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 mm_struct *mm;
  37 };
  38
  39 static void do_up_read(struct irq_work *entry)
  40 {
  41         struct stack_map_irq_work *work;
  42
  43         if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
  44                 return;
  45
  46         work = container_of(entry, struct stack_map_irq_work, irq_work);
  47         mmap_read_unlock_non_owner(work->mm);
  48 }
  49
  50 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
  51
  52 static inline bool stack_map_use_build_id(struct bpf_map *map)
  53 {
  54         return (map->map_flags & BPF_F_STACK_BUILD_ID);
  55 }
  56
  57 static inline int stack_map_data_size(struct bpf_map *map)
  58 {
  59         return stack_map_use_build_id(map) ?
  60                 sizeof(struct bpf_stack_build_id) : sizeof(u64);
  61 }
  62
  63 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
  64 {
  65         u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
  66         int err;
  67
  68         smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
  69                                          smap->map.numa_node);
  70         if (!smap->elems)
  71                 return -ENOMEM;
  72
  73         err = pcpu_freelist_init(&smap->freelist);
  74         if (err)
  75                 goto free_elems;
  76
  77         pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
  78                                smap->map.max_entries);
  79         return 0;
  80
  81 free_elems:
  82         bpf_map_area_free(smap->elems);
  83         return err;
  84 }
  85
  86 /* Called from syscall */
  87 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
  88 {
  89         u32 value_size = attr->value_size;
  90         struct bpf_stack_map *smap;
  91         struct bpf_map_memory mem;
  92         u64 cost, n_buckets;
  93         int err;
  94
  95         if (!bpf_capable())
  96                 return ERR_PTR(-EPERM);
  97
  98         if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
  99                 return ERR_PTR(-EINVAL);
 100
 101         /* check sanity of attributes */
 102         if (attr->max_entries == 0 || attr->key_size != 4 ||
 103             value_size < 8 || value_size % 8)
 104                 return ERR_PTR(-EINVAL);
 105
 106         BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
 107         if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
 108                 if (value_size % sizeof(struct bpf_stack_build_id) ||
 109                     value_size / sizeof(struct bpf_stack_build_id)
 110                     > sysctl_perf_event_max_stack)
 111                         return ERR_PTR(-EINVAL);
 112         } else if (value_size / 8 > sysctl_perf_event_max_stack)
 113                 return ERR_PTR(-EINVAL);
 114
 115         /* hash table size must be power of 2 */
 116         n_buckets = roundup_pow_of_two(attr->max_entries);
 117
 118         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 119         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
 120         err = bpf_map_charge_init(&mem, cost);
 121         if (err)
 122                 return ERR_PTR(err);
 123
 124         smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
 125         if (!smap) {
 126                 bpf_map_charge_finish(&mem);
 127                 return ERR_PTR(-ENOMEM);
 128         }
 129
 130         bpf_map_init_from_attr(&smap->map, attr);
 131         smap->map.value_size = value_size;
 132         smap->n_buckets = n_buckets;
 133
 134         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 135         if (err)
 136                 goto free_charge;
 137
 138         err = prealloc_elems_and_freelist(smap);
 139         if (err)
 140                 goto put_buffers;
 141
 142         bpf_map_charge_move(&smap->map.memory, &mem);
 143
 144         return &smap->map;
 145
 146 put_buffers:
 147         put_callchain_buffers();
 148 free_charge:
 149         bpf_map_charge_finish(&mem);
 150         bpf_map_area_free(smap);
 151         return ERR_PTR(err);
 152 }
 153
 154 #define BPF_BUILD_ID 3
 155 /*
 156  * Parse build id from the note segment. This logic can be shared between
 157  * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
 158  * identical.
