net: netif_tx_queue_stopped too expensive
authorEric Dumazet <dada1@cosmosbay.com>
Tue, 28 Apr 2009 11:43:42 +0000 (04:43 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 28 Apr 2009 11:43:42 +0000 (04:43 -0700)
netif_tx_queue_stopped(txq) is most of the time false.

Yet its cost is very expensive on SMP.

static inline int netif_tx_queue_stopped(const struct netdev_queue *dev_queue)
{
return test_bit(__QUEUE_STATE_XOFF, &dev_queue->state);
}

I saw this on oprofile hunting and bnx2 driver bnx2_tx_int().

We probably should split "struct netdev_queue" in two parts, one
being read mostly.

__netif_tx_lock() touches _xmit_lock & xmit_lock_owner, these
deserve a separate cache line.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/netdevice.h

index f8c3619d551f69ac0b34ed42e96713e622b74a89..505a3c6cb12dcd342d370c435714eedacaa7b34a 100644 (file)
@@ -447,12 +447,18 @@ enum netdev_queue_state_t
 };
 
 struct netdev_queue {
+/*
+ * read mostly part
+ */
        struct net_device       *dev;
        struct Qdisc            *qdisc;
        unsigned long           state;
-       spinlock_t              _xmit_lock;
-       int                     xmit_lock_owner;
        struct Qdisc            *qdisc_sleeping;
+/*
+ * write mostly part
+ */
+       spinlock_t              _xmit_lock ____cacheline_aligned_in_smp;
+       int                     xmit_lock_owner;
 } ____cacheline_aligned_in_smp;