1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * BPF Jit compiler for s390. 4 * 5 * Minimum build requirements: 6 * 7 * - HAVE_MARCH_Z196_FEATURES: laal, laalg 8 * - HAVE_MARCH_Z10_FEATURES: msfi, cgrj, clgrj 9 * - HAVE_MARCH_Z9_109_FEATURES: alfi, llilf, clfi, oilf, nilf 10 * - 64BIT 11 * 12 * Copyright IBM Corp. 2012,2015 13 * 14 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com> 15 * Michael Holzheu <holzheu@linux.vnet.ibm.com> 16 */ 17 18 #define KMSG_COMPONENT "bpf_jit" 19 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 20 21 #include <linux/netdevice.h> 22 #include <linux/filter.h> 23 #include <linux/init.h> 24 #include <linux/bpf.h> 25 #include <linux/mm.h> 26 #include <linux/kernel.h> 27 #include <asm/cacheflush.h> 28 #include <asm/extable.h> 29 #include <asm/dis.h> 30 #include <asm/facility.h> 31 #include <asm/nospec-branch.h> 32 #include <asm/set_memory.h> 33 #include <asm/text-patching.h> 34 #include "bpf_jit.h" 35 36 struct bpf_jit { 37 u32 seen; /* Flags to remember seen eBPF instructions */ 38 u32 seen_reg[16]; /* Array to remember which registers are used */ 39 u32 *addrs; /* Array with relative instruction addresses */ 40 u8 *prg_buf; /* Start of program */ 41 int size; /* Size of program and literal pool */ 42 int size_prg; /* Size of program */ 43 int prg; /* Current position in program */ 44 int lit32_start; /* Start of 32-bit literal pool */ 45 int lit32; /* Current position in 32-bit literal pool */ 46 int lit64_start; /* Start of 64-bit literal pool */ 47 int lit64; /* Current position in 64-bit literal pool */ 48 int base_ip; /* Base address for literal pool */ 49 int exit_ip; /* Address of exit */ 50 int r1_thunk_ip; /* Address of expoline thunk for 'br %r1' */ 51 int r14_thunk_ip; /* Address of expoline thunk for 'br %r14' */ 52 int tail_call_start; /* Tail call start offset */ 53 int excnt; /* Number of exception table entries */ 54 int prologue_plt_ret; /* Return address for prologue hotpatch PLT */ 55 int prologue_plt; /* Start of prologue hotpatch PLT */ 56 }; 57 58 #define SEEN_MEM BIT(0) /* use mem[] for temporary storage */ 59 #define SEEN_LITERAL BIT(1) /* code uses literals */ 60 #define SEEN_FUNC BIT(2) /* calls C functions */ 61 #define SEEN_STACK (SEEN_FUNC | SEEN_MEM) 62 63 /* 64 * s390 registers 65 */ 66 #define REG_W0 (MAX_BPF_JIT_REG + 0) /* Work register 1 (even) */ 67 #define REG_W1 (MAX_BPF_JIT_REG + 1) /* Work register 2 (odd) */ 68 #define REG_L (MAX_BPF_JIT_REG + 2) /* Literal pool register */ 69 #define REG_15 (MAX_BPF_JIT_REG + 3) /* Register 15 */ 70 #define REG_0 REG_W0 /* Register 0 */ 71 #define REG_1 REG_W1 /* Register 1 */ 72 #define REG_2 BPF_REG_1 /* Register 2 */ 73 #define REG_3 BPF_REG_2 /* Register 3 */ 74 #define REG_4 BPF_REG_3 /* Register 4 */ 75 #define REG_7 BPF_REG_6 /* Register 7 */ 76 #define REG_8 BPF_REG_7 /* Register 8 */ 77 #define REG_14 BPF_REG_0 /* Register 14 */ 78 79 /* 80 * Mapping of BPF registers to s390 registers 81 */ 82 static const int reg2hex[] = { 83 /* Return code */ 84 [BPF_REG_0] = 14, 85 /* Function parameters */ 86 [BPF_REG_1] = 2, 87 [BPF_REG_2] = 3, 88 [BPF_REG_3] = 4, 89 [BPF_REG_4] = 5, 90 [BPF_REG_5] = 6, 91 /* Call saved registers */ 92 [BPF_REG_6] = 7, 93 [BPF_REG_7] = 8, 94 [BPF_REG_8] = 9, 95 [BPF_REG_9] = 10, 96 /* BPF stack pointer */ 97 [BPF_REG_FP] = 13, 98 /* Register for blinding */ 99 [BPF_REG_AX] = 12, 100 /* Work registers for s390x backend */ 101 [REG_W0] = 0, 102 [REG_W1] = 1, 103 [REG_L] = 11, 104 [REG_15] = 15, 105 }; 106 107 static inline u32 reg(u32 dst_reg, u32 src_reg) 108 { 109 return reg2hex[dst_reg] << 4 | reg2hex[src_reg]; 110 } 111 112 static inline u32 reg_high(u32 reg) 113 { 114 return reg2hex[reg] << 4; 115 } 116 117 static inline void reg_set_seen(struct bpf_jit *jit, u32 b1) 118 { 119 u32 r1 = reg2hex[b1]; 120 121 if (r1 >= 6 && r1 <= 15 && !jit->seen_reg[r1]) 122 jit->seen_reg[r1] = 1; 123 } 124 125 #define REG_SET_SEEN(b1) \ 126 ({ \ 127 reg_set_seen(jit, b1); \ 128 }) 129 130 #define REG_SEEN(b1) jit->seen_reg[reg2hex[(b1)]] 131 132 /* 133 * EMIT macros for code generation 134 */ 135 136 #define _EMIT2(op) \ 137 ({ \ 138 if (jit->prg_buf) \ 139 *(u16 *) (jit->prg_buf + jit->prg) = (op); \ 140 jit->prg += 2; \ 141 }) 142 143 #define EMIT2(op, b1, b2) \ 144 ({ \ 145 _EMIT2((op) | reg(b1, b2)); \ 146 REG_SET_SEEN(b1); \ 147 REG_SET_SEEN(b2); \ 148 }) 149 150 #define _EMIT4(op) \ 151 ({ \ 152 if (jit->prg_buf) \ 153 *(u32 *) (jit->prg_buf + jit->prg) = (op); \ 154 jit->prg += 4; \ 155 }) 156 157 #define EMIT4(op, b1, b2) \ 158 ({ \ 159 _EMIT4((op) | reg(b1, b2)); \ 160 REG_SET_SEEN(b1); \ 161 REG_SET_SEEN(b2); \ 162 }) 163 164 #define EMIT4_RRF(op, b1, b2, b3) \ 165 ({ \ 166 _EMIT4((op) | reg_high(b3) << 8 | reg(b1, b2)); \ 167 REG_SET_SEEN(b1); \ 168 REG_SET_SEEN(b2); \ 169 REG_SET_SEEN(b3); \ 170 }) 171 172 #define _EMIT4_DISP(op, disp) \ 173 ({ \ 174 unsigned int __disp = (disp) & 0xfff; \ 175 _EMIT4((op) | __disp); \ 176 }) 177 178 #define EMIT4_DISP(op, b1, b2, disp) \ 179 ({ \ 180 _EMIT4_DISP((op) | reg_high(b1) << 16 | \ 181 reg_high(b2) << 8, (disp)); \ 182 REG_SET_SEEN(b1); \ 183 REG_SET_SEEN(b2); \ 184 }) 185 186 #define EMIT4_IMM(op, b1, imm) \ 187 ({ \ 188 unsigned int __imm = (imm) & 0xffff; \ 189 _EMIT4((op) | reg_high(b1) << 16 | __imm); \ 190 REG_SET_SEEN(b1); \ 191 }) 192 193 #define EMIT4_PCREL(op, pcrel) \ 194 ({ \ 195 long __pcrel = ((pcrel) >> 1) & 0xffff; \ 196 _EMIT4((op) | __pcrel); \ 197 }) 198 199 #define EMIT4_PCREL_RIC(op, mask, target) \ 200 ({ \ 201 int __rel = ((target) - jit->prg) / 2; \ 202 _EMIT4((op) | (mask) << 20 | (__rel & 0xffff)); \ 203 }) 204 205 #define _EMIT6(op1, op2) \ 206 ({ \ 207 if (jit->prg_buf) { \ 208 *(u32 *) (jit->prg_buf + jit->prg) = (op1); \ 209 *(u16 *) (jit->prg_buf + jit->prg + 4) = (op2); \ 210 } \ 211 jit->prg += 6; \ 212 }) 213 214 #define _EMIT6_DISP(op1, op2, disp) \ 215 ({ \ 216 unsigned int __disp = (disp) & 0xfff; \ 217 _EMIT6((op1) | __disp, op2); \ 218 }) 219 220 #define _EMIT6_DISP_LH(op1, op2, disp) \ 221 ({ \ 222 u32 _disp = (u32) (disp); \ 223 unsigned int __disp_h = _disp & 0xff000; \ 224 unsigned int __disp_l = _disp & 0x00fff; \ 225 _EMIT6((op1) | __disp_l, (op2) | __disp_h >> 4); \ 226 }) 227 228 #define EMIT6_DISP_LH(op1, op2, b1, b2, b3, disp) \ 229 ({ \ 230 _EMIT6_DISP_LH((op1) | reg(b1, b2) << 16 | \ 231 reg_high(b3) << 8, op2, disp); \ 232 REG_SET_SEEN(b1); \ 233 REG_SET_SEEN(b2); \ 234 REG_SET_SEEN(b3); \ 235 }) 236 237 #define EMIT6_PCREL_RIEB(op1, op2, b1, b2, mask, target) \ 238 ({ \ 239 unsigned int rel = (int)((target) - jit->prg) / 2; \ 240 _EMIT6((op1) | reg(b1, b2) << 16 | (rel & 0xffff), \ 241 (op2) | (mask) << 12); \ 242 REG_SET_SEEN(b1); \ 243 REG_SET_SEEN(b2); \ 244 }) 245 246 #define EMIT6_PCREL_RIEC(op1, op2, b1, imm, mask, target) \ 247 ({ \ 248 unsigned int rel = (int)((target) - jit->prg) / 2; \ 249 _EMIT6((op1) | (reg_high(b1) | (mask)) << 16 | \ 250 (rel & 0xffff), (op2) | ((imm) & 0xff) << 8); \ 251 REG_SET_SEEN(b1); \ 252 BUILD_BUG_ON(((unsigned long) (imm)) > 0xff); \ 253 }) 254 255 #define EMIT6_PCREL(op1, op2, b1, b2, i, off, mask) \ 256 ({ \ 257 int rel = (addrs[(i) + (off) + 1] - jit->prg) / 2; \ 258 _EMIT6((op1) | reg(b1, b2) << 16 | (rel & 0xffff), (op2) | (mask));\ 259 REG_SET_SEEN(b1); \ 260 REG_SET_SEEN(b2); \ 261 }) 262 263 #define EMIT6_PCREL_RILB(op, b, target) \ 264 ({ \ 265 unsigned int rel = (int)((target) - jit->prg) / 2; \ 266 _EMIT6((op) | reg_high(b) << 16 | rel >> 16, rel & 0xffff);\ 267 REG_SET_SEEN(b); \ 268 }) 269 270 #define EMIT6_PCREL_RIL(op, target) \ 271 ({ \ 272 unsigned int rel = (int)((target) - jit->prg) / 2; \ 273 _EMIT6((op) | rel >> 16, rel & 0xffff); \ 274 }) 275 276 #define EMIT6_PCREL_RILC(op, mask, target) \ 277 ({ \ 278 EMIT6_PCREL_RIL((op) | (mask) << 20, (target)); \ 279 }) 280 281 #define _EMIT6_IMM(op, imm) \ 282 ({ \ 283 unsigned int __imm = (imm); \ 284 _EMIT6((op) | (__imm >> 16), __imm & 0xffff); \ 285 }) 286 287 #define EMIT6_IMM(op, b1, imm) \ 288 ({ \ 289 _EMIT6_IMM((op) | reg_high(b1) << 16, imm); \ 290 REG_SET_SEEN(b1); \ 291 }) 292 293 #define _EMIT_CONST_U32(val) \ 294 ({ \ 295 unsigned int ret; \ 296 ret = jit->lit32; \ 297 if (jit->prg_buf) \ 298 *(u32 *)(jit->prg_buf + jit->lit32) = (u32)(val);\ 299 jit->lit32 += 4; \ 300 ret; \ 301 }) 302 303 #define EMIT_CONST_U32(val) \ 304 ({ \ 305 jit->seen |= SEEN_LITERAL; \ 306 _EMIT_CONST_U32(val) - jit->base_ip; \ 307 }) 308 309 #define _EMIT_CONST_U64(val) \ 310 ({ \ 311 unsigned int ret; \ 312 ret = jit->lit64; \ 313 if (jit->prg_buf) \ 314 *(u64 *)(jit->prg_buf + jit->lit64) = (u64)(val);\ 315 jit->lit64 += 8; \ 316 ret; \ 317 }) 318 319 #define EMIT_CONST_U64(val) \ 320 ({ \ 321 jit->seen |= SEEN_LITERAL; \ 322 _EMIT_CONST_U64(val) - jit->base_ip; \ 323 }) 324 325 #define EMIT_ZERO(b1) \ 326 ({ \ 327 if (!fp->aux->verifier_zext) { \ 328 /* llgfr %dst,%dst (zero extend to 64 bit) */ \ 329 EMIT4(0xb9160000, b1, b1); \ 330 REG_SET_SEEN(b1); \ 331 } \ 332 }) 333 334 /* 335 * Return whether this is the first pass. The first pass is special, since we 336 * don't know any sizes yet, and thus must be conservative. 337 */ 338 static bool is_first_pass(struct bpf_jit *jit) 339 { 340 return jit->size == 0; 341 } 342 343 /* 344 * Return whether this is the code generation pass. The code generation pass is 345 * special, since we should change as little as possible. 