1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * bpf_jit_comp64.c: eBPF JIT compiler 4 * 5 * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com> 6 * IBM Corporation 7 * 8 * Based on the powerpc classic BPF JIT compiler by Matt Evans 9 */ 10 #include <linux/moduleloader.h> 11 #include <asm/cacheflush.h> 12 #include <asm/asm-compat.h> 13 #include <linux/netdevice.h> 14 #include <linux/filter.h> 15 #include <linux/if_vlan.h> 16 #include <asm/kprobes.h> 17 #include <linux/bpf.h> 18 #include <asm/security_features.h> 19 20 #include "bpf_jit.h" 21 22 /* 23 * Stack layout: 24 * Ensure the top half (upto local_tmp_var) stays consistent 25 * with our redzone usage. 26 * 27 * [ prev sp ] <------------- 28 * [ nv gpr save area ] 5*8 | 29 * [ tail_call_cnt ] 8 | 30 * [ local_tmp_var ] 16 | 31 * fp (r31) --> [ ebpf stack space ] upto 512 | 32 * [ frame header ] 32/112 | 33 * sp (r1) ---> [ stack pointer ] -------------- 34 */ 35 36 /* for gpr non volatile registers BPG_REG_6 to 10 */ 37 #define BPF_PPC_STACK_SAVE (5*8) 38 /* for bpf JIT code internal usage */ 39 #define BPF_PPC_STACK_LOCALS 24 40 /* stack frame excluding BPF stack, ensure this is quadword aligned */ 41 #define BPF_PPC_STACKFRAME (STACK_FRAME_MIN_SIZE + \ 42 BPF_PPC_STACK_LOCALS + BPF_PPC_STACK_SAVE) 43 44 /* BPF register usage */ 45 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0) 46 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1) 47 48 /* BPF to ppc register mappings */ 49 void bpf_jit_init_reg_mapping(struct codegen_context *ctx) 50 { 51 /* function return value */ 52 ctx->b2p[BPF_REG_0] = _R8; 53 /* function arguments */ 54 ctx->b2p[BPF_REG_1] = _R3; 55 ctx->b2p[BPF_REG_2] = _R4; 56 ctx->b2p[BPF_REG_3] = _R5; 57 ctx->b2p[BPF_REG_4] = _R6; 58 ctx->b2p[BPF_REG_5] = _R7; 59 /* non volatile registers */ 60 ctx->b2p[BPF_REG_6] = _R27; 61 ctx->b2p[BPF_REG_7] = _R28; 62 ctx->b2p[BPF_REG_8] = _R29; 63 ctx->b2p[BPF_REG_9] = _R30; 64 /* frame pointer aka BPF_REG_10 */ 65 ctx->b2p[BPF_REG_FP] = _R31; 66 /* eBPF jit internal registers */ 67 ctx->b2p[BPF_REG_AX] = _R12; 68 ctx->b2p[TMP_REG_1] = _R9; 69 ctx->b2p[TMP_REG_2] = _R10; 70 } 71 72 /* PPC NVR range -- update this if we ever use NVRs below r27 */ 73 #define BPF_PPC_NVR_MIN _R27 74 75 static inline bool bpf_has_stack_frame(struct codegen_context *ctx) 76 { 77 /* 78 * We only need a stack frame if: 79 * - we call other functions (kernel helpers), or 80 * - the bpf program uses its stack area 81 * The latter condition is deduced from the usage of BPF_REG_FP 82 */ 83 return ctx->seen & SEEN_FUNC || bpf_is_seen_register(ctx, bpf_to_ppc(BPF_REG_FP)); 84 } 85 86 /* 87 * When not setting up our own stackframe, the redzone usage is: 88 * 89 * [ prev sp ] <------------- 90 * [ ... ] | 91 * sp (r1) ---> [ stack pointer ] -------------- 92 * [ nv gpr save area ] 5*8 93 * [ tail_call_cnt ] 8 94 * [ local_tmp_var ] 16 95 * [ unused red zone ] 208 bytes protected 96 */ 97 static int bpf_jit_stack_local(struct codegen_context *ctx) 98 { 99 if (bpf_has_stack_frame(ctx)) 100 return STACK_FRAME_MIN_SIZE + ctx->stack_size; 101 else 102 return -(BPF_PPC_STACK_SAVE + 24); 103 } 104 105 static int bpf_jit_stack_tailcallcnt(struct codegen_context *ctx) 106 { 107 return bpf_jit_stack_local(ctx) + 16; 108 } 109 110 static int bpf_jit_stack_offsetof(struct codegen_context *ctx, int reg) 111 { 112 if (reg >= BPF_PPC_NVR_MIN && reg < 32) 113 return (bpf_has_stack_frame(ctx) ? 114 (BPF_PPC_STACKFRAME + ctx->stack_size) : 0) 115 - (8 * (32 - reg)); 116 117 pr_err("BPF JIT is asking about unknown registers"); 118 BUG(); 119 } 120 121 void bpf_jit_realloc_regs(struct codegen_context *ctx) 122 { 123 } 124 125 void bpf_jit_build_prologue(u32 *image, struct codegen_context *ctx) 126 { 127 int i; 128 129 #ifndef CONFIG_PPC_KERNEL_PCREL 130 if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2)) 131 EMIT(PPC_RAW_LD(_R2, _R13, offsetof(struct paca_struct, kernel_toc))); 132 #endif 133 134 /* 135 * Initialize tail_call_cnt if we do tail calls. 136 * Otherwise, put in NOPs so that it can be skipped when we are 137 * invoked through a tail call. 138 */ 139 if (ctx->seen & SEEN_TAILCALL) { 140 EMIT(PPC_RAW_LI(bpf_to_ppc(TMP_REG_1), 0)); 141 /* this goes in the redzone */ 142 EMIT(PPC_RAW_STD(bpf_to_ppc(TMP_REG_1), _R1, -(BPF_PPC_STACK_SAVE + 8))); 143 } else { 144 EMIT(PPC_RAW_NOP()); 145 EMIT(PPC_RAW_NOP()); 146 } 147 148 if (bpf_has_stack_frame(ctx)) { 149 /* 150 * We need a stack frame, but we don't necessarily need to 151 * save/restore LR unless we call other functions 152 */ 153 if (ctx->seen & SEEN_FUNC) { 154 EMIT(PPC_RAW_MFLR(_R0)); 155 EMIT(PPC_RAW_STD(_R0, _R1, PPC_LR_STKOFF)); 156 } 157 158 EMIT(PPC_RAW_STDU(_R1, _R1, -(BPF_PPC_STACKFRAME + ctx->stack_size))); 159 } 160 161 /* 162 * Back up non-volatile regs -- BPF registers 6-10 163 * If we haven't created our own stack frame, we save these 164 * in the protected