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[] __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 */
95 n_buckets = roundup_pow_of_two(attr->max_entries);
96 if (!n_buckets)
97 return ERR_PTR(-E2BIG);
98
99 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
100 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
101 if (!smap)
102 return ERR_PTR(-ENOMEM);
103
104 bpf_map_init_from_attr(&smap->map, attr);
105 smap->n_buckets = n_buckets;
106
107 err = get_callchain_buffers(sysctl_perf_event_max_stack);
108 if (err)
109 goto free_smap;
110
111 err = prealloc_elems_and_freelist(smap);
112 if (err)
113 goto put_buffers;
114
115 return &smap->map;
116
117put_buffers:
118 put_callchain_buffers();
119free_smap:
120 bpf_map_area_free(smap);
121 return ERR_PTR(err);
122}
123
124static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
125 u64 *ips, u32 trace_nr, bool user)
126{
127 int i;
128 struct mmap_unlock_irq_work *work = NULL;
129 bool irq_work_busy = bpf_mmap_unlock_get_irq_work(&work);
130 struct vm_area_struct *vma, *prev_vma = NULL;
131 const char *prev_build_id;
132
133 /* If the irq_work is in use, fall back to report ips. Same
134 * fallback is used for kernel stack (!user) on a stackmap with
135 * build_id.
136 */
137 if (!user || !current || !current->mm || irq_work_busy ||
138 !mmap_read_trylock(current->mm)) {
139 /* cannot access current->mm, fall back to ips */
140 for (i = 0; i < trace_nr; i++) {
141 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
142 id_offs[i].ip = ips[i];
143 memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
144 }
145 return;
146 }
147
148 for (i = 0; i < trace_nr; i++) {
149 if (range_in_vma(prev_vma, ips[i], ips[i])) {
150 vma = prev_vma;
151 memcpy(id_offs[i].build_id, prev_build_id,
152 BUILD_ID_SIZE_MAX);
153 goto build_id_valid;
154 }
155 vma = find_vma(current->mm, ips[i]);
156 if (!vma || build_id_parse(vma, id_offs[i].build_id, NULL)) {
157 /* per entry fall back to ips */
158 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
159 id_offs[i].ip = ips[i];
160 memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
161 continue;
162 }
163build_id_valid:
164 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
165 - vma->vm_start;
166 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
167 prev_vma = vma;
168 prev_build_id = id_offs[i].build_id;
169 }
170 bpf_mmap_unlock_mm(work, current->mm);
171}
172
173static struct perf_callchain_entry *
174get_callchain_entry_for_task(struct task_struct *task, u32 max_depth)
175{
176#ifdef CONFIG_STACKTRACE
177 struct perf_callchain_entry *entry;
178 int rctx;
179
180 entry = get_callchain_entry(&rctx);
181
182 if (!entry)
183 return NULL;
184
185 entry->nr = stack_trace_save_tsk(task, (unsigned long *)entry->ip,
186 max_depth, 0);
187
188 /* stack_trace_save_tsk() works on unsigned long array, while
189 * perf_callchain_entry uses u64 array. For 32-bit systems, it is
190 * necessary to fix this mismatch.
