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 <linux/irq_work.h>
15 #include "percpu_freelist.h"
17 #define STACK_CREATE_FLAG_MASK \
18 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
21 struct stack_map_bucket {
22 struct pcpu_freelist_node fnode;
28 struct bpf_stack_map {
31 struct pcpu_freelist freelist;
33 struct stack_map_bucket *buckets[];
36 /* irq_work to run up_read() for build_id lookup in nmi context */
37 struct stack_map_irq_work {
38 struct irq_work irq_work;
39 struct rw_semaphore *sem;
42 static void do_up_read(struct irq_work *entry)
44 struct stack_map_irq_work *work;
46 work = container_of(entry, struct stack_map_irq_work, irq_work);
51 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
53 static inline bool stack_map_use_build_id(struct bpf_map *map)
55 return (map->map_flags & BPF_F_STACK_BUILD_ID);
58 static inline int stack_map_data_size(struct bpf_map *map)
60 return stack_map_use_build_id(map) ?
61 sizeof(struct bpf_stack_build_id) : sizeof(u64);
64 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
66 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
69 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
74 err = pcpu_freelist_init(&smap->freelist);
78 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
79 smap->map.max_entries);
83 bpf_map_area_free(smap->elems);
87 /* Called from syscall */
88 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
90 u32 value_size = attr->value_size;
91 struct bpf_stack_map *smap;
95 if (!capable(CAP_SYS_ADMIN))
96 return ERR_PTR(-EPERM);
98 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
99 return ERR_PTR(-EINVAL);
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);
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);
115 /* hash table size must be power of 2 */
116 n_buckets = roundup_pow_of_two(attr->max_entries);
118 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
119 if (cost >= U32_MAX - PAGE_SIZE)
120 return ERR_PTR(-E2BIG);
122 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
124 return ERR_PTR(-ENOMEM);
127 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
128 if (cost >= U32_MAX - PAGE_SIZE)
131 bpf_map_init_from_attr(&smap->map, attr);
132 smap->map.value_size = value_size;
133 smap->n_buckets = n_buckets;
134 smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
136 err = bpf_map_precharge_memlock(smap->map.pages);
140 err = get_callchain_buffers(sysctl_perf_event_max_stack);
144 err = prealloc_elems_and_freelist(smap);
151 put_callchain_buffers();
153 bpf_map_area_free(smap);
157 #define BPF_BUILD_ID 3
159 * Parse build id from the note segment. This logic can be shared between
160 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
163 static inline int stack_map_parse_build_id(void *page_addr,
164 unsigned char *build_id,
166 Elf32_Word note_size)
168 Elf32_Word note_offs = 0, new_offs;
170 /* check for overflow */
171 if (note_start < page_addr || note_start + note_size < note_start)
174 /* only supports note that fits in the first page */
175 if (note_start + note_size > page_addr + PAGE_SIZE)
178 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
179 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
181 if (nhdr->n_type == BPF_BUILD_ID &&
182 nhdr->n_namesz == sizeof("GNU") &&
183 nhdr->n_descsz == BPF_BUILD_ID_SIZE) {
185 note_start + note_offs +
186 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
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 */
194 note_offs = new_offs;
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)
203 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
207 /* only supports phdr that fits in one page */
209 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
212 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
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,
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)
226 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
230 /* only supports phdr that fits in one page */
232 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
235 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
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,
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)
254 /* only works for page backed storage */
258 page = find_get_page(vma->vm_file->f_mapping, 0);
260 return -EFAULT; /* page not mapped */
263 page_addr = kmap_atomic(page);
264 ehdr = (Elf32_Ehdr *)page_addr;
266 /* compare magic x7f "ELF" */
267 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
270 /* only support executable file and shared object file */
271 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
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);
279 kunmap_atomic(page_addr);
284 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
285 u64 *ips, u32 trace_nr, bool user)
288 struct vm_area_struct *vma;
289 bool irq_work_busy = false;
290 struct stack_map_irq_work *work = NULL;
293 work = this_cpu_ptr(&up_read_work);
294 if (work->irq_work.flags & IRQ_WORK_BUSY)
295 /* cannot queue more up_read, fallback */
296 irq_work_busy = true;
300 * We cannot do up_read() in nmi context. To do build_id lookup
301 * in nmi context, we need to run up_read() in irq_work. We use
302 * a percpu variable to do the irq_work. If the irq_work is
303 * already used by another lookup, we fall back to report ips.