 159  */
 160 static inline int stack_map_parse_build_id(void *page_addr,
 161                                            unsigned char *build_id,
 162                                            void *note_start,
 163                                            Elf32_Word note_size)
 164 {
 165         Elf32_Word note_offs = 0, new_offs;
 166
 167         /* check for overflow */
 168         if (note_start < page_addr || note_start + note_size < note_start)
 169                 return -EINVAL;
 170
 171         /* only supports note that fits in the first page */
 172         if (note_start + note_size > page_addr + PAGE_SIZE)
 173                 return -EINVAL;
 174
 175         while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
 176                 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
 177
 178                 if (nhdr->n_type == BPF_BUILD_ID &&
 179                     nhdr->n_namesz == sizeof("GNU") &&
 180                     nhdr->n_descsz > 0 &&
 181                     nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
 182                         memcpy(build_id,
 183                                note_start + note_offs +
 184                                ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
 185                                nhdr->n_descsz);
 186                         memset(build_id + nhdr->n_descsz, 0,
 187                                BPF_BUILD_ID_SIZE - nhdr->n_descsz);
 188                         return 0;
 189                 }
 190                 new_offs = note_offs + sizeof(Elf32_Nhdr) +
 191                         ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
 192                 if (new_offs <= note_offs)  /* overflow */
 193                         break;
 194                 note_offs = new_offs;
 195         }
 196         return -EINVAL;
 197 }
 198
 199 /* Parse build ID from 32-bit ELF */
 200 static int stack_map_get_build_id_32(void *page_addr,
 201                                      unsigned char *build_id)
 202 {
 203         Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
 204         Elf32_Phdr *phdr;
 205         int i;
 206
 207         /* only supports phdr that fits in one page */
 208         if (ehdr->e_phnum >
 209             (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
 210                 return -EINVAL;
 211
 212         phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
 213
 214         for (i = 0; i < ehdr->e_phnum; ++i)
 215                 if (phdr[i].p_type == PT_NOTE)
 216                         return stack_map_parse_build_id(page_addr, build_id,
 217                                         page_addr + phdr[i].p_offset,
 218                                         phdr[i].p_filesz);
 219         return -EINVAL;
 220 }
 221
 222 /* Parse build ID from 64-bit ELF */
 223 static int stack_map_get_build_id_64(void *page_addr,
 224                                      unsigned char *build_id)
 225 {
 226         Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
 227         Elf64_Phdr *phdr;
 228         int i;
 229
 230         /* only supports phdr that fits in one page */
 231         if (ehdr->e_phnum >
 232             (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
 233                 return -EINVAL;
 234
 235         phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
 236
 237         for (i = 0; i < ehdr->e_phnum; ++i)
 238                 if (phdr[i].p_type == PT_NOTE)
 239                         return stack_map_parse_build_id(page_addr, build_id,
 240                                         page_addr + phdr[i].p_offset,
 241                                         phdr[i].p_filesz);
 242         return -EINVAL;
 243 }
 244
 245 /* Parse build ID of ELF file mapped to vma */
 246 static int stack_map_get_build_id(struct vm_area_struct *vma,
 247                                   unsigned char *build_id)
 248 {
 249         Elf32_Ehdr *ehdr;
 250         struct page *page;
 251         void *page_addr;
 252         int ret;
 253
 254         /* only works for page backed storage  */
 255         if (!vma->vm_file)
 256                 return -EINVAL;
 257
 258         page = find_get_page(vma->vm_file->f_mapping, 0);
 259         if (!page)
 260                 return -EFAULT; /* page not mapped */
 261
 262         ret = -EINVAL;
 263         page_addr = kmap_atomic(page);
 264         ehdr = (Elf32_Ehdr *)page_addr;
 265
 266         /* compare magic x7f "ELF" */
 267         if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
 268                 goto out;
 269
 270         /* only support executable file and shared object file */
 271         if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
 272                 goto out;
 273
 274         if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
 275                 ret = stack_map_get_build_id_32(page_addr, build_id);
 276         else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
 277                 ret = stack_map_get_build_id_64(page_addr, build_id);
 278 out:
 279         kunmap_atomic(page_addr);
 280         put_page(page);
 281         return ret;
 282 }
 283
 284 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 285                                           u64 *ips, u32 trace_nr, bool user)
 286 {
 287         int i;
 288         struct vm_area_struct *vma;
 289         bool irq_work_busy = false;
 290         struct stack_map_irq_work *work = NULL;
 291
 292         if (irqs_disabled()) {
 293                 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
 294                         work = this_cpu_ptr(&up_read_work);
 295                         if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY) {
 296                                 /* cannot queue more up_read, fallback */
 297                                 irq_work_busy = true;
 298                         }
 299                 } else {
 300                         /*
 301                          * PREEMPT_RT does not allow to trylock mmap sem in
 302                          * interrupt disabled context. Force the fallback code.
 303                          */
 304                         irq_work_busy = true;
 305                 }
 306         }
 307
 308         /*
 309          * We cannot do up_read() when the irq is disabled, because of
 310          * risk to deadlock with rq_lock. To do build_id lookup when the
 311          * irqs are disabled, we need to run up_read() in irq_work. We use
 312          * a percpu variable to do the irq_work. If the irq_work is
 313          * already used by another lookup, we fall back to report ips.
