1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Linux Socket Filter - Kernel level socket filtering 4 * 5 * Based on the design of the Berkeley Packet Filter. The new 6 * internal format has been designed by PLUMgrid: 7 * 8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 9 * 10 * Authors: 11 * 12 * Jay Schulist <jschlst@samba.org> 13 * Alexei Starovoitov <ast@plumgrid.com> 14 * Daniel Borkmann <dborkman@redhat.com> 15 * 16 * Andi Kleen - Fix a few bad bugs and races. 17 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 18 */ 19 20 #include <uapi/linux/btf.h> 21 #include <crypto/sha1.h> 22 #include <linux/filter.h> 23 #include <linux/skbuff.h> 24 #include <linux/vmalloc.h> 25 #include <linux/prandom.h> 26 #include <linux/bpf.h> 27 #include <linux/btf.h> 28 #include <linux/hex.h> 29 #include <linux/objtool.h> 30 #include <linux/overflow.h> 31 #include <linux/rbtree_latch.h> 32 #include <linux/kallsyms.h> 33 #include <linux/rcupdate.h> 34 #include <linux/perf_event.h> 35 #include <linux/extable.h> 36 #include <linux/log2.h> 37 #include <linux/bpf_verifier.h> 38 #include <linux/nodemask.h> 39 #include <linux/nospec.h> 40 #include <linux/bpf_mem_alloc.h> 41 #include <linux/memcontrol.h> 42 #include <linux/execmem.h> 43 #include <crypto/sha2.h> 44 45 #include <asm/barrier.h> 46 #include <linux/unaligned.h> 47 48 /* Registers */ 49 #define BPF_R0 regs[BPF_REG_0] 50 #define BPF_R1 regs[BPF_REG_1] 51 #define BPF_R2 regs[BPF_REG_2] 52 #define BPF_R3 regs[BPF_REG_3] 53 #define BPF_R4 regs[BPF_REG_4] 54 #define BPF_R5 regs[BPF_REG_5] 55 #define BPF_R6 regs[BPF_REG_6] 56 #define BPF_R7 regs[BPF_REG_7] 57 #define BPF_R8 regs[BPF_REG_8] 58 #define BPF_R9 regs[BPF_REG_9] 59 #define BPF_R10 regs[BPF_REG_10] 60 61 /* Named registers */ 62 #define DST regs[insn->dst_reg] 63 #define SRC regs[insn->src_reg] 64 #define FP regs[BPF_REG_FP] 65 #define AX regs[BPF_REG_AX] 66 #define ARG1 regs[BPF_REG_ARG1] 67 #define CTX regs[BPF_REG_CTX] 68 #define OFF insn->off 69 #define IMM insn->imm 70 71 struct bpf_mem_alloc bpf_global_ma; 72 bool bpf_global_ma_set; 73 74 /* No hurry in this branch 75 * 76 * Exported for the bpf jit load helper. 77 */ 78 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 79 { 80 u8 *ptr = NULL; 81 82 if (k >= SKF_NET_OFF) { 83 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 84 } else if (k >= SKF_LL_OFF) { 85 if (unlikely(!skb_mac_header_was_set(skb))) 86 return NULL; 87 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 88 } 89 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 90 return ptr; 91 92 return NULL; 93 } 94 95 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 96 enum page_size_enum { 97 __PAGE_SIZE = PAGE_SIZE 98 }; 99 100 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 101 { 102 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 103 struct bpf_prog_aux *aux; 104 struct bpf_prog *fp; 105 106 size = round_up(size, __PAGE_SIZE); 107 fp = __vmalloc(size, gfp_flags); 108 if (fp == NULL) 109 return NULL; 110 111 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 112 if (aux == NULL) { 113 vfree(fp); 114 return NULL; 115 } 116 fp->active = __alloc_percpu_gfp(sizeof(u8[BPF_NR_CONTEXTS]), 4, 117 bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 118 if (!fp->active) { 119 vfree(fp); 120 kfree(aux); 121 return NULL; 122 } 123 124 fp->pages = size / PAGE_SIZE; 125 fp->aux = aux; 126 fp->aux->main_prog_aux = aux; 127 fp->aux->prog = fp; 128 fp->jit_requested = ebpf_jit_enabled(); 129 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 130 #ifdef CONFIG_CGROUP_BPF 131 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 132 #endif 133 134 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 135 #ifdef CONFIG_FINEIBT 136 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 137 #endif 138 mutex_init(&fp->aux->used_maps_mutex); 139 mutex_init(&fp->aux->ext_mutex); 140 mutex_init(&fp->aux->dst_mutex); 141 mutex_init(&fp->aux->st_ops_assoc_mutex); 142 143 #ifdef CONFIG_BPF_SYSCALL 144 bpf_prog_stream_init(fp); 145 #endif 146 147 return fp; 148 } 149 150 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 151 { 152 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 153 struct bpf_prog *prog; 154 int cpu; 155 156 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 157 if (!prog) 158 return NULL; 159 160 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 161 if (!prog->stats) { 162 free_percpu(prog->active); 163 kfree(prog->aux); 164 vfree(prog); 165 return NULL; 166 } 167 168 for_each_possible_cpu(cpu) { 169 struct bpf_prog_stats *pstats; 170 171 pstats = per_cpu_ptr(prog->stats, cpu); 172 u64_stats_init(&pstats->syncp); 173 } 174 return prog; 175 } 176 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 177 178 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 179 { 180 if (!prog->aux->nr_linfo || !prog->jit_requested) 181 return 0; 182 183 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 184 sizeof(*prog->aux->jited_linfo), 185 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 186 if (!prog->aux->jited_linfo) 187 return -ENOMEM; 188 189 return 0; 190 } 191 192 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 193 { 194 if (prog->aux->jited_linfo && 195 (!prog->jited || !prog->aux->jited_linfo[0])) { 196 kvfree(prog->aux->jited_linfo); 197 prog->aux->jited_linfo = NULL; 198 } 199 200 kfree(prog->aux->kfunc_tab); 201 prog->aux->kfunc_tab = NULL; 202 } 203 204 /* The jit engine is responsible to provide an array 205 * for insn_off to the jited_off mapping (insn_to_jit_off). 206 * 207 * The idx to this array is the insn_off. Hence, the insn_off 208 * here is relative to the prog itself instead of the main prog. 209 * This array has one entry for each xlated bpf insn. 210 * 211 * jited_off is the byte off to the end of the jited insn. 212 * 213 * Hence, with 214 * insn_start: 215 * The first bpf insn off of the prog. The insn off 216 * here is relative to the main prog. 217 * e.g. if prog is a subprog, insn_start > 0 218 * linfo_idx: 219 * The prog's idx to prog->aux->linfo and jited_linfo 220 * 221 * jited_linfo[linfo_idx] = prog->bpf_func 222 * 223 * For i > linfo_idx, 224 * 225 * jited_linfo[i] = prog->bpf_func + 226 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 227 */ 228 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 229 const u32 *insn_to_jit_off) 230 { 231 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 232 const struct bpf_line_info *linfo; 233 void **jited_linfo; 234 235 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 236 /* Userspace did not provide linfo */ 237 return; 238 239 linfo_idx = prog->aux->linfo_idx; 240 linfo = &prog->aux->linfo[linfo_idx]; 241 insn_start = linfo[0].insn_off; 242 insn_end = insn_start + prog->len; 243 244 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 245 jited_linfo[0] = prog->bpf_func; 246 247 nr_linfo = prog->aux->nr_linfo - linfo_idx; 248 249 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 250 /* The verifier ensures that linfo[i].insn_off is 251 * strictly increasing 252 */ 253 jited_linfo[i] = prog->bpf_func + 254 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 255 } 256 257 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 258 gfp_t gfp_extra_flags) 259 { 260 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 261 struct bpf_prog *fp; 262 u32 pages; 263 264 size = round_up(size, PAGE_SIZE); 265 pages = size / PAGE_SIZE; 266 if (pages <= fp_old->pages) 267 return fp_old; 268 269 fp = __vmalloc(size, gfp_flags); 270 if (fp) { 271 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 272 fp->pages = pages; 273 fp->aux->prog = fp; 274 275 /* We keep fp->aux from fp_old around in the new 276 * reallocated structure. 277 */ 278 fp_old->aux = NULL; 279 fp_old->stats = NULL; 280 fp_old->active = NULL; 281 __bpf_prog_free(fp_old); 282 } 283 284 return fp; 285 } 286 287 void __bpf_prog_free(struct bpf_prog *fp) 288 { 289 if (fp->aux) { 290 mutex_destroy(&fp->aux->used_maps_mutex); 291 mutex_destroy(&fp->aux->dst_mutex); 292 mutex_destroy(&fp->aux->st_ops_assoc_mutex); 293 kfree(fp->aux->poke_tab); 294 kfree(fp->aux); 295 } 296 free_percpu(fp->stats); 297 free_percpu(fp->active); 298 vfree(fp); 299 } 300 301 int bpf_prog_calc_tag(struct bpf_prog *fp) 302 { 303 size_t size = bpf_prog_insn_size(fp); 304 struct bpf_insn *dst; 305 bool was_ld_map; 306 u32 i; 307 308 dst = vmalloc(size); 309 if (!dst) 310 return -ENOMEM; 311 312 /* We need to take out the map fd for the digest calculation 313 * since they are unstable from user space side. 314 */ 315 for (i = 0, was_ld_map = false; i < fp->len; i++) { 316 dst[i] = fp->insnsi[i]; 317 if (!was_ld_map && 318 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 319 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 320 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 321 was_ld_map = true; 322 dst[i].imm = 0; 323 } else if (was_ld_map && 324 dst[i].code == 0 && 325 dst[i].dst_reg == 0 && 326 dst[i].src_reg == 0 && 327 dst[i].off == 0) { 328 was_ld_map = false; 329 dst[i].imm = 0; 330 } else { 331 was_ld_map = false; 332 } 333 } 334 sha256((u8 *)dst, size, fp->digest); 335 vfree(dst); 336 return 0; 337 } 338 339 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 340 s32 end_new, s32 curr, const bool probe_pass) 341 { 342 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 343 s32 delta = end_new - end_old; 344 s64 imm = insn->imm; 345 346 if (curr < pos && curr + imm + 1 >= end_old) 347 imm += delta; 348 else if (curr >= end_new && curr + imm + 1 < end_new) 349 imm -= delta; 350 if (imm < imm_min || imm > imm_max) 351 return -ERANGE; 352 if (!probe_pass) 353 insn->imm = imm; 354 return 0; 355 } 356 357 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 358 s32 end_new, s32 curr, const bool probe_pass) 359 { 360 s64 off_min, off_max, off; 361 s32 delta = end_new - end_old; 362 363 if (insn->code == (BPF_JMP32 | BPF_JA)) { 364 off = insn->imm; 365 off_min = S32_MIN; 366 off_max = S32_MAX; 367 } else { 368 off = insn->off; 369 off_min = S16_MIN; 370 off_max = S16_MAX; 371 } 372 373 if (curr < pos && curr + off + 1 >= end_old) 374 off += delta; 375 else if (curr >= end_new && curr + off + 1 < end_new) 376 off -= delta; 377 if (off < off_min || off > off_max) 378 return -ERANGE; 379 if (!probe_pass) { 380 if (insn->code == (BPF_JMP32 | BPF_JA)) 381 insn->imm = off; 382 else 383 insn->off = off; 384 } 385 return 0; 386 } 387 388 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 389 s32 end_new, const bool probe_pass) 390 { 391 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 392 struct bpf_insn *insn = prog->insnsi; 393 int ret = 0; 394 395 for (i = 0; i < insn_cnt; i++, insn++) { 396 u8 code; 397 398 /* In the probing pass we still operate on the original, 399 * unpatched image in order to check overflows before we 400 * do any other adjustments. Therefore skip the patchlet. 401 */ 402 if (probe_pass && i == pos) { 403 i = end_new; 404 insn = prog->insnsi + end_old; 405 } 406 if (bpf_pseudo_func(insn)) { 407 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 408 end_new, i, probe_pass); 409 if (ret) 410 return ret; 411 continue; 412 } 413 code = insn->code; 414 if ((BPF_CLASS(code) != BPF_JMP && 415 BPF_CLASS(code) != BPF_JMP32) || 416 BPF_OP(code) == BPF_EXIT) 417 continue; 418 /* Adjust offset of jmps if we cross patch boundaries. */ 419 if (BPF_OP(code) == BPF_CALL) { 420 if (insn->src_reg != BPF_PSEUDO_CALL) 421 continue; 422 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 423 end_new, i, probe_pass); 424 } else { 425 ret = bpf_adj_delta_to_off(insn, pos, end_old, 426 end_new, i, probe_pass); 427 } 428 if (ret) 429 break; 430 } 431 432 return ret; 433 } 434 435 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 436 { 437 struct bpf_line_info *linfo; 438 u32 i, nr_linfo; 439 440 nr_linfo = prog->aux->nr_linfo; 441 if (!nr_linfo || !delta) 442 return; 443 444 linfo = prog->aux->linfo; 445 446 for (i = 0; i < nr_linfo; i++) 447 if (off < linfo[i].insn_off) 448 break; 449 450 /* Push all off < linfo[i].insn_off by delta */ 451 for (; i < nr_linfo; i++) 452 linfo[i].insn_off += delta; 453 } 454 455 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 456 const struct bpf_insn *patch, u32 len) 457 { 458 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 459 const u32 cnt_max = S16_MAX; 460 struct bpf_prog *prog_adj; 461 int err; 462 463 /* Since our patchlet doesn't expand the image, we're done. */ 464 if (insn_delta == 0) { 465 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 466 return prog; 467 } 468 469 insn_adj_cnt = prog->len + insn_delta; 470 471 /* Reject anything that would potentially let the insn->off 472 * target overflow when we have excessive program expansions. 