1 /* Copyright (c) 2016 Facebook
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.
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"
16 #define STACK_CREATE_FLAG_MASK \
17 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
20 struct stack_map_bucket {
21 struct pcpu_freelist_node fnode;
27 struct bpf_stack_map {
30 struct pcpu_freelist freelist;
32 struct stack_map_bucket *buckets[];
35 static inline bool stack_map_use_build_id(struct bpf_map *map)
37 return (map->map_flags & BPF_F_STACK_BUILD_ID);
40 static inline int stack_map_data_size(struct bpf_map *map)
42 return stack_map_use_build_id(map) ?
43 sizeof(struct bpf_stack_build_id) : sizeof(u64);
46 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
48 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
51 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
56 err = pcpu_freelist_init(&smap->freelist);
60 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
61 smap->map.max_entries);
65 bpf_map_area_free(smap->elems);
69 /* Called from syscall */
70 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
72 u32 value_size = attr->value_size;
73 struct bpf_stack_map *smap;
77 if (!capable(CAP_SYS_ADMIN))
78 return ERR_PTR(-EPERM);
80 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
81 return ERR_PTR(-EINVAL);
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);
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);
97 /* hash table size must be power of 2 */
98 n_buckets = roundup_pow_of_two(attr->max_entries);
100 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
101 if (cost >= U32_MAX - PAGE_SIZE)
102 return ERR_PTR(-E2BIG);
104 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
106 return ERR_PTR(-ENOMEM);
109 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
110 if (cost >= U32_MAX - PAGE_SIZE)
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;
118 err = bpf_map_precharge_memlock(smap->map.pages);
122 err = get_callchain_buffers(sysctl_perf_event_max_stack);
126 err = prealloc_elems_and_freelist(smap);
133 put_callchain_buffers();
135 bpf_map_area_free(smap);
139 #define BPF_BUILD_ID 3
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
145 static inline int stack_map_parse_build_id(void *page_addr,
146 unsigned char *build_id,
148 Elf32_Word note_size)
150 Elf32_Word note_offs = 0, new_offs;
152 /* check for overflow */
153 if (note_start < page_addr || note_start + note_size < note_start)
156 /* only supports note that fits in the first page */
157 if (note_start + note_size > page_addr + PAGE_SIZE)
160 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
161 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
163 if (nhdr->n_type == BPF_BUILD_ID &&
164 nhdr->n_namesz == sizeof("GNU") &&
165 nhdr->n_descsz == BPF_BUILD_ID_SIZE) {
167 note_start + note_offs +
168 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
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 */
176 note_offs = new_offs;
181 /* Parse build ID from 32-bit ELF */
182 static int stack_map_get_build_id_32(void *page_addr,
183 unsigned char *build_id)
185 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
189 /* only supports phdr that fits in one page */
191 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
194 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
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,
204 /* Parse build ID from 64-bit ELF */
205 static int stack_map_get_build_id_64(void *page_addr,
206 unsigned char *build_id)
208 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
212 /* only supports phdr that fits in one page */
214 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
217 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
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,
227 /* Parse build ID of ELF file mapped to vma */
228 static int stack_map_get_build_id(struct vm_area_struct *vma,
229 unsigned char *build_id)
236 /* only works for page backed storage */
240 page = find_get_page(vma->vm_file->f_mapping, 0);
242 return -EFAULT; /* page not mapped */
245 page_addr = page_address(page);
246 ehdr = (Elf32_Ehdr *)page_addr;
248 /* compare magic x7f "ELF" */
249 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
252 /* only support executable file and shared object file */
253 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
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);
265 static 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)
270 struct vm_area_struct *vma;
271 struct bpf_stack_build_id *id_offs;
273 bucket->nr = trace_nr;
274 id_offs = (struct bpf_stack_build_id *)bucket->data;
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.
282 * Same fallback is used for kernel stack (!user) on a stackmap
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];
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];
303 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
305 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
307 up_read(¤t->mm->mmap_sem);
310 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
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;
326 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
327 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
330 trace = get_perf_callchain(regs, init_nr, kernel, user,
331 sysctl_perf_event_max_stack, false, false);
333 if (unlikely(!trace))
334 /* couldn't fetch the stack trace */
337 /* get_perf_callchain() guarantees that trace->nr >= init_nr
338 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
340 trace_nr = trace->nr - init_nr;
342 if (trace_nr <= skip)
343 /* skipping more than usable stack trace */
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]);
353 hash_matches = bucket && bucket->hash == hash;
355 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
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))
364 stack_map_get_build_id_offset(map, new_bucket, ips,
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);
372 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
373 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
377 if (hash_matches && bucket->nr == trace_nr &&
378 memcmp(bucket->data, ips, trace_len) == 0)
380 if (bucket && !(flags & BPF_F_REUSE_STACKID))
383 new_bucket = (struct stack_map_bucket *)
384 pcpu_freelist_pop(&smap->freelist);
385 if (unlikely(!new_bucket))
387 memcpy(new_bucket->data, ips, trace_len);
390 new_bucket->hash = hash;
391 new_bucket->nr = trace_nr;
393 old_bucket = xchg(&smap->buckets[id], new_bucket);
395 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
399 const struct bpf_func_proto bpf_get_stackid_proto = {
400 .func = bpf_get_stackid,
402 .ret_type = RET_INTEGER,
403 .arg1_type = ARG_PTR_TO_CTX,
404 .arg2_type = ARG_CONST_MAP_PTR,
405 .arg3_type = ARG_ANYTHING,
408 /* Called from eBPF program */
409 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
414 /* Called from syscall */
415 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
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;
421 if (unlikely(id >= smap->n_buckets))
424 bucket = xchg(&smap->buckets[id], NULL);
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);
432 old_bucket = xchg(&smap->buckets[id], bucket);
434 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
438 static int stack_map_get_next_key(struct bpf_map *map, void *key,
441 struct bpf_stack_map *smap = container_of(map,
442 struct bpf_stack_map, map);
445 WARN_ON_ONCE(!rcu_read_lock_held());
451 if (id >= smap->n_buckets || !smap->buckets[id])
457 while (id < smap->n_buckets && !smap->buckets[id])
460 if (id >= smap->n_buckets)
463 *(u32 *)next_key = id;
467 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
473 /* Called from syscall or from eBPF program */
474 static int stack_map_delete_elem(struct bpf_map *map, void *key)
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;
480 if (unlikely(id >= smap->n_buckets))
483 old_bucket = xchg(&smap->buckets[id], NULL);
485 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
492 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
493 static void stack_map_free(struct bpf_map *map)
495 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
497 /* wait for bpf programs to complete before freeing stack map */
500 bpf_map_area_free(smap->elems);
501 pcpu_freelist_destroy(&smap->freelist);
502 bpf_map_area_free(smap);
503 put_callchain_buffers();
506 const 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,