Patchwork [net-next,3/4] bnx2: Remove some unnecessary smp_mb() in tx fast path.

login
register
mail settings
Submitter Michael Chan
Date July 20, 2010, 12:15 a.m.
Message ID <1279584905-15084-3-git-send-email-mchan@broadcom.com>
Download mbox | patch
Permalink /patch/59240/
State Accepted
Delegated to: David Miller
Headers show

Comments

Michael Chan - July 20, 2010, 12:15 a.m.
smp_mb() inside bnx2_tx_avail() is used twice in the normal
bnx2_start_xmit() path (see illustration below).  The full memory
barrier is only necessary during race conditions with tx completion.
We can speed up the tx path by replacing smp_mb() in bnx2_tx_avail()
with a compiler barrier.  The compiler barrier is to force the
compiler to fetch the tx_prod and tx_cons from memory.

In the race condition between bnx2_start_xmit() and bnx2_tx_int(),
we have the following situation:

bnx2_start_xmit()                       bnx2_tx_int()
    if (!bnx2_tx_avail())
            BUG();

    ...

    if (!bnx2_tx_avail())
            netif_tx_stop_queue();          update_tx_index();
            smp_mb();                       smp_mb();
            if (bnx2_tx_avail())            if (netif_tx_queue_stopped() &&
                    netif_tx_wake_queue();      bnx2_tx_avail())

With smp_mb() removed from bnx2_tx_avail(), we need to add smp_mb() to
bnx2_start_xmit() as shown above to properly order netif_tx_stop_queue()
and bnx2_tx_avail() to check the ring index.  If it is not strictly
ordered, the tx queue can be stopped forever.

This improves performance by about 5% with 2 ports running bi-directional
64-byte packets.

Reviewed-by: Benjamin Li <benli@broadcom.com>
Reviewed-by: Matt Carlson <mcarlson@broadcom.com>
Signed-off-by: Michael Chan <mchan@broadcom.com>
---
 drivers/net/bnx2.c |   10 +++++++++-
 1 files changed, 9 insertions(+), 1 deletions(-)
David Miller - July 20, 2010, 3:31 a.m.
From: "Michael Chan" <mchan@broadcom.com>
Date: Mon, 19 Jul 2010 17:15:04 -0700

> smp_mb() inside bnx2_tx_avail() is used twice in the normal
> bnx2_start_xmit() path (see illustration below).  The full memory
> barrier is only necessary during race conditions with tx completion.
> We can speed up the tx path by replacing smp_mb() in bnx2_tx_avail()
> with a compiler barrier.  The compiler barrier is to force the
> compiler to fetch the tx_prod and tx_cons from memory.
> 
> In the race condition between bnx2_start_xmit() and bnx2_tx_int(),
> we have the following situation:
> 
> bnx2_start_xmit()                       bnx2_tx_int()
>     if (!bnx2_tx_avail())
>             BUG();
> 
>     ...
> 
>     if (!bnx2_tx_avail())
>             netif_tx_stop_queue();          update_tx_index();
>             smp_mb();                       smp_mb();
>             if (bnx2_tx_avail())            if (netif_tx_queue_stopped() &&
>                     netif_tx_wake_queue();      bnx2_tx_avail())
> 
> With smp_mb() removed from bnx2_tx_avail(), we need to add smp_mb() to
> bnx2_start_xmit() as shown above to properly order netif_tx_stop_queue()
> and bnx2_tx_avail() to check the ring index.  If it is not strictly
> ordered, the tx queue can be stopped forever.
> 
> This improves performance by about 5% with 2 ports running bi-directional
> 64-byte packets.
> 
> Reviewed-by: Benjamin Li <benli@broadcom.com>
> Reviewed-by: Matt Carlson <mcarlson@broadcom.com>
> Signed-off-by: Michael Chan <mchan@broadcom.com>

