1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Linux Socket Filter - Kernel level socket filtering 4 * 5 * Based on the design of the Berkeley Packet Filter. The new 6 * internal format has been designed by PLUMgrid: 7 * 8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 9 * 10 * Authors: 11 * 12 * Jay Schulist <jschlst@samba.org> 13 * Alexei Starovoitov <ast@plumgrid.com> 14 * Daniel Borkmann <dborkman@redhat.com> 15 * 16 * Andi Kleen - Fix a few bad bugs and races. 17 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 18 */ 19 20 #include <uapi/linux/btf.h> 21 #include <crypto/sha1.h> 22 #include <linux/filter.h> 23 #include <linux/skbuff.h> 24 #include <linux/vmalloc.h> 25 #include <linux/prandom.h> 26 #include <linux/bpf.h> 27 #include <linux/btf.h> 28 #include <linux/objtool.h> 29 #include <linux/overflow.h> 30 #include <linux/rbtree_latch.h> 31 #include <linux/kallsyms.h> 32 #include <linux/rcupdate.h> 33 #include <linux/perf_event.h> 34 #include <linux/extable.h> 35 #include <linux/log2.h> 36 #include <linux/bpf_verifier.h> 37 #include <linux/nodemask.h> 38 #include <linux/nospec.h> 39 #include <linux/bpf_mem_alloc.h> 40 #include <linux/memcontrol.h> 41 #include <linux/execmem.h> 42 43 #include <asm/barrier.h> 44 #include <linux/unaligned.h> 45 46 /* Registers */ 47 #define BPF_R0 regs[BPF_REG_0] 48 #define BPF_R1 regs[BPF_REG_1] 49 #define BPF_R2 regs[BPF_REG_2] 50 #define BPF_R3 regs[BPF_REG_3] 51 #define BPF_R4 regs[BPF_REG_4] 52 #define BPF_R5 regs[BPF_REG_5] 53 #define BPF_R6 regs[BPF_REG_6] 54 #define BPF_R7 regs[BPF_REG_7] 55 #define BPF_R8 regs[BPF_REG_8] 56 #define BPF_R9 regs[BPF_REG_9] 57 #define BPF_R10 regs[BPF_REG_10] 58 59 /* Named registers */ 60 #define DST regs[insn->dst_reg] 61 #define SRC regs[insn->src_reg] 62 #define FP regs[BPF_REG_FP] 63 #define AX regs[BPF_REG_AX] 64 #define ARG1 regs[BPF_REG_ARG1] 65 #define CTX regs[BPF_REG_CTX] 66 #define OFF insn->off 67 #define IMM insn->imm 68 69 struct bpf_mem_alloc bpf_global_ma; 70 bool bpf_global_ma_set; 71 72 /* No hurry in this branch 73 * 74 * Exported for the bpf jit load helper. 75 */ 76 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 77 { 78 u8 *ptr = NULL; 79 80 if (k >= SKF_NET_OFF) { 81 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 82 } else if (k >= SKF_LL_OFF) { 83 if (unlikely(!skb_mac_header_was_set(skb))) 84 return NULL; 85 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 86 } 87 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 88 return ptr; 89 90 return NULL; 91 } 92 93 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 94 enum page_size_enum { 95 __PAGE_SIZE = PAGE_SIZE 96 }; 97 98 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 99 { 100 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 101 struct bpf_prog_aux *aux; 102 struct bpf_prog *fp; 103 104 size = round_up(size, __PAGE_SIZE); 105 fp = __vmalloc(size, gfp_flags); 106 if (fp == NULL) 107 return NULL; 108 109 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 110 if (aux == NULL) { 111 vfree(fp); 112 return NULL; 113 } 114 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 115 if (!fp->active) { 116 vfree(fp); 117 kfree(aux); 118 return NULL; 119 } 120 121 fp->pages = size / PAGE_SIZE; 122 fp->aux = aux; 123 fp->aux->main_prog_aux = aux; 124 fp->aux->prog = fp; 125 fp->jit_requested = ebpf_jit_enabled(); 126 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 127 #ifdef CONFIG_CGROUP_BPF 128 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 129 #endif 130 131 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 132 #ifdef CONFIG_FINEIBT 133 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 134 #endif 135 mutex_init(&fp->aux->used_maps_mutex); 136 mutex_init(&fp->aux->ext_mutex); 137 mutex_init(&fp->aux->dst_mutex); 138 139 #ifdef CONFIG_BPF_SYSCALL 140 bpf_prog_stream_init(fp); 141 #endif 142 143 return fp; 144 } 145 146 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 147 { 148 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 149 struct bpf_prog *prog; 150 int cpu; 151 152 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 153 if (!prog) 154 return NULL; 155 156 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 157 if (!prog->stats) { 158 free_percpu(prog->active); 159 kfree(prog->aux); 160 vfree(prog); 161 return NULL; 162 } 163 164 for_each_possible_cpu(cpu) { 165 struct bpf_prog_stats *pstats; 166 167 pstats = per_cpu_ptr(prog->stats, cpu); 168 u64_stats_init(&pstats->syncp); 169 } 170 return prog; 171 } 172 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 173 174 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 175 { 176 if (!prog->aux->nr_linfo || !prog->jit_requested) 177 return 0; 178 179 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 180 sizeof(*prog->aux->jited_linfo), 181 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 182 if (!prog->aux->jited_linfo) 183 return -ENOMEM; 184 185 return 0; 186 } 187 188 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 189 { 190 if (prog->aux->jited_linfo && 191 (!prog->jited || !prog->aux->jited_linfo[0])) { 192 kvfree(prog->aux->jited_linfo); 193 prog->aux->jited_linfo = NULL; 194 } 195 196 kfree(prog->aux->kfunc_tab); 197 prog->aux->kfunc_tab = NULL; 198 } 199 200 /* The jit engine is responsible to provide an array 201 * for insn_off to the jited_off mapping (insn_to_jit_off). 202 * 203 * The idx to this array is the insn_off. Hence, the insn_off 204 * here is relative to the prog itself instead of the main prog. 205 * This array has one entry for each xlated bpf insn. 206 * 207 * jited_off is the byte off to the end of the jited insn. 208 * 209 * Hence, with 210 * insn_start: 211 * The first bpf insn off of the prog. The insn off 212 * here is relative to the main prog. 213 * e.g. if prog is a subprog, insn_start > 0 214 * linfo_idx: 215 * The prog's idx to prog->aux->linfo and jited_linfo 216 * 217 * jited_linfo[linfo_idx] = prog->bpf_func 218 * 219 * For i > linfo_idx, 220 * 221 * jited_linfo[i] = prog->bpf_func + 222 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 223 */ 224 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 225 const u32 *insn_to_jit_off) 226 { 227 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 228 const struct bpf_line_info *linfo; 229 void **jited_linfo; 230 231 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 232 /* Userspace did not provide linfo */ 233 return; 234 235 linfo_idx = prog->aux->linfo_idx; 236 linfo = &prog->aux->linfo[linfo_idx]; 237 insn_start = linfo[0].insn_off; 238 insn_end = insn_start + prog->len; 239 240 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 241 jited_linfo[0] = prog->bpf_func; 242 243 nr_linfo = prog->aux->nr_linfo - linfo_idx; 244 245 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 246 /* The verifier ensures that linfo[i].insn_off is 247 * strictly increasing 248 */ 249 jited_linfo[i] = prog->bpf_func + 250 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 251 } 252 253 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 254 gfp_t gfp_extra_flags) 255 { 256 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 257 struct bpf_prog *fp; 258 u32 pages; 259 260 size = round_up(size, PAGE_SIZE); 261 pages = size / PAGE_SIZE; 262 if (pages <= fp_old->pages) 263 return fp_old; 264 265 fp = __vmalloc(size, gfp_flags); 266 if (fp) { 267 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 268 fp->pages = pages; 269 fp->aux->prog = fp; 270 271 /* We keep fp->aux from fp_old around in the new 272 * reallocated structure. 273 */ 274 fp_old->aux = NULL; 275 fp_old->stats = NULL; 276 fp_old->active = NULL; 277 __bpf_prog_free(fp_old); 278 } 279 280 return fp; 281 } 282 283 void __bpf_prog_free(struct bpf_prog *fp) 284 { 285 if (fp->aux) { 286 mutex_destroy(&fp->aux->used_maps_mutex); 287 mutex_destroy(&fp->aux->dst_mutex); 288 kfree(fp->aux->poke_tab); 289 kfree(fp->aux); 290 } 291 free_percpu(fp->stats); 292 free_percpu(fp->active); 293 vfree(fp); 294 } 295 296 int bpf_prog_calc_tag(struct bpf_prog *fp) 297 { 298 size_t size = bpf_prog_insn_size(fp); 299 u8 digest[SHA1_DIGEST_SIZE]; 300 struct bpf_insn *dst; 301 bool was_ld_map; 302 u32 i; 303 304 dst = vmalloc(size); 305 if (!dst) 306 return -ENOMEM; 307 308 /* We need to take out the map fd for the digest calculation 309 * since they are unstable from user space side. 310 */ 311 for (i = 0, was_ld_map = false; i < fp->len; i++) { 312 dst[i] = fp->insnsi[i]; 313 if (!was_ld_map && 314 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 315 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 316 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 317 was_ld_map = true; 318 dst[i].imm = 0; 319 } else if (was_ld_map && 320 dst[i].code == 0 && 321 dst[i].dst_reg == 0 && 322 dst[i].src_reg == 0 && 323 dst[i].off == 0) { 324 was_ld_map = false; 325 dst[i].imm = 0; 326 } else { 327 was_ld_map = false; 328 } 329 } 330 sha1((const u8 *)dst, size, digest); 331 memcpy(fp->tag, digest, sizeof(fp->tag)); 332 vfree(dst); 333 return 0; 334 } 335 336 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 337 s32 end_new, s32 curr, const bool probe_pass) 338 { 339 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 340 s32 delta = end_new - end_old; 341 s64 imm = insn->imm; 342 343 if (curr < pos && curr + imm + 1 >= end_old) 344 imm += delta; 345 else if (curr >= end_new && curr + imm + 1 < end_new) 346 imm -= delta; 347 if (imm < imm_min || imm > imm_max) 348 return -ERANGE; 349 if (!probe_pass) 350 insn->imm = imm; 351 return 0; 352 } 353 354 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 355 s32 end_new, s32 curr, const bool probe_pass) 356 { 357 s64 off_min, off_max, off; 358 s32 delta = end_new - end_old; 359 360 if (insn->code == (BPF_JMP32 | BPF_JA)) { 361 off = insn->imm; 362 off_min = S32_MIN; 363 off_max = S32_MAX; 364 } else { 365 off = insn->off; 366 off_min = S16_MIN; 367 off_max = S16_MAX; 368 } 369 370 if (curr < pos && curr + off + 1 >= end_old) 371 off += delta; 372 else if (curr >= end_new && curr + off + 1 < end_new) 373 off -= delta; 374 if (off < off_min || off > off_max) 375 return -ERANGE; 376 if (!probe_pass) { 377 if (insn->code == (BPF_JMP32 | BPF_JA)) 378 insn->imm = off; 379 else 380 insn->off = off; 381 } 382 return 0; 383 } 384 385 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 386 s32 end_new, const bool probe_pass) 387 { 388 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 389 struct bpf_insn *insn = prog->insnsi; 390 int ret = 0; 391 392 for (i = 0; i < insn_cnt; i++, insn++) { 393 u8 code; 394 395 /* In the probing pass we still operate on the original, 396 * unpatched image in order to check overflows before we 397 * do any other adjustments. Therefore skip the patchlet. 