473 * We need to probe here before we do any reallocation where 474 * we afterwards may not fail anymore. 475 */ 476 if (insn_adj_cnt > cnt_max && 477 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 478 return ERR_PTR(err); 479 480 /* Several new instructions need to be inserted. Make room 481 * for them. Likely, there's no need for a new allocation as 482 * last page could have large enough tailroom. 483 */ 484 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 485 GFP_USER); 486 if (!prog_adj) 487 return ERR_PTR(-ENOMEM); 488 489 prog_adj->len = insn_adj_cnt; 490 491 /* Patching happens in 3 steps: 492 * 493 * 1) Move over tail of insnsi from next instruction onwards, 494 * so we can patch the single target insn with one or more 495 * new ones (patching is always from 1 to n insns, n > 0). 496 * 2) Inject new instructions at the target location. 497 * 3) Adjust branch offsets if necessary. 498 */ 499 insn_rest = insn_adj_cnt - off - len; 500 501 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 502 sizeof(*patch) * insn_rest); 503 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 504 505 /* We are guaranteed to not fail at this point, otherwise 506 * the ship has sailed to reverse to the original state. An 507 * overflow cannot happen at this point. 508 */ 509 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 510 511 bpf_adj_linfo(prog_adj, off, insn_delta); 512 513 return prog_adj; 514 } 515 516 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 517 { 518 int err; 519 520 /* Branch offsets can't overflow when program is shrinking, no need 521 * to call bpf_adj_branches(..., true) here 522 */ 523 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 524 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 525 prog->len -= cnt; 526 527 err = bpf_adj_branches(prog, off, off + cnt, off, false); 528 WARN_ON_ONCE(err); 529 return err; 530 } 531 532 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 533 { 534 int i; 535 536 for (i = 0; i < fp->aux->real_func_cnt; i++) 537 bpf_prog_kallsyms_del(fp->aux->func[i]); 538 } 539 540 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 541 { 542 bpf_prog_kallsyms_del_subprogs(fp); 543 bpf_prog_kallsyms_del(fp); 544 } 545 546 #ifdef CONFIG_BPF_JIT 547 /* All BPF JIT sysctl knobs here. */ 548 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 549 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 550 int bpf_jit_harden __read_mostly; 551 long bpf_jit_limit __read_mostly; 552 long bpf_jit_limit_max __read_mostly; 553 554 static void 555 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 556 { 557 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 558 559 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 560 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 561 } 562 563 static void 564 bpf_prog_ksym_set_name(struct bpf_prog *prog) 565 { 566 char *sym = prog->aux->ksym.name; 567 const char *end = sym + KSYM_NAME_LEN; 568 const struct btf_type *type; 569 const char *func_name; 570 571 BUILD_BUG_ON(sizeof("bpf_prog_") + 572 sizeof(prog->tag) * 2 + 573 /* name has been null terminated. 574 * We should need +1 for the '_' preceding 575 * the name. However, the null character 576 * is double counted between the name and the 577 * sizeof("bpf_prog_") above, so we omit 578 * the +1 here. 579 */ 580 sizeof(prog->aux->name) > KSYM_NAME_LEN); 581 582 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 583 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 584 585 /* prog->aux->name will be ignored if full btf name is available */ 586 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 587 type = btf_type_by_id(prog->aux->btf, 588 prog->aux->func_info[prog->aux->func_idx].type_id); 589 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 590 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 591 return; 592 } 593 594 if (prog->aux->name[0]) 595 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 596 else 597 *sym = 0; 598 } 599 600 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 601 { 602 return container_of(n, struct bpf_ksym, tnode)->start; 603 } 604 605 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 606 struct latch_tree_node *b) 607 { 608 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 609 } 610 611 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 612 { 613 unsigned long val = (unsigned long)key; 614 const struct bpf_ksym *ksym; 615 616 ksym = container_of(n, struct bpf_ksym, tnode); 617 618 if (val < ksym->start) 619 return -1; 620 /* Ensure that we detect return addresses as part of the program, when 621 * the final instruction is a call for a program part of the stack 622 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 623 */ 624 if (val > ksym->end) 625 return 1; 626 627 return 0; 628 } 629 630 static const struct latch_tree_ops bpf_tree_ops = { 631 .less = bpf_tree_less, 632 .comp = bpf_tree_comp, 633 }; 634 635 static DEFINE_SPINLOCK(bpf_lock); 636 static LIST_HEAD(bpf_kallsyms); 637 static struct latch_tree_root bpf_tree __cacheline_aligned; 638 639 void bpf_ksym_add(struct bpf_ksym *ksym) 640 { 641 spin_lock_bh(&bpf_lock); 642 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 643 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 644 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 645 spin_unlock_bh(&bpf_lock); 646 } 647 648 static void __bpf_ksym_del(struct bpf_ksym *ksym) 649 { 650 if (list_empty(&ksym->lnode)) 651 return; 652 653 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 654 list_del_rcu(&ksym->lnode); 655 } 656 657 void bpf_ksym_del(struct bpf_ksym *ksym) 658 { 659 spin_lock_bh(&bpf_lock); 660 __bpf_ksym_del(ksym); 661 spin_unlock_bh(&bpf_lock); 662 } 663 664 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 665 { 666 return fp->jited && !bpf_prog_was_classic(fp); 667 } 668 669 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 670 { 671 if (!bpf_prog_kallsyms_candidate(fp) || 672 !bpf_token_capable(fp->aux->token, CAP_BPF)) 673 return; 674 675 bpf_prog_ksym_set_addr(fp); 676 bpf_prog_ksym_set_name(fp); 677 fp->aux->ksym.prog = true; 678 679 bpf_ksym_add(&fp->aux->ksym); 680 681 #ifdef CONFIG_FINEIBT 682 /* 683 * When FineIBT, code in the __cfi_foo() symbols can get executed 684 * and hence unwinder needs help. 685 */ 686 if (cfi_mode != CFI_FINEIBT) 687 return; 688 689 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 690 "__cfi_%s", fp->aux->ksym.name); 691 692 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 693 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 694 695 bpf_ksym_add(&fp->aux->ksym_prefix); 696 #endif 697 } 698 699 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 700 { 701 if (!bpf_prog_kallsyms_candidate(fp)) 702 return; 703 704 bpf_ksym_del(&fp->aux->ksym); 705 #ifdef CONFIG_FINEIBT 706 if (cfi_mode != CFI_FINEIBT) 707 return; 708 bpf_ksym_del(&fp->aux->ksym_prefix); 709 #endif 710 } 711 712 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 713 { 714 struct latch_tree_node *n; 715 716 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 717 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 718 } 719 720 int bpf_address_lookup(unsigned long addr, unsigned long *size, 721 unsigned long *off, char *sym) 722 { 723 struct bpf_ksym *ksym; 724 int ret = 0; 725 726 rcu_read_lock(); 727 ksym = bpf_ksym_find(addr); 728 if (ksym) { 729 unsigned long symbol_start = ksym->start; 730 unsigned long symbol_end = ksym->end; 731 732 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 733 734 if (size) 735 *size = symbol_end - symbol_start; 736 if (off) 737 *off = addr - symbol_start; 738 } 739 rcu_read_unlock(); 740 741 return ret; 742 } 743 744 bool is_bpf_text_address(unsigned long addr) 745 { 746 bool ret; 747 748 rcu_read_lock(); 749 ret = bpf_ksym_find(addr) != NULL; 750 rcu_read_unlock(); 751 752 return ret; 753 } 754 755 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 756 { 757 struct bpf_ksym *ksym; 758 759 WARN_ON_ONCE(!rcu_read_lock_held()); 760 ksym = bpf_ksym_find(addr); 761 762 return ksym && ksym->prog ? 763 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 764 NULL; 765 } 766 767 bool bpf_has_frame_pointer(unsigned long ip) 768 { 769 struct bpf_ksym *ksym; 770 unsigned long offset; 771 772 guard(rcu)(); 773 774 ksym = bpf_ksym_find(ip); 775 if (!ksym || !ksym->fp_start || !ksym->fp_end) 776 return false; 777 778 offset = ip - ksym->start; 779 780 return offset >= ksym->fp_start && offset < ksym->fp_end; 781 } 782 783 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 784 { 785 const struct exception_table_entry *e = NULL; 786 struct bpf_prog *prog; 787 788 rcu_read_lock(); 789 prog = bpf_prog_ksym_find(addr); 790 if (!prog) 791 goto out; 792 if (!prog->aux->num_exentries) 793 goto out; 794 795 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 796 out: 797 rcu_read_unlock(); 798 return e; 799 } 800 801 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 802 char *sym) 803 { 804 struct bpf_ksym *ksym; 805 unsigned int it = 0; 806 int ret = -ERANGE; 807 808 if (!bpf_jit_kallsyms_enabled()) 809 return ret; 810 811 rcu_read_lock(); 812 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 813 if (it++ != symnum) 814 continue; 815 816 strscpy(sym, ksym->name, KSYM_NAME_LEN); 817 818 *value = ksym->start; 819 *type = BPF_SYM_ELF_TYPE; 820 821 ret = 0; 822 break; 823 } 824 rcu_read_unlock(); 825 826 return ret; 827 } 828 829 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 830 struct bpf_jit_poke_descriptor *poke) 831 { 832 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 833 static const u32 poke_tab_max = 1024; 834 u32 slot = prog->aux->size_poke_tab; 835 u32 size = slot + 1; 836 837 if (size > poke_tab_max) 838 return -ENOSPC; 839 if (poke->tailcall_target || poke->tailcall_target_stable || 840 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 841 return -EINVAL; 842 843 switch (poke->reason) { 844 case BPF_POKE_REASON_TAIL_CALL: 845 if (!poke->tail_call.map) 846 return -EINVAL; 847 break; 848 default: 849 return -EINVAL; 850 } 851 852 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 853 if (!tab) 854 return -ENOMEM; 855 856 memcpy(&tab[slot], poke, sizeof(*poke)); 857 prog->aux->size_poke_tab = size; 858 prog->aux->poke_tab = tab; 859 860 return slot; 861 } 862 863 /* 864 * BPF program pack allocator. 865 * 866 * Most BPF programs are pretty small. Allocating a hole page for each 867 * program is sometime a waste. Many small bpf program also adds pressure 868 * to instruction TLB. To solve this issue, we introduce a BPF program pack 869 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 870 * to host BPF programs. 