kernel/bpf/reuseport_array.c

   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         /*
 103          * ops->map_*_elem() will not be able to access this
 104          * array now. Hence, this function only races with
 105          * bpf_sk_reuseport_detach() which was triggerred by
 106          * close() or disconnect().
 107          *
 108          * This function and bpf_sk_reuseport_detach() are
 109          * both removing sk from "array".  Who removes it
 110          * first does not matter.
 111          *
 112          * The only concern here is bpf_sk_reuseport_detach()
 113          * may access "array" which is being freed here.
 114          * bpf_sk_reuseport_detach() access this "array"
 115          * through sk->sk_user_data _and_ with sk->sk_callback_lock
 116          * held which is enough because this "array" is not freed
 117          * until all sk->sk_user_data has stopped referencing this "array".
 118          *
 119          * Hence, due to the above, taking "reuseport_lock" is not
 120          * needed here.
 121          */
 122
 123         /*
 124          * Since reuseport_lock is not taken, sk is accessed under
 125          * rcu_read_lock()
 126          */
 127         rcu_read_lock();
 128         for (i = 0; i < map->max_entries; i++) {
 129                 sk = rcu_dereference(array->ptrs[i]);
 130                 if (sk) {
 131                         write_lock_bh(&sk->sk_callback_lock);
 132                         /*
 133                          * No need for WRITE_ONCE(). At this point,
 134                          * no one is reading it without taking the
 135                          * sk->sk_callback_lock.
 136                          */
 137                         sk->sk_user_data = NULL;
 138                         write_unlock_bh(&sk->sk_callback_lock);
 139                         RCU_INIT_POINTER(array->ptrs[i], NULL);
 140                 }
 141         }
 142         rcu_read_unlock();
 143
 144         /*
 145          * Once reaching here, all sk->sk_user_data is not
 146          * referenceing this "array".  "array" can be freed now.
 147          */
 148         bpf_map_area_free(array);
 149 }
 150
 151 static struct bpf_map *reuseport_array_alloc(union bpf_attr *attr)
 152 {
 153         int numa_node = bpf_map_attr_numa_node(attr);
 154         struct reuseport_array *array;
 155         u64 array_size;
 156
 157         if (!bpf_capable())
 158                 return ERR_PTR(-EPERM);
 159
 160         array_size = sizeof(*array);
 161         array_size += (u64)attr->max_entries * sizeof(struct sock *);
 162
 163         /* allocate all map elements and zero-initialize them */
 164         array = bpf_map_area_alloc(array_size, numa_node);
 165         if (!array)
 166                 return ERR_PTR(-ENOMEM);
 167
 168         /* copy mandatory map attributes */
 169         bpf_map_init_from_attr(&array->map, attr);
 170
 171         return &array->map;
 172 }
 173
 174 int bpf_fd_reuseport_array_lookup_elem(struct bpf_map *map, void *key,
 175                                        void *value)
 176 {
 177         struct sock *sk;
 178         int err;
 179
 180         if (map->value_size != sizeof(u64))
 181                 return -ENOSPC;
 182
 183         rcu_read_lock();
 184         sk = reuseport_array_lookup_elem(map, key);
 185         if (sk) {
 186                 *(u64 *)value = __sock_gen_cookie(sk);
 187                 err = 0;
 188         } else {
 189                 err = -ENOENT;
 190         }
 191         rcu_read_unlock();
 192
 193         return err;
 194 }
 195
 196 static int
 197 reuseport_array_update_check(const struct reuseport_array *array,
 198                              const struct sock *nsk,
 199                              const struct sock *osk,
 200                              const struct sock_reuseport *nsk_reuse,
 201                              u32 map_flags)
 202 {
 203         if (osk && map_flags == BPF_NOEXIST)
 204                 return -EEXIST;
 205
 206         if (!osk && map_flags == BPF_EXIST)
 207                 return -ENOENT;
 208
 209         if (nsk->sk_protocol != IPPROTO_UDP && nsk->sk_protocol != IPPROTO_TCP)
 210                 return -ENOTSUPP;
 211
 212         if (nsk->sk_family != AF_INET && nsk->sk_family != AF_INET6)
 213                 return -ENOTSUPP;
 214
 215         if (nsk->sk_type != SOCK_STREAM && nsk->sk_type != SOCK_DGRAM)
 216                 return -ENOTSUPP;
 217
 218         /*
 219          * sk must be hashed (i.e. listening in the TCP case or binded
 220          * in the UDP case) and
 221          * it must also be a SO_REUSEPORT sk (i.e. reuse cannot be NULL).
 222          *
 223          * Also, sk will be used in bpf helper that is protected by
 224          * rcu_read_lock().
