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/stacktrace.h>
8#include <linux/perf_event.h>
9#include <linux/elf.h>
10#include <linux/pagemap.h>
11#include <linux/irq_work.h>
12#include "percpu_freelist.h"
13
14#define STACK_CREATE_FLAG_MASK \
15 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
16 BPF_F_STACK_BUILD_ID)
17
18struct stack_map_bucket {
19 struct pcpu_freelist_node fnode;
20 u32 hash;
21 u32 nr;
22 u64 data[];
23};
24
25struct bpf_stack_map {
26 struct bpf_map map;
27 void *elems;
28 struct pcpu_freelist freelist;
29 u32 n_buckets;
30 struct stack_map_bucket *buckets[];
31};
32
33/* irq_work to run up_read() for build_id lookup in nmi context */
34struct stack_map_irq_work {
35 struct irq_work irq_work;
36 struct rw_semaphore *sem;
37};
38
39static void do_up_read(struct irq_work *entry)
40{
41 struct stack_map_irq_work *work;
42
43 work = container_of(entry, struct stack_map_irq_work, irq_work);
44 up_read_non_owner(work->sem);
45 work->sem = NULL;
46}
47
48static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
49
50static inline bool stack_map_use_build_id(struct bpf_map *map)
51{
52 return (map->map_flags & BPF_F_STACK_BUILD_ID);
53}
54
55static inline int stack_map_data_size(struct bpf_map *map)
56{
57 return stack_map_use_build_id(map) ?
58 sizeof(struct bpf_stack_build_id) : sizeof(u64);
59}
60
61static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
62{
63 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
64 int err;
65
66 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
67 smap->map.numa_node);
68 if (!smap->elems)
69 return -ENOMEM;
70
71 err = pcpu_freelist_init(&smap->freelist);
72 if (err)
73 goto free_elems;
74
75 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
76 smap->map.max_entries);
77 return 0;
78
79free_elems:
80 bpf_map_area_free(smap->elems);
81 return err;
82}
83
84/* Called from syscall */
85static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
86{
87 u32 value_size = attr->value_size;
88 struct bpf_stack_map *smap;
89 struct bpf_map_memory mem;
90 u64 cost, n_buckets;
91 int err;
92
93 if (!capable(CAP_SYS_ADMIN))
94 return ERR_PTR(-EPERM);
95
96 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
97 return ERR_PTR(-EINVAL);
98
99 /* check sanity of attributes */
100 if (attr->max_entries == 0 || attr->key_size != 4 ||
101 value_size < 8 || value_size % 8)
102 return ERR_PTR(-EINVAL);
103
104 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
105 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
106 if (value_size % sizeof(struct bpf_stack_build_id) ||
107 value_size / sizeof(struct bpf_stack_build_id)
108 > sysctl_perf_event_max_stack)
109 return ERR_PTR(-EINVAL);
110 } else if (value_size / 8 > sysctl_perf_event_max_stack)
111 return ERR_PTR(-EINVAL);
112
113 /* hash table size must be power of 2 */
114 n_buckets = roundup_pow_of_two(attr->max_entries);
115
116 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
117 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
118 err = bpf_map_charge_init(&mem, cost);
119 if (err)
120 return ERR_PTR(err);
121
122 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
123 if (!smap) {
124 bpf_map_charge_finish(&mem);
125 return ERR_PTR(-ENOMEM);
126 }
127
128 bpf_map_init_from_attr(&smap->map, attr);
129 smap->map.value_size = value_size;
130 smap->n_buckets = n_buckets;
131
132 err = get_callchain_buffers(sysctl_perf_event_max_stack);
133 if (err)
134 goto free_charge;
135
136 err = prealloc_elems_and_freelist(smap);
137 if (err)
138 goto put_buffers;
139
140 bpf_map_charge_move(&smap->map.memory, &mem);
141
142 return &smap->map;
143
144put_buffers:
145 put_callchain_buffers();
146free_charge:
147 bpf_map_charge_finish(&mem);
148 bpf_map_area_free(smap);
149 return ERR_PTR(err);
150}
151
152#define BPF_BUILD_ID 3
153/*
154 * Parse build id from the note segment. This logic can be shared between
155 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
156 * identical.
