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/kernel.h>
   8 #include <linux/stacktrace.h>
   9 #include <linux/perf_event.h>
  10 #include <linux/btf_ids.h>
  11 #include <linux/buildid.h>
  12 #include "percpu_freelist.h"
  13 #include "mmap_unlock_work.h"
  14
  15 #define STACK_CREATE_FLAG_MASK                                  \
  16         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY |        \
  17          BPF_F_STACK_BUILD_ID)
  18
  19 struct stack_map_bucket {
  20         struct pcpu_freelist_node fnode;
  21         u32 hash;
  22         u32 nr;
  23         u64 data[];
  24 };
  25
  26 struct bpf_stack_map {
  27         struct bpf_map map;
  28         void *elems;
  29         struct pcpu_freelist freelist;
  30         u32 n_buckets;
  31         struct stack_map_bucket *buckets[] __counted_by(n_buckets);
  32 };
  33
  34 static inline bool stack_map_use_build_id(struct bpf_map *map)
  35 {
  36         return (map->map_flags & BPF_F_STACK_BUILD_ID);
  37 }
  38
  39 static inline int stack_map_data_size(struct bpf_map *map)
  40 {
  41         return stack_map_use_build_id(map) ?
  42                 sizeof(struct bpf_stack_build_id) : sizeof(u64);
  43 }
  44
  45 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
  46 {
  47         u64 elem_size = sizeof(struct stack_map_bucket) +
  48                         (u64)smap->map.value_size;
  49         int err;
  50
  51         smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
  52                                          smap->map.numa_node);
  53         if (!smap->elems)
  54                 return -ENOMEM;
  55
  56         err = pcpu_freelist_init(&smap->freelist);
  57         if (err)
  58                 goto free_elems;
  59
  60         pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
  61                                smap->map.max_entries);
  62         return 0;
  63
  64 free_elems:
  65         bpf_map_area_free(smap->elems);
  66         return err;
  67 }
  68
  69 /* Called from syscall */
  70 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
  71 {
  72         u32 value_size = attr->value_size;
  73         struct bpf_stack_map *smap;
  74         u64 cost, n_buckets;
  75         int err;
  76
  77         if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
  78                 return ERR_PTR(-EINVAL);
  79
  80         /* check sanity of attributes */
  81         if (attr->max_entries == 0 || attr->key_size != 4 ||
  82             value_size < 8 || value_size % 8)
  83                 return ERR_PTR(-EINVAL);
  84
  85         BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
  86         if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
  87                 if (value_size % sizeof(struct bpf_stack_build_id) ||
  88                     value_size / sizeof(struct bpf_stack_build_id)
  89                     > sysctl_perf_event_max_stack)
  90                         return ERR_PTR(-EINVAL);
  91         } else if (value_size / 8 > sysctl_perf_event_max_stack)
  92                 return ERR_PTR(-EINVAL);
  93
  94         /* hash table size must be power of 2; roundup_pow_of_two() can overflow
  95          * into UB on 32-bit arches, so check that first
  96          */
  97         if (attr->max_entries > 1UL << 31)
  98                 return ERR_PTR(-E2BIG);
  99
 100         n_buckets = roundup_pow_of_two(attr->max_entries);
 101
 102         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 103         smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
 104         if (!smap)
 105                 return ERR_PTR(-ENOMEM);
 106
 107         bpf_map_init_from_attr(&smap->map, attr);
 108         smap->n_buckets = n_buckets;
 109
 110         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 111         if (err)
 112                 goto free_smap;
 113
 114         err = prealloc_elems_and_freelist(smap);
 115         if (err)
 116                 goto put_buffers;
 117
 118         return &smap->map;
 119
 120 put_buffers:
 121         put_callchain_buffers();
 122 free_smap:
 123         bpf_map_area_free(smap);
 124         return ERR_PTR(err);
 125 }
 126
 127 static int fetch_build_id(struct vm_area_struct *vma, unsigned char *build_id, bool may_fault)
 128 {
 129         return may_fault ? build_id_parse(vma, build_id, NULL)
 130                          : build_id_parse_nofault(vma, build_id, NULL);
 131 }
 132
 133 /*
 134  * Expects all id_offs[i].ip values to be set to correct initial IPs.
