xref: /openbmc/linux/kernel/bpf/reuseport_array.c (revision 11c416e3)
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 
10 struct reuseport_array {
11 	struct bpf_map map;
12 	struct sock __rcu *ptrs[];
13 };
14 
15 static struct reuseport_array *reuseport_array(struct bpf_map *map)
16 {
17 	return (struct reuseport_array *)map;
18 }
19 
20 /* The caller must hold the reuseport_lock */
21 void bpf_sk_reuseport_detach(struct sock *sk)
22 {
23 	uintptr_t sk_user_data;
24 
25 	write_lock_bh(&sk->sk_callback_lock);
26 	sk_user_data = (uintptr_t)sk->sk_user_data;
27 	if (sk_user_data & SK_USER_DATA_BPF) {
28 		struct sock __rcu **socks;
29 
30 		socks = (void *)(sk_user_data & SK_USER_DATA_PTRMASK);
31 		WRITE_ONCE(sk->sk_user_data, NULL);
32 		/*
33 		 * Do not move this NULL assignment outside of
34 		 * sk->sk_callback_lock because there is
35 		 * a race with reuseport_array_free()
36 		 * which does not hold the reuseport_lock.
37 		 */
38 		RCU_INIT_POINTER(*socks, NULL);
39 	}
40 	write_unlock_bh(&sk->sk_callback_lock);
41 }
42 
43 static int reuseport_array_alloc_check(union bpf_attr *attr)
44 {
45 	if (attr->value_size != sizeof(u32) &&
46 	    attr->value_size != sizeof(u64))
47 		return -EINVAL;
48 
49 	return array_map_alloc_check(attr);
50 }
51 
52 static void *reuseport_array_lookup_elem(struct bpf_map *map, void *key)
53 {
54 	struct reuseport_array *array = reuseport_array(map);
55 	u32 index = *(u32 *)key;
56 
57 	if (unlikely(index >= array->map.max_entries))
58 		return NULL;
59 
60 	return rcu_dereference(array->ptrs[index]);
61 }
62 
63 /* Called from syscall only */
64 static int reuseport_array_delete_elem(struct bpf_map *map, void *key)
65 {
66 	struct reuseport_array *array = reuseport_array(map);
67 	u32 index = *(u32 *)key;
68 	struct sock *sk;
69 	int err;
70 
71 	if (index >= map->max_entries)
72 		return -E2BIG;
73 
74 	if (!rcu_access_pointer(array->ptrs[index]))
75 		return -ENOENT;
76 
77 	spin_lock_bh(&reuseport_lock);
78 
79 	sk = rcu_dereference_protected(array->ptrs[index],
80 				       lockdep_is_held(&reuseport_lock));
81 	if (sk) {
82 		write_lock_bh(&sk->sk_callback_lock);
83 		WRITE_ONCE(sk->sk_user_data, NULL);
84 		RCU_INIT_POINTER(array->ptrs[index], NULL);
85 		write_unlock_bh(&sk->sk_callback_lock);
86 		err = 0;
87 	} else {
88 		err = -ENOENT;
89 	}
90 
91 	spin_unlock_bh(&reuseport_lock);
92 
93 	return err;
94 }
95 
96 static void reuseport_array_free(struct bpf_map *map)
97 {
98 	struct reuseport_array *array = reuseport_array(map);
99 	struct sock *sk;
100 	u32 i;
101 
102 	synchronize_rcu();
103 
104 	/*
105 	 * ops->map_*_elem() will not be able to access this
106 	 * array now. Hence, this function only races with
107 	 * bpf_sk_reuseport_detach() which was triggerred by
108 	 * close() or disconnect().
109 	 *
110 	 * This function and bpf_sk_reuseport_detach() are
111 	 * both removing sk from "array".  Who removes it
112 	 * first does not matter.
113 	 *
114 	 * The only concern here is bpf_sk_reuseport_detach()
115 	 * may access "array" which is being freed here.
116 	 * bpf_sk_reuseport_detach() access this "array"
117 	 * through sk->sk_user_data _and_ with sk->sk_callback_lock
118 	 * held which is enough because this "array" is not freed
119 	 * until all sk->sk_user_data has stopped referencing this "array".
120 	 *
121 	 * Hence, due to the above, taking "reuseport_lock" is not
122 	 * needed here.
