; 14920 goto err_free; 14921 } 14922 14923 /* 14924 * Check insn_off to ensure 14925 * 1) strictly increasing AND 14926 * 2) bounded by prog->len 14927 * 14928 * The linfo[0].insn_off == 0 check logically falls into 14929 * the later "missing bpf_line_info for func..." case 14930 * because the first linfo[0].insn_off must be the 14931 * first sub also and the first sub must have 14932 * subprog_info[0].start == 0. 14933 */ 14934 if ((i && linfo[i].insn_off <= prev_offset) || 14935 linfo[i].insn_off >= prog->len) { 14936 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n", 14937 i, linfo[i].insn_off, prev_offset, 14938 prog->len); 14939 err = -EINVAL; 14940 goto err_free; 14941 } 14942 14943 if (!prog->insnsi[linfo[i].insn_off].code) { 14944 verbose(env, 14945 "Invalid insn code at line_info[%u].insn_off\n", 14946 i); 14947 err = -EINVAL; 14948 goto err_free; 14949 } 14950 14951 if (!btf_name_by_offset(btf, linfo[i].line_off) || 14952 !btf_name_by_offset(btf, linfo[i].file_name_off)) { 14953 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i); 14954 err = -EINVAL; 14955 goto err_free; 14956 } 14957 14958 if (s != env->subprog_cnt) { 14959 if (linfo[i].insn_off == sub[s].start) { 14960 sub[s].linfo_idx = i; 14961 s++; 14962 } else if (sub[s].start < linfo[i].insn_off) { 14963 verbose(env, "missing bpf_line_info for func#%u\n", s); 14964 err = -EINVAL; 14965 goto err_free; 14966 } 14967 } 14968 14969 prev_offset = linfo[i].insn_off; 14970 bpfptr_add(&ulinfo, rec_size); 14971 } 14972 14973 if (s != env->subprog_cnt) { 14974 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n", 14975 env->subprog_cnt - s, s); 14976 err = -EINVAL; 14977 goto err_free; 14978 } 14979 14980 prog->aux->linfo = linfo; 14981 prog->aux->nr_linfo = nr_linfo; 14982 14983 return 0; 14984 14985 err_free: 14986 kvfree(linfo); 14987 return err; 14988 } 14989 14990 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo) 14991 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE 14992 14993 static int check_core_relo(struct bpf_verifier_env *env, 14994 const union bpf_attr *attr, 14995 bpfptr_t uattr) 14996 { 14997 u32 i, nr_core_relo, ncopy, expected_size, rec_size; 14998 struct bpf_core_relo core_relo = {}; 14999 struct bpf_prog *prog = env->prog; 15000 const struct btf *btf = prog->aux->btf; 15001 struct bpf_core_ctx ctx = { 15002 .log = &env->log, 15003 .btf = btf, 15004 }; 15005 bpfptr_t u_core_relo; 15006 int err; 15007 15008 nr_core_relo = attr->core_relo_cnt; 15009 if (!nr_core_relo) 15010 return 0; 15011 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo)) 15012 return -EINVAL; 15013 15014 rec_size = attr->core_relo_rec_size; 15015 if (rec_size < MIN_CORE_RELO_SIZE || 15016 rec_size > MAX_CORE_RELO_SIZE || 15017 rec_size % sizeof(u32)) 15018 return -EINVAL; 15019 15020 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel); 15021 expected_size = sizeof(struct bpf_core_relo); 15022 ncopy = min_t(u32, expected_size, rec_size); 15023 15024 /* Unlike func_info and line_info, copy and apply each CO-RE 15025 * relocation record one at a time. 15026 */ 15027 for (i = 0; i < nr_core_relo; i++) { 15028 /* future proofing when sizeof(bpf_core_relo) changes */ 15029 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size); 15030 if (err) { 15031 if (err == -E2BIG) { 15032 verbose(env, "nonzero tailing record in core_relo"); 15033 if (copy_to_bpfptr_offset(uattr, 15034 offsetof(union bpf_attr, core_relo_rec_size), 15035 &expected_size, sizeof(expected_size))) 15036 err = -EFAULT; 15037 } 15038 break; 15039 } 15040 15041 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) { 15042 err = -EFAULT; 15043 break; 15044 } 15045 15046 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) { 15047 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n", 15048 i, core_relo.insn_off, prog->len); 15049 err = -EINVAL; 15050 break; 15051 } 15052 15053 err = bpf_core_apply(&ctx, &core_relo, i, 15054 &prog->insnsi[core_relo.insn_off / 8]); 15055 if (err) 15056 break; 15057 bpfptr_add(&u_core_relo, rec_size); 15058 } 15059 return err; 15060 } 15061 15062 static int check_btf_info(struct bpf_verifier_env *env, 15063 const union bpf_attr *attr, 15064 bpfptr_t uattr) 15065 { 15066 struct btf *btf; 15067 int err; 15068 15069 if (!attr->func_info_cnt && !attr->line_info_cnt) { 15070 if (check_abnormal_return(env)) 15071 return -EINVAL; 15072 return 0; 15073 } 15074 15075 btf = btf_get_by_fd(attr->prog_btf_fd); 15076 if (IS_ERR(btf)) 15077 return PTR_ERR(btf); 15078 if (btf_is_kernel(btf)) { 15079 btf_put(btf); 15080 return -EACCES; 15081 } 15082 env->prog->aux->btf = btf; 15083 15084 err = check_btf_func(env, attr, uattr); 15085 if (err) 15086 return err; 15087 15088 err = check_btf_line(env, attr, uattr); 15089 if (err) 15090 return err; 15091 15092 err = check_core_relo(env, attr, uattr); 15093 if (err) 15094 return err; 15095 15096 return 0; 15097 } 15098 15099 /* check %cur's range satisfies %old's */ 15100 static bool range_within(struct bpf_reg_state *old, 15101 struct bpf_reg_state *cur) 15102 { 15103 return old->umin_value <= cur->umin_value && 15104 old->umax_value >= cur->umax_value && 15105 old->smin_value <= cur->smin_value && 15106 old->smax_value >= cur->smax_value && 15107 old->u32_min_value <= cur->u32_min_value && 15108 old->u32_max_value >= cur->u32_max_value && 15109 old->s32_min_value <= cur->s32_min_value && 15110 old->s32_max_value >= cur->s32_max_value; 15111 } 15112 15113 /* If in the old state two registers had the same id, then they need to have 15114 * the same id in the new state as well. But that id could be different from 15115 * the old state, so we need to track the mapping from old to new ids. 15116 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent 15117 * regs with old id 5 must also have new id 9 for the new state to be safe. But 15118 * regs with a different old id could still have new id 9, we don't care about 15119 * that. 15120 * So we look through our idmap to see if this old id has been seen before. If 15121 * so, we require the new id to match; otherwise, we add the id pair to the map. 15122 */ 15123 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap) 15124 { 15125 struct bpf_id_pair *map = idmap->map; 15126 unsigned int i; 15127 15128 /* either both IDs should be set or both should be zero */ 15129 if (!!old_id != !!cur_id) 15130 return false; 15131 15132 if (old_id == 0) /* cur_id == 0 as well */ 15133 return true; 15134 15135 for (i = 0; i < BPF_ID_MAP_SIZE; i++) { 15136 if (!map[i].old) { 15137 /* Reached an empty slot; haven't seen this id before */ 15138 map[i].old = old_id; 15139 map[i].cur = cur_id; 15140 return true; 15141 } 15142 if (map[i].old == old_id) 15143 return map[i].cur == cur_id; 15144 if (map[i].cur == cur_id) 15145 return false; 15146 } 15147 /* We ran out of idmap slots, which should be impossible */ 15148 WARN_ON_ONCE(1); 15149 return false; 15150 } 15151 15152 /* Similar to check_ids(), but allocate a unique temporary ID 15153 * for 'old_id' or 'cur_id' of zero. 15154 * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid. 15155 */ 15156 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap) 15157 { 15158 old_id = old_id ? old_id : ++idmap->tmp_id_gen; 15159 cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen; 15160 15161 return check_ids(old_id, cur_id, idmap); 15162 } 15163 15164 static void clean_func_state(struct bpf_verifier_env *env, 15165 struct bpf_func_state *st) 15166 { 15167 enum bpf_reg_liveness live; 15168 int i, j; 15169 15170 for (i = 0; i < BPF_REG_FP; i++) { 15171 live = st->regs[i].live; 15172 /* liveness must not touch this register anymore */ 15173 st->regs[i].live |= REG_LIVE_DONE; 15174 if (!(live & REG_LIVE_READ)) 15175 /* since the register is unused, clear its state 15176 * to make further comparison simpler 15177 */ 15178 __mark_reg_not_init(env, &st->regs[i]); 15179 } 15180 15181 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) { 15182 live = st->stack[i].spilled_ptr.live; 15183 /* liveness must not touch this stack slot anymore */ 15184 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE; 15185 if (!(live & REG_LIVE_READ)) { 15186 __mark_reg_not_init(env, &st->stack[i].spilled_ptr); 15187 for (j = 0; j < BPF_REG_SIZE; j++) 15188 st->stack[i].slot_type[j] = STACK_INVALID; 15189 } 15190 } 15191 } 15192 15193 static void clean_verifier_state(struct bpf_verifier_env *env, 15194 struct bpf_verifier_state *st) 15195 { 15196 int i; 15197 15198 if (st->frame[0]->regs[0].live & REG_LIVE_DONE) 15199 /* all regs in this state in all frames were already marked */ 15200 return; 15201 15202 for (i = 0; i <= st->curframe; i++) 15203 clean_func_state(env, st->frame[i]); 15204 } 15205 15206 /* the parentage chains form a tree. 15207 * the verifier states are added to state lists at given insn and 15208 * pushed into state stack for future exploration. 15209 * when the verifier reaches bpf_exit insn some of the verifer states 15210 * stored in the state lists have their final liveness state already, 15211 * but a lot of states will get revised from liveness point of view when 15212 * the verifier explores other branches. 15213 * Example: 15214 * 1: r0 = 1 15215 * 2: if r1 == 100 goto pc+1 15216 * 3: r0 = 2 15217 * 4: exit 15218 * when the verifier reaches exit insn the register r0 in the state list of 15219 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch 15220 * of insn 2 and goes exploring further. At the insn 4 it will walk the 15221 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ. 15222 * 15223 * Since the verifier pushes the branch states as it sees them while exploring 15224 * the program the condition of walking the branch instruction for the second 15225 * time means that all states below this branch were already explored and 15226 * their final liveness marks are already propagated. 15227 * Hence when the verifier completes the search of state list in is_state_visited() 15228 * we can call this clean_live_states() function to mark all liveness states 15229 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state' 15230 * will not be used. 15231 * This function also clears the registers and stack for states that !READ 15232 * to simplify state merging. 15233 * 15234 * Important note here that walking the same branch instruction in the callee 15235 * doesn't meant that the states are DONE. The verifier has to compare 15236 * the callsites 15237 */ 15238 static void clean_live_states(struct bpf_verifier_env *env, int insn, 15239 struct bpf_verifier_state *cur) 15240 { 15241 struct bpf_verifier_state_list *sl; 15242 int i; 15243 15244 sl = *explored_state(env, insn); 15245 while (sl) { 15246 if (sl->state.branches) 15247 goto next; 15248 if (sl->state.insn_idx != insn || 15249 sl->state.curframe != cur->curframe) 15250 goto next; 15251 for (i = 0; i <= cur->curframe; i++) 15252 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite) 15253 goto next; 15254 clean_verifier_state(env, &sl->state); 15255 next: 15256 sl = sl->next; 15257 } 15258 } 15259 15260 static bool regs_exact(const struct bpf_reg_state *rold, 15261 const struct bpf_reg_state *rcur, 15262 struct bpf_idmap *idmap) 15263 { 15264 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 && 15265 check_ids(rold->id, rcur->id, idmap) && 15266 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap); 15267 } 15268 15269 /* Returns true if (rold safe implies rcur safe) */ 15270 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold, 15271 struct bpf_reg_state *rcur, struct bpf_idmap *idmap) 15272 { 15273 if (!(rold->live & REG_LIVE_READ)) 15274 /* explored state didn't use this */ 15275 return true; 15276 if (rold->type == NOT_INIT) 15277 /* explored state can't have used this */ 15278 return true; 15279 if (rcur->type == NOT_INIT) 15280 return false; 15281 15282 /* Enforce that register types have to match exactly, including their 15283 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general 15284 * rule. 15285 * 15286 * One can make a point that using a pointer register as unbounded 15287 * SCALAR would be technically acceptable, but this could lead to 15288 * pointer leaks because scalars are allowed to leak while pointers 15289 * are not. We could make this safe in special cases if root is 15290 * calling us, but it's probably not worth the hassle. 15291 * 15292 * Also, register types that are *not* MAYBE_NULL could technically be 15293 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE 15294 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point 15295 * to the same map). 15296 * However, if the old MAYBE_NULL register then got NULL checked, 15297 * doing so could have affected others with the same id, and we can't 15298 * check for that because we lost the id when we converted to 15299 * a non-MAYBE_NULL variant. 15300 * So, as a general rule we don't allow mixing MAYBE_NULL and 15301 * non-MAYBE_NULL registers as well. 15302 */ 15303 if (rold->type != rcur->type) 15304 return false; 15305 15306 switch (base_type(rold->type)) { 15307 case SCALAR_VALUE: 15308 if (env->explore_alu_limits) { 15309 /* explore_alu_limits disables tnum_in() and range_within() 15310 * logic and requires everything to be strict 15311 */ 15312 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 && 15313 check_scalar_ids(rold->id, rcur->id, idmap); 15314 } 15315 if (!rold->precise) 15316 return true; 15317 /* Why check_ids() for scalar registers? 15318 * 15319 * Consider the following BPF code: 15320 * 1: r6 = ... unbound scalar, ID=a ... 15321 * 2: r7 = ... unbound scalar, ID=b ... 15322 * 3: if (r6 > r7) goto +1 15323 * 4: r6 = r7 15324 * 5: if (r6 > X) goto ... 15325 * 6: ... memory operation using r7 ... 15326 * 15327 * First verification path is [1-6]: 15328 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7; 15329 * - at (5) r6 would be marked <= X, find_equal_scalars() would also mark 15330 * r7 <= X, because r6 and r7 share same id. 15331 * Next verification path is [1-4, 6]. 15332 * 15333 * Instruction (6) would be reached in two states: 15334 * I. r6{.id=b}, r7{.id=b} via path 1-6; 15335 * II. r6{.id=a}, r7{.id=b} via path 1-4, 6. 15336 * 15337 * Use check_ids() to distinguish these states. 15338 * --- 15339 * Also verify that new value satisfies old value range knowledge. 15340 */ 15341 return range_within(rold, rcur) && 15342 tnum_in(rold->var_off, rcur->var_off) && 15343 check_scalar_ids(rold->id, rcur->id, idmap); 15344 case PTR_TO_MAP_KEY: 15345 case PTR_TO_MAP_VALUE: 15346 case PTR_TO_MEM: 15347 case PTR_TO_BUF: 15348 case PTR_TO_TP_BUFFER: 15349 /* If the new min/max/var_off satisfy the old ones and 15350 * everything else matches, we are OK. 15351 */ 15352 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 && 15353 range_within(rold, rcur) && 15354 tnum_in(rold->var_off, rcur->var_off) && 15355 check_ids(rold->id, rcur->id, idmap) && 15356 check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap); 15357 case PTR_TO_PACKET_META: 15358 case PTR_TO_PACKET: 15359 /* We must have at least as much range as the old ptr 15360 * did, so that any accesses which were safe before are 15361 * still safe. This is true even if old range < old off, 15362 * since someone could have accessed through (ptr - k), or 15363 * even done ptr -= k in a register, to get a safe access. 15364 */ 15365 if (rold->range > rcur->range) 15366 return false; 15367 /* If the offsets don't match, we can't trust our alignment; 15368 * nor can we be sure that we won't fall out of range. 15369 */ 15370 if (rold->off != rcur->off) 15371 return false; 15372 /* id relations must be preserved */ 15373 if (!check_ids(rold->id, rcur->id, idmap)) 15374 return false; 15375 /* new val must satisfy old val knowledge */ 15376 return range_within(rold, rcur) && 15377 tnum_in(rold->var_off, rcur->var_off); 15378 case PTR_TO_STACK: 15379 /* two stack pointers are equal only if they're pointing to 15380 * the same stack frame, since fp-8 in foo != fp-8 in bar 15381 */ 15382 return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno; 15383 default: 15384 return regs_exact(rold, rcur, idmap); 15385 } 15386 } 15387 15388 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old, 15389 struct bpf_func_state *cur, struct bpf_idmap *idmap) 15390 { 15391 int i, spi; 15392 15393 /* walk slots of the explored stack and ignore any additional 15394 * slots in the current stack, since explored(safe) state 15395 * didn't use them 15396 */ 15397 for (i = 0; i < old->allocated_stack; i++) { 15398 struct bpf_reg_state *old_reg, *cur_reg; 15399 15400 spi = i / BPF_REG_SIZE; 15401 15402 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) { 15403 i += BPF_REG_SIZE - 1; 15404 /* explored state didn't use this */ 15405 continue; 15406 } 15407 15408 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID) 15409 continue; 15410 15411 if (env->allow_uninit_stack && 15412 old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC) 15413 continue; 15414 15415 /* explored stack has more populated slots than current stack 15416 * and these slots were used 15417 */ 15418 if (i >= cur->allocated_stack) 15419 return false; 15420 15421 /* if old state was safe with misc data in the stack 15422 * it will be safe with zero-initialized stack. 15423 * The opposite is not true 15424 */ 15425 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC && 15426 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO) 15427 continue; 15428 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] != 15429 cur->stack[spi].slot_type[i % BPF_REG_SIZE]) 15430 /* Ex: old explored (safe) state has STACK_SPILL in 15431 * this stack slot, but current has STACK_MISC -> 15432 * this verifier states are not equivalent, 15433 * return false to continue verification of this path 15434 */ 15435 return false; 15436 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1) 15437 continue; 15438 /* Both old and cur are having same slot_type */ 15439 switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) { 15440 case STACK_SPILL: 15441 /* when explored and current stack slot are both storing 15442 * spilled registers, check that stored pointers types 15443 * are the same as well. 15444 * Ex: explored safe path could have stored 15445 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8} 15446 * but current path has stored: 15447 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16} 15448 * such verifier states are not equivalent. 15449 * return false to continue verification of this path 15450 */ 15451 if (!regsafe(env, &old->stack[spi].spilled_ptr, 15452 &cur->stack[spi].spilled_ptr, idmap)) 15453 return false; 15454 break; 15455 case STACK_DYNPTR: 15456 old_reg = &old->stack[spi].spilled_ptr; 15457 cur_reg = &cur->stack[spi].spilled_ptr; 15458 if (old_reg->dynptr.type != cur_reg->dynptr.type || 15459 old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot || 15460 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap)) 15461 return false; 15462 break; 15463 case STACK_ITER: 15464 old_reg = &old->stack[spi].spilled_ptr; 15465 cur_reg = &cur->stack[spi].spilled_ptr; 15466 /* iter.depth is not compared between states as it 15467 * doesn't matter for correctness and would otherwise 15468 * prevent convergence; we maintain it only to prevent 15469 * infinite loop check triggering, see 15470 * iter_active_depths_differ() 15471 */ 15472 if (old_reg->iter.btf != cur_reg->iter.btf || 15473 old_reg->iter.btf_id != cur_reg->iter.btf_id || 15474 old_reg->iter.state != cur_reg->iter.state || 15475 /* ignore {old_reg,cur_reg}->iter.depth, see above */ 15476 !