157 */
158static inline int stack_map_parse_build_id(void *page_addr,
159 unsigned char *build_id,
160 void *note_start,
161 Elf32_Word note_size)
162{
163 Elf32_Word note_offs = 0, new_offs;
164
165 /* check for overflow */
166 if (note_start < page_addr || note_start + note_size < note_start)
167 return -EINVAL;
168
169 /* only supports note that fits in the first page */
170 if (note_start + note_size > page_addr + PAGE_SIZE)
171 return -EINVAL;
172
173 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
174 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
175
176 if (nhdr->n_type == BPF_BUILD_ID &&
177 nhdr->n_namesz == sizeof("GNU") &&
178 nhdr->n_descsz > 0 &&
179 nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
180 memcpy(build_id,
181 note_start + note_offs +
182 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
183 nhdr->n_descsz);
184 memset(build_id + nhdr->n_descsz, 0,
185 BPF_BUILD_ID_SIZE - nhdr->n_descsz);
186 return 0;
187 }
188 new_offs = note_offs + sizeof(Elf32_Nhdr) +
189 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
190 if (new_offs <= note_offs) /* overflow */
191 break;
192 note_offs = new_offs;
193 }
194 return -EINVAL;
195}
196
197/* Parse build ID from 32-bit ELF */
198static int stack_map_get_build_id_32(void *page_addr,
199 unsigned char *build_id)
200{
201 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
202 Elf32_Phdr *phdr;
203 int i;
204
205 /* only supports phdr that fits in one page */
206 if (ehdr->e_phnum >
207 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
208 return -EINVAL;
209
210 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
211
212 for (i = 0; i < ehdr->e_phnum; ++i)
213 if (phdr[i].p_type == PT_NOTE)
214 return stack_map_parse_build_id(page_addr, build_id,
215 page_addr + phdr[i].p_offset,
216 phdr[i].p_filesz);
217 return -EINVAL;
218}
219
220/* Parse build ID from 64-bit ELF */
221static int stack_map_get_build_id_64(void *page_addr,
222 unsigned char *build_id)
223{
224 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
225 Elf64_Phdr *phdr;
226 int i;
227
228 /* only supports phdr that fits in one page */
229 if (ehdr->e_phnum >
230 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
231 return -EINVAL;
232
233 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
234
235 for (i = 0; i < ehdr->e_phnum; ++i)
236 if (phdr[i].p_type == PT_NOTE)
237 return stack_map_parse_build_id(page_addr, build_id,
238 page_addr + phdr[i].p_offset,
239 phdr[i].p_filesz);
240 return -EINVAL;
241}
242
243/* Parse build ID of ELF file mapped to vma */
244static int stack_map_get_build_id(struct vm_area_struct *vma,
245 unsigned char *build_id)
246{
247 Elf32_Ehdr *ehdr;
248 struct page *page;
249 void *page_addr;
250 int ret;
251
252 /* only works for page backed storage */
253 if (!vma->vm_file)
254 return -EINVAL;
255
256 page = find_get_page(vma->vm_file->f_mapping, 0);
257 if (!page)
258 return -EFAULT; /* page not mapped */
259
260 ret = -EINVAL;
261 page_addr = kmap_atomic(page);
262 ehdr = (Elf32_Ehdr *)page_addr;
263
264 /* compare magic x7f "ELF" */
265 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
266 goto out;
267
268 /* only support executable file and shared object file */
269 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
270 goto out;
271
272 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
273 ret = stack_map_get_build_id_32(page_addr, build_id);
274 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
275 ret = stack_map_get_build_id_64(page_addr, build_id);
276out:
277 kunmap_atomic(page_addr);
278 put_page(page);
279 return ret;
280}
281
282static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
283 u64 *ips, u32 trace_nr, bool user)
284{
285 int i;
286 struct vm_area_struct *vma;
287 bool irq_work_busy = false;
288 struct stack_map_irq_work *work = NULL;
289
290 if (in_nmi()) {
291 work = this_cpu_ptr(&up_read_work);
292 if (work->irq_work.flags & IRQ_WORK_BUSY)
293 /* cannot queue more up_read, fallback */
294 irq_work_busy = true;
295 }
296
297 /*
298 * We cannot do up_read() in nmi context. To do build_id lookup
299 * in nmi context, we need to run up_read() in irq_work. We use
300 * a percpu variable to do the irq_work. If the irq_work is
301 * already used by another lookup, we fall back to report ips.
302 *
303 * Same fallback is used for kernel stack (!user) on a stackmap
304 * with build_id.