398 */ 399 if (probe_pass && i == pos) { 400 i = end_new; 401 insn = prog->insnsi + end_old; 402 } 403 if (bpf_pseudo_func(insn)) { 404 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 405 end_new, i, probe_pass); 406 if (ret) 407 return ret; 408 continue; 409 } 410 code = insn->code; 411 if ((BPF_CLASS(code) != BPF_JMP && 412 BPF_CLASS(code) != BPF_JMP32) || 413 BPF_OP(code) == BPF_EXIT) 414 continue; 415 /* Adjust offset of jmps if we cross patch boundaries. */ 416 if (BPF_OP(code) == BPF_CALL) { 417 if (insn->src_reg != BPF_PSEUDO_CALL) 418 continue; 419 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 420 end_new, i, probe_pass); 421 } else { 422 ret = bpf_adj_delta_to_off(insn, pos, end_old, 423 end_new, i, probe_pass); 424 } 425 if (ret) 426 break; 427 } 428 429 return ret; 430 } 431 432 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 433 { 434 struct bpf_line_info *linfo; 435 u32 i, nr_linfo; 436 437 nr_linfo = prog->aux->nr_linfo; 438 if (!nr_linfo || !delta) 439 return; 440 441 linfo = prog->aux->linfo; 442 443 for (i = 0; i < nr_linfo; i++) 444 if (off < linfo[i].insn_off) 445 break; 446 447 /* Push all off < linfo[i].insn_off by delta */ 448 for (; i < nr_linfo; i++) 449 linfo[i].insn_off += delta; 450 } 451 452 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 453 const struct bpf_insn *patch, u32 len) 454 { 455 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 456 const u32 cnt_max = S16_MAX; 457 struct bpf_prog *prog_adj; 458 int err; 459 460 /* Since our patchlet doesn't expand the image, we're done. */ 461 if (insn_delta == 0) { 462 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 463 return prog; 464 } 465 466 insn_adj_cnt = prog->len + insn_delta; 467 468 /* Reject anything that would potentially let the insn->off 469 * target overflow when we have excessive program expansions. 470 * We need to probe here before we do any reallocation where 471 * we afterwards may not fail anymore. 472 */ 473 if (insn_adj_cnt > cnt_max && 474 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 475 return ERR_PTR(err); 476 477 /* Several new instructions need to be inserted. Make room 478 * for them. Likely, there's no need for a new allocation as 479 * last page could have large enough tailroom. 480 */ 481 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 482 GFP_USER); 483 if (!prog_adj) 484 return ERR_PTR(-ENOMEM); 485 486 prog_adj->len = insn_adj_cnt; 487 488 /* Patching happens in 3 steps: 489 * 490 * 1) Move over tail of insnsi from next instruction onwards, 491 * so we can patch the single target insn with one or more 492 * new ones (patching is always from 1 to n insns, n > 0). 493 * 2) Inject new instructions at the target location. 494 * 3) Adjust branch offsets if necessary. 495 */ 496 insn_rest = insn_adj_cnt - off - len; 497 498 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 499 sizeof(*patch) * insn_rest); 500 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 501 502 /* We are guaranteed to not fail at this point, otherwise 503 * the ship has sailed to reverse to the original state. An 504 * overflow cannot happen at this point. 505 */ 506 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 507 508 bpf_adj_linfo(prog_adj, off, insn_delta); 509 510 return prog_adj; 511 } 512 513 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 514 { 515 int err; 516 517 /* Branch offsets can't overflow when program is shrinking, no need 518 * to call bpf_adj_branches(..., true) here 519 */ 520 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 521 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 522 prog->len -= cnt; 523 524 err = bpf_adj_branches(prog, off, off + cnt, off, false); 525 WARN_ON_ONCE(err); 526 return err; 527 } 528 529 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 530 { 531 int i; 532 533 for (i = 0; i < fp->aux->real_func_cnt; i++) 534 bpf_prog_kallsyms_del(fp->aux->func[i]); 535 } 536 537 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 538 { 539 bpf_prog_kallsyms_del_subprogs(fp); 540 bpf_prog_kallsyms_del(fp); 541 } 542 543 #ifdef CONFIG_BPF_JIT 544 /* All BPF JIT sysctl knobs here. */ 545 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 546 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 547 int bpf_jit_harden __read_mostly; 548 long bpf_jit_limit __read_mostly; 549 long bpf_jit_limit_max __read_mostly; 550 551 static void 552 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 553 { 554 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 555 556 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 557 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 558 } 559 560 static void 561 bpf_prog_ksym_set_name(struct bpf_prog *prog) 562 { 563 char *sym = prog->aux->ksym.name; 564 const char *end = sym + KSYM_NAME_LEN; 565 const struct btf_type *type; 566 const char *func_name; 567 568 BUILD_BUG_ON(sizeof("bpf_prog_") + 569 sizeof(prog->tag) * 2 + 570 /* name has been null terminated. 571 * We should need +1 for the '_' preceding 572 * the name. However, the null character 573 * is double counted between the name and the 574 * sizeof("bpf_prog_") above, so we omit 575 * the +1 here. 576 */ 577 sizeof(prog->aux->name) > KSYM_NAME_LEN); 578 579 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 580 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 581 582 /* prog->aux->name will be ignored if full btf name is available */ 583 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 584 type = btf_type_by_id(prog->aux->btf, 585 prog->aux->func_info[prog->aux->func_idx].type_id); 586 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 587 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 588 return; 589 } 590 591 if (prog->aux->name[0]) 592 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 593 else 594 *sym = 0; 595 } 596 597 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 598 { 599 return container_of(n, struct bpf_ksym, tnode)->start; 600 } 601 602 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 603 struct latch_tree_node *b) 604 { 605 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 606 } 607 608 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 609 { 610 unsigned long val = (unsigned long)key; 611 const struct bpf_ksym *ksym; 612 613 ksym = container_of(n, struct bpf_ksym, tnode); 614 615 if (val < ksym->start) 616 return -1; 617 /* Ensure that we detect return addresses as part of the program, when 618 * the final instruction is a call for a program part of the stack 619 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 620 */ 621 if (val > ksym->end) 622 return 1; 623 624 return 0; 625 } 626 627 static const struct latch_tree_ops bpf_tree_ops = { 628 .less = bpf_tree_less, 629 .comp = bpf_tree_comp, 630 }; 631 632 static DEFINE_SPINLOCK(bpf_lock); 633 static LIST_HEAD(bpf_kallsyms); 634 static struct latch_tree_root bpf_tree __cacheline_aligned; 635 636 void bpf_ksym_add(struct bpf_ksym *ksym) 637 { 638 spin_lock_bh(&bpf_lock); 639 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 640 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 641 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 642 spin_unlock_bh(&bpf_lock); 643 } 644 645 static void __bpf_ksym_del(struct bpf_ksym *ksym) 646 { 647 if (list_empty(&ksym->lnode)) 648 return; 649 650 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 651 list_del_rcu(&ksym->lnode); 652 } 653 654 void bpf_ksym_del(struct bpf_ksym *ksym) 655 { 656 spin_lock_bh(&bpf_lock); 657 __bpf_ksym_del(ksym); 658 spin_unlock_bh(&bpf_lock); 659 } 660 661 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 662 { 663 return fp->jited && !bpf_prog_was_classic(fp); 664 } 665 666 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 667 { 668 if (!bpf_prog_kallsyms_candidate(fp) || 669 !bpf_token_capable(fp->aux->token, CAP_BPF)) 670 return; 671 672 bpf_prog_ksym_set_addr(fp); 673 bpf_prog_ksym_set_name(fp); 674 fp->aux->ksym.prog = true; 675 676 bpf_ksym_add(&fp->aux->ksym); 677 678 #ifdef CONFIG_FINEIBT 679 /* 680 * When FineIBT, code in the __cfi_foo() symbols can get executed 681 * and hence unwinder needs help. 682 */ 683 if (cfi_mode != CFI_FINEIBT) 684 return; 685 686 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 687 "__cfi_%s", fp->aux->ksym.name); 688 689 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 690 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 691 692 bpf_ksym_add(&fp->aux->ksym_prefix); 693 #endif 694 } 695 696 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 697 { 698 if (!bpf_prog_kallsyms_candidate(fp)) 699 return; 700 701 bpf_ksym_del(&fp->aux->ksym); 702 #ifdef CONFIG_FINEIBT 703 if (cfi_mode != CFI_FINEIBT) 704 return; 705 bpf_ksym_del(&fp->aux->ksym_prefix); 706 #endif 707 } 708 709 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 710 { 711 struct latch_tree_node *n; 712 713 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 714 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 715 } 716 717 int __bpf_address_lookup(unsigned long addr, unsigned long *size, 718 unsigned long *off, char *sym) 719 { 720 struct bpf_ksym *ksym; 721 int ret = 0; 722 723 rcu_read_lock(); 724 ksym = bpf_ksym_find(addr); 725 if (ksym) { 726 unsigned long symbol_start = ksym->start; 727 unsigned long symbol_end = ksym->end; 728 729 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 730 731 if (size) 732 *size = symbol_end - symbol_start; 733 if (off) 734 *off = addr - symbol_start; 735 } 736 rcu_read_unlock(); 737 738 return ret; 739 } 740 741 bool is_bpf_text_address(unsigned long addr) 742 { 743 bool ret; 744 745 rcu_read_lock(); 746 ret = bpf_ksym_find(addr) != NULL; 747 rcu_read_unlock(); 748 749 return ret; 750 } 751 752 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 753 { 754 struct bpf_ksym *ksym; 755 756 WARN_ON_ONCE(!rcu_read_lock_held()); 757 ksym = bpf_ksym_find(addr); 758 759 return ksym && ksym->prog ? 760 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 761 NULL; 762 } 763 764 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 765 { 766 const struct exception_table_entry *e = NULL; 767 struct bpf_prog *prog; 768 769 rcu_read_lock(); 770 prog = bpf_prog_ksym_find(addr); 771 if (!prog) 772 goto out; 773 if (!prog->aux->num_exentries) 774 goto out; 775 776 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 777 out: 778 rcu_read_unlock(); 779 return e; 780 } 781 782 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 783 char *sym) 784 { 785 struct bpf_ksym *ksym; 786 unsigned int it = 0; 787 int ret = -ERANGE; 788 789 if (!bpf_jit_kallsyms_enabled()) 790 return ret; 791 792 rcu_read_lock(); 793 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 794 if (it++ != symnum) 795 continue; 796 797 strscpy(sym, ksym->name, KSYM_NAME_LEN); 798 799 *value = ksym->start; 800 *type = BPF_SYM_ELF_TYPE; 801 802 ret = 0; 803 break; 804 } 805 rcu_read_unlock(); 806 807 return ret; 808 } 809 810 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 811 struct bpf_jit_poke_descriptor *poke) 812 { 813 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 814 static const u32 poke_tab_max = 1024; 815 u32 slot = prog->aux->size_poke_tab; 816 u32 size = slot + 1; 817 818 if (size > poke_tab_max) 819 return -ENOSPC; 820 if (poke->tailcall_target || poke->tailcall_target_stable || 821 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 822 return -EINVAL; 823 824 switch (poke->reason) { 825 case BPF_POKE_REASON_TAIL_CALL: 826 if (!poke->tail_call.map) 827 return -EINVAL; 828 break; 829 default: 830 return -EINVAL; 831 } 832 833 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 834 if (!tab) 835 return -ENOMEM; 836 837 memcpy(&tab[slot], poke, sizeof(*poke)); 838 prog->aux->size_poke_tab = size; 839 prog->aux->poke_tab = tab; 840 841 return slot; 842 } 843 844 /* 845 * BPF program pack allocator. 846 * 847 * Most BPF programs are pretty small. Allocating a hole page for each 848 * program is sometime a waste. Many small bpf program also adds pressure 849 * to instruction TLB. To solve this issue, we introduce a BPF program pack 850 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 851 * to host BPF programs. 