 225          */
 226         if (!sock_flag(nsk, SOCK_RCU_FREE) || !sk_hashed(nsk) || !nsk_reuse)
 227                 return -EINVAL;
 228
 229         /* READ_ONCE because the sk->sk_callback_lock may not be held here */
 230         if (READ_ONCE(nsk->sk_user_data))
 231                 return -EBUSY;
 232
 233         return 0;
 234 }
 235
 236 /*
 237  * Called from syscall only.
 238  * The "nsk" in the fd refcnt.
 239  * The "osk" and "reuse" are protected by reuseport_lock.
 240  */
 241 int bpf_fd_reuseport_array_update_elem(struct bpf_map *map, void *key,
 242                                        void *value, u64 map_flags)
 243 {
 244         struct reuseport_array *array = reuseport_array(map);
 245         struct sock *free_osk = NULL, *osk, *nsk;
 246         struct sock_reuseport *reuse;
 247         u32 index = *(u32 *)key;
 248         uintptr_t sk_user_data;
 249         struct socket *socket;
 250         int err, fd;
 251
 252         if (map_flags > BPF_EXIST)
 253                 return -EINVAL;
 254
 255         if (index >= map->max_entries)
 256                 return -E2BIG;
 257
 258         if (map->value_size == sizeof(u64)) {
 259                 u64 fd64 = *(u64 *)value;
 260
 261                 if (fd64 > S32_MAX)
 262                         return -EINVAL;
 263                 fd = fd64;
 264         } else {
 265                 fd = *(int *)value;
 266         }
 267
 268         socket = sockfd_lookup(fd, &err);
 269         if (!socket)
 270                 return err;
 271
 272         nsk = socket->sk;
 273         if (!nsk) {
 274                 err = -EINVAL;
 275                 goto put_file;
 276         }
 277
 278         /* Quick checks before taking reuseport_lock */
 279         err = reuseport_array_update_check(array, nsk,
 280                                            rcu_access_pointer(array->ptrs[index]),
 281                                            rcu_access_pointer(nsk->sk_reuseport_cb),
 282                                            map_flags);
 283         if (err)
 284                 goto put_file;
 285
 286         spin_lock_bh(&reuseport_lock);
 287         /*
 288          * Some of the checks only need reuseport_lock
 289          * but it is done under sk_callback_lock also
 290          * for simplicity reason.
 291          */
 292         write_lock_bh(&nsk->sk_callback_lock);
 293
 294         osk = rcu_dereference_protected(array->ptrs[index],
 295                                         lockdep_is_held(&reuseport_lock));
 296         reuse = rcu_dereference_protected(nsk->sk_reuseport_cb,
 297                                           lockdep_is_held(&reuseport_lock));
 298         err = reuseport_array_update_check(array, nsk, osk, reuse, map_flags);
 299         if (err)
 300                 goto put_file_unlock;
 301
 302         sk_user_data = (uintptr_t)&array->ptrs[index] | SK_USER_DATA_NOCOPY |
 303                 SK_USER_DATA_BPF;
 304         WRITE_ONCE(nsk->sk_user_data, (void *)sk_user_data);
 305         rcu_assign_pointer(array->ptrs[index], nsk);
 306         free_osk = osk;
 307         err = 0;
 308
 309 put_file_unlock:
 310         write_unlock_bh(&nsk->sk_callback_lock);
 311
 312         if (free_osk) {
 313                 write_lock_bh(&free_osk->sk_callback_lock);
 314                 WRITE_ONCE(free_osk->sk_user_data, NULL);
 315                 write_unlock_bh(&free_osk->sk_callback_lock);
 316         }
 317
 318         spin_unlock_bh(&reuseport_lock);
 319 put_file:
 320         fput(socket->file);
 321         return err;
 322 }
 323
 324 /* Called from syscall */
 325 static int reuseport_array_get_next_key(struct bpf_map *map, void *key,
 326                                         void *next_key)
 327 {
 328         struct reuseport_array *array = reuseport_array(map);
 329         u32 index = key ? *(u32 *)key : U32_MAX;
 330         u32 *next = (u32 *)next_key;
 331
 332         if (index >= array->map.max_entries) {
 333                 *next = 0;
 334                 return 0;
 335         }
 336
 337         if (index == array->map.max_entries - 1)
 338                 return -ENOENT;
 339
 340         *next = index + 1;
 341         return 0;
 342 }
 343
 344 static int reuseport_array_map_btf_id;
 345 const struct bpf_map_ops reuseport_array_ops = {
 346         .map_meta_equal = bpf_map_meta_equal,
 347         .map_alloc_check = reuseport_array_alloc_check,
 348         .map_alloc = reuseport_array_alloc,
 349         .map_free = reuseport_array_free,
 350         .map_lookup_elem = reuseport_array_lookup_elem,
 351         .map_get_next_key = reuseport_array_get_next_key,
 352         .map_delete_elem = reuseport_array_delete_elem,
 353         .map_btf_name = "reuseport_array",
 354         .map_btf_id = &reuseport_array_map_btf_id,
 355 };