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 <linux/filter.h> 22 #include <linux/skbuff.h> 23 #include <linux/static_call.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_obj(*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->jit_required = IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON); 130 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 131 #ifdef CONFIG_CGROUP_BPF 132 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 133 #endif 134 135 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 136 #ifdef CONFIG_FINEIBT 137 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 138 #endif 139 mutex_init(&fp->aux->used_maps_mutex); 140 mutex_init(&fp->aux->ext_mutex); 141 mutex_init(&fp->aux->dst_mutex); 142 mutex_init(&fp->aux->st_ops_assoc_mutex); 143 144 #ifdef CONFIG_BPF_SYSCALL 145 bpf_prog_stream_init(fp); 146 #endif 147 148 return fp; 149 } 150 151 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 152 { 153 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 154 struct bpf_prog *prog; 155 int cpu; 156 157 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 158 if (!prog) 159 return NULL; 160 161 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 162 if (!prog->stats) { 163 free_percpu(prog->active); 164 kfree(prog->aux); 165 vfree(prog); 166 return NULL; 167 } 168 169 for_each_possible_cpu(cpu) { 170 struct bpf_prog_stats *pstats; 171 172 pstats = per_cpu_ptr(prog->stats, cpu); 173 u64_stats_init(&pstats->syncp); 174 } 175 return prog; 176 } 177 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 178 179 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 180 { 181 if (!prog->aux->nr_linfo || !prog->jit_requested) 182 return 0; 183 184 prog->aux->jited_linfo = kvzalloc_objs(*prog->aux->jited_linfo, 185 prog->aux->nr_linfo, 186 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 187 if (!prog->aux->jited_linfo) 188 return -ENOMEM; 189 190 return 0; 191 } 192 193 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 194 { 195 if (prog->aux->jited_linfo && 196 (!prog->jited || !prog->aux->jited_linfo[0])) { 197 kvfree(prog->aux->jited_linfo); 198 prog->aux->jited_linfo = NULL; 199 } 200 201 kfree(prog->aux->kfunc_tab); 202 prog->aux->kfunc_tab = NULL; 203 } 204 205 /* The jit engine is responsible to provide an array 206 * for insn_off to the jited_off mapping (insn_to_jit_off). 207 * 208 * The idx to this array is the insn_off. Hence, the insn_off 209 * here is relative to the prog itself instead of the main prog. 210 * This array has one entry for each xlated bpf insn. 211 * 212 * jited_off is the byte off to the end of the jited insn. 213 * 214 * Hence, with 215 * insn_start: 216 * The first bpf insn off of the prog. The insn off 217 * here is relative to the main prog. 218 * e.g. if prog is a subprog, insn_start > 0 219 * linfo_idx: 220 * The prog's idx to prog->aux->linfo and jited_linfo 221 * 222 * jited_linfo[linfo_idx] = prog->bpf_func 223 * 224 * For i > linfo_idx, 225 * 226 * jited_linfo[i] = prog->bpf_func + 227 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 228 */ 229 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 230 const u32 *insn_to_jit_off) 231 { 232 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 233 const struct bpf_line_info *linfo; 234 void **jited_linfo; 235 236 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 237 /* Userspace did not provide linfo */ 238 return; 239 240 linfo_idx = prog->aux->linfo_idx; 241 linfo = &prog->aux->linfo[linfo_idx]; 242 insn_start = linfo[0].insn_off; 243 insn_end = insn_start + prog->len; 244 245 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 246 jited_linfo[0] = prog->bpf_func; 247 248 nr_linfo = prog->aux->nr_linfo - linfo_idx; 249 250 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 251 /* The verifier ensures that linfo[i].insn_off is 252 * strictly increasing 253 */ 254 jited_linfo[i] = prog->bpf_func + 255 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 256 } 257 258 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 259 gfp_t gfp_extra_flags) 260 { 261 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 262 struct bpf_prog *fp; 263 u32 pages; 264 265 size = round_up(size, PAGE_SIZE); 266 pages = size / PAGE_SIZE; 267 if (pages <= fp_old->pages) 268 return fp_old; 269 270 fp = __vmalloc(size, gfp_flags); 271 if (fp) { 272 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 273 fp->pages = pages; 274 fp->aux->prog = fp; 275 276 /* We keep fp->aux from fp_old around in the new 277 * reallocated structure. 278 */ 279 fp_old->aux = NULL; 280 fp_old->stats = NULL; 281 fp_old->active = NULL; 282 __bpf_prog_free(fp_old); 283 } 284 285 return fp; 286 } 287 288 void __bpf_prog_free(struct bpf_prog *fp) 289 { 290 if (fp->aux) { 291 mutex_destroy(&fp->aux->used_maps_mutex); 292 mutex_destroy(&fp->aux->dst_mutex); 293 mutex_destroy(&fp->aux->st_ops_assoc_mutex); 294 kfree(fp->aux->poke_tab); 295 kfree(fp->aux); 296 } 297 free_percpu(fp->stats); 298 free_percpu(fp->active); 299 vfree(fp); 300 } 301 302 int bpf_prog_calc_tag(struct bpf_prog *fp) 303 { 304 size_t size = bpf_prog_insn_size(fp); 305 struct bpf_insn *dst; 306 bool was_ld_map; 307 u32 i; 308 309 dst = __vmalloc(size, GFP_KERNEL_ACCOUNT); 310 if (!dst) 311 return -ENOMEM; 312 313 /* We need to take out the map fd for the digest calculation 314 * since they are unstable from user space side. 315 */ 316 for (i = 0, was_ld_map = false; i < fp->len; i++) { 317 dst[i] = fp->insnsi[i]; 318 if (!was_ld_map && 319 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 320 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 321 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 322 was_ld_map = true; 323 dst[i].imm = 0; 324 } else if (was_ld_map && 325 dst[i].code == 0 && 326 dst[i].dst_reg == 0 && 327 dst[i].src_reg == 0 && 328 dst[i].off == 0) { 329 was_ld_map = false; 330 dst[i].imm = 0; 331 } else { 332 was_ld_map = false; 333 } 334 } 335 sha256((u8 *)dst, size, fp->digest); 336 vfree(dst); 337 return 0; 338 } 339 340 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 341 s32 end_new, s32 curr, const bool probe_pass) 342 { 343 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 344 s32 delta = end_new - end_old; 345 s64 imm = insn->imm; 346 347 if (curr < pos && curr + imm + 1 >= end_old) 348 imm += delta; 349 else if (curr >= end_new && curr + imm + 1 < end_new) 350 imm -= delta; 351 if (imm < imm_min || imm > imm_max) 352 return -ERANGE; 353 if (!probe_pass) 354 insn->imm = imm; 355 return 0; 356 } 357 358 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 359 s32 end_new, s32 curr, const bool probe_pass) 360 { 361 s64 off_min, off_max, off; 362 s32 delta = end_new - end_old; 363 364 if (insn->code == (BPF_JMP32 | BPF_JA)) { 365 off = insn->imm; 366 off_min = S32_MIN; 367 off_max = S32_MAX; 368 } else { 369 off = insn->off; 370 off_min = S16_MIN; 371 off_max = S16_MAX; 372 } 373 374 if (curr < pos && curr + off + 1 >= end_old) 375 off += delta; 376 else if (curr >= end_new && curr + off + 1 < end_new) 377 off -= delta; 378 if (off < off_min || off > off_max) 379 return -ERANGE; 380 if (!probe_pass) { 381 if (insn->code == (BPF_JMP32 | BPF_JA)) 382 insn->imm = off; 383 else 384 insn->off = off; 385 } 386 return 0; 387 } 388 389 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 390 s32 end_new, const bool probe_pass) 391 { 392 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 393 struct bpf_insn *insn = prog->insnsi; 394 int ret = 0; 395 396 for (i = 0; i < insn_cnt; i++, insn++) { 397 u8 code; 398 399 /* In the probing pass we still operate on the original, 400 * unpatched image in order to check overflows before we 401 * do any other adjustments. Therefore skip the patchlet. 402 */ 403 if (probe_pass && i == pos) { 404 i = end_new; 405 insn = prog->insnsi + end_old; 406 } 407 if (bpf_pseudo_func(insn)) { 408 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 409 end_new, i, probe_pass); 410 if (ret) 411 return ret; 412 continue; 413 } 414 code = insn->code; 415 if ((BPF_CLASS(code) != BPF_JMP && 416 BPF_CLASS(code) != BPF_JMP32) || 417 BPF_OP(code) == BPF_EXIT) 418 continue; 419 /* Adjust offset of jmps if we cross patch boundaries. */ 420 if (BPF_OP(code) == BPF_CALL) { 421 if (insn->src_reg != BPF_PSEUDO_CALL) 422 continue; 423 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 424 end_new, i, probe_pass); 425 } else { 426 ret = bpf_adj_delta_to_off(insn, pos, end_old, 427 end_new, i, probe_pass); 428 } 429 if (ret) 430 break; 431 } 432 433 return ret; 434 } 435 436 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 437 { 438 struct bpf_line_info *linfo; 439 u32 i, nr_linfo; 440 441 nr_linfo = prog->aux->nr_linfo; 442 if (!nr_linfo || !delta) 443 return; 444 445 linfo = prog->aux->linfo; 446 447 for (i = 0; i < nr_linfo; i++) 448 if (off < linfo[i].insn_off) 449 break; 450 451 /* Push all off < linfo[i].insn_off by delta */ 452 for (; i < nr_linfo; i++) 453 linfo[i].insn_off += delta; 454 } 455 456 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 457 const struct bpf_insn *patch, u32 len) 458 { 459 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 460 const u32 cnt_max = S16_MAX; 461 struct bpf_prog *prog_adj; 462 int err; 463 464 /* Since our patchlet doesn't expand the image, we're done. */ 465 if (insn_delta == 0) { 466 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 467 return prog; 468 } 469 470 insn_adj_cnt = prog->len + insn_delta; 471 472 /* Reject anything that would potentially let the insn->off 473 * target overflow when we have excessive program expansions. 474 * We need to probe here before we do any reallocation where 475 * we afterwards may not fail anymore. 476 */ 477 if (insn_adj_cnt > cnt_max && 478 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 479 return ERR_PTR(err); 480 481 /* Several new instructions need to be inserted. Make room 482 * for them. Likely, there's no need for a new allocation as 483 * last page could have large enough tailroom. 484 */ 485 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 486 GFP_USER); 487 if (!prog_adj) 488 return ERR_PTR(-ENOMEM); 489 490 prog_adj->len = insn_adj_cnt; 491 492 /* Patching happens in 3 steps: 493 * 494 * 1) Move over tail of insnsi from next instruction onwards, 495 * so we can patch the single target insn with one or more 496 * new ones (patching is always from 1 to n insns, n > 0). 497 * 2) Inject new instructions at the target location. 498 * 3) Adjust branch offsets if necessary. 499 */ 500 insn_rest = insn_adj_cnt - off - len; 501 502 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 503 sizeof(*patch) * insn_rest); 504 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 505 506 /* We are guaranteed to not fail at this point, otherwise 507 * the ship has sailed to reverse to the original state. An 508 * overflow cannot happen at this point. 509 */ 510 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 511 512 bpf_adj_linfo(prog_adj, off, insn_delta); 513 514 return prog_adj; 515 } 516 517 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 518 { 519 int err; 520 521 /* Branch offsets can't overflow when program is shrinking, no need 522 * to call bpf_adj_branches(..., true) here 523 */ 524 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 525 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 526 prog->len -= cnt; 527 528 err = bpf_adj_branches(prog, off, off + cnt, off, false); 529 WARN_ON_ONCE(err); 530 return err; 531 } 532 533 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 534 { 535 int i; 536 537 for (i = 0; i < fp->aux->real_func_cnt; i++) 538 bpf_prog_kallsyms_del(fp->aux->func[i]); 539 } 540 541 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 542 { 543 bpf_prog_kallsyms_del_subprogs(fp); 544 bpf_prog_kallsyms_del(fp); 545 } 546 547 #ifdef CONFIG_BPF_JIT 548 /* All BPF JIT sysctl knobs here. */ 549 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 550 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 551 int bpf_jit_harden __read_mostly; 552 long bpf_jit_limit __read_mostly; 553 long bpf_jit_limit_max __read_mostly; 554 555 static void 556 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 557 { 558 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 559 560 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 561 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 562 } 563 564 static void 565 bpf_prog_ksym_set_name(struct bpf_prog *prog) 566 { 567 char *sym = prog->aux->ksym.name; 568 const char *end = sym + KSYM_NAME_LEN; 569 const struct btf_type *type; 570 const char *func_name; 571 572 BUILD_BUG_ON(sizeof("bpf_prog_") + 573 sizeof(prog->tag) * 2 + 574 /* name has been null terminated. 