1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Linux Socket Filter - Kernel level socket filtering 4 * 5 * Based on the design of the Berkeley Packet Filter. The new 6 * internal format has been designed by PLUMgrid: 7 * 8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 9 * 10 * Authors: 11 * 12 * Jay Schulist <jschlst@samba.org> 13 * Alexei Starovoitov <ast@plumgrid.com> 14 * Daniel Borkmann <dborkman@redhat.com> 15 * 16 * Andi Kleen - Fix a few bad bugs and races. 17 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 18 */ 19 20 #include <uapi/linux/btf.h> 21 #include <linux/filter.h> 22 #include <linux/skbuff.h> 23 #include <linux/static_call.h> 24 #include <linux/vmalloc.h> 25 #include <linux/prandom.h> 26 #include <linux/bpf.h> 27 #include <linux/btf.h> 28 #include <linux/hex.h> 29 #include <linux/objtool.h> 30 #include <linux/overflow.h> 31 #include <linux/rbtree_latch.h> 32 #include <linux/kallsyms.h> 33 #include <linux/rcupdate.h> 34 #include <linux/perf_event.h> 35 #include <linux/extable.h> 36 #include <linux/log2.h> 37 #include <linux/bpf_verifier.h> 38 #include <linux/nodemask.h> 39 #include <linux/nospec.h> 40 #include <linux/bpf_mem_alloc.h> 41 #include <linux/memcontrol.h> 42 #include <linux/execmem.h> 43 #include <crypto/sha2.h> 44 45 #include <asm/barrier.h> 46 #include <linux/unaligned.h> 47 48 /* Registers */ 49 #define BPF_R0 regs[BPF_REG_0] 50 #define BPF_R1 regs[BPF_REG_1] 51 #define BPF_R2 regs[BPF_REG_2] 52 #define BPF_R3 regs[BPF_REG_3] 53 #define BPF_R4 regs[BPF_REG_4] 54 #define BPF_R5 regs[BPF_REG_5] 55 #define BPF_R6 regs[BPF_REG_6] 56 #define BPF_R7 regs[BPF_REG_7] 57 #define BPF_R8 regs[BPF_REG_8] 58 #define BPF_R9 regs[BPF_REG_9] 59 #define BPF_R10 regs[BPF_REG_10] 60 61 /* Named registers */ 62 #define DST regs[insn->dst_reg] 63 #define SRC regs[insn->src_reg] 64 #define FP regs[BPF_REG_FP] 65 #define AX regs[BPF_REG_AX] 66 #define ARG1 regs[BPF_REG_ARG1] 67 #define CTX regs[BPF_REG_CTX] 68 #define OFF insn->off 69 #define IMM insn->imm 70 71 struct bpf_mem_alloc bpf_global_ma; 72 bool bpf_global_ma_set; 73 74 /* No hurry in this branch 75 * 76 * Exported for the bpf jit load helper. 77 */ 78 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 79 { 80 u8 *ptr = NULL; 81 82 if (k >= SKF_NET_OFF) { 83 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 84 } else if (k >= SKF_LL_OFF) { 85 if (unlikely(!skb_mac_header_was_set(skb))) 86 return NULL; 87 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 88 } 89 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 90 return ptr; 91 92 return NULL; 93 } 94 95 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 96 enum page_size_enum { 97 __PAGE_SIZE = PAGE_SIZE 98 }; 99 100 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 101 { 102 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 103 struct bpf_prog_aux *aux; 104 struct bpf_prog *fp; 105 106 size = round_up(size, __PAGE_SIZE); 107 fp = __vmalloc(size, gfp_flags); 108 if (fp == NULL) 109 return NULL; 110 111 aux = kzalloc_obj(*aux, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 112 if (aux == NULL) { 113 vfree(fp); 114 return NULL; 115 } 116 fp->active = __alloc_percpu_gfp(sizeof(u8[BPF_NR_CONTEXTS]), 4, 117 bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 118 if (!fp->active) { 119 vfree(fp); 120 kfree(aux); 121 return NULL; 122 } 123 124 fp->pages = size / PAGE_SIZE; 125 fp->aux = aux; 126 fp->aux->main_prog_aux = aux; 127 fp->aux->prog = fp; 128 fp->jit_requested = ebpf_jit_enabled(); 129 fp->jit_required = IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON); 130 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 131 #ifdef CONFIG_CGROUP_BPF 132 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 133 #endif 134 135 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 136 #ifdef CONFIG_FINEIBT 137 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 138 #endif 139 mutex_init(&fp->aux->used_maps_mutex); 140 mutex_init(&fp->aux->ext_mutex); 141 mutex_init(&fp->aux->dst_mutex); 142 mutex_init(&fp->aux->st_ops_assoc_mutex); 143 144 #ifdef CONFIG_BPF_SYSCALL 145 bpf_prog_stream_init(fp); 146 #endif 147 148 return fp; 149 } 150 151 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 152 { 153 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 154 struct bpf_prog *prog; 155 int cpu; 156 157 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 158 if (!prog) 159 return NULL; 160 161 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 162 if (!prog->stats) { 163 free_percpu(prog->active); 164 kfree(prog->aux); 165 vfree(prog); 166 return NULL; 167 } 168 169 for_each_possible_cpu(cpu) { 170 struct bpf_prog_stats *pstats; 171 172 pstats = per_cpu_ptr(prog->stats, cpu); 173 u64_stats_init(&pstats->syncp); 174 } 175 return prog; 176 } 177 EXPORT_SYMBOL_GPL(bpf_prog_alloc); 178 179 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 180 { 181 if (!prog->aux->nr_linfo || !prog->jit_requested) 182 return 0; 183 184 prog->aux->jited_linfo = kvzalloc_objs(*prog->aux->jited_linfo, 185 prog->aux->nr_linfo, 186 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 187 if (!prog->aux->jited_linfo) 188 return -ENOMEM; 189 190 return 0; 191 } 192 193 void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 194 { 195 if (prog->aux->jited_linfo && 196 (!prog->jited || !prog->aux->jited_linfo[0])) { 197 kvfree(prog->aux->jited_linfo); 198 prog->aux->jited_linfo = NULL; 199 } 200 201 kfree(prog->aux->kfunc_tab); 202 prog->aux->kfunc_tab = NULL; 203 } 204 205 /* The jit engine is responsible to provide an array 206 * for insn_off to the jited_off mapping (insn_to_jit_off). 207 * 208 * The idx to this array is the insn_off. Hence, the insn_off 209 * here is relative to the prog itself instead of the main prog. 210 * This array has one entry for each xlated bpf insn. 211 * 212 * jited_off is the byte off to the end of the jited insn. 213 * 214 * Hence, with 215 * insn_start: 216 * The first bpf insn off of the prog. The insn off 217 * here is relative to the main prog. 218 * e.g. if prog is a subprog, insn_start > 0 219 * linfo_idx: 220 * The prog's idx to prog->aux->linfo and jited_linfo 221 * 222 * jited_linfo[linfo_idx] = prog->bpf_func 223 * 224 * For i > linfo_idx, 225 * 226 * jited_linfo[i] = prog->bpf_func + 227 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 228 */ 229 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 230 const u32 *insn_to_jit_off) 231 { 232 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 233 const struct bpf_line_info *linfo; 234 void **jited_linfo; 235 236 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 237 /* Userspace did not provide linfo */ 238 return; 239 240 linfo_idx = prog->aux->linfo_idx; 241 linfo = &prog->aux->linfo[linfo_idx]; 242 insn_start = linfo[0].insn_off; 243 insn_end = insn_start + prog->len; 244 245 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 246 jited_linfo[0] = prog->bpf_func; 247 248 nr_linfo = prog->aux->nr_linfo - linfo_idx; 249 250 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 251 /* The verifier ensures that linfo[i].insn_off is 252 * strictly increasing 253 */ 254 jited_linfo[i] = prog->bpf_func + 255 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 256 } 257 258 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 259 gfp_t gfp_extra_flags) 260 { 261 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 262 struct bpf_prog *fp; 263 u32 pages; 264 265 size = round_up(size, PAGE_SIZE); 266 pages = size / PAGE_SIZE; 267 if (pages <= fp_old->pages) 268 return fp_old; 269 270 fp = __vmalloc(size, gfp_flags); 271 if (fp) { 272 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 273 fp->pages = pages; 274 fp->aux->prog = fp; 275 276 /* We keep fp->aux from fp_old around in the new 277 * reallocated structure. 278 */ 279 fp_old->aux = NULL; 280 fp_old->stats = NULL; 281 fp_old->active = NULL; 282 __bpf_prog_free(fp_old); 283 } 284 285 return fp; 286 } 287 288 void __bpf_prog_free(struct bpf_prog *fp) 289 { 290 if (fp->aux) { 291 mutex_destroy(&fp->aux->used_maps_mutex); 292 mutex_destroy(&fp->aux->dst_mutex); 293 mutex_destroy(&fp->aux->st_ops_assoc_mutex); 294 kfree(fp->aux->poke_tab); 295 kfree(fp->aux); 296 } 297 free_percpu(fp->stats); 298 free_percpu(fp->active); 299 vfree(fp); 300 } 301 302 int bpf_prog_calc_tag(struct bpf_prog *fp) 303 { 304 size_t size = bpf_prog_insn_size(fp); 305 struct bpf_insn *dst; 306 bool was_ld_map; 307 u32 i; 308 309 dst = __vmalloc(size, GFP_KERNEL_ACCOUNT); 310 if (!dst) 311 return -ENOMEM; 312 313 /* We need to take out the map fd for the digest calculation 314 * since they are unstable from user space side. 315 */ 316 for (i = 0, was_ld_map = false; i < fp->len; i++) { 317 dst[i] = fp->insnsi[i]; 318 if (!was_ld_map && 319 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 320 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 321 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 322 was_ld_map = true; 323 dst[i].imm = 0; 324 } else if (was_ld_map && 325 dst[i].code == 0 && 326 dst[i].dst_reg == 0 && 327 dst[i].src_reg == 0 && 328 dst[i].off == 0) { 329 was_ld_map = false; 330 dst[i].imm = 0; 331 } else { 332 was_ld_map = false; 333 } 334 } 335 sha256((u8 *)dst, size, fp->digest); 336 vfree(dst); 337 return 0; 338 } 339 340 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 341 s32 end_new, s32 curr, const bool probe_pass) 342 { 343 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 344 s32 delta = end_new - end_old; 345 s64 imm = insn->imm; 346 347 if (curr < pos && curr + imm + 1 >= end_old) 348 imm += delta; 349 else if (curr >= end_new && curr + imm + 1 < end_new) 350 imm -= delta; 351 if (imm < imm_min || imm > imm_max) 352 return -ERANGE; 353 if (!probe_pass) 354 insn->imm = imm; 355 return 0; 356 } 357 358 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 359 s32 end_new, s32 curr, const bool probe_pass) 360 { 361 s64 off_min, off_max, off; 362 s32 delta = end_new - end_old; 363 364 if (insn->code == (BPF_JMP32 | BPF_JA)) { 365 off = insn->imm; 366 off_min = S32_MIN; 367 off_max = S32_MAX; 368 } else { 369 off = insn->off; 370 off_min = S16_MIN; 371 off_max = S16_MAX; 372 } 373 374 if (curr < pos && curr + off + 1 >= end_old) 375 off += delta; 376 else if (curr >= end_new && curr + off + 1 < end_new) 377 off -= delta; 378 if (off < off_min || off > off_max) 379 return -ERANGE; 380 if (!probe_pass) { 381 if (insn->code == (BPF_JMP32 | BPF_JA)) 382 insn->imm = off; 383 else 384 insn->off = off; 385 } 386 return 0; 387 } 388 389 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 390 s32 end_new, const bool probe_pass) 391 { 392 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 393 struct bpf_insn *insn = prog->insnsi; 394 int ret = 0; 395 396 for (i = 0; i < insn_cnt; i++, insn++) { 397 u8 code; 398 399 /* In the probing pass we still operate on the original, 400 * unpatched image in order to check overflows before we 401 * do any other adjustments. Therefore skip the patchlet. 402 */ 403 if (probe_pass && i == pos) { 404 i = end_new; 405 insn = prog->insnsi + end_old; 406 } 407 if (bpf_pseudo_func(insn)) { 408 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 409 end_new, i, probe_pass); 410 if (ret) 411 return ret; 412 continue; 413 } 414 code = insn->code; 415 if ((BPF_CLASS(code) != BPF_JMP && 416 BPF_CLASS(code) != BPF_JMP32) || 417 BPF_OP(code) == BPF_EXIT) 418 continue; 419 /* Adjust offset of jmps if we cross patch boundaries. */ 420 if (BPF_OP(code) == BPF_CALL) { 421 if (insn->src_reg != BPF_PSEUDO_CALL) 422 continue; 423 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 424 end_new, i, probe_pass); 425 } else { 426 ret = bpf_adj_delta_to_off(insn, pos, end_old, 427 end_new, i, probe_pass); 428 } 429 if (ret) 430 break; 431 } 432 433 return ret; 434 } 435 436 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 437 { 438 struct bpf_line_info *linfo; 439 u32 i, nr_linfo; 440 441 nr_linfo = prog->aux->nr_linfo; 442 if (!nr_linfo || !delta) 443 return; 444 445 linfo = prog->aux->linfo; 446 447 for (i = 0; i < nr_linfo; i++) 448 if (off < linfo[i].insn_off) 449 break; 450 451 /* Push all off < linfo[i].insn_off by delta */ 452 for (; i < nr_linfo; i++) 453 linfo[i].insn_off += delta; 454 } 455 456 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 457 const struct bpf_insn *patch, u32 len) 458 { 459 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 460 const u32 cnt_max = S16_MAX; 461 struct bpf_prog *prog_adj; 462 int err; 463 464 /* Since our patchlet doesn't expand the image, we're done. */ 465 if (insn_delta == 0) { 466 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 467 return prog; 468 } 469 470 insn_adj_cnt = prog->len + insn_delta; 471 472 /* Reject anything that would potentially let the insn->off 473 * target overflow when we have excessive program expansions. 