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);
47 up_read_non_owner(work->sem);
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 u64 elem_size = sizeof(struct stack_map_bucket) +
67 (u64)smap->map.value_size;
70 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
75 err = pcpu_freelist_init(&smap->freelist);
79 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
80 smap->map.max_entries);
84 bpf_map_area_free(smap->elems);
88 /* Called from syscall */
89 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
91 u32 value_size = attr->value_size;
92 struct bpf_stack_map *smap;
96 if (!capable(CAP_SYS_ADMIN))
97 return ERR_PTR(-EPERM);
99 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
100 return ERR_PTR(-EINVAL);
102 /* check sanity of attributes */
103 if (attr->max_entries == 0 || attr->key_size != 4 ||
104 value_size < 8 || value_size % 8)
105 return ERR_PTR(-EINVAL);
107 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
108 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
109 if (value_size % sizeof(struct bpf_stack_build_id) ||
110 value_size / sizeof(struct bpf_stack_build_id)
111 > sysctl_perf_event_max_stack)
112 return ERR_PTR(-EINVAL);
113 } else if (value_size / 8 > sysctl_perf_event_max_stack)
114 return ERR_PTR(-EINVAL);
116 /* hash table size must be power of 2 */
117 n_buckets = roundup_pow_of_two(attr->max_entries);
119 return ERR_PTR(-E2BIG);
121 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
122 if (cost >= U32_MAX - PAGE_SIZE)
123 return ERR_PTR(-E2BIG);
125 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
127 return ERR_PTR(-ENOMEM);
130 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
131 if (cost >= U32_MAX - PAGE_SIZE)
134 bpf_map_init_from_attr(&smap->map, attr);
135 smap->map.value_size = value_size;
136 smap->n_buckets = n_buckets;
137 smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
139 err = bpf_map_precharge_memlock(smap->map.pages);
143 err = get_callchain_buffers(sysctl_perf_event_max_stack);
147 err = prealloc_elems_and_freelist(smap);
154 put_callchain_buffers();
156 bpf_map_area_free(smap);
160 #define BPF_BUILD_ID 3
162 * Parse build id from the note segment. This logic can be shared between
163 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
166 static inline int stack_map_parse_build_id(void *page_addr,
167 unsigned char *build_id,
169 Elf32_Word note_size)
171 Elf32_Word note_offs = 0, new_offs;
173 /* check for overflow */
174 if (note_start < page_addr || note_start + note_size < note_start)
177 /* only supports note that fits in the first page */
178 if (note_start + note_size > page_addr + PAGE_SIZE)
181 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
182 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
184 if (nhdr->n_type == BPF_BUILD_ID &&
185 nhdr->n_namesz == sizeof("GNU") &&
186 nhdr->n_descsz > 0 &&
187 nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
189 note_start + note_offs +
190 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
192 memset(build_id + nhdr->n_descsz, 0,
193 BPF_BUILD_ID_SIZE - nhdr->n_descsz);
196 new_offs = note_offs + sizeof(Elf32_Nhdr) +
197 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
198 if (new_offs <= note_offs) /* overflow */
200 note_offs = new_offs;
205 /* Parse build ID from 32-bit ELF */
206 static int stack_map_get_build_id_32(void *page_addr,
207 unsigned char *build_id)
209 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
213 /* only supports phdr that fits in one page */
215 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
218 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
220 for (i = 0; i < ehdr->e_phnum; ++i)
221 if (phdr[i].p_type == PT_NOTE)
222 return stack_map_parse_build_id(page_addr, build_id,
223 page_addr + phdr[i].p_offset,
228 /* Parse build ID from 64-bit ELF */
229 static int stack_map_get_build_id_64(void *page_addr,
230 unsigned char *build_id)
232 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
236 /* only supports phdr that fits in one page */
238 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
241 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
243 for (i = 0; i < ehdr->e_phnum; ++i)
244 if (phdr[i].p_type == PT_NOTE)
245 return stack_map_parse_build_id(page_addr, build_id,
246 page_addr + phdr[i].p_offset,
251 /* Parse build ID of ELF file mapped to vma */
252 static int stack_map_get_build_id(struct vm_area_struct *vma,
253 unsigned char *build_id)
260 /* only works for page backed storage */
264 page = find_get_page(vma->vm_file->f_mapping, 0);
266 return -EFAULT; /* page not mapped */
269 page_addr = kmap_atomic(page);
270 ehdr = (Elf32_Ehdr *)page_addr;
272 /* compare magic x7f "ELF" */
273 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
276 /* only support executable file and shared object file */
277 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
280 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
281 ret = stack_map_get_build_id_32(page_addr, build_id);
282 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
283 ret = stack_map_get_build_id_64(page_addr, build_id);
285 kunmap_atomic(page_addr);
290 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
291 u64 *ips, u32 trace_nr, bool user)
294 struct vm_area_struct *vma;
295 bool irq_work_busy = false;
296 struct stack_map_irq_work *work = NULL;
298 if (irqs_disabled()) {
299 work = this_cpu_ptr(&up_read_work);
300 if (work->irq_work.flags & IRQ_WORK_BUSY)
301 /* cannot queue more up_read, fallback */
302 irq_work_busy = true;
306 * We cannot do up_read() when the irq is disabled, because of
307 * risk to deadlock with rq_lock. To do build_id lookup when the
308 * irqs are disabled, we need to run up_read() in irq_work. We use
309 * a percpu variable to do the irq_work. If the irq_work is
310 * already used by another lookup, we fall back to report ips.
