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; 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};
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/elf.h>
11#include <linux/pagemap.h>
12#include <linux/irq_work.h>
13#include <linux/btf_ids.h>
14#include "percpu_freelist.h"
15
16#define STACK_CREATE_FLAG_MASK \
17 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
18 BPF_F_STACK_BUILD_ID)
19
20struct stack_map_bucket {
21 struct pcpu_freelist_node fnode;
22 u32 hash;
23 u32 nr;
24 u64 data[];
25};
26
27struct bpf_stack_map {
28 struct bpf_map map;
29 void *elems;
30 struct pcpu_freelist freelist;
31 u32 n_buckets;
32 struct stack_map_bucket *buckets[];
33};
34
35/* irq_work to run up_read() for build_id lookup in nmi context */
36struct stack_map_irq_work {
37 struct irq_work irq_work;
38 struct mm_struct *mm;
39};
40
41static void do_up_read(struct irq_work *entry)
42{
43 struct stack_map_irq_work *work;
44
45 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
46 return;
47
48 work = container_of(entry, struct stack_map_irq_work, irq_work);
49 mmap_read_unlock_non_owner(work->mm);
50}
51
52static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
53
54static inline bool stack_map_use_build_id(struct bpf_map *map)
55{
56 return (map->map_flags & BPF_F_STACK_BUILD_ID);
57}
58
59static inline int stack_map_data_size(struct bpf_map *map)
60{
61 return stack_map_use_build_id(map) ?
62 sizeof(struct bpf_stack_build_id) : sizeof(u64);
63}
64
65static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
66{
67 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
68 int err;
69
70 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
71 smap->map.numa_node);
72 if (!smap->elems)
73 return -ENOMEM;
74
75 err = pcpu_freelist_init(&smap->freelist);
76 if (err)
77 goto free_elems;
78
79 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
80 smap->map.max_entries);
81 return 0;
82
83free_elems:
84 bpf_map_area_free(smap->elems);
85 return err;
86}
87
88/* Called from syscall */
89static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
90{
91 u32 value_size = attr->value_size;
92 struct bpf_stack_map *smap;
93 struct bpf_map_memory mem;
94 u64 cost, n_buckets;
95 int err;
96
97 if (!bpf_capable())
98 return ERR_PTR(-EPERM);
99
100 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
101 return ERR_PTR(-EINVAL);
102
103 /* check sanity of attributes */
104 if (attr->max_entries == 0 || attr->key_size != 4 ||
105 value_size < 8 || value_size % 8)
106 return ERR_PTR(-EINVAL);
107
108 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
109 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
110 if (value_size % sizeof(struct bpf_stack_build_id) ||
111 value_size / sizeof(struct bpf_stack_build_id)
112 > sysctl_perf_event_max_stack)
113 return ERR_PTR(-EINVAL);
114 } else if (value_size / 8 > sysctl_perf_event_max_stack)
115 return ERR_PTR(-EINVAL);
116
117 /* hash table size must be power of 2 */
118 n_buckets = roundup_pow_of_two(attr->max_entries);
119
120 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
121 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
122 err = bpf_map_charge_init(&mem, cost);
123 if (err)
124 return ERR_PTR(err);
125
126 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
127 if (!smap) {
128 bpf_map_charge_finish(&mem);
129 return ERR_PTR(-ENOMEM);
130 }
131
132 bpf_map_init_from_attr(&smap->map, attr);
133 smap->map.value_size = value_size;
134 smap->n_buckets = n_buckets;
135
136 err = get_callchain_buffers(sysctl_perf_event_max_stack);
137 if (err)
138 goto free_charge;
139
140 err = prealloc_elems_and_freelist(smap);
141 if (err)
142 goto put_buffers;
143
144 bpf_map_charge_move(&smap->map.memory, &mem);
145
146 return &smap->map;
147
148put_buffers:
149 put_callchain_buffers();
150free_charge:
151 bpf_map_charge_finish(&mem);
152 bpf_map_area_free(smap);
153 return ERR_PTR(err);
154}
155
156#define BPF_BUILD_ID 3
157/*
158 * Parse build id from the note segment. This logic can be shared between
159 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
160 * identical.