575 * We should need +1 for the '_' preceding 576 * the name. However, the null character 577 * is double counted between the name and the 578 * sizeof("bpf_prog_") above, so we omit 579 * the +1 here. 580 */ 581 sizeof(prog->aux->name) > KSYM_NAME_LEN); 582 583 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 584 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 585 586 /* prog->aux->name will be ignored if full btf name is available */ 587 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 588 type = btf_type_by_id(prog->aux->btf, 589 prog->aux->func_info[prog->aux->func_idx].type_id); 590 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 591 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 592 return; 593 } 594 595 if (prog->aux->name[0]) 596 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 597 else 598 *sym = 0; 599 } 600 601 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 602 { 603 return container_of(n, struct bpf_ksym, tnode)->start; 604 } 605 606 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 607 struct latch_tree_node *b) 608 { 609 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 610 } 611 612 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 613 { 614 unsigned long val = (unsigned long)key; 615 const struct bpf_ksym *ksym; 616 617 ksym = container_of(n, struct bpf_ksym, tnode); 618 619 if (val < ksym->start) 620 return -1; 621 /* Ensure that we detect return addresses as part of the program, when 622 * the final instruction is a call for a program part of the stack 623 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 624 */ 625 if (val > ksym->end) 626 return 1; 627 628 return 0; 629 } 630 631 static const struct latch_tree_ops bpf_tree_ops = { 632 .less = bpf_tree_less, 633 .comp = bpf_tree_comp, 634 }; 635 636 static DEFINE_SPINLOCK(bpf_lock); 637 static LIST_HEAD(bpf_kallsyms); 638 static struct latch_tree_root bpf_tree __cacheline_aligned; 639 640 void bpf_ksym_add(struct bpf_ksym *ksym) 641 { 642 spin_lock_bh(&bpf_lock); 643 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 644 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 645 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 646 spin_unlock_bh(&bpf_lock); 647 } 648 649 static void __bpf_ksym_del(struct bpf_ksym *ksym) 650 { 651 if (list_empty(&ksym->lnode)) 652 return; 653 654 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 655 list_del_rcu(&ksym->lnode); 656 } 657 658 void bpf_ksym_del(struct bpf_ksym *ksym) 659 { 660 spin_lock_bh(&bpf_lock); 661 __bpf_ksym_del(ksym); 662 spin_unlock_bh(&bpf_lock); 663 } 664 665 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 666 { 667 return fp->jited && !bpf_prog_was_classic(fp); 668 } 669 670 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 671 { 672 if (!bpf_prog_kallsyms_candidate(fp) || 673 !bpf_token_capable(fp->aux->token, CAP_BPF)) 674 return; 675 676 bpf_prog_ksym_set_addr(fp); 677 bpf_prog_ksym_set_name(fp); 678 fp->aux->ksym.prog = true; 679 680 bpf_ksym_add(&fp->aux->ksym); 681 682 #ifdef CONFIG_FINEIBT 683 /* 684 * When FineIBT, code in the __cfi_foo() symbols can get executed 685 * and hence unwinder needs help. 686 */ 687 if (cfi_mode != CFI_FINEIBT) 688 return; 689 690 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 691 "__cfi_%s", fp->aux->ksym.name); 692 693 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 694 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 695 696 bpf_ksym_add(&fp->aux->ksym_prefix); 697 #endif 698 } 699 700 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 701 { 702 if (!bpf_prog_kallsyms_candidate(fp)) 703 return; 704 705 bpf_ksym_del(&fp->aux->ksym); 706 #ifdef CONFIG_FINEIBT 707 if (cfi_mode != CFI_FINEIBT) 708 return; 709 bpf_ksym_del(&fp->aux->ksym_prefix); 710 #endif 711 } 712 713 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 714 { 715 struct latch_tree_node *n; 716 717 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 718 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 719 } 720 721 int bpf_address_lookup(unsigned long addr, unsigned long *size, 722 unsigned long *off, char *sym) 723 { 724 struct bpf_ksym *ksym; 725 int ret = 0; 726 727 rcu_read_lock(); 728 ksym = bpf_ksym_find(addr); 729 if (ksym) { 730 unsigned long symbol_start = ksym->start; 731 unsigned long symbol_end = ksym->end; 732 733 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 734 735 if (size) 736 *size = symbol_end - symbol_start; 737 if (off) 738 *off = addr - symbol_start; 739 } 740 rcu_read_unlock(); 741 742 return ret; 743 } 744 745 bool is_bpf_text_address(unsigned long addr) 746 { 747 bool ret; 748 749 rcu_read_lock(); 750 ret = bpf_ksym_find(addr) != NULL; 751 rcu_read_unlock(); 752 753 return ret; 754 } 755 756 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 757 { 758 struct bpf_ksym *ksym; 759 760 WARN_ON_ONCE(!rcu_read_lock_held()); 761 ksym = bpf_ksym_find(addr); 762 763 return ksym && ksym->prog ? 764 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 765 NULL; 766 } 767 768 bool bpf_has_frame_pointer(unsigned long ip) 769 { 770 struct bpf_ksym *ksym; 771 unsigned long offset; 772 773 guard(rcu)(); 774 775 ksym = bpf_ksym_find(ip); 776 if (!ksym || !ksym->fp_start || !ksym->fp_end) 777 return false; 778 779 offset = ip - ksym->start; 780 781 return offset >= ksym->fp_start && offset < ksym->fp_end; 782 } 783 784 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 785 { 786 const struct exception_table_entry *e = NULL; 787 struct bpf_prog *prog; 788 789 rcu_read_lock(); 790 prog = bpf_prog_ksym_find(addr); 791 if (!prog) 792 goto out; 793 if (!prog->aux->num_exentries) 794 goto out; 795 796 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 797 out: 798 rcu_read_unlock(); 799 return e; 800 } 801 802 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 803 char *sym) 804 { 805 struct bpf_ksym *ksym; 806 unsigned int it = 0; 807 int ret = -ERANGE; 808 809 if (!bpf_jit_kallsyms_enabled()) 810 return ret; 811 812 rcu_read_lock(); 813 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 814 if (it++ != symnum) 815 continue; 816 817 strscpy(sym, ksym->name, KSYM_NAME_LEN); 818 819 *value = ksym->start; 820 *type = BPF_SYM_ELF_TYPE; 821 822 ret = 0; 823 break; 824 } 825 rcu_read_unlock(); 826 827 return ret; 828 } 829 830 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 831 struct bpf_jit_poke_descriptor *poke) 832 { 833 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 834 static const u32 poke_tab_max = 1024; 835 u32 slot = prog->aux->size_poke_tab; 836 u32 size = slot + 1; 837 838 if (size > poke_tab_max) 839 return -ENOSPC; 840 if (poke->tailcall_target || poke->tailcall_target_stable || 841 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 842 return -EINVAL; 843 844 switch (poke->reason) { 845 case BPF_POKE_REASON_TAIL_CALL: 846 if (!poke->tail_call.map) 847 return -EINVAL; 848 break; 849 default: 850 return -EINVAL; 851 } 852 853 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 854 if (!tab) 855 return -ENOMEM; 856 857 memcpy(&tab[slot], poke, sizeof(*poke)); 858 prog->aux->size_poke_tab = size; 859 prog->aux->poke_tab = tab; 860 861 return slot; 862 } 863 864 /* 865 * BPF program pack allocator. 866 * 867 * Most BPF programs are pretty small. Allocating a hole page for each 868 * program is sometime a waste. Many small bpf program also adds pressure 869 * to instruction TLB. To solve this issue, we introduce a BPF program pack 870 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 871 * to host BPF programs. 872 */ 873 #define BPF_PROG_CHUNK_SHIFT 6 874 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 875 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 876 877 struct bpf_prog_pack { 878 struct list_head list; 879 void *ptr; 880 bool arch_flush_needed; 881 unsigned long bitmap[]; 882 }; 883 884 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 885 { 886 memset(area, 0, size); 887 } 888 889 DEFINE_STATIC_CALL_NULL(bpf_arch_pred_flush, bpf_arch_pred_flush); 890 891 /* 892 * Enabled once bpf_arch_pred_flush points at a real flush routine. Lets the 893 * pack allocator test "is a predictor flush wired up at all" with a cheap 894 * static branch instead of repeatedly querying the static call target. 895 */ 896 DEFINE_STATIC_KEY_FALSE(bpf_pred_flush_enabled); 897 898 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 899 900 static DEFINE_MUTEX(pack_mutex); 901 static LIST_HEAD(pack_list); 902 903 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 904 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 905 */ 906 #ifdef PMD_SIZE 907 /* PMD_SIZE is really big for some archs. It doesn't make sense to 908 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 909 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 910 * greater than or equal to 2MB. 911 */ 912 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 913 #else 914 #define BPF_PROG_PACK_SIZE PAGE_SIZE 915 #endif 916 917 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 918 919 static bool bpf_jit_mem_is_rox(void) 920 { 921 return execmem_is_rox(EXECMEM_BPF); 922 } 923 924 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 925 { 926 struct bpf_prog_pack *pack; 927 int err; 928 929 pack = kzalloc_flex(*pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)); 930 if (!pack) 931 return NULL; 932 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 933 if (!pack->ptr) 934 goto out; 935 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 936 937 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 938 pack->arch_flush_needed = true; 939 if (!bpf_jit_mem_is_rox()) { 940 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 941 set_vm_flush_reset_perms(pack->ptr); 942 err = set_memory_rox((unsigned long)pack->ptr, 943 BPF_PROG_PACK_SIZE / PAGE_SIZE); 944 if (err) 945 goto out; 946 } 947 list_add_tail(&pack->list, &pack_list); 948 return pack; 949 950 out: 951 bpf_jit_free_exec(pack->ptr); 952 kfree(pack); 953 return NULL; 954 } 955 956 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns, bool was_classic) 957 { 958 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 959 struct bpf_prog_pack *pack, *fallback_pack = NULL; 960 unsigned long pos, fallback_pos = 0; 961 void *ptr = NULL; 962 963 mutex_lock(&pack_mutex); 964 if (size > BPF_PROG_PACK_SIZE) { 965 /* 966 * Allocations larger than a pack get their own pages, and 967 * predictors are not flushed for such allocation. This is only 968 * safe because cBPF programs (the unprivileged attack surface) 969 * are bounded well below a pack size. 970 */ 971 if (was_classic && static_branch_unlikely(&bpf_pred_flush_enabled)) 972 pr_warn_once("BPF: Predictors not flushed for allocations greater than BPF_PROG_PACK_SIZE\n"); 973 size = round_up(size, PAGE_SIZE); 974 ptr = bpf_jit_alloc_exec(size); 975 if (ptr && !bpf_jit_mem_is_rox()) { 976 int err; 977 978 bpf_fill_ill_insns(ptr, size); 979 set_vm_flush_reset_perms(ptr); 980 err = set_memory_rox((unsigned long)ptr, 981 size / PAGE_SIZE); 982 if (err) { 983 bpf_jit_free_exec(ptr); 984 ptr = NULL; 985 } 986 } 987 goto out; 988 } 989 list_for_each_entry(pack, &pack_list, list) { 990 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 991 nbits, 0); 992 if (pos >= BPF_PROG_CHUNK_COUNT) 993 continue; 994 /* Flush not enabled, use any pack */ 995 if (!static_branch_unlikely(&bpf_pred_flush_enabled)) 996 goto found_free_area; 997 /* 998 * cBPF reuse of a dirty pack triggers a flush, so prefer a 999 * clean pack for cBPF. eBPF never flushes, so steer it to a 1000 * dirty pack and keep clean packs free for cBPF. 1001 */ 1002 if (was_classic ^ pack->arch_flush_needed) 1003 goto found_free_area; 1004 if (!fallback_pack) { 1005 fallback_pack = pack; 1006 fallback_pos = pos; 1007 } 1008 } 1009 1010 /* No preferred pack found */ 1011 if (fallback_pack) { 1012 pack = fallback_pack; 1013 pos = fallback_pos; 