1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Facebook
5 #include <linux/jhash.h>
6 #include <linux/filter.h>
7 #include <linux/stacktrace.h>
8 #include <linux/perf_event.h>
10 #include <linux/pagemap.h>
11 #include <linux/irq_work.h>
12 #include "percpu_freelist.h"
14 #define STACK_CREATE_FLAG_MASK \
15 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY | \
18 struct stack_map_bucket {
19 struct pcpu_freelist_node fnode;
25 struct bpf_stack_map {
28 struct pcpu_freelist freelist;
30 struct stack_map_bucket *buckets[];
33 /* irq_work to run up_read() for build_id lookup in nmi context */
34 struct stack_map_irq_work {
35 struct irq_work irq_work;
36 struct rw_semaphore *sem;
39 static void do_up_read(struct irq_work *entry)
41 struct stack_map_irq_work *work;
43 work = container_of(entry, struct stack_map_irq_work, irq_work);
44 up_read_non_owner(work->sem);
48 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
50 static inline bool stack_map_use_build_id(struct bpf_map *map)
52 return (map->map_flags & BPF_F_STACK_BUILD_ID);
55 static inline int stack_map_data_size(struct bpf_map *map)
57 return stack_map_use_build_id(map) ?
58 sizeof(struct bpf_stack_build_id) : sizeof(u64);
61 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
63 u64 elem_size = sizeof(struct stack_map_bucket) +
64 (u64)smap->map.value_size;
67 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
72 err = pcpu_freelist_init(&smap->freelist);
76 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
77 smap->map.max_entries);
81 bpf_map_area_free(smap->elems);
85 /* Called from syscall */
86 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
88 u32 value_size = attr->value_size;
89 struct bpf_stack_map *smap;
90 struct bpf_map_memory mem;
94 if (!capable(CAP_SYS_ADMIN))
95 return ERR_PTR(-EPERM);
97 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
98 return ERR_PTR(-EINVAL);
100 /* check sanity of attributes */
101 if (attr->max_entries == 0 || attr->key_size != 4 ||
102 value_size < 8 || value_size % 8)
103 return ERR_PTR(-EINVAL);
105 BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
106 if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
107 if (value_size % sizeof(struct bpf_stack_build_id) ||
108 value_size / sizeof(struct bpf_stack_build_id)
109 > sysctl_perf_event_max_stack)
110 return ERR_PTR(-EINVAL);
111 } else if (value_size / 8 > sysctl_perf_event_max_stack)
112 return ERR_PTR(-EINVAL);
114 /* hash table size must be power of 2 */
115 n_buckets = roundup_pow_of_two(attr->max_entries);
117 return ERR_PTR(-E2BIG);
119 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
120 err = bpf_map_charge_init(&mem, cost + attr->max_entries *
121 (sizeof(struct stack_map_bucket) + (u64)value_size));
125 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
127 bpf_map_charge_finish(&mem);
128 return ERR_PTR(-ENOMEM);
131 bpf_map_init_from_attr(&smap->map, attr);
132 smap->map.value_size = value_size;
133 smap->n_buckets = n_buckets;
135 err = get_callchain_buffers(sysctl_perf_event_max_stack);
139 err = prealloc_elems_and_freelist(smap);
143 bpf_map_charge_move(&smap->map.memory, &mem);
148 put_callchain_buffers();
150 bpf_map_charge_finish(&mem);
151 bpf_map_area_free(smap);
155 #define BPF_BUILD_ID 3
157 * Parse build id from the note segment. This logic can be shared between
158 * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
161 static inline int stack_map_parse_build_id(void *page_addr,
162 unsigned char *build_id,
164 Elf32_Word note_size)
166 Elf32_Word note_offs = 0, new_offs;
168 /* check for overflow */
169 if (note_start < page_addr || note_start + note_size < note_start)
172 /* only supports note that fits in the first page */
173 if (note_start + note_size > page_addr + PAGE_SIZE)
176 while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
177 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
179 if (nhdr->n_type == BPF_BUILD_ID &&
180 nhdr->n_namesz == sizeof("GNU") &&
181 nhdr->n_descsz > 0 &&
182 nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
184 note_start + note_offs +
185 ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
187 memset(build_id + nhdr->n_descsz, 0,
188 BPF_BUILD_ID_SIZE - nhdr->n_descsz);
191 new_offs = note_offs + sizeof(Elf32_Nhdr) +
192 ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
193 if (new_offs <= note_offs) /* overflow */
195 note_offs = new_offs;
200 /* Parse build ID from 32-bit ELF */
201 static int stack_map_get_build_id_32(void *page_addr,
202 unsigned char *build_id)
204 Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
208 /* only supports phdr that fits in one page */
210 (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
213 phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
215 for (i = 0; i < ehdr->e_phnum; ++i)
216 if (phdr[i].p_type == PT_NOTE)
217 return stack_map_parse_build_id(page_addr, build_id,
218 page_addr + phdr[i].p_offset,
223 /* Parse build ID from 64-bit ELF */