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap)) 15477 return false; 15478 break; 15479 case STACK_MISC: 15480 case STACK_ZERO: 15481 case STACK_INVALID: 15482 continue; 15483 /* Ensure that new unhandled slot types return false by default */ 15484 default: 15485 return false; 15486 } 15487 } 15488 return true; 15489 } 15490 15491 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur, 15492 struct bpf_idmap *idmap) 15493 { 15494 int i; 15495 15496 if (old->acquired_refs != cur->acquired_refs) 15497 return false; 15498 15499 for (i = 0; i < old->acquired_refs; i++) { 15500 if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap)) 15501 return false; 15502 } 15503 15504 return true; 15505 } 15506 15507 /* compare two verifier states 15508 * 15509 * all states stored in state_list are known to be valid, since 15510 * verifier reached 'bpf_exit' instruction through them 15511 * 15512 * this function is called when verifier exploring different branches of 15513 * execution popped from the state stack. If it sees an old state that has 15514 * more strict register state and more strict stack state then this execution 15515 * branch doesn't need to be explored further, since verifier already 15516 * concluded that more strict state leads to valid finish. 15517 * 15518 * Therefore two states are equivalent if register state is more conservative 15519 * and explored stack state is more conservative than the current one. 15520 * Example: 15521 * explored current 15522 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC) 15523 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC) 15524 * 15525 * In other words if current stack state (one being explored) has more 15526 * valid slots than old one that already passed validation, it means 15527 * the verifier can stop exploring and conclude that current state is valid too 15528 * 15529 * Similarly with registers. If explored state has register type as invalid 15530 * whereas register type in current state is meaningful, it means that 15531 * the current state will reach 'bpf_exit' instruction safely 15532 */ 15533 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old, 15534 struct bpf_func_state *cur) 15535 { 15536 int i; 15537 15538 for (i = 0; i < MAX_BPF_REG; i++) 15539 if (!regsafe(env, &old->regs[i], &cur->regs[i], 15540 &env->idmap_scratch)) 15541 return false; 15542 15543 if (!stacksafe(env, old, cur, &env->idmap_scratch)) 15544 return false; 15545 15546 if (!refsafe(old, cur, &env->idmap_scratch)) 15547 return false; 15548 15549 return true; 15550 } 15551 15552 static bool states_equal(struct bpf_verifier_env *env, 15553 struct bpf_verifier_state *old, 15554 struct bpf_verifier_state *cur) 15555 { 15556 int i; 15557 15558 if (old->curframe != cur->curframe) 15559 return false; 15560 15561 env->idmap_scratch.tmp_id_gen = env->id_gen; 15562 memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map)); 15563 15564 /* Verification state from speculative execution simulation 15565 * must never prune a non-speculative execution one. 15566 */ 15567 if (old->speculative && !cur->speculative) 15568 return false; 15569 15570 if (old->active_lock.ptr != cur->active_lock.ptr) 15571 return false; 15572 15573 /* Old and cur active_lock's have to be either both present 15574 * or both absent. 15575 */ 15576 if (!!old->active_lock.id != !!cur->active_lock.id) 15577 return false; 15578 15579 if (old->active_lock.id && 15580 !check_ids(old->active_lock.id, cur->active_lock.id, &env->idmap_scratch)) 15581 return false; 15582 15583 if (old->active_rcu_lock != cur->active_rcu_lock) 15584 return false; 15585 15586 /* for states to be equal callsites have to be the same 15587 * and all frame states need to be equivalent 15588 */ 15589 for (i = 0; i <= old->curframe; i++) { 15590 if (old->frame[i]->callsite != cur->frame[i]->callsite) 15591 return false; 15592 if (!func_states_equal(env, old->frame[i], cur->frame[i])) 15593 return false; 15594 } 15595 return true; 15596 } 15597 15598 /* Return 0 if no propagation happened. Return negative error code if error 15599 * happened. Otherwise, return the propagated bit. 15600 */ 15601 static int propagate_liveness_reg(struct bpf_verifier_env *env, 15602 struct bpf_reg_state *reg, 15603 struct bpf_reg_state *parent_reg) 15604 { 15605 u8 parent_flag = parent_reg->live & REG_LIVE_READ; 15606 u8 flag = reg->live & REG_LIVE_READ; 15607 int err; 15608 15609 /* When comes here, read flags of PARENT_REG or REG could be any of 15610 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need 15611 * of propagation if PARENT_REG has strongest REG_LIVE_READ64. 15612 */ 15613 if (parent_flag == REG_LIVE_READ64 || 15614 /* Or if there is no read flag from REG. */ 15615 !flag || 15616 /* Or if the read flag from REG is the same as PARENT_REG. */ 15617 parent_flag == flag) 15618 return 0; 15619 15620 err = mark_reg_read(env, reg, parent_reg, flag); 15621 if (err) 15622 return err; 15623 15624 return flag; 15625 } 15626 15627 /* A write screens off any subsequent reads; but write marks come from the 15628 * straight-line code between a state and its parent. When we arrive at an 15629 * equivalent state (jump target or such) we didn't arrive by the straight-line 15630 * code, so read marks in the state must propagate to the parent regardless 15631 * of the state's write marks. That's what 'parent == state->parent' comparison 15632 * in mark_reg_read() is for. 15633 */ 15634 static int propagate_liveness(struct bpf_verifier_env *env, 15635 const struct bpf_verifier_state *vstate, 15636 struct bpf_verifier_state *vparent) 15637 { 15638 struct bpf_reg_state *state_reg, *parent_reg; 15639 struct bpf_func_state *state, *parent; 15640 int i, frame, err = 0; 15641 15642 if (vparent->curframe != vstate->curframe) { 15643 WARN(1, "propagate_live: parent frame %d current frame %d\n", 15644 vparent->curframe, vstate->curframe); 15645 return -EFAULT; 15646 } 15647 /* Propagate read liveness of registers... */ 15648 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG); 15649 for (frame = 0; frame <= vstate->curframe; frame++) { 15650 parent = vparent->frame[frame]; 15651 state = vstate->frame[frame]; 15652 parent_reg = parent->regs; 15653 state_reg = state->regs; 15654 /* We don't need to worry about FP liveness, it's read-only */ 15655 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) { 15656 err = propagate_liveness_reg(env, &state_reg[i], 15657 &parent_reg[i]); 15658 if (err < 0) 15659 return err; 15660 if (err == REG_LIVE_READ64) 15661 mark_insn_zext(env, &parent_reg[i]); 15662 } 15663 15664 /* Propagate stack slots. */ 15665 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE && 15666 i < parent->allocated_stack / BPF_REG_SIZE; i++) { 15667 parent_reg = &parent->stack[i].spilled_ptr; 15668 state_reg = &state->stack[i].spilled_ptr; 15669 err = propagate_liveness_reg(env, state_reg, 15670 parent_reg); 15671 if (err < 0) 15672 return err; 15673 } 15674 } 15675 return 0; 15676 } 15677 15678 /* find precise scalars in the previous equivalent state and 15679 * propagate them into the current state 15680 */ 15681 static int propagate_precision(struct bpf_verifier_env *env, 15682 const struct bpf_verifier_state *old) 15683 { 15684 struct bpf_reg_state *state_reg; 15685 struct bpf_func_state *state; 15686 int i, err = 0, fr; 15687 bool first; 15688 15689 for (fr = old->curframe; fr >= 0; fr--) { 15690 state = old->frame[fr]; 15691 state_reg = state->regs; 15692 first = true; 15693 for (i = 0; i < BPF_REG_FP; i++, state_reg++) { 15694 if (state_reg->type != SCALAR_VALUE || 15695 !state_reg->precise || 15696 !(state_reg->live & REG_LIVE_READ)) 15697 continue; 15698 if (env->log.level & BPF_LOG_LEVEL2) { 15699 if (first) 15700 verbose(env, "frame %d: propagating r%d", fr, i); 15701 else 15702 verbose(env, ",r%d", i); 15703 } 15704 bt_set_frame_reg(&env->bt, fr, i); 15705 first = false; 15706 } 15707 15708 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { 15709 if (!is_spilled_reg(&state->stack[i])) 15710 continue; 15711 state_reg = &state->stack[i].spilled_ptr; 15712 if (state_reg->type != SCALAR_VALUE || 15713 !state_reg->precise || 15714 !(state_reg->live & REG_LIVE_READ)) 15715 continue; 15716 if (env->log.level & BPF_LOG_LEVEL2) { 15717 if (first) 15718 verbose(env, "frame %d: propagating fp%d", 15719 fr, (-i - 1) * BPF_REG_SIZE); 15720 else 15721 verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE); 15722 } 15723 bt_set_frame_slot(&env->bt, fr, i); 15724 first = false; 15725 } 15726 if (!first) 15727 verbose(env, "\n"); 15728 } 15729 15730 err = mark_chain_precision_batch(env); 15731 if (err < 0) 15732 return err; 15733 15734 return 0; 15735 } 15736 15737 static bool states_maybe_looping(struct bpf_verifier_state *old, 15738 struct bpf_verifier_state *cur) 15739 { 15740 struct bpf_func_state *fold, *fcur; 15741 int i, fr = cur->curframe; 15742 15743 if (old->curframe != fr) 15744 return false; 15745 15746 fold = old->frame[fr]; 15747 fcur = cur->frame[fr]; 15748 for (i = 0; i < MAX_BPF_REG; i++) 15749 if (memcmp(&fold->regs[i], &fcur->regs[i], 15750 offsetof(struct bpf_reg_state, parent))) 15751 return false; 15752 return true; 15753 } 15754 15755 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx) 15756 { 15757 return env->insn_aux_data[insn_idx].is_iter_next; 15758 } 15759 15760 /* is_state_visited() handles iter_next() (see process_iter_next_call() for 15761 * terminology) calls specially: as opposed to bounded BPF loops, it *expects* 15762 * states to match, which otherwise would look like an infinite loop. So while 15763 * iter_next() calls are taken care of, we still need to be careful and 15764 * prevent erroneous and too eager declaration of "ininite loop", when 15765 * iterators are involved. 15766 * 15767 * Here's a situation in pseudo-BPF assembly form: 15768 * 15769 * 0: again: ; set up iter_next() call args 15770 * 1: r1 = &it ; <CHECKPOINT HERE> 15771 * 2: call bpf_iter_num_next ; this is iter_next() call 15772 * 3: if r0 == 0 goto done 15773 * 4: ... something useful here ... 15774 * 5: goto again ; another iteration 15775 * 6: done: 15776 * 7: r1 = &it 15777 * 8: call bpf_iter_num_destroy ; clean up iter state 15778 * 9: exit 15779 * 15780 * This is a typical loop. Let's assume that we have a prune point at 1:, 15781 * before we get to `call bpf_iter_num_next` (e.g., because of that `goto 15782 * again`, assuming other heuristics don't get in a way). 15783 * 15784 * When we first time come to 1:, let's say we have some state X. We proceed 15785 * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit. 15786 * Now we come back to validate that forked ACTIVE state. We proceed through 15787 * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we 15788 * are converging. But the problem is that we don't know that yet, as this 15789 * convergence has to happen at iter_next() call site only. So if nothing is 15790 * done, at 1: verifier will use bounded loop logic and declare infinite 15791 * looping (and would be *technically* correct, if not for iterator's 15792 * "eventual sticky NULL" contract, see process_iter_next_call()). But we 15793 * don't want that. So what we do in process_iter_next_call() when we go on 15794 * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's 15795 * a different iteration. So when we suspect an infinite loop, we additionally 15796 * check if any of the *ACTIVE* iterator states depths differ. If yes, we 15797 * pretend we are not looping and wait for next iter_next() call. 15798 * 15799 * This only applies to ACTIVE state. In DRAINED state we don't expect to 15800 * loop, because that would actually mean infinite loop, as DRAINED state is 15801 * "sticky", and so we'll keep returning into the same instruction with the 15802 * same state (at least in one of possible code paths). 15803 * 15804 * This approach allows to keep infinite loop heuristic even in the face of 15805 * active iterator. E.g., C snippet below is and will be detected as 15806 * inifintely looping: 15807 * 15808 * struct bpf_iter_num it; 15809 * int *p, x; 15810 * 15811 * bpf_iter_num_new(&it, 0, 10); 15812 * while ((p = bpf_iter_num_next(&t))) { 15813 * x = p; 15814 * while (x--) {} // <<-- infinite loop here 15815 * } 15816 * 15817 */ 15818 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur) 15819 { 15820 struct bpf_reg_state *slot, *cur_slot; 15821 struct bpf_func_state *state; 15822 int i, fr; 15823 15824 for (fr = old->curframe; fr >= 0; fr--) { 15825 state = old->frame[fr]; 15826 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) { 15827 if (state->stack[i].slot_type[0] != STACK_ITER) 15828 continue; 15829 15830 slot = &state->stack[i].spilled_ptr; 15831 if (slot->iter.state != BPF_ITER_STATE_ACTIVE) 15832 continue; 15833 15834 cur_slot = &cur->frame[fr]->stack[i].spilled_ptr; 15835 if (cur_slot->iter.depth != slot->iter.depth) 15836 return true; 15837 } 15838 } 15839 return false; 15840 } 15841 15842 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx) 15843 { 15844 struct bpf_verifier_state_list *new_sl; 15845 struct bpf_verifier_state_list *sl, **pprev; 15846 struct bpf_verifier_state *cur = env->cur_state, *new; 15847 int i, j, err, states_cnt = 0; 15848 bool force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx); 15849 bool add_new_state = force_new_state; 15850 15851 /* bpf progs typically have pruning point every 4 instructions 15852 * http://vger.kernel.org/bpfconf2019.html#session-1 15853 * Do not add new state for future pruning if the verifier hasn't seen 15854 * at least 2 jumps and at least 8 instructions. 15855 * This heuristics helps decrease 'total_states' and 'peak_states' metric. 15856 * In tests that amounts to up to 50% reduction into total verifier 15857 * memory consumption and 20% verifier time speedup. 15858 */ 15859 if (env->jmps_processed - env->prev_jmps_processed >= 2 && 15860 env->insn_processed - env->prev_insn_processed >= 8) 15861 add_new_state = true; 15862 15863 pprev = explored_state(env, insn_idx); 15864 sl = *pprev; 15865 15866 clean_live_states(env, insn_idx, cur); 15867 15868 while (sl) { 15869 states_cnt++; 15870 if (sl->state.insn_idx != insn_idx) 15871 goto next; 15872 15873 if (sl->state.branches) { 15874 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe]; 15875 15876 if (frame->in_async_callback_fn && 15877 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) { 15878 /* Different async_entry_cnt means that the verifier is 15879 * processing another entry into async callback. 15880 * Seeing the same state is not an indication of infinite 15881 * loop or infinite recursion. 15882 * But finding the same state doesn't mean that it's safe 15883 * to stop processing the current state. The previous state 15884 * hasn't yet reached bpf_exit, since state.branches > 0. 15885 * Checking in_async_callback_fn alone is not enough either. 15886 * Since the verifier still needs to catch infinite loops 15887 * inside async callbacks. 15888 */ 15889 goto skip_inf_loop_check; 15890 } 15891 /* BPF open-coded iterators loop detection is special. 15892 * states_maybe_looping() logic is too simplistic in detecting 15893 * states that *might* be equivalent, because it doesn't know 15894 * about ID remapping, so don't even perform it. 15895 * See process_iter_next_call() and iter_active_depths_differ() 15896 * for overview of the logic. When current and one of parent 15897 * states are detected as equivalent, it's a good thing: we prove 15898 * convergence and can stop simulating further iterations. 15899 * It's safe to assume that iterator loop will finish, taking into 15900 * account iter_next() contract of eventually returning 15901 * sticky NULL result. 15902 */ 15903 if (is_iter_next_insn(env, insn_idx)) { 15904 if (states_equal(env, &sl->state, cur)) { 15905 struct bpf_func_state *cur_frame; 15906 struct bpf_reg_state *iter_state, *iter_reg; 15907 int spi; 15908 15909 cur_frame = cur->frame[cur->curframe]; 15910 /* btf_check_iter_kfuncs() enforces that 15911 * iter state pointer is always the first arg 15912 */ 15913 iter_reg = &cur_frame->regs[BPF_REG_1]; 15914 /* current state is valid due to states_equal(), 15915 * so we can assume valid iter and reg state, 15916 * no need for extra (re-)validations 15917 */ 15918 spi = __get_spi(iter_reg->off + iter_reg->var_off.value); 15919 iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr; 15920 if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE) 15921 goto hit; 15922 } 15923 goto skip_inf_loop_check; 15924 } 15925 /* attempt to detect infinite loop to avoid unnecessary doomed work */ 15926 if (states_maybe_looping(&sl->state, cur) && 15927 states_equal(env, &sl->state, cur) && 15928 !iter_active_depths_differ(&sl->state, cur)) { 15929 verbose_linfo(env, insn_idx, "; "); 15930 verbose(env, "infinite loop detected at insn %d\n", insn_idx); 15931 return -EINVAL; 15932 } 15933 /* if the verifier is processing a loop, avoid adding new state 15934 * too often, since different loop iterations have distinct 15935 * states and may not help future pruning. 15936 * This threshold shouldn't be too low to make sure that 15937 * a loop with large bound will be rejected quickly. 15938 * The most abusive loop will be: 15939 * r1 += 1 15940 * if r1 < 1000000 goto pc-2 15941 * 1M insn_procssed limit / 100 == 10k peak states. 15942 * This threshold shouldn't be too high either, since states 15943 * at the end of the loop are likely to be useful in pruning. 15944 */ 15945 skip_inf_loop_check: 15946 if (!force_new_state && 15947 env->jmps_processed - env->prev_jmps_processed < 20 && 15948 env->insn_processed - env->prev_insn_processed < 100) 15949 add_new_state = false; 15950 goto miss; 15951 } 15952 if (states_equal(env, &sl->state, cur)) { 15953 hit: 15954 sl->hit_cnt++; 15955 /* reached equivalent register/stack state, 15956 * prune the search. 15957 * Registers read by the continuation are read by us. 15958 * If we have any write marks in env->cur_state, they 15959 * will prevent corresponding reads in the continuation 15960 * from reaching our parent (an explored_state). Our 15961 * own state will get the read marks recorded, but 15962 * they'll be immediately forgotten as we're pruning 15963 * this state and will pop a new one. 15964 */ 15965 err = propagate_liveness(env, &sl->state, cur); 15966 15967 /* if previous state reached the exit with precision and 15968 * current state is equivalent to it (except precsion marks) 15969 * the precision needs to be propagated back in 15970 * the current state. 15971 */ 15972 err = err ? : push_jmp_history(env, cur); 15973 err = err ? : propagate_precision(env, &sl->state); 15974 if (err) 15975 return err; 15976 return 1; 15977 } 15978 miss: 15979 /* when new state is not going to be added do not increase miss count. 15980 * Otherwise several loop iterations will remove the state 15981 * recorded earlier. The goal of these heuristics is to have 15982 * states from some iterations of the loop (some in the beginning 15983 * and some at the end) to help pruning. 