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/irq_work.h>
11#include <linux/btf_ids.h>
12#include <linux/buildid.h>
13#include "percpu_freelist.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
34/* irq_work to run up_read() for build_id lookup in nmi context */
35struct stack_map_irq_work {
36 struct irq_work irq_work;
37 struct mm_struct *mm;
38};
39
40static void do_up_read(struct irq_work *entry)
41{
42 struct stack_map_irq_work *work;
43
44 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
45 return;
46
47 work = container_of(entry, struct stack_map_irq_work, irq_work);
48 mmap_read_unlock_non_owner(work->mm);
49}
50
51static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
52
53static inline bool stack_map_use_build_id(struct bpf_map *map)
54{
55 return (map->map_flags & BPF_F_STACK_BUILD_ID);
56}
57
58static inline int stack_map_data_size(struct bpf_map *map)
59{
60 return stack_map_use_build_id(map) ?
61 sizeof(struct bpf_stack_build_id) : sizeof(u64);
62}
63
64static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
65{
66 u64 elem_size = sizeof(struct stack_map_bucket) +
67 (u64)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 u64 cost, n_buckets;
94 int err;
95
96 if (!bpf_capable())
97 return ERR_PTR(-EPERM);
98
99 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
100 return ERR_PTR(-EINVAL);
101
102 /* check sanity of attributes */
103 if (attr->max_entries == 0 || attr->key_size != 4 ||
104 value_size < 8 || value_size % 8)
105 return ERR_PTR(-EINVAL);
106
107 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
108 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
109 if (value_size % sizeof(struct bpf_stack_build_id) ||
110 value_size / sizeof(struct bpf_stack_build_id)
111 > sysctl_perf_event_max_stack)
112 return ERR_PTR(-EINVAL);
113 } else if (value_size / 8 > sysctl_perf_event_max_stack)
114 return ERR_PTR(-EINVAL);
115
116 /* hash table size must be power of 2 */
117 n_buckets = roundup_pow_of_two(attr->max_entries);
118 if (!n_buckets)
119 return ERR_PTR(-E2BIG);
120
121 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
122 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
123 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
124 if (!smap)
125 return ERR_PTR(-ENOMEM);
126
127 bpf_map_init_from_attr(&smap->map, attr);
128 smap->map.value_size = value_size;
129 smap->n_buckets = n_buckets;
130
131 err = get_callchain_buffers(sysctl_perf_event_max_stack);
132 if (err)
133 goto free_smap;
134
135 err = prealloc_elems_and_freelist(smap);
136 if (err)
137 goto put_buffers;
138
139 return &smap->map;
140
141put_buffers:
142 put_callchain_buffers();
143free_smap:
144 bpf_map_area_free(smap);
145 return ERR_PTR(err);
146}
147
148static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
149 u64 *ips, u32 trace_nr, bool user)
150{
151 int i;
152 struct vm_area_struct *vma;
153 bool irq_work_busy = false;
154 struct stack_map_irq_work *work = NULL;
155
156 if (irqs_disabled()) {
157 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
158 work = this_cpu_ptr(&up_read_work);
159 if (irq_work_is_busy(&work->irq_work)) {
160 /* cannot queue more up_read, fallback */
161 irq_work_busy = true;
162 }
163 } else {
164 /*
165 * PREEMPT_RT does not allow to trylock mmap sem in
166 * interrupt disabled context. Force the fallback code.
167 */
168 irq_work_busy = true;
169 }
170 }
171
172 /*
173 * We cannot do up_read() when the irq is disabled, because of
174 * risk to deadlock with rq_lock. To do build_id lookup when the
175 * irqs are disabled, we need to run up_read() in irq_work. We use
176 * a percpu variable to do the irq_work. If the irq_work is
177 * already used by another lookup, we fall back to report ips.
178 *
179 * Same fallback is used for kernel stack (!user) on a stackmap
180 * with build_id.
181 */
182 if (!user || !current || !current->mm || irq_work_busy ||
183 !mmap_read_trylock_non_owner(current->mm)) {
184 /* cannot access current->mm, fall back to ips */
185 for (i = 0; i < trace_nr; i++) {
186 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
187 id_offs[i].ip = ips[i];
188 memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
189 }
190 return;
191 }
192
193 for (i = 0; i < trace_nr; i++) {
194 vma = find_vma(current->mm, ips[i]);
195 if (!vma || build_id_parse(vma, id_offs[i].build_id, NULL)) {
196 /* per entry fall back to ips */
197 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
198 id_offs[i].ip = ips[i];
199 memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
200 continue;
201 }
202 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
203 - vma->vm_start;
204 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
205 }
206
207 if (!work) {
208 mmap_read_unlock_non_owner(current->mm);
209 } else {
210 work->mm = current->mm;
211 irq_work_queue(&work->irq_work);
212 }
213}
214
215static struct perf_callchain_entry *
216get_callchain_entry_for_task(struct task_struct *task, u32 init_nr)
217{
218#ifdef CONFIG_STACKTRACE
219 struct perf_callchain_entry *entry;
220 int rctx;
221
222 entry = get_callchain_entry(&rctx);
223
224 if (!entry)
225 return NULL;
226
227 entry->nr = init_nr +
228 stack_trace_save_tsk(task, (unsigned long *)(entry->ip + init_nr),
229 sysctl_perf_event_max_stack - init_nr, 0);
230
231 /* stack_trace_save_tsk() works on unsigned long array, while
232 * perf_callchain_entry uses u64 array. For 32-bit systems, it is
233 * necessary to fix this mismatch.
