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/sched/signal.h> 23 #include <linux/skbuff.h> 24 #include <linux/static_call.h> 25 #include <linux/vmalloc.h> 26 #include <linux/prandom.h> 27 #include <linux/bpf.h> 28 #include <linux/btf.h> 29 #include <linux/hex.h> 30 #include <linux/objtool.h> 31 #include <linux/overflow.h> 32 #include <linux/rbtree_latch.h> 33 #include <linux/kallsyms.h> 34 #include <linux/rcupdate.h> 35 #include <linux/perf_event.h> 36 #include <linux/extable.h> 37 #include <linux/log2.h> 38 #include <linux/bpf_verifier.h> 39 #include <linux/nodemask.h> 40 #include <linux/nospec.h> 41 #include <linux/bpf_mem_alloc.h> 42 #include <linux/memcontrol.h> 43 #include <linux/execmem.h> 44 #include <crypto/sha2.h> 45 46 #include <asm/barrier.h> 47 #include <linux/unaligned.h> 48 49 /* Registers */ 50 #define BPF_R0 regs[BPF_REG_0] 51 #define BPF_R1 regs[BPF_REG_1] 52 #define BPF_R2 regs[BPF_REG_2] 53 #define BPF_R3 regs[BPF_REG_3] 54 #define BPF_R4 regs[BPF_REG_4] 55 #define BPF_R5 regs[BPF_REG_5] 56 #define BPF_R6 regs[BPF_REG_6] 57 #define BPF_R7 regs[BPF_REG_7] 58 #define BPF_R8 regs[BPF_REG_8] 59 #define BPF_R9 regs[BPF_REG_9] 60 #define BPF_R10 regs[BPF_REG_10] 61 62 /* Named registers */ 63 #define DST regs[insn->dst_reg] 64 #define SRC regs[insn->src_reg] 65 #define FP regs[BPF_REG_FP] 66 #define AX regs[BPF_REG_AX] 67 #define ARG1 regs[BPF_REG_ARG1] 68 #define CTX regs[BPF_REG_CTX] 69 #define OFF insn->off 70 #define IMM insn->imm 71 72 struct bpf_mem_alloc bpf_global_ma; 73 bool bpf_global_ma_set; 74 75 /* No hurry in this branch 76 * 77 * Exported for the bpf jit load helper. 78 */ 79 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 80 { 81 u8 *ptr = NULL; 82 83 if (k >= SKF_NET_OFF) { 84 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 85 } else if (k >= SKF_LL_OFF) { 86 if (unlikely(!skb_mac_header_was_set(skb))) 87 return NULL; 88 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 89 } 90 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 91 return ptr; 92 93 return NULL; 94 } 95 96 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 97 enum page_size_enum { 98 __PAGE_SIZE = PAGE_SIZE 99 }; 100 101 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 102 { 103 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 104 struct bpf_prog_aux *aux; 105 struct bpf_prog *fp; 106 107 size = round_up(size, __PAGE_SIZE); 108 fp = __vmalloc(size, gfp_flags); 109 if (fp == NULL) 110 return NULL; 111 112 aux = kzalloc_obj(*aux, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 113 if (aux == NULL) { 114 vfree(fp); 115 return NULL; 116 } 117 fp->active = __alloc_percpu_gfp(sizeof(u8[BPF_NR_CONTEXTS]), 4, 118 bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 119 if (!fp->active) { 120 vfree(fp); 121 kfree(aux); 122 return NULL; 123 } 124 125 fp->pages = size / PAGE_SIZE; 126 fp->aux = aux; 127 fp->aux->main_prog_aux = aux; 128 fp->aux->prog = fp; 129 fp->jit_requested = ebpf_jit_enabled(); 130 fp->jit_required = IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON); 131 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 132 #ifdef CONFIG_CGROUP_BPF 133 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 134 #endif 135 136 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 137 #ifdef CONFIG_FINEIBT 138 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 139 #endif 140 mutex_init(&fp->aux->used_maps_mutex); 141 mutex_init(&fp->aux->ext_mutex); 142 mutex_init(&fp->aux->dst_mutex); 143 mutex_init(&fp->aux->st_ops_assoc_mutex); 144 145 #ifdef CONFIG_BPF_SYSCALL 146 bpf_prog_stream_init(fp); 147 #endif 148 149 return fp; 150 } 151 152 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 153 { 154 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 155 struct bpf_prog *prog; 156 int cpu; 157 158 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 159 if (!prog) 160 return NULL; 161 162 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 163 if (!prog->stats) { 164 free_percpu(prog->active); 165 kfree(prog->aux); 166 vfree(prog); 167 return NULL; 168 } 169 170 for_each_possible_cpu(cpu) { 171 struct bpf_prog_stats *pstats; 172 173 pstats = per_cpu_ptr(prog->stats, cpu); 174 u64_stats_init(&pstats->syncp); 175 } 176 return prog; 177 } 178 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 179 180 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 181 { 182 if (!prog->aux->nr_linfo || !prog->jit_requested) 183 return 0; 184 185 prog->aux->jited_linfo = kvzalloc_objs(*prog->aux->jited_linfo, 186 prog->aux->nr_linfo, 187 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 188 if (!prog->aux->jited_linfo) 189 return -ENOMEM; 190 191 return 0; 192 } 193 194 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 195 { 196 if (prog->aux->jited_linfo && 197 (!prog->jited || !prog->aux->jited_linfo[0])) { 198 kvfree(prog->aux->jited_linfo); 199 prog->aux->jited_linfo = NULL; 200 } 201 202 kfree(prog->aux->kfunc_tab); 203 prog->aux->kfunc_tab = NULL; 204 } 205 206 /* The jit engine is responsible to provide an array 207 * for insn_off to the jited_off mapping (insn_to_jit_off). 208 * 209 * The idx to this array is the insn_off. Hence, the insn_off 210 * here is relative to the prog itself instead of the main prog. 211 * This array has one entry for each xlated bpf insn. 212 * 213 * jited_off is the byte off to the end of the jited insn. 214 * 215 * Hence, with 216 * insn_start: 217 * The first bpf insn off of the prog. The insn off 218 * here is relative to the main prog. 219 * e.g. if prog is a subprog, insn_start > 0 220 * linfo_idx: 221 * The prog's idx to prog->aux->linfo and jited_linfo 222 * 223 * jited_linfo[linfo_idx] = prog->bpf_func 224 * 225 * For i > linfo_idx, 226 * 227 * jited_linfo[i] = prog->bpf_func + 228 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 229 */ 230 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 231 const u32 *insn_to_jit_off) 232 { 233 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 234 const struct bpf_line_info *linfo; 235 void **jited_linfo; 236 237 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 238 /* Userspace did not provide linfo */ 239 return; 240 241 linfo_idx = prog->aux->linfo_idx; 242 linfo = &prog->aux->linfo[linfo_idx]; 243 insn_start = linfo[0].insn_off; 244 insn_end = insn_start + prog->len; 245 246 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 247 jited_linfo[0] = prog->bpf_func; 248 249 nr_linfo = prog->aux->nr_linfo - linfo_idx; 250 251 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 252 /* The verifier ensures that linfo[i].insn_off is 253 * strictly increasing 254 */ 255 jited_linfo[i] = prog->bpf_func + 256 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 257 } 258 259 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 260 gfp_t gfp_extra_flags) 261 { 262 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 263 struct bpf_prog *fp; 264 u32 pages; 265 266 size = round_up(size, PAGE_SIZE); 267 pages = size / PAGE_SIZE; 268 if (pages <= fp_old->pages) 269 return fp_old; 270 271 fp = __vmalloc(size, gfp_flags); 272 if (fp) { 273 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 274 fp->pages = pages; 275 fp->aux->prog = fp; 276 277 /* We keep fp->aux from fp_old around in the new 278 * reallocated structure. 279 */ 280 fp_old->aux = NULL; 281 fp_old->stats = NULL; 282 fp_old->active = NULL; 283 __bpf_prog_free(fp_old); 284 } 285 286 return fp; 287 } 288 289 void __bpf_prog_free(struct bpf_prog *fp) 290 { 291 if (fp->aux) { 292 mutex_destroy(&fp->aux->used_maps_mutex); 293 mutex_destroy(&fp->aux->dst_mutex); 294 mutex_destroy(&fp->aux->st_ops_assoc_mutex); 295 kfree(fp->aux->poke_tab); 296 kfree(fp->aux); 297 } 298 free_percpu(fp->stats); 299 free_percpu(fp->active); 300 vfree(fp); 301 } 302 303 int bpf_prog_calc_tag(struct bpf_prog *fp) 304 { 305 size_t size = bpf_prog_insn_size(fp); 306 struct bpf_insn *dst; 307 bool was_ld_map; 308 u32 i; 309 310 dst = __vmalloc(size, GFP_KERNEL_ACCOUNT); 311 if (!dst) 312 return -ENOMEM; 313 314 /* We need to take out the map fd for the digest calculation 315 * since they are unstable from user space side. 316 */ 317 for (i = 0, was_ld_map = false; i < fp->len; i++) { 318 dst[i] = fp->insnsi[i]; 319 if (!was_ld_map && 320 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 321 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 322 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 323 was_ld_map = true; 324 dst[i].imm = 0; 325 } else if (was_ld_map && 326 dst[i].code == 0 && 327 dst[i].dst_reg == 0 && 328 dst[i].src_reg == 0 && 329 dst[i].off == 0) { 330 was_ld_map = false; 331 dst[i].imm = 0; 332 } else { 333 was_ld_map = false; 334 } 335 } 336 sha256((u8 *)dst, size, fp->digest); 337 vfree(dst); 338 return 0; 339 } 340 341 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 342 s32 end_new, s32 curr, const bool probe_pass) 343 { 344 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 345 s32 delta = end_new - end_old; 346 s64 imm = insn->imm; 347 348 if (curr < pos && curr + imm + 1 >= end_old) 349 imm += delta; 350 else if (curr >= end_new && curr + imm + 1 < end_new) 351 imm -= delta; 352 if (imm < imm_min || imm > imm_max) 353 return -ERANGE; 354 if (!probe_pass) 355 insn->imm = imm; 356 return 0; 357 } 358 359 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 360 s32 end_new, s32 curr, const bool probe_pass) 361 { 362 s64 off_min, off_max, off; 363 s32 delta = end_new - end_old; 364 365 if (insn->code == (BPF_JMP32 | BPF_JA)) { 366 off = insn->imm; 367 off_min = S32_MIN; 368 off_max = S32_MAX; 369 } else { 370 off = insn->off; 371 off_min = S16_MIN; 372 off_max = S16_MAX; 373 } 374 375 if (curr < pos && curr + off + 1 >= end_old) 376 off += delta; 377 else if (curr >= end_new && curr + off + 1 < end_new) 378 off -= delta; 379 if (off < off_min || off > off_max) 380 return -ERANGE; 381 if (!probe_pass) { 382 if (insn->code == (BPF_JMP32 | BPF_JA)) 383 insn->imm = off; 384 else 385 insn->off = off; 386 } 387 return 0; 388 } 389 390 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 391 s32 end_new, const bool probe_pass) 392 { 393 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 394 struct bpf_insn *insn = prog->insnsi; 395 int ret = 0; 396 397 for (i = 0; i < insn_cnt; i++, insn++) { 398 u8 code; 399 400 /* In the probing pass we still operate on the original, 401 * unpatched image in order to check overflows before we 402 * do any other adjustments. Therefore skip the patchlet. 403 */ 404 if (probe_pass && i == pos) { 405 i = end_new; 406 insn = prog->insnsi + end_old; 407 } 408 if (bpf_pseudo_func(insn)) { 409 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 410 end_new, i, probe_pass); 411 if (ret) 412 return ret; 413 continue; 414 } 415 code = insn->code; 416 if ((BPF_CLASS(code) != BPF_JMP && 417 BPF_CLASS(code) != BPF_JMP32) || 418 BPF_OP(code) == BPF_EXIT) 419 continue; 420 /* Adjust offset of jmps if we cross patch boundaries. */ 421 if (BPF_OP(code) == BPF_CALL) { 422 if (insn->src_reg != BPF_PSEUDO_CALL) 423 continue; 424 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 425 end_new, i, probe_pass); 426 } else { 427 ret = bpf_adj_delta_to_off(insn, pos, end_old, 428 end_new, i, probe_pass); 429 } 430 if (ret) 431 break; 432 } 433 434 return ret; 435 } 436 437 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 438 { 439 struct bpf_line_info *linfo; 440 u32 i, nr_linfo; 441 442 nr_linfo = prog->aux->nr_linfo; 443 if (!nr_linfo || !delta) 444 return; 445 446 linfo = prog->aux->linfo; 447 448 for (i = 0; i < nr_linfo; i++) 449 if (off < linfo[i].insn_off) 450 break; 451 452 /* Push all off < linfo[i].insn_off by delta */ 453 for (; i < nr_linfo; i++) 454 linfo[i].insn_off += delta; 455 } 456 457 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 458 const struct bpf_insn *patch, u32 len) 459 { 460 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 461 const u32 cnt_max = S16_MAX; 462 struct bpf_prog *prog_adj; 463 int err; 464 465 /* Since our patchlet doesn't expand the image, we're done. */ 466 if (insn_delta == 0) { 467 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 468 return prog; 469 } 470 471 insn_adj_cnt = prog->len + insn_delta; 472 473 /* Reject anything that would potentially let the insn->off 474 * target overflow when we have excessive program expansions. 475 * We need to probe here before we do any reallocation where 476 * we afterwards may not fail anymore. 477 */ 478 if (insn_adj_cnt > cnt_max && 479 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 480 return ERR_PTR(err); 481 482 /* Several new instructions need to be inserted. Make room 483 * for them. Likely, there's no need for a new allocation as 484 * last page could have large enough tailroom. 