346 */ 347 static bool is_codegen_pass(struct bpf_jit *jit) 348 { 349 return jit->prg_buf; 350 } 351 352 /* 353 * Return whether "rel" can be encoded as a short PC-relative offset 354 */ 355 static bool is_valid_rel(int rel) 356 { 357 return rel >= -65536 && rel <= 65534; 358 } 359 360 /* 361 * Return whether "off" can be reached using a short PC-relative offset 362 */ 363 static bool can_use_rel(struct bpf_jit *jit, int off) 364 { 365 return is_valid_rel(off - jit->prg); 366 } 367 368 /* 369 * Return whether given displacement can be encoded using 370 * Long-Displacement Facility 371 */ 372 static bool is_valid_ldisp(int disp) 373 { 374 return disp >= -524288 && disp <= 524287; 375 } 376 377 /* 378 * Return whether the next 32-bit literal pool entry can be referenced using 379 * Long-Displacement Facility 380 */ 381 static bool can_use_ldisp_for_lit32(struct bpf_jit *jit) 382 { 383 return is_valid_ldisp(jit->lit32 - jit->base_ip); 384 } 385 386 /* 387 * Return whether the next 64-bit literal pool entry can be referenced using 388 * Long-Displacement Facility 389 */ 390 static bool can_use_ldisp_for_lit64(struct bpf_jit *jit) 391 { 392 return is_valid_ldisp(jit->lit64 - jit->base_ip); 393 } 394 395 /* 396 * Fill whole space with illegal instructions 397 */ 398 static void jit_fill_hole(void *area, unsigned int size) 399 { 400 memset(area, 0, size); 401 } 402 403 /* 404 * Save registers from "rs" (register start) to "re" (register end) on stack 405 */ 406 static void save_regs(struct bpf_jit *jit, u32 rs, u32 re) 407 { 408 u32 off = STK_OFF_R6 + (rs - 6) * 8; 409 410 if (rs == re) 411 /* stg %rs,off(%r15) */ 412 _EMIT6(0xe300f000 | rs << 20 | off, 0x0024); 413 else 414 /* stmg %rs,%re,off(%r15) */ 415 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0024, off); 416 } 417 418 /* 419 * Restore registers from "rs" (register start) to "re" (register end) on stack 420 */ 421 static void restore_regs(struct bpf_jit *jit, u32 rs, u32 re, u32 stack_depth) 422 { 423 u32 off = STK_OFF_R6 + (rs - 6) * 8; 424 425 if (jit->seen & SEEN_STACK) 426 off += STK_OFF + stack_depth; 427 428 if (rs == re) 429 /* lg %rs,off(%r15) */ 430 _EMIT6(0xe300f000 | rs << 20 | off, 0x0004); 431 else 432 /* lmg %rs,%re,off(%r15) */ 433 _EMIT6_DISP(0xeb00f000 | rs << 20 | re << 16, 0x0004, off); 434 } 435 436 /* 437 * Return first seen register (from start) 438 */ 439 static int get_start(struct bpf_jit *jit, int start) 440 { 441 int i; 442 443 for (i = start; i <= 15; i++) { 444 if (jit->seen_reg[i]) 445 return i; 446 } 447 return 0; 448 } 449 450 /* 451 * Return last seen register (from start) (gap >= 2) 452 */ 453 static int get_end(struct bpf_jit *jit, int start) 454 { 455 int i; 456 457 for (i = start; i < 15; i++) { 458 if (!jit->seen_reg[i] && !jit->seen_reg[i + 1]) 459 return i - 1; 460 } 461 return jit->seen_reg[15] ? 15 : 14; 462 } 463 464 #define REGS_SAVE 1 465 #define REGS_RESTORE 0 466 /* 467 * Save and restore clobbered registers (6-15) on stack. 468 * We save/restore registers in chunks with gap >= 2 registers. 469 */ 470 static void save_restore_regs(struct bpf_jit *jit, int op, u32 stack_depth) 471 { 472 const int last = 15, save_restore_size = 6; 473 int re = 6, rs; 474 475 if (is_first_pass(jit)) { 476 /* 477 * We don't know yet which registers are used. Reserve space 478 * conservatively. 479 */ 480 jit->prg += (last - re + 1) * save_restore_size; 481 return; 482 } 483 484 do { 485 rs = get_start(jit, re); 486 if (!rs) 487 break; 488 re = get_end(jit, rs + 1); 489 if (op == REGS_SAVE) 490 save_regs(jit, rs, re); 491 else 492 restore_regs(jit, rs, re, stack_depth); 493 re++; 494 } while (re <= last); 495 } 496 497 static void bpf_skip(struct bpf_jit *jit, int size) 498 { 499 if (size >= 6 && !is_valid_rel(size)) { 500 /* brcl 0xf,size */ 501 EMIT6_PCREL_RIL(0xc0f4000000, size); 502 size -= 6; 503 } else if (size >= 4 && is_valid_rel(size)) { 504 /* brc 0xf,size */ 505 EMIT4_PCREL(0xa7f40000, size); 506 size -= 4; 507 } 508 while (size >= 2) { 509 /* bcr 0,%0 */ 510 _EMIT2(0x0700); 511 size -= 2; 512 } 513 } 514 515 /* 516 * PLT for hotpatchable calls. The calling convention is the same as for the 517 * ftrace hotpatch trampolines: %r0 is return address, %r1 is clobbered. 518 */ 519 extern const char bpf_plt[]; 520 extern const char bpf_plt_ret[]; 521 extern const char bpf_plt_target[]; 522 extern const char bpf_plt_end[]; 523 #define BPF_PLT_SIZE 32 524 asm( 525 ".pushsection .rodata\n" 526 " .align 8\n" 527 "bpf_plt:\n" 528 " lgrl %r0,bpf_plt_ret\n" 529 " lgrl %r1,bpf_plt_target\n" 530 " br %r1\n" 531 " .align 8\n" 532 "bpf_plt_ret: .quad 0\n" 533 "bpf_plt_target: .quad 0\n" 534 "bpf_plt_end:\n" 535 " .popsection\n" 536 ); 537 538 static void bpf_jit_plt(void *plt, void *ret, void *target) 539 { 540 memcpy(plt, bpf_plt, BPF_PLT_SIZE); 541 *(void **)((char *)plt + (bpf_plt_ret - bpf_plt)) = ret; 542 *(void **)((char *)plt + (bpf_plt_target - bpf_plt)) = target ?: ret; 543 } 544 545 /* 546 * Emit function prologue 547 * 548 * Save registers and create stack frame if necessary. 549 * See stack frame layout description in "bpf_jit.h"! 550 */ 551 static void bpf_jit_prologue(struct bpf_jit *jit, struct bpf_prog *fp, 552 u32 stack_depth) 553 { 554 /* No-op for hotpatching */ 555 /* brcl 0,prologue_plt */ 556 EMIT6_PCREL_RILC(0xc0040000, 0, jit->prologue_plt); 557 jit->prologue_plt_ret = jit->prg; 558 559 if (fp->aux->func_idx == 0) { 560 /* Initialize the tail call counter in the main program. */ 561 /* xc STK_OFF_TCCNT(4,%r15),STK_OFF_TCCNT(%r15) */ 562 _EMIT6(0xd703f000 | STK_OFF_TCCNT, 0xf000 | STK_OFF_TCCNT); 563 } else { 564 /* 565 * Skip the tail call counter initialization in subprograms. 566 * Insert nops in order to have tail_call_start at a 567 * predictable offset. 568 */ 569 bpf_skip(jit, 6); 570 } 571 /* Tail calls have to skip above initialization */ 572 jit->tail_call_start = jit->prg; 573 /* Save registers */ 574 save_restore_regs(jit, REGS_SAVE, stack_depth); 575 /* Setup literal pool */ 576 if (is_first_pass(jit) || (jit->seen & SEEN_LITERAL)) { 577 if (!is_first_pass(jit) && 578 is_valid_ldisp(jit->size - (jit->prg + 2))) { 579 /* basr %l,0 */ 580 EMIT2(0x0d00, REG_L, REG_0); 581 jit->base_ip = jit->prg; 582 } else { 583 /* larl %l,lit32_start */ 584 EMIT6_PCREL_RILB(0xc0000000, REG_L, jit->lit32_start); 585 jit->base_ip = jit->lit32_start; 586 } 587 } 588 /* Setup stack and backchain */ 589 if (is_first_pass(jit) || (jit->seen & SEEN_STACK)) { 590 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC)) 591 /* lgr %w1,%r15 (backchain) */ 592 EMIT4(0xb9040000, REG_W1, REG_15); 593 /* la %bfp,STK_160_UNUSED(%r15) (BPF frame pointer) */ 594 EMIT4_DISP(0x41000000, BPF_REG_FP, REG_15, STK_160_UNUSED); 595 /* aghi %r15,-STK_OFF */ 596 EMIT4_IMM(0xa70b0000, REG_15, -(STK_OFF + stack_depth)); 597 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC)) 598 /* stg %w1,152(%r15) (backchain) */ 599 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W1, REG_0, 600 REG_15, 152); 601 } 602 } 603 604 /* 605 * Emit an expoline for a jump that follows 606 */ 607 static void emit_expoline(struct bpf_jit *jit) 608 { 609 /* exrl %r0,.+10 */ 610 EMIT6_PCREL_RIL(0xc6000000, jit->prg + 10); 611 /* j . */ 612 EMIT4_PCREL(0xa7f40000, 0); 613 } 614 615 /* 616 * Emit __s390_indirect_jump_r1 thunk if necessary 617 */ 618 static void emit_r1_thunk(struct bpf_jit *jit) 619 { 620 if (nospec_uses_trampoline()) { 621 jit->r1_thunk_ip = jit->prg; 622 emit_expoline(jit); 623 /* br %r1 */ 624 _EMIT2(0x07f1); 625 } 626 } 627 628 /* 629 * Call r1 either directly or via __s390_indirect_jump_r1 thunk 630 */ 631 static void call_r1(struct bpf_jit *jit) 632 { 633 if (nospec_uses_trampoline()) 634 /* brasl %r14,__s390_indirect_jump_r1 */ 635 EMIT6_PCREL_RILB(0xc0050000, REG_14, jit->r1_thunk_ip); 636 else 637 /* basr %r14,%r1 */ 638 EMIT2(0x0d00, REG_14, REG_1); 639 } 640 641 /* 642 * Function epilogue 643 */ 644 static void bpf_jit_epilogue(struct bpf_jit *jit, u32 stack_depth) 645 { 646 jit->exit_ip = jit->prg; 647 /* Load exit code: lgr %r2,%b0 */ 648 EMIT4(0xb9040000, REG_2, BPF_REG_0); 649 /* Restore registers */ 650 save_restore_regs(jit, REGS_RESTORE, stack_depth); 651 if (nospec_uses_trampoline()) { 652 jit->r14_thunk_ip = jit->prg; 653 /* Generate __s390_indirect_jump_r14 thunk */ 654 emit_expoline(jit); 655 } 656 /* br %r14 */ 657 _EMIT2(0x07fe); 658 659 if (is_first_pass(jit) || (jit->seen & SEEN_FUNC)) 660 emit_r1_thunk(jit); 661 662 jit->prg = ALIGN(jit->prg, 8); 663 jit->prologue_plt = jit->prg; 664 if (jit->prg_buf) 665 bpf_jit_plt(jit->prg_buf + jit->prg, 666 jit->prg_buf + jit->prologue_plt_ret, NULL); 667 jit->prg += BPF_PLT_SIZE; 668 } 669 670 static int get_probe_mem_regno(const u8 *insn) 671 { 672 /* 673 * insn must point to llgc, llgh, llgf or lg, which have destination 674 * register at the same position. 