191 */
192 if (__BITS_PER_LONG != 64) {
193 unsigned long *from = (unsigned long *) entry->ip;
194 u64 *to = entry->ip;
195 int i;
196
197 /* copy data from the end to avoid using extra buffer */
198 for (i = entry->nr - 1; i >= 0; i--)
199 to[i] = (u64)(from[i]);
200 }
201
202 put_callchain_entry(rctx);
203
204 return entry;
205#else /* CONFIG_STACKTRACE */
206 return NULL;
207#endif
208}
209
210static long __bpf_get_stackid(struct bpf_map *map,
211 struct perf_callchain_entry *trace, u64 flags)
212{
213 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
214 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
215 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
216 u32 hash, id, trace_nr, trace_len;
217 bool user = flags & BPF_F_USER_STACK;
218 u64 *ips;
219 bool hash_matches;
220
221 if (trace->nr <= skip)
222 /* skipping more than usable stack trace */
223 return -EFAULT;
224
225 trace_nr = trace->nr - skip;
226 trace_len = trace_nr * sizeof(u64);
227 ips = trace->ip + skip;
228 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
229 id = hash & (smap->n_buckets - 1);
230 bucket = READ_ONCE(smap->buckets[id]);
231
232 hash_matches = bucket && bucket->hash == hash;
233 /* fast cmp */
234 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
235 return id;
236
237 if (stack_map_use_build_id(map)) {
238 /* for build_id+offset, pop a bucket before slow cmp */
239 new_bucket = (struct stack_map_bucket *)
240 pcpu_freelist_pop(&smap->freelist);
241 if (unlikely(!new_bucket))
242 return -ENOMEM;
243 new_bucket->nr = trace_nr;
244 stack_map_get_build_id_offset(
245 (struct bpf_stack_build_id *)new_bucket->data,
246 ips, trace_nr, user);
247 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
248 if (hash_matches && bucket->nr == trace_nr &&
249 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
250 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
251 return id;
252 }
253 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
254 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
255 return -EEXIST;
256 }
257 } else {
258 if (hash_matches && bucket->nr == trace_nr &&
259 memcmp(bucket->data, ips, trace_len) == 0)
260 return id;
261 if (bucket && !(flags & BPF_F_REUSE_STACKID))
262 return -EEXIST;
263
264 new_bucket = (struct stack_map_bucket *)
265 pcpu_freelist_pop(&smap->freelist);
266 if (unlikely(!new_bucket))
267 return -ENOMEM;
268 memcpy(new_bucket->data, ips, trace_len);
269 }
270
271 new_bucket->hash = hash;
272 new_bucket->nr = trace_nr;
273
274 old_bucket = xchg(&smap->buckets[id], new_bucket);
275 if (old_bucket)
276 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
277 return id;
278}
279
280BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
281 u64, flags)
282{
283 u32 max_depth = map->value_size / stack_map_data_size(map);
284 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
285 bool user = flags & BPF_F_USER_STACK;
286 struct perf_callchain_entry *trace;
287 bool kernel = !user;
288
289 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
290 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
291 return -EINVAL;
292
293 max_depth += skip;
294 if (max_depth > sysctl_perf_event_max_stack)
295 max_depth = sysctl_perf_event_max_stack;
296
297 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
298 false, false);
299
300 if (unlikely(!trace))
301 /* couldn't fetch the stack trace */
302 return -EFAULT;
303
304 return __bpf_get_stackid(map, trace, flags);
305}
306
307const struct bpf_func_proto bpf_get_stackid_proto = {
308 .func = bpf_get_stackid,
309 .gpl_only = true,
310 .ret_type = RET_INTEGER,
311 .arg1_type = ARG_PTR_TO_CTX,
312 .arg2_type = ARG_CONST_MAP_PTR,
313 .arg3_type = ARG_ANYTHING,
314};
315
316static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
317{
318 __u64 nr_kernel = 0;
319
320 while (nr_kernel < trace->nr) {
321 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
322 break;
323 nr_kernel++;
324 }
325 return nr_kernel;
326}
327
328BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
329 struct bpf_map *, map, u64, flags)
330{
331 struct perf_event *event = ctx->event;
332 struct perf_callchain_entry *trace;
333 bool kernel, user;
334 __u64 nr_kernel;
335 int ret;
336
337 /* perf_sample_data doesn't have callchain, use bpf_get_stackid */
338 if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
339 return bpf_get_stackid((unsigned long)(ctx->regs),
340 (unsigned long) map, flags, 0, 0);
341
342 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
343 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
344 return -EINVAL;
345
346 user = flags & BPF_F_USER_STACK;
347 kernel = !user;
348
349 trace = ctx->data->callchain;
350 if (unlikely(!trace))
351 return -EFAULT;
352
353 nr_kernel = count_kernel_ip(trace);
354
355 if (kernel) {
356 __u64 nr = trace->nr;
357
358 trace->nr = nr_kernel;
359 ret = __bpf_get_stackid(map, trace, flags);
360
361 /* restore nr */
362 trace->nr = nr;
363 } else { /* user */
364 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
365
366 skip += nr_kernel;
367 if (skip > BPF_F_SKIP_FIELD_MASK)
368 return -EFAULT;
369
370 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
371 ret = __bpf_get_stackid(map, trace, flags);
372 }
373 return ret;
374}
375
376const struct bpf_func_proto bpf_get_stackid_proto_pe = {
377 .func = bpf_get_stackid_pe,
378 .gpl_only = false,
379 .ret_type = RET_INTEGER,
380 .arg1_type = ARG_PTR_TO_CTX,
381 .arg2_type = ARG_CONST_MAP_PTR,
382 .arg3_type = ARG_ANYTHING,
383};
384
385static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
386 struct perf_callchain_entry *trace_in,
387 void *buf, u32 size, u64 flags)
388{
389 u32 trace_nr, copy_len, elem_size, num_elem, max_depth;
390 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
391 bool crosstask = task && task != current;
392 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
393 bool user = flags & BPF_F_USER_STACK;
394 struct perf_callchain_entry *trace;
395 bool kernel = !user;
396 int err = -EINVAL;
397 u64 *ips;
398
399 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
400 BPF_F_USER_BUILD_ID)))
401 goto clear;
402 if (kernel && user_build_id)
403 goto clear;
404
405 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
406 : sizeof(u64);
407 if (unlikely(size % elem_size))
408 goto clear;
409
410 /* cannot get valid user stack for task without user_mode regs */
411 if (task && user && !user_mode(regs))
412 goto err_fault;
413
414 /* get_perf_callchain does not support crosstask user stack walking
415 * but returns an empty stack instead of NULL.
