Loading...
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 "percpu_freelist.h"
15
16#define STACK_CREATE_FLAG_MASK \
17 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
18 BPF_F_STACK_BUILD_ID)
19
20struct stack_map_bucket {
21 struct pcpu_freelist_node fnode;
22 u32 hash;
23 u32 nr;
24 u64 data[];
25};
26
27struct bpf_stack_map {
28 struct bpf_map map;
29 void *elems;
30 struct pcpu_freelist freelist;
31 u32 n_buckets;
32 struct stack_map_bucket *buckets[];
33};
34
35static inline bool stack_map_use_build_id(struct bpf_map *map)
36{
37 return (map->map_flags & BPF_F_STACK_BUILD_ID);
38}
39
40static inline int stack_map_data_size(struct bpf_map *map)
41{
42 return stack_map_use_build_id(map) ?
43 sizeof(struct bpf_stack_build_id) : sizeof(u64);
44}
45
46static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
47{
48 u32 elem_size = sizeof(struct stack_map_bucket) + 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
64free_elems:
65 bpf_map_area_free(smap->elems);
66 return err;
67}
68
69/* Called from syscall */
70static 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 (!capable(CAP_SYS_ADMIN))
78 return ERR_PTR(-EPERM);
79
80 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
81 return ERR_PTR(-EINVAL);
82
83 /* check sanity of attributes */
84 if (attr->max_entries == 0 || attr->key_size != 4 ||
85 value_size < 8 || value_size % 8)
86 return ERR_PTR(-EINVAL);
87
88 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
89 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
90 if (value_size % sizeof(struct bpf_stack_build_id) ||
91 value_size / sizeof(struct bpf_stack_build_id)
92 > sysctl_perf_event_max_stack)
93 return ERR_PTR(-EINVAL);
94 } else if (value_size / 8 > sysctl_perf_event_max_stack)
95 return ERR_PTR(-EINVAL);
96
97 /* hash table size must be power of 2 */
98 n_buckets = roundup_pow_of_two(attr->max_entries);
99
100 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
101 if (cost >= U32_MAX - PAGE_SIZE)
102 return ERR_PTR(-E2BIG);
103
104 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
105 if (!smap)
106 return ERR_PTR(-ENOMEM);
107
108 err = -E2BIG;
109 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
110 if (cost >= U32_MAX - PAGE_SIZE)
111 goto free_smap;
112
113 bpf_map_init_from_attr(&smap->map, attr);
114 smap->map.value_size = value_size;
115 smap->n_buckets = n_buckets;
116 smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
117
118 err = bpf_map_precharge_memlock(smap->map.pages);
119 if (err)
120 goto free_smap;
121
122 err = get_callchain_buffers(sysctl_perf_event_max_stack);
123 if (err)
124 goto free_smap;
125
126 err = prealloc_elems_and_freelist(smap);
127 if (err)
128 goto put_buffers;
129
130 return &smap->map;
131
132put_buffers:
133 put_callchain_buffers();
134free_smap:
135 bpf_map_area_free(smap);
136 return ERR_PTR(err);
137}
138
139#define BPF_BUILD_ID 3
140/*
141 * Parse build id from the note segment. This logic can be shared between
142 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
143 * identical.
