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