Loading...
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[];
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 (!bpf_capable())
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 if (!n_buckets)
100 return ERR_PTR(-E2BIG);
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 void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
128 u64 *ips, u32 trace_nr, bool user)
129{
130 int i;
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;
135
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
138 * build_id.
139 */
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);
147 }
148 return;
149 }
150
151 for (i = 0; i < trace_nr; i++) {
152 if (range_in_vma(prev_vma, ips[i], ips[i])) {
153 vma = prev_vma;
154 memcpy(id_offs[i].build_id, prev_build_id,
155 BUILD_ID_SIZE_MAX);
156 goto build_id_valid;
157 }
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);
164 continue;
165 }
166build_id_valid:
167 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
168 - vma->vm_start;
169 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
170 prev_vma = vma;
171 prev_build_id = id_offs[i].build_id;
172 }
173 bpf_mmap_unlock_mm(work, current->mm);
174}
175
176static struct perf_callchain_entry *
177get_callchain_entry_for_task(struct task_struct *task, u32 max_depth)
178{
179#ifdef CONFIG_STACKTRACE
180 struct perf_callchain_entry *entry;
181 int rctx;
182
183 entry = get_callchain_entry(&rctx);
184
185 if (!entry)
186 return NULL;
187
188 entry->nr = stack_trace_save_tsk(task, (unsigned long *)entry->ip,
189 max_depth, 0);
190
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.
194 */
195 if (__BITS_PER_LONG != 64) {
196 unsigned long *from = (unsigned long *) entry->ip;
197 u64 *to = entry->ip;
198 int i;
199
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]);
203 }
204
205 put_callchain_entry(rctx);
206
207 return entry;
208#else /* CONFIG_STACKTRACE */
209 return NULL;
210#endif
211}
212
213static long __bpf_get_stackid(struct bpf_map *map,
214 struct perf_callchain_entry *trace, u64 flags)
215{
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;
221 u64 *ips;
222 bool hash_matches;
223
224 if (trace->nr <= skip)
225 /* skipping more than usable stack trace */
226 return -EFAULT;
227
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]);
234
235 hash_matches = bucket && bucket->hash == hash;
236 /* fast cmp */
237 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
238 return id;
239
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))
245 return -ENOMEM;
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);
254 return id;
255 }
256 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
257 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
258 return -EEXIST;
259 }
260 } else {
261 if (hash_matches && bucket->nr == trace_nr &&
262 memcmp(bucket->data, ips, trace_len) == 0)
263 return id;
264 if (bucket && !(flags & BPF_F_REUSE_STACKID))
265 return -EEXIST;
266
267 new_bucket = (struct stack_map_bucket *)
268 pcpu_freelist_pop(&smap->freelist);
269 if (unlikely(!new_bucket))
270 return -ENOMEM;
271 memcpy(new_bucket->data, ips, trace_len);
272 }
273
274 new_bucket->hash = hash;
275 new_bucket->nr = trace_nr;
276
277 old_bucket = xchg(&smap->buckets[id], new_bucket);
278 if (old_bucket)
279 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
280 return id;
281}
282
283BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
284 u64, flags)
285{
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;
290 bool kernel = !user;
291
292 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
293 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
294 return -EINVAL;
295
296 max_depth += skip;
297 if (max_depth > sysctl_perf_event_max_stack)
298 max_depth = sysctl_perf_event_max_stack;
299
300 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
301 false, false);
302
303 if (unlikely(!trace))
304 /* couldn't fetch the stack trace */
305 return -EFAULT;
306
307 return __bpf_get_stackid(map, trace, flags);
308}
309
310const struct bpf_func_proto bpf_get_stackid_proto = {
311 .func = bpf_get_stackid,
312 .gpl_only = true,
313 .ret_type = RET_INTEGER,
314 .arg1_type = ARG_PTR_TO_CTX,
315 .arg2_type = ARG_CONST_MAP_PTR,
316 .arg3_type = ARG_ANYTHING,
317};
318
319static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
320{
321 __u64 nr_kernel = 0;
322
323 while (nr_kernel < trace->nr) {
324 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
325 break;
326 nr_kernel++;
327 }