zone below the previous stack frame 165 */ 166 for (i = BPF_REG_6; i <= BPF_REG_10; i++) 167 if (bpf_is_seen_register(ctx, bpf_to_ppc(i))) 168 EMIT(PPC_RAW_STD(bpf_to_ppc(i), _R1, bpf_jit_stack_offsetof(ctx, bpf_to_ppc(i)))); 169 170 /* Setup frame pointer to point to the bpf stack area */ 171 if (bpf_is_seen_register(ctx, bpf_to_ppc(BPF_REG_FP))) 172 EMIT(PPC_RAW_ADDI(bpf_to_ppc(BPF_REG_FP), _R1, 173 STACK_FRAME_MIN_SIZE + ctx->stack_size)); 174 } 175 176 static void bpf_jit_emit_common_epilogue(u32 *image, struct codegen_context *ctx) 177 { 178 int i; 179 180 /* Restore NVRs */ 181 for (i = BPF_REG_6; i <= BPF_REG_10; i++) 182 if (bpf_is_seen_register(ctx, bpf_to_ppc(i))) 183 EMIT(PPC_RAW_LD(bpf_to_ppc(i), _R1, bpf_jit_stack_offsetof(ctx, bpf_to_ppc(i)))); 184 185 /* Tear down our stack frame */ 186 if (bpf_has_stack_frame(ctx)) { 187 EMIT(PPC_RAW_ADDI(_R1, _R1, BPF_PPC_STACKFRAME + ctx->stack_size)); 188 if (ctx->seen & SEEN_FUNC) { 189 EMIT(PPC_RAW_LD(_R0, _R1, PPC_LR_STKOFF)); 190 EMIT(PPC_RAW_MTLR(_R0)); 191 } 192 } 193 } 194 195 void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx) 196 { 197 bpf_jit_emit_common_epilogue(image, ctx); 198 199 /* Move result to r3 */ 200 EMIT(PPC_RAW_MR(_R3, bpf_to_ppc(BPF_REG_0))); 201 202 EMIT(PPC_RAW_BLR()); 203 } 204 205 static int bpf_jit_emit_func_call_hlp(u32 *image, struct codegen_context *ctx, u64 func) 206 { 207 unsigned long func_addr = func ? ppc_function_entry((void *)func) : 0; 208 long reladdr; 209 210 if (WARN_ON_ONCE(!core_kernel_text(func_addr))) 211 return -EINVAL; 212 213 if (IS_ENABLED(CONFIG_PPC_KERNEL_PCREL)) { 214 reladdr = func_addr - CTX_NIA(ctx); 215 216 if (reladdr >= (long)SZ_8G || reladdr < -(long)SZ_8G) { 217 pr_err("eBPF: address of %ps out of range of pcrel address.\n", 218 (void *)func); 219 return -ERANGE; 220 } 221 /* pla r12,addr */ 222 EMIT(PPC_PREFIX_MLS | __PPC_PRFX_R(1) | IMM_H18(reladdr)); 223 EMIT(PPC_INST_PADDI | ___PPC_RT(_R12) | IMM_L(reladdr)); 224 EMIT(PPC_RAW_MTCTR(_R12)); 225 EMIT(PPC_RAW_BCTR()); 226 227 } else { 228 reladdr = func_addr - kernel_toc_addr(); 229 if (reladdr > 0x7FFFFFFF || reladdr < -(0x80000000L)) { 230 pr_err("eBPF: address of %ps out of range of kernel_toc.\n", (void *)func); 231 return -ERANGE; 232 } 233 234 EMIT(PPC_RAW_ADDIS(_R12, _R2, PPC_HA(reladdr))); 235 EMIT(PPC_RAW_ADDI(_R12, _R12, PPC_LO(reladdr))); 236 EMIT(PPC_RAW_MTCTR(_R12)); 237 EMIT(PPC_RAW_BCTRL()); 238 } 239 240 return 0; 241 } 242 243 int bpf_jit_emit_func_call_rel(u32 *image, struct codegen_context *ctx, u64 func) 244 { 245 unsigned int i, ctx_idx = ctx->idx; 246 247 if (WARN_ON_ONCE(func && is_module_text_address(func))) 248 return -EINVAL; 249 250 /* skip past descriptor if elf v1 */ 251 func += FUNCTION_DESCR_SIZE; 252 253 /* Load function address into r12 */ 254 PPC_LI64(_R12, func); 255 256 /* For bpf-to-bpf function calls, the callee's address is unknown 257 * until the last extra pass. As seen above, we use PPC_LI64() to 258 * load the callee's address, but this may optimize the number of 259 * instructions required based on the nature of the address. 260 * 261 * Since we don't want the number of instructions emitted to increase, 262 * we pad the optimized PPC_LI64() call with NOPs to guarantee that 263 * we always have a five-instruction sequence, which is the maximum 264 * that PPC_LI64() can emit. 265 */ 266 if (!image) 267 for (i = ctx->idx - ctx_idx; i < 5; i++) 268 EMIT(PPC_RAW_NOP()); 269 270 EMIT(PPC_RAW_MTCTR(_R12)); 271 EMIT(PPC_RAW_BCTRL()); 272 273 return 0; 274 } 275 276 static int bpf_jit_emit_tail_call(u32 *image, struct codegen_context *ctx, u32 out) 277 { 278 /* 279 * By now, the eBPF program has already setup parameters in r3, r4 and r5 280 * r3/BPF_REG_1 - pointer to ctx -- passed as is to the next bpf program 281 * r4/BPF_REG_2 - pointer to bpf_array 282 * r5/BPF_REG_3 - index in bpf_array 283 */ 284 int b2p_bpf_array = bpf_to_ppc(BPF_REG_2); 285 int b2p_index = bpf_to_ppc(BPF_REG_3); 286 int bpf_tailcall_prologue_size = 8; 287 288 if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2)) 289 bpf_tailcall_prologue_size += 4; /* skip past the toc load */ 290 291 /* 292 * if (index >= array->map.max_entries) 293 * goto out; 294 */ 295 EMIT(PPC_RAW_LWZ(bpf_to_ppc(TMP_REG_1), b2p_bpf_array, offsetof(struct bpf_array, map.max_entries))); 296 EMIT(PPC_RAW_RLWINM(b2p_index, b2p_index, 0, 0, 31)); 297 EMIT(PPC_RAW_CMPLW(b2p_index, bpf_to_ppc(TMP_REG_1))); 298 PPC_BCC_SHORT(COND_GE, out); 299 300 /* 301 * if (tail_call_cnt >= MAX_TAIL_CALL_CNT) 302 * goto out; 303 */ 304 EMIT(PPC_RAW_LD(bpf_to_ppc(TMP_REG_1), _R1, bpf_jit_stack_tailcallcnt(ctx))); 305 EMIT(PPC_RAW_CMPLWI(bpf_to_ppc(TMP_REG_1), MAX_TAIL_CALL_CNT)); 306 PPC_BCC_SHORT(COND_GE, out); 307 308 /* 309 * tail_call_cnt++; 310 */ 311 EMIT(PPC_RAW_ADDI(bpf_to_ppc(TMP_REG_1), bpf_to_ppc(TMP_REG_1), 1)); 312 EMIT(PPC_RAW_STD(bpf_to_ppc(TMP_REG_1), _R1, bpf_jit_stack_tailcallcnt(ctx))); 