312 * Same fallback is used for kernel stack (!user) on a stackmap
315 if (!user || !current || !current->mm || irq_work_busy ||
316 down_read_trylock(¤t->mm->mmap_sem) == 0) {
317 /* cannot access current->mm, fall back to ips */
318 for (i = 0; i < trace_nr; i++) {
319 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
320 id_offs[i].ip = ips[i];
321 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
326 for (i = 0; i < trace_nr; i++) {
327 vma = find_vma(current->mm, ips[i]);
328 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
329 /* per entry fall back to ips */
330 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
331 id_offs[i].ip = ips[i];
332 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
335 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
337 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
341 up_read(¤t->mm->mmap_sem);
343 work->sem = ¤t->mm->mmap_sem;
344 irq_work_queue(&work->irq_work);
346 * The irq_work will release the mmap_sem with
347 * up_read_non_owner(). The rwsem_release() is called
348 * here to release the lock from lockdep's perspective.
350 rwsem_release(¤t->mm->mmap_sem.dep_map, 1, _RET_IP_);
354 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
357 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
358 struct perf_callchain_entry *trace;
359 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
360 u32 max_depth = map->value_size / stack_map_data_size(map);
361 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
362 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
363 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
364 u32 hash, id, trace_nr, trace_len;
365 bool user = flags & BPF_F_USER_STACK;
370 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
371 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
374 trace = get_perf_callchain(regs, init_nr, kernel, user,
375 sysctl_perf_event_max_stack, false, false);
377 if (unlikely(!trace))
378 /* couldn't fetch the stack trace */
381 /* get_perf_callchain() guarantees that trace->nr >= init_nr
382 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
384 trace_nr = trace->nr - init_nr;
386 if (trace_nr <= skip)
387 /* skipping more than usable stack trace */
391 trace_len = trace_nr * sizeof(u64);
392 ips = trace->ip + skip + init_nr;
393 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
394 id = hash & (smap->n_buckets - 1);
395 bucket = READ_ONCE(smap->buckets[id]);
397 hash_matches = bucket && bucket->hash == hash;
399 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
402 if (stack_map_use_build_id(map)) {
403 /* for build_id+offset, pop a bucket before slow cmp */
404 new_bucket = (struct stack_map_bucket *)
405 pcpu_freelist_pop(&smap->freelist);
406 if (unlikely(!new_bucket))
408 new_bucket->nr = trace_nr;
409 stack_map_get_build_id_offset(
410 (struct bpf_stack_build_id *)new_bucket->data,
411 ips, trace_nr, user);
412 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
413 if (hash_matches && bucket->nr == trace_nr &&
414 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
415 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
418 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
419 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
423 if (hash_matches && bucket->nr == trace_nr &&
424 memcmp(bucket->data, ips, trace_len) == 0)
426 if (bucket && !(flags & BPF_F_REUSE_STACKID))
429 new_bucket = (struct stack_map_bucket *)
430 pcpu_freelist_pop(&smap->freelist);
431 if (unlikely(!new_bucket))
433 memcpy(new_bucket->data, ips, trace_len);
436 new_bucket->hash = hash;
437 new_bucket->nr = trace_nr;
439 old_bucket = xchg(&smap->buckets[id], new_bucket);
441 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
445 const struct bpf_func_proto bpf_get_stackid_proto = {
446 .func = bpf_get_stackid,
448 .ret_type = RET_INTEGER,
449 .arg1_type = ARG_PTR_TO_CTX,
450 .arg2_type = ARG_CONST_MAP_PTR,
