tcp: Combat per-cpu skew in orphan tests.

As reported by Anton Blanchard when we use
percpu_counter_read_positive() to make our orphan socket limit checks,
the check can be off by up to num_cpus_online() * batch (which is 32
by default) which on a 128 cpu machine can be as large as the default
orphan limit itself.

Fix this by doing the full expensive sum check if the optimized check
triggers.

Reported-by: Anton Blanchard <anton@samba.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>

diff --git a/include/net/tcp.h b/include/net/tcp.h
index df6a2eb20193f935bc72204ca22c7f2f21b5070f..eaa9582779d029a9362c32b3816fed8c157e4d3e 100644 (file)
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -268,11 +268,21 @@ static inline int between(__u32 seq1, __u32 seq2, __u32 seq3)
        return seq3 - seq2 >= seq1 - seq2;
 }
 
-static inline int tcp_too_many_orphans(struct sock *sk, int num)
+static inline bool tcp_too_many_orphans(struct sock *sk, int shift)
 {
-       return (num > sysctl_tcp_max_orphans) ||
-               (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
-                atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2]);
+       struct percpu_counter *ocp = sk->sk_prot->orphan_count;
+       int orphans = percpu_counter_read_positive(ocp);
+
+       if (orphans << shift > sysctl_tcp_max_orphans) {
+               orphans = percpu_counter_sum_positive(ocp);
+               if (orphans << shift > sysctl_tcp_max_orphans)
+                       return true;
+       }
+
+       if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
+           atomic_read(&tcp_memory_allocated) > sysctl_tcp_mem[2])
+               return true;
+       return false;
 }
 
 /* syncookies: remember time of last synqueue overflow */

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 176e11aaea771795b21c0be6b1453b46c6349f0c..197b9b77fa3ed1a3c2fe30755394e250d274bd8f 100644 (file)
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -2011,11 +2011,8 @@ adjudge_to_death:
                }
        }
        if (sk->sk_state != TCP_CLOSE) {
-               int orphan_count = percpu_counter_read_positive(
-                                               sk->sk_prot->orphan_count);
-
                sk_mem_reclaim(sk);
-               if (tcp_too_many_orphans(sk, orphan_count)) {
+               if (tcp_too_many_orphans(sk, 0)) {
                        if (net_ratelimit())
                                printk(KERN_INFO "TCP: too many of orphaned "
                                       "sockets\n");

diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c
index 808bb920c9f5e67249f6747c90cdeb7dddd9ddea..c35b469e851c298814d69583bd593ec1c580dde5 100644 (file)
--- a/net/ipv4/tcp_timer.c
+++ b/net/ipv4/tcp_timer.c
@@ -66,18 +66,18 @@ static void tcp_write_err(struct sock *sk)
 static int tcp_out_of_resources(struct sock *sk, int do_reset)
 {
        struct tcp_sock *tp = tcp_sk(sk);
-       int orphans = percpu_counter_read_positive(&tcp_orphan_count);
+       int shift = 0;
 
        /* If peer does not open window for long time, or did not transmit
         * anything for long time, penalize it. */
        if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
-               orphans <<= 1;
+               shift++;
 
        /* If some dubious ICMP arrived, penalize even more. */
        if (sk->sk_err_soft)
-               orphans <<= 1;
+               shift++;
 
-       if (tcp_too_many_orphans(sk, orphans)) {
+       if (tcp_too_many_orphans(sk, shift)) {
                if (net_ratelimit())
                        printk(KERN_INFO "Out of socket memory\n");