kernel/bpf/stackmap.c

   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         u64 elem_size = sizeof(struct stack_map_bucket) +
  64                         (u64)smap->map.value_size;
  65         int err;
  66
  67         smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
  68                                          smap->map.numa_node);
  69         if (!smap->elems)
  70                 return -ENOMEM;
  71
  72         err = pcpu_freelist_init(&smap->freelist);
  73         if (err)
  74                 goto free_elems;
  75
  76         pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
  77                                smap->map.max_entries);
  78         return 0;
  79
  80 free_elems:
  81         bpf_map_area_free(smap->elems);
  82         return err;
  83 }
  84
  85 /* Called from syscall */
  86 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
  87 {
  88         u32 value_size = attr->value_size;
  89         struct bpf_stack_map *smap;
  90         struct bpf_map_memory mem;
  91         u64 cost, n_buckets;
  92         int err;
  93
  94         if (!capable(CAP_SYS_ADMIN))
  95                 return ERR_PTR(-EPERM);
  96
  97         if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
  98                 return ERR_PTR(-EINVAL);
  99
 100         /* check sanity of attributes */
 101         if (attr->max_entries == 0 || attr->key_size != 4 ||
 102             value_size < 8 || value_size % 8)
 103                 return ERR_PTR(-EINVAL);
 104
 105         BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
 106         if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
 107                 if (value_size % sizeof(struct bpf_stack_build_id) ||
 108                     value_size / sizeof(struct bpf_stack_build_id)
 109                     > sysctl_perf_event_max_stack)
 110                         return ERR_PTR(-EINVAL);
 111         } else if (value_size / 8 > sysctl_perf_event_max_stack)
 112                 return ERR_PTR(-EINVAL);
 113
 114         /* hash table size must be power of 2 */
 115         n_buckets = roundup_pow_of_two(attr->max_entries);
 116         if (!n_buckets)
 117                 return ERR_PTR(-E2BIG);
 118
 119         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 120         err = bpf_map_charge_init(&mem, cost + attr->max_entries *
 121                            (sizeof(struct stack_map_bucket) + (u64)value_size));
 122         if (err)
 123                 return ERR_PTR(err);
 124
 125         smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
 126         if (!smap) {
 127                 bpf_map_charge_finish(&mem);
 128                 return ERR_PTR(-ENOMEM);
 129         }
 130
 131         bpf_map_init_from_attr(&smap->map, attr);
 132         smap->map.value_size = value_size;
 133         smap->n_buckets = n_buckets;
 134
 135         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 136         if (err)
 137                 goto free_charge;
 138
 139         err = prealloc_elems_and_freelist(smap);
 140         if (err)
 141                 goto put_buffers;
 142
 143         bpf_map_charge_move(&smap->map.memory, &mem);
 144
 145         return &smap->map;
 146
 147 put_buffers:
 148         put_callchain_buffers();
 149 free_charge:
 150         bpf_map_charge_finish(&mem);
 151         bpf_map_area_free(smap);
 152         return ERR_PTR(err);
 153 }
 154
 155 #define BPF_BUILD_ID 3
 156 /*
 157  * Parse build id from the note segment. This logic can be shared between
 158  * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
 159  * identical.
 160  */
 161 static inline int stack_map_parse_build_id(void *page_addr,
 162                                            unsigned char *build_id,
 163                                            void *note_start,
 164                                            Elf32_Word note_size)
 165 {
 166         Elf32_Word note_offs = 0, new_offs;
 167
 168         /* check for overflow */
 169         if (note_start < page_addr || note_start + note_size < note_start)
 170                 return -EINVAL;
 171
 172         /* only supports note that fits in the first page */
 173         if (note_start + note_size > page_addr + PAGE_SIZE)
 174                 return -EINVAL;
 175
 176         while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
 177                 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
 178
 179                 if (nhdr->n_type == BPF_BUILD_ID &&
 180                     nhdr->n_namesz == sizeof("GNU") &&
 181                     nhdr->n_descsz > 0 &&
 182                     nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
 183                         memcpy(build_id,
 184                                note_start + note_offs +
 185                                ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
 186                                nhdr->n_descsz);
 187                         memset(build_id + nhdr->n_descsz, 0,
 188                                BPF_BUILD_ID_SIZE - nhdr->n_descsz);
 189                         return 0;
 190                 }
 191                 new_offs = note_offs + sizeof(Elf32_Nhdr) +
 192                         ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
 193                 if (new_offs <= note_offs)  /* overflow */
 194                         break;
 195                 note_offs = new_offs;
 196         }
 197         return -EINVAL;
 198 }
 199
 200 /* Parse build ID from 32-bit ELF */
 201 static int stack_map_get_build_id_32(void *page_addr,
 202                                      unsigned char *build_id)
 203 {
 204         Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
 205         Elf32_Phdr *phdr;
 206         int i;
 207
 208         /* only supports phdr that fits in one page */
