projects/linux·blob v5.8-rc5 · refs

637 lines · 16.6 KiB

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 mm_struct *mm;
37};
38
39static void do_up_read(struct irq_work *entry)
40{
41 struct stack_map_irq_work *work;
42
43 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
44 return;
45
46 work = container_of(entry, struct stack_map_irq_work, irq_work);
47 mmap_read_unlock_non_owner(work->mm);
48}
49
50static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
51
52static inline bool stack_map_use_build_id(struct bpf_map *map)
53{
54 return (map->map_flags & BPF_F_STACK_BUILD_ID);
55}
56
57static inline int stack_map_data_size(struct bpf_map *map)
58{
59 return stack_map_use_build_id(map) ?
60 sizeof(struct bpf_stack_build_id) : sizeof(u64);
61}
62
63static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
64{
65 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
66 int err;
67
68 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
69 smap->map.numa_node);
70 if (!smap->elems)
71 return -ENOMEM;
72
73 err = pcpu_freelist_init(&smap->freelist);
74 if (err)
75 goto free_elems;
76
77 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
78 smap->map.max_entries);
79 return 0;
80
81free_elems:
82 bpf_map_area_free(smap->elems);
83 return err;
84}
85
86/* Called from syscall */
87static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
88{
89 u32 value_size = attr->value_size;
90 struct bpf_stack_map *smap;
91 struct bpf_map_memory mem;
92 u64 cost, n_buckets;
93 int err;
94
95 if (!bpf_capable())
96 return ERR_PTR(-EPERM);
97
98 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
99 return ERR_PTR(-EINVAL);
100
101 /* check sanity of attributes */
102 if (attr->max_entries == 0 || attr->key_size != 4 ||
103 value_size < 8 || value_size % 8)
104 return ERR_PTR(-EINVAL);
105
106 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
107 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
108 if (value_size % sizeof(struct bpf_stack_build_id) ||
109 value_size / sizeof(struct bpf_stack_build_id)
110 > sysctl_perf_event_max_stack)
111 return ERR_PTR(-EINVAL);
112 } else if (value_size / 8 > sysctl_perf_event_max_stack)
113 return ERR_PTR(-EINVAL);
114
115 /* hash table size must be power of 2 */
116 n_buckets = roundup_pow_of_two(attr->max_entries);
117
118 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
119 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
120 err = bpf_map_charge_init(&mem, cost);
121 if (err)
122 return ERR_PTR(err);
123
124 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
125 if (!smap) {
126 bpf_map_charge_finish(&mem);
127 return ERR_PTR(-ENOMEM);
128 }
129
130 bpf_map_init_from_attr(&smap->map, attr);
131 smap->map.value_size = value_size;
132 smap->n_buckets = n_buckets;
133
134 err = get_callchain_buffers(sysctl_perf_event_max_stack);
135 if (err)
136 goto free_charge;
137
138 err = prealloc_elems_and_freelist(smap);
139 if (err)
140 goto put_buffers;
141
142 bpf_map_charge_move(&smap->map.memory, &mem);
143
144 return &smap->map;
145
146put_buffers:
147 put_callchain_buffers();
148free_charge:
149 bpf_map_charge_finish(&mem);
150 bpf_map_area_free(smap);
151 return ERR_PTR(err);
152}
153
154#define BPF_BUILD_ID 3
155/*
156 * Parse build id from the note segment. This logic can be shared between
157 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
158 * identical.