313 314 /* prog = array->ptrs[index]; */ 315 EMIT(PPC_RAW_MULI(bpf_to_ppc(TMP_REG_1), b2p_index, 8)); 316 EMIT(PPC_RAW_ADD(bpf_to_ppc(TMP_REG_1), bpf_to_ppc(TMP_REG_1), b2p_bpf_array)); 317 EMIT(PPC_RAW_LD(bpf_to_ppc(TMP_REG_1), bpf_to_ppc(TMP_REG_1), offsetof(struct bpf_array, ptrs))); 318 319 /* 320 * if (prog == NULL) 321 * goto out; 322 */ 323 EMIT(PPC_RAW_CMPLDI(bpf_to_ppc(TMP_REG_1), 0)); 324 PPC_BCC_SHORT(COND_EQ, out); 325 326 /* goto *(prog->bpf_func + prologue_size); */ 327 EMIT(PPC_RAW_LD(bpf_to_ppc(TMP_REG_1), bpf_to_ppc(TMP_REG_1), offsetof(struct bpf_prog, bpf_func))); 328 EMIT(PPC_RAW_ADDI(bpf_to_ppc(TMP_REG_1), bpf_to_ppc(TMP_REG_1), 329 FUNCTION_DESCR_SIZE + bpf_tailcall_prologue_size)); 330 EMIT(PPC_RAW_MTCTR(bpf_to_ppc(TMP_REG_1))); 331 332 /* tear down stack, restore NVRs, ... */ 333 bpf_jit_emit_common_epilogue(image, ctx); 334 335 EMIT(PPC_RAW_BCTR()); 336 337 /* out: */ 338 return 0; 339 } 340 341 /* 342 * We spill into the redzone always, even if the bpf program has its own stackframe. 343 * Offsets hardcoded based on BPF_PPC_STACK_SAVE -- see bpf_jit_stack_local() 344 */ 345 void bpf_stf_barrier(void); 346 347 asm ( 348 " .global bpf_stf_barrier ;" 349 " bpf_stf_barrier: ;" 350 " std 21,-64(1) ;" 351 " std 22,-56(1) ;" 352 " sync ;" 353 " ld 21,-64(1) ;" 354 " ld 22,-56(1) ;" 355 " ori 31,31,0 ;" 356 " .rept 14 ;" 357 " b 1f ;" 358 " 1: ;" 359 " .endr ;" 360 " blr ;" 361 ); 362 363 /* Assemble the body code between the prologue & epilogue */ 364 int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *ctx, 365 u32 *addrs, int pass, bool extra_pass) 366 { 367 enum stf_barrier_type stf_barrier = stf_barrier_type_get(); 368 const struct bpf_insn *insn = fp->insnsi; 369 int flen = fp->len; 370 int i, ret; 371 372 /* Start of epilogue code - will only be valid 2nd pass onwards */ 373 u32 exit_addr = addrs[flen]; 374 375 for (i = 0; i < flen; i++) { 376 u32 code = insn[i].code; 377 u32 dst_reg = bpf_to_ppc(insn[i].dst_reg); 378 u32 src_reg = bpf_to_ppc(insn[i].src_reg); 379 u32 size = BPF_SIZE(code); 380 u32 tmp1_reg = bpf_to_ppc(TMP_REG_1); 381 u32 tmp2_reg = bpf_to_ppc(TMP_REG_2); 382 u32 save_reg, ret_reg; 383 s16 off = insn[i].off; 384 s32 imm = insn[i].imm; 385 bool func_addr_fixed; 386 u64 func_addr; 387 u64 imm64; 388 u32 true_cond; 389 u32 tmp_idx; 390 int j; 391 392 /* 393 * addrs[] maps a BPF bytecode address into a real offset from 394 * the start of the body code. 395 */ 396 addrs[i] = ctx->idx * 4; 397 398 /* 399 * As an optimization, we note down which non-volatile registers 400 * are used so that we can only save/restore those in our 401 * prologue and epilogue. We do this here regardless of whether 402 * the actual BPF instruction uses src/dst registers or not 403 * (for instance, BPF_CALL does not use them). The expectation 404 * is that those instructions will have src_reg/dst_reg set to 405 * 0. Even otherwise, we just lose some prologue/epilogue 406 * optimization but everything else should work without 407 * any issues. 408 */ 409 if (dst_reg >= BPF_PPC_NVR_MIN && dst_reg < 32) 410 bpf_set_seen_register(ctx, dst_reg); 411 if (src_reg >= BPF_PPC_NVR_MIN && src_reg < 32) 412 bpf_set_seen_register(ctx, src_reg); 413 414 switch (code) { 415 /* 416 * Arithmetic operations: ADD/SUB/MUL/DIV/MOD/NEG 417 */ 418 case BPF_ALU | BPF_ADD | BPF_X: /* (u32) dst += (u32) src */ 419 case BPF_ALU64 | BPF_ADD | BPF_X: /* dst += src */ 420 EMIT(PPC_RAW_ADD(dst_reg, dst_reg, src_reg)); 421 goto bpf_alu32_trunc; 422 case BPF_ALU | BPF_SUB | BPF_X: /* (u32) dst -= (u32) src */ 423 case BPF_ALU64 | BPF_SUB | BPF_X: /* dst -= src */ 424 EMIT(PPC_RAW_SUB(dst_reg, dst_reg, src_reg)); 425 goto bpf_alu32_trunc; 426 case BPF_ALU | BPF_ADD | BPF_K: /* (u32) dst += (u32) imm */ 427 case BPF_ALU64 | BPF_ADD | BPF_K: /* dst += imm */ 428 if (!imm) { 429 goto bpf_alu32_trunc; 430 } else if (imm >= -32768 && imm < 32768) { 431 EMIT(PPC_RAW_ADDI(dst_reg, dst_reg, IMM_L(imm))); 432 } else { 433 PPC_LI32(tmp1_reg, imm); 434 EMIT(PPC_RAW_ADD(dst_reg, dst_reg, tmp1_reg)); 435 } 436 goto bpf_alu32_trunc; 437 case BPF_ALU | BPF_SUB | BPF_K: /* (u32) dst -= (u32) imm */ 438 case BPF_ALU64 | BPF_SUB | BPF_K: /* dst -= imm */ 439 if (!imm) { 440 goto bpf_alu32_trunc; 441 } else if (imm > -32768 && imm <= 32768) { 442 EMIT(PPC_RAW_ADDI(dst_reg, dst_reg, IMM_L(-imm))); 443 } else { 444 PPC_LI32(tmp1_reg, imm); 445 EMIT(PPC_RAW_SUB(dst_reg, dst_reg, tmp1_reg)); 446 } 447 goto bpf_alu32_trunc; 448 case BPF_ALU | BPF_MUL | BPF_X: /* (u32) dst *= (u32) src */ 449 case BPF_ALU64 | BPF_MUL | BPF_X: /* dst *= src */ 450 if (BPF_CLASS(code) == BPF_ALU) 451 EMIT(PPC_RAW_MULW(dst_reg, dst_reg, src_reg)); 452 else 453 EMIT(PPC_RAW_MULD(dst_reg, dst_reg, src_reg)); 454 goto bpf_alu32_trunc; 455 case BPF_ALU | BPF_MUL | BPF_K: /* (u32) dst *= (u32) imm */ 456 case BPF_ALU64 | BPF_MUL | BPF_K: /* dst *= imm */ 457 if (imm >= -32768 && imm < 32768) 458 EMIT(PPC_RAW_MULI(dst_reg, dst_reg, IMM_L(imm))); 459 else { 460 PPC_LI32(tmp1_reg, imm); 461 if (BPF_CLASS(code) == BPF_ALU) 462 EMIT(PPC_RAW_MULW(dst_reg, dst_reg, tmp1_reg)); 463 else 464 EMIT(PPC_RAW_MULD(dst_reg, dst_reg, tmp1_reg)); 465 } 466 goto bpf_alu32_trunc; 467 case BPF_ALU | BPF_DIV | BPF_X: /* (u32) dst /= (u32) src */ 468 case BPF_ALU | BPF_MOD | BPF_X: /* (u32) dst %= (u32) src */ 469 if (BPF_OP(code) == BPF_MOD) { 470 EMIT(PPC_RAW_DIVWU(tmp1_reg, dst_reg, src_reg)); 471 EMIT(PPC_RAW_MULW(tmp1_reg, src_reg, tmp1_reg)); 472 EMIT(PPC_RAW_SUB(dst_reg, dst_reg, tmp1_reg)); 473 } else 474 EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg, src_reg)); 475 goto bpf_alu32_trunc; 476 case BPF_ALU64 | BPF_DIV | BPF_X: /* dst /= src */ 477 case BPF_ALU64 | BPF_MOD | BPF_X: /* dst %= src */ 478 if (BPF_OP(code) == BPF_MOD) { 479 EMIT(PPC_RAW_DIVDU(tmp1_reg, dst_reg, src_reg)); 480 EMIT(PPC_RAW_MULD(tmp1_reg, src_reg, tmp1_reg)); 481 EMIT(PPC_RAW_SUB(dst_reg, dst_reg, tmp1_reg)); 482 } else 483 EMIT(PPC_RAW_DIVDU(dst_reg, dst_reg, src_reg)); 484 break; 485 case BPF_ALU | BPF_MOD | BPF_K: /* (u32) dst %= (u32) imm */ 486 case BPF_ALU | BPF_DIV | BPF_K: /* (u32) dst /= (u32) imm */ 487 case BPF_ALU64 | BPF_MOD | BPF_K: /* dst %= imm */ 488 case BPF_ALU64 | BPF_DIV | BPF_K: /* dst /= imm */ 489 if (imm == 0) 490 return -EINVAL; 491 if (imm == 1) { 492 if (BPF_OP(code) == BPF_DIV) { 493 goto bpf_alu32_trunc; 494 } else { 495 EMIT(PPC_RAW_LI(dst_reg, 0)); 496 break; 497 } 498 } 499 500 PPC_LI32(tmp1_reg, imm); 501 switch (BPF_CLASS(code)) { 502 case BPF_ALU: 503 if (BPF_OP(code) == BPF_MOD) { 504 EMIT(PPC_RAW_DIVWU(tmp2_reg, dst_reg, tmp1_reg)); 505 EMIT(PPC_RAW_MULW(tmp1_reg, tmp1_reg, tmp2_reg)); 506 EMIT(PPC_RAW_SUB(dst_reg, dst_reg, tmp1_reg)); 507 } else 508 EMIT(PPC_RAW_DIVWU(dst_reg, dst_reg, tmp1_reg)); 509 break; 510 case BPF_ALU64: 511 if (BPF_OP(code) == BPF_MOD) { 512 EMIT(PPC_RAW_DIVDU(tmp2_reg, dst_reg, tmp1_reg)); 513 EMIT(PPC_RAW_MULD(tmp1_reg, tmp1_reg, tmp2_reg)); 514 EMIT(PPC_RAW_SUB(dst_reg, dst_reg, tmp1_reg)); 515 } else 516 EMIT(PPC_RAW_DIVDU(dst_reg, dst_reg, tmp1_reg)); 517 break; 518 } 519 goto bpf_alu32_trunc; 520 case BPF_ALU | BPF_NEG: /* (u32) dst = -dst */ 521 case BPF_ALU64 | BPF_NEG: /* dst = -dst */ 522 EMIT(PPC_RAW_NEG(dst_reg, dst_reg)); 523 goto bpf_alu32_trunc; 524 525 /* 526 * Logical operations: AND/OR/XOR/[A]LSH/[A]RSH 527 */ 528 case BPF_ALU | BPF_AND | BPF_X: /* (u32) dst = dst & src */ 529 case BPF_ALU64 | BPF_AND | BPF_X: /* dst = dst & src */ 530 EMIT(PPC_RAW_AND(dst_reg, dst_reg, src_reg)); 531 goto bpf_alu32_trunc; 532 case BPF_ALU | BPF_AND | BPF_K: /* (u32) dst = dst & imm */ 533 case BPF_ALU64 | BPF_AND | BPF_K: /* dst = dst & imm */ 534 if (!IMM_H(imm)) 535 EMIT(PPC_RAW_ANDI(dst_reg, dst_reg, IMM_L(imm))); 536 else { 537 /* Sign-extended */ 538 PPC_LI32(tmp1_reg, imm); 539 EMIT(PPC_RAW_AND(dst_reg, dst_reg, tmp1_reg)); 540 } 541 goto bpf_alu32_trunc; 542 case BPF_ALU | BPF_OR | BPF_X: /* dst = (u32) dst | (u32) src */ 543 case BPF_ALU64 | BPF_OR | BPF_X: /* dst = dst | src */ 544 EMIT(PPC_RAW_OR(dst_reg, dst_reg, src_reg)); 545 goto bpf_alu32_trunc; 546 case BPF_ALU | BPF_OR | BPF_K:/* dst = (u32) dst | (u32) imm */ 547 case BPF_ALU64 | BPF_OR | BPF_K:/* dst = dst | imm */ 548 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) { 549 /* Sign-extended */ 550 PPC_LI32(tmp1_reg, imm); 551 EMIT(PPC_RAW_OR(dst_reg, dst_reg, tmp1_reg)); 552 } else { 553 if (IMM_L(imm)) 554 EMIT(PPC_RAW_ORI(dst_reg, dst_reg, IMM_L(imm))); 555 if (IMM_H(imm)) 556 EMIT(PPC_RAW_ORIS(dst_reg, dst_reg, IMM_H(imm))); 557 } 558 goto bpf_alu32_trunc; 559 case BPF_ALU | BPF_XOR | BPF_X: /* (u32) dst ^= src */ 560 case BPF_ALU64 | BPF_XOR | BPF_X: /* dst ^= src */ 561 EMIT(PPC_RAW_XOR(dst_reg, dst_reg, src_reg)); 562 goto bpf_alu32_trunc; 563 case BPF_ALU | BPF_XOR | BPF_K: /* (u32) dst ^= (u32) imm */ 564 case BPF_ALU64 | BPF_XOR | BPF_K: /* dst ^= imm */ 565 if (imm < 0 && BPF_CLASS(code) == BPF_ALU64) { 566 /* Sign-extended */ 567 PPC_LI32(tmp1_reg, imm); 568 EMIT(PPC_RAW_XOR(dst_reg, dst_reg, tmp1_reg)); 569 } else { 570 if (IMM_L(imm)) 571 EMIT(PPC_RAW_XORI(dst_reg, dst_reg, IMM_L(imm))); 572 if (IMM_H(imm)) 573 EMIT(PPC_RAW_XORIS(dst_reg, dst_reg, IMM_H(imm))); 574 } 575 goto bpf_alu32_trunc; 576 case BPF_ALU | BPF_LSH | BPF_X: /* (u32) dst <<= (u32) src */ 577 /* slw clears top 32 bits */ 578 EMIT(PPC_RAW_SLW(dst_reg, dst_reg, src_reg)); 579 /* skip zero extension move, but set address map. */ 580 if (insn_is_zext(&insn[i + 1])) 581 addrs[++i] = ctx->idx * 4; 582 break; 583 case BPF_ALU64 | BPF_LSH | BPF_X: /* dst <<= src; */ 584 EMIT(PPC_RAW_SLD(dst_reg, dst_reg, src_reg)); 585 break; 586 case BPF_ALU | BPF_LSH | BPF_K: /* (u32) dst <<== (u32) imm */ 587 /* with imm 0, we still need to clear top 32 bits */ 588 EMIT(PPC_RAW_SLWI(dst_reg, dst_reg, imm)); 