1014 goto found_free_area; 1015 } 1016 1017 pack = alloc_new_pack(bpf_fill_ill_insns); 1018 if (!pack) 1019 goto out; 1020 1021 pos = 0; 1022 1023 found_free_area: 1024 /* Flush only for cBPF as it may contain a crafted gadget */ 1025 if (static_branch_unlikely(&bpf_pred_flush_enabled) && 1026 pack->arch_flush_needed && 1027 was_classic) { 1028 struct bpf_prog_pack *p; 1029 1030 static_call_cond(bpf_arch_pred_flush)(); 1031 list_for_each_entry(p, &pack_list, list) 1032 p->arch_flush_needed = false; 1033 } 1034 bitmap_set(pack->bitmap, pos, nbits); 1035 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 1036 1037 out: 1038 mutex_unlock(&pack_mutex); 1039 return ptr; 1040 } 1041 1042 void bpf_prog_pack_free(void *ptr, u32 size) 1043 { 1044 struct bpf_prog_pack *pack = NULL, *tmp; 1045 unsigned int nbits; 1046 unsigned long pos; 1047 1048 mutex_lock(&pack_mutex); 1049 if (size > BPF_PROG_PACK_SIZE) { 1050 bpf_jit_free_exec(ptr); 1051 goto out; 1052 } 1053 1054 list_for_each_entry(tmp, &pack_list, list) { 1055 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 1056 pack = tmp; 1057 break; 1058 } 1059 } 1060 1061 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1062 goto out; 1063 1064 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1065 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1066 1067 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1068 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1069 1070 bitmap_clear(pack->bitmap, pos, nbits); 1071 1072 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 1073 pack->arch_flush_needed = true; 1074 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1075 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1076 list_del(&pack->list); 1077 bpf_jit_free_exec(pack->ptr); 1078 kfree(pack); 1079 } 1080 out: 1081 mutex_unlock(&pack_mutex); 1082 } 1083 1084 static atomic_long_t bpf_jit_current; 1085 1086 /* Can be overridden by an arch's JIT compiler if it has a custom, 1087 * dedicated BPF backend memory area, or if neither of the two 1088 * below apply. 1089 */ 1090 u64 __weak bpf_jit_alloc_exec_limit(void) 1091 { 1092 #if defined(MODULES_VADDR) 1093 return MODULES_END - MODULES_VADDR; 1094 #else 1095 return VMALLOC_END - VMALLOC_START; 1096 #endif 1097 } 1098 1099 static int __init bpf_jit_charge_init(void) 1100 { 1101 /* Only used as heuristic here to derive limit. */ 1102 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1103 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1104 PAGE_SIZE), LONG_MAX); 1105 return 0; 1106 } 1107 pure_initcall(bpf_jit_charge_init); 1108 1109 int bpf_jit_charge_modmem(u32 size) 1110 { 1111 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1112 if (!bpf_capable()) { 1113 atomic_long_sub(size, &bpf_jit_current); 1114 return -EPERM; 1115 } 1116 } 1117 1118 return 0; 1119 } 1120 1121 void bpf_jit_uncharge_modmem(u32 size) 1122 { 1123 atomic_long_sub(size, &bpf_jit_current); 1124 } 1125 1126 void *bpf_jit_alloc_exec(unsigned long size) 1127 { 1128 return execmem_alloc(EXECMEM_BPF, size); 1129 } 1130 1131 void bpf_jit_free_exec(void *addr) 1132 { 1133 execmem_free(addr); 1134 } 1135 1136 struct bpf_binary_header * 1137 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1138 unsigned int alignment, 1139 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1140 { 1141 struct bpf_binary_header *hdr; 1142 u32 size, hole, start; 1143 1144 WARN_ON_ONCE(!is_power_of_2(alignment) || 1145 alignment > BPF_IMAGE_ALIGNMENT); 1146 1147 /* Most of BPF filters are really small, but if some of them 1148 * fill a page, allow at least 128 extra bytes to insert a 1149 * random section of illegal instructions. 1150 */ 1151 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1152 1153 if (bpf_jit_charge_modmem(size)) 1154 return NULL; 1155 hdr = bpf_jit_alloc_exec(size); 1156 if (!hdr) { 1157 bpf_jit_uncharge_modmem(size); 1158 return NULL; 1159 } 1160 1161 /* Fill space with illegal/arch-dep instructions. */ 1162 bpf_fill_ill_insns(hdr, size); 1163 1164 hdr->size = size; 1165 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1166 PAGE_SIZE - sizeof(*hdr)); 1167 start = get_random_u32_below(hole) & ~(alignment - 1); 1168 1169 /* Leave a random number of instructions before BPF code. */ 1170 *image_ptr = &hdr->image[start]; 1171 1172 return hdr; 1173 } 1174 1175 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1176 { 1177 u32 size = hdr->size; 1178 1179 bpf_jit_free_exec(hdr); 1180 bpf_jit_uncharge_modmem(size); 1181 } 1182 1183 /* Allocate jit binary from bpf_prog_pack allocator. 1184 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1185 * to the memory. To solve this problem, a RW buffer is also allocated at 1186 * as the same time. The JIT engine should calculate offsets based on the 1187 * RO memory address, but write JITed program to the RW buffer. Once the 1188 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1189 * the JITed program to the RO memory. 1190 */ 1191 struct bpf_binary_header * 1192 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1193 unsigned int alignment, 1194 struct bpf_binary_header **rw_header, 1195 u8 **rw_image, 1196 bpf_jit_fill_hole_t bpf_fill_ill_insns, 1197 bool was_classic) 1198 { 1199 struct bpf_binary_header *ro_header; 1200 u32 size, hole, start; 1201 1202 WARN_ON_ONCE(!is_power_of_2(alignment) || 1203 alignment > BPF_IMAGE_ALIGNMENT); 1204 1205 /* add 16 bytes for a random section of illegal instructions */ 1206 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1207 1208 if (bpf_jit_charge_modmem(size)) 1209 return NULL; 1210 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns, was_classic); 1211 if (!ro_header) { 1212 bpf_jit_uncharge_modmem(size); 1213 return NULL; 1214 } 1215 1216 *rw_header = kvmalloc(size, GFP_KERNEL); 1217 if (!*rw_header) { 1218 bpf_prog_pack_free(ro_header, size); 1219 bpf_jit_uncharge_modmem(size); 1220 return NULL; 1221 } 1222 1223 /* Fill space with illegal/arch-dep instructions. */ 1224 bpf_fill_ill_insns(*rw_header, size); 1225 (*rw_header)->size = size; 1226 1227 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1228 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1229 start = get_random_u32_below(hole) & ~(alignment - 1); 1230 1231 *image_ptr = &ro_header->image[start]; 1232 *rw_image = &(*rw_header)->image[start]; 1233 1234 return ro_header; 1235 } 1236 1237 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1238 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1239 struct bpf_binary_header *rw_header) 1240 { 1241 void *ptr; 1242 1243 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1244 1245 kvfree(rw_header); 1246 1247 if (IS_ERR(ptr)) { 1248 bpf_prog_pack_free(ro_header, ro_header->size); 1249 return PTR_ERR(ptr); 1250 } 1251 return 0; 1252 } 1253 1254 /* bpf_jit_binary_pack_free is called in two different scenarios: 1255 * 1) when the program is freed after; 1256 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1257 * For case 2), we need to free both the RO memory and the RW buffer. 1258 * 1259 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1260 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1261 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1262 * bpf_arch_text_copy (when jit fails). 1263 */ 1264 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1265 struct bpf_binary_header *rw_header) 1266 { 1267 u32 size = ro_header->size; 1268 1269 bpf_prog_pack_free(ro_header, size); 1270 kvfree(rw_header); 1271 bpf_jit_uncharge_modmem(size); 1272 } 1273 1274 struct bpf_binary_header * 1275 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1276 { 1277 unsigned long real_start = (unsigned long)fp->bpf_func; 1278 unsigned long addr; 1279 1280 addr = real_start & BPF_PROG_CHUNK_MASK; 1281 return (void *)addr; 1282 } 1283 1284 static inline struct bpf_binary_header * 1285 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1286 { 1287 unsigned long real_start = (unsigned long)fp->bpf_func; 1288 unsigned long addr; 1289 1290 addr = real_start & PAGE_MASK; 1291 return (void *)addr; 1292 } 1293 1294 /* This symbol is only overridden by archs that have different 1295 * requirements than the usual eBPF JITs, f.e. when they only 1296 * implement cBPF JIT, do not set images read-only, etc. 1297 */ 1298 void __weak bpf_jit_free(struct bpf_prog *fp) 1299 { 1300 if (fp->jited) { 1301 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1302 1303 bpf_jit_binary_free(hdr); 1304 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1305 } 1306 1307 bpf_prog_unlock_free(fp); 1308 } 1309 1310 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1311 const struct bpf_insn *insn, bool extra_pass, 1312 u64 *func_addr, bool *func_addr_fixed) 1313 { 1314 s16 off = insn->off; 1315 s32 imm = insn->imm; 1316 u8 *addr; 1317 int err; 1318 1319 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1320 if (!*func_addr_fixed) { 1321 /* Place-holder address till the last pass has collected 1322 * all addresses for JITed subprograms in which case we 1323 * can pick them up from prog->aux. 1324 */ 1325 if (!extra_pass) 1326 addr = NULL; 1327 else if (prog->aux->func && 1328 off >= 0 && off < prog->aux->real_func_cnt) 1329 addr = (u8 *)prog->aux->func[off]->bpf_func; 1330 else 1331 return -EINVAL; 1332 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1333 bpf_jit_supports_far_kfunc_call()) { 1334 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1335 if (err) 1336 return err; 1337 } else { 1338 /* Address of a BPF helper call. Since part of the core 1339 * kernel, it's always at a fixed location. __bpf_call_base 1340 * and the helper with imm relative to it are both in core 1341 * kernel. 1342 */ 1343 addr = (u8 *)__bpf_call_base + imm; 1344 } 1345 1346 *func_addr = (unsigned long)addr; 1347 return 0; 1348 } 1349 1350 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1351 { 1352 if (prog->aux->ksym.prog) 1353 return prog->aux->ksym.name; 1354 return prog->aux->name; 1355 } 1356 1357 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1358 const struct bpf_insn *aux, 1359 struct bpf_insn *to_buff, 1360 bool emit_zext) 1361 { 1362 struct bpf_insn *to = to_buff; 1363 u32 imm_rnd = get_random_u32(); 1364 s16 off; 1365 1366 BUILD_BUG_ON(BPF_REG_PARAMS + 2 != MAX_BPF_JIT_REG); 1367 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1368 1369 /* Constraints on AX register: 1370 * 1371 * AX register is inaccessible from user space. It is mapped in 1372 * all JITs, and used here for constant blinding rewrites. It is 1373 * typically "stateless" meaning its contents are only valid within 1374 * the executed instruction, but not across several instructions. 1375 * There are a few exceptions however which are further detailed 1376 * below. 1377 * 1378 * Constant blinding is only used by JITs, not in the interpreter. 1379 * The interpreter uses AX in some occasions as a local temporary 1380 * register e.g. in DIV or MOD instructions. 1381 * 1382 * In restricted circumstances, the verifier can also use the AX 1383 * register for rewrites as long as they do not interfere with 1384 * the above cases! 