[PATCH] ethernet/arc/arc_emac: optimize the Tx/Tx-reclaim paths a bit

[PATCH] ethernet/arc/arc_emac: optimize the Tx/Tx-reclaim paths a bit

[Vineet Gupta]

This came out of staring at code due to recent performance fix.

* TX BD reclaim can call netif_wake_queue() once, outside the loop if
  one/more BDs were freed, NO need to do this each iteration.

* TX need not look at next BD to stop the netif queue. It rather be done
  in the next tx call, when it actually fails as the queue seldom gets
  full but the check nevertheless needs to be done for each packet Tx.
  Profiled this under heavy traffic (big tar file cp, LMBench betworking
  tests) and saw not a single hit to that code.

Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
Cc: Alexey Brodkin <abrodkin@synopsys.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Francois Romieu <romieu@fr.zoreil.com>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: arc-linux-dev@synopsys.com
---
 drivers/net/ethernet/arc/emac_main.c | 13 +++----------
 1 file changed, 3 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/arc/emac_main.c b/drivers/net/ethernet/arc/emac_main.c
index 9e16014..a3dd048 100644
--- a/drivers/net/ethernet/arc/emac_main.c
+++ b/drivers/net/ethernet/arc/emac_main.c
@@ -179,10 +179,10 @@ static void arc_emac_tx_clean(struct net_device *ndev)
 		txbd->info = 0;
 
 		*txbd_dirty = (*txbd_dirty + 1) % TX_BD_NUM;
-
-		if (netif_queue_stopped(ndev))
-			netif_wake_queue(ndev);
 	}
+
+	if (i && netif_queue_stopped(ndev))
+		netif_wake_queue(ndev);
 }
 
 /**
@@ -570,13 +570,6 @@ static int arc_emac_tx(struct sk_buff *skb, struct net_device *ndev)
 	/* Increment index to point to the next BD */
 	*txbd_curr = (*txbd_curr + 1) % TX_BD_NUM;
 
-	/* Get "info" of the next BD */
-	info = &priv->txbd[*txbd_curr].info;
-
-	/* Check if if Tx BD ring is full - next BD is still owned by EMAC */
-	if (unlikely((le32_to_cpu(*info) & OWN_MASK) == FOR_EMAC))
-		netif_stop_queue(ndev);
-
 	arc_reg_set(priv, R_STATUS, TXPL_MASK);
 
 	skb_tx_timestamp(skb);
-- 
1.8.1.2

Re: [PATCH] ethernet/arc/arc_emac: optimize the Tx/Tx-reclaim paths a bit

[David Miller]

From: Vineet Gupta <Vineet.Gupta1@synopsys.com>
Date: Wed, 4 Sep 2013 18:33:11 +0530

> This came out of staring at code due to recent performance fix.
> 
> * TX BD reclaim can call netif_wake_queue() once, outside the loop if
>   one/more BDs were freed, NO need to do this each iteration.
> 
> * TX need not look at next BD to stop the netif queue. It rather be done
>   in the next tx call, when it actually fails as the queue seldom gets
>   full but the check nevertheless needs to be done for each packet Tx.
>   Profiled this under heavy traffic (big tar file cp, LMBench betworking
>   tests) and saw not a single hit to that code.
> 
> Signed-off-by: Vineet Gupta <vgupta@synopsys.com>

You should keep the check in the transmit queueing code as a BUG check,
almost every driver has code of the form (using NIU as an example):

	if (niu_tx_avail(rp) <= (skb_shinfo(skb)->nr_frags + 1)) {
		netif_tx_stop_queue(txq);
		dev_err(np->device, "%s: BUG! Tx ring full when queue awake!\n", dev->name);
		rp->tx_errors++;
		return NETDEV_TX_BUSY;
	}

and arc_emac should too.

Otherwise queue management bugs are incredibly hard to diagnose.

I'm not applying this patch.