161 */
162static inline int stack_map_parse_build_id(void *page_addr,
163 unsigned char *build_id,
164 void *note_start,
165 Elf32_Word note_size)
166{
167 Elf32_Word note_offs = 0, new_offs;
168
169 /* check for overflow */
170 if (note_start < page_addr || note_start + note_size < note_start)
171 return -EINVAL;
172
173 /* only supports note that fits in the first page */
174 if (note_start + note_size > page_addr + PAGE_SIZE)
175 return -EINVAL;
176
177 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
178 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
179
180 if (nhdr->n_type == BPF_BUILD_ID &&
181 nhdr->n_namesz == sizeof("GNU") &&
182 nhdr->n_descsz > 0 &&
183 nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
184 memcpy(build_id,
185 note_start + note_offs +
186 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
187 nhdr->n_descsz);
188 memset(build_id + nhdr->n_descsz, 0,
189 BPF_BUILD_ID_SIZE - nhdr->n_descsz);
190 return 0;
191 }
192 new_offs = note_offs + sizeof(Elf32_Nhdr) +
193 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
194 if (new_offs <= note_offs) /* overflow */
195 break;
196 note_offs = new_offs;
197 }
198 return -EINVAL;
199}
200
201/* Parse build ID from 32-bit ELF */
202static int stack_map_get_build_id_32(void *page_addr,
203 unsigned char *build_id)
204{
205 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
206 Elf32_Phdr *phdr;
207 int i;
208
209 /* only supports phdr that fits in one page */
210 if (ehdr->e_phnum >
211 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
212 return -EINVAL;
213
214 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
215
216 for (i = 0; i < ehdr->e_phnum; ++i) {
217 if (phdr[i].p_type == PT_NOTE &&
218 !stack_map_parse_build_id(page_addr, build_id,
219 page_addr + phdr[i].p_offset,
220 phdr[i].p_filesz))
221 return 0;
222 }
223 return -EINVAL;
224}
225
226/* Parse build ID from 64-bit ELF */
227static int stack_map_get_build_id_64(void *page_addr,
228 unsigned char *build_id)
229{
230 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
231 Elf64_Phdr *phdr;
232 int i;
233
234 /* only supports phdr that fits in one page */
235 if (ehdr->e_phnum >
236 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
237 return -EINVAL;
238
239 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
240
241 for (i = 0; i < ehdr->e_phnum; ++i) {
242 if (phdr[i].p_type == PT_NOTE &&
243 !stack_map_parse_build_id(page_addr, build_id,
244 page_addr + phdr[i].p_offset,
245 phdr[i].p_filesz))
246 return 0;
247 }
248 return -EINVAL;
249}
250
251/* Parse build ID of ELF file mapped to vma */
252static int stack_map_get_build_id(struct vm_area_struct *vma,
253 unsigned char *build_id)
254{
255 Elf32_Ehdr *ehdr;
256 struct page *page;
257 void *page_addr;
258 int ret;
259
260 /* only works for page backed storage */
261 if (!vma->vm_file)
262 return -EINVAL;
263
264 page = find_get_page(vma->vm_file->f_mapping, 0);
265 if (!page)
266 return -EFAULT; /* page not mapped */
267
268 ret = -EINVAL;
269 page_addr = kmap_atomic(page);
270 ehdr = (Elf32_Ehdr *)page_addr;
271
272 /* compare magic x7f "ELF" */
273 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
274 goto out;
275
276 /* only support executable file and shared object file */
277 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
278 goto out;
279
280 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
281 ret = stack_map_get_build_id_32(page_addr, build_id);
282 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
283 ret = stack_map_get_build_id_64(page_addr, build_id);
284out:
285 kunmap_atomic(page_addr);
286 put_page(page);
287 return ret;
288}
289
290static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
291 u64 *ips, u32 trace_nr, bool user)
292{
293 int i;
294 struct vm_area_struct *vma;
295 bool irq_work_busy = false;
296 struct stack_map_irq_work *work = NULL;
297
298 if (irqs_disabled()) {
299 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
300 work = this_cpu_ptr(&up_read_work);
301 if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY) {
302 /* cannot queue more up_read, fallback */
303 irq_work_busy = true;
304 }
305 } else {
306 /*
307 * PREEMPT_RT does not allow to trylock mmap sem in
308 * interrupt disabled context. Force the fallback code.