1385 */ 1386 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1387 goto out; 1388 1389 if (from->imm == 0 && 1390 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1391 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1392 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1393 goto out; 1394 } 1395 1396 switch (from->code) { 1397 case BPF_ALU | BPF_ADD | BPF_K: 1398 case BPF_ALU | BPF_SUB | BPF_K: 1399 case BPF_ALU | BPF_AND | BPF_K: 1400 case BPF_ALU | BPF_OR | BPF_K: 1401 case BPF_ALU | BPF_XOR | BPF_K: 1402 case BPF_ALU | BPF_MUL | BPF_K: 1403 case BPF_ALU | BPF_MOV | BPF_K: 1404 case BPF_ALU | BPF_DIV | BPF_K: 1405 case BPF_ALU | BPF_MOD | BPF_K: 1406 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1407 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1408 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1409 break; 1410 1411 case BPF_ALU64 | BPF_ADD | BPF_K: 1412 case BPF_ALU64 | BPF_SUB | BPF_K: 1413 case BPF_ALU64 | BPF_AND | BPF_K: 1414 case BPF_ALU64 | BPF_OR | BPF_K: 1415 case BPF_ALU64 | BPF_XOR | BPF_K: 1416 case BPF_ALU64 | BPF_MUL | BPF_K: 1417 case BPF_ALU64 | BPF_MOV | BPF_K: 1418 case BPF_ALU64 | BPF_DIV | BPF_K: 1419 case BPF_ALU64 | BPF_MOD | BPF_K: 1420 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1421 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1422 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1423 break; 1424 1425 case BPF_JMP | BPF_JEQ | BPF_K: 1426 case BPF_JMP | BPF_JNE | BPF_K: 1427 case BPF_JMP | BPF_JGT | BPF_K: 1428 case BPF_JMP | BPF_JLT | BPF_K: 1429 case BPF_JMP | BPF_JGE | BPF_K: 1430 case BPF_JMP | BPF_JLE | BPF_K: 1431 case BPF_JMP | BPF_JSGT | BPF_K: 1432 case BPF_JMP | BPF_JSLT | BPF_K: 1433 case BPF_JMP | BPF_JSGE | BPF_K: 1434 case BPF_JMP | BPF_JSLE | BPF_K: 1435 case BPF_JMP | BPF_JSET | BPF_K: 1436 /* Accommodate for extra offset in case of a backjump. */ 1437 off = from->off; 1438 if (off < 0) 1439 off -= 2; 1440 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1441 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1442 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1443 break; 1444 1445 case BPF_JMP32 | BPF_JEQ | BPF_K: 1446 case BPF_JMP32 | BPF_JNE | BPF_K: 1447 case BPF_JMP32 | BPF_JGT | BPF_K: 1448 case BPF_JMP32 | BPF_JLT | BPF_K: 1449 case BPF_JMP32 | BPF_JGE | BPF_K: 1450 case BPF_JMP32 | BPF_JLE | BPF_K: 1451 case BPF_JMP32 | BPF_JSGT | BPF_K: 1452 case BPF_JMP32 | BPF_JSLT | BPF_K: 1453 case BPF_JMP32 | BPF_JSGE | BPF_K: 1454 case BPF_JMP32 | BPF_JSLE | BPF_K: 1455 case BPF_JMP32 | BPF_JSET | BPF_K: 1456 /* Accommodate for extra offset in case of a backjump. */ 1457 off = from->off; 1458 if (off < 0) 1459 off -= 2; 1460 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1461 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1462 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1463 off); 1464 break; 1465 1466 case BPF_LD | BPF_IMM | BPF_DW: 1467 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1468 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1469 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1470 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1471 break; 1472 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1473 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1474 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1475 if (emit_zext) 1476 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1477 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1478 break; 1479 1480 case BPF_ST | BPF_MEM | BPF_DW: 1481 case BPF_ST | BPF_MEM | BPF_W: 1482 case BPF_ST | BPF_MEM | BPF_H: 1483 case BPF_ST | BPF_MEM | BPF_B: 1484 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1485 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1486 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1487 break; 1488 1489 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW: 1490 case BPF_ST | BPF_PROBE_MEM32 | BPF_W: 1491 case BPF_ST | BPF_PROBE_MEM32 | BPF_H: 1492 case BPF_ST | BPF_PROBE_MEM32 | BPF_B: 1493 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ 1494 from->imm); 1495 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1496 /* 1497 * Cannot use BPF_STX_MEM() macro here as it 1498 * hardcodes BPF_MEM mode, losing PROBE_MEM32 1499 * and breaking arena addressing in the JIT. 1500 */ 1501 *to++ = (struct bpf_insn) { 1502 .code = BPF_STX | BPF_PROBE_MEM32 | 1503 BPF_SIZE(from->code), 1504 .dst_reg = from->dst_reg, 1505 .src_reg = BPF_REG_AX, 1506 .off = from->off, 1507 }; 1508 break; 1509 } 1510 out: 1511 return to - to_buff; 1512 } 1513 1514 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1515 gfp_t gfp_extra_flags) 1516 { 1517 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1518 struct bpf_prog *fp; 1519 1520 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1521 if (fp != NULL) { 1522 /* aux->prog still points to the fp_other one, so 1523 * when promoting the clone to the real program, 1524 * this still needs to be adapted. 1525 */ 1526 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1527 } 1528 1529 return fp; 1530 } 1531 1532 static void bpf_prog_clone_free(struct bpf_prog *fp) 1533 { 1534 /* aux was stolen by the other clone, so we cannot free 1535 * it from this path! It will be freed eventually by the 1536 * other program on release. 1537 * 1538 * At this point, we don't need a deferred release since 1539 * clone is guaranteed to not be locked. 1540 */ 1541 fp->aux = NULL; 1542 fp->stats = NULL; 1543 fp->active = NULL; 1544 __bpf_prog_free(fp); 1545 } 1546 1547 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1548 { 1549 /* We have to repoint aux->prog to self, as we don't 1550 * know whether fp here is the clone or the original. 1551 */ 1552 fp->aux->prog = fp; 1553 if (fp->aux->offload) 1554 fp->aux->offload->prog = fp; 1555 bpf_prog_clone_free(fp_other); 1556 } 1557 1558 /* 1559 * Now this function is used only to blind the main prog and must be invoked only when 1560 * bpf_prog_need_blind() returns true. 1561 */ 1562 struct bpf_prog *bpf_jit_blind_constants(struct bpf_verifier_env *env, struct bpf_prog *prog) 1563 { 1564 struct bpf_insn insn_buff[16], aux[2]; 1565 struct bpf_prog *clone, *tmp; 1566 int insn_delta, insn_cnt; 1567 struct bpf_insn *insn; 1568 int i, rewritten; 1569 1570 if (WARN_ON_ONCE(env && env->prog != prog)) 1571 return ERR_PTR(-EINVAL); 1572 1573 clone = bpf_prog_clone_create(prog, GFP_USER); 1574 if (!clone) 1575 return ERR_PTR(-ENOMEM); 1576 1577 /* make sure bpf_patch_insn_data() patches the correct prog */ 1578 if (env) 1579 env->prog = clone; 1580 1581 insn_cnt = clone->len; 1582 insn = clone->insnsi; 1583 1584 for (i = 0; i < insn_cnt; i++, insn++) { 1585 if (bpf_pseudo_func(insn)) { 1586 /* ld_imm64 with an address of bpf subprog is not 1587 * a user controlled constant. Don't randomize it, 1588 * since it will conflict with jit_subprogs() logic. 1589 */ 1590 insn++; 1591 i++; 1592 continue; 1593 } 1594 1595 /* We temporarily need to hold the original ld64 insn 1596 * so that we can still access the first part in the 1597 * second blinding run. 1598 */ 1599 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1600 insn[1].code == 0) 1601 memcpy(aux, insn, sizeof(aux)); 1602 1603 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1604 clone->aux->verifier_zext); 1605 if (!rewritten) 1606 continue; 1607 1608 if (env) 1609 tmp = bpf_patch_insn_data(env, i, insn_buff, rewritten); 1610 else 1611 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1612 1613 if (IS_ERR_OR_NULL(tmp)) { 1614 if (env) 1615 /* restore the original prog */ 1616 env->prog = prog; 1617 /* Patching may have repointed aux->prog during 1618 * realloc from the original one, so we need to 1619 * fix it up here on error. 1620 */ 1621 bpf_jit_prog_release_other(prog, clone); 1622 return IS_ERR(tmp) ? tmp : ERR_PTR(-ENOMEM); 1623 } 1624 1625 clone = tmp; 1626 insn_delta = rewritten - 1; 1627 1628 if (env) 1629 env->prog = clone; 1630 1631 /* Walk new program and skip insns we just inserted. */ 1632 insn = clone->insnsi + i + insn_delta; 1633 insn_cnt += insn_delta; 1634 i += insn_delta; 1635 } 1636 1637 clone->blinded = 1; 1638 return clone; 1639 } 1640 1641 bool bpf_insn_is_indirect_target(const struct bpf_verifier_env *env, const struct bpf_prog *prog, 1642 int insn_idx) 1643 { 1644 if (!env) 1645 return false; 1646 insn_idx += prog->aux->subprog_start; 1647 return env->insn_aux_data[insn_idx].indirect_target; 1648 } 1649 1650 u16 bpf_out_stack_arg_cnt(const struct bpf_verifier_env *env, const struct bpf_prog *prog) 1651 { 1652 const struct bpf_subprog_info *sub; 1653 1654 if (!env) 1655 return 0; 1656 sub = &env->subprog_info[prog->aux->func_idx]; 1657 return sub->stack_arg_cnt - bpf_in_stack_arg_cnt(sub); 1658 } 1659 #endif /* CONFIG_BPF_JIT */ 1660 1661 /* Base function for offset calculation. Needs to go into .text section, 1662 * therefore keeping it non-static as well; will also be used by JITs 1663 * anyway later on, so do not let the compiler omit it. This also needs 1664 * to go into kallsyms for correlation from e.g. bpftool, so naming 1665 * must not change. 1666 */ 1667 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1668 { 1669 return 0; 1670 } 1671 EXPORT_SYMBOL_GPL(__bpf_call_base); 1672 1673 /* All UAPI available opcodes. */ 1674 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1675 /* 32 bit ALU operations. */ \ 1676 /* Register based. */ \ 1677 INSN_3(ALU, ADD, X), \ 1678 INSN_3(ALU, SUB, X), \ 1679 INSN_3(ALU, AND, X), \ 1680 INSN_3(ALU, OR, X), \ 1681 INSN_3(ALU, LSH, X), \ 1682 INSN_3(ALU, RSH, X), \ 1683 INSN_3(ALU, XOR, X), \ 1684 INSN_3(ALU, MUL, X), \ 1685 INSN_3(ALU, MOV, X), \ 1686 INSN_3(ALU, ARSH, X), \ 1687 INSN_3(ALU, DIV, X), \ 1688 INSN_3(ALU, MOD, X), \ 1689 INSN_2(ALU, NEG), \ 1690 INSN_3(ALU, END, TO_BE), \ 1691 INSN_3(ALU, END, TO_LE), \ 1692 /* Immediate based. */ \ 1693 INSN_3(ALU, ADD, K), \ 1694 INSN_3(ALU, SUB, K), \ 1695 INSN_3(ALU, AND, K), \ 1696 INSN_3(ALU, OR, K), \ 1697 INSN_3(ALU, LSH, K), \ 1698 INSN_3(ALU, RSH, K), \ 1699 INSN_3(ALU, XOR, K), \ 1700 INSN_3(ALU, MUL, K), \ 1701 INSN_3(ALU, MOV, K), \ 1702 INSN_3(ALU, ARSH, K), \ 1703 INSN_3(ALU, DIV, K), \ 1704 INSN_3(ALU, MOD, K), \ 1705 /* 64 bit ALU operations. */ \ 1706 /* Register based. */ \ 1707 INSN_3(ALU64, ADD, X), \ 1708 INSN_3(ALU64, SUB, X), \ 1709 INSN_3(ALU64, AND, X), \ 1710 INSN_3(ALU64, OR, X), \ 1711 INSN_3(ALU64, LSH, X), \ 1712 INSN_3(ALU64, RSH, X), \ 1713 INSN_3(ALU64, XOR, X), \ 1714 INSN_3(ALU64, MUL, X), \ 1715 INSN_3(ALU64, MOV, X), \ 1716 INSN_3(ALU64, ARSH, X), \ 1717 INSN_3(ALU64, DIV, X), \ 1718 INSN_3(ALU64, MOD, X), \ 1719 INSN_2(ALU64, NEG), \ 1720 INSN_3(ALU64, END, TO_LE), \ 1721 /* Immediate based. */ \ 1722 INSN_3(ALU64, ADD, K), \ 1723 INSN_3(ALU64, SUB, K), \ 1724 INSN_3(ALU64, AND, K), \ 1725 INSN_3(ALU64, OR, K), \ 1726 INSN_3(ALU64, LSH, K), \ 1727 INSN_3(ALU64, RSH, K), \ 1728 INSN_3(ALU64, XOR, K), \ 1729 INSN_3(ALU64, MUL, K), \ 1730 INSN_3(ALU64, MOV, K), \ 1731 INSN_3(ALU64, ARSH, K), \ 1732 INSN_3(ALU64, DIV, K), \ 1733 INSN_3(ALU64, MOD, K), \ 1734 /* Call instruction. */ \ 1735 INSN_2(JMP, CALL), \ 1736 /* Exit instruction. */ \ 1737 INSN_2(JMP, EXIT), \ 1738 /* 32-bit Jump instructions. */ \ 1739 /* Register based. */ \ 1740 INSN_3(JMP32, JEQ, X), \ 1741 INSN_3(JMP32, JNE, X), \ 1742 INSN_3(JMP32, JGT, X), \ 1743 INSN_3(JMP32, JLT, X), \ 1744 INSN_3(JMP32, JGE, X), \ 1745 INSN_3(JMP32, JLE, X), \ 1746 INSN_3(JMP32, JSGT, X), \ 1747 INSN_3(JMP32, JSLT, X), \ 1748 INSN_3(JMP32, JSGE, X), \ 1749 INSN_3(JMP32, JSLE, X), \ 1750 INSN_3(JMP32, JSET, X), \ 1751 /* Immediate based. */ \ 1752 INSN_3(JMP32, JEQ, K), \ 1753 INSN_3(JMP32, JNE, K), \ 1754 INSN_3(JMP32, JGT, K), \ 1755 INSN_3(JMP32, JLT, K), \ 1756 INSN_3(JMP32, JGE, K), \ 1757 INSN_3(JMP32, JLE, K), \ 1758 INSN_3(JMP32, JSGT, K), \ 1759 INSN_3(JMP32, JSLT, K), \ 1760 INSN_3(JMP32, JSGE, K), \ 1761 INSN_3(JMP32, JSLE, K), \ 1762 INSN_3(JMP32, JSET, K), \ 1763 /* Jump instructions. */ \ 1764 /* Register based. */ \ 1765 INSN_3(JMP, JEQ, X), \ 1766 INSN_3(JMP, JNE, X), \ 1767 INSN_3(JMP, JGT, X), \ 1768 INSN_3(JMP, JLT, X), \ 1769 INSN_3(JMP, JGE, X), \ 1770 INSN_3(JMP, JLE, X), \ 1771 INSN_3(JMP, JSGT, X), \ 1772 INSN_3(JMP, JSLT, X), \ 1773 INSN_3(JMP, JSGE, X), \ 1774 INSN_3(JMP, JSLE, X), \ 1775 INSN_3(JMP, JSET, X), \ 1776 /* Immediate based. */ \ 1777 INSN_3(JMP, JEQ, K), \ 1778 INSN_3(JMP, JNE, K), \ 1779 INSN_3(JMP, JGT, K), \ 1780 INSN_3(JMP, JLT, K), \ 1781 INSN_3(JMP, JGE, K), \ 1782 INSN_3(JMP, JLE, K), \ 1783 INSN_3(JMP, JSGT, K), \ 1784 INSN_3(JMP, JSLT, K), \ 1785 INSN_3(JMP, JSGE, K), \ 1786 INSN_3(JMP, JSLE, K), \ 1787 INSN_3(JMP, JSET, K), \ 1788 INSN_2(JMP, JA), \ 1789 INSN_2(JMP32, JA), \ 1790 /* Atomic operations. */ \ 1791 INSN_3(STX, ATOMIC, B), \ 1792 INSN_3(STX, ATOMIC, H), \ 1793 INSN_3(STX, ATOMIC, W), \ 1794 INSN_3(STX, ATOMIC, DW), \ 1795 /* Store instructions. */ \ 1796 /* Register based. */ \ 1797 INSN_3(STX, MEM, B), \ 1798 INSN_3(STX, MEM, H), \ 1799 INSN_3(STX, MEM, W), \ 1800 INSN_3(STX, MEM, DW), \ 1801 /* Immediate based. */ \ 1802 INSN_3(ST, MEM, B), \ 1803 INSN_3(ST, MEM, H), \ 1804 INSN_3(ST, MEM, W), \ 1805 INSN_3(ST, MEM, DW), \ 1806 /* Load instructions. */ \ 1807 /* Register based. */ \ 1808 INSN_3(LDX, MEM, B), \ 1809 INSN_3(LDX, MEM, H), \ 1810 INSN_3(LDX, MEM, W), \ 1811 INSN_3(LDX, MEM, DW), \ 1812 INSN_3(LDX, MEMSX, B), \ 1813 INSN_3(LDX, MEMSX, H), \ 1814 INSN_3(LDX, MEMSX, W), \ 1815 /* Immediate based. */ \ 1816 INSN_3(LD, IMM, DW) 1817 1818 bool bpf_opcode_in_insntable(u8 code) 1819 { 1820 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1821 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1822 static const bool public_insntable[256] = { 1823 [0 ... 