159 */
160static inline int stack_map_parse_build_id(void *page_addr,
161 unsigned char *build_id,
162 void *note_start,
163 Elf32_Word note_size)
164{
165 Elf32_Word note_offs = 0, new_offs;
166
167 /* check for overflow */
168 if (note_start < page_addr || note_start + note_size < note_start)
169 return -EINVAL;
170
171 /* only supports note that fits in the first page */
172 if (note_start + note_size > page_addr + PAGE_SIZE)
173 return -EINVAL;
174
175 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
176 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
177
178 if (nhdr->n_type == BPF_BUILD_ID &&
179 nhdr->n_namesz == sizeof("GNU") &&
180 nhdr->n_descsz > 0 &&
181 nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
182 memcpy(build_id,
183 note_start + note_offs +
184 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
185 nhdr->n_descsz);
186 memset(build_id + nhdr->n_descsz, 0,
187 BPF_BUILD_ID_SIZE - nhdr->n_descsz);
188 return 0;
189 }
190 new_offs = note_offs + sizeof(Elf32_Nhdr) +
191 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
192 if (new_offs <= note_offs) /* overflow */
193 break;
194 note_offs = new_offs;
195 }
196 return -EINVAL;
197}
198
199/* Parse build ID from 32-bit ELF */
200static int stack_map_get_build_id_32(void *page_addr,
201 unsigned char *build_id)
202{
203 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
204 Elf32_Phdr *phdr;
205 int i;
206
207 /* only supports phdr that fits in one page */
208 if (ehdr->e_phnum >
209 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
210 return -EINVAL;
211
212 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
213
214 for (i = 0; i < ehdr->e_phnum; ++i)
215 if (phdr[i].p_type == PT_NOTE)
216 return stack_map_parse_build_id(page_addr, build_id,
217 page_addr + phdr[i].p_offset,
218 phdr[i].p_filesz);
219 return -EINVAL;
220}
221
222/* Parse build ID from 64-bit ELF */
223static int stack_map_get_build_id_64(void *page_addr,
224 unsigned char *build_id)
225{
226 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
227 Elf64_Phdr *phdr;
228 int i;
229
230 /* only supports phdr that fits in one page */
231 if (ehdr->e_phnum >
232 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
233 return -EINVAL;
234
235 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
236
237 for (i = 0; i < ehdr->e_phnum; ++i)
238 if (phdr[i].p_type == PT_NOTE)
239 return stack_map_parse_build_id(page_addr, build_id,
240 page_addr + phdr[i].p_offset,
241 phdr[i].p_filesz);
242 return -EINVAL;
243}
244
245/* Parse build ID of ELF file mapped to vma */
246static int stack_map_get_build_id(struct vm_area_struct *vma,
247 unsigned char *build_id)
248{
249 Elf32_Ehdr *ehdr;
250 struct page *page;
251 void *page_addr;
252 int ret;
253
254 /* only works for page backed storage */
255 if (!vma->vm_file)
256 return -EINVAL;
257
258 page = find_get_page(vma->vm_file->f_mapping, 0);
259 if (!page)
260 return -EFAULT; /* page not mapped */
261
262 ret = -EINVAL;
263 page_addr = kmap_atomic(page);
264 ehdr = (Elf32_Ehdr *)page_addr;
265
266 /* compare magic x7f "ELF" */
267 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
268 goto out;
269
270 /* only support executable file and shared object file */
271 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
272 goto out;
273
274 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
275 ret = stack_map_get_build_id_32(page_addr, build_id);
276 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
277 ret = stack_map_get_build_id_64(page_addr, build_id);
278out:
279 kunmap_atomic(page_addr);
280 put_page(page);
281 return ret;
282}
283
284static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
285 u64 *ips, u32 trace_nr, bool user)
286{
287 int i;
288 struct vm_area_struct *vma;
289 bool irq_work_busy = false;
290 struct stack_map_irq_work *work = NULL;
291
292 if (irqs_disabled()) {
293 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
294 work = this_cpu_ptr(&up_read_work);
295 if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY) {
296 /* cannot queue more up_read, fallback */
297 irq_work_busy = true;
298 }
299 } else {
300 /*
301 * PREEMPT_RT does not allow to trylock mmap sem in
302 * interrupt disabled context. Force the fallback code.
303 */
304 irq_work_busy = true;
305 }
306 }
307
308 /*
309 * We cannot do up_read() when the irq is disabled, because of
310 * risk to deadlock with rq_lock. To do build_id lookup when the
311 * irqs are disabled, we need to run up_read() in irq_work. We use
312 * a percpu variable to do the irq_work. If the irq_work is
313 * already used by another lookup, we fall back to report ips.
314 *
315 * Same fallback is used for kernel stack (!user) on a stackmap
316 * with build_id.