852 */ 853 #define BPF_PROG_CHUNK_SHIFT 6 854 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 855 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 856 857 struct bpf_prog_pack { 858 struct list_head list; 859 void *ptr; 860 unsigned long bitmap[]; 861 }; 862 863 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 864 { 865 memset(area, 0, size); 866 } 867 868 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 869 870 static DEFINE_MUTEX(pack_mutex); 871 static LIST_HEAD(pack_list); 872 873 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 874 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 875 */ 876 #ifdef PMD_SIZE 877 /* PMD_SIZE is really big for some archs. It doesn't make sense to 878 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 879 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 880 * greater than or equal to 2MB. 881 */ 882 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 883 #else 884 #define BPF_PROG_PACK_SIZE PAGE_SIZE 885 #endif 886 887 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 888 889 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 890 { 891 struct bpf_prog_pack *pack; 892 int err; 893 894 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 895 GFP_KERNEL); 896 if (!pack) 897 return NULL; 898 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 899 if (!pack->ptr) 900 goto out; 901 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 902 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 903 904 set_vm_flush_reset_perms(pack->ptr); 905 err = set_memory_rox((unsigned long)pack->ptr, 906 BPF_PROG_PACK_SIZE / PAGE_SIZE); 907 if (err) 908 goto out; 909 list_add_tail(&pack->list, &pack_list); 910 return pack; 911 912 out: 913 bpf_jit_free_exec(pack->ptr); 914 kfree(pack); 915 return NULL; 916 } 917 918 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 919 { 920 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 921 struct bpf_prog_pack *pack; 922 unsigned long pos; 923 void *ptr = NULL; 924 925 mutex_lock(&pack_mutex); 926 if (size > BPF_PROG_PACK_SIZE) { 927 size = round_up(size, PAGE_SIZE); 928 ptr = bpf_jit_alloc_exec(size); 929 if (ptr) { 930 int err; 931 932 bpf_fill_ill_insns(ptr, size); 933 set_vm_flush_reset_perms(ptr); 934 err = set_memory_rox((unsigned long)ptr, 935 size / PAGE_SIZE); 936 if (err) { 937 bpf_jit_free_exec(ptr); 938 ptr = NULL; 939 } 940 } 941 goto out; 942 } 943 list_for_each_entry(pack, &pack_list, list) { 944 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 945 nbits, 0); 946 if (pos < BPF_PROG_CHUNK_COUNT) 947 goto found_free_area; 948 } 949 950 pack = alloc_new_pack(bpf_fill_ill_insns); 951 if (!pack) 952 goto out; 953 954 pos = 0; 955 956 found_free_area: 957 bitmap_set(pack->bitmap, pos, nbits); 958 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 959 960 out: 961 mutex_unlock(&pack_mutex); 962 return ptr; 963 } 964 965 void bpf_prog_pack_free(void *ptr, u32 size) 966 { 967 struct bpf_prog_pack *pack = NULL, *tmp; 968 unsigned int nbits; 969 unsigned long pos; 970 971 mutex_lock(&pack_mutex); 972 if (size > BPF_PROG_PACK_SIZE) { 973 bpf_jit_free_exec(ptr); 974 goto out; 975 } 976 977 list_for_each_entry(tmp, &pack_list, list) { 978 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 979 pack = tmp; 980 break; 981 } 982 } 983 984 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 985 goto out; 986 987 nbits = BPF_PROG_SIZE_TO_NBITS(size); 988 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 989 990 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 991 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 992 993 bitmap_clear(pack->bitmap, pos, nbits); 994 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 995 BPF_PROG_CHUNK_COUNT, 0) == 0) { 996 list_del(&pack->list); 997 bpf_jit_free_exec(pack->ptr); 998 kfree(pack); 999 } 1000 out: 1001 mutex_unlock(&pack_mutex); 1002 } 1003 1004 static atomic_long_t bpf_jit_current; 1005 1006 /* Can be overridden by an arch's JIT compiler if it has a custom, 1007 * dedicated BPF backend memory area, or if neither of the two 1008 * below apply. 1009 */ 1010 u64 __weak bpf_jit_alloc_exec_limit(void) 1011 { 1012 #if defined(MODULES_VADDR) 1013 return MODULES_END - MODULES_VADDR; 1014 #else 1015 return VMALLOC_END - VMALLOC_START; 1016 #endif 1017 } 1018 1019 static int __init bpf_jit_charge_init(void) 1020 { 1021 /* Only used as heuristic here to derive limit. */ 1022 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1023 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1024 PAGE_SIZE), LONG_MAX); 1025 return 0; 1026 } 1027 pure_initcall(bpf_jit_charge_init); 1028 1029 int bpf_jit_charge_modmem(u32 size) 1030 { 1031 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1032 if (!bpf_capable()) { 1033 atomic_long_sub(size, &bpf_jit_current); 1034 return -EPERM; 1035 } 1036 } 1037 1038 return 0; 1039 } 1040 1041 void bpf_jit_uncharge_modmem(u32 size) 1042 { 1043 atomic_long_sub(size, &bpf_jit_current); 1044 } 1045 1046 void *__weak bpf_jit_alloc_exec(unsigned long size) 1047 { 1048 return execmem_alloc(EXECMEM_BPF, size); 1049 } 1050 1051 void __weak bpf_jit_free_exec(void *addr) 1052 { 1053 execmem_free(addr); 1054 } 1055 1056 struct bpf_binary_header * 1057 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1058 unsigned int alignment, 1059 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1060 { 1061 struct bpf_binary_header *hdr; 1062 u32 size, hole, start; 1063 1064 WARN_ON_ONCE(!is_power_of_2(alignment) || 1065 alignment > BPF_IMAGE_ALIGNMENT); 1066 1067 /* Most of BPF filters are really small, but if some of them 1068 * fill a page, allow at least 128 extra bytes to insert a 1069 * random section of illegal instructions. 1070 */ 1071 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1072 1073 if (bpf_jit_charge_modmem(size)) 1074 return NULL; 1075 hdr = bpf_jit_alloc_exec(size); 1076 if (!hdr) { 1077 bpf_jit_uncharge_modmem(size); 1078 return NULL; 1079 } 1080 1081 /* Fill space with illegal/arch-dep instructions. */ 1082 bpf_fill_ill_insns(hdr, size); 1083 1084 hdr->size = size; 1085 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1086 PAGE_SIZE - sizeof(*hdr)); 1087 start = get_random_u32_below(hole) & ~(alignment - 1); 1088 1089 /* Leave a random number of instructions before BPF code. */ 1090 *image_ptr = &hdr->image[start]; 1091 1092 return hdr; 1093 } 1094 1095 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1096 { 1097 u32 size = hdr->size; 1098 1099 bpf_jit_free_exec(hdr); 1100 bpf_jit_uncharge_modmem(size); 1101 } 1102 1103 /* Allocate jit binary from bpf_prog_pack allocator. 1104 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1105 * to the memory. To solve this problem, a RW buffer is also allocated at 1106 * as the same time. The JIT engine should calculate offsets based on the 1107 * RO memory address, but write JITed program to the RW buffer. Once the 1108 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1109 * the JITed program to the RO memory. 1110 */ 1111 struct bpf_binary_header * 1112 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1113 unsigned int alignment, 1114 struct bpf_binary_header **rw_header, 1115 u8 **rw_image, 1116 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1117 { 1118 struct bpf_binary_header *ro_header; 1119 u32 size, hole, start; 1120 1121 WARN_ON_ONCE(!is_power_of_2(alignment) || 1122 alignment > BPF_IMAGE_ALIGNMENT); 1123 1124 /* add 16 bytes for a random section of illegal instructions */ 1125 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1126 1127 if (bpf_jit_charge_modmem(size)) 1128 return NULL; 1129 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1130 if (!ro_header) { 1131 bpf_jit_uncharge_modmem(size); 1132 return NULL; 1133 } 1134 1135 *rw_header = kvmalloc(size, GFP_KERNEL); 1136 if (!*rw_header) { 1137 bpf_prog_pack_free(ro_header, size); 1138 bpf_jit_uncharge_modmem(size); 1139 return NULL; 1140 } 1141 1142 /* Fill space with illegal/arch-dep instructions. */ 1143 bpf_fill_ill_insns(*rw_header, size); 1144 (*rw_header)->size = size; 1145 1146 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1147 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1148 start = get_random_u32_below(hole) & ~(alignment - 1); 1149 1150 *image_ptr = &ro_header->image[start]; 1151 *rw_image = &(*rw_header)->image[start]; 1152 1153 return ro_header; 1154 } 1155 1156 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1157 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1158 struct bpf_binary_header *rw_header) 1159 { 1160 void *ptr; 1161 1162 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1163 1164 kvfree(rw_header); 1165 1166 if (IS_ERR(ptr)) { 1167 bpf_prog_pack_free(ro_header, ro_header->size); 1168 return PTR_ERR(ptr); 1169 } 1170 return 0; 1171 } 1172 1173 /* bpf_jit_binary_pack_free is called in two different scenarios: 1174 * 1) when the program is freed after; 1175 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1176 * For case 2), we need to free both the RO memory and the RW buffer. 1177 * 1178 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1179 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1180 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1181 * bpf_arch_text_copy (when jit fails). 1182 */ 1183 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1184 struct bpf_binary_header *rw_header) 1185 { 1186 u32 size = ro_header->size; 1187 1188 bpf_prog_pack_free(ro_header, size); 1189 kvfree(rw_header); 1190 bpf_jit_uncharge_modmem(size); 1191 } 1192 1193 struct bpf_binary_header * 1194 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1195 { 1196 unsigned long real_start = (unsigned long)fp->bpf_func; 1197 unsigned long addr; 1198 1199 addr = real_start & BPF_PROG_CHUNK_MASK; 1200 return (void *)addr; 1201 } 1202 1203 static inline struct bpf_binary_header * 1204 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1205 { 1206 unsigned long real_start = (unsigned long)fp->bpf_func; 1207 unsigned long addr; 1208 1209 addr = real_start & PAGE_MASK; 1210 return (void *)addr; 1211 } 1212 1213 /* This symbol is only overridden by archs that have different 1214 * requirements than the usual eBPF JITs, f.e. when they only 1215 * implement cBPF JIT, do not set images read-only, etc. 1216 */ 1217 void __weak bpf_jit_free(struct bpf_prog *fp) 1218 { 1219 if (fp->jited) { 1220 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1221 1222 bpf_jit_binary_free(hdr); 1223 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1224 } 1225 1226 bpf_prog_unlock_free(fp); 1227 } 1228 1229 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1230 const struct bpf_insn *insn, bool extra_pass, 1231 u64 *func_addr, bool *func_addr_fixed) 1232 { 1233 s16 off = insn->off; 1234 s32 imm = insn->imm; 1235 u8 *addr; 1236 int err; 1237 1238 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1239 if (!*func_addr_fixed) { 1240 /* Place-holder address till the last pass has collected 1241 * all addresses for JITed subprograms in which case we 1242 * can pick them up from prog->aux. 1243 */ 1244 if (!extra_pass) 1245 addr = NULL; 1246 else if (prog->aux->func && 1247 off >= 0 && off < prog->aux->real_func_cnt) 1248 addr = (u8 *)prog->aux->func[off]->bpf_func; 1249 else 1250 return -EINVAL; 1251 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1252 bpf_jit_supports_far_kfunc_call()) { 1253 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1254 if (err) 1255 return err; 1256 } else { 1257 /* Address of a BPF helper call. Since part of the core 1258 * kernel, it's always at a fixed location. __bpf_call_base 1259 * and the helper with imm relative to it are both in core 1260 * kernel. 