309 */
310 irq_work_busy = true;
311 }
312 }
313
314 /*
315 * We cannot do up_read() when the irq is disabled, because of
316 * risk to deadlock with rq_lock. To do build_id lookup when the
317 * irqs are disabled, we need to run up_read() in irq_work. We use
318 * a percpu variable to do the irq_work. If the irq_work is
319 * already used by another lookup, we fall back to report ips.
320 *
321 * Same fallback is used for kernel stack (!user) on a stackmap
322 * with build_id.
323 */
324 if (!user || !current || !current->mm || irq_work_busy ||
325 !mmap_read_trylock_non_owner(current->mm)) {
326 /* cannot access current->mm, fall back to ips */
327 for (i = 0; i < trace_nr; i++) {
328 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
329 id_offs[i].ip = ips[i];
330 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
331 }
332 return;
333 }
334
335 for (i = 0; i < trace_nr; i++) {
336 vma = find_vma(current->mm, ips[i]);
337 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
338 /* per entry fall back to ips */
339 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
340 id_offs[i].ip = ips[i];
341 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
342 continue;
343 }
344 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
345 - vma->vm_start;
346 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
347 }
348
349 if (!work) {
350 mmap_read_unlock_non_owner(current->mm);
351 } else {
352 work->mm = current->mm;
353 irq_work_queue(&work->irq_work);
354 }
355}
356
357static struct perf_callchain_entry *
358get_callchain_entry_for_task(struct task_struct *task, u32 init_nr)
359{
360#ifdef CONFIG_STACKTRACE
361 struct perf_callchain_entry *entry;
362 int rctx;
363
364 entry = get_callchain_entry(&rctx);
365
366 if (!entry)
367 return NULL;
368
369 entry->nr = init_nr +
370 stack_trace_save_tsk(task, (unsigned long *)(entry->ip + init_nr),
371 sysctl_perf_event_max_stack - init_nr, 0);
372
373 /* stack_trace_save_tsk() works on unsigned long array, while
374 * perf_callchain_entry uses u64 array. For 32-bit systems, it is
375 * necessary to fix this mismatch.
376 */
377 if (__BITS_PER_LONG != 64) {
378 unsigned long *from = (unsigned long *) entry->ip;
379 u64 *to = entry->ip;
380 int i;
381
382 /* copy data from the end to avoid using extra buffer */
383 for (i = entry->nr - 1; i >= (int)init_nr; i--)
384 to[i] = (u64)(from[i]);
385 }
386
387 put_callchain_entry(rctx);
388
389 return entry;
390#else /* CONFIG_STACKTRACE */
391 return NULL;
392#endif
393}
394
395static long __bpf_get_stackid(struct bpf_map *map,
396 struct perf_callchain_entry *trace, u64 flags)
397{
398 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
399 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
400 u32 max_depth = map->value_size / stack_map_data_size(map);
401 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
402 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
403 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
404 u32 hash, id, trace_nr, trace_len;
405 bool user = flags & BPF_F_USER_STACK;
406 u64 *ips;
407 bool hash_matches;
408
409 /* get_perf_callchain() guarantees that trace->nr >= init_nr
410 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
411 */
412 trace_nr = trace->nr - init_nr;
413
414 if (trace_nr <= skip)
415 /* skipping more than usable stack trace */
416 return -EFAULT;
417
418 trace_nr -= skip;
419 trace_len = trace_nr * sizeof(u64);
420 ips = trace->ip + skip + init_nr;
421 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
422 id = hash & (smap->n_buckets - 1);
423 bucket = READ_ONCE(smap->buckets[id]);
424
425 hash_matches = bucket && bucket->hash == hash;
426 /* fast cmp */
427 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
428 return id;
429
430 if (stack_map_use_build_id(map)) {
431 /* for build_id+offset, pop a bucket before slow cmp */
432 new_bucket = (struct stack_map_bucket *)
433 pcpu_freelist_pop(&smap->freelist);