328 return nr_kernel;
329}
330
331BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
332 struct bpf_map *, map, u64, flags)
333{
334 struct perf_event *event = ctx->event;
335 struct perf_callchain_entry *trace;
336 bool kernel, user;
337 __u64 nr_kernel;
338 int ret;
339
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);
344
345 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
346 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
347 return -EINVAL;
348
349 user = flags & BPF_F_USER_STACK;
350 kernel = !user;
351
352 trace = ctx->data->callchain;
353 if (unlikely(!trace))
354 return -EFAULT;
355
356 nr_kernel = count_kernel_ip(trace);
357
358 if (kernel) {
359 __u64 nr = trace->nr;
360
361 trace->nr = nr_kernel;
362 ret = __bpf_get_stackid(map, trace, flags);
363
364 /* restore nr */
365 trace->nr = nr;
366 } else { /* user */
367 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
368
369 skip += nr_kernel;
370 if (skip > BPF_F_SKIP_FIELD_MASK)
371 return -EFAULT;
372
373 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
374 ret = __bpf_get_stackid(map, trace, flags);
375 }
376 return ret;
377}
378
379const struct bpf_func_proto bpf_get_stackid_proto_pe = {
380 .func = bpf_get_stackid_pe,
381 .gpl_only = false,
382 .ret_type = RET_INTEGER,
383 .arg1_type = ARG_PTR_TO_CTX,
384 .arg2_type = ARG_CONST_MAP_PTR,
385 .arg3_type = ARG_ANYTHING,
386};
387
388static 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)
391{
392 u32 trace_nr, copy_len, elem_size, num_elem, max_depth;
393 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
394 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
395 bool user = flags & BPF_F_USER_STACK;
396 struct perf_callchain_entry *trace;
397 bool kernel = !user;
398 int err = -EINVAL;
399 u64 *ips;
400
401 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
402 BPF_F_USER_BUILD_ID)))
403 goto clear;
404 if (kernel && user_build_id)
405 goto clear;
406
407 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
408 : sizeof(u64);
409 if (unlikely(size % elem_size))
410 goto clear;
411
412 /* cannot get valid user stack for task without user_mode regs */
413 if (task && user && !user_mode(regs))
414 goto err_fault;
415
416 num_elem = size / elem_size;
417 max_depth = num_elem + skip;
418 if (sysctl_perf_event_max_stack < max_depth)
419 max_depth = sysctl_perf_event_max_stack;
420
421 if (trace_in)
422 trace = trace_in;
423 else if (kernel && task)
424 trace = get_callchain_entry_for_task(task, max_depth);
425 else
426 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
427 false, false);
428 if (unlikely(!trace))
429 goto err_fault;
430
431 if (trace->nr < skip)
432 goto err_fault;
433
434 trace_nr = trace->nr - skip;
435 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
436 copy_len = trace_nr * elem_size;
437
438 ips = trace->ip + skip;
439 if (user && user_build_id)
440 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
441 else
442 memcpy(buf, ips, copy_len);
443
444 if (size > copy_len)
445 memset(buf + copy_len, 0, size - copy_len);
446 return copy_len;
447
448err_fault:
449 err = -EFAULT;
450clear:
451 memset(buf, 0, size);
452 return err;
453}
454
455BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
456 u64, flags)
457{
458 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
459}
460
461const struct bpf_func_proto bpf_get_stack_proto = {
462 .func = bpf_get_stack,
463 .gpl_only = true,
464 .ret_type = RET_INTEGER,
465 .arg1_type = ARG_PTR_TO_CTX,
466 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
467 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
468 .arg4_type = ARG_ANYTHING,
469};
470
471BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
472 u32, size, u64, flags)
473{
474 struct pt_regs *regs;
475 long res = -EINVAL;
476
477 if (!try_get_task_stack(task))
478 return -EFAULT;
479
480 regs = task_pt_regs(task);
481 if (regs)
482 res = __bpf_get_stack(regs, task, NULL, buf, size, flags);
483 put_task_stack(task);
484
485 return res;
486}
487
488const struct bpf_func_proto bpf_get_task_stack_proto = {
489 .func = bpf_get_task_stack,
490 .gpl_only = false,
491 .ret_type = RET_INTEGER,
492 .arg1_type = ARG_PTR_TO_BTF_ID,
493 .arg1_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
494 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
495 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
496 .arg4_type = ARG_ANYTHING,
497};
498
499BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