144 */
145static inline int stack_map_parse_build_id(void *page_addr,
146 unsigned char *build_id,
147 void *note_start,
148 Elf32_Word note_size)
149{
150 Elf32_Word note_offs = 0, new_offs;
151
152 /* check for overflow */
153 if (note_start < page_addr || note_start + note_size < note_start)
154 return -EINVAL;
155
156 /* only supports note that fits in the first page */
157 if (note_start + note_size > page_addr + PAGE_SIZE)
158 return -EINVAL;
159
160 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
161 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
162
163 if (nhdr->n_type == BPF_BUILD_ID &&
164 nhdr->n_namesz == sizeof("GNU") &&
165 nhdr->n_descsz == BPF_BUILD_ID_SIZE) {
166 memcpy(build_id,
167 note_start + note_offs +
168 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
169 BPF_BUILD_ID_SIZE);
170 return 0;
171 }
172 new_offs = note_offs + sizeof(Elf32_Nhdr) +
173 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
174 if (new_offs <= note_offs) /* overflow */
175 break;
176 note_offs = new_offs;
177 }
178 return -EINVAL;
179}
180
181/* Parse build ID from 32-bit ELF */
182static int stack_map_get_build_id_32(void *page_addr,
183 unsigned char *build_id)
184{
185 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
186 Elf32_Phdr *phdr;
187 int i;
188
189 /* only supports phdr that fits in one page */
190 if (ehdr->e_phnum >
191 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
192 return -EINVAL;
193
194 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
195
196 for (i = 0; i < ehdr->e_phnum; ++i)
197 if (phdr[i].p_type == PT_NOTE)
198 return stack_map_parse_build_id(page_addr, build_id,
199 page_addr + phdr[i].p_offset,
200 phdr[i].p_filesz);
201 return -EINVAL;
202}
203
204/* Parse build ID from 64-bit ELF */
205static int stack_map_get_build_id_64(void *page_addr,
206 unsigned char *build_id)
207{
208 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
209 Elf64_Phdr *phdr;
210 int i;
211
212 /* only supports phdr that fits in one page */
213 if (ehdr->e_phnum >
214 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
215 return -EINVAL;
216
217 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
218
219 for (i = 0; i < ehdr->e_phnum; ++i)
220 if (phdr[i].p_type == PT_NOTE)
221 return stack_map_parse_build_id(page_addr, build_id,
222 page_addr + phdr[i].p_offset,
223 phdr[i].p_filesz);
224 return -EINVAL;
225}
226
227/* Parse build ID of ELF file mapped to vma */
228static int stack_map_get_build_id(struct vm_area_struct *vma,
229 unsigned char *build_id)
230{
231 Elf32_Ehdr *ehdr;
232 struct page *page;
233 void *page_addr;
234 int ret;
235
236 /* only works for page backed storage */
237 if (!vma->vm_file)
238 return -EINVAL;
239
240 page = find_get_page(vma->vm_file->f_mapping, 0);
241 if (!page)
242 return -EFAULT; /* page not mapped */
243
244 ret = -EINVAL;
245 page_addr = page_address(page);
246 ehdr = (Elf32_Ehdr *)page_addr;
247
248 /* compare magic x7f "ELF" */
249 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
250 goto out;
251
252 /* only support executable file and shared object file */
253 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
254 goto out;
255
256 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
257 ret = stack_map_get_build_id_32(page_addr, build_id);
258 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
259 ret = stack_map_get_build_id_64(page_addr, build_id);
260out:
261 put_page(page);
262 return ret;
263}
264
265static void stack_map_get_build_id_offset(struct bpf_map *map,
266 struct stack_map_bucket *bucket,
267 u64 *ips, u32 trace_nr, bool user)
268{
269 int i;
270 struct vm_area_struct *vma;
271 struct bpf_stack_build_id *id_offs;
272
273 bucket->nr = trace_nr;
274 id_offs = (struct bpf_stack_build_id *)bucket->data;
275
276 /*
277 * We cannot do up_read() in nmi context, so build_id lookup is
278 * only supported for non-nmi events. If at some point, it is
279 * possible to run find_vma() without taking the semaphore, we
280 * would like to allow build_id lookup in nmi context.
281 *
282 * Same fallback is used for kernel stack (!user) on a stackmap
283 * with build_id.
