kernel/bpf/stackmap.c

   1 /* Copyright (c) 2016 Facebook
   2  *
   3  * This program is free software; you can redistribute it and/or
   4  * modify it under the terms of version 2 of the GNU General Public
   5  * License as published by the Free Software Foundation.
   6  */
   7 #include <linux/bpf.h>
   8 #include <linux/jhash.h>
   9 #include <linux/filter.h>
  10 #include <linux/stacktrace.h>
  11 #include <linux/perf_event.h>
  12 #include <linux/elf.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/irq_work.h>
  15 #include "percpu_freelist.h"
  16
  17 #define STACK_CREATE_FLAG_MASK                                  \
  18         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY |        \
  19          BPF_F_STACK_BUILD_ID)
  20
  21 struct stack_map_bucket {
  22         struct pcpu_freelist_node fnode;
  23         u32 hash;
  24         u32 nr;
  25         u64 data[];
  26 };
  27
  28 struct bpf_stack_map {
  29         struct bpf_map map;
  30         void *elems;
  31         struct pcpu_freelist freelist;
  32         u32 n_buckets;
  33         struct stack_map_bucket *buckets[];
  34 };
  35
  36 /* irq_work to run up_read() for build_id lookup in nmi context */
  37 struct stack_map_irq_work {
  38         struct irq_work irq_work;
  39         struct rw_semaphore *sem;
  40 };
  41
  42 static void do_up_read(struct irq_work *entry)
  43 {
  44         struct stack_map_irq_work *work;
  45
  46         work = container_of(entry, struct stack_map_irq_work, irq_work);
  47         up_read_non_owner(work->sem);
  48         work->sem = NULL;
  49 }
  50
  51 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
  52
  53 static inline bool stack_map_use_build_id(struct bpf_map *map)
  54 {
  55         return (map->map_flags & BPF_F_STACK_BUILD_ID);
  56 }
  57
  58 static inline int stack_map_data_size(struct bpf_map *map)
  59 {
  60         return stack_map_use_build_id(map) ?
  61                 sizeof(struct bpf_stack_build_id) : sizeof(u64);
  62 }
  63
  64 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
  65 {
  66         u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
  67         int err;
  68
  69         smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
  70                                          smap->map.numa_node);
  71         if (!smap->elems)
  72                 return -ENOMEM;
  73
  74         err = pcpu_freelist_init(&smap->freelist);
  75         if (err)
  76                 goto free_elems;
  77
  78         pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
  79                                smap->map.max_entries);
  80         return 0;
  81
  82 free_elems:
  83         bpf_map_area_free(smap->elems);
  84         return err;
  85 }
  86
  87 /* Called from syscall */
  88 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
  89 {
  90         u32 value_size = attr->value_size;
  91         struct bpf_stack_map *smap;
  92         struct bpf_map_memory mem;
  93         u64 cost, n_buckets;
  94         int err;
  95
  96         if (!capable(CAP_SYS_ADMIN))
  97                 return ERR_PTR(-EPERM);
  98
  99         if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
 100                 return ERR_PTR(-EINVAL);
 101
 102         /* check sanity of attributes */
 103         if (attr->max_entries == 0 || attr->key_size != 4 ||
 104             value_size < 8 || value_size % 8)
 105                 return ERR_PTR(-EINVAL);
 106
 107         BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
 108         if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
 109                 if (value_size % sizeof(struct bpf_stack_build_id) ||
 110                     value_size / sizeof(struct bpf_stack_build_id)
 111                     > sysctl_perf_event_max_stack)
 112                         return ERR_PTR(-EINVAL);
 113         } else if (value_size / 8 > sysctl_perf_event_max_stack)
 114                 return ERR_PTR(-EINVAL);
 115
 116         /* hash table size must be power of 2 */
 117         n_buckets = roundup_pow_of_two(attr->max_entries);
 118
 119         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 120         if (cost >= U32_MAX - PAGE_SIZE)
 121                 return ERR_PTR(-E2BIG);
 122         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
 123         if (cost >= U32_MAX - PAGE_SIZE)
 124                 return ERR_PTR(-E2BIG);
 125
 126         err = bpf_map_charge_init(&mem,
 127                                   round_up(cost, PAGE_SIZE) >> PAGE_SHIFT);
 128         if (err)
 129                 return ERR_PTR(err);
 130
 131         smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
 132         if (!smap) {
 133                 bpf_map_charge_finish(&mem);
 134                 return ERR_PTR(-ENOMEM);
 135         }
 136
 137         bpf_map_init_from_attr(&smap->map, attr);
 138         smap->map.value_size = value_size;
 139         smap->n_buckets = n_buckets;
 140
 141         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 142         if (err)
 143                 goto free_charge;
 144
 145         err = prealloc_elems_and_freelist(smap);
 146         if (err)
 147                 goto put_buffers;
 148
 149         bpf_map_charge_move(&smap->map.memory, &mem);
 150
 151         return &smap->map;
 152
 153 put_buffers:
 154         put_callchain_buffers();
 155 free_charge:
 156         bpf_map_charge_finish(&mem);
 157         bpf_map_area_free(smap);
 158         return ERR_PTR(err);
 159 }
 160
 161 #define BPF_BUILD_ID 3
 162 /*
 163  * Parse build id from the note segment. This logic can be shared between
 164  * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
 165  * identical.
