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(sizeof(u8[BPF_NR_CONTEXTS]), 4, 116 bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 117 if (!fp->active) { 118 vfree(fp); 119 kfree(aux); 120 return NULL; 121 } 122 123 fp->pages = size / PAGE_SIZE; 124 fp->aux = aux; 125 fp->aux->main_prog_aux = aux; 126 fp->aux->prog = fp; 127 fp->jit_requested = ebpf_jit_enabled(); 128 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 129 #ifdef CONFIG_CGROUP_BPF 130 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 131 #endif 132 133 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 134 #ifdef CONFIG_FINEIBT 135 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 136 #endif 137 mutex_init(&fp->aux->used_maps_mutex); 138 mutex_init(&fp->aux->ext_mutex); 139 mutex_init(&fp->aux->dst_mutex); 140 mutex_init(&fp->aux->st_ops_assoc_mutex); 141 142 #ifdef CONFIG_BPF_SYSCALL 143 bpf_prog_stream_init(fp); 144 #endif 145 146 return fp; 147 } 148 149 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 150 { 151 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 152 struct bpf_prog *prog; 153 int cpu; 154 155 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 156 if (!prog) 157 return NULL; 158 159 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 160 if (!prog->stats) { 161 free_percpu(prog->active); 162 kfree(prog->aux); 163 vfree(prog); 164 return NULL; 165 } 166 167 for_each_possible_cpu(cpu) { 168 struct bpf_prog_stats *pstats; 169 170 pstats = per_cpu_ptr(prog->stats, cpu); 171 u64_stats_init(&pstats->syncp); 172 } 173 return prog; 174 } 175 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 176 177 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 178 { 179 if (!prog->aux->nr_linfo || !prog->jit_requested) 180 return 0; 181 182 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 183 sizeof(*prog->aux->jited_linfo), 184 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 185 if (!prog->aux->jited_linfo) 186 return -ENOMEM; 187 188 return 0; 189 } 190 191 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 192 { 193 if (prog->aux->jited_linfo && 194 (!prog->jited || !prog->aux->jited_linfo[0])) { 195 kvfree(prog->aux->jited_linfo); 196 prog->aux->jited_linfo = NULL; 197 } 198 199 kfree(prog->aux->kfunc_tab); 200 prog->aux->kfunc_tab = NULL; 201 } 202 203 /* The jit engine is responsible to provide an array 204 * for insn_off to the jited_off mapping (insn_to_jit_off). 205 * 206 * The idx to this array is the insn_off. Hence, the insn_off 207 * here is relative to the prog itself instead of the main prog. 208 * This array has one entry for each xlated bpf insn. 209 * 210 * jited_off is the byte off to the end of the jited insn. 211 * 212 * Hence, with 213 * insn_start: 214 * The first bpf insn off of the prog. The insn off 215 * here is relative to the main prog. 216 * e.g. if prog is a subprog, insn_start > 0 217 * linfo_idx: 218 * The prog's idx to prog->aux->linfo and jited_linfo 219 * 220 * jited_linfo[linfo_idx] = prog->bpf_func 221 * 222 * For i > linfo_idx, 223 * 224 * jited_linfo[i] = prog->bpf_func + 225 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 226 */ 227 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 228 const u32 *insn_to_jit_off) 229 { 230 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 231 const struct bpf_line_info *linfo; 232 void **jited_linfo; 233 234 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 235 /* Userspace did not provide linfo */ 236 return; 237 238 linfo_idx = prog->aux->linfo_idx; 239 linfo = &prog->aux->linfo[linfo_idx]; 240 insn_start = linfo[0].insn_off; 241 insn_end = insn_start + prog->len; 242 243 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 244 jited_linfo[0] = prog->bpf_func; 245 246 nr_linfo = prog->aux->nr_linfo - linfo_idx; 247 248 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 249 /* The verifier ensures that linfo[i].insn_off is 250 * strictly increasing 251 */ 252 jited_linfo[i] = prog->bpf_func + 253 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 254 } 255 256 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 257 gfp_t gfp_extra_flags) 258 { 259 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 260 struct bpf_prog *fp; 261 u32 pages; 262 263 size = round_up(size, PAGE_SIZE); 264 pages = size / PAGE_SIZE; 265 if (pages <= fp_old->pages) 266 return fp_old; 267 268 fp = __vmalloc(size, gfp_flags); 269 if (fp) { 270 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 271 fp->pages = pages; 272 fp->aux->prog = fp; 273 274 /* We keep fp->aux from fp_old around in the new 275 * reallocated structure. 276 */ 277 fp_old->aux = NULL; 278 fp_old->stats = NULL; 279 fp_old->active = NULL; 280 __bpf_prog_free(fp_old); 281 } 282 283 return fp; 284 } 285 286 void __bpf_prog_free(struct bpf_prog *fp) 287 { 288 if (fp->aux) { 289 mutex_destroy(&fp->aux->used_maps_mutex); 290 mutex_destroy(&fp->aux->dst_mutex); 291 mutex_destroy(&fp->aux->st_ops_assoc_mutex); 292 kfree(fp->aux->poke_tab); 293 kfree(fp->aux); 294 } 295 free_percpu(fp->stats); 296 free_percpu(fp->active); 297 vfree(fp); 298 } 299 300 int bpf_prog_calc_tag(struct bpf_prog *fp) 301 { 302 size_t size = bpf_prog_insn_size(fp); 303 struct bpf_insn *dst; 304 bool was_ld_map; 305 u32 i; 306 307 dst = vmalloc(size); 308 if (!dst) 309 return -ENOMEM; 310 311 /* We need to take out the map fd for the digest calculation 312 * since they are unstable from user space side. 313 */ 314 for (i = 0, was_ld_map = false; i < fp->len; i++) { 315 dst[i] = fp->insnsi[i]; 316 if (!was_ld_map && 317 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 318 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 319 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 320 was_ld_map = true; 321 dst[i].imm = 0; 322 } else if (was_ld_map && 323 dst[i].code == 0 && 324 dst[i].dst_reg == 0 && 325 dst[i].src_reg == 0 && 326 dst[i].off == 0) { 327 was_ld_map = false; 328 dst[i].imm = 0; 329 } else { 330 was_ld_map = false; 331 } 332 } 333 sha256((u8 *)dst, size, fp->digest); 334 vfree(dst); 335 return 0; 336 } 337 338 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 339 s32 end_new, s32 curr, const bool probe_pass) 340 { 341 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 342 s32 delta = end_new - end_old; 343 s64 imm = insn->imm; 344 345 if (curr < pos && curr + imm + 1 >= end_old) 346 imm += delta; 347 else if (curr >= end_new && curr + imm + 1 < end_new) 348 imm -= delta; 349 if (imm < imm_min || imm > imm_max) 350 return -ERANGE; 351 if (!probe_pass) 352 insn->imm = imm; 353 return 0; 354 } 355 356 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 357 s32 end_new, s32 curr, const bool probe_pass) 358 { 359 s64 off_min, off_max, off; 360 s32 delta = end_new - end_old; 361 362 if (insn->code == (BPF_JMP32 | BPF_JA)) { 363 off = insn->imm; 364 off_min = S32_MIN; 365 off_max = S32_MAX; 366 } else { 367 off = insn->off; 368 off_min = S16_MIN; 369 off_max = S16_MAX; 370 } 371 372 if (curr < pos && curr + off + 1 >= end_old) 373 off += delta; 374 else if (curr >= end_new && curr + off + 1 < end_new) 375 off -= delta; 376 if (off < off_min || off > off_max) 377 return -ERANGE; 378 if (!probe_pass) { 379 if (insn->code == (BPF_JMP32 | BPF_JA)) 380 insn->imm = off; 381 else 382 insn->off = off; 383 } 384 return 0; 385 } 386 387 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 388 s32 end_new, const bool probe_pass) 389 { 390 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 391 struct bpf_insn *insn = prog->insnsi; 392 int ret = 0; 393 394 for (i = 0; i < insn_cnt; i++, insn++) { 395 u8 code; 396 397 /* In the probing pass we still operate on the original, 398 * unpatched image in order to check overflows before we 399 * do any other adjustments. Therefore skip the patchlet. 400 */ 401 if (probe_pass && i == pos) { 402 i = end_new; 403 insn = prog->insnsi + end_old; 404 } 405 if (bpf_pseudo_func(insn)) { 406 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 407 end_new, i, probe_pass); 408 if (ret) 409 return ret; 410 continue; 411 } 412 code = insn->code; 413 if ((BPF_CLASS(code) != BPF_JMP && 414 BPF_CLASS(code) != BPF_JMP32) || 415 BPF_OP(code) == BPF_EXIT) 416 continue; 417 /* Adjust offset of jmps if we cross patch boundaries. */ 418 if (BPF_OP(code) == BPF_CALL) { 419 if (insn->src_reg != BPF_PSEUDO_CALL) 420 continue; 421 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 422 end_new, i, probe_pass); 423 } else { 424 ret = bpf_adj_delta_to_off(insn, pos, end_old, 425 end_new, i, probe_pass); 426 } 427 if (ret) 428 break; 429 } 430 431 return ret; 432 } 433 434 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 435 { 436 struct bpf_line_info *linfo; 437 u32 i, nr_linfo; 438 439 nr_linfo = prog->aux->nr_linfo; 440 if (!nr_linfo || !delta) 441 return; 442 443 linfo = prog->aux->linfo; 444 445 for (i = 0; i < nr_linfo; i++) 446 if (off < linfo[i].insn_off) 447 break; 448 449 /* Push all off < linfo[i].insn_off by delta */ 450 for (; i < nr_linfo; i++) 451 linfo[i].insn_off += delta; 452 } 453 454 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 455 const struct bpf_insn *patch, u32 len) 456 { 457 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 458 const u32 cnt_max = S16_MAX; 459 struct bpf_prog *prog_adj; 460 int err; 461 462 /* Since our patchlet doesn't expand the image, we're done. */ 463 if (insn_delta == 0) { 464 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 465 return prog; 466 } 467 468 insn_adj_cnt = prog->len + insn_delta; 469 470 /* Reject anything that would potentially let the insn->off 471 * target overflow when we have excessive program expansions. 472 * We need to probe here before we do any reallocation where 473 * we afterwards may not fail anymore. 474 */ 475 if (insn_adj_cnt > cnt_max && 476 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 477 return ERR_PTR(err); 478 479 /* Several new instructions need to be inserted. Make room 480 * for them. Likely, there's no need for a new allocation as 481 * last page could have large enough tailroom. 482 */ 483 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 484 GFP_USER); 485 if (!prog_adj) 486 return ERR_PTR(-ENOMEM); 487 488 prog_adj->len = insn_adj_cnt; 489 490 /* Patching happens in 3 steps: 491 * 492 * 1) Move over tail of insnsi from next instruction onwards, 493 * so we can patch the single target insn with one or more 494 * new ones (patching is always from 1 to n insns, n > 0). 495 * 2) Inject new instructions at the target location. 496 * 3) Adjust branch offsets if necessary. 497 */ 498 insn_rest = insn_adj_cnt - off - len; 499 500 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 501 sizeof(*patch) * insn_rest); 502 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 503 504 /* We are guaranteed to not fail at this point, otherwise 505 * the ship has sailed to reverse to the original state. An 506 * overflow cannot happen at this point. 