 135  * They will be subsequently:
 136  *   - either adjusted in place to a file offset, if build ID fetching
 137  *     succeeds; in this case id_offs[i].build_id is set to correct build ID,
 138  *     and id_offs[i].status is set to BPF_STACK_BUILD_ID_VALID;
 139  *   - or IP will be kept intact, if build ID fetching failed; in this case
 140  *     id_offs[i].build_id is zeroed out and id_offs[i].status is set to
 141  *     BPF_STACK_BUILD_ID_IP.
 142  */
 143 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 144                                           u32 trace_nr, bool user, bool may_fault)
 145 {
 146         int i;
 147         struct mmap_unlock_irq_work *work = NULL;
 148         bool irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
 149         struct vm_area_struct *vma, *prev_vma = NULL;
 150         const char *prev_build_id;
 151
 152         /* If the irq_work is in use, fall back to report ips. Same
 153          * fallback is used for kernel stack (!user) on a stackmap with
 154          * build_id.
 155          */
 156         if (!user || !current || !current->mm || irq_work_busy ||
 157             !mmap_read_trylock(current->mm)) {
 158                 /* cannot access current->mm, fall back to ips */
 159                 for (i = 0; i < trace_nr; i++) {
 160                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 161                         memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
 162                 }
 163                 return;
 164         }
 165
 166         for (i = 0; i < trace_nr; i++) {
 167                 u64 ip = READ_ONCE(id_offs[i].ip);
 168
 169                 if (range_in_vma(prev_vma, ip, ip)) {
 170                         vma = prev_vma;
 171                         memcpy(id_offs[i].build_id, prev_build_id, BUILD_ID_SIZE_MAX);
 172                         goto build_id_valid;
 173                 }
 174                 vma = find_vma(current->mm, ip);
 175                 if (!vma || fetch_build_id(vma, id_offs[i].build_id, may_fault)) {
 176                         /* per entry fall back to ips */
 177                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 178                         memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
 179                         continue;
 180                 }
 181 build_id_valid:
 182                 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ip - vma->vm_start;
 183                 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
 184                 prev_vma = vma;
 185                 prev_build_id = id_offs[i].build_id;
 186         }
 187         bpf_mmap_unlock_mm(work, current->mm);
 188 }
 189
 190 static struct perf_callchain_entry *
 191 get_callchain_entry_for_task(struct task_struct *task, u32 max_depth)
 192 {
 193 #ifdef CONFIG_STACKTRACE
 194         struct perf_callchain_entry *entry;
 195         int rctx;
 196
 197         entry = get_callchain_entry(&rctx);
 198
 199         if (!entry)
 200                 return NULL;
 201
 202         entry->nr = stack_trace_save_tsk(task, (unsigned long *)entry->ip,
 203                                          max_depth, 0);
 204
 205         /* stack_trace_save_tsk() works on unsigned long array, while
 206          * perf_callchain_entry uses u64 array. For 32-bit systems, it is
 207          * necessary to fix this mismatch.
 208          */
 209         if (__BITS_PER_LONG != 64) {
 210                 unsigned long *from = (unsigned long *) entry->ip;
 211                 u64 *to = entry->ip;
 212                 int i;
 213
 214                 /* copy data from the end to avoid using extra buffer */
 215                 for (i = entry->nr - 1; i >= 0; i--)
 216                         to[i] = (u64)(from[i]);
 217         }
 218
 219         put_callchain_entry(rctx);
 220
 221         return entry;
 222 #else /* CONFIG_STACKTRACE */
 223         return NULL;
 224 #endif
 225 }
 226
 227 static long __bpf_get_stackid(struct bpf_map *map,
 228                               struct perf_callchain_entry *trace, u64 flags)
 229 {
 230         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 231         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 232         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 233         u32 hash, id, trace_nr, trace_len, i;
 234         bool user = flags & BPF_F_USER_STACK;
 235         u64 *ips;
 236         bool hash_matches;
 237
 238         if (trace->nr <= skip)
 239                 /* skipping more than usable stack trace */
 240                 return -EFAULT;
 241
 242         trace_nr = trace->nr - skip;
 243         trace_len = trace_nr * sizeof(u64);
 244         ips = trace->ip + skip;
 245         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 246         id = hash & (smap->n_buckets - 1);