871 */ 872 #define BPF_PROG_CHUNK_SHIFT 6 873 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 874 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 875 876 struct bpf_prog_pack { 877 struct list_head list; 878 void *ptr; 879 unsigned long bitmap[]; 880 }; 881 882 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 883 { 884 memset(area, 0, size); 885 } 886 887 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 888 889 static DEFINE_MUTEX(pack_mutex); 890 static LIST_HEAD(pack_list); 891 892 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 893 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 894 */ 895 #ifdef PMD_SIZE 896 /* PMD_SIZE is really big for some archs. It doesn't make sense to 897 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 898 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 899 * greater than or equal to 2MB. 900 */ 901 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 902 #else 903 #define BPF_PROG_PACK_SIZE PAGE_SIZE 904 #endif 905 906 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 907 908 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 909 { 910 struct bpf_prog_pack *pack; 911 int err; 912 913 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 914 GFP_KERNEL); 915 if (!pack) 916 return NULL; 917 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 918 if (!pack->ptr) 919 goto out; 920 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 921 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 922 923 set_vm_flush_reset_perms(pack->ptr); 924 err = set_memory_rox((unsigned long)pack->ptr, 925 BPF_PROG_PACK_SIZE / PAGE_SIZE); 926 if (err) 927 goto out; 928 list_add_tail(&pack->list, &pack_list); 929 return pack; 930 931 out: 932 bpf_jit_free_exec(pack->ptr); 933 kfree(pack); 934 return NULL; 935 } 936 937 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 938 { 939 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 940 struct bpf_prog_pack *pack; 941 unsigned long pos; 942 void *ptr = NULL; 943 944 mutex_lock(&pack_mutex); 945 if (size > BPF_PROG_PACK_SIZE) { 946 size = round_up(size, PAGE_SIZE); 947 ptr = bpf_jit_alloc_exec(size); 948 if (ptr) { 949 int err; 950 951 bpf_fill_ill_insns(ptr, size); 952 set_vm_flush_reset_perms(ptr); 953 err = set_memory_rox((unsigned long)ptr, 954 size / PAGE_SIZE); 955 if (err) { 956 bpf_jit_free_exec(ptr); 957 ptr = NULL; 958 } 959 } 960 goto out; 961 } 962 list_for_each_entry(pack, &pack_list, list) { 963 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 964 nbits, 0); 965 if (pos < BPF_PROG_CHUNK_COUNT) 966 goto found_free_area; 967 } 968 969 pack = alloc_new_pack(bpf_fill_ill_insns); 970 if (!pack) 971 goto out; 972 973 pos = 0; 974 975 found_free_area: 976 bitmap_set(pack->bitmap, pos, nbits); 977 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 978 979 out: 980 mutex_unlock(&pack_mutex); 981 return ptr; 982 } 983 984 void bpf_prog_pack_free(void *ptr, u32 size) 985 { 986 struct bpf_prog_pack *pack = NULL, *tmp; 987 unsigned int nbits; 988 unsigned long pos; 989 990 mutex_lock(&pack_mutex); 991 if (size > BPF_PROG_PACK_SIZE) { 992 bpf_jit_free_exec(ptr); 993 goto out; 994 } 995 996 list_for_each_entry(tmp, &pack_list, list) { 997 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 998 pack = tmp; 999 break; 1000 } 1001 } 1002 1003 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1004 goto out; 1005 1006 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1007 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1008 1009 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1010 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1011 1012 bitmap_clear(pack->bitmap, pos, nbits); 1013 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1014 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1015 list_del(&pack->list); 1016 bpf_jit_free_exec(pack->ptr); 1017 kfree(pack); 1018 } 1019 out: 1020 mutex_unlock(&pack_mutex); 1021 } 1022 1023 static atomic_long_t bpf_jit_current; 1024 1025 /* Can be overridden by an arch's JIT compiler if it has a custom, 1026 * dedicated BPF backend memory area, or if neither of the two 1027 * below apply. 1028 */ 1029 u64 __weak bpf_jit_alloc_exec_limit(void) 1030 { 1031 #if defined(MODULES_VADDR) 1032 return MODULES_END - MODULES_VADDR; 1033 #else 1034 return VMALLOC_END - VMALLOC_START; 1035 #endif 1036 } 1037 1038 static int __init bpf_jit_charge_init(void) 1039 { 1040 /* Only used as heuristic here to derive limit. */ 1041 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1042 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1043 PAGE_SIZE), LONG_MAX); 1044 return 0; 1045 } 1046 pure_initcall(bpf_jit_charge_init); 1047 1048 int bpf_jit_charge_modmem(u32 size) 1049 { 1050 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1051 if (!bpf_capable()) { 1052 atomic_long_sub(size, &bpf_jit_current); 1053 return -EPERM; 1054 } 1055 } 1056 1057 return 0; 1058 } 1059 1060 void bpf_jit_uncharge_modmem(u32 size) 1061 { 1062 atomic_long_sub(size, &bpf_jit_current); 1063 } 1064 1065 void *__weak bpf_jit_alloc_exec(unsigned long size) 1066 { 1067 return execmem_alloc(EXECMEM_BPF, size); 1068 } 1069 1070 void __weak bpf_jit_free_exec(void *addr) 1071 { 1072 execmem_free(addr); 1073 } 1074 1075 struct bpf_binary_header * 1076 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1077 unsigned int alignment, 1078 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1079 { 1080 struct bpf_binary_header *hdr; 1081 u32 size, hole, start; 1082 1083 WARN_ON_ONCE(!is_power_of_2(alignment) || 1084 alignment > BPF_IMAGE_ALIGNMENT); 1085 1086 /* Most of BPF filters are really small, but if some of them 1087 * fill a page, allow at least 128 extra bytes to insert a 1088 * random section of illegal instructions. 1089 */ 1090 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1091 1092 if (bpf_jit_charge_modmem(size)) 1093 return NULL; 1094 hdr = bpf_jit_alloc_exec(size); 1095 if (!hdr) { 1096 bpf_jit_uncharge_modmem(size); 1097 return NULL; 1098 } 1099 1100 /* Fill space with illegal/arch-dep instructions. */ 1101 bpf_fill_ill_insns(hdr, size); 1102 1103 hdr->size = size; 1104 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1105 PAGE_SIZE - sizeof(*hdr)); 1106 start = get_random_u32_below(hole) & ~(alignment - 1); 1107 1108 /* Leave a random number of instructions before BPF code. */ 1109 *image_ptr = &hdr->image[start]; 1110 1111 return hdr; 1112 } 1113 1114 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1115 { 1116 u32 size = hdr->size; 1117 1118 bpf_jit_free_exec(hdr); 1119 bpf_jit_uncharge_modmem(size); 1120 } 1121 1122 /* Allocate jit binary from bpf_prog_pack allocator. 1123 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1124 * to the memory. To solve this problem, a RW buffer is also allocated at 1125 * as the same time. The JIT engine should calculate offsets based on the 1126 * RO memory address, but write JITed program to the RW buffer. Once the 1127 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1128 * the JITed program to the RO memory. 1129 */ 1130 struct bpf_binary_header * 1131 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1132 unsigned int alignment, 1133 struct bpf_binary_header **rw_header, 1134 u8 **rw_image, 1135 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1136 { 1137 struct bpf_binary_header *ro_header; 1138 u32 size, hole, start; 1139 1140 WARN_ON_ONCE(!is_power_of_2(alignment) || 1141 alignment > BPF_IMAGE_ALIGNMENT); 1142 1143 /* add 16 bytes for a random section of illegal instructions */ 1144 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1145 1146 if (bpf_jit_charge_modmem(size)) 1147 return NULL; 1148 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1149 if (!ro_header) { 1150 bpf_jit_uncharge_modmem(size); 1151 return NULL; 1152 } 1153 1154 *rw_header = kvmalloc(size, GFP_KERNEL); 1155 if (!*rw_header) { 1156 bpf_prog_pack_free(ro_header, size); 1157 bpf_jit_uncharge_modmem(size); 1158 return NULL; 1159 } 1160 1161 /* Fill space with illegal/arch-dep instructions. */ 1162 bpf_fill_ill_insns(*rw_header, size); 1163 (*rw_header)->size = size; 1164 1165 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1166 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1167 start = get_random_u32_below(hole) & ~(alignment - 1); 1168 1169 *image_ptr = &ro_header->image[start]; 1170 *rw_image = &(*rw_header)->image[start]; 1171 1172 return ro_header; 1173 } 1174 1175 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1176 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1177 struct bpf_binary_header *rw_header) 1178 { 1179 void *ptr; 1180 1181 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1182 1183 kvfree(rw_header); 1184 1185 if (IS_ERR(ptr)) { 1186 bpf_prog_pack_free(ro_header, ro_header->size); 1187 return PTR_ERR(ptr); 1188 } 1189 return 0; 1190 } 1191 1192 /* bpf_jit_binary_pack_free is called in two different scenarios: 1193 * 1) when the program is freed after; 1194 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1195 * For case 2), we need to free both the RO memory and the RW buffer. 1196 * 1197 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1198 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1199 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1200 * bpf_arch_text_copy (when jit fails). 1201 */ 1202 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1203 struct bpf_binary_header *rw_header) 1204 { 1205 u32 size = ro_header->size; 1206 1207 bpf_prog_pack_free(ro_header, size); 1208 kvfree(rw_header); 1209 bpf_jit_uncharge_modmem(size); 1210 } 1211 1212 struct bpf_binary_header * 1213 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1214 { 1215 unsigned long real_start = (unsigned long)fp->bpf_func; 1216 unsigned long addr; 1217 1218 addr = real_start & BPF_PROG_CHUNK_MASK; 1219 return (void *)addr; 1220 } 1221 1222 static inline struct bpf_binary_header * 1223 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1224 { 1225 unsigned long real_start = (unsigned long)fp->bpf_func; 1226 unsigned long addr; 1227 1228 addr = real_start & PAGE_MASK; 1229 return (void *)addr; 1230 } 1231 1232 /* This symbol is only overridden by archs that have different 1233 * requirements than the usual eBPF JITs, f.e. when they only 1234 * implement cBPF JIT, do not set images read-only, etc. 1235 */ 1236 void __weak bpf_jit_free(struct bpf_prog *fp) 1237 { 1238 if (fp->jited) { 1239 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1240 1241 bpf_jit_binary_free(hdr); 1242 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1243 } 1244 1245 bpf_prog_unlock_free(fp); 1246 } 1247 1248 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1249 const struct bpf_insn *insn, bool extra_pass, 1250 u64 *func_addr, bool *func_addr_fixed) 1251 { 1252 s16 off = insn->off; 1253 s32 imm = insn->imm; 1254 u8 *addr; 1255 int err; 1256 1257 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1258 if (!*func_addr_fixed) { 1259 /* Place-holder address till the last pass has collected 1260 * all addresses for JITed subprograms in which case we 1261 * can pick them up from prog->aux. 1262 */ 1263 if (!extra_pass) 1264 addr = NULL; 1265 else if (prog->aux->func && 1266 off >= 0 && off < prog->aux->real_func_cnt) 1267 addr = (u8 *)prog->aux->func[off]->bpf_func; 1268 else 1269 return -EINVAL; 1270 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1271 bpf_jit_supports_far_kfunc_call()) { 1272 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1273 if (err) 1274 return err; 1275 } else { 1276 /* Address of a BPF helper call. Since part of the core 1277 * kernel, it's always at a fixed location. __bpf_call_base 1278 * and the helper with imm relative to it are both in core 1279 * kernel. 