451 .arg3_type = ARG_ANYTHING,
454 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
457 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
458 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
459 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
460 bool user = flags & BPF_F_USER_STACK;
461 struct perf_callchain_entry *trace;
466 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
467 BPF_F_USER_BUILD_ID)))
469 if (kernel && user_build_id)
472 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
474 if (unlikely(size % elem_size))
477 num_elem = size / elem_size;
478 if (sysctl_perf_event_max_stack < num_elem)
481 init_nr = sysctl_perf_event_max_stack - num_elem;
482 trace = get_perf_callchain(regs, init_nr, kernel, user,
483 sysctl_perf_event_max_stack, false, false);
484 if (unlikely(!trace))
487 trace_nr = trace->nr - init_nr;
492 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
493 copy_len = trace_nr * elem_size;
494 ips = trace->ip + skip + init_nr;
495 if (user && user_build_id)
496 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
498 memcpy(buf, ips, copy_len);
501 memset(buf + copy_len, 0, size - copy_len);
507 memset(buf, 0, size);
511 const struct bpf_func_proto bpf_get_stack_proto = {
512 .func = bpf_get_stack,
514 .ret_type = RET_INTEGER,
515 .arg1_type = ARG_PTR_TO_CTX,
516 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
517 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
518 .arg4_type = ARG_ANYTHING,
521 /* Called from eBPF program */
522 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
527 /* Called from syscall */
528 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
530 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
531 struct stack_map_bucket *bucket, *old_bucket;
532 u32 id = *(u32 *)key, trace_len;
534 if (unlikely(id >= smap->n_buckets))
537 bucket = xchg(&smap->buckets[id], NULL);
541 trace_len = bucket->nr * stack_map_data_size(map);
542 memcpy(value, bucket->data, trace_len);
543 memset(value + trace_len, 0, map->value_size - trace_len);
545 old_bucket = xchg(&smap->buckets[id], bucket);
547 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
551 static int stack_map_get_next_key(struct bpf_map *map, void *key,
554 struct bpf_stack_map *smap = container_of(map,
555 struct bpf_stack_map, map);
558 WARN_ON_ONCE(!rcu_read_lock_held());
564 if (id >= smap->n_buckets || !smap->buckets[id])
570 while (id < smap->n_buckets && !smap->buckets[id])
573 if (id >= smap->n_buckets)
576 *(u32 *)next_key = id;
580 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
586 /* Called from syscall or from eBPF program */
587 static int stack_map_delete_elem(struct bpf_map *map, void *key)
589 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
590 struct stack_map_bucket *old_bucket;
591 u32 id = *(u32 *)key;
593 if (unlikely(id >= smap->n_buckets))
596 old_bucket = xchg(&smap->buckets[id], NULL);
598 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
605 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
606 static void stack_map_free(struct bpf_map *map)
608 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
610 /* wait for bpf programs to complete before freeing stack map */
613 bpf_map_area_free(smap->elems);
614 pcpu_freelist_destroy(&smap->freelist);
615 bpf_map_area_free(smap);
616 put_callchain_buffers();
619 const struct bpf_map_ops stack_map_ops = {
620 .map_alloc = stack_map_alloc,
621 .map_free = stack_map_free,
622 .map_get_next_key = stack_map_get_next_key,
623 .map_lookup_elem = stack_map_lookup_elem,
624 .map_update_elem = stack_map_update_elem,
625 .map_delete_elem = stack_map_delete_elem,
626 .map_check_btf = map_check_no_btf,
629 static int __init stack_map_init(void)
632 struct stack_map_irq_work *work;
634 for_each_possible_cpu(cpu) {
635 work = per_cpu_ptr(&up_read_work, cpu);
636 init_irq_work(&work->irq_work, do_up_read);
640 subsys_initcall(stack_map_init);