474 * We need to probe here before we do any reallocation where 475 * we afterwards may not fail anymore. 476 */ 477 if (insn_adj_cnt > cnt_max && 478 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 479 return ERR_PTR(err); 480 481 /* Several new instructions need to be inserted. Make room 482 * for them. Likely, there's no need for a new allocation as 483 * last page could have large enough tailroom. 484 */ 485 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 486 GFP_USER); 487 if (!prog_adj) 488 return ERR_PTR(-ENOMEM); 489 490 prog_adj->len = insn_adj_cnt; 491 492 /* Patching happens in 3 steps: 493 * 494 * 1) Move over tail of insnsi from next instruction onwards, 495 * so we can patch the single target insn with one or more 496 * new ones (patching is always from 1 to n insns, n > 0). 497 * 2) Inject new instructions at the target location. 498 * 3) Adjust branch offsets if necessary. 499 */ 500 insn_rest = insn_adj_cnt - off - len; 501 502 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 503 sizeof(*patch) * insn_rest); 504 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 505 506 /* We are guaranteed to not fail at this point, otherwise 507 * the ship has sailed to reverse to the original state. An 508 * overflow cannot happen at this point. 509 */ 510 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 511 512 bpf_adj_linfo(prog_adj, off, insn_delta); 513 514 return prog_adj; 515 } 516 517 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 518 { 519 int err; 520 521 /* Branch offsets can't overflow when program is shrinking, no need 522 * to call bpf_adj_branches(..., true) here 523 */ 524 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 525 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 526 prog->len -= cnt; 527 528 err = bpf_adj_branches(prog, off, off + cnt, off, false); 529 WARN_ON_ONCE(err); 530 return err; 531 } 532 533 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 534 { 535 int i; 536 537 for (i = 0; i < fp->aux->real_func_cnt; i++) 538 bpf_prog_kallsyms_del(fp->aux->func[i]); 539 } 540 541 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 542 { 543 bpf_prog_kallsyms_del_subprogs(fp); 544 bpf_prog_kallsyms_del(fp); 545 } 546 547 #ifdef CONFIG_BPF_JIT 548 /* All BPF JIT sysctl knobs here. */ 549 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 550 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 551 int bpf_jit_harden __read_mostly; 552 long bpf_jit_limit __read_mostly; 553 long bpf_jit_limit_max __read_mostly; 554 555 static void 556 bpf_prog_ksym_set_addr(struct bpf_prog *prog) 557 { 558 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 559 560 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 561 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 562 } 563 564 static void 565 bpf_prog_ksym_set_name(struct bpf_prog *prog) 566 { 567 char *sym = prog->aux->ksym.name; 568 const char *end = sym + KSYM_NAME_LEN; 569 const struct btf_type *type; 570 const char *func_name; 571 572 BUILD_BUG_ON(sizeof("bpf_prog_") + 573 sizeof(prog->tag) * 2 + 574 /* name has been null terminated. 575 * We should need +1 for the '_' preceding 576 * the name. However, the null character 577 * is double counted between the name and the 578 * sizeof("bpf_prog_") above, so we omit 579 * the +1 here. 580 */ 581 sizeof(prog->aux->name) > KSYM_NAME_LEN); 582 583 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 584 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 585 586 /* prog->aux->name will be ignored if full btf name is available */ 587 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 588 type = btf_type_by_id(prog->aux->btf, 589 prog->aux->func_info[prog->aux->func_idx].type_id); 590 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 591 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 592 return; 593 } 594 595 if (prog->aux->name[0]) 596 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 597 else 598 *sym = 0; 599 } 600 601 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 602 { 603 return container_of(n, struct bpf_ksym, tnode)->start; 604 } 605 606 static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 607 struct latch_tree_node *b) 608 { 609 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 610 } 611 612 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 613 { 614 unsigned long val = (unsigned long)key; 615 const struct bpf_ksym *ksym; 616 617 ksym = container_of(n, struct bpf_ksym, tnode); 618 619 if (val < ksym->start) 620 return -1; 621 /* Ensure that we detect return addresses as part of the program, when 622 * the final instruction is a call for a program part of the stack 623 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 624 */ 625 if (val > ksym->end) 626 return 1; 627 628 return 0; 629 } 630 631 static const struct latch_tree_ops bpf_tree_ops = { 632 .less = bpf_tree_less, 633 .comp = bpf_tree_comp, 634 }; 635 636 static DEFINE_SPINLOCK(bpf_lock); 637 static LIST_HEAD(bpf_kallsyms); 638 static struct latch_tree_root bpf_tree __cacheline_aligned; 639 640 void bpf_ksym_add(struct bpf_ksym *ksym) 641 { 642 spin_lock_bh(&bpf_lock); 643 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 644 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 645 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 646 spin_unlock_bh(&bpf_lock); 647 } 648 649 static void __bpf_ksym_del(struct bpf_ksym *ksym) 650 { 651 if (list_empty(&ksym->lnode)) 652 return; 653 654 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 655 list_del_rcu(&ksym->lnode); 656 } 657 658 void bpf_ksym_del(struct bpf_ksym *ksym) 659 { 660 spin_lock_bh(&bpf_lock); 661 __bpf_ksym_del(ksym); 662 spin_unlock_bh(&bpf_lock); 663 } 664 665 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 666 { 667 return fp->jited && !bpf_prog_was_classic(fp); 668 } 669 670 void bpf_prog_kallsyms_add(struct bpf_prog *fp) 671 { 672 if (!bpf_prog_kallsyms_candidate(fp) || 673 !bpf_token_capable(fp->aux->token, CAP_BPF)) 674 return; 675 676 bpf_prog_ksym_set_addr(fp); 677 bpf_prog_ksym_set_name(fp); 678 fp->aux->ksym.prog = true; 679 680 bpf_ksym_add(&fp->aux->ksym); 681 682 #ifdef CONFIG_FINEIBT 683 /* 684 * When FineIBT, code in the __cfi_foo() symbols can get executed 685 * and hence unwinder needs help. 686 */ 687 if (cfi_mode != CFI_FINEIBT) 688 return; 689 690 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 691 "__cfi_%s", fp->aux->ksym.name); 692 693 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 694 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 695 696 bpf_ksym_add(&fp->aux->ksym_prefix); 697 #endif 698 } 699 700 void bpf_prog_kallsyms_del(struct bpf_prog *fp) 701 { 702 if (!bpf_prog_kallsyms_candidate(fp)) 703 return; 704 705 bpf_ksym_del(&fp->aux->ksym); 706 #ifdef CONFIG_FINEIBT 707 if (cfi_mode != CFI_FINEIBT) 708 return; 709 bpf_ksym_del(&fp->aux->ksym_prefix); 710 #endif 711 } 712 713 static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 714 { 715 struct latch_tree_node *n; 716 717 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 718 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 719 } 720 721 int bpf_address_lookup(unsigned long addr, unsigned long *size, 722 unsigned long *off, char *sym) 723 { 724 struct bpf_ksym *ksym; 725 int ret = 0; 726 727 rcu_read_lock(); 728 ksym = bpf_ksym_find(addr); 729 if (ksym) { 730 unsigned long symbol_start = ksym->start; 731 unsigned long symbol_end = ksym->end; 732 733 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 734 735 if (size) 736 *size = symbol_end - symbol_start; 737 if (off) 738 *off = addr - symbol_start; 739 } 740 rcu_read_unlock(); 741 742 return ret; 743 } 744 745 bool is_bpf_text_address(unsigned long addr) 746 { 747 bool ret; 748 749 rcu_read_lock(); 750 ret = bpf_ksym_find(addr) != NULL; 751 rcu_read_unlock(); 752 753 return ret; 754 } 755 756 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 757 { 758 struct bpf_ksym *ksym; 759 760 WARN_ON_ONCE(!rcu_read_lock_held()); 761 ksym = bpf_ksym_find(addr); 762 763 return ksym && ksym->prog ? 764 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 765 NULL; 766 } 767 768 bool bpf_has_frame_pointer(unsigned long ip) 769 { 770 struct bpf_ksym *ksym; 771 unsigned long offset; 772 773 guard(rcu)(); 774 775 ksym = bpf_ksym_find(ip); 776 if (!ksym || !ksym->fp_start || !ksym->fp_end) 777 return false; 778 779 offset = ip - ksym->start; 780 781 return offset >= ksym->fp_start && offset < ksym->fp_end; 782 } 783 784 const struct exception_table_entry *search_bpf_extables(unsigned long addr) 785 { 786 const struct exception_table_entry *e = NULL; 787 struct bpf_prog *prog; 788 789 rcu_read_lock(); 790 prog = bpf_prog_ksym_find(addr); 791 if (!prog) 792 goto out; 793 if (!prog->aux->num_exentries) 794 goto out; 795 796 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 797 out: 798 rcu_read_unlock(); 799 return e; 800 } 801 802 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 803 char *sym) 804 { 805 struct bpf_ksym *ksym; 806 unsigned int it = 0; 807 int ret = -ERANGE; 808 809 if (!bpf_jit_kallsyms_enabled()) 810 return ret; 811 812 rcu_read_lock(); 813 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 814 if (it++ != symnum) 815 continue; 816 817 strscpy(sym, ksym->name, KSYM_NAME_LEN); 818 819 *value = ksym->start; 820 *type = BPF_SYM_ELF_TYPE; 821 822 ret = 0; 823 break; 824 } 825 rcu_read_unlock(); 826 827 return ret; 828 } 829 830 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 831 struct bpf_jit_poke_descriptor *poke) 832 { 833 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 834 static const u32 poke_tab_max = 1024; 835 u32 slot = prog->aux->size_poke_tab; 836 u32 size = slot + 1; 837 838 if (size > poke_tab_max) 839 return -ENOSPC; 840 if (poke->tailcall_target || poke->tailcall_target_stable || 841 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 842 return -EINVAL; 843 844 switch (poke->reason) { 845 case BPF_POKE_REASON_TAIL_CALL: 846 if (!poke->tail_call.map) 847 return -EINVAL; 848 break; 849 default: 850 return -EINVAL; 851 } 852 853 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 854 if (!tab) 855 return -ENOMEM; 856 857 memcpy(&tab[slot], poke, sizeof(*poke)); 858 prog->aux->size_poke_tab = size; 859 prog->aux->poke_tab = tab; 860 861 return slot; 862 } 863 864 /* 865 * BPF program pack allocator. 866 * 867 * Most BPF programs are pretty small. Allocating a hole page for each 868 * program is sometime a waste. Many small bpf program also adds pressure 869 * to instruction TLB. To solve this issue, we introduce a BPF program pack 870 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 871 * to host BPF programs. 872 */ 873 #define BPF_PROG_CHUNK_SHIFT 6 874 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 875 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 876 877 struct bpf_prog_pack { 878 struct list_head list; 879 void *ptr; 880 bool arch_flush_needed; 881 unsigned long bitmap[]; 882 }; 883 884 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 885 { 886 memset(area, 0, size); 887 } 888 889 DEFINE_STATIC_CALL_NULL(bpf_arch_pred_flush, bpf_arch_pred_flush); 890 891 /* 892 * Enabled once bpf_arch_pred_flush points at a real flush routine. Lets the 893 * pack allocator test "is a predictor flush wired up at all" with a cheap 894 * static branch instead of repeatedly querying the static call target. 