 209         if (ehdr->e_phnum >
 210             (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
 211                 return -EINVAL;
 212
 213         phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
 214
 215         for (i = 0; i < ehdr->e_phnum; ++i)
 216                 if (phdr[i].p_type == PT_NOTE)
 217                         return stack_map_parse_build_id(page_addr, build_id,
 218                                         page_addr + phdr[i].p_offset,
 219                                         phdr[i].p_filesz);
 220         return -EINVAL;
 221 }
 222
 223 /* Parse build ID from 64-bit ELF */
 224 static int stack_map_get_build_id_64(void *page_addr,
 225                                      unsigned char *build_id)
 226 {
 227         Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
 228         Elf64_Phdr *phdr;
 229         int i;
 230
 231         /* only supports phdr that fits in one page */
 232         if (ehdr->e_phnum >
 233             (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
 234                 return -EINVAL;
 235
 236         phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
 237
 238         for (i = 0; i < ehdr->e_phnum; ++i)
 239                 if (phdr[i].p_type == PT_NOTE)
 240                         return stack_map_parse_build_id(page_addr, build_id,
 241                                         page_addr + phdr[i].p_offset,
 242                                         phdr[i].p_filesz);
 243         return -EINVAL;
 244 }
 245
 246 /* Parse build ID of ELF file mapped to vma */
 247 static int stack_map_get_build_id(struct vm_area_struct *vma,
 248                                   unsigned char *build_id)
 249 {
 250         Elf32_Ehdr *ehdr;
 251         struct page *page;
 252         void *page_addr;
 253         int ret;
 254
 255         /* only works for page backed storage  */
 256         if (!vma->vm_file)
 257                 return -EINVAL;
 258
 259         page = find_get_page(vma->vm_file->f_mapping, 0);
 260         if (!page)
 261                 return -EFAULT; /* page not mapped */
 262
 263         ret = -EINVAL;
 264         page_addr = kmap_atomic(page);
 265         ehdr = (Elf32_Ehdr *)page_addr;
 266
 267         /* compare magic x7f "ELF" */
 268         if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
 269                 goto out;
 270
 271         /* only support executable file and shared object file */
 272         if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
 273                 goto out;
 274
 275         if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
 276                 ret = stack_map_get_build_id_32(page_addr, build_id);
 277         else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
 278                 ret = stack_map_get_build_id_64(page_addr, build_id);
 279 out:
 280         kunmap_atomic(page_addr);
 281         put_page(page);
 282         return ret;
 283 }
 284
 285 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 286                                           u64 *ips, u32 trace_nr, bool user)
 287 {
 288         int i;
 289         struct vm_area_struct *vma;
 290         bool irq_work_busy = false;
 291         struct stack_map_irq_work *work = NULL;
 292
 293         if (irqs_disabled()) {
 294                 work = this_cpu_ptr(&up_read_work);
 295                 if (work->irq_work.flags & IRQ_WORK_BUSY)
 296                         /* cannot queue more up_read, fallback */
 297                         irq_work_busy = true;
 298         }
 299
 300         /*
 301          * We cannot do up_read() when the irq is disabled, because of
 302          * risk to deadlock with rq_lock. To do build_id lookup when the
 303          * irqs are disabled, we need to run up_read() in irq_work. We use
 304          * a percpu variable to do the irq_work. If the irq_work is
 305          * already used by another lookup, we fall back to report ips.
 306          *
 307          * Same fallback is used for kernel stack (!user) on a stackmap
 308          * with build_id.
 309          */
 310         if (!user || !current || !current->mm || irq_work_busy ||
 311             down_read_trylock(&current->mm->mmap_sem) == 0) {
 312                 /* cannot access current->mm, fall back to ips */
 313                 for (i = 0; i < trace_nr; i++) {
 314                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 315                         id_offs[i].ip = ips[i];
 316                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 317                 }
 318                 return;
 319         }
 320
 321         for (i = 0; i < trace_nr; i++) {
 322                 vma = find_vma(current->mm, ips[i]);
 323                 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
 324                         /* per entry fall back to ips */
 325                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 326                         id_offs[i].ip = ips[i];
 327                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 328                         continue;
 329                 }
 330                 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
 331                         - vma->vm_start;
 332                 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
 333         }
 334
 335         if (!work) {
 336                 up_read(&current->mm->mmap_sem);
 337         } else {
 338                 work->sem = &current->mm->mmap_sem;
 339                 irq_work_queue(&work->irq_work);
 340                 /*
 341                  * The irq_work will release the mmap_sem with
 342                  * up_read_non_owner(). The rwsem_release() is called
 343                  * here to release the lock from lockdep's perspective.