 314          *
 315          * Same fallback is used for kernel stack (!user) on a stackmap
 316          * with build_id.
 317          */
 318         if (!user || !current || !current->mm || irq_work_busy ||
 319             !mmap_read_trylock_non_owner(current->mm)) {
 320                 /* cannot access current->mm, fall back to ips */
 321                 for (i = 0; i < trace_nr; i++) {
 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                 }
 326                 return;
 327         }
 328
 329         for (i = 0; i < trace_nr; i++) {
 330                 vma = find_vma(current->mm, ips[i]);
 331                 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
 332                         /* per entry fall back to ips */
 333                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 334                         id_offs[i].ip = ips[i];
 335                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 336                         continue;
 337                 }
 338                 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
 339                         - vma->vm_start;
 340                 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
 341         }
 342
 343         if (!work) {
 344                 mmap_read_unlock_non_owner(current->mm);
 345         } else {
 346                 work->mm = current->mm;
 347                 irq_work_queue(&work->irq_work);
 348         }
 349 }
 350
 351 static struct perf_callchain_entry *
 352 get_callchain_entry_for_task(struct task_struct *task, u32 init_nr)
 353 {
 354 #ifdef CONFIG_STACKTRACE
 355         struct perf_callchain_entry *entry;
 356         int rctx;
 357
 358         entry = get_callchain_entry(&rctx);
 359
 360         if (!entry)
 361                 return NULL;
 362
 363         entry->nr = init_nr +
 364                 stack_trace_save_tsk(task, (unsigned long *)(entry->ip + init_nr),
 365                                      sysctl_perf_event_max_stack - init_nr, 0);
 366
 367         /* stack_trace_save_tsk() works on unsigned long array, while
 368          * perf_callchain_entry uses u64 array. For 32-bit systems, it is
 369          * necessary to fix this mismatch.
 370          */
 371         if (__BITS_PER_LONG != 64) {
 372                 unsigned long *from = (unsigned long *) entry->ip;
 373                 u64 *to = entry->ip;
 374                 int i;
 375
 376                 /* copy data from the end to avoid using extra buffer */
 377                 for (i = entry->nr - 1; i >= (int)init_nr; i--)
 378                         to[i] = (u64)(from[i]);
 379         }
 380
 381         put_callchain_entry(rctx);
 382
 383         return entry;
 384 #else /* CONFIG_STACKTRACE */
 385         return NULL;
 386 #endif
 387 }
 388
 389 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 390            u64, flags)
 391 {
 392         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 393         struct perf_callchain_entry *trace;
 394         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 395         u32 max_depth = map->value_size / stack_map_data_size(map);
 396         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 397         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 398         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 399         u32 hash, id, trace_nr, trace_len;
 400         bool user = flags & BPF_F_USER_STACK;
 401         bool kernel = !user;
 402         u64 *ips;
 403         bool hash_matches;
 404
 405         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 406                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 407                 return -EINVAL;
 408
 409         trace = get_perf_callchain(regs, init_nr, kernel, user,
 410                                    sysctl_perf_event_max_stack, false, false);
 411
 412         if (unlikely(!trace))
 413                 /* couldn't fetch the stack trace */
 414                 return -EFAULT;
 415
 416         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 417          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 418          */
 419         trace_nr = trace->nr - init_nr;
 420
 421         if (trace_nr <= skip)
 422                 /* skipping more than usable stack trace */
 423                 return -EFAULT;
 424
 425         trace_nr -= skip;
 426         trace_len = trace_nr * sizeof(u64);
 427         ips = trace->ip + skip + init_nr;
 428         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 429         id = hash & (smap->n_buckets - 1);
 430         bucket = READ_ONCE(smap->buckets[id]);
 431
 432         hash_matches = bucket && bucket->hash == hash;
 433         /* fast cmp */
 434         if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
 435                 return id;
 436
 437         if (stack_map_use_build_id(map)) {
 438                 /* for build_id+offset, pop a bucket before slow cmp */
 439                 new_bucket = (struct stack_map_bucket *)