1280 */ 1281 addr = (u8 *)__bpf_call_base + imm; 1282 } 1283 1284 *func_addr = (unsigned long)addr; 1285 return 0; 1286 } 1287 1288 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1289 { 1290 if (prog->aux->ksym.prog) 1291 return prog->aux->ksym.name; 1292 return prog->aux->name; 1293 } 1294 1295 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1296 const struct bpf_insn *aux, 1297 struct bpf_insn *to_buff, 1298 bool emit_zext) 1299 { 1300 struct bpf_insn *to = to_buff; 1301 u32 imm_rnd = get_random_u32(); 1302 s16 off; 1303 1304 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1305 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1306 1307 /* Constraints on AX register: 1308 * 1309 * AX register is inaccessible from user space. It is mapped in 1310 * all JITs, and used here for constant blinding rewrites. It is 1311 * typically "stateless" meaning its contents are only valid within 1312 * the executed instruction, but not across several instructions. 1313 * There are a few exceptions however which are further detailed 1314 * below. 1315 * 1316 * Constant blinding is only used by JITs, not in the interpreter. 1317 * The interpreter uses AX in some occasions as a local temporary 1318 * register e.g. in DIV or MOD instructions. 1319 * 1320 * In restricted circumstances, the verifier can also use the AX 1321 * register for rewrites as long as they do not interfere with 1322 * the above cases! 1323 */ 1324 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1325 goto out; 1326 1327 if (from->imm == 0 && 1328 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1329 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1330 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1331 goto out; 1332 } 1333 1334 switch (from->code) { 1335 case BPF_ALU | BPF_ADD | BPF_K: 1336 case BPF_ALU | BPF_SUB | BPF_K: 1337 case BPF_ALU | BPF_AND | BPF_K: 1338 case BPF_ALU | BPF_OR | BPF_K: 1339 case BPF_ALU | BPF_XOR | BPF_K: 1340 case BPF_ALU | BPF_MUL | BPF_K: 1341 case BPF_ALU | BPF_MOV | BPF_K: 1342 case BPF_ALU | BPF_DIV | BPF_K: 1343 case BPF_ALU | BPF_MOD | BPF_K: 1344 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1345 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1346 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1347 break; 1348 1349 case BPF_ALU64 | BPF_ADD | BPF_K: 1350 case BPF_ALU64 | BPF_SUB | BPF_K: 1351 case BPF_ALU64 | BPF_AND | BPF_K: 1352 case BPF_ALU64 | BPF_OR | BPF_K: 1353 case BPF_ALU64 | BPF_XOR | BPF_K: 1354 case BPF_ALU64 | BPF_MUL | BPF_K: 1355 case BPF_ALU64 | BPF_MOV | BPF_K: 1356 case BPF_ALU64 | BPF_DIV | BPF_K: 1357 case BPF_ALU64 | BPF_MOD | BPF_K: 1358 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1359 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1360 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1361 break; 1362 1363 case BPF_JMP | BPF_JEQ | BPF_K: 1364 case BPF_JMP | BPF_JNE | BPF_K: 1365 case BPF_JMP | BPF_JGT | BPF_K: 1366 case BPF_JMP | BPF_JLT | BPF_K: 1367 case BPF_JMP | BPF_JGE | BPF_K: 1368 case BPF_JMP | BPF_JLE | BPF_K: 1369 case BPF_JMP | BPF_JSGT | BPF_K: 1370 case BPF_JMP | BPF_JSLT | BPF_K: 1371 case BPF_JMP | BPF_JSGE | BPF_K: 1372 case BPF_JMP | BPF_JSLE | BPF_K: 1373 case BPF_JMP | BPF_JSET | BPF_K: 1374 /* Accommodate for extra offset in case of a backjump. */ 1375 off = from->off; 1376 if (off < 0) 1377 off -= 2; 1378 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1379 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1380 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1381 break; 1382 1383 case BPF_JMP32 | BPF_JEQ | BPF_K: 1384 case BPF_JMP32 | BPF_JNE | BPF_K: 1385 case BPF_JMP32 | BPF_JGT | BPF_K: 1386 case BPF_JMP32 | BPF_JLT | BPF_K: 1387 case BPF_JMP32 | BPF_JGE | BPF_K: 1388 case BPF_JMP32 | BPF_JLE | BPF_K: 1389 case BPF_JMP32 | BPF_JSGT | BPF_K: 1390 case BPF_JMP32 | BPF_JSLT | BPF_K: 1391 case BPF_JMP32 | BPF_JSGE | BPF_K: 1392 case BPF_JMP32 | BPF_JSLE | BPF_K: 1393 case BPF_JMP32 | BPF_JSET | BPF_K: 1394 /* Accommodate for extra offset in case of a backjump. */ 1395 off = from->off; 1396 if (off < 0) 1397 off -= 2; 1398 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1399 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1400 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1401 off); 1402 break; 1403 1404 case BPF_LD | BPF_IMM | BPF_DW: 1405 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1406 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1407 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1408 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1409 break; 1410 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1411 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1412 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1413 if (emit_zext) 1414 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1415 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1416 break; 1417 1418 case BPF_ST | BPF_MEM | BPF_DW: 1419 case BPF_ST | BPF_MEM | BPF_W: 1420 case BPF_ST | BPF_MEM | BPF_H: 1421 case BPF_ST | BPF_MEM | BPF_B: 1422 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1423 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1424 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1425 break; 1426 } 1427 out: 1428 return to - to_buff; 1429 } 1430 1431 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1432 gfp_t gfp_extra_flags) 1433 { 1434 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1435 struct bpf_prog *fp; 1436 1437 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1438 if (fp != NULL) { 1439 /* aux->prog still points to the fp_other one, so 1440 * when promoting the clone to the real program, 1441 * this still needs to be adapted. 1442 */ 1443 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1444 } 1445 1446 return fp; 1447 } 1448 1449 static void bpf_prog_clone_free(struct bpf_prog *fp) 1450 { 1451 /* aux was stolen by the other clone, so we cannot free 1452 * it from this path! It will be freed eventually by the 1453 * other program on release. 1454 * 1455 * At this point, we don't need a deferred release since 1456 * clone is guaranteed to not be locked. 1457 */ 1458 fp->aux = NULL; 1459 fp->stats = NULL; 1460 fp->active = NULL; 1461 __bpf_prog_free(fp); 1462 } 1463 1464 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1465 { 1466 /* We have to repoint aux->prog to self, as we don't 1467 * know whether fp here is the clone or the original. 1468 */ 1469 fp->aux->prog = fp; 1470 bpf_prog_clone_free(fp_other); 1471 } 1472 1473 static void adjust_insn_arrays(struct bpf_prog *prog, u32 off, u32 len) 1474 { 1475 #ifdef CONFIG_BPF_SYSCALL 1476 struct bpf_map *map; 1477 int i; 1478 1479 if (len <= 1) 1480 return; 1481 1482 for (i = 0; i < prog->aux->used_map_cnt; i++) { 1483 map = prog->aux->used_maps[i]; 1484 if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY) 1485 bpf_insn_array_adjust(map, off, len); 1486 } 1487 #endif 1488 } 1489 1490 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1491 { 1492 struct bpf_insn insn_buff[16], aux[2]; 1493 struct bpf_prog *clone, *tmp; 1494 int insn_delta, insn_cnt; 1495 struct bpf_insn *insn; 1496 int i, rewritten; 1497 1498 if (!prog->blinding_requested || prog->blinded) 1499 return prog; 1500 1501 clone = bpf_prog_clone_create(prog, GFP_USER); 1502 if (!clone) 1503 return ERR_PTR(-ENOMEM); 1504 1505 insn_cnt = clone->len; 1506 insn = clone->insnsi; 1507 1508 for (i = 0; i < insn_cnt; i++, insn++) { 1509 if (bpf_pseudo_func(insn)) { 1510 /* ld_imm64 with an address of bpf subprog is not 1511 * a user controlled constant. Don't randomize it, 1512 * since it will conflict with jit_subprogs() logic. 1513 */ 1514 insn++; 1515 i++; 1516 continue; 1517 } 1518 1519 /* We temporarily need to hold the original ld64 insn 1520 * so that we can still access the first part in the 1521 * second blinding run. 1522 */ 1523 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1524 insn[1].code == 0) 1525 memcpy(aux, insn, sizeof(aux)); 1526 1527 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1528 clone->aux->verifier_zext); 1529 if (!rewritten) 1530 continue; 1531 1532 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1533 if (IS_ERR(tmp)) { 1534 /* Patching may have repointed aux->prog during 1535 * realloc from the original one, so we need to 1536 * fix it up here on error. 1537 */ 1538 bpf_jit_prog_release_other(prog, clone); 1539 return tmp; 1540 } 1541 1542 clone = tmp; 1543 insn_delta = rewritten - 1; 1544 1545 /* Instructions arrays must be updated using absolute xlated offsets */ 1546 adjust_insn_arrays(clone, prog->aux->subprog_start + i, rewritten); 1547 1548 /* Walk new program and skip insns we just inserted. */ 1549 insn = clone->insnsi + i + insn_delta; 1550 insn_cnt += insn_delta; 1551 i += insn_delta; 1552 } 1553 1554 clone->blinded = 1; 1555 return clone; 1556 } 1557 #endif /* CONFIG_BPF_JIT */ 1558 1559 /* Base function for offset calculation. Needs to go into .text section, 1560 * therefore keeping it non-static as well; will also be used by JITs 1561 * anyway later on, so do not let the compiler omit it. This also needs 1562 * to go into kallsyms for correlation from e.g. bpftool, so naming 1563 * must not change. 1564 */ 1565 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1566 { 1567 return 0; 1568 } 1569 EXPORT_SYMBOL_GPL(__bpf_call_base); 1570 1571 /* All UAPI available opcodes. */ 1572 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1573 /* 32 bit ALU operations. */ \ 1574 /* Register based. */ \ 1575 INSN_3(ALU, ADD, X), \ 1576 INSN_3(ALU, SUB, X), \ 1577 INSN_3(ALU, AND, X), \ 1578 INSN_3(ALU, OR, X), \ 1579 INSN_3(ALU, LSH, X), \ 1580 INSN_3(ALU, RSH, X), \ 1581 INSN_3(ALU, XOR, X), \ 1582 INSN_3(ALU, MUL, X), \ 1583 INSN_3(ALU, MOV, X), \ 1584 INSN_3(ALU, ARSH, X), \ 1585 INSN_3(ALU, DIV, X), \ 1586 INSN_3(ALU, MOD, X), \ 1587 INSN_2(ALU, NEG), \ 1588 INSN_3(ALU, END, TO_BE), \ 1589 INSN_3(ALU, END, TO_LE), \ 1590 /* Immediate based. */ \ 1591 INSN_3(ALU, ADD, K), \ 1592 INSN_3(ALU, SUB, K), \ 1593 INSN_3(ALU, AND, K), \ 1594 INSN_3(ALU, OR, K), \ 1595 INSN_3(ALU, LSH, K), \ 1596 INSN_3(ALU, RSH, K), \ 1597 INSN_3(ALU, XOR, K), \ 1598 INSN_3(ALU, MUL, K), \ 1599 INSN_3(ALU, MOV, K), \ 1600 INSN_3(ALU, ARSH, K), \ 1601 INSN_3(ALU, DIV, K), \ 1602 INSN_3(ALU, MOD, K), \ 1603 /* 64 bit ALU operations. */ \ 1604 /* Register based. */ \ 1605 INSN_3(ALU64, ADD, X), \ 1606 INSN_3(ALU64, SUB, X), \ 1607 INSN_3(ALU64, AND, X), \ 1608 INSN_3(ALU64, OR, X), \ 1609 INSN_3(ALU64, LSH, X), \ 1610 INSN_3(ALU64, RSH, X), \ 1611 INSN_3(ALU64, XOR, X), \ 1612 INSN_3(ALU64, MUL, X), \ 1613 INSN_3(ALU64, MOV, X), \ 1614 INSN_3(ALU64, ARSH, X), \ 1615 INSN_3(ALU64, DIV, X), \ 1616 INSN_3(ALU64, MOD, X), \ 1617 INSN_2(ALU64, NEG), \ 1618 INSN_3(ALU64, END, TO_LE), \ 1619 /* Immediate based. */ \ 1620 INSN_3(ALU64, ADD, K), \ 1621 INSN_3(ALU64, SUB, K), \ 1622 INSN_3(ALU64, AND, K), \ 1623 INSN_3(ALU64, OR, K), \ 1624 INSN_3(ALU64, LSH, K), \ 1625 INSN_3(ALU64, RSH, K), \ 1626 INSN_3(ALU64, XOR, K), \ 1627 INSN_3(ALU64, MUL, K), \ 1628 INSN_3(ALU64, MOV, K), \ 1629 INSN_3(ALU64, ARSH, K), \ 1630 INSN_3(ALU64, DIV, K), \ 1631 INSN_3(ALU64, MOD, K), \ 1632 /* Call instruction. */ \ 1633 INSN_2(JMP, CALL), \ 1634 /* Exit instruction. */ \ 1635 INSN_2(JMP, EXIT), \ 1636 /* 32-bit Jump instructions. */ \ 1637 /* Register based. */ \ 1638 INSN_3(JMP32, JEQ, X), \ 1639 INSN_3(JMP32, JNE, X), \ 1640 INSN_3(JMP32, JGT, X), \ 1641 INSN_3(JMP32, JLT, X), \ 1642 INSN_3(JMP32, JGE, X), \ 1643 INSN_3(JMP32, JLE, X), \ 1644 INSN_3(JMP32, JSGT, X), \ 1645 INSN_3(JMP32, JSLT, X), \ 1646 INSN_3(JMP32, JSGE, X), \ 1647 INSN_3(JMP32, JSLE, X), \ 1648 INSN_3(JMP32, JSET, X), \ 1649 /* Immediate based. */ \ 1650 INSN_3(JMP32, JEQ, K), \ 1651 INSN_3(JMP32, JNE, K), \ 1652 INSN_3(JMP32, JGT, K), \ 1653 INSN_3(JMP32, JLT, K), \ 1654 INSN_3(JMP32, JGE, K), \ 1655 INSN_3(JMP32, JLE, K), \ 1656 INSN_3(JMP32, JSGT, K), \ 1657 INSN_3(JMP32, JSLT, K), \ 1658 INSN_3(JMP32, JSGE, K), \ 1659 INSN_3(JMP32, JSLE, K), \ 1660 INSN_3(JMP32, JSET, K), \ 1661 /* Jump instructions. */ \ 1662 /* Register based. */ \ 1663 INSN_3(JMP, JEQ, X), \ 1664 INSN_3(JMP, JNE, X), \ 1665 INSN_3(JMP, JGT, X), \ 1666 INSN_3(JMP, JLT, X), \ 1667 INSN_3(JMP, JGE, X), \ 1668 INSN_3(JMP, JLE, X), \ 1669 INSN_3(JMP, JSGT, X), \ 1670 INSN_3(JMP, JSLT, X), \ 1671 INSN_3(JMP, JSGE, X), \ 1672 INSN_3(JMP, JSLE, X), \ 1673 INSN_3(JMP, JSET, X), \ 1674 /* Immediate based. */ \ 1675 INSN_3(JMP, JEQ, K), \ 1676 INSN_3(JMP, JNE, K), \ 1677 INSN_3(JMP, JGT, K), \ 1678 INSN_3(JMP, JLT, K), \ 1679 INSN_3(JMP, JGE, K), \ 1680 INSN_3(JMP, JLE, K), \ 1681 INSN_3(JMP, JSGT, K), \ 1682 INSN_3(JMP, JSLT, K), \ 1683 INSN_3(JMP, JSGE, K), \ 1684 INSN_3(JMP, JSLE, K), \ 1685 INSN_3(JMP, JSET, K), \ 1686 INSN_2(JMP, JA), \ 1687 INSN_2(JMP32, JA), \ 1688 /* Atomic operations. */ \ 1689 INSN_3(STX, ATOMIC, B), \ 1690 INSN_3(STX, ATOMIC, H), \ 1691 INSN_3(STX, ATOMIC, W), \ 1692 INSN_3(STX, ATOMIC, DW), \ 1693 /* Store instructions. */ \ 1694 /* Register based. */ \ 1695 INSN_3(STX, MEM, B), \ 1696 INSN_3(STX, MEM, H), \ 1697 INSN_3(STX, MEM, W), \ 1698 INSN_3(STX, MEM, DW), \ 1699 /* Immediate based. */ \ 1700 INSN_3(ST, MEM, B), \ 1701 INSN_3(ST, MEM, H), \ 1702 INSN_3(ST, MEM, W), \ 1703 INSN_3(ST, MEM, DW), \ 1704 /* Load instructions. */ \ 1705 /* Register based. */ \ 1706 INSN_3(LDX, MEM, B), \ 1707 INSN_3(LDX, MEM, H), \ 1708 INSN_3(LDX, MEM, W), \ 1709 INSN_3(LDX, MEM, DW), \ 1710 INSN_3(LDX, MEMSX, B), \ 1711 INSN_3(LDX, MEMSX, H), \ 1712 INSN_3(LDX, MEMSX, W), \ 1713 /* Immediate based. */ \ 1714 INSN_3(LD, IMM, DW) 1715 1716 bool bpf_opcode_in_insntable(u8 code) 1717 { 1718 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1719 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1720 static const bool public_insntable[256] = { 1721 [0 ... 