1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2016 Facebook 3 */ 4 #include <linux/bpf.h> 5 #include <linux/jhash.h> 6 #include <linux/filter.h> 7 #include <linux/stacktrace.h> 8 #include <linux/perf_event.h> 9 #include <linux/elf.h> 10 #include <linux/pagemap.h> 11 #include <linux/irq_work.h> 12 #include "percpu_freelist.h" 13 14 #define STACK_CREATE_FLAG_MASK \ 15 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \ 16 BPF_F_STACK_BUILD_ID) 17 18 struct stack_map_bucket { 19 struct pcpu_freelist_node fnode; 20 u32 hash; 21 u32 nr; 22 u64 data[]; 23 }; 24 25 struct bpf_stack_map { 26 struct bpf_map map; 27 void *elems; 28 struct pcpu_freelist freelist; 29 u32 n_buckets; 30 struct stack_map_bucket *buckets[]; 31 }; 32 33 /* irq_work to run up_read() for build_id lookup in nmi context */ 34 struct stack_map_irq_work { 35 struct irq_work irq_work; 36 struct rw_semaphore *sem; 37 }; 38 39 static void do_up_read(struct irq_work *entry) 40 { 41 struct stack_map_irq_work *work; 42 43 work = container_of(entry, struct stack_map_irq_work, irq_work); 44 up_read_non_owner(work->sem); 45 work->sem = NULL; 46 } 47 48 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work); 49 50 static inline bool stack_map_use_build_id(struct bpf_map *map) 51 { 52 return (map->map_flags & BPF_F_STACK_BUILD_ID); 53 } 54 55 static inline int stack_map_data_size(struct bpf_map *map) 56 { 57 return stack_map_use_build_id(map) ? 58 sizeof(struct bpf_stack_build_id) : sizeof(u64); 59 } 60 61 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap) 62 { 63 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size; 64 int err; 65 66 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries, 67 smap->map.numa_node); 68 if (!smap->elems) 69 return -ENOMEM; 70 71 err = pcpu_freelist_init(&smap->freelist); 72 if (err) 73 goto free_elems; 74 75 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size, 76 smap->map.max_entries); 77 return 0; 78 79 free_elems: 80 bpf_map_area_free(smap->elems); 81 return err; 82 } 83 84 /* Called from syscall */ 85 static struct bpf_map *stack_map_alloc(union bpf_attr *attr) 86 { 87 u32 value_size = attr->value_size; 88 struct bpf_stack_map *smap; 89 struct bpf_map_memory mem; 90 u64 cost, n_buckets; 91 int err; 92 93 if (!capable(CAP_SYS_ADMIN)) 94 return ERR_PTR(-EPERM); 95 96 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK) 97 return ERR_PTR(-EINVAL); 98 99 /* check sanity of attributes */ 100 if (attr->max_entries == 0 || attr->key_size != 4 || 101 value_size < 8 || value_size % 8) 102 return ERR_PTR(-EINVAL); 103 104 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64)); 105 if (attr->map_flags & BPF_F_STACK_BUILD_ID) { 106 if (value_size % sizeof(struct bpf_stack_build_id) || 107 value_size / sizeof(struct bpf_stack_build_id) 108 > sysctl_perf_event_max_stack) 109 return ERR_PTR(-EINVAL); 110 } else if (value_size / 8 > sysctl_perf_event_max_stack) 111 return ERR_PTR(-EINVAL); 112 113 /* hash table size must be power of 2 */ 114 n_buckets = roundup_pow_of_two(attr->max_entries); 115 116 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap); 117 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket)); 118 err = bpf_map_charge_init(&mem, cost); 119 if (err) 120 return ERR_PTR(err); 121 122 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr)); 123 if (!smap) { 124 bpf_map_charge_finish(&mem); 125 return ERR_PTR(-ENOMEM); 126 } 127 128 bpf_map_init_from_attr(&smap->map, attr); 129 smap->map.value_size = value_size; 130 smap->n_buckets = n_buckets; 131 132 err = get_callchain_buffers(sysctl_perf_event_max_stack); 133 if (err) 134 goto free_charge; 135 136 err = prealloc_elems_and_freelist(smap); 137 if (err) 138 goto put_buffers; 139 140 bpf_map_charge_move(&smap->map.memory, &mem); 141 142 return &smap->map; 143 144 put_buffers: 145 put_callchain_buffers(); 146 free_charge: 147 bpf_map_charge_finish(&mem); 148 bpf_map_area_free(smap); 149 return ERR_PTR(err); 150 } 151 152 #define BPF_BUILD_ID 3 153 /* 154 * Parse build id from the note segment. This logic can be shared between 155 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are 156 * identical. 