895 */ 896 DEFINE_STATIC_KEY_FALSE(bpf_pred_flush_enabled); 897 898 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 899 900 static DEFINE_MUTEX(pack_mutex); 901 static LIST_HEAD(pack_list); 902 903 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 904 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 905 */ 906 #ifdef PMD_SIZE 907 /* PMD_SIZE is really big for some archs. It doesn't make sense to 908 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 909 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 910 * greater than or equal to 2MB. 911 */ 912 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 913 #else 914 #define BPF_PROG_PACK_SIZE PAGE_SIZE 915 #endif 916 917 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 918 919 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 920 { 921 struct bpf_prog_pack *pack; 922 int err; 923 924 pack = kzalloc_flex(*pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)); 925 if (!pack) 926 return NULL; 927 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 928 if (!pack->ptr) 929 goto out; 930 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 931 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 932 933 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 934 pack->arch_flush_needed = true; 935 set_vm_flush_reset_perms(pack->ptr); 936 err = set_memory_rox((unsigned long)pack->ptr, 937 BPF_PROG_PACK_SIZE / PAGE_SIZE); 938 if (err) 939 goto out; 940 list_add_tail(&pack->list, &pack_list); 941 return pack; 942 943 out: 944 bpf_jit_free_exec(pack->ptr); 945 kfree(pack); 946 return NULL; 947 } 948 949 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns, bool was_classic) 950 { 951 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 952 struct bpf_prog_pack *pack, *fallback_pack = NULL; 953 unsigned long pos, fallback_pos = 0; 954 void *ptr = NULL; 955 956 mutex_lock(&pack_mutex); 957 if (size > BPF_PROG_PACK_SIZE) { 958 /* 959 * Allocations larger than a pack get their own pages, and 960 * predictors are not flushed for such allocation. This is only 961 * safe because cBPF programs (the unprivileged attack surface) 962 * are bounded well below a pack size. 963 */ 964 if (was_classic && static_branch_unlikely(&bpf_pred_flush_enabled)) 965 pr_warn_once("BPF: Predictors not flushed for allocations greater than BPF_PROG_PACK_SIZE\n"); 966 size = round_up(size, PAGE_SIZE); 967 ptr = bpf_jit_alloc_exec(size); 968 if (ptr) { 969 int err; 970 971 bpf_fill_ill_insns(ptr, size); 972 set_vm_flush_reset_perms(ptr); 973 err = set_memory_rox((unsigned long)ptr, 974 size / PAGE_SIZE); 975 if (err) { 976 bpf_jit_free_exec(ptr); 977 ptr = NULL; 978 } 979 } 980 goto out; 981 } 982 list_for_each_entry(pack, &pack_list, list) { 983 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 984 nbits, 0); 985 if (pos >= BPF_PROG_CHUNK_COUNT) 986 continue; 987 /* Flush not enabled, use any pack */ 988 if (!static_branch_unlikely(&bpf_pred_flush_enabled)) 989 goto found_free_area; 990 /* 991 * cBPF reuse of a dirty pack triggers a flush, so prefer a 992 * clean pack for cBPF. eBPF never flushes, so steer it to a 993 * dirty pack and keep clean packs free for cBPF. 994 */ 995 if (was_classic ^ pack->arch_flush_needed) 996 goto found_free_area; 997 if (!fallback_pack) { 998 fallback_pack = pack; 999 fallback_pos = pos; 1000 } 1001 } 1002 1003 /* No preferred pack found */ 1004 if (fallback_pack) { 1005 pack = fallback_pack; 1006 pos = fallback_pos; 1007 goto found_free_area; 1008 } 1009 1010 pack = alloc_new_pack(bpf_fill_ill_insns); 1011 if (!pack) 1012 goto out; 1013 1014 pos = 0; 1015 1016 found_free_area: 1017 /* Flush only for cBPF as it may contain a crafted gadget */ 1018 if (static_branch_unlikely(&bpf_pred_flush_enabled) && 1019 pack->arch_flush_needed && 1020 was_classic) { 1021 struct bpf_prog_pack *p; 1022 1023 static_call_cond(bpf_arch_pred_flush)(); 1024 list_for_each_entry(p, &pack_list, list) 1025 p->arch_flush_needed = false; 1026 } 1027 bitmap_set(pack->bitmap, pos, nbits); 1028 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 1029 1030 out: 1031 mutex_unlock(&pack_mutex); 1032 return ptr; 1033 } 1034 1035 void bpf_prog_pack_free(void *ptr, u32 size) 1036 { 1037 struct bpf_prog_pack *pack = NULL, *tmp; 1038 unsigned int nbits; 1039 unsigned long pos; 1040 1041 mutex_lock(&pack_mutex); 1042 if (size > BPF_PROG_PACK_SIZE) { 1043 bpf_jit_free_exec(ptr); 1044 goto out; 1045 } 1046 1047 list_for_each_entry(tmp, &pack_list, list) { 1048 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 1049 pack = tmp; 1050 break; 1051 } 1052 } 1053 1054 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1055 goto out; 1056 1057 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1058 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1059 1060 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1061 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1062 1063 bitmap_clear(pack->bitmap, pos, nbits); 1064 1065 if (static_branch_unlikely(&bpf_pred_flush_enabled)) 1066 pack->arch_flush_needed = true; 1067 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1068 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1069 list_del(&pack->list); 1070 bpf_jit_free_exec(pack->ptr); 1071 kfree(pack); 1072 } 1073 out: 1074 mutex_unlock(&pack_mutex); 1075 } 1076 1077 static atomic_long_t bpf_jit_current; 1078 1079 /* Can be overridden by an arch's JIT compiler if it has a custom, 1080 * dedicated BPF backend memory area, or if neither of the two 1081 * below apply. 1082 */ 1083 u64 __weak bpf_jit_alloc_exec_limit(void) 1084 { 1085 #if defined(MODULES_VADDR) 1086 return MODULES_END - MODULES_VADDR; 1087 #else 1088 return VMALLOC_END - VMALLOC_START; 1089 #endif 1090 } 1091 1092 static int __init bpf_jit_charge_init(void) 1093 { 1094 /* Only used as heuristic here to derive limit. */ 1095 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1096 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1097 PAGE_SIZE), LONG_MAX); 1098 return 0; 1099 } 1100 pure_initcall(bpf_jit_charge_init); 1101 1102 int bpf_jit_charge_modmem(u32 size) 1103 { 1104 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1105 if (!bpf_capable()) { 1106 atomic_long_sub(size, &bpf_jit_current); 1107 return -EPERM; 1108 } 1109 } 1110 1111 return 0; 1112 } 1113 1114 void bpf_jit_uncharge_modmem(u32 size) 1115 { 1116 atomic_long_sub(size, &bpf_jit_current); 1117 } 1118 1119 void *__weak bpf_jit_alloc_exec(unsigned long size) 1120 { 1121 return execmem_alloc(EXECMEM_BPF, size); 1122 } 1123 1124 void __weak bpf_jit_free_exec(void *addr) 1125 { 1126 execmem_free(addr); 1127 } 1128 1129 struct bpf_binary_header * 1130 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1131 unsigned int alignment, 1132 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1133 { 1134 struct bpf_binary_header *hdr; 1135 u32 size, hole, start; 1136 1137 WARN_ON_ONCE(!is_power_of_2(alignment) || 1138 alignment > BPF_IMAGE_ALIGNMENT); 1139 1140 /* Most of BPF filters are really small, but if some of them 1141 * fill a page, allow at least 128 extra bytes to insert a 1142 * random section of illegal instructions. 1143 */ 1144 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1145 1146 if (bpf_jit_charge_modmem(size)) 1147 return NULL; 1148 hdr = bpf_jit_alloc_exec(size); 1149 if (!hdr) { 1150 bpf_jit_uncharge_modmem(size); 1151 return NULL; 1152 } 1153 1154 /* Fill space with illegal/arch-dep instructions. */ 1155 bpf_fill_ill_insns(hdr, size); 1156 1157 hdr->size = size; 1158 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1159 PAGE_SIZE - sizeof(*hdr)); 1160 start = get_random_u32_below(hole) & ~(alignment - 1); 1161 1162 /* Leave a random number of instructions before BPF code. */ 1163 *image_ptr = &hdr->image[start]; 1164 1165 return hdr; 1166 } 1167 1168 void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1169 { 1170 u32 size = hdr->size; 1171 1172 bpf_jit_free_exec(hdr); 1173 bpf_jit_uncharge_modmem(size); 1174 } 1175 1176 /* Allocate jit binary from bpf_prog_pack allocator. 1177 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1178 * to the memory. To solve this problem, a RW buffer is also allocated at 1179 * as the same time. The JIT engine should calculate offsets based on the 1180 * RO memory address, but write JITed program to the RW buffer. Once the 1181 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1182 * the JITed program to the RO memory. 1183 */ 1184 struct bpf_binary_header * 1185 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1186 unsigned int alignment, 1187 struct bpf_binary_header **rw_header, 1188 u8 **rw_image, 1189 bpf_jit_fill_hole_t bpf_fill_ill_insns, 1190 bool was_classic) 1191 { 1192 struct bpf_binary_header *ro_header; 1193 u32 size, hole, start; 1194 1195 WARN_ON_ONCE(!is_power_of_2(alignment) || 1196 alignment > BPF_IMAGE_ALIGNMENT); 1197 1198 /* add 16 bytes for a random section of illegal instructions */ 1199 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1200 1201 if (bpf_jit_charge_modmem(size)) 1202 return NULL; 1203 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns, was_classic); 1204 if (!ro_header) { 1205 bpf_jit_uncharge_modmem(size); 1206 return NULL; 1207 } 1208 1209 *rw_header = kvmalloc(size, GFP_KERNEL); 1210 if (!*rw_header) { 1211 bpf_prog_pack_free(ro_header, size); 1212 bpf_jit_uncharge_modmem(size); 1213 return NULL; 1214 } 1215 1216 /* Fill space with illegal/arch-dep instructions. */ 1217 bpf_fill_ill_insns(*rw_header, size); 1218 (*rw_header)->size = size; 1219 1220 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1221 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1222 start = get_random_u32_below(hole) & ~(alignment - 1); 1223 1224 *image_ptr = &ro_header->image[start]; 1225 *rw_image = &(*rw_header)->image[start]; 1226 1227 return ro_header; 1228 } 1229 1230 /* Copy JITed text from rw_header to its final location, the ro_header. */ 1231 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1232 struct bpf_binary_header *rw_header) 1233 { 1234 void *ptr; 1235 1236 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1237 1238 kvfree(rw_header); 1239 1240 if (IS_ERR(ptr)) { 1241 bpf_prog_pack_free(ro_header, ro_header->size); 1242 return PTR_ERR(ptr); 1243 } 1244 return 0; 1245 } 1246 1247 /* bpf_jit_binary_pack_free is called in two different scenarios: 1248 * 1) when the program is freed after; 1249 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1250 * For case 2), we need to free both the RO memory and the RW buffer. 1251 * 1252 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1253 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1254 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1255 * bpf_arch_text_copy (when jit fails). 1256 */ 1257 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1258 struct bpf_binary_header *rw_header) 1259 { 1260 u32 size = ro_header->size; 1261 1262 bpf_prog_pack_free(ro_header, size); 1263 kvfree(rw_header); 1264 bpf_jit_uncharge_modmem(size); 1265 } 1266 1267 struct bpf_binary_header * 1268 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1269 { 1270 unsigned long real_start = (unsigned long)fp->bpf_func; 1271 unsigned long addr; 1272 1273 addr = real_start & BPF_PROG_CHUNK_MASK; 1274 return (void *)addr; 1275 } 1276 1277 static inline struct bpf_binary_header * 1278 bpf_jit_binary_hdr(const struct bpf_prog *fp) 1279 { 1280 unsigned long real_start = (unsigned long)fp->bpf_func; 1281 unsigned long addr; 1282 1283 addr = real_start & PAGE_MASK; 1284 return (void *)addr; 1285 } 1286 1287 /* This symbol is only overridden by archs that have different 1288 * requirements than the usual eBPF JITs, f.e. when they only 1289 * implement cBPF JIT, do not set images read-only, etc. 1290 */ 1291 void __weak bpf_jit_free(struct bpf_prog *fp) 1292 { 1293 if (fp->jited) { 1294 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1295 1296 bpf_jit_binary_free(hdr); 1297 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1298 } 1299 1300 bpf_prog_unlock_free(fp); 1301 } 1302 1303 int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1304 const struct bpf_insn *insn, bool extra_pass, 1305 u64 *func_addr, bool *func_addr_fixed) 1306 { 1307 s16 off = insn->off; 1308 s32 imm = insn->imm; 1309 u8 *addr; 1310 int err; 1311 1312 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1313 if (!*func_addr_fixed) { 1314 /* Place-holder address till the last pass has collected 1315 * all addresses for JITed subprograms in which case we 1316 * can pick them up from prog->aux. 1317 */ 1318 if (!extra_pass) 1319 addr = NULL; 1320 else if (prog->aux->func && 1321 off >= 0 && off < prog->aux->real_func_cnt) 1322 addr = (u8 *)prog->aux->func[off]->bpf_func; 1323 else 1324 return -EINVAL; 1325 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1326 bpf_jit_supports_far_kfunc_call()) { 1327 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1328 if (err) 1329 return err; 1330 } else { 1331 /* Address of a BPF helper call. Since part of the core 1332 * kernel, it's always at a fixed location. __bpf_call_base 1333 * and the helper with imm relative to it are both in core 1334 * kernel. 1335 */ 1336 addr = (u8 *)__bpf_call_base + imm; 1337 } 1338 1339 *func_addr = (unsigned long)addr; 1340 return 0; 1341 } 1342 1343 const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1344 { 1345 if (prog->aux->ksym.prog) 1346 return prog->aux->ksym.name; 1347 return prog->aux->name; 1348 } 1349 1350 static int bpf_jit_blind_insn(const struct bpf_insn *from, 1351 const struct bpf_insn *aux, 1352 struct bpf_insn *to_buff, 1353 bool emit_zext) 1354 { 1355 struct bpf_insn *to = to_buff; 1356 u32 imm_rnd = get_random_u32(); 1357 s16 off; 1358 1359 BUILD_BUG_ON(BPF_REG_PARAMS + 2 != MAX_BPF_JIT_REG); 1360 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1361 1362 /* Constraints on AX register: 1363 * 1364 * AX register is inaccessible from user space. It is mapped in 1365 * all JITs, and used here for constant blinding rewrites. It is 1366 * typically "stateless" meaning its contents are only valid within 1367 * the executed instruction, but not across several instructions. 