507 */ 508 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 509 510 bpf_adj_linfo(prog_adj, off, insn_delta); 511 512 return prog_adj; 513 } 514 515 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 516 { 517 int err; 518 519 /* Branch offsets can't overflow when program is shrinking, no need 520 * to call bpf_adj_branches(..., true) here 521 */ 522 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 523 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 524 prog->len -= cnt; 525 526 err = bpf_adj_branches(prog, off, off + cnt, off, false); 527 WARN_ON_ONCE(err); 528 return err; 529 } 530 531 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 532 { 533 int i; 534 535 for (i = 0; i < fp->aux->real_func_cnt; i++) 536 bpf_prog_kallsyms_del(fp->aux->func[i]); 537 } 538 539 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 540 { 541 bpf_prog_kallsyms_del_subprogs(fp); 542 bpf_prog_kallsyms_del(fp); 543 } 544 545 #ifdef CONFIG_BPF_JIT 546 /* All BPF JIT sysctl knobs here. */ 547 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 548 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 549 int bpf_jit_harden __read_mostly; 550 long bpf_jit_limit __read_mostly; 551 long bpf_jit_limit_max __read_mostly; 552 553 static void 554 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 555 { 556 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 557 558 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 559 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 560 } 561 562 static void 563 bpf_prog_ksym_set_name(struct bpf_prog *prog) 564 { 565 char *sym = prog->aux->ksym.name; 566 const char *end = sym + KSYM_NAME_LEN; 567 const struct btf_type *type; 568 const char *func_name; 569 570 BUILD_BUG_ON(sizeof("bpf_prog_") + 571 sizeof(prog->tag) * 2 + 572 /* name has been null terminated. 573 * We should need +1 for the '_' preceding 574 * the name. However, the null character 575 * is double counted between the name and the 576 * sizeof("bpf_prog_") above, so we omit 577 * the +1 here. 578 */ 579 sizeof(prog->aux->name) > KSYM_NAME_LEN); 580 581 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 582 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 583 584 /* prog->aux->name will be ignored if full btf name is available */ 585 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 586 type = btf_type_by_id(prog->aux->btf, 587 prog->aux->func_info[prog->aux->func_idx].type_id); 588 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 589 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 590 return; 591 } 592 593 if (prog->aux->name[0]) 594 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 595 else 596 *sym = 0; 597 } 598 599 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 600 { 601 return container_of(n, struct bpf_ksym, tnode)->start; 602 } 603 604 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 605 struct latch_tree_node *b) 606 { 607 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 608 } 609 610 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 611 { 612 unsigned long val = (unsigned long)key; 613 const struct bpf_ksym *ksym; 614 615 ksym = container_of(n, struct bpf_ksym, tnode); 616 617 if (val < ksym->start) 618 return -1; 619 /* Ensure that we detect return addresses as part of the program, when 620 * the final instruction is a call for a program part of the stack 621 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 622 */ 623 if (val > ksym->end) 624 return 1; 625 626 return 0; 627 } 628 629 static const struct latch_tree_ops bpf_tree_ops = { 630 .less = bpf_tree_less, 631 .comp = bpf_tree_comp, 632 }; 633 634 static DEFINE_SPINLOCK(bpf_lock); 635 static LIST_HEAD(bpf_kallsyms); 636 static struct latch_tree_root bpf_tree __cacheline_aligned; 637 638 void bpf_ksym_add(struct bpf_ksym *ksym) 639 { 640 spin_lock_bh(&bpf_lock); 641 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 642 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 643 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 644 spin_unlock_bh(&bpf_lock); 645 } 646 647 static void __bpf_ksym_del(struct bpf_ksym *ksym) 648 { 649 if (list_empty(&ksym->lnode)) 650 return; 651 652 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 653 list_del_rcu(&ksym->lnode); 654 } 655 656 void bpf_ksym_del(struct bpf_ksym *ksym) 657 { 658 spin_lock_bh(&bpf_lock); 659 __bpf_ksym_del(ksym); 660 spin_unlock_bh(&bpf_lock); 661 } 662 663 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 664 { 665 return fp->jited && !bpf_prog_was_classic(fp); 666 } 667 668 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 669 { 670 if (!bpf_prog_kallsyms_candidate(fp) || 671 !bpf_token_capable(fp->aux->token, CAP_BPF)) 672 return; 673 674 bpf_prog_ksym_set_addr(fp); 675 bpf_prog_ksym_set_name(fp); 676 fp->aux->ksym.prog = true; 677 678 bpf_ksym_add(&fp->aux->ksym); 679 680 #ifdef CONFIG_FINEIBT 681 /* 682 * When FineIBT, code in the __cfi_foo() symbols can get executed 683 * and hence unwinder needs help. 684 */ 685 if (cfi_mode != CFI_FINEIBT) 686 return; 687 688 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 689 "__cfi_%s", fp->aux->ksym.name); 690 691 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 692 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 693 694 bpf_ksym_add(&fp->aux->ksym_prefix); 695 #endif 696 } 697 698 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 699 { 700 if (!bpf_prog_kallsyms_candidate(fp)) 701 return; 702 703 bpf_ksym_del(&fp->aux->ksym); 704 #ifdef CONFIG_FINEIBT 705 if (cfi_mode != CFI_FINEIBT) 706 return; 707 bpf_ksym_del(&fp->aux->ksym_prefix); 708 #endif 709 } 710 711 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 712 { 713 struct latch_tree_node *n; 714 715 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 716 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 717 } 718 719 int __bpf_address_lookup(unsigned long addr, unsigned long *size, 720 unsigned long *off, char *sym) 721 { 722 struct bpf_ksym *ksym; 723 int ret = 0; 724 725 rcu_read_lock(); 726 ksym = bpf_ksym_find(addr); 727 if (ksym) { 728 unsigned long symbol_start = ksym->start; 729 unsigned long symbol_end = ksym->end; 730 731 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 732 733 if (size) 734 *size = symbol_end - symbol_start; 735 if (off) 736 *off = addr - symbol_start; 737 } 738 rcu_read_unlock(); 739 740 return ret; 741 } 742 743 bool is_bpf_text_address(unsigned long addr) 744 { 745 bool ret; 746 747 rcu_read_lock(); 748 ret = bpf_ksym_find(addr) != NULL; 749 rcu_read_unlock(); 750 751 return ret; 752 } 753 754 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 755 { 756 struct bpf_ksym *ksym; 757 758 WARN_ON_ONCE(!rcu_read_lock_held()); 759 ksym = bpf_ksym_find(addr); 760 761 return ksym && ksym->prog ? 762 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 763 NULL; 764 } 765 766 bool bpf_has_frame_pointer(unsigned long ip) 767 { 768 struct bpf_ksym *ksym; 769 unsigned long offset; 770 771 guard(rcu)(); 772 773 ksym = bpf_ksym_find(ip); 774 if (!ksym || !ksym->fp_start || !ksym->fp_end) 775 return false; 776 777 offset = ip - ksym->start; 778 779 return offset >= ksym->fp_start && offset < ksym->fp_end; 780 } 781 782 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 783 { 784 const struct exception_table_entry *e = NULL; 785 struct bpf_prog *prog; 786 787 rcu_read_lock(); 788 prog = bpf_prog_ksym_find(addr); 789 if (!prog) 790 goto out; 791 if (!prog->aux->num_exentries) 792 goto out; 793 794 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 795 out: 796 rcu_read_unlock(); 797 return e; 798 } 799 800 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 801 char *sym) 802 { 803 struct bpf_ksym *ksym; 804 unsigned int it = 0; 805 int ret = -ERANGE; 806 807 if (!bpf_jit_kallsyms_enabled()) 808 return ret; 809 810 rcu_read_lock(); 811 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 812 if (it++ != symnum) 813 continue; 814 815 strscpy(sym, ksym->name, KSYM_NAME_LEN); 816 817 *value = ksym->start; 818 *type = BPF_SYM_ELF_TYPE; 819 820 ret = 0; 821 break; 822 } 823 rcu_read_unlock(); 824 825 return ret; 826 } 827 828 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 829 struct bpf_jit_poke_descriptor *poke) 830 { 831 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 832 static const u32 poke_tab_max = 1024; 833 u32 slot = prog->aux->size_poke_tab; 834 u32 size = slot + 1; 835 836 if (size > poke_tab_max) 837 return -ENOSPC; 838 if (poke->tailcall_target || poke->tailcall_target_stable || 839 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 840 return -EINVAL; 841 842 switch (poke->reason) { 843 case BPF_POKE_REASON_TAIL_CALL: 844 if (!poke->tail_call.map) 845 return -EINVAL; 846 break; 847 default: 848 return -EINVAL; 849 } 850 851 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 852 if (!tab) 853 return -ENOMEM; 854 855 memcpy(&tab[slot], poke, sizeof(*poke)); 856 prog->aux->size_poke_tab = size; 857 prog->aux->poke_tab = tab; 858 859 return slot; 860 } 861 862 /* 863 * BPF program pack allocator. 864 * 865 * Most BPF programs are pretty small. Allocating a hole page for each 866 * program is sometime a waste. Many small bpf program also adds pressure 867 * to instruction TLB. To solve this issue, we introduce a BPF program pack 868 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 869 * to host BPF programs. 870 */ 871 #define BPF_PROG_CHUNK_SHIFT 6 872 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 873 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 874 875 struct bpf_prog_pack { 876 struct list_head list; 877 void *ptr; 878 unsigned long bitmap[]; 879 }; 880 881 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 882 { 883 memset(area, 0, size); 884 } 885 886 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 887 888 static DEFINE_MUTEX(pack_mutex); 889 static LIST_HEAD(pack_list); 890 891 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 892 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 893 */ 894 #ifdef PMD_SIZE 895 /* PMD_SIZE is really big for some archs. It doesn't make sense to 896 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 897 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 898 * greater than or equal to 2MB. 899 */ 900 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 901 #else 902 #define BPF_PROG_PACK_SIZE PAGE_SIZE 903 #endif 904 905 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 906 907 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 908 { 909 struct bpf_prog_pack *pack; 910 int err; 911 912 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 913 GFP_KERNEL); 914 if (!pack) 915 return NULL; 916 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 917 if (!pack->ptr) 918 goto out; 919 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 920 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 921 922 set_vm_flush_reset_perms(pack->ptr); 923 err = set_memory_rox((unsigned long)pack->ptr, 924 BPF_PROG_PACK_SIZE / PAGE_SIZE); 925 if (err) 926 goto out; 927 list_add_tail(&pack->list, &pack_list); 928 return pack; 929 930 out: 931 bpf_jit_free_exec(pack->ptr); 932 kfree(pack); 933 return NULL; 934 } 935 936 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 937 { 938 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 939 struct bpf_prog_pack *pack; 940 unsigned long pos; 941 void *ptr = NULL; 942 943 mutex_lock(&pack_mutex); 944 if (size > BPF_PROG_PACK_SIZE) { 945 size = round_up(size, PAGE_SIZE); 946 ptr = bpf_jit_alloc_exec(size); 947 if (ptr) { 948 int err; 949 950 bpf_fill_ill_insns(ptr, size); 951 set_vm_flush_reset_perms(ptr); 952 err = set_memory_rox((unsigned long)ptr, 953 size / PAGE_SIZE); 954 if (err) { 955 bpf_jit_free_exec(ptr); 956 ptr = NULL; 957 } 958 } 959 goto out; 960 } 961 list_for_each_entry(pack, &pack_list, list) { 962 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 963 nbits, 0); 964 if (pos < BPF_PROG_CHUNK_COUNT) 965 goto found_free_area; 966 } 967 968 pack = alloc_new_pack(bpf_fill_ill_insns); 969 if (!pack) 970 goto out; 971 972 pos = 0; 973 974 found_free_area: 975 bitmap_set(pack->bitmap, pos, nbits); 976 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 977 978 out: 979 mutex_unlock(&pack_mutex); 980 return ptr; 981 } 982 983 void bpf_prog_pack_free(void *ptr, u32 size) 984 { 985 struct bpf_prog_pack *pack = NULL, *tmp; 986 unsigned int nbits; 987 unsigned long pos; 988 989 mutex_lock(&pack_mutex); 990 if (size > BPF_PROG_PACK_SIZE) { 991 bpf_jit_free_exec(ptr); 992 goto out; 993 } 994 995 list_for_each_entry(tmp, &pack_list, list) { 996 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 997 pack = tmp; 998 break; 999 } 1000 } 1001 1002 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1003 goto out; 1004 1005 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1006 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1007 1008 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1009 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1010 1011 bitmap_clear(pack->bitmap, pos, nbits); 1012 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1013 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1014 list_del(&pack->list); 1015 bpf_jit_free_exec(pack->ptr); 1016 kfree(pack); 1017 } 1018 out: 1019 mutex_unlock(&pack_mutex); 1020 } 1021 1022 static atomic_long_t bpf_jit_current; 1023 1024 /* Can be overridden by an arch's JIT compiler if it has a custom, 1025 * dedicated BPF backend memory area, or if neither of the two 1026 * below apply. 