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