234 */
235 if (__BITS_PER_LONG != 64) {
236 unsigned long *from = (unsigned long *) entry->ip;
237 u64 *to = entry->ip;
238 int i;
239
240 /* copy data from the end to avoid using extra buffer */
241 for (i = entry->nr - 1; i >= (int)init_nr; i--)
242 to[i] = (u64)(from[i]);
243 }
244
245 put_callchain_entry(rctx);
246
247 return entry;
248#else /* CONFIG_STACKTRACE */
249 return NULL;
250#endif
251}
252
253static long __bpf_get_stackid(struct bpf_map *map,
254 struct perf_callchain_entry *trace, u64 flags)
255{
256 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
257 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
258 u32 max_depth = map->value_size / stack_map_data_size(map);
259 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
260 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
261 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
262 u32 hash, id, trace_nr, trace_len;
263 bool user = flags & BPF_F_USER_STACK;
264 u64 *ips;
265 bool hash_matches;
266
267 /* get_perf_callchain() guarantees that trace->nr >= init_nr
268 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
269 */
270 trace_nr = trace->nr - init_nr;
271
272 if (trace_nr <= skip)
273 /* skipping more than usable stack trace */
274 return -EFAULT;
275
276 trace_nr -= skip;
277 trace_len = trace_nr * sizeof(u64);
278 ips = trace->ip + skip + init_nr;
279 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
280 id = hash & (smap->n_buckets - 1);
281 bucket = READ_ONCE(smap->buckets[id]);
282
283 hash_matches = bucket && bucket->hash == hash;
284 /* fast cmp */
285 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
286 return id;
287
288 if (stack_map_use_build_id(map)) {
289 /* for build_id+offset, pop a bucket before slow cmp */
290 new_bucket = (struct stack_map_bucket *)
291 pcpu_freelist_pop(&smap->freelist);
292 if (unlikely(!new_bucket))
293 return -ENOMEM;
294 new_bucket->nr = trace_nr;
295 stack_map_get_build_id_offset(
296 (struct bpf_stack_build_id *)new_bucket->data,
297 ips, trace_nr, user);
298 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
299 if (hash_matches && bucket->nr == trace_nr &&
300 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
301 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
302 return id;
303 }
304 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
305 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
306 return -EEXIST;
307 }
308 } else {
309 if (hash_matches && bucket->nr == trace_nr &&
310 memcmp(bucket->data, ips, trace_len) == 0)
311 return id;
312 if (bucket && !(flags & BPF_F_REUSE_STACKID))
313 return -EEXIST;
314
315 new_bucket = (struct stack_map_bucket *)
316 pcpu_freelist_pop(&smap->freelist);
317 if (unlikely(!new_bucket))
318 return -ENOMEM;
319 memcpy(new_bucket->data, ips, trace_len);
320 }
321
322 new_bucket->hash = hash;
323 new_bucket->nr = trace_nr;
324
325 old_bucket = xchg(&smap->buckets[id], new_bucket);
326 if (old_bucket)
327 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
328 return id;
329}
330
331BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
332 u64, flags)
333{
334 u32 max_depth = map->value_size / stack_map_data_size(map);
335 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
336 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
337 bool user = flags & BPF_F_USER_STACK;
338 struct perf_callchain_entry *trace;
339 bool kernel = !user;
340
341 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
342 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
343 return -EINVAL;
344
345 trace = get_perf_callchain(regs, init_nr, kernel, user,
346 sysctl_perf_event_max_stack, false, false);
347
348 if (unlikely(!trace))
349 /* couldn't fetch the stack trace */
350 return -EFAULT;
351
352 return __bpf_get_stackid(map, trace, flags);
353}
354
355const struct bpf_func_proto bpf_get_stackid_proto = {
356 .func = bpf_get_stackid,
357 .gpl_only = true,
358 .ret_type = RET_INTEGER,
359 .arg1_type = ARG_PTR_TO_CTX,
360 .arg2_type = ARG_CONST_MAP_PTR,
361 .arg3_type = ARG_ANYTHING,
362};
363
364static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
365{
366 __u64 nr_kernel = 0;
367
368 while (nr_kernel < trace->nr) {
369 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
370 break;
371 nr_kernel++;
372 }
373 return nr_kernel;
374}
375
376BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
377 struct bpf_map *, map, u64, flags)
378{
379 struct perf_event *event = ctx->event;