485 */ 486 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 487 GFP_USER); 488 if (!prog_adj) 489 return ERR_PTR(-ENOMEM); 490 491 prog_adj->len = insn_adj_cnt; 492 493 /* Patching happens in 3 steps: 494 * 495 * 1) Move over tail of insnsi from next instruction onwards, 496 * so we can patch the single target insn with one or more 497 * new ones (patching is always from 1 to n insns, n > 0). 498 * 2) Inject new instructions at the target location. 499 * 3) Adjust branch offsets if necessary. 500 */ 501 insn_rest = insn_adj_cnt - off - len; 502 503 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 504 sizeof(*patch) * insn_rest); 505 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 506 507 /* We are guaranteed to not fail at this point, otherwise 508 * the ship has sailed to reverse to the original state. An 509 * overflow cannot happen at this point. 510 */ 511 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 512 513 bpf_adj_linfo(prog_adj, off, insn_delta); 514 515 return prog_adj; 516 } 517 518 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 519 { 520 int err; 521 522 /* Branch offsets can't overflow when program is shrinking, no need 523 * to call bpf_adj_branches(..., true) here 524 */ 525 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 526 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 527 prog->len -= cnt; 528 529 err = bpf_adj_branches(prog, off, off + cnt, off, false); 530 WARN_ON_ONCE(err); 531 return err; 532 } 533 534 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 535 { 536 int i; 537 538 for (i = 0; i < fp->aux->real_func_cnt; i++) 539 bpf_prog_kallsyms_del(fp->aux->func[i]); 540 } 541 542 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 543 { 544 bpf_prog_kallsyms_del_subprogs(fp); 545 bpf_prog_kallsyms_del(fp); 546 } 547 548 #ifdef CONFIG_BPF_JIT 549 /* All BPF JIT sysctl knobs here. */ 550 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 551 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 552 int bpf_jit_harden __read_mostly; 553 long bpf_jit_limit __read_mostly; 554 long bpf_jit_limit_max __read_mostly; 555 556 static void 557 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 558 { 559 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 560 561 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 562 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 563 } 564 565 static void 566 bpf_prog_ksym_set_name(struct bpf_prog *prog) 567 { 568 char *sym = prog->aux->ksym.name; 569 const char *end = sym + KSYM_NAME_LEN; 570 const struct btf_type *type; 571 const char *func_name; 572 573 BUILD_BUG_ON(sizeof("bpf_prog_") + 574 sizeof(prog->tag) * 2 + 575 /* name has been null terminated. 576 * We should need +1 for the '_' preceding 577 * the name. However, the null character 578 * is double counted between the name and the 579 * sizeof("bpf_prog_") above, so we omit 580 * the +1 here. 581 */ 582 sizeof(prog->aux->name) > KSYM_NAME_LEN); 583 584 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 585 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 586 587 /* prog->aux->name will be ignored if full btf name is available */ 588 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 589 type = btf_type_by_id(prog->aux->btf, 590 prog->aux->func_info[prog->aux->func_idx].type_id); 591 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 592 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 593 return; 594 } 595 596 if (prog->aux->name[0]) 597 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 598 else 599 *sym = 0; 600 } 601 602 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 603 { 604 return container_of(n, struct bpf_ksym, tnode)->start; 605 } 606 607 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 608 struct latch_tree_node *b) 609 { 610 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 611 } 612 613 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 614 { 615 unsigned long val = (unsigned long)key; 616 const struct bpf_ksym *ksym; 617 618 ksym = container_of(n, struct bpf_ksym, tnode); 619 620 if (val < ksym->start) 621 return -1; 622 /* Ensure that we detect return addresses as part of the program, when 623 * the final instruction is a call for a program part of the stack 624 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 625 */ 626 if (val > ksym->end) 627 return 1; 628 629 return 0; 630 } 631 632 static const struct latch_tree_ops bpf_tree_ops = { 633 .less = bpf_tree_less, 634 .comp = bpf_tree_comp, 635 }; 636 637 static DEFINE_SPINLOCK(bpf_lock); 638 static LIST_HEAD(bpf_kallsyms); 639 static struct latch_tree_root bpf_tree __cacheline_aligned; 640 641 void bpf_ksym_add(struct bpf_ksym *ksym) 642 { 643 spin_lock_bh(&bpf_lock); 644 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 645 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 646 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 647 spin_unlock_bh(&bpf_lock); 648 } 649 650 static void __bpf_ksym_del(struct bpf_ksym *ksym) 651 { 652 if (list_empty(&ksym->lnode)) 653 return; 654 655 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 656 list_del_rcu(&ksym->lnode); 657 } 658 659 void bpf_ksym_del(struct bpf_ksym *ksym) 660 { 661 spin_lock_bh(&bpf_lock); 662 __bpf_ksym_del(ksym); 663 spin_unlock_bh(&bpf_lock); 664 } 665 666 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 667 { 668 return fp->jited && !bpf_prog_was_classic(fp); 669 } 670 671 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 672 { 673 if (!bpf_prog_kallsyms_candidate(fp) || 674 !bpf_token_capable(fp->aux->token, CAP_BPF)) 675 return; 676 677 bpf_prog_ksym_set_addr(fp); 678 bpf_prog_ksym_set_name(fp); 679 fp->aux->ksym.prog = true; 680 681 bpf_ksym_add(&fp->aux->ksym); 682 683 #ifdef CONFIG_FINEIBT 684 /* 685 * When FineIBT, code in the __cfi_foo() symbols can get executed 686 * and hence unwinder needs help. 687 */ 688 if (cfi_mode != CFI_FINEIBT) 689 return; 690 691 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 692 "__cfi_%s", fp->aux->ksym.name); 693 694 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 695 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 696 697 bpf_ksym_add(&fp->aux->ksym_prefix); 698 #endif 699 } 700 701 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 702 { 703 if (!bpf_prog_kallsyms_candidate(fp)) 704 return; 705 706 bpf_ksym_del(&fp->aux->ksym); 707 #ifdef CONFIG_FINEIBT 708 if (cfi_mode != CFI_FINEIBT) 709 return; 710 bpf_ksym_del(&fp->aux->ksym_prefix); 711 #endif 712 } 713 714 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 715 { 716 struct latch_tree_node *n; 717 718 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 719 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 720 } 721 722 int bpf_address_lookup(unsigned long addr, unsigned long *size, 723 unsigned long *off, char *sym) 724 { 725 struct bpf_ksym *ksym; 726 int ret = 0; 727 728 rcu_read_lock(); 729 ksym = bpf_ksym_find(addr); 730 if (ksym) { 731 unsigned long symbol_start = ksym->start; 732 unsigned long symbol_end = ksym->end; 733 734 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 735 736 if (size) 737 *size = symbol_end - symbol_start; 738 if (off) 739 *off = addr - symbol_start; 740 } 741 rcu_read_unlock(); 742 743 return ret; 744 } 745 746 bool is_bpf_text_address(unsigned long addr) 747 { 748 bool ret; 749 750 rcu_read_lock(); 751 ret = bpf_ksym_find(addr) != NULL; 752 rcu_read_unlock(); 753 754 return ret; 755 } 756 757 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 758 { 759 struct bpf_ksym *ksym; 760 761 WARN_ON_ONCE(!rcu_read_lock_held()); 762 ksym = bpf_ksym_find(addr); 763 764 return ksym && ksym->prog ? 765 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 766 NULL; 767 } 768 769 bool bpf_has_frame_pointer(unsigned long ip) 770 { 771 struct bpf_ksym *ksym; 772 unsigned long offset; 773 774 guard(rcu)(); 775 776 ksym = bpf_ksym_find(ip); 777 if (!ksym || !ksym->fp_start || !ksym->fp_end) 778 return false; 779 780 offset = ip - ksym->start; 781 782 return offset >= ksym->fp_start && offset < ksym->fp_end; 783 } 784 785 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 786 { 787 const struct exception_table_entry *e = NULL; 788 struct bpf_prog *prog; 789 790 rcu_read_lock(); 791 prog = bpf_prog_ksym_find(addr); 792 if (!prog) 793 goto out; 794 if (!prog->aux->num_exentries) 795 goto out; 796 797 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 798 out: 799 rcu_read_unlock(); 800 return e; 801 } 802 803 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 804 char *sym) 805 { 806 struct bpf_ksym *ksym; 807 unsigned int it = 0; 808 int ret = -ERANGE; 809 810 if (!bpf_jit_kallsyms_enabled()) 811 return ret; 812 813 rcu_read_lock(); 814 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 815 if (it++ != symnum) 816 continue; 817 818 strscpy(sym, ksym->name, KSYM_NAME_LEN); 819 820 *value = ksym->start; 821 *type = BPF_SYM_ELF_TYPE; 822 823 ret = 0; 824 break; 825 } 826 rcu_read_unlock(); 827 828 return ret; 829 } 830 831 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 832 struct bpf_jit_poke_descriptor *poke) 833 { 834 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 835 static const u32 poke_tab_max = 1024; 836 u32 slot = prog->aux->size_poke_tab; 837 u32 size = slot + 1; 838 839 if (size > poke_tab_max) 840 return -ENOSPC; 841 if (poke->tailcall_target || poke->tailcall_target_stable || 842 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 843 return -EINVAL; 844 845 switch (poke->reason) { 846 case BPF_POKE_REASON_TAIL_CALL: 847 if (!poke->tail_call.map) 848 return -EINVAL; 849 break; 850 default: 851 return -EINVAL; 852 } 853 854 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 855 if (!tab) 856 return -ENOMEM; 857 858 memcpy(&tab[slot], poke, sizeof(*poke)); 859 prog->aux->size_poke_tab = size; 860 prog->aux->poke_tab = tab; 861 862 return slot; 863 } 864 865 /* 866 * BPF program pack allocator. 867 * 868 * Most BPF programs are pretty small. Allocating a hole page for each 869 * program is sometime a waste. Many small bpf program also adds pressure 870 * to instruction TLB. To solve this issue, we introduce a BPF program pack 871 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 872 * to host BPF programs. 873 */ 874 #define BPF_PROG_CHUNK_SHIFT 6 875 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 876 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 877 878 struct bpf_prog_pack { 879 struct list_head list; 880 void *ptr; 881 bool arch_flush_needed; 882 unsigned long bitmap[]; 883 }; 884 885 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 886 { 887 memset(area, 0, size); 888 } 889 890 DEFINE_STATIC_CALL_NULL(bpf_arch_pred_flush, bpf_arch_pred_flush); 891 892 /* 893 * Enabled once bpf_arch_pred_flush points at a real flush routine. Lets the 894 * pack allocator test "is a predictor flush wired up at all" with a cheap 895 * static branch instead of repeatedly querying the static call target. 