1027 */ 1028 u64 __weak bpf_jit_alloc_exec_limit(void) 1029 { 1030 #if defined(MODULES_VADDR) 1031 return MODULES_END - MODULES_VADDR; 1032 #else 1033 return VMALLOC_END - VMALLOC_START; 1034 #endif 1035 } 1036 1037 static int __init bpf_jit_charge_init(void) 1038 { 1039 /* Only used as heuristic here to derive limit. */ 1040 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1041 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1042 PAGE_SIZE), LONG_MAX); 1043 return 0; 1044 } 1045 pure_initcall(bpf_jit_charge_init); 1046 1047 int bpf_jit_charge_modmem(u32 size) 1048 { 1049 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1050 if (!bpf_capable()) { 1051 atomic_long_sub(size, &bpf_jit_current); 1052 return -EPERM; 1053 } 1054 } 1055 1056 return 0; 1057 } 1058 1059 void bpf_jit_uncharge_modmem(u32 size) 1060 { 1061 atomic_long_sub(size, &bpf_jit_current); 1062 } 1063 1064 void *__weak bpf_jit_alloc_exec(unsigned long size) 1065 { 1066 return execmem_alloc(EXECMEM_BPF, size); 1067 } 1068 1069 void __weak bpf_jit_free_exec(void *addr) 1070 { 1071 execmem_free(addr); 1072 } 1073 1074 struct bpf_binary_header * 1075 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1076 unsigned int alignment, 1077 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1078 { 1079 struct bpf_binary_header *hdr; 1080 u32 size, hole, start; 1081 1082 WARN_ON_ONCE(!is_power_of_2(alignment) || 1083 alignment > BPF_IMAGE_ALIGNMENT); 1084 1085 /* Most of BPF filters are really small, but if some of them 1086 * fill a page, allow at least 128 extra bytes to insert a 1087 * random section of illegal instructions. 1088 */ 1089 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1090 1091 if (bpf_jit_charge_modmem(size)) 1092 return NULL; 1093 hdr = bpf_jit_alloc_exec(size); 1094 if (!hdr) { 1095 bpf_jit_uncharge_modmem(size); 1096 return NULL; 1097 } 1098 1099 /* Fill space with illegal/arch-dep instructions. */ 1100 bpf_fill_ill_insns(hdr, size); 1101 1102 hdr->size = size; 1103 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1104 PAGE_SIZE - sizeof(*hdr)); 1105 start = get_random_u32_below(hole) & ~(alignment - 1); 1106 1107 /* Leave a random number of instructions before BPF code. */ 1108 *image_ptr = &hdr->image[start]; 1109 1110 return hdr; 1111 } 1112 1113 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1114 { 1115 u32 size = hdr->size; 1116 1117 bpf_jit_free_exec(hdr); 1118 bpf_jit_uncharge_modmem(size); 1119 } 1120 1121 /* Allocate jit binary from bpf_prog_pack allocator. 1122 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1123 * to the memory. To solve this problem, a RW buffer is also allocated at 1124 * as the same time. The JIT engine should calculate offsets based on the 1125 * RO memory address, but write JITed program to the RW buffer. Once the 1126 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1127 * the JITed program to the RO memory. 1128 */ 1129 struct bpf_binary_header * 1130 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1131 unsigned int alignment, 1132 struct bpf_binary_header **rw_header, 1133 u8 **rw_image, 1134 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1135 { 1136 struct bpf_binary_header *ro_header; 1137 u32 size, hole, start; 1138 1139 WARN_ON_ONCE(!is_power_of_2(alignment) || 1140 alignment > BPF_IMAGE_ALIGNMENT); 1141 1142 /* add 16 bytes for a random section of illegal instructions */ 1143 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1144 1145 if (bpf_jit_charge_modmem(size)) 1146 return NULL; 1147 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1148 if (!ro_header) { 1149 bpf_jit_uncharge_modmem(size); 1150 return NULL; 1151 } 1152 1153 *rw_header = kvmalloc(size, GFP_KERNEL); 1154 if (!*rw_header) { 1155 bpf_prog_pack_free(ro_header, size); 1156 bpf_jit_uncharge_modmem(size); 1157 return NULL; 1158 } 1159 1160 /* Fill space with illegal/arch-dep instructions. */ 1161 bpf_fill_ill_insns(*rw_header, size); 1162 (*rw_header)->size = size; 1163 1164 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1165 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1166 start = get_random_u32_below(hole) & ~(alignment - 1); 1167 1168 *image_ptr = &ro_header->image[start]; 1169 *rw_image = &(*rw_header)->image[start]; 1170 1171 return ro_header; 1172 } 1173 1174 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1175 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1176 struct bpf_binary_header *rw_header) 1177 { 1178 void *ptr; 1179 1180 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1181 1182 kvfree(rw_header); 1183 1184 if (IS_ERR(ptr)) { 1185 bpf_prog_pack_free(ro_header, ro_header->size); 1186 return PTR_ERR(ptr); 1187 } 1188 return 0; 1189 } 1190 1191 /* bpf_jit_binary_pack_free is called in two different scenarios: 1192 * 1) when the program is freed after; 1193 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1194 * For case 2), we need to free both the RO memory and the RW buffer. 1195 * 1196 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1197 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1198 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1199 * bpf_arch_text_copy (when jit fails). 1200 */ 1201 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1202 struct bpf_binary_header *rw_header) 1203 { 1204 u32 size = ro_header->size; 1205 1206 bpf_prog_pack_free(ro_header, size); 1207 kvfree(rw_header); 1208 bpf_jit_uncharge_modmem(size); 1209 } 1210 1211 struct bpf_binary_header * 1212 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1213 { 1214 unsigned long real_start = (unsigned long)fp->bpf_func; 1215 unsigned long addr; 1216 1217 addr = real_start & BPF_PROG_CHUNK_MASK; 1218 return (void *)addr; 1219 } 1220 1221 static inline struct bpf_binary_header * 1222 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1223 { 1224 unsigned long real_start = (unsigned long)fp->bpf_func; 1225 unsigned long addr; 1226 1227 addr = real_start & PAGE_MASK; 1228 return (void *)addr; 1229 } 1230 1231 /* This symbol is only overridden by archs that have different 1232 * requirements than the usual eBPF JITs, f.e. when they only 1233 * implement cBPF JIT, do not set images read-only, etc. 1234 */ 1235 void __weak bpf_jit_free(struct bpf_prog *fp) 1236 { 1237 if (fp->jited) { 1238 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1239 1240 bpf_jit_binary_free(hdr); 1241 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1242 } 1243 1244 bpf_prog_unlock_free(fp); 1245 } 1246 1247 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1248 const struct bpf_insn *insn, bool extra_pass, 1249 u64 *func_addr, bool *func_addr_fixed) 1250 { 1251 s16 off = insn->off; 1252 s32 imm = insn->imm; 1253 u8 *addr; 1254 int err; 1255 1256 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1257 if (!*func_addr_fixed) { 1258 /* Place-holder address till the last pass has collected 1259 * all addresses for JITed subprograms in which case we 1260 * can pick them up from prog->aux. 1261 */ 1262 if (!extra_pass) 1263 addr = NULL; 1264 else if (prog->aux->func && 1265 off >= 0 && off < prog->aux->real_func_cnt) 1266 addr = (u8 *)prog->aux->func[off]->bpf_func; 1267 else 1268 return -EINVAL; 1269 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1270 bpf_jit_supports_far_kfunc_call()) { 1271 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1272 if (err) 1273 return err; 1274 } else { 1275 /* Address of a BPF helper call. Since part of the core 1276 * kernel, it's always at a fixed location. __bpf_call_base 1277 * and the helper with imm relative to it are both in core 1278 * kernel. 1279 */ 1280 addr = (u8 *)__bpf_call_base + imm; 1281 } 1282 1283 *func_addr = (unsigned long)addr; 1284 return 0; 1285 } 1286 1287 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1288 { 1289 if (prog->aux->ksym.prog) 1290 return prog->aux->ksym.name; 1291 return prog->aux->name; 1292 } 1293 1294 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1295 const struct bpf_insn *aux, 1296 struct bpf_insn *to_buff, 1297 bool emit_zext) 1298 { 1299 struct bpf_insn *to = to_buff; 1300 u32 imm_rnd = get_random_u32(); 1301 s16 off; 1302 1303 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1304 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1305 1306 /* Constraints on AX register: 1307 * 1308 * AX register is inaccessible from user space. It is mapped in 1309 * all JITs, and used here for constant blinding rewrites. It is 1310 * typically "stateless" meaning its contents are only valid within 1311 * the executed instruction, but not across several instructions. 1312 * There are a few exceptions however which are further detailed 1313 * below. 1314 * 1315 * Constant blinding is only used by JITs, not in the interpreter. 1316 * The interpreter uses AX in some occasions as a local temporary 1317 * register e.g. in DIV or MOD instructions. 1318 * 1319 * In restricted circumstances, the verifier can also use the AX 1320 * register for rewrites as long as they do not interfere with 1321 * the above cases! 