1261 */ 1262 addr = (u8 *)__bpf_call_base + imm; 1263 } 1264 1265 *func_addr = (unsigned long)addr; 1266 return 0; 1267 } 1268 1269 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1270 { 1271 if (prog->aux->ksym.prog) 1272 return prog->aux->ksym.name; 1273 return prog->aux->name; 1274 } 1275 1276 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1277 const struct bpf_insn *aux, 1278 struct bpf_insn *to_buff, 1279 bool emit_zext) 1280 { 1281 struct bpf_insn *to = to_buff; 1282 u32 imm_rnd = get_random_u32(); 1283 s16 off; 1284 1285 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1286 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1287 1288 /* Constraints on AX register: 1289 * 1290 * AX register is inaccessible from user space. It is mapped in 1291 * all JITs, and used here for constant blinding rewrites. It is 1292 * typically "stateless" meaning its contents are only valid within 1293 * the executed instruction, but not across several instructions. 1294 * There are a few exceptions however which are further detailed 1295 * below. 1296 * 1297 * Constant blinding is only used by JITs, not in the interpreter. 1298 * The interpreter uses AX in some occasions as a local temporary 1299 * register e.g. in DIV or MOD instructions. 1300 * 1301 * In restricted circumstances, the verifier can also use the AX 1302 * register for rewrites as long as they do not interfere with 1303 * the above cases! 1304 */ 1305 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1306 goto out; 1307 1308 if (from->imm == 0 && 1309 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1310 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1311 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1312 goto out; 1313 } 1314 1315 switch (from->code) { 1316 case BPF_ALU | BPF_ADD | BPF_K: 1317 case BPF_ALU | BPF_SUB | BPF_K: 1318 case BPF_ALU | BPF_AND | BPF_K: 1319 case BPF_ALU | BPF_OR | BPF_K: 1320 case BPF_ALU | BPF_XOR | BPF_K: 1321 case BPF_ALU | BPF_MUL | BPF_K: 1322 case BPF_ALU | BPF_MOV | BPF_K: 1323 case BPF_ALU | BPF_DIV | BPF_K: 1324 case BPF_ALU | BPF_MOD | BPF_K: 1325 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1326 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1327 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1328 break; 1329 1330 case BPF_ALU64 | BPF_ADD | BPF_K: 1331 case BPF_ALU64 | BPF_SUB | BPF_K: 1332 case BPF_ALU64 | BPF_AND | BPF_K: 1333 case BPF_ALU64 | BPF_OR | BPF_K: 1334 case BPF_ALU64 | BPF_XOR | BPF_K: 1335 case BPF_ALU64 | BPF_MUL | BPF_K: 1336 case BPF_ALU64 | BPF_MOV | BPF_K: 1337 case BPF_ALU64 | BPF_DIV | BPF_K: 1338 case BPF_ALU64 | BPF_MOD | BPF_K: 1339 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1340 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1341 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1342 break; 1343 1344 case BPF_JMP | BPF_JEQ | BPF_K: 1345 case BPF_JMP | BPF_JNE | BPF_K: 1346 case BPF_JMP | BPF_JGT | BPF_K: 1347 case BPF_JMP | BPF_JLT | BPF_K: 1348 case BPF_JMP | BPF_JGE | BPF_K: 1349 case BPF_JMP | BPF_JLE | BPF_K: 1350 case BPF_JMP | BPF_JSGT | BPF_K: 1351 case BPF_JMP | BPF_JSLT | BPF_K: 1352 case BPF_JMP | BPF_JSGE | BPF_K: 1353 case BPF_JMP | BPF_JSLE | BPF_K: 1354 case BPF_JMP | BPF_JSET | BPF_K: 1355 /* Accommodate for extra offset in case of a backjump. */ 1356 off = from->off; 1357 if (off < 0) 1358 off -= 2; 1359 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1360 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1361 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1362 break; 1363 1364 case BPF_JMP32 | BPF_JEQ | BPF_K: 1365 case BPF_JMP32 | BPF_JNE | BPF_K: 1366 case BPF_JMP32 | BPF_JGT | BPF_K: 1367 case BPF_JMP32 | BPF_JLT | BPF_K: 1368 case BPF_JMP32 | BPF_JGE | BPF_K: 1369 case BPF_JMP32 | BPF_JLE | BPF_K: 1370 case BPF_JMP32 | BPF_JSGT | BPF_K: 1371 case BPF_JMP32 | BPF_JSLT | BPF_K: 1372 case BPF_JMP32 | BPF_JSGE | BPF_K: 1373 case BPF_JMP32 | BPF_JSLE | BPF_K: 1374 case BPF_JMP32 | BPF_JSET | BPF_K: 1375 /* Accommodate for extra offset in case of a backjump. */ 1376 off = from->off; 1377 if (off < 0) 1378 off -= 2; 1379 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1380 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1381 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1382 off); 1383 break; 1384 1385 case BPF_LD | BPF_IMM | BPF_DW: 1386 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1387 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1388 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1389 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1390 break; 1391 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1392 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1393 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1394 if (emit_zext) 1395 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1396 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1397 break; 1398 1399 case BPF_ST | BPF_MEM | BPF_DW: 1400 case BPF_ST | BPF_MEM | BPF_W: 1401 case BPF_ST | BPF_MEM | BPF_H: 1402 case BPF_ST | BPF_MEM | BPF_B: 1403 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1404 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1405 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1406 break; 1407 } 1408 out: 1409 return to - to_buff; 1410 } 1411 1412 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1413 gfp_t gfp_extra_flags) 1414 { 1415 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1416 struct bpf_prog *fp; 1417 1418 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1419 if (fp != NULL) { 1420 /* aux->prog still points to the fp_other one, so 1421 * when promoting the clone to the real program, 1422 * this still needs to be adapted. 1423 */ 1424 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1425 } 1426 1427 return fp; 1428 } 1429 1430 static void bpf_prog_clone_free(struct bpf_prog *fp) 1431 { 1432 /* aux was stolen by the other clone, so we cannot free 1433 * it from this path! It will be freed eventually by the 1434 * other program on release. 1435 * 1436 * At this point, we don't need a deferred release since 1437 * clone is guaranteed to not be locked. 1438 */ 1439 fp->aux = NULL; 1440 fp->stats = NULL; 1441 fp->active = NULL; 1442 __bpf_prog_free(fp); 1443 } 1444 1445 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1446 { 1447 /* We have to repoint aux->prog to self, as we don't 1448 * know whether fp here is the clone or the original. 1449 */ 1450 fp->aux->prog = fp; 1451 bpf_prog_clone_free(fp_other); 1452 } 1453 1454 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1455 { 1456 struct bpf_insn insn_buff[16], aux[2]; 1457 struct bpf_prog *clone, *tmp; 1458 int insn_delta, insn_cnt; 1459 struct bpf_insn *insn; 1460 int i, rewritten; 1461 1462 if (!prog->blinding_requested || prog->blinded) 1463 return prog; 1464 1465 clone = bpf_prog_clone_create(prog, GFP_USER); 1466 if (!clone) 1467 return ERR_PTR(-ENOMEM); 1468 1469 insn_cnt = clone->len; 1470 insn = clone->insnsi; 1471 1472 for (i = 0; i < insn_cnt; i++, insn++) { 1473 if (bpf_pseudo_func(insn)) { 1474 /* ld_imm64 with an address of bpf subprog is not 1475 * a user controlled constant. Don't randomize it, 1476 * since it will conflict with jit_subprogs() logic. 1477 */ 1478 insn++; 1479 i++; 1480 continue; 1481 } 1482 1483 /* We temporarily need to hold the original ld64 insn 1484 * so that we can still access the first part in the 1485 * second blinding run. 1486 */ 1487 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1488 insn[1].code == 0) 1489 memcpy(aux, insn, sizeof(aux)); 1490 1491 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1492 clone->aux->verifier_zext); 1493 if (!rewritten) 1494 continue; 1495 1496 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1497 if (IS_ERR(tmp)) { 1498 /* Patching may have repointed aux->prog during 1499 * realloc from the original one, so we need to 1500 * fix it up here on error. 1501 */ 1502 bpf_jit_prog_release_other(prog, clone); 1503 return tmp; 1504 } 1505 1506 clone = tmp; 1507 insn_delta = rewritten - 1; 1508 1509 /* Walk new program and skip insns we just inserted. */ 1510 insn = clone->insnsi + i + insn_delta; 1511 insn_cnt += insn_delta; 1512 i += insn_delta; 1513 } 1514 1515 clone->blinded = 1; 1516 return clone; 1517 } 1518 #endif /* CONFIG_BPF_JIT */ 1519 1520 /* Base function for offset calculation. Needs to go into .text section, 1521 * therefore keeping it non-static as well; will also be used by JITs 1522 * anyway later on, so do not let the compiler omit it. This also needs 1523 * to go into kallsyms for correlation from e.g. bpftool, so naming 1524 * must not change. 1525 */ 1526 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1527 { 1528 return 0; 1529 } 1530 EXPORT_SYMBOL_GPL(__bpf_call_base); 1531 1532 /* All UAPI available opcodes. */ 1533 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1534 /* 32 bit ALU operations. */ \ 1535 /* Register based. */ \ 1536 INSN_3(ALU, ADD, X), \ 1537 INSN_3(ALU, SUB, X), \ 1538 INSN_3(ALU, AND, X), \ 1539 INSN_3(ALU, OR, X), \ 1540 INSN_3(ALU, LSH, X), \ 1541 INSN_3(ALU, RSH, X), \ 1542 INSN_3(ALU, XOR, X), \ 1543 INSN_3(ALU, MUL, X), \ 1544 INSN_3(ALU, MOV, X), \ 1545 INSN_3(ALU, ARSH, X), \ 1546 INSN_3(ALU, DIV, X), \ 1547 INSN_3(ALU, MOD, X), \ 1548 INSN_2(ALU, NEG), \ 1549 INSN_3(ALU, END, TO_BE), \ 1550 INSN_3(ALU, END, TO_LE), \ 1551 /* Immediate based. */ \ 1552 INSN_3(ALU, ADD, K), \ 1553 INSN_3(ALU, SUB, K), \ 1554 INSN_3(ALU, AND, K), \ 1555 INSN_3(ALU, OR, K), \ 1556 INSN_3(ALU, LSH, K), \ 1557 INSN_3(ALU, RSH, K), \ 1558 INSN_3(ALU, XOR, K), \ 1559 INSN_3(ALU, MUL, K), \ 1560 INSN_3(ALU, MOV, K), \ 1561 INSN_3(ALU, ARSH, K), \ 1562 INSN_3(ALU, DIV, K), \ 1563 INSN_3(ALU, MOD, K), \ 1564 /* 64 bit ALU operations. */ \ 1565 /* Register based. */ \ 1566 INSN_3(ALU64, ADD, X), \ 1567 INSN_3(ALU64, SUB, X), \ 1568 INSN_3(ALU64, AND, X), \ 1569 INSN_3(ALU64, OR, X), \ 1570 INSN_3(ALU64, LSH, X), \ 1571 INSN_3(ALU64, RSH, X), \ 1572 INSN_3(ALU64, XOR, X), \ 1573 INSN_3(ALU64, MUL, X), \ 1574 INSN_3(ALU64, MOV, X), \ 1575 INSN_3(ALU64, ARSH, X), \ 1576 INSN_3(ALU64, DIV, X), \ 1577 INSN_3(ALU64, MOD, X), \ 1578 INSN_2(ALU64, NEG), \ 1579 INSN_3(ALU64, END, TO_LE), \ 1580 /* Immediate based. */ \ 1581 INSN_3(ALU64, ADD, K), \ 1582 INSN_3(ALU64, SUB, K), \ 1583 INSN_3(ALU64, AND, K), \ 1584 INSN_3(ALU64, OR, K), \ 1585 INSN_3(ALU64, LSH, K), \ 1586 INSN_3(ALU64, RSH, K), \ 1587 INSN_3(ALU64, XOR, K), \ 1588 INSN_3(ALU64, MUL, K), \ 1589 INSN_3(ALU64, MOV, K), \ 1590 INSN_3(ALU64, ARSH, K), \ 1591 INSN_3(ALU64, DIV, K), \ 1592 INSN_3(ALU64, MOD, K), \ 1593 /* Call instruction. */ \ 1594 INSN_2(JMP, CALL), \ 1595 /* Exit instruction. */ \ 1596 INSN_2(JMP, EXIT), \ 1597 /* 32-bit Jump instructions. */ \ 1598 /* Register based. */ \ 1599 INSN_3(JMP32, JEQ, X), \ 1600 INSN_3(JMP32, JNE, X), \ 1601 INSN_3(JMP32, JGT, X), \ 1602 INSN_3(JMP32, JLT, X), \ 1603 INSN_3(JMP32, JGE, X), \ 1604 INSN_3(JMP32, JLE, X), \ 1605 INSN_3(JMP32, JSGT, X), \ 1606 INSN_3(JMP32, JSLT, X), \ 1607 INSN_3(JMP32, JSGE, X), \ 1608 INSN_3(JMP32, JSLE, X), \ 1609 INSN_3(JMP32, JSET, X), \ 1610 /* Immediate based. */ \ 1611 INSN_3(JMP32, JEQ, K), \ 1612 INSN_3(JMP32, JNE, K), \ 1613 INSN_3(JMP32, JGT, K), \ 1614 INSN_3(JMP32, JLT, K), \ 1615 INSN_3(JMP32, JGE, K), \ 1616 INSN_3(JMP32, JLE, K), \ 1617 INSN_3(JMP32, JSGT, K), \ 1618 INSN_3(JMP32, JSLT, K), \ 1619 INSN_3(JMP32, JSGE, K), \ 1620 INSN_3(JMP32, JSLE, K), \ 1621 INSN_3(JMP32, JSET, K), \ 1622 /* Jump instructions. */ \ 1623 /* Register based. */ \ 1624 INSN_3(JMP, JEQ, X), \ 1625 INSN_3(JMP, JNE, X), \ 1626 INSN_3(JMP, JGT, X), \ 1627 INSN_3(JMP, JLT, X), \ 1628 INSN_3(JMP, JGE, X), \ 1629 INSN_3(JMP, JLE, X), \ 1630 INSN_3(JMP, JSGT, X), \ 1631 INSN_3(JMP, JSLT, X), \ 1632 INSN_3(JMP, JSGE, X), \ 1633 INSN_3(JMP, JSLE, X), \ 1634 INSN_3(JMP, JSET, X), \ 1635 /* Immediate based. */ \ 1636 INSN_3(JMP, JEQ, K), \ 1637 INSN_3(JMP, JNE, K), \ 1638 INSN_3(JMP, JGT, K), \ 1639 INSN_3(JMP, JLT, K), \ 1640 INSN_3(JMP, JGE, K), \ 1641 INSN_3(JMP, JLE, K), \ 1642 INSN_3(JMP, JSGT, K), \ 1643 INSN_3(JMP, JSLT, K), \ 1644 INSN_3(JMP, JSGE, K), \ 1645 INSN_3(JMP, JSLE, K), \ 1646 INSN_3(JMP, JSET, K), \ 1647 INSN_2(JMP, JA), \ 1648 INSN_2(JMP32, JA), \ 1649 /* Atomic operations. */ \ 1650 INSN_3(STX, ATOMIC, B), \ 1651 INSN_3(STX, ATOMIC, H), \ 1652 INSN_3(STX, ATOMIC, W), \ 1653 INSN_3(STX, ATOMIC, DW), \ 1654 /* Store instructions. */ \ 1655 /* Register based. */ \ 1656 INSN_3(STX, MEM, B), \ 1657 INSN_3(STX, MEM, H), \ 1658 INSN_3(STX, MEM, W), \ 1659 INSN_3(STX, MEM, DW), \ 1660 /* Immediate based. */ \ 1661 INSN_3(ST, MEM, B), \ 1662 INSN_3(ST, MEM, H), \ 1663 INSN_3(ST, MEM, W), \ 1664 INSN_3(ST, MEM, DW), \ 1665 /* Load instructions. */ \ 1666 /* Register based. */ \ 1667 INSN_3(LDX, MEM, B), \ 1668 INSN_3(LDX, MEM, H), \ 1669 INSN_3(LDX, MEM, W), \ 1670 INSN_3(LDX, MEM, DW), \ 1671 INSN_3(LDX, MEMSX, B), \ 1672 INSN_3(LDX, MEMSX, H), \ 1673 INSN_3(LDX, MEMSX, W), \ 1674 /* Immediate based. */ \ 1675 INSN_3(LD, IMM, DW) 1676 1677 bool bpf_opcode_in_insntable(u8 code) 1678 { 1679 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1680 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1681 static const bool public_insntable[256] = { 1682 [0 ... 