896 */ 897 DEFINE_STATIC_KEY_FALSE(bpf_pred_flush_enabled); 898 899 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 900 901 static DEFINE_MUTEX(pack_mutex); 902 static LIST_HEAD(pack_list); 903 904 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 905 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 906 */ 907 #ifdef PMD_SIZE 908 /* PMD_SIZE is really big for some archs. It doesn't make sense to 909 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 910 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 911 * greater than or equal to 2MB. 912 */ 913 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 914 #else 915 #define BPF_PROG_PACK_SIZE PAGE_SIZE 916 #endif 917 918 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 919 920 static bool bpf_jit_mem_is_rox(void) 921 { 922 return execmem_is_rox(EXECMEM_BPF); 923 } 924 925 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 926 { 927 struct bpf_prog_pack *pack; 928 int err; 929 930 pack = kzalloc_flex(*pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)); 931 if (!pack) 932 return NULL; 933 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 934 if (!pack->ptr) 935 goto out; 936 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 937 938 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 939 pack->arch_flush_needed = true; 940 if (!bpf_jit_mem_is_rox()) { 941 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 942 set_vm_flush_reset_perms(pack->ptr); 943 err = set_memory_rox((unsigned long)pack->ptr, 944 BPF_PROG_PACK_SIZE / PAGE_SIZE); 945 if (err) 946 goto out; 947 } 948 list_add_tail(&pack->list, &pack_list); 949 return pack; 950 951 out: 952 bpf_jit_free_exec(pack->ptr); 953 kfree(pack); 954 return NULL; 955 } 956 957 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns, bool was_classic) 958 { 959 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 960 struct bpf_prog_pack *pack, *fallback_pack = NULL; 961 unsigned long pos, fallback_pos = 0; 962 void *ptr = NULL; 963 964 mutex_lock(&pack_mutex); 965 if (size > BPF_PROG_PACK_SIZE) { 966 /* 967 * Allocations larger than a pack get their own pages, and 968 * predictors are not flushed for such allocation. This is only 969 * safe because cBPF programs (the unprivileged attack surface) 970 * are bounded well below a pack size. 971 */ 972 if (was_classic && static_branch_unlikely(&bpf_pred_flush_enabled)) 973 pr_warn_once("BPF: Predictors not flushed for allocations greater than BPF_PROG_PACK_SIZE\n"); 974 size = round_up(size, PAGE_SIZE); 975 ptr = bpf_jit_alloc_exec(size); 976 if (ptr && !bpf_jit_mem_is_rox()) { 977 int err; 978 979 bpf_fill_ill_insns(ptr, size); 980 set_vm_flush_reset_perms(ptr); 981 err = set_memory_rox((unsigned long)ptr, 982 size / PAGE_SIZE); 983 if (err) { 984 bpf_jit_free_exec(ptr); 985 ptr = NULL; 986 } 987 } 988 goto out; 989 } 990 list_for_each_entry(pack, &pack_list, list) { 991 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 992 nbits, 0); 993 if (pos >= BPF_PROG_CHUNK_COUNT) 994 continue; 995 /* Flush not enabled, use any pack */ 996 if (!static_branch_unlikely(&bpf_pred_flush_enabled)) 997 goto found_free_area; 998 /* 999 * cBPF reuse of a dirty pack triggers a flush, so prefer a 1000 * clean pack for cBPF. eBPF never flushes, so steer it to a 1001 * dirty pack and keep clean packs free for cBPF. 1002 */ 1003 if (was_classic ^ pack->arch_flush_needed) 1004 goto found_free_area; 1005 if (!fallback_pack) { 1006 fallback_pack = pack; 1007 fallback_pos = pos; 1008 } 1009 } 1010 1011 /* No preferred pack found */ 1012 if (fallback_pack) { 1013 pack = fallback_pack; 1014 pos = fallback_pos; 1015 goto found_free_area; 1016 } 1017 1018 pack = alloc_new_pack(bpf_fill_ill_insns); 1019 if (!pack) 1020 goto out; 1021 1022 pos = 0; 1023 1024 found_free_area: 1025 /* Flush only for cBPF as it may contain a crafted gadget */ 1026 if (static_branch_unlikely(&bpf_pred_flush_enabled) && 1027 pack->arch_flush_needed && 1028 was_classic) { 1029 struct bpf_prog_pack *p; 1030 1031 static_call_cond(bpf_arch_pred_flush)(); 1032 list_for_each_entry(p, &pack_list, list) 1033 p->arch_flush_needed = false; 1034 } 1035 bitmap_set(pack->bitmap, pos, nbits); 1036 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 1037 1038 out: 1039 mutex_unlock(&pack_mutex); 1040 return ptr; 1041 } 1042 1043 void bpf_prog_pack_free(void *ptr, u32 size) 1044 { 1045 struct bpf_prog_pack *pack = NULL, *tmp; 1046 unsigned int nbits; 1047 unsigned long pos; 1048 1049 mutex_lock(&pack_mutex); 1050 if (size > BPF_PROG_PACK_SIZE) { 1051 bpf_jit_free_exec(ptr); 1052 goto out; 1053 } 1054 1055 list_for_each_entry(tmp, &pack_list, list) { 1056 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 1057 pack = tmp; 1058 break; 1059 } 1060 } 1061 1062 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1063 goto out; 1064 1065 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1066 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1067 1068 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1069 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1070 1071 bitmap_clear(pack->bitmap, pos, nbits); 1072 1073 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 1074 pack->arch_flush_needed = true; 1075 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1076 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1077 list_del(&pack->list); 1078 bpf_jit_free_exec(pack->ptr); 1079 kfree(pack); 1080 } 1081 out: 1082 mutex_unlock(&pack_mutex); 1083 } 1084 1085 static atomic_long_t bpf_jit_current; 1086 1087 /* Can be overridden by an arch's JIT compiler if it has a custom, 1088 * dedicated BPF backend memory area, or if neither of the two 1089 * below apply. 1090 */ 1091 u64 __weak bpf_jit_alloc_exec_limit(void) 1092 { 1093 #if defined(MODULES_VADDR) 1094 return MODULES_END - MODULES_VADDR; 1095 #else 1096 return VMALLOC_END - VMALLOC_START; 1097 #endif 1098 } 1099 1100 static int __init bpf_jit_charge_init(void) 1101 { 1102 /* Only used as heuristic here to derive limit. */ 1103 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1104 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1105 PAGE_SIZE), LONG_MAX); 1106 return 0; 1107 } 1108 pure_initcall(bpf_jit_charge_init); 1109 1110 int bpf_jit_charge_modmem(u32 size) 1111 { 1112 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1113 if (!bpf_capable()) { 1114 atomic_long_sub(size, &bpf_jit_current); 1115 return -EPERM; 1116 } 1117 } 1118 1119 return 0; 1120 } 1121 1122 void bpf_jit_uncharge_modmem(u32 size) 1123 { 1124 atomic_long_sub(size, &bpf_jit_current); 1125 } 1126 1127 void *bpf_jit_alloc_exec(unsigned long size) 1128 { 1129 return execmem_alloc(EXECMEM_BPF, size); 1130 } 1131 1132 void bpf_jit_free_exec(void *addr) 1133 { 1134 execmem_free(addr); 1135 } 1136 1137 struct bpf_binary_header * 1138 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1139 unsigned int alignment, 1140 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1141 { 1142 struct bpf_binary_header *hdr; 1143 u32 size, hole, start; 1144 1145 WARN_ON_ONCE(!is_power_of_2(alignment) || 1146 alignment > BPF_IMAGE_ALIGNMENT); 1147 1148 /* Most of BPF filters are really small, but if some of them 1149 * fill a page, allow at least 128 extra bytes to insert a 1150 * random section of illegal instructions. 1151 */ 1152 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1153 1154 if (bpf_jit_charge_modmem(size)) 1155 return NULL; 1156 hdr = bpf_jit_alloc_exec(size); 1157 if (!hdr) { 1158 bpf_jit_uncharge_modmem(size); 1159 return NULL; 1160 } 1161 1162 /* Fill space with illegal/arch-dep instructions. */ 1163 bpf_fill_ill_insns(hdr, size); 1164 1165 hdr->size = size; 1166 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1167 PAGE_SIZE - sizeof(*hdr)); 1168 start = get_random_u32_below(hole) & ~(alignment - 1); 1169 1170 /* Leave a random number of instructions before BPF code. */ 1171 *image_ptr = &hdr->image[start]; 1172 1173 return hdr; 1174 } 1175 1176 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1177 { 1178 u32 size = hdr->size; 1179 1180 bpf_jit_free_exec(hdr); 1181 bpf_jit_uncharge_modmem(size); 1182 } 1183 1184 /* Allocate jit binary from bpf_prog_pack allocator. 1185 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1186 * to the memory. To solve this problem, a RW buffer is also allocated at 1187 * as the same time. The JIT engine should calculate offsets based on the 1188 * RO memory address, but write JITed program to the RW buffer. Once the 1189 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1190 * the JITed program to the RO memory. 1191 */ 1192 struct bpf_binary_header * 1193 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1194 unsigned int alignment, 1195 struct bpf_binary_header **rw_header, 1196 u8 **rw_image, 1197 bpf_jit_fill_hole_t bpf_fill_ill_insns, 1198 bool was_classic) 1199 { 1200 struct bpf_binary_header *ro_header; 1201 u32 size, hole, start; 1202 1203 WARN_ON_ONCE(!is_power_of_2(alignment) || 1204 alignment > BPF_IMAGE_ALIGNMENT); 1205 1206 /* add 16 bytes for a random section of illegal instructions */ 1207 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1208 1209 if (bpf_jit_charge_modmem(size)) 1210 return NULL; 1211 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns, was_classic); 1212 if (!ro_header) { 1213 bpf_jit_uncharge_modmem(size); 1214 return NULL; 1215 } 1216 1217 *rw_header = kvmalloc(size, GFP_KERNEL); 1218 if (!*rw_header) { 1219 bpf_prog_pack_free(ro_header, size); 1220 bpf_jit_uncharge_modmem(size); 1221 return NULL; 1222 } 1223 1224 /* Fill space with illegal/arch-dep instructions. */ 1225 bpf_fill_ill_insns(*rw_header, size); 1226 (*rw_header)->size = size; 1227 1228 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1229 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1230 start = get_random_u32_below(hole) & ~(alignment - 1); 1231 1232 *image_ptr = &ro_header->image[start]; 1233 *rw_image = &(*rw_header)->image[start]; 1234 1235 return ro_header; 1236 } 1237 1238 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1239 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1240 struct bpf_binary_header *rw_header) 1241 { 1242 void *ptr; 1243 1244 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1245 1246 kvfree(rw_header); 1247 1248 if (IS_ERR(ptr)) { 1249 bpf_prog_pack_free(ro_header, ro_header->size); 1250 return PTR_ERR(ptr); 1251 } 1252 return 0; 1253 } 1254 1255 /* bpf_jit_binary_pack_free is called in two different scenarios: 1256 * 1) when the program is freed after; 1257 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1258 * For case 2), we need to free both the RO memory and the RW buffer. 1259 * 1260 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1261 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1262 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1263 * bpf_arch_text_copy (when jit fails). 1264 */ 1265 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1266 struct bpf_binary_header *rw_header) 1267 { 1268 u32 size = ro_header->size; 1269 1270 bpf_prog_pack_free(ro_header, size); 1271 kvfree(rw_header); 1272 bpf_jit_uncharge_modmem(size); 1273 } 1274 1275 struct bpf_binary_header * 1276 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1277 { 1278 unsigned long real_start = (unsigned long)fp->bpf_func; 1279 unsigned long addr; 1280 1281 addr = real_start & BPF_PROG_CHUNK_MASK; 1282 return (void *)addr; 1283 } 1284 1285 static inline struct bpf_binary_header * 1286 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1287 { 1288 unsigned long real_start = (unsigned long)fp->bpf_func; 1289 unsigned long addr; 1290 1291 addr = real_start & PAGE_MASK; 1292 return (void *)addr; 1293 } 1294 1295 /* This symbol is only overridden by archs that have different 1296 * requirements than the usual eBPF JITs, f.e. when they only 1297 * implement cBPF JIT, do not set images read-only, etc. 1298 */ 1299 void __weak bpf_jit_free(struct bpf_prog *fp) 1300 { 1301 if (fp->jited) { 1302 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1303 1304 bpf_jit_binary_free(hdr); 1305 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1306 } 1307 1308 bpf_prog_unlock_free(fp); 1309 } 1310 1311 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1312 const struct bpf_insn *insn, bool extra_pass, 1313 u64 *func_addr, bool *func_addr_fixed) 1314 { 1315 s16 off = insn->off; 1316 s32 imm = insn->imm; 1317 u8 *addr; 1318 int err; 1319 1320 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1321 if (!*func_addr_fixed) { 1322 /* Place-holder address till the last pass has collected 1323 * all addresses for JITed subprograms in which case we 1324 * can pick them up from prog->aux. 1325 */ 1326 if (!extra_pass) 1327 addr = NULL; 1328 else if (prog->aux->func && 1329 off >= 0 && off < prog->aux->real_func_cnt) 1330 addr = (u8 *)prog->aux->func[off]->bpf_func; 1331 else 1332 return -EINVAL; 1333 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1334 bpf_jit_supports_far_kfunc_call()) { 1335 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1336 if (err) 1337 return err; 1338 } else { 1339 /* Address of a BPF helper call. Since part of the core 1340 * kernel, it's always at a fixed location. __bpf_call_base 1341 * and the helper with imm relative to it are both in core 1342 * kernel. 1343 */ 1344 addr = (u8 *)__bpf_call_base + imm; 1345 } 1346 1347 *func_addr = (unsigned long)addr; 1348 return 0; 1349 } 1350 1351 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1352 { 1353 if (prog->aux->ksym.prog) 1354 return prog->aux->ksym.name; 1355 return prog->aux->name; 1356 } 1357 1358 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1359 const struct bpf_insn *aux, 1360 struct bpf_insn *to_buff, 1361 bool emit_zext) 1362 { 1363 struct bpf_insn *to = to_buff; 1364 u32 imm_rnd = get_random_u32(); 1365 int off; 1366 1367 BUILD_BUG_ON(BPF_REG_PARAMS + 2 != MAX_BPF_JIT_REG); 1368 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1369 1370 /* Constraints on AX register: 1371 * 1372 * AX register is inaccessible from user space. It is mapped in 1373 * all JITs, and used here for constant blinding rewrites. It is 1374 * typically "stateless" meaning its contents are only valid within 1375 * the executed instruction, but not across several instructions. 