305 */
306 if (!user || !current || !current->mm || irq_work_busy ||
307 down_read_trylock(¤t->mm->mmap_sem) == 0) {
308 /* cannot access current->mm, fall back to ips */
309 for (i = 0; i < trace_nr; i++) {
310 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
311 id_offs[i].ip = ips[i];
312 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
313 }
314 return;
315 }
316
317 for (i = 0; i < trace_nr; i++) {
318 vma = find_vma(current->mm, ips[i]);
319 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
320 /* per entry fall back to ips */
321 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
322 id_offs[i].ip = ips[i];
323 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
324 continue;
325 }
326 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
327 - vma->vm_start;
328 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
329 }
330
331 if (!work) {
332 up_read(¤t->mm->mmap_sem);
333 } else {
334 work->sem = ¤t->mm->mmap_sem;
335 irq_work_queue(&work->irq_work);
336 /*
337 * The irq_work will release the mmap_sem with
338 * up_read_non_owner(). The rwsem_release() is called
339 * here to release the lock from lockdep's perspective.
340 */
341 rwsem_release(¤t->mm->mmap_sem.dep_map, 1, _RET_IP_);
342 }
343}
344
345BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
346 u64, flags)
347{
348 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
349 struct perf_callchain_entry *trace;
350 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
351 u32 max_depth = map->value_size / stack_map_data_size(map);
352 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
353 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
354 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
355 u32 hash, id, trace_nr, trace_len;
356 bool user = flags & BPF_F_USER_STACK;
357 bool kernel = !user;
358 u64 *ips;
359 bool hash_matches;
360
361 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
362 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
363 return -EINVAL;
364
365 trace = get_perf_callchain(regs, init_nr, kernel, user,
366 sysctl_perf_event_max_stack, false, false);
367
368 if (unlikely(!trace))
369 /* couldn't fetch the stack trace */
370 return -EFAULT;
371
372 /* get_perf_callchain() guarantees that trace->nr >= init_nr
373 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
374 */
375 trace_nr = trace->nr - init_nr;
376
377 if (trace_nr <= skip)
378 /* skipping more than usable stack trace */
379 return -EFAULT;
380
381 trace_nr -= skip;
382 trace_len = trace_nr * sizeof(u64);
383 ips = trace->ip + skip + init_nr;
384 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
385 id = hash & (smap->n_buckets - 1);
386 bucket = READ_ONCE(smap->buckets[id]);
387
388 hash_matches = bucket && bucket->hash == hash;
389 /* fast cmp */
390 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
391 return id;
392
393 if (stack_map_use_build_id(map)) {
394 /* for build_id+offset, pop a bucket before slow cmp */
395 new_bucket = (struct stack_map_bucket *)
396 pcpu_freelist_pop(&smap->freelist);
397 if (unlikely(!new_bucket))
398 return -ENOMEM;
399 new_bucket->nr = trace_nr;
400 stack_map_get_build_id_offset(
401 (struct bpf_stack_build_id *)new_bucket->data,
402 ips, trace_nr, user);
403 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
404 if (hash_matches && bucket->nr == trace_nr &&
405 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
406 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
407 return id;
408 }
409 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
410 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
411 return -EEXIST;
412 }
413 } else {
414 if (hash_matches && bucket->nr == trace_nr &&
415 memcmp(bucket->data, ips, trace_len) == 0)
416 return id;
417 if (bucket && !(flags & BPF_F_REUSE_STACKID))
418 return -EEXIST;
419
420 new_bucket = (struct stack_map_bucket *)
421 pcpu_freelist_pop(&smap->freelist);
422 if (unlikely(!new_bucket))
423 return -ENOMEM;
424 memcpy(new_bucket->data, ips, trace_len);
425 }
426
427 new_bucket->hash = hash;
428 new_bucket->nr = trace_nr;
429
430 old_bucket = xchg(&smap->buckets[id], new_bucket);
431 if (old_bucket)
432 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
433 return id;
434}
435
436const struct bpf_func_proto bpf_get_stackid_proto = {
437 .func = bpf_get_stackid,
438 .gpl_only = true,
439 .ret_type = RET_INTEGER,
440 .arg1_type = ARG_PTR_TO_CTX,
441 .arg2_type = ARG_CONST_MAP_PTR,
442 .arg3_type = ARG_ANYTHING,
443};
444
445BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
446 u64, flags)
447{
448 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