15984 */ 15985 if (add_new_state) 15986 sl->miss_cnt++; 15987 /* heuristic to determine whether this state is beneficial 15988 * to keep checking from state equivalence point of view. 15989 * Higher numbers increase max_states_per_insn and verification time, 15990 * but do not meaningfully decrease insn_processed. 15991 */ 15992 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) { 15993 /* the state is unlikely to be useful. Remove it to 15994 * speed up verification 15995 */ 15996 *pprev = sl->next; 15997 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) { 15998 u32 br = sl->state.branches; 15999 16000 WARN_ONCE(br, 16001 "BUG live_done but branches_to_explore %d\n", 16002 br); 16003 free_verifier_state(&sl->state, false); 16004 kfree(sl); 16005 env->peak_states--; 16006 } else { 16007 /* cannot free this state, since parentage chain may 16008 * walk it later. Add it for free_list instead to 16009 * be freed at the end of verification 16010 */ 16011 sl->next = env->free_list; 16012 env->free_list = sl; 16013 } 16014 sl = *pprev; 16015 continue; 16016 } 16017 next: 16018 pprev = &sl->next; 16019 sl = *pprev; 16020 } 16021 16022 if (env->max_states_per_insn < states_cnt) 16023 env->max_states_per_insn = states_cnt; 16024 16025 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES) 16026 return 0; 16027 16028 if (!add_new_state) 16029 return 0; 16030 16031 /* There were no equivalent states, remember the current one. 16032 * Technically the current state is not proven to be safe yet, 16033 * but it will either reach outer most bpf_exit (which means it's safe) 16034 * or it will be rejected. When there are no loops the verifier won't be 16035 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx) 16036 * again on the way to bpf_exit. 16037 * When looping the sl->state.branches will be > 0 and this state 16038 * will not be considered for equivalence until branches == 0. 16039 */ 16040 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL); 16041 if (!new_sl) 16042 return -ENOMEM; 16043 env->total_states++; 16044 env->peak_states++; 16045 env->prev_jmps_processed = env->jmps_processed; 16046 env->prev_insn_processed = env->insn_processed; 16047 16048 /* forget precise markings we inherited, see __mark_chain_precision */ 16049 if (env->bpf_capable) 16050 mark_all_scalars_imprecise(env, cur); 16051 16052 /* add new state to the head of linked list */ 16053 new = &new_sl->state; 16054 err = copy_verifier_state(new, cur); 16055 if (err) { 16056 free_verifier_state(new, false); 16057 kfree(new_sl); 16058 return err; 16059 } 16060 new->insn_idx = insn_idx; 16061 WARN_ONCE(new->branches != 1, 16062 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx); 16063 16064 cur->parent = new; 16065 cur->first_insn_idx = insn_idx; 16066 clear_jmp_history(cur); 16067 new_sl->next = *explored_state(env, insn_idx); 16068 *explored_state(env, insn_idx) = new_sl; 16069 /* connect new state to parentage chain. Current frame needs all 16070 * registers connected. Only r6 - r9 of the callers are alive (pushed 16071 * to the stack implicitly by JITs) so in callers' frames connect just 16072 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to 16073 * the state of the call instruction (with WRITTEN set), and r0 comes 16074 * from callee with its full parentage chain, anyway. 16075 */ 16076 /* clear write marks in current state: the writes we did are not writes 16077 * our child did, so they don't screen off its reads from us. 16078 * (There are no read marks in current state, because reads always mark 16079 * their parent and current state never has children yet. Only 16080 * explored_states can get read marks.) 16081 */ 16082 for (j = 0; j <= cur->curframe; j++) { 16083 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) 16084 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i]; 16085 for (i = 0; i < BPF_REG_FP; i++) 16086 cur->frame[j]->regs[i].live = REG_LIVE_NONE; 16087 } 16088 16089 /* all stack frames are accessible from callee, clear them all */ 16090 for (j = 0; j <= cur->curframe; j++) { 16091 struct bpf_func_state *frame = cur->frame[j]; 16092 struct bpf_func_state *newframe = new->frame[j]; 16093 16094 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) { 16095 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE; 16096 frame->stack[i].spilled_ptr.parent = 16097 &newframe->stack[i].spilled_ptr; 16098 } 16099 } 16100 return 0; 16101 } 16102 16103 /* Return true if it's OK to have the same insn return a different type. */ 16104 static bool reg_type_mismatch_ok(enum bpf_reg_type type) 16105 { 16106 switch (base_type(type)) { 16107 case PTR_TO_CTX: 16108 case PTR_TO_SOCKET: 16109 case PTR_TO_SOCK_COMMON: 16110 case PTR_TO_TCP_SOCK: 16111 case PTR_TO_XDP_SOCK: 16112 case PTR_TO_BTF_ID: 16113 return false; 16114 default: 16115 return true; 16116 } 16117 } 16118 16119 /* If an instruction was previously used with particular pointer types, then we 16120 * need to be careful to avoid cases such as the below, where it may be ok 16121 * for one branch accessing the pointer, but not ok for the other branch: 16122 * 16123 * R1 = sock_ptr 16124 * goto X; 16125 * ... 16126 * R1 = some_other_valid_ptr; 16127 * goto X; 16128 * ... 16129 * R2 = *(u32 *)(R1 + 0); 16130 */ 16131 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev) 16132 { 16133 return src != prev && (!reg_type_mismatch_ok(src) || 16134 !reg_type_mismatch_ok(prev)); 16135 } 16136 16137 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type, 16138 bool allow_trust_missmatch) 16139 { 16140 enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type; 16141 16142 if (*prev_type == NOT_INIT) { 16143 /* Saw a valid insn 16144 * dst_reg = *(u32 *)(src_reg + off) 16145 * save type to validate intersecting paths 16146 */ 16147 *prev_type = type; 16148 } else if (reg_type_mismatch(type, *prev_type)) { 16149 /* Abuser program is trying to use the same insn 16150 * dst_reg = *(u32*) (src_reg + off) 16151 * with different pointer types: 16152 * src_reg == ctx in one branch and 16153 * src_reg == stack|map in some other branch. 16154 * Reject it. 16155 */ 16156 if (allow_trust_missmatch && 16157 base_type(type) == PTR_TO_BTF_ID && 16158 base_type(*prev_type) == PTR_TO_BTF_ID) { 16159 /* 16160 * Have to support a use case when one path through 16161 * the program yields TRUSTED pointer while another 16162 * is UNTRUSTED. Fallback to UNTRUSTED to generate 16163 * BPF_PROBE_MEM. 16164 */ 16165 *prev_type = PTR_TO_BTF_ID | PTR_UNTRUSTED; 16166 } else { 16167 verbose(env, "same insn cannot be used with different pointers\n"); 16168 return -EINVAL; 16169 } 16170 } 16171 16172 return 0; 16173 } 16174 16175 static int do_check(struct bpf_verifier_env *env) 16176 { 16177 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2); 16178 struct bpf_verifier_state *state = env->cur_state; 16179 struct bpf_insn *insns = env->prog->insnsi; 16180 struct bpf_reg_state *regs; 16181 int insn_cnt = env->prog->len; 16182 bool do_print_state = false; 16183 int prev_insn_idx = -1; 16184 16185 for (;;) { 16186 struct bpf_insn *insn; 16187 u8 class; 16188 int err; 16189 16190 env->prev_insn_idx = prev_insn_idx; 16191 if (env->insn_idx >= insn_cnt) { 16192 verbose(env, "invalid insn idx %d insn_cnt %d\n", 16193 env->insn_idx, insn_cnt); 16194 return -EFAULT; 16195 } 16196 16197 insn = &insns[env->insn_idx]; 16198 class = BPF_CLASS(insn->code); 16199 16200 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) { 16201 verbose(env, 16202 "BPF program is too large. Processed %d insn\n", 16203 env->insn_processed); 16204 return -E2BIG; 16205 } 16206 16207 state->last_insn_idx = env->prev_insn_idx; 16208 16209 if (is_prune_point(env, env->insn_idx)) { 16210 err = is_state_visited(env, env->insn_idx); 16211 if (err < 0) 16212 return err; 16213 if (err == 1) { 16214 /* found equivalent state, can prune the search */ 16215 if (env->log.level & BPF_LOG_LEVEL) { 16216 if (do_print_state) 16217 verbose(env, "\nfrom %d to %d%s: safe\n", 16218 env->prev_insn_idx, env->insn_idx, 16219 env->cur_state->speculative ? 16220 " (speculative execution)" : ""); 16221 else 16222 verbose(env, "%d: safe\n", env->insn_idx); 16223 } 16224 goto process_bpf_exit; 16225 } 16226 } 16227 16228 if (is_jmp_point(env, env->insn_idx)) { 16229 err = push_jmp_history(env, state); 16230 if (err) 16231 return err; 16232 } 16233 16234 if (signal_pending(current)) 16235 return -EAGAIN; 16236 16237 if (need_resched()) 16238 cond_resched(); 16239 16240 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) { 16241 verbose(env, "\nfrom %d to %d%s:", 16242 env->prev_insn_idx, env->insn_idx, 16243 env->cur_state->speculative ? 16244 " (speculative execution)" : ""); 16245 print_verifier_state(env, state->frame[state->curframe], true); 16246 do_print_state = false; 16247 } 16248 16249 if (env->log.level & BPF_LOG_LEVEL) { 16250 const struct bpf_insn_cbs cbs = { 16251 .cb_call = disasm_kfunc_name, 16252 .cb_print = verbose, 16253 .private_data = env, 16254 }; 16255 16256 if (verifier_state_scratched(env)) 16257 print_insn_state(env, state->frame[state->curframe]); 16258 16259 verbose_linfo(env, env->insn_idx, "; "); 16260 env->prev_log_pos = env->log.end_pos; 16261 verbose(env, "%d: ", env->insn_idx); 16262 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks); 16263 env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos; 16264 env->prev_log_pos = env->log.end_pos; 16265 } 16266 16267 if (bpf_prog_is_offloaded(env->prog->aux)) { 16268 err = bpf_prog_offload_verify_insn(env, env->insn_idx, 16269 env->prev_insn_idx); 16270 if (err) 16271 return err; 16272 } 16273 16274 regs = cur_regs(env); 16275 sanitize_mark_insn_seen(env); 16276 prev_insn_idx = env->insn_idx; 16277 16278 if (class == BPF_ALU || class == BPF_ALU64) { 16279 err = check_alu_op(env, insn); 16280 if (err) 16281 return err; 16282 16283 } else if (class == BPF_LDX) { 16284 enum bpf_reg_type src_reg_type; 16285 16286 /* check for reserved fields is already done */ 16287 16288 /* check src operand */ 16289 err = check_reg_arg(env, insn->src_reg, SRC_OP); 16290 if (err) 16291 return err; 16292 16293 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK); 16294 if (err) 16295 return err; 16296 16297 src_reg_type = regs[insn->src_reg].type; 16298 16299 /* check that memory (src_reg + off) is readable, 16300 * the state of dst_reg will be updated by this func 16301 */ 16302 err = check_mem_access(env, env->insn_idx, insn->src_reg, 16303 insn->off, BPF_SIZE(insn->code), 16304 BPF_READ, insn->dst_reg, false); 16305 if (err) 16306 return err; 16307 16308 err = save_aux_ptr_type(env, src_reg_type, true); 16309 if (err) 16310 return err; 16311 } else if (class == BPF_STX) { 16312 enum bpf_reg_type dst_reg_type; 16313 16314 if (BPF_MODE(insn->code) == BPF_ATOMIC) { 16315 err = check_atomic(env, env->insn_idx, insn); 16316 if (err) 16317 return err; 16318 env->insn_idx++; 16319 continue; 16320 } 16321 16322 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) { 16323 verbose(env, "BPF_STX uses reserved fields\n"); 16324 return -EINVAL; 16325 } 16326 16327 /* check src1 operand */ 16328 err = check_reg_arg(env, insn->src_reg, SRC_OP); 16329 if (err) 16330 return err; 16331 /* check src2 operand */ 16332 err = check_reg_arg(env, insn->dst_reg, SRC_OP); 16333 if (err) 16334 return err; 16335 16336 dst_reg_type = regs[insn->dst_reg].type; 16337 16338 /* check that memory (dst_reg + off) is writeable */ 16339 err = check_mem_access(env, env->insn_idx, insn->dst_reg, 16340 insn->off, BPF_SIZE(insn->code), 16341 BPF_WRITE, insn->src_reg, false); 16342 if (err) 16343 return err; 16344 16345 err = save_aux_ptr_type(env, dst_reg_type, false); 16346 if (err) 16347 return err; 16348 } else if (class == BPF_ST) { 16349 enum bpf_reg_type dst_reg_type; 16350 16351 if (BPF_MODE(insn->code) != BPF_MEM || 16352 insn->src_reg != BPF_REG_0) { 16353 verbose(env, "BPF_ST uses reserved fields\n"); 16354 return -EINVAL; 16355 } 16356 /* check src operand */ 16357 err = check_reg_arg(env, insn->dst_reg, SRC_OP); 16358 if (err) 16359 return err; 16360 16361 dst_reg_type = regs[insn->dst_reg].type; 16362 16363 /* check that memory (dst_reg + off) is writeable */ 16364 err = check_mem_access(env, env->insn_idx, insn->dst_reg, 16365 insn->off, BPF_SIZE(insn->code), 16366 BPF_WRITE, -1, false); 16367 if (err) 16368 return err; 16369 16370 err = save_aux_ptr_type(env, dst_reg_type, false); 16371 if (err) 16372 return err; 16373 } else if (class == BPF_JMP || class == BPF_JMP32) { 16374 u8 opcode = BPF_OP(insn->code); 16375 16376 env->jmps_processed++; 16377 if (opcode == BPF_CALL) { 16378 if (BPF_SRC(insn->code) != BPF_K || 16379 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL 16380 && insn->off != 0) || 16381 (insn->src_reg != BPF_REG_0 && 16382 insn->src_reg != BPF_PSEUDO_CALL && 16383 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) || 16384 insn->dst_reg != BPF_REG_0 || 16385 class == BPF_JMP32) { 16386 verbose(env, "BPF_CALL uses reserved fields\n"); 16387 return -EINVAL; 16388 } 16389 16390 if (env->cur_state->active_lock.ptr) { 16391 if ((insn->src_reg == BPF_REG_0 && insn->imm != BPF_FUNC_spin_unlock) || 16392 (insn->src_reg == BPF_PSEUDO_CALL) || 16393 (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 16394 (insn->off != 0 || !is_bpf_graph_api_kfunc(insn->imm)))) { 16395 verbose(env, "function calls are not allowed while holding a lock\n"); 16396 return -EINVAL; 16397 } 16398 } 16399 if (insn->src_reg == BPF_PSEUDO_CALL) 16400 err = check_func_call(env, insn, &env->insn_idx); 16401 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) 16402 err = check_kfunc_call(env, insn, &env->insn_idx); 16403 else 16404 err = check_helper_call(env, insn, &env->insn_idx); 16405 if (err) 16406 return err; 16407 16408 mark_reg_scratched(env, BPF_REG_0); 16409 } else if (opcode == BPF_JA) { 16410 if (BPF_SRC(insn->code) != BPF_K || 16411 insn->imm != 0 || 16412 insn->src_reg != BPF_REG_0 || 16413 insn->dst_reg != BPF_REG_0 || 16414 class == BPF_JMP32) { 16415 verbose(env, "BPF_JA uses reserved fields\n"); 16416 return -EINVAL; 16417 } 16418 16419 env->insn_idx += insn->off + 1; 16420 continue; 16421 16422 } else if (opcode == BPF_EXIT) { 16423 if (BPF_SRC(insn->code) != BPF_K || 16424 insn->imm != 0 || 16425 insn->src_reg != BPF_REG_0 || 16426 insn->dst_reg != BPF_REG_0 || 16427 class == BPF_JMP32) { 16428 verbose(env, "BPF_EXIT uses reserved fields\n"); 16429 return -EINVAL; 16430 } 16431 16432 if (env->cur_state->active_lock.ptr && 16433 !in_rbtree_lock_required_cb(env)) { 16434 verbose(env, "bpf_spin_unlock is missing\n"); 16435 return -EINVAL; 16436 } 16437 16438 if (env->cur_state->active_rcu_lock) { 16439 verbose(env, "bpf_rcu_read_unlock is missing\n"); 16440 return -EINVAL; 16441 } 16442 16443 /* We must do check_reference_leak here before 16444 * prepare_func_exit to handle the case when 16445 * state->curframe > 0, it may be a callback 16446 * function, for which reference_state must 16447 * match caller reference state when it exits. 16448 */ 16449 err = check_reference_leak(env); 16450 if (err) 16451 return err; 16452 16453 if (state->curframe) { 16454 /* exit from nested function */ 16455 err = prepare_func_exit(env, &env->insn_idx); 16456 if (err) 16457 return err; 16458 do_print_state = true; 16459 continue; 16460 } 16461 16462 err = check_return_code(env); 16463 if (err) 16464 return err; 16465 process_bpf_exit: 16466 mark_verifier_state_scratched(env); 16467 update_branch_counts(env, env->cur_state); 16468 err = pop_stack(env, &prev_insn_idx, 16469 &env->insn_idx, pop_log); 16470 if (err < 0) { 16471 if (err != -ENOENT) 16472 return err; 16473 break; 16474 } else { 16475 do_print_state = true; 16476 continue; 16477 } 16478 } else { 16479 err = check_cond_jmp_op(env, insn, &env->insn_idx); 16480 if (err) 16481 return err; 16482 } 16483 } else if (class == BPF_LD) { 16484 u8 mode = BPF_MODE(insn->code); 16485 16486 if (mode == BPF_ABS || mode == BPF_IND) { 16487 err = check_ld_abs(env, insn); 16488 if (err) 16489 return err; 16490 16491 } else if (mode == BPF_IMM) { 16492 err = check_ld_imm(env, insn); 16493 if (err) 16494 return err; 16495 16496 env->insn_idx++; 16497 sanitize_mark_insn_seen(env); 16498 } else { 16499 verbose(env, "invalid BPF_LD mode\n"); 16500 return -EINVAL; 16501 } 16502 } else { 16503 verbose(env, "unknown insn class %d\n", class); 16504 return -EINVAL; 16505 } 16506 16507 env->insn_idx++; 16508 } 16509 16510 return 0; 16511 } 16512 16513 static int find_btf_percpu_datasec(struct btf *btf) 16514 { 16515 const struct btf_type *t; 16516 const char *tname; 16517 int i, n; 16518 16519 /* 16520 * Both vmlinux and module each have their own ".data..percpu" 16521 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF 16522 * types to look at only module's own BTF types. 16523 */ 16524 n = btf_nr_types(btf); 16525 if (btf_is_module(btf)) 16526 i = btf_nr_types(btf_vmlinux); 16527 else 16528 i = 1; 16529 16530 for(; i < n; i++) { 16531 t = btf_type_by_id(btf, i); 16532 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC) 16533 continue; 16534 16535 tname = btf_name_by_offset(btf, t->name_off); 16536 if (!strcmp(tname, ".data..percpu")) 16537 return i; 16538 } 16539 16540 return -ENOENT; 16541 } 16542 16543 /* replace pseudo btf_id with kernel symbol address */ 16544 static int check_pseudo_btf_id(struct bpf_verifier_env *env, 16545 struct bpf_insn *insn, 16546 struct bpf_insn_aux_data *aux) 16547 { 16548 const struct btf_var_secinfo *vsi; 16549 const struct btf_type *datasec; 16550 struct btf_mod_pair *btf_mod; 16551 const struct btf_type *t; 16552 const char *sym_name; 16553 bool percpu = false; 16554 u32 type, id = insn->imm; 16555 struct btf *btf; 16556 s32 datasec_id; 16557 u64 addr; 16558 int i, btf_fd, err; 16559 16560 btf_fd = insn[1].imm; 16561 if (btf_fd) { 16562 btf = btf_get_by_fd(btf_fd); 16563 if (IS_ERR(btf)) { 16564 verbose(env, "invalid module BTF object FD specified.