net/core/sock_reuseport.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * To speed up listener socket lookup, create an array to store all sockets
   4  * listening on the same port.  This allows a decision to be made after finding
   5  * the first socket.  An optional BPF program can also be configured for
   6  * selecting the socket index from the array of available sockets.
   7  */
   8
   9 #include <net/sock_reuseport.h>
  10 #include <linux/bpf.h>
  11 #include <linux/idr.h>
  12 #include <linux/filter.h>
  13 #include <linux/rcupdate.h>
  14
  15 #define INIT_SOCKS 128
  16
  17 DEFINE_SPINLOCK(reuseport_lock);
  18
  19 #define REUSEPORT_MIN_ID 1
  20 static DEFINE_IDA(reuseport_ida);
  21
  22 int reuseport_get_id(struct sock_reuseport *reuse)
  23 {
  24         int id;
  25
  26         if (reuse->reuseport_id)
  27                 return reuse->reuseport_id;
  28
  29         id = ida_simple_get(&reuseport_ida, REUSEPORT_MIN_ID, 0,
  30                             /* Called under reuseport_lock */
  31                             GFP_ATOMIC);
  32         if (id < 0)
  33                 return id;
  34
  35         reuse->reuseport_id = id;
  36
  37         return reuse->reuseport_id;
  38 }
  39
  40 static struct sock_reuseport *__reuseport_alloc(unsigned int max_socks)
  41 {
  42         unsigned int size = sizeof(struct sock_reuseport) +
  43                       sizeof(struct sock *) * max_socks;
  44         struct sock_reuseport *reuse = kzalloc(size, GFP_ATOMIC);
  45
  46         if (!reuse)
  47                 return NULL;
  48
  49         reuse->max_socks = max_socks;
  50
  51         RCU_INIT_POINTER(reuse->prog, NULL);
  52         return reuse;
  53 }
  54
  55 int reuseport_alloc(struct sock *sk, bool bind_inany)
  56 {
  57         struct sock_reuseport *reuse;
  58
  59         /* bh lock used since this function call may precede hlist lock in
  60          * soft irq of receive path or setsockopt from process context
  61          */
  62         spin_lock_bh(&reuseport_lock);
  63
  64         /* Allocation attempts can occur concurrently via the setsockopt path
  65          * and the bind/hash path.  Nothing to do when we lose the race.
  66          */
  67         reuse = rcu_dereference_protected(sk->sk_reuseport_cb,
  68                                           lockdep_is_held(&reuseport_lock));
  69         if (reuse) {
  70                 /* Only set reuse->bind_inany if the bind_inany is true.
  71                  * Otherwise, it will overwrite the reuse->bind_inany
  72                  * which was set by the bind/hash path.
  73                  */
  74                 if (bind_inany)
  75                         reuse->bind_inany = bind_inany;
  76                 goto out;
  77         }
  78
  79         reuse = __reuseport_alloc(INIT_SOCKS);
  80         if (!reuse) {
  81                 spin_unlock_bh(&reuseport_lock);
  82                 return -ENOMEM;
  83         }
  84
  85         reuse->socks[0] = sk;
  86         reuse->num_socks = 1;
  87         reuse->bind_inany = bind_inany;
  88         rcu_assign_pointer(sk->sk_reuseport_cb, reuse);
  89
  90 out:
  91         spin_unlock_bh(&reuseport_lock);
  92
  93         return 0;
  94 }
  95 EXPORT_SYMBOL(reuseport_alloc);
  96
  97 static struct sock_reuseport *reuseport_grow(struct sock_reuseport *reuse)
  98 {
  99         struct sock_reuseport *more_reuse;
 100         u32 more_socks_size, i;
 101
 102         more_socks_size = reuse->max_socks * 2U;
 103         if (more_socks_size > U16_MAX)
 104                 return NULL;
 105
 106         more_reuse = __reuseport_alloc(more_socks_size);
 107         if (!more_reuse)
 108                 return NULL;
 109
 110         more_reuse->max_socks = more_socks_size;
 111         more_reuse->num_socks = reuse->num_socks;
 112         more_reuse->prog = reuse->prog;
 113         more_reuse->reuseport_id = reuse->reuseport_id;
 114         more_reuse->bind_inany = reuse->bind_inany;
 115
 116         memcpy(more_reuse->socks, reuse->socks,
 117                reuse->num_socks * sizeof(struct sock *));
 118         more_reuse->synq_overflow_ts = READ_ONCE(reuse->synq_overflow_ts);
 119
 120         for (i = 0; i < reuse->num_socks; ++i)
 121                 rcu_assign_pointer(reuse->socks[i]->sk_reuseport_cb,
 122                                    more_reuse);
 123
 124         /* Note: we use kfree_rcu here instead of reuseport_free_rcu so
 125          * that reuse and more_reuse can temporarily share a reference
 126          * to prog.