416 */
417 if (crosstask && user) {
418 err = -EOPNOTSUPP;
419 goto clear;
420 }
421
422 num_elem = size / elem_size;
423 max_depth = num_elem + skip;
424 if (sysctl_perf_event_max_stack < max_depth)
425 max_depth = sysctl_perf_event_max_stack;
426
427 if (trace_in)
428 trace = trace_in;
429 else if (kernel && task)
430 trace = get_callchain_entry_for_task(task, max_depth);
431 else
432 trace = get_perf_callchain(regs, 0, kernel, user, max_depth,
433 crosstask, false);
434 if (unlikely(!trace))
435 goto err_fault;
436
437 if (trace->nr < skip)
438 goto err_fault;
439
440 trace_nr = trace->nr - skip;
441 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
442 copy_len = trace_nr * elem_size;
443
444 ips = trace->ip + skip;
445 if (user && user_build_id)
446 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
447 else
448 memcpy(buf, ips, copy_len);
449
450 if (size > copy_len)
451 memset(buf + copy_len, 0, size - copy_len);
452 return copy_len;
453
454err_fault:
455 err = -EFAULT;
456clear:
457 memset(buf, 0, size);
458 return err;
459}
460
461BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
462 u64, flags)
463{
464 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
465}
466
467const struct bpf_func_proto bpf_get_stack_proto = {
468 .func = bpf_get_stack,
469 .gpl_only = true,
470 .ret_type = RET_INTEGER,
471 .arg1_type = ARG_PTR_TO_CTX,
472 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
473 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
474 .arg4_type = ARG_ANYTHING,
475};
476
477BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
478 u32, size, u64, flags)
479{
480 struct pt_regs *regs;
481 long res = -EINVAL;
482
483 if (!try_get_task_stack(task))
484 return -EFAULT;
485
486 regs = task_pt_regs(task);
487 if (regs)
488 res = __bpf_get_stack(regs, task, NULL, buf, size, flags);
489 put_task_stack(task);
490
491 return res;
492}
493
494const struct bpf_func_proto bpf_get_task_stack_proto = {
495 .func = bpf_get_task_stack,
496 .gpl_only = false,
497 .ret_type = RET_INTEGER,
498 .arg1_type = ARG_PTR_TO_BTF_ID,
499 .arg1_btf_id = &btf_tracing_ids[BTF_TRACING_TYPE_TASK],
500 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
501 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
502 .arg4_type = ARG_ANYTHING,
503};
504
505BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
506 void *, buf, u32, size, u64, flags)
507{
508 struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
509 struct perf_event *event = ctx->event;
510 struct perf_callchain_entry *trace;
511 bool kernel, user;
512 int err = -EINVAL;
513 __u64 nr_kernel;
514
515 if (!(event->attr.sample_type & PERF_SAMPLE_CALLCHAIN))
516 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
517
518 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
519 BPF_F_USER_BUILD_ID)))
520 goto clear;
521
522 user = flags & BPF_F_USER_STACK;
523 kernel = !user;
524
525 err = -EFAULT;
526 trace = ctx->data->callchain;
527 if (unlikely(!trace))
528 goto clear;
529
530 nr_kernel = count_kernel_ip(trace);
531
532 if (kernel) {
533 __u64 nr = trace->nr;
534
535 trace->nr = nr_kernel;
536 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
537
538 /* restore nr */
539 trace->nr = nr;
540 } else { /* user */
541 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
542
543 skip += nr_kernel;
544 if (skip > BPF_F_SKIP_FIELD_MASK)
545 goto clear;
546
547 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
548 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
549 }
550 return err;
551
552clear:
553 memset(buf, 0, size);
554 return err;
555
556}
557
558const struct bpf_func_proto bpf_get_stack_proto_pe = {
559 .func = bpf_get_stack_pe,