\n"); 16565 return -EINVAL; 16566 } 16567 } else { 16568 if (!btf_vmlinux) { 16569 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n"); 16570 return -EINVAL; 16571 } 16572 btf = btf_vmlinux; 16573 btf_get(btf); 16574 } 16575 16576 t = btf_type_by_id(btf, id); 16577 if (!t) { 16578 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id); 16579 err = -ENOENT; 16580 goto err_put; 16581 } 16582 16583 if (!btf_type_is_var(t) && !btf_type_is_func(t)) { 16584 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id); 16585 err = -EINVAL; 16586 goto err_put; 16587 } 16588 16589 sym_name = btf_name_by_offset(btf, t->name_off); 16590 addr = kallsyms_lookup_name(sym_name); 16591 if (!addr) { 16592 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n", 16593 sym_name); 16594 err = -ENOENT; 16595 goto err_put; 16596 } 16597 insn[0].imm = (u32)addr; 16598 insn[1].imm = addr >> 32; 16599 16600 if (btf_type_is_func(t)) { 16601 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY; 16602 aux->btf_var.mem_size = 0; 16603 goto check_btf; 16604 } 16605 16606 datasec_id = find_btf_percpu_datasec(btf); 16607 if (datasec_id > 0) { 16608 datasec = btf_type_by_id(btf, datasec_id); 16609 for_each_vsi(i, datasec, vsi) { 16610 if (vsi->type == id) { 16611 percpu = true; 16612 break; 16613 } 16614 } 16615 } 16616 16617 type = t->type; 16618 t = btf_type_skip_modifiers(btf, type, NULL); 16619 if (percpu) { 16620 aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU; 16621 aux->btf_var.btf = btf; 16622 aux->btf_var.btf_id = type; 16623 } else if (!btf_type_is_struct(t)) { 16624 const struct btf_type *ret; 16625 const char *tname; 16626 u32 tsize; 16627 16628 /* resolve the type size of ksym. */ 16629 ret = btf_resolve_size(btf, t, &tsize); 16630 if (IS_ERR(ret)) { 16631 tname = btf_name_by_offset(btf, t->name_off); 16632 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n", 16633 tname, PTR_ERR(ret)); 16634 err = -EINVAL; 16635 goto err_put; 16636 } 16637 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY; 16638 aux->btf_var.mem_size = tsize; 16639 } else { 16640 aux->btf_var.reg_type = PTR_TO_BTF_ID; 16641 aux->btf_var.btf = btf; 16642 aux->btf_var.btf_id = type; 16643 } 16644 check_btf: 16645 /* check whether we recorded this BTF (and maybe module) already */ 16646 for (i = 0; i < env->used_btf_cnt; i++) { 16647 if (env->used_btfs[i].btf == btf) { 16648 btf_put(btf); 16649 return 0; 16650 } 16651 } 16652 16653 if (env->used_btf_cnt >= MAX_USED_BTFS) { 16654 err = -E2BIG; 16655 goto err_put; 16656 } 16657 16658 btf_mod = &env->used_btfs[env->used_btf_cnt]; 16659 btf_mod->btf = btf; 16660 btf_mod->module = NULL; 16661 16662 /* if we reference variables from kernel module, bump its refcount */ 16663 if (btf_is_module(btf)) { 16664 btf_mod->module = btf_try_get_module(btf); 16665 if (!btf_mod->module) { 16666 err = -ENXIO; 16667 goto err_put; 16668 } 16669 } 16670 16671 env->used_btf_cnt++; 16672 16673 return 0; 16674 err_put: 16675 btf_put(btf); 16676 return err; 16677 } 16678 16679 static bool is_tracing_prog_type(enum bpf_prog_type type) 16680 { 16681 switch (type) { 16682 case BPF_PROG_TYPE_KPROBE: 16683 case BPF_PROG_TYPE_TRACEPOINT: 16684 case BPF_PROG_TYPE_PERF_EVENT: 16685 case BPF_PROG_TYPE_RAW_TRACEPOINT: 16686 case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE: 16687 return true; 16688 default: 16689 return false; 16690 } 16691 } 16692 16693 static int check_map_prog_compatibility(struct bpf_verifier_env *env, 16694 struct bpf_map *map, 16695 struct bpf_prog *prog) 16696 16697 { 16698 enum bpf_prog_type prog_type = resolve_prog_type(prog); 16699 16700 if (btf_record_has_field(map->record, BPF_LIST_HEAD) || 16701 btf_record_has_field(map->record, BPF_RB_ROOT)) { 16702 if (is_tracing_prog_type(prog_type)) { 16703 verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n"); 16704 return -EINVAL; 16705 } 16706 } 16707 16708 if (btf_record_has_field(map->record, BPF_SPIN_LOCK)) { 16709 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) { 16710 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n"); 16711 return -EINVAL; 16712 } 16713 16714 if (is_tracing_prog_type(prog_type)) { 16715 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n"); 16716 return -EINVAL; 16717 } 16718 16719 if (prog->aux->sleepable) { 16720 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n"); 16721 return -EINVAL; 16722 } 16723 } 16724 16725 if (btf_record_has_field(map->record, BPF_TIMER)) { 16726 if (is_tracing_prog_type(prog_type)) { 16727 verbose(env, "tracing progs cannot use bpf_timer yet\n"); 16728 return -EINVAL; 16729 } 16730 } 16731 16732 if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) && 16733 !bpf_offload_prog_map_match(prog, map)) { 16734 verbose(env, "offload device mismatch between prog and map\n"); 16735 return -EINVAL; 16736 } 16737 16738 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) { 16739 verbose(env, "bpf_struct_ops map cannot be used in prog\n"); 16740 return -EINVAL; 16741 } 16742 16743 if (prog->aux->sleepable) 16744 switch (map->map_type) { 16745 case BPF_MAP_TYPE_HASH: 16746 case BPF_MAP_TYPE_LRU_HASH: 16747 case BPF_MAP_TYPE_ARRAY: 16748 case BPF_MAP_TYPE_PERCPU_HASH: 16749 case BPF_MAP_TYPE_PERCPU_ARRAY: 16750 case BPF_MAP_TYPE_LRU_PERCPU_HASH: 16751 case BPF_MAP_TYPE_ARRAY_OF_MAPS: 16752 case BPF_MAP_TYPE_HASH_OF_MAPS: 16753 case BPF_MAP_TYPE_RINGBUF: 16754 case BPF_MAP_TYPE_USER_RINGBUF: 16755 case BPF_MAP_TYPE_INODE_STORAGE: 16756 case BPF_MAP_TYPE_SK_STORAGE: 16757 case BPF_MAP_TYPE_TASK_STORAGE: 16758 case BPF_MAP_TYPE_CGRP_STORAGE: 16759 break; 16760 default: 16761 verbose(env, 16762 "Sleepable programs can only use array, hash, ringbuf and local storage maps\n"); 16763 return -EINVAL; 16764 } 16765 16766 return 0; 16767 } 16768 16769 static bool bpf_map_is_cgroup_storage(struct bpf_map *map) 16770 { 16771 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE || 16772 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE); 16773 } 16774 16775 /* find and rewrite pseudo imm in ld_imm64 instructions: 16776 * 16777 * 1. if it accesses map FD, replace it with actual map pointer. 16778 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var. 16779 * 16780 * NOTE: btf_vmlinux is required for converting pseudo btf_id. 16781 */ 16782 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env) 16783 { 16784 struct bpf_insn *insn = env->prog->insnsi; 16785 int insn_cnt = env->prog->len; 16786 int i, j, err; 16787 16788 err = bpf_prog_calc_tag(env->prog); 16789 if (err) 16790 return err; 16791 16792 for (i = 0; i < insn_cnt; i++, insn++) { 16793 if (BPF_CLASS(insn->code) == BPF_LDX && 16794 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) { 16795 verbose(env, "BPF_LDX uses reserved fields\n"); 16796 return -EINVAL; 16797 } 16798 16799 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) { 16800 struct bpf_insn_aux_data *aux; 16801 struct bpf_map *map; 16802 struct fd f; 16803 u64 addr; 16804 u32 fd; 16805 16806 if (i == insn_cnt - 1 || insn[1].code != 0 || 16807 insn[1].dst_reg != 0 || insn[1].src_reg != 0 || 16808 insn[1].off != 0) { 16809 verbose(env, "invalid bpf_ld_imm64 insn\n"); 16810 return -EINVAL; 16811 } 16812 16813 if (insn[0].src_reg == 0) 16814 /* valid generic load 64-bit imm */ 16815 goto next_insn; 16816 16817 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) { 16818 aux = &env->insn_aux_data[i]; 16819 err = check_pseudo_btf_id(env, insn, aux); 16820 if (err) 16821 return err; 16822 goto next_insn; 16823 } 16824 16825 if (insn[0].src_reg == BPF_PSEUDO_FUNC) { 16826 aux = &env->insn_aux_data[i]; 16827 aux->ptr_type = PTR_TO_FUNC; 16828 goto next_insn; 16829 } 16830 16831 /* In final convert_pseudo_ld_imm64() step, this is 16832 * converted into regular 64-bit imm load insn. 16833 */ 16834 switch (insn[0].src_reg) { 16835 case BPF_PSEUDO_MAP_VALUE: 16836 case BPF_PSEUDO_MAP_IDX_VALUE: 16837 break; 16838 case BPF_PSEUDO_MAP_FD: 16839 case BPF_PSEUDO_MAP_IDX: 16840 if (insn[1].imm == 0) 16841 break; 16842 fallthrough; 16843 default: 16844 verbose(env, "unrecognized bpf_ld_imm64 insn\n"); 16845 return -EINVAL; 16846 } 16847 16848 switch (insn[0].src_reg) { 16849 case BPF_PSEUDO_MAP_IDX_VALUE: 16850 case BPF_PSEUDO_MAP_IDX: 16851 if (bpfptr_is_null(env->fd_array)) { 16852 verbose(env, "fd_idx without fd_array is invalid\n"); 16853 return -EPROTO; 16854 } 16855 if (copy_from_bpfptr_offset(&fd, env->fd_array, 16856 insn[0].imm * sizeof(fd), 16857 sizeof(fd))) 16858 return -EFAULT; 16859 break; 16860 default: 16861 fd = insn[0].imm; 16862 break; 16863 } 16864 16865 f = fdget(fd); 16866 map = __bpf_map_get(f); 16867 if (IS_ERR(map)) { 16868 verbose(env, "fd %d is not pointing to valid bpf_map\n", 16869 insn[0].imm); 16870 return PTR_ERR(map); 16871 } 16872 16873 err = check_map_prog_compatibility(env, map, env->prog); 16874 if (err) { 16875 fdput(f); 16876 return err; 16877 } 16878 16879 aux = &env->insn_aux_data[i]; 16880 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD || 16881 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) { 16882 addr = (unsigned long)map; 16883 } else { 16884 u32 off = insn[1].imm; 16885 16886 if (off >= BPF_MAX_VAR_OFF) { 16887 verbose(env, "direct value offset of %u is not allowed\n", off); 16888 fdput(f); 16889 return -EINVAL; 16890 } 16891 16892 if (!map->ops->map_direct_value_addr) { 16893 verbose(env, "no direct value access support for this map type\n"); 16894 fdput(f); 16895 return -EINVAL; 16896 } 16897 16898 err = map->ops->map_direct_value_addr(map, &addr, off); 16899 if (err) { 16900 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n", 16901 map->value_size, off); 16902 fdput(f); 16903 return err; 16904 } 16905 16906 aux->map_off = off; 16907 addr += off; 16908 } 16909 16910 insn[0].imm = (u32)addr; 16911 insn[1].imm = addr >> 32; 16912 16913 /* check whether we recorded this map already */ 16914 for (j = 0; j < env->used_map_cnt; j++) { 16915 if (env->used_maps[j] == map) { 16916 aux->map_index = j; 16917 fdput(f); 16918 goto next_insn; 16919 } 16920 } 16921 16922 if (env->used_map_cnt >= MAX_USED_MAPS) { 16923 fdput(f); 16924 return -E2BIG; 16925 } 16926 16927 /* hold the map. If the program is rejected by verifier, 16928 * the map will be released by release_maps() or it 16929 * will be used by the valid program until it's unloaded 16930 * and all maps are released in free_used_maps() 16931 */ 16932 bpf_map_inc(map); 16933 16934 aux->map_index = env->used_map_cnt; 16935 env->used_maps[env->used_map_cnt++] = map; 16936 16937 if (bpf_map_is_cgroup_storage(map) && 16938 bpf_cgroup_storage_assign(env->prog->aux, map)) { 16939 verbose(env, "only one cgroup storage of each type is allowed\n"); 16940 fdput(f); 16941 return -EBUSY; 16942 } 16943 16944 fdput(f); 16945 next_insn: 16946 insn++; 16947 i++; 16948 continue; 16949 } 16950 16951 /* Basic sanity check before we invest more work here. */ 16952 if (!bpf_opcode_in_insntable(insn->code)) { 16953 verbose(env, "unknown opcode %02x\n", insn->code); 16954 return -EINVAL; 16955 } 16956 } 16957 16958 /* now all pseudo BPF_LD_IMM64 instructions load valid 16959 * 'struct bpf_map *' into a register instead of user map_fd. 16960 * These pointers will be used later by verifier to validate map access. 16961 */ 16962 return 0; 16963 } 16964 16965 /* drop refcnt of maps used by the rejected program */ 16966 static void release_maps(struct bpf_verifier_env *env) 16967 { 16968 __bpf_free_used_maps(env->prog->aux, env->used_maps, 16969 env->used_map_cnt); 16970 } 16971 16972 /* drop refcnt of maps used by the rejected program */ 16973 static void release_btfs(struct bpf_verifier_env *env) 16974 { 16975 __bpf_free_used_btfs(env->prog->aux, env->used_btfs, 16976 env->used_btf_cnt); 16977 } 16978 16979 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */ 16980 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env) 16981 { 16982 struct bpf_insn *insn = env->prog->insnsi; 16983 int insn_cnt = env->prog->len; 16984 int i; 16985 16986 for (i = 0; i < insn_cnt; i++, insn++) { 16987 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW)) 16988 continue; 16989 if (insn->src_reg == BPF_PSEUDO_FUNC) 16990 continue; 16991 insn->src_reg = 0; 16992 } 16993 } 16994 16995 /* single env->prog->insni[off] instruction was replaced with the range 16996 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying 16997 * [0, off) and [off, end) to new locations, so the patched range stays zero 16998 */ 16999 static void adjust_insn_aux_data(struct bpf_verifier_env *env, 17000 struct bpf_insn_aux_data *new_data, 17001 struct bpf_prog *new_prog, u32 off, u32 cnt) 17002 { 17003 struct bpf_insn_aux_data *old_data = env->insn_aux_data; 17004 struct bpf_insn *insn = new_prog->insnsi; 17005 u32 old_seen = old_data[off].seen; 17006 u32 prog_len; 17007 int i; 17008 17009 /* aux info at OFF always needs adjustment, no matter fast path 17010 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the 17011 * original insn at old prog. 17012 */ 17013 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1); 17014 17015 if (cnt == 1) 17016 return; 17017 prog_len = new_prog->len; 17018 17019 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off); 17020 memcpy(new_data + off + cnt - 1, old_data + off, 17021 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1)); 17022 for (i = off; i < off + cnt - 1; i++) { 17023 /* Expand insni[off]'s seen count to the patched range. */ 17024 new_data[i].seen = old_seen; 17025 new_data[i].zext_dst = insn_has_def32(env, insn + i); 17026 } 17027 env->insn_aux_data = new_data; 17028 vfree(old_data); 17029 } 17030 17031 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len) 17032 { 17033 int i; 17034 17035 if (len == 1) 17036 return; 17037 /* NOTE: fake 'exit' subprog should be updated as well. */ 17038 for (i = 0; i <= env->subprog_cnt; i++) { 17039 if (env->subprog_info[i].start <= off) 17040 continue; 17041 env->subprog_info[i].start += len - 1; 17042 } 17043 } 17044 17045 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len) 17046 { 17047 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 17048 int i, sz = prog->aux->size_poke_tab; 17049 struct bpf_jit_poke_descriptor *desc; 17050 17051 for (i = 0; i < sz; i++) { 17052 desc = &tab[i]; 17053 if (desc->insn_idx <= off) 17054 continue; 17055 desc->insn_idx += len - 1; 17056 } 17057 } 17058 17059 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off, 17060 const struct bpf_insn *patch, u32 len) 17061 { 17062 struct bpf_prog *new_prog; 17063 struct bpf_insn_aux_data *new_data = NULL; 17064 17065 if (len > 1) { 17066 new_data = vzalloc(array_size(env->prog->len + len - 1, 17067 sizeof(struct bpf_insn_aux_data))); 17068 if (!new_data) 17069 return NULL; 17070 } 17071 17072 new_prog = bpf_patch_insn_single(env->prog, off, patch, len); 17073 if (IS_ERR(new_prog)) { 17074 if (PTR_ERR(new_prog) == -ERANGE) 17075 verbose(env, 17076 "insn %d cannot be patched due to 16-bit range\n", 17077 env->insn_aux_data[off].orig_idx); 17078 vfree(new_data); 17079 return NULL; 17080 } 17081 adjust_insn_aux_data(env, new_data, new_prog, off, len); 17082 adjust_subprog_starts(env, off, len); 17083 adjust_poke_descs(new_prog, off, len); 17084 return new_prog; 17085 } 17086 17087 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env, 17088 u32 off, u32 cnt) 17089 { 17090 int i, j; 17091 17092 /* find first prog starting at or after off (first to remove) */ 17093 for (i = 0; i < env->subprog_cnt; i++) 17094 if (env->subprog_info[i].start >= off) 17095 break; 17096 /* find first prog starting at or after off + cnt (first to stay) */ 17097 for (j = i; j < env->subprog_cnt; j++) 17098 if (env->subprog_info[j].start >= off + cnt) 17099 break; 17100 /* if j doesn't start exactly at off + cnt, we are just removing 17101 * the front of previous prog 17102 */ 17103 if (env->subprog_info[j].start != off + cnt) 17104 j--; 17105 17106 if (j > i) { 17107 struct bpf_prog_aux *aux = env->prog->aux; 17108 int move; 17109 17110 /* move fake 'exit' subprog as well */ 17111 move = env->subprog_cnt + 1 - j; 17112 17113 memmove(env->subprog_info + i, 17114 env->subprog_info + j, 17115 sizeof(*env->subprog_info) * move); 17116 env->subprog_cnt -= j - i; 17117 17118 /* remove func_info */ 17119 if (aux->func_info) { 17120 move = aux->func_info_cnt - j; 17121 17122 memmove(aux->func_info + i, 17123 aux->func_info + j, 17124 sizeof(*aux->func_info) * move); 17125 aux->func_info_cnt -= j - i; 17126 /* func_info->insn_off is set after all code rewrites, 17127 * in adjust_btf_func() - no need to adjust 17128 */ 17129 } 17130 } else { 17131 /* convert i from "first prog to remove" to "first to adjust" */ 17132 if (env->subprog_info[i].start == off) 17133 i++; 17134 } 17135 17136 /* update fake 'exit' subprog as well */ 17137 for (; i <= env->subprog_cnt; i++) 17138 env->subprog_info[i].start -= cnt; 17139 17140 return 0; 17141 } 17142 17143 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off, 17144 u32 cnt) 17145 { 17146 struct bpf_prog *prog = env->prog; 17147 u32 i, l_off, l_cnt, nr_linfo; 17148 struct bpf_line_info *linfo; 17149 17150 nr_linfo = prog->aux->nr_linfo; 17151 if (!nr_linfo) 17152 return 0; 17153 17154 linfo = prog->aux->linfo; 17155 17156 /* find first line info to remove, count lines to be removed */ 17157 for (i = 0; i < nr_linfo; i++) 17158 if (linfo[i].insn_off >= off) 17159 break; 17160 17161 l_off = i; 17162 l_cnt = 0; 17163 for (; i < nr_linfo; i++) 17164 if (linfo[i].insn_off < off + cnt) 17165 l_cnt++; 17166 else 17167 break; 17168 17169 /* First live insn doesn't match first live linfo, it needs to "inherit" 17170 * last removed linfo. prog is already modified, so prog->len == off 17171 * means no live instructions after (tail of the program was removed). 17172 */ 17173 if (prog->len != off && l_cnt && 17174 (i == nr_linfo || linfo[i].insn_off != off + cnt)) { 17175 l_cnt--; 17176 linfo[--i].insn_off = off + cnt; 17177 } 17178 17179 /* remove the line info which refer to the removed instructions */ 17180 if (l_cnt) { 17181 memmove(linfo + l_off, linfo + i, 17182 sizeof(*linfo) * (nr_linfo - i)); 17183 17184 prog->aux->nr_linfo -= l_cnt; 17185 nr_linfo = prog->aux->nr_linfo; 17186 } 17187 17188 /* pull all linfo[i].insn_off >= off + cnt in by cnt */ 17189 for (i = l_off; i < nr_linfo; i++) 17190 linfo[i].insn_off -= cnt; 17191 17192 /* fix up all subprogs (incl. 'exit') which start >= off */ 17193 for (i = 0; i <= env->subprog_cnt; i++) 17194 if (env->subprog_info[i].linfo_idx > l_off) { 17195 /* program may have started in the removed region but 17196 * may not be fully removed 17197 */ 17198 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt) 17199 env->subprog_info[i].linfo_idx -= l_cnt; 17200 else 17201 env->subprog_info[i].linfo_idx = l_off; 17202 } 17203 17204 return 0; 17205 } 17206 17207 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt) 17208 { 17209 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 17210 unsigned int orig_prog_len = env->prog->len; 17211 int err; 17212 17213 if (bpf_prog_is_offloaded(env->prog->aux)) 17214 bpf_prog_offload_remove_insns(env, off, cnt); 17215 17216 err = bpf_remove_insns(env->prog, off, cnt); 17217 if (err) 17218 return err; 17219 17220 err = adjust_subprog_starts_after_remove(env, off, cnt); 17221 if (err) 17222 return err; 17223 17224 err = bpf_adj_linfo_after_remove(env, off, cnt); 17225 if (err) 17226 return err; 17227 17228 memmove(aux_data + off, aux_data + off + cnt, 17229 sizeof(*aux_data) * (orig_prog_len - off - cnt)); 17230 17231 return 0; 17232 } 17233 17234 /* The verifier does more data flow analysis than llvm and will not 17235 * explore branches that are dead at run time. Malicious programs can 17236 * have dead code too. Therefore replace all dead at-run-time code 17237 * with 'ja -1'. 17238 * 17239 * Just nops are not optimal, e.g. if they would sit at the end of the 17240 * program and through another bug we would manage to jump there, then 17241 * we'd execute beyond program memory otherwise. Returning exception 17242 * code also wouldn't work since we can have subprogs where the dead 17243 * code could be located. 