1368 * There are a few exceptions however which are further detailed 1369 * below. 1370 * 1371 * Constant blinding is only used by JITs, not in the interpreter. 1372 * The interpreter uses AX in some occasions as a local temporary 1373 * register e.g. in DIV or MOD instructions. 1374 * 1375 * In restricted circumstances, the verifier can also use the AX 1376 * register for rewrites as long as they do not interfere with 1377 * the above cases! 1378 */ 1379 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1380 goto out; 1381 1382 if (from->imm == 0 && 1383 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1384 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1385 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1386 goto out; 1387 } 1388 1389 switch (from->code) { 1390 case BPF_ALU | BPF_ADD | BPF_K: 1391 case BPF_ALU | BPF_SUB | BPF_K: 1392 case BPF_ALU | BPF_AND | BPF_K: 1393 case BPF_ALU | BPF_OR | BPF_K: 1394 case BPF_ALU | BPF_XOR | BPF_K: 1395 case BPF_ALU | BPF_MUL | BPF_K: 1396 case BPF_ALU | BPF_MOV | BPF_K: 1397 case BPF_ALU | BPF_DIV | BPF_K: 1398 case BPF_ALU | BPF_MOD | BPF_K: 1399 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1400 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1401 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1402 break; 1403 1404 case BPF_ALU64 | BPF_ADD | BPF_K: 1405 case BPF_ALU64 | BPF_SUB | BPF_K: 1406 case BPF_ALU64 | BPF_AND | BPF_K: 1407 case BPF_ALU64 | BPF_OR | BPF_K: 1408 case BPF_ALU64 | BPF_XOR | BPF_K: 1409 case BPF_ALU64 | BPF_MUL | BPF_K: 1410 case BPF_ALU64 | BPF_MOV | BPF_K: 1411 case BPF_ALU64 | BPF_DIV | BPF_K: 1412 case BPF_ALU64 | BPF_MOD | BPF_K: 1413 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1414 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1415 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1416 break; 1417 1418 case BPF_JMP | BPF_JEQ | BPF_K: 1419 case BPF_JMP | BPF_JNE | BPF_K: 1420 case BPF_JMP | BPF_JGT | BPF_K: 1421 case BPF_JMP | BPF_JLT | BPF_K: 1422 case BPF_JMP | BPF_JGE | BPF_K: 1423 case BPF_JMP | BPF_JLE | BPF_K: 1424 case BPF_JMP | BPF_JSGT | BPF_K: 1425 case BPF_JMP | BPF_JSLT | BPF_K: 1426 case BPF_JMP | BPF_JSGE | BPF_K: 1427 case BPF_JMP | BPF_JSLE | BPF_K: 1428 case BPF_JMP | BPF_JSET | BPF_K: 1429 /* Accommodate for extra offset in case of a backjump. */ 1430 off = from->off; 1431 if (off < 0) 1432 off -= 2; 1433 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1434 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1435 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1436 break; 1437 1438 case BPF_JMP32 | BPF_JEQ | BPF_K: 1439 case BPF_JMP32 | BPF_JNE | BPF_K: 1440 case BPF_JMP32 | BPF_JGT | BPF_K: 1441 case BPF_JMP32 | BPF_JLT | BPF_K: 1442 case BPF_JMP32 | BPF_JGE | BPF_K: 1443 case BPF_JMP32 | BPF_JLE | BPF_K: 1444 case BPF_JMP32 | BPF_JSGT | BPF_K: 1445 case BPF_JMP32 | BPF_JSLT | BPF_K: 1446 case BPF_JMP32 | BPF_JSGE | BPF_K: 1447 case BPF_JMP32 | BPF_JSLE | BPF_K: 1448 case BPF_JMP32 | BPF_JSET | BPF_K: 1449 /* Accommodate for extra offset in case of a backjump. */ 1450 off = from->off; 1451 if (off < 0) 1452 off -= 2; 1453 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1454 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1455 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1456 off); 1457 break; 1458 1459 case BPF_LD | BPF_IMM | BPF_DW: 1460 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1461 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1462 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1463 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1464 break; 1465 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1466 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1467 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1468 if (emit_zext) 1469 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1470 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1471 break; 1472 1473 case BPF_ST | BPF_MEM | BPF_DW: 1474 case BPF_ST | BPF_MEM | BPF_W: 1475 case BPF_ST | BPF_MEM | BPF_H: 1476 case BPF_ST | BPF_MEM | BPF_B: 1477 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1478 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1479 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1480 break; 1481 1482 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW: 1483 case BPF_ST | BPF_PROBE_MEM32 | BPF_W: 1484 case BPF_ST | BPF_PROBE_MEM32 | BPF_H: 1485 case BPF_ST | BPF_PROBE_MEM32 | BPF_B: 1486 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ 1487 from->imm); 1488 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1489 /* 1490 * Cannot use BPF_STX_MEM() macro here as it 1491 * hardcodes BPF_MEM mode, losing PROBE_MEM32 1492 * and breaking arena addressing in the JIT. 1493 */ 1494 *to++ = (struct bpf_insn) { 1495 .code = BPF_STX | BPF_PROBE_MEM32 | 1496 BPF_SIZE(from->code), 1497 .dst_reg = from->dst_reg, 1498 .src_reg = BPF_REG_AX, 1499 .off = from->off, 1500 }; 1501 break; 1502 } 1503 out: 1504 return to - to_buff; 1505 } 1506 1507 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1508 gfp_t gfp_extra_flags) 1509 { 1510 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1511 struct bpf_prog *fp; 1512 1513 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1514 if (fp != NULL) { 1515 /* aux->prog still points to the fp_other one, so 1516 * when promoting the clone to the real program, 1517 * this still needs to be adapted. 1518 */ 1519 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1520 } 1521 1522 return fp; 1523 } 1524 1525 static void bpf_prog_clone_free(struct bpf_prog *fp) 1526 { 1527 /* aux was stolen by the other clone, so we cannot free 1528 * it from this path! It will be freed eventually by the 1529 * other program on release. 1530 * 1531 * At this point, we don't need a deferred release since 1532 * clone is guaranteed to not be locked. 1533 */ 1534 fp->aux = NULL; 1535 fp->stats = NULL; 1536 fp->active = NULL; 1537 __bpf_prog_free(fp); 1538 } 1539 1540 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1541 { 1542 /* We have to repoint aux->prog to self, as we don't 1543 * know whether fp here is the clone or the original. 1544 */ 1545 fp->aux->prog = fp; 1546 if (fp->aux->offload) 1547 fp->aux->offload->prog = fp; 1548 bpf_prog_clone_free(fp_other); 1549 } 1550 1551 /* 1552 * Now this function is used only to blind the main prog and must be invoked only when 1553 * bpf_prog_need_blind() returns true. 1554 */ 1555 struct bpf_prog *bpf_jit_blind_constants(struct bpf_verifier_env *env, struct bpf_prog *prog) 1556 { 1557 struct bpf_insn insn_buff[16], aux[2]; 1558 struct bpf_prog *clone, *tmp; 1559 int insn_delta, insn_cnt; 1560 struct bpf_insn *insn; 1561 int i, rewritten; 1562 1563 if (WARN_ON_ONCE(env && env->prog != prog)) 1564 return ERR_PTR(-EINVAL); 1565 1566 clone = bpf_prog_clone_create(prog, GFP_USER); 1567 if (!clone) 1568 return ERR_PTR(-ENOMEM); 1569 1570 /* make sure bpf_patch_insn_data() patches the correct prog */ 1571 if (env) 1572 env->prog = clone; 1573 1574 insn_cnt = clone->len; 1575 insn = clone->insnsi; 1576 1577 for (i = 0; i < insn_cnt; i++, insn++) { 1578 if (bpf_pseudo_func(insn)) { 1579 /* ld_imm64 with an address of bpf subprog is not 1580 * a user controlled constant. Don't randomize it, 1581 * since it will conflict with jit_subprogs() logic. 1582 */ 1583 insn++; 1584 i++; 1585 continue; 1586 } 1587 1588 /* We temporarily need to hold the original ld64 insn 1589 * so that we can still access the first part in the 1590 * second blinding run. 1591 */ 1592 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1593 insn[1].code == 0) 1594 memcpy(aux, insn, sizeof(aux)); 1595 1596 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1597 clone->aux->verifier_zext); 1598 if (!rewritten) 1599 continue; 1600 1601 if (env) 1602 tmp = bpf_patch_insn_data(env, i, insn_buff, rewritten); 1603 else 1604 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1605 1606 if (IS_ERR_OR_NULL(tmp)) { 1607 if (env) 1608 /* restore the original prog */ 1609 env->prog = prog; 1610 /* Patching may have repointed aux->prog during 1611 * realloc from the original one, so we need to 1612 * fix it up here on error. 1613 */ 1614 bpf_jit_prog_release_other(prog, clone); 1615 return IS_ERR(tmp) ? tmp : ERR_PTR(-ENOMEM); 1616 } 1617 1618 clone = tmp; 1619 insn_delta = rewritten - 1; 1620 1621 if (env) 1622 env->prog = clone; 1623 1624 /* Walk new program and skip insns we just inserted. */ 1625 insn = clone->insnsi + i + insn_delta; 1626 insn_cnt += insn_delta; 1627 i += insn_delta; 1628 } 1629 1630 clone->blinded = 1; 1631 return clone; 1632 } 1633 1634 bool bpf_insn_is_indirect_target(const struct bpf_verifier_env *env, const struct bpf_prog *prog, 1635 int insn_idx) 1636 { 1637 if (!env) 1638 return false; 1639 insn_idx += prog->aux->subprog_start; 1640 return env->insn_aux_data[insn_idx].indirect_target; 1641 } 1642 1643 u16 bpf_out_stack_arg_cnt(const struct bpf_verifier_env *env, const struct bpf_prog *prog) 1644 { 1645 const struct bpf_subprog_info *sub; 1646 1647 if (!env) 1648 return 0; 1649 sub = &env->subprog_info[prog->aux->func_idx]; 1650 return sub->stack_arg_cnt - bpf_in_stack_arg_cnt(sub); 1651 } 1652 #endif /* CONFIG_BPF_JIT */ 1653 1654 /* Base function for offset calculation. Needs to go into .text section, 1655 * therefore keeping it non-static as well; will also be used by JITs 1656 * anyway later on, so do not let the compiler omit it. This also needs 1657 * to go into kallsyms for correlation from e.g. bpftool, so naming 1658 * must not change. 1659 */ 1660 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1661 { 1662 return 0; 1663 } 1664 EXPORT_SYMBOL_GPL(__bpf_call_base); 1665 1666 /* All UAPI available opcodes. */ 1667 #define BPF_INSN_MAP(INSN_2, INSN_3) \ 1668 /* 32 bit ALU operations. */ \ 1669 /* Register based. */ \ 1670 INSN_3(ALU, ADD, X), \ 1671 INSN_3(ALU, SUB, X), \ 1672 INSN_3(ALU, AND, X), \ 1673 INSN_3(ALU, OR, X), \ 1674 INSN_3(ALU, LSH, X), \ 1675 INSN_3(ALU, RSH, X), \ 1676 INSN_3(ALU, XOR, X), \ 1677 INSN_3(ALU, MUL, X), \ 1678 INSN_3(ALU, MOV, X), \ 1679 INSN_3(ALU, ARSH, X), \ 1680 INSN_3(ALU, DIV, X), \ 1681 INSN_3(ALU, MOD, X), \ 1682 INSN_2(ALU, NEG), \ 1683 INSN_3(ALU, END, TO_BE), \ 1684 INSN_3(ALU, END, TO_LE), \ 1685 /* Immediate based. */ \ 1686 INSN_3(ALU, ADD, K), \ 1687 INSN_3(ALU, SUB, K), \ 1688 INSN_3(ALU, AND, K), \ 1689 INSN_3(ALU, OR, K), \ 1690 INSN_3(ALU, LSH, K), \ 1691 INSN_3(ALU, RSH, K), \ 1692 INSN_3(ALU, XOR, K), \ 1693 INSN_3(ALU, MUL, K), \ 1694 INSN_3(ALU, MOV, K), \ 1695 INSN_3(ALU, ARSH, K), \ 1696 INSN_3(ALU, DIV, K), \ 1697 INSN_3(ALU, MOD, K), \ 1698 /* 64 bit ALU operations. */ \ 1699 /* Register based. */ \ 1700 INSN_3(ALU64, ADD, X), \ 1701 INSN_3(ALU64, SUB, X), \ 1702 INSN_3(ALU64, AND, X), \ 1703 INSN_3(ALU64, OR, X), \ 1704 INSN_3(ALU64, LSH, X), \ 1705 INSN_3(ALU64, RSH, X), \ 1706 INSN_3(ALU64, XOR, X), \ 1707 INSN_3(ALU64, MUL, X), \ 1708 INSN_3(ALU64, MOV, X), \ 1709 INSN_3(ALU64, ARSH, X), \ 1710 INSN_3(ALU64, DIV, X), \ 1711 INSN_3(ALU64, MOD, X), \ 1712 INSN_2(ALU64, NEG), \ 1713 INSN_3(ALU64, END, TO_LE), \ 1714 /* Immediate based. */ \ 1715 INSN_3(ALU64, ADD, K), \ 1716 INSN_3(ALU64, SUB, K), \ 1717 INSN_3(ALU64, AND, K), \ 1718 INSN_3(ALU64, OR, K), \ 1719 INSN_3(ALU64, LSH, K), \ 1720 INSN_3(ALU64, RSH, K), \ 1721 INSN_3(ALU64, XOR, K), \ 1722 INSN_3(ALU64, MUL, K), \ 1723 INSN_3(ALU64, MOV, K), \ 1724 INSN_3(ALU64, ARSH, K), \ 1725 INSN_3(ALU64, DIV, K), \ 1726 INSN_3(ALU64, MOD, K), \ 1727 /* Call instruction. */ \ 1728 INSN_2(JMP, CALL), \ 1729 /* Exit instruction. */ \ 1730 INSN_2(JMP, EXIT), \ 1731 /* 32-bit Jump instructions. */ \ 1732 /* Register based. */ \ 1733 INSN_3(JMP32, JEQ, X), \ 1734 INSN_3(JMP32, JNE, X), \ 1735 INSN_3(JMP32, JGT, X), \ 1736 INSN_3(JMP32, JLT, X), \ 1737 INSN_3(JMP32, JGE, X), \ 1738 INSN_3(JMP32, JLE, X), \ 1739 INSN_3(JMP32, JSGT, X), \ 1740 INSN_3(JMP32, JSLT, X), \ 1741 INSN_3(JMP32, JSGE, X), \ 1742 INSN_3(JMP32, JSLE, X), \ 1743 INSN_3(JMP32, JSET, X), \ 1744 /* Immediate based. */ \ 1745 INSN_3(JMP32, JEQ, K), \ 1746 INSN_3(JMP32, JNE, K), \ 1747 INSN_3(JMP32, JGT, K), \ 1748 INSN_3(JMP32, JLT, K), \ 1749 INSN_3(JMP32, JGE, K), \ 1750 INSN_3(JMP32, JLE, K), \ 1751 INSN_3(JMP32, JSGT, K), \ 1752 INSN_3(JMP32, JSLT, K), \ 1753 INSN_3(JMP32, JSGE, K), \ 1754 INSN_3(JMP32, JSLE, K), \ 1755 INSN_3(JMP32, JSET, K), \ 1756 /* Jump instructions. */ \ 1757 /* Register based. */ \ 1758 INSN_3(JMP, JEQ, X), \ 1759 INSN_3(JMP, JNE, X), \ 1760 INSN_3(JMP, JGT, X), \ 1761 INSN_3(JMP, JLT, X), \ 1762 INSN_3(JMP, JGE, X), \ 1763 INSN_3(JMP, JLE, X), \ 1764 INSN_3(JMP, JSGT, X), \ 1765 INSN_3(JMP, JSLT, X), \ 1766 INSN_3(JMP, JSGE, X), \ 1767 INSN_3(JMP, JSLE, X), \ 1768 INSN_3(JMP, JSET, X), \ 1769 /* Immediate based. */ \ 1770 INSN_3(JMP, JEQ, K), \ 1771 INSN_3(JMP, JNE, K), \ 1772 INSN_3(JMP, JGT, K), \ 1773 INSN_3(JMP, JLT, K), \ 1774 INSN_3(JMP, JGE, K), \ 1775 INSN_3(JMP, JLE, K), \ 1776 INSN_3(JMP, JSGT, K), \ 1777 INSN_3(JMP, JSLT, K), \ 1778 INSN_3(JMP, JSGE, K), \ 1779 INSN_3(JMP, JSLE, K), \ 1780 INSN_3(JMP, JSET, K), \ 1781 INSN_2(JMP, JA), \ 1782 INSN_2(JMP32, JA), \ 1783 /* Atomic operations. */ \ 1784 INSN_3(STX, ATOMIC, B), \ 1785 INSN_3(STX, ATOMIC, H), \ 1786 INSN_3(STX, ATOMIC, W), \ 1787 INSN_3(STX, ATOMIC, DW), \ 1788 /* Store instructions. */ \ 1789 /* Register based. */ \ 1790 INSN_3(STX, MEM, B), \ 1791 INSN_3(STX, MEM, H), \ 1792 INSN_3(STX, MEM, W), \ 1793 INSN_3(STX, MEM, DW), \ 1794 /* Immediate based. */ \ 1795 INSN_3(ST, MEM, B), \ 1796 INSN_3(ST, MEM, H), \ 1797 INSN_3(ST, MEM, W), \ 1798 INSN_3(ST, MEM, DW), \ 1799 /* Load instructions. */ \ 1800 /* Register based. */ \ 1801 INSN_3(LDX, MEM, B), \ 1802 INSN_3(LDX, MEM, H), \ 1803 INSN_3(LDX, MEM, W), \ 1804 INSN_3(LDX, MEM, DW), \ 1805 INSN_3(LDX, MEMSX, B), \ 1806 INSN_3(LDX, MEMSX, H), \ 1807 INSN_3(LDX, MEMSX, W), \ 1808 /* Immediate based. */ \ 1809 INSN_3(LD, IMM, DW) 1810 1811 bool bpf_opcode_in_insntable(u8 code) 1812 { 1813 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1814 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1815 static const bool public_insntable[256] = { 1816 [0 ... 