 247         bucket = READ_ONCE(smap->buckets[id]);
 248
 249         hash_matches = bucket && bucket->hash == hash;
 250         /* fast cmp */
 251         if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
 252                 return id;
 253
 254         if (stack_map_use_build_id(map)) {
 255                 struct bpf_stack_build_id *id_offs;
 256
 257                 /* for build_id+offset, pop a bucket before slow cmp */
 258                 new_bucket = (struct stack_map_bucket *)
 259                         pcpu_freelist_pop(&smap->freelist);
 260                 if (unlikely(!new_bucket))
 261                         return -ENOMEM;
 262                 new_bucket->nr = trace_nr;
 263                 id_offs = (struct bpf_stack_build_id *)new_bucket->data;
 264                 for (i = 0; i < trace_nr; i++)
 265                         id_offs[i].ip = ips[i];
 266                 stack_map_get_build_id_offset(id_offs, trace_nr, user, false /* !may_fault */);
 267                 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
 268                 if (hash_matches && bucket->nr == trace_nr &&
 269                     memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
 270                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 271                         return id;
 272                 }
 273                 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
 274                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 275                         return -EEXIST;
 276                 }
 277         } else {
 278                 if (hash_matches && bucket->nr == trace_nr &&
 279                     memcmp(bucket->data, ips, trace_len) == 0)
 280                         return id;
 281                 if (bucket && !(flags & BPF_F_REUSE_STACKID))
 282                         return -EEXIST;
 283
 284                 new_bucket = (struct stack_map_bucket *)
 285                         pcpu_freelist_pop(&smap->freelist);
 286                 if (unlikely(!new_bucket))
 287                         return -ENOMEM;
 288                 memcpy(new_bucket->data, ips, trace_len);
 289         }
 290
 291         new_bucket->hash = hash;
 292         new_bucket->nr = trace_nr;
 293
 294         old_bucket = xchg(&smap->buckets[id], new_bucket);
 295         if (old_bucket)
 296                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 297         return id;
 298 }
 299
 300 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 301            u64, flags)
 302 {
 303         u32 max_depth = map->value_size / stack_map_data_size(map);
 304         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 305         bool user = flags & BPF_F_USER_STACK;
 306         struct perf_callchain_entry *trace;
 307         bool kernel = !user;
 308
 309         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 310                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 311                 return -EINVAL;
 312
 313         max_depth += skip;
 314         if (max_depth > sysctl_perf_event_max_stack)
 315                 max_depth = sysctl_perf_event_max_stack;
 316
 317         trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
 318                                    false, false);
 319
 320         if (unlikely(!trace))
 321                 /* couldn't fetch the stack trace */
 322                 return -EFAULT;
 323
 324         return __bpf_get_stackid(map, trace, flags);
 325 }
 326
 327 const struct bpf_func_proto bpf_get_stackid_proto = {
 328         .func           = bpf_get_stackid,
 329         .gpl_only       = true,
 330         .ret_type       = RET_INTEGER,
 331         .arg1_type      = ARG_PTR_TO_CTX,
 332         .arg2_type      = ARG_CONST_MAP_PTR,
 333         .arg3_type      = ARG_ANYTHING,
 334 };
 335
 336 static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
 337 {
 338         __u64 nr_kernel = 0;
 339
 340         while (nr_kernel < trace->nr) {
 341                 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
 342                         break;
 343                 nr_kernel++;
 344         }
 345         return nr_kernel;
 346 }
 347
 348 BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
 349            struct bpf_map *, map, u64, flags)
 350 {
 351         struct perf_event *event = ctx->event;
 352         struct perf_callchain_entry *trace;
 353         bool kernel, user;
 354         __u64 nr_kernel;
 355         int ret;
 356
 357         /* perf_sample_data doesn't have callchain, use bpf_get_stackid */
 358         if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
 359                 return bpf_get_stackid((unsigned long)(ctx->regs),
 360                                        (unsigned long) map, flags, 0, 0);
 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         user = flags & BPF_F_USER_STACK;
