projects/linux·blob v4.15-rc6 · refs

280 lines · 7.3 KiB

1/* Copyright (c) 2016 Facebook
2 *
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of version 2 of the GNU General Public
5 * License as published by the Free Software Foundation.
6 */
7#include <linux/bpf.h>
8#include <linux/jhash.h>
9#include <linux/filter.h>
10#include <linux/stacktrace.h>
11#include <linux/perf_event.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
17struct stack_map_bucket {
18 struct pcpu_freelist_node fnode;
19 u32 hash;
20 u32 nr;
21 u64 ip[];
22};
23
24struct bpf_stack_map {
25 struct bpf_map map;
26 void *elems;
27 struct pcpu_freelist freelist;
28 u32 n_buckets;
29 struct stack_map_bucket *buckets[];
30};
31
32static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
33{
34 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
35 int err;
36
37 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
38 smap->map.numa_node);
39 if (!smap->elems)
40 return -ENOMEM;
41
42 err = pcpu_freelist_init(&smap->freelist);
43 if (err)
44 goto free_elems;
45
46 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
47 smap->map.max_entries);
48 return 0;
49
50free_elems:
51 bpf_map_area_free(smap->elems);
52 return err;
53}
54
55/* Called from syscall */
56static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
57{
58 u32 value_size = attr->value_size;
59 struct bpf_stack_map *smap;
60 u64 cost, n_buckets;
61 int err;
62
63 if (!capable(CAP_SYS_ADMIN))
64 return ERR_PTR(-EPERM);
65
66 if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
67 return ERR_PTR(-EINVAL);
68
69 /* check sanity of attributes */
70 if (attr->max_entries == 0 || attr->key_size != 4 ||
71 value_size < 8 || value_size % 8 ||
72 value_size / 8 > sysctl_perf_event_max_stack)
73 return ERR_PTR(-EINVAL);
74
75 /* hash table size must be power of 2 */
76 n_buckets = roundup_pow_of_two(attr->max_entries);
77
78 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
79 if (cost >= U32_MAX - PAGE_SIZE)
80 return ERR_PTR(-E2BIG);
81
82 smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
83 if (!smap)
84 return ERR_PTR(-ENOMEM);
85
86 err = -E2BIG;
87 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
88 if (cost >= U32_MAX - PAGE_SIZE)
89 goto free_smap;
90
91 smap->map.map_type = attr->map_type;
92 smap->map.key_size = attr->key_size;
93 smap->map.value_size = value_size;
94 smap->map.max_entries = attr->max_entries;
95 smap->map.map_flags = attr->map_flags;
96 smap->n_buckets = n_buckets;
97 smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
98 smap->map.numa_node = bpf_map_attr_numa_node(attr);
99
100 err = bpf_map_precharge_memlock(smap->map.pages);
101 if (err)
102 goto free_smap;
103
104 err = get_callchain_buffers(sysctl_perf_event_max_stack);
105 if (err)
106 goto free_smap;
107
108 err = prealloc_elems_and_freelist(smap);
109 if (err)
110 goto put_buffers;
111
112 return &smap->map;
113
114put_buffers:
115 put_callchain_buffers();
116free_smap:
117 bpf_map_area_free(smap);
118 return ERR_PTR(err);
119}
120
121BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
122 u64, flags)
123{
124 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
125 struct perf_callchain_entry *trace;
126 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
127 u32 max_depth = map->value_size / 8;
128 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
129 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
130 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
131 u32 hash, id, trace_nr, trace_len;
132 bool user = flags & BPF_F_USER_STACK;
133 bool kernel = !user;
134 u64 *ips;
135
136 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
137 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
138 return -EINVAL;
139
140 trace = get_perf_callchain(regs, init_nr, kernel, user,
141 sysctl_perf_event_max_stack, false, false);