 166  */
 167 static inline int stack_map_parse_build_id(void *page_addr,
 168                                            unsigned char *build_id,
 169                                            void *note_start,
 170                                            Elf32_Word note_size)
 171 {
 172         Elf32_Word note_offs = 0, new_offs;
 173
 174         /* check for overflow */
 175         if (note_start < page_addr || note_start + note_size < note_start)
 176                 return -EINVAL;
 177
 178         /* only supports note that fits in the first page */
 179         if (note_start + note_size > page_addr + PAGE_SIZE)
 180                 return -EINVAL;
 181
 182         while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
 183                 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
 184
 185                 if (nhdr->n_type == BPF_BUILD_ID &&
 186                     nhdr->n_namesz == sizeof("GNU") &&
 187                     nhdr->n_descsz > 0 &&
 188                     nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
 189                         memcpy(build_id,
 190                                note_start + note_offs +
 191                                ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
 192                                nhdr->n_descsz);
 193                         memset(build_id + nhdr->n_descsz, 0,
 194                                BPF_BUILD_ID_SIZE - nhdr->n_descsz);
 195                         return 0;
 196                 }
 197                 new_offs = note_offs + sizeof(Elf32_Nhdr) +
 198                         ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
 199                 if (new_offs <= note_offs)  /* overflow */
 200                         break;
 201                 note_offs = new_offs;
 202         }
 203         return -EINVAL;
 204 }
 205
 206 /* Parse build ID from 32-bit ELF */
 207 static int stack_map_get_build_id_32(void *page_addr,
 208                                      unsigned char *build_id)
 209 {
 210         Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
 211         Elf32_Phdr *phdr;
 212         int i;
 213
 214         /* only supports phdr that fits in one page */
 215         if (ehdr->e_phnum >
 216             (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
 217                 return -EINVAL;
 218
 219         phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
 220
 221         for (i = 0; i < ehdr->e_phnum; ++i)
 222                 if (phdr[i].p_type == PT_NOTE)
 223                         return stack_map_parse_build_id(page_addr, build_id,
 224                                         page_addr + phdr[i].p_offset,
 225                                         phdr[i].p_filesz);
 226         return -EINVAL;
 227 }
 228
 229 /* Parse build ID from 64-bit ELF */
 230 static int stack_map_get_build_id_64(void *page_addr,
 231                                      unsigned char *build_id)
 232 {
 233         Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
 234         Elf64_Phdr *phdr;
 235         int i;
 236
 237         /* only supports phdr that fits in one page */
 238         if (ehdr->e_phnum >
 239             (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
 240                 return -EINVAL;
 241
 242         phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
 243
 244         for (i = 0; i < ehdr->e_phnum; ++i)
 245                 if (phdr[i].p_type == PT_NOTE)
 246                         return stack_map_parse_build_id(page_addr, build_id,
 247                                         page_addr + phdr[i].p_offset,
 248                                         phdr[i].p_filesz);
 249         return -EINVAL;
 250 }
 251
 252 /* Parse build ID of ELF file mapped to vma */
 253 static int stack_map_get_build_id(struct vm_area_struct *vma,
 254                                   unsigned char *build_id)
 255 {
 256         Elf32_Ehdr *ehdr;