 367         kernel = !user;
 368
 369         trace = ctx->data->callchain;
 370         if (unlikely(!trace))
 371                 return -EFAULT;
 372
 373         nr_kernel = count_kernel_ip(trace);
 374
 375         if (kernel) {
 376                 __u64 nr = trace->nr;
 377
 378                 trace->nr = nr_kernel;
 379                 ret = __bpf_get_stackid(map, trace, flags);
 380
 381                 /* restore nr */
 382                 trace->nr = nr;
 383         } else { /* user */
 384                 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
 385
 386                 skip += nr_kernel;
 387                 if (skip > BPF_F_SKIP_FIELD_MASK)
 388                         return -EFAULT;
 389
 390                 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
 391                 ret = __bpf_get_stackid(map, trace, flags);
 392         }
 393         return ret;
 394 }
 395
 396 const struct bpf_func_proto bpf_get_stackid_proto_pe = {
 397         .func           = bpf_get_stackid_pe,
 398         .gpl_only       = false,
 399         .ret_type       = RET_INTEGER,
 400         .arg1_type      = ARG_PTR_TO_CTX,
 401         .arg2_type      = ARG_CONST_MAP_PTR,
 402         .arg3_type      = ARG_ANYTHING,
 403 };
 404
 405 static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 406                             struct perf_callchain_entry *trace_in,
 407                             void *buf, u32 size, u64 flags, bool may_fault)
 408 {
 409         u32 trace_nr, copy_len, elem_size, num_elem, max_depth;
 410         bool user_build_id = flags & BPF_F_USER_BUILD_ID;
 411         bool crosstask = task && task != current;
 412         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 413         bool user = flags & BPF_F_USER_STACK;
 414         struct perf_callchain_entry *trace;
 415         bool kernel = !user;
 416         int err = -EINVAL;
 417         u64 *ips;
 418
 419         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 420                                BPF_F_USER_BUILD_ID)))
 421                 goto clear;
 422         if (kernel && user_build_id)
 423                 goto clear;
 424
 425         elem_size = user_build_id ? sizeof(struct bpf_stack_build_id) : sizeof(u64);
 426         if (unlikely(size % elem_size))
 427                 goto clear;
 428
 429         /* cannot get valid user stack for task without user_mode regs */
 430         if (task && user && !user_mode(regs))
 431                 goto err_fault;
 432
 433         /* get_perf_callchain does not support crosstask user stack walking
 434          * but returns an empty stack instead of NULL.
 435          */
 436         if (crosstask && user) {
 437                 err = -EOPNOTSUPP;
 438                 goto clear;
 439         }
 440
 441         num_elem = size / elem_size;
 442         max_depth = num_elem + skip;
 443         if (sysctl_perf_event_max_stack < max_depth)
 444                 max_depth = sysctl_perf_event_max_stack;
 445
 446         if (may_fault)
 447                 rcu_read_lock(); /* need RCU for perf's callchain below */
 448
 449         if (trace_in)
 450                 trace = trace_in;
 451         else if (kernel && task)
 452                 trace = get_callchain_entry_for_task(task, max_depth);
 453         else
 454                 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
 455                                            crosstask, false);
 456
 457         if (unlikely(!trace) || trace->nr < skip) {
 458                 if (may_fault)
 459                         rcu_read_unlock();
 460                 goto err_fault;
 461         }
 462
 463         trace_nr = trace->nr - skip;
 464         trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
 465         copy_len = trace_nr * elem_size;
 466
 467         ips = trace->ip + skip;
 468         if (user_build_id) {
 469                 struct bpf_stack_build_id *id_offs = buf;
 470                 u32 i;
 471
 472                 for (i = 0; i < trace_nr; i++)
 473                         id_offs[i].ip = ips[i];
 474         } else {
 475                 memcpy(buf, ips, copy_len);
 476         }
 477
 478         /* trace/ips should not be dereferenced after this point */
 479         if (may_fault)
 480                 rcu_read_unlock();
 481
 482         if (user_build_id)
 483                 stack_map_get_build_id_offset(buf, trace_nr, user, may_fault);
 484
 485         if (size > copy_len)
 486                 memset(buf + copy_len, 0, size - copy_len);
 487         return copy_len;
 488
 489 err_fault:
 490         err = -EFAULT;
 491 clear:
 492         memset(buf, 0, size);
 493         return err;
 494 }
 495
 496 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
 497            u64, flags)
 498 {
 499         return __bpf_get_stack(regs, NULL, NULL, buf, size, flags, false /* !may_fault */);
