1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Facebook
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"
15 #define STACK_CREATE_FLAG_MASK \
16 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
19 struct stack_map_bucket {
20 struct pcpu_freelist_node fnode;
26 struct bpf_stack_map {
29 struct pcpu_freelist freelist;
31 struct stack_map_bucket *buckets[] __counted_by(n_buckets);
34 static inline bool stack_map_use_build_id(struct bpf_map *map)
36 return (map->map_flags & BPF_F_STACK_BUILD_ID);
39 static inline int stack_map_data_size(struct bpf_map *map)
41 return stack_map_use_build_id(map) ?
42 sizeof(struct bpf_stack_build_id) : sizeof(u64);
45 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
47 u64 elem_size = sizeof(struct stack_map_bucket) +
48 (u64)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 (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
78 return ERR_PTR(-EINVAL);
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);
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);
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
97 if (attr->max_entries > 1UL << 31)
98 return ERR_PTR(-E2BIG);
100 n_buckets = roundup_pow_of_two(attr->max_entries);
102 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
103 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
105 return ERR_PTR(-ENOMEM);
107 bpf_map_init_from_attr(&smap->map, attr);
108 smap->n_buckets = n_buckets;
110 err = get_callchain_buffers(sysctl_perf_event_max_stack);
114 err = prealloc_elems_and_freelist(smap);
121 put_callchain_buffers();
123 bpf_map_area_free(smap);
127 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
128 u64 *ips, u32 trace_nr, bool user)
131 struct mmap_unlock_irq_work *work = NULL;
132 bool irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
133 struct vm_area_struct *vma, *prev_vma = NULL;
134 const char *prev_build_id;
136 /* If the irq_work is in use, fall back to report ips. Same
137 * fallback is used for kernel stack (!user) on a stackmap with
140 if (!user || !current || !current->mm || irq_work_busy ||
141 !mmap_read_trylock(current->mm)) {
142 /* cannot access current->mm, fall back to ips */
143 for (i = 0; i < trace_nr; i++) {
144 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
145 id_offs[i].ip = ips[i];
146 memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
151 for (i = 0; i < trace_nr; i++) {
152 if (range_in_vma(prev_vma, ips[i], ips[i])) {
154 memcpy(id_offs[i].build_id, prev_build_id,
158 vma = find_vma(current->mm, ips[i]);
159 if (!vma || build_id_parse(vma, id_offs[i].build_id, NULL)) {
160 /* per entry fall back to ips */
161 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
162 id_offs[i].ip = ips[i];
163 memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
167 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
169 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
171 prev_build_id = id_offs[i].build_id;
173 bpf_mmap_unlock_mm(work, current->mm);
176 static struct perf_callchain_entry *
177 get_callchain_entry_for_task(struct task_struct *task, u32 max_depth)
179 #ifdef CONFIG_STACKTRACE
180 struct perf_callchain_entry *entry;
183 entry = get_callchain_entry(&rctx);
188 entry->nr = stack_trace_save_tsk(task, (unsigned long *)entry->ip,
191 /* stack_trace_save_tsk() works on unsigned long array, while
192 * perf_callchain_entry uses u64 array. For 32-bit systems, it is
193 * necessary to fix this mismatch.
195 if (__BITS_PER_LONG != 64) {
196 unsigned long *from = (unsigned long *) entry->ip;
200 /* copy data from the end to avoid using extra buffer */
201 for (i = entry->nr - 1; i >= 0; i--)
202 to[i] = (u64)(from[i]);
205 put_callchain_entry(rctx);
208 #else /* CONFIG_STACKTRACE */
213 static long __bpf_get_stackid(struct bpf_map *map,
214 struct perf_callchain_entry *trace, u64 flags)
216 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
217 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
218 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
219 u32 hash, id, trace_nr, trace_len;
220 bool user = flags & BPF_F_USER_STACK;
224 if (trace->nr <= skip)
225 /* skipping more than usable stack trace */
228 trace_nr = trace->nr - skip;
229 trace_len = trace_nr * sizeof(u64);
230 ips = trace->ip + skip;
231 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
232 id = hash & (smap->n_buckets - 1);
233 bucket = READ_ONCE(smap->buckets[id]);