1322 */ 1323 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1324 goto out; 1325 1326 if (from->imm == 0 && 1327 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1328 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1329 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1330 goto out; 1331 } 1332 1333 switch (from->code) { 1334 case BPF_ALU | BPF_ADD | BPF_K: 1335 case BPF_ALU | BPF_SUB | BPF_K: 1336 case BPF_ALU | BPF_AND | BPF_K: 1337 case BPF_ALU | BPF_OR | BPF_K: 1338 case BPF_ALU | BPF_XOR | BPF_K: 1339 case BPF_ALU | BPF_MUL | BPF_K: 1340 case BPF_ALU | BPF_MOV | BPF_K: 1341 case BPF_ALU | BPF_DIV | BPF_K: 1342 case BPF_ALU | BPF_MOD | BPF_K: 1343 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1344 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1345 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1346 break; 1347 1348 case BPF_ALU64 | BPF_ADD | BPF_K: 1349 case BPF_ALU64 | BPF_SUB | BPF_K: 1350 case BPF_ALU64 | BPF_AND | BPF_K: 1351 case BPF_ALU64 | BPF_OR | BPF_K: 1352 case BPF_ALU64 | BPF_XOR | BPF_K: 1353 case BPF_ALU64 | BPF_MUL | BPF_K: 1354 case BPF_ALU64 | BPF_MOV | BPF_K: 1355 case BPF_ALU64 | BPF_DIV | BPF_K: 1356 case BPF_ALU64 | BPF_MOD | BPF_K: 1357 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1358 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1359 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1360 break; 1361 1362 case BPF_JMP | BPF_JEQ | BPF_K: 1363 case BPF_JMP | BPF_JNE | BPF_K: 1364 case BPF_JMP | BPF_JGT | BPF_K: 1365 case BPF_JMP | BPF_JLT | BPF_K: 1366 case BPF_JMP | BPF_JGE | BPF_K: 1367 case BPF_JMP | BPF_JLE | BPF_K: 1368 case BPF_JMP | BPF_JSGT | BPF_K: 1369 case BPF_JMP | BPF_JSLT | BPF_K: 1370 case BPF_JMP | BPF_JSGE | BPF_K: 1371 case BPF_JMP | BPF_JSLE | BPF_K: 1372 case BPF_JMP | BPF_JSET | BPF_K: 1373 /* Accommodate for extra offset in case of a backjump. */ 1374 off = from->off; 1375 if (off < 0) 1376 off -= 2; 1377 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1378 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1379 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1380 break; 1381 1382 case BPF_JMP32 | BPF_JEQ | BPF_K: 1383 case BPF_JMP32 | BPF_JNE | BPF_K: 1384 case BPF_JMP32 | BPF_JGT | BPF_K: 1385 case BPF_JMP32 | BPF_JLT | BPF_K: 1386 case BPF_JMP32 | BPF_JGE | BPF_K: 1387 case BPF_JMP32 | BPF_JLE | BPF_K: 1388 case BPF_JMP32 | BPF_JSGT | BPF_K: 1389 case BPF_JMP32 | BPF_JSLT | BPF_K: 1390 case BPF_JMP32 | BPF_JSGE | BPF_K: 1391 case BPF_JMP32 | BPF_JSLE | BPF_K: 1392 case BPF_JMP32 | BPF_JSET | BPF_K: 1393 /* Accommodate for extra offset in case of a backjump. */ 1394 off = from->off; 1395 if (off < 0) 1396 off -= 2; 1397 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1398 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1399 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1400 off); 1401 break; 1402 1403 case BPF_LD | BPF_IMM | BPF_DW: 1404 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1405 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1406 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1407 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1408 break; 1409 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1410 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1411 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1412 if (emit_zext) 1413 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1414 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1415 break; 1416 1417 case BPF_ST | BPF_MEM | BPF_DW: 1418 case BPF_ST | BPF_MEM | BPF_W: 1419 case BPF_ST | BPF_MEM | BPF_H: 1420 case BPF_ST | BPF_MEM | BPF_B: 1421 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1422 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1423 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1424 break; 1425 } 1426 out: 1427 return to - to_buff; 1428 } 1429 1430 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1431 gfp_t gfp_extra_flags) 1432 { 1433 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1434 struct bpf_prog *fp; 1435 1436 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1437 if (fp != NULL) { 1438 /* aux->prog still points to the fp_other one, so 1439 * when promoting the clone to the real program, 1440 * this still needs to be adapted. 1441 */ 1442 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1443 } 1444 1445 return fp; 1446 } 1447 1448 static void bpf_prog_clone_free(struct bpf_prog *fp) 1449 { 1450 /* aux was stolen by the other clone, so we cannot free 1451 * it from this path! It will be freed eventually by the 1452 * other program on release. 1453 * 1454 * At this point, we don't need a deferred release since 1455 * clone is guaranteed to not be locked. 1456 */ 1457 fp->aux = NULL; 1458 fp->stats = NULL; 1459 fp->active = NULL; 1460 __bpf_prog_free(fp); 1461 } 1462 1463 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1464 { 1465 /* We have to repoint aux->prog to self, as we don't 1466 * know whether fp here is the clone or the original. 1467 */ 1468 fp->aux->prog = fp; 1469 bpf_prog_clone_free(fp_other); 1470 } 1471 1472 static void adjust_insn_arrays(struct bpf_prog *prog, u32 off, u32 len) 1473 { 1474 #ifdef CONFIG_BPF_SYSCALL 1475 struct bpf_map *map; 1476 int i; 1477 1478 if (len <= 1) 1479 return; 1480 1481 for (i = 0; i < prog->aux->used_map_cnt; i++) { 1482 map = prog->aux->used_maps[i]; 1483 if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY) 1484 bpf_insn_array_adjust(map, off, len); 1485 } 1486 #endif 1487 } 1488 1489 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1490 { 1491 struct bpf_insn insn_buff[16], aux[2]; 1492 struct bpf_prog *clone, *tmp; 1493 int insn_delta, insn_cnt; 1494 struct bpf_insn *insn; 1495 int i, rewritten; 1496 1497 if (!prog->blinding_requested || prog->blinded) 1498 return prog; 1499 1500 clone = bpf_prog_clone_create(prog, GFP_USER); 1501 if (!clone) 1502 return ERR_PTR(-ENOMEM); 1503 1504 insn_cnt = clone->len; 1505 insn = clone->insnsi; 1506 1507 for (i = 0; i < insn_cnt; i++, insn++) { 1508 if (bpf_pseudo_func(insn)) { 1509 /* ld_imm64 with an address of bpf subprog is not 1510 * a user controlled constant. Don't randomize it, 1511 * since it will conflict with jit_subprogs() logic. 1512 */ 1513 insn++; 1514 i++; 1515 continue; 1516 } 1517 1518 /* We temporarily need to hold the original ld64 insn 1519 * so that we can still access the first part in the 1520 * second blinding run. 1521 */ 1522 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1523 insn[1].code == 0) 1524 memcpy(aux, insn, sizeof(aux)); 1525 1526 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1527 clone->aux->verifier_zext); 1528 if (!rewritten) 1529 continue; 1530 1531 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1532 if (IS_ERR(tmp)) { 1533 /* Patching may have repointed aux->prog during 1534 * realloc from the original one, so we need to 1535 * fix it up here on error. 1536 */ 1537 bpf_jit_prog_release_other(prog, clone); 1538 return tmp; 1539 } 1540 1541 clone = tmp; 1542 insn_delta = rewritten - 1; 1543 1544 /* Instructions arrays must be updated using absolute xlated offsets */ 1545 adjust_insn_arrays(clone, prog->aux->subprog_start + i, rewritten); 1546 1547 /* Walk new program and skip insns we just inserted. */ 1548 insn = clone->insnsi + i + insn_delta; 1549 insn_cnt += insn_delta; 1550 i += insn_delta; 1551 } 1552 1553 clone->blinded = 1; 1554 return clone; 1555 } 1556 #endif /* CONFIG_BPF_JIT */ 1557 1558 /* Base function for offset calculation. Needs to go into .text section, 1559 * therefore keeping it non-static as well; will also be used by JITs 1560 * anyway later on, so do not let the compiler omit it. This also needs 1561 * to go into kallsyms for correlation from e.g. bpftool, so naming 1562 * must not change. 1563 */ 1564 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1565 { 1566 return 0; 1567 } 1568 EXPORT_SYMBOL_GPL(__bpf_call_base); 1569 1570 /* All UAPI available opcodes. */ 1571 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1572 /* 32 bit ALU operations. */ \ 1573 /* Register based. */ \ 1574 INSN_3(ALU, ADD, X), \ 1575 INSN_3(ALU, SUB, X), \ 1576 INSN_3(ALU, AND, X), \ 1577 INSN_3(ALU, OR, X), \ 1578 INSN_3(ALU, LSH, X), \ 1579 INSN_3(ALU, RSH, X), \ 1580 INSN_3(ALU, XOR, X), \ 1581 INSN_3(ALU, MUL, X), \ 1582 INSN_3(ALU, MOV, X), \ 1583 INSN_3(ALU, ARSH, X), \ 1584 INSN_3(ALU, DIV, X), \ 1585 INSN_3(ALU, MOD, X), \ 1586 INSN_2(ALU, NEG), \ 1587 INSN_3(ALU, END, TO_BE), \ 1588 INSN_3(ALU, END, TO_LE), \ 1589 /* Immediate based. */ \ 1590 INSN_3(ALU, ADD, K), \ 1591 INSN_3(ALU, SUB, K), \ 1592 INSN_3(ALU, AND, K), \ 1593 INSN_3(ALU, OR, K), \ 1594 INSN_3(ALU, LSH, K), \ 1595 INSN_3(ALU, RSH, K), \ 1596 INSN_3(ALU, XOR, K), \ 1597 INSN_3(ALU, MUL, K), \ 1598 INSN_3(ALU, MOV, K), \ 1599 INSN_3(ALU, ARSH, K), \ 1600 INSN_3(ALU, DIV, K), \ 1601 INSN_3(ALU, MOD, K), \ 1602 /* 64 bit ALU operations. */ \ 1603 /* Register based. */ \ 1604 INSN_3(ALU64, ADD, X), \ 1605 INSN_3(ALU64, SUB, X), \ 1606 INSN_3(ALU64, AND, X), \ 1607 INSN_3(ALU64, OR, X), \ 1608 INSN_3(ALU64, LSH, X), \ 1609 INSN_3(ALU64, RSH, X), \ 1610 INSN_3(ALU64, XOR, X), \ 1611 INSN_3(ALU64, MUL, X), \ 1612 INSN_3(ALU64, MOV, X), \ 1613 INSN_3(ALU64, ARSH, X), \ 1614 INSN_3(ALU64, DIV, X), \ 1615 INSN_3(ALU64, MOD, X), \ 1616 INSN_2(ALU64, NEG), \ 1617 INSN_3(ALU64, END, TO_LE), \ 1618 /* Immediate based. */ \ 1619 INSN_3(ALU64, ADD, K), \ 1620 INSN_3(ALU64, SUB, K), \ 1621 INSN_3(ALU64, AND, K), \ 1622 INSN_3(ALU64, OR, K), \ 1623 INSN_3(ALU64, LSH, K), \ 1624 INSN_3(ALU64, RSH, K), \ 1625 INSN_3(ALU64, XOR, K), \ 1626 INSN_3(ALU64, MUL, K), \ 1627 INSN_3(ALU64, MOV, K), \ 1628 INSN_3(ALU64, ARSH, K), \ 1629 INSN_3(ALU64, DIV, K), \ 1630 INSN_3(ALU64, MOD, K), \ 1631 /* Call instruction. */ \ 1632 INSN_2(JMP, CALL), \ 1633 /* Exit instruction. */ \ 1634 INSN_2(JMP, EXIT), \ 1635 /* 32-bit Jump instructions. */ \ 1636 /* Register based. */ \ 1637 INSN_3(JMP32, JEQ, X), \ 1638 INSN_3(JMP32, JNE, X), \ 1639 INSN_3(JMP32, JGT, X), \ 1640 INSN_3(JMP32, JLT, X), \ 1641 INSN_3(JMP32, JGE, X), \ 1642 INSN_3(JMP32, JLE, X), \ 1643 INSN_3(JMP32, JSGT, X), \ 1644 INSN_3(JMP32, JSLT, X), \ 1645 INSN_3(JMP32, JSGE, X), \ 1646 INSN_3(JMP32, JSLE, X), \ 1647 INSN_3(JMP32, JSET, X), \ 1648 /* Immediate based. */ \ 1649 INSN_3(JMP32, JEQ, K), \ 1650 INSN_3(JMP32, JNE, K), \ 1651 INSN_3(JMP32, JGT, K), \ 1652 INSN_3(JMP32, JLT, K), \ 1653 INSN_3(JMP32, JGE, K), \ 1654 INSN_3(JMP32, JLE, K), \ 1655 INSN_3(JMP32, JSGT, K), \ 1656 INSN_3(JMP32, JSLT, K), \ 1657 INSN_3(JMP32, JSGE, K), \ 1658 INSN_3(JMP32, JSLE, K), \ 1659 INSN_3(JMP32, JSET, K), \ 1660 /* Jump instructions. */ \ 1661 /* Register based. */ \ 1662 INSN_3(JMP, JEQ, X), \ 1663 INSN_3(JMP, JNE, X), \ 1664 INSN_3(JMP, JGT, X), \ 1665 INSN_3(JMP, JLT, X), \ 1666 INSN_3(JMP, JGE, X), \ 1667 INSN_3(JMP, JLE, X), \ 1668 INSN_3(JMP, JSGT, X), \ 1669 INSN_3(JMP, JSLT, X), \ 1670 INSN_3(JMP, JSGE, X), \ 1671 INSN_3(JMP, JSLE, X), \ 1672 INSN_3(JMP, JSET, X), \ 1673 /* Immediate based. */ \ 1674 INSN_3(JMP, JEQ, K), \ 1675 INSN_3(JMP, JNE, K), \ 1676 INSN_3(JMP, JGT, K), \ 1677 INSN_3(JMP, JLT, K), \ 1678 INSN_3(JMP, JGE, K), \ 1679 INSN_3(JMP, JLE, K), \ 1680 INSN_3(JMP, JSGT, K), \ 1681 INSN_3(JMP, JSLT, K), \ 1682 INSN_3(JMP, JSGE, K), \ 1683 INSN_3(JMP, JSLE, K), \ 1684 INSN_3(JMP, JSET, K), \ 1685 INSN_2(JMP, JA), \ 1686 INSN_2(JMP32, JA), \ 1687 /* Atomic operations. */ \ 1688 INSN_3(STX, ATOMIC, B), \ 1689 INSN_3(STX, ATOMIC, H), \ 1690 INSN_3(STX, ATOMIC, W), \ 1691 INSN_3(STX, ATOMIC, DW), \ 1692 /* Store instructions. */ \ 1693 /* Register based. */ \ 1694 INSN_3(STX, MEM, B), \ 1695 INSN_3(STX, MEM, H), \ 1696 INSN_3(STX, MEM, W), \ 1697 INSN_3(STX, MEM, DW), \ 1698 /* Immediate based. */ \ 1699 INSN_3(ST, MEM, B), \ 1700 INSN_3(ST, MEM, H), \ 1701 INSN_3(ST, MEM, W), \ 1702 INSN_3(ST, MEM, DW), \ 1703 /* Load instructions. */ \ 1704 /* Register based. */ \ 1705 INSN_3(LDX, MEM, B), \ 1706 INSN_3(LDX, MEM, H), \ 1707 INSN_3(LDX, MEM, W), \ 1708 INSN_3(LDX, MEM, DW), \ 1709 INSN_3(LDX, MEMSX, B), \ 1710 INSN_3(LDX, MEMSX, H), \ 1711 INSN_3(LDX, MEMSX, W), \ 1712 /* Immediate based. */ \ 1713 INSN_3(LD, IMM, DW) 1714 1715 bool bpf_opcode_in_insntable(u8 code) 1716 { 1717 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1718 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1719 static const bool public_insntable[256] = { 1720 [0 ... 