FILE
core.c
File preview with syntax highlighting.
linux-6.19.14/kernel/bpf/core.c
1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * Linux Socket Filter - Kernel level socket filtering 4 * 5 * Based on the design of the Berkeley Packet Filter. The new 6 * internal format has been designed by PLUMgrid: 7 * 8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com 9 * 10 * Authors: 11 * 12 * Jay Schulist <jschlst@samba.org> 13 * Alexei Starovoitov <ast@plumgrid.com> 14 * Daniel Borkmann <dborkman@redhat.com> 15 * 16 * Andi Kleen - Fix a few bad bugs and races. 17 * Kris Katterjohn - Added many additional checks in bpf_check_classic() 18 */ 19 20#include <uapi/linux/btf.h> 21#include <crypto/sha1.h> 22#include <linux/filter.h> 23#include <linux/skbuff.h> 24#include <linux/vmalloc.h> 25#include <linux/prandom.h> 26#include <linux/bpf.h> 27#include <linux/btf.h> 28#include <linux/objtool.h> 29#include <linux/overflow.h> 30#include <linux/rbtree_latch.h> 31#include <linux/kallsyms.h> 32#include <linux/rcupdate.h> 33#include <linux/perf_event.h> 34#include <linux/extable.h> 35#include <linux/log2.h> 36#include <linux/bpf_verifier.h> 37#include <linux/nodemask.h> 38#include <linux/nospec.h> 39#include <linux/bpf_mem_alloc.h> 40#include <linux/memcontrol.h> 41#include <linux/execmem.h> 42#include <crypto/sha2.h> 43 44#include <asm/barrier.h> 45#include <linux/unaligned.h> 46 47/* Registers */ 48#define BPF_R0 regs[BPF_REG_0] 49#define BPF_R1 regs[BPF_REG_1] 50#define BPF_R2 regs[BPF_REG_2] 51#define BPF_R3 regs[BPF_REG_3] 52#define BPF_R4 regs[BPF_REG_4] 53#define BPF_R5 regs[BPF_REG_5] 54#define BPF_R6 regs[BPF_REG_6] 55#define BPF_R7 regs[BPF_REG_7] 56#define BPF_R8 regs[BPF_REG_8] 57#define BPF_R9 regs[BPF_REG_9] 58#define BPF_R10 regs[BPF_REG_10] 59 60/* Named registers */ 61#define DST regs[insn->dst_reg] 62#define SRC regs[insn->src_reg] 63#define FP regs[BPF_REG_FP] 64#define AX regs[BPF_REG_AX] 65#define ARG1 regs[BPF_REG_ARG1] 66#define CTX regs[BPF_REG_CTX] 67#define OFF insn->off 68#define IMM insn->imm 69 70struct bpf_mem_alloc bpf_global_ma; 71bool bpf_global_ma_set; 72 73/* No hurry in this branch 74 * 75 * Exported for the bpf jit load helper. 76 */ 77void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size) 78{ 79 u8 *ptr = NULL; 80 81 if (k >= SKF_NET_OFF) { 82 ptr = skb_network_header(skb) + k - SKF_NET_OFF; 83 } else if (k >= SKF_LL_OFF) { 84 if (unlikely(!skb_mac_header_was_set(skb))) 85 return NULL; 86 ptr = skb_mac_header(skb) + k - SKF_LL_OFF; 87 } 88 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb)) 89 return ptr; 90 91 return NULL; 92} 93 94/* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */ 95enum page_size_enum { 96 __PAGE_SIZE = PAGE_SIZE 97}; 98 99struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags) 100{ 101 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 102 struct bpf_prog_aux *aux; 103 struct bpf_prog *fp; 104 105 size = round_up(size, __PAGE_SIZE); 106 fp = __vmalloc(size, gfp_flags); 107 if (fp == NULL) 108 return NULL; 109 110 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 111 if (aux == NULL) { 112 vfree(fp); 113 return NULL; 114 } 115 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags)); 116 if (!fp->active) { 117 vfree(fp); 118 kfree(aux); 119 return NULL; 120 } 121 122 fp->pages = size / PAGE_SIZE; 123 fp->aux = aux; 124 fp->aux->main_prog_aux = aux; 125 fp->aux->prog = fp; 126 fp->jit_requested = ebpf_jit_enabled(); 127 fp->blinding_requested = bpf_jit_blinding_enabled(fp); 128#ifdef CONFIG_CGROUP_BPF 129 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID; 130#endif 131 132 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode); 133#ifdef CONFIG_FINEIBT 134 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode); 135#endif 136 mutex_init(&fp->aux->used_maps_mutex); 137 mutex_init(&fp->aux->ext_mutex); 138 mutex_init(&fp->aux->dst_mutex); 139 140#ifdef CONFIG_BPF_SYSCALL 141 bpf_prog_stream_init(fp); 142#endif 143 144 return fp; 145} 146 147struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags) 148{ 149 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 150 struct bpf_prog *prog; 151 int cpu; 152 153 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags); 154 if (!prog) 155 return NULL; 156 157 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags); 158 if (!prog->stats) { 159 free_percpu(prog->active); 160 kfree(prog->aux); 161 vfree(prog); 162 return NULL; 163 } 164 165 for_each_possible_cpu(cpu) { 166 struct bpf_prog_stats *pstats; 167 168 pstats = per_cpu_ptr(prog->stats, cpu); 169 u64_stats_init(&pstats->syncp); 170 } 171 return prog; 172} 173EXPORT_SYMBOL_GPL(bpf_prog_alloc); 174 175int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog) 176{ 177 if (!prog->aux->nr_linfo || !prog->jit_requested) 178 return 0; 179 180 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo, 181 sizeof(*prog->aux->jited_linfo), 182 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN)); 183 if (!prog->aux->jited_linfo) 184 return -ENOMEM; 185 186 return 0; 187} 188 189void bpf_prog_jit_attempt_done(struct bpf_prog *prog) 190{ 191 if (prog->aux->jited_linfo && 192 (!prog->jited || !prog->aux->jited_linfo[0])) { 193 kvfree(prog->aux->jited_linfo); 194 prog->aux->jited_linfo = NULL; 195 } 196 197 kfree(prog->aux->kfunc_tab); 198 prog->aux->kfunc_tab = NULL; 199} 200 201/* The jit engine is responsible to provide an array 202 * for insn_off to the jited_off mapping (insn_to_jit_off). 203 * 204 * The idx to this array is the insn_off. Hence, the insn_off 205 * here is relative to the prog itself instead of the main prog. 206 * This array has one entry for each xlated bpf insn. 207 * 208 * jited_off is the byte off to the end of the jited insn. 209 * 210 * Hence, with 211 * insn_start: 212 * The first bpf insn off of the prog. The insn off 213 * here is relative to the main prog. 214 * e.g. if prog is a subprog, insn_start > 0 215 * linfo_idx: 216 * The prog's idx to prog->aux->linfo and jited_linfo 217 * 218 * jited_linfo[linfo_idx] = prog->bpf_func 219 * 220 * For i > linfo_idx, 221 * 222 * jited_linfo[i] = prog->bpf_func + 223 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1] 224 */ 225void bpf_prog_fill_jited_linfo(struct bpf_prog *prog, 226 const u32 *insn_to_jit_off) 227{ 228 u32 linfo_idx, insn_start, insn_end, nr_linfo, i; 229 const struct bpf_line_info *linfo; 230 void **jited_linfo; 231 232 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt) 233 /* Userspace did not provide linfo */ 234 return; 235 236 linfo_idx = prog->aux->linfo_idx; 237 linfo = &prog->aux->linfo[linfo_idx]; 238 insn_start = linfo[0].insn_off; 239 insn_end = insn_start + prog->len; 240 241 jited_linfo = &prog->aux->jited_linfo[linfo_idx]; 242 jited_linfo[0] = prog->bpf_func; 243 244 nr_linfo = prog->aux->nr_linfo - linfo_idx; 245 246 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++) 247 /* The verifier ensures that linfo[i].insn_off is 248 * strictly increasing 249 */ 250 jited_linfo[i] = prog->bpf_func + 251 insn_to_jit_off[linfo[i].insn_off - insn_start - 1]; 252} 253 254struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size, 255 gfp_t gfp_extra_flags) 256{ 257 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags); 258 struct bpf_prog *fp; 259 u32 pages; 260 261 size = round_up(size, PAGE_SIZE); 262 pages = size / PAGE_SIZE; 263 if (pages <= fp_old->pages) 264 return fp_old; 265 266 fp = __vmalloc(size, gfp_flags); 267 if (fp) { 268 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE); 269 fp->pages = pages; 270 fp->aux->prog = fp; 271 272 /* We keep fp->aux from fp_old around in the new 273 * reallocated structure. 274 */ 275 fp_old->aux = NULL; 276 fp_old->stats = NULL; 277 fp_old->active = NULL; 278 __bpf_prog_free(fp_old); 279 } 280 281 return fp; 282} 283 284void __bpf_prog_free(struct bpf_prog *fp) 285{ 286 if (fp->aux) { 287 mutex_destroy(&fp->aux->used_maps_mutex); 288 mutex_destroy(&fp->aux->dst_mutex); 289 kfree(fp->aux->poke_tab); 290 kfree(fp->aux); 291 } 292 free_percpu(fp->stats); 293 free_percpu(fp->active); 294 vfree(fp); 295} 296 297int bpf_prog_calc_tag(struct bpf_prog *fp) 298{ 299 size_t size = bpf_prog_insn_size(fp); 300 struct bpf_insn *dst; 301 bool was_ld_map; 302 u32 i; 303 304 dst = vmalloc(size); 305 if (!dst) 306 return -ENOMEM; 307 308 /* We need to take out the map fd for the digest calculation 309 * since they are unstable from user space side. 310 */ 311 for (i = 0, was_ld_map = false; i < fp->len; i++) { 312 dst[i] = fp->insnsi[i]; 313 if (!was_ld_map && 314 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) && 315 (dst[i].src_reg == BPF_PSEUDO_MAP_FD || 316 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) { 317 was_ld_map = true; 318 dst[i].imm = 0; 319 } else if (was_ld_map && 320 dst[i].code == 0 && 321 dst[i].dst_reg == 0 && 322 dst[i].src_reg == 0 && 323 dst[i].off == 0) { 324 was_ld_map = false; 325 dst[i].imm = 0; 326 } else { 327 was_ld_map = false; 328 } 329 } 330 sha256((u8 *)dst, size, fp->digest); 331 vfree(dst); 332 return 0; 333} 334 335static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old, 336 s32 end_new, s32 curr, const bool probe_pass) 337{ 338 const s64 imm_min = S32_MIN, imm_max = S32_MAX; 339 s32 delta = end_new - end_old; 340 s64 imm = insn->imm; 341 342 if (curr < pos && curr + imm + 1 >= end_old) 343 imm += delta; 344 else if (curr >= end_new && curr + imm + 1 < end_new) 345 imm -= delta; 346 if (imm < imm_min || imm > imm_max) 347 return -ERANGE; 348 if (!probe_pass) 349 insn->imm = imm; 350 return 0; 351} 352 353static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old, 354 s32 end_new, s32 curr, const bool probe_pass) 355{ 356 s64 off_min, off_max, off; 357 s32 delta = end_new - end_old; 358 359 if (insn->code == (BPF_JMP32 | BPF_JA)) { 360 off = insn->imm; 361 off_min = S32_MIN; 362 off_max = S32_MAX; 363 } else { 364 off = insn->off; 365 off_min = S16_MIN; 366 off_max = S16_MAX; 367 } 368 369 if (curr < pos && curr + off + 1 >= end_old) 370 off += delta; 371 else if (curr >= end_new && curr + off + 1 < end_new) 372 off -= delta; 373 if (off < off_min || off > off_max) 374 return -ERANGE; 375 if (!probe_pass) { 376 if (insn->code == (BPF_JMP32 | BPF_JA)) 377 insn->imm = off; 378 else 379 insn->off = off; 380 } 381 return 0; 382} 383 384static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old, 385 s32 end_new, const bool probe_pass) 386{ 387 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0); 388 struct bpf_insn *insn = prog->insnsi; 389 int ret = 0; 390 391 for (i = 0; i < insn_cnt; i++, insn++) { 392 u8 code; 393 394 /* In the probing pass we still operate on the original, 395 * unpatched image in order to check overflows before we 396 * do any other adjustments. Therefore skip the patchlet. 397 */ 398 if (probe_pass && i == pos) { 399 i = end_new; 400 insn = prog->insnsi + end_old; 401 } 402 if (bpf_pseudo_func(insn)) { 403 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 404 end_new, i, probe_pass); 405 if (ret) 406 return ret; 407 continue; 408 } 409 code = insn->code; 410 if ((BPF_CLASS(code) != BPF_JMP && 411 BPF_CLASS(code) != BPF_JMP32) || 412 BPF_OP(code) == BPF_EXIT) 413 continue; 414 /* Adjust offset of jmps if we cross patch boundaries. */ 415 if (BPF_OP(code) == BPF_CALL) { 416 if (insn->src_reg != BPF_PSEUDO_CALL) 417 continue; 418 ret = bpf_adj_delta_to_imm(insn, pos, end_old, 419 end_new, i, probe_pass); 420 } else { 421 ret = bpf_adj_delta_to_off(insn, pos, end_old, 422 end_new, i, probe_pass); 423 } 424 if (ret) 425 break; 426 } 427 428 return ret; 429} 430 431static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta) 432{ 433 struct bpf_line_info *linfo; 434 u32 i, nr_linfo; 435 436 nr_linfo = prog->aux->nr_linfo; 437 if (!nr_linfo || !delta) 438 return; 439 440 linfo = prog->aux->linfo; 441 442 for (i = 0; i < nr_linfo; i++) 443 if (off < linfo[i].insn_off) 444 break; 445 446 /* Push all off < linfo[i].insn_off by delta */ 447 for (; i < nr_linfo; i++) 448 linfo[i].insn_off += delta; 449} 450 451struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off, 452 const struct bpf_insn *patch, u32 len) 453{ 454 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1; 455 const u32 cnt_max = S16_MAX; 456 struct bpf_prog *prog_adj; 457 int err; 458 459 /* Since our patchlet doesn't expand the image, we're done. */ 460 if (insn_delta == 0) { 461 memcpy(prog->insnsi + off, patch, sizeof(*patch)); 462 return prog; 463 } 464 465 insn_adj_cnt = prog->len + insn_delta; 466 467 /* Reject anything that would potentially let the insn->off 468 * target overflow when we have excessive program expansions. 