224 static int stack_map_get_build_id_64(void *page_addr,
225 unsigned char *build_id)
227 Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
231 /* only supports phdr that fits in one page */
233 (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
236 phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
238 for (i = 0; i < ehdr->e_phnum; ++i)
239 if (phdr[i].p_type == PT_NOTE)
240 return stack_map_parse_build_id(page_addr, build_id,
241 page_addr + phdr[i].p_offset,
246 /* Parse build ID of ELF file mapped to vma */
247 static int stack_map_get_build_id(struct vm_area_struct *vma,
248 unsigned char *build_id)
255 /* only works for page backed storage */
259 page = find_get_page(vma->vm_file->f_mapping, 0);
261 return -EFAULT; /* page not mapped */
264 page_addr = kmap_atomic(page);
265 ehdr = (Elf32_Ehdr *)page_addr;
267 /* compare magic x7f "ELF" */
268 if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
271 /* only support executable file and shared object file */
272 if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
275 if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
276 ret = stack_map_get_build_id_32(page_addr, build_id);
277 else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
278 ret = stack_map_get_build_id_64(page_addr, build_id);
280 kunmap_atomic(page_addr);
285 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
286 u64 *ips, u32 trace_nr, bool user)
289 struct vm_area_struct *vma;
290 bool irq_work_busy = false;
291 struct stack_map_irq_work *work = NULL;
293 if (irqs_disabled()) {
294 work = this_cpu_ptr(&up_read_work);
295 if (work->irq_work.flags & IRQ_WORK_BUSY)
296 /* cannot queue more up_read, fallback */
297 irq_work_busy = true;
301 * We cannot do up_read() when the irq is disabled, because of
302 * risk to deadlock with rq_lock. To do build_id lookup when the
303 * irqs are disabled, we need to run up_read() in irq_work. We use
304 * a percpu variable to do the irq_work. If the irq_work is
305 * already used by another lookup, we fall back to report ips.
307 * Same fallback is used for kernel stack (!user) on a stackmap
310 if (!user || !current || !current->mm || irq_work_busy ||
311 down_read_trylock(¤t->mm->mmap_sem) == 0) {
312 /* cannot access current->mm, fall back to ips */
313 for (i = 0; i < trace_nr; i++) {
314 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
315 id_offs[i].ip = ips[i];
316 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
321 for (i = 0; i < trace_nr; i++) {
322 vma = find_vma(current->mm, ips[i]);
323 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
324 /* per entry fall back to ips */
325 id_offs[i].status = BPF_STACK_BUILD_ID_IP;
326 id_offs[i].ip = ips[i];
327 memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
330 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
332 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
336 up_read(¤t->mm->mmap_sem);
338 work->sem = ¤t->mm->mmap_sem;
339 irq_work_queue(&work->irq_work);
341 * The irq_work will release the mmap_sem with
342 * up_read_non_owner(). The rwsem_release() is called
343 * here to release the lock from lockdep's perspective.
345 rwsem_release(¤t->mm->mmap_sem.dep_map, 1, _RET_IP_);
349 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
352 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
353 struct perf_callchain_entry *trace;
354 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
355 u32 max_depth = map->value_size / stack_map_data_size(map);
356 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
357 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
358 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
359 u32 hash, id, trace_nr, trace_len;
360 bool user = flags & BPF_F_USER_STACK;
365 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
366 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
369 trace = get_perf_callchain(regs, init_nr, kernel, user,
370 sysctl_perf_event_max_stack, false, false);
372 if (unlikely(!trace))
373 /* couldn't fetch the stack trace */
376 /* get_perf_callchain() guarantees that trace->nr >= init_nr
377 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
379 trace_nr = trace->nr - init_nr;
381 if (trace_nr <= skip)
382 /* skipping more than usable stack trace */
386 trace_len = trace_nr * sizeof(u64);
387 ips = trace->ip + skip + init_nr;
388 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
389 id = hash & (smap->n_buckets - 1);
390 bucket = READ_ONCE(smap->buckets[id]);
392 hash_matches = bucket && bucket->hash == hash;
394 if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
397 if (stack_map_use_build_id(map)) {
398 /* for build_id+offset, pop a bucket before slow cmp */
399 new_bucket = (struct stack_map_bucket *)
400 pcpu_freelist_pop(&smap->freelist);
401 if (unlikely(!new_bucket))
403 new_bucket->nr = trace_nr;
404 stack_map_get_build_id_offset(
405 (struct bpf_stack_build_id *)new_bucket->data,
406 ips, trace_nr, user);