255] = false, 1817 /* Now overwrite non-defaults ... */ 1818 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1819 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1820 [BPF_LD | BPF_ABS | BPF_B] = true, 1821 [BPF_LD | BPF_ABS | BPF_H] = true, 1822 [BPF_LD | BPF_ABS | BPF_W] = true, 1823 [BPF_LD | BPF_IND | BPF_B] = true, 1824 [BPF_LD | BPF_IND | BPF_H] = true, 1825 [BPF_LD | BPF_IND | BPF_W] = true, 1826 [BPF_JMP | BPF_JA | BPF_X] = true, 1827 [BPF_JMP | BPF_JCOND] = true, 1828 }; 1829 #undef BPF_INSN_3_TBL 1830 #undef BPF_INSN_2_TBL 1831 return public_insntable[code]; 1832 } 1833 1834 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 1835 /* Absolute value of s32 without undefined behavior for S32_MIN */ 1836 static u32 abs_s32(s32 x) 1837 { 1838 return x >= 0 ? (u32)x : -(u32)x; 1839 } 1840 1841 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 1842 const struct bpf_insn *insn); 1843 1844 /** 1845 * ___bpf_prog_run - run eBPF program on a given context 1846 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1847 * @insn: is the array of eBPF instructions 1848 * 1849 * Decode and execute eBPF instructions. 1850 * 1851 * Return: whatever value is in %BPF_R0 at program exit 1852 */ 1853 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1854 { 1855 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1856 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1857 static const void * const jumptable[256] __annotate_jump_table = { 1858 [0 ... 255] = &&default_label, 1859 /* Now overwrite non-defaults ... */ 1860 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1861 /* Non-UAPI available opcodes. */ 1862 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1863 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1864 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1865 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1866 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1867 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1868 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1869 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1870 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1871 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1872 }; 1873 #undef BPF_INSN_3_LBL 1874 #undef BPF_INSN_2_LBL 1875 u32 tail_call_cnt = 0; 1876 1877 #define CONT ({ insn++; goto select_insn; }) 1878 #define CONT_JMP ({ insn++; goto select_insn; }) 1879 1880 select_insn: 1881 goto *jumptable[insn->code]; 1882 1883 /* Explicitly mask the register-based shift amounts with 63 or 31 1884 * to avoid undefined behavior. Normally this won't affect the 1885 * generated code, for example, in case of native 64 bit archs such 1886 * as x86-64 or arm64, the compiler is optimizing the AND away for 1887 * the interpreter. In case of JITs, each of the JIT backends compiles 1888 * the BPF shift operations to machine instructions which produce 1889 * implementation-defined results in such a case; the resulting 1890 * contents of the register may be arbitrary, but program behaviour 1891 * as a whole remains defined. In other words, in case of JIT backends, 1892 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1893 */ 1894 /* ALU (shifts) */ 1895 #define SHT(OPCODE, OP) \ 1896 ALU64_##OPCODE##_X: \ 1897 DST = DST OP (SRC & 63); \ 1898 CONT; \ 1899 ALU_##OPCODE##_X: \ 1900 DST = (u32) DST OP ((u32) SRC & 31); \ 1901 CONT; \ 1902 ALU64_##OPCODE##_K: \ 1903 DST = DST OP IMM; \ 1904 CONT; \ 1905 ALU_##OPCODE##_K: \ 1906 DST = (u32) DST OP (u32) IMM; \ 1907 CONT; 1908 /* ALU (rest) */ 1909 #define ALU(OPCODE, OP) \ 1910 ALU64_##OPCODE##_X: \ 1911 DST = DST OP SRC; \ 1912 CONT; \ 1913 ALU_##OPCODE##_X: \ 1914 DST = (u32) DST OP (u32) SRC; \ 1915 CONT; \ 1916 ALU64_##OPCODE##_K: \ 1917 DST = DST OP IMM; \ 1918 CONT; \ 1919 ALU_##OPCODE##_K: \ 1920 DST = (u32) DST OP (u32) IMM; \ 1921 CONT; 1922 ALU(ADD, +) 1923 ALU(SUB, -) 1924 ALU(AND, &) 1925 ALU(OR, |) 1926 ALU(XOR, ^) 1927 ALU(MUL, *) 1928 SHT(LSH, <<) 1929 SHT(RSH, >>) 1930 #undef SHT 1931 #undef ALU 1932 ALU_NEG: 1933 DST = (u32) -DST; 1934 CONT; 1935 ALU64_NEG: 1936 DST = -DST; 1937 CONT; 1938 ALU_MOV_X: 1939 switch (OFF) { 1940 case 0: 1941 DST = (u32) SRC; 1942 break; 1943 case 8: 1944 DST = (u32)(s8) SRC; 1945 break; 1946 case 16: 1947 DST = (u32)(s16) SRC; 1948 break; 1949 } 1950 CONT; 1951 ALU_MOV_K: 1952 DST = (u32) IMM; 1953 CONT; 1954 ALU64_MOV_X: 1955 switch (OFF) { 1956 case 0: 1957 DST = SRC; 1958 break; 1959 case 8: 1960 DST = (s8) SRC; 1961 break; 1962 case 16: 1963 DST = (s16) SRC; 1964 break; 1965 case 32: 1966 DST = (s32) SRC; 1967 break; 1968 } 1969 CONT; 1970 ALU64_MOV_K: 1971 DST = IMM; 1972 CONT; 1973 LD_IMM_DW: 1974 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1975 insn++; 1976 CONT; 1977 ALU_ARSH_X: 1978 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1979 CONT; 1980 ALU_ARSH_K: 1981 DST = (u64) (u32) (((s32) DST) >> IMM); 1982 CONT; 1983 ALU64_ARSH_X: 1984 (*(s64 *) &DST) >>= (SRC & 63); 1985 CONT; 1986 ALU64_ARSH_K: 1987 (*(s64 *) &DST) >>= IMM; 1988 CONT; 1989 ALU64_MOD_X: 1990 switch (OFF) { 1991 case 0: 1992 div64_u64_rem(DST, SRC, &AX); 1993 DST = AX; 1994 break; 1995 case 1: 1996 AX = div64_s64(DST, SRC); 1997 DST = DST - AX * SRC; 1998 break; 1999 } 2000 CONT; 2001 ALU_MOD_X: 2002 switch (OFF) { 2003 case 0: 2004 AX = (u32) DST; 2005 DST = do_div(AX, (u32) SRC); 2006 break; 2007 case 1: 2008 AX = abs_s32((s32)DST); 2009 AX = do_div(AX, abs_s32((s32)SRC)); 2010 if ((s32)DST < 0) 2011 DST = (u32)-AX; 2012 else 2013 DST = (u32)AX; 2014 break; 2015 } 2016 CONT; 2017 ALU64_MOD_K: 2018 switch (OFF) { 2019 case 0: 2020 div64_u64_rem(DST, IMM, &AX); 2021 DST = AX; 2022 break; 2023 case 1: 2024 AX = div64_s64(DST, IMM); 2025 DST = DST - AX * IMM; 2026 break; 2027 } 2028 CONT; 2029 ALU_MOD_K: 2030 switch (OFF) { 2031 case 0: 2032 AX = (u32) DST; 2033 DST = do_div(AX, (u32) IMM); 2034 break; 2035 case 1: 2036 AX = abs_s32((s32)DST); 2037 AX = do_div(AX, abs_s32((s32)IMM)); 2038 if ((s32)DST < 0) 2039 DST = (u32)-AX; 2040 else 2041 DST = (u32)AX; 2042 break; 2043 } 2044 CONT; 2045 ALU64_DIV_X: 2046 switch (OFF) { 2047 case 0: 2048 DST = div64_u64(DST, SRC); 2049 break; 2050 case 1: 2051 DST = div64_s64(DST, SRC); 2052 break; 2053 } 2054 CONT; 2055 ALU_DIV_X: 2056 switch (OFF) { 2057 case 0: 2058 AX = (u32) DST; 2059 do_div(AX, (u32) SRC); 2060 DST = (u32) AX; 2061 break; 2062 case 1: 2063 AX = abs_s32((s32)DST); 2064 do_div(AX, abs_s32((s32)SRC)); 2065 if (((s32)DST < 0) == ((s32)SRC < 0)) 2066 DST = (u32)AX; 2067 else 2068 DST = (u32)-AX; 2069 break; 2070 } 2071 CONT; 2072 ALU64_DIV_K: 2073 switch (OFF) { 2074 case 0: 2075 DST = div64_u64(DST, IMM); 2076 break; 2077 case 1: 2078 DST = div64_s64(DST, IMM); 2079 break; 2080 } 2081 CONT; 2082 ALU_DIV_K: 2083 switch (OFF) { 2084 case 0: 2085 AX = (u32) DST; 2086 do_div(AX, (u32) IMM); 2087 DST = (u32) AX; 2088 break; 2089 case 1: 2090 AX = abs_s32((s32)DST); 2091 do_div(AX, abs_s32((s32)IMM)); 2092 if (((s32)DST < 0) == ((s32)IMM < 0)) 2093 DST = (u32)AX; 2094 else 2095 DST = (u32)-AX; 2096 break; 2097 } 2098 CONT; 2099 ALU_END_TO_BE: 2100 switch (IMM) { 2101 case 16: 2102 DST = (__force u16) cpu_to_be16(DST); 2103 break; 2104 case 32: 2105 DST = (__force u32) cpu_to_be32(DST); 2106 break; 2107 case 64: 2108 DST = (__force u64) cpu_to_be64(DST); 2109 break; 2110 } 2111 CONT; 2112 ALU_END_TO_LE: 2113 switch (IMM) { 2114 case 16: 2115 DST = (__force u16) cpu_to_le16(DST); 2116 break; 2117 case 32: 2118 DST = (__force u32) cpu_to_le32(DST); 2119 break; 2120 case 64: 2121 DST = (__force u64) cpu_to_le64(DST); 2122 break; 2123 } 2124 CONT; 2125 ALU64_END_TO_LE: 2126 switch (IMM) { 2127 case 16: 2128 DST = (__force u16) __swab16(DST); 2129 break; 2130 case 32: 2131 DST = (__force u32) __swab32(DST); 2132 break; 2133 case 64: 2134 DST = (__force u64) __swab64(DST); 2135 break; 2136 } 2137 CONT; 2138 2139 /* CALL */ 2140 JMP_CALL: 2141 /* Function call scratches BPF_R1-BPF_R5 registers, 2142 * preserves BPF_R6-BPF_R9, and stores return value 2143 * into BPF_R0. 2144 */ 2145 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2146 BPF_R4, BPF_R5); 2147 CONT; 2148 2149 JMP_CALL_ARGS: 2150 BPF_R0 = interpreters_args[insn->off](BPF_R1, BPF_R2, BPF_R3, 2151 BPF_R4, BPF_R5, 2152 insn + insn->imm + 1); 2153 CONT; 2154 2155 JMP_TAIL_CALL: { 2156 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2157 struct bpf_array *array = container_of(map, struct bpf_array, map); 2158 struct bpf_prog *prog; 2159 u32 index = BPF_R3; 2160 2161 if (unlikely(index >= array->map.max_entries)) 2162 goto out; 2163 2164 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2165 goto out; 2166 2167 prog = READ_ONCE(array->ptrs[index]); 2168 if (!prog) 2169 goto out; 2170 2171 tail_call_cnt++; 2172 2173 /* ARG1 at this point is guaranteed to point to CTX from 2174 * the verifier side due to the fact that the tail call is 2175 * handled like a helper, that is, bpf_tail_call_proto, 2176 * where arg1_type is ARG_PTR_TO_CTX. 2177 */ 2178 insn = prog->insnsi; 2179 goto select_insn; 2180 out: 2181 CONT; 2182 } 2183 JMP_JA: 2184 insn += insn->off; 2185 CONT; 2186 JMP32_JA: 2187 insn += insn->imm; 2188 CONT; 2189 JMP_EXIT: 2190 return BPF_R0; 2191 /* JMP */ 2192 #define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2193 JMP_##OPCODE##_X: \ 2194 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2195 insn += insn->off; \ 2196 CONT_JMP; \ 2197 } \ 2198 CONT; \ 2199 JMP32_##OPCODE##_X: \ 2200 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2201 insn += insn->off; \ 2202 CONT_JMP; \ 2203 } \ 2204 CONT; \ 2205 JMP_##OPCODE##_K: \ 2206 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2207 insn += insn->off; \ 2208 CONT_JMP; \ 2209 } \ 2210 CONT; \ 2211 JMP32_##OPCODE##_K: \ 2212 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2213 insn += insn->off; \ 2214 CONT_JMP; \ 2215 } \ 2216 CONT; 2217 COND_JMP(u, JEQ, ==) 2218 COND_JMP(u, JNE, !