255] = false, 1683 /* Now overwrite non-defaults ... */ 1684 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1685 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1686 [BPF_LD | BPF_ABS | BPF_B] = true, 1687 [BPF_LD | BPF_ABS | BPF_H] = true, 1688 [BPF_LD | BPF_ABS | BPF_W] = true, 1689 [BPF_LD | BPF_IND | BPF_B] = true, 1690 [BPF_LD | BPF_IND | BPF_H] = true, 1691 [BPF_LD | BPF_IND | BPF_W] = true, 1692 [BPF_JMP | BPF_JCOND] = true, 1693 }; 1694 #undef BPF_INSN_3_TBL 1695 #undef BPF_INSN_2_TBL 1696 return public_insntable[code]; 1697 } 1698 1699 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1700 /** 1701 * ___bpf_prog_run - run eBPF program on a given context 1702 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1703 * @insn: is the array of eBPF instructions 1704 * 1705 * Decode and execute eBPF instructions. 1706 * 1707 * Return: whatever value is in %BPF_R0 at program exit 1708 */ 1709 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1710 { 1711 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1712 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1713 static const void * const jumptable[256] __annotate_jump_table = { 1714 [0 ... 255] = &&default_label, 1715 /* Now overwrite non-defaults ... */ 1716 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1717 /* Non-UAPI available opcodes. */ 1718 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1719 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1720 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1721 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1722 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1723 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1724 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1725 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1726 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1727 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1728 }; 1729 #undef BPF_INSN_3_LBL 1730 #undef BPF_INSN_2_LBL 1731 u32 tail_call_cnt = 0; 1732 1733 #define CONT ({ insn++; goto select_insn; }) 1734 #define CONT_JMP ({ insn++; goto select_insn; }) 1735 1736 select_insn: 1737 goto *jumptable[insn->code]; 1738 1739 /* Explicitly mask the register-based shift amounts with 63 or 31 1740 * to avoid undefined behavior. Normally this won't affect the 1741 * generated code, for example, in case of native 64 bit archs such 1742 * as x86-64 or arm64, the compiler is optimizing the AND away for 1743 * the interpreter. In case of JITs, each of the JIT backends compiles 1744 * the BPF shift operations to machine instructions which produce 1745 * implementation-defined results in such a case; the resulting 1746 * contents of the register may be arbitrary, but program behaviour 1747 * as a whole remains defined. In other words, in case of JIT backends, 1748 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1749 */ 1750 /* ALU (shifts) */ 1751 #define SHT(OPCODE, OP) \ 1752 ALU64_##OPCODE##_X: \ 1753 DST = DST OP (SRC & 63); \ 1754 CONT; \ 1755 ALU_##OPCODE##_X: \ 1756 DST = (u32) DST OP ((u32) SRC & 31); \ 1757 CONT; \ 1758 ALU64_##OPCODE##_K: \ 1759 DST = DST OP IMM; \ 1760 CONT; \ 1761 ALU_##OPCODE##_K: \ 1762 DST = (u32) DST OP (u32) IMM; \ 1763 CONT; 1764 /* ALU (rest) */ 1765 #define ALU(OPCODE, OP) \ 1766 ALU64_##OPCODE##_X: \ 1767 DST = DST OP SRC; \ 1768 CONT; \ 1769 ALU_##OPCODE##_X: \ 1770 DST = (u32) DST OP (u32) SRC; \ 1771 CONT; \ 1772 ALU64_##OPCODE##_K: \ 1773 DST = DST OP IMM; \ 1774 CONT; \ 1775 ALU_##OPCODE##_K: \ 1776 DST = (u32) DST OP (u32) IMM; \ 1777 CONT; 1778 ALU(ADD, +) 1779 ALU(SUB, -) 1780 ALU(AND, &) 1781 ALU(OR, |) 1782 ALU(XOR, ^) 1783 ALU(MUL, *) 1784 SHT(LSH, <<) 1785 SHT(RSH, >>) 1786 #undef SHT 1787 #undef ALU 1788 ALU_NEG: 1789 DST = (u32) -DST; 1790 CONT; 1791 ALU64_NEG: 1792 DST = -DST; 1793 CONT; 1794 ALU_MOV_X: 1795 switch (OFF) { 1796 case 0: 1797 DST = (u32) SRC; 1798 break; 1799 case 8: 1800 DST = (u32)(s8) SRC; 1801 break; 1802 case 16: 1803 DST = (u32)(s16) SRC; 1804 break; 1805 } 1806 CONT; 1807 ALU_MOV_K: 1808 DST = (u32) IMM; 1809 CONT; 1810 ALU64_MOV_X: 1811 switch (OFF) { 1812 case 0: 1813 DST = SRC; 1814 break; 1815 case 8: 1816 DST = (s8) SRC; 1817 break; 1818 case 16: 1819 DST = (s16) SRC; 1820 break; 1821 case 32: 1822 DST = (s32) SRC; 1823 break; 1824 } 1825 CONT; 1826 ALU64_MOV_K: 1827 DST = IMM; 1828 CONT; 1829 LD_IMM_DW: 1830 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1831 insn++; 1832 CONT; 1833 ALU_ARSH_X: 1834 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1835 CONT; 1836 ALU_ARSH_K: 1837 DST = (u64) (u32) (((s32) DST) >> IMM); 1838 CONT; 1839 ALU64_ARSH_X: 1840 (*(s64 *) &DST) >>= (SRC & 63); 1841 CONT; 1842 ALU64_ARSH_K: 1843 (*(s64 *) &DST) >>= IMM; 1844 CONT; 1845 ALU64_MOD_X: 1846 switch (OFF) { 1847 case 0: 1848 div64_u64_rem(DST, SRC, &AX); 1849 DST = AX; 1850 break; 1851 case 1: 1852 AX = div64_s64(DST, SRC); 1853 DST = DST - AX * SRC; 1854 break; 1855 } 1856 CONT; 1857 ALU_MOD_X: 1858 switch (OFF) { 1859 case 0: 1860 AX = (u32) DST; 1861 DST = do_div(AX, (u32) SRC); 1862 break; 1863 case 1: 1864 AX = abs((s32)DST); 1865 AX = do_div(AX, abs((s32)SRC)); 1866 if ((s32)DST < 0) 1867 DST = (u32)-AX; 1868 else 1869 DST = (u32)AX; 1870 break; 1871 } 1872 CONT; 1873 ALU64_MOD_K: 1874 switch (OFF) { 1875 case 0: 1876 div64_u64_rem(DST, IMM, &AX); 1877 DST = AX; 1878 break; 1879 case 1: 1880 AX = div64_s64(DST, IMM); 1881 DST = DST - AX * IMM; 1882 break; 1883 } 1884 CONT; 1885 ALU_MOD_K: 1886 switch (OFF) { 1887 case 0: 1888 AX = (u32) DST; 1889 DST = do_div(AX, (u32) IMM); 1890 break; 1891 case 1: 1892 AX = abs((s32)DST); 1893 AX = do_div(AX, abs((s32)IMM)); 1894 if ((s32)DST < 0) 1895 DST = (u32)-AX; 1896 else 1897 DST = (u32)AX; 1898 break; 1899 } 1900 CONT; 1901 ALU64_DIV_X: 1902 switch (OFF) { 1903 case 0: 1904 DST = div64_u64(DST, SRC); 1905 break; 1906 case 1: 1907 DST = div64_s64(DST, SRC); 1908 break; 1909 } 1910 CONT; 1911 ALU_DIV_X: 1912 switch (OFF) { 1913 case 0: 1914 AX = (u32) DST; 1915 do_div(AX, (u32) SRC); 1916 DST = (u32) AX; 1917 break; 1918 case 1: 1919 AX = abs((s32)DST); 1920 do_div(AX, abs((s32)SRC)); 1921 if (((s32)DST < 0) == ((s32)SRC < 0)) 1922 DST = (u32)AX; 1923 else 1924 DST = (u32)-AX; 1925 break; 1926 } 1927 CONT; 1928 ALU64_DIV_K: 1929 switch (OFF) { 1930 case 0: 1931 DST = div64_u64(DST, IMM); 1932 break; 1933 case 1: 1934 DST = div64_s64(DST, IMM); 1935 break; 1936 } 1937 CONT; 1938 ALU_DIV_K: 1939 switch (OFF) { 1940 case 0: 1941 AX = (u32) DST; 1942 do_div(AX, (u32) IMM); 1943 DST = (u32) AX; 1944 break; 1945 case 1: 1946 AX = abs((s32)DST); 1947 do_div(AX, abs((s32)IMM)); 1948 if (((s32)DST < 0) == ((s32)IMM < 0)) 1949 DST = (u32)AX; 1950 else 1951 DST = (u32)-AX; 1952 break; 1953 } 1954 CONT; 1955 ALU_END_TO_BE: 1956 switch (IMM) { 1957 case 16: 1958 DST = (__force u16) cpu_to_be16(DST); 1959 break; 1960 case 32: 1961 DST = (__force u32) cpu_to_be32(DST); 1962 break; 1963 case 64: 1964 DST = (__force u64) cpu_to_be64(DST); 1965 break; 1966 } 1967 CONT; 1968 ALU_END_TO_LE: 1969 switch (IMM) { 1970 case 16: 1971 DST = (__force u16) cpu_to_le16(DST); 1972 break; 1973 case 32: 1974 DST = (__force u32) cpu_to_le32(DST); 1975 break; 1976 case 64: 1977 DST = (__force u64) cpu_to_le64(DST); 1978 break; 1979 } 1980 CONT; 1981 ALU64_END_TO_LE: 1982 switch (IMM) { 1983 case 16: 1984 DST = (__force u16) __swab16(DST); 1985 break; 1986 case 32: 1987 DST = (__force u32) __swab32(DST); 1988 break; 1989 case 64: 1990 DST = (__force u64) __swab64(DST); 1991 break; 1992 } 1993 CONT; 1994 1995 /* CALL */ 1996 JMP_CALL: 1997 /* Function call scratches BPF_R1-BPF_R5 registers, 1998 * preserves BPF_R6-BPF_R9, and stores return value 1999 * into BPF_R0. 2000 */ 2001 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2002 BPF_R4, BPF_R5); 2003 CONT; 2004 2005 JMP_CALL_ARGS: 2006 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2007 BPF_R3, BPF_R4, 2008 BPF_R5, 2009 insn + insn->off + 1); 2010 CONT; 2011 2012 JMP_TAIL_CALL: { 2013 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2014 struct bpf_array *array = container_of(map, struct bpf_array, map); 2015 struct bpf_prog *prog; 2016 u32 index = BPF_R3; 2017 2018 if (unlikely(index >= array->map.max_entries)) 2019 goto out; 2020 2021 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2022 goto out; 2023 2024 tail_call_cnt++; 2025 2026 prog = READ_ONCE(array->ptrs[index]); 2027 if (!prog) 2028 goto out; 2029 2030 /* ARG1 at this point is guaranteed to point to CTX from 2031 * the verifier side due to the fact that the tail call is 2032 * handled like a helper, that is, bpf_tail_call_proto, 2033 * where arg1_type is ARG_PTR_TO_CTX. 2034 */ 2035 insn = prog->insnsi; 2036 goto select_insn; 2037 out: 2038 CONT; 2039 } 2040 JMP_JA: 2041 insn += insn->off; 2042 CONT; 2043 JMP32_JA: 2044 insn += insn->imm; 2045 CONT; 2046 JMP_EXIT: 2047 return BPF_R0; 2048 /* JMP */ 2049 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2050 JMP_##OPCODE##_X: \ 2051 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2052 insn += insn->off; \ 2053 CONT_JMP; \ 2054 } \ 2055 CONT; \ 2056 JMP32_##OPCODE##_X: \ 2057 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2058 insn += insn->off; \ 2059 CONT_JMP; \ 2060 } \ 2061 CONT; \ 2062 JMP_##OPCODE##_K: \ 2063 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2064 insn += insn->off; \ 2065 CONT_JMP; \ 2066 } \ 2067 CONT; \ 2068 JMP32_##OPCODE##_K: \ 2069 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2070 insn += insn->off; \ 2071 CONT_JMP; \ 2072 } \ 2073 CONT; 2074 COND_JMP(u, JEQ, ==) 2075 COND_JMP(u, JNE, !