407 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
408 if (hash_matches && bucket->nr == trace_nr &&
409 memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
410 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
413 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
414 pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
418 if (hash_matches && bucket->nr == trace_nr &&
419 memcmp(bucket->data, ips, trace_len) == 0)
421 if (bucket && !(flags & BPF_F_REUSE_STACKID))
424 new_bucket = (struct stack_map_bucket *)
425 pcpu_freelist_pop(&smap->freelist);
426 if (unlikely(!new_bucket))
428 memcpy(new_bucket->data, ips, trace_len);
431 new_bucket->hash = hash;
432 new_bucket->nr = trace_nr;
434 old_bucket = xchg(&smap->buckets[id], new_bucket);
436 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
440 const struct bpf_func_proto bpf_get_stackid_proto = {
441 .func = bpf_get_stackid,
443 .ret_type = RET_INTEGER,
444 .arg1_type = ARG_PTR_TO_CTX,
445 .arg2_type = ARG_CONST_MAP_PTR,
446 .arg3_type = ARG_ANYTHING,
449 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
452 u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
453 bool user_build_id = flags & BPF_F_USER_BUILD_ID;
454 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
455 bool user = flags & BPF_F_USER_STACK;
456 struct perf_callchain_entry *trace;
461 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
462 BPF_F_USER_BUILD_ID)))
464 if (kernel && user_build_id)
467 elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
469 if (unlikely(size % elem_size))
472 num_elem = size / elem_size;
473 if (sysctl_perf_event_max_stack < num_elem)
476 init_nr = sysctl_perf_event_max_stack - num_elem;
477 trace = get_perf_callchain(regs, init_nr, kernel, user,
478 sysctl_perf_event_max_stack, false, false);
479 if (unlikely(!trace))
482 trace_nr = trace->nr - init_nr;
487 trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
488 copy_len = trace_nr * elem_size;
489 ips = trace->ip + skip + init_nr;
490 if (user && user_build_id)
491 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
493 memcpy(buf, ips, copy_len);
496 memset(buf + copy_len, 0, size - copy_len);
502 memset(buf, 0, size);
506 const struct bpf_func_proto bpf_get_stack_proto = {
507 .func = bpf_get_stack,
509 .ret_type = RET_INTEGER,
510 .arg1_type = ARG_PTR_TO_CTX,
511 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
512 .arg3_type = ARG_CONST_SIZE_OR_ZERO,
513 .arg4_type = ARG_ANYTHING,
516 /* Called from eBPF program */
517 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
519 return ERR_PTR(-EOPNOTSUPP);
522 /* Called from syscall */
523 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
525 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
526 struct stack_map_bucket *bucket, *old_bucket;
527 u32 id = *(u32 *)key, trace_len;
529 if (unlikely(id >= smap->n_buckets))
532 bucket = xchg(&smap->buckets[id], NULL);
536 trace_len = bucket->nr * stack_map_data_size(map);
537 memcpy(value, bucket->data, trace_len);
538 memset(value + trace_len, 0, map->value_size - trace_len);
540 old_bucket = xchg(&smap->buckets[id], bucket);
542 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
546 static int stack_map_get_next_key(struct bpf_map *map, void *key,
549 struct bpf_stack_map *smap = container_of(map,
550 struct bpf_stack_map, map);
553 WARN_ON_ONCE(!rcu_read_lock_held());
559 if (id >= smap->n_buckets || !smap->buckets[id])
565 while (id < smap->n_buckets && !smap->buckets[id])
568 if (id >= smap->n_buckets)
571 *(u32 *)next_key = id;
575 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
581 /* Called from syscall or from eBPF program */
582 static int stack_map_delete_elem(struct bpf_map *map, void *key)
584 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
585 struct stack_map_bucket *old_bucket;
586 u32 id = *(u32 *)key;
588 if (unlikely(id >= smap->n_buckets))
591 old_bucket = xchg(&smap->buckets[id], NULL);
593 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
600 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
601 static void stack_map_free(struct bpf_map *map)
603 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
605 /* wait for bpf programs to complete before freeing stack map */
608 bpf_map_area_free(smap->elems);
609 pcpu_freelist_destroy(&smap->freelist);
610 bpf_map_area_free(smap);
611 put_callchain_buffers();
614 const struct bpf_map_ops stack_trace_map_ops = {
615 .map_alloc = stack_map_alloc,
616 .map_free = stack_map_free,
617 .map_get_next_key = stack_map_get_next_key,
618 .map_lookup_elem = stack_map_lookup_elem,
619 .map_update_elem = stack_map_update_elem,
620 .map_delete_elem = stack_map_delete_elem,
621 .map_check_btf = map_check_no_btf,
624 static int __init stack_map_init(void)
627 struct stack_map_irq_work *work;
629 for_each_possible_cpu(cpu) {
630 work = per_cpu_ptr(&up_read_work, cpu);
631 init_irq_work(&work->irq_work, do_up_read);
635 subsys_initcall(stack_map_init);