255] = false, 1824 /* Now overwrite non-defaults ... */ 1825 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1826 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1827 [BPF_LD | BPF_ABS | BPF_B] = true, 1828 [BPF_LD | BPF_ABS | BPF_H] = true, 1829 [BPF_LD | BPF_ABS | BPF_W] = true, 1830 [BPF_LD | BPF_IND | BPF_B] = true, 1831 [BPF_LD | BPF_IND | BPF_H] = true, 1832 [BPF_LD | BPF_IND | BPF_W] = true, 1833 [BPF_JMP | BPF_JA | BPF_X] = true, 1834 [BPF_JMP | BPF_JCOND] = true, 1835 }; 1836 #undef BPF_INSN_3_TBL 1837 #undef BPF_INSN_2_TBL 1838 return public_insntable[code]; 1839 } 1840 1841 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1842 /* Absolute value of s32 without undefined behavior for S32_MIN */ 1843 static u32 abs_s32(s32 x) 1844 { 1845 return x >= 0 ? (u32)x : -(u32)x; 1846 } 1847 1848 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1849 const struct bpf_insn *insn); 1850 1851 /** 1852 * ___bpf_prog_run - run eBPF program on a given context 1853 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1854 * @insn: is the array of eBPF instructions 1855 * 1856 * Decode and execute eBPF instructions. 1857 * 1858 * Return: whatever value is in %BPF_R0 at program exit 1859 */ 1860 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1861 { 1862 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1863 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1864 static const void * const jumptable[256] __annotate_jump_table = { 1865 [0 ... 255] = &&default_label, 1866 /* Now overwrite non-defaults ... */ 1867 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1868 /* Non-UAPI available opcodes. */ 1869 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1870 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1871 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1872 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1873 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1874 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1875 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1876 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1877 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1878 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1879 }; 1880 #undef BPF_INSN_3_LBL 1881 #undef BPF_INSN_2_LBL 1882 u32 tail_call_cnt = 0; 1883 1884 #define CONT ({ insn++; goto select_insn; }) 1885 #define CONT_JMP ({ insn++; goto select_insn; }) 1886 1887 select_insn: 1888 goto *jumptable[insn->code]; 1889 1890 /* Explicitly mask the register-based shift amounts with 63 or 31 1891 * to avoid undefined behavior. Normally this won't affect the 1892 * generated code, for example, in case of native 64 bit archs such 1893 * as x86-64 or arm64, the compiler is optimizing the AND away for 1894 * the interpreter. In case of JITs, each of the JIT backends compiles 1895 * the BPF shift operations to machine instructions which produce 1896 * implementation-defined results in such a case; the resulting 1897 * contents of the register may be arbitrary, but program behaviour 1898 * as a whole remains defined. In other words, in case of JIT backends, 1899 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1900 */ 1901 /* ALU (shifts) */ 1902 #define SHT(OPCODE, OP) \ 1903 ALU64_##OPCODE##_X: \ 1904 DST = DST OP (SRC & 63); \ 1905 CONT; \ 1906 ALU_##OPCODE##_X: \ 1907 DST = (u32) DST OP ((u32) SRC & 31); \ 1908 CONT; \ 1909 ALU64_##OPCODE##_K: \ 1910 DST = DST OP IMM; \ 1911 CONT; \ 1912 ALU_##OPCODE##_K: \ 1913 DST = (u32) DST OP (u32) IMM; \ 1914 CONT; 1915 /* ALU (rest) */ 1916 #define ALU(OPCODE, OP) \ 1917 ALU64_##OPCODE##_X: \ 1918 DST = DST OP SRC; \ 1919 CONT; \ 1920 ALU_##OPCODE##_X: \ 1921 DST = (u32) DST OP (u32) SRC; \ 1922 CONT; \ 1923 ALU64_##OPCODE##_K: \ 1924 DST = DST OP IMM; \ 1925 CONT; \ 1926 ALU_##OPCODE##_K: \ 1927 DST = (u32) DST OP (u32) IMM; \ 1928 CONT; 1929 ALU(ADD, +) 1930 ALU(SUB, -) 1931 ALU(AND, &) 1932 ALU(OR, |) 1933 ALU(XOR, ^) 1934 ALU(MUL, *) 1935 SHT(LSH, <<) 1936 SHT(RSH, >>) 1937 #undef SHT 1938 #undef ALU 1939 ALU_NEG: 1940 DST = (u32) -DST; 1941 CONT; 1942 ALU64_NEG: 1943 DST = -DST; 1944 CONT; 1945 ALU_MOV_X: 1946 switch (OFF) { 1947 case 0: 1948 DST = (u32) SRC; 1949 break; 1950 case 8: 1951 DST = (u32)(s8) SRC; 1952 break; 1953 case 16: 1954 DST = (u32)(s16) SRC; 1955 break; 1956 } 1957 CONT; 1958 ALU_MOV_K: 1959 DST = (u32) IMM; 1960 CONT; 1961 ALU64_MOV_X: 1962 switch (OFF) { 1963 case 0: 1964 DST = SRC; 1965 break; 1966 case 8: 1967 DST = (s8) SRC; 1968 break; 1969 case 16: 1970 DST = (s16) SRC; 1971 break; 1972 case 32: 1973 DST = (s32) SRC; 1974 break; 1975 } 1976 CONT; 1977 ALU64_MOV_K: 1978 DST = IMM; 1979 CONT; 1980 LD_IMM_DW: 1981 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1982 insn++; 1983 CONT; 1984 ALU_ARSH_X: 1985 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1986 CONT; 1987 ALU_ARSH_K: 1988 DST = (u64) (u32) (((s32) DST) >> IMM); 1989 CONT; 1990 ALU64_ARSH_X: 1991 (*(s64 *) &DST) >>= (SRC & 63); 1992 CONT; 1993 ALU64_ARSH_K: 1994 (*(s64 *) &DST) >>= IMM; 1995 CONT; 1996 ALU64_MOD_X: 1997 switch (OFF) { 1998 case 0: 1999 div64_u64_rem(DST, SRC, &AX); 2000 DST = AX; 2001 break; 2002 case 1: 2003 AX = div64_s64(DST, SRC); 2004 DST = DST - AX * SRC; 2005 break; 2006 } 2007 CONT; 2008 ALU_MOD_X: 2009 switch (OFF) { 2010 case 0: 2011 AX = (u32) DST; 2012 DST = do_div(AX, (u32) SRC); 2013 break; 2014 case 1: 2015 AX = abs_s32((s32)DST); 2016 AX = do_div(AX, abs_s32((s32)SRC)); 2017 if ((s32)DST < 0) 2018 DST = (u32)-AX; 2019 else 2020 DST = (u32)AX; 2021 break; 2022 } 2023 CONT; 2024 ALU64_MOD_K: 2025 switch (OFF) { 2026 case 0: 2027 div64_u64_rem(DST, IMM, &AX); 2028 DST = AX; 2029 break; 2030 case 1: 2031 AX = div64_s64(DST, IMM); 2032 DST = DST - AX * IMM; 2033 break; 2034 } 2035 CONT; 2036 ALU_MOD_K: 2037 switch (OFF) { 2038 case 0: 2039 AX = (u32) DST; 2040 DST = do_div(AX, (u32) IMM); 2041 break; 2042 case 1: 2043 AX = abs_s32((s32)DST); 2044 AX = do_div(AX, abs_s32((s32)IMM)); 2045 if ((s32)DST < 0) 2046 DST = (u32)-AX; 2047 else 2048 DST = (u32)AX; 2049 break; 2050 } 2051 CONT; 2052 ALU64_DIV_X: 2053 switch (OFF) { 2054 case 0: 2055 DST = div64_u64(DST, SRC); 2056 break; 2057 case 1: 2058 DST = div64_s64(DST, SRC); 2059 break; 2060 } 2061 CONT; 2062 ALU_DIV_X: 2063 switch (OFF) { 2064 case 0: 2065 AX = (u32) DST; 2066 do_div(AX, (u32) SRC); 2067 DST = (u32) AX; 2068 break; 2069 case 1: 2070 AX = abs_s32((s32)DST); 2071 do_div(AX, abs_s32((s32)SRC)); 2072 if (((s32)DST < 0) == ((s32)SRC < 0)) 2073 DST = (u32)AX; 2074 else 2075 DST = (u32)-AX; 2076 break; 2077 } 2078 CONT; 2079 ALU64_DIV_K: 2080 switch (OFF) { 2081 case 0: 2082 DST = div64_u64(DST, IMM); 2083 break; 2084 case 1: 2085 DST = div64_s64(DST, IMM); 2086 break; 2087 } 2088 CONT; 2089 ALU_DIV_K: 2090 switch (OFF) { 2091 case 0: 2092 AX = (u32) DST; 2093 do_div(AX, (u32) IMM); 2094 DST = (u32) AX; 2095 break; 2096 case 1: 2097 AX = abs_s32((s32)DST); 2098 do_div(AX, abs_s32((s32)IMM)); 2099 if (((s32)DST < 0) == ((s32)IMM < 0)) 2100 DST = (u32)AX; 2101 else 2102 DST = (u32)-AX; 2103 break; 2104 } 2105 CONT; 2106 ALU_END_TO_BE: 2107 switch (IMM) { 2108 case 16: 2109 DST = (__force u16) cpu_to_be16(DST); 2110 break; 2111 case 32: 2112 DST = (__force u32) cpu_to_be32(DST); 2113 break; 2114 case 64: 2115 DST = (__force u64) cpu_to_be64(DST); 2116 break; 2117 } 2118 CONT; 2119 ALU_END_TO_LE: 2120 switch (IMM) { 2121 case 16: 2122 DST = (__force u16) cpu_to_le16(DST); 2123 break; 2124 case 32: 2125 DST = (__force u32) cpu_to_le32(DST); 2126 break; 2127 case 64: 2128 DST = (__force u64) cpu_to_le64(DST); 2129 break; 2130 } 2131 CONT; 2132 ALU64_END_TO_LE: 2133 switch (IMM) { 2134 case 16: 2135 DST = (__force u16) __swab16(DST); 2136 break; 2137 case 32: 2138 DST = (__force u32) __swab32(DST); 2139 break; 2140 case 64: 2141 DST = (__force u64) __swab64(DST); 2142 break; 2143 } 2144 CONT; 2145 2146 /* CALL */ 2147 JMP_CALL: 2148 /* Function call scratches BPF_R1-BPF_R5 registers, 2149 * preserves BPF_R6-BPF_R9, and stores return value 2150 * into BPF_R0. 2151 */ 2152 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2153 BPF_R4, BPF_R5); 2154 CONT; 2155 2156 JMP_CALL_ARGS: 2157 BPF_R0 = interpreters_args[insn->off](BPF_R1, BPF_R2, BPF_R3, 2158 BPF_R4, BPF_R5, 2159 insn + insn->imm + 1); 2160 CONT; 2161 2162 JMP_TAIL_CALL: { 2163 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2164 struct bpf_array *array = container_of(map, struct bpf_array, map); 2165 struct bpf_prog *prog; 2166 u32 index = BPF_R3; 2167 2168 if (unlikely(index >= array->map.max_entries)) 2169 goto out; 2170 2171 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2172 goto out; 2173 2174 prog = READ_ONCE(array->ptrs[index]); 2175 if (!prog) 2176 goto out; 2177 2178 tail_call_cnt++; 2179 2180 /* ARG1 at this point is guaranteed to point to CTX from 2181 * the verifier side due to the fact that the tail call is 2182 * handled like a helper, that is, bpf_tail_call_proto, 2183 * where arg1_type is ARG_PTR_TO_CTX. 2184 */ 2185 insn = prog->insnsi; 2186 goto select_insn; 2187 out: 2188 CONT; 2189 } 2190 JMP_JA: 2191 insn += insn->off; 2192 CONT; 2193 JMP32_JA: 2194 insn += insn->imm; 2195 CONT; 2196 JMP_EXIT: 2197 return BPF_R0; 2198 /* JMP */ 2199 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2200 JMP_##OPCODE##_X: \ 2201 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2202 insn += insn->off; \ 2203 CONT_JMP; \ 2204 } \ 2205 CONT; \ 2206 JMP32_##OPCODE##_X: \ 2207 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2208 insn += insn->off; \ 2209 CONT_JMP; \ 2210 } \ 2211 CONT; \ 2212 JMP_##OPCODE##_K: \ 2213 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2214 insn += insn->off; \ 2215 CONT_JMP; \ 2216 } \ 2217 CONT; \ 2218 JMP32_##OPCODE##_K: \ 2219 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2220 insn += insn->off; \ 2221 CONT_JMP; \ 2222 } \ 2223 CONT; 2224 COND_JMP(u, JEQ, ==) 2225 COND_JMP(u, JNE, !=) 2226 COND_JMP(u, JGT, >) 2227 COND_JMP(u, JLT, <) 2228 COND_JMP(u, JGE, >=) 2229 COND_JMP(u, JLE, <=) 2230 COND_JMP(u, JSET, &) 2231 COND_JMP(s, JSGT, >) 2232 COND_JMP(s, JSLT, <) 2233 COND_JMP(s, JSGE, >=) 2234 COND_JMP(s, JSLE, <=) 2235 #undef COND_JMP 2236 /* ST, STX and LDX*/ 2237 ST_NOSPEC: 2238 /* Speculation barrier for mitigating Speculative Store Bypass, 2239 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2240 * rely on the firmware mitigation as controlled via the ssbd 2241 * kernel parameter. Whenever the mitigation is enabled, it 2242 * works for all of the kernel code with no need to provide any 2243 * additional instructions here. In case of x86, we use 'lfence' 2244 * insn for mitigation. We reuse preexisting logic from Spectre 2245 * v1 mitigation that happens to produce the required code on 2246 * x86 for v4 as well. 2247 */ 2248 barrier_nospec(); 2249 CONT; 2250 #define LDST(SIZEOP, SIZE) \ 2251 STX_MEM_##SIZEOP: \ 2252 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2253 CONT; \ 2254 ST_MEM_##SIZEOP: \ 2255 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2256 CONT; \ 2257 LDX_MEM_##SIZEOP: \ 2258 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2259 CONT; \ 2260 LDX_PROBE_MEM_##SIZEOP: \ 2261 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2262 (const void *)(long) (SRC + insn->off)); \ 2263 DST = *((SIZE *)&DST); \ 2264 CONT; 2265 2266 LDST(B, u8) 2267 LDST(H, u16) 2268 LDST(W, u32) 2269 LDST(DW, u64) 2270 #undef LDST 2271 2272 #define LDSX(SIZEOP, SIZE) \ 2273 LDX_MEMSX_##SIZEOP: \ 2274 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2275 CONT; \ 2276 LDX_PROBE_MEMSX_##SIZEOP: \ 2277 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2278 (const void *)(long) (SRC + insn->off)); \ 2279 DST = *((SIZE *)&DST); \ 2280 CONT; 2281 2282 LDSX(B, s8) 2283 LDSX(H, s16) 2284 LDSX(W, s32) 2285 #undef LDSX 2286 2287 #define ATOMIC_ALU_OP(BOP, KOP) \ 2288 case BOP: \ 2289 if (BPF_SIZE(insn->code) == BPF_W) \ 2290 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2291 (DST + insn->off)); \ 2292 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2293 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2294 (DST + insn->off)); \ 2295 else \ 2296 goto default_label; \ 2297 break; \ 2298 case BOP | BPF_FETCH: \ 2299 if (BPF_SIZE(insn->code) == BPF_W) \ 2300 SRC = (u32) atomic_fetch_##KOP( \ 2301 (u32) SRC, \ 2302 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2303 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2304 SRC = (u64) atomic64_fetch_##KOP( \ 2305 (u64) SRC, \ 2306 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2307 else \ 2308 goto default_label; \ 2309 break; 2310 2311 STX_ATOMIC_DW: 2312 STX_ATOMIC_W: 2313 STX_ATOMIC_H: 2314 STX_ATOMIC_B: 2315 switch (IMM) { 2316 /* Atomic read-modify-write instructions support only W and DW 2317 * size modifiers. 