 127          */
 128         kfree_rcu(reuse, rcu);
 129         return more_reuse;
 130 }
 131
 132 static void reuseport_free_rcu(struct rcu_head *head)
 133 {
 134         struct sock_reuseport *reuse;
 135
 136         reuse = container_of(head, struct sock_reuseport, rcu);
 137         sk_reuseport_prog_free(rcu_dereference_protected(reuse->prog, 1));
 138         if (reuse->reuseport_id)
 139                 ida_simple_remove(&reuseport_ida, reuse->reuseport_id);
 140         kfree(reuse);
 141 }
 142
 143 /**
 144  *  reuseport_add_sock - Add a socket to the reuseport group of another.
 145  *  @sk:  New socket to add to the group.
 146  *  @sk2: Socket belonging to the existing reuseport group.
 147  *  @bind_inany: Whether or not the group is bound to a local INANY address.
 148  *
 149  *  May return ENOMEM and not add socket to group under memory pressure.
 150  */
 151 int reuseport_add_sock(struct sock *sk, struct sock *sk2, bool bind_inany)
 152 {
 153         struct sock_reuseport *old_reuse, *reuse;
 154
 155         if (!rcu_access_pointer(sk2->sk_reuseport_cb)) {
 156                 int err = reuseport_alloc(sk2, bind_inany);
 157
 158                 if (err)
 159                         return err;
 160         }
 161
 162         spin_lock_bh(&reuseport_lock);
 163         reuse = rcu_dereference_protected(sk2->sk_reuseport_cb,
 164                                           lockdep_is_held(&reuseport_lock));
 165         old_reuse = rcu_dereference_protected(sk->sk_reuseport_cb,
 166                                              lockdep_is_held(&reuseport_lock));
 167         if (old_reuse && old_reuse->num_socks != 1) {
 168                 spin_unlock_bh(&reuseport_lock);
 169                 return -EBUSY;
 170         }
 171
 172         if (reuse->num_socks == reuse->max_socks) {
 173                 reuse = reuseport_grow(reuse);
 174                 if (!reuse) {
 175                         spin_unlock_bh(&reuseport_lock);
 176                         return -ENOMEM;
 177                 }
 178         }
 179
 180         reuse->socks[reuse->num_socks] = sk;
 181         /* paired with smp_rmb() in reuseport_select_sock() */
 182         smp_wmb();
 183         reuse->num_socks++;
 184         rcu_assign_pointer(sk->sk_reuseport_cb, reuse);
 185
 186         spin_unlock_bh(&reuseport_lock);
 187
 188         if (old_reuse)
 189                 call_rcu(&old_reuse->rcu, reuseport_free_rcu);
 190         return 0;
 191 }
 192 EXPORT_SYMBOL(reuseport_add_sock);
 193
 194 void reuseport_detach_sock(struct sock *sk)
 195 {
 196         struct sock_reuseport *reuse;
 197         int i;
 198
 199         spin_lock_bh(&reuseport_lock);
 200         reuse = rcu_dereference_protected(sk->sk_reuseport_cb,
 201                                           lockdep_is_held(&reuseport_lock));
 202
 203         /* At least one of the sk in this reuseport group is added to
 204          * a bpf map.  Notify the bpf side.  The bpf map logic will
 205          * remove the sk if it is indeed added to a bpf map.