 344                  */
 345                 rwsem_release(&current->mm->mmap_sem.dep_map, 1, _RET_IP_);
 346         }
 347 }
 348
 349 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 350            u64, flags)
 351 {
 352         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 353         struct perf_callchain_entry *trace;
 354         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 355         u32 max_depth = map->value_size / stack_map_data_size(map);
 356         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 357         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 358         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 359         u32 hash, id, trace_nr, trace_len;
 360         bool user = flags & BPF_F_USER_STACK;
 361         bool kernel = !user;
 362         u64 *ips;
 363         bool hash_matches;
 364
 365         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 366                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 367                 return -EINVAL;
 368
 369         trace = get_perf_callchain(regs, init_nr, kernel, user,
 370                                    sysctl_perf_event_max_stack, false, false);
 371
 372         if (unlikely(!trace))
 373                 /* couldn't fetch the stack trace */
 374                 return -EFAULT;
 375
 376         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 377          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 378          */
 379         trace_nr = trace->nr - init_nr;
 380
 381         if (trace_nr <= skip)
 382                 /* skipping more than usable stack trace */
 383                 return -EFAULT;
 384
 385         trace_nr -= skip;
 386         trace_len = trace_nr * sizeof(u64);
 387         ips = trace->ip + skip + init_nr;
 388         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 389         id = hash & (smap->n_buckets - 1);
 390         bucket = READ_ONCE(smap->buckets[id]);
 391
 392         hash_matches = bucket && bucket->hash == hash;
 393         /* fast cmp */
 394         if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
 395                 return id;
 396
 397         if (stack_map_use_build_id(map)) {
 398                 /* for build_id+offset, pop a bucket before slow cmp */
 399                 new_bucket = (struct stack_map_bucket *)
 400                         pcpu_freelist_pop(&smap->freelist);
 401                 if (unlikely(!new_bucket))
 402                         return -ENOMEM;
 403                 new_bucket->nr = trace_nr;
 404                 stack_map_get_build_id_offset(
 405                         (struct bpf_stack_build_id *)new_bucket->data,
 406                         ips, trace_nr, user);
 407                 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
 408                 if (hash_matches && bucket->nr == trace_nr &&
 409                     memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
 410                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 411                         return id;
 412                 }
 413                 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
 414                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 415                         return -EEXIST;
 416                 }
 417         } else {
 418                 if (hash_matches && bucket->nr == trace_nr &&
 419                     memcmp(bucket->data, ips, trace_len) == 0)
 420                         return id;
 421                 if (bucket && !(flags & BPF_F_REUSE_STACKID))
 422                         return -EEXIST;
 423
 424                 new_bucket = (struct stack_map_bucket *)
 425                         pcpu_freelist_pop(&smap->freelist);
 426                 if (unlikely(!new_bucket))
 427                         return -ENOMEM;
 428                 memcpy(new_bucket->data, ips, trace_len);
 429         }
 430
 431         new_bucket->hash = hash;
 432         new_bucket->nr = trace_nr;
 433
 434         old_bucket = xchg(&smap->buckets[id], new_bucket);
 435         if (old_bucket)
 436                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 437         return id;
 438 }
 439
 440 const struct bpf_func_proto bpf_get_stackid_proto = {
 441         .func           = bpf_get_stackid,
 442         .gpl_only       = true,
 443         .ret_type       = RET_INTEGER,
 444         .arg1_type      = ARG_PTR_TO_CTX,
 445         .arg2_type      = ARG_CONST_MAP_PTR,
 446         .arg3_type      = ARG_ANYTHING,
 447 };
 448
 449 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
 450            u64, flags)
 451 {
 452         u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
 453         bool user_build_id = flags & BPF_F_USER_BUILD_ID;
 454         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 455         bool user = flags & BPF_F_USER_STACK;
 456         struct perf_callchain_entry *trace;
 457         bool kernel = !user;
 458         int err = -EINVAL;
 459         u64 *ips;
 460
 461         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 462                                BPF_F_USER_BUILD_ID)))
 463                 goto clear;
 464         if (kernel && user_build_id)
 465                 goto clear;
 466
 467         elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
 468                                             : sizeof(u64);
 469         if (unlikely(size % elem_size))
 470                 goto clear;
 471
 472         num_elem = size / elem_size;
 473         if (sysctl_perf_event_max_stack < num_elem)