675 */ 676 if (insn[0] != 0xe3) /* common llgc, llgh, llgf and lg prefix */ 677 return -1; 678 if (insn[5] != 0x90 && /* llgc */ 679 insn[5] != 0x91 && /* llgh */ 680 insn[5] != 0x16 && /* llgf */ 681 insn[5] != 0x04) /* lg */ 682 return -1; 683 return insn[1] >> 4; 684 } 685 686 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs) 687 { 688 regs->psw.addr = extable_fixup(x); 689 regs->gprs[x->data] = 0; 690 return true; 691 } 692 693 static int bpf_jit_probe_mem(struct bpf_jit *jit, struct bpf_prog *fp, 694 int probe_prg, int nop_prg) 695 { 696 struct exception_table_entry *ex; 697 int reg, prg; 698 s64 delta; 699 u8 *insn; 700 int i; 701 702 if (!fp->aux->extable) 703 /* Do nothing during early JIT passes. */ 704 return 0; 705 insn = jit->prg_buf + probe_prg; 706 reg = get_probe_mem_regno(insn); 707 if (WARN_ON_ONCE(reg < 0)) 708 /* JIT bug - unexpected probe instruction. */ 709 return -1; 710 if (WARN_ON_ONCE(probe_prg + insn_length(*insn) != nop_prg)) 711 /* JIT bug - gap between probe and nop instructions. */ 712 return -1; 713 for (i = 0; i < 2; i++) { 714 if (WARN_ON_ONCE(jit->excnt >= fp->aux->num_exentries)) 715 /* Verifier bug - not enough entries. */ 716 return -1; 717 ex = &fp->aux->extable[jit->excnt]; 718 /* Add extable entries for probe and nop instructions. */ 719 prg = i == 0 ? probe_prg : nop_prg; 720 delta = jit->prg_buf + prg - (u8 *)&ex->insn; 721 if (WARN_ON_ONCE(delta < INT_MIN || delta > INT_MAX)) 722 /* JIT bug - code and extable must be close. */ 723 return -1; 724 ex->insn = delta; 725 /* 726 * Always land on the nop. Note that extable infrastructure 727 * ignores fixup field, it is handled by ex_handler_bpf(). 728 */ 729 delta = jit->prg_buf + nop_prg - (u8 *)&ex->fixup; 730 if (WARN_ON_ONCE(delta < INT_MIN || delta > INT_MAX)) 731 /* JIT bug - landing pad and extable must be close. */ 732 return -1; 733 ex->fixup = delta; 734 ex->type = EX_TYPE_BPF; 735 ex->data = reg; 736 jit->excnt++; 737 } 738 return 0; 739 } 740 741 /* 742 * Sign-extend the register if necessary 743 */ 744 static int sign_extend(struct bpf_jit *jit, int r, u8 size, u8 flags) 745 { 746 if (!(flags & BTF_FMODEL_SIGNED_ARG)) 747 return 0; 748 749 switch (size) { 750 case 1: 751 /* lgbr %r,%r */ 752 EMIT4(0xb9060000, r, r); 753 return 0; 754 case 2: 755 /* lghr %r,%r */ 756 EMIT4(0xb9070000, r, r); 757 return 0; 758 case 4: 759 /* lgfr %r,%r */ 760 EMIT4(0xb9140000, r, r); 761 return 0; 762 case 8: 763 return 0; 764 default: 765 return -1; 766 } 767 } 768 769 /* 770 * Compile one eBPF instruction into s390x code 771 * 772 * NOTE: Use noinline because for gcov (-fprofile-arcs) gcc allocates a lot of 773 * stack space for the large switch statement. 774 */ 775 static noinline int bpf_jit_insn(struct bpf_jit *jit, struct bpf_prog *fp, 776 int i, bool extra_pass, u32 stack_depth) 777 { 778 struct bpf_insn *insn = &fp->insnsi[i]; 779 u32 dst_reg = insn->dst_reg; 780 u32 src_reg = insn->src_reg; 781 int last, insn_count = 1; 782 u32 *addrs = jit->addrs; 783 s32 imm = insn->imm; 784 s16 off = insn->off; 785 int probe_prg = -1; 786 unsigned int mask; 787 int nop_prg; 788 int err; 789 790 if (BPF_CLASS(insn->code) == BPF_LDX && 791 BPF_MODE(insn->code) == BPF_PROBE_MEM) 792 probe_prg = jit->prg; 793 794 switch (insn->code) { 795 /* 796 * BPF_MOV 797 */ 798 case BPF_ALU | BPF_MOV | BPF_X: /* dst = (u32) src */ 799 /* llgfr %dst,%src */ 800 EMIT4(0xb9160000, dst_reg, src_reg); 801 if (insn_is_zext(&insn[1])) 802 insn_count = 2; 803 break; 804 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */ 805 /* lgr %dst,%src */ 806 EMIT4(0xb9040000, dst_reg, src_reg); 807 break; 808 case BPF_ALU | BPF_MOV | BPF_K: /* dst = (u32) imm */ 809 /* llilf %dst,imm */ 810 EMIT6_IMM(0xc00f0000, dst_reg, imm); 811 if (insn_is_zext(&insn[1])) 812 insn_count = 2; 813 break; 814 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = imm */ 815 /* lgfi %dst,imm */ 816 EMIT6_IMM(0xc0010000, dst_reg, imm); 817 break; 818 /* 819 * BPF_LD 64 820 */ 821 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */ 822 { 823 /* 16 byte instruction that uses two 'struct bpf_insn' */ 824 u64 imm64; 825 826 imm64 = (u64)(u32) insn[0].imm | ((u64)(u32) insn[1].imm) << 32; 827 /* lgrl %dst,imm */ 828 EMIT6_PCREL_RILB(0xc4080000, dst_reg, _EMIT_CONST_U64(imm64)); 829 insn_count = 2; 830 break; 831 } 832 /* 833 * BPF_ADD 834 */ 835 case BPF_ALU | BPF_ADD | BPF_X: /* dst = (u32) dst + (u32) src */ 836 /* ar %dst,%src */ 837 EMIT2(0x1a00, dst_reg, src_reg); 838 EMIT_ZERO(dst_reg); 839 break; 840 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst = dst + src */ 841 /* agr %dst,%src */ 842 EMIT4(0xb9080000, dst_reg, src_reg); 843 break; 844 case BPF_ALU | BPF_ADD | BPF_K: /* dst = (u32) dst + (u32) imm */ 845 if (imm != 0) { 846 /* alfi %dst,imm */ 847 EMIT6_IMM(0xc20b0000, dst_reg, imm); 848 } 849 EMIT_ZERO(dst_reg); 850 break; 851 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst = dst + imm */ 852 if (!imm) 853 break; 854 /* agfi %dst,imm */ 855 EMIT6_IMM(0xc2080000, dst_reg, imm); 856 break; 857 /* 858 * BPF_SUB 859 */ 860 case BPF_ALU | BPF_SUB | BPF_X: /* dst = (u32) dst - (u32) src */ 861 /* sr %dst,%src */ 862 EMIT2(0x1b00, dst_reg, src_reg); 863 EMIT_ZERO(dst_reg); 864 break; 865 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst = dst - src */ 866 /* sgr %dst,%src */ 867 EMIT4(0xb9090000, dst_reg, src_reg); 868 break; 869 case BPF_ALU | BPF_SUB | BPF_K: /* dst = (u32) dst - (u32) imm */ 870 if (imm != 0) { 871 /* alfi %dst,-imm */ 872 EMIT6_IMM(0xc20b0000, dst_reg, -imm); 873 } 874 EMIT_ZERO(dst_reg); 875 break; 876 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst = dst - imm */ 877 if (!imm) 878 break; 879 if (imm == -0x80000000) { 880 /* algfi %dst,0x80000000 */ 881 EMIT6_IMM(0xc20a0000, dst_reg, 0x80000000); 882 } else { 883 /* agfi %dst,-imm */ 884 EMIT6_IMM(0xc2080000, dst_reg, -imm); 885 } 886 break; 887 /* 888 * BPF_MUL 889 */ 890 case BPF_ALU | BPF_MUL | BPF_X: /* dst = (u32) dst * (u32) src */ 891 /* msr %dst,%src */ 892 EMIT4(0xb2520000, dst_reg, src_reg); 893 EMIT_ZERO(dst_reg); 894 break; 895 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst = dst * src */ 896 /* msgr %dst,%src */ 897 EMIT4(0xb90c0000, dst_reg, src_reg); 898 break; 899 case BPF_ALU | BPF_MUL | BPF_K: /* dst = (u32) dst * (u32) imm */ 900 if (imm != 1) { 901 /* msfi %r5,imm */ 902 EMIT6_IMM(0xc2010000, dst_reg, imm); 903 } 904 EMIT_ZERO(dst_reg); 905 break; 906 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst = dst * imm */ 907 if (imm == 1) 908 break; 909 /* msgfi %dst,imm */ 910 EMIT6_IMM(0xc2000000, dst_reg, imm); 911 break; 912 /* 913 * BPF_DIV / BPF_MOD 914 */ 915 case BPF_ALU | BPF_DIV | BPF_X: /* dst = (u32) dst / (u32) src */ 916 case BPF_ALU | BPF_MOD | BPF_X: /* dst = (u32) dst % (u32) src */ 917 { 918 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0; 919 920 /* lhi %w0,0 */ 921 EMIT4_IMM(0xa7080000, REG_W0, 0); 922 /* lr %w1,%dst */ 923 EMIT2(0x1800, REG_W1, dst_reg); 924 /* dlr %w0,%src */ 925 EMIT4(0xb9970000, REG_W0, src_reg); 926 /* llgfr %dst,%rc */ 927 EMIT4(0xb9160000, dst_reg, rc_reg); 928 if (insn_is_zext(&insn[1])) 929 insn_count = 2; 930 break; 931 } 932 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst = dst / src */ 933 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst = dst % src */ 934 { 935 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0; 936 937 /* lghi %w0,0 */ 938 EMIT4_IMM(0xa7090000, REG_W0, 0); 939 /* lgr %w1,%dst */ 940 EMIT4(0xb9040000, REG_W1, dst_reg); 941 /* dlgr %w0,%dst */ 942 EMIT4(0xb9870000, REG_W0, src_reg); 943 /* lgr %dst,%rc */ 944 EMIT4(0xb9040000, dst_reg, rc_reg); 945 break; 946 } 947 case BPF_ALU | BPF_DIV | BPF_K: /* dst = (u32) dst / (u32) imm */ 948 case BPF_ALU | BPF_MOD | BPF_K: /* dst = (u32) dst % (u32) imm */ 949 { 950 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0; 951 952 if (imm == 1) { 953 if (BPF_OP(insn->code) == BPF_MOD) 954 /* lhgi %dst,0 */ 955 EMIT4_IMM(0xa7090000, dst_reg, 0); 956 else 957 EMIT_ZERO(dst_reg); 958 break; 959 } 960 /* lhi %w0,0 */ 961 EMIT4_IMM(0xa7080000, REG_W0, 0); 962 /* lr %w1,%dst */ 963 EMIT2(0x1800, REG_W1, dst_reg); 964 if (!is_first_pass(jit) && can_use_ldisp_for_lit32(jit)) { 965 /* dl %w0,<d(imm)>(%l) */ 966 EMIT6_DISP_LH(0xe3000000, 0x0097, REG_W0, REG_0, REG_L, 967 EMIT_CONST_U32(imm)); 968 } else { 969 /* lgfrl %dst,imm */ 970 EMIT6_PCREL_RILB(0xc40c0000, dst_reg, 971 _EMIT_CONST_U32(imm)); 972 jit->seen |= SEEN_LITERAL; 973 /* dlr %w0,%dst */ 974 EMIT4(0xb9970000, REG_W0, dst_reg); 975 } 976 /* llgfr %dst,%rc */ 977 EMIT4(0xb9160000, dst_reg, rc_reg); 978 if (insn_is_zext(&insn[1])) 979 insn_count = 2; 980 break; 981 } 982 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst = dst / imm */ 983 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst = dst % imm */ 984 { 985 int rc_reg = BPF_OP(insn->code) == BPF_DIV ? REG_W1 : REG_W0; 986 987 if (imm == 1) { 988 if (BPF_OP(insn->code) == BPF_MOD) 989 /* lhgi %dst,0 */ 990 EMIT4_IMM(0xa7090000, dst_reg, 0); 991 break; 992 } 993 /* lghi %w0,0 */ 994 EMIT4_IMM(0xa7090000, REG_W0, 0); 995 /* lgr %w1,%dst */ 996 EMIT4(0xb9040000, REG_W1, dst_reg); 997 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) { 998 /* dlg %w0,<d(imm)>(%l) */ 999 EMIT6_DISP_LH(0xe3000000, 0x0087, REG_W0, REG_0, REG_L, 1000 EMIT_CONST_U64(imm)); 1001 } else { 1002 /* lgrl %dst,imm */ 1003 EMIT6_PCREL_RILB(0xc4080000, dst_reg, 1004 _EMIT_CONST_U64(imm)); 1005 jit->seen |= SEEN_LITERAL; 1006 /* dlgr %w0,%dst */ 1007 EMIT4(0xb9870000, REG_W0, dst_reg); 1008 } 1009 /* lgr %dst,%rc */ 1010 EMIT4(0xb9040000, dst_reg, rc_reg); 1011 break; 1012 } 1013 /* 1014 * BPF_AND 1015 */ 1016 case BPF_ALU | BPF_AND | BPF_X: /* dst = (u32) dst & (u32) src */ 1017 /* nr %dst,%src */ 1018 EMIT2(0x1400, dst_reg, src_reg); 1019 EMIT_ZERO(dst_reg); 1020 break; 1021 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */ 1022 /* ngr %dst,%src */ 1023 EMIT4(0xb9800000, dst_reg, src_reg); 1024 break; 1025 case BPF_ALU | BPF_AND | BPF_K: /* dst = (u32) dst & (u32) imm */ 1026 /* nilf %dst,imm */ 1027 EMIT6_IMM(0xc00b0000, dst_reg, imm); 1028 EMIT_ZERO(dst_reg); 1029 break; 1030 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */ 1031 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) { 1032 /* ng %dst,<d(imm)>(%l) */ 1033 EMIT6_DISP_LH(0xe3000000, 0x0080, 1034 dst_reg, REG_0, REG_L, 1035 EMIT_CONST_U64(imm)); 1036 } else { 1037 /* lgrl %w0,imm */ 1038 EMIT6_PCREL_RILB(0xc4080000, REG_W0, 1039 _EMIT_CONST_U64(imm)); 1040 jit->seen |= SEEN_LITERAL; 1041 /* ngr %dst,%w0 */ 1042 EMIT4(0xb9800000, dst_reg, REG_W0); 1043 } 1044 break; 1045 /* 1046 * BPF_OR 1047 */ 1048 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */ 1049 /* or %dst,%src */ 1050 EMIT2(0x1600, dst_reg, src_reg); 1051 EMIT_ZERO(dst_reg); 1052 break; 1053 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */ 1054 /* ogr %dst,%src */ 1055 EMIT4(0xb9810000, dst_reg, src_reg); 1056 break; 1057 case BPF_ALU | BPF_OR | BPF_K: /* dst = (u32) dst | (u32) imm */ 1058 /* oilf %dst,imm */ 1059 EMIT6_IMM(0xc00d0000, dst_reg, imm); 1060 EMIT_ZERO(dst_reg); 1061 break; 1062 case BPF_ALU64 | BPF_OR | BPF_K: /* dst = dst | imm */ 1063 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) { 1064 /* og %dst,<d(imm)>(%l) */ 1065 EMIT6_DISP_LH(0xe3000000, 0x0081, 1066 dst_reg, REG_0, REG_L, 1067 EMIT_CONST_U64(imm)); 1068 } else { 1069 /* lgrl %w0,imm */ 1070 EMIT6_PCREL_RILB(0xc4080000, REG_W0, 1071 _EMIT_CONST_U64(imm)); 1072 jit->seen |= SEEN_LITERAL; 1073 /* ogr %dst,%w0 */ 1074 EMIT4(0xb9810000, dst_reg, REG_W0); 1075 } 1076 break; 1077 /* 1078 * BPF_XOR 1079 */ 1080 case BPF_ALU | BPF_XOR | BPF_X: /* dst = (u32) dst ^ (u32) src */ 1081 /* xr %dst,%src */ 1082 EMIT2(0x1700, dst_reg, src_reg); 1083 EMIT_ZERO(dst_reg); 1084 break; 1085 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst = dst ^ src */ 1086 /* xgr %dst,%src */ 1087 EMIT4(0xb9820000, dst_reg, src_reg); 1088 break; 1089 case BPF_ALU | BPF_XOR | BPF_K: /* dst = (u32) dst ^ (u32) imm */ 1090 if (imm != 0) { 1091 /* xilf %dst,imm */ 1092 EMIT6_IMM(0xc0070000, dst_reg, imm); 1093 } 1094 EMIT_ZERO(dst_reg); 1095 break; 1096 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst = dst ^ imm */ 1097 if (!is_first_pass(jit) && can_use_ldisp_for_lit64(jit)) { 1098 /* xg %dst,<d(imm)>(%l) */ 1099 EMIT6_DISP_LH(0xe3000000, 0x0082, 1100 dst_reg, REG_0, REG_L, 1101 EMIT_CONST_U64(imm)); 1102 } else { 1103 /* lgrl %w0,imm */ 1104 EMIT6_PCREL_RILB(0xc4080000, REG_W0, 1105 _EMIT_CONST_U64(imm)); 1106 jit->seen |= SEEN_LITERAL; 1107 /* xgr %dst,%w0 */ 1108 EMIT4(0xb9820000, dst_reg, REG_W0); 1109 } 1110 break; 1111 /* 1112 * BPF_LSH 1113 */ 1114 case BPF_ALU | BPF_LSH | BPF_X: /* dst = (u32) dst << (u32) src */ 1115 /* sll %dst,0(%src) */ 1116 EMIT4_DISP(0x89000000, dst_reg, src_reg, 0); 1117 EMIT_ZERO(dst_reg); 1118 break; 1119 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst = dst << src */ 1120 /* sllg %dst,%dst,0(%src) */ 1121 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, src_reg, 0); 1122 break; 1123 case BPF_ALU | BPF_LSH | BPF_K: /* dst = (u32) dst << (u32) imm */ 1124 if (imm != 0) { 1125 /* sll %dst,imm(%r0) */ 1126 EMIT4_DISP(0x89000000, dst_reg, REG_0, imm); 1127 } 1128 EMIT_ZERO(dst_reg); 1129 break; 1130 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst = dst << imm */ 1131 if (imm == 0) 1132 break; 1133 /* sllg %dst,%dst,imm(%r0) */ 1134 EMIT6_DISP_LH(0xeb000000, 0x000d, dst_reg, dst_reg, REG_0, imm); 1135 break; 1136 /* 1137 * BPF_RSH 1138 */ 1139 case BPF_ALU | BPF_RSH | BPF_X: /* dst = (u32) dst >> (u32) src */ 1140 /* srl %dst,0(%src) */ 1141 EMIT4_DISP(0x88000000, dst_reg, src_reg, 0); 1142 EMIT_ZERO(dst_reg); 1143 break; 1144 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst = dst >> src */ 1145 /* srlg %dst,%dst,0(%src) */ 1146 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, src_reg, 0); 1147 break; 1148 case BPF_ALU | BPF_RSH | BPF_K: /* dst = (u32) dst >> (u32) imm */ 1149 if (imm != 0) { 1150 /* srl %dst,imm(%r0) */ 1151 EMIT4_DISP(0x88000000, dst_reg, REG_0, imm); 1152 } 1153 EMIT_ZERO(dst_reg); 1154 break; 1155 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst = dst >> imm */ 1156 if (imm == 0) 1157 break; 1158 /* srlg %dst,%dst,imm(%r0) */ 1159 EMIT6_DISP_LH(0xeb000000, 0x000c, dst_reg, dst_reg, REG_0, imm); 1160 break; 1161 /* 1162 * BPF_ARSH 1163 */ 1164 case BPF_ALU | BPF_ARSH | BPF_X: /* ((s32) dst) >>= src */ 1165 /* sra %dst,%dst,0(%src) */ 1166 EMIT4_DISP(0x8a000000, dst_reg, src_reg, 0); 1167 EMIT_ZERO(dst_reg); 1168 break; 1169 case BPF_ALU64 | BPF_ARSH | BPF_X: /* ((s64) dst) >>= src */ 1170 /* srag %dst,%dst,0(%src) */ 1171 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, src_reg, 0); 1172 break; 1173 case BPF_ALU | BPF_ARSH | BPF_K: /* ((s32) dst >> imm */ 1174 if (imm != 0) { 1175 /* sra %dst,imm(%r0) */ 1176 EMIT4_DISP(0x8a000000, dst_reg, REG_0, imm); 1177 } 1178 EMIT_ZERO(dst_reg); 1179 break; 1180 case BPF_ALU64 | BPF_ARSH | BPF_K: /* ((s64) dst) >>= imm */ 1181 if (imm == 0) 1182 break; 1183 /* srag %dst,%dst,imm(%r0) */ 1184 EMIT6_DISP_LH(0xeb000000, 0x000a, dst_reg, dst_reg, REG_0, imm); 1185 break; 1186 /* 1187 * BPF_NEG 1188 */ 1189 case BPF_ALU | BPF_NEG: /* dst = (u32) -dst */ 1190 /* lcr %dst,%dst */ 1191 EMIT2(0x1300, dst_reg, dst_reg); 1192 EMIT_ZERO(dst_reg); 1193 break; 1194 case BPF_ALU64 | BPF_NEG: /* dst = -dst */ 1195 /* lcgr %dst,%dst */ 1196 EMIT4(0xb9030000, dst_reg, dst_reg); 1197 break; 1198 /* 1199 * BPF_FROM_BE/LE 1200 */ 1201 case BPF_ALU | BPF_END | BPF_FROM_BE: 1202 /* s390 is big endian, therefore only clear high order bytes */ 1203 switch (imm) { 1204 case 16: /* dst = (u16) cpu_to_be16(dst) */ 1205 /* llghr %dst,%dst */ 1206 EMIT4(0xb9850000, dst_reg, dst_reg); 1207 if (insn_is_zext(&insn[1])) 1208 insn_count = 2; 1209 break; 1210 case 32: /* dst = (u32) cpu_to_be32(dst) */ 1211 if (!fp->aux->verifier_zext) 1212 /* llgfr %dst,%dst */ 1213 EMIT4(0xb9160000, dst_reg, dst_reg); 1214 break; 1215 case 64: /* dst = (u64) cpu_to_be64(dst) */ 1216 break; 1217 } 1218 break; 1219 case BPF_ALU | BPF_END | BPF_FROM_LE: 1220 switch (imm) { 1221 case 16: /* dst = (u16) cpu_to_le16(dst) */ 1222 /* lrvr %dst,%dst */ 1223 EMIT4(0xb91f0000, dst_reg, dst_reg); 1224 /* srl %dst,16(%r0) */ 1225 EMIT4_DISP(0x88000000, dst_reg, REG_0, 16); 1226 /* llghr %dst,%dst */ 1227 EMIT4(0xb9850000, dst_reg, dst_reg); 1228 if (insn_is_zext(&insn[1])) 1229 insn_count = 2; 1230 break; 1231 case 32: /* dst = (u32) cpu_to_le32(dst) */ 1232 /* lrvr %dst,%dst */ 1233 EMIT4(0xb91f0000, dst_reg, dst_reg); 1234 if (!fp->aux->verifier_zext) 1235 /* llgfr %dst,%dst */ 1236 EMIT4(0xb9160000, dst_reg, dst_reg); 1237 break; 1238 case 64: /* dst = (u64) cpu_to_le64(dst) */ 1239 /* lrvgr %dst,%dst */ 1240 EMIT4(0xb90f0000, dst_reg, dst_reg); 1241 break; 1242 } 1243 break; 1244 /* 1245 * BPF_NOSPEC (speculation barrier) 1246 */ 1247 case BPF_ST | BPF_NOSPEC: 1248 break; 1249 /* 1250 * BPF_ST(X) 1251 */ 1252 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src_reg */ 1253 /* stcy %src,off(%dst) */ 1254 EMIT6_DISP_LH(0xe3000000, 0x0072, src_reg, dst_reg, REG_0, off); 1255 jit->seen |= SEEN_MEM; 1256 break; 1257 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */ 1258 /* sthy %src,off(%dst) */ 1259 EMIT6_DISP_LH(0xe3000000, 0x0070, src_reg, dst_reg, REG_0, off); 1260 jit->seen |= SEEN_MEM; 1261 break; 1262 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */ 1263 /* sty %src,off(%dst) */ 1264 EMIT6_DISP_LH(0xe3000000, 0x0050, src_reg, dst_reg, REG_0, off); 1265 jit->seen |= SEEN_MEM; 1266 break; 1267 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */ 1268 /* stg %src,off(%dst) */ 1269 EMIT6_DISP_LH(0xe3000000, 0x0024, src_reg, dst_reg, REG_0, off); 1270 jit->seen |= SEEN_MEM; 1271 break; 1272 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */ 1273 /* lhi %w0,imm */ 1274 EMIT4_IMM(0xa7080000, REG_W0, (u8) imm); 1275 /* stcy %w0,off(dst) */ 1276 EMIT6_DISP_LH(0xe3000000, 0x0072, REG_W0, dst_reg, REG_0, off); 1277 jit->seen |= SEEN_MEM; 1278 break; 1279 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */ 1280 /* lhi %w0,imm */ 1281 EMIT4_IMM(0xa7080000, REG_W0, (u16) imm); 1282 /* sthy %w0,off(dst) */ 1283 EMIT6_DISP_LH(0xe3000000, 0x0070, REG_W0, dst_reg, REG_0, off); 1284 jit->seen |= SEEN_MEM; 1285 break; 1286 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */ 1287 /* llilf %w0,imm */ 1288 EMIT6_IMM(0xc00f0000, REG_W0, (u32) imm); 1289 /* sty %w0,off(%dst) */ 1290 EMIT6_DISP_LH(0xe3000000, 0x0050, REG_W0, dst_reg, REG_0, off); 1291 jit->seen |= SEEN_MEM; 1292 break; 1293 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */ 1294 /* lgfi %w0,imm */ 1295 EMIT6_IMM(0xc0010000, REG_W0, imm); 1296 /* stg %w0,off(%dst) */ 1297 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, dst_reg, REG_0, off); 1298 jit->seen |= SEEN_MEM; 1299 break; 1300 /* 1301 * BPF_ATOMIC 1302 */ 1303 case BPF_STX | BPF_ATOMIC | BPF_DW: 1304 case BPF_STX | BPF_ATOMIC | BPF_W: 1305 { 1306 bool is32 = BPF_SIZE(insn->code) == BPF_W; 1307 1308 switch (insn->imm) { 1309 /* {op32|op64} {%w0|%src},%src,off(%dst) */ 1310 #define EMIT_ATOMIC(op32, op64) do { \ 1311 EMIT6_DISP_LH(0xeb000000, is32 ? (op32) : (op64), \ 1312 (insn->imm & BPF_FETCH) ? src_reg : REG_W0, \ 1313 src_reg, dst_reg, off); \ 1314 if (is32 && (insn->imm & BPF_FETCH)) \ 1315 EMIT_ZERO(src_reg); \ 1316 } while (0) 1317 case BPF_ADD: 1318 case BPF_ADD | BPF_FETCH: 1319 /* {laal|laalg} */ 1320 EMIT_ATOMIC(0x00fa, 0x00ea); 1321 break; 1322 case BPF_AND: 1323 case BPF_AND | BPF_FETCH: 1324 /* {lan|lang} */ 1325 EMIT_ATOMIC(0x00f4, 0x00e4); 1326 break; 1327 case BPF_OR: 1328 case BPF_OR | BPF_FETCH: 1329 /* {lao|laog} */ 1330 EMIT_ATOMIC(0x00f6, 0x00e6); 1331 break; 1332 case BPF_XOR: 1333 case BPF_XOR | BPF_FETCH: 1334 /* {lax|laxg} */ 1335 EMIT_ATOMIC(0x00f7, 0x00e7); 1336 break; 1337 #undef EMIT_ATOMIC 1338 case BPF_XCHG: 1339 /* {ly|lg} %w0,off(%dst) */ 1340 EMIT6_DISP_LH(0xe3000000, 1341 is32 ? 