235 hash_matches = bucket && bucket->hash == hash;
237 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
240 if (stack_map_use_build_id(map)) {
241 /* for build_id+offset, pop a bucket before slow cmp */
242 new_bucket = (struct stack_map_bucket *)
243 pcpu_freelist_pop(&smap->freelist);
244 if (unlikely(!new_bucket))
246 new_bucket->nr = trace_nr;
247 stack_map_get_build_id_offset(
248 (struct bpf_stack_build_id *)new_bucket->data,
249 ips, trace_nr, user);
250 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
251 if (hash_matches && bucket->nr == trace_nr &&
252 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
253 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
256 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
257 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
261 if (hash_matches && bucket->nr == trace_nr &&
262 memcmp(bucket->data, ips, trace_len) == 0)
264 if (bucket && !(flags & BPF_F_REUSE_STACKID))
267 new_bucket = (struct stack_map_bucket *)
268 pcpu_freelist_pop(&smap->freelist);
269 if (unlikely(!new_bucket))
271 memcpy(new_bucket->data, ips, trace_len);
274 new_bucket->hash = hash;
275 new_bucket->nr = trace_nr;
277 old_bucket = xchg(&smap->buckets[id], new_bucket);
279 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
283 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
286 u32 max_depth = map->value_size / stack_map_data_size(map);
287 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
288 bool user = flags & BPF_F_USER_STACK;
289 struct perf_callchain_entry *trace;
292 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
293 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
297 if (max_depth > sysctl_perf_event_max_stack)
298 max_depth = sysctl_perf_event_max_stack;
300 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
303 if (unlikely(!trace))
304 /* couldn't fetch the stack trace */
307 return __bpf_get_stackid(map, trace, flags);
310 const struct bpf_func_proto bpf_get_stackid_proto = {
311 .func = bpf_get_stackid,
313 .ret_type = RET_INTEGER,
314 .arg1_type = ARG_PTR_TO_CTX,
315 .arg2_type = ARG_CONST_MAP_PTR,
316 .arg3_type = ARG_ANYTHING,
319 static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
323 while (nr_kernel < trace->nr) {
324 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
331 BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
332 struct bpf_map *, map, u64, flags)
334 struct perf_event *event = ctx->event;
335 struct perf_callchain_entry *trace;
340 /* perf_sample_data doesn't have callchain, use bpf_get_stackid */
341 if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
342 return bpf_get_stackid((unsigned long)(ctx->regs),
343 (unsigned long) map, flags, 0, 0);
345 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
346 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
349 user = flags & BPF_F_USER_STACK;
352 trace = ctx->data->callchain;
353 if (unlikely(!trace))
356 nr_kernel = count_kernel_ip(trace);
359 __u64 nr = trace->nr;
361 trace->nr = nr_kernel;
362 ret = __bpf_get_stackid(map, trace, flags);
367 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
370 if (skip > BPF_F_SKIP_FIELD_MASK)
373 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
374 ret = __bpf_get_stackid(map, trace, flags);
379 const struct bpf_func_proto bpf_get_stackid_proto_pe = {
380 .func = bpf_get_stackid_pe,
382 .ret_type = RET_INTEGER,
383 .arg1_type = ARG_PTR_TO_CTX,
384 .arg2_type = ARG_CONST_MAP_PTR,
385 .arg3_type = ARG_ANYTHING,
388 static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
389 struct perf_callchain_entry *trace_in,
390 void *buf, u32 size, u64 flags)
392 u32 trace_nr, copy_len, elem_size, num_elem, max_depth;
393 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
394 bool crosstask = task && task != current;
395 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
396 bool user = flags & BPF_F_USER_STACK;
397 struct perf_callchain_entry *trace;
402 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
403 BPF_F_USER_BUILD_ID)))
405 if (kernel && user_build_id)
408 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
410 if (unlikely(size % elem_size))
413 /* cannot get valid user stack for task without user_mode regs */
414 if (task && user && !user_mode(regs))
417 /* get_perf_callchain does not support crosstask user stack walking
418 * but returns an empty stack instead of NULL.