284 */
285 if (!user || !current || !current->mm || in_nmi() ||
286 down_read_trylock(¤t->mm->mmap_sem) == 0) {
287 /* cannot access current->mm, fall back to ips */
288 for (i = 0; i < trace_nr; i++) {
289 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
290 id_offs[i].ip = ips[i];
291 }
292 return;
293 }
294
295 for (i = 0; i < trace_nr; i++) {
296 vma = find_vma(current->mm, ips[i]);
297 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
298 /* per entry fall back to ips */
299 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
300 id_offs[i].ip = ips[i];
301 continue;
302 }
303 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
304 - vma->vm_start;
305 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
306 }
307 up_read(¤t->mm->mmap_sem);
308}
309
310BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
311 u64, flags)
312{
313 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
314 struct perf_callchain_entry *trace;
315 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
316 u32 max_depth = map->value_size / stack_map_data_size(map);
317 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
318 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
319 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
320 u32 hash, id, trace_nr, trace_len;
321 bool user = flags & BPF_F_USER_STACK;
322 bool kernel = !user;
323 u64 *ips;
324 bool hash_matches;
325
326 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
327 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
328 return -EINVAL;
329
330 trace = get_perf_callchain(regs, init_nr, kernel, user,
331 sysctl_perf_event_max_stack, false, false);
332
333 if (unlikely(!trace))
334 /* couldn't fetch the stack trace */
335 return -EFAULT;
336
337 /* get_perf_callchain() guarantees that trace->nr >= init_nr
338 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
339 */
340 trace_nr = trace->nr - init_nr;
341
342 if (trace_nr <= skip)
343 /* skipping more than usable stack trace */
344 return -EFAULT;
345
346 trace_nr -= skip;
347 trace_len = trace_nr * sizeof(u64);
348 ips = trace->ip + skip + init_nr;
349 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
350 id = hash & (smap->n_buckets - 1);
351 bucket = READ_ONCE(smap->buckets[id]);
352
353 hash_matches = bucket && bucket->hash == hash;
354 /* fast cmp */
355 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
356 return id;
357
358 if (stack_map_use_build_id(map)) {
359 /* for build_id+offset, pop a bucket before slow cmp */
360 new_bucket = (struct stack_map_bucket *)
361 pcpu_freelist_pop(&smap->freelist);
362 if (unlikely(!new_bucket))
363 return -ENOMEM;
364 stack_map_get_build_id_offset(map, new_bucket, ips,
365 trace_nr, user);
366 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
367 if (hash_matches && bucket->nr == trace_nr &&
368 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
369 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
370 return id;
371 }
372 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
373 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
374 return -EEXIST;
375 }
376 } else {
377 if (hash_matches && bucket->nr == trace_nr &&
378 memcmp(bucket->data, ips, trace_len) == 0)
379 return id;
380 if (bucket && !(flags & BPF_F_REUSE_STACKID))
381 return -EEXIST;
382
383 new_bucket = (struct stack_map_bucket *)
384 pcpu_freelist_pop(&smap->freelist);
385 if (unlikely(!new_bucket))
386 return -ENOMEM;
387 memcpy(new_bucket->data, ips, trace_len);
388 }
389
390 new_bucket->hash = hash;
391 new_bucket->nr = trace_nr;
392
393 old_bucket = xchg(&smap->buckets[id], new_bucket);
394 if (old_bucket)
395 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
396 return id;
397}
398
399const struct bpf_func_proto bpf_get_stackid_proto = {
400 .func = bpf_get_stackid,
401 .gpl_only = true,
402 .ret_type = RET_INTEGER,
403 .arg1_type = ARG_PTR_TO_CTX,
404 .arg2_type = ARG_CONST_MAP_PTR,
405 .arg3_type = ARG_ANYTHING,
406};
407
408/* Called from eBPF program */
409static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
410{
411 return NULL;
412}
413
414/* Called from syscall */
415int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
416{
417 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
418 struct stack_map_bucket *bucket, *old_bucket;
419 u32 id = *(u32 *)key, trace_len;
420
421 if (unlikely(id >= smap->n_buckets))
422 return -ENOENT;
423
424 bucket = xchg(&smap->buckets[id], NULL);
425 if (!bucket)
426 return -ENOENT;
427
428 trace_len = bucket->nr * stack_map_data_size(map);
429 memcpy(value, bucket->data, trace_len);
430 memset(value + trace_len, 0, map->value_size - trace_len);
431
432 old_bucket = xchg(&smap->buckets[id], bucket);
433 if (old_bucket)
434 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
435 return 0;
436}
437
438static int stack_map_get_next_key(struct bpf_map *map, void *key,