469 * We need to probe here before we do any reallocation where 470 * we afterwards may not fail anymore. 471 */ 472 if (insn_adj_cnt > cnt_max && 473 (err = bpf_adj_branches(prog, off, off + 1, off + len, true))) 474 return ERR_PTR(err); 475 476 /* Several new instructions need to be inserted. Make room 477 * for them. Likely, there's no need for a new allocation as 478 * last page could have large enough tailroom. 479 */ 480 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt), 481 GFP_USER); 482 if (!prog_adj) 483 return ERR_PTR(-ENOMEM); 484 485 prog_adj->len = insn_adj_cnt; 486 487 /* Patching happens in 3 steps: 488 * 489 * 1) Move over tail of insnsi from next instruction onwards, 490 * so we can patch the single target insn with one or more 491 * new ones (patching is always from 1 to n insns, n > 0). 492 * 2) Inject new instructions at the target location. 493 * 3) Adjust branch offsets if necessary. 494 */ 495 insn_rest = insn_adj_cnt - off - len; 496 497 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1, 498 sizeof(*patch) * insn_rest); 499 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len); 500 501 /* We are guaranteed to not fail at this point, otherwise 502 * the ship has sailed to reverse to the original state. An 503 * overflow cannot happen at this point. 504 */ 505 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false)); 506 507 bpf_adj_linfo(prog_adj, off, insn_delta); 508 509 return prog_adj; 510} 511 512int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt) 513{ 514 int err; 515 516 /* Branch offsets can't overflow when program is shrinking, no need 517 * to call bpf_adj_branches(..., true) here 518 */ 519 memmove(prog->insnsi + off, prog->insnsi + off + cnt, 520 sizeof(struct bpf_insn) * (prog->len - off - cnt)); 521 prog->len -= cnt; 522 523 err = bpf_adj_branches(prog, off, off + cnt, off, false); 524 WARN_ON_ONCE(err); 525 return err; 526} 527 528static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp) 529{ 530 int i; 531 532 for (i = 0; i < fp->aux->real_func_cnt; i++) 533 bpf_prog_kallsyms_del(fp->aux->func[i]); 534} 535 536void bpf_prog_kallsyms_del_all(struct bpf_prog *fp) 537{ 538 bpf_prog_kallsyms_del_subprogs(fp); 539 bpf_prog_kallsyms_del(fp); 540} 541 542#ifdef CONFIG_BPF_JIT 543/* All BPF JIT sysctl knobs here. */ 544int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 545int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON); 546int bpf_jit_harden __read_mostly; 547long bpf_jit_limit __read_mostly; 548long bpf_jit_limit_max __read_mostly; 549 550static void 551bpf_prog_ksym_set_addr(struct bpf_prog *prog) 552{ 553 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog)); 554 555 prog->aux->ksym.start = (unsigned long) prog->bpf_func; 556 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len; 557} 558 559static void 560bpf_prog_ksym_set_name(struct bpf_prog *prog) 561{ 562 char *sym = prog->aux->ksym.name; 563 const char *end = sym + KSYM_NAME_LEN; 564 const struct btf_type *type; 565 const char *func_name; 566 567 BUILD_BUG_ON(sizeof("bpf_prog_") + 568 sizeof(prog->tag) * 2 + 569 /* name has been null terminated. 570 * We should need +1 for the '_' preceding 571 * the name. However, the null character 572 * is double counted between the name and the 573 * sizeof("bpf_prog_") above, so we omit 574 * the +1 here. 575 */ 576 sizeof(prog->aux->name) > KSYM_NAME_LEN); 577 578 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_"); 579 sym = bin2hex(sym, prog->tag, sizeof(prog->tag)); 580 581 /* prog->aux->name will be ignored if full btf name is available */ 582 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) { 583 type = btf_type_by_id(prog->aux->btf, 584 prog->aux->func_info[prog->aux->func_idx].type_id); 585 func_name = btf_name_by_offset(prog->aux->btf, type->name_off); 586 snprintf(sym, (size_t)(end - sym), "_%s", func_name); 587 return; 588 } 589 590 if (prog->aux->name[0]) 591 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name); 592 else 593 *sym = 0; 594} 595 596static unsigned long bpf_get_ksym_start(struct latch_tree_node *n) 597{ 598 return container_of(n, struct bpf_ksym, tnode)->start; 599} 600 601static __always_inline bool bpf_tree_less(struct latch_tree_node *a, 602 struct latch_tree_node *b) 603{ 604 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b); 605} 606 607static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n) 608{ 609 unsigned long val = (unsigned long)key; 610 const struct bpf_ksym *ksym; 611 612 ksym = container_of(n, struct bpf_ksym, tnode); 613 614 if (val < ksym->start) 615 return -1; 616 /* Ensure that we detect return addresses as part of the program, when 617 * the final instruction is a call for a program part of the stack 618 * trace. Therefore, do val > ksym->end instead of val >= ksym->end. 619 */ 620 if (val > ksym->end) 621 return 1; 622 623 return 0; 624} 625 626static const struct latch_tree_ops bpf_tree_ops = { 627 .less = bpf_tree_less, 628 .comp = bpf_tree_comp, 629}; 630 631static DEFINE_SPINLOCK(bpf_lock); 632static LIST_HEAD(bpf_kallsyms); 633static struct latch_tree_root bpf_tree __cacheline_aligned; 634 635void bpf_ksym_add(struct bpf_ksym *ksym) 636{ 637 spin_lock_bh(&bpf_lock); 638 WARN_ON_ONCE(!list_empty(&ksym->lnode)); 639 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms); 640 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 641 spin_unlock_bh(&bpf_lock); 642} 643 644static void __bpf_ksym_del(struct bpf_ksym *ksym) 645{ 646 if (list_empty(&ksym->lnode)) 647 return; 648 649 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops); 650 list_del_rcu(&ksym->lnode); 651} 652 653void bpf_ksym_del(struct bpf_ksym *ksym) 654{ 655 spin_lock_bh(&bpf_lock); 656 __bpf_ksym_del(ksym); 657 spin_unlock_bh(&bpf_lock); 658} 659 660static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp) 661{ 662 return fp->jited && !bpf_prog_was_classic(fp); 663} 664 665void bpf_prog_kallsyms_add(struct bpf_prog *fp) 666{ 667 if (!bpf_prog_kallsyms_candidate(fp) || 668 !bpf_token_capable(fp->aux->token, CAP_BPF)) 669 return; 670 671 bpf_prog_ksym_set_addr(fp); 672 bpf_prog_ksym_set_name(fp); 673 fp->aux->ksym.prog = true; 674 675 bpf_ksym_add(&fp->aux->ksym); 676 677#ifdef CONFIG_FINEIBT 678 /* 679 * When FineIBT, code in the __cfi_foo() symbols can get executed 680 * and hence unwinder needs help. 681 */ 682 if (cfi_mode != CFI_FINEIBT) 683 return; 684 685 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN, 686 "__cfi_%s", fp->aux->ksym.name); 687 688 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16; 689 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func; 690 691 bpf_ksym_add(&fp->aux->ksym_prefix); 692#endif 693} 694 695void bpf_prog_kallsyms_del(struct bpf_prog *fp) 696{ 697 if (!bpf_prog_kallsyms_candidate(fp)) 698 return; 699 700 bpf_ksym_del(&fp->aux->ksym); 701#ifdef CONFIG_FINEIBT 702 if (cfi_mode != CFI_FINEIBT) 703 return; 704 bpf_ksym_del(&fp->aux->ksym_prefix); 705#endif 706} 707 708static struct bpf_ksym *bpf_ksym_find(unsigned long addr) 709{ 710 struct latch_tree_node *n; 711 712 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops); 713 return n ? container_of(n, struct bpf_ksym, tnode) : NULL; 714} 715 716int bpf_address_lookup(unsigned long addr, unsigned long *size, 717 unsigned long *off, char *sym) 718{ 719 struct bpf_ksym *ksym; 720 int ret = 0; 721 722 rcu_read_lock(); 723 ksym = bpf_ksym_find(addr); 724 if (ksym) { 725 unsigned long symbol_start = ksym->start; 726 unsigned long symbol_end = ksym->end; 727 728 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN); 729 730 if (size) 731 *size = symbol_end - symbol_start; 732 if (off) 733 *off = addr - symbol_start; 734 } 735 rcu_read_unlock(); 736 737 return ret; 738} 739 740bool is_bpf_text_address(unsigned long addr) 741{ 742 bool ret; 743 744 rcu_read_lock(); 745 ret = bpf_ksym_find(addr) != NULL; 746 rcu_read_unlock(); 747 748 return ret; 749} 750 751struct bpf_prog *bpf_prog_ksym_find(unsigned long addr) 752{ 753 struct bpf_ksym *ksym; 754 755 WARN_ON_ONCE(!rcu_read_lock_held()); 756 ksym = bpf_ksym_find(addr); 757 758 return ksym && ksym->prog ? 759 container_of(ksym, struct bpf_prog_aux, ksym)->prog : 760 NULL; 761} 762 763bool bpf_has_frame_pointer(unsigned long ip) 764{ 765 struct bpf_ksym *ksym; 766 unsigned long offset; 767 768 guard(rcu)(); 769 770 ksym = bpf_ksym_find(ip); 771 if (!ksym || !ksym->fp_start || !ksym->fp_end) 772 return false; 773 774 offset = ip - ksym->start; 775 776 return offset >= ksym->fp_start && offset < ksym->fp_end; 777} 778 779const struct exception_table_entry *search_bpf_extables(unsigned long addr) 780{ 781 const struct exception_table_entry *e = NULL; 782 struct bpf_prog *prog; 783 784 rcu_read_lock(); 785 prog = bpf_prog_ksym_find(addr); 786 if (!prog) 787 goto out; 788 if (!prog->aux->num_exentries) 789 goto out; 790 791 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr); 792out: 793 rcu_read_unlock(); 794 return e; 795} 796 797int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type, 798 char *sym) 799{ 800 struct bpf_ksym *ksym; 801 unsigned int it = 0; 802 int ret = -ERANGE; 803 804 if (!bpf_jit_kallsyms_enabled()) 805 return ret; 806 807 rcu_read_lock(); 808 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) { 809 if (it++ != symnum) 810 continue; 811 812 strscpy(sym, ksym->name, KSYM_NAME_LEN); 813 814 *value = ksym->start; 815 *type = BPF_SYM_ELF_TYPE; 816 817 ret = 0; 818 break; 819 } 820 rcu_read_unlock(); 821 822 return ret; 823} 824 825int bpf_jit_add_poke_descriptor(struct bpf_prog *prog, 826 struct bpf_jit_poke_descriptor *poke) 827{ 828 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab; 829 static const u32 poke_tab_max = 1024; 830 u32 slot = prog->aux->size_poke_tab; 831 u32 size = slot + 1; 832 833 if (size > poke_tab_max) 834 return -ENOSPC; 835 if (poke->tailcall_target || poke->tailcall_target_stable || 836 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr) 837 return -EINVAL; 838 839 switch (poke->reason) { 840 case BPF_POKE_REASON_TAIL_CALL: 841 if (!poke->tail_call.map) 842 return -EINVAL; 843 break; 844 default: 845 return -EINVAL; 846 } 847 848 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL); 849 if (!tab) 850 return -ENOMEM; 851 852 memcpy(&tab[slot], poke, sizeof(*poke)); 853 prog->aux->size_poke_tab = size; 854 prog->aux->poke_tab = tab; 855 856 return slot; 857} 858 859/* 860 * BPF program pack allocator. 861 * 862 * Most BPF programs are pretty small. Allocating a hole page for each 863 * program is sometime a waste. Many small bpf program also adds pressure 864 * to instruction TLB. To solve this issue, we introduce a BPF program pack 865 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86) 866 * to host BPF programs. 867 */ 868#define BPF_PROG_CHUNK_SHIFT 6 869#define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT) 870#define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1)) 871 872struct bpf_prog_pack { 873 struct list_head list; 874 void *ptr; 875 unsigned long bitmap[]; 876}; 877 878void bpf_jit_fill_hole_with_zero(void *area, unsigned int size) 879{ 880 memset(area, 0, size); 881} 882 883#define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE) 884 885static DEFINE_MUTEX(pack_mutex); 886static LIST_HEAD(pack_list); 887 888/* PMD_SIZE is not available in some special config, e.g. ARCH=arm with 889 * CONFIG_MMU=n. Use PAGE_SIZE in these cases. 