 206          */
 207         if (reuse->reuseport_id)
 208                 bpf_sk_reuseport_detach(sk);
 209
 210         rcu_assign_pointer(sk->sk_reuseport_cb, NULL);
 211
 212         for (i = 0; i < reuse->num_socks; i++) {
 213                 if (reuse->socks[i] == sk) {
 214                         reuse->socks[i] = reuse->socks[reuse->num_socks - 1];
 215                         reuse->num_socks--;
 216                         if (reuse->num_socks == 0)
 217                                 call_rcu(&reuse->rcu, reuseport_free_rcu);
 218                         break;
 219                 }
 220         }
 221         spin_unlock_bh(&reuseport_lock);
 222 }
 223 EXPORT_SYMBOL(reuseport_detach_sock);
 224
 225 static struct sock *run_bpf_filter(struct sock_reuseport *reuse, u16 socks,
 226                                    struct bpf_prog *prog, struct sk_buff *skb,
 227                                    int hdr_len)
 228 {
 229         struct sk_buff *nskb = NULL;
 230         u32 index;
 231
 232         if (skb_shared(skb)) {
 233                 nskb = skb_clone(skb, GFP_ATOMIC);
 234                 if (!nskb)
 235                         return NULL;
 236                 skb = nskb;
 237         }
 238
 239         /* temporarily advance data past protocol header */
 240         if (!pskb_pull(skb, hdr_len)) {
 241                 kfree_skb(nskb);
 242                 return NULL;
 243         }
 244         index = bpf_prog_run_save_cb(prog, skb);
 245         __skb_push(skb, hdr_len);
 246
 247         consume_skb(nskb);
 248
 249         if (index >= socks)
 250                 return NULL;
 251
 252         return reuse->socks[index];
 253 }
 254
 255 /**
 256  *  reuseport_select_sock - Select a socket from an SO_REUSEPORT group.
 257  *  @sk: First socket in the group.
 258  *  @hash: When no BPF filter is available, use this hash to select.
 259  *  @skb: skb to run through BPF filter.
 260  *  @hdr_len: BPF filter expects skb data pointer at payload data.  If
 261  *    the skb does not yet point at the payload, this parameter represents
 262  *    how far the pointer needs to advance to reach the payload.
 263  *  Returns a socket that should receive the packet (or NULL on error).
 264  */
 265 struct sock *reuseport_select_sock(struct sock *sk,
 266                                    u32 hash,
 267                                    struct sk_buff *skb,
 268                                    int hdr_len)
 269 {
 270         struct sock_reuseport *reuse;
 271         struct bpf_prog *prog;
 272         struct sock *sk2 = NULL;
 273         u16 socks;
 274
 275         rcu_read_lock();
 276         reuse = rcu_dereference(sk->sk_reuseport_cb);
 277
 278         /* if memory allocation failed or add call is not yet complete */
 279         if (!reuse)
 280                 goto out;
 281
 282         prog = rcu_dereference(reuse->prog);
 283         socks = READ_ONCE(reuse->num_socks);
 284         if (likely(socks)) {
 285                 /* paired with smp_wmb() in reuseport_add_sock() */
 286                 smp_rmb();
 287
 288                 if (!prog || !skb)
 289                         goto select_by_hash;
 290
 291                 if (prog->type == BPF_PROG_TYPE_SK_REUSEPORT)
 292                         sk2 = bpf_run_sk_reuseport(reuse, sk, prog, skb, hash);
 293                 else
 294                         sk2 = run_bpf_filter(reuse, socks, prog, skb, hdr_len);
 295
 296 select_by_hash:
 297                 /* no bpf or invalid bpf result: fall back to hash usage */
 298                 if (!sk2)
 299                         sk2 = reuse->socks[reciprocal_scale(hash, socks)];
 300         }
 301
 302 out:
 303         rcu_read_unlock();
 304         return sk2;
 305 }
 306 EXPORT_SYMBOL(reuseport_select_sock);
 307
 308 int reuseport_attach_prog(struct sock *sk, struct bpf_prog *prog)
 309 {
 310         struct sock_reuseport *reuse;
 311         struct bpf_prog *old_prog;
 312
 313         if (sk_unhashed(sk) && sk->sk_reuseport) {
 314                 int err = reuseport_alloc(sk, false);
 315
 316                 if (err)
 317                         return err;
 318         } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
 319                 /* The socket wasn't bound with SO_REUSEPORT */
 320                 return -EINVAL;
 321         }
 322
 323         spin_lock_bh(&reuseport_lock);
 324         reuse = rcu_dereference_protected(sk->sk_reuseport_cb,
 325                                           lockdep_is_held(&reuseport_lock));
 326         old_prog = rcu_dereference_protected(reuse->prog,
 327                                              lockdep_is_held(&reuseport_lock));
 328         rcu_assign_pointer(reuse->prog, prog);
 329         spin_unlock_bh(&reuseport_lock);
 330
 331         sk_reuseport_prog_free(old_prog);
 332         return 0;
 333 }
 334 EXPORT_SYMBOL(reuseport_attach_prog);