net/core/sock_reuseport.c

   1 /*
   2  * To speed up listener socket lookup, create an array to store all sockets
   3  * listening on the same port.  This allows a decision to be made after finding
   4  * the first socket.  An optional BPF program can also be configured for
   5  * selecting the socket index from the array of available sockets.
   6  */
   7
   8 #include <net/sock_reuseport.h>
   9 #include <linux/bpf.h>
  10 #include <linux/rcupdate.h>
  11
  12 #define INIT_SOCKS 128
  13
  14 static DEFINE_SPINLOCK(reuseport_lock);
  15
  16 static struct sock_reuseport *__reuseport_alloc(u16 max_socks)
  17 {
  18         size_t size = sizeof(struct sock_reuseport) +
  19                       sizeof(struct sock *) * max_socks;
  20         struct sock_reuseport *reuse = kzalloc(size, GFP_ATOMIC);
  21
  22         if (!reuse)
  23                 return NULL;
  24
  25         reuse->max_socks = max_socks;
  26
  27         RCU_INIT_POINTER(reuse->prog, NULL);
  28         return reuse;
  29 }
  30
  31 int reuseport_alloc(struct sock *sk)
  32 {
  33         struct sock_reuseport *reuse;
  34
  35         /* bh lock used since this function call may precede hlist lock in
  36          * soft irq of receive path or setsockopt from process context
  37          */
  38         spin_lock_bh(&reuseport_lock);
  39         WARN_ONCE(rcu_dereference_protected(sk->sk_reuseport_cb,
  40                                             lockdep_is_held(&reuseport_lock)),
  41                   "multiple allocations for the same socket");
  42         reuse = __reuseport_alloc(INIT_SOCKS);
  43         if (!reuse) {
  44                 spin_unlock_bh(&reuseport_lock);
  45                 return -ENOMEM;
  46         }
  47
  48         reuse->socks[0] = sk;
  49         reuse->num_socks = 1;
  50         rcu_assign_pointer(sk->sk_reuseport_cb, reuse);
  51
  52         spin_unlock_bh(&reuseport_lock);
  53
  54         return 0;
  55 }
  56 EXPORT_SYMBOL(reuseport_alloc);
  57
  58 static struct sock_reuseport *reuseport_grow(struct sock_reuseport *reuse)
  59 {
  60         struct sock_reuseport *more_reuse;
  61         u32 more_socks_size, i;
  62
  63         more_socks_size = reuse->max_socks * 2U;
  64         if (more_socks_size > U16_MAX)
  65                 return NULL;
  66
  67         more_reuse = __reuseport_alloc(more_socks_size);
  68         if (!more_reuse)
  69                 return NULL;
  70
  71         more_reuse->max_socks = more_socks_size;
  72         more_reuse->num_socks = reuse->num_socks;
  73         more_reuse->prog = reuse->prog;
  74
  75         memcpy(more_reuse->socks, reuse->socks,
  76                reuse->num_socks * sizeof(struct sock *));
  77
  78         for (i = 0; i < reuse->num_socks; ++i)
  79                 rcu_assign_pointer(reuse->socks[i]->sk_reuseport_cb,
  80                                    more_reuse);
  81
  82         /* Note: we use kfree_rcu here instead of reuseport_free_rcu so
  83          * that reuse and more_reuse can temporarily share a reference
  84          * to prog.
  85          */
  86         kfree_rcu(reuse, rcu);
  87         return more_reuse;
  88 }
  89
  90 /**
  91  *  reuseport_add_sock - Add a socket to the reuseport group of another.
  92  *  @sk:  New socket to add to the group.
  93  *  @sk2: Socket belonging to the existing reuseport group.
  94  *  May return ENOMEM and not add socket to group under memory pressure.