890 */ 891#ifdef PMD_SIZE 892/* PMD_SIZE is really big for some archs. It doesn't make sense to 893 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to 894 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be 895 * greater than or equal to 2MB. 896 */ 897#define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes()) 898#else 899#define BPF_PROG_PACK_SIZE PAGE_SIZE 900#endif 901 902#define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE) 903 904static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns) 905{ 906 struct bpf_prog_pack *pack; 907 int err; 908 909 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)), 910 GFP_KERNEL); 911 if (!pack) 912 return NULL; 913 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE); 914 if (!pack->ptr) 915 goto out; 916 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE); 917 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE); 918 919 set_vm_flush_reset_perms(pack->ptr); 920 err = set_memory_rox((unsigned long)pack->ptr, 921 BPF_PROG_PACK_SIZE / PAGE_SIZE); 922 if (err) 923 goto out; 924 list_add_tail(&pack->list, &pack_list); 925 return pack; 926 927out: 928 bpf_jit_free_exec(pack->ptr); 929 kfree(pack); 930 return NULL; 931} 932 933void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns) 934{ 935 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size); 936 struct bpf_prog_pack *pack; 937 unsigned long pos; 938 void *ptr = NULL; 939 940 mutex_lock(&pack_mutex); 941 if (size > BPF_PROG_PACK_SIZE) { 942 size = round_up(size, PAGE_SIZE); 943 ptr = bpf_jit_alloc_exec(size); 944 if (ptr) { 945 int err; 946 947 bpf_fill_ill_insns(ptr, size); 948 set_vm_flush_reset_perms(ptr); 949 err = set_memory_rox((unsigned long)ptr, 950 size / PAGE_SIZE); 951 if (err) { 952 bpf_jit_free_exec(ptr); 953 ptr = NULL; 954 } 955 } 956 goto out; 957 } 958 list_for_each_entry(pack, &pack_list, list) { 959 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 960 nbits, 0); 961 if (pos < BPF_PROG_CHUNK_COUNT) 962 goto found_free_area; 963 } 964 965 pack = alloc_new_pack(bpf_fill_ill_insns); 966 if (!pack) 967 goto out; 968 969 pos = 0; 970 971found_free_area: 972 bitmap_set(pack->bitmap, pos, nbits); 973 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT); 974 975out: 976 mutex_unlock(&pack_mutex); 977 return ptr; 978} 979 980void bpf_prog_pack_free(void *ptr, u32 size) 981{ 982 struct bpf_prog_pack *pack = NULL, *tmp; 983 unsigned int nbits; 984 unsigned long pos; 985 986 mutex_lock(&pack_mutex); 987 if (size > BPF_PROG_PACK_SIZE) { 988 bpf_jit_free_exec(ptr); 989 goto out; 990 } 991 992 list_for_each_entry(tmp, &pack_list, list) { 993 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) { 994 pack = tmp; 995 break; 996 } 997 } 998 999 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n")) 1000 goto out; 1001 1002 nbits = BPF_PROG_SIZE_TO_NBITS(size); 1003 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT; 1004 1005 WARN_ONCE(bpf_arch_text_invalidate(ptr, size), 1006 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n"); 1007 1008 bitmap_clear(pack->bitmap, pos, nbits); 1009 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0, 1010 BPF_PROG_CHUNK_COUNT, 0) == 0) { 1011 list_del(&pack->list); 1012 bpf_jit_free_exec(pack->ptr); 1013 kfree(pack); 1014 } 1015out: 1016 mutex_unlock(&pack_mutex); 1017} 1018 1019static atomic_long_t bpf_jit_current; 1020 1021/* Can be overridden by an arch's JIT compiler if it has a custom, 1022 * dedicated BPF backend memory area, or if neither of the two 1023 * below apply. 1024 */ 1025u64 __weak bpf_jit_alloc_exec_limit(void) 1026{ 1027#if defined(MODULES_VADDR) 1028 return MODULES_END - MODULES_VADDR; 1029#else 1030 return VMALLOC_END - VMALLOC_START; 1031#endif 1032} 1033 1034static int __init bpf_jit_charge_init(void) 1035{ 1036 /* Only used as heuristic here to derive limit. */ 1037 bpf_jit_limit_max = bpf_jit_alloc_exec_limit(); 1038 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1, 1039 PAGE_SIZE), LONG_MAX); 1040 return 0; 1041} 1042pure_initcall(bpf_jit_charge_init); 1043 1044int bpf_jit_charge_modmem(u32 size) 1045{ 1046 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) { 1047 if (!bpf_capable()) { 1048 atomic_long_sub(size, &bpf_jit_current); 1049 return -EPERM; 1050 } 1051 } 1052 1053 return 0; 1054} 1055 1056void bpf_jit_uncharge_modmem(u32 size) 1057{ 1058 atomic_long_sub(size, &bpf_jit_current); 1059} 1060 1061void *__weak bpf_jit_alloc_exec(unsigned long size) 1062{ 1063 return execmem_alloc(EXECMEM_BPF, size); 1064} 1065 1066void __weak bpf_jit_free_exec(void *addr) 1067{ 1068 execmem_free(addr); 1069} 1070 1071struct bpf_binary_header * 1072bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr, 1073 unsigned int alignment, 1074 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1075{ 1076 struct bpf_binary_header *hdr; 1077 u32 size, hole, start; 1078 1079 WARN_ON_ONCE(!is_power_of_2(alignment) || 1080 alignment > BPF_IMAGE_ALIGNMENT); 1081 1082 /* Most of BPF filters are really small, but if some of them 1083 * fill a page, allow at least 128 extra bytes to insert a 1084 * random section of illegal instructions. 1085 */ 1086 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE); 1087 1088 if (bpf_jit_charge_modmem(size)) 1089 return NULL; 1090 hdr = bpf_jit_alloc_exec(size); 1091 if (!hdr) { 1092 bpf_jit_uncharge_modmem(size); 1093 return NULL; 1094 } 1095 1096 /* Fill space with illegal/arch-dep instructions. */ 1097 bpf_fill_ill_insns(hdr, size); 1098 1099 hdr->size = size; 1100 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)), 1101 PAGE_SIZE - sizeof(*hdr)); 1102 start = get_random_u32_below(hole) & ~(alignment - 1); 1103 1104 /* Leave a random number of instructions before BPF code. */ 1105 *image_ptr = &hdr->image[start]; 1106 1107 return hdr; 1108} 1109 1110void bpf_jit_binary_free(struct bpf_binary_header *hdr) 1111{ 1112 u32 size = hdr->size; 1113 1114 bpf_jit_free_exec(hdr); 1115 bpf_jit_uncharge_modmem(size); 1116} 1117 1118/* Allocate jit binary from bpf_prog_pack allocator. 1119 * Since the allocated memory is RO+X, the JIT engine cannot write directly 1120 * to the memory. To solve this problem, a RW buffer is also allocated at 1121 * as the same time. The JIT engine should calculate offsets based on the 1122 * RO memory address, but write JITed program to the RW buffer. Once the 1123 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies 1124 * the JITed program to the RO memory. 1125 */ 1126struct bpf_binary_header * 1127bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr, 1128 unsigned int alignment, 1129 struct bpf_binary_header **rw_header, 1130 u8 **rw_image, 1131 bpf_jit_fill_hole_t bpf_fill_ill_insns) 1132{ 1133 struct bpf_binary_header *ro_header; 1134 u32 size, hole, start; 1135 1136 WARN_ON_ONCE(!is_power_of_2(alignment) || 1137 alignment > BPF_IMAGE_ALIGNMENT); 1138 1139 /* add 16 bytes for a random section of illegal instructions */ 1140 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE); 1141 1142 if (bpf_jit_charge_modmem(size)) 1143 return NULL; 1144 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns); 1145 if (!ro_header) { 1146 bpf_jit_uncharge_modmem(size); 1147 return NULL; 1148 } 1149 1150 *rw_header = kvmalloc(size, GFP_KERNEL); 1151 if (!*rw_header) { 1152 bpf_prog_pack_free(ro_header, size); 1153 bpf_jit_uncharge_modmem(size); 1154 return NULL; 1155 } 1156 1157 /* Fill space with illegal/arch-dep instructions. */ 1158 bpf_fill_ill_insns(*rw_header, size); 1159 (*rw_header)->size = size; 1160 1161 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)), 1162 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header)); 1163 start = get_random_u32_below(hole) & ~(alignment - 1); 1164 1165 *image_ptr = &ro_header->image[start]; 1166 *rw_image = &(*rw_header)->image[start]; 1167 1168 return ro_header; 1169} 1170 1171/* Copy JITed text from rw_header to its final location, the ro_header. */ 1172int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header, 1173 struct bpf_binary_header *rw_header) 1174{ 1175 void *ptr; 1176 1177 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size); 1178 1179 kvfree(rw_header); 1180 1181 if (IS_ERR(ptr)) { 1182 bpf_prog_pack_free(ro_header, ro_header->size); 1183 return PTR_ERR(ptr); 1184 } 1185 return 0; 1186} 1187 1188/* bpf_jit_binary_pack_free is called in two different scenarios: 1189 * 1) when the program is freed after; 1190 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize). 1191 * For case 2), we need to free both the RO memory and the RW buffer. 1192 * 1193 * bpf_jit_binary_pack_free requires proper ro_header->size. However, 1194 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size 1195 * must be set with either bpf_jit_binary_pack_finalize (normal path) or 1196 * bpf_arch_text_copy (when jit fails). 1197 */ 1198void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header, 1199 struct bpf_binary_header *rw_header) 1200{ 1201 u32 size = ro_header->size; 1202 1203 bpf_prog_pack_free(ro_header, size); 1204 kvfree(rw_header); 1205 bpf_jit_uncharge_modmem(size); 1206} 1207 1208struct bpf_binary_header * 1209bpf_jit_binary_pack_hdr(const struct bpf_prog *fp) 1210{ 1211 unsigned long real_start = (unsigned long)fp->bpf_func; 1212 unsigned long addr; 1213 1214 addr = real_start & BPF_PROG_CHUNK_MASK; 1215 return (void *)addr; 1216} 1217 1218static inline struct bpf_binary_header * 1219bpf_jit_binary_hdr(const struct bpf_prog *fp) 1220{ 1221 unsigned long real_start = (unsigned long)fp->bpf_func; 1222 unsigned long addr; 1223 1224 addr = real_start & PAGE_MASK; 1225 return (void *)addr; 1226} 1227 1228/* This symbol is only overridden by archs that have different 1229 * requirements than the usual eBPF JITs, f.e. when they only 1230 * implement cBPF JIT, do not set images read-only, etc. 1231 */ 1232void __weak bpf_jit_free(struct bpf_prog *fp) 1233{ 1234 if (fp->jited) { 1235 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp); 1236 1237 bpf_jit_binary_free(hdr); 1238 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp)); 1239 } 1240 1241 bpf_prog_unlock_free(fp); 1242} 1243 1244int bpf_jit_get_func_addr(const struct bpf_prog *prog, 1245 const struct bpf_insn *insn, bool extra_pass, 1246 u64 *func_addr, bool *func_addr_fixed) 1247{ 1248 s16 off = insn->off; 1249 s32 imm = insn->imm; 1250 u8 *addr; 1251 int err; 1252 1253 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL; 1254 if (!*func_addr_fixed) { 1255 /* Place-holder address till the last pass has collected 1256 * all addresses for JITed subprograms in which case we 1257 * can pick them up from prog->aux. 1258 */ 1259 if (!extra_pass) 1260 addr = NULL; 1261 else if (prog->aux->func && 1262 off >= 0 && off < prog->aux->real_func_cnt) 1263 addr = (u8 *)prog->aux->func[off]->bpf_func; 1264 else 1265 return -EINVAL; 1266 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && 1267 bpf_jit_supports_far_kfunc_call()) { 1268 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr); 1269 if (err) 1270 return err; 1271 } else { 1272 /* Address of a BPF helper call. Since part of the core 1273 * kernel, it's always at a fixed location. __bpf_call_base 1274 * and the helper with imm relative to it are both in core 1275 * kernel. 1276 */ 1277 addr = (u8 *)__bpf_call_base + imm; 1278 } 1279 1280 *func_addr = (unsigned long)addr; 1281 return 0; 1282} 1283 1284const char *bpf_jit_get_prog_name(struct bpf_prog *prog) 1285{ 1286 if (prog->aux->ksym.prog) 1287 return prog->aux->ksym.name; 1288 return prog->aux->name; 1289} 1290 1291static int bpf_jit_blind_insn(const struct bpf_insn *from, 1292 const struct bpf_insn *aux, 1293 struct bpf_insn *to_buff, 1294 bool emit_zext) 1295{ 1296 struct bpf_insn *to = to_buff; 1297 u32 imm_rnd = get_random_u32(); 1298 s16 off; 1299 1300 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG); 1301 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG); 1302 1303 /* Constraints on AX register: 1304 * 1305 * AX register is inaccessible from user space. It is mapped in 1306 * all JITs, and used here for constant blinding rewrites. It is 1307 * typically "stateless" meaning its contents are only valid within 1308 * the executed instruction, but not across several instructions. 