1376 * There are a few exceptions however which are further detailed 1377 * below. 1378 * 1379 * Constant blinding is only used by JITs, not in the interpreter. 1380 * The interpreter uses AX in some occasions as a local temporary 1381 * register e.g. in DIV or MOD instructions. 1382 * 1383 * In restricted circumstances, the verifier can also use the AX 1384 * register for rewrites as long as they do not interfere with 1385 * the above cases! 1386 */ 1387 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1388 goto out; 1389 1390 if (from->imm == 0 && 1391 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1392 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1393 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1394 goto out; 1395 } 1396 1397 switch (from->code) { 1398 case BPF_ALU | BPF_ADD | BPF_K: 1399 case BPF_ALU | BPF_SUB | BPF_K: 1400 case BPF_ALU | BPF_AND | BPF_K: 1401 case BPF_ALU | BPF_OR | BPF_K: 1402 case BPF_ALU | BPF_XOR | BPF_K: 1403 case BPF_ALU | BPF_MUL | BPF_K: 1404 case BPF_ALU | BPF_MOV | BPF_K: 1405 case BPF_ALU | BPF_DIV | BPF_K: 1406 case BPF_ALU | BPF_MOD | BPF_K: 1407 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1408 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1409 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1410 break; 1411 1412 case BPF_ALU64 | BPF_ADD | BPF_K: 1413 case BPF_ALU64 | BPF_SUB | BPF_K: 1414 case BPF_ALU64 | BPF_AND | BPF_K: 1415 case BPF_ALU64 | BPF_OR | BPF_K: 1416 case BPF_ALU64 | BPF_XOR | BPF_K: 1417 case BPF_ALU64 | BPF_MUL | BPF_K: 1418 case BPF_ALU64 | BPF_MOV | BPF_K: 1419 case BPF_ALU64 | BPF_DIV | BPF_K: 1420 case BPF_ALU64 | BPF_MOD | BPF_K: 1421 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1422 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1423 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1424 break; 1425 1426 case BPF_JMP | BPF_JEQ | BPF_K: 1427 case BPF_JMP | BPF_JNE | BPF_K: 1428 case BPF_JMP | BPF_JGT | BPF_K: 1429 case BPF_JMP | BPF_JLT | BPF_K: 1430 case BPF_JMP | BPF_JGE | BPF_K: 1431 case BPF_JMP | BPF_JLE | BPF_K: 1432 case BPF_JMP | BPF_JSGT | BPF_K: 1433 case BPF_JMP | BPF_JSLT | BPF_K: 1434 case BPF_JMP | BPF_JSGE | BPF_K: 1435 case BPF_JMP | BPF_JSLE | BPF_K: 1436 case BPF_JMP | BPF_JSET | BPF_K: 1437 /* Accommodate for extra offset in case of a backjump. */ 1438 off = from->off; 1439 if (off < 0) 1440 off -= 2; 1441 if (off < S16_MIN) 1442 return -ERANGE; 1443 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1444 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1445 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1446 break; 1447 1448 case BPF_JMP32 | BPF_JEQ | BPF_K: 1449 case BPF_JMP32 | BPF_JNE | BPF_K: 1450 case BPF_JMP32 | BPF_JGT | BPF_K: 1451 case BPF_JMP32 | BPF_JLT | BPF_K: 1452 case BPF_JMP32 | BPF_JGE | BPF_K: 1453 case BPF_JMP32 | BPF_JLE | BPF_K: 1454 case BPF_JMP32 | BPF_JSGT | BPF_K: 1455 case BPF_JMP32 | BPF_JSLT | BPF_K: 1456 case BPF_JMP32 | BPF_JSGE | BPF_K: 1457 case BPF_JMP32 | BPF_JSLE | BPF_K: 1458 case BPF_JMP32 | BPF_JSET | BPF_K: 1459 /* Accommodate for extra offset in case of a backjump. */ 1460 off = from->off; 1461 if (off < 0) 1462 off -= 2; 1463 if (off < S16_MIN) 1464 return -ERANGE; 1465 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1466 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1467 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1468 off); 1469 break; 1470 1471 case BPF_LD | BPF_IMM | BPF_DW: 1472 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1473 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1474 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1475 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1476 break; 1477 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1478 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1479 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1480 if (emit_zext) 1481 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1482 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1483 break; 1484 1485 case BPF_ST | BPF_MEM | BPF_DW: 1486 case BPF_ST | BPF_MEM | BPF_W: 1487 case BPF_ST | BPF_MEM | BPF_H: 1488 case BPF_ST | BPF_MEM | BPF_B: 1489 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1490 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1491 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1492 break; 1493 1494 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW: 1495 case BPF_ST | BPF_PROBE_MEM32 | BPF_W: 1496 case BPF_ST | BPF_PROBE_MEM32 | BPF_H: 1497 case BPF_ST | BPF_PROBE_MEM32 | BPF_B: 1498 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ 1499 from->imm); 1500 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1501 /* 1502 * Cannot use BPF_STX_MEM() macro here as it 1503 * hardcodes BPF_MEM mode, losing PROBE_MEM32 1504 * and breaking arena addressing in the JIT. 1505 */ 1506 *to++ = (struct bpf_insn) { 1507 .code = BPF_STX | BPF_PROBE_MEM32 | 1508 BPF_SIZE(from->code), 1509 .dst_reg = from->dst_reg, 1510 .src_reg = BPF_REG_AX, 1511 .off = from->off, 1512 }; 1513 break; 1514 } 1515 out: 1516 return to - to_buff; 1517 } 1518 1519 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1520 gfp_t gfp_extra_flags) 1521 { 1522 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1523 struct bpf_prog *fp; 1524 1525 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1526 if (fp != NULL) { 1527 /* aux->prog still points to the fp_other one, so 1528 * when promoting the clone to the real program, 1529 * this still needs to be adapted. 1530 */ 1531 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1532 } 1533 1534 return fp; 1535 } 1536 1537 static void bpf_prog_clone_free(struct bpf_prog *fp) 1538 { 1539 /* aux was stolen by the other clone, so we cannot free 1540 * it from this path! It will be freed eventually by the 1541 * other program on release. 1542 * 1543 * At this point, we don't need a deferred release since 1544 * clone is guaranteed to not be locked. 1545 */ 1546 fp->aux = NULL; 1547 fp->stats = NULL; 1548 fp->active = NULL; 1549 __bpf_prog_free(fp); 1550 } 1551 1552 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1553 { 1554 /* We have to repoint aux->prog to self, as we don't 1555 * know whether fp here is the clone or the original. 1556 */ 1557 fp->aux->prog = fp; 1558 if (fp->aux->offload) 1559 fp->aux->offload->prog = fp; 1560 bpf_prog_clone_free(fp_other); 1561 } 1562 1563 /* 1564 * Now this function is used only to blind the main prog and must be invoked only when 1565 * bpf_prog_need_blind() returns true. 1566 */ 1567 struct bpf_prog *bpf_jit_blind_constants(struct bpf_verifier_env *env, struct bpf_prog *prog) 1568 { 1569 struct bpf_insn insn_buff[16], aux[2]; 1570 struct bpf_prog *clone, *tmp; 1571 int insn_delta, insn_cnt; 1572 struct bpf_insn *insn; 1573 int i, rewritten; 1574 1575 if (WARN_ON_ONCE(env && env->prog != prog)) 1576 return ERR_PTR(-EINVAL); 1577 1578 clone = bpf_prog_clone_create(prog, GFP_USER); 1579 if (!clone) 1580 return ERR_PTR(-ENOMEM); 1581 1582 /* make sure bpf_patch_insn_data() patches the correct prog */ 1583 if (env) 1584 env->prog = clone; 1585 1586 insn_cnt = clone->len; 1587 insn = clone->insnsi; 1588 1589 for (i = 0; i < insn_cnt; i++, insn++) { 1590 if (bpf_pseudo_func(insn)) { 1591 /* ld_imm64 with an address of bpf subprog is not 1592 * a user controlled constant. Don't randomize it, 1593 * since it will conflict with jit_subprogs() logic. 1594 */ 1595 insn++; 1596 i++; 1597 continue; 1598 } 1599 1600 /* We temporarily need to hold the original ld64 insn 1601 * so that we can still access the first part in the 1602 * second blinding run. 1603 */ 1604 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1605 insn[1].code == 0) 1606 memcpy(aux, insn, sizeof(aux)); 1607 1608 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1609 clone->aux->verifier_zext); 1610 if (!rewritten) 1611 continue; 1612 1613 if (rewritten < 0) 1614 tmp = ERR_PTR(rewritten); 1615 else if (env) 1616 tmp = bpf_patch_insn_data(env, i, insn_buff, rewritten); 1617 else 1618 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1619 1620 if (IS_ERR_OR_NULL(tmp)) { 1621 if (env) 1622 /* restore the original prog */ 1623 env->prog = prog; 1624 /* Patching may have repointed aux->prog during 1625 * realloc from the original one, so we need to 1626 * fix it up here on error. 1627 */ 1628 bpf_jit_prog_release_other(prog, clone); 1629 if (env && fatal_signal_pending(current)) 1630 return ERR_PTR(-EINTR); 1631 return IS_ERR(tmp) ? tmp : ERR_PTR(-ENOMEM); 1632 } 1633 1634 clone = tmp; 1635 insn_delta = rewritten - 1; 1636 1637 if (env) 1638 env->prog = clone; 1639 1640 /* Walk new program and skip insns we just inserted. */ 1641 insn = clone->insnsi + i + insn_delta; 1642 insn_cnt += insn_delta; 1643 i += insn_delta; 1644 } 1645 1646 clone->blinded = 1; 1647 return clone; 1648 } 1649 1650 bool bpf_insn_is_indirect_target(const struct bpf_verifier_env *env, const struct bpf_prog *prog, 1651 int insn_idx) 1652 { 1653 if (!env) 1654 return false; 1655 insn_idx += prog->aux->subprog_start; 1656 return env->insn_aux_data[insn_idx].indirect_target; 1657 } 1658 1659 u16 bpf_out_stack_arg_cnt(const struct bpf_verifier_env *env, const struct bpf_prog *prog) 1660 { 1661 const struct bpf_subprog_info *sub; 1662 1663 if (!env) 1664 return 0; 1665 sub = &env->subprog_info[prog->aux->func_idx]; 1666 return sub->stack_arg_cnt - bpf_in_stack_arg_cnt(sub); 1667 } 1668 #endif /* CONFIG_BPF_JIT */ 1669 1670 /* Base function for offset calculation. Needs to go into .text section, 1671 * therefore keeping it non-static as well; will also be used by JITs 1672 * anyway later on, so do not let the compiler omit it. This also needs 1673 * to go into kallsyms for correlation from e.g. bpftool, so naming 1674 * must not change. 1675 */ 1676 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1677 { 1678 return 0; 1679 } 1680 EXPORT_SYMBOL_GPL(__bpf_call_base); 1681 1682 /* All UAPI available opcodes. */ 1683 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1684 /* 32 bit ALU operations. */ \ 1685 /* Register based. */ \ 1686 INSN_3(ALU, ADD, X), \ 1687 INSN_3(ALU, SUB, X), \ 1688 INSN_3(ALU, AND, X), \ 1689 INSN_3(ALU, OR, X), \ 1690 INSN_3(ALU, LSH, X), \ 1691 INSN_3(ALU, RSH, X), \ 1692 INSN_3(ALU, XOR, X), \ 1693 INSN_3(ALU, MUL, X), \ 1694 INSN_3(ALU, MOV, X), \ 1695 INSN_3(ALU, ARSH, X), \ 1696 INSN_3(ALU, DIV, X), \ 1697 INSN_3(ALU, MOD, X), \ 1698 INSN_2(ALU, NEG), \ 1699 INSN_3(ALU, END, TO_BE), \ 1700 INSN_3(ALU, END, TO_LE), \ 1701 /* Immediate based. */ \ 1702 INSN_3(ALU, ADD, K), \ 1703 INSN_3(ALU, SUB, K), \ 1704 INSN_3(ALU, AND, K), \ 1705 INSN_3(ALU, OR, K), \ 1706 INSN_3(ALU, LSH, K), \ 1707 INSN_3(ALU, RSH, K), \ 1708 INSN_3(ALU, XOR, K), \ 1709 INSN_3(ALU, MUL, K), \ 1710 INSN_3(ALU, MOV, K), \ 1711 INSN_3(ALU, ARSH, K), \ 1712 INSN_3(ALU, DIV, K), \ 1713 INSN_3(ALU, MOD, K), \ 1714 /* 64 bit ALU operations. */ \ 1715 /* Register based. */ \ 1716 INSN_3(ALU64, ADD, X), \ 1717 INSN_3(ALU64, SUB, X), \ 1718 INSN_3(ALU64, AND, X), \ 1719 INSN_3(ALU64, OR, X), \ 1720 INSN_3(ALU64, LSH, X), \ 1721 INSN_3(ALU64, RSH, X), \ 1722 INSN_3(ALU64, XOR, X), \ 1723 INSN_3(ALU64, MUL, X), \ 1724 INSN_3(ALU64, MOV, X), \ 1725 INSN_3(ALU64, ARSH, X), \ 1726 INSN_3(ALU64, DIV, X), \ 1727 INSN_3(ALU64, MOD, X), \ 1728 INSN_2(ALU64, NEG), \ 1729 INSN_3(ALU64, END, TO_LE), \ 1730 /* Immediate based. */ \ 1731 INSN_3(ALU64, ADD, K), \ 1732 INSN_3(ALU64, SUB, K), \ 1733 INSN_3(ALU64, AND, K), \ 1734 INSN_3(ALU64, OR, K), \ 1735 INSN_3(ALU64, LSH, K), \ 1736 INSN_3(ALU64, RSH, K), \ 1737 INSN_3(ALU64, XOR, K), \ 1738 INSN_3(ALU64, MUL, K), \ 1739 INSN_3(ALU64, MOV, K), \ 1740 INSN_3(ALU64, ARSH, K), \ 1741 INSN_3(ALU64, DIV, K), \ 1742 INSN_3(ALU64, MOD, K), \ 1743 /* Call instruction. */ \ 1744 INSN_2(JMP, CALL), \ 1745 /* Exit instruction. */ \ 1746 INSN_2(JMP, EXIT), \ 1747 /* 32-bit Jump instructions. */ \ 1748 /* Register based. */ \ 1749 INSN_3(JMP32, JEQ, X), \ 1750 INSN_3(JMP32, JNE, X), \ 1751 INSN_3(JMP32, JGT, X), \ 1752 INSN_3(JMP32, JLT, X), \ 1753 INSN_3(JMP32, JGE, X), \ 1754 INSN_3(JMP32, JLE, X), \ 1755 INSN_3(JMP32, JSGT, X), \ 1756 INSN_3(JMP32, JSLT, X), \ 1757 INSN_3(JMP32, JSGE, X), \ 1758 INSN_3(JMP32, JSLE, X), \ 1759 INSN_3(JMP32, JSET, X), \ 1760 /* Immediate based. */ \ 1761 INSN_3(JMP32, JEQ, K), \ 1762 INSN_3(JMP32, JNE, K), \ 1763 INSN_3(JMP32, JGT, K), \ 1764 INSN_3(JMP32, JLT, K), \ 1765 INSN_3(JMP32, JGE, K), \ 1766 INSN_3(JMP32, JLE, K), \ 1767 INSN_3(JMP32, JSGT, K), \ 1768 INSN_3(JMP32, JSLT, K), \ 1769 INSN_3(JMP32, JSGE, K), \ 1770 INSN_3(JMP32, JSLE, K), \ 1771 INSN_3(JMP32, JSET, K), \ 1772 /* Jump instructions. */ \ 1773 /* Register based. */ \ 1774 INSN_3(JMP, JEQ, X), \ 1775 INSN_3(JMP, JNE, X), \ 1776 INSN_3(JMP, JGT, X), \ 1777 INSN_3(JMP, JLT, X), \ 1778 INSN_3(JMP, JGE, X), \ 1779 INSN_3(JMP, JLE, X), \ 1780 INSN_3(JMP, JSGT, X), \ 1781 INSN_3(JMP, JSLT, X), \ 1782 INSN_3(JMP, JSGE, X), \ 1783 INSN_3(JMP, JSLE, X), \ 1784 INSN_3(JMP, JSET, X), \ 1785 /* Immediate based. */ \ 1786 INSN_3(JMP, JEQ, K), \ 1787 INSN_3(JMP, JNE, K), \ 1788 INSN_3(JMP, JGT, K), \ 1789 INSN_3(JMP, JLT, K), \ 1790 INSN_3(JMP, JGE, K), \ 1791 INSN_3(JMP, JLE, K), \ 1792 INSN_3(JMP, JSGT, K), \ 1793 INSN_3(JMP, JSLT, K), \ 1794 INSN_3(JMP, JSGE, K), \ 1795 INSN_3(JMP, JSLE, K), \ 1796 INSN_3(JMP, JSET, K), \ 1797 INSN_2(JMP, JA), \ 1798 INSN_2(JMP32, JA), \ 1799 /* Atomic operations. */ \ 1800 INSN_3(STX, ATOMIC, B), \ 1801 INSN_3(STX, ATOMIC, H), \ 1802 INSN_3(STX, ATOMIC, W), \ 1803 INSN_3(STX, ATOMIC, DW), \ 1804 /* Store instructions. */ \ 1805 /* Register based. */ \ 1806 INSN_3(STX, MEM, B), \ 1807 INSN_3(STX, MEM, H), \ 1808 INSN_3(STX, MEM, W), \ 1809 INSN_3(STX, MEM, DW), \ 1810 /* Immediate based. */ \ 1811 INSN_3(ST, MEM, B), \ 1812 INSN_3(ST, MEM, H), \ 1813 INSN_3(ST, MEM, W), \ 1814 INSN_3(ST, MEM, DW), \ 1815 /* Load instructions. */ \ 1816 /* Register based. */ \ 1817 INSN_3(LDX, MEM, B), \ 1818 INSN_3(LDX, MEM, H), \ 1819 INSN_3(LDX, MEM, W), \ 1820 INSN_3(LDX, MEM, DW), \ 1821 INSN_3(LDX, MEMSX, B), \ 1822 INSN_3(LDX, MEMSX, H), \ 1823 INSN_3(LDX, MEMSX, W), \ 1824 /* Immediate based. */ \ 1825 INSN_3(LD, IMM, DW) 1826 1827 bool bpf_opcode_in_insntable(u8 code) 1828 { 1829 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1830 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1831 static const bool public_insntable[256] = { 1832 [0 ... 