udp: udp_rmem_release() should touch sk_rmem_alloc later
authorEric Dumazet <edumazet@google.com>
Thu, 8 Dec 2016 19:41:57 +0000 (11:41 -0800)
committerDavid S. Miller <davem@davemloft.net>
Sat, 10 Dec 2016 03:12:21 +0000 (22:12 -0500)
In flood situations, keeping sk_rmem_alloc at a high value
prevents producers from touching the socket.

It makes sense to lower sk_rmem_alloc only at the end
of udp_rmem_release() after the thread draining receive
queue in udp_recvmsg() finished the writes to sk_forward_alloc.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/udp.c

index 5a38faa12cde7fdcd5b6d86cdc0f4bc33de427bb..9ca279b130d51f6feaa97785b1c90677581080a5 100644 (file)
@@ -1191,13 +1191,14 @@ static void udp_rmem_release(struct sock *sk, int size, int partial)
        }
        up->forward_deficit = 0;
 
-       atomic_sub(size, &sk->sk_rmem_alloc);
        sk->sk_forward_alloc += size;
        amt = (sk->sk_forward_alloc - partial) & ~(SK_MEM_QUANTUM - 1);
        sk->sk_forward_alloc -= amt;
 
        if (amt)
                __sk_mem_reduce_allocated(sk, amt >> SK_MEM_QUANTUM_SHIFT);
+
+       atomic_sub(size, &sk->sk_rmem_alloc);
 }
 
 /* Note: called with sk_receive_queue.lock held.