17244 */ 17245 static void sanitize_dead_code(struct bpf_verifier_env *env) 17246 { 17247 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 17248 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1); 17249 struct bpf_insn *insn = env->prog->insnsi; 17250 const int insn_cnt = env->prog->len; 17251 int i; 17252 17253 for (i = 0; i < insn_cnt; i++) { 17254 if (aux_data[i].seen) 17255 continue; 17256 memcpy(insn + i, &trap, sizeof(trap)); 17257 aux_data[i].zext_dst = false; 17258 } 17259 } 17260 17261 static bool insn_is_cond_jump(u8 code) 17262 { 17263 u8 op; 17264 17265 if (BPF_CLASS(code) == BPF_JMP32) 17266 return true; 17267 17268 if (BPF_CLASS(code) != BPF_JMP) 17269 return false; 17270 17271 op = BPF_OP(code); 17272 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL; 17273 } 17274 17275 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env) 17276 { 17277 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 17278 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0); 17279 struct bpf_insn *insn = env->prog->insnsi; 17280 const int insn_cnt = env->prog->len; 17281 int i; 17282 17283 for (i = 0; i < insn_cnt; i++, insn++) { 17284 if (!insn_is_cond_jump(insn->code)) 17285 continue; 17286 17287 if (!aux_data[i + 1].seen) 17288 ja.off = insn->off; 17289 else if (!aux_data[i + 1 + insn->off].seen) 17290 ja.off = 0; 17291 else 17292 continue; 17293 17294 if (bpf_prog_is_offloaded(env->prog->aux)) 17295 bpf_prog_offload_replace_insn(env, i, &ja); 17296 17297 memcpy(insn, &ja, sizeof(ja)); 17298 } 17299 } 17300 17301 static int opt_remove_dead_code(struct bpf_verifier_env *env) 17302 { 17303 struct bpf_insn_aux_data *aux_data = env->insn_aux_data; 17304 int insn_cnt = env->prog->len; 17305 int i, err; 17306 17307 for (i = 0; i < insn_cnt; i++) { 17308 int j; 17309 17310 j = 0; 17311 while (i + j < insn_cnt && !aux_data[i + j].seen) 17312 j++; 17313 if (!j) 17314 continue; 17315 17316 err = verifier_remove_insns(env, i, j); 17317 if (err) 17318 return err; 17319 insn_cnt = env->prog->len; 17320 } 17321 17322 return 0; 17323 } 17324 17325 static int opt_remove_nops(struct bpf_verifier_env *env) 17326 { 17327 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0); 17328 struct bpf_insn *insn = env->prog->insnsi; 17329 int insn_cnt = env->prog->len; 17330 int i, err; 17331 17332 for (i = 0; i < insn_cnt; i++) { 17333 if (memcmp(&insn[i], &ja, sizeof(ja))) 17334 continue; 17335 17336 err = verifier_remove_insns(env, i, 1); 17337 if (err) 17338 return err; 17339 insn_cnt--; 17340 i--; 17341 } 17342 17343 return 0; 17344 } 17345 17346 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env, 17347 const union bpf_attr *attr) 17348 { 17349 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4]; 17350 struct bpf_insn_aux_data *aux = env->insn_aux_data; 17351 int i, patch_len, delta = 0, len = env->prog->len; 17352 struct bpf_insn *insns = env->prog->insnsi; 17353 struct bpf_prog *new_prog; 17354 bool rnd_hi32; 17355 17356 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32; 17357 zext_patch[1] = BPF_ZEXT_REG(0); 17358 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0); 17359 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 17360 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX); 17361 for (i = 0; i < len; i++) { 17362 int adj_idx = i + delta; 17363 struct bpf_insn insn; 17364 int load_reg; 17365 17366 insn = insns[adj_idx]; 17367 load_reg = insn_def_regno(&insn); 17368 if (!aux[adj_idx].zext_dst) { 17369 u8 code, class; 17370 u32 imm_rnd; 17371 17372 if (!rnd_hi32) 17373 continue; 17374 17375 code = insn.code; 17376 class = BPF_CLASS(code); 17377 if (load_reg == -1) 17378 continue; 17379 17380 /* NOTE: arg "reg" (the fourth one) is only used for 17381 * BPF_STX + SRC_OP, so it is safe to pass NULL 17382 * here. 17383 */ 17384 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) { 17385 if (class == BPF_LD && 17386 BPF_MODE(code) == BPF_IMM) 17387 i++; 17388 continue; 17389 } 17390 17391 /* ctx load could be transformed into wider load. */ 17392 if (class == BPF_LDX && 17393 aux[adj_idx].ptr_type == PTR_TO_CTX) 17394 continue; 17395 17396 imm_rnd = get_random_u32(); 17397 rnd_hi32_patch[0] = insn; 17398 rnd_hi32_patch[1].imm = imm_rnd; 17399 rnd_hi32_patch[3].dst_reg = load_reg; 17400 patch = rnd_hi32_patch; 17401 patch_len = 4; 17402 goto apply_patch_buffer; 17403 } 17404 17405 /* Add in an zero-extend instruction if a) the JIT has requested 17406 * it or b) it's a CMPXCHG. 17407 * 17408 * The latter is because: BPF_CMPXCHG always loads a value into 17409 * R0, therefore always zero-extends. However some archs' 17410 * equivalent instruction only does this load when the 17411 * comparison is successful. This detail of CMPXCHG is 17412 * orthogonal to the general zero-extension behaviour of the 17413 * CPU, so it's treated independently of bpf_jit_needs_zext. 17414 */ 17415 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn)) 17416 continue; 17417 17418 /* Zero-extension is done by the caller. */ 17419 if (bpf_pseudo_kfunc_call(&insn)) 17420 continue; 17421 17422 if (WARN_ON(load_reg == -1)) { 17423 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n"); 17424 return -EFAULT; 17425 } 17426 17427 zext_patch[0] = insn; 17428 zext_patch[1].dst_reg = load_reg; 17429 zext_patch[1].src_reg = load_reg; 17430 patch = zext_patch; 17431 patch_len = 2; 17432 apply_patch_buffer: 17433 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len); 17434 if (!new_prog) 17435 return -ENOMEM; 17436 env->prog = new_prog; 17437 insns = new_prog->insnsi; 17438 aux = env->insn_aux_data; 17439 delta += patch_len - 1; 17440 } 17441 17442 return 0; 17443 } 17444 17445 /* convert load instructions that access fields of a context type into a 17446 * sequence of instructions that access fields of the underlying structure: 17447 * struct __sk_buff -> struct sk_buff 17448 * struct bpf_sock_ops -> struct sock 17449 */ 17450 static int convert_ctx_accesses(struct bpf_verifier_env *env) 17451 { 17452 const struct bpf_verifier_ops *ops = env->ops; 17453 int i, cnt, size, ctx_field_size, delta = 0; 17454 const int insn_cnt = env->prog->len; 17455 struct bpf_insn insn_buf[16], *insn; 17456 u32 target_size, size_default, off; 17457 struct bpf_prog *new_prog; 17458 enum bpf_access_type type; 17459 bool is_narrower_load; 17460 17461 if (ops->gen_prologue || env->seen_direct_write) { 17462 if (!ops->gen_prologue) { 17463 verbose(env, "bpf verifier is misconfigured\n"); 17464 return -EINVAL; 17465 } 17466 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write, 17467 env->prog); 17468 if (cnt >= ARRAY_SIZE(insn_buf)) { 17469 verbose(env, "bpf verifier is misconfigured\n"); 17470 return -EINVAL; 17471 } else if (cnt) { 17472 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt); 17473 if (!new_prog) 17474 return -ENOMEM; 17475 17476 env->prog = new_prog; 17477 delta += cnt - 1; 17478 } 17479 } 17480 17481 if (bpf_prog_is_offloaded(env->prog->aux)) 17482 return 0; 17483 17484 insn = env->prog->insnsi + delta; 17485 17486 for (i = 0; i < insn_cnt; i++, insn++) { 17487 bpf_convert_ctx_access_t convert_ctx_access; 17488 17489 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) || 17490 insn->code == (BPF_LDX | BPF_MEM | BPF_H) || 17491 insn->code == (BPF_LDX | BPF_MEM | BPF_W) || 17492 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) { 17493 type = BPF_READ; 17494 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) || 17495 insn->code == (BPF_STX | BPF_MEM | BPF_H) || 17496 insn->code == (BPF_STX | BPF_MEM | BPF_W) || 17497 insn->code == (BPF_STX | BPF_MEM | BPF_DW) || 17498 insn->code == (BPF_ST | BPF_MEM | BPF_B) || 17499 insn->code == (BPF_ST | BPF_MEM | BPF_H) || 17500 insn->code == (BPF_ST | BPF_MEM | BPF_W) || 17501 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) { 17502 type = BPF_WRITE; 17503 } else { 17504 continue; 17505 } 17506 17507 if (type == BPF_WRITE && 17508 env->insn_aux_data[i + delta].sanitize_stack_spill) { 17509 struct bpf_insn patch[] = { 17510 *insn, 17511 BPF_ST_NOSPEC(), 17512 }; 17513 17514 cnt = ARRAY_SIZE(patch); 17515 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt); 17516 if (!new_prog) 17517 return -ENOMEM; 17518 17519 delta += cnt - 1; 17520 env->prog = new_prog; 17521 insn = new_prog->insnsi + i + delta; 17522 continue; 17523 } 17524 17525 switch ((int)env->insn_aux_data[i + delta].ptr_type) { 17526 case PTR_TO_CTX: 17527 if (!ops->convert_ctx_access) 17528 continue; 17529 convert_ctx_access = ops->convert_ctx_access; 17530 break; 17531 case PTR_TO_SOCKET: 17532 case PTR_TO_SOCK_COMMON: 17533 convert_ctx_access = bpf_sock_convert_ctx_access; 17534 break; 17535 case PTR_TO_TCP_SOCK: 17536 convert_ctx_access = bpf_tcp_sock_convert_ctx_access; 17537 break; 17538 case PTR_TO_XDP_SOCK: 17539 convert_ctx_access = bpf_xdp_sock_convert_ctx_access; 17540 break; 17541 case PTR_TO_BTF_ID: 17542 case PTR_TO_BTF_ID | PTR_UNTRUSTED: 17543 /* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike 17544 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot 17545 * be said once it is marked PTR_UNTRUSTED, hence we must handle 17546 * any faults for loads into such types. BPF_WRITE is disallowed 17547 * for this case. 17548 */ 17549 case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED: 17550 if (type == BPF_READ) { 17551 insn->code = BPF_LDX | BPF_PROBE_MEM | 17552 BPF_SIZE((insn)->code); 17553 env->prog->aux->num_exentries++; 17554 } 17555 continue; 17556 default: 17557 continue; 17558 } 17559 17560 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size; 17561 size = BPF_LDST_BYTES(insn); 17562 17563 /* If the read access is a narrower load of the field, 17564 * convert to a 4/8-byte load, to minimum program type specific 17565 * convert_ctx_access changes. If conversion is successful, 17566 * we will apply proper mask to the result. 17567 */ 17568 is_narrower_load = size < ctx_field_size; 17569 size_default = bpf_ctx_off_adjust_machine(ctx_field_size); 17570 off = insn->off; 17571 if (is_narrower_load) { 17572 u8 size_code; 17573 17574 if (type == BPF_WRITE) { 17575 verbose(env, "bpf verifier narrow ctx access misconfigured\n"); 17576 return -EINVAL; 17577 } 17578 17579 size_code = BPF_H; 17580 if (ctx_field_size == 4) 17581 size_code = BPF_W; 17582 else if (ctx_field_size == 8) 17583 size_code = BPF_DW; 17584 17585 insn->off = off & ~(size_default - 1); 17586 insn->code = BPF_LDX | BPF_MEM | size_code; 17587 } 17588 17589 target_size = 0; 17590 cnt = convert_ctx_access(type, insn, insn_buf, env->prog, 17591 &target_size); 17592 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) || 17593 (ctx_field_size && !target_size)) { 17594 verbose(env, "bpf verifier is misconfigured\n"); 17595 return -EINVAL; 17596 } 17597 17598 if (is_narrower_load && size < target_size) { 17599 u8 shift = bpf_ctx_narrow_access_offset( 17600 off, size, size_default) * 8; 17601 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) { 17602 verbose(env, "bpf verifier narrow ctx load misconfigured\n"); 17603 return -EINVAL; 17604 } 17605 if (ctx_field_size <= 4) { 17606 if (shift) 17607 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH, 17608 insn->dst_reg, 17609 shift); 17610 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg, 17611 (1 << size * 8) - 1); 17612 } else { 17613 if (shift) 17614 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH, 17615 insn->dst_reg, 17616 shift); 17617 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg, 17618 (1ULL << size * 8) - 1); 17619 } 17620 } 17621 17622 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 17623 if (!new_prog) 17624 return -ENOMEM; 17625 17626 delta += cnt - 1; 17627 17628 /* keep walking new program and skip insns we just inserted */ 17629 env->prog = new_prog; 17630 insn = new_prog->insnsi + i + delta; 17631 } 17632 17633 return 0; 17634 } 17635 17636 static int jit_subprogs(struct bpf_verifier_env *env) 17637 { 17638 struct bpf_prog *prog = env->prog, **func, *tmp; 17639 int i, j, subprog_start, subprog_end = 0, len, subprog; 17640 struct bpf_map *map_ptr; 17641 struct bpf_insn *insn; 17642 void *old_bpf_func; 17643 int err, num_exentries; 17644 17645 if (env->subprog_cnt <= 1) 17646 return 0; 17647 17648 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { 17649 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn)) 17650 continue; 17651 17652 /* Upon error here we cannot fall back to interpreter but 17653 * need a hard reject of the program. Thus -EFAULT is 17654 * propagated in any case. 17655 */ 17656 subprog = find_subprog(env, i + insn->imm + 1); 17657 if (subprog < 0) { 17658 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n", 17659 i + insn->imm + 1); 17660 return -EFAULT; 17661 } 17662 /* temporarily remember subprog id inside insn instead of 17663 * aux_data, since next loop will split up all insns into funcs 17664 */ 17665 insn->off = subprog; 17666 /* remember original imm in case JIT fails and fallback 17667 * to interpreter will be needed 17668 */ 17669 env->insn_aux_data[i].call_imm = insn->imm; 17670 /* point imm to __bpf_call_base+1 from JITs point of view */ 17671 insn->imm = 1; 17672 if (bpf_pseudo_func(insn)) 17673 /* jit (e.g. x86_64) may emit fewer instructions 17674 * if it learns a u32 imm is the same as a u64 imm. 17675 * Force a non zero here. 17676 */ 17677 insn[1].imm = 1; 17678 } 17679 17680 err = bpf_prog_alloc_jited_linfo(prog); 17681 if (err) 17682 goto out_undo_insn; 17683 17684 err = -ENOMEM; 17685 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL); 17686 if (!func) 17687 goto out_undo_insn; 17688 17689 for (i = 0; i < env->subprog_cnt; i++) { 17690 subprog_start = subprog_end; 17691 subprog_end = env->subprog_info[i + 1].start; 17692 17693 len = subprog_end - subprog_start; 17694 /* bpf_prog_run() doesn't call subprogs directly, 17695 * hence main prog stats include the runtime of subprogs. 17696 * subprogs don't have IDs and not reachable via prog_get_next_id 17697 * func[i]->stats will never be accessed and stays NULL 17698 */ 17699 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER); 17700 if (!func[i]) 17701 goto out_free; 17702 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start], 17703 len * sizeof(struct bpf_insn)); 17704 func[i]->type = prog->type; 17705 func[i]->len = len; 17706 if (bpf_prog_calc_tag(func[i])) 17707 goto out_free; 17708 func[i]->is_func = 1; 17709 func[i]->aux->func_idx = i; 17710 /* Below members will be freed only at prog->aux */ 17711 func[i]->aux->btf = prog->aux->btf; 17712 func[i]->aux->func_info = prog->aux->func_info; 17713 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt; 17714 func[i]->aux->poke_tab = prog->aux->poke_tab; 17715 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab; 17716 17717 for (j = 0; j < prog->aux->size_poke_tab; j++) { 17718 struct bpf_jit_poke_descriptor *poke; 17719 17720 poke = &prog->aux->poke_tab[j]; 17721 if (poke->insn_idx < subprog_end && 17722 poke->insn_idx >= subprog_start) 17723 poke->aux = func[i]->aux; 17724 } 17725 17726 func[i]->aux->name[0] = 'F'; 17727 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth; 17728 func[i]->jit_requested = 1; 17729 func[i]->blinding_requested = prog->blinding_requested; 17730 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab; 17731 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab; 17732 func[i]->aux->linfo = prog->aux->linfo; 17733 func[i]->aux->nr_linfo = prog->aux->nr_linfo; 17734 func[i]->aux->jited_linfo = prog->aux->jited_linfo; 17735 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx; 17736 num_exentries = 0; 17737 insn = func[i]->insnsi; 17738 for (j = 0; j < func[i]->len; j++, insn++) { 17739 if (BPF_CLASS(insn->code) == BPF_LDX && 17740 BPF_MODE(insn->code) == BPF_PROBE_MEM) 17741 num_exentries++; 17742 } 17743 func[i]->aux->num_exentries = num_exentries; 17744 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable; 17745 func[i] = bpf_int_jit_compile(func[i]); 17746 if (!func[i]->jited) { 17747 err = -ENOTSUPP; 17748 goto out_free; 17749 } 17750 cond_resched(); 17751 } 17752 17753 /* at this point all bpf functions were successfully JITed 17754 * now populate all bpf_calls with correct addresses and 17755 * run last pass of JIT 17756 */ 17757 for (i = 0; i < env->subprog_cnt; i++) { 17758 insn = func[i]->insnsi; 17759 for (j = 0; j < func[i]->len; j++, insn++) { 17760 if (bpf_pseudo_func(insn)) { 17761 subprog = insn->off; 17762 insn[0].imm = (u32)(long)func[subprog]->bpf_func; 17763 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32; 17764 continue; 17765 } 17766 if (!bpf_pseudo_call(insn)) 17767 continue; 17768 subprog = insn->off; 17769 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func); 17770 } 17771 17772 /* we use the aux data to keep a list of the start addresses 17773 * of the JITed images for each function in the program 17774 * 17775 * for some architectures, such as powerpc64, the imm field 17776 * might not be large enough to hold the offset of the start 17777 * address of the callee's JITed image from __bpf_call_base 17778 * 17779 * in such cases, we can lookup the start address of a callee 17780 * by using its subprog id, available from the off field of 17781 * the call instruction, as an index for this list 17782 */ 17783 func[i]->aux->func = func; 17784 func[i]->aux->func_cnt = env->subprog_cnt; 17785 } 17786 for (i = 0; i < env->subprog_cnt; i++) { 17787 old_bpf_func = func[i]->bpf_func; 17788 tmp = bpf_int_jit_compile(func[i]); 17789 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) { 17790 verbose(env, "JIT doesn't support bpf-to-bpf calls\n"); 17791 err = -ENOTSUPP; 17792 goto out_free; 17793 } 17794 cond_resched(); 17795 } 17796 17797 /* finally lock prog and jit images for all functions and 17798 * populate kallsysm. Begin at the first subprogram, since 17799 * bpf_prog_load will add the kallsyms for the main program. 17800 */ 17801 for (i = 1; i < env->subprog_cnt; i++) { 17802 bpf_prog_lock_ro(func[i]); 17803 bpf_prog_kallsyms_add(func[i]); 17804 } 17805 17806 /* Last step: make now unused interpreter insns from main 17807 * prog consistent for later dump requests, so they can 17808 * later look the same as if they were interpreted only. 17809 */ 17810 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { 17811 if (bpf_pseudo_func(insn)) { 17812 insn[0].imm = env->insn_aux_data[i].call_imm; 17813 insn[1].imm = insn->off; 17814 insn->off = 0; 17815 continue; 17816 } 17817 if (!bpf_pseudo_call(insn)) 17818 continue; 17819 insn->off = env->insn_aux_data[i].call_imm; 17820 subprog = find_subprog(env, i + insn->off + 1); 17821 insn->imm = subprog; 17822 } 17823 17824 prog->jited = 1; 17825 prog->bpf_func = func[0]->bpf_func; 17826 prog->jited_len = func[0]->jited_len; 17827 prog->aux->extable = func[0]->aux->extable; 17828 prog->aux->num_exentries = func[0]->aux->num_exentries; 17829 prog->aux->func = func; 17830 prog->aux->func_cnt = env->subprog_cnt; 17831 bpf_prog_jit_attempt_done(prog); 17832 return 0; 17833 out_free: 17834 /* We failed JIT'ing, so at this point we need to unregister poke 17835 * descriptors from subprogs, so that kernel is not attempting to 17836 * patch it anymore as we're freeing the subprog JIT memory. 