2318 */ 2319 ATOMIC_ALU_OP(BPF_ADD, add) 2320 ATOMIC_ALU_OP(BPF_AND, and) 2321 ATOMIC_ALU_OP(BPF_OR, or) 2322 ATOMIC_ALU_OP(BPF_XOR, xor) 2323 #undef ATOMIC_ALU_OP 2324 2325 case BPF_XCHG: 2326 if (BPF_SIZE(insn->code) == BPF_W) 2327 SRC = (u32) atomic_xchg( 2328 (atomic_t *)(unsigned long) (DST + insn->off), 2329 (u32) SRC); 2330 else if (BPF_SIZE(insn->code) == BPF_DW) 2331 SRC = (u64) atomic64_xchg( 2332 (atomic64_t *)(unsigned long) (DST + insn->off), 2333 (u64) SRC); 2334 else 2335 goto default_label; 2336 break; 2337 case BPF_CMPXCHG: 2338 if (BPF_SIZE(insn->code) == BPF_W) 2339 BPF_R0 = (u32) atomic_cmpxchg( 2340 (atomic_t *)(unsigned long) (DST + insn->off), 2341 (u32) BPF_R0, (u32) SRC); 2342 else if (BPF_SIZE(insn->code) == BPF_DW) 2343 BPF_R0 = (u64) atomic64_cmpxchg( 2344 (atomic64_t *)(unsigned long) (DST + insn->off), 2345 (u64) BPF_R0, (u64) SRC); 2346 else 2347 goto default_label; 2348 break; 2349 /* Atomic load and store instructions support all size 2350 * modifiers. 2351 */ 2352 case BPF_LOAD_ACQ: 2353 switch (BPF_SIZE(insn->code)) { 2354 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2355 case BPF_##SIZEOP: \ 2356 DST = (SIZE)smp_load_acquire( \ 2357 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2358 break; 2359 LOAD_ACQUIRE(B, u8) 2360 LOAD_ACQUIRE(H, u16) 2361 LOAD_ACQUIRE(W, u32) 2362 #ifdef CONFIG_64BIT 2363 LOAD_ACQUIRE(DW, u64) 2364 #endif 2365 #undef LOAD_ACQUIRE 2366 default: 2367 goto default_label; 2368 } 2369 break; 2370 case BPF_STORE_REL: 2371 switch (BPF_SIZE(insn->code)) { 2372 #define STORE_RELEASE(SIZEOP, SIZE) \ 2373 case BPF_##SIZEOP: \ 2374 smp_store_release( \ 2375 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2376 break; 2377 STORE_RELEASE(B, u8) 2378 STORE_RELEASE(H, u16) 2379 STORE_RELEASE(W, u32) 2380 #ifdef CONFIG_64BIT 2381 STORE_RELEASE(DW, u64) 2382 #endif 2383 #undef STORE_RELEASE 2384 default: 2385 goto default_label; 2386 } 2387 break; 2388 2389 default: 2390 goto default_label; 2391 } 2392 CONT; 2393 2394 default_label: 2395 /* If we ever reach this, we have a bug somewhere. Die hard here 2396 * instead of just returning 0; we could be somewhere in a subprog, 2397 * so execution could continue otherwise which we do /not/ want. 2398 * 2399 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2400 */ 2401 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2402 insn->code, insn->imm); 2403 BUG_ON(1); 2404 return 0; 2405 } 2406 2407 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2408 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2409 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2410 { \ 2411 u64 stack[stack_size / sizeof(u64)]; \ 2412 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2413 \ 2414 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2415 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2416 ARG1 = (u64) (unsigned long) ctx; \ 2417 return ___bpf_prog_run(regs, insn); \ 2418 } 2419 2420 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2421 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2422 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2423 const struct bpf_insn *insn) \ 2424 { \ 2425 u64 stack[stack_size / sizeof(u64)]; \ 2426 u64 regs[MAX_BPF_EXT_REG]; \ 2427 \ 2428 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2429 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2430 BPF_R1 = r1; \ 2431 BPF_R2 = r2; \ 2432 BPF_R3 = r3; \ 2433 BPF_R4 = r4; \ 2434 BPF_R5 = r5; \ 2435 return ___bpf_prog_run(regs, insn); \ 2436 } 2437 2438 #define EVAL1(FN, X) FN(X) 2439 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2440 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2441 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2442 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2443 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2444 2445 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2446 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2447 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2448 2449 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2450 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2451 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2452 2453 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2454 2455 static unsigned int (*interpreters[])(const void *ctx, 2456 const struct bpf_insn *insn) = { 2457 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2458 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2459 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2460 }; 2461 #undef PROG_NAME_LIST 2462 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2463 static __maybe_unused 2464 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2465 const struct bpf_insn *insn) = { 2466 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2467 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2468 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2469 }; 2470 #undef PROG_NAME_LIST 2471 2472 #ifdef CONFIG_BPF_SYSCALL 2473 int bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2474 { 2475 stack_depth = max_t(u32, stack_depth, 1); 2476 /* Prevent out-of-bounds read to interpreters_args */ 2477 if (stack_depth > MAX_BPF_STACK) 2478 return -EINVAL; 2479 insn->off = (round_up(stack_depth, 32) / 32) - 1; 2480 insn->code = BPF_JMP | BPF_CALL_ARGS; 2481 return 0; 2482 } 2483 2484 s32 bpf_call_args_imm(s16 idx) 2485 { 2486 if (WARN_ON_ONCE(idx < 0 || idx >= ARRAY_SIZE(interpreters_args))) 2487 return 0; 2488 return BPF_CALL_IMM(interpreters_args[idx]); 2489 } 2490 #endif 2491 #endif 2492 2493 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2494 const struct bpf_insn *insn) 2495 { 2496 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2497 * is not working properly, so warn about it! 2498 */ 2499 WARN_ON_ONCE(1); 2500 return 0; 2501 } 2502 2503 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2504 const struct bpf_prog *fp) 2505 { 2506 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2507 struct bpf_prog_aux *aux = fp->aux; 2508 enum bpf_cgroup_storage_type i; 2509 bool ret = false; 2510 u64 cookie; 2511 2512 if (fp->kprobe_override) 2513 return ret; 2514 2515 spin_lock(&map->owner_lock); 2516 /* There's no owner yet where we could check for compatibility. */ 2517 if (!map->owner) { 2518 map->owner = bpf_map_owner_alloc(map); 2519 if (!map->owner) 2520 goto err; 2521 map->owner->type = prog_type; 2522 map->owner->jited = fp->jited; 2523 map->owner->xdp_has_frags = aux->xdp_has_frags; 2524 map->owner->sleepable = fp->sleepable; 2525 map->owner->expected_attach_type = fp->expected_attach_type; 2526 map->owner->attach_func_proto = aux->attach_func_proto; 2527 for_each_cgroup_storage_type(i) { 2528 map->owner->storage_cookie[i] = 2529 aux->cgroup_storage[i] ? 2530 aux->cgroup_storage[i]->cookie : 0; 2531 } 2532 ret = true; 2533 } else { 2534 ret = map->owner->type == prog_type && 2535 map->owner->jited == fp->jited && 2536 map->owner->xdp_has_frags == aux->xdp_has_frags && 2537 map->owner->sleepable == fp->sleepable; 2538 if (ret && 2539 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2540 map->owner->expected_attach_type != fp->expected_attach_type) 2541 ret = false; 2542 for_each_cgroup_storage_type(i) { 2543 if (!ret) 2544 break; 2545 cookie = aux->cgroup_storage[i] ? 2546 aux->cgroup_storage[i]->cookie : 0; 2547 ret = map->owner->storage_cookie[i] == cookie || 2548 (!cookie && !aux->tail_call_reachable); 2549 } 2550 if (ret && 2551 map->owner->attach_func_proto != aux->attach_func_proto) { 2552 switch (prog_type) { 2553 case BPF_PROG_TYPE_TRACING: 2554 case BPF_PROG_TYPE_LSM: 2555 case BPF_PROG_TYPE_EXT: 2556 case BPF_PROG_TYPE_STRUCT_OPS: 2557 ret = false; 2558 break; 2559 default: 2560 break; 2561 } 2562 } 2563 } 2564 err: 2565 spin_unlock(&map->owner_lock); 2566 return ret; 2567 } 2568 2569 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2570 { 2571 /* XDP programs inserted into maps are not guaranteed to run on 2572 * a particular netdev (and can run outside driver context entirely 2573 * in the case of devmap and cpumap). Until device checks 2574 * are implemented, prohibit adding dev-bound programs to program maps. 2575 */ 2576 if (bpf_prog_is_dev_bound(fp->aux)) 2577 return false; 2578 2579 return __bpf_prog_map_compatible(map, fp); 2580 } 2581 2582 static int bpf_check_tail_call(const struct bpf_prog *fp) 2583 { 2584 struct bpf_prog_aux *aux = fp->aux; 2585 int i, ret = 0; 2586 2587 mutex_lock(&aux->used_maps_mutex); 2588 for (i = 0; i < aux->used_map_cnt; i++) { 2589 struct bpf_map *map = aux->used_maps[i]; 2590 2591 if (!map_type_contains_progs(map)) 2592 continue; 2593 2594 if (!__bpf_prog_map_compatible(map, fp)) { 2595 ret = -EINVAL; 2596 goto out; 2597 } 2598 } 2599 2600 out: 2601 mutex_unlock(&aux->used_maps_mutex); 2602 return ret; 2603 } 2604 2605 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2606 { 2607 bool select_interpreter = false; 2608 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2609 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2610 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2611 2612 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2613 * But for non-JITed programs, we don't need bpf_func, so no bounds 2614 * check needed. 2615 */ 2616 if (idx < ARRAY_SIZE(interpreters)) { 2617 fp->bpf_func = interpreters[idx]; 2618 select_interpreter = true; 2619 } else { 2620 fp->bpf_func = __bpf_prog_ret0_warn; 2621 } 2622 #else 2623 fp->bpf_func = __bpf_prog_ret0_warn; 2624 #endif 2625 return select_interpreter; 2626 } 2627 2628 static struct bpf_prog *bpf_prog_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 2629 { 2630 #ifdef CONFIG_BPF_JIT 2631 struct bpf_prog *orig_prog; 2632 2633 if (!bpf_prog_need_blind(prog)) 2634 return bpf_int_jit_compile(env, prog); 2635 2636 orig_prog = prog; 2637 prog = bpf_jit_blind_constants(env, prog); 2638 /* 2639 * If blinding was requested and we failed during blinding, we must fall 2640 * back to the interpreter. 2641 */ 2642 if (IS_ERR(prog)) 2643 goto out_restore; 2644 2645 prog = bpf_int_jit_compile(env, prog); 2646 if (prog->jited) { 2647 bpf_jit_prog_release_other(prog, orig_prog); 2648 return prog; 2649 } 2650 2651 bpf_jit_prog_release_other(orig_prog, prog); 2652 2653 out_restore: 2654 prog = orig_prog; 2655 #endif 2656 return prog; 2657 } 2658 2659 struct bpf_prog *__bpf_prog_select_runtime(struct bpf_verifier_env *env, struct bpf_prog *fp, 2660 int *err) 2661 { 2662 /* In case of BPF to BPF calls, verifier did all the prep 2663 * work with regards to JITing, etc. 2664 */ 2665 bool jit_needed = fp->jit_required; 2666 2667 if (fp->bpf_func) 2668 goto finalize; 2669 2670 if (!bpf_prog_select_interpreter(fp)) 2671 jit_needed = true; 2672 2673 /* eBPF JITs can rewrite the program in case constant 2674 * blinding is active. However, in case of error during 2675 * blinding, bpf_int_jit_compile() must always return a 2676 * valid program, which in this case would simply not 2677 * be JITed, but falls back to the interpreter. 2678 */ 2679 if (!bpf_prog_is_offloaded(fp->aux)) { 2680 *err = bpf_prog_alloc_jited_linfo(fp); 2681 if (*err) 2682 return fp; 2683 2684 fp = bpf_prog_jit_compile(env, fp); 2685 bpf_prog_jit_attempt_done(fp); 2686 if (!fp->jited && jit_needed) { 2687 *err = -ENOTSUPP; 2688 return fp; 2689 } 2690 } else { 2691 *err = bpf_prog_offload_compile(fp); 2692 if (*err) 2693 return fp; 2694 } 2695 2696 finalize: 2697 *err = bpf_prog_lock_ro(fp); 2698 if (*err) 2699 return fp; 2700 2701 /* The tail call compatibility check can only be done at 2702 * this late stage as we need to determine, if we deal 2703 * with JITed or non JITed program concatenations and not 2704 * all eBPF JITs might immediately support all features. 