 257         struct page *page;
 258         void *page_addr;
 259         int ret;
 260
 261         /* only works for page backed storage  */
 262         if (!vma->vm_file)
 263                 return -EINVAL;
 264
 265         page = find_get_page(vma->vm_file->f_mapping, 0);
 266         if (!page)
 267                 return -EFAULT; /* page not mapped */
 268
 269         ret = -EINVAL;
 270         page_addr = kmap_atomic(page);
 271         ehdr = (Elf32_Ehdr *)page_addr;
 272
 273         /* compare magic x7f "ELF" */
 274         if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
 275                 goto out;
 276
 277         /* only support executable file and shared object file */
 278         if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
 279                 goto out;
 280
 281         if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
 282                 ret = stack_map_get_build_id_32(page_addr, build_id);
 283         else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
 284                 ret = stack_map_get_build_id_64(page_addr, build_id);
 285 out:
 286         kunmap_atomic(page_addr);
 287         put_page(page);
 288         return ret;
 289 }
 290
 291 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 292                                           u64 *ips, u32 trace_nr, bool user)
 293 {
 294         int i;
 295         struct vm_area_struct *vma;
 296         bool irq_work_busy = false;
 297         struct stack_map_irq_work *work = NULL;
 298
 299         if (in_nmi()) {
 300                 work = this_cpu_ptr(&up_read_work);
 301                 if (work->irq_work.flags & IRQ_WORK_BUSY)
 302                         /* cannot queue more up_read, fallback */
 303                         irq_work_busy = true;
 304         }
 305
 306         /*
 307          * We cannot do up_read() in nmi context. To do build_id lookup
 308          * in nmi context, we need to run up_read() in irq_work. We use
 309          * a percpu variable to do the irq_work. If the irq_work is
 310          * already used by another lookup, we fall back to report ips.
 311          *
 312          * Same fallback is used for kernel stack (!user) on a stackmap
 313          * with build_id.
 314          */
 315         if (!user || !current || !current->mm || irq_work_busy ||
 316             down_read_trylock(&current->mm->mmap_sem) == 0) {
 317                 /* cannot access current->mm, fall back to ips */
 318                 for (i = 0; i < trace_nr; i++) {
 319                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 320                         id_offs[i].ip = ips[i];
 321                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 322                 }
 323                 return;
 324         }
 325
 326         for (i = 0; i < trace_nr; i++) {
 327                 vma = find_vma(current->mm, ips[i]);
 328                 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
 329                         /* per entry fall back to ips */
 330                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 331                         id_offs[i].ip = ips[i];
 332                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 333                         continue;
 334                 }
 335                 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
 336                         - vma->vm_start;
 337                 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
 338         }
 339
 340         if (!work) {
 341                 up_read(&current->mm->mmap_sem);
 342         } else {
 343                 work->sem = &current->mm->mmap_sem;
 344                 irq_work_queue(&work->irq_work);
 345                 /*
 346                  * The irq_work will release the mmap_sem with
 347                  * up_read_non_owner(). The rwsem_release() is called
 348                  * here to release the lock from lockdep's perspective.