 440                         pcpu_freelist_pop(&smap->freelist);
 441                 if (unlikely(!new_bucket))
 442                         return -ENOMEM;
 443                 new_bucket->nr = trace_nr;
 444                 stack_map_get_build_id_offset(
 445                         (struct bpf_stack_build_id *)new_bucket->data,
 446                         ips, trace_nr, user);
 447                 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
 448                 if (hash_matches && bucket->nr == trace_nr &&
 449                     memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
 450                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 451                         return id;
 452                 }
 453                 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
 454                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 455                         return -EEXIST;
 456                 }
 457         } else {
 458                 if (hash_matches && bucket->nr == trace_nr &&
 459                     memcmp(bucket->data, ips, trace_len) == 0)
 460                         return id;
 461                 if (bucket && !(flags & BPF_F_REUSE_STACKID))
 462                         return -EEXIST;
 463
 464                 new_bucket = (struct stack_map_bucket *)
 465                         pcpu_freelist_pop(&smap->freelist);
 466                 if (unlikely(!new_bucket))
 467                         return -ENOMEM;
 468                 memcpy(new_bucket->data, ips, trace_len);
 469         }
 470
 471         new_bucket->hash = hash;
 472         new_bucket->nr = trace_nr;
 473
 474         old_bucket = xchg(&smap->buckets[id], new_bucket);
 475         if (old_bucket)
 476                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 477         return id;
 478 }
 479
 480 const struct bpf_func_proto bpf_get_stackid_proto = {
 481         .func           = bpf_get_stackid,
 482         .gpl_only       = true,
 483         .ret_type       = RET_INTEGER,
 484         .arg1_type      = ARG_PTR_TO_CTX,
 485         .arg2_type      = ARG_CONST_MAP_PTR,
 486         .arg3_type      = ARG_ANYTHING,
 487 };
 488
 489 static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 490                             void *buf, u32 size, u64 flags)
 491 {
 492         u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
 493         bool user_build_id = flags & BPF_F_USER_BUILD_ID;
 494         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 495         bool user = flags & BPF_F_USER_STACK;
 496         struct perf_callchain_entry *trace;
 497         bool kernel = !user;
 498         int err = -EINVAL;
 499         u64 *ips;
 500
 501         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 502                                BPF_F_USER_BUILD_ID)))
 503                 goto clear;
 504         if (kernel && user_build_id)
 505                 goto clear;
 506
 507         elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
 508                                             : sizeof(u64);
 509         if (unlikely(size % elem_size))
 510                 goto clear;
 511
 512         /* cannot get valid user stack for task without user_mode regs */
 513         if (task && user && !user_mode(regs))
 514                 goto err_fault;
 515
 516         num_elem = size / elem_size;
 517         if (sysctl_perf_event_max_stack < num_elem)
 518                 init_nr = 0;
 519         else
 520                 init_nr = sysctl_perf_event_max_stack - num_elem;
 521
 522         if (kernel && task)
 523                 trace = get_callchain_entry_for_task(task, init_nr);
 524         else
 525                 trace = get_perf_callchain(regs, init_nr, kernel, user,
 526                                            sysctl_perf_event_max_stack,
 527                                            false, false);
 528         if (unlikely(!trace))
 529                 goto err_fault;
 530
 531         trace_nr = trace->nr - init_nr;
 532         if (trace_nr < skip)
 533                 goto err_fault;
 534
 535         trace_nr -= skip;
 536         trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
 537         copy_len = trace_nr * elem_size;
 538         ips = trace->ip + skip + init_nr;
 539         if (user && user_build_id)
 540                 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
 541         else
 542                 memcpy(buf, ips, copy_len);
 543
 544         if (size > copy_len)
 545                 memset(buf + copy_len, 0, size - copy_len);
 546         return copy_len;
 547
 548 err_fault:
 549         err = -EFAULT;
 550 clear:
 551         memset(buf, 0, size);
 552         return err;
 553 }
 554
 555 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
 556            u64, flags)
 557 {
 558         return __bpf_get_stack(regs, NULL, buf, size, flags);
 559 }
 560
 561 const struct bpf_func_proto bpf_get_stack_proto = {
 562         .func           = bpf_get_stack,
 563         .gpl_only       = true,
 564         .ret_type       = RET_INTEGER,