  95  */
  96 int reuseport_add_sock(struct sock *sk, struct sock *sk2)
  97 {
  98         struct sock_reuseport *reuse;
  99
 100         if (!rcu_access_pointer(sk2->sk_reuseport_cb)) {
 101                 int err = reuseport_alloc(sk2);
 102
 103                 if (err)
 104                         return err;
 105         }
 106
 107         spin_lock_bh(&reuseport_lock);
 108         reuse = rcu_dereference_protected(sk2->sk_reuseport_cb,
 109                                           lockdep_is_held(&reuseport_lock)),
 110         WARN_ONCE(rcu_dereference_protected(sk->sk_reuseport_cb,
 111                                             lockdep_is_held(&reuseport_lock)),
 112                   "socket already in reuseport group");
 113
 114         if (reuse->num_socks == reuse->max_socks) {
 115                 reuse = reuseport_grow(reuse);
 116                 if (!reuse) {
 117                         spin_unlock_bh(&reuseport_lock);
 118                         return -ENOMEM;
 119                 }
 120         }
 121
 122         reuse->socks[reuse->num_socks] = sk;
 123         /* paired with smp_rmb() in reuseport_select_sock() */
 124         smp_wmb();
 125         reuse->num_socks++;
 126         rcu_assign_pointer(sk->sk_reuseport_cb, reuse);
 127
 128         spin_unlock_bh(&reuseport_lock);
 129
 130         return 0;
 131 }
 132
 133 static void reuseport_free_rcu(struct rcu_head *head)
 134 {
 135         struct sock_reuseport *reuse;
 136
 137         reuse = container_of(head, struct sock_reuseport, rcu);
 138         if (reuse->prog)
 139                 bpf_prog_destroy(reuse->prog);
 140         kfree(reuse);
 141 }
 142
 143 void reuseport_detach_sock(struct sock *sk)
 144 {
 145         struct sock_reuseport *reuse;
 146         int i;
 147
 148         spin_lock_bh(&reuseport_lock);
 149         reuse = rcu_dereference_protected(sk->sk_reuseport_cb,
 150                                           lockdep_is_held(&reuseport_lock));
 151         rcu_assign_pointer(sk->sk_reuseport_cb, NULL);
 152
 153         for (i = 0; i < reuse->num_socks; i++) {
 154                 if (reuse->socks[i] == sk) {
 155                         reuse->socks[i] = reuse->socks[reuse->num_socks - 1];
 156                         reuse->num_socks--;
 157                         if (reuse->num_socks == 0)
 158                                 call_rcu(&reuse->rcu, reuseport_free_rcu);
 159                         break;
 160                 }
 161         }
 162         spin_unlock_bh(&reuseport_lock);
 163 }
 164 EXPORT_SYMBOL(reuseport_detach_sock);
 165
 166 static struct sock *run_bpf(struct sock_reuseport *reuse, u16 socks,
 167                             struct bpf_prog *prog, struct sk_buff *skb,
 168                             int hdr_len)
 169 {
 170         struct sk_buff *nskb = NULL;
 171         u32 index;
 172
 173         if (skb_shared(skb)) {
 174                 nskb = skb_clone(skb, GFP_ATOMIC);
 175                 if (!nskb)
 176                         return NULL;
 177                 skb = nskb;
 178         }
 179
 180         /* temporarily advance data past protocol header */
 181         if (!pskb_pull(skb, hdr_len)) {
 182                 kfree_skb(nskb);
 183                 return NULL;
 184         }
 185         index = bpf_prog_run_save_cb(prog, skb);
 186         __skb_push(skb, hdr_len);
 187
 188         consume_skb(nskb);
 189
 190         if (index >= socks)
 191                 return NULL;
 192
 193         return reuse->socks[index];
 194 }
 195
 196 /**
 197  *  reuseport_select_sock - Select a socket from an SO_REUSEPORT group.
 198  *  @sk: First socket in the group.
 199  *  @hash: When no BPF filter is available, use this hash to select.
 200  *  @skb: skb to run through BPF filter.
 201  *  @hdr_len: BPF filter expects skb data pointer at payload data.  If
 202  *    the skb does not yet point at the payload, this parameter represents
 203  *    how far the pointer needs to advance to reach the payload.
 204  *  Returns a socket that should receive the packet (or NULL on error).
 205  */
 206 struct sock *reuseport_select_sock(struct sock *sk,
 207                                    u32 hash,
 208                                    struct sk_buff *skb,
 209                                    int hdr_len)
 210 {
 211         struct sock_reuseport *reuse;
 212         struct bpf_prog *prog;
 213         struct sock *sk2 = NULL;
 214         u16 socks;
 215
 216         rcu_read_lock();
 217         reuse = rcu_dereference(sk->sk_reuseport_cb);
 218
 219         /* if memory allocation failed or add call is not yet complete */
 220         if (!reuse)
 221                 goto out;
 222
 223         prog = rcu_dereference(reuse->prog);
 224         socks = READ_ONCE(reuse->num_socks);
 225         if (likely(socks)) {
 226                 /* paired with smp_wmb() in reuseport_add_sock() */
 227                 smp_rmb();
 228
 229                 if (prog && skb)
 230                         sk2 = run_bpf(reuse, socks, prog, skb, hdr_len);
 231                 else
 232                         sk2 = reuse->socks[reciprocal_scale(hash, socks)];
 233         }
 234
 235 out:
 236         rcu_read_unlock();
 237         return sk2;
 238 }
 239 EXPORT_SYMBOL(reuseport_select_sock);
 240
 241 struct bpf_prog *
 242 reuseport_attach_prog(struct sock *sk, struct bpf_prog *prog)
 243 {
 244         struct sock_reuseport *reuse;
 245         struct bpf_prog *old_prog;
 246
 247         spin_lock_bh(&reuseport_lock);
 248         reuse = rcu_dereference_protected(sk->sk_reuseport_cb,
 249                                           lockdep_is_held(&reuseport_lock));
 250         old_prog = rcu_dereference_protected(reuse->prog,
 251                                              lockdep_is_held(&reuseport_lock));
 252         rcu_assign_pointer(reuse->prog, prog);
 253         spin_unlock_bh(&reuseport_lock);
 254
 255         return old_prog;
 256 }
 257 EXPORT_SYMBOL(reuseport_attach_prog);