1309 * There are a few exceptions however which are further detailed 1310 * below. 1311 * 1312 * Constant blinding is only used by JITs, not in the interpreter. 1313 * The interpreter uses AX in some occasions as a local temporary 1314 * register e.g. in DIV or MOD instructions. 1315 * 1316 * In restricted circumstances, the verifier can also use the AX 1317 * register for rewrites as long as they do not interfere with 1318 * the above cases! 1319 */ 1320 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX) 1321 goto out; 1322 1323 if (from->imm == 0 && 1324 (from->code == (BPF_ALU | BPF_MOV | BPF_K) || 1325 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) { 1326 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg); 1327 goto out; 1328 } 1329 1330 switch (from->code) { 1331 case BPF_ALU | BPF_ADD | BPF_K: 1332 case BPF_ALU | BPF_SUB | BPF_K: 1333 case BPF_ALU | BPF_AND | BPF_K: 1334 case BPF_ALU | BPF_OR | BPF_K: 1335 case BPF_ALU | BPF_XOR | BPF_K: 1336 case BPF_ALU | BPF_MUL | BPF_K: 1337 case BPF_ALU | BPF_MOV | BPF_K: 1338 case BPF_ALU | BPF_DIV | BPF_K: 1339 case BPF_ALU | BPF_MOD | BPF_K: 1340 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1341 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1342 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1343 break; 1344 1345 case BPF_ALU64 | BPF_ADD | BPF_K: 1346 case BPF_ALU64 | BPF_SUB | BPF_K: 1347 case BPF_ALU64 | BPF_AND | BPF_K: 1348 case BPF_ALU64 | BPF_OR | BPF_K: 1349 case BPF_ALU64 | BPF_XOR | BPF_K: 1350 case BPF_ALU64 | BPF_MUL | BPF_K: 1351 case BPF_ALU64 | BPF_MOV | BPF_K: 1352 case BPF_ALU64 | BPF_DIV | BPF_K: 1353 case BPF_ALU64 | BPF_MOD | BPF_K: 1354 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1355 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1356 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off); 1357 break; 1358 1359 case BPF_JMP | BPF_JEQ | BPF_K: 1360 case BPF_JMP | BPF_JNE | BPF_K: 1361 case BPF_JMP | BPF_JGT | BPF_K: 1362 case BPF_JMP | BPF_JLT | BPF_K: 1363 case BPF_JMP | BPF_JGE | BPF_K: 1364 case BPF_JMP | BPF_JLE | BPF_K: 1365 case BPF_JMP | BPF_JSGT | BPF_K: 1366 case BPF_JMP | BPF_JSLT | BPF_K: 1367 case BPF_JMP | BPF_JSGE | BPF_K: 1368 case BPF_JMP | BPF_JSLE | BPF_K: 1369 case BPF_JMP | BPF_JSET | BPF_K: 1370 /* Accommodate for extra offset in case of a backjump. */ 1371 off = from->off; 1372 if (off < 0) 1373 off -= 2; 1374 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1375 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1376 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off); 1377 break; 1378 1379 case BPF_JMP32 | BPF_JEQ | BPF_K: 1380 case BPF_JMP32 | BPF_JNE | BPF_K: 1381 case BPF_JMP32 | BPF_JGT | BPF_K: 1382 case BPF_JMP32 | BPF_JLT | BPF_K: 1383 case BPF_JMP32 | BPF_JGE | BPF_K: 1384 case BPF_JMP32 | BPF_JLE | BPF_K: 1385 case BPF_JMP32 | BPF_JSGT | BPF_K: 1386 case BPF_JMP32 | BPF_JSLT | BPF_K: 1387 case BPF_JMP32 | BPF_JSGE | BPF_K: 1388 case BPF_JMP32 | BPF_JSLE | BPF_K: 1389 case BPF_JMP32 | BPF_JSET | BPF_K: 1390 /* Accommodate for extra offset in case of a backjump. */ 1391 off = from->off; 1392 if (off < 0) 1393 off -= 2; 1394 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1395 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1396 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX, 1397 off); 1398 break; 1399 1400 case BPF_LD | BPF_IMM | BPF_DW: 1401 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm); 1402 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1403 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32); 1404 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX); 1405 break; 1406 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */ 1407 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm); 1408 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1409 if (emit_zext) 1410 *to++ = BPF_ZEXT_REG(BPF_REG_AX); 1411 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX); 1412 break; 1413 1414 case BPF_ST | BPF_MEM | BPF_DW: 1415 case BPF_ST | BPF_MEM | BPF_W: 1416 case BPF_ST | BPF_MEM | BPF_H: 1417 case BPF_ST | BPF_MEM | BPF_B: 1418 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm); 1419 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1420 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off); 1421 break; 1422 1423 case BPF_ST | BPF_PROBE_MEM32 | BPF_DW: 1424 case BPF_ST | BPF_PROBE_MEM32 | BPF_W: 1425 case BPF_ST | BPF_PROBE_MEM32 | BPF_H: 1426 case BPF_ST | BPF_PROBE_MEM32 | BPF_B: 1427 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ 1428 from->imm); 1429 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd); 1430 /* 1431 * Cannot use BPF_STX_MEM() macro here as it 1432 * hardcodes BPF_MEM mode, losing PROBE_MEM32 1433 * and breaking arena addressing in the JIT. 1434 */ 1435 *to++ = (struct bpf_insn) { 1436 .code = BPF_STX | BPF_PROBE_MEM32 | 1437 BPF_SIZE(from->code), 1438 .dst_reg = from->dst_reg, 1439 .src_reg = BPF_REG_AX, 1440 .off = from->off, 1441 }; 1442 break; 1443 } 1444out: 1445 return to - to_buff; 1446} 1447 1448static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other, 1449 gfp_t gfp_extra_flags) 1450{ 1451 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags; 1452 struct bpf_prog *fp; 1453 1454 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags); 1455 if (fp != NULL) { 1456 /* aux->prog still points to the fp_other one, so 1457 * when promoting the clone to the real program, 1458 * this still needs to be adapted. 1459 */ 1460 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE); 1461 } 1462 1463 return fp; 1464} 1465 1466static void bpf_prog_clone_free(struct bpf_prog *fp) 1467{ 1468 /* aux was stolen by the other clone, so we cannot free 1469 * it from this path! It will be freed eventually by the 1470 * other program on release. 1471 * 1472 * At this point, we don't need a deferred release since 1473 * clone is guaranteed to not be locked. 1474 */ 1475 fp->aux = NULL; 1476 fp->stats = NULL; 1477 fp->active = NULL; 1478 __bpf_prog_free(fp); 1479} 1480 1481void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other) 1482{ 1483 /* We have to repoint aux->prog to self, as we don't 1484 * know whether fp here is the clone or the original. 1485 */ 1486 fp->aux->prog = fp; 1487 bpf_prog_clone_free(fp_other); 1488} 1489 1490static void adjust_insn_arrays(struct bpf_prog *prog, u32 off, u32 len) 1491{ 1492#ifdef CONFIG_BPF_SYSCALL 1493 struct bpf_map *map; 1494 int i; 1495 1496 if (len <= 1) 1497 return; 1498 1499 for (i = 0; i < prog->aux->used_map_cnt; i++) { 1500 map = prog->aux->used_maps[i]; 1501 if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY) 1502 bpf_insn_array_adjust(map, off, len); 1503 } 1504#endif 1505} 1506 1507struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog) 1508{ 1509 struct bpf_insn insn_buff[16], aux[2]; 1510 struct bpf_prog *clone, *tmp; 1511 int insn_delta, insn_cnt; 1512 struct bpf_insn *insn; 1513 int i, rewritten; 1514 1515 if (!prog->blinding_requested || prog->blinded) 1516 return prog; 1517 1518 clone = bpf_prog_clone_create(prog, GFP_USER); 1519 if (!clone) 1520 return ERR_PTR(-ENOMEM); 1521 1522 insn_cnt = clone->len; 1523 insn = clone->insnsi; 1524 1525 for (i = 0; i < insn_cnt; i++, insn++) { 1526 if (bpf_pseudo_func(insn)) { 1527 /* ld_imm64 with an address of bpf subprog is not 1528 * a user controlled constant. Don't randomize it, 1529 * since it will conflict with jit_subprogs() logic. 1530 */ 1531 insn++; 1532 i++; 1533 continue; 1534 } 1535 1536 /* We temporarily need to hold the original ld64 insn 1537 * so that we can still access the first part in the 1538 * second blinding run. 1539 */ 1540 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) && 1541 insn[1].code == 0) 1542 memcpy(aux, insn, sizeof(aux)); 1543 1544 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff, 1545 clone->aux->verifier_zext); 1546 if (!rewritten) 1547 continue; 1548 1549 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten); 1550 if (IS_ERR(tmp)) { 1551 /* Patching may have repointed aux->prog during 1552 * realloc from the original one, so we need to 1553 * fix it up here on error. 1554 */ 1555 bpf_jit_prog_release_other(prog, clone); 1556 return tmp; 1557 } 1558 1559 clone = tmp; 1560 insn_delta = rewritten - 1; 1561 1562 /* Instructions arrays must be updated using absolute xlated offsets */ 1563 adjust_insn_arrays(clone, prog->aux->subprog_start + i, rewritten); 1564 1565 /* Walk new program and skip insns we just inserted. */ 1566 insn = clone->insnsi + i + insn_delta; 1567 insn_cnt += insn_delta; 1568 i += insn_delta; 1569 } 1570 1571 clone->blinded = 1; 1572 return clone; 1573} 1574#endif /* CONFIG_BPF_JIT */ 1575 1576/* Base function for offset calculation. Needs to go into .text section, 1577 * therefore keeping it non-static as well; will also be used by JITs 1578 * anyway later on, so do not let the compiler omit it. This also needs 1579 * to go into kallsyms for correlation from e.g. bpftool, so naming 1580 * must not change. 1581 */ 1582noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5) 1583{ 1584 return 0; 1585} 1586EXPORT_SYMBOL_GPL(__bpf_call_base); 1587 1588/* All UAPI available opcodes. */ 1589#define BPF_INSN_MAP(INSN_2, INSN_3) \ 1590 /* 32 bit ALU operations. */ \ 1591 /* Register based. */ \ 1592 INSN_3(ALU, ADD, X), \ 1593 INSN_3(ALU, SUB, X), \ 1594 INSN_3(ALU, AND, X), \ 1595 INSN_3(ALU, OR, X), \ 1596 INSN_3(ALU, LSH, X), \ 1597 INSN_3(ALU, RSH, X), \ 1598 INSN_3(ALU, XOR, X), \ 1599 INSN_3(ALU, MUL, X), \ 1600 INSN_3(ALU, MOV, X), \ 1601 INSN_3(ALU, ARSH, X), \ 1602 INSN_3(ALU, DIV, X), \ 1603 INSN_3(ALU, MOD, X), \ 1604 INSN_2(ALU, NEG), \ 1605 INSN_3(ALU, END, TO_BE), \ 1606 INSN_3(ALU, END, TO_LE), \ 1607 /* Immediate based. */ \ 1608 INSN_3(ALU, ADD, K), \ 1609 INSN_3(ALU, SUB, K), \ 1610 INSN_3(ALU, AND, K), \ 1611 INSN_3(ALU, OR, K), \ 1612 INSN_3(ALU, LSH, K), \ 1613 INSN_3(ALU, RSH, K), \ 1614 INSN_3(ALU, XOR, K), \ 1615 INSN_3(ALU, MUL, K), \ 1616 INSN_3(ALU, MOV, K), \ 1617 INSN_3(ALU, ARSH, K), \ 1618 INSN_3(ALU, DIV, K), \ 1619 INSN_3(ALU, MOD, K), \ 1620 /* 64 bit ALU operations. */ \ 1621 /* Register based. */ \ 1622 INSN_3(ALU64, ADD, X), \ 1623 INSN_3(ALU64, SUB, X), \ 1624 INSN_3(ALU64, AND, X), \ 1625 INSN_3(ALU64, OR, X), \ 1626 INSN_3(ALU64, LSH, X), \ 1627 INSN_3(ALU64, RSH, X), \ 1628 INSN_3(ALU64, XOR, X), \ 1629 INSN_3(ALU64, MUL, X), \ 1630 INSN_3(ALU64, MOV, X), \ 1631 INSN_3(ALU64, ARSH, X), \ 1632 INSN_3(ALU64, DIV, X), \ 1633 INSN_3(ALU64, MOD, X), \ 1634 INSN_2(ALU64, NEG), \ 1635 INSN_3(ALU64, END, TO_LE), \ 1636 /* Immediate based. */ \ 1637 INSN_3(ALU64, ADD, K), \ 1638 INSN_3(ALU64, SUB, K), \ 1639 INSN_3(ALU64, AND, K), \ 1640 INSN_3(ALU64, OR, K), \ 1641 INSN_3(ALU64, LSH, K), \ 1642 INSN_3(ALU64, RSH, K), \ 1643 INSN_3(ALU64, XOR, K), \ 1644 INSN_3(ALU64, MUL, K), \ 1645 INSN_3(ALU64, MOV, K), \ 1646 INSN_3(ALU64, ARSH, K), \ 1647 INSN_3(ALU64, DIV, K), \ 1648 INSN_3(ALU64, MOD, K), \ 1649 /* Call instruction. */ \ 1650 INSN_2(JMP, CALL), \ 1651 /* Exit instruction. */ \ 1652 INSN_2(JMP, EXIT), \ 1653 /* 32-bit Jump instructions. */ \ 1654 /* Register based. */ \ 1655 INSN_3(JMP32, JEQ, X), \ 1656 INSN_3(JMP32, JNE, X), \ 1657 INSN_3(JMP32, JGT, X), \ 1658 INSN_3(JMP32, JLT, X), \ 1659 INSN_3(JMP32, JGE, X), \ 1660 INSN_3(JMP32, JLE, X), \ 1661 INSN_3(JMP32, JSGT, X), \ 1662 INSN_3(JMP32, JSLT, X), \ 1663 INSN_3(JMP32, JSGE, X), \ 1664 INSN_3(JMP32, JSLE, X), \ 1665 INSN_3(JMP32, JSET, X), \ 1666 /* Immediate based. */ \ 1667 INSN_3(JMP32, JEQ, K), \ 1668 INSN_3(JMP32, JNE, K), \ 1669 INSN_3(JMP32, JGT, K), \ 1670 INSN_3(JMP32, JLT, K), \ 1671 INSN_3(JMP32, JGE, K), \ 1672 INSN_3(JMP32, JLE, K), \ 1673 INSN_3(JMP32, JSGT, K), \ 1674 INSN_3(JMP32, JSLT, K), \ 1675 INSN_3(JMP32, JSGE, K), \ 1676 INSN_3(JMP32, JSLE, K), \ 1677 INSN_3(JMP32, JSET, K), \ 