1 /* Copyright (c) 2016 Facebook
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.
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"
14 struct stack_map_bucket {
15 struct pcpu_freelist_node fnode;
21 struct bpf_stack_map {
24 struct pcpu_freelist freelist;
26 struct stack_map_bucket *buckets[];
29 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
31 u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
34 smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries);
38 err = pcpu_freelist_init(&smap->freelist);
42 pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
43 smap->map.max_entries);
47 bpf_map_area_free(smap->elems);
51 /* Called from syscall */
52 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
54 u32 value_size = attr->value_size;
55 struct bpf_stack_map *smap;
59 if (!capable(CAP_SYS_ADMIN))
60 return ERR_PTR(-EPERM);
63 return ERR_PTR(-EINVAL);
65 /* check sanity of attributes */
66 if (attr->max_entries == 0 || attr->key_size != 4 ||
67 value_size < 8 || value_size % 8 ||
68 value_size / 8 > sysctl_perf_event_max_stack)
69 return ERR_PTR(-EINVAL);
71 /* hash table size must be power of 2 */
72 n_buckets = roundup_pow_of_two(attr->max_entries);
74 return ERR_PTR(-E2BIG);
76 cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
77 if (cost >= U32_MAX - PAGE_SIZE)
78 return ERR_PTR(-E2BIG);
80 smap = bpf_map_area_alloc(cost);
82 return ERR_PTR(-ENOMEM);
85 cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
86 if (cost >= U32_MAX - PAGE_SIZE)
89 smap->map.map_type = attr->map_type;
90 smap->map.key_size = attr->key_size;
91 smap->map.value_size = value_size;
92 smap->map.max_entries = attr->max_entries;
93 smap->map.map_flags = attr->map_flags;
94 smap->n_buckets = n_buckets;
95 smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
97 err = bpf_map_precharge_memlock(smap->map.pages);
101 err = get_callchain_buffers(sysctl_perf_event_max_stack);
105 err = prealloc_elems_and_freelist(smap);
112 put_callchain_buffers();
114 bpf_map_area_free(smap);
118 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
121 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
122 struct perf_callchain_entry *trace;
123 struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
124 u32 max_depth = map->value_size / 8;
125 /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
126 u32 init_nr = sysctl_perf_event_max_stack - max_depth;
127 u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
128 u32 hash, id, trace_nr, trace_len;
129 bool user = flags & BPF_F_USER_STACK;
133 if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
134 BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
137 trace = get_perf_callchain(regs, init_nr, kernel, user,
138 sysctl_perf_event_max_stack, false, false);
140 if (unlikely(!trace))
141 /* couldn't fetch the stack trace */
144 /* get_perf_callchain() guarantees that trace->nr >= init_nr
145 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
147 trace_nr = trace->nr - init_nr;
149 if (trace_nr <= skip)
150 /* skipping more than usable stack trace */
154 trace_len = trace_nr * sizeof(u64);
155 ips = trace->ip + skip + init_nr;
156 hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
157 id = hash & (smap->n_buckets - 1);
158 bucket = READ_ONCE(smap->buckets[id]);
160 if (bucket && bucket->hash == hash) {
161 if (flags & BPF_F_FAST_STACK_CMP)
163 if (bucket->nr == trace_nr &&
164 memcmp(bucket->ip, ips, trace_len) == 0)
168 /* this call stack is not in the map, try to add it */
169 if (bucket && !(flags & BPF_F_REUSE_STACKID))
172 new_bucket = (struct stack_map_bucket *)
173 pcpu_freelist_pop(&smap->freelist);
174 if (unlikely(!new_bucket))
177 memcpy(new_bucket->ip, ips, trace_len);
178 new_bucket->hash = hash;
179 new_bucket->nr = trace_nr;
181 old_bucket = xchg(&smap->buckets[id], new_bucket);
183 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
187 const struct bpf_func_proto bpf_get_stackid_proto = {
188 .func = bpf_get_stackid,
190 .ret_type = RET_INTEGER,
191 .arg1_type = ARG_PTR_TO_CTX,
192 .arg2_type = ARG_CONST_MAP_PTR,
193 .arg3_type = ARG_ANYTHING,
196 /* Called from eBPF program */
197 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
202 /* Called from syscall */
203 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
205 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
206 struct stack_map_bucket *bucket, *old_bucket;
207 u32 id = *(u32 *)key, trace_len;
209 if (unlikely(id >= smap->n_buckets))
212 bucket = xchg(&smap->buckets[id], NULL);
216 trace_len = bucket->nr * sizeof(u64);
217 memcpy(value, bucket->ip, trace_len);
218 memset(value + trace_len, 0, map->value_size - trace_len);
220 old_bucket = xchg(&smap->buckets[id], bucket);
222 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
226 static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
231 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
237 /* Called from syscall or from eBPF program */
238 static int stack_map_delete_elem(struct bpf_map *map, void *key)
240 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
241 struct stack_map_bucket *old_bucket;
242 u32 id = *(u32 *)key;
244 if (unlikely(id >= smap->n_buckets))
247 old_bucket = xchg(&smap->buckets[id], NULL);
249 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
256 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
257 static void stack_map_free(struct bpf_map *map)
259 struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
261 /* wait for bpf programs to complete before freeing stack map */
264 bpf_map_area_free(smap->elems);
265 pcpu_freelist_destroy(&smap->freelist);
266 bpf_map_area_free(smap);
267 put_callchain_buffers();
270 static const struct bpf_map_ops stack_map_ops = {
271 .map_alloc = stack_map_alloc,
272 .map_free = stack_map_free,
273 .map_get_next_key = stack_map_get_next_key,
274 .map_lookup_elem = stack_map_lookup_elem,
275 .map_update_elem = stack_map_update_elem,
276 .map_delete_elem = stack_map_delete_elem,
279 static struct bpf_map_type_list stack_map_type __read_mostly = {
280 .ops = &stack_map_ops,
281 .type = BPF_MAP_TYPE_STACK_TRACE,
284 static int __init register_stack_map(void)
286 bpf_register_map_type(&stack_map_type);
289 late_initcall(register_stack_map);