255] = false, 1722 /* Now overwrite non-defaults ... */ 1723 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1724 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1725 [BPF_LD | BPF_ABS | BPF_B] = true, 1726 [BPF_LD | BPF_ABS | BPF_H] = true, 1727 [BPF_LD | BPF_ABS | BPF_W] = true, 1728 [BPF_LD | BPF_IND | BPF_B] = true, 1729 [BPF_LD | BPF_IND | BPF_H] = true, 1730 [BPF_LD | BPF_IND | BPF_W] = true, 1731 [BPF_JMP | BPF_JA | BPF_X] = true, 1732 [BPF_JMP | BPF_JCOND] = true, 1733 }; 1734 #undef BPF_INSN_3_TBL 1735 #undef BPF_INSN_2_TBL 1736 return public_insntable[code]; 1737 } 1738 1739 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1740 /** 1741 * ___bpf_prog_run - run eBPF program on a given context 1742 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1743 * @insn: is the array of eBPF instructions 1744 * 1745 * Decode and execute eBPF instructions. 1746 * 1747 * Return: whatever value is in %BPF_R0 at program exit 1748 */ 1749 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1750 { 1751 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1752 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1753 static const void * const jumptable[256] __annotate_jump_table = { 1754 [0 ... 255] = &&default_label, 1755 /* Now overwrite non-defaults ... */ 1756 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1757 /* Non-UAPI available opcodes. */ 1758 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1759 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1760 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1761 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1762 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1763 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1764 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1765 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1766 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1767 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1768 }; 1769 #undef BPF_INSN_3_LBL 1770 #undef BPF_INSN_2_LBL 1771 u32 tail_call_cnt = 0; 1772 1773 #define CONT ({ insn++; goto select_insn; }) 1774 #define CONT_JMP ({ insn++; goto select_insn; }) 1775 1776 select_insn: 1777 goto *jumptable[insn->code]; 1778 1779 /* Explicitly mask the register-based shift amounts with 63 or 31 1780 * to avoid undefined behavior. Normally this won't affect the 1781 * generated code, for example, in case of native 64 bit archs such 1782 * as x86-64 or arm64, the compiler is optimizing the AND away for 1783 * the interpreter. In case of JITs, each of the JIT backends compiles 1784 * the BPF shift operations to machine instructions which produce 1785 * implementation-defined results in such a case; the resulting 1786 * contents of the register may be arbitrary, but program behaviour 1787 * as a whole remains defined. In other words, in case of JIT backends, 1788 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1789 */ 1790 /* ALU (shifts) */ 1791 #define SHT(OPCODE, OP) \ 1792 ALU64_##OPCODE##_X: \ 1793 DST = DST OP (SRC & 63); \ 1794 CONT; \ 1795 ALU_##OPCODE##_X: \ 1796 DST = (u32) DST OP ((u32) SRC & 31); \ 1797 CONT; \ 1798 ALU64_##OPCODE##_K: \ 1799 DST = DST OP IMM; \ 1800 CONT; \ 1801 ALU_##OPCODE##_K: \ 1802 DST = (u32) DST OP (u32) IMM; \ 1803 CONT; 1804 /* ALU (rest) */ 1805 #define ALU(OPCODE, OP) \ 1806 ALU64_##OPCODE##_X: \ 1807 DST = DST OP SRC; \ 1808 CONT; \ 1809 ALU_##OPCODE##_X: \ 1810 DST = (u32) DST OP (u32) SRC; \ 1811 CONT; \ 1812 ALU64_##OPCODE##_K: \ 1813 DST = DST OP IMM; \ 1814 CONT; \ 1815 ALU_##OPCODE##_K: \ 1816 DST = (u32) DST OP (u32) IMM; \ 1817 CONT; 1818 ALU(ADD, +) 1819 ALU(SUB, -) 1820 ALU(AND, &) 1821 ALU(OR, |) 1822 ALU(XOR, ^) 1823 ALU(MUL, *) 1824 SHT(LSH, <<) 1825 SHT(RSH, >>) 1826 #undef SHT 1827 #undef ALU 1828 ALU_NEG: 1829 DST = (u32) -DST; 1830 CONT; 1831 ALU64_NEG: 1832 DST = -DST; 1833 CONT; 1834 ALU_MOV_X: 1835 switch (OFF) { 1836 case 0: 1837 DST = (u32) SRC; 1838 break; 1839 case 8: 1840 DST = (u32)(s8) SRC; 1841 break; 1842 case 16: 1843 DST = (u32)(s16) SRC; 1844 break; 1845 } 1846 CONT; 1847 ALU_MOV_K: 1848 DST = (u32) IMM; 1849 CONT; 1850 ALU64_MOV_X: 1851 switch (OFF) { 1852 case 0: 1853 DST = SRC; 1854 break; 1855 case 8: 1856 DST = (s8) SRC; 1857 break; 1858 case 16: 1859 DST = (s16) SRC; 1860 break; 1861 case 32: 1862 DST = (s32) SRC; 1863 break; 1864 } 1865 CONT; 1866 ALU64_MOV_K: 1867 DST = IMM; 1868 CONT; 1869 LD_IMM_DW: 1870 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1871 insn++; 1872 CONT; 1873 ALU_ARSH_X: 1874 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1875 CONT; 1876 ALU_ARSH_K: 1877 DST = (u64) (u32) (((s32) DST) >> IMM); 1878 CONT; 1879 ALU64_ARSH_X: 1880 (*(s64 *) &DST) >>= (SRC & 63); 1881 CONT; 1882 ALU64_ARSH_K: 1883 (*(s64 *) &DST) >>= IMM; 1884 CONT; 1885 ALU64_MOD_X: 1886 switch (OFF) { 1887 case 0: 1888 div64_u64_rem(DST, SRC, &AX); 1889 DST = AX; 1890 break; 1891 case 1: 1892 AX = div64_s64(DST, SRC); 1893 DST = DST - AX * SRC; 1894 break; 1895 } 1896 CONT; 1897 ALU_MOD_X: 1898 switch (OFF) { 1899 case 0: 1900 AX = (u32) DST; 1901 DST = do_div(AX, (u32) SRC); 1902 break; 1903 case 1: 1904 AX = abs((s32)DST); 1905 AX = do_div(AX, abs((s32)SRC)); 1906 if ((s32)DST < 0) 1907 DST = (u32)-AX; 1908 else 1909 DST = (u32)AX; 1910 break; 1911 } 1912 CONT; 1913 ALU64_MOD_K: 1914 switch (OFF) { 1915 case 0: 1916 div64_u64_rem(DST, IMM, &AX); 1917 DST = AX; 1918 break; 1919 case 1: 1920 AX = div64_s64(DST, IMM); 1921 DST = DST - AX * IMM; 1922 break; 1923 } 1924 CONT; 1925 ALU_MOD_K: 1926 switch (OFF) { 1927 case 0: 1928 AX = (u32) DST; 1929 DST = do_div(AX, (u32) IMM); 1930 break; 1931 case 1: 1932 AX = abs((s32)DST); 1933 AX = do_div(AX, abs((s32)IMM)); 1934 if ((s32)DST < 0) 1935 DST = (u32)-AX; 1936 else 1937 DST = (u32)AX; 1938 break; 1939 } 1940 CONT; 1941 ALU64_DIV_X: 1942 switch (OFF) { 1943 case 0: 1944 DST = div64_u64(DST, SRC); 1945 break; 1946 case 1: 1947 DST = div64_s64(DST, SRC); 1948 break; 1949 } 1950 CONT; 1951 ALU_DIV_X: 1952 switch (OFF) { 1953 case 0: 1954 AX = (u32) DST; 1955 do_div(AX, (u32) SRC); 1956 DST = (u32) AX; 1957 break; 1958 case 1: 1959 AX = abs((s32)DST); 1960 do_div(AX, abs((s32)SRC)); 1961 if (((s32)DST < 0) == ((s32)SRC < 0)) 1962 DST = (u32)AX; 1963 else 1964 DST = (u32)-AX; 1965 break; 1966 } 1967 CONT; 1968 ALU64_DIV_K: 1969 switch (OFF) { 1970 case 0: 1971 DST = div64_u64(DST, IMM); 1972 break; 1973 case 1: 1974 DST = div64_s64(DST, IMM); 1975 break; 1976 } 1977 CONT; 1978 ALU_DIV_K: 1979 switch (OFF) { 1980 case 0: 1981 AX = (u32) DST; 1982 do_div(AX, (u32) IMM); 1983 DST = (u32) AX; 1984 break; 1985 case 1: 1986 AX = abs((s32)DST); 1987 do_div(AX, abs((s32)IMM)); 1988 if (((s32)DST < 0) == ((s32)IMM < 0)) 1989 DST = (u32)AX; 1990 else 1991 DST = (u32)-AX; 1992 break; 1993 } 1994 CONT; 1995 ALU_END_TO_BE: 1996 switch (IMM) { 1997 case 16: 1998 DST = (__force u16) cpu_to_be16(DST); 1999 break; 2000 case 32: 2001 DST = (__force u32) cpu_to_be32(DST); 2002 break; 2003 case 64: 2004 DST = (__force u64) cpu_to_be64(DST); 2005 break; 2006 } 2007 CONT; 2008 ALU_END_TO_LE: 2009 switch (IMM) { 2010 case 16: 2011 DST = (__force u16) cpu_to_le16(DST); 2012 break; 2013 case 32: 2014 DST = (__force u32) cpu_to_le32(DST); 2015 break; 2016 case 64: 2017 DST = (__force u64) cpu_to_le64(DST); 2018 break; 2019 } 2020 CONT; 2021 ALU64_END_TO_LE: 2022 switch (IMM) { 2023 case 16: 2024 DST = (__force u16) __swab16(DST); 2025 break; 2026 case 32: 2027 DST = (__force u32) __swab32(DST); 2028 break; 2029 case 64: 2030 DST = (__force u64) __swab64(DST); 2031 break; 2032 } 2033 CONT; 2034 2035 /* CALL */ 2036 JMP_CALL: 2037 /* Function call scratches BPF_R1-BPF_R5 registers, 2038 * preserves BPF_R6-BPF_R9, and stores return value 2039 * into BPF_R0. 2040 */ 2041 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2042 BPF_R4, BPF_R5); 2043 CONT; 2044 2045 JMP_CALL_ARGS: 2046 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2047 BPF_R3, BPF_R4, 2048 BPF_R5, 2049 insn + insn->off + 1); 2050 CONT; 2051 2052 JMP_TAIL_CALL: { 2053 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2054 struct bpf_array *array = container_of(map, struct bpf_array, map); 2055 struct bpf_prog *prog; 2056 u32 index = BPF_R3; 2057 2058 if (unlikely(index >= array->map.max_entries)) 2059 goto out; 2060 2061 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2062 goto out; 2063 2064 tail_call_cnt++; 2065 2066 prog = READ_ONCE(array->ptrs[index]); 2067 if (!prog) 2068 goto out; 2069 2070 /* ARG1 at this point is guaranteed to point to CTX from 2071 * the verifier side due to the fact that the tail call is 2072 * handled like a helper, that is, bpf_tail_call_proto, 2073 * where arg1_type is ARG_PTR_TO_CTX. 2074 */ 2075 insn = prog->insnsi; 2076 goto select_insn; 2077 out: 2078 CONT; 2079 } 2080 JMP_JA: 2081 insn += insn->off; 2082 CONT; 2083 JMP32_JA: 2084 insn += insn->imm; 2085 CONT; 2086 JMP_EXIT: 2087 return BPF_R0; 2088 /* JMP */ 2089 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2090 JMP_##OPCODE##_X: \ 2091 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2092 insn += insn->off; \ 2093 CONT_JMP; \ 2094 } \ 2095 CONT; \ 2096 JMP32_##OPCODE##_X: \ 2097 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2098 insn += insn->off; \ 2099 CONT_JMP; \ 2100 } \ 2101 CONT; \ 2102 JMP_##OPCODE##_K: \ 2103 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2104 insn += insn->off; \ 2105 CONT_JMP; \ 2106 } \ 2107 CONT; \ 2108 JMP32_##OPCODE##_K: \ 2109 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2110 insn += insn->off; \ 2111 CONT_JMP; \ 2112 } \ 2113 CONT; 2114 COND_JMP(u, JEQ, ==) 2115 COND_JMP(u, JNE, !