434 if (unlikely(!new_bucket))
435 return -ENOMEM;
436 new_bucket->nr = trace_nr;
437 stack_map_get_build_id_offset(
438 (struct bpf_stack_build_id *)new_bucket->data,
439 ips, trace_nr, user);
440 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
441 if (hash_matches && bucket->nr == trace_nr &&
442 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
443 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
444 return id;
445 }
446 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
447 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
448 return -EEXIST;
449 }
450 } else {
451 if (hash_matches && bucket->nr == trace_nr &&
452 memcmp(bucket->data, ips, trace_len) == 0)
453 return id;
454 if (bucket && !(flags & BPF_F_REUSE_STACKID))
455 return -EEXIST;
456
457 new_bucket = (struct stack_map_bucket *)
458 pcpu_freelist_pop(&smap->freelist);
459 if (unlikely(!new_bucket))
460 return -ENOMEM;
461 memcpy(new_bucket->data, ips, trace_len);
462 }
463
464 new_bucket->hash = hash;
465 new_bucket->nr = trace_nr;
466
467 old_bucket = xchg(&smap->buckets[id], new_bucket);
468 if (old_bucket)
469 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
470 return id;
471}
472
473BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
474 u64, flags)
475{
476 u32 max_depth = map->value_size / stack_map_data_size(map);
477 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
478 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
479 bool user = flags & BPF_F_USER_STACK;
480 struct perf_callchain_entry *trace;
481 bool kernel = !user;
482
483 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
484 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
485 return -EINVAL;
486
487 trace = get_perf_callchain(regs, init_nr, kernel, user,
488 sysctl_perf_event_max_stack, false, false);
489
490 if (unlikely(!trace))
491 /* couldn't fetch the stack trace */
492 return -EFAULT;
493
494 return __bpf_get_stackid(map, trace, flags);
495}
496
497const struct bpf_func_proto bpf_get_stackid_proto = {
498 .func = bpf_get_stackid,
499 .gpl_only = true,
500 .ret_type = RET_INTEGER,
501 .arg1_type = ARG_PTR_TO_CTX,
502 .arg2_type = ARG_CONST_MAP_PTR,
503 .arg3_type = ARG_ANYTHING,
504};
505
506static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
507{
508 __u64 nr_kernel = 0;
509
510 while (nr_kernel < trace->nr) {
511 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
512 break;
513 nr_kernel++;
514 }
515 return nr_kernel;
516}
517
518BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
519 struct bpf_map *, map, u64, flags)
520{
521 struct perf_event *event = ctx->event;
522 struct perf_callchain_entry *trace;
523 bool kernel, user;
524 __u64 nr_kernel;
525 int ret;
526
527 /* perf_sample_data doesn't have callchain, use bpf_get_stackid */
528 if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
529 return bpf_get_stackid((unsigned long)(ctx->regs),
530 (unsigned long) map, flags, 0, 0);
531
532 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
533 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
534 return -EINVAL;
535
536 user = flags & BPF_F_USER_STACK;
537 kernel = !user;
538
539 trace = ctx->data->callchain;
540 if (unlikely(!trace))
541 return -EFAULT;
542
543 nr_kernel = count_kernel_ip(trace);
544
545 if (kernel) {
546 __u64 nr = trace->nr;
547
548 trace->nr = nr_kernel;
549 ret = __bpf_get_stackid(map, trace, flags);
550
551 /* restore nr */
552 trace->nr = nr;
553 } else { /* user */
554 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
555
556 skip += nr_kernel;
557 if (skip > BPF_F_SKIP_FIELD_MASK)
558 return -EFAULT;
559
560 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
561 ret = __bpf_get_stackid(map, trace, flags);
562 }
563 return ret;
564}
565
566const struct bpf_func_proto bpf_get_stackid_proto_pe = {
567 .func = bpf_get_stackid_pe,
568 .gpl_only = false,
569 .ret_type = RET_INTEGER,
570 .arg1_type = ARG_PTR_TO_CTX,