Re: [PATCH] ethernet/arc/arc_emac: optimize the Tx/Tx-reclaim paths a bit

[Vineet Gupta]

Hi David,

On 09/05/2013 11:54 PM, David Miller wrote:
> From: Vineet Gupta <Vineet.Gupta1@synopsys.com>
> Date: Wed, 4 Sep 2013 18:33:11 +0530
>
>> This came out of staring at code due to recent performance fix.
>>
>> * TX BD reclaim can call netif_wake_queue() once, outside the loop if
>>   one/more BDs were freed, NO need to do this each iteration.
>>
>> * TX need not look at next BD to stop the netif queue. It rather be done
>>   in the next tx call, when it actually fails as the queue seldom gets
>>   full but the check nevertheless needs to be done for each packet Tx.
>>   Profiled this under heavy traffic (big tar file cp, LMBench betworking
>>   tests) and saw not a single hit to that code.
>>
>> Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
> You should keep the check in the transmit queueing code as a BUG check,
> almost every driver has code of the form (using NIU as an example):
>
> 	if (niu_tx_avail(rp) <= (skb_shinfo(skb)->nr_frags + 1)) {
> 		netif_tx_stop_queue(txq);
> 		dev_err(np->device, "%s: BUG! Tx ring full when queue awake!\n", dev->name);
> 		rp->tx_errors++;
> 		return NETDEV_TX_BUSY;
> 	}
>
> and arc_emac should too.
>
> Otherwise queue management bugs are incredibly hard to diagnose.
>
> I'm not applying this patch.

The check is already there for current BD. What I removed was checking for next BD
too (please see below). IMHO this is useless since it will be done in next
iteration anyways. In my tests, the next check never got hit, so it was waste of
cycles.

static int arc_emac_tx(struct sk_buff *skb, struct net_device *ndev)
{
    if (unlikely((le32_to_cpu(*info) & OWN_MASK) == FOR_EMAC)) {
        netif_stop_queue(ndev);
        return NETDEV_TX_BUSY;
    }

...
        *txbd_curr = (*txbd_curr + 1) % TX_BD_NUM;

-       /* Get "info" of the next BD */
-       info = &priv->txbd[*txbd_curr].info;
-
-       /* Check if if Tx BD ring is full - next BD is still owned by EMAC */
-       if (unlikely((le32_to_cpu(*info) & OWN_MASK) == FOR_EMAC))
-               netif_stop_queue(ndev);

OTOH, I do see a slight stats update issue - if the queue is stopped (but pkt not
dropped) we are failing to increment tx_errors. But that would be a separate patch.

-Vineet

Re: [PATCH] ethernet/arc/arc_emac: optimize the Tx/Tx-reclaim paths a bit

[David Miller]

From: Vineet Gupta <Vineet.Gupta1@synopsys.com>
Date: Fri, 6 Sep 2013 04:24:39 +0000

> On 09/05/2013 11:54 PM, David Miller wrote:
>> You should keep the check in the transmit queueing code as a BUG check,
>> almost every driver has code of the form (using NIU as an example):
 ...
>> Otherwise queue management bugs are incredibly hard to diagnose.
>>
>> I'm not applying this patch.
> 
> The check is already there for current BD. What I removed was checking for next BD
> too (please see below). IMHO this is useless since it will be done in next
> iteration anyways. In my tests, the next check never got hit, so it was waste of
> cycles.
> 
> static int arc_emac_tx(struct sk_buff *skb, struct net_device *ndev)
> {
>     if (unlikely((le32_to_cpu(*info) & OWN_MASK) == FOR_EMAC)) {
>         netif_stop_queue(ndev);
>         return NETDEV_TX_BUSY;
>     }
> 
> ...
>         *txbd_curr = (*txbd_curr + 1) % TX_BD_NUM;
> 
> -       /* Get "info" of the next BD */
> -       info = &priv->txbd[*txbd_curr].info;
> -
> -       /* Check if if Tx BD ring is full - next BD is still owned by EMAC */
> -       if (unlikely((le32_to_cpu(*info) & OWN_MASK) == FOR_EMAC))
> -               netif_stop_queue(ndev);
> 
> OTOH, I do see a slight stats update issue - if the queue is stopped (but pkt not
> dropped) we are failing to increment tx_errors. But that would be a separate patch.

It is exactly the correct thing to do.  The driver should _NEVER_
return NETDEV_TX_BUSY under normal circumstances.  The queue should
always be stopped by the ->ndo_start_xmit() method when it fills the
queue.

Again, when ->ndo_start_xmit() is invoked, it should never see the
queue full.  When that happens it is a bug.

You are deleting exactly the correct part of this function, what it is
doing right now is precisely the correct way to manage netif queue
state.

The only valid change you can make here is to make the:

	if (unlikely((le32_to_cpu(*info) & OWN_MASK) == FOR_EMAC)) {
		netif_stop_queue(ndev);
		return NETDEV_TX_BUSY;
	}

print out an error message and increment tx_errors.