589 if (insn_is_zext(&insn[i + 1])) 590 addrs[++i] = ctx->idx * 4; 591 break; 592 case BPF_ALU64 | BPF_LSH | BPF_K: /* dst <<== imm */ 593 if (imm != 0) 594 EMIT(PPC_RAW_SLDI(dst_reg, dst_reg, imm)); 595 break; 596 case BPF_ALU | BPF_RSH | BPF_X: /* (u32) dst >>= (u32) src */ 597 EMIT(PPC_RAW_SRW(dst_reg, dst_reg, src_reg)); 598 if (insn_is_zext(&insn[i + 1])) 599 addrs[++i] = ctx->idx * 4; 600 break; 601 case BPF_ALU64 | BPF_RSH | BPF_X: /* dst >>= src */ 602 EMIT(PPC_RAW_SRD(dst_reg, dst_reg, src_reg)); 603 break; 604 case BPF_ALU | BPF_RSH | BPF_K: /* (u32) dst >>= (u32) imm */ 605 EMIT(PPC_RAW_SRWI(dst_reg, dst_reg, imm)); 606 if (insn_is_zext(&insn[i + 1])) 607 addrs[++i] = ctx->idx * 4; 608 break; 609 case BPF_ALU64 | BPF_RSH | BPF_K: /* dst >>= imm */ 610 if (imm != 0) 611 EMIT(PPC_RAW_SRDI(dst_reg, dst_reg, imm)); 612 break; 613 case BPF_ALU | BPF_ARSH | BPF_X: /* (s32) dst >>= src */ 614 EMIT(PPC_RAW_SRAW(dst_reg, dst_reg, src_reg)); 615 goto bpf_alu32_trunc; 616 case BPF_ALU64 | BPF_ARSH | BPF_X: /* (s64) dst >>= src */ 617 EMIT(PPC_RAW_SRAD(dst_reg, dst_reg, src_reg)); 618 break; 619 case BPF_ALU | BPF_ARSH | BPF_K: /* (s32) dst >>= imm */ 620 EMIT(PPC_RAW_SRAWI(dst_reg, dst_reg, imm)); 621 goto bpf_alu32_trunc; 622 case BPF_ALU64 | BPF_ARSH | BPF_K: /* (s64) dst >>= imm */ 623 if (imm != 0) 624 EMIT(PPC_RAW_SRADI(dst_reg, dst_reg, imm)); 625 break; 626 627 /* 628 * MOV 629 */ 630 case BPF_ALU | BPF_MOV | BPF_X: /* (u32) dst = src */ 631 case BPF_ALU64 | BPF_MOV | BPF_X: /* dst = src */ 632 if (imm == 1) { 633 /* special mov32 for zext */ 634 EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 0, 31)); 635 break; 636 } 637 EMIT(PPC_RAW_MR(dst_reg, src_reg)); 638 goto bpf_alu32_trunc; 639 case BPF_ALU | BPF_MOV | BPF_K: /* (u32) dst = imm */ 640 case BPF_ALU64 | BPF_MOV | BPF_K: /* dst = (s64) imm */ 641 PPC_LI32(dst_reg, imm); 642 if (imm < 0) 643 goto bpf_alu32_trunc; 644 else if (insn_is_zext(&insn[i + 1])) 645 addrs[++i] = ctx->idx * 4; 646 break; 647 648 bpf_alu32_trunc: 649 /* Truncate to 32-bits */ 650 if (BPF_CLASS(code) == BPF_ALU && !fp->aux->verifier_zext) 651 EMIT(PPC_RAW_RLWINM(dst_reg, dst_reg, 0, 0, 31)); 652 break; 653 654 /* 655 * BPF_FROM_BE/LE 656 */ 657 case BPF_ALU | BPF_END | BPF_FROM_LE: 658 case BPF_ALU | BPF_END | BPF_FROM_BE: 659 #ifdef __BIG_ENDIAN__ 660 if (BPF_SRC(code) == BPF_FROM_BE) 661 goto emit_clear; 662 #else /* !__BIG_ENDIAN__ */ 663 if (BPF_SRC(code) == BPF_FROM_LE) 664 goto emit_clear; 665 #endif 666 switch (imm) { 667 case 16: 668 /* Rotate 8 bits left & mask with 0x0000ff00 */ 669 EMIT(PPC_RAW_RLWINM(tmp1_reg, dst_reg, 8, 16, 23)); 670 /* Rotate 8 bits right & insert LSB to reg */ 671 EMIT(PPC_RAW_RLWIMI(tmp1_reg, dst_reg, 24, 24, 31)); 672 /* Move result back to dst_reg */ 673 EMIT(PPC_RAW_MR(dst_reg, tmp1_reg)); 674 break; 675 case 32: 676 /* 677 * Rotate word left by 8 bits: 678 * 2 bytes are already in their final position 679 * -- byte 2 and 4 (of bytes 1, 2, 3 and 4) 680 */ 681 EMIT(PPC_RAW_RLWINM(tmp1_reg, dst_reg, 8, 0, 31)); 682 /* Rotate 24 bits and insert byte 1 */ 683 EMIT(PPC_RAW_RLWIMI(tmp1_reg, dst_reg, 24, 0, 7)); 684 /* Rotate 24 bits and insert byte 3 */ 685 EMIT(PPC_RAW_RLWIMI(tmp1_reg, dst_reg, 24, 16, 23)); 686 EMIT(PPC_RAW_MR(dst_reg, tmp1_reg)); 687 break; 688 case 64: 689 /* Store the value to stack and then use byte-reverse loads */ 690 EMIT(PPC_RAW_STD(dst_reg, _R1, bpf_jit_stack_local(ctx))); 691 EMIT(PPC_RAW_ADDI(tmp1_reg, _R1, bpf_jit_stack_local(ctx))); 692 if (cpu_has_feature(CPU_FTR_ARCH_206)) { 693 EMIT(PPC_RAW_LDBRX(dst_reg, 0, tmp1_reg)); 694 } else { 695 EMIT(PPC_RAW_LWBRX(dst_reg, 0, tmp1_reg)); 696 if (IS_ENABLED(CONFIG_CPU_LITTLE_ENDIAN)) 697 EMIT(PPC_RAW_SLDI(dst_reg, dst_reg, 32)); 698 EMIT(PPC_RAW_LI(tmp2_reg, 4)); 699 EMIT(PPC_RAW_LWBRX(tmp2_reg, tmp2_reg, tmp1_reg)); 700 if (IS_ENABLED(CONFIG_CPU_BIG_ENDIAN)) 701 EMIT(PPC_RAW_SLDI(tmp2_reg, tmp2_reg, 32)); 702 EMIT(PPC_RAW_OR(dst_reg, dst_reg, tmp2_reg)); 703 } 704 break; 705 } 706 break; 707 708 emit_clear: 709 switch (imm) { 710 case 16: 711 /* zero-extend 16 bits into 64 bits */ 712 EMIT(PPC_RAW_RLDICL(dst_reg, dst_reg, 0, 48)); 713 if (insn_is_zext(&insn[i + 1])) 714 addrs[++i] = ctx->idx * 4; 715 break; 716 case 32: 717 if (!fp->aux->verifier_zext) 718 /* zero-extend 32 bits into 64 bits */ 719 EMIT(PPC_RAW_RLDICL(dst_reg, dst_reg, 0, 32)); 720 break; 721 case 64: 722 /* nop */ 723 break; 724 } 725 break; 726 727 /* 728 * BPF_ST NOSPEC (speculation barrier) 729 */ 730 case BPF_ST | BPF_NOSPEC: 731 if (!security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) || 732 !security_ftr_enabled(SEC_FTR_STF_BARRIER)) 733 break; 734 735 switch (stf_barrier) { 736 case STF_BARRIER_EIEIO: 737 EMIT(PPC_RAW_EIEIO() | 0x02000000); 738 break; 739 case STF_BARRIER_SYNC_ORI: 740 EMIT(PPC_RAW_SYNC()); 741 EMIT(PPC_RAW_LD(tmp1_reg, _R13, 0)); 742 EMIT(PPC_RAW_ORI(_R31, _R31, 