xref: /openbmc/linux/kernel/bpf/reuseport_array.c (revision 53a2a90d)
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 	uintptr_t sk_user_data;
25 
26 	write_lock_bh(&sk->sk_callback_lock);
27 	sk_user_data = (uintptr_t)sk->sk_user_data;
28 	if (sk_user_data & SK_USER_DATA_BPF) {
29 		struct sock __rcu **socks;
30 
31 		socks = (void *)(sk_user_data & SK_USER_DATA_PTRMASK);
32 		WRITE_ONCE(sk->sk_user_data, NULL);
33 		/*
34 		 * Do not move this NULL assignment outside of
35 		 * sk->sk_callback_lock because there is
36 		 * a race with reuseport_array_free()
37 		 * which does not hold the reuseport_lock.
38 		 */
39 		RCU_INIT_POINTER(*socks, NULL);
40 	}
41 	write_unlock_bh(&sk->sk_callback_lock);
42 }
43 
44 static int reuseport_array_alloc_check(union bpf_attr *attr)
45 {
46 	if (attr->value_size != sizeof(u32) &&
47 	    attr->value_size != sizeof(u64))
48 		return -EINVAL;
49 
50 	return array_map_alloc_check(attr);
51 }
52 
53 static void *reuseport_array_lookup_elem(struct bpf_map *map, void *key)
54 {
55 	struct reuseport_array *array = reuseport_array(map);
56 	u32 index = *(u32 *)key;
57 
58 	if (unlikely(index >= array->map.max_entries))
59 		return NULL;
60 
61 	return rcu_dereference(array->ptrs[index]);
62 }
63 
64 /* Called from syscall only */
65 static int reuseport_array_delete_elem(struct bpf_map *map, void *key)
66 {
67 	struct reuseport_array *array = reuseport_array(map);
68 	u32 index = *(u32 *)key;
69 	struct sock *sk;
70 	int err;
71 
72 	if (index >= map->max_entries)
73 		return -E2BIG;
74 
75 	if (!rcu_access_pointer(array->ptrs[index]))
76 		return -ENOENT;
77 
78 	spin_lock_bh(&reuseport_lock);
79 
80 	sk = rcu_dereference_protected(array->ptrs[index],
81 				       lockdep_is_held(&reuseport_lock));
82 	if (sk) {
83 		write_lock_bh(&sk->sk_callback_lock);
84 		WRITE_ONCE(sk->sk_user_data, NULL);
85 		RCU_INIT_POINTER(array->ptrs[index], NULL);
86 		write_unlock_bh(&sk->sk_callback_lock);
87 		err = 0;
88 	} else {
89 		err = -ENOENT;
90 	}
91 
92 	spin_unlock_bh(&reuseport_lock);
93 
94 	return err;
95 }
96 
97 static void reuseport_array_free(struct bpf_map *map)
98 {
99 	struct reuseport_array *array = reuseport_array(map);
100 	struct sock *sk;
101 	u32 i;
102 
103 	/*
104 	 * ops->map_*_elem() will not be able to access this
105 	 * array now. Hence, this function only races with
106 	 * bpf_sk_reuseport_detach() which was triggered by
107 	 * close() or disconnect().
108 	 *
109 	 * This function and bpf_sk_reuseport_detach() are
110 	 * both removing sk from "array".  Who removes it
111 	 * first does not matter.
112 	 *
113 	 * The only concern here is bpf_sk_reuseport_detach()
114 	 * may access "array" which is being freed here.
115 	 * bpf_sk_reuseport_detach() access this "array"
116 	 * through sk->sk_user_data _and_ with sk->sk_callback_lock
117 	 * held which is enough because this "array" is not freed
118 	 * until all sk->sk_user_data has stopped referencing this "array".
119 	 *
120 	 * Hence, due to the above, taking "reuseport_lock" is not
121 	 * needed here.
122 	 */
123 
124 	/*
125 	 * Since reuseport_lock is not taken, sk is accessed under
126 	 * rcu_read_lock()
127 	 */
128 	rcu_read_lock();
129 	for (i = 0; i < map->max_entries; i++) {
130 		sk = rcu_dereference(array->ptrs[i]);
131 		if (sk) {
132 			write_lock_bh(&sk->sk_callback_lock);
133 			/*
134 			 * No need for WRITE_ONCE(). At this point,
135 			 * no one is reading it without taking the
136 			 * sk->sk_callback_lock.