560 .gpl_only = true,
561 .ret_type = RET_INTEGER,
562 .arg1_type = ARG_PTR_TO_CTX,
563 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
564 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
565 .arg4_type = ARG_ANYTHING,
566};
567
568/* Called from eBPF program */
569static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
570{
571 return ERR_PTR(-EOPNOTSUPP);
572}
573
574/* Called from syscall */
575int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
576{
577 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
578 struct stack_map_bucket *bucket, *old_bucket;
579 u32 id = *(u32 *)key, trace_len;
580
581 if (unlikely(id >= smap->n_buckets))
582 return -ENOENT;
583
584 bucket = xchg(&smap->buckets[id], NULL);
585 if (!bucket)
586 return -ENOENT;
587
588 trace_len = bucket->nr * stack_map_data_size(map);
589 memcpy(value, bucket->data, trace_len);
590 memset(value + trace_len, 0, map->value_size - trace_len);
591
592 old_bucket = xchg(&smap->buckets[id], bucket);
593 if (old_bucket)
594 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
595 return 0;
596}
597
598static int stack_map_get_next_key(struct bpf_map *map, void *key,
599 void *next_key)
600{
601 struct bpf_stack_map *smap = container_of(map,
602 struct bpf_stack_map, map);
603 u32 id;
604
605 WARN_ON_ONCE(!rcu_read_lock_held());
606
607 if (!key) {
608 id = 0;
609 } else {
610 id = *(u32 *)key;
611 if (id >= smap->n_buckets || !smap->buckets[id])
612 id = 0;
613 else
614 id++;
615 }
616
617 while (id < smap->n_buckets && !smap->buckets[id])
618 id++;
619
620 if (id >= smap->n_buckets)
621 return -ENOENT;
622
623 *(u32 *)next_key = id;
624 return 0;
625}
626
627static long stack_map_update_elem(struct bpf_map *map, void *key, void *value,
628 u64 map_flags)
629{
630 return -EINVAL;
631}
632
633/* Called from syscall or from eBPF program */
634static long stack_map_delete_elem(struct bpf_map *map, void *key)
635{
636 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
637 struct stack_map_bucket *old_bucket;
638 u32 id = *(u32 *)key;
639
640 if (unlikely(id >= smap->n_buckets))
641 return -E2BIG;
642
643 old_bucket = xchg(&smap->buckets[id], NULL);
644 if (old_bucket) {
645 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
646 return 0;
647 } else {
648 return -ENOENT;
649 }
650}
651
652/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
653static void stack_map_free(struct bpf_map *map)
654{
655 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
656
657 bpf_map_area_free(smap->elems);
658 pcpu_freelist_destroy(&smap->freelist);
659 bpf_map_area_free(smap);
660 put_callchain_buffers();
661}
662
663static u64 stack_map_mem_usage(const struct bpf_map *map)
664{
665 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
666 u64 value_size = map->value_size;
667 u64 n_buckets = smap->n_buckets;
668 u64 enties = map->max_entries;
669 u64 usage = sizeof(*smap);
670
671 usage += n_buckets * sizeof(struct stack_map_bucket *);
672 usage += enties * (sizeof(struct stack_map_bucket) + value_size);
673 return usage;
674}
675
676BTF_ID_LIST_SINGLE(stack_trace_map_btf_ids, struct, bpf_stack_map)
677const struct bpf_map_ops stack_trace_map_ops = {
678 .map_meta_equal = bpf_map_meta_equal,
679 .map_alloc = stack_map_alloc,
680 .map_free = stack_map_free,
681 .map_get_next_key = stack_map_get_next_key,
682 .map_lookup_elem = stack_map_lookup_elem,
683 .map_update_elem = stack_map_update_elem,
684 .map_delete_elem = stack_map_delete_elem,
685 .map_check_btf = map_check_no_btf,
686 .map_mem_usage = stack_map_mem_usage,
687 .map_btf_id = &stack_trace_map_btf_ids[0],
688};
1/* Copyright (c) 2016 Facebook
2 *
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.
