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
 14struct stack_map_bucket {
 15	struct pcpu_freelist_node fnode;
 16	u32 hash;
 17	u32 nr;
 18	u64 ip[];
 19};
 20
 21struct bpf_stack_map {
 22	struct bpf_map map;
 23	void *elems;
 24	struct pcpu_freelist freelist;
 25	u32 n_buckets;
 26	struct stack_map_bucket *buckets[];
 27};
 28
 29static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
 30{
 31	u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
 32	int err;
 33
 34	smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries);
 35	if (!smap->elems)
 36		return -ENOMEM;
 37
 38	err = pcpu_freelist_init(&smap->freelist);
 39	if (err)
 40		goto free_elems;
 41
 42	pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
 43			       smap->map.max_entries);
 44	return 0;
 45
 46free_elems:
 47	bpf_map_area_free(smap->elems);
 48	return err;
 49}
 50
 51/* Called from syscall */
 52static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
 53{
 54	u32 value_size = attr->value_size;
 55	struct bpf_stack_map *smap;
 56	u64 cost, n_buckets;
 57	int err;
 58
 59	if (!capable(CAP_SYS_ADMIN))
 60		return ERR_PTR(-EPERM);
 61
 62	if (attr->map_flags)
 63		return ERR_PTR(-EINVAL);
 64
 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);
 70
 71	/* hash table size must be power of 2 */
 72	n_buckets = roundup_pow_of_two(attr->max_entries);
 73
 74	cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 75	if (cost >= U32_MAX - PAGE_SIZE)
 76		return ERR_PTR(-E2BIG);
 77
 78	smap = bpf_map_area_alloc(cost);
 79	if (!smap)
 80		return ERR_PTR(-ENOMEM);
 81
 82	err = -E2BIG;
 83	cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
 84	if (cost >= U32_MAX - PAGE_SIZE)
 85		goto free_smap;
 86
 87	smap->map.map_type = attr->map_type;
 88	smap->map.key_size = attr->key_size;
 89	smap->map.value_size = value_size;
 90	smap->map.max_entries = attr->max_entries;
 91	smap->n_buckets = n_buckets;
 92	smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 93
 94	err = bpf_map_precharge_memlock(smap->map.pages);
 95	if (err)
 96		goto free_smap;
 97
 98	err = get_callchain_buffers(sysctl_perf_event_max_stack);
 99	if (err)
100		goto free_smap;
101
102	err = prealloc_elems_and_freelist(smap);
103	if (err)
104		goto put_buffers;
105
106	return &smap->map;
107
108put_buffers:
109	put_callchain_buffers();
110free_smap:
111	bpf_map_area_free(smap);
112	return ERR_PTR(err);
113}
114
115BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
116	   u64, flags)
117{
118	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
119	struct perf_callchain_entry *trace;
120	struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
121	u32 max_depth = map->value_size / 8;
122	/* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
123	u32 init_nr = sysctl_perf_event_max_stack - max_depth;
124	u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
125	u32 hash, id, trace_nr, trace_len;
126	bool user = flags & BPF_F_USER_STACK;
127	bool kernel = !user;
128	u64 *ips;
129
130	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
131			       BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
132		return -EINVAL;
133
134	trace = get_perf_callchain(regs, init_nr, kernel, user,
135				   sysctl_perf_event_max_stack, false, false);
136
137	if (unlikely(!trace))
138		/* couldn't fetch the stack trace */
139		return -EFAULT;
140
141	/* get_perf_callchain() guarantees that trace->nr >= init_nr
142	 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
143	 */
144	trace_nr = trace->nr - init_nr;
145
146	if (trace_nr <= skip)
147		/* skipping more than usable stack trace */
148		return -EFAULT;
149
150	trace_nr -= skip;
151	trace_len = trace_nr * sizeof(u64);
152	ips = trace->ip + skip + init_nr;