 565         .arg1_type      = ARG_PTR_TO_CTX,
 566         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 567         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 568         .arg4_type      = ARG_ANYTHING,
 569 };
 570
 571 BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
 572            u32, size, u64, flags)
 573 {
 574         struct pt_regs *regs = task_pt_regs(task);
 575
 576         return __bpf_get_stack(regs, task, buf, size, flags);
 577 }
 578
 579 static int bpf_get_task_stack_btf_ids[5];
 580 const struct bpf_func_proto bpf_get_task_stack_proto = {
 581         .func           = bpf_get_task_stack,
 582         .gpl_only       = false,
 583         .ret_type       = RET_INTEGER,
 584         .arg1_type      = ARG_PTR_TO_BTF_ID,
 585         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 586         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 587         .arg4_type      = ARG_ANYTHING,
 588         .btf_id         = bpf_get_task_stack_btf_ids,
 589 };
 590
 591 /* Called from eBPF program */
 592 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 593 {
 594         return ERR_PTR(-EOPNOTSUPP);
 595 }
 596
 597 /* Called from syscall */
 598 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 599 {
 600         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 601         struct stack_map_bucket *bucket, *old_bucket;
 602         u32 id = *(u32 *)key, trace_len;
 603
 604         if (unlikely(id >= smap->n_buckets))
 605                 return -ENOENT;
 606
 607         bucket = xchg(&smap->buckets[id], NULL);
 608         if (!bucket)
 609                 return -ENOENT;
 610
 611         trace_len = bucket->nr * stack_map_data_size(map);
 612         memcpy(value, bucket->data, trace_len);
 613         memset(value + trace_len, 0, map->value_size - trace_len);
 614
 615         old_bucket = xchg(&smap->buckets[id], bucket);
 616         if (old_bucket)
 617                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 618         return 0;
 619 }
 620
 621 static int stack_map_get_next_key(struct bpf_map *map, void *key,
 622                                   void *next_key)
 623 {
 624         struct bpf_stack_map *smap = container_of(map,
 625                                                   struct bpf_stack_map, map);
 626         u32 id;
 627
 628         WARN_ON_ONCE(!rcu_read_lock_held());
 629
 630         if (!key) {
 631                 id = 0;
 632         } else {
 633                 id = *(u32 *)key;
 634                 if (id >= smap->n_buckets || !smap->buckets[id])
 635                         id = 0;
 636                 else
 637                         id++;
 638         }
 639
 640         while (id < smap->n_buckets && !smap->buckets[id])
 641                 id++;
 642
 643         if (id >= smap->n_buckets)
 644                 return -ENOENT;
 645
 646         *(u32 *)next_key = id;
 647         return 0;
 648 }
 649
 650 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 651                                  u64 map_flags)
 652 {
 653         return -EINVAL;
 654 }
 655
 656 /* Called from syscall or from eBPF program */
 657 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 658 {
 659         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 660         struct stack_map_bucket *old_bucket;
 661         u32 id = *(u32 *)key;
 662
 663         if (unlikely(id >= smap->n_buckets))
 664                 return -E2BIG;
 665
 666         old_bucket = xchg(&smap->buckets[id], NULL);
 667         if (old_bucket) {
 668                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 669                 return 0;
 670         } else {
 671                 return -ENOENT;
 672         }
 673 }
 674
 675 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 676 static void stack_map_free(struct bpf_map *map)
 677 {
 678         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 679
 680         bpf_map_area_free(smap->elems);
 681         pcpu_freelist_destroy(&smap->freelist);
 682         bpf_map_area_free(smap);
 683         put_callchain_buffers();
 684 }
 685
 686 static int stack_trace_map_btf_id;
 687 const struct bpf_map_ops stack_trace_map_ops = {
 688         .map_alloc = stack_map_alloc,
 689         .map_free = stack_map_free,
 690         .map_get_next_key = stack_map_get_next_key,
 691         .map_lookup_elem = stack_map_lookup_elem,
 692         .map_update_elem = stack_map_update_elem,
 693         .map_delete_elem = stack_map_delete_elem,
 694         .map_check_btf = map_check_no_btf,
 695         .map_btf_name = "bpf_stack_map",
 696         .map_btf_id = &stack_trace_map_btf_id,
 697 };
 698
 699 static int __init stack_map_init(void)
 700 {
 701         int cpu;
 702         struct stack_map_irq_work *work;
 703
 704         for_each_possible_cpu(cpu) {
 705                 work = per_cpu_ptr(&up_read_work, cpu);
 706                 init_irq_work(&work->irq_work, do_up_read);
 707         }
 708         return 0;
 709 }
 710 subsys_initcall(stack_map_init);