449 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
450 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
451 bool user = flags & BPF_F_USER_STACK;
452 struct perf_callchain_entry *trace;
453 bool kernel = !user;
454 int err = -EINVAL;
455 u64 *ips;
456
457 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
458 BPF_F_USER_BUILD_ID)))
459 goto clear;
460 if (kernel && user_build_id)
461 goto clear;
462
463 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
464 : sizeof(u64);
465 if (unlikely(size % elem_size))
466 goto clear;
467
468 num_elem = size / elem_size;
469 if (sysctl_perf_event_max_stack < num_elem)
470 init_nr = 0;
471 else
472 init_nr = sysctl_perf_event_max_stack - num_elem;
473 trace = get_perf_callchain(regs, init_nr, kernel, user,
474 sysctl_perf_event_max_stack, false, false);
475 if (unlikely(!trace))
476 goto err_fault;
477
478 trace_nr = trace->nr - init_nr;
479 if (trace_nr < skip)
480 goto err_fault;
481
482 trace_nr -= skip;
483 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
484 copy_len = trace_nr * elem_size;
485 ips = trace->ip + skip + init_nr;
486 if (user && user_build_id)
487 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
488 else
489 memcpy(buf, ips, copy_len);
490
491 if (size > copy_len)
492 memset(buf + copy_len, 0, size - copy_len);
493 return copy_len;
494
495err_fault:
496 err = -EFAULT;
497clear:
498 memset(buf, 0, size);
499 return err;
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
512/* Called from eBPF program */
513static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
514{
515 return ERR_PTR(-EOPNOTSUPP);
516}
517
518/* Called from syscall */
519int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
520{
521 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
522 struct stack_map_bucket *bucket, *old_bucket;
523 u32 id = *(u32 *)key, trace_len;
524
525 if (unlikely(id >= smap->n_buckets))
526 return -ENOENT;
527
528 bucket = xchg(&smap->buckets[id], NULL);
529 if (!bucket)
530 return -ENOENT;
531
532 trace_len = bucket->nr * stack_map_data_size(map);
533 memcpy(value, bucket->data, trace_len);
534 memset(value + trace_len, 0, map->value_size - trace_len);
535
536 old_bucket = xchg(&smap->buckets[id], bucket);
537 if (old_bucket)
538 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
539 return 0;
540}
541
542static int stack_map_get_next_key(struct bpf_map *map, void *key,
543 void *next_key)
544{
545 struct bpf_stack_map *smap = container_of(map,
546 struct bpf_stack_map, map);
547 u32 id;
548
549 WARN_ON_ONCE(!rcu_read_lock_held());
550
551 if (!key) {
552 id = 0;
553 } else {
554 id = *(u32 *)key;
555 if (id >= smap->n_buckets || !smap->buckets[id])
556 id = 0;
557 else
558 id++;
559 }
560
561 while (id < smap->n_buckets && !smap->buckets[id])
562 id++;
563
564 if (id >= smap->n_buckets)
565 return -ENOENT;
566
567 *(u32 *)next_key = id;
568 return 0;
569}
570
571static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
572 u64 map_flags)
573{
574 return -EINVAL;
575}
576
577/* Called from syscall or from eBPF program */
578static int stack_map_delete_elem(struct bpf_map *map, void *key)
579{
580 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
581 struct stack_map_bucket *old_bucket;
582 u32 id = *(u32 *)key;
583
584 if (unlikely(id >= smap->n_buckets))
585 return -E2BIG;
586
587 old_bucket = xchg(&smap->buckets[id], NULL);
588 if (old_bucket) {
589 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
590 return 0;
591 } else {
592 return -ENOENT;
593 }
594}
595
596/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
597static void stack_map_free(struct bpf_map *map)
598{
599 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
600
601 /* wait for bpf programs to complete before freeing stack map */
602 synchronize_rcu();
603
604 bpf_map_area_free(smap->elems);
605 pcpu_freelist_destroy(&smap->freelist);
606 bpf_map_area_free(smap);
607 put_callchain_buffers();
608}
609
610const struct bpf_map_ops stack_trace_map_ops = {
611 .map_alloc = stack_map_alloc,
612 .map_free = stack_map_free,
613 .map_get_next_key = stack_map_get_next_key,
614 .map_lookup_elem = stack_map_lookup_elem,
615 .map_update_elem = stack_map_update_elem,
616 .map_delete_elem = stack_map_delete_elem,
617 .map_check_btf = map_check_no_btf,
618};
619
620static int __init stack_map_init(void)
621{
622 int cpu;
623 struct stack_map_irq_work *work;
624
625 for_each_possible_cpu(cpu) {
626 work = per_cpu_ptr(&up_read_work, cpu);
627 init_irq_work(&work->irq_work, do_up_read);
628 }
629 return 0;
630}
631subsys_initcall(stack_map_init);