1678 /* Jump instructions. */ \ 1679 /* Register based. */ \ 1680 INSN_3(JMP, JEQ, X), \ 1681 INSN_3(JMP, JNE, X), \ 1682 INSN_3(JMP, JGT, X), \ 1683 INSN_3(JMP, JLT, X), \ 1684 INSN_3(JMP, JGE, X), \ 1685 INSN_3(JMP, JLE, X), \ 1686 INSN_3(JMP, JSGT, X), \ 1687 INSN_3(JMP, JSLT, X), \ 1688 INSN_3(JMP, JSGE, X), \ 1689 INSN_3(JMP, JSLE, X), \ 1690 INSN_3(JMP, JSET, X), \ 1691 /* Immediate based. */ \ 1692 INSN_3(JMP, JEQ, K), \ 1693 INSN_3(JMP, JNE, K), \ 1694 INSN_3(JMP, JGT, K), \ 1695 INSN_3(JMP, JLT, K), \ 1696 INSN_3(JMP, JGE, K), \ 1697 INSN_3(JMP, JLE, K), \ 1698 INSN_3(JMP, JSGT, K), \ 1699 INSN_3(JMP, JSLT, K), \ 1700 INSN_3(JMP, JSGE, K), \ 1701 INSN_3(JMP, JSLE, K), \ 1702 INSN_3(JMP, JSET, K), \ 1703 INSN_2(JMP, JA), \ 1704 INSN_2(JMP32, JA), \ 1705 /* Atomic operations. */ \ 1706 INSN_3(STX, ATOMIC, B), \ 1707 INSN_3(STX, ATOMIC, H), \ 1708 INSN_3(STX, ATOMIC, W), \ 1709 INSN_3(STX, ATOMIC, DW), \ 1710 /* Store instructions. */ \ 1711 /* Register based. */ \ 1712 INSN_3(STX, MEM, B), \ 1713 INSN_3(STX, MEM, H), \ 1714 INSN_3(STX, MEM, W), \ 1715 INSN_3(STX, MEM, DW), \ 1716 /* Immediate based. */ \ 1717 INSN_3(ST, MEM, B), \ 1718 INSN_3(ST, MEM, H), \ 1719 INSN_3(ST, MEM, W), \ 1720 INSN_3(ST, MEM, DW), \ 1721 /* Load instructions. */ \ 1722 /* Register based. */ \ 1723 INSN_3(LDX, MEM, B), \ 1724 INSN_3(LDX, MEM, H), \ 1725 INSN_3(LDX, MEM, W), \ 1726 INSN_3(LDX, MEM, DW), \ 1727 INSN_3(LDX, MEMSX, B), \ 1728 INSN_3(LDX, MEMSX, H), \ 1729 INSN_3(LDX, MEMSX, W), \ 1730 /* Immediate based. */ \ 1731 INSN_3(LD, IMM, DW) 1732 1733bool bpf_opcode_in_insntable(u8 code) 1734{ 1735#define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true 1736#define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true 1737 static const bool public_insntable[256] = { 1738 [0 ... 255] = false, 1739 /* Now overwrite non-defaults ... */ 1740 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL), 1741 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */ 1742 [BPF_LD | BPF_ABS | BPF_B] = true, 1743 [BPF_LD | BPF_ABS | BPF_H] = true, 1744 [BPF_LD | BPF_ABS | BPF_W] = true, 1745 [BPF_LD | BPF_IND | BPF_B] = true, 1746 [BPF_LD | BPF_IND | BPF_H] = true, 1747 [BPF_LD | BPF_IND | BPF_W] = true, 1748 [BPF_JMP | BPF_JA | BPF_X] = true, 1749 [BPF_JMP | BPF_JCOND] = true, 1750 }; 1751#undef BPF_INSN_3_TBL 1752#undef BPF_INSN_2_TBL 1753 return public_insntable[code]; 1754} 1755 1756#ifndef CONFIG_BPF_JIT_ALWAYS_ON 1757/* Absolute value of s32 without undefined behavior for S32_MIN */ 1758static u32 abs_s32(s32 x) 1759{ 1760 return x >= 0 ? (u32)x : -(u32)x; 1761} 1762 1763/** 1764 * ___bpf_prog_run - run eBPF program on a given context 1765 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers 1766 * @insn: is the array of eBPF instructions 1767 * 1768 * Decode and execute eBPF instructions. 1769 * 1770 * Return: whatever value is in %BPF_R0 at program exit 1771 */ 1772static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn) 1773{ 1774#define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y 1775#define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z 1776 static const void * const jumptable[256] __annotate_jump_table = { 1777 [0 ... 255] = &&default_label, 1778 /* Now overwrite non-defaults ... */ 1779 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL), 1780 /* Non-UAPI available opcodes. */ 1781 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS, 1782 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL, 1783 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC, 1784 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B, 1785 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H, 1786 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W, 1787 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW, 1788 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B, 1789 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H, 1790 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W, 1791 }; 1792#undef BPF_INSN_3_LBL 1793#undef BPF_INSN_2_LBL 1794 u32 tail_call_cnt = 0; 1795 1796#define CONT ({ insn++; goto select_insn; }) 1797#define CONT_JMP ({ insn++; goto select_insn; }) 1798 1799select_insn: 1800 goto *jumptable[insn->code]; 1801 1802 /* Explicitly mask the register-based shift amounts with 63 or 31 1803 * to avoid undefined behavior. Normally this won't affect the 1804 * generated code, for example, in case of native 64 bit archs such 1805 * as x86-64 or arm64, the compiler is optimizing the AND away for 1806 * the interpreter. In case of JITs, each of the JIT backends compiles 1807 * the BPF shift operations to machine instructions which produce 1808 * implementation-defined results in such a case; the resulting 1809 * contents of the register may be arbitrary, but program behaviour 1810 * as a whole remains defined. In other words, in case of JIT backends, 1811 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation. 1812 */ 1813 /* ALU (shifts) */ 1814#define SHT(OPCODE, OP) \ 1815 ALU64_##OPCODE##_X: \ 1816 DST = DST OP (SRC & 63); \ 1817 CONT; \ 1818 ALU_##OPCODE##_X: \ 1819 DST = (u32) DST OP ((u32) SRC & 31); \ 1820 CONT; \ 1821 ALU64_##OPCODE##_K: \ 1822 DST = DST OP IMM; \ 1823 CONT; \ 1824 ALU_##OPCODE##_K: \ 1825 DST = (u32) DST OP (u32) IMM; \ 1826 CONT; 1827 /* ALU (rest) */ 1828#define ALU(OPCODE, OP) \ 1829 ALU64_##OPCODE##_X: \ 1830 DST = DST OP SRC; \ 1831 CONT; \ 1832 ALU_##OPCODE##_X: \ 1833 DST = (u32) DST OP (u32) SRC; \ 1834 CONT; \ 1835 ALU64_##OPCODE##_K: \ 1836 DST = DST OP IMM; \ 1837 CONT; \ 1838 ALU_##OPCODE##_K: \ 1839 DST = (u32) DST OP (u32) IMM; \ 1840 CONT; 1841 ALU(ADD, +) 1842 ALU(SUB, -) 1843 ALU(AND, &) 1844 ALU(OR, |) 1845 ALU(XOR, ^) 1846 ALU(MUL, *) 1847 SHT(LSH, <<) 1848 SHT(RSH, >>) 1849#undef SHT 1850#undef ALU 1851 ALU_NEG: 1852 DST = (u32) -DST; 1853 CONT; 1854 ALU64_NEG: 1855 DST = -DST; 1856 CONT; 1857 ALU_MOV_X: 1858 switch (OFF) { 1859 case 0: 1860 DST = (u32) SRC; 1861 break; 1862 case 8: 1863 DST = (u32)(s8) SRC; 1864 break; 1865 case 16: 1866 DST = (u32)(s16) SRC; 1867 break; 1868 } 1869 CONT; 1870 ALU_MOV_K: 1871 DST = (u32) IMM; 1872 CONT; 1873 ALU64_MOV_X: 1874 switch (OFF) { 1875 case 0: 1876 DST = SRC; 1877 break; 1878 case 8: 1879 DST = (s8) SRC; 1880 break; 1881 case 16: 1882 DST = (s16) SRC; 1883 break; 1884 case 32: 1885 DST = (s32) SRC; 1886 break; 1887 } 1888 CONT; 1889 ALU64_MOV_K: 1890 DST = IMM; 1891 CONT; 1892 LD_IMM_DW: 1893 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32; 1894 insn++; 1895 CONT; 1896 ALU_ARSH_X: 1897 DST = (u64) (u32) (((s32) DST) >> (SRC & 31)); 1898 CONT; 1899 ALU_ARSH_K: 1900 DST = (u64) (u32) (((s32) DST) >> IMM); 1901 CONT; 1902 ALU64_ARSH_X: 1903 (*(s64 *) &DST) >>= (SRC & 63); 1904 CONT; 1905 ALU64_ARSH_K: 1906 (*(s64 *) &DST) >>= IMM; 1907 CONT; 1908 ALU64_MOD_X: 1909 switch (OFF) { 1910 case 0: 1911 div64_u64_rem(DST, SRC, &AX); 1912 DST = AX; 1913 break; 1914 case 1: 1915 AX = div64_s64(DST, SRC); 1916 DST = DST - AX * SRC; 1917 break; 1918 } 1919 CONT; 1920 ALU_MOD_X: 1921 switch (OFF) { 1922 case 0: 1923 AX = (u32) DST; 1924 DST = do_div(AX, (u32) SRC); 1925 break; 1926 case 1: 1927 AX = abs_s32((s32)DST); 1928 AX = do_div(AX, abs_s32((s32)SRC)); 1929 if ((s32)DST < 0) 1930 DST = (u32)-AX; 1931 else 1932 DST = (u32)AX; 1933 break; 1934 } 1935 CONT; 1936 ALU64_MOD_K: 1937 switch (OFF) { 1938 case 0: 1939 div64_u64_rem(DST, IMM, &AX); 1940 DST = AX; 1941 break; 1942 case 1: 1943 AX = div64_s64(DST, IMM); 1944 DST = DST - AX * IMM; 1945 break; 1946 } 1947 CONT; 1948 ALU_MOD_K: 1949 switch (OFF) { 1950 case 0: 1951 AX = (u32) DST; 1952 DST = do_div(AX, (u32) IMM); 1953 break; 1954 case 1: 1955 AX = abs_s32((s32)DST); 1956 AX = do_div(AX, abs_s32((s32)IMM)); 1957 if ((s32)DST < 0) 1958 DST = (u32)-AX; 1959 else 1960 DST = (u32)AX; 1961 break; 1962 } 1963 CONT; 1964 ALU64_DIV_X: 1965 switch (OFF) { 1966 case 0: 1967 DST = div64_u64(DST, SRC); 1968 break; 1969 case 1: 1970 DST = div64_s64(DST, SRC); 1971 break; 1972 } 1973 CONT; 1974 ALU_DIV_X: 1975 switch (OFF) { 1976 case 0: 1977 AX = (u32) DST; 1978 do_div(AX, (u32) SRC); 1979 DST = (u32) AX; 1980 break; 1981 case 1: 1982 AX = abs_s32((s32)DST); 1983 do_div(AX, abs_s32((s32)SRC)); 1984 if (((s32)DST < 0) == ((s32)SRC < 0)) 1985 DST = (u32)AX; 1986 else 1987 DST = (u32)-AX; 1988 break; 1989 } 1990 CONT; 1991 ALU64_DIV_K: 1992 switch (OFF) { 1993 case 0: 1994 DST = div64_u64(DST, IMM); 1995 break; 1996 case 1: 1997 DST = div64_s64(DST, IMM); 1998 break; 1999 } 2000 CONT; 2001 ALU_DIV_K: 2002 switch (OFF) { 2003 case 0: 2004 AX = (u32) DST; 2005 do_div(AX, (u32) IMM); 2006 DST = (u32) AX; 2007 break; 2008 case 1: 2009 AX = abs_s32((s32)DST); 2010 do_div(AX, abs_s32((s32)IMM)); 2011 if (((s32)DST < 0) == ((s32)IMM < 0)) 2012 DST = (u32)AX; 2013 else 2014 DST = (u32)-AX; 2015 break; 2016 } 2017 CONT; 2018 ALU_END_TO_BE: 2019 switch (IMM) { 2020 case 16: 2021 DST = (__force u16) cpu_to_be16(DST); 2022 break; 2023 case 32: 2024 DST = (__force u32) cpu_to_be32(DST); 2025 break; 2026 case 64: 2027 DST = (__force u64) cpu_to_be64(DST); 2028 break; 2029 } 2030 CONT; 2031 ALU_END_TO_LE: 2032 switch (IMM) { 2033 case 16: 2034 DST = (__force u16) cpu_to_le16(DST); 2035 break; 2036 case 32: 2037 DST = (__force u32) cpu_to_le32(DST); 2038 break; 2039 case 64: 2040 DST = (__force u64) cpu_to_le64(DST); 2041 break; 2042 } 2043 CONT; 2044 ALU64_END_TO_LE: 2045 switch (IMM) { 2046 case 16: 2047 DST = (__force u16) __swab16(DST); 2048 break; 2049 case 32: 2050 DST = (__force u32) __swab32(DST); 2051 break; 2052 case 64: 2053 DST = (__force u64) __swab64(DST); 2054 break; 2055 } 2056 CONT; 2057 2058 /* CALL */ 2059 JMP_CALL: 2060 /* Function call scratches BPF_R1-BPF_R5 registers, 2061 * preserves BPF_R6-BPF_R9, and stores return value 2062 * into BPF_R0. 2063 */ 2064 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3, 2065 BPF_R4, BPF_R5); 2066 CONT; 2067 2068 JMP_CALL_ARGS: 2069 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2, 2070 BPF_R3, BPF_R4, 2071 BPF_R5, 2072 insn + insn->off + 1); 2073 CONT; 2074 2075 JMP_TAIL_CALL: { 2076 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2; 2077 struct bpf_array *array = container_of(map, struct bpf_array, map); 2078 struct bpf_prog *prog; 2079 u32 index = BPF_R3; 2080 2081 if (unlikely(index >= array->map.max_entries)) 2082 goto out; 2083 2084 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT)) 2085 goto out; 2086 2087 tail_call_cnt++; 2088 2089 prog = READ_ONCE(array->ptrs[index]); 2090 if (!prog) 2091 goto out; 2092 2093 /* ARG1 at this point is guaranteed to point to CTX from 2094 * the verifier side due to the fact that the tail call is 2095 * handled like a helper, that is, bpf_tail_call_proto, 2096 * where arg1_type is ARG_PTR_TO_CTX. 2097 */ 2098 insn = prog->insnsi; 2099 goto select_insn; 2100out: 2101 CONT; 2102 } 2103 JMP_JA: 2104 insn += insn->off; 2105 CONT; 2106 JMP32_JA: 2107 insn += insn->imm; 2108 CONT; 2109 JMP_EXIT: 2110 return BPF_R0; 2111 /* JMP */ 2112#define COND_JMP(SIGN, OPCODE, CMP_OP) \ 2113 JMP_##OPCODE##_X: \ 2114 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \ 2115 insn += insn->off; \ 2116 CONT_JMP; \ 2117 } \ 2118 CONT; \ 2119 JMP32_##OPCODE##_X: \ 2120 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \ 2121 insn += insn->off; \ 2122 CONT_JMP; \ 2123 } \ 2124 CONT; \ 2125 JMP_##OPCODE##_K: \ 2126 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \ 2127 insn += insn->off; \ 2128 CONT_JMP; \ 2129 } \ 2130 CONT; \ 2131 JMP32_##OPCODE##_K: \ 2132 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \ 2133 insn += insn->off; \ 2134 CONT_JMP; \ 2135 } \ 2136 CONT; 2137 COND_JMP(u, JEQ, ==) 2138 COND_JMP(u, JNE, !