0x0058 : 0x0004, REG_W0, REG_0, 1342 dst_reg, off); 1343 /* 0: {csy|csg} %w0,%src,off(%dst) */ 1344 EMIT6_DISP_LH(0xeb000000, is32 ? 0x0014 : 0x0030, 1345 REG_W0, src_reg, dst_reg, off); 1346 /* brc 4,0b */ 1347 EMIT4_PCREL_RIC(0xa7040000, 4, jit->prg - 6); 1348 /* {llgfr|lgr} %src,%w0 */ 1349 EMIT4(is32 ? 0xb9160000 : 0xb9040000, src_reg, REG_W0); 1350 if (is32 && insn_is_zext(&insn[1])) 1351 insn_count = 2; 1352 break; 1353 case BPF_CMPXCHG: 1354 /* 0: {csy|csg} %b0,%src,off(%dst) */ 1355 EMIT6_DISP_LH(0xeb000000, is32 ? 0x0014 : 0x0030, 1356 BPF_REG_0, src_reg, dst_reg, off); 1357 break; 1358 default: 1359 pr_err("Unknown atomic operation %02x\n", insn->imm); 1360 return -1; 1361 } 1362 1363 jit->seen |= SEEN_MEM; 1364 break; 1365 } 1366 /* 1367 * BPF_LDX 1368 */ 1369 case BPF_LDX | BPF_MEM | BPF_B: /* dst = *(u8 *)(ul) (src + off) */ 1370 case BPF_LDX | BPF_PROBE_MEM | BPF_B: 1371 /* llgc %dst,0(off,%src) */ 1372 EMIT6_DISP_LH(0xe3000000, 0x0090, dst_reg, src_reg, REG_0, off); 1373 jit->seen |= SEEN_MEM; 1374 if (insn_is_zext(&insn[1])) 1375 insn_count = 2; 1376 break; 1377 case BPF_LDX | BPF_MEM | BPF_H: /* dst = *(u16 *)(ul) (src + off) */ 1378 case BPF_LDX | BPF_PROBE_MEM | BPF_H: 1379 /* llgh %dst,0(off,%src) */ 1380 EMIT6_DISP_LH(0xe3000000, 0x0091, dst_reg, src_reg, REG_0, off); 1381 jit->seen |= SEEN_MEM; 1382 if (insn_is_zext(&insn[1])) 1383 insn_count = 2; 1384 break; 1385 case BPF_LDX | BPF_MEM | BPF_W: /* dst = *(u32 *)(ul) (src + off) */ 1386 case BPF_LDX | BPF_PROBE_MEM | BPF_W: 1387 /* llgf %dst,off(%src) */ 1388 jit->seen |= SEEN_MEM; 1389 EMIT6_DISP_LH(0xe3000000, 0x0016, dst_reg, src_reg, REG_0, off); 1390 if (insn_is_zext(&insn[1])) 1391 insn_count = 2; 1392 break; 1393 case BPF_LDX | BPF_MEM | BPF_DW: /* dst = *(u64 *)(ul) (src + off) */ 1394 case BPF_LDX | BPF_PROBE_MEM | BPF_DW: 1395 /* lg %dst,0(off,%src) */ 1396 jit->seen |= SEEN_MEM; 1397 EMIT6_DISP_LH(0xe3000000, 0x0004, dst_reg, src_reg, REG_0, off); 1398 break; 1399 /* 1400 * BPF_JMP / CALL 1401 */ 1402 case BPF_JMP | BPF_CALL: 1403 { 1404 const struct btf_func_model *m; 1405 bool func_addr_fixed; 1406 int j, ret; 1407 u64 func; 1408 1409 ret = bpf_jit_get_func_addr(fp, insn, extra_pass, 1410 &func, &func_addr_fixed); 1411 if (ret < 0) 1412 return -1; 1413 1414 REG_SET_SEEN(BPF_REG_5); 1415 jit->seen |= SEEN_FUNC; 1416 /* 1417 * Copy the tail call counter to where the callee expects it. 1418 * 1419 * Note 1: The callee can increment the tail call counter, but 1420 * we do not load it back, since the x86 JIT does not do this 1421 * either. 1422 * 1423 * Note 2: We assume that the verifier does not let us call the 1424 * main program, which clears the tail call counter on entry. 1425 */ 1426 /* mvc STK_OFF_TCCNT(4,%r15),N(%r15) */ 1427 _EMIT6(0xd203f000 | STK_OFF_TCCNT, 1428 0xf000 | (STK_OFF_TCCNT + STK_OFF + stack_depth)); 1429 1430 /* Sign-extend the kfunc arguments. */ 1431 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) { 1432 m = bpf_jit_find_kfunc_model(fp, insn); 1433 if (!m) 1434 return -1; 1435 1436 for (j = 0; j < m->nr_args; j++) { 1437 if (sign_extend(jit, BPF_REG_1 + j, 1438 m->arg_size[j], 1439 m->arg_flags[j])) 1440 return -1; 1441 } 1442 } 1443 1444 /* lgrl %w1,func */ 1445 EMIT6_PCREL_RILB(0xc4080000, REG_W1, _EMIT_CONST_U64(func)); 1446 /* %r1() */ 1447 call_r1(jit); 1448 /* lgr %b0,%r2: load return value into %b0 */ 1449 EMIT4(0xb9040000, BPF_REG_0, REG_2); 1450 break; 1451 } 1452 case BPF_JMP | BPF_TAIL_CALL: { 1453 int patch_1_clrj, patch_2_clij, patch_3_brc; 1454 1455 /* 1456 * Implicit input: 1457 * B1: pointer to ctx 1458 * B2: pointer to bpf_array 1459 * B3: index in bpf_array 1460 * 1461 * if (index >= array->map.max_entries) 1462 * goto out; 1463 */ 1464 1465 /* llgf %w1,map.max_entries(%b2) */ 1466 EMIT6_DISP_LH(0xe3000000, 0x0016, REG_W1, REG_0, BPF_REG_2, 1467 offsetof(struct bpf_array, map.max_entries)); 1468 /* if ((u32)%b3 >= (u32)%w1) goto out; */ 1469 /* clrj %b3,%w1,0xa,out */ 1470 patch_1_clrj = jit->prg; 1471 EMIT6_PCREL_RIEB(0xec000000, 0x0077, BPF_REG_3, REG_W1, 0xa, 1472 jit->prg); 1473 1474 /* 1475 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT) 1476 * goto out; 1477 */ 1478 1479 if (jit->seen & SEEN_STACK) 1480 off = STK_OFF_TCCNT + STK_OFF + stack_depth; 1481 else 1482 off = STK_OFF_TCCNT; 1483 /* lhi %w0,1 */ 1484 EMIT4_IMM(0xa7080000, REG_W0, 1); 1485 /* laal %w1,%w0,off(%r15) */ 1486 EMIT6_DISP_LH(0xeb000000, 0x00fa, REG_W1, REG_W0, REG_15, off); 1487 /* clij %w1,MAX_TAIL_CALL_CNT-1,0x2,out */ 1488 patch_2_clij = jit->prg; 1489 EMIT6_PCREL_RIEC(0xec000000, 0x007f, REG_W1, MAX_TAIL_CALL_CNT - 1, 1490 2, jit->prg); 1491 1492 /* 1493 * prog = array->ptrs[index]; 1494 * if (prog == NULL) 1495 * goto out; 1496 */ 1497 1498 /* llgfr %r1,%b3: %r1 = (u32) index */ 1499 EMIT4(0xb9160000, REG_1, BPF_REG_3); 1500 /* sllg %r1,%r1,3: %r1 *= 8 */ 1501 EMIT6_DISP_LH(0xeb000000, 0x000d, REG_1, REG_1, REG_0, 3); 1502 /* ltg %r1,prog(%b2,%r1) */ 1503 EMIT6_DISP_LH(0xe3000000, 0x0002, REG_1, BPF_REG_2, 1504 REG_1, offsetof(struct bpf_array, ptrs)); 1505 /* brc 0x8,out */ 1506 patch_3_brc = jit->prg; 1507 EMIT4_PCREL_RIC(0xa7040000, 8, jit->prg); 1508 1509 /* 1510 * Restore registers before calling function 1511 */ 1512 save_restore_regs(jit, REGS_RESTORE, stack_depth); 1513 1514 /* 1515 * goto *(prog->bpf_func + tail_call_start); 1516 */ 1517 1518 /* lg %r1,bpf_func(%r1) */ 1519 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_1, REG_0, 1520 offsetof(struct bpf_prog, bpf_func)); 1521 if (nospec_uses_trampoline()) { 1522 jit->seen |= SEEN_FUNC; 1523 /* aghi %r1,tail_call_start */ 1524 EMIT4_IMM(0xa70b0000, REG_1, jit->tail_call_start); 1525 /* brcl 0xf,__s390_indirect_jump_r1 */ 1526 EMIT6_PCREL_RILC(0xc0040000, 0xf, jit->r1_thunk_ip); 1527 } else { 1528 /* bc 0xf,tail_call_start(%r1) */ 1529 _EMIT4(0x47f01000 + jit->tail_call_start); 1530 } 1531 /* out: */ 1532 if (jit->prg_buf) { 1533 *(u16 *)(jit->prg_buf + patch_1_clrj + 2) = 1534 (jit->prg - patch_1_clrj) >> 1; 1535 *(u16 *)(jit->prg_buf + patch_2_clij + 2) = 1536 (jit->prg - patch_2_clij) >> 1; 1537 *(u16 *)(jit->prg_buf + patch_3_brc + 2) = 1538 (jit->prg - patch_3_brc) >> 1; 1539 } 1540 break; 1541 } 1542 case BPF_JMP | BPF_EXIT: /* return b0 */ 1543 last = (i == fp->len - 1) ? 1 : 0; 1544 if (last) 1545 break; 1546 if (!is_first_pass(jit) && can_use_rel(jit, jit->exit_ip)) 1547 /* brc 0xf, <exit> */ 1548 EMIT4_PCREL_RIC(0xa7040000, 0xf, jit->exit_ip); 1549 else 1550 /* brcl 0xf, <exit> */ 1551 EMIT6_PCREL_RILC(0xc0040000, 0xf, jit->exit_ip); 1552 break; 1553 /* 1554 * Branch relative (number of skipped instructions) to offset on 1555 * condition. 1556 * 1557 * Condition code to mask mapping: 1558 * 1559 * CC | Description | Mask 1560 * ------------------------------ 1561 * 0 | Operands equal | 8 1562 * 1 | First operand low | 4 1563 * 2 | First operand high | 2 1564 * 3 | Unused | 1 1565 * 1566 * For s390x relative branches: ip = ip + off_bytes 1567 * For BPF relative branches: insn = insn + off_insns + 1 1568 * 1569 * For example for s390x with offset 0 we jump to the branch 1570 * instruction itself (loop) and for BPF with offset 0 we 1571 * branch to the instruction behind the branch. 