255] = false, 1721 /* Now overwrite non-defaults ... */ 1722 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1723 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1724 [BPF_LD | BPF_ABS | BPF_B] = true, 1725 [BPF_LD | BPF_ABS | BPF_H] = true, 1726 [BPF_LD | BPF_ABS | BPF_W] = true, 1727 [BPF_LD | BPF_IND | BPF_B] = true, 1728 [BPF_LD | BPF_IND | BPF_H] = true, 1729 [BPF_LD | BPF_IND | BPF_W] = true, 1730 [BPF_JMP | BPF_JA | BPF_X] = true, 1731 [BPF_JMP | BPF_JCOND] = true, 1732 }; 1733 #undef BPF_INSN_3_TBL 1734 #undef BPF_INSN_2_TBL 1735 return public_insntable[code]; 1736 } 1737 1738 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1739 /** 1740 * ___bpf_prog_run - run eBPF program on a given context 1741 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1742 * @insn: is the array of eBPF instructions 1743 * 1744 * Decode and execute eBPF instructions. 1745 * 1746 * Return: whatever value is in %BPF_R0 at program exit 1747 */ 1748 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1749 { 1750 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1751 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1752 static const void * const jumptable[256] __annotate_jump_table = { 1753 [0 ... 255] = &&default_label, 1754 /* Now overwrite non-defaults ... */ 1755 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1756 /* Non-UAPI available opcodes. */ 1757 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1758 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1759 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1760 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1761 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1762 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1763 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1764 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1765 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1766 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1767 }; 1768 #undef BPF_INSN_3_LBL 1769 #undef BPF_INSN_2_LBL 1770 u32 tail_call_cnt = 0; 1771 1772 #define CONT ({ insn++; goto select_insn; }) 1773 #define CONT_JMP ({ insn++; goto select_insn; }) 1774 1775 select_insn: 1776 goto *jumptable[insn->code]; 1777 1778 /* Explicitly mask the register-based shift amounts with 63 or 31 1779 * to avoid undefined behavior. Normally this won't affect the 1780 * generated code, for example, in case of native 64 bit archs such 1781 * as x86-64 or arm64, the compiler is optimizing the AND away for 1782 * the interpreter. In case of JITs, each of the JIT backends compiles 1783 * the BPF shift operations to machine instructions which produce 1784 * implementation-defined results in such a case; the resulting 1785 * contents of the register may be arbitrary, but program behaviour 1786 * as a whole remains defined. In other words, in case of JIT backends, 1787 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1788 */ 1789 /* ALU (shifts) */ 1790 #define SHT(OPCODE, OP) \ 1791 ALU64_##OPCODE##_X: \ 1792 DST = DST OP (SRC & 63); \ 1793 CONT; \ 1794 ALU_##OPCODE##_X: \ 1795 DST = (u32) DST OP ((u32) SRC & 31); \ 1796 CONT; \ 1797 ALU64_##OPCODE##_K: \ 1798 DST = DST OP IMM; \ 1799 CONT; \ 1800 ALU_##OPCODE##_K: \ 1801 DST = (u32) DST OP (u32) IMM; \ 1802 CONT; 1803 /* ALU (rest) */ 1804 #define ALU(OPCODE, OP) \ 1805 ALU64_##OPCODE##_X: \ 1806 DST = DST OP SRC; \ 1807 CONT; \ 1808 ALU_##OPCODE##_X: \ 1809 DST = (u32) DST OP (u32) SRC; \ 1810 CONT; \ 1811 ALU64_##OPCODE##_K: \ 1812 DST = DST OP IMM; \ 1813 CONT; \ 1814 ALU_##OPCODE##_K: \ 1815 DST = (u32) DST OP (u32) IMM; \ 1816 CONT; 1817 ALU(ADD, +) 1818 ALU(SUB, -) 1819 ALU(AND, &) 1820 ALU(OR, |) 1821 ALU(XOR, ^) 1822 ALU(MUL, *) 1823 SHT(LSH, <<) 1824 SHT(RSH, >>) 1825 #undef SHT 1826 #undef ALU 1827 ALU_NEG: 1828 DST = (u32) -DST; 1829 CONT; 1830 ALU64_NEG: 1831 DST = -DST; 1832 CONT; 1833 ALU_MOV_X: 1834 switch (OFF) { 1835 case 0: 1836 DST = (u32) SRC; 1837 break; 1838 case 8: 1839 DST = (u32)(s8) SRC; 1840 break; 1841 case 16: 1842 DST = (u32)(s16) SRC; 1843 break; 1844 } 1845 CONT; 1846 ALU_MOV_K: 1847 DST = (u32) IMM; 1848 CONT; 1849 ALU64_MOV_X: 1850 switch (OFF) { 1851 case 0: 1852 DST = SRC; 1853 break; 1854 case 8: 1855 DST = (s8) SRC; 1856 break; 1857 case 16: 1858 DST = (s16) SRC; 1859 break; 1860 case 32: 1861 DST = (s32) SRC; 1862 break; 1863 } 1864 CONT; 1865 ALU64_MOV_K: 1866 DST = IMM; 1867 CONT; 1868 LD_IMM_DW: 1869 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1870 insn++; 1871 CONT; 1872 ALU_ARSH_X: 1873 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1874 CONT; 1875 ALU_ARSH_K: 1876 DST = (u64) (u32) (((s32) DST) >> IMM); 1877 CONT; 1878 ALU64_ARSH_X: 1879 (*(s64 *) &DST) >>= (SRC & 63); 1880 CONT; 1881 ALU64_ARSH_K: 1882 (*(s64 *) &DST) >>= IMM; 1883 CONT; 1884 ALU64_MOD_X: 1885 switch (OFF) { 1886 case 0: 1887 div64_u64_rem(DST, SRC, &AX); 1888 DST = AX; 1889 break; 1890 case 1: 1891 AX = div64_s64(DST, SRC); 1892 DST = DST - AX * SRC; 1893 break; 1894 } 1895 CONT; 1896 ALU_MOD_X: 1897 switch (OFF) { 1898 case 0: 1899 AX = (u32) DST; 1900 DST = do_div(AX, (u32) SRC); 1901 break; 1902 case 1: 1903 AX = abs((s32)DST); 1904 AX = do_div(AX, abs((s32)SRC)); 1905 if ((s32)DST < 0) 1906 DST = (u32)-AX; 1907 else 1908 DST = (u32)AX; 1909 break; 1910 } 1911 CONT; 1912 ALU64_MOD_K: 1913 switch (OFF) { 1914 case 0: 1915 div64_u64_rem(DST, IMM, &AX); 1916 DST = AX; 1917 break; 1918 case 1: 1919 AX = div64_s64(DST, IMM); 1920 DST = DST - AX * IMM; 1921 break; 1922 } 1923 CONT; 1924 ALU_MOD_K: 1925 switch (OFF) { 1926 case 0: 1927 AX = (u32) DST; 1928 DST = do_div(AX, (u32) IMM); 1929 break; 1930 case 1: 1931 AX = abs((s32)DST); 1932 AX = do_div(AX, abs((s32)IMM)); 1933 if ((s32)DST < 0) 1934 DST = (u32)-AX; 1935 else 1936 DST = (u32)AX; 1937 break; 1938 } 1939 CONT; 1940 ALU64_DIV_X: 1941 switch (OFF) { 1942 case 0: 1943 DST = div64_u64(DST, SRC); 1944 break; 1945 case 1: 1946 DST = div64_s64(DST, SRC); 1947 break; 1948 } 1949 CONT; 1950 ALU_DIV_X: 1951 switch (OFF) { 1952 case 0: 1953 AX = (u32) DST; 1954 do_div(AX, (u32) SRC); 1955 DST = (u32) AX; 1956 break; 1957 case 1: 1958 AX = abs((s32)DST); 1959 do_div(AX, abs((s32)SRC)); 1960 if (((s32)DST < 0) == ((s32)SRC < 0)) 1961 DST = (u32)AX; 1962 else 1963 DST = (u32)-AX; 1964 break; 1965 } 1966 CONT; 1967 ALU64_DIV_K: 1968 switch (OFF) { 1969 case 0: 1970 DST = div64_u64(DST, IMM); 1971 break; 1972 case 1: 1973 DST = div64_s64(DST, IMM); 1974 break; 1975 } 1976 CONT; 1977 ALU_DIV_K: 1978 switch (OFF) { 1979 case 0: 1980 AX = (u32) DST; 1981 do_div(AX, (u32) IMM); 1982 DST = (u32) AX; 1983 break; 1984 case 1: 1985 AX = abs((s32)DST); 1986 do_div(AX, abs((s32)IMM)); 1987 if (((s32)DST < 0) == ((s32)IMM < 0)) 1988 DST = (u32)AX; 1989 else 1990 DST = (u32)-AX; 1991 break; 1992 } 1993 CONT; 1994 ALU_END_TO_BE: 1995 switch (IMM) { 1996 case 16: 1997 DST = (__force u16) cpu_to_be16(DST); 1998 break; 1999 case 32: 2000 DST = (__force u32) cpu_to_be32(DST); 2001 break; 2002 case 64: 2003 DST = (__force u64) cpu_to_be64(DST); 2004 break; 2005 } 2006 CONT; 2007 ALU_END_TO_LE: 2008 switch (IMM) { 2009 case 16: 2010 DST = (__force u16) cpu_to_le16(DST); 2011 break; 2012 case 32: 2013 DST = (__force u32) cpu_to_le32(DST); 2014 break; 2015 case 64: 2016 DST = (__force u64) cpu_to_le64(DST); 2017 break; 2018 } 2019 CONT; 2020 ALU64_END_TO_LE: 2021 switch (IMM) { 2022 case 16: 2023 DST = (__force u16) __swab16(DST); 2024 break; 2025 case 32: 2026 DST = (__force u32) __swab32(DST); 2027 break; 2028 case 64: 2029 DST = (__force u64) __swab64(DST); 2030 break; 2031 } 2032 CONT; 2033 2034 /* CALL */ 2035 JMP_CALL: 2036 /* Function call scratches BPF_R1-BPF_R5 registers, 2037 * preserves BPF_R6-BPF_R9, and stores return value 2038 * into BPF_R0. 2039 */ 2040 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2041 BPF_R4, BPF_R5); 2042 CONT; 2043 2044 JMP_CALL_ARGS: 2045 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2046 BPF_R3, BPF_R4, 2047 BPF_R5, 2048 insn + insn->off + 1); 2049 CONT; 2050 2051 JMP_TAIL_CALL: { 2052 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2053 struct bpf_array *array = container_of(map, struct bpf_array, map); 2054 struct bpf_prog *prog; 2055 u32 index = BPF_R3; 2056 2057 if (unlikely(index >= array->map.max_entries)) 2058 goto out; 2059 2060 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2061 goto out; 2062 2063 tail_call_cnt++; 2064 2065 prog = READ_ONCE(array->ptrs[index]); 2066 if (!prog) 2067 goto out; 2068 2069 /* ARG1 at this point is guaranteed to point to CTX from 2070 * the verifier side due to the fact that the tail call is 2071 * handled like a helper, that is, bpf_tail_call_proto, 2072 * where arg1_type is ARG_PTR_TO_CTX. 2073 */ 2074 insn = prog->insnsi; 2075 goto select_insn; 2076 out: 2077 CONT; 2078 } 2079 JMP_JA: 2080 insn += insn->off; 2081 CONT; 2082 JMP32_JA: 2083 insn += insn->imm; 2084 CONT; 2085 JMP_EXIT: 2086 return BPF_R0; 2087 /* JMP */ 2088 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2089 JMP_##OPCODE##_X: \ 2090 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2091 insn += insn->off; \ 2092 CONT_JMP; \ 2093 } \ 2094 CONT; \ 2095 JMP32_##OPCODE##_X: \ 2096 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2097 insn += insn->off; \ 2098 CONT_JMP; \ 2099 } \ 2100 CONT; \ 2101 JMP_##OPCODE##_K: \ 2102 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2103 insn += insn->off; \ 2104 CONT_JMP; \ 2105 } \ 2106 CONT; \ 2107 JMP32_##OPCODE##_K: \ 2108 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2109 insn += insn->off; \ 2110 CONT_JMP; \ 2111 } \ 2112 CONT; 2113 COND_JMP(u, JEQ, ==) 2114 COND_JMP(u, JNE, !=) 2115 COND_JMP(u, JGT, >) 2116 COND_JMP(u, JLT, <) 2117 COND_JMP(u, JGE, >=) 2118 COND_JMP(u, JLE, <=) 2119 COND_JMP(u, JSET, &) 2120 COND_JMP(s, JSGT, >) 2121 COND_JMP(s, JSLT, <) 2122 COND_JMP(s, JSGE, >=) 2123 COND_JMP(s, JSLE, <=) 2124 #undef COND_JMP 2125 /* ST, STX and LDX*/ 2126 ST_NOSPEC: 2127 /* Speculation barrier for mitigating Speculative Store Bypass, 2128 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2129 * rely on the firmware mitigation as controlled via the ssbd 2130 * kernel parameter. Whenever the mitigation is enabled, it 2131 * works for all of the kernel code with no need to provide any 2132 * additional instructions here. In case of x86, we use 'lfence' 2133 * insn for mitigation. We reuse preexisting logic from Spectre 2134 * v1 mitigation that happens to produce the required code on 2135 * x86 for v4 as well. 2136 */ 2137 barrier_nospec(); 2138 CONT; 2139 #define LDST(SIZEOP, SIZE) \ 2140 STX_MEM_##SIZEOP: \ 2141 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2142 CONT; \ 2143 ST_MEM_##SIZEOP: \ 2144 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2145 CONT; \ 2146 LDX_MEM_##SIZEOP: \ 2147 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2148 CONT; \ 2149 LDX_PROBE_MEM_##SIZEOP: \ 2150 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2151 (const void *)(long) (SRC + insn->off)); \ 2152 DST = *((SIZE *)&DST); \ 2153 CONT; 2154 2155 LDST(B, u8) 2156 LDST(H, u16) 2157 LDST(W, u32) 2158 LDST(DW, u64) 2159 #undef LDST 2160 2161 #define LDSX(SIZEOP, SIZE) \ 2162 LDX_MEMSX_##SIZEOP: \ 2163 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2164 CONT; \ 2165 LDX_PROBE_MEMSX_##SIZEOP: \ 2166 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2167 (const void *)(long) (SRC + insn->off)); \ 2168 DST = *((SIZE *)&DST); \ 2169 CONT; 2170 2171 LDSX(B, s8) 2172 LDSX(H, s16) 2173 LDSX(W, s32) 2174 #undef LDSX 2175 2176 #define ATOMIC_ALU_OP(BOP, KOP) \ 2177 case BOP: \ 2178 if (BPF_SIZE(insn->code) == BPF_W) \ 2179 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2180 (DST + insn->off)); \ 2181 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2182 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2183 (DST + insn->off)); \ 2184 else \ 2185 goto default_label; \ 2186 break; \ 2187 case BOP | BPF_FETCH: \ 2188 if (BPF_SIZE(insn->code) == BPF_W) \ 2189 SRC = (u32) atomic_fetch_##KOP( \ 2190 (u32) SRC, \ 2191 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2192 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2193 SRC = (u64) atomic64_fetch_##KOP( \ 2194 (u64) SRC, \ 2195 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2196 else \ 2197 goto default_label; \ 2198 break; 2199 2200 STX_ATOMIC_DW: 2201 STX_ATOMIC_W: 2202 STX_ATOMIC_H: 2203 STX_ATOMIC_B: 2204 switch (IMM) { 2205 /* Atomic read-modify-write instructions support only W and DW 2206 * size modifiers. 