 500 }
 501
 502 const struct bpf_func_proto bpf_get_stack_proto = {
 503         .func           = bpf_get_stack,
 504         .gpl_only       = true,
 505         .ret_type       = RET_INTEGER,
 506         .arg1_type      = ARG_PTR_TO_CTX,
 507         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 508         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 509         .arg4_type      = ARG_ANYTHING,
 510 };
 511
 512 BPF_CALL_4(bpf_get_stack_sleepable, struct pt_regs *, regs, void *, buf, u32, size,
 513            u64, flags)
 514 {
 515         return __bpf_get_stack(regs, NULL, NULL, buf, size, flags, true /* may_fault */);
 516 }
 517
 518 const struct bpf_func_proto bpf_get_stack_sleepable_proto = {
 519         .func           = bpf_get_stack_sleepable,
 520         .gpl_only       = true,
 521         .ret_type       = RET_INTEGER,
 522         .arg1_type      = ARG_PTR_TO_CTX,
 523         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 524         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 525         .arg4_type      = ARG_ANYTHING,
 526 };
 527
 528 static long __bpf_get_task_stack(struct task_struct *task, void *buf, u32 size,
 529                                  u64 flags, bool may_fault)
 530 {
 531         struct pt_regs *regs;
 532         long res = -EINVAL;
 533
 534         if (!try_get_task_stack(task))
 535                 return -EFAULT;
 536
 537         regs = task_pt_regs(task);
 538         if (regs)
 539                 res = __bpf_get_stack(regs, task, NULL, buf, size, flags, may_fault);
 540         put_task_stack(task);
 541
 542         return res;
 543 }
 544
 545 BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
 546            u32, size, u64, flags)
 547 {
 548         return __bpf_get_task_stack(task, buf, size, flags, false /* !may_fault */);
 549 }
 550
 551 const struct bpf_func_proto bpf_get_task_stack_proto = {
 552         .func           = bpf_get_task_stack,
 553         .gpl_only       = false,
 554         .ret_type       = RET_INTEGER,
 555         .arg1_type      = ARG_PTR_TO_BTF_ID,
 556         .arg1_btf_id    = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 557         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 558         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 559         .arg4_type      = ARG_ANYTHING,
 560 };
 561
 562 BPF_CALL_4(bpf_get_task_stack_sleepable, struct task_struct *, task, void *, buf,
 563            u32, size, u64, flags)
 564 {
 565         return __bpf_get_task_stack(task, buf, size, flags, true /* !may_fault */);
 566 }
 567
 568 const struct bpf_func_proto bpf_get_task_stack_sleepable_proto = {
 569         .func           = bpf_get_task_stack_sleepable,
 570         .gpl_only       = false,
 571         .ret_type       = RET_INTEGER,
 572         .arg1_type      = ARG_PTR_TO_BTF_ID,
 573         .arg1_btf_id    = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
 574         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 575         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 576         .arg4_type      = ARG_ANYTHING,
 577 };
 578
 579 BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
 580            void *, buf, u32, size, u64, flags)
 581 {
 582         struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
 583         struct perf_event *event = ctx->event;
 584         struct perf_callchain_entry *trace;
 585         bool kernel, user;
 586         int err = -EINVAL;
 587         __u64 nr_kernel;
 588
 589         if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
 590                 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags, false /* !may_fault */);
 591
 592         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 593                                BPF_F_USER_BUILD_ID)))
 594                 goto clear;
 595
 596         user = flags & BPF_F_USER_STACK;
 597         kernel = !user;
 598
 599         err = -EFAULT;
 600         trace = ctx->data->callchain;
 601         if (unlikely(!trace))
 602                 goto clear;
 603
 604         nr_kernel = count_kernel_ip(trace);
 605
 606         if (kernel) {
 607                 __u64 nr = trace->nr;
 608
 609                 trace->nr = nr_kernel;
 610                 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags, false /* !may_fault */);
 611
 612                 /* restore nr */
 613                 trace->nr = nr;
 614         } else { /* user */
 615                 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
 616
 617                 skip += nr_kernel;
 618                 if (skip > BPF_F_SKIP_FIELD_MASK)
 619                         goto clear;
 620
 621                 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
 622                 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags, false /* !may_fault */);