 349                  */
 350                 rwsem_release(&current->mm->mmap_sem.dep_map, 1, _RET_IP_);
 351         }
 352 }
 353
 354 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 355            u64, flags)
 356 {
 357         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 358         struct perf_callchain_entry *trace;
 359         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 360         u32 max_depth = map->value_size / stack_map_data_size(map);
 361         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 362         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 363         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 364         u32 hash, id, trace_nr, trace_len;
 365         bool user = flags & BPF_F_USER_STACK;
 366         bool kernel = !user;
 367         u64 *ips;
 368         bool hash_matches;
 369
 370         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 371                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 372                 return -EINVAL;
 373
 374         trace = get_perf_callchain(regs, init_nr, kernel, user,
 375                                    sysctl_perf_event_max_stack, false, false);
 376
 377         if (unlikely(!trace))
 378                 /* couldn't fetch the stack trace */
 379                 return -EFAULT;
 380
 381         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 382          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 383          */
 384         trace_nr = trace->nr - init_nr;
 385
 386         if (trace_nr <= skip)
 387                 /* skipping more than usable stack trace */
 388                 return -EFAULT;
 389
 390         trace_nr -= skip;
 391         trace_len = trace_nr * sizeof(u64);
 392         ips = trace->ip + skip + init_nr;
 393         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 394         id = hash & (smap->n_buckets - 1);
 395         bucket = READ_ONCE(smap->buckets[id]);
 396
 397         hash_matches = bucket && bucket->hash == hash;
 398         /* fast cmp */
 399         if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
 400                 return id;
 401
 402         if (stack_map_use_build_id(map)) {
 403                 /* for build_id+offset, pop a bucket before slow cmp */
 404                 new_bucket = (struct stack_map_bucket *)
 405                         pcpu_freelist_pop(&smap->freelist);
 406                 if (unlikely(!new_bucket))
 407                         return -ENOMEM;
 408                 new_bucket->nr = trace_nr;
 409                 stack_map_get_build_id_offset(
 410                         (struct bpf_stack_build_id *)new_bucket->data,
 411                         ips, trace_nr, user);
 412                 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
 413                 if (hash_matches && bucket->nr == trace_nr &&
 414                     memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
 415                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 416                         return id;
 417                 }
 418                 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
 419                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 420                         return -EEXIST;
 421                 }
 422         } else {
 423                 if (hash_matches && bucket->nr == trace_nr &&
 424                     memcmp(bucket->data, ips, trace_len) == 0)
 425                         return id;
 426                 if (bucket && !(flags & BPF_F_REUSE_STACKID))
 427                         return -EEXIST;
 428
 429                 new_bucket = (struct stack_map_bucket *)
 430                         pcpu_freelist_pop(&smap->freelist);
 431                 if (unlikely(!new_bucket))
 432                         return -ENOMEM;
 433                 memcpy(new_bucket->data, ips, trace_len);
 434         }
 435
 436         new_bucket->hash = hash;
 437         new_bucket->nr = trace_nr;
 438
 439         old_bucket = xchg(&smap->buckets[id], new_bucket);
 440         if (old_bucket)
 441                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 442         return id;
 443 }
 444
 445 const struct bpf_func_proto bpf_get_stackid_proto = {
 446         .func           = bpf_get_stackid,
 447         .gpl_only       = true,
 448         .ret_type       = RET_INTEGER,
 449         .arg1_type      = ARG_PTR_TO_CTX,
 450         .arg2_type      = ARG_CONST_MAP_PTR,
 451         .arg3_type      = ARG_ANYTHING,
 452 };
 453
 454 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
 455            u64, flags)
 456 {
 457         u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
 458         bool user_build_id = flags & BPF_F_USER_BUILD_ID;
 459         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 460         bool user = flags & BPF_F_USER_STACK;
 461         struct perf_callchain_entry *trace;
 462         bool kernel = !user;
 463         int err = -EINVAL;
 464         u64 *ips;
 465
 466         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 467                                BPF_F_USER_BUILD_ID)))
 468                 goto clear;
 469         if (kernel && user_build_id)
 470                 goto clear;
 471
 472         elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
 473                                             : sizeof(u64);
 474         if (unlikely(size % elem_size))
 475                 goto clear;
 476
 477         num_elem = size / elem_size;
 478         if (sysctl_perf_event_max_stack < num_elem)