=) 2116 COND_JMP(u, JGT, >) 2117 COND_JMP(u, JLT, <) 2118 COND_JMP(u, JGE, >=) 2119 COND_JMP(u, JLE, <=) 2120 COND_JMP(u, JSET, &) 2121 COND_JMP(s, JSGT, >) 2122 COND_JMP(s, JSLT, <) 2123 COND_JMP(s, JSGE, >=) 2124 COND_JMP(s, JSLE, <=) 2125 #undef COND_JMP 2126 /* ST, STX and LDX*/ 2127 ST_NOSPEC: 2128 /* Speculation barrier for mitigating Speculative Store Bypass, 2129 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2130 * rely on the firmware mitigation as controlled via the ssbd 2131 * kernel parameter. Whenever the mitigation is enabled, it 2132 * works for all of the kernel code with no need to provide any 2133 * additional instructions here. In case of x86, we use 'lfence' 2134 * insn for mitigation. We reuse preexisting logic from Spectre 2135 * v1 mitigation that happens to produce the required code on 2136 * x86 for v4 as well. 2137 */ 2138 barrier_nospec(); 2139 CONT; 2140 #define LDST(SIZEOP, SIZE) \ 2141 STX_MEM_##SIZEOP: \ 2142 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2143 CONT; \ 2144 ST_MEM_##SIZEOP: \ 2145 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2146 CONT; \ 2147 LDX_MEM_##SIZEOP: \ 2148 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2149 CONT; \ 2150 LDX_PROBE_MEM_##SIZEOP: \ 2151 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2152 (const void *)(long) (SRC + insn->off)); \ 2153 DST = *((SIZE *)&DST); \ 2154 CONT; 2155 2156 LDST(B, u8) 2157 LDST(H, u16) 2158 LDST(W, u32) 2159 LDST(DW, u64) 2160 #undef LDST 2161 2162 #define LDSX(SIZEOP, SIZE) \ 2163 LDX_MEMSX_##SIZEOP: \ 2164 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2165 CONT; \ 2166 LDX_PROBE_MEMSX_##SIZEOP: \ 2167 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2168 (const void *)(long) (SRC + insn->off)); \ 2169 DST = *((SIZE *)&DST); \ 2170 CONT; 2171 2172 LDSX(B, s8) 2173 LDSX(H, s16) 2174 LDSX(W, s32) 2175 #undef LDSX 2176 2177 #define ATOMIC_ALU_OP(BOP, KOP) \ 2178 case BOP: \ 2179 if (BPF_SIZE(insn->code) == BPF_W) \ 2180 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2181 (DST + insn->off)); \ 2182 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2183 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2184 (DST + insn->off)); \ 2185 else \ 2186 goto default_label; \ 2187 break; \ 2188 case BOP | BPF_FETCH: \ 2189 if (BPF_SIZE(insn->code) == BPF_W) \ 2190 SRC = (u32) atomic_fetch_##KOP( \ 2191 (u32) SRC, \ 2192 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2193 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2194 SRC = (u64) atomic64_fetch_##KOP( \ 2195 (u64) SRC, \ 2196 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2197 else \ 2198 goto default_label; \ 2199 break; 2200 2201 STX_ATOMIC_DW: 2202 STX_ATOMIC_W: 2203 STX_ATOMIC_H: 2204 STX_ATOMIC_B: 2205 switch (IMM) { 2206 /* Atomic read-modify-write instructions support only W and DW 2207 * size modifiers. 2208 */ 2209 ATOMIC_ALU_OP(BPF_ADD, add) 2210 ATOMIC_ALU_OP(BPF_AND, and) 2211 ATOMIC_ALU_OP(BPF_OR, or) 2212 ATOMIC_ALU_OP(BPF_XOR, xor) 2213 #undef ATOMIC_ALU_OP 2214 2215 case BPF_XCHG: 2216 if (BPF_SIZE(insn->code) == BPF_W) 2217 SRC = (u32) atomic_xchg( 2218 (atomic_t *)(unsigned long) (DST + insn->off), 2219 (u32) SRC); 2220 else if (BPF_SIZE(insn->code) == BPF_DW) 2221 SRC = (u64) atomic64_xchg( 2222 (atomic64_t *)(unsigned long) (DST + insn->off), 2223 (u64) SRC); 2224 else 2225 goto default_label; 2226 break; 2227 case BPF_CMPXCHG: 2228 if (BPF_SIZE(insn->code) == BPF_W) 2229 BPF_R0 = (u32) atomic_cmpxchg( 2230 (atomic_t *)(unsigned long) (DST + insn->off), 2231 (u32) BPF_R0, (u32) SRC); 2232 else if (BPF_SIZE(insn->code) == BPF_DW) 2233 BPF_R0 = (u64) atomic64_cmpxchg( 2234 (atomic64_t *)(unsigned long) (DST + insn->off), 2235 (u64) BPF_R0, (u64) SRC); 2236 else 2237 goto default_label; 2238 break; 2239 /* Atomic load and store instructions support all size 2240 * modifiers. 2241 */ 2242 case BPF_LOAD_ACQ: 2243 switch (BPF_SIZE(insn->code)) { 2244 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2245 case BPF_##SIZEOP: \ 2246 DST = (SIZE)smp_load_acquire( \ 2247 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2248 break; 2249 LOAD_ACQUIRE(B, u8) 2250 LOAD_ACQUIRE(H, u16) 2251 LOAD_ACQUIRE(W, u32) 2252 #ifdef CONFIG_64BIT 2253 LOAD_ACQUIRE(DW, u64) 2254 #endif 2255 #undef LOAD_ACQUIRE 2256 default: 2257 goto default_label; 2258 } 2259 break; 2260 case BPF_STORE_REL: 2261 switch (BPF_SIZE(insn->code)) { 2262 #define STORE_RELEASE(SIZEOP, SIZE) \ 2263 case BPF_##SIZEOP: \ 2264 smp_store_release( \ 2265 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2266 break; 2267 STORE_RELEASE(B, u8) 2268 STORE_RELEASE(H, u16) 2269 STORE_RELEASE(W, u32) 2270 #ifdef CONFIG_64BIT 2271 STORE_RELEASE(DW, u64) 2272 #endif 2273 #undef STORE_RELEASE 2274 default: 2275 goto default_label; 2276 } 2277 break; 2278 2279 default: 2280 goto default_label; 2281 } 2282 CONT; 2283 2284 default_label: 2285 /* If we ever reach this, we have a bug somewhere. Die hard here 2286 * instead of just returning 0; we could be somewhere in a subprog, 2287 * so execution could continue otherwise which we do /not/ want. 2288 * 2289 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2290 */ 2291 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2292 insn->code, insn->imm); 2293 BUG_ON(1); 2294 return 0; 2295 } 2296 2297 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2298 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2299 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2300 { \ 2301 u64 stack[stack_size / sizeof(u64)]; \ 2302 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2303 \ 2304 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2305 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2306 ARG1 = (u64) (unsigned long) ctx; \ 2307 return ___bpf_prog_run(regs, insn); \ 2308 } 2309 2310 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2311 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2312 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2313 const struct bpf_insn *insn) \ 2314 { \ 2315 u64 stack[stack_size / sizeof(u64)]; \ 2316 u64 regs[MAX_BPF_EXT_REG]; \ 2317 \ 2318 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2319 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2320 BPF_R1 = r1; \ 2321 BPF_R2 = r2; \ 2322 BPF_R3 = r3; \ 2323 BPF_R4 = r4; \ 2324 BPF_R5 = r5; \ 2325 return ___bpf_prog_run(regs, insn); \ 2326 } 2327 2328 #define EVAL1(FN, X) FN(X) 2329 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2330 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2331 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2332 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2333 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2334 2335 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2336 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2337 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2338 2339 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2340 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2341 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2342 2343 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2344 2345 static unsigned int (*interpreters[])(const void *ctx, 2346 const struct bpf_insn *insn) = { 2347 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2348 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2349 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2350 }; 2351 #undef PROG_NAME_LIST 2352 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2353 static __maybe_unused 2354 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2355 const struct bpf_insn *insn) = { 2356 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2357 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2358 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2359 }; 2360 #undef PROG_NAME_LIST 2361 2362 #ifdef CONFIG_BPF_SYSCALL 2363 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2364 { 2365 stack_depth = max_t(u32, stack_depth, 1); 2366 insn->off = (s16) insn->imm; 2367 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2368 __bpf_call_base_args; 2369 insn->code = BPF_JMP | BPF_CALL_ARGS; 2370 } 2371 #endif 2372 #endif 2373 2374 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2375 const struct bpf_insn *insn) 2376 { 2377 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2378 * is not working properly, so warn about it! 2379 */ 2380 WARN_ON_ONCE(1); 2381 return 0; 2382 } 2383 2384 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2385 const struct bpf_prog *fp) 2386 { 2387 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2388 struct bpf_prog_aux *aux = fp->aux; 2389 enum bpf_cgroup_storage_type i; 2390 bool ret = false; 2391 u64 cookie; 2392 2393 if (fp->kprobe_override) 2394 return ret; 2395 2396 spin_lock(&map->owner_lock); 2397 /* There's no owner yet where we could check for compatibility. */ 2398 if (!map->owner) { 2399 map->owner = bpf_map_owner_alloc(map); 2400 if (!map->owner) 2401 goto err; 2402 map->owner->type = prog_type; 2403 map->owner->jited = fp->jited; 2404 map->owner->xdp_has_frags = aux->xdp_has_frags; 2405 map->owner->sleepable = fp->sleepable; 2406 map->owner->expected_attach_type = fp->expected_attach_type; 2407 map->owner->attach_func_proto = aux->attach_func_proto; 2408 for_each_cgroup_storage_type(i) { 2409 map->owner->storage_cookie[i] = 2410 aux->cgroup_storage[i] ? 2411 aux->cgroup_storage[i]->cookie : 0; 2412 } 2413 ret = true; 2414 } else { 2415 ret = map->owner->type == prog_type && 2416 map->owner->jited == fp->jited && 2417 map->owner->xdp_has_frags == aux->xdp_has_frags && 2418 map->owner->sleepable == fp->sleepable; 2419 if (ret && 2420 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2421 map->owner->expected_attach_type != fp->expected_attach_type) 2422 ret = false; 2423 for_each_cgroup_storage_type(i) { 2424 if (!ret) 2425 break; 2426 cookie = aux->cgroup_storage[i] ? 2427 aux->cgroup_storage[i]->cookie : 0; 2428 ret = map->owner->storage_cookie[i] == cookie || 2429 !cookie; 2430 } 2431 if (ret && 2432 map->owner->attach_func_proto != aux->attach_func_proto) { 2433 switch (prog_type) { 2434 case BPF_PROG_TYPE_TRACING: 2435 case BPF_PROG_TYPE_LSM: 2436 case BPF_PROG_TYPE_EXT: 2437 case BPF_PROG_TYPE_STRUCT_OPS: 2438 ret = false; 2439 break; 2440 default: 2441 break; 2442 } 2443 } 2444 } 2445 err: 2446 spin_unlock(&map->owner_lock); 2447 return ret; 2448 } 2449 2450 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2451 { 2452 /* XDP programs inserted into maps are not guaranteed to run on 2453 * a particular netdev (and can run outside driver context entirely 2454 * in the case of devmap and cpumap). Until device checks 2455 * are implemented, prohibit adding dev-bound programs to program maps. 2456 */ 2457 if (bpf_prog_is_dev_bound(fp->aux)) 2458 return false; 2459 2460 return __bpf_prog_map_compatible(map, fp); 2461 } 2462 2463 static int bpf_check_tail_call(const struct bpf_prog *fp) 2464 { 2465 struct bpf_prog_aux *aux = fp->aux; 2466 int i, ret = 0; 2467 2468 mutex_lock(&aux->used_maps_mutex); 2469 for (i = 0; i < aux->used_map_cnt; i++) { 2470 struct bpf_map *map = aux->used_maps[i]; 2471 2472 if (!map_type_contains_progs(map)) 2473 continue; 2474 2475 if (!__bpf_prog_map_compatible(map, fp)) { 2476 ret = -EINVAL; 2477 goto out; 2478 } 2479 } 2480 2481 out: 2482 mutex_unlock(&aux->used_maps_mutex); 2483 return ret; 2484 } 2485 2486 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2487 { 2488 bool select_interpreter = false; 2489 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2490 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2491 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2492 2493 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2494 * But for non-JITed programs, we don't need bpf_func, so no bounds 2495 * check needed. 2496 */ 2497 if (idx < ARRAY_SIZE(interpreters)) { 2498 fp->bpf_func = interpreters[idx]; 2499 select_interpreter = true; 2500 } else { 2501 fp->bpf_func = __bpf_prog_ret0_warn; 2502 } 2503 #else 2504 fp->bpf_func = __bpf_prog_ret0_warn; 2505 #endif 2506 return select_interpreter; 2507 } 2508 2509 /** 2510 * bpf_prog_select_runtime - select exec runtime for BPF program 2511 * @fp: bpf_prog populated with BPF program 2512 * @err: pointer to error variable 2513 * 2514 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2515 * The BPF program will be executed via bpf_prog_run() function. 