571 .arg2_type = ARG_CONST_MAP_PTR,
572 .arg3_type = ARG_ANYTHING,
573};
574
575static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
576 struct perf_callchain_entry *trace_in,
577 void *buf, u32 size, u64 flags)
578{
579 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
580 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
581 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
582 bool user = flags & BPF_F_USER_STACK;
583 struct perf_callchain_entry *trace;
584 bool kernel = !user;
585 int err = -EINVAL;
586 u64 *ips;
587
588 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
589 BPF_F_USER_BUILD_ID)))
590 goto clear;
591 if (kernel && user_build_id)
592 goto clear;
593
594 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
595 : sizeof(u64);
596 if (unlikely(size % elem_size))
597 goto clear;
598
599 /* cannot get valid user stack for task without user_mode regs */
600 if (task && user && !user_mode(regs))
601 goto err_fault;
602
603 num_elem = size / elem_size;
604 if (sysctl_perf_event_max_stack < num_elem)
605 init_nr = 0;
606 else
607 init_nr = sysctl_perf_event_max_stack - num_elem;
608
609 if (trace_in)
610 trace = trace_in;
611 else if (kernel && task)
612 trace = get_callchain_entry_for_task(task, init_nr);
613 else
614 trace = get_perf_callchain(regs, init_nr, kernel, user,
615 sysctl_perf_event_max_stack,
616 false, false);
617 if (unlikely(!trace))
618 goto err_fault;
619
620 trace_nr = trace->nr - init_nr;
621 if (trace_nr < skip)
622 goto err_fault;
623
624 trace_nr -= skip;
625 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
626 copy_len = trace_nr * elem_size;
627 ips = trace->ip + skip + init_nr;
628 if (user && user_build_id)
629 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
630 else
631 memcpy(buf, ips, copy_len);
632
633 if (size > copy_len)
634 memset(buf + copy_len, 0, size - copy_len);
635 return copy_len;
636
637err_fault:
638 err = -EFAULT;
639clear:
640 memset(buf, 0, size);
641 return err;
642}
643
644BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
645 u64, flags)
646{
647 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
648}
649
650const struct bpf_func_proto bpf_get_stack_proto = {
651 .func = bpf_get_stack,
652 .gpl_only = true,
653 .ret_type = RET_INTEGER,
654 .arg1_type = ARG_PTR_TO_CTX,
655 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
656 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
657 .arg4_type = ARG_ANYTHING,
658};
659
660BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
661 u32, size, u64, flags)
662{
663 struct pt_regs *regs = task_pt_regs(task);
664
665 return __bpf_get_stack(regs, task, NULL, buf, size, flags);
666}
667
668BTF_ID_LIST(bpf_get_task_stack_btf_ids)
669BTF_ID(struct, task_struct)
670
671const struct bpf_func_proto bpf_get_task_stack_proto = {
672 .func = bpf_get_task_stack,
673 .gpl_only = false,
674 .ret_type = RET_INTEGER,
675 .arg1_type = ARG_PTR_TO_BTF_ID,
676 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
677 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
678 .arg4_type = ARG_ANYTHING,
679 .btf_id = bpf_get_task_stack_btf_ids,
680};
681
682BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
683 void *, buf, u32, size, u64, flags)
684{
685 struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
686 struct perf_event *event = ctx->event;
687 struct perf_callchain_entry *trace;
688 bool kernel, user;
689 int err = -EINVAL;
690 __u64 nr_kernel;
691
692 if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
693 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
694
695 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
696 BPF_F_USER_BUILD_ID)))
697 goto clear;
698
699 user = flags & BPF_F_USER_STACK;
700 kernel = !user;
701
702 err = -EFAULT;
703 trace = ctx->data->callchain;
704 if (unlikely(!trace))
705 goto clear;
706
707 nr_kernel = count_kernel_ip(trace);
708
709 if (kernel) {
710 __u64 nr = trace->nr;
711
712 trace->nr = nr_kernel;