157 */ 158 static inline int stack_map_parse_build_id(void *page_addr, 159 unsigned char *build_id, 160 void *note_start, 161 Elf32_Word note_size) 162 { 163 Elf32_Word note_offs = 0, new_offs; 164 165 /* check for overflow */ 166 if (note_start < page_addr || note_start + note_size < note_start) 167 return -EINVAL; 168 169 /* only supports note that fits in the first page */ 170 if (note_start + note_size > page_addr + PAGE_SIZE) 171 return -EINVAL; 172 173 while (note_offs + sizeof(Elf32_Nhdr) < note_size) { 174 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs); 175 176 if (nhdr->n_type == BPF_BUILD_ID && 177 nhdr->n_namesz == sizeof("GNU") && 178 nhdr->n_descsz > 0 && 179 nhdr->n_descsz <= BPF_BUILD_ID_SIZE) { 180 memcpy(build_id, 181 note_start + note_offs + 182 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr), 183 nhdr->n_descsz); 184 memset(build_id + nhdr->n_descsz, 0, 185 BPF_BUILD_ID_SIZE - nhdr->n_descsz); 186 return 0; 187 } 188 new_offs = note_offs + sizeof(Elf32_Nhdr) + 189 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4); 190 if (new_offs <= note_offs) /* overflow */ 191 break; 192 note_offs = new_offs; 193 } 194 return -EINVAL; 195 } 196 197 /* Parse build ID from 32-bit ELF */ 198 static int stack_map_get_build_id_32(void *page_addr, 199 unsigned char *build_id) 200 { 201 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr; 202 Elf32_Phdr *phdr; 203 int i; 204 205 /* only supports phdr that fits in one page */ 206 if (ehdr->e_phnum > 207 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr)) 208 return -EINVAL; 209 210 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr)); 211 212 for (i = 0; i < ehdr->e_phnum; ++i) 213 if (phdr[i].p_type == PT_NOTE) 214 return stack_map_parse_build_id(page_addr, build_id, 215 page_addr + phdr[i].p_offset, 216 phdr[i].p_filesz); 217 return -EINVAL; 218 } 219 220 /* Parse build ID from 64-bit ELF */ 221 static int stack_map_get_build_id_64(void *page_addr, 222 unsigned char *build_id) 223 { 224 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr; 225 Elf64_Phdr *phdr; 226 int i; 227 228 /* only supports phdr that fits in one page */ 229 if (ehdr->e_phnum > 230 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr)) 231 return -EINVAL; 232 233 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr)); 234 235 for (i = 0; i < ehdr->e_phnum; ++i) 236 if (phdr[i].p_type == PT_NOTE) 237 return stack_map_parse_build_id(page_addr, build_id, 238 page_addr + phdr[i].p_offset, 239 phdr[i].p_filesz); 240 return -EINVAL; 241 } 242 243 /* Parse build ID of ELF file mapped to vma */ 244 static int stack_map_get_build_id(struct vm_area_struct *vma, 245 unsigned char *build_id) 246 { 247 Elf32_Ehdr *ehdr; 248 struct page *page; 249 void *page_addr; 250 int ret; 251 252 /* only works for page backed storage */ 253 if (!vma->vm_file) 254 return -EINVAL; 255 256 page = find_get_page(vma->vm_file->f_mapping, 0); 257 if (!page) 258 return -EFAULT; /* page not mapped */ 259 260 ret = -EINVAL; 261 page_addr = kmap_atomic(page); 262 ehdr = (Elf32_Ehdr *)page_addr; 263 264 /* compare magic x7f "ELF" */ 265 