439 void *next_key)
440{
441 struct bpf_stack_map *smap = container_of(map,
442 struct bpf_stack_map, map);
443 u32 id;
444
445 WARN_ON_ONCE(!rcu_read_lock_held());
446
447 if (!key) {
448 id = 0;
449 } else {
450 id = *(u32 *)key;
451 if (id >= smap->n_buckets || !smap->buckets[id])
452 id = 0;
453 else
454 id++;
455 }
456
457 while (id < smap->n_buckets && !smap->buckets[id])
458 id++;
459
460 if (id >= smap->n_buckets)
461 return -ENOENT;
462
463 *(u32 *)next_key = id;
464 return 0;
465}
466
467static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
468 u64 map_flags)
469{
470 return -EINVAL;
471}
472
473/* Called from syscall or from eBPF program */
474static int stack_map_delete_elem(struct bpf_map *map, void *key)
475{
476 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
477 struct stack_map_bucket *old_bucket;
478 u32 id = *(u32 *)key;
479
480 if (unlikely(id >= smap->n_buckets))
481 return -E2BIG;
482
483 old_bucket = xchg(&smap->buckets[id], NULL);
484 if (old_bucket) {
485 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
486 return 0;
487 } else {
488 return -ENOENT;
489 }
490}
491
492/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
493static void stack_map_free(struct bpf_map *map)
494{
495 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
496
497 /* wait for bpf programs to complete before freeing stack map */
498 synchronize_rcu();
499
500 bpf_map_area_free(smap->elems);
501 pcpu_freelist_destroy(&smap->freelist);
502 bpf_map_area_free(smap);
503 put_callchain_buffers();
504}
505
506const struct bpf_map_ops stack_map_ops = {
507 .map_alloc = stack_map_alloc,
508 .map_free = stack_map_free,
509 .map_get_next_key = stack_map_get_next_key,
510 .map_lookup_elem = stack_map_lookup_elem,
511 .map_update_elem = stack_map_update_elem,
512 .map_delete_elem = stack_map_delete_elem,
513};
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
19struct stack_map_bucket {
20 struct pcpu_freelist_node fnode;
21 u32 hash;
22 u32 nr;
23 u64 data[];
24};
25
26struct 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
34static inline bool stack_map_use_build_id(struct bpf_map *map)
35{
36 return (map->map_flags & BPF_F_STACK_BUILD_ID);
37}
38
39static 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
45static 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
64free_elems:
65 bpf_map_area_free(smap->elems);
66 return err;
67}
68
69/* Called from syscall */
70static 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
120put_buffers:
121 put_callchain_buffers();
122free_smap:
123 bpf_map_area_free(smap);
124 return ERR_PTR(err);
125}
126
127static 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 */
143static 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 }
181build_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
190static struct perf_callchain_entry *
191get_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
227static 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
300BPF_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
327const 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
336static __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
348BPF_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
396const 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
405static 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
489err_fault:
490 err = -EFAULT;
491clear:
492 memset(buf, 0, size);
493 return err;
494}
495
496BPF_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
502const 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
512BPF_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
518const 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
528static 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
545BPF_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
551const 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
562BPF_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
568const 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
579BPF_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
626clear:
627 memset(buf, 0, size);
628 return err;
629
630}
631
632const 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 */
643static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
644{
645 return ERR_PTR(-EOPNOTSUPP);
646}
647
648/* Called from syscall */
649int 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
672static 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
701static 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 */
708static 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 */
727static 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
737static 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
750BTF_ID_LIST_SINGLE(stack_trace_map_btf_ids, struct, bpf_stack_map)
751const 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};