xref: /linux/kernel/bpf/reuseport_array.c (revision 6562c9acb43ac69ba5a956b0c3911b883d90541f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2018 Facebook
4  */
5 #include <linux/bpf.h>
6 #include <linux/err.h>
7 #include <linux/sock_diag.h>
8 #include <net/sock_reuseport.h>
9 #include <linux/btf_ids.h>
10 
11 struct reuseport_array {
12 	struct bpf_map map;
13 	struct sock __rcu *ptrs[];
14 };
15 
16 static struct reuseport_array *reuseport_array(struct bpf_map *map)
17 {
18 	return (struct reuseport_array *)map;
19 }
20 
21 /* The caller must hold the reuseport_lock */
22 void bpf_sk_reuseport_detach(struct sock *sk)
23 {
24 	struct sock __rcu **socks;
25 
26 	write_lock_bh(&sk->sk_callback_lock);
27 	socks = __locked_read_sk_user_data_with_flags(sk, SK_USER_DATA_BPF);
28 	if (socks) {
29 		WRITE_ONCE(sk->sk_user_data, NULL);
30 		/*
31 		 * Do not move this NULL assignment outside of
32 		 * sk->sk_callback_lock because there is
33 		 * a race with reuseport_array_free()
34 		 * which does not hold the reuseport_lock.
35 		 */
36 		RCU_INIT_POINTER(*socks, NULL);
37 	}
38 	write_unlock_bh(&sk->sk_callback_lock);
39 }
40 
41 static int reuseport_array_alloc_check(union bpf_attr *attr)
42 {
43 	if (attr->value_size != sizeof(u32) &&
44 	    attr->value_size != sizeof(u64))
45 		return -EINVAL;
46 
47 	return array_map_alloc_check(attr);
48 }
49 
50 static void *reuseport_array_lookup_elem(struct bpf_map *map, void *key)
51 {
52 	struct reuseport_array *array = reuseport_array(map);
53 	u32 index = *(u32 *)key;
54 
55 	if (unlikely(index >= array->map.max_entries))
56 		return NULL;
57 
58 	return rcu_dereference(array->ptrs[index]);
59 }
60 
61 /* Called from syscall only */
62 static int reuseport_array_delete_elem(struct bpf_map *map, void *key)
63 {
64 	struct reuseport_array *array = reuseport_array(map);
65 	u32 index = *(u32 *)key;
66 	struct sock *sk;
67 	int err;
68 
69 	if (index >= map->max_entries)
70 		return -E2BIG;
71 
72 	if (!rcu_access_pointer(array->ptrs[index]))
73 		return -ENOENT;
74 
75 	spin_lock_bh(&reuseport_lock);
76 
77 	sk = rcu_dereference_protected(array->ptrs[index],
78 				       lockdep_is_held(&reuseport_lock));
79 	if (sk) {
80 		write_lock_bh(&sk->sk_callback_lock);
81 		WRITE_ONCE(sk->sk_user_data, NULL);
82 		RCU_INIT_POINTER(array->ptrs[index], NULL);
83 		write_unlock_bh(&sk->sk_callback_lock);
84 		err = 0;
85 	} else {
86 		err = -ENOENT;
87 	}
88 
89 	spin_unlock_bh(&reuseport_lock);
90 
91 	return err;
92 }
93 
94 static void reuseport_array_free(struct bpf_map *map)
95 {
96 	struct reuseport_array *array = reuseport_array(map);
97 	struct sock *sk;
98 	u32 i;
99 
100 	/*
101 	 * ops->map_*_elem() will not be able to access this
102 	 * array now. Hence, this function only races with
103 	 * bpf_sk_reuseport_detach() which was triggered by
104 	 * close() or disconnect().
105 	 *
106 	 * This function and bpf_sk_reuseport_detach() are
107 	 * both removing sk from "array".  Who removes it
108 	 * first does not matter.
109 	 *
110 	 * The only concern here is bpf_sk_reuseport_detach()
111 	 * may access "array" which is being freed here.
112 	 * bpf_sk_reuseport_detach() access this "array"
113 	 * through sk->sk_user_data _and_ with sk->sk_callback_lock
114 	 * held which is enough because this "array" is not freed
115 	 * until all sk->sk_user_data has stopped referencing this "array".
116 	 *
117 	 * Hence, due to the above, taking "reuseport_lock" is not
118 	 * needed here.