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0) 266 goto out; 267 268 /* only support executable file and shared object file */ 269 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN) 270 goto out; 271 272 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32) 273 ret = stack_map_get_build_id_32(page_addr, build_id); 274 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64) 275 ret = stack_map_get_build_id_64(page_addr, build_id); 276 out: 277 kunmap_atomic(page_addr); 278 put_page(page); 279 return ret; 280 } 281 282 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs, 283 u64 *ips, u32 trace_nr, bool user) 284 { 285 int i; 286 struct vm_area_struct *vma; 287 bool irq_work_busy = false; 288 struct stack_map_irq_work *work = NULL; 289 290 if (irqs_disabled()) { 291 work = this_cpu_ptr(&up_read_work); 292 if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY) 293 /* cannot queue more up_read, fallback */ 294 irq_work_busy = true; 295 } 296 297 /* 298 * We cannot do up_read() when the irq is disabled, because of 299 * risk to deadlock with rq_lock. To do build_id lookup when the 300 * irqs are disabled, we need to run up_read() in irq_work. We use 301 * a percpu variable to do the irq_work. If the irq_work is 302 * already used by another lookup, we fall back to report ips. 303 * 304 * Same fallback is used for kernel stack (!user) on a stackmap 305 * with build_id. 306 */ 307 if (!user || !current || !current->mm || irq_work_busy || 308 down_read_trylock(¤t->mm->mmap_sem) == 0) { 309 /* cannot access current->mm, fall back to ips */ 310 for (i = 0; i < trace_nr; i++) { 311 id_offs[i].status = BPF_STACK_BUILD_ID_IP; 312 id_offs[i].ip = ips[i]; 313 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE); 314 } 315 return; 316 } 317 318 for (i = 0; i < trace_nr; i++) { 319 vma = find_vma(current->mm, ips[i]); 320 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) { 321 /* per entry fall back to ips */ 322 id_offs[i].status = BPF_STACK_BUILD_ID_IP; 323 id_offs[i].ip = ips[i]; 324 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE); 325 continue; 326 } 327 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i] 328 - vma->vm_start; 329 id_offs[i].status = BPF_STACK_BUILD_ID_VALID; 330 } 331 332 if (!work) { 333 up_read(¤t->mm->mmap_sem); 334 } else { 335 work->sem = ¤t->mm->mmap_sem; 336 irq_work_queue(&work->irq_work); 337 /* 338 * The irq_work will release the mmap_sem with 339 * up_read_non_owner(). The rwsem_release() is called 340 * here to release the lock from lockdep's perspective. 341 */ 342 rwsem_release(¤t->mm->mmap_sem.dep_map, _RET_IP_); 343 } 344 } 345 346 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map, 347 u64, flags) 348 { 349 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map); 350 struct perf_callchain_entry *trace; 351 struct stack_map_bucket *bucket, *new_bucket, *old_bucket; 352 u32 max_depth = map->value_size / stack_map_data_size(map); 353 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */ 354 u32 init_nr = sysctl_perf_event_max_stack - max_depth; 355 u32 skip = flags & BPF_F_SKIP_FIELD_MASK; 356 u32 hash, id, trace_nr, trace_len; 357 bool user = flags & BPF_F_USER_STACK; 358 bool kernel = !user; 359 u64 *ips; 360 bool hash_matches; 361 362 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK | 363 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID))) 364 return -EINVAL; 365 366 trace = get_perf_callchain(regs, init_nr, kernel, user, 367 sysctl_perf_event_max_stack, false, false); 368 369 if (unlikely(!trace)) 370 /* couldn't fetch the stack trace */ 371 return -EFAULT; 372 373 /* get_perf_callchain() guarantees that trace->nr >= init_nr 374 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth 375 */ 376 trace_nr = trace->nr - init_nr; 377 378 if (trace_nr <= skip) 379 /* skipping more than usable stack trace */ 380 return -EFAULT; 381 382 trace_nr -= skip; 383 trace_len = trace_nr * sizeof(u64); 384 ips = trace->ip + skip + init_nr; 385 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0); 386 id = hash & (smap->n_buckets - 1); 387 bucket = READ_ONCE(smap->buckets[id]); 388 389 hash_matches = bucket && bucket->hash == hash; 390 /* fast cmp */ 391 if (hash_matches && flags & BPF_F_FAST_STACK_CMP) 392 return id; 393 394 if (stack_map_use_build_id(map)) { 395 /* for build_id+offset, pop a bucket before slow cmp */ 396 new_bucket = (struct stack_map_bucket *) 397 pcpu_freelist_pop(&smap->freelist); 398 if (unlikely(!new_bucket)) 399 return -ENOMEM; 400 new_bucket->nr = trace_nr; 401 stack_map_get_build_id_offset( 402 (struct bpf_stack_build_id *)new_bucket->data, 403 ips, trace_nr, user); 404 trace_len = trace_nr * sizeof(struct bpf_stack_build_id); 405 if (hash_matches && bucket->nr == trace_nr && 406 memcmp(bucket->data, new_bucket->data, trace_len) == 0) { 407 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode); 408 return id; 409 } 410 if (bucket && !(flags & BPF_F_REUSE_STACKID)) { 411 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode); 412 return -EEXIST; 413 } 414 } else { 415 if (hash_matches && bucket->nr == trace_nr && 416 memcmp(bucket->data, ips, trace_len) == 0) 417 return id; 418 if (bucket && !(flags & BPF_F_REUSE_STACKID)) 419 return -EEXIST; 420 421 new_bucket = (struct stack_map_bucket *) 422 pcpu_freelist_pop(&smap->freelist); 423 if (unlikely(!new_bucket)) 424 return -ENOMEM; 425 memcpy(new_bucket->data, ips, trace_len); 426 } 427 428 new_bucket->hash = hash; 429 new_bucket->nr = trace_nr; 430 431 old_bucket = xchg(&smap->buckets[id], new_bucket); 432 if (old_bucket) 433 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode); 434 return id; 435 } 436 437 const struct bpf_func_proto bpf_get_stackid_proto = { 438 .func = bpf_get_stackid, 439 .gpl_only = true, 440 .ret_type = RET_INTEGER, 441 .arg1_type = ARG_PTR_TO_CTX, 442 .arg2_type = ARG_CONST_MAP_PTR, 443 .arg3_type = ARG_ANYTHING, 444 }; 445 446 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size, 447 u64, flags) 448 { 449 u32 init_nr, trace_nr, copy_len, elem_size, num_elem; 450 bool user_build_id = flags & BPF_F_USER_BUILD_ID; 451 u32 skip = flags & BPF_F_SKIP_FIELD_MASK; 452 bool user = flags & BPF_F_USER_STACK; 453 struct perf_callchain_entry *trace; 454 bool kernel = !user; 455 int err = -EINVAL; 456 u64 *ips; 457 458 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK | 459 BPF_F_USER_BUILD_ID))) 460 goto clear; 461 if (kernel && user_build_id) 462 goto clear; 463 464 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id) 465 : sizeof(u64); 466 if (unlikely(size % elem_size)) 467 goto clear; 468 469 num_elem = size / elem_size; 470 if (sysctl_perf_event_max_stack < num_elem) 471 init_nr = 0; 472 else 473 init_nr = sysctl_perf_event_max_stack - num_elem; 474 trace = get_perf_callchain(regs, init_nr, kernel, user, 475 sysctl_perf_event_max_stack, false, false); 476 if (unlikely(!trace)) 477 goto err_fault; 478 479 trace_nr = trace->nr - init_nr; 480 if (trace_nr < skip) 481 goto err_fault; 482 483 trace_nr -= skip; 484 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem; 485 copy_len = trace_nr * elem_size; 486 ips = trace->ip + skip + init_nr; 487 if (user && user_build_id) 488 stack_map_get_build_id_offset(buf, ips, trace_nr, user); 489 else 490 memcpy(buf, ips, copy_len); 491 492 if (size > copy_len) 493 memset(buf + copy_len, 0, size - copy_len); 494 return copy_len; 495 496 err_fault: 497 err = -EFAULT; 498 clear: 499 memset(buf, 0, size); 500 return err; 501 } 502 503 const struct bpf_func_proto bpf_get_stack_proto = { 504 .func = bpf_get_stack, 505 .gpl_only = true, 506 .ret_type = RET_INTEGER, 507 .arg1_type = ARG_PTR_TO_CTX, 508 .arg2_type = ARG_PTR_TO_UNINIT_MEM, 509 .arg3_type = ARG_CONST_SIZE_OR_ZERO, 510 .arg4_type = ARG_ANYTHING, 511 }; 512 513 /* Called from eBPF program */ 514 static void *stack_map_lookup_elem(struct bpf_map *map, void *key) 515 { 516 return ERR_PTR(-EOPNOTSUPP); 517 } 518 519 /* Called from syscall */ 520 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value) 521 { 522 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map); 523 struct stack_map_bucket *bucket, *old_bucket; 524 u32 id = *(u32 *)key, trace_len; 525 526 if (unlikely(id >= smap->n_buckets)) 527 return -ENOENT; 528 529 bucket = xchg(&smap->buckets[id], NULL); 530 if (!bucket) 531 return -ENOENT; 532 533 trace_len = bucket->nr * stack_map_data_size(map); 534 memcpy(value, bucket->data, trace_len); 535 memset(value + trace_len, 0, map->value_size - trace_len); 536 537 old_bucket = xchg(&smap->buckets[id], bucket); 538 if (old_bucket) 539 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode); 540 return 0; 541 } 542 543 static int stack_map_get_next_key(struct bpf_map *map, void *key, 544 void *next_key) 545 { 546 struct bpf_stack_map *smap = container_of(map, 547 struct bpf_stack_map, map); 548 u32 id; 549 550 WARN_ON_ONCE(!rcu_read_lock_held()); 551 552 if (!key) { 553 id = 0; 554 } else { 555 id = *(u32 *)key; 556 if (id >= smap->n_buckets || !smap->buckets[id]) 557 id = 0; 558 else 559 id++; 560 } 561 562 while (id < smap->n_buckets && !smap->buckets[id]) 563 id++; 564 565 if (id >= smap->n_buckets) 566 return -ENOENT; 567 568 *(u32 *)next_key = id; 569 return 0; 570 } 571 572 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value, 573 u64 map_flags) 574 { 575 return -EINVAL; 576 } 577 578 /* Called from syscall or from eBPF program */ 579 static int stack_map_delete_elem(struct bpf_map *map, void *key) 580 { 581 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map); 582 struct stack_map_bucket *old_bucket; 583 u32 id = *(u32 *)key; 584 585 if (unlikely(id >= smap->n_buckets)) 586 return -E2BIG; 587 588 old_bucket = xchg(&smap->buckets[id], NULL); 589 if (old_bucket) { 590 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode); 591 return 0; 592 } else { 593 return -ENOENT; 594 } 595 } 596 597 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */ 598 static void stack_map_free(struct bpf_map *map) 599 { 600 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map); 601 602 /* wait for bpf programs to complete before freeing stack map */ 603 synchronize_rcu(); 604 605 bpf_map_area_free(smap->elems); 606 pcpu_freelist_destroy(&smap->freelist); 607 bpf_map_area_free(smap); 608 put_callchain_buffers(); 609 } 610 611 const struct bpf_map_ops stack_trace_map_ops = { 612 .map_alloc = stack_map_alloc, 613 .map_free = stack_map_free, 614 .map_get_next_key = stack_map_get_next_key, 615 .map_lookup_elem = stack_map_lookup_elem, 616 .map_update_elem = stack_map_update_elem, 617 .map_delete_elem = stack_map_delete_elem, 618 .map_check_btf = map_check_no_btf, 619 }; 620 621 static int __init stack_map_init(void) 622 { 623 int cpu; 624 struct stack_map_irq_work *work; 625 626 for_each_possible_cpu(cpu) { 627 work = per_cpu_ptr(&up_read_work, cpu); 628 init_irq_work(&work->irq_work, do_up_read); 629 } 630 return 0; 631 } 632 subsys_initcall(stack_map_init); 633