255] = false, 1833 /* Now overwrite non-defaults ... */ 1834 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1835 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1836 [BPF_LD | BPF_ABS | BPF_B] = true, 1837 [BPF_LD | BPF_ABS | BPF_H] = true, 1838 [BPF_LD | BPF_ABS | BPF_W] = true, 1839 [BPF_LD | BPF_IND | BPF_B] = true, 1840 [BPF_LD | BPF_IND | BPF_H] = true, 1841 [BPF_LD | BPF_IND | BPF_W] = true, 1842 [BPF_JMP | BPF_JA | BPF_X] = true, 1843 [BPF_JMP | BPF_JCOND] = true, 1844 }; 1845 #undef BPF_INSN_3_TBL 1846 #undef BPF_INSN_2_TBL 1847 return public_insntable[code]; 1848 } 1849 1850 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1851 /* Absolute value of s32 without undefined behavior for S32_MIN */ 1852 static u32 abs_s32(s32 x) 1853 { 1854 return x >= 0 ? (u32)x : -(u32)x; 1855 } 1856 1857 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1858 const struct bpf_insn *insn); 1859 1860 /** 1861 * ___bpf_prog_run - run eBPF program on a given context 1862 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1863 * @insn: is the array of eBPF instructions 1864 * 1865 * Decode and execute eBPF instructions. 1866 * 1867 * Return: whatever value is in %BPF_R0 at program exit 1868 */ 1869 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1870 { 1871 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1872 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1873 static const void * const jumptable[256] __annotate_jump_table = { 1874 [0 ... 255] = &&default_label, 1875 /* Now overwrite non-defaults ... */ 1876 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1877 /* Non-UAPI available opcodes. */ 1878 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1879 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1880 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1881 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1882 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1883 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1884 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1885 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1886 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1887 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1888 }; 1889 #undef BPF_INSN_3_LBL 1890 #undef BPF_INSN_2_LBL 1891 u32 tail_call_cnt = 0; 1892 1893 #define CONT ({ insn++; goto select_insn; }) 1894 #define CONT_JMP ({ insn++; goto select_insn; }) 1895 1896 select_insn: 1897 goto *jumptable[insn->code]; 1898 1899 /* Explicitly mask the register-based shift amounts with 63 or 31 1900 * to avoid undefined behavior. Normally this won't affect the 1901 * generated code, for example, in case of native 64 bit archs such 1902 * as x86-64 or arm64, the compiler is optimizing the AND away for 1903 * the interpreter. In case of JITs, each of the JIT backends compiles 1904 * the BPF shift operations to machine instructions which produce 1905 * implementation-defined results in such a case; the resulting 1906 * contents of the register may be arbitrary, but program behaviour 1907 * as a whole remains defined. In other words, in case of JIT backends, 1908 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1909 */ 1910 /* ALU (shifts) */ 1911 #define SHT(OPCODE, OP) \ 1912 ALU64_##OPCODE##_X: \ 1913 DST = DST OP (SRC & 63); \ 1914 CONT; \ 1915 ALU_##OPCODE##_X: \ 1916 DST = (u32) DST OP ((u32) SRC & 31); \ 1917 CONT; \ 1918 ALU64_##OPCODE##_K: \ 1919 DST = DST OP IMM; \ 1920 CONT; \ 1921 ALU_##OPCODE##_K: \ 1922 DST = (u32) DST OP (u32) IMM; \ 1923 CONT; 1924 /* ALU (rest) */ 1925 #define ALU(OPCODE, OP) \ 1926 ALU64_##OPCODE##_X: \ 1927 DST = DST OP SRC; \ 1928 CONT; \ 1929 ALU_##OPCODE##_X: \ 1930 DST = (u32) DST OP (u32) SRC; \ 1931 CONT; \ 1932 ALU64_##OPCODE##_K: \ 1933 DST = DST OP IMM; \ 1934 CONT; \ 1935 ALU_##OPCODE##_K: \ 1936 DST = (u32) DST OP (u32) IMM; \ 1937 CONT; 1938 ALU(ADD, +) 1939 ALU(SUB, -) 1940 ALU(AND, &) 1941 ALU(OR, |) 1942 ALU(XOR, ^) 1943 ALU(MUL, *) 1944 SHT(LSH, <<) 1945 SHT(RSH, >>) 1946 #undef SHT 1947 #undef ALU 1948 ALU_NEG: 1949 DST = (u32) -DST; 1950 CONT; 1951 ALU64_NEG: 1952 DST = -DST; 1953 CONT; 1954 ALU_MOV_X: 1955 switch (OFF) { 1956 case 0: 1957 DST = (u32) SRC; 1958 break; 1959 case 8: 1960 DST = (u32)(s8) SRC; 1961 break; 1962 case 16: 1963 DST = (u32)(s16) SRC; 1964 break; 1965 } 1966 CONT; 1967 ALU_MOV_K: 1968 DST = (u32) IMM; 1969 CONT; 1970 ALU64_MOV_X: 1971 switch (OFF) { 1972 case 0: 1973 DST = SRC; 1974 break; 1975 case 8: 1976 DST = (s8) SRC; 1977 break; 1978 case 16: 1979 DST = (s16) SRC; 1980 break; 1981 case 32: 1982 DST = (s32) SRC; 1983 break; 1984 } 1985 CONT; 1986 ALU64_MOV_K: 1987 DST = IMM; 1988 CONT; 1989 LD_IMM_DW: 1990 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1991 insn++; 1992 CONT; 1993 ALU_ARSH_X: 1994 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1995 CONT; 1996 ALU_ARSH_K: 1997 DST = (u64) (u32) (((s32) DST) >> IMM); 1998 CONT; 1999 ALU64_ARSH_X: 2000 (*(s64 *) &DST) >>= (SRC & 63); 2001 CONT; 2002 ALU64_ARSH_K: 2003 (*(s64 *) &DST) >>= IMM; 2004 CONT; 2005 ALU64_MOD_X: 2006 switch (OFF) { 2007 case 0: 2008 div64_u64_rem(DST, SRC, &AX); 2009 DST = AX; 2010 break; 2011 case 1: 2012 AX = div64_s64(DST, SRC); 2013 DST = DST - AX * SRC; 2014 break; 2015 } 2016 CONT; 2017 ALU_MOD_X: 2018 switch (OFF) { 2019 case 0: 2020 AX = (u32) DST; 2021 DST = do_div(AX, (u32) SRC); 2022 break; 2023 case 1: 2024 AX = abs_s32((s32)DST); 2025 AX = do_div(AX, abs_s32((s32)SRC)); 2026 if ((s32)DST < 0) 2027 DST = (u32)-AX; 2028 else 2029 DST = (u32)AX; 2030 break; 2031 } 2032 CONT; 2033 ALU64_MOD_K: 2034 switch (OFF) { 2035 case 0: 2036 div64_u64_rem(DST, IMM, &AX); 2037 DST = AX; 2038 break; 2039 case 1: 2040 AX = div64_s64(DST, IMM); 2041 DST = DST - AX * IMM; 2042 break; 2043 } 2044 CONT; 2045 ALU_MOD_K: 2046 switch (OFF) { 2047 case 0: 2048 AX = (u32) DST; 2049 DST = do_div(AX, (u32) IMM); 2050 break; 2051 case 1: 2052 AX = abs_s32((s32)DST); 2053 AX = do_div(AX, abs_s32((s32)IMM)); 2054 if ((s32)DST < 0) 2055 DST = (u32)-AX; 2056 else 2057 DST = (u32)AX; 2058 break; 2059 } 2060 CONT; 2061 ALU64_DIV_X: 2062 switch (OFF) { 2063 case 0: 2064 DST = div64_u64(DST, SRC); 2065 break; 2066 case 1: 2067 DST = div64_s64(DST, SRC); 2068 break; 2069 } 2070 CONT; 2071 ALU_DIV_X: 2072 switch (OFF) { 2073 case 0: 2074 AX = (u32) DST; 2075 do_div(AX, (u32) SRC); 2076 DST = (u32) AX; 2077 break; 2078 case 1: 2079 AX = abs_s32((s32)DST); 2080 do_div(AX, abs_s32((s32)SRC)); 2081 if (((s32)DST < 0) == ((s32)SRC < 0)) 2082 DST = (u32)AX; 2083 else 2084 DST = (u32)-AX; 2085 break; 2086 } 2087 CONT; 2088 ALU64_DIV_K: 2089 switch (OFF) { 2090 case 0: 2091 DST = div64_u64(DST, IMM); 2092 break; 2093 case 1: 2094 DST = div64_s64(DST, IMM); 2095 break; 2096 } 2097 CONT; 2098 ALU_DIV_K: 2099 switch (OFF) { 2100 case 0: 2101 AX = (u32) DST; 2102 do_div(AX, (u32) IMM); 2103 DST = (u32) AX; 2104 break; 2105 case 1: 2106 AX = abs_s32((s32)DST); 2107 do_div(AX, abs_s32((s32)IMM)); 2108 if (((s32)DST < 0) == ((s32)IMM < 0)) 2109 DST = (u32)AX; 2110 else 2111 DST = (u32)-AX; 2112 break; 2113 } 2114 CONT; 2115 ALU_END_TO_BE: 2116 switch (IMM) { 2117 case 16: 2118 DST = (__force u16) cpu_to_be16(DST); 2119 break; 2120 case 32: 2121 DST = (__force u32) cpu_to_be32(DST); 2122 break; 2123 case 64: 2124 DST = (__force u64) cpu_to_be64(DST); 2125 break; 2126 } 2127 CONT; 2128 ALU_END_TO_LE: 2129 switch (IMM) { 2130 case 16: 2131 DST = (__force u16) cpu_to_le16(DST); 2132 break; 2133 case 32: 2134 DST = (__force u32) cpu_to_le32(DST); 2135 break; 2136 case 64: 2137 DST = (__force u64) cpu_to_le64(DST); 2138 break; 2139 } 2140 CONT; 2141 ALU64_END_TO_LE: 2142 switch (IMM) { 2143 case 16: 2144 DST = (__force u16) __swab16(DST); 2145 break; 2146 case 32: 2147 DST = (__force u32) __swab32(DST); 2148 break; 2149 case 64: 2150 DST = (__force u64) __swab64(DST); 2151 break; 2152 } 2153 CONT; 2154 2155 /* CALL */ 2156 JMP_CALL: 2157 /* Function call scratches BPF_R1-BPF_R5 registers, 2158 * preserves BPF_R6-BPF_R9, and stores return value 2159 * into BPF_R0. 2160 */ 2161 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2162 BPF_R4, BPF_R5); 2163 CONT; 2164 2165 JMP_CALL_ARGS: 2166 BPF_R0 = interpreters_args[insn->off](BPF_R1, BPF_R2, BPF_R3, 2167 BPF_R4, BPF_R5, 2168 insn + insn->imm + 1); 2169 CONT; 2170 2171 JMP_TAIL_CALL: { 2172 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2173 struct bpf_array *array = container_of(map, struct bpf_array, map); 2174 struct bpf_prog *prog; 2175 u32 index = BPF_R3; 2176 2177 if (unlikely(index >= array->map.max_entries)) 2178 goto out; 2179 2180 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2181 goto out; 2182 2183 prog = READ_ONCE(array->ptrs[index]); 2184 if (!prog) 2185 goto out; 2186 2187 tail_call_cnt++; 2188 2189 /* ARG1 at this point is guaranteed to point to CTX from 2190 * the verifier side due to the fact that the tail call is 2191 * handled like a helper, that is, bpf_tail_call_proto, 2192 * where arg1_type is ARG_PTR_TO_CTX. 2193 */ 2194 insn = prog->insnsi; 2195 goto select_insn; 2196 out: 2197 CONT; 2198 } 2199 JMP_JA: 2200 insn += insn->off; 2201 CONT; 2202 JMP32_JA: 2203 insn += insn->imm; 2204 CONT; 2205 JMP_EXIT: 2206 return BPF_R0; 2207 /* JMP */ 2208 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2209 JMP_##OPCODE##_X: \ 2210 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2211 insn += insn->off; \ 2212 CONT_JMP; \ 2213 } \ 2214 CONT; \ 2215 JMP32_##OPCODE##_X: \ 2216 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2217 insn += insn->off; \ 2218 CONT_JMP; \ 2219 } \ 2220 CONT; \ 2221 JMP_##OPCODE##_K: \ 2222 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2223 insn += insn->off; \ 2224 CONT_JMP; \ 2225 } \ 2226 CONT; \ 2227 JMP32_##OPCODE##_K: \ 2228 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2229 insn += insn->off; \ 2230 CONT_JMP; \ 2231 } \ 2232 CONT; 2233 COND_JMP(u, JEQ, ==) 2234 COND_JMP(u, JNE, !