=) 2076 COND_JMP(u, JGT, >) 2077 COND_JMP(u, JLT, <) 2078 COND_JMP(u, JGE, >=) 2079 COND_JMP(u, JLE, <=) 2080 COND_JMP(u, JSET, &) 2081 COND_JMP(s, JSGT, >) 2082 COND_JMP(s, JSLT, <) 2083 COND_JMP(s, JSGE, >=) 2084 COND_JMP(s, JSLE, <=) 2085 #undef COND_JMP 2086 /* ST, STX and LDX*/ 2087 ST_NOSPEC: 2088 /* Speculation barrier for mitigating Speculative Store Bypass, 2089 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2090 * rely on the firmware mitigation as controlled via the ssbd 2091 * kernel parameter. Whenever the mitigation is enabled, it 2092 * works for all of the kernel code with no need to provide any 2093 * additional instructions here. In case of x86, we use 'lfence' 2094 * insn for mitigation. We reuse preexisting logic from Spectre 2095 * v1 mitigation that happens to produce the required code on 2096 * x86 for v4 as well. 2097 */ 2098 barrier_nospec(); 2099 CONT; 2100 #define LDST(SIZEOP, SIZE) \ 2101 STX_MEM_##SIZEOP: \ 2102 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2103 CONT; \ 2104 ST_MEM_##SIZEOP: \ 2105 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2106 CONT; \ 2107 LDX_MEM_##SIZEOP: \ 2108 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2109 CONT; \ 2110 LDX_PROBE_MEM_##SIZEOP: \ 2111 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2112 (const void *)(long) (SRC + insn->off)); \ 2113 DST = *((SIZE *)&DST); \ 2114 CONT; 2115 2116 LDST(B, u8) 2117 LDST(H, u16) 2118 LDST(W, u32) 2119 LDST(DW, u64) 2120 #undef LDST 2121 2122 #define LDSX(SIZEOP, SIZE) \ 2123 LDX_MEMSX_##SIZEOP: \ 2124 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2125 CONT; \ 2126 LDX_PROBE_MEMSX_##SIZEOP: \ 2127 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2128 (const void *)(long) (SRC + insn->off)); \ 2129 DST = *((SIZE *)&DST); \ 2130 CONT; 2131 2132 LDSX(B, s8) 2133 LDSX(H, s16) 2134 LDSX(W, s32) 2135 #undef LDSX 2136 2137 #define ATOMIC_ALU_OP(BOP, KOP) \ 2138 case BOP: \ 2139 if (BPF_SIZE(insn->code) == BPF_W) \ 2140 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2141 (DST + insn->off)); \ 2142 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2143 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2144 (DST + insn->off)); \ 2145 else \ 2146 goto default_label; \ 2147 break; \ 2148 case BOP | BPF_FETCH: \ 2149 if (BPF_SIZE(insn->code) == BPF_W) \ 2150 SRC = (u32) atomic_fetch_##KOP( \ 2151 (u32) SRC, \ 2152 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2153 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2154 SRC = (u64) atomic64_fetch_##KOP( \ 2155 (u64) SRC, \ 2156 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2157 else \ 2158 goto default_label; \ 2159 break; 2160 2161 STX_ATOMIC_DW: 2162 STX_ATOMIC_W: 2163 STX_ATOMIC_H: 2164 STX_ATOMIC_B: 2165 switch (IMM) { 2166 /* Atomic read-modify-write instructions support only W and DW 2167 * size modifiers. 2168 */ 2169 ATOMIC_ALU_OP(BPF_ADD, add) 2170 ATOMIC_ALU_OP(BPF_AND, and) 2171 ATOMIC_ALU_OP(BPF_OR, or) 2172 ATOMIC_ALU_OP(BPF_XOR, xor) 2173 #undef ATOMIC_ALU_OP 2174 2175 case BPF_XCHG: 2176 if (BPF_SIZE(insn->code) == BPF_W) 2177 SRC = (u32) atomic_xchg( 2178 (atomic_t *)(unsigned long) (DST + insn->off), 2179 (u32) SRC); 2180 else if (BPF_SIZE(insn->code) == BPF_DW) 2181 SRC = (u64) atomic64_xchg( 2182 (atomic64_t *)(unsigned long) (DST + insn->off), 2183 (u64) SRC); 2184 else 2185 goto default_label; 2186 break; 2187 case BPF_CMPXCHG: 2188 if (BPF_SIZE(insn->code) == BPF_W) 2189 BPF_R0 = (u32) atomic_cmpxchg( 2190 (atomic_t *)(unsigned long) (DST + insn->off), 2191 (u32) BPF_R0, (u32) SRC); 2192 else if (BPF_SIZE(insn->code) == BPF_DW) 2193 BPF_R0 = (u64) atomic64_cmpxchg( 2194 (atomic64_t *)(unsigned long) (DST + insn->off), 2195 (u64) BPF_R0, (u64) SRC); 2196 else 2197 goto default_label; 2198 break; 2199 /* Atomic load and store instructions support all size 2200 * modifiers. 2201 */ 2202 case BPF_LOAD_ACQ: 2203 switch (BPF_SIZE(insn->code)) { 2204 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2205 case BPF_##SIZEOP: \ 2206 DST = (SIZE)smp_load_acquire( \ 2207 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2208 break; 2209 LOAD_ACQUIRE(B, u8) 2210 LOAD_ACQUIRE(H, u16) 2211 LOAD_ACQUIRE(W, u32) 2212 #ifdef CONFIG_64BIT 2213 LOAD_ACQUIRE(DW, u64) 2214 #endif 2215 #undef LOAD_ACQUIRE 2216 default: 2217 goto default_label; 2218 } 2219 break; 2220 case BPF_STORE_REL: 2221 switch (BPF_SIZE(insn->code)) { 2222 #define STORE_RELEASE(SIZEOP, SIZE) \ 2223 case BPF_##SIZEOP: \ 2224 smp_store_release( \ 2225 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2226 break; 2227 STORE_RELEASE(B, u8) 2228 STORE_RELEASE(H, u16) 2229 STORE_RELEASE(W, u32) 2230 #ifdef CONFIG_64BIT 2231 STORE_RELEASE(DW, u64) 2232 #endif 2233 #undef STORE_RELEASE 2234 default: 2235 goto default_label; 2236 } 2237 break; 2238 2239 default: 2240 goto default_label; 2241 } 2242 CONT; 2243 2244 default_label: 2245 /* If we ever reach this, we have a bug somewhere. Die hard here 2246 * instead of just returning 0; we could be somewhere in a subprog, 2247 * so execution could continue otherwise which we do /not/ want. 2248 * 2249 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2250 */ 2251 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2252 insn->code, insn->imm); 2253 BUG_ON(1); 2254 return 0; 2255 } 2256 2257 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2258 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2259 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2260 { \ 2261 u64 stack[stack_size / sizeof(u64)]; \ 2262 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2263 \ 2264 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2265 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2266 ARG1 = (u64) (unsigned long) ctx; \ 2267 return ___bpf_prog_run(regs, insn); \ 2268 } 2269 2270 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2271 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2272 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2273 const struct bpf_insn *insn) \ 2274 { \ 2275 u64 stack[stack_size / sizeof(u64)]; \ 2276 u64 regs[MAX_BPF_EXT_REG]; \ 2277 \ 2278 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2279 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2280 BPF_R1 = r1; \ 2281 BPF_R2 = r2; \ 2282 BPF_R3 = r3; \ 2283 BPF_R4 = r4; \ 2284 BPF_R5 = r5; \ 2285 return ___bpf_prog_run(regs, insn); \ 2286 } 2287 2288 #define EVAL1(FN, X) FN(X) 2289 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2290 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2291 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2292 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2293 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2294 2295 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2296 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2297 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2298 2299 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2300 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2301 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2302 2303 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2304 2305 static unsigned int (*interpreters[])(const void *ctx, 2306 const struct bpf_insn *insn) = { 2307 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2308 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2309 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2310 }; 2311 #undef PROG_NAME_LIST 2312 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2313 static __maybe_unused 2314 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2315 const struct bpf_insn *insn) = { 2316 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2317 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2318 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2319 }; 2320 #undef PROG_NAME_LIST 2321 2322 #ifdef CONFIG_BPF_SYSCALL 2323 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2324 { 2325 stack_depth = max_t(u32, stack_depth, 1); 2326 insn->off = (s16) insn->imm; 2327 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2328 __bpf_call_base_args; 2329 insn->code = BPF_JMP | BPF_CALL_ARGS; 2330 } 2331 #endif 2332 #endif 2333 2334 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2335 const struct bpf_insn *insn) 2336 { 2337 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2338 * is not working properly, so warn about it! 2339 */ 2340 WARN_ON_ONCE(1); 2341 return 0; 2342 } 2343 2344 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2345 const struct bpf_prog *fp) 2346 { 2347 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2348 struct bpf_prog_aux *aux = fp->aux; 2349 enum bpf_cgroup_storage_type i; 2350 bool ret = false; 2351 u64 cookie; 2352 2353 if (fp->kprobe_override) 2354 return ret; 2355 2356 spin_lock(&map->owner_lock); 2357 /* There's no owner yet where we could check for compatibility. */ 2358 if (!map->owner) { 2359 map->owner = bpf_map_owner_alloc(map); 2360 if (!map->owner) 2361 goto err; 2362 map->owner->type = prog_type; 2363 map->owner->jited = fp->jited; 2364 map->owner->xdp_has_frags = aux->xdp_has_frags; 2365 map->owner->attach_func_proto = aux->attach_func_proto; 2366 for_each_cgroup_storage_type(i) { 2367 map->owner->storage_cookie[i] = 2368 aux->cgroup_storage[i] ? 2369 aux->cgroup_storage[i]->cookie : 0; 2370 } 2371 ret = true; 2372 } else { 2373 ret = map->owner->type == prog_type && 2374 map->owner->jited == fp->jited && 2375 map->owner->xdp_has_frags == aux->xdp_has_frags; 2376 for_each_cgroup_storage_type(i) { 2377 if (!ret) 2378 break; 2379 cookie = aux->cgroup_storage[i] ? 2380 aux->cgroup_storage[i]->cookie : 0; 2381 ret = map->owner->storage_cookie[i] == cookie || 2382 !cookie; 2383 } 2384 if (ret && 2385 map->owner->attach_func_proto != aux->attach_func_proto) { 2386 switch (prog_type) { 2387 case BPF_PROG_TYPE_TRACING: 2388 case BPF_PROG_TYPE_LSM: 2389 case BPF_PROG_TYPE_EXT: 2390 case BPF_PROG_TYPE_STRUCT_OPS: 2391 ret = false; 2392 break; 2393 default: 2394 break; 2395 } 2396 } 2397 } 2398 err: 2399 spin_unlock(&map->owner_lock); 2400 return ret; 2401 } 2402 2403 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2404 { 2405 /* XDP programs inserted into maps are not guaranteed to run on 2406 * a particular netdev (and can run outside driver context entirely 2407 * in the case of devmap and cpumap). Until device checks 2408 * are implemented, prohibit adding dev-bound programs to program maps. 2409 */ 2410 if (bpf_prog_is_dev_bound(fp->aux)) 2411 return false; 2412 2413 return __bpf_prog_map_compatible(map, fp); 2414 } 2415 2416 static int bpf_check_tail_call(const struct bpf_prog *fp) 2417 { 2418 struct bpf_prog_aux *aux = fp->aux; 2419 int i, ret = 0; 2420 2421 mutex_lock(&aux->used_maps_mutex); 2422 for (i = 0; i < aux->used_map_cnt; i++) { 2423 struct bpf_map *map = aux->used_maps[i]; 2424 2425 if (!map_type_contains_progs(map)) 2426 continue; 2427 2428 if (!__bpf_prog_map_compatible(map, fp)) { 2429 ret = -EINVAL; 2430 goto out; 2431 } 2432 } 2433 2434 out: 2435 mutex_unlock(&aux->used_maps_mutex); 2436 return ret; 2437 } 2438 2439 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2440 { 2441 bool select_interpreter = false; 2442 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2443 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2444 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2445 2446 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2447 * But for non-JITed programs, we don't need bpf_func, so no bounds 2448 * check needed. 2449 */ 2450 if (idx < ARRAY_SIZE(interpreters)) { 2451 fp->bpf_func = interpreters[idx]; 2452 select_interpreter = true; 2453 } else { 2454 fp->bpf_func = __bpf_prog_ret0_warn; 2455 } 2456 #else 2457 fp->bpf_func = __bpf_prog_ret0_warn; 2458 #endif 2459 return select_interpreter; 2460 } 2461 2462 /** 2463 * bpf_prog_select_runtime - select exec runtime for BPF program 2464 * @fp: bpf_prog populated with BPF program 2465 * @err: pointer to error variable 2466 * 2467 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2468 * The BPF program will be executed via bpf_prog_run() function. 