17837 */ 17838 for (i = 0; i < prog->aux->size_poke_tab; i++) { 17839 map_ptr = prog->aux->poke_tab[i].tail_call.map; 17840 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux); 17841 } 17842 /* At this point we're guaranteed that poke descriptors are not 17843 * live anymore. We can just unlink its descriptor table as it's 17844 * released with the main prog. 17845 */ 17846 for (i = 0; i < env->subprog_cnt; i++) { 17847 if (!func[i]) 17848 continue; 17849 func[i]->aux->poke_tab = NULL; 17850 bpf_jit_free(func[i]); 17851 } 17852 kfree(func); 17853 out_undo_insn: 17854 /* cleanup main prog to be interpreted */ 17855 prog->jit_requested = 0; 17856 prog->blinding_requested = 0; 17857 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) { 17858 if (!bpf_pseudo_call(insn)) 17859 continue; 17860 insn->off = 0; 17861 insn->imm = env->insn_aux_data[i].call_imm; 17862 } 17863 bpf_prog_jit_attempt_done(prog); 17864 return err; 17865 } 17866 17867 static int fixup_call_args(struct bpf_verifier_env *env) 17868 { 17869 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 17870 struct bpf_prog *prog = env->prog; 17871 struct bpf_insn *insn = prog->insnsi; 17872 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog); 17873 int i, depth; 17874 #endif 17875 int err = 0; 17876 17877 if (env->prog->jit_requested && 17878 !bpf_prog_is_offloaded(env->prog->aux)) { 17879 err = jit_subprogs(env); 17880 if (err == 0) 17881 return 0; 17882 if (err == -EFAULT) 17883 return err; 17884 } 17885 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 17886 if (has_kfunc_call) { 17887 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n"); 17888 return -EINVAL; 17889 } 17890 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) { 17891 /* When JIT fails the progs with bpf2bpf calls and tail_calls 17892 * have to be rejected, since interpreter doesn't support them yet. 17893 */ 17894 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n"); 17895 return -EINVAL; 17896 } 17897 for (i = 0; i < prog->len; i++, insn++) { 17898 if (bpf_pseudo_func(insn)) { 17899 /* When JIT fails the progs with callback calls 17900 * have to be rejected, since interpreter doesn't support them yet. 17901 */ 17902 verbose(env, "callbacks are not allowed in non-JITed programs\n"); 17903 return -EINVAL; 17904 } 17905 17906 if (!bpf_pseudo_call(insn)) 17907 continue; 17908 depth = get_callee_stack_depth(env, insn, i); 17909 if (depth < 0) 17910 return depth; 17911 bpf_patch_call_args(insn, depth); 17912 } 17913 err = 0; 17914 #endif 17915 return err; 17916 } 17917 17918 /* replace a generic kfunc with a specialized version if necessary */ 17919 static void specialize_kfunc(struct bpf_verifier_env *env, 17920 u32 func_id, u16 offset, unsigned long *addr) 17921 { 17922 struct bpf_prog *prog = env->prog; 17923 bool seen_direct_write; 17924 void *xdp_kfunc; 17925 bool is_rdonly; 17926 17927 if (bpf_dev_bound_kfunc_id(func_id)) { 17928 xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id); 17929 if (xdp_kfunc) { 17930 *addr = (unsigned long)xdp_kfunc; 17931 return; 17932 } 17933 /* fallback to default kfunc when not supported by netdev */ 17934 } 17935 17936 if (offset) 17937 return; 17938 17939 if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) { 17940 seen_direct_write = env->seen_direct_write; 17941 is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE); 17942 17943 if (is_rdonly) 17944 *addr = (unsigned long)bpf_dynptr_from_skb_rdonly; 17945 17946 /* restore env->seen_direct_write to its original value, since 17947 * may_access_direct_pkt_data mutates it 17948 */ 17949 env->seen_direct_write = seen_direct_write; 17950 } 17951 } 17952 17953 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux, 17954 u16 struct_meta_reg, 17955 u16 node_offset_reg, 17956 struct bpf_insn *insn, 17957 struct bpf_insn *insn_buf, 17958 int *cnt) 17959 { 17960 struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta; 17961 struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) }; 17962 17963 insn_buf[0] = addr[0]; 17964 insn_buf[1] = addr[1]; 17965 insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off); 17966 insn_buf[3] = *insn; 17967 *cnt = 4; 17968 } 17969 17970 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn, 17971 struct bpf_insn *insn_buf, int insn_idx, int *cnt) 17972 { 17973 const struct bpf_kfunc_desc *desc; 17974 17975 if (!insn->imm) { 17976 verbose(env, "invalid kernel function call not eliminated in verifier pass\n"); 17977 return -EINVAL; 17978 } 17979 17980 *cnt = 0; 17981 17982 /* insn->imm has the btf func_id. Replace it with an offset relative to 17983 * __bpf_call_base, unless the JIT needs to call functions that are 17984 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()). 17985 */ 17986 desc = find_kfunc_desc(env->prog, insn->imm, insn->off); 17987 if (!desc) { 17988 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n", 17989 insn->imm); 17990 return -EFAULT; 17991 } 17992 17993 if (!bpf_jit_supports_far_kfunc_call()) 17994 insn->imm = BPF_CALL_IMM(desc->addr); 17995 if (insn->off) 17996 return 0; 17997 if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl]) { 17998 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta; 17999 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) }; 18000 u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size; 18001 18002 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size); 18003 insn_buf[1] = addr[0]; 18004 insn_buf[2] = addr[1]; 18005 insn_buf[3] = *insn; 18006 *cnt = 4; 18007 } else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] || 18008 desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) { 18009 struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta; 18010 struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) }; 18011 18012 insn_buf[0] = addr[0]; 18013 insn_buf[1] = addr[1]; 18014 insn_buf[2] = *insn; 18015 *cnt = 3; 18016 } else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] || 18017 desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] || 18018 desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) { 18019 int struct_meta_reg = BPF_REG_3; 18020 int node_offset_reg = BPF_REG_4; 18021 18022 /* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */ 18023 if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) { 18024 struct_meta_reg = BPF_REG_4; 18025 node_offset_reg = BPF_REG_5; 18026 } 18027 18028 __fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg, 18029 node_offset_reg, insn, insn_buf, cnt); 18030 } else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] || 18031 desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) { 18032 insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1); 18033 *cnt = 1; 18034 } 18035 return 0; 18036 } 18037 18038 /* Do various post-verification rewrites in a single program pass. 18039 * These rewrites simplify JIT and interpreter implementations. 18040 */ 18041 static int do_misc_fixups(struct bpf_verifier_env *env) 18042 { 18043 struct bpf_prog *prog = env->prog; 18044 enum bpf_attach_type eatype = prog->expected_attach_type; 18045 enum bpf_prog_type prog_type = resolve_prog_type(prog); 18046 struct bpf_insn *insn = prog->insnsi; 18047 const struct bpf_func_proto *fn; 18048 const int insn_cnt = prog->len; 18049 const struct bpf_map_ops *ops; 18050 struct bpf_insn_aux_data *aux; 18051 struct bpf_insn insn_buf[16]; 18052 struct bpf_prog *new_prog; 18053 struct bpf_map *map_ptr; 18054 int i, ret, cnt, delta = 0; 18055 18056 for (i = 0; i < insn_cnt; i++, insn++) { 18057 /* Make divide-by-zero exceptions impossible. */ 18058 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) || 18059 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) || 18060 insn->code == (BPF_ALU | BPF_MOD | BPF_X) || 18061 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) { 18062 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; 18063 bool isdiv = BPF_OP(insn->code) == BPF_DIV; 18064 struct bpf_insn *patchlet; 18065 struct bpf_insn chk_and_div[] = { 18066 /* [R,W]x div 0 -> 0 */ 18067 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 18068 BPF_JNE | BPF_K, insn->src_reg, 18069 0, 2, 0), 18070 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg), 18071 BPF_JMP_IMM(BPF_JA, 0, 0, 1), 18072 *insn, 18073 }; 18074 struct bpf_insn chk_and_mod[] = { 18075 /* [R,W]x mod 0 -> [R,W]x */ 18076 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) | 18077 BPF_JEQ | BPF_K, insn->src_reg, 18078 0, 1 + (is64 ? 0 : 1), 0), 18079 *insn, 18080 BPF_JMP_IMM(BPF_JA, 0, 0, 1), 18081 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg), 18082 }; 18083 18084 patchlet = isdiv ? chk_and_div : chk_and_mod; 18085 cnt = isdiv ? ARRAY_SIZE(chk_and_div) : 18086 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0); 18087 18088 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt); 18089 if (!new_prog) 18090 return -ENOMEM; 18091 18092 delta += cnt - 1; 18093 env->prog = prog = new_prog; 18094 insn = new_prog->insnsi + i + delta; 18095 continue; 18096 } 18097 18098 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */ 18099 if (BPF_CLASS(insn->code) == BPF_LD && 18100 (BPF_MODE(insn->code) == BPF_ABS || 18101 BPF_MODE(insn->code) == BPF_IND)) { 18102 cnt = env->ops->gen_ld_abs(insn, insn_buf); 18103 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) { 18104 verbose(env, "bpf verifier is misconfigured\n"); 18105 return -EINVAL; 18106 } 18107 18108 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18109 if (!new_prog) 18110 return -ENOMEM; 18111 18112 delta += cnt - 1; 18113 env->prog = prog = new_prog; 18114 insn = new_prog->insnsi + i + delta; 18115 continue; 18116 } 18117 18118 /* Rewrite pointer arithmetic to mitigate speculation attacks. */ 18119 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) || 18120 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) { 18121 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X; 18122 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X; 18123 struct bpf_insn *patch = &insn_buf[0]; 18124 bool issrc, isneg, isimm; 18125 u32 off_reg; 18126 18127 aux = &env->insn_aux_data[i + delta]; 18128 if (!aux->alu_state || 18129 aux->alu_state == BPF_ALU_NON_POINTER) 18130 continue; 18131 18132 isneg = aux->alu_state & BPF_ALU_NEG_VALUE; 18133 issrc = (aux->alu_state & BPF_ALU_SANITIZE) == 18134 BPF_ALU_SANITIZE_SRC; 18135 isimm = aux->alu_state & BPF_ALU_IMMEDIATE; 18136 18137 off_reg = issrc ? insn->src_reg : insn->dst_reg; 18138 if (isimm) { 18139 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit); 18140 } else { 18141 if (isneg) 18142 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1); 18143 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit); 18144 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg); 18145 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg); 18146 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0); 18147 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63); 18148 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg); 18149 } 18150 if (!issrc) 18151 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg); 18152 insn->src_reg = BPF_REG_AX; 18153 if (isneg) 18154 insn->code = insn->code == code_add ? 18155 code_sub : code_add; 18156 *patch++ = *insn; 18157 if (issrc && isneg && !isimm) 18158 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1); 18159 cnt = patch - insn_buf; 18160 18161 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18162 if (!new_prog) 18163 return -ENOMEM; 18164 18165 delta += cnt - 1; 18166 env->prog = prog = new_prog; 18167 insn = new_prog->insnsi + i + delta; 18168 continue; 18169 } 18170 18171 if (insn->code != (BPF_JMP | BPF_CALL)) 18172 continue; 18173 if (insn->src_reg == BPF_PSEUDO_CALL) 18174 continue; 18175 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) { 18176 ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt); 18177 if (ret) 18178 return ret; 18179 if (cnt == 0) 18180 continue; 18181 18182 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18183 if (!new_prog) 18184 return -ENOMEM; 18185 18186 delta += cnt - 1; 18187 env->prog = prog = new_prog; 18188 insn = new_prog->insnsi + i + delta; 18189 continue; 18190 } 18191 18192 if (insn->imm == BPF_FUNC_get_route_realm) 18193 prog->dst_needed = 1; 18194 if (insn->imm == BPF_FUNC_get_prandom_u32) 18195 bpf_user_rnd_init_once(); 18196 if (insn->imm == BPF_FUNC_override_return) 18197 prog->kprobe_override = 1; 18198 if (insn->imm == BPF_FUNC_tail_call) { 18199 /* If we tail call into other programs, we 18200 * cannot make any assumptions since they can 18201 * be replaced dynamically during runtime in 18202 * the program array. 18203 */ 18204 prog->cb_access = 1; 18205 if (!allow_tail_call_in_subprogs(env)) 18206 prog->aux->stack_depth = MAX_BPF_STACK; 18207 prog->aux->max_pkt_offset = MAX_PACKET_OFF; 18208 18209 /* mark bpf_tail_call as different opcode to avoid 18210 * conditional branch in the interpreter for every normal 18211 * call and to prevent accidental JITing by JIT compiler 18212 * that doesn't support bpf_tail_call yet 18213 */ 18214 insn->imm = 0; 18215 insn->code = BPF_JMP | BPF_TAIL_CALL; 18216 18217 aux = &env->insn_aux_data[i + delta]; 18218 if (env->bpf_capable && !prog->blinding_requested && 18219 prog->jit_requested && 18220 !bpf_map_key_poisoned(aux) && 18221 !bpf_map_ptr_poisoned(aux) && 18222 !bpf_map_ptr_unpriv(aux)) { 18223 struct bpf_jit_poke_descriptor desc = { 18224 .reason = BPF_POKE_REASON_TAIL_CALL, 18225 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state), 18226 .tail_call.key = bpf_map_key_immediate(aux), 18227 .insn_idx = i + delta, 18228 }; 18229 18230 ret = bpf_jit_add_poke_descriptor(prog, &desc); 18231 if (ret < 0) { 18232 verbose(env, "adding tail call poke descriptor failed\n"); 18233 return ret; 18234 } 18235 18236 insn->imm = ret + 1; 18237 continue; 18238 } 18239 18240 if (!bpf_map_ptr_unpriv(aux)) 18241 continue; 18242 18243 /* instead of changing every JIT dealing with tail_call 18244 * emit two extra insns: 18245 * if (index >= max_entries) goto out; 18246 * index &= array->index_mask; 18247 * to avoid out-of-bounds cpu speculation 18248 */ 18249 if (bpf_map_ptr_poisoned(aux)) { 18250 verbose(env, "tail_call abusing map_ptr\n"); 18251 return -EINVAL; 18252 } 18253 18254 map_ptr = BPF_MAP_PTR(aux->map_ptr_state); 18255 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3, 18256 map_ptr->max_entries, 2); 18257 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3, 18258 container_of(map_ptr, 18259 struct bpf_array, 18260 map)->index_mask); 18261 insn_buf[2] = *insn; 18262 cnt = 3; 18263 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18264 if (!new_prog) 18265 return -ENOMEM; 18266 18267 delta += cnt - 1; 18268 env->prog = prog = new_prog; 18269 insn = new_prog->insnsi + i + delta; 18270 continue; 18271 } 18272 18273 if (insn->imm == BPF_FUNC_timer_set_callback) { 18274 /* The verifier will process callback_fn as many times as necessary 18275 * with different maps and the register states prepared by 18276 * set_timer_callback_state will be accurate. 18277 * 18278 * The following use case is valid: 18279 * map1 is shared by prog1, prog2, prog3. 18280 * prog1 calls bpf_timer_init for some map1 elements 18281 * prog2 calls bpf_timer_set_callback for some map1 elements. 18282 * Those that were not bpf_timer_init-ed will return -EINVAL. 18283 * prog3 calls bpf_timer_start for some map1 elements. 18284 * Those that were not both bpf_timer_init-ed and 18285 * bpf_timer_set_callback-ed will return -EINVAL. 18286 */ 18287 struct bpf_insn ld_addrs[2] = { 18288 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux), 18289 }; 18290 18291 insn_buf[0] = ld_addrs[0]; 18292 insn_buf[1] = ld_addrs[1]; 18293 insn_buf[2] = *insn; 18294 cnt = 3; 18295 18296 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18297 if (!new_prog) 18298 return -ENOMEM; 18299 18300 delta += cnt - 1; 18301 env->prog = prog = new_prog; 18302 insn = new_prog->insnsi + i + delta; 18303 goto patch_call_imm; 18304 } 18305 18306 if (is_storage_get_function(insn->imm)) { 18307 if (!env->prog->aux->sleepable || 18308 env->insn_aux_data[i + delta].storage_get_func_atomic) 18309 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC); 18310 else 18311 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL); 18312 insn_buf[1] = *insn; 18313 cnt = 2; 18314 18315 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18316 if (!new_prog) 18317 return -ENOMEM; 18318 18319 delta += cnt - 1; 18320 env->prog = prog = new_prog; 18321 insn = new_prog->insnsi + i + delta; 18322 goto patch_call_imm; 18323 } 18324 18325 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup 18326 * and other inlining handlers are currently limited to 64 bit 18327 * only. 18328 */ 18329 if (prog->jit_requested && BITS_PER_LONG == 64 && 18330 (insn->imm == BPF_FUNC_map_lookup_elem || 18331 insn->imm == BPF_FUNC_map_update_elem || 18332 insn->imm == BPF_FUNC_map_delete_elem || 18333 insn->imm == BPF_FUNC_map_push_elem || 18334 insn->imm == BPF_FUNC_map_pop_elem || 18335 insn->imm == BPF_FUNC_map_peek_elem || 18336 insn->imm == BPF_FUNC_redirect_map || 18337 insn->imm == BPF_FUNC_for_each_map_elem || 18338 insn->imm == BPF_FUNC_map_lookup_percpu_elem)) { 18339 aux = &env->insn_aux_data[i + delta]; 18340 if (bpf_map_ptr_poisoned(aux)) 18341 goto patch_call_imm; 18342 18343 map_ptr = BPF_MAP_PTR(aux->map_ptr_state); 18344 ops = map_ptr->ops; 18345 if (insn->imm == BPF_FUNC_map_lookup_elem && 18346 ops->map_gen_lookup) { 18347 cnt = ops->map_gen_lookup(map_ptr, insn_buf); 18348 if (cnt == -EOPNOTSUPP) 18349 goto patch_map_ops_generic; 18350 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) { 18351 verbose(env, "bpf verifier is misconfigured\n"); 18352 return -EINVAL; 18353 } 18354 18355 new_prog = bpf_patch_insn_data(env, i + delta, 18356 insn_buf, cnt); 18357 if (!new_prog) 18358 return -ENOMEM; 18359 18360 delta += cnt - 1; 18361 env->prog = prog = new_prog; 18362 insn = new_prog->insnsi + i + delta; 18363 continue; 18364 } 18365 18366 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem, 18367 (void *(*)(struct bpf_map *map, void *key))NULL)); 18368 BUILD_BUG_ON(!__same_type(ops->map_delete_elem, 18369 (long (*)(struct bpf_map *map, void *key))NULL)); 18370 BUILD_BUG_ON(!__same_type(ops->map_update_elem, 18371 (long (*)(struct bpf_map *map, void *key, void *value, 18372 u64 flags))NULL)); 18373 BUILD_BUG_ON(!__same_type(ops->map_push_elem, 18374 (long (*)(struct bpf_map *map, void *value, 18375 u64 flags))NULL)); 18376 BUILD_BUG_ON(!__same_type(ops->map_pop_elem, 18377 (long (*)(struct bpf_map *map, void *value))NULL)); 18378 BUILD_BUG_ON(!__same_type(ops->map_peek_elem, 18379 (long (*)(struct bpf_map *map, void *value))NULL)); 18380 BUILD_BUG_ON(!