380 struct perf_callchain_entry *trace;
381 bool kernel, user;
382 __u64 nr_kernel;
383 int ret;
384
385 /* perf_sample_data doesn't have callchain, use bpf_get_stackid */
386 if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
387 return bpf_get_stackid((unsigned long)(ctx->regs),
388 (unsigned long) map, flags, 0, 0);
389
390 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
391 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
392 return -EINVAL;
393
394 user = flags & BPF_F_USER_STACK;
395 kernel = !user;
396
397 trace = ctx->data->callchain;
398 if (unlikely(!trace))
399 return -EFAULT;
400
401 nr_kernel = count_kernel_ip(trace);
402
403 if (kernel) {
404 __u64 nr = trace->nr;
405
406 trace->nr = nr_kernel;
407 ret = __bpf_get_stackid(map, trace, flags);
408
409 /* restore nr */
410 trace->nr = nr;
411 } else { /* user */
412 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
413
414 skip += nr_kernel;
415 if (skip > BPF_F_SKIP_FIELD_MASK)
416 return -EFAULT;
417
418 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
419 ret = __bpf_get_stackid(map, trace, flags);
420 }
421 return ret;
422}
423
424const struct bpf_func_proto bpf_get_stackid_proto_pe = {
425 .func = bpf_get_stackid_pe,
426 .gpl_only = false,
427 .ret_type = RET_INTEGER,
428 .arg1_type = ARG_PTR_TO_CTX,
429 .arg2_type = ARG_CONST_MAP_PTR,
430 .arg3_type = ARG_ANYTHING,
431};
432
433static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
434 struct perf_callchain_entry *trace_in,
435 void *buf, u32 size, u64 flags)
436{
437 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
438 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
439 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
440 bool user = flags & BPF_F_USER_STACK;
441 struct perf_callchain_entry *trace;
442 bool kernel = !user;
443 int err = -EINVAL;
444 u64 *ips;
445
446 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
447 BPF_F_USER_BUILD_ID)))
448 goto clear;
449 if (kernel && user_build_id)
450 goto clear;
451
452 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
453 : sizeof(u64);
454 if (unlikely(size % elem_size))
455 goto clear;
456
457 /* cannot get valid user stack for task without user_mode regs */
458 if (task && user && !user_mode(regs))
459 goto err_fault;
460
461 num_elem = size / elem_size;
462 if (sysctl_perf_event_max_stack < num_elem)
463 init_nr = 0;
464 else
465 init_nr = sysctl_perf_event_max_stack - num_elem;
466
467 if (trace_in)
468 trace = trace_in;
469 else if (kernel && task)
470 trace = get_callchain_entry_for_task(task, init_nr);
471 else
472 trace = get_perf_callchain(regs, init_nr, kernel, user,
473 sysctl_perf_event_max_stack,
474 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
502BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
503 u64, flags)
504{
505 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
506}
507
508const struct bpf_func_proto bpf_get_stack_proto = {
509 .func = bpf_get_stack,
510 .gpl_only = true,
511 .ret_type = RET_INTEGER,
512 .arg1_type = ARG_PTR_TO_CTX,
513 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
514 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
515 .arg4_type = ARG_ANYTHING,
516};
517
518BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
519 u32, size, u64, flags)
520{
521 struct pt_regs *regs;
522 long res;
523
524 if (!try_get_task_stack(task))
525 return -EFAULT;
526
527 regs = task_pt_regs(task);
528 res = __bpf_get_stack(regs, task, NULL, buf, size, flags);
529 put_task_stack(task);
530
531 return res;
532}
533
534BTF_ID_LIST_SINGLE(bpf_get_task_stack_btf_ids, struct, task_struct)
535
536const struct bpf_func_proto bpf_get_task_stack_proto = {
537 .func = bpf_get_task_stack,
538 .gpl_only = false,
539 .ret_type = RET_INTEGER,
540 .arg1_type = ARG_PTR_TO_BTF_ID,
541 .arg1_btf_id = &bpf_get_task_stack_btf_ids[0],
542 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
543 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
544 .arg4_type = ARG_ANYTHING,
545};
546
547BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
548 void *, buf, u32, size, u64, flags)
549{
550 struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
551 struct perf_event *event = ctx->event;
552 struct perf_callchain_entry *trace;
553 bool kernel, user;
554 int err = -EINVAL;
555 __u64 nr_kernel;
556
557 if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
558 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