 474                 init_nr = 0;
 475         else
 476                 init_nr = sysctl_perf_event_max_stack - num_elem;
 477         trace = get_perf_callchain(regs, init_nr, kernel, user,
 478                                    sysctl_perf_event_max_stack, false, false);
 479         if (unlikely(!trace))
 480                 goto err_fault;
 481
 482         trace_nr = trace->nr - init_nr;
 483         if (trace_nr < skip)
 484                 goto err_fault;
 485
 486         trace_nr -= skip;
 487         trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
 488         copy_len = trace_nr * elem_size;
 489         ips = trace->ip + skip + init_nr;
 490         if (user && user_build_id)
 491                 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
 492         else
 493                 memcpy(buf, ips, copy_len);
 494
 495         if (size > copy_len)
 496                 memset(buf + copy_len, 0, size - copy_len);
 497         return copy_len;
 498
 499 err_fault:
 500         err = -EFAULT;
 501 clear:
 502         memset(buf, 0, size);
 503         return err;
 504 }
 505
 506 const struct bpf_func_proto bpf_get_stack_proto = {
 507         .func           = bpf_get_stack,
 508         .gpl_only       = true,
 509         .ret_type       = RET_INTEGER,
 510         .arg1_type      = ARG_PTR_TO_CTX,
 511         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 512         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 513         .arg4_type      = ARG_ANYTHING,
 514 };
 515
 516 /* Called from eBPF program */
 517 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 518 {
 519         return ERR_PTR(-EOPNOTSUPP);
 520 }
 521
 522 /* Called from syscall */
 523 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 524 {
 525         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 526         struct stack_map_bucket *bucket, *old_bucket;
 527         u32 id = *(u32 *)key, trace_len;
 528
 529         if (unlikely(id >= smap->n_buckets))
 530                 return -ENOENT;
 531
 532         bucket = xchg(&smap->buckets[id], NULL);
 533         if (!bucket)
 534                 return -ENOENT;
 535
 536         trace_len = bucket->nr * stack_map_data_size(map);
 537         memcpy(value, bucket->data, trace_len);
 538         memset(value + trace_len, 0, map->value_size - trace_len);
 539
 540         old_bucket = xchg(&smap->buckets[id], bucket);
 541         if (old_bucket)
 542                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 543         return 0;
 544 }
 545
 546 static int stack_map_get_next_key(struct bpf_map *map, void *key,
 547                                   void *next_key)
 548 {
 549         struct bpf_stack_map *smap = container_of(map,
 550                                                   struct bpf_stack_map, map);
 551         u32 id;
 552
 553         WARN_ON_ONCE(!rcu_read_lock_held());
 554
 555         if (!key) {
 556                 id = 0;
 557         } else {
 558                 id = *(u32 *)key;
 559                 if (id >= smap->n_buckets || !smap->buckets[id])
 560                         id = 0;
 561                 else
 562                         id++;
 563         }
 564
 565         while (id < smap->n_buckets && !smap->buckets[id])
 566                 id++;
 567
 568         if (id >= smap->n_buckets)
 569                 return -ENOENT;
 570
 571         *(u32 *)next_key = id;
 572         return 0;
 573 }
 574
 575 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 576                                  u64 map_flags)
 577 {
 578         return -EINVAL;
 579 }
 580
 581 /* Called from syscall or from eBPF program */
 582 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 583 {
 584         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 585         struct stack_map_bucket *old_bucket;
 586         u32 id = *(u32 *)key;
 587
 588         if (unlikely(id >= smap->n_buckets))
 589                 return -E2BIG;
 590
 591         old_bucket = xchg(&smap->buckets[id], NULL);
 592         if (old_bucket) {
 593                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 594                 return 0;
 595         } else {
 596                 return -ENOENT;
 597         }
 598 }
 599
 600 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 601 static void stack_map_free(struct bpf_map *map)
 602 {
 603         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 604
 605         /* wait for bpf programs to complete before freeing stack map */
 606         synchronize_rcu();
 607
 608         bpf_map_area_free(smap->elems);
 609         pcpu_freelist_destroy(&smap->freelist);
 610         bpf_map_area_free(smap);
 611         put_callchain_buffers();
 612 }
 613
 614 const struct bpf_map_ops stack_trace_map_ops = {
 615         .map_alloc = stack_map_alloc,
 616         .map_free = stack_map_free,
 617         .map_get_next_key = stack_map_get_next_key,
 618         .map_lookup_elem = stack_map_lookup_elem,
 619         .map_update_elem = stack_map_update_elem,
 620         .map_delete_elem = stack_map_delete_elem,
 621         .map_check_btf = map_check_no_btf,
 622 };
 623
 624 static int __init stack_map_init(void)
 625 {
 626         int cpu;
 627         struct stack_map_irq_work *work;
 628
 629         for_each_possible_cpu(cpu) {
 630                 work = per_cpu_ptr(&up_read_work, cpu);
 631                 init_irq_work(&work->irq_work, do_up_read);
 632         }
 633         return 0;
 634 }
 635 subsys_initcall(stack_map_init);