317 */
318 if (!user || !current || !current->mm || irq_work_busy ||
319 !mmap_read_trylock_non_owner(current->mm)) {
320 /* cannot access current->mm, fall back to ips */
321 for (i = 0; i < trace_nr; i++) {
322 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
323 id_offs[i].ip = ips[i];
324 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
325 }
326 return;
327 }
328
329 for (i = 0; i < trace_nr; i++) {
330 vma = find_vma(current->mm, ips[i]);
331 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
332 /* per entry fall back to ips */
333 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
334 id_offs[i].ip = ips[i];
335 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
336 continue;
337 }
338 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
339 - vma->vm_start;
340 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
341 }
342
343 if (!work) {
344 mmap_read_unlock_non_owner(current->mm);
345 } else {
346 work->mm = current->mm;
347 irq_work_queue(&work->irq_work);
348 }
349}
350
351BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
352 u64, flags)
353{
354 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
355 struct perf_callchain_entry *trace;
356 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
357 u32 max_depth = map->value_size / stack_map_data_size(map);
358 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
359 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
360 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
361 u32 hash, id, trace_nr, trace_len;
362 bool user = flags & BPF_F_USER_STACK;
363 bool kernel = !user;
364 u64 *ips;
365 bool hash_matches;
366
367 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
368 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
369 return -EINVAL;
370
371 trace = get_perf_callchain(regs, init_nr, kernel, user,
372 sysctl_perf_event_max_stack, false, false);
373
374 if (unlikely(!trace))
375 /* couldn't fetch the stack trace */
376 return -EFAULT;
377
378 /* get_perf_callchain() guarantees that trace->nr >= init_nr
379 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
380 */
381 trace_nr = trace->nr - init_nr;
382
383 if (trace_nr <= skip)
384 /* skipping more than usable stack trace */
385 return -EFAULT;
386
387 trace_nr -= skip;
388 trace_len = trace_nr * sizeof(u64);
389 ips = trace->ip + skip + init_nr;
390 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
391 id = hash & (smap->n_buckets - 1);
392 bucket = READ_ONCE(smap->buckets[id]);
393
394 hash_matches = bucket && bucket->hash == hash;
395 /* fast cmp */
396 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
397 return id;
398
399 if (stack_map_use_build_id(map)) {
400 /* for build_id+offset, pop a bucket before slow cmp */
401 new_bucket = (struct stack_map_bucket *)
402 pcpu_freelist_pop(&smap->freelist);
403 if (unlikely(!new_bucket))
404 return -ENOMEM;
405 new_bucket->nr = trace_nr;
406 stack_map_get_build_id_offset(
407 (struct bpf_stack_build_id *)new_bucket->data,
408 ips, trace_nr, user);
409 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
410 if (hash_matches && bucket->nr == trace_nr &&
411 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
412 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
413 return id;
414 }
415 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
416 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
417 return -EEXIST;
418 }
419 } else {
420 if (hash_matches && bucket->nr == trace_nr &&
421 memcmp(bucket->data, ips, trace_len) == 0)
422 return id;
423 if (bucket && !(flags & BPF_F_REUSE_STACKID))
424 return -EEXIST;
425
426 new_bucket = (struct stack_map_bucket *)
427 pcpu_freelist_pop(&smap->freelist);
428 if (unlikely(!new_bucket))
429 return -ENOMEM;
430 memcpy(new_bucket->data, ips, trace_len);
431 }
432
433 new_bucket->hash = hash;
434 new_bucket->nr = trace_nr;
435
436 old_bucket = xchg(&smap->buckets[id], new_bucket);
437 if (old_bucket)
438 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
439 return id;
440}
441
442const struct bpf_func_proto bpf_get_stackid_proto = {
443 .func = bpf_get_stackid,
444 .gpl_only = true,
445 .ret_type = RET_INTEGER,
446 .arg1_type = ARG_PTR_TO_CTX,
447 .arg2_type = ARG_CONST_MAP_PTR,
448 .arg3_type = ARG_ANYTHING,
449};
450
451BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
452 u64, flags)
453{
454 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
455 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
456 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