=) 2235 COND_JMP(u, JGT, >) 2236 COND_JMP(u, JLT, <) 2237 COND_JMP(u, JGE, >=) 2238 COND_JMP(u, JLE, <=) 2239 COND_JMP(u, JSET, &) 2240 COND_JMP(s, JSGT, >) 2241 COND_JMP(s, JSLT, <) 2242 COND_JMP(s, JSGE, >=) 2243 COND_JMP(s, JSLE, <=) 2244 #undef COND_JMP 2245 /* ST, STX and LDX*/ 2246 ST_NOSPEC: 2247 /* Speculation barrier for mitigating Speculative Store Bypass, 2248 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2249 * rely on the firmware mitigation as controlled via the ssbd 2250 * kernel parameter. Whenever the mitigation is enabled, it 2251 * works for all of the kernel code with no need to provide any 2252 * additional instructions here. In case of x86, we use 'lfence' 2253 * insn for mitigation. We reuse preexisting logic from Spectre 2254 * v1 mitigation that happens to produce the required code on 2255 * x86 for v4 as well. 2256 */ 2257 barrier_nospec(); 2258 CONT; 2259 #define LDST(SIZEOP, SIZE) \ 2260 STX_MEM_##SIZEOP: \ 2261 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2262 CONT; \ 2263 ST_MEM_##SIZEOP: \ 2264 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2265 CONT; \ 2266 LDX_MEM_##SIZEOP: \ 2267 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2268 CONT; \ 2269 LDX_PROBE_MEM_##SIZEOP: \ 2270 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2271 (const void *)(long) (SRC + insn->off)); \ 2272 DST = *((SIZE *)&DST); \ 2273 CONT; 2274 2275 LDST(B, u8) 2276 LDST(H, u16) 2277 LDST(W, u32) 2278 LDST(DW, u64) 2279 #undef LDST 2280 2281 #define LDSX(SIZEOP, SIZE) \ 2282 LDX_MEMSX_##SIZEOP: \ 2283 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2284 CONT; \ 2285 LDX_PROBE_MEMSX_##SIZEOP: \ 2286 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2287 (const void *)(long) (SRC + insn->off)); \ 2288 DST = *((SIZE *)&DST); \ 2289 CONT; 2290 2291 LDSX(B, s8) 2292 LDSX(H, s16) 2293 LDSX(W, s32) 2294 #undef LDSX 2295 2296 #define ATOMIC_ALU_OP(BOP, KOP) \ 2297 case BOP: \ 2298 if (BPF_SIZE(insn->code) == BPF_W) \ 2299 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2300 (DST + insn->off)); \ 2301 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2302 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2303 (DST + insn->off)); \ 2304 else \ 2305 goto default_label; \ 2306 break; \ 2307 case BOP | BPF_FETCH: \ 2308 if (BPF_SIZE(insn->code) == BPF_W) \ 2309 SRC = (u32) atomic_fetch_##KOP( \ 2310 (u32) SRC, \ 2311 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2312 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2313 SRC = (u64) atomic64_fetch_##KOP( \ 2314 (u64) SRC, \ 2315 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2316 else \ 2317 goto default_label; \ 2318 break; 2319 2320 STX_ATOMIC_DW: 2321 STX_ATOMIC_W: 2322 STX_ATOMIC_H: 2323 STX_ATOMIC_B: 2324 switch (IMM) { 2325 /* Atomic read-modify-write instructions support only W and DW 2326 * size modifiers. 2327 */ 2328 ATOMIC_ALU_OP(BPF_ADD, add) 2329 ATOMIC_ALU_OP(BPF_AND, and) 2330 ATOMIC_ALU_OP(BPF_OR, or) 2331 ATOMIC_ALU_OP(BPF_XOR, xor) 2332 #undef ATOMIC_ALU_OP 2333 2334 case BPF_XCHG: 2335 if (BPF_SIZE(insn->code) == BPF_W) 2336 SRC = (u32) atomic_xchg( 2337 (atomic_t *)(unsigned long) (DST + insn->off), 2338 (u32) SRC); 2339 else if (BPF_SIZE(insn->code) == BPF_DW) 2340 SRC = (u64) atomic64_xchg( 2341 (atomic64_t *)(unsigned long) (DST + insn->off), 2342 (u64) SRC); 2343 else 2344 goto default_label; 2345 break; 2346 case BPF_CMPXCHG: 2347 if (BPF_SIZE(insn->code) == BPF_W) 2348 BPF_R0 = (u32) atomic_cmpxchg( 2349 (atomic_t *)(unsigned long) (DST + insn->off), 2350 (u32) BPF_R0, (u32) SRC); 2351 else if (BPF_SIZE(insn->code) == BPF_DW) 2352 BPF_R0 = (u64) atomic64_cmpxchg( 2353 (atomic64_t *)(unsigned long) (DST + insn->off), 2354 (u64) BPF_R0, (u64) SRC); 2355 else 2356 goto default_label; 2357 break; 2358 /* Atomic load and store instructions support all size 2359 * modifiers. 2360 */ 2361 case BPF_LOAD_ACQ: 2362 switch (BPF_SIZE(insn->code)) { 2363 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2364 case BPF_##SIZEOP: \ 2365 DST = (SIZE)smp_load_acquire( \ 2366 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2367 break; 2368 LOAD_ACQUIRE(B, u8) 2369 LOAD_ACQUIRE(H, u16) 2370 LOAD_ACQUIRE(W, u32) 2371 #ifdef CONFIG_64BIT 2372 LOAD_ACQUIRE(DW, u64) 2373 #endif 2374 #undef LOAD_ACQUIRE 2375 default: 2376 goto default_label; 2377 } 2378 break; 2379 case BPF_STORE_REL: 2380 switch (BPF_SIZE(insn->code)) { 2381 #define STORE_RELEASE(SIZEOP, SIZE) \ 2382 case BPF_##SIZEOP: \ 2383 smp_store_release( \ 2384 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2385 break; 2386 STORE_RELEASE(B, u8) 2387 STORE_RELEASE(H, u16) 2388 STORE_RELEASE(W, u32) 2389 #ifdef CONFIG_64BIT 2390 STORE_RELEASE(DW, u64) 2391 #endif 2392 #undef STORE_RELEASE 2393 default: 2394 goto default_label; 2395 } 2396 break; 2397 2398 default: 2399 goto default_label; 2400 } 2401 CONT; 2402 2403 default_label: 2404 /* If we ever reach this, we have a bug somewhere. Die hard here 2405 * instead of just returning 0; we could be somewhere in a subprog, 2406 * so execution could continue otherwise which we do /not/ want. 2407 * 2408 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2409 */ 2410 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2411 insn->code, insn->imm); 2412 BUG_ON(1); 2413 return 0; 2414 } 2415 2416 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2417 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2418 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2419 { \ 2420 u64 stack[stack_size / sizeof(u64)]; \ 2421 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2422 \ 2423 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2424 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2425 ARG1 = (u64) (unsigned long) ctx; \ 2426 return ___bpf_prog_run(regs, insn); \ 2427 } 2428 2429 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2430 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2431 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2432 const struct bpf_insn *insn) \ 2433 { \ 2434 u64 stack[stack_size / sizeof(u64)]; \ 2435 u64 regs[MAX_BPF_EXT_REG]; \ 2436 \ 2437 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2438 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2439 BPF_R1 = r1; \ 2440 BPF_R2 = r2; \ 2441 BPF_R3 = r3; \ 2442 BPF_R4 = r4; \ 2443 BPF_R5 = r5; \ 2444 return ___bpf_prog_run(regs, insn); \ 2445 } 2446 2447 #define EVAL1(FN, X) FN(X) 2448 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2449 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2450 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2451 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2452 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2453 2454 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2455 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2456 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2457 2458 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2459 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2460 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2461 2462 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2463 2464 static unsigned int (*interpreters[])(const void *ctx, 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 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2472 static __maybe_unused 2473 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2474 const struct bpf_insn *insn) = { 2475 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2476 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2477 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2478 }; 2479 #undef PROG_NAME_LIST 2480 2481 #ifdef CONFIG_BPF_SYSCALL 2482 int bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2483 { 2484 stack_depth = max_t(u32, stack_depth, 1); 2485 /* Prevent out-of-bounds read to interpreters_args */ 2486 if (stack_depth > MAX_BPF_STACK) 2487 return -EINVAL; 2488 insn->off = (round_up(stack_depth, 32) / 32) - 1; 2489 insn->code = BPF_JMP | BPF_CALL_ARGS; 2490 return 0; 2491 } 2492 2493 s32 bpf_call_args_imm(s16 idx) 2494 { 2495 if (WARN_ON_ONCE(idx < 0 || idx >= ARRAY_SIZE(interpreters_args))) 2496 return 0; 2497 return BPF_CALL_IMM(interpreters_args[idx]); 2498 } 2499 #endif 2500 #endif 2501 2502 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2503 const struct bpf_insn *insn) 2504 { 2505 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2506 * is not working properly, so warn about it! 2507 */ 2508 WARN_ON_ONCE(1); 2509 return 0; 2510 } 2511 2512 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2513 const struct bpf_prog *fp) 2514 { 2515 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2516 struct bpf_prog_aux *aux = fp->aux; 2517 enum bpf_cgroup_storage_type i; 2518 bool ret = false; 2519 u64 cookie; 2520 2521 if (fp->kprobe_override) 2522 return ret; 2523 2524 spin_lock(&map->owner_lock); 2525 /* There's no owner yet where we could check for compatibility. */ 2526 if (!map->owner) { 2527 map->owner = bpf_map_owner_alloc(map); 2528 if (!map->owner) 2529 goto err; 2530 map->owner->type = prog_type; 2531 map->owner->jited = fp->jited; 2532 map->owner->xdp_has_frags = aux->xdp_has_frags; 2533 map->owner->sleepable = fp->sleepable; 2534 map->owner->expected_attach_type = fp->expected_attach_type; 2535 map->owner->attach_func_proto = aux->attach_func_proto; 2536 for_each_cgroup_storage_type(i) { 2537 map->owner->storage_cookie[i] = 2538 aux->cgroup_storage[i] ? 2539 aux->cgroup_storage[i]->cookie : 0; 2540 } 2541 ret = true; 2542 } else { 2543 ret = map->owner->type == prog_type && 2544 map->owner->jited == fp->jited && 2545 map->owner->xdp_has_frags == aux->xdp_has_frags && 2546 map->owner->sleepable == fp->sleepable; 2547 if (ret && 2548 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2549 map->owner->expected_attach_type != fp->expected_attach_type) 2550 ret = false; 2551 for_each_cgroup_storage_type(i) { 2552 if (!ret) 2553 break; 2554 cookie = aux->cgroup_storage[i] ? 2555 aux->cgroup_storage[i]->cookie : 0; 2556 ret = map->owner->storage_cookie[i] == cookie || 2557 (!cookie && !aux->tail_call_reachable); 2558 } 2559 if (ret && 2560 map->owner->attach_func_proto != aux->attach_func_proto) { 2561 switch (prog_type) { 2562 case BPF_PROG_TYPE_TRACING: 2563 case BPF_PROG_TYPE_LSM: 2564 case BPF_PROG_TYPE_EXT: 2565 case BPF_PROG_TYPE_STRUCT_OPS: 2566 ret = false; 2567 break; 2568 default: 2569 break; 2570 } 2571 } 2572 } 2573 err: 2574 spin_unlock(&map->owner_lock); 2575 return ret; 2576 } 2577 2578 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2579 { 2580 /* XDP programs inserted into maps are not guaranteed to run on 2581 * a particular netdev (and can run outside driver context entirely 2582 * in the case of devmap and cpumap). Until device checks 2583 * are implemented, prohibit adding dev-bound programs to program maps. 2584 */ 2585 if (bpf_prog_is_dev_bound(fp->aux)) 2586 return false; 2587 2588 return __bpf_prog_map_compatible(map, fp); 2589 } 2590 2591 static int bpf_check_tail_call(const struct bpf_prog *fp) 2592 { 2593 struct bpf_prog_aux *aux = fp->aux; 2594 int i, ret = 0; 2595 2596 mutex_lock(&aux->used_maps_mutex); 2597 for (i = 0; i < aux->used_map_cnt; i++) { 2598 struct bpf_map *map = aux->used_maps[i]; 2599 2600 if (!map_type_contains_progs(map)) 2601 continue; 2602 2603 if (!__bpf_prog_map_compatible(map, fp)) { 2604 ret = -EINVAL; 2605 goto out; 2606 } 2607 } 2608 2609 out: 2610 mutex_unlock(&aux->used_maps_mutex); 2611 return ret; 2612 } 2613 2614 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2615 { 2616 bool select_interpreter = false; 2617 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2618 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2619 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2620 2621 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2622 * But for non-JITed programs, we don't need bpf_func, so no bounds 2623 * check needed. 2624 */ 2625 if (idx < ARRAY_SIZE(interpreters)) { 2626 fp->bpf_func = interpreters[idx]; 2627 select_interpreter = true; 2628 } else { 2629 fp->bpf_func = __bpf_prog_ret0_warn; 2630 } 2631 #else 2632 fp->bpf_func = __bpf_prog_ret0_warn; 2633 #endif 2634 return select_interpreter; 2635 } 2636 2637 static struct bpf_prog *bpf_prog_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 2638 { 2639 #ifdef CONFIG_BPF_JIT 2640 struct bpf_prog *orig_prog; 2641 2642 if (!bpf_prog_need_blind(prog)) 2643 return bpf_int_jit_compile(env, prog); 2644 2645 orig_prog = prog; 2646 prog = bpf_jit_blind_constants(env, prog); 2647 /* 2648 * Fall back to the interpreter after blinding failures, except when 2649 * the loader was killed. 2650 */ 2651 if (IS_ERR(prog)) { 2652 if (PTR_ERR(prog) == -EINTR) 2653 return prog; 2654 goto out_restore; 2655 } 2656 2657 prog = bpf_int_jit_compile(env, prog); 2658 if (prog->jited) { 2659 bpf_jit_prog_release_other(prog, orig_prog); 2660 return prog; 2661 } 2662 2663 bpf_jit_prog_release_other(orig_prog, prog); 2664 2665 out_restore: 2666 prog = orig_prog; 2667 #endif 2668 return prog; 2669 } 2670 2671 struct bpf_prog *__bpf_prog_select_runtime(struct bpf_verifier_env *env, struct bpf_prog *fp, 2672 int *err) 2673 { 2674 struct bpf_prog *jit_prog; 2675 2676 /* In case of BPF to BPF calls, verifier did all the prep 2677 * work with regards to JITing, etc. 2678 */ 2679 bool jit_needed = fp->jit_required; 2680 2681 if (fp->bpf_func) 2682 goto finalize; 2683 2684 if (!bpf_prog_select_interpreter(fp)) 2685 jit_needed = true; 2686 2687 /* eBPF JITs can rewrite the program in case constant 2688 * blinding is active. However, in case of error during 2689 * blinding, bpf_int_jit_compile() must always return a 2690 * valid program, which in this case would simply not 2691 * be JITed, but falls back to the interpreter. 