__same_type(ops->map_redirect, 18381 (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL)); 18382 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback, 18383 (long (*)(struct bpf_map *map, 18384 bpf_callback_t callback_fn, 18385 void *callback_ctx, 18386 u64 flags))NULL)); 18387 BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem, 18388 (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL)); 18389 18390 patch_map_ops_generic: 18391 switch (insn->imm) { 18392 case BPF_FUNC_map_lookup_elem: 18393 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem); 18394 continue; 18395 case BPF_FUNC_map_update_elem: 18396 insn->imm = BPF_CALL_IMM(ops->map_update_elem); 18397 continue; 18398 case BPF_FUNC_map_delete_elem: 18399 insn->imm = BPF_CALL_IMM(ops->map_delete_elem); 18400 continue; 18401 case BPF_FUNC_map_push_elem: 18402 insn->imm = BPF_CALL_IMM(ops->map_push_elem); 18403 continue; 18404 case BPF_FUNC_map_pop_elem: 18405 insn->imm = BPF_CALL_IMM(ops->map_pop_elem); 18406 continue; 18407 case BPF_FUNC_map_peek_elem: 18408 insn->imm = BPF_CALL_IMM(ops->map_peek_elem); 18409 continue; 18410 case BPF_FUNC_redirect_map: 18411 insn->imm = BPF_CALL_IMM(ops->map_redirect); 18412 continue; 18413 case BPF_FUNC_for_each_map_elem: 18414 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback); 18415 continue; 18416 case BPF_FUNC_map_lookup_percpu_elem: 18417 insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem); 18418 continue; 18419 } 18420 18421 goto patch_call_imm; 18422 } 18423 18424 /* Implement bpf_jiffies64 inline. */ 18425 if (prog->jit_requested && BITS_PER_LONG == 64 && 18426 insn->imm == BPF_FUNC_jiffies64) { 18427 struct bpf_insn ld_jiffies_addr[2] = { 18428 BPF_LD_IMM64(BPF_REG_0, 18429 (unsigned long)&jiffies), 18430 }; 18431 18432 insn_buf[0] = ld_jiffies_addr[0]; 18433 insn_buf[1] = ld_jiffies_addr[1]; 18434 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, 18435 BPF_REG_0, 0); 18436 cnt = 3; 18437 18438 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 18439 cnt); 18440 if (!new_prog) 18441 return -ENOMEM; 18442 18443 delta += cnt - 1; 18444 env->prog = prog = new_prog; 18445 insn = new_prog->insnsi + i + delta; 18446 continue; 18447 } 18448 18449 /* Implement bpf_get_func_arg inline. */ 18450 if (prog_type == BPF_PROG_TYPE_TRACING && 18451 insn->imm == BPF_FUNC_get_func_arg) { 18452 /* Load nr_args from ctx - 8 */ 18453 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8); 18454 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6); 18455 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3); 18456 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1); 18457 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0); 18458 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0); 18459 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0); 18460 insn_buf[7] = BPF_JMP_A(1); 18461 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL); 18462 cnt = 9; 18463 18464 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18465 if (!new_prog) 18466 return -ENOMEM; 18467 18468 delta += cnt - 1; 18469 env->prog = prog = new_prog; 18470 insn = new_prog->insnsi + i + delta; 18471 continue; 18472 } 18473 18474 /* Implement bpf_get_func_ret inline. */ 18475 if (prog_type == BPF_PROG_TYPE_TRACING && 18476 insn->imm == BPF_FUNC_get_func_ret) { 18477 if (eatype == BPF_TRACE_FEXIT || 18478 eatype == BPF_MODIFY_RETURN) { 18479 /* Load nr_args from ctx - 8 */ 18480 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8); 18481 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3); 18482 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1); 18483 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0); 18484 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0); 18485 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0); 18486 cnt = 6; 18487 } else { 18488 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP); 18489 cnt = 1; 18490 } 18491 18492 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); 18493 if (!new_prog) 18494 return -ENOMEM; 18495 18496 delta += cnt - 1; 18497 env->prog = prog = new_prog; 18498 insn = new_prog->insnsi + i + delta; 18499 continue; 18500 } 18501 18502 /* Implement get_func_arg_cnt inline. */ 18503 if (prog_type == BPF_PROG_TYPE_TRACING && 18504 insn->imm == BPF_FUNC_get_func_arg_cnt) { 18505 /* Load nr_args from ctx - 8 */ 18506 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8); 18507 18508 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1); 18509 if (!new_prog) 18510 return -ENOMEM; 18511 18512 env->prog = prog = new_prog; 18513 insn = new_prog->insnsi + i + delta; 18514 continue; 18515 } 18516 18517 /* Implement bpf_get_func_ip inline. */ 18518 if (prog_type == BPF_PROG_TYPE_TRACING && 18519 insn->imm == BPF_FUNC_get_func_ip) { 18520 /* Load IP address from ctx - 16 */ 18521 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16); 18522 18523 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1); 18524 if (!new_prog) 18525 return -ENOMEM; 18526 18527 env->prog = prog = new_prog; 18528 insn = new_prog->insnsi + i + delta; 18529 continue; 18530 } 18531 18532 patch_call_imm: 18533 fn = env->ops->get_func_proto(insn->imm, env->prog); 18534 /* all functions that have prototype and verifier allowed 18535 * programs to call them, must be real in-kernel functions 18536 */ 18537 if (!fn->func) { 18538 verbose(env, 18539 "kernel subsystem misconfigured func %s#%d\n", 18540 func_id_name(insn->imm), insn->imm); 18541 return -EFAULT; 18542 } 18543 insn->imm = fn->func - __bpf_call_base; 18544 } 18545 18546 /* Since poke tab is now finalized, publish aux to tracker. */ 18547 for (i = 0; i < prog->aux->size_poke_tab; i++) { 18548 map_ptr = prog->aux->poke_tab[i].tail_call.map; 18549 if (!map_ptr->ops->map_poke_track || 18550 !map_ptr->ops->map_poke_untrack || 18551 !map_ptr->ops->map_poke_run) { 18552 verbose(env, "bpf verifier is misconfigured\n"); 18553 return -EINVAL; 18554 } 18555 18556 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux); 18557 if (ret < 0) { 18558 verbose(env, "tracking tail call prog failed\n"); 18559 return ret; 18560 } 18561 } 18562 18563 sort_kfunc_descs_by_imm_off(env->prog); 18564 18565 return 0; 18566 } 18567 18568 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env, 18569 int position, 18570 s32 stack_base, 18571 u32 callback_subprogno, 18572 u32 *cnt) 18573 { 18574 s32 r6_offset = stack_base + 0 * BPF_REG_SIZE; 18575 s32 r7_offset = stack_base + 1 * BPF_REG_SIZE; 18576 s32 r8_offset = stack_base + 2 * BPF_REG_SIZE; 18577 int reg_loop_max = BPF_REG_6; 18578 int reg_loop_cnt = BPF_REG_7; 18579 int reg_loop_ctx = BPF_REG_8; 18580 18581 struct bpf_prog *new_prog; 18582 u32 callback_start; 18583 u32 call_insn_offset; 18584 s32 callback_offset; 18585 18586 /* This represents an inlined version of bpf_iter.c:bpf_loop, 18587 * be careful to modify this code in sync. 18588 */ 18589 struct bpf_insn insn_buf[] = { 18590 /* Return error and jump to the end of the patch if 18591 * expected number of iterations is too big. 18592 */ 18593 BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2), 18594 BPF_MOV32_IMM(BPF_REG_0, -E2BIG), 18595 BPF_JMP_IMM(BPF_JA, 0, 0, 16), 18596 /* spill R6, R7, R8 to use these as loop vars */ 18597 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset), 18598 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset), 18599 BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset), 18600 /* initialize loop vars */ 18601 BPF_MOV64_REG(reg_loop_max, BPF_REG_1), 18602 BPF_MOV32_IMM(reg_loop_cnt, 0), 18603 BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3), 18604 /* loop header, 18605 * if reg_loop_cnt >= reg_loop_max skip the loop body 18606 */ 18607 BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5), 18608 /* callback call, 18609 * correct callback offset would be set after patching 18610 */ 18611 BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt), 18612 BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx), 18613 BPF_CALL_REL(0), 18614 /* increment loop counter */ 18615 BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1), 18616 /* jump to loop header if callback returned 0 */ 18617 BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6), 18618 /* return value of bpf_loop, 18619 * set R0 to the number of iterations 18620 */ 18621 BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt), 18622 /* restore original values of R6, R7, R8 */ 18623 BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset), 18624 BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset), 18625 BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset), 18626 }; 18627 18628 *cnt = ARRAY_SIZE(insn_buf); 18629 new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt); 18630 if (!new_prog) 18631 return new_prog; 18632 18633 /* callback start is known only after patching */ 18634 callback_start = env->subprog_info[callback_subprogno].start; 18635 /* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */ 18636 call_insn_offset = position + 12; 18637 callback_offset = callback_start - call_insn_offset - 1; 18638 new_prog->insnsi[call_insn_offset].imm = callback_offset; 18639 18640 return new_prog; 18641 } 18642 18643 static bool is_bpf_loop_call(struct bpf_insn *insn) 18644 { 18645 return insn->code == (BPF_JMP | BPF_CALL) && 18646 insn->src_reg == 0 && 18647 insn->imm == BPF_FUNC_loop; 18648 } 18649 18650 /* For all sub-programs in the program (including main) check 18651 * insn_aux_data to see if there are bpf_loop calls that require 18652 * inlining. If such calls are found the calls are replaced with a 18653 * sequence of instructions produced by `inline_bpf_loop` function and 18654 * subprog stack_depth is increased by the size of 3 registers. 18655 * This stack space is used to spill values of the R6, R7, R8. These 18656 * registers are used to store the loop bound, counter and context 18657 * variables. 18658 */ 18659 static int optimize_bpf_loop(struct bpf_verifier_env *env) 18660 { 18661 struct bpf_subprog_info *subprogs = env->subprog_info; 18662 int i, cur_subprog = 0, cnt, delta = 0; 18663 struct bpf_insn *insn = env->prog->insnsi; 18664 int insn_cnt = env->prog->len; 18665 u16 stack_depth = subprogs[cur_subprog].stack_depth; 18666 u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth; 18667 u16 stack_depth_extra = 0; 18668 18669 for (i = 0; i < insn_cnt; i++, insn++) { 18670 struct bpf_loop_inline_state *inline_state = 18671 &env->insn_aux_data[i + delta].loop_inline_state; 18672 18673 if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) { 18674 struct bpf_prog *new_prog; 18675 18676 stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup; 18677 new_prog = inline_bpf_loop(env, 18678 i + delta, 18679 -(stack_depth + stack_depth_extra), 18680 inline_state->callback_subprogno, 18681 &cnt); 18682 if (!new_prog) 18683 return -ENOMEM; 18684 18685 delta += cnt - 1; 18686 env->prog = new_prog; 18687 insn = new_prog->insnsi + i + delta; 18688 } 18689 18690 if (subprogs[cur_subprog + 1].start == i + delta + 1) { 18691 subprogs[cur_subprog].stack_depth += stack_depth_extra; 18692 cur_subprog++; 18693 stack_depth = subprogs[cur_subprog].stack_depth; 18694 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth; 18695 stack_depth_extra = 0; 18696 } 18697 } 18698 18699 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth; 18700 18701 return 0; 18702 } 18703 18704 static void free_states(struct bpf_verifier_env *env) 18705 { 18706 struct bpf_verifier_state_list *sl, *sln; 18707 int i; 18708 18709 sl = env->free_list; 18710 while (sl) { 18711 sln = sl->next; 18712 free_verifier_state(&sl->state, false); 18713 kfree(sl); 18714 sl = sln; 18715 } 18716 env->free_list = NULL; 18717 18718 if (!env->explored_states) 18719 return; 18720 18721 for (i = 0; i < state_htab_size(env); i++) { 18722 sl = env->explored_states[i]; 18723 18724 while (sl) { 18725 sln = sl->next; 18726 free_verifier_state(&sl->state, false); 18727 kfree(sl); 18728 sl = sln; 18729 } 18730 env->explored_states[i] = NULL; 18731 } 18732 } 18733 18734 static int do_check_common(struct bpf_verifier_env *env, int subprog) 18735 { 18736 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2); 18737 struct bpf_verifier_state *state; 18738 struct bpf_reg_state *regs; 18739 int ret, i; 18740 18741 env->prev_linfo = NULL; 18742 env->pass_cnt++; 18743 18744 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL); 18745 if (!state) 18746 return -ENOMEM; 18747 state->curframe = 0; 18748 state->speculative = false; 18749 state->branches = 1; 18750 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL); 18751 if (!state->frame[0]) { 18752 kfree(state); 18753 return -ENOMEM; 18754 } 18755 env->cur_state = state; 18756 init_func_state(env, state->frame[0], 18757 BPF_MAIN_FUNC /* callsite */, 18758 0 /* frameno */, 18759 subprog); 18760 state->first_insn_idx = env->subprog_info[subprog].start; 18761 state->last_insn_idx = -1; 18762 18763 regs = state->frame[state->curframe]->regs; 18764 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) { 18765 ret = btf_prepare_func_args(env, subprog, regs); 18766 if (ret) 18767 goto out; 18768 for (i = BPF_REG_1; i <= BPF_REG_5; i++) { 18769 if (regs[i].type == PTR_TO_CTX) 18770 mark_reg_known_zero(env, regs, i); 18771 else if (regs[i].type == SCALAR_VALUE) 18772 mark_reg_unknown(env, regs, i); 18773 else if (base_type(regs[i].type) == PTR_TO_MEM) { 18774 const u32 mem_size = regs[i].mem_size; 18775 18776 mark_reg_known_zero(env, regs, i); 18777 regs[i].mem_size = mem_size; 18778 regs[i].id = ++env->id_gen; 18779 } 18780 } 18781 } else { 18782 /* 1st arg to a function */ 18783 regs[BPF_REG_1].type = PTR_TO_CTX; 18784 mark_reg_known_zero(env, regs, BPF_REG_1); 18785 ret = btf_check_subprog_arg_match(env, subprog, regs); 18786 if (ret == -EFAULT) 18787 /* unlikely verifier bug. abort. 18788 * ret == 0 and ret < 0 are sadly acceptable for 18789 * main() function due to backward compatibility. 18790 * Like socket filter program may be written as: 18791 * int bpf_prog(struct pt_regs *ctx) 18792 * and never dereference that ctx in the program. 18793 * 'struct pt_regs' is a type mismatch for socket 18794 * filter that should be using 'struct __sk_buff'. 18795 */ 18796 goto out; 18797 } 18798 18799 ret = do_check(env); 18800 out: 18801 /* check for NULL is necessary, since cur_state can be freed inside 18802 * do_check() under memory pressure. 18803 */ 18804 if (env->cur_state) { 18805 free_verifier_state(env->cur_state, true); 18806 env->cur_state = NULL; 18807 } 18808 while (!pop_stack(env, NULL, NULL, false)); 18809 if (!ret && pop_log) 18810 bpf_vlog_reset(&env->log, 0); 18811 free_states(env); 18812 return ret; 18813 } 18814 18815 /* Verify all global functions in a BPF program one by one based on their BTF. 18816 * All global functions must pass verification. Otherwise the whole program is rejected. 18817 * Consider: 18818 * int bar(int); 18819 * int foo(int f) 18820 * { 18821 * return bar(f); 18822 * } 18823 * int bar(int b) 18824 * { 18825 * ... 18826 * } 18827 * foo() will be verified first for R1=any_scalar_value. During verification it 18828 * will be assumed that bar() already verified successfully and call to bar() 18829 * from foo() will be checked for type match only. Later bar() will be verified 18830 * independently to check that it's safe for R1=any_scalar_value. 18831 */ 18832 static int do_check_subprogs(struct bpf_verifier_env *env) 18833 { 18834 struct bpf_prog_aux *aux = env->prog->aux; 18835 int i, ret; 18836 18837 if (!aux->func_info) 18838 return 0; 18839 18840 for (i = 1; i < env->subprog_cnt; i++) { 18841 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL) 18842 continue; 18843 env->insn_idx = env->subprog_info[i].start; 18844 WARN_ON_ONCE(env->insn_idx == 0); 18845 ret = do_check_common(env, i); 18846 if (ret) { 18847 return ret; 18848 } else if (env->log.level & BPF_LOG_LEVEL) { 18849 verbose(env, 18850 "Func#%d is safe for any args that match its prototype\n", 18851 i); 18852 } 18853 } 18854 return 0; 18855 } 18856 18857 static int do_check_main(struct bpf_verifier_env *env) 18858 { 18859 int ret; 18860 18861 env->insn_idx = 0; 18862 ret = do_check_common(env, 0); 18863 if (!ret) 18864 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth; 18865 return ret; 18866 } 18867 18868 18869 static void print_verification_stats(struct bpf_verifier_env *env) 18870 { 18871 int i; 18872 18873 if (env->log.level & BPF_LOG_STATS) { 18874 verbose(env, "verification time %lld usec\n", 18875 div_u64(env->verification_time, 1000)); 18876 verbose(env, "stack depth "); 18877 for (i = 0; i < env->subprog_cnt; i++) { 18878 u32 depth = env->subprog_info[i].stack_depth; 18879 18880 verbose(env, "%d", depth); 18881 if (i + 1 < env->subprog_cnt) 18882 verbose(env, "+"); 18883 } 18884 verbose(env, "\n"); 18885 } 18886 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d " 18887 "total_states %d peak_states %d mark_read %d\n", 18888 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS, 18889 env->max_states_per_insn, env->total_states, 18890 env->peak_states, env->longest_mark_read_walk); 18891 } 18892 18893 static int check_struct_ops_btf_id(struct bpf_verifier_env *env) 18894 { 18895 const struct btf_type *t, *func_proto; 18896 const struct bpf_struct_ops *st_ops; 18897 const struct btf_member *member; 18898 struct bpf_prog *prog = env->prog; 18899 u32 btf_id, member_idx; 18900 const char *mname; 18901 18902 if (!prog->gpl_compatible) { 18903 verbose(env, "struct ops programs must have a GPL compatible license\n"); 18904 return -EINVAL; 18905 } 18906 18907 btf_id = prog->aux->attach_btf_id; 18908 st_ops = bpf_struct_ops_find(btf_id); 18909 if (!st_ops) { 18910 verbose(env, "attach_btf_id %u is not a supported struct\n", 18911 btf_id); 18912 return -ENOTSUPP; 18913 } 18914 18915 t = st_ops->type; 18916 member_idx = prog->expected_attach_type; 18917 if (member_idx >= btf_type_vlen(t)) { 18918 verbose(env, "attach to invalid member idx %u of struct %s\n", 18919 member_idx, st_ops->name); 18920 return -EINVAL; 18921 } 18922 18923 member = &btf_type_member(t)[member_idx]; 18924 mname = btf_name_by_offset(btf_vmlinux, member->name_off); 18925 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type, 18926 NULL); 18927 if (!func_proto) { 18928 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n", 18929 mname, member_idx, st_ops->name); 18930 return -EINVAL; 18931 } 18932 18933 if (st_ops->check_member) { 18934 int err = st_ops->check_member(t, member, prog); 18935 18936 if (err) { 18937 verbose(env, "attach to unsupported member %s of struct %s\n", 18938 mname, st_ops->name); 18939 return err; 18940 } 18941 } 18942 18943 prog->aux->attach_func_proto = func_proto; 18944 prog->aux->attach_func_name = mname; 18945 env->ops = st_ops->verifier_ops; 18946 18947 return 0; 18948 } 18949 #define SECURITY_PREFIX "security_" 18950 18951 static int check_attach_modify_return(unsigned long addr, const char *func_name) 18952 { 18953 if (within_error_injection_list(addr) || 18954 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1)) 18955 return 0; 18956 18957 return -EINVAL; 18958 } 18959 18960 /* list of non-sleepable functions that are otherwise on 18961 * ALLOW_ERROR_INJECTION list 18962 */ 18963 BTF_SET_START(btf_non_sleepable_error_inject) 18964 /* Three functions below can be called from sleepable and non-sleepable context. 