420 if (crosstask && user) {
425 num_elem = size / elem_size;
426 max_depth = num_elem + skip;
427 if (sysctl_perf_event_max_stack < max_depth)
428 max_depth = sysctl_perf_event_max_stack;
432 else if (kernel && task)
433 trace = get_callchain_entry_for_task(task, max_depth);
435 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
437 if (unlikely(!trace))
440 if (trace->nr < skip)
443 trace_nr = trace->nr - skip;
444 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
445 copy_len = trace_nr * elem_size;
447 ips = trace->ip + skip;
448 if (user && user_build_id)
449 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
451 memcpy(buf, ips, copy_len);
454 memset(buf + copy_len, 0, size - copy_len);
460 memset(buf, 0, size);
464 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
467 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
470 const struct bpf_func_proto bpf_get_stack_proto = {
471 .func = bpf_get_stack,
473 .ret_type = RET_INTEGER,
474 .arg1_type = ARG_PTR_TO_CTX,
475 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
476 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
477 .arg4_type = ARG_ANYTHING,
480 BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
481 u32, size, u64, flags)
483 struct pt_regs *regs;
486 if (!try_get_task_stack(task))
489 regs = task_pt_regs(task);
491 res = __bpf_get_stack(regs, task, NULL, buf, size, flags);
492 put_task_stack(task);
497 const struct bpf_func_proto bpf_get_task_stack_proto = {
498 .func = bpf_get_task_stack,
500 .ret_type = RET_INTEGER,
501 .arg1_type = ARG_PTR_TO_BTF_ID,
502 .arg1_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
503 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
504 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
505 .arg4_type = ARG_ANYTHING,
508 BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
509 void *, buf, u32, size, u64, flags)
511 struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
512 struct perf_event *event = ctx->event;
513 struct perf_callchain_entry *trace;
518 if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
519 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
521 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
522 BPF_F_USER_BUILD_ID)))
525 user = flags & BPF_F_USER_STACK;
529 trace = ctx->data->callchain;
530 if (unlikely(!trace))
533 nr_kernel = count_kernel_ip(trace);
536 __u64 nr = trace->nr;
538 trace->nr = nr_kernel;
539 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
544 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
547 if (skip > BPF_F_SKIP_FIELD_MASK)
550 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
551 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
556 memset(buf, 0, size);
561 const struct bpf_func_proto bpf_get_stack_proto_pe = {
562 .func = bpf_get_stack_pe,
564 .ret_type = RET_INTEGER,
565 .arg1_type = ARG_PTR_TO_CTX,
566 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
567 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
568 .arg4_type = ARG_ANYTHING,
571 /* Called from eBPF program */
572 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
574 return ERR_PTR(-EOPNOTSUPP);
577 /* Called from syscall */
578 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
580 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
581 struct stack_map_bucket *bucket, *old_bucket;
582 u32 id = *(u32 *)key, trace_len;
584 if (unlikely(id >= smap->n_buckets))
587 bucket = xchg(&smap->buckets[id], NULL);
591 trace_len = bucket->nr * stack_map_data_size(map);
592 memcpy(value, bucket->data, trace_len);
593 memset(value + trace_len, 0, map->value_size - trace_len);
595 old_bucket = xchg(&smap->buckets[id], bucket);
597 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
601 static int stack_map_get_next_key(struct bpf_map *map, void *key,
604 struct bpf_stack_map *smap = container_of(map,
605 struct bpf_stack_map, map);
608 WARN_ON_ONCE(!rcu_read_lock_held());
614 if (id >= smap->n_buckets || !smap->buckets[id])
620 while (id < smap->n_buckets && !smap->buckets[id])
623 if (id >= smap->n_buckets)
626 *(u32 *)next_key = id;
630 static long stack_map_update_elem(struct bpf_map *map, void *key, void *value,
636 /* Called from syscall or from eBPF program */
637 static long stack_map_delete_elem(struct bpf_map *map, void *key)
639 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
640 struct stack_map_bucket *old_bucket;
641 u32 id = *(u32 *)key;
643 if (unlikely(id >= smap->n_buckets))
646 old_bucket = xchg(&smap->buckets[id], NULL);
648 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
655 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
656 static void stack_map_free(struct bpf_map *map)
658 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
660 bpf_map_area_free(smap->elems);
661 pcpu_freelist_destroy(&smap->freelist);
662 bpf_map_area_free(smap);
663 put_callchain_buffers();
666 static u64 stack_map_mem_usage(const struct bpf_map *map)
668 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
669 u64 value_size = map->value_size;
670 u64 n_buckets = smap->n_buckets;
671 u64 enties = map->max_entries;
672 u64 usage = sizeof(*smap);
674 usage += n_buckets * sizeof(struct stack_map_bucket *);
675 usage += enties * (sizeof(struct stack_map_bucket) + value_size);
679 BTF_ID_LIST_SINGLE(stack_trace_map_btf_ids, struct, bpf_stack_map)
680 const struct bpf_map_ops stack_trace_map_ops = {
681 .map_meta_equal = bpf_map_meta_equal,
682 .map_alloc = stack_map_alloc,
683 .map_free = stack_map_free,
684 .map_get_next_key = stack_map_get_next_key,
685 .map_lookup_elem = stack_map_lookup_elem,
686 .map_update_elem = stack_map_update_elem,
687 .map_delete_elem = stack_map_delete_elem,
688 .map_check_btf = map_check_no_btf,
689 .map_mem_usage = stack_map_mem_usage,
690 .map_btf_id = &stack_trace_map_btf_ids[0],