0)); 743 break; 744 case STF_BARRIER_FALLBACK: 745 ctx->seen |= SEEN_FUNC; 746 PPC_LI64(_R12, dereference_kernel_function_descriptor(bpf_stf_barrier)); 747 EMIT(PPC_RAW_MTCTR(_R12)); 748 EMIT(PPC_RAW_BCTRL()); 749 break; 750 case STF_BARRIER_NONE: 751 break; 752 } 753 break; 754 755 /* 756 * BPF_ST(X) 757 */ 758 case BPF_STX | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = src */ 759 case BPF_ST | BPF_MEM | BPF_B: /* *(u8 *)(dst + off) = imm */ 760 if (BPF_CLASS(code) == BPF_ST) { 761 EMIT(PPC_RAW_LI(tmp1_reg, imm)); 762 src_reg = tmp1_reg; 763 } 764 EMIT(PPC_RAW_STB(src_reg, dst_reg, off)); 765 break; 766 case BPF_STX | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = src */ 767 case BPF_ST | BPF_MEM | BPF_H: /* (u16 *)(dst + off) = imm */ 768 if (BPF_CLASS(code) == BPF_ST) { 769 EMIT(PPC_RAW_LI(tmp1_reg, imm)); 770 src_reg = tmp1_reg; 771 } 772 EMIT(PPC_RAW_STH(src_reg, dst_reg, off)); 773 break; 774 case BPF_STX | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = src */ 775 case BPF_ST | BPF_MEM | BPF_W: /* *(u32 *)(dst + off) = imm */ 776 if (BPF_CLASS(code) == BPF_ST) { 777 PPC_LI32(tmp1_reg, imm); 778 src_reg = tmp1_reg; 779 } 780 EMIT(PPC_RAW_STW(src_reg, dst_reg, off)); 781 break; 782 case BPF_STX | BPF_MEM | BPF_DW: /* (u64 *)(dst + off) = src */ 783 case BPF_ST | BPF_MEM | BPF_DW: /* *(u64 *)(dst + off) = imm */ 784 if (BPF_CLASS(code) == BPF_ST) { 785 PPC_LI32(tmp1_reg, imm); 786 src_reg = tmp1_reg; 787 } 788 if (off % 4) { 789 EMIT(PPC_RAW_LI(tmp2_reg, off)); 790 EMIT(PPC_RAW_STDX(src_reg, dst_reg, tmp2_reg)); 791 } else { 792 EMIT(PPC_RAW_STD(src_reg, dst_reg, off)); 793 } 794 break; 795 796 /* 797 * BPF_STX ATOMIC (atomic ops) 798 */ 799 case BPF_STX | BPF_ATOMIC | BPF_W: 800 case BPF_STX | BPF_ATOMIC | BPF_DW: 801 save_reg = tmp2_reg; 802 ret_reg = src_reg; 803 804 /* Get offset into TMP_REG_1 */ 805 EMIT(PPC_RAW_LI(tmp1_reg, off)); 806 /* 807 * Enforce full ordering for operations with BPF_FETCH by emitting a 'sync' 808 * before and after the operation. 809 * 810 * This is a requirement in the Linux Kernel Memory Model. 811 * See __cmpxchg_u64() in asm/cmpxchg.h as an example. 812 */ 813 if ((imm & BPF_FETCH) && IS_ENABLED(CONFIG_SMP)) 814 EMIT(PPC_RAW_SYNC()); 815 tmp_idx = ctx->idx * 4; 816 /* load value from memory into TMP_REG_2 */ 817 if (size == BPF_DW) 818 EMIT(PPC_RAW_LDARX(tmp2_reg, tmp1_reg, dst_reg, 0)); 819 else 820 EMIT(PPC_RAW_LWARX(tmp2_reg, tmp1_reg, dst_reg, 0)); 821 822 /* Save old value in _R0 */ 823 if (imm & BPF_FETCH) 824 EMIT(PPC_RAW_MR(_R0, tmp2_reg)); 825 826 switch (imm) { 827 case BPF_ADD: 828 case BPF_ADD | BPF_FETCH: 829 EMIT(PPC_RAW_ADD(tmp2_reg, tmp2_reg, src_reg)); 830 break; 831 case BPF_AND: 832 case BPF_AND | BPF_FETCH: 833 EMIT(PPC_RAW_AND(tmp2_reg, tmp2_reg, src_reg)); 834 break; 835 case BPF_OR: 836 case BPF_OR | BPF_FETCH: 837 EMIT(PPC_RAW_OR(tmp2_reg, tmp2_reg, src_reg)); 838 break; 839 case BPF_XOR: 840 case BPF_XOR | BPF_FETCH: 841 EMIT(PPC_RAW_XOR(tmp2_reg, tmp2_reg, src_reg)); 842 break; 843 case BPF_CMPXCHG: 844 /* 845 * Return old value in BPF_REG_0 for BPF_CMPXCHG & 846 * in src_reg for other cases. 847 */ 848 ret_reg = bpf_to_ppc(BPF_REG_0); 849 850 /* Compare with old value in BPF_R0 */ 851 if (size == BPF_DW) 852 EMIT(PPC_RAW_CMPD(bpf_to_ppc(BPF_REG_0), tmp2_reg)); 853 else 854 EMIT(PPC_RAW_CMPW(bpf_to_ppc(BPF_REG_0), tmp2_reg)); 855 /* Don't set if different from old value */ 856 PPC_BCC_SHORT(COND_NE, (ctx->idx + 3) * 4); 857 fallthrough; 858 case BPF_XCHG: 859 save_reg = src_reg; 860 break; 861 default: 862 pr_err_ratelimited( 863 "eBPF filter atomic op code %02x (@%d) unsupported\n", 864 code, i); 865 return -EOPNOTSUPP; 866 } 867 868 /* store new value */ 869 if (size == BPF_DW) 870 EMIT(PPC_RAW_STDCX(save_reg, tmp1_reg, dst_reg)); 871 else 872 EMIT(PPC_RAW_STWCX(save_reg, tmp1_reg, dst_reg)); 873 /* we're done if this succeeded */ 874 PPC_BCC_SHORT(COND_NE, tmp_idx); 875 876 if (imm & BPF_FETCH) { 877 /* Emit 'sync' to enforce full ordering */ 878 if (IS_ENABLED(CONFIG_SMP)) 879 EMIT(PPC_RAW_SYNC()); 880 EMIT(PPC_RAW_MR(ret_reg, _R0)); 881 /* 882 * Skip unnecessary zero-extension for 32-bit cmpxchg. 883 * For context, see commit 39491867ace5. 884 */ 885 if (size != BPF_DW && imm == BPF_CMPXCHG && 886 insn_is_zext(&insn[i + 1])) 887 addrs[++i] = ctx->idx * 4; 888 } 889 break; 890 891 /* 892 * BPF_LDX 893 */ 894 /* dst = *(u8 *)(ul) (src + off) */ 895 case BPF_LDX | BPF_MEM | BPF_B: 896 case BPF_LDX | BPF_PROBE_MEM | BPF_B: 897 /* dst = *(u16 *)(ul) (src + off) */ 898 case BPF_LDX | BPF_MEM | BPF_H: 899 case BPF_LDX | BPF_PROBE_MEM | BPF_H: 900 /* dst = *(u32 *)(ul) (src + off) */ 901 case BPF_LDX | BPF_MEM | BPF_W: 902 case BPF_LDX | BPF_PROBE_MEM | BPF_W: 903 /* dst = *(u64 *)(ul) (src + off) */ 904 case BPF_LDX | BPF_MEM | BPF_DW: 905 case BPF_LDX | BPF_PROBE_MEM | BPF_DW: 906 /* 907 * As PTR_TO_BTF_ID that uses BPF_PROBE_MEM mode could either be a valid 908 * kernel pointer or NULL but not a userspace address, execute BPF_PROBE_MEM 909 * load only if addr is kernel address (see is_kernel_addr()), otherwise 910 * set dst_reg=0 and move on. 911 */ 912 if (BPF_MODE(code) == BPF_PROBE_MEM) { 913 EMIT(PPC_RAW_ADDI(tmp1_reg, src_reg, off)); 914 if (IS_ENABLED(CONFIG_PPC_BOOK3E_64)) 915 PPC_LI64(tmp2_reg, 0x8000000000000000ul); 916 else /* BOOK3S_64 */ 917 PPC_LI64(tmp2_reg, PAGE_OFFSET); 918 EMIT(PPC_RAW_CMPLD(tmp1_reg, tmp2_reg)); 919 PPC_BCC_SHORT(COND_GT, (ctx->idx + 3) * 4); 920 EMIT(PPC_RAW_LI(dst_reg, 0)); 921 /* 922 * Check if 'off' is word aligned for BPF_DW, because 923 * we might generate two instructions. 