 479                 init_nr = 0;
 480         else
 481                 init_nr = sysctl_perf_event_max_stack - num_elem;
 482         trace = get_perf_callchain(regs, init_nr, kernel, user,
 483                                    sysctl_perf_event_max_stack, false, false);
 484         if (unlikely(!trace))
 485                 goto err_fault;
 486
 487         trace_nr = trace->nr - init_nr;
 488         if (trace_nr < skip)
 489                 goto err_fault;
 490
 491         trace_nr -= skip;
 492         trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
 493         copy_len = trace_nr * elem_size;
 494         ips = trace->ip + skip + init_nr;
 495         if (user && user_build_id)
 496                 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
 497         else
 498                 memcpy(buf, ips, copy_len);
 499
 500         if (size > copy_len)
 501                 memset(buf + copy_len, 0, size - copy_len);
 502         return copy_len;
 503
 504 err_fault:
 505         err = -EFAULT;
 506 clear:
 507         memset(buf, 0, size);
 508         return err;
 509 }
 510
 511 const struct bpf_func_proto bpf_get_stack_proto = {
 512         .func           = bpf_get_stack,
 513         .gpl_only       = true,
 514         .ret_type       = RET_INTEGER,
 515         .arg1_type      = ARG_PTR_TO_CTX,
 516         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 517         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 518         .arg4_type      = ARG_ANYTHING,
 519 };
 520
 521 /* Called from eBPF program */
 522 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 523 {
 524         return ERR_PTR(-EOPNOTSUPP);
 525 }
 526
 527 /* Called from syscall */
 528 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 529 {
 530         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 531         struct stack_map_bucket *bucket, *old_bucket;
 532         u32 id = *(u32 *)key, trace_len;
 533
 534         if (unlikely(id >= smap->n_buckets))
 535                 return -ENOENT;
 536
 537         bucket = xchg(&smap->buckets[id], NULL);
 538         if (!bucket)
 539                 return -ENOENT;
 540
 541         trace_len = bucket->nr * stack_map_data_size(map);
 542         memcpy(value, bucket->data, trace_len);
 543         memset(value + trace_len, 0, map->value_size - trace_len);
 544
 545         old_bucket = xchg(&smap->buckets[id], bucket);
 546         if (old_bucket)
 547                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 548         return 0;
 549 }
 550
 551 static int stack_map_get_next_key(struct bpf_map *map, void *key,
 552                                   void *next_key)
 553 {
 554         struct bpf_stack_map *smap = container_of(map,
 555                                                   struct bpf_stack_map, map);
 556         u32 id;
 557
 558         WARN_ON_ONCE(!rcu_read_lock_held());
 559
 560         if (!key) {
 561                 id = 0;
 562         } else {
 563                 id = *(u32 *)key;
 564                 if (id >= smap->n_buckets || !smap->buckets[id])
 565                         id = 0;
 566                 else
 567                         id++;
 568         }
 569
 570         while (id < smap->n_buckets && !smap->buckets[id])
 571                 id++;
 572
 573         if (id >= smap->n_buckets)
 574                 return -ENOENT;
 575
 576         *(u32 *)next_key = id;
 577         return 0;
 578 }
 579
 580 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 581                                  u64 map_flags)
 582 {
 583         return -EINVAL;
 584 }
 585
 586 /* Called from syscall or from eBPF program */
 587 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 588 {
 589         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 590         struct stack_map_bucket *old_bucket;
 591         u32 id = *(u32 *)key;
 592
 593         if (unlikely(id >= smap->n_buckets))
 594                 return -E2BIG;
 595
 596         old_bucket = xchg(&smap->buckets[id], NULL);
 597         if (old_bucket) {
 598                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 599                 return 0;
 600         } else {
 601                 return -ENOENT;
 602         }
 603 }
 604
 605 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 606 static void stack_map_free(struct bpf_map *map)
 607 {
 608         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 609
 610         /* wait for bpf programs to complete before freeing stack map */
 611         synchronize_rcu();
 612
 613         bpf_map_area_free(smap->elems);
 614         pcpu_freelist_destroy(&smap->freelist);
 615         bpf_map_area_free(smap);
 616         put_callchain_buffers();
 617 }
 618
 619 const struct bpf_map_ops stack_trace_map_ops = {
 620         .map_alloc = stack_map_alloc,
 621         .map_free = stack_map_free,
 622         .map_get_next_key = stack_map_get_next_key,
 623         .map_lookup_elem = stack_map_lookup_elem,
 624         .map_update_elem = stack_map_update_elem,
 625         .map_delete_elem = stack_map_delete_elem,
 626         .map_check_btf = map_check_no_btf,
 627 };
 628
 629 static int __init stack_map_init(void)
 630 {
 631         int cpu;
 632         struct stack_map_irq_work *work;
 633
 634         for_each_possible_cpu(cpu) {
 635                 work = per_cpu_ptr(&up_read_work, cpu);
 636                 init_irq_work(&work->irq_work, do_up_read);
 637         }
 638         return 0;
 639 }
 640 subsys_initcall(stack_map_init);