=) 2139 COND_JMP(u, JGT, >) 2140 COND_JMP(u, JLT, <) 2141 COND_JMP(u, JGE, >=) 2142 COND_JMP(u, JLE, <=) 2143 COND_JMP(u, JSET, &) 2144 COND_JMP(s, JSGT, >) 2145 COND_JMP(s, JSLT, <) 2146 COND_JMP(s, JSGE, >=) 2147 COND_JMP(s, JSLE, <=) 2148#undef COND_JMP 2149 /* ST, STX and LDX*/ 2150 ST_NOSPEC: 2151 /* Speculation barrier for mitigating Speculative Store Bypass, 2152 * Bounds-Check Bypass and Type Confusion. In case of arm64, we 2153 * rely on the firmware mitigation as controlled via the ssbd 2154 * kernel parameter. Whenever the mitigation is enabled, it 2155 * works for all of the kernel code with no need to provide any 2156 * additional instructions here. In case of x86, we use 'lfence' 2157 * insn for mitigation. We reuse preexisting logic from Spectre 2158 * v1 mitigation that happens to produce the required code on 2159 * x86 for v4 as well. 2160 */ 2161 barrier_nospec(); 2162 CONT; 2163#define LDST(SIZEOP, SIZE) \ 2164 STX_MEM_##SIZEOP: \ 2165 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \ 2166 CONT; \ 2167 ST_MEM_##SIZEOP: \ 2168 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \ 2169 CONT; \ 2170 LDX_MEM_##SIZEOP: \ 2171 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2172 CONT; \ 2173 LDX_PROBE_MEM_##SIZEOP: \ 2174 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2175 (const void *)(long) (SRC + insn->off)); \ 2176 DST = *((SIZE *)&DST); \ 2177 CONT; 2178 2179 LDST(B, u8) 2180 LDST(H, u16) 2181 LDST(W, u32) 2182 LDST(DW, u64) 2183#undef LDST 2184 2185#define LDSX(SIZEOP, SIZE) \ 2186 LDX_MEMSX_##SIZEOP: \ 2187 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \ 2188 CONT; \ 2189 LDX_PROBE_MEMSX_##SIZEOP: \ 2190 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \ 2191 (const void *)(long) (SRC + insn->off)); \ 2192 DST = *((SIZE *)&DST); \ 2193 CONT; 2194 2195 LDSX(B, s8) 2196 LDSX(H, s16) 2197 LDSX(W, s32) 2198#undef LDSX 2199 2200#define ATOMIC_ALU_OP(BOP, KOP) \ 2201 case BOP: \ 2202 if (BPF_SIZE(insn->code) == BPF_W) \ 2203 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \ 2204 (DST + insn->off)); \ 2205 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2206 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \ 2207 (DST + insn->off)); \ 2208 else \ 2209 goto default_label; \ 2210 break; \ 2211 case BOP | BPF_FETCH: \ 2212 if (BPF_SIZE(insn->code) == BPF_W) \ 2213 SRC = (u32) atomic_fetch_##KOP( \ 2214 (u32) SRC, \ 2215 (atomic_t *)(unsigned long) (DST + insn->off)); \ 2216 else if (BPF_SIZE(insn->code) == BPF_DW) \ 2217 SRC = (u64) atomic64_fetch_##KOP( \ 2218 (u64) SRC, \ 2219 (atomic64_t *)(unsigned long) (DST + insn->off)); \ 2220 else \ 2221 goto default_label; \ 2222 break; 2223 2224 STX_ATOMIC_DW: 2225 STX_ATOMIC_W: 2226 STX_ATOMIC_H: 2227 STX_ATOMIC_B: 2228 switch (IMM) { 2229 /* Atomic read-modify-write instructions support only W and DW 2230 * size modifiers. 2231 */ 2232 ATOMIC_ALU_OP(BPF_ADD, add) 2233 ATOMIC_ALU_OP(BPF_AND, and) 2234 ATOMIC_ALU_OP(BPF_OR, or) 2235 ATOMIC_ALU_OP(BPF_XOR, xor) 2236#undef ATOMIC_ALU_OP 2237 2238 case BPF_XCHG: 2239 if (BPF_SIZE(insn->code) == BPF_W) 2240 SRC = (u32) atomic_xchg( 2241 (atomic_t *)(unsigned long) (DST + insn->off), 2242 (u32) SRC); 2243 else if (BPF_SIZE(insn->code) == BPF_DW) 2244 SRC = (u64) atomic64_xchg( 2245 (atomic64_t *)(unsigned long) (DST + insn->off), 2246 (u64) SRC); 2247 else 2248 goto default_label; 2249 break; 2250 case BPF_CMPXCHG: 2251 if (BPF_SIZE(insn->code) == BPF_W) 2252 BPF_R0 = (u32) atomic_cmpxchg( 2253 (atomic_t *)(unsigned long) (DST + insn->off), 2254 (u32) BPF_R0, (u32) SRC); 2255 else if (BPF_SIZE(insn->code) == BPF_DW) 2256 BPF_R0 = (u64) atomic64_cmpxchg( 2257 (atomic64_t *)(unsigned long) (DST + insn->off), 2258 (u64) BPF_R0, (u64) SRC); 2259 else 2260 goto default_label; 2261 break; 2262 /* Atomic load and store instructions support all size 2263 * modifiers. 2264 */ 2265 case BPF_LOAD_ACQ: 2266 switch (BPF_SIZE(insn->code)) { 2267#define LOAD_ACQUIRE(SIZEOP, SIZE) \ 2268 case BPF_##SIZEOP: \ 2269 DST = (SIZE)smp_load_acquire( \ 2270 (SIZE *)(unsigned long)(SRC + insn->off)); \ 2271 break; 2272 LOAD_ACQUIRE(B, u8) 2273 LOAD_ACQUIRE(H, u16) 2274 LOAD_ACQUIRE(W, u32) 2275#ifdef CONFIG_64BIT 2276 LOAD_ACQUIRE(DW, u64) 2277#endif 2278#undef LOAD_ACQUIRE 2279 default: 2280 goto default_label; 2281 } 2282 break; 2283 case BPF_STORE_REL: 2284 switch (BPF_SIZE(insn->code)) { 2285#define STORE_RELEASE(SIZEOP, SIZE) \ 2286 case BPF_##SIZEOP: \ 2287 smp_store_release( \ 2288 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \ 2289 break; 2290 STORE_RELEASE(B, u8) 2291 STORE_RELEASE(H, u16) 2292 STORE_RELEASE(W, u32) 2293#ifdef CONFIG_64BIT 2294 STORE_RELEASE(DW, u64) 2295#endif 2296#undef STORE_RELEASE 2297 default: 2298 goto default_label; 2299 } 2300 break; 2301 2302 default: 2303 goto default_label; 2304 } 2305 CONT; 2306 2307 default_label: 2308 /* If we ever reach this, we have a bug somewhere. Die hard here 2309 * instead of just returning 0; we could be somewhere in a subprog, 2310 * so execution could continue otherwise which we do /not/ want. 2311 * 2312 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable(). 2313 */ 2314 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n", 2315 insn->code, insn->imm); 2316 BUG_ON(1); 2317 return 0; 2318} 2319 2320#define PROG_NAME(stack_size) __bpf_prog_run##stack_size 2321#define DEFINE_BPF_PROG_RUN(stack_size) \ 2322static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \ 2323{ \ 2324 u64 stack[stack_size / sizeof(u64)]; \ 2325 u64 regs[MAX_BPF_EXT_REG] = {}; \ 2326\ 2327 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2328 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2329 ARG1 = (u64) (unsigned long) ctx; \ 2330 return ___bpf_prog_run(regs, insn); \ 2331} 2332 2333#define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size 2334#define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \ 2335static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \ 2336 const struct bpf_insn *insn) \ 2337{ \ 2338 u64 stack[stack_size / sizeof(u64)]; \ 2339 u64 regs[MAX_BPF_EXT_REG]; \ 2340\ 2341 kmsan_unpoison_memory(stack, sizeof(stack)); \ 2342 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \ 2343 BPF_R1 = r1; \ 2344 BPF_R2 = r2; \ 2345 BPF_R3 = r3; \ 2346 BPF_R4 = r4; \ 2347 BPF_R5 = r5; \ 2348 return ___bpf_prog_run(regs, insn); \ 2349} 2350 2351#define EVAL1(FN, X) FN(X) 2352#define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y) 2353#define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y) 2354#define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y) 2355#define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y) 2356#define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y) 2357 2358EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192); 2359EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384); 2360EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512); 2361 2362EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192); 2363EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384); 2364EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512); 2365 2366#define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size), 2367 2368static unsigned int (*interpreters[])(const void *ctx, 2369 const struct bpf_insn *insn) = { 2370EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2371EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2372EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2373}; 2374#undef PROG_NAME_LIST 2375#define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size), 2376static __maybe_unused 2377u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, 2378 const struct bpf_insn *insn) = { 2379EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192) 2380EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384) 2381EVAL4(PROG_NAME_LIST, 416, 448, 480, 512) 2382}; 2383#undef PROG_NAME_LIST 2384 2385#ifdef CONFIG_BPF_SYSCALL 2386void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth) 2387{ 2388 stack_depth = max_t(u32, stack_depth, 1); 2389 insn->off = (s16) insn->imm; 2390 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] - 2391 __bpf_call_base_args; 2392 insn->code = BPF_JMP | BPF_CALL_ARGS; 2393} 2394#endif 2395#endif 2396 2397static unsigned int __bpf_prog_ret0_warn(const void *ctx, 2398 const struct bpf_insn *insn) 2399{ 2400 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON 2401 * is not working properly, so warn about it! 2402 */ 2403 WARN_ON_ONCE(1); 2404 return 0; 2405} 2406 2407static bool __bpf_prog_map_compatible(struct bpf_map *map, 2408 const struct bpf_prog *fp) 2409{ 2410 enum bpf_prog_type prog_type = resolve_prog_type(fp); 2411 struct bpf_prog_aux *aux = fp->aux; 2412 enum bpf_cgroup_storage_type i; 2413 bool ret = false; 2414 u64 cookie; 2415 2416 if (fp->kprobe_override) 2417 return ret; 2418 2419 spin_lock(&map->owner_lock); 2420 /* There's no owner yet where we could check for compatibility. */ 2421 if (!map->owner) { 2422 map->owner = bpf_map_owner_alloc(map); 2423 if (!map->owner) 2424 goto err; 2425 map->owner->type = prog_type; 2426 map->owner->jited = fp->jited; 2427 map->owner->xdp_has_frags = aux->xdp_has_frags; 2428 map->owner->expected_attach_type = fp->expected_attach_type; 2429 map->owner->attach_func_proto = aux->attach_func_proto; 2430 for_each_cgroup_storage_type(i) { 2431 map->owner->storage_cookie[i] = 2432 aux->cgroup_storage[i] ? 2433 aux->cgroup_storage[i]->cookie : 0; 2434 } 2435 ret = true; 2436 } else { 2437 ret = map->owner->type == prog_type && 2438 map->owner->jited == fp->jited && 2439 map->owner->xdp_has_frags == aux->xdp_has_frags; 2440 if (ret && 2441 map->map_type == BPF_MAP_TYPE_PROG_ARRAY && 2442 map->owner->expected_attach_type != fp->expected_attach_type) 2443 ret = false; 2444 for_each_cgroup_storage_type(i) { 2445 if (!ret) 2446 break; 2447 cookie = aux->cgroup_storage[i] ? 2448 aux->cgroup_storage[i]->cookie : 0; 2449 ret = map->owner->storage_cookie[i] == cookie || 2450 !cookie; 2451 } 2452 if (ret && 2453 map->owner->attach_func_proto != aux->attach_func_proto) { 2454 switch (prog_type) { 2455 case BPF_PROG_TYPE_TRACING: 2456 case BPF_PROG_TYPE_LSM: 2457 case BPF_PROG_TYPE_EXT: 2458 case BPF_PROG_TYPE_STRUCT_OPS: 2459 ret = false; 2460 break; 2461 default: 2462 break; 2463 } 2464 } 2465 } 2466err: 2467 spin_unlock(&map->owner_lock); 2468 return ret; 2469} 2470 2471bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp) 2472{ 2473 /* XDP programs inserted into maps are not guaranteed to run on 2474 * a particular netdev (and can run outside driver context entirely 2475 * in the case of devmap and cpumap). Until device checks 2476 * are implemented, prohibit adding dev-bound programs to program maps. 2477 */ 2478 if (bpf_prog_is_dev_bound(fp->aux)) 2479 return false; 2480 2481 return __bpf_prog_map_compatible(map, fp); 2482} 2483 2484static int bpf_check_tail_call(const struct bpf_prog *fp) 2485{ 2486 struct bpf_prog_aux *aux = fp->aux; 2487 int i, ret = 0; 2488 2489 mutex_lock(&aux->used_maps_mutex); 2490 for (i = 0; i < aux->used_map_cnt; i++) { 2491 struct bpf_map *map = aux->used_maps[i]; 2492 2493 if (!map_type_contains_progs(map)) 2494 continue; 2495 2496 if (!__bpf_prog_map_compatible(map, fp)) { 2497 ret = -EINVAL; 2498 goto out; 2499 } 2500 } 2501 2502out: 2503 mutex_unlock(&aux->used_maps_mutex); 2504 return ret; 2505} 2506 2507static bool bpf_prog_select_interpreter(struct bpf_prog *fp) 2508{ 2509 bool select_interpreter = false; 2510#ifndef CONFIG_BPF_JIT_ALWAYS_ON 2511 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1); 2512 u32 idx = (round_up(stack_depth, 32) / 32) - 1; 2513 2514 /* may_goto may cause stack size > 512, leading to idx out-of-bounds. 2515 * But for non-JITed programs, we don't need bpf_func, so no bounds 2516 * check needed. 2517 */ 2518 if (idx < ARRAY_SIZE(interpreters)) { 2519 fp->bpf_func = interpreters[idx]; 2520 select_interpreter = true; 2521 } else { 2522 fp->bpf_func = __bpf_prog_ret0_warn; 2523 } 2524#else 2525 fp->bpf_func = __bpf_prog_ret0_warn; 2526#endif 2527 return select_interpreter; 2528} 2529 2530/** 2531 * bpf_prog_select_runtime - select exec runtime for BPF program 2532 * @fp: bpf_prog populated with BPF program 2533 * @err: pointer to error variable 2534 * 2535 * Try to JIT eBPF program, if JIT is not available, use interpreter. 