2469 * 2470 * Return: the &fp argument along with &err set to 0 for success or 2471 * a negative errno code on failure 2472 */ 2473 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2474 { 2475 /* In case of BPF to BPF calls, verifier did all the prep 2476 * work with regards to JITing, etc. 2477 */ 2478 bool jit_needed = false; 2479 2480 if (fp->bpf_func) 2481 goto finalize; 2482 2483 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2484 bpf_prog_has_kfunc_call(fp)) 2485 jit_needed = true; 2486 2487 if (!bpf_prog_select_interpreter(fp)) 2488 jit_needed = true; 2489 2490 /* eBPF JITs can rewrite the program in case constant 2491 * blinding is active. However, in case of error during 2492 * blinding, bpf_int_jit_compile() must always return a 2493 * valid program, which in this case would simply not 2494 * be JITed, but falls back to the interpreter. 2495 */ 2496 if (!bpf_prog_is_offloaded(fp->aux)) { 2497 *err = bpf_prog_alloc_jited_linfo(fp); 2498 if (*err) 2499 return fp; 2500 2501 fp = bpf_int_jit_compile(fp); 2502 bpf_prog_jit_attempt_done(fp); 2503 if (!fp->jited && jit_needed) { 2504 *err = -ENOTSUPP; 2505 return fp; 2506 } 2507 } else { 2508 *err = bpf_prog_offload_compile(fp); 2509 if (*err) 2510 return fp; 2511 } 2512 2513 finalize: 2514 *err = bpf_prog_lock_ro(fp); 2515 if (*err) 2516 return fp; 2517 2518 /* The tail call compatibility check can only be done at 2519 * this late stage as we need to determine, if we deal 2520 * with JITed or non JITed program concatenations and not 2521 * all eBPF JITs might immediately support all features. 2522 */ 2523 *err = bpf_check_tail_call(fp); 2524 2525 return fp; 2526 } 2527 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2528 2529 static unsigned int __bpf_prog_ret1(const void *ctx, 2530 const struct bpf_insn *insn) 2531 { 2532 return 1; 2533 } 2534 2535 static struct bpf_prog_dummy { 2536 struct bpf_prog prog; 2537 } dummy_bpf_prog = { 2538 .prog = { 2539 .bpf_func = __bpf_prog_ret1, 2540 }, 2541 }; 2542 2543 struct bpf_empty_prog_array bpf_empty_prog_array = { 2544 .null_prog = NULL, 2545 }; 2546 EXPORT_SYMBOL(bpf_empty_prog_array); 2547 2548 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2549 { 2550 struct bpf_prog_array *p; 2551 2552 if (prog_cnt) 2553 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags); 2554 else 2555 p = &bpf_empty_prog_array.hdr; 2556 2557 return p; 2558 } 2559 2560 void bpf_prog_array_free(struct bpf_prog_array *progs) 2561 { 2562 if (!progs || progs == &bpf_empty_prog_array.hdr) 2563 return; 2564 kfree_rcu(progs, rcu); 2565 } 2566 2567 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2568 { 2569 struct bpf_prog_array *progs; 2570 2571 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2572 * no need to call kfree_rcu(), just call kfree() directly. 2573 */ 2574 progs = container_of(rcu, struct bpf_prog_array, rcu); 2575 if (rcu_trace_implies_rcu_gp()) 2576 kfree(progs); 2577 else 2578 kfree_rcu(progs, rcu); 2579 } 2580 2581 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2582 { 2583 if (!progs || progs == &bpf_empty_prog_array.hdr) 2584 return; 2585 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2586 } 2587 2588 int bpf_prog_array_length(struct bpf_prog_array *array) 2589 { 2590 struct bpf_prog_array_item *item; 2591 u32 cnt = 0; 2592 2593 for (item = array->items; item->prog; item++) 2594 if (item->prog != &dummy_bpf_prog.prog) 2595 cnt++; 2596 return cnt; 2597 } 2598 2599 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2600 { 2601 struct bpf_prog_array_item *item; 2602 2603 for (item = array->items; item->prog; item++) 2604 if (item->prog != &dummy_bpf_prog.prog) 2605 return false; 2606 return true; 2607 } 2608 2609 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2610 u32 *prog_ids, 2611 u32 request_cnt) 2612 { 2613 struct bpf_prog_array_item *item; 2614 int i = 0; 2615 2616 for (item = array->items; item->prog; item++) { 2617 if (item->prog == &dummy_bpf_prog.prog) 2618 continue; 2619 prog_ids[i] = item->prog->aux->id; 2620 if (++i == request_cnt) { 2621 item++; 2622 break; 2623 } 2624 } 2625 2626 return !!(item->prog); 2627 } 2628 2629 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2630 __u32 __user *prog_ids, u32 cnt) 2631 { 2632 unsigned long err = 0; 2633 bool nospc; 2634 u32 *ids; 2635 2636 /* users of this function are doing: 2637 * cnt = bpf_prog_array_length(); 2638 * if (cnt > 0) 2639 * bpf_prog_array_copy_to_user(..., cnt); 2640 * so below kcalloc doesn't need extra cnt > 0 check. 2641 */ 2642 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2643 if (!ids) 2644 return -ENOMEM; 2645 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2646 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2647 kfree(ids); 2648 if (err) 2649 return -EFAULT; 2650 if (nospc) 2651 return -ENOSPC; 2652 return 0; 2653 } 2654 2655 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2656 struct bpf_prog *old_prog) 2657 { 2658 struct bpf_prog_array_item *item; 2659 2660 for (item = array->items; item->prog; item++) 2661 if (item->prog == old_prog) { 2662 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2663 break; 2664 } 2665 } 2666 2667 /** 2668 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2669 * index into the program array with 2670 * a dummy no-op program. 2671 * @array: a bpf_prog_array 2672 * @index: the index of the program to replace 2673 * 2674 * Skips over dummy programs, by not counting them, when calculating 2675 * the position of the program to replace. 2676 * 2677 * Return: 2678 * * 0 - Success 2679 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2680 * * -ENOENT - Index out of range 2681 */ 2682 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2683 { 2684 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2685 } 2686 2687 /** 2688 * bpf_prog_array_update_at() - Updates the program at the given index 2689 * into the program array. 2690 * @array: a bpf_prog_array 2691 * @index: the index of the program to update 2692 * @prog: the program to insert into the array 2693 * 2694 * Skips over dummy programs, by not counting them, when calculating 2695 * the position of the program to update. 2696 * 2697 * Return: 2698 * * 0 - Success 2699 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2700 * * -ENOENT - Index out of range 2701 */ 2702 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2703 struct bpf_prog *prog) 2704 { 2705 struct bpf_prog_array_item *item; 2706 2707 if (unlikely(index < 0)) 2708 return -EINVAL; 2709 2710 for (item = array->items; item->prog; item++) { 2711 if (item->prog == &dummy_bpf_prog.prog) 2712 continue; 2713 if (!index) { 2714 WRITE_ONCE(item->prog, prog); 2715 return 0; 2716 } 2717 index--; 2718 } 2719 return -ENOENT; 2720 } 2721 2722 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2723 struct bpf_prog *exclude_prog, 2724 struct bpf_prog *include_prog, 2725 u64 bpf_cookie, 2726 struct bpf_prog_array **new_array) 2727 { 2728 int new_prog_cnt, carry_prog_cnt = 0; 2729 struct bpf_prog_array_item *existing, *new; 2730 struct bpf_prog_array *array; 2731 bool found_exclude = false; 2732 2733 /* Figure out how many existing progs we need to carry over to 2734 * the new array. 2735 */ 2736 if (old_array) { 2737 existing = old_array->items; 2738 for (; existing->prog; existing++) { 2739 if (existing->prog == exclude_prog) { 2740 found_exclude = true; 2741 continue; 2742 } 2743 if (existing->prog != &dummy_bpf_prog.prog) 2744 carry_prog_cnt++; 2745 if (existing->prog == include_prog) 2746 return -EEXIST; 2747 } 2748 } 2749 2750 if (exclude_prog && !found_exclude) 2751 return -ENOENT; 2752 2753 /* How many progs (not NULL) will be in the new array? */ 2754 new_prog_cnt = carry_prog_cnt; 2755 if (include_prog) 2756 new_prog_cnt += 1; 2757 2758 /* Do we have any prog (not NULL) in the new array? */ 2759 if (!new_prog_cnt) { 2760 *new_array = NULL; 2761 return 0; 2762 } 2763 2764 /* +1 as the end of prog_array is marked with NULL */ 2765 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2766 if (!array) 2767 return -ENOMEM; 2768 new = array->items; 2769 2770 /* Fill in the new prog array */ 2771 if (carry_prog_cnt) { 2772 existing = old_array->items; 2773 for (; existing->prog; existing++) { 2774 if (existing->prog == exclude_prog || 2775 existing->prog == &dummy_bpf_prog.prog) 2776 continue; 2777 2778 new->prog = existing->prog; 2779 new->bpf_cookie = existing->bpf_cookie; 2780 new++; 2781 } 2782 } 2783 if (include_prog) { 2784 new->prog = include_prog; 2785 new->bpf_cookie = bpf_cookie; 2786 new++; 2787 } 2788 new->prog = NULL; 2789 *new_array = array; 2790 return 0; 2791 } 2792 2793 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2794 u32 *prog_ids, u32 request_cnt, 2795 u32 *prog_cnt) 2796 { 2797 u32 cnt = 0; 2798 2799 if (array) 2800 cnt = bpf_prog_array_length(array); 2801 2802 *prog_cnt = cnt; 2803 2804 /* return early if user requested only program count or nothing to copy */ 2805 if (!request_cnt || !cnt) 2806 return 0; 2807 2808 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2809 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2810 : 0; 2811 } 2812 2813 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2814 struct bpf_map **used_maps, u32 len) 2815 { 2816 struct bpf_map *map; 2817 bool sleepable; 2818 u32 i; 2819 2820 sleepable = aux->prog->sleepable; 2821 for (i = 0; i < len; i++) { 2822 map = used_maps[i]; 2823 if (map->ops->map_poke_untrack) 2824 map->ops->map_poke_untrack(map, aux); 2825 if (sleepable) 2826 atomic64_dec(&map->sleepable_refcnt); 2827 bpf_map_put(map); 2828 } 2829 } 2830 2831 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2832 { 2833 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2834 kfree(aux->used_maps); 2835 } 2836 2837 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 2838 { 2839 #ifdef CONFIG_BPF_SYSCALL 2840 struct btf_mod_pair *btf_mod; 2841 u32 i; 2842 2843 for (i = 0; i < len; i++) { 2844 btf_mod = &used_btfs[i]; 2845 if (btf_mod->module) 2846 module_put(btf_mod->module); 2847 btf_put(btf_mod->btf); 2848 } 2849 #endif 2850 } 2851 2852 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2853 { 2854 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 2855 kfree(aux->used_btfs); 2856 } 2857 2858 static void bpf_prog_free_deferred(struct work_struct *work) 2859 { 2860 struct bpf_prog_aux *aux; 2861 int i; 2862 2863 aux = container_of(work, struct bpf_prog_aux, work); 2864 #ifdef CONFIG_BPF_SYSCALL 2865 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2866 bpf_prog_stream_free(aux->prog); 2867 #endif 2868 #ifdef CONFIG_CGROUP_BPF 2869 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2870 bpf_cgroup_atype_put(aux->cgroup_atype); 2871 #endif 2872 bpf_free_used_maps(aux); 2873 bpf_free_used_btfs(aux); 2874 if (bpf_prog_is_dev_bound(aux)) 2875 bpf_prog_dev_bound_destroy(aux->prog); 2876 #ifdef CONFIG_PERF_EVENTS 2877 if (aux->prog->has_callchain_buf) 2878 put_callchain_buffers(); 2879 #endif 2880 if (aux->dst_trampoline) 2881 bpf_trampoline_put(aux->dst_trampoline); 2882 for (i = 0; i < aux->real_func_cnt; i++) { 2883 /* We can just unlink the subprog poke descriptor table as 2884 * it was originally linked to the main program and is also 2885 * released along with it. 2886 */ 2887 aux->func[i]->aux->poke_tab = NULL; 2888 bpf_jit_free(aux->func[i]); 2889 } 2890 if (aux->real_func_cnt) { 2891 kfree(aux->func); 2892 bpf_prog_unlock_free(aux->prog); 2893 } else { 2894 bpf_jit_free(aux->prog); 2895 } 2896 } 2897 2898 void bpf_prog_free(struct bpf_prog *fp) 2899 { 2900 struct bpf_prog_aux *aux = fp->aux; 2901 2902 if (aux->dst_prog) 2903 bpf_prog_put(aux->dst_prog); 2904 bpf_token_put(aux->token); 2905 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2906 schedule_work(&aux->work); 2907 } 2908 EXPORT_SYMBOL_GPL(bpf_prog_free); 