137 			 */
138 			sk->sk_user_data = NULL;
139 			write_unlock_bh(&sk->sk_callback_lock);
140 			RCU_INIT_POINTER(array->ptrs[i], NULL);
141 		}
142 	}
143 	rcu_read_unlock();
144 
145 	/*
146 	 * Once reaching here, all sk->sk_user_data is not
147 	 * referencing this "array". "array" can be freed now.
148 	 */
149 	bpf_map_area_free(array);
150 }
151 
152 static struct bpf_map *reuseport_array_alloc(union bpf_attr *attr)
153 {
154 	int numa_node = bpf_map_attr_numa_node(attr);
155 	struct reuseport_array *array;
156 
157 	if (!bpf_capable())
158 		return ERR_PTR(-EPERM);
159 
160 	/* allocate all map elements and zero-initialize them */
161 	array = bpf_map_area_alloc(struct_size(array, ptrs, attr->max_entries), numa_node);
162 	if (!array)
163 		return ERR_PTR(-ENOMEM);
164 
165 	/* copy mandatory map attributes */
166 	bpf_map_init_from_attr(&array->map, attr);
167 
168 	return &array->map;
169 }
170 
171 int bpf_fd_reuseport_array_lookup_elem(struct bpf_map *map, void *key,
172 				       void *value)
173 {
174 	struct sock *sk;
175 	int err;
176 
177 	if (map->value_size != sizeof(u64))
178 		return -ENOSPC;
179 
180 	rcu_read_lock();
181 	sk = reuseport_array_lookup_elem(map, key);
182 	if (sk) {
183 		*(u64 *)value = __sock_gen_cookie(sk);
184 		err = 0;
185 	} else {
186 		err = -ENOENT;
187 	}
188 	rcu_read_unlock();
189 
190 	return err;
191 }
192 
193 static int
194 reuseport_array_update_check(const struct reuseport_array *array,
195 			     const struct sock *nsk,
196 			     const struct sock *osk,
197 			     const struct sock_reuseport *nsk_reuse,
198 			     u32 map_flags)
199 {
200 	if (osk && map_flags == BPF_NOEXIST)
201 		return -EEXIST;
202 
203 	if (!osk && map_flags == BPF_EXIST)
204 		return -ENOENT;
205 
206 	if (nsk->sk_protocol != IPPROTO_UDP && nsk->sk_protocol != IPPROTO_TCP)
207 		return -ENOTSUPP;
208 
209 	if (nsk->sk_family != AF_INET && nsk->sk_family != AF_INET6)
210 		return -ENOTSUPP;
211 
212 	if (nsk->sk_type != SOCK_STREAM && nsk->sk_type != SOCK_DGRAM)
213 		return -ENOTSUPP;
214 
215 	/*
216 	 * sk must be hashed (i.e. listening in the TCP case or binded
217 	 * in the UDP case) and
218 	 * it must also be a SO_REUSEPORT sk (i.e. reuse cannot be NULL).
219 	 *
220 	 * Also, sk will be used in bpf helper that is protected by
221 	 * rcu_read_lock().
222 	 */
223 	if (!sock_flag(nsk, SOCK_RCU_FREE) || !sk_hashed(nsk) || !nsk_reuse)
224 		return -EINVAL;
225 
226 	/* READ_ONCE because the sk->sk_callback_lock may not be held here */
227 	if (READ_ONCE(nsk->sk_user_data))
228 		return -EBUSY;
229 
230 	return 0;
231 }
232 
233 /*
234  * Called from syscall only.
235  * The "nsk" in the fd refcnt.
236  * The "osk" and "reuse" are protected by reuseport_lock.