924 */ 925 if (BPF_SIZE(code) == BPF_DW && (off & 3)) 926 PPC_JMP((ctx->idx + 3) * 4); 927 else 928 PPC_JMP((ctx->idx + 2) * 4); 929 } 930 931 switch (size) { 932 case BPF_B: 933 EMIT(PPC_RAW_LBZ(dst_reg, src_reg, off)); 934 break; 935 case BPF_H: 936 EMIT(PPC_RAW_LHZ(dst_reg, src_reg, off)); 937 break; 938 case BPF_W: 939 EMIT(PPC_RAW_LWZ(dst_reg, src_reg, off)); 940 break; 941 case BPF_DW: 942 if (off % 4) { 943 EMIT(PPC_RAW_LI(tmp1_reg, off)); 944 EMIT(PPC_RAW_LDX(dst_reg, src_reg, tmp1_reg)); 945 } else { 946 EMIT(PPC_RAW_LD(dst_reg, src_reg, off)); 947 } 948 break; 949 } 950 951 if (size != BPF_DW && insn_is_zext(&insn[i + 1])) 952 addrs[++i] = ctx->idx * 4; 953 954 if (BPF_MODE(code) == BPF_PROBE_MEM) { 955 ret = bpf_add_extable_entry(fp, image, pass, ctx, ctx->idx - 1, 956 4, dst_reg); 957 if (ret) 958 return ret; 959 } 960 break; 961 962 /* 963 * Doubleword load 964 * 16 byte instruction that uses two 'struct bpf_insn' 965 */ 966 case BPF_LD | BPF_IMM | BPF_DW: /* dst = (u64) imm */ 967 imm64 = ((u64)(u32) insn[i].imm) | 968 (((u64)(u32) insn[i+1].imm) << 32); 969 tmp_idx = ctx->idx; 970 PPC_LI64(dst_reg, imm64); 971 /* padding to allow full 5 instructions for later patching */ 972 if (!image) 973 for (j = ctx->idx - tmp_idx; j < 5; j++) 974 EMIT(PPC_RAW_NOP()); 975 /* Adjust for two bpf instructions */ 976 addrs[++i] = ctx->idx * 4; 977 break; 978 979 /* 980 * Return/Exit 981 */ 982 case BPF_JMP | BPF_EXIT: 983 /* 984 * If this isn't the very last instruction, branch to 985 * the epilogue. If we _are_ the last instruction, 986 * we'll just fall through to the epilogue. 987 */ 988 if (i != flen - 1) { 989 ret = bpf_jit_emit_exit_insn(image, ctx, tmp1_reg, exit_addr); 990 if (ret) 991 return ret; 992 } 993 /* else fall through to the epilogue */ 994 break; 995 996 /* 997 * Call kernel helper or bpf function 998 */ 999 case BPF_JMP | BPF_CALL: 1000 ctx->seen |= SEEN_FUNC; 1001 1002 ret = bpf_jit_get_func_addr(fp, &insn[i], extra_pass, 1003 &func_addr, &func_addr_fixed); 1004 if (ret < 0) 1005 return ret; 1006 1007 if (func_addr_fixed) 1008 ret = bpf_jit_emit_func_call_hlp(image, ctx, func_addr); 1009 else 1010 ret = bpf_jit_emit_func_call_rel(image, ctx, func_addr); 1011 1012 if (ret) 1013 return ret; 1014 1015 /* move return value from r3 to BPF_REG_0 */ 1016 EMIT(PPC_RAW_MR(bpf_to_ppc(BPF_REG_0), _R3)); 1017 break; 1018 1019 /* 1020 * Jumps and branches 1021 */ 1022 case BPF_JMP | BPF_JA: 1023 PPC_JMP(addrs[i + 1 + off]); 1024 break; 1025 1026 case BPF_JMP | BPF_JGT | BPF_K: 1027 case BPF_JMP | BPF_JGT | BPF_X: 1028 case BPF_JMP | BPF_JSGT | BPF_K: 1029 case BPF_JMP | BPF_JSGT | BPF_X: 1030 case BPF_JMP32 | BPF_JGT | BPF_K: 1031 case BPF_JMP32 | BPF_JGT | BPF_X: 1032 case BPF_JMP32 | BPF_JSGT | BPF_K: 1033 case BPF_JMP32 | BPF_JSGT | BPF_X: 1034 true_cond = COND_GT; 1035 goto cond_branch; 1036 case BPF_JMP | BPF_JLT | BPF_K: 1037 case BPF_JMP | BPF_JLT | BPF_X: 1038 case BPF_JMP | BPF_JSLT | BPF_K: 1039 case BPF_JMP | BPF_JSLT | BPF_X: 1040 case BPF_JMP32 | BPF_JLT | BPF_K: 1041 case BPF_JMP32 | BPF_JLT | BPF_X: 1042 case BPF_JMP32 | BPF_JSLT | BPF_K: 1043 case BPF_JMP32 | BPF_JSLT | BPF_X: 1044 true_cond = COND_LT; 1045 goto cond_branch; 1046 case BPF_JMP | BPF_JGE | BPF_K: 1047 case BPF_JMP | BPF_JGE | BPF_X: 1048 case BPF_JMP | BPF_JSGE | BPF_K: 1049 case BPF_JMP | BPF_JSGE | BPF_X: 1050 case BPF_JMP32 | BPF_JGE | BPF_K: 1051 case BPF_JMP32 | BPF_JGE | BPF_X: 1052 case BPF_JMP32 | BPF_JSGE | BPF_K: 1053 case BPF_JMP32 | BPF_JSGE | BPF_X: 1054 true_cond = COND_GE; 1055 goto cond_branch; 1056 case BPF_JMP | BPF_JLE | BPF_K: 1057 case BPF_JMP | BPF_JLE | BPF_X: 1058 case BPF_JMP | BPF_JSLE | BPF_K: 1059 case BPF_JMP | BPF_JSLE | BPF_X: 1060 case BPF_JMP32 | BPF_JLE | BPF_K: 1061 case BPF_JMP32 | BPF_JLE | BPF_X: 1062 case BPF_JMP32 | BPF_JSLE | BPF_K: 1063 case BPF_JMP32 | BPF_JSLE | BPF_X: 1064 true_cond = COND_LE; 1065 goto cond_branch; 1066 case BPF_JMP | BPF_JEQ | BPF_K: 1067 case BPF_JMP | BPF_JEQ | BPF_X: 1068 case BPF_JMP32 | BPF_JEQ | BPF_K: 1069 case BPF_JMP32 | BPF_JEQ | BPF_X: 1070 true_cond = COND_EQ; 1071 goto cond_branch; 1072 case BPF_JMP | BPF_JNE | BPF_K: 1073 case BPF_JMP | BPF_JNE | BPF_X: 1074 case BPF_JMP32 | BPF_JNE | BPF_K: 1075 case BPF_JMP32 | BPF_JNE | BPF_X: 1076 true_cond = COND_NE; 1077 goto cond_branch; 1078 case BPF_JMP | BPF_JSET | BPF_K: 1079 case BPF_JMP | BPF_JSET | BPF_X: 1080 case BPF_JMP32 | BPF_JSET | BPF_K: 1081 case BPF_JMP32 | BPF_JSET | BPF_X: 1082 true_cond = COND_NE; 1083 /* Fall through */ 1084 1085 cond_branch: 1086 switch (code) { 1087 case BPF_JMP | BPF_JGT | BPF_X: 1088 case BPF_JMP | BPF_JLT | BPF_X: 1089 case BPF_JMP | BPF_JGE | BPF_X: 1090 case BPF_JMP | BPF_JLE | BPF_X: 1091 case BPF_JMP | BPF_JEQ | BPF_X: 1092 case BPF_JMP | BPF_JNE | BPF_X: 1093 case BPF_JMP32 | BPF_JGT | BPF_X: 1094 case BPF_JMP32 | BPF_JLT | BPF_X: 1095 case BPF_JMP32 | BPF_JGE | BPF_X: 1096 case BPF_JMP32 | BPF_JLE | BPF_X: 1097 case BPF_JMP32 | BPF_JEQ | BPF_X: 1098 case BPF_JMP32 | BPF_JNE | BPF_X: 1099 /* unsigned comparison */ 1100 if (BPF_CLASS(code) == BPF_JMP32) 1101 EMIT(PPC_RAW_CMPLW(dst_reg, src_reg)); 1102 else 1103 EMIT(PPC_RAW_CMPLD(dst_reg, src_reg)); 1104 break; 1105 case BPF_JMP | BPF_JSGT | BPF_X: 1106 case BPF_JMP | BPF_JSLT | BPF_X: 1107 case BPF_JMP | BPF_JSGE | BPF_X: 1108 case BPF_JMP | BPF_JSLE | BPF_X: 1109 case BPF_JMP32 | BPF_JSGT | BPF_X: 1110 case BPF_JMP32 | BPF_JSLT | BPF_X: 1111 case BPF_JMP32 | BPF_JSGE | BPF_X: 1112 case BPF_JMP32 | BPF_JSLE | BPF_X: 1113 /* signed comparison */ 1114 if (BPF_CLASS(code) == BPF_JMP32) 1115 EMIT(PPC_RAW_CMPW(dst_reg, src_reg)); 1116 else 1117 EMIT(PPC_RAW_CMPD(dst_reg, src_reg)); 1118 break; 1119 case BPF_JMP | BPF_JSET | BPF_X: 1120 case BPF_JMP32 | BPF_JSET | BPF_X: 1121 if (BPF_CLASS(code) == BPF_JMP) { 1122 EMIT(PPC_RAW_AND_DOT(tmp1_reg, dst_reg, src_reg)); 1123 } else { 1124 EMIT(PPC_RAW_AND(tmp1_reg, dst_reg, src_reg)); 1125 EMIT(PPC_RAW_RLWINM_DOT(tmp1_reg, tmp1_reg, 0, 0, 31)); 1126 } 1127 break; 1128 case BPF_JMP | BPF_JNE | BPF_K: 1129 case BPF_JMP | BPF_JEQ | BPF_K: 1130 case BPF_JMP | BPF_JGT | BPF_K: 1131 case BPF_JMP | BPF_JLT | BPF_K: 1132 case BPF_JMP | BPF_JGE | BPF_K: 1133 case BPF_JMP | BPF_JLE | BPF_K: 1134 case BPF_JMP32 | BPF_JNE | BPF_K: 1135 case BPF_JMP32 | BPF_JEQ | BPF_K: 1136 case BPF_JMP32 | BPF_JGT | BPF_K: 1137 case BPF_JMP32 | BPF_JLT | BPF_K: 1138 case BPF_JMP32 | BPF_JGE | BPF_K: 1139 case BPF_JMP32 | BPF_JLE | BPF_K: 1140 { 1141 bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32; 1142 1143 /* 1144 * Need sign-extended load, so only positive 1145 * values can be used as imm in cmpldi 1146 */ 1147 if (imm >= 0 && imm < 32768) { 1148 if (is_jmp32) 1149 EMIT(PPC_RAW_CMPLWI(dst_reg, imm)); 1150 else 1151 EMIT(PPC_RAW_CMPLDI(dst_reg, imm)); 1152 } else { 1153 /* sign-extending load */ 1154 PPC_LI32(tmp1_reg, imm); 1155 /* ... but unsigned comparison */ 1156 if (is_jmp32) 1157 EMIT(PPC_RAW_CMPLW(dst_reg, tmp1_reg)); 1158 else 1159 EMIT(PPC_RAW_CMPLD(dst_reg, tmp1_reg)); 1160 } 1161 break; 1162 } 1163 case BPF_JMP | BPF_JSGT | BPF_K: 1164 case BPF_JMP | BPF_JSLT | BPF_K: 1165 case BPF_JMP | BPF_JSGE | BPF_K: 1166 case BPF_JMP | BPF_JSLE | BPF_K: 1167 case BPF_JMP32 | BPF_JSGT | BPF_K: 1168 case BPF_JMP32 | BPF_JSLT | BPF_K: 1169 case BPF_JMP32 | BPF_JSGE | BPF_K: 1170 case BPF_JMP32 | BPF_JSLE | BPF_K: 1171 { 1172 bool is_jmp32 = BPF_CLASS(code) == BPF_JMP32; 1173 1174 /* 1175 * signed comparison, so any 16-bit value 1176 * can be used in cmpdi 1177 */ 1178 if (imm >= -32768 && imm < 32768) { 1179 if (is_jmp32) 1180 EMIT(PPC_RAW_CMPWI(dst_reg, imm)); 1181 else 1182 EMIT(PPC_RAW_CMPDI(dst_reg, imm)); 1183 } else { 1184 PPC_LI32(tmp1_reg, imm); 1185 if (is_jmp32) 1186 EMIT(PPC_RAW_CMPW(dst_reg, tmp1_reg)); 1187 else 1188 EMIT(PPC_RAW_CMPD(dst_reg, tmp1_reg)); 1189 } 1190 break; 1191 } 1192 case BPF_JMP | BPF_JSET | BPF_K: 1193 case BPF_JMP32 | BPF_JSET | BPF_K: 1194 /* andi does not sign-extend the immediate */ 1195 if (imm >= 0 && imm < 32768) 1196 /* PPC_ANDI is _only/always_ dot-form */ 1197 EMIT(PPC_RAW_ANDI(tmp1_reg, dst_reg, imm)); 1198 else { 1199 PPC_LI32(tmp1_reg, imm); 1200 if (BPF_CLASS(code) == BPF_JMP) { 1201 EMIT(PPC_RAW_AND_DOT(tmp1_reg, dst_reg, 1202 tmp1_reg)); 1203 } else { 1204 EMIT(PPC_RAW_AND(tmp1_reg, dst_reg, tmp1_reg)); 1205 EMIT(PPC_RAW_RLWINM_DOT(tmp1_reg, tmp1_reg, 1206 0, 0, 31)); 1207 } 1208 } 1209 break; 1210 } 1211 PPC_BCC(true_cond, addrs[i + 1 + off]); 1212 break; 1213 1214 /* 1215 * Tail call 1216 */ 1217 case BPF_JMP | BPF_TAIL_CALL: 1218 ctx->seen |= SEEN_TAILCALL; 1219 ret = bpf_jit_emit_tail_call(image, ctx, addrs[i + 1]); 1220 if (ret < 0) 1221 return ret; 1222 break; 1223 1224 default: 1225 /* 1226 * The filter contains something cruel & unusual. 1227 * We don't handle it, but also there shouldn't be 1228 * anything missing from our list. 1229 */ 1230 pr_err_ratelimited("eBPF filter opcode %04x (@%d) unsupported\n", 1231 code, i); 1232 return -ENOTSUPP; 1233 } 1234 } 1235 1236 /* Set end-of-body-code address for exit. */ 1237 addrs[i] = ctx->idx * 4; 1238 1239 return 0; 1240 } 1241