2207 */ 2208 ATOMIC_ALU_OP(BPF_ADD, add) 2209 ATOMIC_ALU_OP(BPF_AND, and) 2210 ATOMIC_ALU_OP(BPF_OR, or) 2211 ATOMIC_ALU_OP(BPF_XOR, xor) 2212 #undef ATOMIC_ALU_OP 2213 2214 case BPF_XCHG: 2215 if (BPF_SIZE(insn->code) == BPF_W) 2216 SRC = (u32) atomic_xchg( 2217 (atomic_t *)(unsigned long) (DST + insn->off), 2218 (u32) SRC); 2219 else if (BPF_SIZE(insn->code) == BPF_DW) 2220 SRC = (u64) atomic64_xchg( 2221 (atomic64_t *)(unsigned long) (DST + insn->off), 2222 (u64) SRC); 2223 else 2224 goto default_label; 2225 break; 2226 case BPF_CMPXCHG: 2227 if (BPF_SIZE(insn->code) == BPF_W) 2228 BPF_R0 = (u32) atomic_cmpxchg( 2229 (atomic_t *)(unsigned long) (DST + insn->off), 2230 (u32) BPF_R0, (u32) SRC); 2231 else if (BPF_SIZE(insn->code) == BPF_DW) 2232 BPF_R0 = (u64) atomic64_cmpxchg( 2233 (atomic64_t *)(unsigned long) (DST + insn->off), 2234 (u64) BPF_R0, (u64) SRC); 2235 else 2236 goto default_label; 2237 break; 2238 /* Atomic load and store instructions support all size 2239 * modifiers. 2240 */ 2241 case BPF_LOAD_ACQ: 2242 switch (BPF_SIZE(insn->code)) { 2243 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2244 case BPF_##SIZEOP: \ 2245 DST = (SIZE)smp_load_acquire( \ 2246 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2247 break; 2248 LOAD_ACQUIRE(B, u8) 2249 LOAD_ACQUIRE(H, u16) 2250 LOAD_ACQUIRE(W, u32) 2251 #ifdef CONFIG_64BIT 2252 LOAD_ACQUIRE(DW, u64) 2253 #endif 2254 #undef LOAD_ACQUIRE 2255 default: 2256 goto default_label; 2257 } 2258 break; 2259 case BPF_STORE_REL: 2260 switch (BPF_SIZE(insn->code)) { 2261 #define STORE_RELEASE(SIZEOP, SIZE) \ 2262 case BPF_##SIZEOP: \ 2263 smp_store_release( \ 2264 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2265 break; 2266 STORE_RELEASE(B, u8) 2267 STORE_RELEASE(H, u16) 2268 STORE_RELEASE(W, u32) 2269 #ifdef CONFIG_64BIT 2270 STORE_RELEASE(DW, u64) 2271 #endif 2272 #undef STORE_RELEASE 2273 default: 2274 goto default_label; 2275 } 2276 break; 2277 2278 default: 2279 goto default_label; 2280 } 2281 CONT; 2282 2283 default_label: 2284 /* If we ever reach this, we have a bug somewhere. Die hard here 2285 * instead of just returning 0; we could be somewhere in a subprog, 2286 * so execution could continue otherwise which we do /not/ want. 2287 * 2288 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2289 */ 2290 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2291 insn->code, insn->imm); 2292 BUG_ON(1); 2293 return 0; 2294 } 2295 2296 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2297 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2298 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2299 { \ 2300 u64 stack[stack_size / sizeof(u64)]; \ 2301 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2302 \ 2303 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2304 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2305 ARG1 = (u64) (unsigned long) ctx; \ 2306 return ___bpf_prog_run(regs, insn); \ 2307 } 2308 2309 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2310 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2311 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2312 const struct bpf_insn *insn) \ 2313 { \ 2314 u64 stack[stack_size / sizeof(u64)]; \ 2315 u64 regs[MAX_BPF_EXT_REG]; \ 2316 \ 2317 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2318 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2319 BPF_R1 = r1; \ 2320 BPF_R2 = r2; \ 2321 BPF_R3 = r3; \ 2322 BPF_R4 = r4; \ 2323 BPF_R5 = r5; \ 2324 return ___bpf_prog_run(regs, insn); \ 2325 } 2326 2327 #define EVAL1(FN, X) FN(X) 2328 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2329 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2330 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2331 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2332 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2333 2334 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2335 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2336 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2337 2338 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2339 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2340 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2341 2342 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2343 2344 static unsigned int (*interpreters[])(const void *ctx, 2345 const struct bpf_insn *insn) = { 2346 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2347 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2348 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2349 }; 2350 #undef PROG_NAME_LIST 2351 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2352 static __maybe_unused 2353 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2354 const struct bpf_insn *insn) = { 2355 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2356 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2357 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2358 }; 2359 #undef PROG_NAME_LIST 2360 2361 #ifdef CONFIG_BPF_SYSCALL 2362 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2363 { 2364 stack_depth = max_t(u32, stack_depth, 1); 2365 insn->off = (s16) insn->imm; 2366 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2367 __bpf_call_base_args; 2368 insn->code = BPF_JMP | BPF_CALL_ARGS; 2369 } 2370 #endif 2371 #endif 2372 2373 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2374 const struct bpf_insn *insn) 2375 { 2376 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2377 * is not working properly, so warn about it! 2378 */ 2379 WARN_ON_ONCE(1); 2380 return 0; 2381 } 2382 2383 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2384 const struct bpf_prog *fp) 2385 { 2386 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2387 struct bpf_prog_aux *aux = fp->aux; 2388 enum bpf_cgroup_storage_type i; 2389 bool ret = false; 2390 u64 cookie; 2391 2392 if (fp->kprobe_override) 2393 return ret; 2394 2395 spin_lock(&map->owner_lock); 2396 /* There's no owner yet where we could check for compatibility. */ 2397 if (!map->owner) { 2398 map->owner = bpf_map_owner_alloc(map); 2399 if (!map->owner) 2400 goto err; 2401 map->owner->type = prog_type; 2402 map->owner->jited = fp->jited; 2403 map->owner->xdp_has_frags = aux->xdp_has_frags; 2404 map->owner->sleepable = fp->sleepable; 2405 map->owner->expected_attach_type = fp->expected_attach_type; 2406 map->owner->attach_func_proto = aux->attach_func_proto; 2407 for_each_cgroup_storage_type(i) { 2408 map->owner->storage_cookie[i] = 2409 aux->cgroup_storage[i] ? 2410 aux->cgroup_storage[i]->cookie : 0; 2411 } 2412 ret = true; 2413 } else { 2414 ret = map->owner->type == prog_type && 2415 map->owner->jited == fp->jited && 2416 map->owner->xdp_has_frags == aux->xdp_has_frags && 2417 map->owner->sleepable == fp->sleepable; 2418 if (ret && 2419 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2420 map->owner->expected_attach_type != fp->expected_attach_type) 2421 ret = false; 2422 for_each_cgroup_storage_type(i) { 2423 if (!ret) 2424 break; 2425 cookie = aux->cgroup_storage[i] ? 2426 aux->cgroup_storage[i]->cookie : 0; 2427 ret = map->owner->storage_cookie[i] == cookie || 2428 !cookie; 2429 } 2430 if (ret && 2431 map->owner->attach_func_proto != aux->attach_func_proto) { 2432 switch (prog_type) { 2433 case BPF_PROG_TYPE_TRACING: 2434 case BPF_PROG_TYPE_LSM: 2435 case BPF_PROG_TYPE_EXT: 2436 case BPF_PROG_TYPE_STRUCT_OPS: 2437 ret = false; 2438 break; 2439 default: 2440 break; 2441 } 2442 } 2443 } 2444 err: 2445 spin_unlock(&map->owner_lock); 2446 return ret; 2447 } 2448 2449 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2450 { 2451 /* XDP programs inserted into maps are not guaranteed to run on 2452 * a particular netdev (and can run outside driver context entirely 2453 * in the case of devmap and cpumap). Until device checks 2454 * are implemented, prohibit adding dev-bound programs to program maps. 2455 */ 2456 if (bpf_prog_is_dev_bound(fp->aux)) 2457 return false; 2458 2459 return __bpf_prog_map_compatible(map, fp); 2460 } 2461 2462 static int bpf_check_tail_call(const struct bpf_prog *fp) 2463 { 2464 struct bpf_prog_aux *aux = fp->aux; 2465 int i, ret = 0; 2466 2467 mutex_lock(&aux->used_maps_mutex); 2468 for (i = 0; i < aux->used_map_cnt; i++) { 2469 struct bpf_map *map = aux->used_maps[i]; 2470 2471 if (!map_type_contains_progs(map)) 2472 continue; 2473 2474 if (!__bpf_prog_map_compatible(map, fp)) { 2475 ret = -EINVAL; 2476 goto out; 2477 } 2478 } 2479 2480 out: 2481 mutex_unlock(&aux->used_maps_mutex); 2482 return ret; 2483 } 2484 2485 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2486 { 2487 bool select_interpreter = false; 2488 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2489 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2490 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2491 2492 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2493 * But for non-JITed programs, we don't need bpf_func, so no bounds 2494 * check needed. 2495 */ 2496 if (idx < ARRAY_SIZE(interpreters)) { 2497 fp->bpf_func = interpreters[idx]; 2498 select_interpreter = true; 2499 } else { 2500 fp->bpf_func = __bpf_prog_ret0_warn; 2501 } 2502 #else 2503 fp->bpf_func = __bpf_prog_ret0_warn; 2504 #endif 2505 return select_interpreter; 2506 } 2507 2508 /** 2509 * bpf_prog_select_runtime - select exec runtime for BPF program 2510 * @fp: bpf_prog populated with BPF program 2511 * @err: pointer to error variable 2512 * 2513 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2514 * The BPF program will be executed via bpf_prog_run() function. 2515 * 2516 * Return: the &fp argument along with &err set to 0 for success or 2517 * a negative errno code on failure 2518 */ 2519 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2520 { 2521 /* In case of BPF to BPF calls, verifier did all the prep 2522 * work with regards to JITing, etc. 2523 */ 2524 bool jit_needed = false; 2525 2526 if (fp->bpf_func) 2527 goto finalize; 2528 2529 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2530 bpf_prog_has_kfunc_call(fp)) 2531 jit_needed = true; 2532 2533 if (!bpf_prog_select_interpreter(fp)) 2534 jit_needed = true; 2535 2536 /* eBPF JITs can rewrite the program in case constant 2537 * blinding is active. However, in case of error during 2538 * blinding, bpf_int_jit_compile() must always return a 2539 * valid program, which in this case would simply not 2540 * be JITed, but falls back to the interpreter. 