237  */
238 int bpf_fd_reuseport_array_update_elem(struct bpf_map *map, void *key,
239 				       void *value, u64 map_flags)
240 {
241 	struct reuseport_array *array = reuseport_array(map);
242 	struct sock *free_osk = NULL, *osk, *nsk;
243 	struct sock_reuseport *reuse;
244 	u32 index = *(u32 *)key;
245 	uintptr_t sk_user_data;
246 	struct socket *socket;
247 	int err, fd;
248 
249 	if (map_flags > BPF_EXIST)
250 		return -EINVAL;
251 
252 	if (index >= map->max_entries)
253 		return -E2BIG;
254 
255 	if (map->value_size == sizeof(u64)) {
256 		u64 fd64 = *(u64 *)value;
257 
258 		if (fd64 > S32_MAX)
259 			return -EINVAL;
260 		fd = fd64;
261 	} else {
262 		fd = *(int *)value;
263 	}
264 
265 	socket = sockfd_lookup(fd, &err);
266 	if (!socket)
267 		return err;
268 
269 	nsk = socket->sk;
270 	if (!nsk) {
271 		err = -EINVAL;
272 		goto put_file;
273 	}
274 
275 	/* Quick checks before taking reuseport_lock */
276 	err = reuseport_array_update_check(array, nsk,
277 					   rcu_access_pointer(array->ptrs[index]),
278 					   rcu_access_pointer(nsk->sk_reuseport_cb),
279 					   map_flags);
280 	if (err)
281 		goto put_file;
282 
283 	spin_lock_bh(&reuseport_lock);
284 	/*
285 	 * Some of the checks only need reuseport_lock
286 	 * but it is done under sk_callback_lock also
287 	 * for simplicity reason.
288 	 */
289 	write_lock_bh(&nsk->sk_callback_lock);
290 
291 	osk = rcu_dereference_protected(array->ptrs[index],
292 					lockdep_is_held(&reuseport_lock));
293 	reuse = rcu_dereference_protected(nsk->sk_reuseport_cb,
294 					  lockdep_is_held(&reuseport_lock));
295 	err = reuseport_array_update_check(array, nsk, osk, reuse, map_flags);
296 	if (err)
297 		goto put_file_unlock;
298 
299 	sk_user_data = (uintptr_t)&array->ptrs[index] | SK_USER_DATA_NOCOPY |
300 		SK_USER_DATA_BPF;
301 	WRITE_ONCE(nsk->sk_user_data, (void *)sk_user_data);
302 	rcu_assign_pointer(array->ptrs[index], nsk);
303 	free_osk = osk;
304 	err = 0;
305 
306 put_file_unlock:
307 	write_unlock_bh(&nsk->sk_callback_lock);
308 
309 	if (free_osk) {
310 		write_lock_bh(&free_osk->sk_callback_lock);
311 		WRITE_ONCE(free_osk->sk_user_data, NULL);
312 		write_unlock_bh(&free_osk->sk_callback_lock);
313 	}
314 
315 	spin_unlock_bh(&reuseport_lock);
316 put_file:
317 	fput(socket->file);
318 	return err;
319 }
320 
321 /* Called from syscall */
322 static int reuseport_array_get_next_key(struct bpf_map *map, void *key,
323 					void *next_key)
324 {
325 	struct reuseport_array *array = reuseport_array(map);
326 	u32 index = key ? *(u32 *)key : U32_MAX;
327 	u32 *next = (u32 *)next_key;
328 
329 	if (index >= array->map.max_entries) {
330 		*next = 0;
331 		return 0;
332 	}
333 
334 	if (index == array->map.max_entries - 1)
335 		return -ENOENT;
336 
337 	*next = index + 1;
338 	return 0;
339 }
340 
341 BTF_ID_LIST_SINGLE(reuseport_array_map_btf_ids, struct, reuseport_array)
342 const struct bpf_map_ops reuseport_array_ops = {
343 	.map_meta_equal = bpf_map_meta_equal,
344 	.map_alloc_check = reuseport_array_alloc_check,
345 	.map_alloc = reuseport_array_alloc,
346 	.map_free = reuseport_array_free,
347 	.map_lookup_elem = reuseport_array_lookup_elem,
348 	.map_get_next_key = reuseport_array_get_next_key,
349 	.map_delete_elem = reuseport_array_delete_elem,
350 	.map_btf_id = &reuseport_array_map_btf_ids[0],
351 };
352