123 	 */
124 
125 	/*
126 	 * Since reuseport_lock is not taken, sk is accessed under
127 	 * rcu_read_lock()
128 	 */
129 	rcu_read_lock();
130 	for (i = 0; i < map->max_entries; i++) {
131 		sk = rcu_dereference(array->ptrs[i]);
132 		if (sk) {
133 			write_lock_bh(&sk->sk_callback_lock);
134 			/*
135 			 * No need for WRITE_ONCE(). At this point,
136 			 * no one is reading it without taking the
137 			 * sk->sk_callback_lock.
138 			 */
139 			sk->sk_user_data = NULL;
140 			write_unlock_bh(&sk->sk_callback_lock);
141 			RCU_INIT_POINTER(array->ptrs[i], NULL);
142 		}
143 	}
144 	rcu_read_unlock();
145 
146 	/*
147 	 * Once reaching here, all sk->sk_user_data is not
148 	 * referenceing this "array".  "array" can be freed now.
149 	 */
150 	bpf_map_area_free(array);
151 }
152 
153 static struct bpf_map *reuseport_array_alloc(union bpf_attr *attr)
154 {
155 	int err, numa_node = bpf_map_attr_numa_node(attr);
156 	struct reuseport_array *array;
157 	struct bpf_map_memory mem;
158 	u64 array_size;
159 
160 	if (!bpf_capable())
161 		return ERR_PTR(-EPERM);
162 
163 	array_size = sizeof(*array);
164 	array_size += (u64)attr->max_entries * sizeof(struct sock *);
165 
166 	err = bpf_map_charge_init(&mem, array_size);
167 	if (err)
168 		return ERR_PTR(err);
169 
170 	/* allocate all map elements and zero-initialize them */
171 	array = bpf_map_area_alloc(array_size, numa_node);
172 	if (!array) {
173 		bpf_map_charge_finish(&mem);
174 		return ERR_PTR(-ENOMEM);
175 	}
176 
177 	/* copy mandatory map attributes */
178 	bpf_map_init_from_attr(&array->map, attr);
179 	bpf_map_charge_move(&array->map.memory, &mem);
180 
181 	return &array->map;
182 }
183 
184 int bpf_fd_reuseport_array_lookup_elem(struct bpf_map *map, void *key,
185 				       void *value)
186 {
187 	struct sock *sk;
188 	int err;
189 
190 	if (map->value_size != sizeof(u64))
191 		return -ENOSPC;
192 
193 	rcu_read_lock();
194 	sk = reuseport_array_lookup_elem(map, key);
195 	if (sk) {
196 		*(u64 *)value = sock_gen_cookie(sk);
197 		err = 0;
198 	} else {
199 		err = -ENOENT;
200 	}
201 	rcu_read_unlock();
202 
203 	return err;
204 }
205 
206 static int
207 reuseport_array_update_check(const struct reuseport_array *array,
208 			     const struct sock *nsk,
209 			     const struct sock *osk,
210 			     const struct sock_reuseport *nsk_reuse,
211 			     u32 map_flags)
212 {
213 	if (osk && map_flags == BPF_NOEXIST)
214 		return -EEXIST;
215 
216 	if (!osk && map_flags == BPF_EXIST)
217 		return -ENOENT;
218 
219 	if (nsk->sk_protocol != IPPROTO_UDP && nsk->sk_protocol != IPPROTO_TCP)
220 		return -ENOTSUPP;
221 
222 	if (nsk->sk_family != AF_INET && nsk->sk_family != AF_INET6)
223 		return -ENOTSUPP;
224 
225 	if (nsk->sk_type != SOCK_STREAM && nsk->sk_type != SOCK_DGRAM)
226 		return -ENOTSUPP;
227 
228 	/*
229 	 * sk must be hashed (i.e. listening in the TCP case or binded
230 	 * in the UDP case) and
231 	 * it must also be a SO_REUSEPORT sk (i.e. reuse cannot be NULL).
232 	 *
233 	 * Also, sk will be used in bpf helper that is protected by
234 	 * rcu_read_lock().
235 	 */
236 	if (!sock_flag(nsk, SOCK_RCU_FREE) || !sk_hashed(nsk) || !nsk_reuse)
237 		return -EINVAL;
238 
239 	/* READ_ONCE because the sk->sk_callback_lock may not be held here */
240 	if (READ_ONCE(nsk->sk_user_data))
241 		return -EBUSY;
242 
243 	return 0;
244 }
245 
246 /*
247  * Called from syscall only.
