1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2018 Facebook
7 #include <linux/sock_diag.h>
8 #include <net/sock_reuseport.h>
10 struct reuseport_array {
12 struct sock __rcu *ptrs[];
15 static struct reuseport_array *reuseport_array(struct bpf_map *map)
17 return (struct reuseport_array *)map;
20 /* The caller must hold the reuseport_lock */
21 void bpf_sk_reuseport_detach(struct sock *sk)
23 uintptr_t sk_user_data;
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;
30 socks = (void *)(sk_user_data & SK_USER_DATA_PTRMASK);
31 WRITE_ONCE(sk->sk_user_data, NULL);
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.
38 RCU_INIT_POINTER(*socks, NULL);
40 write_unlock_bh(&sk->sk_callback_lock);
43 static int reuseport_array_alloc_check(union bpf_attr *attr)
45 if (attr->value_size != sizeof(u32) &&
46 attr->value_size != sizeof(u64))
49 return array_map_alloc_check(attr);
52 static void *reuseport_array_lookup_elem(struct bpf_map *map, void *key)
54 struct reuseport_array *array = reuseport_array(map);
55 u32 index = *(u32 *)key;
57 if (unlikely(index >= array->map.max_entries))
60 return rcu_dereference(array->ptrs[index]);
63 /* Called from syscall only */
64 static int reuseport_array_delete_elem(struct bpf_map *map, void *key)
66 struct reuseport_array *array = reuseport_array(map);
67 u32 index = *(u32 *)key;
71 if (index >= map->max_entries)
74 if (!rcu_access_pointer(array->ptrs[index]))
77 spin_lock_bh(&reuseport_lock);
79 sk = rcu_dereference_protected(array->ptrs[index],
80 lockdep_is_held(&reuseport_lock));
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);
91 spin_unlock_bh(&reuseport_lock);
96 static void reuseport_array_free(struct bpf_map *map)
98 struct reuseport_array *array = reuseport_array(map);
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().
110 * This function and bpf_sk_reuseport_detach() are
111 * both removing sk from "array". Who removes it
112 * first does not matter.
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".
121 * Hence, due to the above, taking "reuseport_lock" is not
126 * Since reuseport_lock is not taken, sk is accessed under
130 for (i = 0; i < map->max_entries; i++) {
131 sk = rcu_dereference(array->ptrs[i]);
133 write_lock_bh(&sk->sk_callback_lock);
135 * No need for WRITE_ONCE(). At this point,
136 * no one is reading it without taking the
137 * sk->sk_callback_lock.
139 sk->sk_user_data = NULL;
140 write_unlock_bh(&sk->sk_callback_lock);
141 RCU_INIT_POINTER(array->ptrs[i], NULL);
147 * Once reaching here, all sk->sk_user_data is not
148 * referenceing this "array". "array" can be freed now.
150 bpf_map_area_free(array);
153 static struct bpf_map *reuseport_array_alloc(union bpf_attr *attr)
155 int err, numa_node = bpf_map_attr_numa_node(attr);
156 struct reuseport_array *array;
157 struct bpf_map_memory mem;
161 return ERR_PTR(-EPERM);
163 array_size = sizeof(*array);
164 array_size += (u64)attr->max_entries * sizeof(struct sock *);
166 err = bpf_map_charge_init(&mem, array_size);
170 /* allocate all map elements and zero-initialize them */
171 array = bpf_map_area_alloc(array_size, numa_node);
173 bpf_map_charge_finish(&mem);
174 return ERR_PTR(-ENOMEM);
177 /* copy mandatory map attributes */
178 bpf_map_init_from_attr(&array->map, attr);
179 bpf_map_charge_move(&array->map.memory, &mem);
184 int bpf_fd_reuseport_array_lookup_elem(struct bpf_map *map, void *key,
190 if (map->value_size != sizeof(u64))
194 sk = reuseport_array_lookup_elem(map, key);
196 *(u64 *)value = sock_gen_cookie(sk);
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,
213 if (osk && map_flags == BPF_NOEXIST)
216 if (!osk && map_flags == BPF_EXIST)
219 if (nsk->sk_protocol != IPPROTO_UDP && nsk->sk_protocol != IPPROTO_TCP)
222 if (nsk->sk_family != AF_INET && nsk->sk_family != AF_INET6)
225 if (nsk->sk_type != SOCK_STREAM && nsk->sk_type != SOCK_DGRAM)
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).
233 * Also, sk will be used in bpf helper that is protected by
236 if (!sock_flag(nsk, SOCK_RCU_FREE) || !sk_hashed(nsk) || !nsk_reuse)
239 /* READ_ONCE because the sk->sk_callback_lock may not be held here */
240 if (READ_ONCE(nsk->sk_user_data))
247 * Called from syscall only.
248 * The "nsk" in the fd refcnt.
249 * The "osk" and "reuse" are protected by reuseport_lock.
251 int bpf_fd_reuseport_array_update_elem(struct bpf_map *map, void *key,
252 void *value, u64 map_flags)
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;
262 if (map_flags > BPF_EXIST)
265 if (index >= map->max_entries)
268 if (map->value_size == sizeof(u64)) {
269 u64 fd64 = *(u64 *)value;
278 socket = sockfd_lookup(fd, &err);
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),
296 spin_lock_bh(&reuseport_lock);
298 * Some of the checks only need reuseport_lock
299 * but it is done under sk_callback_lock also
300 * for simplicity reason.
302 write_lock_bh(&nsk->sk_callback_lock);
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);
310 goto put_file_unlock;
312 sk_user_data = (uintptr_t)&array->ptrs[index] | SK_USER_DATA_NOCOPY |
314 WRITE_ONCE(nsk->sk_user_data, (void *)sk_user_data);
315 rcu_assign_pointer(array->ptrs[index], nsk);
320 write_unlock_bh(&nsk->sk_callback_lock);
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);
328 spin_unlock_bh(&reuseport_lock);
334 /* Called from syscall */
335 static int reuseport_array_get_next_key(struct bpf_map *map, void *key,
338 struct reuseport_array *array = reuseport_array(map);
339 u32 index = key ? *(u32 *)key : U32_MAX;
340 u32 *next = (u32 *)next_key;
342 if (index >= array->map.max_entries) {
347 if (index == array->map.max_entries - 1)
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,