=) 2219 COND_JMP(u, JGT, >) 2220 COND_JMP(u, JLT, <) 2221 COND_JMP(u, JGE, >=) 2222 COND_JMP(u, JLE, <=) 2223 COND_JMP(u, JSET, &) 2224 COND_JMP(s, JSGT, >) 2225 COND_JMP(s, JSLT, <) 2226 COND_JMP(s, JSGE, >=) 2227 COND_JMP(s, JSLE, <=) 2228 #undef COND_JMP 2229 /* ST, STX and LDX*/ 2230 ST_NOSPEC: 2231 /* Speculation barrier for mitigating Speculative Store Bypass, 2232 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2233 * rely on the firmware mitigation as controlled via the ssbd 2234 * kernel parameter. Whenever the mitigation is enabled, it 2235 * works for all of the kernel code with no need to provide any 2236 * additional instructions here. In case of x86, we use 'lfence' 2237 * insn for mitigation. We reuse preexisting logic from Spectre 2238 * v1 mitigation that happens to produce the required code on 2239 * x86 for v4 as well. 2240 */ 2241 barrier_nospec(); 2242 CONT; 2243 #define LDST(SIZEOP, SIZE) \ 2244 STX_MEM_##SIZEOP: \ 2245 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2246 CONT; \ 2247 ST_MEM_##SIZEOP: \ 2248 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2249 CONT; \ 2250 LDX_MEM_##SIZEOP: \ 2251 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2252 CONT; \ 2253 LDX_PROBE_MEM_##SIZEOP: \ 2254 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2255 (const void *)(long) (SRC + insn->off)); \ 2256 DST = *((SIZE *)&DST); \ 2257 CONT; 2258 2259 LDST(B, u8) 2260 LDST(H, u16) 2261 LDST(W, u32) 2262 LDST(DW, u64) 2263 #undef LDST 2264 2265 #define LDSX(SIZEOP, SIZE) \ 2266 LDX_MEMSX_##SIZEOP: \ 2267 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2268 CONT; \ 2269 LDX_PROBE_MEMSX_##SIZEOP: \ 2270 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2271 (const void *)(long) (SRC + insn->off)); \ 2272 DST = *((SIZE *)&DST); \ 2273 CONT; 2274 2275 LDSX(B, s8) 2276 LDSX(H, s16) 2277 LDSX(W, s32) 2278 #undef LDSX 2279 2280 #define ATOMIC_ALU_OP(BOP, KOP) \ 2281 case BOP: \ 2282 if (BPF_SIZE(insn->code) == BPF_W) \ 2283 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2284 (DST + insn->off)); \ 2285 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2286 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2287 (DST + insn->off)); \ 2288 else \ 2289 goto default_label; \ 2290 break; \ 2291 case BOP | BPF_FETCH: \ 2292 if (BPF_SIZE(insn->code) == BPF_W) \ 2293 SRC = (u32) atomic_fetch_##KOP( \ 2294 (u32) SRC, \ 2295 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2296 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2297 SRC = (u64) atomic64_fetch_##KOP( \ 2298 (u64) SRC, \ 2299 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2300 else \ 2301 goto default_label; \ 2302 break; 2303 2304 STX_ATOMIC_DW: 2305 STX_ATOMIC_W: 2306 STX_ATOMIC_H: 2307 STX_ATOMIC_B: 2308 switch (IMM) { 2309 /* Atomic read-modify-write instructions support only W and DW 2310 * size modifiers. 2311 */ 2312 ATOMIC_ALU_OP(BPF_ADD, add) 2313 ATOMIC_ALU_OP(BPF_AND, and) 2314 ATOMIC_ALU_OP(BPF_OR, or) 2315 ATOMIC_ALU_OP(BPF_XOR, xor) 2316 #undef ATOMIC_ALU_OP 2317 2318 case BPF_XCHG: 2319 if (BPF_SIZE(insn->code) == BPF_W) 2320 SRC = (u32) atomic_xchg( 2321 (atomic_t *)(unsigned long) (DST + insn->off), 2322 (u32) SRC); 2323 else if (BPF_SIZE(insn->code) == BPF_DW) 2324 SRC = (u64) atomic64_xchg( 2325 (atomic64_t *)(unsigned long) (DST + insn->off), 2326 (u64) SRC); 2327 else 2328 goto default_label; 2329 break; 2330 case BPF_CMPXCHG: 2331 if (BPF_SIZE(insn->code) == BPF_W) 2332 BPF_R0 = (u32) atomic_cmpxchg( 2333 (atomic_t *)(unsigned long) (DST + insn->off), 2334 (u32) BPF_R0, (u32) SRC); 2335 else if (BPF_SIZE(insn->code) == BPF_DW) 2336 BPF_R0 = (u64) atomic64_cmpxchg( 2337 (atomic64_t *)(unsigned long) (DST + insn->off), 2338 (u64) BPF_R0, (u64) SRC); 2339 else 2340 goto default_label; 2341 break; 2342 /* Atomic load and store instructions support all size 2343 * modifiers. 2344 */ 2345 case BPF_LOAD_ACQ: 2346 switch (BPF_SIZE(insn->code)) { 2347 #define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2348 case BPF_##SIZEOP: \ 2349 DST = (SIZE)smp_load_acquire( \ 2350 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2351 break; 2352 LOAD_ACQUIRE(B, u8) 2353 LOAD_ACQUIRE(H, u16) 2354 LOAD_ACQUIRE(W, u32) 2355 #ifdef CONFIG_64BIT 2356 LOAD_ACQUIRE(DW, u64) 2357 #endif 2358 #undef LOAD_ACQUIRE 2359 default: 2360 goto default_label; 2361 } 2362 break; 2363 case BPF_STORE_REL: 2364 switch (BPF_SIZE(insn->code)) { 2365 #define STORE_RELEASE(SIZEOP, SIZE) \ 2366 case BPF_##SIZEOP: \ 2367 smp_store_release( \ 2368 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2369 break; 2370 STORE_RELEASE(B, u8) 2371 STORE_RELEASE(H, u16) 2372 STORE_RELEASE(W, u32) 2373 #ifdef CONFIG_64BIT 2374 STORE_RELEASE(DW, u64) 2375 #endif 2376 #undef STORE_RELEASE 2377 default: 2378 goto default_label; 2379 } 2380 break; 2381 2382 default: 2383 goto default_label; 2384 } 2385 CONT; 2386 2387 default_label: 2388 /* If we ever reach this, we have a bug somewhere. Die hard here 2389 * instead of just returning 0; we could be somewhere in a subprog, 2390 * so execution could continue otherwise which we do /not/ want. 2391 * 2392 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2393 */ 2394 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2395 insn->code, insn->imm); 2396 BUG_ON(1); 2397 return 0; 2398 } 2399 2400 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2401 #define DEFINE_BPF_PROG_RUN(stack_size) \ 2402 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2403 { \ 2404 u64 stack[stack_size / sizeof(u64)]; \ 2405 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2406 \ 2407 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2408 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2409 ARG1 = (u64) (unsigned long) ctx; \ 2410 return ___bpf_prog_run(regs, insn); \ 2411 } 2412 2413 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2414 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2415 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2416 const struct bpf_insn *insn) \ 2417 { \ 2418 u64 stack[stack_size / sizeof(u64)]; \ 2419 u64 regs[MAX_BPF_EXT_REG]; \ 2420 \ 2421 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2422 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2423 BPF_R1 = r1; \ 2424 BPF_R2 = r2; \ 2425 BPF_R3 = r3; \ 2426 BPF_R4 = r4; \ 2427 BPF_R5 = r5; \ 2428 return ___bpf_prog_run(regs, insn); \ 2429 } 2430 2431 #define EVAL1(FN, X) FN(X) 2432 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2433 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2434 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2435 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2436 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2437 2438 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2439 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2440 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2441 2442 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2443 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2444 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2445 2446 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2447 2448 static unsigned int (*interpreters[])(const void *ctx, 2449 const struct bpf_insn *insn) = { 2450 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2451 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2452 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2453 }; 2454 #undef PROG_NAME_LIST 2455 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2456 static __maybe_unused 2457 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2458 const struct bpf_insn *insn) = { 2459 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2460 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2461 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2462 }; 2463 #undef PROG_NAME_LIST 2464 2465 #ifdef CONFIG_BPF_SYSCALL 2466 int bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2467 { 2468 stack_depth = max_t(u32, stack_depth, 1); 2469 /* Prevent out-of-bounds read to interpreters_args */ 2470 if (stack_depth > MAX_BPF_STACK) 2471 return -EINVAL; 2472 insn->off = (round_up(stack_depth, 32) / 32) - 1; 2473 insn->code = BPF_JMP | BPF_CALL_ARGS; 2474 return 0; 2475 } 2476 2477 s32 bpf_call_args_imm(s16 idx) 2478 { 2479 if (WARN_ON_ONCE(idx < 0 || idx >= ARRAY_SIZE(interpreters_args))) 2480 return 0; 2481 return BPF_CALL_IMM(interpreters_args[idx]); 2482 } 2483 #endif 2484 #endif 2485 2486 static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2487 const struct bpf_insn *insn) 2488 { 2489 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2490 * is not working properly, so warn about it! 2491 */ 2492 WARN_ON_ONCE(1); 2493 return 0; 2494 } 2495 2496 static bool __bpf_prog_map_compatible(struct bpf_map *map, 2497 const struct bpf_prog *fp) 2498 { 2499 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2500 struct bpf_prog_aux *aux = fp->aux; 2501 enum bpf_cgroup_storage_type i; 2502 bool ret = false; 2503 u64 cookie; 2504 2505 if (fp->kprobe_override) 2506 return ret; 2507 2508 spin_lock(&map->owner_lock); 2509 /* There's no owner yet where we could check for compatibility. */ 2510 if (!map->owner) { 2511 map->owner = bpf_map_owner_alloc(map); 2512 if (!map->owner) 2513 goto err; 2514 map->owner->type = prog_type; 2515 map->owner->jited = fp->jited; 2516 map->owner->xdp_has_frags = aux->xdp_has_frags; 2517 map->owner->sleepable = fp->sleepable; 2518 map->owner->expected_attach_type = fp->expected_attach_type; 2519 map->owner->attach_func_proto = aux->attach_func_proto; 2520 for_each_cgroup_storage_type(i) { 2521 map->owner->storage_cookie[i] = 2522 aux->cgroup_storage[i] ? 2523 aux->cgroup_storage[i]->cookie : 0; 2524 } 2525 ret = true; 2526 } else { 2527 ret = map->owner->type == prog_type && 2528 map->owner->jited == fp->jited && 2529 map->owner->xdp_has_frags == aux->xdp_has_frags && 2530 map->owner->sleepable == fp->sleepable; 2531 if (ret && 2532 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2533 map->owner->expected_attach_type != fp->expected_attach_type) 2534 ret = false; 2535 for_each_cgroup_storage_type(i) { 2536 if (!ret) 2537 break; 2538 cookie = aux->cgroup_storage[i] ? 2539 aux->cgroup_storage[i]->cookie : 0; 2540 ret = map->owner->storage_cookie[i] == cookie || 2541 (!cookie && !aux->tail_call_reachable); 2542 } 2543 if (ret && 2544 map->owner->attach_func_proto != aux->attach_func_proto) { 2545 switch (prog_type) { 2546 case BPF_PROG_TYPE_TRACING: 2547 case BPF_PROG_TYPE_LSM: 2548 case BPF_PROG_TYPE_EXT: 2549 case BPF_PROG_TYPE_STRUCT_OPS: 2550 ret = false; 2551 break; 2552 default: 2553 break; 2554 } 2555 } 2556 } 2557 err: 2558 spin_unlock(&map->owner_lock); 2559 return ret; 2560 } 2561 2562 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2563 { 2564 /* XDP programs inserted into maps are not guaranteed to run on 2565 * a particular netdev (and can run outside driver context entirely 2566 * in the case of devmap and cpumap). Until device checks 2567 * are implemented, prohibit adding dev-bound programs to program maps. 2568 */ 2569 if (bpf_prog_is_dev_bound(fp->aux)) 2570 return false; 2571 2572 return __bpf_prog_map_compatible(map, fp); 2573 } 2574 2575 static int bpf_check_tail_call(const struct bpf_prog *fp) 2576 { 2577 struct bpf_prog_aux *aux = fp->aux; 2578 int i, ret = 0; 2579 2580 mutex_lock(&aux->used_maps_mutex); 2581 for (i = 0; i < aux->used_map_cnt; i++) { 2582 struct bpf_map *map = aux->used_maps[i]; 2583 2584 if (!map_type_contains_progs(map)) 2585 continue; 2586 2587 if (!__bpf_prog_map_compatible(map, fp)) { 2588 ret = -EINVAL; 2589 goto out; 2590 } 2591 } 2592 2593 out: 2594 mutex_unlock(&aux->used_maps_mutex); 2595 return ret; 2596 } 2597 2598 static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2599 { 2600 bool select_interpreter = false; 2601 #ifndef CONFIG_BPF_JIT_ALWAYS_ON 2602 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2603 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2604 2605 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2606 * But for non-JITed programs, we don't need bpf_func, so no bounds 2607 * check needed. 2608 */ 2609 if (idx < ARRAY_SIZE(interpreters)) { 2610 fp->bpf_func = interpreters[idx]; 2611 select_interpreter = true; 2612 } else { 2613 fp->bpf_func = __bpf_prog_ret0_warn; 2614 } 2615 #else 2616 fp->bpf_func = __bpf_prog_ret0_warn; 2617 #endif 2618 return select_interpreter; 2619 } 2620 2621 static struct bpf_prog *bpf_prog_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 2622 { 2623 #ifdef CONFIG_BPF_JIT 2624 struct bpf_prog *orig_prog; 2625 struct bpf_insn_aux_data *orig_insn_aux; 2626 2627 if (!bpf_prog_need_blind(prog)) 2628 return bpf_int_jit_compile(env, prog); 2629 2630 if (env) { 2631 /* 2632 * If env is not NULL, we are called from the end of bpf_check(), at this 2633 * point, only insn_aux_data is used after failure, so it should be restored 2634 * on failure. 2635 */ 2636 orig_insn_aux = bpf_dup_insn_aux_data(env); 2637 if (!orig_insn_aux) 2638 return prog; 2639 } 2640 2641 orig_prog = prog; 2642 prog = bpf_jit_blind_constants(env, prog); 2643 /* 2644 * If blinding was requested and we failed during blinding, we must fall 2645 * back to the interpreter. 2646 */ 2647 if (IS_ERR(prog)) 2648 goto out_restore; 2649 2650 prog = bpf_int_jit_compile(env, prog); 2651 if (prog->jited) { 2652 bpf_jit_prog_release_other(prog, orig_prog); 2653 if (env) 2654 vfree(orig_insn_aux); 2655 return prog; 2656 } 2657 2658 bpf_jit_prog_release_other(orig_prog, prog); 2659 2660 out_restore: 2661 prog = orig_prog; 2662 if (env) 2663 bpf_restore_insn_aux_data(env, orig_insn_aux); 2664 #endif 2665 return prog; 2666 } 2667 2668 struct bpf_prog *__bpf_prog_select_runtime(struct bpf_verifier_env *env, struct bpf_prog *fp, 2669 int *err) 2670 { 2671 /* In case of BPF to BPF calls, verifier did all the prep 2672 * work with regards to JITing, etc. 2673 */ 2674 bool jit_needed = fp->jit_required; 2675 2676 if (fp->bpf_func) 2677 goto finalize; 2678 2679 if (!bpf_prog_select_interpreter(fp)) 2680 jit_needed = true; 2681 2682 /* eBPF JITs can rewrite the program in case constant 2683 * blinding is active. However, in case of error during 2684 * blinding, bpf_int_jit_compile() must always return a 2685 * valid program, which in this case would simply not 2686 * be JITed, but falls back to the interpreter. 