153	hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
154	id = hash & (smap->n_buckets - 1);
155	bucket = READ_ONCE(smap->buckets[id]);
156
157	if (bucket && bucket->hash == hash) {
158		if (flags & BPF_F_FAST_STACK_CMP)
159			return id;
160		if (bucket->nr == trace_nr &&
161		    memcmp(bucket->ip, ips, trace_len) == 0)
162			return id;
163	}
164
165	/* this call stack is not in the map, try to add it */
166	if (bucket && !(flags & BPF_F_REUSE_STACKID))
167		return -EEXIST;
168
169	new_bucket = (struct stack_map_bucket *)
170		pcpu_freelist_pop(&smap->freelist);
171	if (unlikely(!new_bucket))
172		return -ENOMEM;
173
174	memcpy(new_bucket->ip, ips, trace_len);
175	new_bucket->hash = hash;
176	new_bucket->nr = trace_nr;
177
178	old_bucket = xchg(&smap->buckets[id], new_bucket);
179	if (old_bucket)
180		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
181	return id;
182}
183
184const struct bpf_func_proto bpf_get_stackid_proto = {
185	.func		= bpf_get_stackid,
186	.gpl_only	= true,
187	.ret_type	= RET_INTEGER,
188	.arg1_type	= ARG_PTR_TO_CTX,
189	.arg2_type	= ARG_CONST_MAP_PTR,
190	.arg3_type	= ARG_ANYTHING,
191};
192
193/* Called from eBPF program */
194static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
195{
196	return NULL;
197}
198
199/* Called from syscall */
200int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
201{
202	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
203	struct stack_map_bucket *bucket, *old_bucket;
204	u32 id = *(u32 *)key, trace_len;
205
206	if (unlikely(id >= smap->n_buckets))
207		return -ENOENT;
208
209	bucket = xchg(&smap->buckets[id], NULL);
210	if (!bucket)
211		return -ENOENT;
212
213	trace_len = bucket->nr * sizeof(u64);
214	memcpy(value, bucket->ip, trace_len);
215	memset(value + trace_len, 0, map->value_size - trace_len);
216
217	old_bucket = xchg(&smap->buckets[id], bucket);
218	if (old_bucket)
219		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
220	return 0;
221}
222
223static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
224{
225	return -EINVAL;
226}
227
228static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
229				 u64 map_flags)
230{
231	return -EINVAL;
232}
233
234/* Called from syscall or from eBPF program */
235static int stack_map_delete_elem(struct bpf_map *map, void *key)
236{
237	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
238	struct stack_map_bucket *old_bucket;
239	u32 id = *(u32 *)key;
240
241	if (unlikely(id >= smap->n_buckets))
242		return -E2BIG;
243
244	old_bucket = xchg(&smap->buckets[id], NULL);
245	if (old_bucket) {
246		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
247		return 0;
248	} else {
249		return -ENOENT;
250	}
251}
252
253/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
254static void stack_map_free(struct bpf_map *map)
255{
256	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
257
258	/* wait for bpf programs to complete before freeing stack map */
259	synchronize_rcu();
260
261	bpf_map_area_free(smap->elems);
262	pcpu_freelist_destroy(&smap->freelist);
263	bpf_map_area_free(smap);
264	put_callchain_buffers();
265}
266
267static const struct bpf_map_ops stack_map_ops = {
268	.map_alloc = stack_map_alloc,
269	.map_free = stack_map_free,
270	.map_get_next_key = stack_map_get_next_key,
271	.map_lookup_elem = stack_map_lookup_elem,
272	.map_update_elem = stack_map_update_elem,
273	.map_delete_elem = stack_map_delete_elem,
274};
275
276static struct bpf_map_type_list stack_map_type __read_mostly = {
277	.ops = &stack_map_ops,
278	.type = BPF_MAP_TYPE_STACK_TRACE,
279};
280
281static int __init register_stack_map(void)
282{
283	bpf_register_map_type(&stack_map_type);
284	return 0;
285}
286late_initcall(register_stack_map);