2536 * The BPF program will be executed via bpf_prog_run() function. 2537 * 2538 * Return: the &fp argument along with &err set to 0 for success or 2539 * a negative errno code on failure 2540 */ 2541struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err) 2542{ 2543 /* In case of BPF to BPF calls, verifier did all the prep 2544 * work with regards to JITing, etc. 2545 */ 2546 bool jit_needed = false; 2547 2548 if (fp->bpf_func) 2549 goto finalize; 2550 2551 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) || 2552 bpf_prog_has_kfunc_call(fp)) 2553 jit_needed = true; 2554 2555 if (!bpf_prog_select_interpreter(fp)) 2556 jit_needed = true; 2557 2558 /* eBPF JITs can rewrite the program in case constant 2559 * blinding is active. However, in case of error during 2560 * blinding, bpf_int_jit_compile() must always return a 2561 * valid program, which in this case would simply not 2562 * be JITed, but falls back to the interpreter. 2563 */ 2564 if (!bpf_prog_is_offloaded(fp->aux)) { 2565 *err = bpf_prog_alloc_jited_linfo(fp); 2566 if (*err) 2567 return fp; 2568 2569 fp = bpf_int_jit_compile(fp); 2570 bpf_prog_jit_attempt_done(fp); 2571 if (!fp->jited && jit_needed) { 2572 *err = -ENOTSUPP; 2573 return fp; 2574 } 2575 } else { 2576 *err = bpf_prog_offload_compile(fp); 2577 if (*err) 2578 return fp; 2579 } 2580 2581finalize: 2582 *err = bpf_prog_lock_ro(fp); 2583 if (*err) 2584 return fp; 2585 2586 /* The tail call compatibility check can only be done at 2587 * this late stage as we need to determine, if we deal 2588 * with JITed or non JITed program concatenations and not 2589 * all eBPF JITs might immediately support all features. 2590 */ 2591 *err = bpf_check_tail_call(fp); 2592 2593 return fp; 2594} 2595EXPORT_SYMBOL_GPL(bpf_prog_select_runtime); 2596 2597static unsigned int __bpf_prog_ret1(const void *ctx, 2598 const struct bpf_insn *insn) 2599{ 2600 return 1; 2601} 2602 2603static struct bpf_prog_dummy { 2604 struct bpf_prog prog; 2605} dummy_bpf_prog = { 2606 .prog = { 2607 .bpf_func = __bpf_prog_ret1, 2608 }, 2609}; 2610 2611struct bpf_empty_prog_array bpf_empty_prog_array = { 2612 .null_prog = NULL, 2613}; 2614EXPORT_SYMBOL(bpf_empty_prog_array); 2615 2616struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags) 2617{ 2618 struct bpf_prog_array *p; 2619 2620 if (prog_cnt) 2621 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags); 2622 else 2623 p = &bpf_empty_prog_array.hdr; 2624 2625 return p; 2626} 2627 2628void bpf_prog_array_free(struct bpf_prog_array *progs) 2629{ 2630 if (!progs || progs == &bpf_empty_prog_array.hdr) 2631 return; 2632 kfree_rcu(progs, rcu); 2633} 2634 2635static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu) 2636{ 2637 struct bpf_prog_array *progs; 2638 2639 /* If RCU Tasks Trace grace period implies RCU grace period, there is 2640 * no need to call kfree_rcu(), just call kfree() directly. 2641 */ 2642 progs = container_of(rcu, struct bpf_prog_array, rcu); 2643 if (rcu_trace_implies_rcu_gp()) 2644 kfree(progs); 2645 else 2646 kfree_rcu(progs, rcu); 2647} 2648 2649void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs) 2650{ 2651 if (!progs || progs == &bpf_empty_prog_array.hdr) 2652 return; 2653 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb); 2654} 2655 2656int bpf_prog_array_length(struct bpf_prog_array *array) 2657{ 2658 struct bpf_prog_array_item *item; 2659 u32 cnt = 0; 2660 2661 for (item = array->items; item->prog; item++) 2662 if (item->prog != &dummy_bpf_prog.prog) 2663 cnt++; 2664 return cnt; 2665} 2666 2667bool bpf_prog_array_is_empty(struct bpf_prog_array *array) 2668{ 2669 struct bpf_prog_array_item *item; 2670 2671 for (item = array->items; item->prog; item++) 2672 if (item->prog != &dummy_bpf_prog.prog) 2673 return false; 2674 return true; 2675} 2676 2677static bool bpf_prog_array_copy_core(struct bpf_prog_array *array, 2678 u32 *prog_ids, 2679 u32 request_cnt) 2680{ 2681 struct bpf_prog_array_item *item; 2682 int i = 0; 2683 2684 for (item = array->items; item->prog; item++) { 2685 if (item->prog == &dummy_bpf_prog.prog) 2686 continue; 2687 prog_ids[i] = item->prog->aux->id; 2688 if (++i == request_cnt) { 2689 item++; 2690 break; 2691 } 2692 } 2693 2694 return !!(item->prog); 2695} 2696 2697int bpf_prog_array_copy_to_user(struct bpf_prog_array *array, 2698 __u32 __user *prog_ids, u32 cnt) 2699{ 2700 unsigned long err = 0; 2701 bool nospc; 2702 u32 *ids; 2703 2704 /* users of this function are doing: 2705 * cnt = bpf_prog_array_length(); 2706 * if (cnt > 0) 2707 * bpf_prog_array_copy_to_user(..., cnt); 2708 * so below kcalloc doesn't need extra cnt > 0 check. 2709 */ 2710 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN); 2711 if (!ids) 2712 return -ENOMEM; 2713 nospc = bpf_prog_array_copy_core(array, ids, cnt); 2714 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32)); 2715 kfree(ids); 2716 if (err) 2717 return -EFAULT; 2718 if (nospc) 2719 return -ENOSPC; 2720 return 0; 2721} 2722 2723void bpf_prog_array_delete_safe(struct bpf_prog_array *array, 2724 struct bpf_prog *old_prog) 2725{ 2726 struct bpf_prog_array_item *item; 2727 2728 for (item = array->items; item->prog; item++) 2729 if (item->prog == old_prog) { 2730 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog); 2731 break; 2732 } 2733} 2734 2735/** 2736 * bpf_prog_array_delete_safe_at() - Replaces the program at the given 2737 * index into the program array with 2738 * a dummy no-op program. 2739 * @array: a bpf_prog_array 2740 * @index: the index of the program to replace 2741 * 2742 * Skips over dummy programs, by not counting them, when calculating 2743 * the position of the program to replace. 2744 * 2745 * Return: 2746 * * 0 - Success 2747 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2748 * * -ENOENT - Index out of range 2749 */ 2750int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index) 2751{ 2752 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog); 2753} 2754 2755/** 2756 * bpf_prog_array_update_at() - Updates the program at the given index 2757 * into the program array. 2758 * @array: a bpf_prog_array 2759 * @index: the index of the program to update 2760 * @prog: the program to insert into the array 2761 * 2762 * Skips over dummy programs, by not counting them, when calculating 2763 * the position of the program to update. 2764 * 2765 * Return: 2766 * * 0 - Success 2767 * * -EINVAL - Invalid index value. Must be a non-negative integer. 2768 * * -ENOENT - Index out of range 2769 */ 2770int bpf_prog_array_update_at(struct bpf_prog_array *array, int index, 2771 struct bpf_prog *prog) 2772{ 2773 struct bpf_prog_array_item *item; 2774 2775 if (unlikely(index < 0)) 2776 return -EINVAL; 2777 2778 for (item = array->items; item->prog; item++) { 2779 if (item->prog == &dummy_bpf_prog.prog) 2780 continue; 2781 if (!index) { 2782 WRITE_ONCE(item->prog, prog); 2783 return 0; 2784 } 2785 index--; 2786 } 2787 return -ENOENT; 2788} 2789 2790int bpf_prog_array_copy(struct bpf_prog_array *old_array, 2791 struct bpf_prog *exclude_prog, 2792 struct bpf_prog *include_prog, 2793 u64 bpf_cookie, 2794 struct bpf_prog_array **new_array) 2795{ 2796 int new_prog_cnt, carry_prog_cnt = 0; 2797 struct bpf_prog_array_item *existing, *new; 2798 struct bpf_prog_array *array; 2799 bool found_exclude = false; 2800 2801 /* Figure out how many existing progs we need to carry over to 2802 * the new array. 2803 */ 2804 if (old_array) { 2805 existing = old_array->items; 2806 for (; existing->prog; existing++) { 2807 if (existing->prog == exclude_prog) { 2808 found_exclude = true; 2809 continue; 2810 } 2811 if (existing->prog != &dummy_bpf_prog.prog) 2812 carry_prog_cnt++; 2813 if (existing->prog == include_prog) 2814 return -EEXIST; 2815 } 2816 } 2817 2818 if (exclude_prog && !found_exclude) 2819 return -ENOENT; 2820 2821 /* How many progs (not NULL) will be in the new array? */ 2822 new_prog_cnt = carry_prog_cnt; 2823 if (include_prog) 2824 new_prog_cnt += 1; 2825 2826 /* Do we have any prog (not NULL) in the new array? */ 2827 if (!new_prog_cnt) { 2828 *new_array = NULL; 2829 return 0; 2830 } 2831 2832 /* +1 as the end of prog_array is marked with NULL */ 2833 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL); 2834 if (!array) 2835 return -ENOMEM; 2836 new = array->items; 2837 2838 /* Fill in the new prog array */ 2839 if (carry_prog_cnt) { 2840 existing = old_array->items; 2841 for (; existing->prog; existing++) { 2842 if (existing->prog == exclude_prog || 2843 existing->prog == &dummy_bpf_prog.prog) 2844 continue; 2845 2846 new->prog = existing->prog; 2847 new->bpf_cookie = existing->bpf_cookie; 2848 new++; 2849 } 2850 } 2851 if (include_prog) { 2852 new->prog = include_prog; 2853 new->bpf_cookie = bpf_cookie; 2854 new++; 2855 } 2856 new->prog = NULL; 2857 *new_array = array; 2858 return 0; 2859} 2860 2861int bpf_prog_array_copy_info(struct bpf_prog_array *array, 2862 u32 *prog_ids, u32 request_cnt, 2863 u32 *prog_cnt) 2864{ 2865 u32 cnt = 0; 2866 2867 if (array) 2868 cnt = bpf_prog_array_length(array); 2869 2870 *prog_cnt = cnt; 2871 2872 /* return early if user requested only program count or nothing to copy */ 2873 if (!request_cnt || !cnt) 2874 return 0; 2875 2876 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */ 2877 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC 2878 : 0; 2879} 2880 2881void __bpf_free_used_maps(struct bpf_prog_aux *aux, 2882 struct bpf_map **used_maps, u32 len) 2883{ 2884 struct bpf_map *map; 2885 bool sleepable; 2886 u32 i; 2887 2888 sleepable = aux->prog->sleepable; 2889 for (i = 0; i < len; i++) { 2890 map = used_maps[i]; 2891 if (map->ops->map_poke_untrack) 2892 map->ops->map_poke_untrack(map, aux); 2893 if (sleepable) 2894 atomic64_dec(&map->sleepable_refcnt); 2895 bpf_map_put(map); 2896 } 2897} 2898 2899static void bpf_free_used_maps(struct bpf_prog_aux *aux) 2900{ 2901 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt); 2902 kfree(aux->used_maps); 2903} 2904 2905void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len) 2906{ 2907#ifdef CONFIG_BPF_SYSCALL 2908 struct btf_mod_pair *btf_mod; 2909 u32 i; 2910 2911 for (i = 0; i < len; i++) { 2912 btf_mod = &used_btfs[i]; 2913 if (btf_mod->module) 2914 module_put(btf_mod->module); 2915 btf_put(btf_mod->btf); 2916 } 2917#endif 2918} 2919 2920static void bpf_free_used_btfs(struct bpf_prog_aux *aux) 2921{ 2922 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt); 2923 kfree(aux->used_btfs); 2924} 2925 2926static void bpf_prog_free_deferred(struct work_struct *work) 2927{ 2928 struct bpf_prog_aux *aux; 2929 int i; 2930 2931 aux = container_of(work, struct bpf_prog_aux, work); 2932#ifdef CONFIG_BPF_SYSCALL 2933 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab); 2934 bpf_prog_stream_free(aux->prog); 2935#endif 2936#ifdef CONFIG_CGROUP_BPF 2937 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID) 2938 bpf_cgroup_atype_put(aux->cgroup_atype); 2939#endif 2940 bpf_free_used_maps(aux); 2941 bpf_free_used_btfs(aux); 2942 if (bpf_prog_is_dev_bound(aux)) 2943 bpf_prog_dev_bound_destroy(aux->prog); 2944#ifdef CONFIG_PERF_EVENTS 2945 if (aux->prog->has_callchain_buf) 2946 put_callchain_buffers(); 2947#endif 2948 if (aux->dst_trampoline) 2949 bpf_trampoline_put(aux->dst_trampoline); 2950 for (i = 0; i < aux->real_func_cnt; i++) { 2951 /* We can just unlink the subprog poke descriptor table as 2952 * it was originally linked to the main program and is also 2953 * released along with it. 2954 */ 2955 aux->func[i]->aux->poke_tab = NULL; 2956 bpf_jit_free(aux->func[i]); 2957 } 2958 if (aux->real_func_cnt) { 2959 kfree(aux->func); 2960 bpf_prog_unlock_free(aux->prog); 2961 } else { 2962 bpf_jit_free(aux->prog); 2963 } 2964} 2965 2966void bpf_prog_free(struct bpf_prog *fp) 2967{ 2968 struct bpf_prog_aux *aux = fp->aux; 2969 2970 if (aux->dst_prog) 2971 bpf_prog_put(aux->dst_prog); 2972 bpf_token_put(aux->token); 2973 INIT_WORK(&aux->work, bpf_prog_free_deferred); 2974 schedule_work(&aux->work); 2975} 2976EXPORT_SYMBOL_GPL(bpf_prog_free); 2977 2978/* RNG for unprivileged user space with