18965 * Assume non-sleepable from bpf safety point of view. 18966 */ 18967 BTF_ID(func, __filemap_add_folio) 18968 BTF_ID(func, should_fail_alloc_page) 18969 BTF_ID(func, should_failslab) 18970 BTF_SET_END(btf_non_sleepable_error_inject) 18971 18972 static int check_non_sleepable_error_inject(u32 btf_id) 18973 { 18974 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id); 18975 } 18976 18977 int bpf_check_attach_target(struct bpf_verifier_log *log, 18978 const struct bpf_prog *prog, 18979 const struct bpf_prog *tgt_prog, 18980 u32 btf_id, 18981 struct bpf_attach_target_info *tgt_info) 18982 { 18983 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT; 18984 const char prefix[] = "btf_trace_"; 18985 int ret = 0, subprog = -1, i; 18986 const struct btf_type *t; 18987 bool conservative = true; 18988 const char *tname; 18989 struct btf *btf; 18990 long addr = 0; 18991 struct module *mod = NULL; 18992 18993 if (!btf_id) { 18994 bpf_log(log, "Tracing programs must provide btf_id\n"); 18995 return -EINVAL; 18996 } 18997 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf; 18998 if (!btf) { 18999 bpf_log(log, 19000 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n"); 19001 return -EINVAL; 19002 } 19003 t = btf_type_by_id(btf, btf_id); 19004 if (!t) { 19005 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id); 19006 return -EINVAL; 19007 } 19008 tname = btf_name_by_offset(btf, t->name_off); 19009 if (!tname) { 19010 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id); 19011 return -EINVAL; 19012 } 19013 if (tgt_prog) { 19014 struct bpf_prog_aux *aux = tgt_prog->aux; 19015 19016 if (bpf_prog_is_dev_bound(prog->aux) && 19017 !bpf_prog_dev_bound_match(prog, tgt_prog)) { 19018 bpf_log(log, "Target program bound device mismatch"); 19019 return -EINVAL; 19020 } 19021 19022 for (i = 0; i < aux->func_info_cnt; i++) 19023 if (aux->func_info[i].type_id == btf_id) { 19024 subprog = i; 19025 break; 19026 } 19027 if (subprog == -1) { 19028 bpf_log(log, "Subprog %s doesn't exist\n", tname); 19029 return -EINVAL; 19030 } 19031 conservative = aux->func_info_aux[subprog].unreliable; 19032 if (prog_extension) { 19033 if (conservative) { 19034 bpf_log(log, 19035 "Cannot replace static functions\n"); 19036 return -EINVAL; 19037 } 19038 if (!prog->jit_requested) { 19039 bpf_log(log, 19040 "Extension programs should be JITed\n"); 19041 return -EINVAL; 19042 } 19043 } 19044 if (!tgt_prog->jited) { 19045 bpf_log(log, "Can attach to only JITed progs\n"); 19046 return -EINVAL; 19047 } 19048 if (tgt_prog->type == prog->type) { 19049 /* Cannot fentry/fexit another fentry/fexit program. 19050 * Cannot attach program extension to another extension. 19051 * It's ok to attach fentry/fexit to extension program. 19052 */ 19053 bpf_log(log, "Cannot recursively attach\n"); 19054 return -EINVAL; 19055 } 19056 if (tgt_prog->type == BPF_PROG_TYPE_TRACING && 19057 prog_extension && 19058 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY || 19059 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) { 19060 /* Program extensions can extend all program types 19061 * except fentry/fexit. The reason is the following. 19062 * The fentry/fexit programs are used for performance 19063 * analysis, stats and can be attached to any program 19064 * type except themselves. When extension program is 19065 * replacing XDP function it is necessary to allow 19066 * performance analysis of all functions. Both original 19067 * XDP program and its program extension. Hence 19068 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is 19069 * allowed. If extending of fentry/fexit was allowed it 19070 * would be possible to create long call chain 19071 * fentry->extension->fentry->extension beyond 19072 * reasonable stack size. Hence extending fentry is not 19073 * allowed. 19074 */ 19075 bpf_log(log, "Cannot extend fentry/fexit\n"); 19076 return -EINVAL; 19077 } 19078 } else { 19079 if (prog_extension) { 19080 bpf_log(log, "Cannot replace kernel functions\n"); 19081 return -EINVAL; 19082 } 19083 } 19084 19085 switch (prog->expected_attach_type) { 19086 case BPF_TRACE_RAW_TP: 19087 if (tgt_prog) { 19088 bpf_log(log, 19089 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n"); 19090 return -EINVAL; 19091 } 19092 if (!btf_type_is_typedef(t)) { 19093 bpf_log(log, "attach_btf_id %u is not a typedef\n", 19094 btf_id); 19095 return -EINVAL; 19096 } 19097 if (strncmp(prefix, tname, sizeof(prefix) - 1)) { 19098 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n", 19099 btf_id, tname); 19100 return -EINVAL; 19101 } 19102 tname += sizeof(prefix) - 1; 19103 t = btf_type_by_id(btf, t->type); 19104 if (!btf_type_is_ptr(t)) 19105 /* should never happen in valid vmlinux build */ 19106 return -EINVAL; 19107 t = btf_type_by_id(btf, t->type); 19108 if (!btf_type_is_func_proto(t)) 19109 /* should never happen in valid vmlinux build */ 19110 return -EINVAL; 19111 19112 break; 19113 case BPF_TRACE_ITER: 19114 if (!btf_type_is_func(t)) { 19115 bpf_log(log, "attach_btf_id %u is not a function\n", 19116 btf_id); 19117 return -EINVAL; 19118 } 19119 t = btf_type_by_id(btf, t->type); 19120 if (!btf_type_is_func_proto(t)) 19121 return -EINVAL; 19122 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel); 19123 if (ret) 19124 return ret; 19125 break; 19126 default: 19127 if (!prog_extension) 19128 return -EINVAL; 19129 fallthrough; 19130 case BPF_MODIFY_RETURN: 19131 case BPF_LSM_MAC: 19132 case BPF_LSM_CGROUP: 19133 case BPF_TRACE_FENTRY: 19134 case BPF_TRACE_FEXIT: 19135 if (!btf_type_is_func(t)) { 19136 bpf_log(log, "attach_btf_id %u is not a function\n", 19137 btf_id); 19138 return -EINVAL; 19139 } 19140 if (prog_extension && 19141 btf_check_type_match(log, prog, btf, t)) 19142 return -EINVAL; 19143 t = btf_type_by_id(btf, t->type); 19144 if (!btf_type_is_func_proto(t)) 19145 return -EINVAL; 19146 19147 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) && 19148 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type || 19149 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type)) 19150 return -EINVAL; 19151 19152 if (tgt_prog && conservative) 19153 t = NULL; 19154 19155 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel); 19156 if (ret < 0) 19157 return ret; 19158 19159 if (tgt_prog) { 19160 if (subprog == 0) 19161 addr = (long) tgt_prog->bpf_func; 19162 else 19163 addr = (long) tgt_prog->aux->func[subprog]->bpf_func; 19164 } else { 19165 if (btf_is_module(btf)) { 19166 mod = btf_try_get_module(btf); 19167 if (mod) 19168 addr = find_kallsyms_symbol_value(mod, tname); 19169 else 19170 addr = 0; 19171 } else { 19172 addr = kallsyms_lookup_name(tname); 19173 } 19174 if (!addr) { 19175 module_put(mod); 19176 bpf_log(log, 19177 "The address of function %s cannot be found\n", 19178 tname); 19179 return -ENOENT; 19180 } 19181 } 19182 19183 if (prog->aux->sleepable) { 19184 ret = -EINVAL; 19185 switch (prog->type) { 19186 case BPF_PROG_TYPE_TRACING: 19187 19188 /* fentry/fexit/fmod_ret progs can be sleepable if they are 19189 * attached to ALLOW_ERROR_INJECTION and are not in denylist. 19190 */ 19191 if (!check_non_sleepable_error_inject(btf_id) && 19192 within_error_injection_list(addr)) 19193 ret = 0; 19194 /* fentry/fexit/fmod_ret progs can also be sleepable if they are 19195 * in the fmodret id set with the KF_SLEEPABLE flag. 19196 */ 19197 else { 19198 u32 *flags = btf_kfunc_is_modify_return(btf, btf_id, 19199 prog); 19200 19201 if (flags && (*flags & KF_SLEEPABLE)) 19202 ret = 0; 19203 } 19204 break; 19205 case BPF_PROG_TYPE_LSM: 19206 /* LSM progs check that they are attached to bpf_lsm_*() funcs. 19207 * Only some of them are sleepable. 19208 */ 19209 if (bpf_lsm_is_sleepable_hook(btf_id)) 19210 ret = 0; 19211 break; 19212 default: 19213 break; 19214 } 19215 if (ret) { 19216 module_put(mod); 19217 bpf_log(log, "%s is not sleepable\n", tname); 19218 return ret; 19219 } 19220 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) { 19221 if (tgt_prog) { 19222 module_put(mod); 19223 bpf_log(log, "can't modify return codes of BPF programs\n"); 19224 return -EINVAL; 19225 } 19226 ret = -EINVAL; 19227 if (btf_kfunc_is_modify_return(btf, btf_id, prog) || 19228 !check_attach_modify_return(addr, tname)) 19229 ret = 0; 19230 if (ret) { 19231 module_put(mod); 19232 bpf_log(log, "%s() is not modifiable\n", tname); 19233 return ret; 19234 } 19235 } 19236 19237 break; 19238 } 19239 tgt_info->tgt_addr = addr; 19240 tgt_info->tgt_name = tname; 19241 tgt_info->tgt_type = t; 19242 tgt_info->tgt_mod = mod; 19243 return 0; 19244 } 19245 19246 BTF_SET_START(btf_id_deny) 19247 BTF_ID_UNUSED 19248 #ifdef CONFIG_SMP 19249 BTF_ID(func, migrate_disable) 19250 BTF_ID(func, migrate_enable) 19251 #endif 19252 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU 19253 BTF_ID(func, rcu_read_unlock_strict) 19254 #endif 19255 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE) 19256 BTF_ID(func, preempt_count_add) 19257 BTF_ID(func, preempt_count_sub) 19258 #endif 19259 #ifdef CONFIG_PREEMPT_RCU 19260 BTF_ID(func, __rcu_read_lock) 19261 BTF_ID(func, __rcu_read_unlock) 19262 #endif 19263 BTF_SET_END(btf_id_deny) 19264 19265 static bool can_be_sleepable(struct bpf_prog *prog) 19266 { 19267 if (prog->type == BPF_PROG_TYPE_TRACING) { 19268 switch (prog->expected_attach_type) { 19269 case BPF_TRACE_FENTRY: 19270 case BPF_TRACE_FEXIT: 19271 case BPF_MODIFY_RETURN: 19272 case BPF_TRACE_ITER: 19273 return true; 19274 default: 19275 return false; 19276 } 19277 } 19278 return prog->type == BPF_PROG_TYPE_LSM || 19279 prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ || 19280 prog->type == BPF_PROG_TYPE_STRUCT_OPS; 19281 } 19282 19283 static int check_attach_btf_id(struct bpf_verifier_env *env) 19284 { 19285 struct bpf_prog *prog = env->prog; 19286 struct bpf_prog *tgt_prog = prog->aux->dst_prog; 19287 struct bpf_attach_target_info tgt_info = {}; 19288 u32 btf_id = prog->aux->attach_btf_id; 19289 struct bpf_trampoline *tr; 19290 int ret; 19291 u64 key; 19292 19293 if (prog->type == BPF_PROG_TYPE_SYSCALL) { 19294 if (prog->aux->sleepable) 19295 /* attach_btf_id checked to be zero already */ 19296 return 0; 19297 verbose(env, "Syscall programs can only be sleepable\n"); 19298 return -EINVAL; 19299 } 19300 19301 if (prog->aux->sleepable && !can_be_sleepable(prog)) { 19302 verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n"); 19303 return -EINVAL; 19304 } 19305 19306 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS) 19307 return check_struct_ops_btf_id(env); 19308 19309 if (prog->type != BPF_PROG_TYPE_TRACING && 19310 prog->type != BPF_PROG_TYPE_LSM && 19311 prog->type != BPF_PROG_TYPE_EXT) 19312 return 0; 19313 19314 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info); 19315 if (ret) 19316 return ret; 19317 19318 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) { 19319 /* to make freplace equivalent to their targets, they need to 19320 * inherit env->ops and expected_attach_type for the rest of the 19321 * verification 19322 */ 19323 env->ops = bpf_verifier_ops[tgt_prog->type]; 19324 prog->expected_attach_type = tgt_prog->expected_attach_type; 19325 } 19326 19327 /* store info about the attachment target that will be used later */ 19328 prog->aux->attach_func_proto = tgt_info.tgt_type; 19329 prog->aux->attach_func_name = tgt_info.tgt_name; 19330 prog->aux->mod = tgt_info.tgt_mod; 19331 19332 if (tgt_prog) { 19333 prog->aux->saved_dst_prog_type = tgt_prog->type; 19334 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type; 19335 } 19336 19337 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) { 19338 prog->aux->attach_btf_trace = true; 19339 return 0; 19340 } else if (prog->expected_attach_type == BPF_TRACE_ITER) { 19341 if (!bpf_iter_prog_supported(prog)) 19342 return -EINVAL; 19343 return 0; 19344 } 19345 19346 if (prog->type == BPF_PROG_TYPE_LSM) { 19347 ret = bpf_lsm_verify_prog(&env->log, prog); 19348 if (ret < 0) 19349 return ret; 19350 } else if (prog->type == BPF_PROG_TYPE_TRACING && 19351 btf_id_set_contains(&btf_id_deny, btf_id)) { 19352 return -EINVAL; 19353 } 19354 19355 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id); 19356 tr = bpf_trampoline_get(key, &tgt_info); 19357 if (!tr) 19358 return -ENOMEM; 19359 19360 prog->aux->dst_trampoline = tr; 19361 return 0; 19362 } 19363 19364 struct btf *bpf_get_btf_vmlinux(void) 19365 { 19366 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) { 19367 mutex_lock(&bpf_verifier_lock); 19368 if (!btf_vmlinux) 19369 btf_vmlinux = btf_parse_vmlinux(); 19370 mutex_unlock(&bpf_verifier_lock); 19371 } 19372 return btf_vmlinux; 19373 } 19374 19375 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size) 19376 { 19377 u64 start_time = ktime_get_ns(); 19378 struct bpf_verifier_env *env; 19379 int i, len, ret = -EINVAL, err; 19380 u32 log_true_size; 19381 bool is_priv; 19382 19383 /* no program is valid */ 19384 if (ARRAY_SIZE(bpf_verifier_ops) == 0) 19385 return -EINVAL; 19386 19387 /* 'struct bpf_verifier_env' can be global, but since it's not small, 19388 * allocate/free it every time bpf_check() is called 19389 */ 19390 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL); 19391 if (!env) 19392 return -ENOMEM; 19393 19394 env->bt.env = env; 19395 19396 len = (*prog)->len; 19397 env->insn_aux_data = 19398 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len)); 19399 ret = -ENOMEM; 19400 if (!env->insn_aux_data) 19401 goto err_free_env; 19402 for (i = 0; i < len; i++) 19403 env->insn_aux_data[i].orig_idx = i; 19404 env->prog = *prog; 19405 env->ops = bpf_verifier_ops[env->prog->type]; 19406 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel); 19407 is_priv = bpf_capable(); 19408 19409 bpf_get_btf_vmlinux(); 19410 19411 /* grab the mutex to protect few globals used by verifier */ 19412 if (!is_priv) 19413 mutex_lock(&bpf_verifier_lock); 19414 19415 /* user could have requested verbose verifier output 19416 * and supplied buffer to store the verification trace 19417 */ 19418 ret = bpf_vlog_init(&env->log, attr->log_level, 19419 (char __user *) (unsigned long) attr->log_buf, 19420 attr->log_size); 19421 if (ret) 19422 goto err_unlock; 19423 19424 mark_verifier_state_clean(env); 19425 19426 if (IS_ERR(btf_vmlinux)) { 19427 /* Either gcc or pahole or kernel are broken. */ 19428 verbose(env, "in-kernel BTF is malformed\n"); 19429 ret = PTR_ERR(btf_vmlinux); 19430 goto skip_full_check; 19431 } 19432 19433 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT); 19434 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)) 19435 env->strict_alignment = true; 19436 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT) 19437 env->strict_alignment = false; 19438 19439 env->allow_ptr_leaks = bpf_allow_ptr_leaks(); 19440 env->allow_uninit_stack = bpf_allow_uninit_stack(); 19441 env->bypass_spec_v1 = bpf_bypass_spec_v1(); 19442 env->bypass_spec_v4 = bpf_bypass_spec_v4(); 19443 env->bpf_capable = bpf_capable(); 19444 19445 if (is_priv) 19446 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ; 19447 19448 env->explored_states = kvcalloc(state_htab_size(env), 19449 sizeof(struct bpf_verifier_state_list *), 19450 GFP_USER); 19451 ret = -ENOMEM; 19452 if (!env->explored_states) 19453 goto skip_full_check; 19454 19455 ret = add_subprog_and_kfunc(env); 19456 if (ret < 0) 19457 goto skip_full_check; 19458 19459 ret = check_subprogs(env); 19460 if (ret < 0) 19461 goto skip_full_check; 19462 19463 ret = check_btf_info(env, attr, uattr); 19464 if (ret < 0) 19465 goto skip_full_check; 19466 19467 ret = check_attach_btf_id(env); 19468 if (ret) 19469 goto skip_full_check; 19470 19471 ret = resolve_pseudo_ldimm64(env); 19472 if (ret < 0) 19473 goto skip_full_check; 19474 19475 if (bpf_prog_is_offloaded(env->prog->aux)) { 19476 ret = bpf_prog_offload_verifier_prep(env->prog); 19477 if (ret) 19478 goto skip_full_check; 19479 } 19480 19481 ret = check_cfg(env); 19482 if (ret < 0) 19483 goto skip_full_check; 19484 19485 ret = do_check_subprogs(env); 19486 ret = ret ?: do_check_main(env); 19487 19488 if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux)) 19489 ret = bpf_prog_offload_finalize(env); 19490 19491 skip_full_check: 19492 kvfree(env->explored_states); 19493 19494 if (ret == 0) 19495 ret = check_max_stack_depth(env); 19496 19497 /* instruction rewrites happen after this point */ 19498 if (ret == 0) 19499 ret = optimize_bpf_loop(env); 19500 19501 if (is_priv) { 19502 if (ret == 0) 19503 opt_hard_wire_dead_code_branches(env); 19504 if (ret == 0) 19505 ret = opt_remove_dead_code(env); 19506 if (ret == 0) 19507 ret = opt_remove_nops(env); 19508 } else { 19509 if (ret == 0) 19510 sanitize_dead_code(env); 19511 } 19512 19513 if (ret == 0) 19514 /* program is valid, convert *(u32*)(ctx + off) accesses */ 19515 ret = convert_ctx_accesses(env); 19516 19517 if (ret == 0) 19518 ret = do_misc_fixups(env); 19519 19520 /* do 32-bit optimization after insn patching has done so those patched 19521 * insns could be handled correctly. 19522 */ 19523 if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) { 19524 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr); 19525 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret 19526 : false; 19527 } 19528 19529 if (ret == 0) 19530 ret = fixup_call_args(env); 19531 19532 env->verification_time = ktime_get_ns() - start_time; 19533 print_verification_stats(env); 19534 env->prog->aux->verified_insns = env->insn_processed; 19535 19536 /* preserve original error even if log finalization is successful */ 19537 err = bpf_vlog_finalize(&env->log, &log_true_size); 19538 if (err) 19539 ret = err; 19540 19541 if (uattr_size >= offsetofend(union bpf_attr, log_true_size) && 19542 copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size), 19543 &log_true_size, sizeof(log_true_size))) { 19544 ret = -EFAULT; 19545 goto err_release_maps; 19546 } 19547 19548 if (ret) 19549 goto err_release_maps; 19550 19551 if (env->used_map_cnt) { 19552 /* if program passed verifier, update used_maps in bpf_prog_info */ 19553 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt, 19554 sizeof(env->used_maps[0]), 19555 GFP_KERNEL); 19556 19557 if (!env->prog->aux->used_maps) { 19558 ret = -ENOMEM; 19559 goto err_release_maps; 19560 } 19561 19562 memcpy(env->prog->aux->used_maps, env->used_maps, 19563 sizeof(env->used_maps[0]) * env->used_map_cnt); 19564 env->prog->aux->used_map_cnt = env->used_map_cnt; 19565 } 19566 if (env->used_btf_cnt) { 19567 /* if program passed verifier, update used_btfs in bpf_prog_aux */ 19568 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt, 19569 sizeof(env->used_btfs[0]), 19570 GFP_KERNEL); 19571 if (!env->prog->aux->used_btfs) { 19572 ret = -ENOMEM; 19573 goto err_release_maps; 19574 } 19575 19576 memcpy(env->prog->aux->used_btfs, env->used_btfs, 19577 sizeof(env->used_btfs[0]) * env->used_btf_cnt); 19578 env->prog->aux->used_btf_cnt = env->used_btf_cnt; 19579 } 19580 if (env->used_map_cnt || env->used_btf_cnt) { 19581 /* program is valid. Convert pseudo bpf_ld_imm64 into generic 19582 * bpf_ld_imm64 instructions 19583 */ 19584 convert_pseudo_ld_imm64(env); 19585 } 19586 19587 adjust_btf_func(env); 19588 19589 err_release_maps: 19590 if (!env->prog->aux->used_maps) 19591 /* if we didn't copy map pointers into bpf_prog_info, release 19592 * them now. Otherwise free_used_maps() will release them. 19593 */ 19594 release_maps(env); 19595 if (!env->prog->aux->used_btfs) 19596 release_btfs(env); 19597 19598 /* extension progs temporarily inherit the attach_type of their targets 19599 for verification purposes, so set it back to zero before returning 19600 */ 19601 if (env->prog->type == BPF_PROG_TYPE_EXT) 19602 env->prog->expected_attach_type = 0; 19603 19604 *prog = env->prog; 19605 err_unlock: 19606 if (!is_priv) 19607 mutex_unlock(&bpf_verifier_lock); 19608 vfree(env->insn_aux_data); 19609 err_free_env: 19610 kfree(env); 19611 return ret; 19612 } 19613