2541 */ 2542 if (!bpf_prog_is_offloaded(fp->aux)) { 2543 *err = bpf_prog_alloc_jited_linfo(fp); 2544 if (*err) 2545 return fp; 2546 2547 fp = bpf_int_jit_compile(fp); 2548 bpf_prog_jit_attempt_done(fp); 2549 if (!fp->jited && jit_needed) { 2550 *err = -ENOTSUPP; 2551 return fp; 2552 } 2553 } else { 2554 *err = bpf_prog_offload_compile(fp); 2555 if (*err) 2556 return fp; 2557 } 2558 2559 finalize: 2560 *err = bpf_prog_lock_ro(fp); 2561 if (*err) 2562 return fp; 2563 2564 /* The tail call compatibility check can only be done at 2565 * this late stage as we need to determine, if we deal 2566 * with JITed or non JITed program concatenations and not 2567 * all eBPF JITs might immediately support all features. 2568 */ 2569 *err = bpf_check_tail_call(fp); 2570 2571 return fp; 2572 } 2573 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2574 2575 static unsigned int __bpf_prog_ret1(const void *ctx, 2576 const struct bpf_insn *insn) 2577 { 2578 return 1; 2579 } 2580 2581 static struct bpf_prog_dummy { 2582 struct bpf_prog prog; 2583 } dummy_bpf_prog = { 2584 .prog = { 2585 .bpf_func = __bpf_prog_ret1, 2586 }, 2587 }; 2588 2589 struct bpf_empty_prog_array bpf_empty_prog_array = { 2590 .null_prog = NULL, 2591 }; 2592 EXPORT_SYMBOL(bpf_empty_prog_array); 2593 2594 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2595 { 2596 struct bpf_prog_array *p; 2597 2598 if (prog_cnt) 2599 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags); 2600 else 2601 p = &bpf_empty_prog_array.hdr; 2602 2603 return p; 2604 } 2605 2606 void bpf_prog_array_free(struct bpf_prog_array *progs) 2607 { 2608 if (!progs || progs == &bpf_empty_prog_array.hdr) 2609 return; 2610 kfree_rcu(progs, rcu); 2611 } 2612 2613 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2614 { 2615 struct bpf_prog_array *progs; 2616 2617 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2618 * no need to call kfree_rcu(), just call kfree() directly. 2619 */ 2620 progs = container_of(rcu, struct bpf_prog_array, rcu); 2621 if (rcu_trace_implies_rcu_gp()) 2622 kfree(progs); 2623 else 2624 kfree_rcu(progs, rcu); 2625 } 2626 2627 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2628 { 2629 if (!progs || progs == &bpf_empty_prog_array.hdr) 2630 return; 2631 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2632 } 2633 2634 int bpf_prog_array_length(struct bpf_prog_array *array) 2635 { 2636 struct bpf_prog_array_item *item; 2637 u32 cnt = 0; 2638 2639 for (item = array->items; item->prog; item++) 2640 if (item->prog != &dummy_bpf_prog.prog) 2641 cnt++; 2642 return cnt; 2643 } 2644 2645 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2646 { 2647 struct bpf_prog_array_item *item; 2648 2649 for (item = array->items; item->prog; item++) 2650 if (item->prog != &dummy_bpf_prog.prog) 2651 return false; 2652 return true; 2653 } 2654 2655 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2656 u32 *prog_ids, 2657 u32 request_cnt) 2658 { 2659 struct bpf_prog_array_item *item; 2660 int i = 0; 2661 2662 for (item = array->items; item->prog; item++) { 2663 if (item->prog == &dummy_bpf_prog.prog) 2664 continue; 2665 prog_ids[i] = item->prog->aux->id; 2666 if (++i == request_cnt) { 2667 item++; 2668 break; 2669 } 2670 } 2671 2672 return !!(item->prog); 2673 } 2674 2675 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2676 __u32 __user *prog_ids, u32 cnt) 2677 { 2678 unsigned long err = 0; 2679 bool nospc; 2680 u32 *ids; 2681 2682 /* users of this function are doing: 2683 * cnt = bpf_prog_array_length(); 2684 * if (cnt > 0) 2685 * bpf_prog_array_copy_to_user(..., cnt); 2686 * so below kcalloc doesn't need extra cnt > 0 check. 2687 */ 2688 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2689 if (!ids) 2690 return -ENOMEM; 2691 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2692 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2693 kfree(ids); 2694 if (err) 2695 return -EFAULT; 2696 if (nospc) 2697 return -ENOSPC; 2698 return 0; 2699 } 2700 2701 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2702 struct bpf_prog *old_prog) 2703 { 2704 struct bpf_prog_array_item *item; 2705 2706 for (item = array->items; item->prog; item++) 2707 if (item->prog == old_prog) { 2708 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2709 break; 2710 } 2711 } 2712 2713 /** 2714 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2715 * index into the program array with 2716 * a dummy no-op program. 2717 * @array: a bpf_prog_array 2718 * @index: the index of the program to replace 2719 * 2720 * Skips over dummy programs, by not counting them, when calculating 2721 * the position of the program to replace. 2722 * 2723 * Return: 2724 * * 0 - Success 2725 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2726 * * -ENOENT - Index out of range 2727 */ 2728 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2729 { 2730 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2731 } 2732 2733 /** 2734 * bpf_prog_array_update_at() - Updates the program at the given index 2735 * into the program array. 2736 * @array: a bpf_prog_array 2737 * @index: the index of the program to update 2738 * @prog: the program to insert into the array 2739 * 2740 * Skips over dummy programs, by not counting them, when calculating 2741 * the position of the program to update. 2742 * 2743 * Return: 2744 * * 0 - Success 2745 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2746 * * -ENOENT - Index out of range 2747 */ 2748 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2749 struct bpf_prog *prog) 2750 { 2751 struct bpf_prog_array_item *item; 2752 2753 if (unlikely(index < 0)) 2754 return -EINVAL; 2755 2756 for (item = array->items; item->prog; item++) { 2757 if (item->prog == &dummy_bpf_prog.prog) 2758 continue; 2759 if (!index) { 2760 WRITE_ONCE(item->prog, prog); 2761 return 0; 2762 } 2763 index--; 2764 } 2765 return -ENOENT; 2766 } 2767 2768 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2769 struct bpf_prog *exclude_prog, 2770 struct bpf_prog *include_prog, 2771 u64 bpf_cookie, 2772 struct bpf_prog_array **new_array) 2773 { 2774 int new_prog_cnt, carry_prog_cnt = 0; 2775 struct bpf_prog_array_item *existing, *new; 2776 struct bpf_prog_array *array; 2777 bool found_exclude = false; 2778 2779 /* Figure out how many existing progs we need to carry over to 2780 * the new array. 2781 */ 2782 if (old_array) { 2783 existing = old_array->items; 2784 for (; existing->prog; existing++) { 2785 if (existing->prog == exclude_prog) { 2786 found_exclude = true; 2787 continue; 2788 } 2789 if (existing->prog != &dummy_bpf_prog.prog) 2790 carry_prog_cnt++; 2791 if (existing->prog == include_prog) 2792 return -EEXIST; 2793 } 2794 } 2795 2796 if (exclude_prog && !found_exclude) 2797 return -ENOENT; 2798 2799 /* How many progs (not NULL) will be in the new array? */ 2800 new_prog_cnt = carry_prog_cnt; 2801 if (include_prog) 2802 new_prog_cnt += 1; 2803 2804 /* Do we have any prog (not NULL) in the new array? */ 2805 if (!new_prog_cnt) { 2806 *new_array = NULL; 2807 return 0; 2808 } 2809 2810 /* +1 as the end of prog_array is marked with NULL */ 2811 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2812 if (!array) 2813 return -ENOMEM; 2814 new = array->items; 2815 2816 /* Fill in the new prog array */ 2817 if (carry_prog_cnt) { 2818 existing = old_array->items; 2819 for (; existing->prog; existing++) { 2820 if (existing->prog == exclude_prog || 2821 existing->prog == &dummy_bpf_prog.prog) 2822 continue; 2823 2824 new->prog = existing->prog; 2825 new->bpf_cookie = existing->bpf_cookie; 2826 new++; 2827 } 2828 } 2829 if (include_prog) { 2830 new->prog = include_prog; 2831 new->bpf_cookie = bpf_cookie; 2832 new++; 2833 } 2834 new->prog = NULL; 2835 *new_array = array; 2836 return 0; 2837 } 2838 2839 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2840 u32 *prog_ids, u32 request_cnt, 2841 u32 *prog_cnt) 2842 { 2843 u32 cnt = 0; 2844 2845 if (array) 2846 cnt = bpf_prog_array_length(array); 2847 2848 *prog_cnt = cnt; 2849 2850 /* return early if user requested only program count or nothing to copy */ 2851 if (!request_cnt || !cnt) 2852 return 0; 2853 2854 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2855 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2856 : 0; 2857 } 2858 2859 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2860 struct bpf_map **used_maps, u32 len) 2861 { 2862 struct bpf_map *map; 2863 bool sleepable; 2864 u32 i; 2865 2866 sleepable = aux->prog->sleepable; 2867 for (i = 0; i < len; i++) { 2868 map = used_maps[i]; 2869 if (map->ops->map_poke_untrack) 2870 map->ops->map_poke_untrack(map, aux); 2871 if (sleepable) 2872 atomic64_dec(&map->sleepable_refcnt); 2873 bpf_map_put(map); 2874 } 2875 } 2876 2877 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2878 { 2879 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2880 kfree(aux->used_maps); 2881 } 2882 2883 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 2884 { 2885 #ifdef CONFIG_BPF_SYSCALL 2886 struct btf_mod_pair *btf_mod; 2887 u32 i; 2888 2889 for (i = 0; i < len; i++) { 2890 btf_mod = &used_btfs[i]; 2891 if (btf_mod->module) 2892 module_put(btf_mod->module); 2893 btf_put(btf_mod->btf); 2894 } 2895 #endif 2896 } 2897 2898 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2899 { 2900 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 2901 kfree(aux->used_btfs); 2902 } 2903 2904 static void bpf_prog_free_deferred(struct work_struct *work) 2905 { 2906 struct bpf_prog_aux *aux; 2907 int i; 2908 2909 aux = container_of(work, struct bpf_prog_aux, work); 2910 #ifdef CONFIG_BPF_SYSCALL 2911 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2912 bpf_prog_stream_free(aux->prog); 2913 #endif 2914 #ifdef CONFIG_CGROUP_BPF 2915 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2916 bpf_cgroup_atype_put(aux->cgroup_atype); 2917 #endif 2918 bpf_free_used_maps(aux); 2919 bpf_free_used_btfs(aux); 2920 bpf_prog_disassoc_struct_ops(aux->prog); 2921 if (bpf_prog_is_dev_bound(aux)) 2922 bpf_prog_dev_bound_destroy(aux->prog); 2923 #ifdef CONFIG_PERF_EVENTS 2924 if (aux->prog->has_callchain_buf) 2925 put_callchain_buffers(); 2926 #endif 2927 if (aux->dst_trampoline) 2928 bpf_trampoline_put(aux->dst_trampoline); 2929 for (i = 0; i < aux->real_func_cnt; i++) { 2930 /* We can just unlink the subprog poke descriptor table as 2931 * it was originally linked to the main program and is also 2932 * released along with it. 2933 */ 2934 aux->func[i]->aux->poke_tab = NULL; 2935 bpf_jit_free(aux->func[i]); 2936 } 2937 if (aux->real_func_cnt) { 2938 kfree(aux->func); 2939 bpf_prog_unlock_free(aux->prog); 2940 } else { 2941 bpf_jit_free(aux->prog); 2942 } 2943 } 2944 2945 void bpf_prog_free(struct bpf_prog *fp) 2946 { 2947 struct bpf_prog_aux *aux = fp->aux; 2948 2949 if (aux->dst_prog) 2950 bpf_prog_put(aux->dst_prog); 2951 bpf_token_put(aux->token); 2952 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2953 schedule_work(&aux->work); 2954 } 2955 EXPORT_SYMBOL_GPL(bpf_prog_free); 2956 2957 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 2958 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2959 2960 void bpf_user_rnd_init_once(void) 2961 { 2962 prandom_init_once(&bpf_user_rnd_state); 2963 } 2964 2965 BPF_CALL_0(bpf_user_rnd_u32) 2966 { 2967 /* Should