6 */
7#include <linux/bpf.h>
8#include <linux/jhash.h>
9#include <linux/filter.h>
10#include <linux/stacktrace.h>
11#include <linux/perf_event.h>
12#include "percpu_freelist.h"
13
14struct stack_map_bucket {
15 struct pcpu_freelist_node fnode;
16 u32 hash;
17 u32 nr;
18 u64 ip[];
19};
20
21struct bpf_stack_map {
22 struct bpf_map map;
23 void *elems;
24 struct pcpu_freelist freelist;
25 u32 n_buckets;
26 struct stack_map_bucket *buckets[];
27};
28
29static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
30{
31 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
32 int err;
33
34 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries);
35 if (!smap->elems)
36 return -ENOMEM;
37
38 err = pcpu_freelist_init(&smap->freelist);
39 if (err)
40 goto free_elems;
41
42 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
43 smap->map.max_entries);
44 return 0;
45
46free_elems:
47 bpf_map_area_free(smap->elems);
48 return err;
49}
50
51/* Called from syscall */
52static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
53{
54 u32 value_size = attr->value_size;
55 struct bpf_stack_map *smap;
56 u64 cost, n_buckets;
57 int err;
58
59 if (!capable(CAP_SYS_ADMIN))
60 return ERR_PTR(-EPERM);
61
62 if (attr->map_flags)
63 return ERR_PTR(-EINVAL);
64
65 /* check sanity of attributes */
66 if (attr->max_entries == 0 || attr->key_size != 4 ||
67 value_size < 8 || value_size % 8 ||
68 value_size / 8 > sysctl_perf_event_max_stack)
69 return ERR_PTR(-EINVAL);
70
71 /* hash table size must be power of 2 */
72 n_buckets = roundup_pow_of_two(attr->max_entries);
73
74 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
75 if (cost >= U32_MAX - PAGE_SIZE)
76 return ERR_PTR(-E2BIG);
77
78 smap = bpf_map_area_alloc(cost);
79 if (!smap)
80 return ERR_PTR(-ENOMEM);
81
82 err = -E2BIG;
83 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
84 if (cost >= U32_MAX - PAGE_SIZE)
85 goto free_smap;
86
87 smap->map.map_type = attr->map_type;
88 smap->map.key_size = attr->key_size;
89 smap->map.value_size = value_size;
90 smap->map.max_entries = attr->max_entries;
91 smap->n_buckets = n_buckets;
92 smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
93
94 err = bpf_map_precharge_memlock(smap->map.pages);
95 if (err)
96 goto free_smap;
97
98 err = get_callchain_buffers(sysctl_perf_event_max_stack);
99 if (err)
100 goto free_smap;
101
102 err = prealloc_elems_and_freelist(smap);
103 if (err)
104 goto put_buffers;
105
106 return &smap->map;
107
108put_buffers:
109 put_callchain_buffers();
110free_smap:
111 bpf_map_area_free(smap);
112 return ERR_PTR(err);
113}
114
115BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
116 u64, flags)
117{
118 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
119 struct perf_callchain_entry *trace;
120 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
121 u32 max_depth = map->value_size / 8;
122 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
123 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
124 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
125 u32 hash, id, trace_nr, trace_len;
126 bool user = flags & BPF_F_USER_STACK;
127 bool kernel = !user;
128 u64 *ips;
129
130 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
131 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
132 return -EINVAL;
133
134 trace = get_perf_callchain(regs, init_nr, kernel, user,
135 sysctl_perf_event_max_stack, false, false);
136
137 if (unlikely(!trace))
138 /* couldn't fetch the stack trace */
139 return -EFAULT;
140
141 /* get_perf_callchain() guarantees that trace->nr >= init_nr
142 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
143 */
144 trace_nr = trace->nr - init_nr;
145
146 if (trace_nr <= skip)
147 /* skipping more than usable stack trace */
148 return -EFAULT;
149
150 trace_nr -= skip;