2705 */ 2706 *err = bpf_check_tail_call(fp); 2707 2708 return fp; 2709 } 2710 2711 /** 2712 * bpf_prog_select_runtime - select exec runtime for BPF program 2713 * @fp: bpf_prog populated with BPF program 2714 * @err: pointer to error variable 2715 * 2716 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2717 * The BPF program will be executed via bpf_prog_run() function. 2718 * 2719 * Return: the &fp argument along with &err set to 0 for success or 2720 * a negative errno code on failure 2721 */ 2722 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2723 { 2724 return __bpf_prog_select_runtime(NULL, fp, err); 2725 } 2726 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2727 2728 static unsigned int __bpf_prog_ret1(const void *ctx, 2729 const struct bpf_insn *insn) 2730 { 2731 return 1; 2732 } 2733 2734 static struct bpf_prog_dummy { 2735 struct bpf_prog prog; 2736 } dummy_bpf_prog = { 2737 .prog = { 2738 .bpf_func = __bpf_prog_ret1, 2739 }, 2740 }; 2741 2742 struct bpf_prog_array bpf_empty_prog_array = { 2743 .items = { 2744 { .prog = NULL }, 2745 }, 2746 }; 2747 EXPORT_SYMBOL(bpf_empty_prog_array); 2748 2749 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2750 { 2751 struct bpf_prog_array *p; 2752 2753 if (prog_cnt) 2754 p = kzalloc_flex(*p, items, prog_cnt + 1, flags); 2755 else 2756 p = &bpf_empty_prog_array; 2757 2758 return p; 2759 } 2760 2761 void bpf_prog_array_free(struct bpf_prog_array *progs) 2762 { 2763 if (!progs || progs == &bpf_empty_prog_array) 2764 return; 2765 kfree_rcu(progs, rcu); 2766 } 2767 2768 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2769 { 2770 struct bpf_prog_array *progs; 2771 2772 /* 2773 * RCU Tasks Trace grace period implies RCU grace period, there is no 2774 * need to call kfree_rcu(), just call kfree() directly. 2775 */ 2776 progs = container_of(rcu, struct bpf_prog_array, rcu); 2777 kfree(progs); 2778 } 2779 2780 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2781 { 2782 if (!progs || progs == &bpf_empty_prog_array) 2783 return; 2784 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2785 } 2786 2787 int bpf_prog_array_length(struct bpf_prog_array *array) 2788 { 2789 struct bpf_prog_array_item *item; 2790 u32 cnt = 0; 2791 2792 for (item = array->items; item->prog; item++) 2793 if (item->prog != &dummy_bpf_prog.prog) 2794 cnt++; 2795 return cnt; 2796 } 2797 2798 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2799 { 2800 struct bpf_prog_array_item *item; 2801 2802 for (item = array->items; item->prog; item++) 2803 if (item->prog != &dummy_bpf_prog.prog) 2804 return false; 2805 return true; 2806 } 2807 2808 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2809 u32 *prog_ids, 2810 u32 request_cnt) 2811 { 2812 struct bpf_prog_array_item *item; 2813 int i = 0; 2814 2815 for (item = array->items; item->prog; item++) { 2816 if (item->prog == &dummy_bpf_prog.prog) 2817 continue; 2818 prog_ids[i] = item->prog->aux->id; 2819 if (++i == request_cnt) { 2820 item++; 2821 break; 2822 } 2823 } 2824 2825 return !!(item->prog); 2826 } 2827 2828 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2829 __u32 __user *prog_ids, u32 cnt) 2830 { 2831 unsigned long err = 0; 2832 bool nospc; 2833 u32 *ids; 2834 2835 /* users of this function are doing: 2836 * cnt = bpf_prog_array_length(); 2837 * if (cnt > 0) 2838 * bpf_prog_array_copy_to_user(..., cnt); 2839 * so below kcalloc doesn't need extra cnt > 0 check. 2840 */ 2841 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2842 if (!ids) 2843 return -ENOMEM; 2844 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2845 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2846 kfree(ids); 2847 if (err) 2848 return -EFAULT; 2849 if (nospc) 2850 return -ENOSPC; 2851 return 0; 2852 } 2853 2854 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2855 struct bpf_prog *old_prog) 2856 { 2857 struct bpf_prog_array_item *item; 2858 2859 for (item = array->items; item->prog; item++) 2860 if (item->prog == old_prog) { 2861 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2862 break; 2863 } 2864 } 2865 2866 /** 2867 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2868 * index into the program array with 2869 * a dummy no-op program. 2870 * @array: a bpf_prog_array 2871 * @index: the index of the program to replace 2872 * 2873 * Skips over dummy programs, by not counting them, when calculating 2874 * the position of the program to replace. 2875 * 2876 * Return: 2877 * * 0 - Success 2878 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2879 * * -ENOENT - Index out of range 2880 */ 2881 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2882 { 2883 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2884 } 2885 2886 /** 2887 * bpf_prog_array_update_at() - Updates the program at the given index 2888 * into the program array. 2889 * @array: a bpf_prog_array 2890 * @index: the index of the program to update 2891 * @prog: the program to insert into the array 2892 * 2893 * Skips over dummy programs, by not counting them, when calculating 2894 * the position of the program to update. 2895 * 2896 * Return: 2897 * * 0 - Success 2898 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2899 * * -ENOENT - Index out of range 2900 */ 2901 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2902 struct bpf_prog *prog) 2903 { 2904 struct bpf_prog_array_item *item; 2905 2906 if (unlikely(index < 0)) 2907 return -EINVAL; 2908 2909 for (item = array->items; item->prog; item++) { 2910 if (item->prog == &dummy_bpf_prog.prog) 2911 continue; 2912 if (!index) { 2913 WRITE_ONCE(item->prog, prog); 2914 return 0; 2915 } 2916 index--; 2917 } 2918 return -ENOENT; 2919 } 2920 2921 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2922 struct bpf_prog *exclude_prog, 2923 struct bpf_prog *include_prog, 2924 u64 bpf_cookie, 2925 struct bpf_prog_array **new_array) 2926 { 2927 int new_prog_cnt, carry_prog_cnt = 0; 2928 struct bpf_prog_array_item *existing, *new; 2929 struct bpf_prog_array *array; 2930 bool found_exclude = false; 2931 2932 /* Figure out how many existing progs we need to carry over to 2933 * the new array. 2934 */ 2935 if (old_array) { 2936 existing = old_array->items; 2937 for (; existing->prog; existing++) { 2938 if (existing->prog == exclude_prog) { 2939 found_exclude = true; 2940 continue; 2941 } 2942 if (existing->prog != &dummy_bpf_prog.prog) 2943 carry_prog_cnt++; 2944 if (existing->prog == include_prog) 2945 return -EEXIST; 2946 } 2947 } 2948 2949 if (exclude_prog && !found_exclude) 2950 return -ENOENT; 2951 2952 /* How many progs (not NULL) will be in the new array? */ 2953 new_prog_cnt = carry_prog_cnt; 2954 if (include_prog) 2955 new_prog_cnt += 1; 2956 2957 /* Do we have any prog (not NULL) in the new array? */ 2958 if (!new_prog_cnt) { 2959 *new_array = NULL; 2960 return 0; 2961 } 2962 2963 /* +1 as the end of prog_array is marked with NULL */ 2964 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2965 if (!array) 2966 return -ENOMEM; 2967 new = array->items; 2968 2969 /* Fill in the new prog array */ 2970 if (carry_prog_cnt) { 2971 existing = old_array->items; 2972 for (; existing->prog; existing++) { 2973 if (existing->prog == exclude_prog || 2974 existing->prog == &dummy_bpf_prog.prog) 2975 continue; 2976 2977 new->prog = existing->prog; 2978 new->bpf_cookie = existing->bpf_cookie; 2979 new++; 2980 } 2981 } 2982 if (include_prog) { 2983 new->prog = include_prog; 2984 new->bpf_cookie = bpf_cookie; 2985 new++; 2986 } 2987 new->prog = NULL; 2988 *new_array = array; 2989 return 0; 2990 } 2991 2992 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2993 u32 *prog_ids, u32 request_cnt, 2994 u32 *prog_cnt) 2995 { 2996 u32 cnt = 0; 2997 2998 if (array) 2999 cnt = bpf_prog_array_length(array); 3000 3001 *prog_cnt = cnt; 3002 3003 /* return early if user requested only program count or nothing to copy */ 3004 if (!request_cnt || !cnt) 3005 return 0; 3006 3007 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 3008 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 3009 : 0; 3010 } 3011 3012 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 3013 struct bpf_map **used_maps, u32 len) 3014 { 3015 struct bpf_map *map; 3016 bool sleepable; 3017 u32 i; 3018 3019 sleepable = aux->prog->sleepable; 3020 for (i = 0; i < len; i++) { 3021 map = used_maps[i]; 3022 if (map->ops->map_poke_untrack) 3023 map->ops->map_poke_untrack(map, aux); 3024 if (sleepable) 3025 atomic64_dec(&map->sleepable_refcnt); 3026 bpf_map_put(map); 3027 } 3028 } 3029 3030 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 3031 { 3032 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 3033 kfree(aux->used_maps); 3034 } 3035 3036 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 3037 { 3038 #ifdef CONFIG_BPF_SYSCALL 3039 struct btf_mod_pair *btf_mod; 3040 u32 i; 3041 3042 for (i = 0; i < len; i++) { 3043 btf_mod = &used_btfs[i]; 3044 if (btf_mod->module) 3045 module_put(btf_mod->module); 3046 btf_put(btf_mod->btf); 3047 } 3048 #endif 3049 } 3050 3051 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 3052 { 3053 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 3054 kfree(aux->used_btfs); 3055 } 3056 3057 static void bpf_prog_free_deferred(struct work_struct *work) 3058 { 3059 struct bpf_prog_aux *aux; 3060 int i; 3061 3062 aux = container_of(work, struct bpf_prog_aux, work); 3063 #ifdef CONFIG_BPF_SYSCALL 3064 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 3065 bpf_prog_stream_free(aux->prog); 3066 #endif 3067 #ifdef CONFIG_CGROUP_BPF 3068 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 3069 bpf_cgroup_atype_put(aux->cgroup_atype); 3070 #endif 3071 bpf_free_used_maps(aux); 3072 bpf_free_used_btfs(aux); 3073 bpf_prog_disassoc_struct_ops(aux->prog); 3074 if (bpf_prog_is_dev_bound(aux)) 3075 bpf_prog_dev_bound_destroy(aux->prog); 3076 #ifdef CONFIG_PERF_EVENTS 3077 if (aux->prog->has_callchain_buf) 3078 put_callchain_buffers(); 3079 #endif 3080 if (aux->dst_trampoline) 3081 bpf_trampoline_put(aux->dst_trampoline); 3082 for (i = 0; i < aux->real_func_cnt; i++) { 3083 /* We can just unlink the subprog poke descriptor table as 3084 * it was originally linked to the main program and is also 3085 * released along with it. 3086 */ 3087 aux->func[i]->aux->poke_tab = NULL; 3088 bpf_jit_free(aux->func[i]); 3089 } 3090 if (aux->real_func_cnt) { 3091 kfree(aux->func); 3092 bpf_prog_unlock_free(aux->prog); 3093 } else { 3094 bpf_jit_free(aux->prog); 3095 } 3096 } 3097 3098 void bpf_prog_free(struct bpf_prog *fp) 3099 { 3100 struct bpf_prog_aux *aux = fp->aux; 3101 3102 if (aux->dst_prog) 3103 bpf_prog_put(aux->dst_prog); 3104 bpf_token_put(aux->token); 3105 INIT_WORK(&aux->work, bpf_prog_free_deferred); 3106 schedule_work(&aux->work); 3107 } 3108 EXPORT_SYMBOL_GPL(bpf_prog_free); 3109 3110 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 3111 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 3112 3113 void bpf_user_rnd_init_once(void) 3114 { 3115 prandom_init_once(&bpf_user_rnd_state); 3116 } 3117 3118 BPF_CALL_0(bpf_user_rnd_u32) 3119 { 3120 /* Should someone ever have the rather unwise idea to use some 3121 * of the registers passed into this function, then note that 3122 * this function is called from native eBPF and classic-to-eBPF 3123 * transformations. Register assignments from both sides are 3124 * different, f.e. classic always sets fn(ctx, A, X) here. 3125 */ 3126 struct rnd_state *state; 3127 u32 res; 3128 3129 state = &get_cpu_var(bpf_user_rnd_state); 3130 res = prandom_u32_state(state); 3131 put_cpu_var(bpf_user_rnd_state); 3132 3133 return res; 3134 } 3135 3136 BPF_CALL_0(bpf_get_raw_cpu_id) 3137 { 3138 return raw_smp_processor_id(); 3139 } 3140 3141 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 3142 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 3143 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 3144 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 3145 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 3146 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 3147 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 3148 