305 * Same fallback is used for kernel stack (!user) on a stackmap
308 if (!user || !current || !current->mm || irq_work_busy ||
309 down_read_trylock(¤t->mm->mmap_sem) == 0) {
310 /* cannot access current->mm, fall back to ips */
311 for (i = 0; i < trace_nr; i++) {
312 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
313 id_offs[i].ip = ips[i];
318 for (i = 0; i < trace_nr; i++) {
319 vma = find_vma(current->mm, ips[i]);
320 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
321 /* per entry fall back to ips */
322 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
323 id_offs[i].ip = ips[i];
326 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
328 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
332 up_read(¤t->mm->mmap_sem);
334 work->sem = ¤t->mm->mmap_sem;
335 irq_work_queue(&work->irq_work);
339 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
342 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
343 struct perf_callchain_entry *trace;
344 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
345 u32 max_depth = map->value_size / stack_map_data_size(map);
346 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
347 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
348 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
349 u32 hash, id, trace_nr, trace_len;
350 bool user = flags & BPF_F_USER_STACK;
355 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
356 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
359 trace = get_perf_callchain(regs, init_nr, kernel, user,
360 sysctl_perf_event_max_stack, false, false);
362 if (unlikely(!trace))
363 /* couldn't fetch the stack trace */
366 /* get_perf_callchain() guarantees that trace->nr >= init_nr
367 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
369 trace_nr = trace->nr - init_nr;
371 if (trace_nr <= skip)
372 /* skipping more than usable stack trace */
376 trace_len = trace_nr * sizeof(u64);
377 ips = trace->ip + skip + init_nr;
378 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
379 id = hash & (smap->n_buckets - 1);
380 bucket = READ_ONCE(smap->buckets[id]);
382 hash_matches = bucket && bucket->hash == hash;
384 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
387 if (stack_map_use_build_id(map)) {
388 /* for build_id+offset, pop a bucket before slow cmp */
389 new_bucket = (struct stack_map_bucket *)
390 pcpu_freelist_pop(&smap->freelist);
391 if (unlikely(!new_bucket))
393 new_bucket->nr = trace_nr;
394 stack_map_get_build_id_offset(
395 (struct bpf_stack_build_id *)new_bucket->data,
396 ips, trace_nr, user);
397 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
398 if (hash_matches && bucket->nr == trace_nr &&
399 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
400 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
403 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
404 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
408 if (hash_matches && bucket->nr == trace_nr &&
409 memcmp(bucket->data, ips, trace_len) == 0)
411 if (bucket && !(flags & BPF_F_REUSE_STACKID))
414 new_bucket = (struct stack_map_bucket *)
415 pcpu_freelist_pop(&smap->freelist);
416 if (unlikely(!new_bucket))
418 memcpy(new_bucket->data, ips, trace_len);
421 new_bucket->hash = hash;
422 new_bucket->nr = trace_nr;
424 old_bucket = xchg(&smap->buckets[id], new_bucket);
426 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
430 const struct bpf_func_proto bpf_get_stackid_proto = {
431 .func = bpf_get_stackid,
433 .ret_type = RET_INTEGER,
434 .arg1_type = ARG_PTR_TO_CTX,
435 .arg2_type = ARG_CONST_MAP_PTR,
436 .arg3_type = ARG_ANYTHING,
439 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
442 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
443 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
444 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
445 bool user = flags & BPF_F_USER_STACK;
446 struct perf_callchain_entry *trace;
451 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
452 BPF_F_USER_BUILD_ID)))
454 if (kernel && user_build_id)
457 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
459 if (unlikely(size % elem_size))
462 num_elem = size / elem_size;
463 if (sysctl_perf_event_max_stack < num_elem)
466 init_nr = sysctl_perf_event_max_stack - num_elem;
467 trace = get_perf_callchain(regs, init_nr, kernel, user,
468 sysctl_perf_event_max_stack, false, false);
469 if (unlikely(!trace))
472 trace_nr = trace->nr - init_nr;
477 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
478 copy_len = trace_nr * elem_size;
479 ips = trace->ip + skip + init_nr;
480 if (user && user_build_id)
481 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
483 memcpy(buf, ips, copy_len);
486 memset(buf + copy_len, 0, size - copy_len);
492 memset(buf, 0, size);
496 const struct bpf_func_proto bpf_get_stack_proto = {
497 .func = bpf_get_stack,
499 .ret_type = RET_INTEGER,
500 .arg1_type = ARG_PTR_TO_CTX,
501 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
502 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
503 .arg4_type = ARG_ANYTHING,
506 /* Called from eBPF program */
507 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
509 return ERR_PTR(-EOPNOTSUPP);
512 /* Called from syscall */
513 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
515 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
516 struct stack_map_bucket *bucket, *old_bucket;
517 u32 id = *(u32 *)key, trace_len;
519 if (unlikely(id >= smap->n_buckets))
522 bucket = xchg(&smap->buckets[id], NULL);
526 trace_len = bucket->nr * stack_map_data_size(map);
527 memcpy(value, bucket->data, trace_len);
528 memset(value + trace_len, 0, map->value_size - trace_len);
530 old_bucket = xchg(&smap->buckets[id], bucket);
532 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
536 static int stack_map_get_next_key(struct bpf_map *map, void *key,
539 struct bpf_stack_map *smap = container_of(map,
540 struct bpf_stack_map, map);
543 WARN_ON_ONCE(!rcu_read_lock_held());
549 if (id >= smap->n_buckets || !smap->buckets[id])
555 while (id < smap->n_buckets && !smap->buckets[id])
558 if (id >= smap->n_buckets)
561 *(u32 *)next_key = id;
565 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
571 /* Called from syscall or from eBPF program */
572 static int stack_map_delete_elem(struct bpf_map *map, void *key)
574 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
575 struct stack_map_bucket *old_bucket;
576 u32 id = *(u32 *)key;
578 if (unlikely(id >= smap->n_buckets))
581 old_bucket = xchg(&smap->buckets[id], NULL);
583 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
590 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
591 static void stack_map_free(struct bpf_map *map)
593 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
595 /* wait for bpf programs to complete before freeing stack map */
598 bpf_map_area_free(smap->elems);
599 pcpu_freelist_destroy(&smap->freelist);
600 bpf_map_area_free(smap);
601 put_callchain_buffers();
604 const struct bpf_map_ops stack_trace_map_ops = {
605 .map_alloc = stack_map_alloc,
606 .map_free = stack_map_free,
607 .map_get_next_key = stack_map_get_next_key,
608 .map_lookup_elem = stack_map_lookup_elem,
609 .map_update_elem = stack_map_update_elem,
610 .map_delete_elem = stack_map_delete_elem,
611 .map_check_btf = map_check_no_btf,
614 static int __init stack_map_init(void)
617 struct stack_map_irq_work *work;
619 for_each_possible_cpu(cpu) {
620 work = per_cpu_ptr(&up_read_work, cpu);
621 init_irq_work(&work->irq_work, do_up_read);
625 subsys_initcall(stack_map_init);