119 	 */
120 
121 	/*
122 	 * Since reuseport_lock is not taken, sk is accessed under
123 	 * rcu_read_lock()
124 	 */
125 	rcu_read_lock();
126 	for (i = 0; i < map->max_entries; i++) {
127 		sk = rcu_dereference(array->ptrs[i]);
128 		if (sk) {
129 			write_lock_bh(&sk->sk_callback_lock);
130 			/*
131 			 * No need for WRITE_ONCE(). At this point,
132 			 * no one is reading it without taking the
133 			 * sk->sk_callback_lock.
134 			 */
135 			sk->sk_user_data = NULL;
136 			write_unlock_bh(&sk->sk_callback_lock);
137 			RCU_INIT_POINTER(array->ptrs[i], NULL);
138 		}
139 	}
140 	rcu_read_unlock();
141 
142 	/*
143 	 * Once reaching here, all sk->sk_user_data is not
144 	 * referencing this "array". "array" can be freed now.
145 	 */
146 	bpf_map_area_free(array);
147 }
148 
149 static struct bpf_map *reuseport_array_alloc(union bpf_attr *attr)
150 {
151 	int numa_node = bpf_map_attr_numa_node(attr);
152 	struct reuseport_array *array;
153 
154 	if (!bpf_capable())
155 		return ERR_PTR(-EPERM);
156 
157 	/* allocate all map elements and zero-initialize them */
158 	array = bpf_map_area_alloc(struct_size(array, ptrs, attr->max_entries), numa_node);
159 	if (!array)
160 		return ERR_PTR(-ENOMEM);
161 
162 	/* copy mandatory map attributes */
163 	bpf_map_init_from_attr(&array->map, attr);
164 
165 	return &array->map;
166 }
167 
168 int bpf_fd_reuseport_array_lookup_elem(struct bpf_map *map, void *key,
169 				       void *value)
170 {
171 	struct sock *sk;
172 	int err;
173 
174 	if (map->value_size != sizeof(u64))
175 		return -ENOSPC;
176 
177 	rcu_read_lock();
178 	sk = reuseport_array_lookup_elem(map, key);
179 	if (sk) {
180 		*(u64 *)value = __sock_gen_cookie(sk);
181 		err = 0;
182 	} else {
183 		err = -ENOENT;
184 	}
185 	rcu_read_unlock();
186 
187 	return err;
188 }
189 
190 static int
191 reuseport_array_update_check(const struct reuseport_array *array,
192 			     const struct sock *nsk,
193 			     const struct sock *osk,
194 			     const struct sock_reuseport *nsk_reuse,
195 			     u32 map_flags)
196 {
197 	if (osk && map_flags == BPF_NOEXIST)
198 		return -EEXIST;
199 
200 	if (!osk && map_flags == BPF_EXIST)
201 		return -ENOENT;
202 
203 	if (nsk->sk_protocol != IPPROTO_UDP && nsk->sk_protocol != IPPROTO_TCP)
204 		return -ENOTSUPP;
205 
206 	if (nsk->sk_family != AF_INET && nsk->sk_family != AF_INET6)
207 		return -ENOTSUPP;
208 
209 	if (nsk->sk_type != SOCK_STREAM && nsk->sk_type != SOCK_DGRAM)
210 		return -ENOTSUPP;
211 
212 	/*
213 	 * sk must be hashed (i.e. listening in the TCP case or binded
214 	 * in the UDP case) and
215 	 * it must also be a SO_REUSEPORT sk (i.e. reuse cannot be NULL).
216 	 *
217 	 * Also, sk will be used in bpf helper that is protected by
218 	 * rcu_read_lock().
219 	 */
220 	if (!sock_flag(nsk, SOCK_RCU_FREE) || !sk_hashed(nsk) || !nsk_reuse)
221 		return -EINVAL;
222 
223 	/* READ_ONCE because the sk->sk_callback_lock may not be held here */
224 	if (READ_ONCE(nsk->sk_user_data))
225 		return -EBUSY;
226 
227 	return 0;
228 }
229 
230 /*
231  * Called from syscall only.
232  * The "nsk" in the fd refcnt.
233  * The "osk" and "reuse" are protected by reuseport_lock.