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 3149 const struct bpf_func_proto bpf_spin_lock_proto __weak; 3150 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 3151 const struct bpf_func_proto bpf_jiffies64_proto __weak; 3152 3153 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3154 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3155 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3156 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3157 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3158 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3159 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3160 3161 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3162 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3163 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3164 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3165 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3166 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3167 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3168 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3169 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3170 const struct bpf_func_proto bpf_set_retval_proto __weak; 3171 const struct bpf_func_proto bpf_get_retval_proto __weak; 3172 3173 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3174 { 3175 return NULL; 3176 } 3177 3178 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3179 { 3180 return NULL; 3181 } 3182 3183 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3184 { 3185 return NULL; 3186 } 3187 3188 u64 __weak 3189 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3190 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3191 { 3192 return -ENOTSUPP; 3193 } 3194 EXPORT_SYMBOL_GPL(bpf_event_output); 3195 3196 /* Always built-in helper functions. */ 3197 const struct bpf_func_proto bpf_tail_call_proto = { 3198 /* func is unused for tail_call, we set it to pass the 3199 * get_helper_proto check 3200 */ 3201 .func = BPF_PTR_POISON, 3202 .gpl_only = false, 3203 .ret_type = RET_VOID, 3204 .arg1_type = ARG_PTR_TO_CTX, 3205 .arg2_type = ARG_CONST_MAP_PTR, 3206 .arg3_type = ARG_ANYTHING, 3207 }; 3208 3209 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3210 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3211 * eBPF and implicitly also cBPF can get JITed! 3212 */ 3213 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 3214 { 3215 return prog; 3216 } 3217 3218 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3219 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3220 */ 3221 void __weak bpf_jit_compile(struct bpf_prog *prog) 3222 { 3223 } 3224 3225 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3226 { 3227 return false; 3228 } 3229 3230 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3231 * analysis code and wants explicit zero extension inserted by verifier. 3232 * Otherwise, return FALSE. 3233 * 3234 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3235 * you don't override this. JITs that don't want these extra insns can detect 3236 * them using insn_is_zext. 3237 */ 3238 bool __weak bpf_jit_needs_zext(void) 3239 { 3240 return false; 3241 } 3242 3243 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3244 * all archs. The value returned must not change at runtime as there is 3245 * currently no support for reloading programs that were loaded without 3246 * mitigations. 3247 */ 3248 bool __weak bpf_jit_bypass_spec_v1(void) 3249 { 3250 return false; 3251 } 3252 3253 bool __weak bpf_jit_bypass_spec_v4(void) 3254 { 3255 return false; 3256 } 3257 3258 /* Return true if the JIT inlines the call to the helper corresponding to 3259 * the imm. 3260 * 3261 * The verifier will not patch the insn->imm for the call to the helper if 3262 * this returns true. 3263 */ 3264 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3265 { 3266 return false; 3267 } 3268 3269 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3270 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3271 { 3272 return false; 3273 } 3274 3275 bool __weak bpf_jit_supports_percpu_insn(void) 3276 { 3277 return false; 3278 } 3279 3280 bool __weak bpf_jit_supports_kfunc_call(void) 3281 { 3282 return false; 3283 } 3284 3285 bool __weak bpf_jit_supports_stack_args(void) 3286 { 3287 return false; 3288 } 3289 3290 bool __weak bpf_jit_supports_arena_args(void) 3291 { 3292 return false; 3293 } 3294 3295 bool __weak bpf_jit_supports_far_kfunc_call(void) 3296 { 3297 return false; 3298 } 3299 3300 bool __weak bpf_jit_supports_arena(void) 3301 { 3302 return false; 3303 } 3304 3305 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3306 { 3307 return false; 3308 } 3309 3310 bool __weak bpf_jit_supports_fsession(void) 3311 { 3312 return false; 3313 } 3314 3315 u64 __weak bpf_arch_uaddress_limit(void) 3316 { 3317 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3318 return TASK_SIZE; 3319 #else 3320 return 0; 3321 #endif 3322 } 3323 3324 /* Return TRUE if the JIT backend satisfies the following two conditions: 3325 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3326 * 2) Under the specific arch, the implementation of xchg() is the same 3327 * as atomic_xchg() on pointer-sized words. 3328 */ 3329 bool __weak bpf_jit_supports_ptr_xchg(void) 3330 { 3331 return false; 3332 } 3333 3334 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3335 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3336 */ 3337 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3338 int len) 3339 { 3340 return -EFAULT; 3341 } 3342 3343 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3344 enum bpf_text_poke_type new_t, void *old_addr, 3345 void *new_addr) 3346 { 3347 return -ENOTSUPP; 3348 } 3349 3350 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3351 { 3352 return ERR_PTR(-ENOTSUPP); 3353 } 3354 3355 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3356 { 3357 return -ENOTSUPP; 3358 } 3359 3360 bool __weak bpf_jit_supports_exceptions(void) 3361 { 3362 return false; 3363 } 3364 3365 bool __weak bpf_jit_supports_private_stack(void) 3366 { 3367 return false; 3368 } 3369 3370 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3371 { 3372 } 3373 3374 bool __weak bpf_jit_supports_timed_may_goto(void) 3375 { 3376 return false; 3377 } 3378 3379 u64 __weak arch_bpf_timed_may_goto(void) 3380 { 3381 return 0; 3382 } 3383 3384 static noinline void bpf_prog_report_may_goto_violation(void) 3385 { 3386 #ifdef CONFIG_BPF_SYSCALL 3387 struct bpf_stream_stage ss; 3388 struct bpf_prog *prog; 3389 3390 prog = bpf_prog_find_from_stack(); 3391 if (!prog) 3392 return; 3393 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3394 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3395 bpf_stream_dump_stack(ss); 3396 })); 3397 #endif 3398 } 3399 3400 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3401 { 3402 u64 time = ktime_get_mono_fast_ns(); 3403 3404 /* Populate the timestamp for this stack frame, and refresh count. */ 3405 if (!p->timestamp) { 3406 p->timestamp = time; 3407 return BPF_MAX_TIMED_LOOPS; 3408 } 3409 /* Check if we've exhausted our time slice, and zero count. */ 3410 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3411 bpf_prog_report_may_goto_violation(); 3412 return 0; 3413 } 3414 /* Refresh the count for the stack frame. */ 3415 return BPF_MAX_TIMED_LOOPS; 3416 } 3417 3418 /* for configs without MMU or 32-bit */ 3419 __weak const struct bpf_map_ops arena_map_ops; 3420 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3421 { 3422 return 0; 3423 } 3424 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3425 { 3426 return 0; 3427 } 3428 3429 #ifdef CONFIG_BPF_SYSCALL 3430 __weak bool bpf_arena_handle_page_fault(unsigned long addr, bool is_write, 3431 unsigned long fault_ip) 3432 { 3433 return false; 3434 } 3435 3436 static int __init bpf_global_ma_init(void) 3437 { 3438 int ret; 3439 3440 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3441 bpf_global_ma_set = !ret; 3442 return ret; 3443 } 3444 late_initcall(bpf_global_ma_init); 3445 #endif 3446 3447 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3448 EXPORT_SYMBOL(bpf_stats_enabled_key); 3449 3450 /* All definitions of tracepoints related to BPF. */ 3451 #define CREATE_TRACE_POINTS 3452 #include <linux/bpf_trace.h> 3453 3454 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3455 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3456 3457 #ifdef CONFIG_BPF_SYSCALL 3458 3459 void bpf_get_linfo_source(struct btf *btf, const struct bpf_line_info *linfo, 3460 struct bpf_linfo_source *src) 3461 { 3462 src->file = kbasename(btf_name_by_offset(btf, linfo->file_name_off)); 3463 src->line = btf_name_by_offset(btf, linfo->line_off); 3464 src->file_name_off = linfo->file_name_off; 3465 src->line_num = BPF_LINE_INFO_LINE_NUM(linfo->line_col); 3466 src->line_col = BPF_LINE_INFO_LINE_COL(linfo->line_col); 3467 } 3468 3469 const struct bpf_line_info *bpf_find_linfo(const struct bpf_prog *prog, u32 insn_off) 3470 { 3471 const struct bpf_line_info *linfo; 3472 u32 nr_linfo; 3473 int l, r, m; 3474 3475 nr_linfo = prog->aux->nr_linfo; 3476 if (!nr_linfo || insn_off >= prog->len) 3477 return NULL; 3478 3479 linfo = prog->aux->linfo; 3480 /* Loop invariant: linfo[l].insn_off <= insns_off. 3481 * linfo[0].insn_off == 0 which always satisfies above condition. 3482 * Binary search is searching for rightmost linfo entry that satisfies 3483 * the above invariant, giving us the desired record that covers given 3484 * instruction offset. 3485 */ 3486 l = 0; 3487 r = nr_linfo - 1; 3488 while (l < r) { 3489 /* (r - l + 1) / 2 means we break a tie to the right, so if: 3490 * l=1, r=2, linfo[l].insn_off <= insn_off, linfo[r].insn_off > insn_off, 3491 * then m=2, we see that linfo[m].insn_off > insn_off, and so 3492 * r becomes 1 and we exit the loop with correct l==1. 3493 * If the tie was broken to the left, m=1 would end us up in 3494 * an endless loop where l and m stay at 1 and r stays at 2. 3495 */ 3496 m = l + (r - l + 1) / 2; 3497 if (linfo[m].insn_off <= insn_off) 3498 l = m; 3499 else 3500 r = m - 1; 3501 } 3502 3503 return &linfo[l]; 3504 } 3505 3506 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3507 const char **linep, int *nump) 3508 { 3509 struct bpf_linfo_source src; 3510 int idx = -1, insn_start, insn_end, len; 3511 struct bpf_line_info *linfo; 3512 void **jited_linfo; 3513 struct btf *btf; 3514 int nr_linfo; 3515 3516 btf = prog->aux->btf; 3517 linfo = prog->aux->linfo; 3518 jited_linfo = prog->aux->jited_linfo; 3519 3520 if (!btf || !linfo || !jited_linfo) 3521 return -EINVAL; 3522 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3523 3524 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3525 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3526 3527 insn_start = linfo[0].insn_off; 3528 insn_end = insn_start + len; 3529 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3530 3531 for (int i = 0; i < nr_linfo && 3532 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3533 if (jited_linfo[i] >= (void *)ip) 3534 break; 3535 idx = i; 3536 } 3537 3538 if (idx == -1) 3539 return -ENOENT; 3540 3541 bpf_get_linfo_source(btf, &linfo[idx], &src); 3542 while (isspace(*src.line)) 3543 src.line++; 3544 if (filep) 3545 *filep = src.file; 3546 if (linep) 3547 *linep = src.line; 3548 if (nump) 3549 *nump = src.line_num; 3550 return 0; 3551 } 3552 3553 struct walk_stack_ctx { 3554 struct bpf_prog *prog; 3555 }; 3556 3557 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3558 { 3559 struct walk_stack_ctx *ctxp = cookie; 3560 struct bpf_prog *prog; 3561 3562 /* 3563 * The RCU read lock is held to safely traverse the latch tree, but we 3564 * don't need its protection when accessing the prog, since it has an 3565 * active stack frame on the current stack trace, and won't disappear. 3566 */ 3567 rcu_read_lock(); 3568 prog = bpf_prog_ksym_find(ip); 3569 rcu_read_unlock(); 3570 if (!prog) 3571 return true; 3572 /* Make sure we return the main prog if we found a subprog */ 3573 ctxp->prog = prog->aux->main_prog_aux->prog; 3574 return false; 3575 } 3576 3577 struct bpf_prog *bpf_prog_find_from_stack(void) 3578 { 3579 struct walk_stack_ctx ctx = {}; 3580 3581 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3582 return ctx.prog; 3583 } 3584 3585 #endif 3586