1572 */ 1573 case BPF_JMP | BPF_JA: /* if (true) */ 1574 mask = 0xf000; /* j */ 1575 goto branch_oc; 1576 case BPF_JMP | BPF_JSGT | BPF_K: /* ((s64) dst > (s64) imm) */ 1577 case BPF_JMP32 | BPF_JSGT | BPF_K: /* ((s32) dst > (s32) imm) */ 1578 mask = 0x2000; /* jh */ 1579 goto branch_ks; 1580 case BPF_JMP | BPF_JSLT | BPF_K: /* ((s64) dst < (s64) imm) */ 1581 case BPF_JMP32 | BPF_JSLT | BPF_K: /* ((s32) dst < (s32) imm) */ 1582 mask = 0x4000; /* jl */ 1583 goto branch_ks; 1584 case BPF_JMP | BPF_JSGE | BPF_K: /* ((s64) dst >= (s64) imm) */ 1585 case BPF_JMP32 | BPF_JSGE | BPF_K: /* ((s32) dst >= (s32) imm) */ 1586 mask = 0xa000; /* jhe */ 1587 goto branch_ks; 1588 case BPF_JMP | BPF_JSLE | BPF_K: /* ((s64) dst <= (s64) imm) */ 1589 case BPF_JMP32 | BPF_JSLE | BPF_K: /* ((s32) dst <= (s32) imm) */ 1590 mask = 0xc000; /* jle */ 1591 goto branch_ks; 1592 case BPF_JMP | BPF_JGT | BPF_K: /* (dst_reg > imm) */ 1593 case BPF_JMP32 | BPF_JGT | BPF_K: /* ((u32) dst_reg > (u32) imm) */ 1594 mask = 0x2000; /* jh */ 1595 goto branch_ku; 1596 case BPF_JMP | BPF_JLT | BPF_K: /* (dst_reg < imm) */ 1597 case BPF_JMP32 | BPF_JLT | BPF_K: /* ((u32) dst_reg < (u32) imm) */ 1598 mask = 0x4000; /* jl */ 1599 goto branch_ku; 1600 case BPF_JMP | BPF_JGE | BPF_K: /* (dst_reg >= imm) */ 1601 case BPF_JMP32 | BPF_JGE | BPF_K: /* ((u32) dst_reg >= (u32) imm) */ 1602 mask = 0xa000; /* jhe */ 1603 goto branch_ku; 1604 case BPF_JMP | BPF_JLE | BPF_K: /* (dst_reg <= imm) */ 1605 case BPF_JMP32 | BPF_JLE | BPF_K: /* ((u32) dst_reg <= (u32) imm) */ 1606 mask = 0xc000; /* jle */ 1607 goto branch_ku; 1608 case BPF_JMP | BPF_JNE | BPF_K: /* (dst_reg != imm) */ 1609 case BPF_JMP32 | BPF_JNE | BPF_K: /* ((u32) dst_reg != (u32) imm) */ 1610 mask = 0x7000; /* jne */ 1611 goto branch_ku; 1612 case BPF_JMP | BPF_JEQ | BPF_K: /* (dst_reg == imm) */ 1613 case BPF_JMP32 | BPF_JEQ | BPF_K: /* ((u32) dst_reg == (u32) imm) */ 1614 mask = 0x8000; /* je */ 1615 goto branch_ku; 1616 case BPF_JMP | BPF_JSET | BPF_K: /* (dst_reg & imm) */ 1617 case BPF_JMP32 | BPF_JSET | BPF_K: /* ((u32) dst_reg & (u32) imm) */ 1618 mask = 0x7000; /* jnz */ 1619 if (BPF_CLASS(insn->code) == BPF_JMP32) { 1620 /* llilf %w1,imm (load zero extend imm) */ 1621 EMIT6_IMM(0xc00f0000, REG_W1, imm); 1622 /* nr %w1,%dst */ 1623 EMIT2(0x1400, REG_W1, dst_reg); 1624 } else { 1625 /* lgfi %w1,imm (load sign extend imm) */ 1626 EMIT6_IMM(0xc0010000, REG_W1, imm); 1627 /* ngr %w1,%dst */ 1628 EMIT4(0xb9800000, REG_W1, dst_reg); 1629 } 1630 goto branch_oc; 1631 1632 case BPF_JMP | BPF_JSGT | BPF_X: /* ((s64) dst > (s64) src) */ 1633 case BPF_JMP32 | BPF_JSGT | BPF_X: /* ((s32) dst > (s32) src) */ 1634 mask = 0x2000; /* jh */ 1635 goto branch_xs; 1636 case BPF_JMP | BPF_JSLT | BPF_X: /* ((s64) dst < (s64) src) */ 1637 case BPF_JMP32 | BPF_JSLT | BPF_X: /* ((s32) dst < (s32) src) */ 1638 mask = 0x4000; /* jl */ 1639 goto branch_xs; 1640 case BPF_JMP | BPF_JSGE | BPF_X: /* ((s64) dst >= (s64) src) */ 1641 case BPF_JMP32 | BPF_JSGE | BPF_X: /* ((s32) dst >= (s32) src) */ 1642 mask = 0xa000; /* jhe */ 1643 goto branch_xs; 1644 case BPF_JMP | BPF_JSLE | BPF_X: /* ((s64) dst <= (s64) src) */ 1645 case BPF_JMP32 | BPF_JSLE | BPF_X: /* ((s32) dst <= (s32) src) */ 1646 mask = 0xc000; /* jle */ 1647 goto branch_xs; 1648 case BPF_JMP | BPF_JGT | BPF_X: /* (dst > src) */ 1649 case BPF_JMP32 | BPF_JGT | BPF_X: /* ((u32) dst > (u32) src) */ 1650 mask = 0x2000; /* jh */ 1651 goto branch_xu; 1652 case BPF_JMP | BPF_JLT | BPF_X: /* (dst < src) */ 1653 case BPF_JMP32 | BPF_JLT | BPF_X: /* ((u32) dst < (u32) src) */ 1654 mask = 0x4000; /* jl */ 1655 goto branch_xu; 1656 case BPF_JMP | BPF_JGE | BPF_X: /* (dst >= src) */ 1657 case BPF_JMP32 | BPF_JGE | BPF_X: /* ((u32) dst >= (u32) src) */ 1658 mask = 0xa000; /* jhe */ 1659 goto branch_xu; 1660 case BPF_JMP | BPF_JLE | BPF_X: /* (dst <= src) */ 1661 case BPF_JMP32 | BPF_JLE | BPF_X: /* ((u32) dst <= (u32) src) */ 1662 mask = 0xc000; /* jle */ 1663 goto branch_xu; 1664 case BPF_JMP | BPF_JNE | BPF_X: /* (dst != src) */ 1665 case BPF_JMP32 | BPF_JNE | BPF_X: /* ((u32) dst != (u32) src) */ 1666 mask = 0x7000; /* jne */ 1667 goto branch_xu; 1668 case BPF_JMP | BPF_JEQ | BPF_X: /* (dst == src) */ 1669 case BPF_JMP32 | BPF_JEQ | BPF_X: /* ((u32) dst == (u32) src) */ 1670 mask = 0x8000; /* je */ 1671 goto branch_xu; 1672 case BPF_JMP | BPF_JSET | BPF_X: /* (dst & src) */ 1673 case BPF_JMP32 | BPF_JSET | BPF_X: /* ((u32) dst & (u32) src) */ 1674 { 1675 bool is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32; 1676 1677 mask = 0x7000; /* jnz */ 1678 /* nrk or ngrk %w1,%dst,%src */ 1679 EMIT4_RRF((is_jmp32 ? 0xb9f40000 : 0xb9e40000), 1680 REG_W1, dst_reg, src_reg); 1681 goto branch_oc; 1682 branch_ks: 1683 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32; 1684 /* cfi or cgfi %dst,imm */ 1685 EMIT6_IMM(is_jmp32 ? 0xc20d0000 : 0xc20c0000, 1686 dst_reg, imm); 1687 if (!is_first_pass(jit) && 1688 can_use_rel(jit, addrs[i + off + 1])) { 1689 /* brc mask,off */ 1690 EMIT4_PCREL_RIC(0xa7040000, 1691 mask >> 12, addrs[i + off + 1]); 1692 } else { 1693 /* brcl mask,off */ 1694 EMIT6_PCREL_RILC(0xc0040000, 1695 mask >> 12, addrs[i + off + 1]); 1696 } 1697 break; 1698 branch_ku: 1699 /* lgfi %w1,imm (load sign extend imm) */ 1700 src_reg = REG_1; 1701 EMIT6_IMM(0xc0010000, src_reg, imm); 1702 goto branch_xu; 1703 branch_xs: 1704 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32; 1705 if (!is_first_pass(jit) && 1706 can_use_rel(jit, addrs[i + off + 1])) { 1707 /* crj or cgrj %dst,%src,mask,off */ 1708 EMIT6_PCREL(0xec000000, (is_jmp32 ? 0x0076 : 0x0064), 1709 dst_reg, src_reg, i, off, mask); 1710 } else { 1711 /* cr or cgr %dst,%src */ 1712 if (is_jmp32) 1713 EMIT2(0x1900, dst_reg, src_reg); 1714 else 1715 EMIT4(0xb9200000, dst_reg, src_reg); 1716 /* brcl mask,off */ 1717 EMIT6_PCREL_RILC(0xc0040000, 1718 mask >> 12, addrs[i + off + 1]); 1719 } 1720 break; 1721 branch_xu: 1722 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32; 1723 if (!is_first_pass(jit) && 1724 can_use_rel(jit, addrs[i + off + 1])) { 1725 /* clrj or clgrj %dst,%src,mask,off */ 1726 EMIT6_PCREL(0xec000000, (is_jmp32 ? 0x0077 : 0x0065), 1727 dst_reg, src_reg, i, off, mask); 1728 } else { 1729 /* clr or clgr %dst,%src */ 1730 if (is_jmp32) 1731 EMIT2(0x1500, dst_reg, src_reg); 1732 else 1733 EMIT4(0xb9210000, dst_reg, src_reg); 1734 /* brcl mask,off */ 1735 EMIT6_PCREL_RILC(0xc0040000, 1736 mask >> 12, addrs[i + off + 1]); 1737 } 1738 break; 1739 branch_oc: 1740 if (!is_first_pass(jit) && 1741 can_use_rel(jit, addrs[i + off + 1])) { 1742 /* brc mask,off */ 1743 EMIT4_PCREL_RIC(0xa7040000, 1744 mask >> 12, addrs[i + off + 1]); 1745 } else { 1746 /* brcl mask,off */ 1747 EMIT6_PCREL_RILC(0xc0040000, 1748 mask >> 12, addrs[i + off + 1]); 1749 } 1750 break; 1751 } 1752 default: /* too complex, give up */ 1753 pr_err("Unknown opcode %02x\n", insn->code); 1754 return -1; 1755 } 1756 1757 if (probe_prg != -1) { 1758 /* 1759 * Handlers of certain exceptions leave psw.addr pointing to 1760 * the instruction directly after the failing one. Therefore, 1761 * create two exception table entries and also add a nop in 1762 * case two probing instructions come directly after each 1763 * other. 1764 */ 1765 nop_prg = jit->prg; 1766 /* bcr 0,%0 */ 1767 _EMIT2(0x0700); 1768 err = bpf_jit_probe_mem(jit, fp, probe_prg, nop_prg); 1769 if (err < 0) 1770 return err; 1771 } 1772 1773 return insn_count; 1774 } 1775 1776 /* 1777 * Return whether new i-th instruction address does not violate any invariant 1778 */ 1779 static bool bpf_is_new_addr_sane(struct bpf_jit *jit, int i) 1780 { 1781 /* On the first pass anything goes */ 1782 if (is_first_pass(jit)) 1783 return true; 1784 1785 /* The codegen pass must not change anything */ 1786 if (is_codegen_pass(jit)) 1787 return jit->addrs[i] == jit->prg; 1788 1789 /* Passes in between must not increase code size */ 1790 return jit->addrs[i] >= jit->prg; 1791 } 1792 1793 /* 1794 * Update the address of i-th instruction 1795 */ 1796 static int bpf_set_addr(struct bpf_jit *jit, int i) 1797 { 1798 int delta; 1799 1800 if (is_codegen_pass(jit)) { 1801 delta = jit->prg - jit->addrs[i]; 1802 if (delta < 0) 1803 bpf_skip(jit, -delta); 1804 } 1805 if (WARN_ON_ONCE(!bpf_is_new_addr_sane(jit, i))) 1806 return -1; 1807 jit->addrs[i] = jit->prg; 1808 return 0; 1809 } 1810 1811 /* 1812 * Compile eBPF program into s390x code 1813 */ 1814 static int bpf_jit_prog(struct bpf_jit *jit, struct bpf_prog *fp, 1815 bool extra_pass, u32 stack_depth) 1816 { 1817 int i, insn_count, lit32_size, lit64_size; 1818 1819 jit->lit32 = jit->lit32_start; 1820 jit->lit64 = jit->lit64_start; 1821 jit->prg = 0; 1822 jit->excnt = 0; 1823 1824 bpf_jit_prologue(jit, fp, stack_depth); 1825 if (bpf_set_addr(jit, 0) < 0) 1826 return -1; 1827 for (i = 0; i < fp->len; i += insn_count) { 1828 insn_count = bpf_jit_insn(jit, fp, i, extra_pass, stack_depth); 1829 if (insn_count < 0) 1830 return -1; 1831 /* Next instruction address */ 1832 if (bpf_set_addr(jit, i + insn_count) < 0) 1833 return -1; 1834 } 1835 bpf_jit_epilogue(jit, stack_depth); 1836 1837 lit32_size = jit->lit32 - jit->lit32_start; 1838 lit64_size = jit->lit64 - jit->lit64_start; 1839 jit->lit32_start = jit->prg; 1840 if (lit32_size) 1841 jit->lit32_start = ALIGN(jit->lit32_start, 4); 1842 jit->lit64_start = jit->lit32_start + lit32_size; 1843 if (lit64_size) 1844 jit->lit64_start = ALIGN(jit->lit64_start, 8); 1845 jit->size = jit->lit64_start + lit64_size; 1846 jit->size_prg = jit->prg; 1847 1848 if (WARN_ON_ONCE(fp->aux->extable && 1849 jit->excnt != fp->aux->num_exentries)) 1850 /* Verifier bug - too many entries. */ 1851 return -1; 1852 1853 return 0; 1854 } 1855 1856 bool bpf_jit_needs_zext(void) 1857 { 1858 return true; 1859 } 1860 1861 struct s390_jit_data { 1862 struct bpf_binary_header *header; 1863 struct bpf_jit ctx; 1864 int pass; 1865 }; 1866 1867 static struct bpf_binary_header *bpf_jit_alloc(struct bpf_jit *jit, 1868 struct bpf_prog *fp) 1869 { 1870 struct bpf_binary_header *header; 1871 u32 extable_size; 1872 u32 code_size; 1873 1874 /* We need two entries per insn. */ 1875 fp->aux->num_exentries *= 2; 1876 1877 code_size = roundup(jit->size, 1878 __alignof__(struct exception_table_entry)); 1879 extable_size = fp->aux->num_exentries * 1880 sizeof(struct exception_table_entry); 1881 header = bpf_jit_binary_alloc(code_size + extable_size, &jit->prg_buf, 1882 8, jit_fill_hole); 1883 if (!header) 1884 return NULL; 1885 fp->aux->extable = (struct exception_table_entry *) 1886 (jit->prg_buf + code_size); 1887 return header; 1888 } 1889 1890 /* 1891 * Compile eBPF program "fp" 1892 */ 1893 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp) 1894 { 1895 u32 stack_depth = round_up(fp->aux->stack_depth, 8); 1896 struct bpf_prog *tmp, *orig_fp = fp; 1897 struct bpf_binary_header *header; 1898 struct s390_jit_data *jit_data; 1899 bool tmp_blinded = false; 1900 bool extra_pass = false; 1901 struct bpf_jit jit; 1902 int pass; 1903 1904 if (WARN_ON_ONCE(bpf_plt_end - bpf_plt != BPF_PLT_SIZE)) 1905 return orig_fp; 1906 1907 if (!fp->jit_requested) 1908 return orig_fp; 1909 1910 tmp = bpf_jit_blind_constants(fp); 1911 /* 1912 * If blinding was requested and we failed during blinding, 1913 * we must fall back to the interpreter. 