someone ever have the rather unwise idea to use some 2968 * of the registers passed into this function, then note that 2969 * this function is called from native eBPF and classic-to-eBPF 2970 * transformations. Register assignments from both sides are 2971 * different, f.e. classic always sets fn(ctx, A, X) here. 2972 */ 2973 struct rnd_state *state; 2974 u32 res; 2975 2976 state = &get_cpu_var(bpf_user_rnd_state); 2977 res = prandom_u32_state(state); 2978 put_cpu_var(bpf_user_rnd_state); 2979 2980 return res; 2981 } 2982 2983 BPF_CALL_0(bpf_get_raw_cpu_id) 2984 { 2985 return raw_smp_processor_id(); 2986 } 2987 2988 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 2989 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 2990 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 2991 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 2992 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 2993 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 2994 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 2995 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 2996 const struct bpf_func_proto bpf_spin_lock_proto __weak; 2997 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 2998 const struct bpf_func_proto bpf_jiffies64_proto __weak; 2999 3000 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3001 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3002 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3003 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3004 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3005 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3006 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3007 3008 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3009 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3010 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3011 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3012 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3013 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3014 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3015 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3016 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3017 const struct bpf_func_proto bpf_set_retval_proto __weak; 3018 const struct bpf_func_proto bpf_get_retval_proto __weak; 3019 3020 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3021 { 3022 return NULL; 3023 } 3024 3025 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3026 { 3027 return NULL; 3028 } 3029 3030 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3031 { 3032 return NULL; 3033 } 3034 3035 u64 __weak 3036 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3037 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3038 { 3039 return -ENOTSUPP; 3040 } 3041 EXPORT_SYMBOL_GPL(bpf_event_output); 3042 3043 /* Always built-in helper functions. */ 3044 const struct bpf_func_proto bpf_tail_call_proto = { 3045 /* func is unused for tail_call, we set it to pass the 3046 * get_helper_proto check 3047 */ 3048 .func = BPF_PTR_POISON, 3049 .gpl_only = false, 3050 .ret_type = RET_VOID, 3051 .arg1_type = ARG_PTR_TO_CTX, 3052 .arg2_type = ARG_CONST_MAP_PTR, 3053 .arg3_type = ARG_ANYTHING, 3054 }; 3055 3056 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3057 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3058 * eBPF and implicitly also cBPF can get JITed! 3059 */ 3060 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 3061 { 3062 return prog; 3063 } 3064 3065 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3066 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3067 */ 3068 void __weak bpf_jit_compile(struct bpf_prog *prog) 3069 { 3070 } 3071 3072 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3073 { 3074 return false; 3075 } 3076 3077 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3078 * analysis code and wants explicit zero extension inserted by verifier. 3079 * Otherwise, return FALSE. 3080 * 3081 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3082 * you don't override this. JITs that don't want these extra insns can detect 3083 * them using insn_is_zext. 3084 */ 3085 bool __weak bpf_jit_needs_zext(void) 3086 { 3087 return false; 3088 } 3089 3090 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3091 * all archs. The value returned must not change at runtime as there is 3092 * currently no support for reloading programs that were loaded without 3093 * mitigations. 3094 */ 3095 bool __weak bpf_jit_bypass_spec_v1(void) 3096 { 3097 return false; 3098 } 3099 3100 bool __weak bpf_jit_bypass_spec_v4(void) 3101 { 3102 return false; 3103 } 3104 3105 /* Return true if the JIT inlines the call to the helper corresponding to 3106 * the imm. 3107 * 3108 * The verifier will not patch the insn->imm for the call to the helper if 3109 * this returns true. 3110 */ 3111 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3112 { 3113 return false; 3114 } 3115 3116 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3117 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3118 { 3119 return false; 3120 } 3121 3122 bool __weak bpf_jit_supports_percpu_insn(void) 3123 { 3124 return false; 3125 } 3126 3127 bool __weak bpf_jit_supports_kfunc_call(void) 3128 { 3129 return false; 3130 } 3131 3132 bool __weak bpf_jit_supports_far_kfunc_call(void) 3133 { 3134 return false; 3135 } 3136 3137 bool __weak bpf_jit_supports_arena(void) 3138 { 3139 return false; 3140 } 3141 3142 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3143 { 3144 return false; 3145 } 3146 3147 bool __weak bpf_jit_supports_fsession(void) 3148 { 3149 return false; 3150 } 3151 3152 u64 __weak bpf_arch_uaddress_limit(void) 3153 { 3154 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3155 return TASK_SIZE; 3156 #else 3157 return 0; 3158 #endif 3159 } 3160 3161 /* Return TRUE if the JIT backend satisfies the following two conditions: 3162 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3163 * 2) Under the specific arch, the implementation of xchg() is the same 3164 * as atomic_xchg() on pointer-sized words. 3165 */ 3166 bool __weak bpf_jit_supports_ptr_xchg(void) 3167 { 3168 return false; 3169 } 3170 3171 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3172 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3173 */ 3174 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3175 int len) 3176 { 3177 return -EFAULT; 3178 } 3179 3180 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3181 enum bpf_text_poke_type new_t, void *old_addr, 3182 void *new_addr) 3183 { 3184 return -ENOTSUPP; 3185 } 3186 3187 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3188 { 3189 return ERR_PTR(-ENOTSUPP); 3190 } 3191 3192 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3193 { 3194 return -ENOTSUPP; 3195 } 3196 3197 bool __weak bpf_jit_supports_exceptions(void) 3198 { 3199 return false; 3200 } 3201 3202 bool __weak bpf_jit_supports_private_stack(void) 3203 { 3204 return false; 3205 } 3206 3207 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3208 { 3209 } 3210 3211 bool __weak bpf_jit_supports_timed_may_goto(void) 3212 { 3213 return false; 3214 } 3215 3216 u64 __weak arch_bpf_timed_may_goto(void) 3217 { 3218 return 0; 3219 } 3220 3221 static noinline void bpf_prog_report_may_goto_violation(void) 3222 { 3223 #ifdef CONFIG_BPF_SYSCALL 3224 struct bpf_stream_stage ss; 3225 struct bpf_prog *prog; 3226 3227 prog = bpf_prog_find_from_stack(); 3228 if (!prog) 3229 return; 3230 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3231 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3232 bpf_stream_dump_stack(ss); 3233 })); 3234 #endif 3235 } 3236 3237 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3238 { 3239 u64 time = ktime_get_mono_fast_ns(); 3240 3241 /* Populate the timestamp for this stack frame, and refresh count. */ 3242 if (!p->timestamp) { 3243 p->timestamp = time; 3244 return BPF_MAX_TIMED_LOOPS; 3245 } 3246 /* Check if we've exhausted our time slice, and zero count. */ 3247 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3248 bpf_prog_report_may_goto_violation(); 3249 return 0; 3250 } 3251 /* Refresh the count for the stack frame. */ 3252 return BPF_MAX_TIMED_LOOPS; 3253 } 3254 3255 /* for configs without MMU or 32-bit */ 3256 __weak const struct bpf_map_ops arena_map_ops; 3257 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3258 { 3259 return 0; 3260 } 3261 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3262 { 3263 return 0; 3264 } 3265 3266 #ifdef CONFIG_BPF_SYSCALL 3267 static int __init bpf_global_ma_init(void) 3268 { 3269 int ret; 3270 3271 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3272 bpf_global_ma_set = !ret; 3273 return ret; 3274 } 3275 late_initcall(bpf_global_ma_init); 3276 #endif 3277 3278 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3279 EXPORT_SYMBOL(bpf_stats_enabled_key); 3280 3281 /* All definitions of tracepoints related to BPF. */ 3282 #define CREATE_TRACE_POINTS 3283 #include <linux/bpf_trace.h> 3284 3285 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3286 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3287 3288 #ifdef CONFIG_BPF_SYSCALL 3289 3290 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3291 const char **linep, int *nump) 3292 { 3293 int idx = -1, insn_start, insn_end, len; 3294 struct bpf_line_info *linfo; 3295 void **jited_linfo; 3296 struct btf *btf; 3297 int nr_linfo; 3298 3299 btf = prog->aux->btf; 3300 linfo = prog->aux->linfo; 3301 jited_linfo = prog->aux->jited_linfo; 3302 3303 if (!btf || !linfo || !jited_linfo) 3304 return -EINVAL; 3305 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3306 3307 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3308 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3309 3310 insn_start = linfo[0].insn_off; 3311 insn_end = insn_start + len; 3312 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3313 3314 for (int i = 0; i < nr_linfo && 3315 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3316 if (jited_linfo[i] >= (void *)ip) 3317 break; 3318 idx = i; 3319 } 3320 3321 if (idx == -1) 3322 return -ENOENT; 3323 3324 /* Get base component of the file path. */ 3325 *filep = btf_name_by_offset(btf, linfo[idx].file_name_off); 3326 *filep = kbasename(*filep); 3327 /* Obtain the source line, and strip whitespace in prefix. */ 3328 *linep = btf_name_by_offset(btf, linfo[idx].line_off); 3329 while (isspace(**linep)) 3330 *linep += 1; 3331 *nump = BPF_LINE_INFO_LINE_NUM(linfo[idx].line_col); 3332 return 0; 3333 } 3334 3335 struct walk_stack_ctx { 3336 struct bpf_prog *prog; 3337 }; 3338 3339 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3340 { 3341 struct walk_stack_ctx *ctxp = cookie; 3342 struct bpf_prog *prog; 3343 3344 /* 3345 * The RCU read lock is held to safely traverse the latch tree, but we 3346 * don't need its protection when accessing the prog, since it has an 3347 * active stack frame on the current stack trace, and won't disappear. 3348 */ 3349 rcu_read_lock(); 3350 prog = bpf_prog_ksym_find(ip); 3351 rcu_read_unlock(); 3352 if (!prog) 3353 return true; 3354 /* Make sure we return the main prog if we found a subprog */ 3355 ctxp->prog = prog->aux->main_prog_aux->prog; 3356 return false; 3357 } 3358 3359 struct bpf_prog *bpf_prog_find_from_stack(void) 3360 { 3361 struct walk_stack_ctx ctx = {}; 3362 3363 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3364 return ctx.prog; 3365 } 3366 3367 #endif 3368