2687 */ 2688 if (!bpf_prog_is_offloaded(fp->aux)) { 2689 *err = bpf_prog_alloc_jited_linfo(fp); 2690 if (*err) 2691 return fp; 2692 2693 fp = bpf_prog_jit_compile(env, fp); 2694 bpf_prog_jit_attempt_done(fp); 2695 if (!fp->jited && jit_needed) { 2696 *err = -ENOTSUPP; 2697 return fp; 2698 } 2699 } else { 2700 *err = bpf_prog_offload_compile(fp); 2701 if (*err) 2702 return fp; 2703 } 2704 2705 finalize: 2706 *err = bpf_prog_lock_ro(fp); 2707 if (*err) 2708 return fp; 2709 2710 /* The tail call compatibility check can only be done at 2711 * this late stage as we need to determine, if we deal 2712 * with JITed or non JITed program concatenations and not 2713 * all eBPF JITs might immediately support all features. 2714 */ 2715 *err = bpf_check_tail_call(fp); 2716 2717 return fp; 2718 } 2719 2720 /** 2721 * bpf_prog_select_runtime - select exec runtime for BPF program 2722 * @fp: bpf_prog populated with BPF program 2723 * @err: pointer to error variable 2724 * 2725 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2726 * The BPF program will be executed via bpf_prog_run() function. 2727 * 2728 * Return: the &fp argument along with &err set to 0 for success or 2729 * a negative errno code on failure 2730 */ 2731 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2732 { 2733 return __bpf_prog_select_runtime(NULL, fp, err); 2734 } 2735 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2736 2737 static unsigned int __bpf_prog_ret1(const void *ctx, 2738 const struct bpf_insn *insn) 2739 { 2740 return 1; 2741 } 2742 2743 static struct bpf_prog_dummy { 2744 struct bpf_prog prog; 2745 } dummy_bpf_prog = { 2746 .prog = { 2747 .bpf_func = __bpf_prog_ret1, 2748 }, 2749 }; 2750 2751 struct bpf_prog_array bpf_empty_prog_array = { 2752 .items = { 2753 { .prog = NULL }, 2754 }, 2755 }; 2756 EXPORT_SYMBOL(bpf_empty_prog_array); 2757 2758 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2759 { 2760 struct bpf_prog_array *p; 2761 2762 if (prog_cnt) 2763 p = kzalloc_flex(*p, items, prog_cnt + 1, flags); 2764 else 2765 p = &bpf_empty_prog_array; 2766 2767 return p; 2768 } 2769 2770 void bpf_prog_array_free(struct bpf_prog_array *progs) 2771 { 2772 if (!progs || progs == &bpf_empty_prog_array) 2773 return; 2774 kfree_rcu(progs, rcu); 2775 } 2776 2777 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2778 { 2779 struct bpf_prog_array *progs; 2780 2781 /* 2782 * RCU Tasks Trace grace period implies RCU grace period, there is no 2783 * need to call kfree_rcu(), just call kfree() directly. 2784 */ 2785 progs = container_of(rcu, struct bpf_prog_array, rcu); 2786 kfree(progs); 2787 } 2788 2789 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2790 { 2791 if (!progs || progs == &bpf_empty_prog_array) 2792 return; 2793 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2794 } 2795 2796 int bpf_prog_array_length(struct bpf_prog_array *array) 2797 { 2798 struct bpf_prog_array_item *item; 2799 u32 cnt = 0; 2800 2801 for (item = array->items; item->prog; item++) 2802 if (item->prog != &dummy_bpf_prog.prog) 2803 cnt++; 2804 return cnt; 2805 } 2806 2807 bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2808 { 2809 struct bpf_prog_array_item *item; 2810 2811 for (item = array->items; item->prog; item++) 2812 if (item->prog != &dummy_bpf_prog.prog) 2813 return false; 2814 return true; 2815 } 2816 2817 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2818 u32 *prog_ids, 2819 u32 request_cnt) 2820 { 2821 struct bpf_prog_array_item *item; 2822 int i = 0; 2823 2824 for (item = array->items; item->prog; item++) { 2825 if (item->prog == &dummy_bpf_prog.prog) 2826 continue; 2827 prog_ids[i] = item->prog->aux->id; 2828 if (++i == request_cnt) { 2829 item++; 2830 break; 2831 } 2832 } 2833 2834 return !!(item->prog); 2835 } 2836 2837 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2838 __u32 __user *prog_ids, u32 cnt) 2839 { 2840 unsigned long err = 0; 2841 bool nospc; 2842 u32 *ids; 2843 2844 /* users of this function are doing: 2845 * cnt = bpf_prog_array_length(); 2846 * if (cnt > 0) 2847 * bpf_prog_array_copy_to_user(..., cnt); 2848 * so below kcalloc doesn't need extra cnt > 0 check. 2849 */ 2850 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2851 if (!ids) 2852 return -ENOMEM; 2853 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2854 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2855 kfree(ids); 2856 if (err) 2857 return -EFAULT; 2858 if (nospc) 2859 return -ENOSPC; 2860 return 0; 2861 } 2862 2863 void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2864 struct bpf_prog *old_prog) 2865 { 2866 struct bpf_prog_array_item *item; 2867 2868 for (item = array->items; item->prog; item++) 2869 if (item->prog == old_prog) { 2870 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2871 break; 2872 } 2873 } 2874 2875 /** 2876 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2877 * index into the program array with 2878 * a dummy no-op program. 2879 * @array: a bpf_prog_array 2880 * @index: the index of the program to replace 2881 * 2882 * Skips over dummy programs, by not counting them, when calculating 2883 * the position of the program to replace. 2884 * 2885 * Return: 2886 * * 0 - Success 2887 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2888 * * -ENOENT - Index out of range 2889 */ 2890 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2891 { 2892 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2893 } 2894 2895 /** 2896 * bpf_prog_array_update_at() - Updates the program at the given index 2897 * into the program array. 2898 * @array: a bpf_prog_array 2899 * @index: the index of the program to update 2900 * @prog: the program to insert into the array 2901 * 2902 * Skips over dummy programs, by not counting them, when calculating 2903 * the position of the program to update. 2904 * 2905 * Return: 2906 * * 0 - Success 2907 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2908 * * -ENOENT - Index out of range 2909 */ 2910 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2911 struct bpf_prog *prog) 2912 { 2913 struct bpf_prog_array_item *item; 2914 2915 if (unlikely(index < 0)) 2916 return -EINVAL; 2917 2918 for (item = array->items; item->prog; item++) { 2919 if (item->prog == &dummy_bpf_prog.prog) 2920 continue; 2921 if (!index) { 2922 WRITE_ONCE(item->prog, prog); 2923 return 0; 2924 } 2925 index--; 2926 } 2927 return -ENOENT; 2928 } 2929 2930 int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2931 struct bpf_prog *exclude_prog, 2932 struct bpf_prog *include_prog, 2933 u64 bpf_cookie, 2934 struct bpf_prog_array **new_array) 2935 { 2936 int new_prog_cnt, carry_prog_cnt = 0; 2937 struct bpf_prog_array_item *existing, *new; 2938 struct bpf_prog_array *array; 2939 bool found_exclude = false; 2940 2941 /* Figure out how many existing progs we need to carry over to 2942 * the new array. 2943 */ 2944 if (old_array) { 2945 existing = old_array->items; 2946 for (; existing->prog; existing++) { 2947 if (existing->prog == exclude_prog) { 2948 found_exclude = true; 2949 continue; 2950 } 2951 if (existing->prog != &dummy_bpf_prog.prog) 2952 carry_prog_cnt++; 2953 if (existing->prog == include_prog) 2954 return -EEXIST; 2955 } 2956 } 2957 2958 if (exclude_prog && !found_exclude) 2959 return -ENOENT; 2960 2961 /* How many progs (not NULL) will be in the new array? */ 2962 new_prog_cnt = carry_prog_cnt; 2963 if (include_prog) 2964 new_prog_cnt += 1; 2965 2966 /* Do we have any prog (not NULL) in the new array? */ 2967 if (!new_prog_cnt) { 2968 *new_array = NULL; 2969 return 0; 2970 } 2971 2972 /* +1 as the end of prog_array is marked with NULL */ 2973 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2974 if (!array) 2975 return -ENOMEM; 2976 new = array->items; 2977 2978 /* Fill in the new prog array */ 2979 if (carry_prog_cnt) { 2980 existing = old_array->items; 2981 for (; existing->prog; existing++) { 2982 if (existing->prog == exclude_prog || 2983 existing->prog == &dummy_bpf_prog.prog) 2984 continue; 2985 2986 new->prog = existing->prog; 2987 new->bpf_cookie = existing->bpf_cookie; 2988 new++; 2989 } 2990 } 2991 if (include_prog) { 2992 new->prog = include_prog; 2993 new->bpf_cookie = bpf_cookie; 2994 new++; 2995 } 2996 new->prog = NULL; 2997 *new_array = array; 2998 return 0; 2999 } 3000 3001 int bpf_prog_array_copy_info(struct bpf_prog_array *array, 3002 u32 *prog_ids, u32 request_cnt, 3003 u32 *prog_cnt) 3004 { 3005 u32 cnt = 0; 3006 3007 if (array) 3008 cnt = bpf_prog_array_length(array); 3009 3010 *prog_cnt = cnt; 3011 3012 /* return early if user requested only program count or nothing to copy */ 3013 if (!request_cnt || !cnt) 3014 return 0; 3015 3016 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 3017 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 3018 : 0; 3019 } 3020 3021 void __bpf_free_used_maps(struct bpf_prog_aux *aux, 3022 struct bpf_map **used_maps, u32 len) 3023 { 3024 struct bpf_map *map; 3025 bool sleepable; 3026 u32 i; 3027 3028 sleepable = aux->prog->sleepable; 3029 for (i = 0; i < len; i++) { 3030 map = used_maps[i]; 3031 if (map->ops->map_poke_untrack) 3032 map->ops->map_poke_untrack(map, aux); 3033 if (sleepable) 3034 atomic64_dec(&map->sleepable_refcnt); 3035 bpf_map_put(map); 3036 } 3037 } 3038 3039 static void bpf_free_used_maps(struct bpf_prog_aux *aux) 3040 { 3041 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 3042 kfree(aux->used_maps); 3043 } 3044 3045 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 3046 { 3047 #ifdef CONFIG_BPF_SYSCALL 3048 struct btf_mod_pair *btf_mod; 3049 u32 i; 3050 3051 for (i = 0; i < len; i++) { 3052 btf_mod = &used_btfs[i]; 3053 if (btf_mod->module) 3054 module_put(btf_mod->module); 3055 btf_put(btf_mod->btf); 3056 } 3057 #endif 3058 } 3059 3060 static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 3061 { 3062 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 3063 kfree(aux->used_btfs); 3064 } 3065 3066 static void bpf_prog_free_deferred(struct work_struct *work) 3067 { 3068 struct bpf_prog_aux *aux; 3069 int i; 3070 3071 aux = container_of(work, struct bpf_prog_aux, work); 3072 #ifdef CONFIG_BPF_SYSCALL 3073 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 3074 bpf_prog_stream_free(aux->prog); 3075 #endif 3076 #ifdef CONFIG_CGROUP_BPF 3077 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 3078 bpf_cgroup_atype_put(aux->cgroup_atype); 3079 #endif 3080 bpf_free_used_maps(aux); 3081 bpf_free_used_btfs(aux); 3082 bpf_prog_disassoc_struct_ops(aux->prog); 3083 if (bpf_prog_is_dev_bound(aux)) 3084 bpf_prog_dev_bound_destroy(aux->prog); 3085 #ifdef CONFIG_PERF_EVENTS 3086 if (aux->prog->has_callchain_buf) 3087 put_callchain_buffers(); 3088 #endif 3089 if (aux->dst_trampoline) 3090 bpf_trampoline_put(aux->dst_trampoline); 3091 for (i = 0; i < aux->real_func_cnt; i++) { 3092 /* We can just unlink the subprog poke descriptor table as 3093 * it was originally linked to the main program and is also 3094 * released along with it. 3095 */ 3096 aux->func[i]->aux->poke_tab = NULL; 3097 bpf_jit_free(aux->func[i]); 3098 } 3099 if (aux->real_func_cnt) { 3100 kfree(aux->func); 3101 bpf_prog_unlock_free(aux->prog); 3102 } else { 3103 bpf_jit_free(aux->prog); 3104 } 3105 } 3106 3107 void bpf_prog_free(struct bpf_prog *fp) 3108 { 3109 struct bpf_prog_aux *aux = fp->aux; 3110 3111 if (aux->dst_prog) 3112 bpf_prog_put(aux->dst_prog); 3113 bpf_token_put(aux->token); 3114 INIT_WORK(&aux->work, bpf_prog_free_deferred); 3115 schedule_work(&aux->work); 3116 } 3117 EXPORT_SYMBOL_GPL(bpf_prog_free); 3118 3119 /* RNG for unprivileged user space with separated state from prandom_u32(). */ 3120 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 3121 3122 void bpf_user_rnd_init_once(void) 3123 { 3124 prandom_init_once(&bpf_user_rnd_state); 3125 } 3126 3127 BPF_CALL_0(bpf_user_rnd_u32) 3128 { 3129 /* Should someone ever have the rather unwise idea to use some 3130 * of the registers passed into this function, then note that 3131 * this function is called from native eBPF and classic-to-eBPF 3132 * transformations. Register assignments from both sides are 3133 * different, f.e. classic always sets fn(ctx, A, X) here. 3134 */ 3135 struct rnd_state *state; 3136 u32 res; 3137 3138 state = &get_cpu_var(bpf_user_rnd_state); 3139 res = prandom_u32_state(state); 3140 put_cpu_var(bpf_user_rnd_state); 3141 3142 return res; 3143 } 3144 3145 BPF_CALL_0(bpf_get_raw_cpu_id) 