Applied.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Dumazet - July 20, 2010, 5:38 a.m.
Le lundi 19 juillet 2010 à 17:15 -0700, Michael Chan a écrit :
> smp_mb() inside bnx2_tx_avail() is used twice in the normal
> bnx2_start_xmit() path (see illustration below).  The full memory
> barrier is only necessary during race conditions with tx completion.
> We can speed up the tx path by replacing smp_mb() in bnx2_tx_avail()
> with a compiler barrier.  The compiler barrier is to force the
> compiler to fetch the tx_prod and tx_cons from memory.
> 
> In the race condition between bnx2_start_xmit() and bnx2_tx_int(),
> we have the following situation:
> 
> bnx2_start_xmit()                       bnx2_tx_int()
>     if (!bnx2_tx_avail())
>             BUG();
> 
>     ...
> 
>     if (!bnx2_tx_avail())
>             netif_tx_stop_queue();          update_tx_index();
>             smp_mb();                       smp_mb();
>             if (bnx2_tx_avail())            if (netif_tx_queue_stopped() &&
>                     netif_tx_wake_queue();      bnx2_tx_avail())
> 
> With smp_mb() removed from bnx2_tx_avail(), we need to add smp_mb() to
> bnx2_start_xmit() as shown above to properly order netif_tx_stop_queue()
> and bnx2_tx_avail() to check the ring index.  If it is not strictly
> ordered, the tx queue can be stopped forever.
> 
> This improves performance by about 5% with 2 ports running bi-directional
> 64-byte packets.
> 
> Reviewed-by: Benjamin Li <benli@broadcom.com>
> Reviewed-by: Matt Carlson <mcarlson@broadcom.com>
> Signed-off-by: Michael Chan <mchan@broadcom.com>
> ---
>  drivers/net/bnx2.c |   10 +++++++++-
>  1 files changed, 9 insertions(+), 1 deletions(-)
> 
> diff --git a/drivers/net/bnx2.c b/drivers/net/bnx2.c
> index d44ecc3..2af570d 100644
> --- a/drivers/net/bnx2.c
> +++ b/drivers/net/bnx2.c
> @@ -253,7 +253,8 @@ static inline u32 bnx2_tx_avail(struct bnx2 *bp, struct bnx2_tx_ring_info *txr)
>  {
>  	u32 diff;
>  
> -	smp_mb();
> +	/* Tell compiler to fetch tx_prod and tx_cons from memory. */
> +	barrier();
>  
>  	/* The ring uses 256 indices for 255 entries, one of them
>  	 * needs to be skipped.
> @@ -6534,6 +6535,13 @@ bnx2_start_xmit(struct sk_buff *skb, struct net_device *dev)
>  
>  	if (unlikely(bnx2_tx_avail(bp, txr) <= MAX_SKB_FRAGS)) {
>  		netif_tx_stop_queue(txq);
> +
> +		/* netif_tx_stop_queue() must be done before checking
> +		 * tx index in bnx2_tx_avail() below, because in
> +		 * bnx2_tx_int(), we update tx index before checking for
> +		 * netif_tx_queue_stopped().
> +		 */
> +		smp_mb();
>  		if (bnx2_tx_avail(bp, txr) > bp->tx_wake_thresh)
>  			netif_tx_wake_queue(txq);
>  	}

Excellent,

Is similar patch for tg3 planned ?

Thanks



--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Michael Chan - July 20, 2010, 5:48 a.m.
Eric Dumazet wrote:

> Excellent,
>
> Is similar patch for tg3 planned ?
>

Yes, this performance issue was actually first discovered when
profiling a new tg3 device.  So Matt should be sending out a
very similar patch as part of his next patch-set.

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/drivers/net/bnx2.c b/drivers/net/bnx2.c
index d44ecc3..2af570d 100644
--- a/drivers/net/bnx2.c
+++ b/drivers/net/bnx2.c
@@ -253,7 +253,8 @@  static inline u32 bnx2_tx_avail(struct bnx2 *bp, struct bnx2_tx_ring_info *txr)
 {
 	u32 diff;
 
-	smp_mb();
+	/* Tell compiler to fetch tx_prod and tx_cons from memory. */
+	barrier();
 
 	/* The ring uses 256 indices for 255 entries, one of them
 	 * needs to be skipped.
@@ -6534,6 +6535,13 @@  bnx2_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	if (unlikely(bnx2_tx_avail(bp, txr) <= MAX_SKB_FRAGS)) {
 		netif_tx_stop_queue(txq);
+
+		/* netif_tx_stop_queue() must be done before checking
+		 * tx index in bnx2_tx_avail() below, because in
+		 * bnx2_tx_int(), we update tx index before checking for
+		 * netif_tx_queue_stopped().
+		 */
+		smp_mb();
 		if (bnx2_tx_avail(bp, txr) > bp->tx_wake_thresh)
 			netif_tx_wake_queue(txq);
 	}