500 void *, buf, u32, size, u64, flags)
501{
502 struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
503 struct perf_event *event = ctx->event;
504 struct perf_callchain_entry *trace;
505 bool kernel, user;
506 int err = -EINVAL;
507 __u64 nr_kernel;
508
509 if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
510 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
511
512 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
513 BPF_F_USER_BUILD_ID)))
514 goto clear;
515
516 user = flags & BPF_F_USER_STACK;
517 kernel = !user;
518
519 err = -EFAULT;
520 trace = ctx->data->callchain;
521 if (unlikely(!trace))
522 goto clear;
523
524 nr_kernel = count_kernel_ip(trace);
525
526 if (kernel) {
527 __u64 nr = trace->nr;
528
529 trace->nr = nr_kernel;
530 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
531
532 /* restore nr */
533 trace->nr = nr;
534 } else { /* user */
535 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
536
537 skip += nr_kernel;
538 if (skip > BPF_F_SKIP_FIELD_MASK)
539 goto clear;
540
541 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
542 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
543 }
544 return err;
545
546clear:
547 memset(buf, 0, size);
548 return err;
549
550}
551
552const struct bpf_func_proto bpf_get_stack_proto_pe = {
553 .func = bpf_get_stack_pe,
554 .gpl_only = true,
555 .ret_type = RET_INTEGER,
556 .arg1_type = ARG_PTR_TO_CTX,
557 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
558 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
559 .arg4_type = ARG_ANYTHING,
560};
561
562/* Called from eBPF program */
563static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
564{
565 return ERR_PTR(-EOPNOTSUPP);
566}
567
568/* Called from syscall */
569int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
570{
571 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
572 struct stack_map_bucket *bucket, *old_bucket;
573 u32 id = *(u32 *)key, trace_len;
574
575 if (unlikely(id >= smap->n_buckets))
576 return -ENOENT;
577
578 bucket = xchg(&smap->buckets[id], NULL);
579 if (!bucket)
580 return -ENOENT;
581
582 trace_len = bucket->nr * stack_map_data_size(map);
583 memcpy(value, bucket->data, trace_len);
584 memset(value + trace_len, 0, map->value_size - trace_len);
585
586 old_bucket = xchg(&smap->buckets[id], bucket);
587 if (old_bucket)
588 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
589 return 0;
590}
591
592static int stack_map_get_next_key(struct bpf_map *map, void *key,
593 void *next_key)
594{
595 struct bpf_stack_map *smap = container_of(map,
596 struct bpf_stack_map, map);
597 u32 id;
598
599 WARN_ON_ONCE(!rcu_read_lock_held());
600
601 if (!key) {
602 id = 0;
603 } else {
604 id = *(u32 *)key;
605 if (id >= smap->n_buckets || !smap->buckets[id])
606 id = 0;
607 else
608 id++;
609 }
610
611 while (id < smap->n_buckets && !smap->buckets[id])
612 id++;
613
614 if (id >= smap->n_buckets)
615 return -ENOENT;
616
617 *(u32 *)next_key = id;
618 return 0;
619}
620
621static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
622 u64 map_flags)
623{
624 return -EINVAL;
625}
626
627/* Called from syscall or from eBPF program */
628static int stack_map_delete_elem(struct bpf_map *map, void *key)
629{
630 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
631 struct stack_map_bucket *old_bucket;
632 u32 id = *(u32 *)key;
633
634 if (unlikely(id >= smap->n_buckets))
635 return -E2BIG;
636
637 old_bucket = xchg(&smap->buckets[id], NULL);
638 if (old_bucket) {
639 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
640 return 0;
641 } else {
642 return -ENOENT;
643 }
644}
645
646/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
647static void stack_map_free(struct bpf_map *map)
648{
649 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
650
651 bpf_map_area_free(smap->elems);
652 pcpu_freelist_destroy(&smap->freelist);
653 bpf_map_area_free(smap);
654 put_callchain_buffers();
655}
656
657BTF_ID_LIST_SINGLE(stack_trace_map_btf_ids, struct, bpf_stack_map)
658const struct bpf_map_ops stack_trace_map_ops = {
659 .map_meta_equal = bpf_map_meta_equal,
660 .map_alloc = stack_map_alloc,
661 .map_free = stack_map_free,
662 .map_get_next_key = stack_map_get_next_key,
663 .map_lookup_elem = stack_map_lookup_elem,
664 .map_update_elem = stack_map_update_elem,
665 .map_delete_elem = stack_map_delete_elem,
666 .map_check_btf = map_check_no_btf,
667 .map_btf_id = &stack_trace_map_btf_ids[0],
668};
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};