tcp: improve REUSEADDR/NOREUSEADDR cohabitation
authorEric Dumazet <edumazet@google.com>
Wed, 20 May 2015 17:59:02 +0000 (10:59 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 21 May 2015 22:55:32 +0000 (18:55 -0400)
inet_csk_get_port() randomization effort tends to spread
sockets on all the available range (ip_local_port_range)

This is unfortunate because SO_REUSEADDR sockets have
less requirements than non SO_REUSEADDR ones.

If an application uses SO_REUSEADDR hint, it is to try to
allow source ports being shared.

So instead of picking a random port number in ip_local_port_range,
lets try first in first half of the range.

This gives more chances to use upper half of the range for the
sockets with strong requirements (not using SO_REUSEADDR)

Note this patch does not add a new sysctl, and only changes
the way we try to pick port number.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Marcelo Ricardo Leitner <mleitner@redhat.com>
Cc: Flavio Leitner <fbl@redhat.com>
Acked-by: Flavio Leitner <fbl@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/inet_connection_sock.c

index b95fb263a13f81132d78125b82a85880e55f2ee5..60021d0d9326ac691dcef21e1f9c20de5f8fe7c6 100644 (file)
@@ -99,6 +99,7 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
        struct net *net = sock_net(sk);
        int smallest_size = -1, smallest_rover;
        kuid_t uid = sock_i_uid(sk);
+       int attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 
        local_bh_disable();
        if (!snum) {
@@ -106,6 +107,14 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum)
 
 again:
                inet_get_local_port_range(net, &low, &high);
+               if (attempt_half) {
+                       int half = low + ((high - low) >> 1);
+
+                       if (attempt_half == 1)
+                               high = half;
+                       else
+                               low = half;
+               }
                remaining = (high - low) + 1;
                smallest_rover = rover = prandom_u32() % remaining + low;
 
@@ -154,6 +163,11 @@ again:
                                snum = smallest_rover;
                                goto have_snum;
                        }
+                       if (attempt_half == 1) {
+                               /* OK we now try the upper half of the range */
+                               attempt_half = 2;
+                               goto again;
+                       }
                        goto fail;
                }
                /* OK, here is the one we will use.  HEAD is