234  */
235 int bpf_fd_reuseport_array_update_elem(struct bpf_map *map, void *key,
236 				       void *value, u64 map_flags)
237 {
238 	struct reuseport_array *array = reuseport_array(map);
239 	struct sock *free_osk = NULL, *osk, *nsk;
240 	struct sock_reuseport *reuse;
241 	u32 index = *(u32 *)key;
242 	uintptr_t sk_user_data;
243 	struct socket *socket;
244 	int err, fd;
245 
246 	if (map_flags > BPF_EXIST)
247 		return -EINVAL;
248 
249 	if (index >= map->max_entries)
250 		return -E2BIG;
251 
252 	if (map->value_size == sizeof(u64)) {
253 		u64 fd64 = *(u64 *)value;
254 
255 		if (fd64 > S32_MAX)
256 			return -EINVAL;
257 		fd = fd64;
258 	} else {
259 		fd = *(int *)value;
260 	}
261 
262 	socket = sockfd_lookup(fd, &err);
263 	if (!socket)
264 		return err;
265 
266 	nsk = socket->sk;
267 	if (!nsk) {
268 		err = -EINVAL;
269 		goto put_file;
270 	}
271 
272 	/* Quick checks before taking reuseport_lock */
273 	err = reuseport_array_update_check(array, nsk,
274 					   rcu_access_pointer(array->ptrs[index]),
275 					   rcu_access_pointer(nsk->sk_reuseport_cb),
276 					   map_flags);
277 	if (err)
278 		goto put_file;
279 
280 	spin_lock_bh(&reuseport_lock);
281 	/*
282 	 * Some of the checks only need reuseport_lock
283 	 * but it is done under sk_callback_lock also
284 	 * for simplicity reason.
285 	 */
286 	write_lock_bh(&nsk->sk_callback_lock);
287 
288 	osk = rcu_dereference_protected(array->ptrs[index],
289 					lockdep_is_held(&reuseport_lock));
290 	reuse = rcu_dereference_protected(nsk->sk_reuseport_cb,
291 					  lockdep_is_held(&reuseport_lock));
292 	err = reuseport_array_update_check(array, nsk, osk, reuse, map_flags);
293 	if (err)
294 		goto put_file_unlock;
295 
296 	sk_user_data = (uintptr_t)&array->ptrs[index] | SK_USER_DATA_NOCOPY |
297 		SK_USER_DATA_BPF;
298 	WRITE_ONCE(nsk->sk_user_data, (void *)sk_user_data);
299 	rcu_assign_pointer(array->ptrs[index], nsk);
300 	free_osk = osk;
301 	err = 0;
302 
303 put_file_unlock:
304 	write_unlock_bh(&nsk->sk_callback_lock);
305 
306 	if (free_osk) {
307 		write_lock_bh(&free_osk->sk_callback_lock);
308 		WRITE_ONCE(free_osk->sk_user_data, NULL);
309 		write_unlock_bh(&free_osk->sk_callback_lock);
310 	}
311 
312 	spin_unlock_bh(&reuseport_lock);
313 put_file:
314 	fput(socket->file);
315 	return err;
316 }
317 
318 /* Called from syscall */
319 static int reuseport_array_get_next_key(struct bpf_map *map, void *key,
320 					void *next_key)
321 {
322 	struct reuseport_array *array = reuseport_array(map);
323 	u32 index = key ? *(u32 *)key : U32_MAX;
324 	u32 *next = (u32 *)next_key;
325 
326 	if (index >= array->map.max_entries) {
327 		*next = 0;
328 		return 0;
329 	}
330 
331 	if (index == array->map.max_entries - 1)
332 		return -ENOENT;
333 
334 	*next = index + 1;
335 	return 0;
336 }
337 
338 BTF_ID_LIST_SINGLE(reuseport_array_map_btf_ids, struct, reuseport_array)
339 const struct bpf_map_ops reuseport_array_ops = {
340 	.map_meta_equal = bpf_map_meta_equal,
341 	.map_alloc_check = reuseport_array_alloc_check,
342 	.map_alloc = reuseport_array_alloc,
343 	.map_free = reuseport_array_free,
344 	.map_lookup_elem = reuseport_array_lookup_elem,
345 	.map_get_next_key = reuseport_array_get_next_key,
346 	.map_delete_elem = reuseport_array_delete_elem,
347 	.map_btf_id = &reuseport_array_map_btf_ids[0],
348 };
349