3146 { 3147 return raw_smp_processor_id(); 3148 } 3149 3150 /* Weak definitions of helper functions in case we don't have bpf syscall. */ 3151 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 3152 const struct bpf_func_proto bpf_map_update_elem_proto __weak; 3153 const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 3154 const struct bpf_func_proto bpf_map_push_elem_proto __weak; 3155 const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 3156 const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 3157 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 3158 const struct bpf_func_proto bpf_spin_lock_proto __weak; 3159 const struct bpf_func_proto bpf_spin_unlock_proto __weak; 3160 const struct bpf_func_proto bpf_jiffies64_proto __weak; 3161 3162 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3163 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3164 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3165 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3166 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3167 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3168 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3169 3170 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3171 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3172 const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3173 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3174 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3175 const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3176 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3177 const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3178 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3179 const struct bpf_func_proto bpf_set_retval_proto __weak; 3180 const struct bpf_func_proto bpf_get_retval_proto __weak; 3181 3182 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3183 { 3184 return NULL; 3185 } 3186 3187 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3188 { 3189 return NULL; 3190 } 3191 3192 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3193 { 3194 return NULL; 3195 } 3196 3197 u64 __weak 3198 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3199 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3200 { 3201 return -ENOTSUPP; 3202 } 3203 EXPORT_SYMBOL_GPL(bpf_event_output); 3204 3205 /* Always built-in helper functions. */ 3206 const struct bpf_func_proto bpf_tail_call_proto = { 3207 /* func is unused for tail_call, we set it to pass the 3208 * get_helper_proto check 3209 */ 3210 .func = BPF_PTR_POISON, 3211 .gpl_only = false, 3212 .ret_type = RET_VOID, 3213 .arg1_type = ARG_PTR_TO_CTX, 3214 .arg2_type = ARG_CONST_MAP_PTR, 3215 .arg3_type = ARG_ANYTHING, 3216 }; 3217 3218 /* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3219 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3220 * eBPF and implicitly also cBPF can get JITed! 3221 */ 3222 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog) 3223 { 3224 return prog; 3225 } 3226 3227 /* Stub for JITs that support eBPF. All cBPF code gets transformed into 3228 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3229 */ 3230 void __weak bpf_jit_compile(struct bpf_prog *prog) 3231 { 3232 } 3233 3234 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3235 { 3236 return false; 3237 } 3238 3239 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3240 * analysis code and wants explicit zero extension inserted by verifier. 3241 * Otherwise, return FALSE. 3242 * 3243 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3244 * you don't override this. JITs that don't want these extra insns can detect 3245 * them using insn_is_zext. 3246 */ 3247 bool __weak bpf_jit_needs_zext(void) 3248 { 3249 return false; 3250 } 3251 3252 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3253 * all archs. The value returned must not change at runtime as there is 3254 * currently no support for reloading programs that were loaded without 3255 * mitigations. 3256 */ 3257 bool __weak bpf_jit_bypass_spec_v1(void) 3258 { 3259 return false; 3260 } 3261 3262 bool __weak bpf_jit_bypass_spec_v4(void) 3263 { 3264 return false; 3265 } 3266 3267 /* Return true if the JIT inlines the call to the helper corresponding to 3268 * the imm. 3269 * 3270 * The verifier will not patch the insn->imm for the call to the helper if 3271 * this returns true. 3272 */ 3273 bool __weak bpf_jit_inlines_helper_call(s32 imm) 3274 { 3275 return false; 3276 } 3277 3278 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3279 bool __weak bpf_jit_supports_subprog_tailcalls(void) 3280 { 3281 return false; 3282 } 3283 3284 bool __weak bpf_jit_supports_percpu_insn(void) 3285 { 3286 return false; 3287 } 3288 3289 bool __weak bpf_jit_supports_kfunc_call(void) 3290 { 3291 return false; 3292 } 3293 3294 bool __weak bpf_jit_supports_stack_args(void) 3295 { 3296 return false; 3297 } 3298 3299 bool __weak bpf_jit_supports_far_kfunc_call(void) 3300 { 3301 return false; 3302 } 3303 3304 bool __weak bpf_jit_supports_arena(void) 3305 { 3306 return false; 3307 } 3308 3309 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3310 { 3311 return false; 3312 } 3313 3314 bool __weak bpf_jit_supports_fsession(void) 3315 { 3316 return false; 3317 } 3318 3319 u64 __weak bpf_arch_uaddress_limit(void) 3320 { 3321 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3322 return TASK_SIZE; 3323 #else 3324 return 0; 3325 #endif 3326 } 3327 3328 /* Return TRUE if the JIT backend satisfies the following two conditions: 3329 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3330 * 2) Under the specific arch, the implementation of xchg() is the same 3331 * as atomic_xchg() on pointer-sized words. 3332 */ 3333 bool __weak bpf_jit_supports_ptr_xchg(void) 3334 { 3335 return false; 3336 } 3337 3338 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3339 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3340 */ 3341 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3342 int len) 3343 { 3344 return -EFAULT; 3345 } 3346 3347 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3348 enum bpf_text_poke_type new_t, void *old_addr, 3349 void *new_addr) 3350 { 3351 return -ENOTSUPP; 3352 } 3353 3354 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3355 { 3356 return ERR_PTR(-ENOTSUPP); 3357 } 3358 3359 int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3360 { 3361 return -ENOTSUPP; 3362 } 3363 3364 bool __weak bpf_jit_supports_exceptions(void) 3365 { 3366 return false; 3367 } 3368 3369 bool __weak bpf_jit_supports_private_stack(void) 3370 { 3371 return false; 3372 } 3373 3374 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3375 { 3376 } 3377 3378 bool __weak bpf_jit_supports_timed_may_goto(void) 3379 { 3380 return false; 3381 } 3382 3383 u64 __weak arch_bpf_timed_may_goto(void) 3384 { 3385 return 0; 3386 } 3387 3388 static noinline void bpf_prog_report_may_goto_violation(void) 3389 { 3390 #ifdef CONFIG_BPF_SYSCALL 3391 struct bpf_stream_stage ss; 3392 struct bpf_prog *prog; 3393 3394 prog = bpf_prog_find_from_stack(); 3395 if (!prog) 3396 return; 3397 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3398 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3399 bpf_stream_dump_stack(ss); 3400 })); 3401 #endif 3402 } 3403 3404 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3405 { 3406 u64 time = ktime_get_mono_fast_ns(); 3407 3408 /* Populate the timestamp for this stack frame, and refresh count. */ 3409 if (!p->timestamp) { 3410 p->timestamp = time; 3411 return BPF_MAX_TIMED_LOOPS; 3412 } 3413 /* Check if we've exhausted our time slice, and zero count. */ 3414 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3415 bpf_prog_report_may_goto_violation(); 3416 return 0; 3417 } 3418 /* Refresh the count for the stack frame. */ 3419 return BPF_MAX_TIMED_LOOPS; 3420 } 3421 3422 /* for configs without MMU or 32-bit */ 3423 __weak const struct bpf_map_ops arena_map_ops; 3424 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3425 { 3426 return 0; 3427 } 3428 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3429 { 3430 return 0; 3431 } 3432 3433 #ifdef CONFIG_BPF_SYSCALL 3434 __weak bool bpf_arena_handle_page_fault(unsigned long addr, bool is_write, 3435 unsigned long fault_ip) 3436 { 3437 return false; 3438 } 3439 3440 static int __init bpf_global_ma_init(void) 3441 { 3442 int ret; 3443 3444 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3445 bpf_global_ma_set = !ret; 3446 return ret; 3447 } 3448 late_initcall(bpf_global_ma_init); 3449 #endif 3450 3451 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3452 EXPORT_SYMBOL(bpf_stats_enabled_key); 3453 3454 /* All definitions of tracepoints related to BPF. */ 3455 #define CREATE_TRACE_POINTS 3456 #include <linux/bpf_trace.h> 3457 3458 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3459 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3460 3461 #ifdef CONFIG_BPF_SYSCALL 3462 3463 void bpf_get_linfo_file_line(struct btf *btf, const struct bpf_line_info *linfo, 3464 const char **filep, const char **linep, int *nump) 3465 { 3466 /* Get base component of the file path. */ 3467 if (filep) { 3468 *filep = btf_name_by_offset(btf, linfo->file_name_off); 3469 *filep = kbasename(*filep); 3470 } 3471 3472 /* Obtain the source line, and strip whitespace in prefix. */ 3473 if (linep) { 3474 *linep = btf_name_by_offset(btf, linfo->line_off); 3475 while (isspace(**linep)) 3476 *linep += 1; 3477 } 3478 3479 if (nump) 3480 *nump = BPF_LINE_INFO_LINE_NUM(linfo->line_col); 3481 } 3482 3483 const struct bpf_line_info *bpf_find_linfo(const struct bpf_prog *prog, u32 insn_off) 3484 { 3485 const struct bpf_line_info *linfo; 3486 u32 nr_linfo; 3487 int l, r, m; 3488 3489 nr_linfo = prog->aux->nr_linfo; 3490 if (!nr_linfo || insn_off >= prog->len) 3491 return NULL; 3492 3493 linfo = prog->aux->linfo; 3494 /* Loop invariant: linfo[l].insn_off <= insns_off. 3495 * linfo[0].insn_off == 0 which always satisfies above condition. 3496 * Binary search is searching for rightmost linfo entry that satisfies 3497 * the above invariant, giving us the desired record that covers given 3498 * instruction offset. 3499 */ 3500 l = 0; 3501 r = nr_linfo - 1; 3502 while (l < r) { 3503 /* (r - l + 1) / 2 means we break a tie to the right, so if: 3504 * l=1, r=2, linfo[l].insn_off <= insn_off, linfo[r].insn_off > insn_off, 3505 * then m=2, we see that linfo[m].insn_off > insn_off, and so 3506 * r becomes 1 and we exit the loop with correct l==1. 3507 * If the tie was broken to the left, m=1 would end us up in 3508 * an endless loop where l and m stay at 1 and r stays at 2. 3509 */ 3510 m = l + (r - l + 1) / 2; 3511 if (linfo[m].insn_off <= insn_off) 3512 l = m; 3513 else 3514 r = m - 1; 3515 } 3516 3517 return &linfo[l]; 3518 } 3519 3520 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3521 const char **linep, int *nump) 3522 { 3523 int idx = -1, insn_start, insn_end, len; 3524 struct bpf_line_info *linfo; 3525 void **jited_linfo; 3526 struct btf *btf; 3527 int nr_linfo; 3528 3529 btf = prog->aux->btf; 3530 linfo = prog->aux->linfo; 3531 jited_linfo = prog->aux->jited_linfo; 3532 3533 if (!btf || !linfo || !jited_linfo) 3534 return -EINVAL; 3535 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3536 3537 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3538 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3539 3540 insn_start = linfo[0].insn_off; 3541 insn_end = insn_start + len; 3542 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3543 3544 for (int i = 0; i < nr_linfo && 3545 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3546 if (jited_linfo[i] >= (void *)ip) 3547 break; 3548 idx = i; 3549 } 3550 3551 if (idx == -1) 3552 return -ENOENT; 3553 3554 bpf_get_linfo_file_line(btf, &linfo[idx], filep, linep, nump); 3555 return 0; 3556 } 3557 3558 struct walk_stack_ctx { 3559 struct bpf_prog *prog; 3560 }; 3561 3562 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3563 { 3564 struct walk_stack_ctx *ctxp = cookie; 3565 struct bpf_prog *prog; 3566 3567 /* 3568 * The RCU read lock is held to safely traverse the latch tree, but we 3569 * don't need its protection when accessing the prog, since it has an 3570 * active stack frame on the current stack trace, and won't disappear. 3571 */ 3572 rcu_read_lock(); 3573 prog = bpf_prog_ksym_find(ip); 3574 rcu_read_unlock(); 3575 if (!prog) 3576 return true; 3577 /* Make sure we return the main prog if we found a subprog */ 3578 ctxp->prog = prog->aux->main_prog_aux->prog; 3579 return false; 3580 } 3581 3582 struct bpf_prog *bpf_prog_find_from_stack(void) 3583 { 3584 struct walk_stack_ctx ctx = {}; 3585 3586 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3587 return ctx.prog; 3588 } 3589 3590 #endif 3591