151 trace_len = trace_nr * sizeof(u64);
152 ips = trace->ip + skip + init_nr;
153 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
154 id = hash & (smap->n_buckets - 1);
155 bucket = READ_ONCE(smap->buckets[id]);
156
157 if (bucket && bucket->hash == hash) {
158 if (flags & BPF_F_FAST_STACK_CMP)
159 return id;
160 if (bucket->nr == trace_nr &&
161 memcmp(bucket->ip, ips, trace_len) == 0)
162 return id;
163 }
164
165 /* this call stack is not in the map, try to add it */
166 if (bucket && !(flags & BPF_F_REUSE_STACKID))
167 return -EEXIST;
168
169 new_bucket = (struct stack_map_bucket *)
170 pcpu_freelist_pop(&smap->freelist);
171 if (unlikely(!new_bucket))
172 return -ENOMEM;
173
174 memcpy(new_bucket->ip, ips, trace_len);
175 new_bucket->hash = hash;
176 new_bucket->nr = trace_nr;
177
178 old_bucket = xchg(&smap->buckets[id], new_bucket);
179 if (old_bucket)
180 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
181 return id;
182}
183
184const struct bpf_func_proto bpf_get_stackid_proto = {
185 .func = bpf_get_stackid,
186 .gpl_only = true,
187 .ret_type = RET_INTEGER,
188 .arg1_type = ARG_PTR_TO_CTX,
189 .arg2_type = ARG_CONST_MAP_PTR,
190 .arg3_type = ARG_ANYTHING,
191};
192
193/* Called from eBPF program */
194static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
195{
196 return NULL;
197}
198
199/* Called from syscall */
200int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
201{
202 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
203 struct stack_map_bucket *bucket, *old_bucket;
204 u32 id = *(u32 *)key, trace_len;
205
206 if (unlikely(id >= smap->n_buckets))
207 return -ENOENT;
208
209 bucket = xchg(&smap->buckets[id], NULL);
210 if (!bucket)
211 return -ENOENT;
212
213 trace_len = bucket->nr * sizeof(u64);
214 memcpy(value, bucket->ip, trace_len);
215 memset(value + trace_len, 0, map->value_size - trace_len);
216
217 old_bucket = xchg(&smap->buckets[id], bucket);
218 if (old_bucket)
219 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
220 return 0;
221}
222
223static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
224{
225 return -EINVAL;
226}
227
228static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
229 u64 map_flags)
230{
231 return -EINVAL;
232}
233
234/* Called from syscall or from eBPF program */
235static int stack_map_delete_elem(struct bpf_map *map, void *key)
236{
237 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
238 struct stack_map_bucket *old_bucket;
239 u32 id = *(u32 *)key;
240
241 if (unlikely(id >= smap->n_buckets))
242 return -E2BIG;
243
244 old_bucket = xchg(&smap->buckets[id], NULL);
245 if (old_bucket) {
246 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
247 return 0;
248 } else {
249 return -ENOENT;
250 }
251}
252
253/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
254static void stack_map_free(struct bpf_map *map)
255{
256 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
257
258 /* wait for bpf programs to complete before freeing stack map */
259 synchronize_rcu();
260
261 bpf_map_area_free(smap->elems);
262 pcpu_freelist_destroy(&smap->freelist);
263 bpf_map_area_free(smap);
264 put_callchain_buffers();
265}
266
267static const struct bpf_map_ops stack_map_ops = {
268 .map_alloc = stack_map_alloc,
269 .map_free = stack_map_free,
270 .map_get_next_key = stack_map_get_next_key,
271 .map_lookup_elem = stack_map_lookup_elem,
272 .map_update_elem = stack_map_update_elem,
273 .map_delete_elem = stack_map_delete_elem,
274};
275
276static struct bpf_map_type_list stack_map_type __read_mostly = {
277 .ops = &stack_map_ops,
278 .type = BPF_MAP_TYPE_STACK_TRACE,
279};
280
281static int __init register_stack_map(void)
282{
283 bpf_register_map_type(&stack_map_type);
284 return 0;
285}
286late_initcall(register_stack_map);