2692 */ 2693 if (!bpf_prog_is_offloaded(fp->aux)) { 2694 *err = bpf_prog_alloc_jited_linfo(fp); 2695 if (*err) 2696 return fp; 2697 2698 jit_prog = bpf_prog_jit_compile(env, fp); 2699 if (IS_ERR(jit_prog)) { 2700 *err = PTR_ERR(jit_prog); 2701 return fp; 2702 } 2703 fp = jit_prog; 2704 bpf_prog_jit_attempt_done(fp); 2705 if (!fp->jited && jit_needed) { 2706 *err = -ENOTSUPP; 2707 return fp; 2708 } 2709 } else { 2710 *err = bpf_prog_offload_compile(fp); 2711 if (*err) 2712 return fp; 2713 } 2714 2715 finalize: 2716 *err = bpf_prog_lock_ro(fp); 2717 if (*err) 2718 return fp; 2719 2720 /* The tail call compatibility check can only be done at 2721 * this late stage as we need to determine, if we deal 2722 * with JITed or non JITed program concatenations and not 2723 * all eBPF JITs might immediately support all features. 2724 */ 2725 *err = bpf_check_tail_call(fp); 2726 2727 return fp; 2728 } 2729 2730 /** 2731 * bpf_prog_select_runtime - select exec runtime for BPF program 2732 * @fp: bpf_prog populated with BPF program 2733 * @err: pointer to error variable 2734 * 2735 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2736 * The BPF program will be executed via bpf_prog_run() function. 2737 * 2738 * Return: the &fp argument along with &err set to 0 for success or 2739 * a negative errno code on failure 2740 */ 2741 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2742 { 2743 return __bpf_prog_select_runtime(NULL, fp, err); 2744 } 2745 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2746 2747 static unsigned int __bpf_prog_ret1(const void *ctx, 2748 const struct bpf_insn *insn) 2749 { 2750 return 1; 2751 } 2752 2753 static struct bpf_prog_dummy { 2754 struct bpf_prog prog; 2755 } dummy_bpf_prog = { 2756 .prog = { 2757 .bpf_func = __bpf_prog_ret1, 2758 }, 2759 }; 2760 2761 struct bpf_prog_array bpf_empty_prog_array = { 2762 .items = { 2763 { .prog = NULL }, 2764 }, 2765 }; 2766 EXPORT_SYMBOL(bpf_empty_prog_array); 2767 2768 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2769 { 2770 struct bpf_prog_array *p; 2771 2772 if (prog_cnt) 2773 p = kzalloc_flex(*p, items, prog_cnt + 1, flags); 2774 else 2775 p = &bpf_empty_prog_array; 2776 2777 return p; 2778 } 2779 2780 void bpf_prog_array_free(struct bpf_prog_array *progs) 2781 { 2782 if (!progs || progs == &bpf_empty_prog_array) 2783 return; 2784 kfree_rcu(progs, rcu); 2785 } 2786 2787 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2788 { 2789 struct bpf_prog_array *progs; 2790 2791 /* 2792 * RCU Tasks Trace grace period implies RCU grace period, there is no 2793 * need to call kfree_rcu(), just call kfree() directly. 2794 */ 2795 progs = container_of(rcu, struct bpf_prog_array, rcu); 2796 kfree(progs); 2797 } 2798 2799 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2800 { 2801 if (!progs || progs == &bpf_empty_prog_array) 2802 return; 2803 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2804 } 2805 2806 int bpf_prog_array_length(struct bpf_prog_array *array) 2807 { 2808 struct bpf_prog_array_item *item; 2809 u32 cnt = 0; 2810 2811 for (item = array->items; item->prog; item++) 2812 if (item->prog != &dummy_bpf_prog.prog) 2813 cnt++; 2814 return cnt; 2815 } 2816 2817 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2818 { 2819 struct bpf_prog_array_item *item; 2820 2821 for (item = array->items; item->prog; item++) 2822 if (item->prog != &dummy_bpf_prog.prog) 2823 return false; 2824 return true; 2825 } 2826 2827 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2828 u32 *prog_ids, 2829 u32 request_cnt) 2830 { 2831 struct bpf_prog_array_item *item; 2832 int i = 0; 2833 2834 for (item = array->items; item->prog; item++) { 2835 if (item->prog == &dummy_bpf_prog.prog) 2836 continue; 2837 prog_ids[i] = item->prog->aux->id; 2838 if (++i == request_cnt) { 2839 item++; 2840 break; 2841 } 2842 } 2843 2844 return !!(item->prog); 2845 } 2846 2847 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2848 __u32 __user *prog_ids, u32 cnt) 2849 { 2850 unsigned long err = 0; 2851 bool nospc; 2852 u32 *ids; 2853 2854 /* users of this function are doing: 2855 * cnt = bpf_prog_array_length(); 2856 * if (cnt > 0) 2857 * bpf_prog_array_copy_to_user(..., cnt); 2858 * so below kcalloc doesn't need extra cnt > 0 check. 2859 */ 2860 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2861 if (!ids) 2862 return -ENOMEM; 2863 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2864 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2865 kfree(ids); 2866 if (err) 2867 return -EFAULT; 2868 if (nospc) 2869 return -ENOSPC; 2870 return 0; 2871 } 2872 2873 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2874 struct bpf_prog *old_prog) 2875 { 2876 struct bpf_prog_array_item *item; 2877 2878 for (item = array->items; item->prog; item++) 2879 if (item->prog == old_prog) { 2880 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2881 break; 2882 } 2883 } 2884 2885 /** 2886 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2887 * index into the program array with 2888 * a dummy no-op program. 2889 * @array: a bpf_prog_array 2890 * @index: the index of the program to replace 2891 * 2892 * Skips over dummy programs, by not counting them, when calculating 2893 * the position of the program to replace. 2894 * 2895 * Return: 2896 * * 0 - Success 2897 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2898 * * -ENOENT - Index out of range 2899 */ 2900 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2901 { 2902 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2903 } 2904 2905 /** 2906 * bpf_prog_array_update_at() - Updates the program at the given index 2907 * into the program array. 2908 * @array: a bpf_prog_array 2909 * @index: the index of the program to update 2910 * @prog: the program to insert into the array 2911 * 2912 * Skips over dummy programs, by not counting them, when calculating 2913 * the position of the program to update. 2914 * 2915 * Return: 2916 * * 0 - Success 2917 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2918 * * -ENOENT - Index out of range 2919 */ 2920 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2921 struct bpf_prog *prog) 2922 { 2923 struct bpf_prog_array_item *item; 2924 2925 if (unlikely(index < 0)) 2926 return -EINVAL; 2927 2928 for (item = array->items; item->prog; item++) { 2929 if (item->prog == &dummy_bpf_prog.prog) 2930 continue; 2931 if (!index) { 2932 WRITE_ONCE(item->prog, prog); 2933 return 0; 2934 } 2935 index--; 2936 } 2937 return -ENOENT; 2938 } 2939 2940 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2941 struct bpf_prog *exclude_prog, 2942 struct bpf_prog *include_prog, 2943 u64 bpf_cookie, 2944 struct bpf_prog_array **new_array) 2945 { 2946 int new_prog_cnt, carry_prog_cnt = 0; 2947 struct bpf_prog_array_item *existing, *new; 2948 struct bpf_prog_array *array; 2949 bool found_exclude = false; 2950 2951 /* Figure out how many existing progs we need to carry over to 2952 * the new array. 2953 */ 2954 if (old_array) { 2955 existing = old_array->items; 2956 for (; existing->prog; existing++) { 2957 if (existing->prog == exclude_prog) { 2958 found_exclude = true; 2959 continue; 2960 } 2961 if (existing->prog != &dummy_bpf_prog.prog) 2962 carry_prog_cnt++; 2963 if (existing->prog == include_prog) 2964 return -EEXIST; 2965 } 2966 } 2967 2968 if (exclude_prog && !found_exclude) 2969 return -ENOENT; 2970 2971 /* How many progs (not NULL) will be in the new array? */ 2972 new_prog_cnt = carry_prog_cnt; 2973 if (include_prog) 2974 new_prog_cnt += 1; 2975 2976 /* Do we have any prog (not NULL) in the new array? */ 2977 if (!new_prog_cnt) { 2978 *new_array = NULL; 2979 return 0; 2980 } 2981 2982 /* +1 as the end of prog_array is marked with NULL */ 2983 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2984 if (!array) 2985 return -ENOMEM; 2986 new = array->items; 2987 2988 /* Fill in the new prog array */ 2989 if (carry_prog_cnt) { 2990 existing = old_array->items; 2991 for (; existing->prog; existing++) { 2992 if (existing->prog == exclude_prog || 2993 existing->prog == &dummy_bpf_prog.prog) 2994 continue; 2995 2996 new->prog = existing->prog; 2997 new->bpf_cookie = existing->bpf_cookie; 2998 new++; 2999 } 3000 } 3001 if (include_prog) { 3002 new->prog = include_prog; 3003 new->bpf_cookie = bpf_cookie; 3004 new++; 3005 } 3006 new->prog = NULL; 3007 *new_array = array; 3008 return 0; 3009 } 3010 3011 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 3012 u32 *prog_ids, u32 request_cnt, 3013 u32 *prog_cnt) 3014 { 3015 u32 cnt = 0; 3016 3017 if (array) 3018 cnt = bpf_prog_array_length(array); 3019 3020 *prog_cnt = cnt; 3021 3022 /* return early if user requested only program count or nothing to copy */ 3023 if (!request_cnt || !cnt) 3024 return 0; 3025 3026 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 3027 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 3028 : 0; 3029 } 3030 3031 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 3032 struct bpf_map **used_maps, u32 len) 3033 { 3034 struct bpf_map *map; 3035 bool sleepable; 3036 u32 i; 3037 3038 sleepable = aux->prog->sleepable; 3039 for (i = 0; i < len; i++) { 3040 map = used_maps[i]; 3041 if (map->ops->map_poke_untrack) 3042 map->ops->map_poke_untrack(map, aux); 3043 if (sleepable) 3044 atomic64_dec(&map->sleepable_refcnt); 3045 bpf_map_put(map); 3046 } 3047 } 3048 3049 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 3050 { 3051 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 3052 kfree(aux->used_maps); 3053 } 3054 3055 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 3056 { 3057 #ifdef CONFIG_BPF_SYSCALL 3058 struct btf_mod_pair *btf_mod; 3059 u32 i; 3060 3061 for (i = 0; i < len; i++) { 3062 btf_mod = &used_btfs[i]; 3063 if (btf_mod->module) 3064 module_put(btf_mod->module); 3065 btf_put(btf_mod->btf); 3066 } 3067 #endif 3068 } 3069 3070 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 3071 { 3072 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 3073 kfree(aux->used_btfs); 3074 } 3075 3076 static void bpf_prog_free_deferred(struct work_struct *work) 3077 { 3078 struct bpf_prog_aux *aux; 3079 int i; 3080 3081 aux = container_of(work, struct bpf_prog_aux, work); 3082 #ifdef CONFIG_BPF_SYSCALL 3083 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 3084 bpf_prog_stream_free(aux->prog); 3085 #endif 3086 #ifdef CONFIG_CGROUP_BPF 3087 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 3088 bpf_cgroup_atype_put(aux->cgroup_atype); 3089 #endif 3090 bpf_free_used_maps(aux); 3091 bpf_free_used_btfs(aux); 3092 bpf_prog_disassoc_struct_ops(aux->prog); 3093 if (bpf_prog_is_dev_bound(aux)) 3094 bpf_prog_dev_bound_destroy(aux->prog); 3095 #ifdef CONFIG_PERF_EVENTS 3096 if (aux->prog->has_callchain_buf) 3097 put_callchain_buffers(); 3098 #endif 3099 if (aux->dst_trampoline) 3100 bpf_trampoline_put(aux->dst_trampoline); 3101 for (i = 0; i < aux->real_func_cnt; i++) { 3102 /* We can just unlink the subprog poke descriptor table as 3103 * it was originally linked to the main program and is also 3104 * released along with it. 3105 */ 3106 aux->func[i]->aux->poke_tab = NULL; 3107 bpf_jit_free(aux->func[i]); 3108 } 3109 if (aux->real_func_cnt) { 3110 kfree(aux->func); 3111 bpf_prog_unlock_free(aux->prog); 3112 } else { 3113 bpf_jit_free(aux->prog); 3114 } 3115 } 3116 3117 void bpf_prog_free(struct bpf_prog *fp) 3118 { 3119 struct bpf_prog_aux *aux = fp->aux; 3120 3121 if (aux->dst_prog) 3122 bpf_prog_put(aux->dst_prog); 3123 bpf_token_put(aux->token); 3124 INIT_WORK(&aux->work, bpf_prog_free_deferred); 3125 schedule_work(&aux->work); 3126 } 3127 EXPORT_SYMBOL_GPL(bpf_prog_free); 3128 3129 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 3130 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 3131 3132 void bpf_user_rnd_init_once(void) 3133 { 3134 prandom_init_once(&bpf_user_rnd_state); 3135 } 3136 3137 BPF_CALL_0(bpf_user_rnd_u32) 3138 { 3139 /* Should someone ever have the rather unwise idea to use some 3140 * of the registers passed into this function, then note that 3141 * this function is called from native eBPF and classic-to-eBPF 3142 * transformations. Register assignments from both sides are 3143 * different, f.e. classic always sets fn(ctx, A, X) here. 3144 */ 3145 struct rnd_state *state; 3146 u32 res; 3147 3148 state = &get_cpu_var(bpf_user_rnd_state); 3149 res = prandom_u32_state(state); 3150 put_cpu_var(bpf_user_rnd_state); 3151 3152 return res; 3153 } 3154 3155 BPF_CALL_0(bpf_get_raw_cpu_id) 3156 { 3157 return raw_smp_processor_id(); 3158 } 3159 3160 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 