457 bool user = flags & BPF_F_USER_STACK;
458 struct perf_callchain_entry *trace;
459 bool kernel = !user;
460 int err = -EINVAL;
461 u64 *ips;
462
463 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
464 BPF_F_USER_BUILD_ID)))
465 goto clear;
466 if (kernel && user_build_id)
467 goto clear;
468
469 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
470 : sizeof(u64);
471 if (unlikely(size % elem_size))
472 goto clear;
473
474 num_elem = size / elem_size;
475 if (sysctl_perf_event_max_stack < num_elem)
476 init_nr = 0;
477 else
478 init_nr = sysctl_perf_event_max_stack - num_elem;
479 trace = get_perf_callchain(regs, init_nr, kernel, user,
480 sysctl_perf_event_max_stack, false, false);
481 if (unlikely(!trace))
482 goto err_fault;
483
484 trace_nr = trace->nr - init_nr;
485 if (trace_nr < skip)
486 goto err_fault;
487
488 trace_nr -= skip;
489 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
490 copy_len = trace_nr * elem_size;
491 ips = trace->ip + skip + init_nr;
492 if (user && user_build_id)
493 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
494 else
495 memcpy(buf, ips, copy_len);
496
497 if (size > copy_len)
498 memset(buf + copy_len, 0, size - copy_len);
499 return copy_len;
500
501err_fault:
502 err = -EFAULT;
503clear:
504 memset(buf, 0, size);
505 return err;
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
518/* Called from eBPF program */
519static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
520{
521 return ERR_PTR(-EOPNOTSUPP);
522}
523
524/* Called from syscall */
525int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
526{
527 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
528 struct stack_map_bucket *bucket, *old_bucket;
529 u32 id = *(u32 *)key, trace_len;
530
531 if (unlikely(id >= smap->n_buckets))
532 return -ENOENT;
533
534 bucket = xchg(&smap->buckets[id], NULL);
535 if (!bucket)
536 return -ENOENT;
537
538 trace_len = bucket->nr * stack_map_data_size(map);
539 memcpy(value, bucket->data, trace_len);
540 memset(value + trace_len, 0, map->value_size - trace_len);
541
542 old_bucket = xchg(&smap->buckets[id], bucket);
543 if (old_bucket)
544 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
545 return 0;
546}
547
548static int stack_map_get_next_key(struct bpf_map *map, void *key,
549 void *next_key)
550{
551 struct bpf_stack_map *smap = container_of(map,
552 struct bpf_stack_map, map);
553 u32 id;
554
555 WARN_ON_ONCE(!rcu_read_lock_held());
556
557 if (!key) {
558 id = 0;
559 } else {
560 id = *(u32 *)key;
561 if (id >= smap->n_buckets || !smap->buckets[id])
562 id = 0;
563 else
564 id++;
565 }
566
567 while (id < smap->n_buckets && !smap->buckets[id])
568 id++;
569
570 if (id >= smap->n_buckets)
571 return -ENOENT;
572
573 *(u32 *)next_key = id;
574 return 0;
575}
576
577static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
578 u64 map_flags)
579{
580 return -EINVAL;
581}
582
583/* Called from syscall or from eBPF program */
584static int stack_map_delete_elem(struct bpf_map *map, void *key)
585{
586 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
587 struct stack_map_bucket *old_bucket;
588 u32 id = *(u32 *)key;
589
590 if (unlikely(id >= smap->n_buckets))
591 return -E2BIG;
592
593 old_bucket = xchg(&smap->buckets[id], NULL);
594 if (old_bucket) {
595 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
596 return 0;
597 } else {
598 return -ENOENT;
599 }
600}
601
602/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
603static void stack_map_free(struct bpf_map *map)
604{
605 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
606
607 /* wait for bpf programs to complete before freeing stack map */
608 synchronize_rcu();
609
610 bpf_map_area_free(smap->elems);
611 pcpu_freelist_destroy(&smap->freelist);
612 bpf_map_area_free(smap);
613 put_callchain_buffers();
614}
615
616const struct bpf_map_ops stack_trace_map_ops = {
617 .map_alloc = stack_map_alloc,
618 .map_free = stack_map_free,
619 .map_get_next_key = stack_map_get_next_key,
620 .map_lookup_elem = stack_map_lookup_elem,
621 .map_update_elem = stack_map_update_elem,
622 .map_delete_elem = stack_map_delete_elem,
623 .map_check_btf = map_check_no_btf,
624};
625
626static int __init stack_map_init(void)
627{
628 int cpu;
629 struct stack_map_irq_work *work;
630
631 for_each_possible_cpu(cpu) {
632 work = per_cpu_ptr(&up_read_work, cpu);
633 init_irq_work(&work->irq_work, do_up_read);
634 }
635 return 0;
636}
637subsys_initcall(stack_map_init);