713 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
714
715 /* restore nr */
716 trace->nr = nr;
717 } else { /* user */
718 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
719
720 skip += nr_kernel;
721 if (skip > BPF_F_SKIP_FIELD_MASK)
722 goto clear;
723
724 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
725 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
726 }
727 return err;
728
729clear:
730 memset(buf, 0, size);
731 return err;
732
733}
734
735const struct bpf_func_proto bpf_get_stack_proto_pe = {
736 .func = bpf_get_stack_pe,
737 .gpl_only = true,
738 .ret_type = RET_INTEGER,
739 .arg1_type = ARG_PTR_TO_CTX,
740 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
741 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
742 .arg4_type = ARG_ANYTHING,
743};
744
745/* Called from eBPF program */
746static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
747{
748 return ERR_PTR(-EOPNOTSUPP);
749}
750
751/* Called from syscall */
752int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
753{
754 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
755 struct stack_map_bucket *bucket, *old_bucket;
756 u32 id = *(u32 *)key, trace_len;
757
758 if (unlikely(id >= smap->n_buckets))
759 return -ENOENT;
760
761 bucket = xchg(&smap->buckets[id], NULL);
762 if (!bucket)
763 return -ENOENT;
764
765 trace_len = bucket->nr * stack_map_data_size(map);
766 memcpy(value, bucket->data, trace_len);
767 memset(value + trace_len, 0, map->value_size - trace_len);
768
769 old_bucket = xchg(&smap->buckets[id], bucket);
770 if (old_bucket)
771 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
772 return 0;
773}
774
775static int stack_map_get_next_key(struct bpf_map *map, void *key,
776 void *next_key)
777{
778 struct bpf_stack_map *smap = container_of(map,
779 struct bpf_stack_map, map);
780 u32 id;
781
782 WARN_ON_ONCE(!rcu_read_lock_held());
783
784 if (!key) {
785 id = 0;
786 } else {
787 id = *(u32 *)key;
788 if (id >= smap->n_buckets || !smap->buckets[id])
789 id = 0;
790 else
791 id++;
792 }
793
794 while (id < smap->n_buckets && !smap->buckets[id])
795 id++;
796
797 if (id >= smap->n_buckets)
798 return -ENOENT;
799
800 *(u32 *)next_key = id;
801 return 0;
802}
803
804static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
805 u64 map_flags)
806{
807 return -EINVAL;
808}
809
810/* Called from syscall or from eBPF program */
811static int stack_map_delete_elem(struct bpf_map *map, void *key)
812{
813 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
814 struct stack_map_bucket *old_bucket;
815 u32 id = *(u32 *)key;
816
817 if (unlikely(id >= smap->n_buckets))
818 return -E2BIG;
819
820 old_bucket = xchg(&smap->buckets[id], NULL);
821 if (old_bucket) {
822 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
823 return 0;
824 } else {
825 return -ENOENT;
826 }
827}
828
829/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
830static void stack_map_free(struct bpf_map *map)
831{
832 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
833
834 bpf_map_area_free(smap->elems);
835 pcpu_freelist_destroy(&smap->freelist);
836 bpf_map_area_free(smap);
837 put_callchain_buffers();
838}
839
840static int stack_trace_map_btf_id;
841const struct bpf_map_ops stack_trace_map_ops = {
842 .map_alloc = stack_map_alloc,
843 .map_free = stack_map_free,
844 .map_get_next_key = stack_map_get_next_key,
845 .map_lookup_elem = stack_map_lookup_elem,
846 .map_update_elem = stack_map_update_elem,
847 .map_delete_elem = stack_map_delete_elem,
848 .map_check_btf = map_check_no_btf,
849 .map_btf_name = "bpf_stack_map",
850 .map_btf_id = &stack_trace_map_btf_id,
851};
852
853static int __init stack_map_init(void)
854{
855 int cpu;
856 struct stack_map_irq_work *work;
857
858 for_each_possible_cpu(cpu) {
859 work = per_cpu_ptr(&up_read_work, cpu);
860 init_irq_work(&work->irq_work, do_up_read);
861 }
862 return 0;
863}
864subsys_initcall(stack_map_init);