1914 */ 1915 if (IS_ERR(tmp)) 1916 return orig_fp; 1917 if (tmp != fp) { 1918 tmp_blinded = true; 1919 fp = tmp; 1920 } 1921 1922 jit_data = fp->aux->jit_data; 1923 if (!jit_data) { 1924 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL); 1925 if (!jit_data) { 1926 fp = orig_fp; 1927 goto out; 1928 } 1929 fp->aux->jit_data = jit_data; 1930 } 1931 if (jit_data->ctx.addrs) { 1932 jit = jit_data->ctx; 1933 header = jit_data->header; 1934 extra_pass = true; 1935 pass = jit_data->pass + 1; 1936 goto skip_init_ctx; 1937 } 1938 1939 memset(&jit, 0, sizeof(jit)); 1940 jit.addrs = kvcalloc(fp->len + 1, sizeof(*jit.addrs), GFP_KERNEL); 1941 if (jit.addrs == NULL) { 1942 fp = orig_fp; 1943 goto free_addrs; 1944 } 1945 /* 1946 * Three initial passes: 1947 * - 1/2: Determine clobbered registers 1948 * - 3: Calculate program size and addrs array 1949 */ 1950 for (pass = 1; pass <= 3; pass++) { 1951 if (bpf_jit_prog(&jit, fp, extra_pass, stack_depth)) { 1952 fp = orig_fp; 1953 goto free_addrs; 1954 } 1955 } 1956 /* 1957 * Final pass: Allocate and generate program 1958 */ 1959 header = bpf_jit_alloc(&jit, fp); 1960 if (!header) { 1961 fp = orig_fp; 1962 goto free_addrs; 1963 } 1964 skip_init_ctx: 1965 if (bpf_jit_prog(&jit, fp, extra_pass, stack_depth)) { 1966 bpf_jit_binary_free(header); 1967 fp = orig_fp; 1968 goto free_addrs; 1969 } 1970 if (bpf_jit_enable > 1) { 1971 bpf_jit_dump(fp->len, jit.size, pass, jit.prg_buf); 1972 print_fn_code(jit.prg_buf, jit.size_prg); 1973 } 1974 if (!fp->is_func || extra_pass) { 1975 bpf_jit_binary_lock_ro(header); 1976 } else { 1977 jit_data->header = header; 1978 jit_data->ctx = jit; 1979 jit_data->pass = pass; 1980 } 1981 fp->bpf_func = (void *) jit.prg_buf; 1982 fp->jited = 1; 1983 fp->jited_len = jit.size; 1984 1985 if (!fp->is_func || extra_pass) { 1986 bpf_prog_fill_jited_linfo(fp, jit.addrs + 1); 1987 free_addrs: 1988 kvfree(jit.addrs); 1989 kfree(jit_data); 1990 fp->aux->jit_data = NULL; 1991 } 1992 out: 1993 if (tmp_blinded) 1994 bpf_jit_prog_release_other(fp, fp == orig_fp ? 1995 tmp : orig_fp); 1996 return fp; 1997 } 1998 1999 bool bpf_jit_supports_kfunc_call(void) 2000 { 2001 return true; 2002 } 2003 2004 bool bpf_jit_supports_far_kfunc_call(void) 2005 { 2006 return true; 2007 } 2008 2009 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 2010 void *old_addr, void *new_addr) 2011 { 2012 struct { 2013 u16 opc; 2014 s32 disp; 2015 } __packed insn; 2016 char expected_plt[BPF_PLT_SIZE]; 2017 char current_plt[BPF_PLT_SIZE]; 2018 char new_plt[BPF_PLT_SIZE]; 2019 char *plt; 2020 char *ret; 2021 int err; 2022 2023 /* Verify the branch to be patched. */ 2024 err = copy_from_kernel_nofault(&insn, ip, sizeof(insn)); 2025 if (err < 0) 2026 return err; 2027 if (insn.opc != (0xc004 | (old_addr ? 0xf0 : 0))) 2028 return -EINVAL; 2029 2030 if (t == BPF_MOD_JUMP && 2031 insn.disp == ((char *)new_addr - (char *)ip) >> 1) { 2032 /* 2033 * The branch already points to the destination, 2034 * there is no PLT. 2035 */ 2036 } else { 2037 /* Verify the PLT. */ 2038 plt = (char *)ip + (insn.disp << 1); 2039 err = copy_from_kernel_nofault(current_plt, plt, BPF_PLT_SIZE); 2040 if (err < 0) 2041 return err; 2042 ret = (char *)ip + 6; 2043 bpf_jit_plt(expected_plt, ret, old_addr); 2044 if (memcmp(current_plt, expected_plt, BPF_PLT_SIZE)) 2045 return -EINVAL; 2046 /* Adjust the call address. */ 2047 bpf_jit_plt(new_plt, ret, new_addr); 2048 s390_kernel_write(plt + (bpf_plt_target - bpf_plt), 2049 new_plt + (bpf_plt_target - bpf_plt), 2050 sizeof(void *)); 2051 } 2052 2053 /* Adjust the mask of the branch. */ 2054 insn.opc = 0xc004 | (new_addr ? 0xf0 : 0); 2055 s390_kernel_write((char *)ip + 1, (char *)&insn.opc + 1, 1); 2056 2057 /* Make the new code visible to the other CPUs. */ 2058 text_poke_sync_lock(); 2059 2060 return 0; 2061 } 2062 2063 struct bpf_tramp_jit { 2064 struct bpf_jit common; 2065 int orig_stack_args_off;/* Offset of arguments placed on stack by the 2066 * func_addr's original caller 2067 */ 2068 int stack_size; /* Trampoline stack size */ 2069 int stack_args_off; /* Offset of stack arguments for calling 2070 * func_addr, has to be at the top 2071 */ 2072 int reg_args_off; /* Offset of register arguments for calling 2073 * func_addr 2074 */ 2075 int ip_off; /* For bpf_get_func_ip(), has to be at 2076 * (ctx - 16) 2077 */ 2078 int arg_cnt_off; /* For bpf_get_func_arg_cnt(), has to be at 2079 * (ctx - 8) 2080 */ 2081 int bpf_args_off; /* Offset of BPF_PROG context, which consists 2082 * of BPF arguments followed by return value 2083 */ 2084 int retval_off; /* Offset of return value (see above) */ 2085 int r7_r8_off; /* Offset of saved %r7 and %r8, which are used 2086 * for __bpf_prog_enter() return value and 2087 * func_addr respectively 2088 */ 2089 int r14_off; /* Offset of saved %r14 */ 2090 int run_ctx_off; /* Offset of struct bpf_tramp_run_ctx */ 2091 int do_fexit; /* do_fexit: label */ 2092 }; 2093 2094 static void load_imm64(struct bpf_jit *jit, int dst_reg, u64 val) 2095 { 2096 /* llihf %dst_reg,val_hi */ 2097 EMIT6_IMM(0xc00e0000, dst_reg, (val >> 32)); 2098 /* oilf %rdst_reg,val_lo */ 2099 EMIT6_IMM(0xc00d0000, dst_reg, val); 2100 } 2101 2102 static int invoke_bpf_prog(struct bpf_tramp_jit *tjit, 2103 const struct btf_func_model *m, 2104 struct bpf_tramp_link *tlink, bool save_ret) 2105 { 2106 struct bpf_jit *jit = &tjit->common; 2107 int cookie_off = tjit->run_ctx_off + 2108 offsetof(struct bpf_tramp_run_ctx, bpf_cookie); 2109 struct bpf_prog *p = tlink->link.prog; 2110 int patch; 2111 2112 /* 2113 * run_ctx.cookie = tlink->cookie; 2114 */ 2115 2116 /* %r0 = tlink->cookie */ 2117 load_imm64(jit, REG_W0, tlink->cookie); 2118 /* stg %r0,cookie_off(%r15) */ 2119 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_W0, REG_0, REG_15, cookie_off); 2120 2121 /* 2122 * if ((start = __bpf_prog_enter(p, &run_ctx)) == 0) 2123 * goto skip; 2124 */ 2125 2126 /* %r1 = __bpf_prog_enter */ 2127 load_imm64(jit, REG_1, (u64)bpf_trampoline_enter(p)); 2128 /* %r2 = p */ 2129 load_imm64(jit, REG_2, (u64)p); 2130 /* la %r3,run_ctx_off(%r15) */ 2131 EMIT4_DISP(0x41000000, REG_3, REG_15, tjit->run_ctx_off); 2132 /* %r1() */ 2133 call_r1(jit); 2134 /* ltgr %r7,%r2 */ 2135 EMIT4(0xb9020000, REG_7, REG_2); 2136 /* brcl 8,skip */ 2137 patch = jit->prg; 2138 EMIT6_PCREL_RILC(0xc0040000, 8, 0); 2139 2140 /* 2141 * retval = bpf_func(args, p->insnsi); 2142 */ 2143 2144 /* %r1 = p->bpf_func */ 2145 load_imm64(jit, REG_1, (u64)p->bpf_func); 2146 /* la %r2,bpf_args_off(%r15) */ 2147 EMIT4_DISP(0x41000000, REG_2, REG_15, tjit->bpf_args_off); 2148 /* %r3 = p->insnsi */ 2149 if (!p->jited) 2150 load_imm64(jit, REG_3, (u64)p->insnsi); 2151 /* %r1() */ 2152 call_r1(jit); 2153 /* stg %r2,retval_off(%r15) */ 2154 if (save_ret) { 2155 if (sign_extend(jit, REG_2, m->ret_size, m->ret_flags)) 2156 return -1; 2157 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_2, REG_0, REG_15, 2158 tjit->retval_off); 2159 } 2160 2161 /* skip: */ 2162 if (jit->prg_buf) 2163 *(u32 *)&jit->prg_buf[patch + 2] = (jit->prg - patch) >> 1; 2164 2165 /* 2166 * __bpf_prog_exit(p, start, &run_ctx); 2167 */ 2168 2169 /* %r1 = __bpf_prog_exit */ 2170 load_imm64(jit, REG_1, (u64)bpf_trampoline_exit(p)); 2171 /* %r2 = p */ 2172 load_imm64(jit, REG_2, (u64)p); 2173 /* lgr %r3,%r7 */ 2174 EMIT4(0xb9040000, REG_3, REG_7); 2175 /* la %r4,run_ctx_off(%r15) */ 2176 EMIT4_DISP(0x41000000, REG_4, REG_15, tjit->run_ctx_off); 2177 /* %r1() */ 2178 call_r1(jit); 2179 2180 return 0; 2181 } 2182 2183 static int alloc_stack(struct bpf_tramp_jit *tjit, size_t size) 2184 { 2185 int stack_offset = tjit->stack_size; 2186 2187 tjit->stack_size += size; 2188 return stack_offset; 2189 } 2190 2191 /* ABI uses %r2 - %r6 for parameter passing. */ 2192 #define MAX_NR_REG_ARGS 5 2193 2194 /* The "L" field of the "mvc" instruction is 8 bits. */ 2195 #define MAX_MVC_SIZE 256 2196 #define MAX_NR_STACK_ARGS (MAX_MVC_SIZE / sizeof(u64)) 2197 2198 /* -mfentry generates a 6-byte nop on s390x. */ 2199 #define S390X_PATCH_SIZE 6 2200 2201 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, 2202 struct bpf_tramp_jit *tjit, 2203 const struct btf_func_model *m, 2204 u32 flags, 2205 struct bpf_tramp_links *tlinks, 2206 void *func_addr) 2207 { 2208 struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN]; 2209 struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY]; 2210 struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT]; 2211 int nr_bpf_args, nr_reg_args, nr_stack_args; 2212 struct bpf_jit *jit = &tjit->common; 2213 int arg, bpf_arg_off; 2214 int i, j; 2215 2216 /* Support as many stack arguments as "mvc" instruction can handle. */ 2217 nr_reg_args = min_t(int, m->nr_args, MAX_NR_REG_ARGS); 2218 nr_stack_args = m->nr_args - nr_reg_args; 2219 if (nr_stack_args > MAX_NR_STACK_ARGS) 2220 return -ENOTSUPP; 2221 2222 /* Return to %r14, since func_addr and %r0 are not available. */ 2223 if (!func_addr && !