 623         }
 624         return err;
 625
 626 clear:
 627         memset(buf, 0, size);
 628         return err;
 629
 630 }
 631
 632 const struct bpf_func_proto bpf_get_stack_proto_pe = {
 633         .func           = bpf_get_stack_pe,
 634         .gpl_only       = true,
 635         .ret_type       = RET_INTEGER,
 636         .arg1_type      = ARG_PTR_TO_CTX,
 637         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 638         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 639         .arg4_type      = ARG_ANYTHING,
 640 };
 641
 642 /* Called from eBPF program */
 643 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 644 {
 645         return ERR_PTR(-EOPNOTSUPP);
 646 }
 647
 648 /* Called from syscall */
 649 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 650 {
 651         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 652         struct stack_map_bucket *bucket, *old_bucket;
 653         u32 id = *(u32 *)key, trace_len;
 654
 655         if (unlikely(id >= smap->n_buckets))
 656                 return -ENOENT;
 657
 658         bucket = xchg(&smap->buckets[id], NULL);
 659         if (!bucket)
 660                 return -ENOENT;
 661
 662         trace_len = bucket->nr * stack_map_data_size(map);
 663         memcpy(value, bucket->data, trace_len);
 664         memset(value + trace_len, 0, map->value_size - trace_len);
 665
 666         old_bucket = xchg(&smap->buckets[id], bucket);
 667         if (old_bucket)
 668                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 669         return 0;
 670 }
 671
 672 static int stack_map_get_next_key(struct bpf_map *map, void *key,
 673                                   void *next_key)
 674 {
 675         struct bpf_stack_map *smap = container_of(map,
 676                                                   struct bpf_stack_map, map);
 677         u32 id;
 678
 679         WARN_ON_ONCE(!rcu_read_lock_held());
 680
 681         if (!key) {
 682                 id = 0;
 683         } else {
 684                 id = *(u32 *)key;
 685                 if (id >= smap->n_buckets || !smap->buckets[id])
 686                         id = 0;
 687                 else
 688                         id++;
 689         }
 690
 691         while (id < smap->n_buckets && !smap->buckets[id])
 692                 id++;
 693
 694         if (id >= smap->n_buckets)
 695                 return -ENOENT;
 696
 697         *(u32 *)next_key = id;
 698         return 0;
 699 }
 700
 701 static long stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 702                                   u64 map_flags)
 703 {
 704         return -EINVAL;
 705 }
 706
 707 /* Called from syscall or from eBPF program */
 708 static long stack_map_delete_elem(struct bpf_map *map, void *key)
 709 {
 710         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 711         struct stack_map_bucket *old_bucket;
 712         u32 id = *(u32 *)key;
 713
 714         if (unlikely(id >= smap->n_buckets))
 715                 return -E2BIG;
 716
 717         old_bucket = xchg(&smap->buckets[id], NULL);
 718         if (old_bucket) {
 719                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 720                 return 0;
 721         } else {
 722                 return -ENOENT;
 723         }
 724 }
 725
 726 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 727 static void stack_map_free(struct bpf_map *map)
 728 {
 729         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 730
 731         bpf_map_area_free(smap->elems);
 732         pcpu_freelist_destroy(&smap->freelist);
 733         bpf_map_area_free(smap);
 734         put_callchain_buffers();
 735 }
 736
 737 static u64 stack_map_mem_usage(const struct bpf_map *map)
 738 {
 739         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 740         u64 value_size = map->value_size;
 741         u64 n_buckets = smap->n_buckets;
 742         u64 enties = map->max_entries;
 743         u64 usage = sizeof(*smap);
 744
 745         usage += n_buckets * sizeof(struct stack_map_bucket *);
 746         usage += enties * (sizeof(struct stack_map_bucket) + value_size);
 747         return usage;
 748 }
 749
 750 BTF_ID_LIST_SINGLE(stack_trace_map_btf_ids, struct, bpf_stack_map)
 751 const struct bpf_map_ops stack_trace_map_ops = {
 752         .map_meta_equal = bpf_map_meta_equal,
 753         .map_alloc = stack_map_alloc,
 754         .map_free = stack_map_free,
 755         .map_get_next_key = stack_map_get_next_key,
 756         .map_lookup_elem = stack_map_lookup_elem,
 757         .map_update_elem = stack_map_update_elem,
 758         .map_delete_elem = stack_map_delete_elem,
 759         .map_check_btf = map_check_no_btf,
 760         .map_mem_usage = stack_map_mem_usage,
 761         .map_btf_id = &stack_trace_map_btf_ids[0],
 762 };