3161 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 3162 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 3163 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 3164 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 3165 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 3166 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 3167 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 3168 const struct bpf_func_proto bpf_spin_lock_proto __weak; 3169 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 3170 const struct bpf_func_proto bpf_jiffies64_proto __weak; 3171 3172 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3173 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3174 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3175 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3176 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3177 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3178 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3179 3180 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3181 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3182 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3183 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3184 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3185 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3186 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3187 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3188 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3189 const struct bpf_func_proto bpf_set_retval_proto __weak; 3190 const struct bpf_func_proto bpf_get_retval_proto __weak; 3191 3192 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3193 { 3194 return NULL; 3195 } 3196 3197 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3198 { 3199 return NULL; 3200 } 3201 3202 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3203 { 3204 return NULL; 3205 } 3206 3207 u64 __weak 3208 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3209 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3210 { 3211 return -ENOTSUPP; 3212 } 3213 EXPORT_SYMBOL_GPL(bpf_event_output); 3214 3215 /* Always built-in helper functions. */ 3216 const struct bpf_func_proto bpf_tail_call_proto = { 3217 /* func is unused for tail_call, we set it to pass the 3218 * get_helper_proto check 3219 */ 3220 .func = BPF_PTR_POISON, 3221 .gpl_only = false, 3222 .ret_type = RET_VOID, 3223 .arg1_type = ARG_PTR_TO_CTX, 3224 .arg2_type = ARG_CONST_MAP_PTR, 3225 .arg3_type = ARG_ANYTHING, 3226 }; 3227 3228 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3229 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3230 * eBPF and implicitly also cBPF can get JITed! 3231 */ 3232 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 3233 { 3234 return prog; 3235 } 3236 3237 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3238 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3239 */ 3240 void __weak bpf_jit_compile(struct bpf_prog *prog) 3241 { 3242 } 3243 3244 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3245 { 3246 return false; 3247 } 3248 3249 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3250 * analysis code and wants explicit zero extension inserted by verifier. 3251 * Otherwise, return FALSE. 3252 * 3253 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3254 * you don't override this. JITs that don't want these extra insns can detect 3255 * them using insn_is_zext. 3256 */ 3257 bool __weak bpf_jit_needs_zext(void) 3258 { 3259 return false; 3260 } 3261 3262 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3263 * all archs. The value returned must not change at runtime as there is 3264 * currently no support for reloading programs that were loaded without 3265 * mitigations. 3266 */ 3267 bool __weak bpf_jit_bypass_spec_v1(void) 3268 { 3269 return false; 3270 } 3271 3272 bool __weak bpf_jit_bypass_spec_v4(void) 3273 { 3274 return false; 3275 } 3276 3277 /* Return true if the JIT inlines the call to the helper corresponding to 3278 * the imm. 3279 * 3280 * The verifier will not patch the insn->imm for the call to the helper if 3281 * this returns true. 3282 */ 3283 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3284 { 3285 return false; 3286 } 3287 3288 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3289 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3290 { 3291 return false; 3292 } 3293 3294 bool __weak bpf_jit_supports_percpu_insn(void) 3295 { 3296 return false; 3297 } 3298 3299 bool __weak bpf_jit_supports_kfunc_call(void) 3300 { 3301 return false; 3302 } 3303 3304 bool __weak bpf_jit_supports_stack_args(void) 3305 { 3306 return false; 3307 } 3308 3309 bool __weak bpf_jit_supports_arena_args(void) 3310 { 3311 return false; 3312 } 3313 3314 bool __weak bpf_jit_supports_far_kfunc_call(void) 3315 { 3316 return false; 3317 } 3318 3319 bool __weak bpf_jit_supports_arena(void) 3320 { 3321 return false; 3322 } 3323 3324 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3325 { 3326 return false; 3327 } 3328 3329 bool __weak bpf_jit_supports_fsession(void) 3330 { 3331 return false; 3332 } 3333 3334 u64 __weak bpf_arch_uaddress_limit(void) 3335 { 3336 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3337 return TASK_SIZE; 3338 #else 3339 return 0; 3340 #endif 3341 } 3342 3343 /* Return TRUE if the JIT backend satisfies the following two conditions: 3344 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3345 * 2) Under the specific arch, the implementation of xchg() is the same 3346 * as atomic_xchg() on pointer-sized words. 3347 */ 3348 bool __weak bpf_jit_supports_ptr_xchg(void) 3349 { 3350 return false; 3351 } 3352 3353 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3354 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3355 */ 3356 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3357 int len) 3358 { 3359 return -EFAULT; 3360 } 3361 3362 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3363 enum bpf_text_poke_type new_t, void *old_addr, 3364 void *new_addr) 3365 { 3366 return -ENOTSUPP; 3367 } 3368 3369 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3370 { 3371 return ERR_PTR(-ENOTSUPP); 3372 } 3373 3374 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3375 { 3376 return -ENOTSUPP; 3377 } 3378 3379 bool __weak bpf_jit_supports_exceptions(void) 3380 { 3381 return false; 3382 } 3383 3384 bool __weak bpf_jit_supports_private_stack(void) 3385 { 3386 return false; 3387 } 3388 3389 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3390 { 3391 } 3392 3393 bool __weak bpf_jit_supports_timed_may_goto(void) 3394 { 3395 return false; 3396 } 3397 3398 u64 __weak arch_bpf_timed_may_goto(void) 3399 { 3400 return 0; 3401 } 3402 3403 static noinline void bpf_prog_report_may_goto_violation(void) 3404 { 3405 #ifdef CONFIG_BPF_SYSCALL 3406 struct bpf_stream_stage ss; 3407 struct bpf_prog *prog; 3408 3409 prog = bpf_prog_find_from_stack(); 3410 if (!prog) 3411 return; 3412 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3413 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3414 bpf_stream_dump_stack(ss); 3415 })); 3416 #endif 3417 } 3418 3419 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3420 { 3421 u64 time = ktime_get_mono_fast_ns(); 3422 3423 /* Populate the timestamp for this stack frame, and refresh count. */ 3424 if (!p->timestamp) { 3425 p->timestamp = time; 3426 return BPF_MAX_TIMED_LOOPS; 3427 } 3428 /* Check if we've exhausted our time slice, and zero count. */ 3429 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3430 bpf_prog_report_may_goto_violation(); 3431 return 0; 3432 } 3433 /* Refresh the count for the stack frame. */ 3434 return BPF_MAX_TIMED_LOOPS; 3435 } 3436 3437 /* for configs without MMU or 32-bit */ 3438 __weak const struct bpf_map_ops arena_map_ops; 3439 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3440 { 3441 return 0; 3442 } 3443 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3444 { 3445 return 0; 3446 } 3447 3448 #ifdef CONFIG_BPF_SYSCALL 3449 __weak bool bpf_arena_handle_page_fault(unsigned long addr, bool is_write, 3450 unsigned long fault_ip) 3451 { 3452 return false; 3453 } 3454 3455 static int __init bpf_global_ma_init(void) 3456 { 3457 int ret; 3458 3459 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3460 bpf_global_ma_set = !ret; 3461 return ret; 3462 } 3463 late_initcall(bpf_global_ma_init); 3464 #endif 3465 3466 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3467 EXPORT_SYMBOL(bpf_stats_enabled_key); 3468 3469 /* All definitions of tracepoints related to BPF. */ 3470 #define CREATE_TRACE_POINTS 3471 #include <linux/bpf_trace.h> 3472 3473 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3474 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3475 3476 #ifdef CONFIG_BPF_SYSCALL 3477 3478 void bpf_get_linfo_source(struct btf *btf, const struct bpf_line_info *linfo, 3479 struct bpf_linfo_source *src) 3480 { 3481 src->file = kbasename(btf_name_by_offset(btf, linfo->file_name_off)); 3482 src->line = btf_name_by_offset(btf, linfo->line_off); 3483 src->file_name_off = linfo->file_name_off; 3484 src->line_num = BPF_LINE_INFO_LINE_NUM(linfo->line_col); 3485 src->line_col = BPF_LINE_INFO_LINE_COL(linfo->line_col); 3486 } 3487 3488 const struct bpf_line_info *bpf_find_linfo(const struct bpf_prog *prog, u32 insn_off) 3489 { 3490 const struct bpf_line_info *linfo; 3491 u32 nr_linfo; 3492 int l, r, m; 3493 3494 nr_linfo = prog->aux->nr_linfo; 3495 if (!nr_linfo || insn_off >= prog->len) 3496 return NULL; 3497 3498 linfo = prog->aux->linfo; 3499 /* Loop invariant: linfo[l].insn_off <= insns_off. 3500 * linfo[0].insn_off == 0 which always satisfies above condition. 3501 * Binary search is searching for rightmost linfo entry that satisfies 3502 * the above invariant, giving us the desired record that covers given 3503 * instruction offset. 3504 */ 3505 l = 0; 3506 r = nr_linfo - 1; 3507 while (l < r) { 3508 /* (r - l + 1) / 2 means we break a tie to the right, so if: 3509 * l=1, r=2, linfo[l].insn_off <= insn_off, linfo[r].insn_off > insn_off, 3510 * then m=2, we see that linfo[m].insn_off > insn_off, and so 3511 * r becomes 1 and we exit the loop with correct l==1. 3512 * If the tie was broken to the left, m=1 would end us up in 3513 * an endless loop where l and m stay at 1 and r stays at 2. 3514 */ 3515 m = l + (r - l + 1) / 2; 3516 if (linfo[m].insn_off <= insn_off) 3517 l = m; 3518 else 3519 r = m - 1; 3520 } 3521 3522 return &linfo[l]; 3523 } 3524 3525 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3526 const char **linep, int *nump) 3527 { 3528 struct bpf_linfo_source src; 3529 int idx = -1, insn_start, insn_end, len; 3530 struct bpf_line_info *linfo; 3531 void **jited_linfo; 3532 struct btf *btf; 3533 int nr_linfo; 3534 3535 btf = prog->aux->btf; 3536 linfo = prog->aux->linfo; 3537 jited_linfo = prog->aux->jited_linfo; 3538 3539 if (!btf || !linfo || !jited_linfo) 3540 return -EINVAL; 3541 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3542 3543 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3544 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3545 3546 insn_start = linfo[0].insn_off; 3547 insn_end = insn_start + len; 3548 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3549 3550 for (int i = 0; i < nr_linfo && 3551 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3552 if (jited_linfo[i] >= (void *)ip) 3553 break; 3554 idx = i; 3555 } 3556 3557 if (idx == -1) 3558 return -ENOENT; 3559 3560 bpf_get_linfo_source(btf, &linfo[idx], &src); 3561 while (isspace(*src.line)) 3562 src.line++; 3563 if (filep) 3564 *filep = src.file; 3565 if (linep) 3566 *linep = src.line; 3567 if (nump) 3568 *nump = src.line_num; 3569 return 0; 3570 } 3571 3572 struct walk_stack_ctx { 3573 struct bpf_prog *prog; 3574 }; 3575 3576 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3577 { 3578 struct walk_stack_ctx *ctxp = cookie; 3579 struct bpf_prog *prog; 3580 3581 /* 3582 * The RCU read lock is held to safely traverse the latch tree, but we 3583 * don't need its protection when accessing the prog, since it has an 3584 * active stack frame on the current stack trace, and won't disappear. 3585 */ 3586 rcu_read_lock(); 3587 prog = bpf_prog_ksym_find(ip); 3588 rcu_read_unlock(); 3589 if (!prog) 3590 return true; 3591 /* Make sure we return the main prog if we found a subprog */ 3592 ctxp->prog = prog->aux->main_prog_aux->prog; 3593 return false; 3594 } 3595 3596 struct bpf_prog *bpf_prog_find_from_stack(void) 3597 { 3598 struct walk_stack_ctx ctx = {}; 3599 3600 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3601 return ctx.prog; 3602 } 3603 3604 #endif 3605