142
143 if (unlikely(!trace))
144 /* couldn't fetch the stack trace */
145 return -EFAULT;
146
147 /* get_perf_callchain() guarantees that trace->nr >= init_nr
148 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
149 */
150 trace_nr = trace->nr - init_nr;
151
152 if (trace_nr <= skip)
153 /* skipping more than usable stack trace */
154 return -EFAULT;
155
156 trace_nr -= skip;
157 trace_len = trace_nr * sizeof(u64);
158 ips = trace->ip + skip + init_nr;
159 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
160 id = hash & (smap->n_buckets - 1);
161 bucket = READ_ONCE(smap->buckets[id]);
162
163 if (bucket && bucket->hash == hash) {
164 if (flags & BPF_F_FAST_STACK_CMP)
165 return id;
166 if (bucket->nr == trace_nr &&
167 memcmp(bucket->ip, ips, trace_len) == 0)
168 return id;
169 }
170
171 /* this call stack is not in the map, try to add it */
172 if (bucket && !(flags & BPF_F_REUSE_STACKID))
173 return -EEXIST;
174
175 new_bucket = (struct stack_map_bucket *)
176 pcpu_freelist_pop(&smap->freelist);
177 if (unlikely(!new_bucket))
178 return -ENOMEM;
179
180 memcpy(new_bucket->ip, ips, trace_len);
181 new_bucket->hash = hash;
182 new_bucket->nr = trace_nr;
183
184 old_bucket = xchg(&smap->buckets[id], new_bucket);
185 if (old_bucket)
186 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
187 return id;
188}
189
190const struct bpf_func_proto bpf_get_stackid_proto = {
191 .func = bpf_get_stackid,
192 .gpl_only = true,
193 .ret_type = RET_INTEGER,
194 .arg1_type = ARG_PTR_TO_CTX,
195 .arg2_type = ARG_CONST_MAP_PTR,
196 .arg3_type = ARG_ANYTHING,
197};
198
199/* Called from eBPF program */
200static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
201{
202 return NULL;
203}
204
205/* Called from syscall */
206int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
207{
208 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
209 struct stack_map_bucket *bucket, *old_bucket;
210 u32 id = *(u32 *)key, trace_len;
211
212 if (unlikely(id >= smap->n_buckets))
213 return -ENOENT;
214
215 bucket = xchg(&smap->buckets[id], NULL);
216 if (!bucket)
217 return -ENOENT;
218
219 trace_len = bucket->nr * sizeof(u64);
220 memcpy(value, bucket->ip, trace_len);
221 memset(value + trace_len, 0, map->value_size - trace_len);
222
223 old_bucket = xchg(&smap->buckets[id], bucket);
224 if (old_bucket)
225 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
226 return 0;
227}
228
229static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
230{
231 return -EINVAL;
232}
233
234static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
235 u64 map_flags)
236{
237 return -EINVAL;
238}
239
240/* Called from syscall or from eBPF program */
241static int stack_map_delete_elem(struct bpf_map *map, void *key)
242{
243 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
244 struct stack_map_bucket *old_bucket;
245 u32 id = *(u32 *)key;
246
247 if (unlikely(id >= smap->n_buckets))
248 return -E2BIG;
249
250 old_bucket = xchg(&smap->buckets[id], NULL);
251 if (old_bucket) {
252 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
253 return 0;
254 } else {
255 return -ENOENT;
256 }
257}
258
259/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
260static void stack_map_free(struct bpf_map *map)
261{
262 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
263
264 /* wait for bpf programs to complete before freeing stack map */
265 synchronize_rcu();
266
267 bpf_map_area_free(smap->elems);
268 pcpu_freelist_destroy(&smap->freelist);
269 bpf_map_area_free(smap);
270 put_callchain_buffers();
271}
272
273const struct bpf_map_ops stack_map_ops = {
274 .map_alloc = stack_map_alloc,
275 .map_free = stack_map_free,
276 .map_get_next_key = stack_map_get_next_key,
277 .map_lookup_elem = stack_map_lookup_elem,
278 .map_update_elem = stack_map_update_elem,
279 .map_delete_elem = stack_map_delete_elem,
280};