248  * The "nsk" in the fd refcnt.
249  * The "osk" and "reuse" are protected by reuseport_lock.
250  */
251 int bpf_fd_reuseport_array_update_elem(struct bpf_map *map, void *key,
252 				       void *value, u64 map_flags)
253 {
254 	struct reuseport_array *array = reuseport_array(map);
255 	struct sock *free_osk = NULL, *osk, *nsk;
256 	struct sock_reuseport *reuse;
257 	u32 index = *(u32 *)key;
258 	uintptr_t sk_user_data;
259 	struct socket *socket;
260 	int err, fd;
261 
262 	if (map_flags > BPF_EXIST)
263 		return -EINVAL;
264 
265 	if (index >= map->max_entries)
266 		return -E2BIG;
267 
268 	if (map->value_size == sizeof(u64)) {
269 		u64 fd64 = *(u64 *)value;
270 
271 		if (fd64 > S32_MAX)
272 			return -EINVAL;
273 		fd = fd64;
274 	} else {
275 		fd = *(int *)value;
276 	}
277 
278 	socket = sockfd_lookup(fd, &err);
279 	if (!socket)
280 		return err;
281 
282 	nsk = socket->sk;
283 	if (!nsk) {
284 		err = -EINVAL;
285 		goto put_file;
286 	}
287 
288 	/* Quick checks before taking reuseport_lock */
289 	err = reuseport_array_update_check(array, nsk,
290 					   rcu_access_pointer(array->ptrs[index]),
291 					   rcu_access_pointer(nsk->sk_reuseport_cb),
292 					   map_flags);
293 	if (err)
294 		goto put_file;
295 
296 	spin_lock_bh(&reuseport_lock);
297 	/*
298 	 * Some of the checks only need reuseport_lock
299 	 * but it is done under sk_callback_lock also
300 	 * for simplicity reason.
301 	 */
302 	write_lock_bh(&nsk->sk_callback_lock);
303 
304 	osk = rcu_dereference_protected(array->ptrs[index],
305 					lockdep_is_held(&reuseport_lock));
306 	reuse = rcu_dereference_protected(nsk->sk_reuseport_cb,
307 					  lockdep_is_held(&reuseport_lock));
308 	err = reuseport_array_update_check(array, nsk, osk, reuse, map_flags);
309 	if (err)
310 		goto put_file_unlock;
311 
312 	sk_user_data = (uintptr_t)&array->ptrs[index] | SK_USER_DATA_NOCOPY |
313 		SK_USER_DATA_BPF;
314 	WRITE_ONCE(nsk->sk_user_data, (void *)sk_user_data);
315 	rcu_assign_pointer(array->ptrs[index], nsk);
316 	free_osk = osk;
317 	err = 0;
318 
319 put_file_unlock:
320 	write_unlock_bh(&nsk->sk_callback_lock);
321 
322 	if (free_osk) {
323 		write_lock_bh(&free_osk->sk_callback_lock);
324 		WRITE_ONCE(free_osk->sk_user_data, NULL);
325 		write_unlock_bh(&free_osk->sk_callback_lock);
326 	}
327 
328 	spin_unlock_bh(&reuseport_lock);
329 put_file:
330 	fput(socket->file);
331 	return err;
332 }
333 
334 /* Called from syscall */
335 static int reuseport_array_get_next_key(struct bpf_map *map, void *key,
336 					void *next_key)
337 {
338 	struct reuseport_array *array = reuseport_array(map);
339 	u32 index = key ? *(u32 *)key : U32_MAX;
340 	u32 *next = (u32 *)next_key;
341 
342 	if (index >= array->map.max_entries) {
343 		*next = 0;
344 		return 0;
345 	}
346 
347 	if (index == array->map.max_entries - 1)
348 		return -ENOENT;
349 
350 	*next = index + 1;
351 	return 0;
352 }
353 
354 const struct bpf_map_ops reuseport_array_ops = {
355 	.map_alloc_check = reuseport_array_alloc_check,
356 	.map_alloc = reuseport_array_alloc,
357 	.map_free = reuseport_array_free,
358 	.map_lookup_elem = reuseport_array_lookup_elem,
359 	.map_get_next_key = reuseport_array_get_next_key,
360 	.map_delete_elem = reuseport_array_delete_elem,
361 };
362