separated state from prandom_u32(). */ 2979static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state); 2980 2981void bpf_user_rnd_init_once(void) 2982{ 2983 prandom_init_once(&bpf_user_rnd_state); 2984} 2985 2986BPF_CALL_0(bpf_user_rnd_u32) 2987{ 2988 /* Should someone ever have the rather unwise idea to use some 2989 * of the registers passed into this function, then note that 2990 * this function is called from native eBPF and classic-to-eBPF 2991 * transformations. Register assignments from both sides are 2992 * different, f.e. classic always sets fn(ctx, A, X) here. 2993 */ 2994 struct rnd_state *state; 2995 u32 res; 2996 2997 state = &get_cpu_var(bpf_user_rnd_state); 2998 res = prandom_u32_state(state); 2999 put_cpu_var(bpf_user_rnd_state); 3000 3001 return res; 3002} 3003 3004BPF_CALL_0(bpf_get_raw_cpu_id) 3005{ 3006 return raw_smp_processor_id(); 3007} 3008 3009/* Weak definitions of helper functions in case we don't have bpf syscall. */ 3010const struct bpf_func_proto bpf_map_lookup_elem_proto __weak; 3011const struct bpf_func_proto bpf_map_update_elem_proto __weak; 3012const struct bpf_func_proto bpf_map_delete_elem_proto __weak; 3013const struct bpf_func_proto bpf_map_push_elem_proto __weak; 3014const struct bpf_func_proto bpf_map_pop_elem_proto __weak; 3015const struct bpf_func_proto bpf_map_peek_elem_proto __weak; 3016const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak; 3017const struct bpf_func_proto bpf_spin_lock_proto __weak; 3018const struct bpf_func_proto bpf_spin_unlock_proto __weak; 3019const struct bpf_func_proto bpf_jiffies64_proto __weak; 3020 3021const struct bpf_func_proto bpf_get_prandom_u32_proto __weak; 3022const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak; 3023const struct bpf_func_proto bpf_get_numa_node_id_proto __weak; 3024const struct bpf_func_proto bpf_ktime_get_ns_proto __weak; 3025const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak; 3026const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak; 3027const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak; 3028 3029const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak; 3030const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak; 3031const struct bpf_func_proto bpf_get_current_comm_proto __weak; 3032const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak; 3033const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak; 3034const struct bpf_func_proto bpf_get_local_storage_proto __weak; 3035const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak; 3036const struct bpf_func_proto bpf_snprintf_btf_proto __weak; 3037const struct bpf_func_proto bpf_seq_printf_btf_proto __weak; 3038const struct bpf_func_proto bpf_set_retval_proto __weak; 3039const struct bpf_func_proto bpf_get_retval_proto __weak; 3040 3041const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void) 3042{ 3043 return NULL; 3044} 3045 3046const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void) 3047{ 3048 return NULL; 3049} 3050 3051const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void) 3052{ 3053 return NULL; 3054} 3055 3056u64 __weak 3057bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size, 3058 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy) 3059{ 3060 return -ENOTSUPP; 3061} 3062EXPORT_SYMBOL_GPL(bpf_event_output); 3063 3064/* Always built-in helper functions. */ 3065const struct bpf_func_proto bpf_tail_call_proto = { 3066 /* func is unused for tail_call, we set it to pass the 3067 * get_helper_proto check 3068 */ 3069 .func = BPF_PTR_POISON, 3070 .gpl_only = false, 3071 .ret_type = RET_VOID, 3072 .arg1_type = ARG_PTR_TO_CTX, 3073 .arg2_type = ARG_CONST_MAP_PTR, 3074 .arg3_type = ARG_ANYTHING, 3075}; 3076 3077/* Stub for JITs that only support cBPF. eBPF programs are interpreted. 3078 * It is encouraged to implement bpf_int_jit_compile() instead, so that 3079 * eBPF and implicitly also cBPF can get JITed! 3080 */ 3081struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog) 3082{ 3083 return prog; 3084} 3085 3086/* Stub for JITs that support eBPF. All cBPF code gets transformed into 3087 * eBPF by the kernel and is later compiled by bpf_int_jit_compile(). 3088 */ 3089void __weak bpf_jit_compile(struct bpf_prog *prog) 3090{ 3091} 3092 3093bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id) 3094{ 3095 return false; 3096} 3097 3098/* Return TRUE if the JIT backend wants verifier to enable sub-register usage 3099 * analysis code and wants explicit zero extension inserted by verifier. 3100 * Otherwise, return FALSE. 3101 * 3102 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if 3103 * you don't override this. JITs that don't want these extra insns can detect 3104 * them using insn_is_zext. 3105 */ 3106bool __weak bpf_jit_needs_zext(void) 3107{ 3108 return false; 3109} 3110 3111/* By default, enable the verifier's mitigations against Spectre v1 and v4 for 3112 * all archs. The value returned must not change at runtime as there is 3113 * currently no support for reloading programs that were loaded without 3114 * mitigations. 3115 */ 3116bool __weak bpf_jit_bypass_spec_v1(void) 3117{ 3118 return false; 3119} 3120 3121bool __weak bpf_jit_bypass_spec_v4(void) 3122{ 3123 return false; 3124} 3125 3126/* Return true if the JIT inlines the call to the helper corresponding to 3127 * the imm. 3128 * 3129 * The verifier will not patch the insn->imm for the call to the helper if 3130 * this returns true. 3131 */ 3132bool __weak bpf_jit_inlines_helper_call(s32 imm) 3133{ 3134 return false; 3135} 3136 3137/* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */ 3138bool __weak bpf_jit_supports_subprog_tailcalls(void) 3139{ 3140 return false; 3141} 3142 3143bool __weak bpf_jit_supports_percpu_insn(void) 3144{ 3145 return false; 3146} 3147 3148bool __weak bpf_jit_supports_kfunc_call(void) 3149{ 3150 return false; 3151} 3152 3153bool __weak bpf_jit_supports_far_kfunc_call(void) 3154{ 3155 return false; 3156} 3157 3158bool __weak bpf_jit_supports_arena(void) 3159{ 3160 return false; 3161} 3162 3163bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena) 3164{ 3165 return false; 3166} 3167 3168u64 __weak bpf_arch_uaddress_limit(void) 3169{ 3170#if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE) 3171 return TASK_SIZE; 3172#else 3173 return 0; 3174#endif 3175} 3176 3177/* Return TRUE if the JIT backend satisfies the following two conditions: 3178 * 1) JIT backend supports atomic_xchg() on pointer-sized words. 3179 * 2) Under the specific arch, the implementation of xchg() is the same 3180 * as atomic_xchg() on pointer-sized words. 3181 */ 3182bool __weak bpf_jit_supports_ptr_xchg(void) 3183{ 3184 return false; 3185} 3186 3187/* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call 3188 * skb_copy_bits(), so provide a weak definition of it for NET-less config. 3189 */ 3190int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to, 3191 int len) 3192{ 3193 return -EFAULT; 3194} 3195 3196int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t, 3197 enum bpf_text_poke_type new_t, void *old_addr, 3198 void *new_addr) 3199{ 3200 return -ENOTSUPP; 3201} 3202 3203void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len) 3204{ 3205 return ERR_PTR(-ENOTSUPP); 3206} 3207 3208int __weak bpf_arch_text_invalidate(void *dst, size_t len) 3209{ 3210 return -ENOTSUPP; 3211} 3212 3213bool __weak bpf_jit_supports_exceptions(void) 3214{ 3215 return false; 3216} 3217 3218bool __weak bpf_jit_supports_private_stack(void) 3219{ 3220 return false; 3221} 3222 3223void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie) 3224{ 3225} 3226 3227bool __weak bpf_jit_supports_timed_may_goto(void) 3228{ 3229 return false; 3230} 3231 3232u64 __weak arch_bpf_timed_may_goto(void) 3233{ 3234 return 0; 3235} 3236 3237static noinline void bpf_prog_report_may_goto_violation(void) 3238{ 3239#ifdef CONFIG_BPF_SYSCALL 3240 struct bpf_stream_stage ss; 3241 struct bpf_prog *prog; 3242 3243 prog = bpf_prog_find_from_stack(); 3244 if (!prog) 3245 return; 3246 bpf_stream_stage(ss, prog, BPF_STDERR, ({ 3247 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n"); 3248 bpf_stream_dump_stack(ss); 3249 })); 3250#endif 3251} 3252 3253u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p) 3254{ 3255 u64 time = ktime_get_mono_fast_ns(); 3256 3257 /* Populate the timestamp for this stack frame, and refresh count. */ 3258 if (!p->timestamp) { 3259 p->timestamp = time; 3260 return BPF_MAX_TIMED_LOOPS; 3261 } 3262 /* Check if we've exhausted our time slice, and zero count. */ 3263 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) { 3264 bpf_prog_report_may_goto_violation(); 3265 return 0; 3266 } 3267 /* Refresh the count for the stack frame. */ 3268 return BPF_MAX_TIMED_LOOPS; 3269} 3270 3271/* for configs without MMU or 32-bit */ 3272__weak const struct bpf_map_ops arena_map_ops; 3273__weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena) 3274{ 3275 return 0; 3276} 3277__weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena) 3278{ 3279 return 0; 3280} 3281 3282#ifdef CONFIG_BPF_SYSCALL 3283static int __init bpf_global_ma_init(void) 3284{ 3285 int ret; 3286 3287 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false); 3288 bpf_global_ma_set = !ret; 3289 return ret; 3290} 3291late_initcall(bpf_global_ma_init); 3292#endif 3293 3294DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key); 3295EXPORT_SYMBOL(bpf_stats_enabled_key); 3296 3297/* All definitions of tracepoints related to BPF. */ 3298#define CREATE_TRACE_POINTS 3299#include <linux/bpf_trace.h> 3300 3301EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception); 3302EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx); 3303 3304#ifdef CONFIG_BPF_SYSCALL 3305 3306int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep, 3307 const char **linep, int *nump) 3308{ 3309 int idx = -1, insn_start, insn_end, len; 3310 struct bpf_line_info *linfo; 3311 void **jited_linfo; 3312 struct btf *btf; 3313 int nr_linfo; 3314 3315 btf = prog->aux->btf; 3316 linfo = prog->aux->linfo; 3317 jited_linfo = prog->aux->jited_linfo; 3318 3319 if (!btf || !linfo || !jited_linfo) 3320 return -EINVAL; 3321 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len; 3322 3323 linfo = &prog->aux->linfo[prog->aux->linfo_idx]; 3324 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx]; 3325 3326 insn_start = linfo[0].insn_off; 3327 insn_end = insn_start + len; 3328 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx; 3329 3330 for (int i = 0; i < nr_linfo && 3331 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) { 3332 if (jited_linfo[i] >= (void *)ip) 3333 break; 3334 idx = i; 3335 } 3336 3337 if (idx == -1) 3338 return -ENOENT; 3339 3340 /* Get base component of the file path. */ 3341 *filep = btf_name_by_offset(btf, linfo[idx].file_name_off); 3342 *filep = kbasename(*filep); 3343 /* Obtain the source line, and strip whitespace in prefix. */ 3344 *linep = btf_name_by_offset(btf, linfo[idx].line_off); 3345 while (isspace(**linep)) 3346 *linep += 1; 3347 *nump = BPF_LINE_INFO_LINE_NUM(linfo[idx].line_col); 3348 return 0; 3349} 3350 3351struct walk_stack_ctx { 3352 struct bpf_prog *prog; 3353}; 3354 3355static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp) 3356{ 3357 struct walk_stack_ctx *ctxp = cookie; 3358 struct bpf_prog *prog; 3359 3360 /* 3361 * The RCU read lock is held to safely traverse the latch tree, but we 3362 * don't need its protection when accessing the prog, since it has an 3363 * active stack frame on the current stack trace, and won't disappear. 3364 */ 3365 rcu_read_lock(); 3366 prog = bpf_prog_ksym_find(ip); 3367 rcu_read_unlock(); 3368 if (!prog) 3369 return true; 3370 /* Make sure we return the main prog if we found a subprog */ 3371 ctxp->prog = prog->aux->main_prog_aux->prog; 3372 return false; 3373} 3374 3375struct bpf_prog *bpf_prog_find_from_stack(void) 3376{ 3377 struct walk_stack_ctx ctx = {}; 3378 3379 arch_bpf_stack_walk(find_from_stack_cb, &ctx); 3380 return ctx.prog; 3381} 3382 3383#endif
~/
Read-only file preview.