2909 2910 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 2911 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2912 2913 void bpf_user_rnd_init_once(void) 2914 { 2915 prandom_init_once(&bpf_user_rnd_state); 2916 } 2917 2918 BPF_CALL_0(bpf_user_rnd_u32) 2919 { 2920 /* Should someone ever have the rather unwise idea to use some 2921 * of the registers passed into this function, then note that 2922 * this function is called from native eBPF and classic-to-eBPF 2923 * transformations. Register assignments from both sides are 2924 * different, f.e. classic always sets fn(ctx, A, X) here. 2925 */ 2926 struct rnd_state *state; 2927 u32 res; 2928 2929 state = &get_cpu_var(bpf_user_rnd_state); 2930 res = prandom_u32_state(state); 2931 put_cpu_var(bpf_user_rnd_state); 2932 2933 return res; 2934 } 2935 2936 BPF_CALL_0(bpf_get_raw_cpu_id) 2937 { 2938 return raw_smp_processor_id(); 2939 } 2940 2941 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2942 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2943 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2944 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2945 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2946 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2947 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2948 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2949 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2950 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2951 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2952 2953 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 2954 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 2955 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 2956 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 2957 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 2958 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 2959 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 2960 2961 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 2962 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 2963 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 2964 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 2965 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 2966 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 2967 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 2968 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 2969 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 2970 const struct bpf_func_proto bpf_set_retval_proto __weak; 2971 const struct bpf_func_proto bpf_get_retval_proto __weak; 2972 2973 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 2974 { 2975 return NULL; 2976 } 2977 2978 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 2979 { 2980 return NULL; 2981 } 2982 2983 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 2984 { 2985 return NULL; 2986 } 2987 2988 u64 __weak 2989 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 2990 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 2991 { 2992 return -ENOTSUPP; 2993 } 2994 EXPORT_SYMBOL_GPL(bpf_event_output); 2995 2996 /* Always built-in helper functions. */ 2997 const struct bpf_func_proto bpf_tail_call_proto = { 2998 /* func is unused for tail_call, we set it to pass the 2999 * get_helper_proto check 3000 */ 3001 .func = BPF_PTR_POISON, 3002 .gpl_only = false, 3003 .ret_type = RET_VOID, 3004 .arg1_type = ARG_PTR_TO_CTX, 3005 .arg2_type = ARG_CONST_MAP_PTR, 3006 .arg3_type = ARG_ANYTHING, 3007 }; 3008 3009 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3010 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3011 * eBPF and implicitly also cBPF can get JITed! 3012 */ 3013 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 3014 { 3015 return prog; 3016 } 3017 3018 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3019 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3020 */ 3021 void __weak bpf_jit_compile(struct bpf_prog *prog) 3022 { 3023 } 3024 3025 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3026 { 3027 return false; 3028 } 3029 3030 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3031 * analysis code and wants explicit zero extension inserted by verifier. 3032 * Otherwise, return FALSE. 3033 * 3034 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3035 * you don't override this. JITs that don't want these extra insns can detect 3036 * them using insn_is_zext. 3037 */ 3038 bool __weak bpf_jit_needs_zext(void) 3039 { 3040 return false; 3041 } 3042 3043 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3044 * all archs. The value returned must not change at runtime as there is 3045 * currently no support for reloading programs that were loaded without 3046 * mitigations. 3047 */ 3048 bool __weak bpf_jit_bypass_spec_v1(void) 3049 { 3050 return false; 3051 } 3052 3053 bool __weak bpf_jit_bypass_spec_v4(void) 3054 { 3055 return false; 3056 } 3057 3058 /* Return true if the JIT inlines the call to the helper corresponding to 3059 * the imm. 3060 * 3061 * The verifier will not patch the insn->imm for the call to the helper if 3062 * this returns true. 3063 */ 3064 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3065 { 3066 return false; 3067 } 3068 3069 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3070 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3071 { 3072 return false; 3073 } 3074 3075 bool __weak bpf_jit_supports_percpu_insn(void) 3076 { 3077 return false; 3078 } 3079 3080 bool __weak bpf_jit_supports_kfunc_call(void) 3081 { 3082 return false; 3083 } 3084 3085 bool __weak bpf_jit_supports_far_kfunc_call(void) 3086 { 3087 return false; 3088 } 3089 3090 bool __weak bpf_jit_supports_arena(void) 3091 { 3092 return false; 3093 } 3094 3095 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3096 { 3097 return false; 3098 } 3099 3100 u64 __weak bpf_arch_uaddress_limit(void) 3101 { 3102 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3103 return TASK_SIZE; 3104 #else 3105 return 0; 3106 #endif 3107 } 3108 3109 /* Return TRUE if the JIT backend satisfies the following two conditions: 3110 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3111 * 2) Under the specific arch, the implementation of xchg() is the same 3112 * as atomic_xchg() on pointer-sized words. 3113 */ 3114 bool __weak bpf_jit_supports_ptr_xchg(void) 3115 { 3116 return false; 3117 } 3118 3119 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3120 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3121 */ 3122 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3123 int len) 3124 { 3125 return -EFAULT; 3126 } 3127 3128 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t, 3129 void *addr1, void *addr2) 3130 { 3131 return -ENOTSUPP; 3132 } 3133 3134 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3135 { 3136 return ERR_PTR(-ENOTSUPP); 3137 } 3138 3139 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3140 { 3141 return -ENOTSUPP; 3142 } 3143 3144 bool __weak bpf_jit_supports_exceptions(void) 3145 { 3146 return false; 3147 } 3148 3149 bool __weak bpf_jit_supports_private_stack(void) 3150 { 3151 return false; 3152 } 3153 3154 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3155 { 3156 } 3157 3158 bool __weak bpf_jit_supports_timed_may_goto(void) 3159 { 3160 return false; 3161 } 3162 3163 u64 __weak arch_bpf_timed_may_goto(void) 3164 { 3165 return 0; 3166 } 3167 3168 static noinline void bpf_prog_report_may_goto_violation(void) 3169 { 3170 #ifdef CONFIG_BPF_SYSCALL 3171 struct bpf_stream_stage ss; 3172 struct bpf_prog *prog; 3173 3174 prog = bpf_prog_find_from_stack(); 3175 if (!prog) 3176 return; 3177 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3178 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3179 bpf_stream_dump_stack(ss); 3180 })); 3181 #endif 3182 } 3183 3184 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3185 { 3186 u64 time = ktime_get_mono_fast_ns(); 3187 3188 /* Populate the timestamp for this stack frame, and refresh count. */ 3189 if (!p->timestamp) { 3190 p->timestamp = time; 3191 return BPF_MAX_TIMED_LOOPS; 3192 } 3193 /* Check if we've exhausted our time slice, and zero count. */ 3194 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3195 bpf_prog_report_may_goto_violation(); 3196 return 0; 3197 } 3198 /* Refresh the count for the stack frame. */ 3199 return BPF_MAX_TIMED_LOOPS; 3200 } 3201 3202 /* for configs without MMU or 32-bit */ 3203 __weak const struct bpf_map_ops arena_map_ops; 3204 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3205 { 3206 return 0; 3207 } 3208 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3209 { 3210 return 0; 3211 } 3212 3213 #ifdef CONFIG_BPF_SYSCALL 3214 static int __init bpf_global_ma_init(void) 3215 { 3216 int ret; 3217 3218 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3219 bpf_global_ma_set = !ret; 3220 return ret; 3221 } 3222 late_initcall(bpf_global_ma_init); 3223 #endif 3224 3225 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3226 EXPORT_SYMBOL(bpf_stats_enabled_key); 3227 3228 /* All definitions of tracepoints related to BPF. */ 3229 #define CREATE_TRACE_POINTS 3230 #include <linux/bpf_trace.h> 3231 3232 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3233 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3234 3235 #ifdef CONFIG_BPF_SYSCALL 3236 3237 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3238 const char **linep, int *nump) 3239 { 3240 int idx = -1, insn_start, insn_end, len; 3241 struct bpf_line_info *linfo; 3242 void **jited_linfo; 3243 struct btf *btf; 3244 int nr_linfo; 3245 3246 btf = prog->aux->btf; 3247 linfo = prog->aux->linfo; 3248 jited_linfo = prog->aux->jited_linfo; 3249 3250 if (!btf || !linfo || !jited_linfo) 3251 return -EINVAL; 3252 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3253 3254 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3255 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3256 3257 insn_start = linfo[0].insn_off; 3258 insn_end = insn_start + len; 3259 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3260 3261 for (int i = 0; i < nr_linfo && 3262 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3263 if (jited_linfo[i] >= (void *)ip) 3264 break; 3265 idx = i; 3266 } 3267 3268 if (idx == -1) 3269 return -ENOENT; 3270 3271 /* Get base component of the file path. */ 3272 *filep = btf_name_by_offset(btf, linfo[idx].file_name_off); 3273 *filep = kbasename(*filep); 3274 /* Obtain the source line, and strip whitespace in prefix. */ 3275 *linep = btf_name_by_offset(btf, linfo[idx].line_off); 3276 while (isspace(**linep)) 3277 *linep += 1; 3278 *nump = BPF_LINE_INFO_LINE_NUM(linfo[idx].line_col); 3279 return 0; 3280 } 3281 3282 struct walk_stack_ctx { 3283 struct bpf_prog *prog; 3284 }; 3285 3286 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3287 { 3288 struct walk_stack_ctx *ctxp = cookie; 3289 struct bpf_prog *prog; 3290 3291 /* 3292 * The RCU read lock is held to safely traverse the latch tree, but we 3293 * don't need its protection when accessing the prog, since it has an 3294 * active stack frame on the current stack trace, and won't disappear. 3295 */ 3296 rcu_read_lock(); 3297 prog = bpf_prog_ksym_find(ip); 3298 rcu_read_unlock(); 3299 if (!prog) 3300 return true; 3301 /* Make sure we return the main prog if we found a subprog */ 3302 ctxp->prog = prog->aux->main_prog_aux->prog; 3303 return false; 3304 } 3305 3306 struct bpf_prog *bpf_prog_find_from_stack(void) 3307 { 3308 struct walk_stack_ctx ctx = {}; 3309 3310 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3311 return ctx.prog; 3312 } 3313 3314 #endif 3315