2516 * 2517 * Return: the &fp argument along with &err set to 0 for success or 2518 * a negative errno code on failure 2519 */ 2520 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2521 { 2522 /* In case of BPF to BPF calls, verifier did all the prep 2523 * work with regards to JITing, etc. 2524 */ 2525 bool jit_needed = false; 2526 2527 if (fp->bpf_func) 2528 goto finalize; 2529 2530 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2531 bpf_prog_has_kfunc_call(fp)) 2532 jit_needed = true; 2533 2534 if (!bpf_prog_select_interpreter(fp)) 2535 jit_needed = true; 2536 2537 /* eBPF JITs can rewrite the program in case constant 2538 * blinding is active. However, in case of error during 2539 * blinding, bpf_int_jit_compile() must always return a 2540 * valid program, which in this case would simply not 2541 * be JITed, but falls back to the interpreter. 2542 */ 2543 if (!bpf_prog_is_offloaded(fp->aux)) { 2544 *err = bpf_prog_alloc_jited_linfo(fp); 2545 if (*err) 2546 return fp; 2547 2548 fp = bpf_int_jit_compile(fp); 2549 bpf_prog_jit_attempt_done(fp); 2550 if (!fp->jited && jit_needed) { 2551 *err = -ENOTSUPP; 2552 return fp; 2553 } 2554 } else { 2555 *err = bpf_prog_offload_compile(fp); 2556 if (*err) 2557 return fp; 2558 } 2559 2560 finalize: 2561 *err = bpf_prog_lock_ro(fp); 2562 if (*err) 2563 return fp; 2564 2565 /* The tail call compatibility check can only be done at 2566 * this late stage as we need to determine, if we deal 2567 * with JITed or non JITed program concatenations and not 2568 * all eBPF JITs might immediately support all features. 2569 */ 2570 *err = bpf_check_tail_call(fp); 2571 2572 return fp; 2573 } 2574 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2575 2576 static unsigned int __bpf_prog_ret1(const void *ctx, 2577 const struct bpf_insn *insn) 2578 { 2579 return 1; 2580 } 2581 2582 static struct bpf_prog_dummy { 2583 struct bpf_prog prog; 2584 } dummy_bpf_prog = { 2585 .prog = { 2586 .bpf_func = __bpf_prog_ret1, 2587 }, 2588 }; 2589 2590 struct bpf_empty_prog_array bpf_empty_prog_array = { 2591 .null_prog = NULL, 2592 }; 2593 EXPORT_SYMBOL(bpf_empty_prog_array); 2594 2595 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2596 { 2597 struct bpf_prog_array *p; 2598 2599 if (prog_cnt) 2600 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags); 2601 else 2602 p = &bpf_empty_prog_array.hdr; 2603 2604 return p; 2605 } 2606 2607 void bpf_prog_array_free(struct bpf_prog_array *progs) 2608 { 2609 if (!progs || progs == &bpf_empty_prog_array.hdr) 2610 return; 2611 kfree_rcu(progs, rcu); 2612 } 2613 2614 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2615 { 2616 struct bpf_prog_array *progs; 2617 2618 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2619 * no need to call kfree_rcu(), just call kfree() directly. 2620 */ 2621 progs = container_of(rcu, struct bpf_prog_array, rcu); 2622 if (rcu_trace_implies_rcu_gp()) 2623 kfree(progs); 2624 else 2625 kfree_rcu(progs, rcu); 2626 } 2627 2628 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2629 { 2630 if (!progs || progs == &bpf_empty_prog_array.hdr) 2631 return; 2632 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2633 } 2634 2635 int bpf_prog_array_length(struct bpf_prog_array *array) 2636 { 2637 struct bpf_prog_array_item *item; 2638 u32 cnt = 0; 2639 2640 for (item = array->items; item->prog; item++) 2641 if (item->prog != &dummy_bpf_prog.prog) 2642 cnt++; 2643 return cnt; 2644 } 2645 2646 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2647 { 2648 struct bpf_prog_array_item *item; 2649 2650 for (item = array->items; item->prog; item++) 2651 if (item->prog != &dummy_bpf_prog.prog) 2652 return false; 2653 return true; 2654 } 2655 2656 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2657 u32 *prog_ids, 2658 u32 request_cnt) 2659 { 2660 struct bpf_prog_array_item *item; 2661 int i = 0; 2662 2663 for (item = array->items; item->prog; item++) { 2664 if (item->prog == &dummy_bpf_prog.prog) 2665 continue; 2666 prog_ids[i] = item->prog->aux->id; 2667 if (++i == request_cnt) { 2668 item++; 2669 break; 2670 } 2671 } 2672 2673 return !!(item->prog); 2674 } 2675 2676 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2677 __u32 __user *prog_ids, u32 cnt) 2678 { 2679 unsigned long err = 0; 2680 bool nospc; 2681 u32 *ids; 2682 2683 /* users of this function are doing: 2684 * cnt = bpf_prog_array_length(); 2685 * if (cnt > 0) 2686 * bpf_prog_array_copy_to_user(..., cnt); 2687 * so below kcalloc doesn't need extra cnt > 0 check. 2688 */ 2689 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2690 if (!ids) 2691 return -ENOMEM; 2692 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2693 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2694 kfree(ids); 2695 if (err) 2696 return -EFAULT; 2697 if (nospc) 2698 return -ENOSPC; 2699 return 0; 2700 } 2701 2702 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2703 struct bpf_prog *old_prog) 2704 { 2705 struct bpf_prog_array_item *item; 2706 2707 for (item = array->items; item->prog; item++) 2708 if (item->prog == old_prog) { 2709 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2710 break; 2711 } 2712 } 2713 2714 /** 2715 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2716 * index into the program array with 2717 * a dummy no-op program. 2718 * @array: a bpf_prog_array 2719 * @index: the index of the program to replace 2720 * 2721 * Skips over dummy programs, by not counting them, when calculating 2722 * the position of the program to replace. 2723 * 2724 * Return: 2725 * * 0 - Success 2726 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2727 * * -ENOENT - Index out of range 2728 */ 2729 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2730 { 2731 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2732 } 2733 2734 /** 2735 * bpf_prog_array_update_at() - Updates the program at the given index 2736 * into the program array. 2737 * @array: a bpf_prog_array 2738 * @index: the index of the program to update 2739 * @prog: the program to insert into the array 2740 * 2741 * Skips over dummy programs, by not counting them, when calculating 2742 * the position of the program to update. 2743 * 2744 * Return: 2745 * * 0 - Success 2746 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2747 * * -ENOENT - Index out of range 2748 */ 2749 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2750 struct bpf_prog *prog) 2751 { 2752 struct bpf_prog_array_item *item; 2753 2754 if (unlikely(index < 0)) 2755 return -EINVAL; 2756 2757 for (item = array->items; item->prog; item++) { 2758 if (item->prog == &dummy_bpf_prog.prog) 2759 continue; 2760 if (!index) { 2761 WRITE_ONCE(item->prog, prog); 2762 return 0; 2763 } 2764 index--; 2765 } 2766 return -ENOENT; 2767 } 2768 2769 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2770 struct bpf_prog *exclude_prog, 2771 struct bpf_prog *include_prog, 2772 u64 bpf_cookie, 2773 struct bpf_prog_array **new_array) 2774 { 2775 int new_prog_cnt, carry_prog_cnt = 0; 2776 struct bpf_prog_array_item *existing, *new; 2777 struct bpf_prog_array *array; 2778 bool found_exclude = false; 2779 2780 /* Figure out how many existing progs we need to carry over to 2781 * the new array. 2782 */ 2783 if (old_array) { 2784 existing = old_array->items; 2785 for (; existing->prog; existing++) { 2786 if (existing->prog == exclude_prog) { 2787 found_exclude = true; 2788 continue; 2789 } 2790 if (existing->prog != &dummy_bpf_prog.prog) 2791 carry_prog_cnt++; 2792 if (existing->prog == include_prog) 2793 return -EEXIST; 2794 } 2795 } 2796 2797 if (exclude_prog && !found_exclude) 2798 return -ENOENT; 2799 2800 /* How many progs (not NULL) will be in the new array? */ 2801 new_prog_cnt = carry_prog_cnt; 2802 if (include_prog) 2803 new_prog_cnt += 1; 2804 2805 /* Do we have any prog (not NULL) in the new array? */ 2806 if (!new_prog_cnt) { 2807 *new_array = NULL; 2808 return 0; 2809 } 2810 2811 /* +1 as the end of prog_array is marked with NULL */ 2812 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2813 if (!array) 2814 return -ENOMEM; 2815 new = array->items; 2816 2817 /* Fill in the new prog array */ 2818 if (carry_prog_cnt) { 2819 existing = old_array->items; 2820 for (; existing->prog; existing++) { 2821 if (existing->prog == exclude_prog || 2822 existing->prog == &dummy_bpf_prog.prog) 2823 continue; 2824 2825 new->prog = existing->prog; 2826 new->bpf_cookie = existing->bpf_cookie; 2827 new++; 2828 } 2829 } 2830 if (include_prog) { 2831 new->prog = include_prog; 2832 new->bpf_cookie = bpf_cookie; 2833 new++; 2834 } 2835 new->prog = NULL; 2836 *new_array = array; 2837 return 0; 2838 } 2839 2840 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2841 u32 *prog_ids, u32 request_cnt, 2842 u32 *prog_cnt) 2843 { 2844 u32 cnt = 0; 2845 2846 if (array) 2847 cnt = bpf_prog_array_length(array); 2848 2849 *prog_cnt = cnt; 2850 2851 /* return early if user requested only program count or nothing to copy */ 2852 if (!request_cnt || !cnt) 2853 return 0; 2854 2855 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2856 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2857 : 0; 2858 } 2859 2860 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2861 struct bpf_map **used_maps, u32 len) 2862 { 2863 struct bpf_map *map; 2864 bool sleepable; 2865 u32 i; 2866 2867 sleepable = aux->prog->sleepable; 2868 for (i = 0; i < len; i++) { 2869 map = used_maps[i]; 2870 if (map->ops->map_poke_untrack) 2871 map->ops->map_poke_untrack(map, aux); 2872 if (sleepable) 2873 atomic64_dec(&map->sleepable_refcnt); 2874 bpf_map_put(map); 2875 } 2876 } 2877 2878 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2879 { 2880 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2881 kfree(aux->used_maps); 2882 } 2883 2884 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 2885 { 2886 #ifdef CONFIG_BPF_SYSCALL 2887 struct btf_mod_pair *btf_mod; 2888 u32 i; 2889 2890 for (i = 0; i < len; i++) { 2891 btf_mod = &used_btfs[i]; 2892 if (btf_mod->module) 2893 module_put(btf_mod->module); 2894 btf_put(btf_mod->btf); 2895 } 2896 #endif 2897 } 2898 2899 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2900 { 2901 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 2902 kfree(aux->used_btfs); 2903 } 2904 2905 static void bpf_prog_free_deferred(struct work_struct *work) 2906 { 2907 struct bpf_prog_aux *aux; 2908 int i; 2909 2910 aux = container_of(work, struct bpf_prog_aux, work); 2911 #ifdef CONFIG_BPF_SYSCALL 2912 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2913 bpf_prog_stream_free(aux->prog); 2914 #endif 2915 #ifdef CONFIG_CGROUP_BPF 2916 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2917 bpf_cgroup_atype_put(aux->cgroup_atype); 2918 #endif 2919 bpf_free_used_maps(aux); 2920 bpf_free_used_btfs(aux); 2921 bpf_prog_disassoc_struct_ops(aux->prog); 2922 if (bpf_prog_is_dev_bound(aux)) 2923 bpf_prog_dev_bound_destroy(aux->prog); 2924 #ifdef CONFIG_PERF_EVENTS 2925 if (aux->prog->has_callchain_buf) 2926 put_callchain_buffers(); 2927 #endif 2928 if (aux->dst_trampoline) 2929 bpf_trampoline_put(aux->dst_trampoline); 2930 for (i = 0; i < aux->real_func_cnt; i++) { 2931 /* We can just unlink the subprog poke descriptor table as 2932 * it was originally linked to the main program and is also 2933 * released along with it. 2934 */ 2935 aux->func[i]->aux->poke_tab = NULL; 2936 bpf_jit_free(aux->func[i]); 2937 } 2938 if (aux->real_func_cnt) { 2939 kfree(aux->func); 2940 bpf_prog_unlock_free(aux->prog); 2941 } else { 2942 bpf_jit_free(aux->prog); 2943 } 2944 } 2945 2946 void bpf_prog_free(struct bpf_prog *fp) 2947 { 2948 struct bpf_prog_aux *aux = fp->aux; 2949 2950 if (aux->dst_prog) 2951 bpf_prog_put(aux->dst_prog); 2952 bpf_token_put(aux->token); 2953 