559
560 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
561 BPF_F_USER_BUILD_ID)))
562 goto clear;
563
564 user = flags & BPF_F_USER_STACK;
565 kernel = !user;
566
567 err = -EFAULT;
568 trace = ctx->data->callchain;
569 if (unlikely(!trace))
570 goto clear;
571
572 nr_kernel = count_kernel_ip(trace);
573
574 if (kernel) {
575 __u64 nr = trace->nr;
576
577 trace->nr = nr_kernel;
578 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
579
580 /* restore nr */
581 trace->nr = nr;
582 } else { /* user */
583 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
584
585 skip += nr_kernel;
586 if (skip > BPF_F_SKIP_FIELD_MASK)
587 goto clear;
588
589 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
590 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
591 }
592 return err;
593
594clear:
595 memset(buf, 0, size);
596 return err;
597
598}
599
600const struct bpf_func_proto bpf_get_stack_proto_pe = {
601 .func = bpf_get_stack_pe,
602 .gpl_only = true,
603 .ret_type = RET_INTEGER,
604 .arg1_type = ARG_PTR_TO_CTX,
605 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
606 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
607 .arg4_type = ARG_ANYTHING,
608};
609
610/* Called from eBPF program */
611static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
612{
613 return ERR_PTR(-EOPNOTSUPP);
614}
615
616/* Called from syscall */
617int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
618{
619 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
620 struct stack_map_bucket *bucket, *old_bucket;
621 u32 id = *(u32 *)key, trace_len;
622
623 if (unlikely(id >= smap->n_buckets))
624 return -ENOENT;
625
626 bucket = xchg(&smap->buckets[id], NULL);
627 if (!bucket)
628 return -ENOENT;
629
630 trace_len = bucket->nr * stack_map_data_size(map);
631 memcpy(value, bucket->data, trace_len);
632 memset(value + trace_len, 0, map->value_size - trace_len);
633
634 old_bucket = xchg(&smap->buckets[id], bucket);
635 if (old_bucket)
636 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
637 return 0;
638}
639
640static int stack_map_get_next_key(struct bpf_map *map, void *key,
641 void *next_key)
642{
643 struct bpf_stack_map *smap = container_of(map,
644 struct bpf_stack_map, map);
645 u32 id;
646
647 WARN_ON_ONCE(!rcu_read_lock_held());
648
649 if (!key) {
650 id = 0;
651 } else {
652 id = *(u32 *)key;
653 if (id >= smap->n_buckets || !smap->buckets[id])
654 id = 0;
655 else
656 id++;
657 }
658
659 while (id < smap->n_buckets && !smap->buckets[id])
660 id++;
661
662 if (id >= smap->n_buckets)
663 return -ENOENT;
664
665 *(u32 *)next_key = id;
666 return 0;
667}
668
669static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
670 u64 map_flags)
671{
672 return -EINVAL;
673}
674
675/* Called from syscall or from eBPF program */
676static int stack_map_delete_elem(struct bpf_map *map, void *key)
677{
678 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
679 struct stack_map_bucket *old_bucket;
680 u32 id = *(u32 *)key;
681
682 if (unlikely(id >= smap->n_buckets))
683 return -E2BIG;
684
685 old_bucket = xchg(&smap->buckets[id], NULL);
686 if (old_bucket) {
687 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
688 return 0;
689 } else {
690 return -ENOENT;
691 }
692}
693
694/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
695static void stack_map_free(struct bpf_map *map)
696{
697 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
698
699 bpf_map_area_free(smap->elems);
700 pcpu_freelist_destroy(&smap->freelist);
701 bpf_map_area_free(smap);
702 put_callchain_buffers();
703}
704
705static int stack_trace_map_btf_id;
706const struct bpf_map_ops stack_trace_map_ops = {
707 .map_meta_equal = bpf_map_meta_equal,
708 .map_alloc = stack_map_alloc,
709 .map_free = stack_map_free,
710 .map_get_next_key = stack_map_get_next_key,
711 .map_lookup_elem = stack_map_lookup_elem,
712 .map_update_elem = stack_map_update_elem,
713 .map_delete_elem = stack_map_delete_elem,
714 .map_check_btf = map_check_no_btf,
715 .map_btf_name = "bpf_stack_map",
716 .map_btf_id = &stack_trace_map_btf_id,
717};
718
719static int __init stack_map_init(void)
720{
721 int cpu;
722 struct stack_map_irq_work *work;
723
724 for_each_possible_cpu(cpu) {
725 work = per_cpu_ptr(&up_read_work, cpu);
726 init_irq_work(&work->irq_work, do_up_read);
727 }
728 return 0;
729}
730subsys_initcall(stack_map_init);
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);