(flags & BPF_TRAMP_F_ORIG_STACK)) 2224 flags |= BPF_TRAMP_F_SKIP_FRAME; 2225 2226 /* 2227 * Compute how many arguments we need to pass to BPF programs. 2228 * BPF ABI mirrors that of x86_64: arguments that are 16 bytes or 2229 * smaller are packed into 1 or 2 registers; larger arguments are 2230 * passed via pointers. 2231 * In s390x ABI, arguments that are 8 bytes or smaller are packed into 2232 * a register; larger arguments are passed via pointers. 2233 * We need to deal with this difference. 2234 */ 2235 nr_bpf_args = 0; 2236 for (i = 0; i < m->nr_args; i++) { 2237 if (m->arg_size[i] <= 8) 2238 nr_bpf_args += 1; 2239 else if (m->arg_size[i] <= 16) 2240 nr_bpf_args += 2; 2241 else 2242 return -ENOTSUPP; 2243 } 2244 2245 /* 2246 * Calculate the stack layout. 2247 */ 2248 2249 /* Reserve STACK_FRAME_OVERHEAD bytes for the callees. */ 2250 tjit->stack_size = STACK_FRAME_OVERHEAD; 2251 tjit->stack_args_off = alloc_stack(tjit, nr_stack_args * sizeof(u64)); 2252 tjit->reg_args_off = alloc_stack(tjit, nr_reg_args * sizeof(u64)); 2253 tjit->ip_off = alloc_stack(tjit, sizeof(u64)); 2254 tjit->arg_cnt_off = alloc_stack(tjit, sizeof(u64)); 2255 tjit->bpf_args_off = alloc_stack(tjit, nr_bpf_args * sizeof(u64)); 2256 tjit->retval_off = alloc_stack(tjit, sizeof(u64)); 2257 tjit->r7_r8_off = alloc_stack(tjit, 2 * sizeof(u64)); 2258 tjit->r14_off = alloc_stack(tjit, sizeof(u64)); 2259 tjit->run_ctx_off = alloc_stack(tjit, 2260 sizeof(struct bpf_tramp_run_ctx)); 2261 /* The caller has already reserved STACK_FRAME_OVERHEAD bytes. */ 2262 tjit->stack_size -= STACK_FRAME_OVERHEAD; 2263 tjit->orig_stack_args_off = tjit->stack_size + STACK_FRAME_OVERHEAD; 2264 2265 /* aghi %r15,-stack_size */ 2266 EMIT4_IMM(0xa70b0000, REG_15, -tjit->stack_size); 2267 /* stmg %r2,%rN,fwd_reg_args_off(%r15) */ 2268 if (nr_reg_args) 2269 EMIT6_DISP_LH(0xeb000000, 0x0024, REG_2, 2270 REG_2 + (nr_reg_args - 1), REG_15, 2271 tjit->reg_args_off); 2272 for (i = 0, j = 0; i < m->nr_args; i++) { 2273 if (i < MAX_NR_REG_ARGS) 2274 arg = REG_2 + i; 2275 else 2276 arg = tjit->orig_stack_args_off + 2277 (i - MAX_NR_REG_ARGS) * sizeof(u64); 2278 bpf_arg_off = tjit->bpf_args_off + j * sizeof(u64); 2279 if (m->arg_size[i] <= 8) { 2280 if (i < MAX_NR_REG_ARGS) 2281 /* stg %arg,bpf_arg_off(%r15) */ 2282 EMIT6_DISP_LH(0xe3000000, 0x0024, arg, 2283 REG_0, REG_15, bpf_arg_off); 2284 else 2285 /* mvc bpf_arg_off(8,%r15),arg(%r15) */ 2286 _EMIT6(0xd207f000 | bpf_arg_off, 2287 0xf000 | arg); 2288 j += 1; 2289 } else { 2290 if (i < MAX_NR_REG_ARGS) { 2291 /* mvc bpf_arg_off(16,%r15),0(%arg) */ 2292 _EMIT6(0xd20ff000 | bpf_arg_off, 2293 reg2hex[arg] << 12); 2294 } else { 2295 /* lg %r1,arg(%r15) */ 2296 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_1, REG_0, 2297 REG_15, arg); 2298 /* mvc bpf_arg_off(16,%r15),0(%r1) */ 2299 _EMIT6(0xd20ff000 | bpf_arg_off, 0x1000); 2300 } 2301 j += 2; 2302 } 2303 } 2304 /* stmg %r7,%r8,r7_r8_off(%r15) */ 2305 EMIT6_DISP_LH(0xeb000000, 0x0024, REG_7, REG_8, REG_15, 2306 tjit->r7_r8_off); 2307 /* stg %r14,r14_off(%r15) */ 2308 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_14, REG_0, REG_15, tjit->r14_off); 2309 2310 if (flags & BPF_TRAMP_F_ORIG_STACK) { 2311 /* 2312 * The ftrace trampoline puts the return address (which is the 2313 * address of the original function + S390X_PATCH_SIZE) into 2314 * %r0; see ftrace_shared_hotpatch_trampoline_br and 2315 * ftrace_init_nop() for details. 2316 */ 2317 2318 /* lgr %r8,%r0 */ 2319 EMIT4(0xb9040000, REG_8, REG_0); 2320 } else { 2321 /* %r8 = func_addr + S390X_PATCH_SIZE */ 2322 load_imm64(jit, REG_8, (u64)func_addr + S390X_PATCH_SIZE); 2323 } 2324 2325 /* 2326 * ip = func_addr; 2327 * arg_cnt = m->nr_args; 2328 */ 2329 2330 if (flags & BPF_TRAMP_F_IP_ARG) { 2331 /* %r0 = func_addr */ 2332 load_imm64(jit, REG_0, (u64)func_addr); 2333 /* stg %r0,ip_off(%r15) */ 2334 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_0, REG_0, REG_15, 2335 tjit->ip_off); 2336 } 2337 /* lghi %r0,nr_bpf_args */ 2338 EMIT4_IMM(0xa7090000, REG_0, nr_bpf_args); 2339 /* stg %r0,arg_cnt_off(%r15) */ 2340 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_0, REG_0, REG_15, 2341 tjit->arg_cnt_off); 2342 2343 if (flags & BPF_TRAMP_F_CALL_ORIG) { 2344 /* 2345 * __bpf_tramp_enter(im); 2346 */ 2347 2348 /* %r1 = __bpf_tramp_enter */ 2349 load_imm64(jit, REG_1, (u64)__bpf_tramp_enter); 2350 /* %r2 = im */ 2351 load_imm64(jit, REG_2, (u64)im); 2352 /* %r1() */ 2353 call_r1(jit); 2354 } 2355 2356 for (i = 0; i < fentry->nr_links; i++) 2357 if (invoke_bpf_prog(tjit, m, fentry->links[i], 2358 flags & BPF_TRAMP_F_RET_FENTRY_RET)) 2359 return -EINVAL; 2360 2361 if (fmod_ret->nr_links) { 2362 /* 2363 * retval = 0; 2364 */ 2365 2366 /* xc retval_off(8,%r15),retval_off(%r15) */ 2367 _EMIT6(0xd707f000 | tjit->retval_off, 2368 0xf000 | tjit->retval_off); 2369 2370 for (i = 0; i < fmod_ret->nr_links; i++) { 2371 if (invoke_bpf_prog(tjit, m, fmod_ret->links[i], true)) 2372 return -EINVAL; 2373 2374 /* 2375 * if (retval) 2376 * goto do_fexit; 2377 */ 2378 2379 /* ltg %r0,retval_off(%r15) */ 2380 EMIT6_DISP_LH(0xe3000000, 0x0002, REG_0, REG_0, REG_15, 2381 tjit->retval_off); 2382 /* brcl 7,do_fexit */ 2383 EMIT6_PCREL_RILC(0xc0040000, 7, tjit->do_fexit); 2384 } 2385 } 2386 2387 if (flags & BPF_TRAMP_F_CALL_ORIG) { 2388 /* 2389 * retval = func_addr(args); 2390 */ 2391 2392 /* lmg %r2,%rN,reg_args_off(%r15) */ 2393 if (nr_reg_args) 2394 EMIT6_DISP_LH(0xeb000000, 0x0004, REG_2, 2395 REG_2 + (nr_reg_args - 1), REG_15, 2396 tjit->reg_args_off); 2397 /* mvc stack_args_off(N,%r15),orig_stack_args_off(%r15) */ 2398 if (nr_stack_args) 2399 _EMIT6(0xd200f000 | 2400 (nr_stack_args * sizeof(u64) - 1) << 16 | 2401 tjit->stack_args_off, 2402 0xf000 | tjit->orig_stack_args_off); 2403 /* lgr %r1,%r8 */ 2404 EMIT4(0xb9040000, REG_1, REG_8); 2405 /* %r1() */ 2406 call_r1(jit); 2407 /* stg %r2,retval_off(%r15) */ 2408 EMIT6_DISP_LH(0xe3000000, 0x0024, REG_2, REG_0, REG_15, 2409 tjit->retval_off); 2410 2411 im->ip_after_call = jit->prg_buf + jit->prg; 2412 2413 /* 2414 * The following nop will be patched by bpf_tramp_image_put(). 2415 */ 2416 2417 /* brcl 0,im->ip_epilogue */ 2418 EMIT6_PCREL_RILC(0xc0040000, 0, (u64)im->ip_epilogue); 2419 } 2420 2421 /* do_fexit: */ 2422 tjit->do_fexit = jit->prg; 2423 for (i = 0; i < fexit->nr_links; i++) 2424 if (invoke_bpf_prog(tjit, m, fexit->links[i], false)) 2425 return -EINVAL; 2426 2427 if (flags & BPF_TRAMP_F_CALL_ORIG) { 2428 im->ip_epilogue = jit->prg_buf + jit->prg; 2429 2430 /* 2431 * __bpf_tramp_exit(im); 2432 */ 2433 2434 /* %r1 = __bpf_tramp_exit */ 2435 load_imm64(jit, REG_1, (u64)__bpf_tramp_exit); 2436 /* %r2 = im */ 2437 load_imm64(jit, REG_2, (u64)im); 2438 /* %r1() */ 2439 call_r1(jit); 2440 } 2441 2442 /* lmg %r2,%rN,reg_args_off(%r15) */ 2443 if ((flags & BPF_TRAMP_F_RESTORE_REGS) && nr_reg_args) 2444 EMIT6_DISP_LH(0xeb000000, 0x0004, REG_2, 2445 REG_2 + (nr_reg_args - 1), REG_15, 2446 tjit->reg_args_off); 2447 /* lgr %r1,%r8 */ 2448 if (!(flags & BPF_TRAMP_F_SKIP_FRAME)) 2449 EMIT4(0xb9040000, REG_1, REG_8); 2450 /* lmg %r7,%r8,r7_r8_off(%r15) */ 2451 EMIT6_DISP_LH(0xeb000000, 0x0004, REG_7, REG_8, REG_15, 2452 tjit->r7_r8_off); 2453 /* lg %r14,r14_off(%r15) */ 2454 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_14, REG_0, REG_15, tjit->r14_off); 2455 /* lg %r2,retval_off(%r15) */ 2456 if (flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET)) 2457 EMIT6_DISP_LH(0xe3000000, 0x0004, REG_2, REG_0, REG_15, 2458 tjit->retval_off); 2459 /* aghi %r15,stack_size */ 2460 EMIT4_IMM(0xa70b0000, REG_15, tjit->stack_size); 2461 /* Emit an expoline for the following indirect jump. */ 2462 if (nospec_uses_trampoline()) 2463 emit_expoline(jit); 2464 if (flags & BPF_TRAMP_F_SKIP_FRAME) 2465 /* br %r14 */ 2466 _EMIT2(0x07fe); 2467 else 2468 /* br %r1 */ 2469 _EMIT2(0x07f1); 2470 2471 emit_r1_thunk(jit); 2472 2473 return 0; 2474 } 2475 2476 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, 2477 void *image_end, const struct btf_func_model *m, 2478 u32 flags, struct bpf_tramp_links *tlinks, 2479 void *func_addr) 2480 { 2481 struct bpf_tramp_jit tjit; 2482 int ret; 2483 int i; 2484 2485 for (i = 0; i < 2; i++) { 2486 if (i == 0) { 2487 /* Compute offsets, check whether the code fits. */ 2488 memset(&tjit, 0, sizeof(tjit)); 2489 } else { 2490 /* Generate the code. */ 2491 tjit.common.prg = 0; 2492 tjit.common.prg_buf = image; 2493 } 2494 ret = __arch_prepare_bpf_trampoline(im, &tjit, m, flags, 2495 tlinks, func_addr); 2496 if (ret < 0) 2497 return ret; 2498 if (tjit.common.prg > (char *)image_end - (char *)image) 2499 /* 2500 * Use the same error code as for exceeding 2501 * BPF_MAX_TRAMP_LINKS. 2502 */ 2503 return -E2BIG; 2504 } 2505 2506 return ret; 2507 } 2508 2509 bool bpf_jit_supports_subprog_tailcalls(void) 2510 { 2511 return true; 2512 } 2513