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2954 schedule_work(&aux->work); 2955 } 2956 EXPORT_SYMBOL_GPL(bpf_prog_free); 2957 2958 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 2959 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2960 2961 void bpf_user_rnd_init_once(void) 2962 { 2963 prandom_init_once(&bpf_user_rnd_state); 2964 } 2965 2966 BPF_CALL_0(bpf_user_rnd_u32) 2967 { 2968 /* Should someone ever have the rather unwise idea to use some 2969 * of the registers passed into this function, then note that 2970 * this function is called from native eBPF and classic-to-eBPF 2971 * transformations. Register assignments from both sides are 2972 * different, f.e. classic always sets fn(ctx, A, X) here. 2973 */ 2974 struct rnd_state *state; 2975 u32 res; 2976 2977 state = &get_cpu_var(bpf_user_rnd_state); 2978 res = prandom_u32_state(state); 2979 put_cpu_var(bpf_user_rnd_state); 2980 2981 return res; 2982 } 2983 2984 BPF_CALL_0(bpf_get_raw_cpu_id) 2985 { 2986 return raw_smp_processor_id(); 2987 } 2988 2989 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2990 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2991 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2992 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2993 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2994 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2995 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2996 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2997 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2998 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2999 const struct bpf_func_proto bpf_jiffies64_proto __weak; 3000 3001 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3002 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3003 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3004 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3005 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3006 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3007 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3008 3009 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3010 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3011 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3012 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3013 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3014 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3015 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3016 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3017 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3018 const struct bpf_func_proto bpf_set_retval_proto __weak; 3019 const struct bpf_func_proto bpf_get_retval_proto __weak; 3020 3021 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3022 { 3023 return NULL; 3024 } 3025 3026 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3027 { 3028 return NULL; 3029 } 3030 3031 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3032 { 3033 return NULL; 3034 } 3035 3036 u64 __weak 3037 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3038 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3039 { 3040 return -ENOTSUPP; 3041 } 3042 EXPORT_SYMBOL_GPL(bpf_event_output); 3043 3044 /* Always built-in helper functions. */ 3045 const struct bpf_func_proto bpf_tail_call_proto = { 3046 /* func is unused for tail_call, we set it to pass the 3047 * get_helper_proto check 3048 */ 3049 .func = BPF_PTR_POISON, 3050 .gpl_only = false, 3051 .ret_type = RET_VOID, 3052 .arg1_type = ARG_PTR_TO_CTX, 3053 .arg2_type = ARG_CONST_MAP_PTR, 3054 .arg3_type = ARG_ANYTHING, 3055 }; 3056 3057 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3058 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3059 * eBPF and implicitly also cBPF can get JITed! 3060 */ 3061 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 3062 { 3063 return prog; 3064 } 3065 3066 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3067 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3068 */ 3069 void __weak bpf_jit_compile(struct bpf_prog *prog) 3070 { 3071 } 3072 3073 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3074 { 3075 return false; 3076 } 3077 3078 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3079 * analysis code and wants explicit zero extension inserted by verifier. 3080 * Otherwise, return FALSE. 3081 * 3082 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3083 * you don't override this. JITs that don't want these extra insns can detect 3084 * them using insn_is_zext. 3085 */ 3086 bool __weak bpf_jit_needs_zext(void) 3087 { 3088 return false; 3089 } 3090 3091 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3092 * all archs. The value returned must not change at runtime as there is 3093 * currently no support for reloading programs that were loaded without 3094 * mitigations. 3095 */ 3096 bool __weak bpf_jit_bypass_spec_v1(void) 3097 { 3098 return false; 3099 } 3100 3101 bool __weak bpf_jit_bypass_spec_v4(void) 3102 { 3103 return false; 3104 } 3105 3106 /* Return true if the JIT inlines the call to the helper corresponding to 3107 * the imm. 3108 * 3109 * The verifier will not patch the insn->imm for the call to the helper if 3110 * this returns true. 3111 */ 3112 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3113 { 3114 return false; 3115 } 3116 3117 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3118 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3119 { 3120 return false; 3121 } 3122 3123 bool __weak bpf_jit_supports_percpu_insn(void) 3124 { 3125 return false; 3126 } 3127 3128 bool __weak bpf_jit_supports_kfunc_call(void) 3129 { 3130 return false; 3131 } 3132 3133 bool __weak bpf_jit_supports_far_kfunc_call(void) 3134 { 3135 return false; 3136 } 3137 3138 bool __weak bpf_jit_supports_arena(void) 3139 { 3140 return false; 3141 } 3142 3143 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3144 { 3145 return false; 3146 } 3147 3148 bool __weak bpf_jit_supports_fsession(void) 3149 { 3150 return false; 3151 } 3152 3153 u64 __weak bpf_arch_uaddress_limit(void) 3154 { 3155 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3156 return TASK_SIZE; 3157 #else 3158 return 0; 3159 #endif 3160 } 3161 3162 /* Return TRUE if the JIT backend satisfies the following two conditions: 3163 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3164 * 2) Under the specific arch, the implementation of xchg() is the same 3165 * as atomic_xchg() on pointer-sized words. 3166 */ 3167 bool __weak bpf_jit_supports_ptr_xchg(void) 3168 { 3169 return false; 3170 } 3171 3172 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3173 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3174 */ 3175 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3176 int len) 3177 { 3178 return -EFAULT; 3179 } 3180 3181 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3182 enum bpf_text_poke_type new_t, void *old_addr, 3183 void *new_addr) 3184 { 3185 return -ENOTSUPP; 3186 } 3187 3188 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3189 { 3190 return ERR_PTR(-ENOTSUPP); 3191 } 3192 3193 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3194 { 3195 return -ENOTSUPP; 3196 } 3197 3198 bool __weak bpf_jit_supports_exceptions(void) 3199 { 3200 return false; 3201 } 3202 3203 bool __weak bpf_jit_supports_private_stack(void) 3204 { 3205 return false; 3206 } 3207 3208 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3209 { 3210 } 3211 3212 bool __weak bpf_jit_supports_timed_may_goto(void) 3213 { 3214 return false; 3215 } 3216 3217 u64 __weak arch_bpf_timed_may_goto(void) 3218 { 3219 return 0; 3220 } 3221 3222 static noinline void bpf_prog_report_may_goto_violation(void) 3223 { 3224 #ifdef CONFIG_BPF_SYSCALL 3225 struct bpf_stream_stage ss; 3226 struct bpf_prog *prog; 3227 3228 prog = bpf_prog_find_from_stack(); 3229 if (!prog) 3230 return; 3231 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3232 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3233 bpf_stream_dump_stack(ss); 3234 })); 3235 #endif 3236 } 3237 3238 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3239 { 3240 u64 time = ktime_get_mono_fast_ns(); 3241 3242 /* Populate the timestamp for this stack frame, and refresh count. */ 3243 if (!p->timestamp) { 3244 p->timestamp = time; 3245 return BPF_MAX_TIMED_LOOPS; 3246 } 3247 /* Check if we've exhausted our time slice, and zero count. */ 3248 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3249 bpf_prog_report_may_goto_violation(); 3250 return 0; 3251 } 3252 /* Refresh the count for the stack frame. */ 3253 return BPF_MAX_TIMED_LOOPS; 3254 } 3255 3256 /* for configs without MMU or 32-bit */ 3257 __weak const struct bpf_map_ops arena_map_ops; 3258 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3259 { 3260 return 0; 3261 } 3262 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3263 { 3264 return 0; 3265 } 3266 3267 #ifdef CONFIG_BPF_SYSCALL 3268 static int __init bpf_global_ma_init(void) 3269 { 3270 int ret; 3271 3272 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3273 bpf_global_ma_set = !ret; 3274 return ret; 3275 } 3276 late_initcall(bpf_global_ma_init); 3277 #endif 3278 3279 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3280 EXPORT_SYMBOL(bpf_stats_enabled_key); 3281 3282 /* All definitions of tracepoints related to BPF. */ 3283 #define CREATE_TRACE_POINTS 3284 #include <linux/bpf_trace.h> 3285 3286 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3287 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3288 3289 #ifdef CONFIG_BPF_SYSCALL 3290 3291 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3292 const char **linep, int *nump) 3293 { 3294 int idx = -1, insn_start, insn_end, len; 3295 struct bpf_line_info *linfo; 3296 void **jited_linfo; 3297 struct btf *btf; 3298 int nr_linfo; 3299 3300 btf = prog->aux->btf; 3301 linfo = prog->aux->linfo; 3302 jited_linfo = prog->aux->jited_linfo; 3303 3304 if (!btf || !linfo || !jited_linfo) 3305 return -EINVAL; 3306 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3307 3308 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3309 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3310 3311 insn_start = linfo[0].insn_off; 3312 insn_end = insn_start + len; 3313 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3314 3315 for (int i = 0; i < nr_linfo && 3316 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3317 if (jited_linfo[i] >= (void *)ip) 3318 break; 3319 idx = i; 3320 } 3321 3322 if (idx == -1) 3323 return -ENOENT; 3324 3325 /* Get base component of the file path. */ 3326 *filep = btf_name_by_offset(btf, linfo[idx].file_name_off); 3327 *filep = kbasename(*filep); 3328 /* Obtain the source line, and strip whitespace in prefix. */ 3329 *linep = btf_name_by_offset(btf, linfo[idx].line_off); 3330 while (isspace(**linep)) 3331 *linep += 1; 3332 *nump = BPF_LINE_INFO_LINE_NUM(linfo[idx].line_col); 3333 return 0; 3334 } 3335 3336 struct walk_stack_ctx { 3337 struct bpf_prog *prog; 3338 }; 3339 3340 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3341 { 3342 struct walk_stack_ctx *ctxp = cookie; 3343 struct bpf_prog *prog; 3344 3345 /* 3346 * The RCU read lock is held to safely traverse the latch tree, but we 3347 * don't need its protection when accessing the prog, since it has an 3348 * active stack frame on the current stack trace, and won't disappear. 3349 */ 3350 rcu_read_lock(); 3351 prog = bpf_prog_ksym_find(ip); 3352 rcu_read_unlock(); 3353 if (!prog) 3354 return true; 3355 /* Make sure we return the main prog if we found a subprog */ 3356 ctxp->prog = prog->aux->main_prog_aux->prog; 3357 return false; 3358 } 3359 3360 struct bpf_prog *bpf_prog_find_from_stack(void) 3361 { 3362 struct walk_stack_ctx ctx = {}; 3363 3364 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3365 return ctx.prog; 3366 } 3367 3368 #endif 3369