Re: [PATCH net-next] net: systemport: Do not block interrupts in TX reclaim

[Florian Fainelli <florian.fainelli@broadcom.com>]

On 1/24/2020 5:35 PM, Eric Dumazet wrote:
> 
> 
> On 1/24/20 3:59 PM, Florian Fainelli wrote:
>> There is no need to disable interrupts with a spin_lock_irqsave() in
>> bcm_sysport_tx_poll() since we are in softIRQ context already. Leave
>> interrupts enabled, thus giving a chance for the RX interrupts to be
>> processed.
>>
>> This now makes bcm_sysport_tx_reclaim() equivalent to
>> bcm_sysport_tx_clean(), thus remove the former, and make
>> bcm_sysport_tx_reclaim_all() to use the latter.
>>
>> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
>> ---
>>  drivers/net/ethernet/broadcom/bcmsysport.c | 30 ++++++----------------
>>  1 file changed, 8 insertions(+), 22 deletions(-)
>>
>> diff --git a/drivers/net/ethernet/broadcom/bcmsysport.c b/drivers/net/ethernet/broadcom/bcmsysport.c
>> index f07ac0e0af59..dfff0657ce8f 100644
>> --- a/drivers/net/ethernet/broadcom/bcmsysport.c
>> +++ b/drivers/net/ethernet/broadcom/bcmsysport.c
>> @@ -925,26 +925,6 @@ static unsigned int __bcm_sysport_tx_reclaim(struct bcm_sysport_priv *priv,
>>  	return pkts_compl;
>>  }
>>  
>> -/* Locked version of the per-ring TX reclaim routine */
>> -static unsigned int bcm_sysport_tx_reclaim(struct bcm_sysport_priv *priv,
>> -					   struct bcm_sysport_tx_ring *ring)
>> -{
>> -	struct netdev_queue *txq;
>> -	unsigned int released;
>> -	unsigned long flags;
>> -
>> -	txq = netdev_get_tx_queue(priv->netdev, ring->index);
>> -
>> -	spin_lock_irqsave(&ring->lock, flags);
>> -	released = __bcm_sysport_tx_reclaim(priv, ring);
>> -	if (released)
>> -		netif_tx_wake_queue(txq);
>> -
>> -	spin_unlock_irqrestore(&ring->lock, flags);
>> -
>> -	return released;
>> -}
>> -
>>  /* Locked version of the per-ring TX reclaim, but does not wake the queue */
>>  static void bcm_sysport_tx_clean(struct bcm_sysport_priv *priv,
>>  				 struct bcm_sysport_tx_ring *ring)
>> @@ -960,9 +940,15 @@ static int bcm_sysport_tx_poll(struct napi_struct *napi, int budget)
>>  {
>>  	struct bcm_sysport_tx_ring *ring =
>>  		container_of(napi, struct bcm_sysport_tx_ring, napi);
>> +	struct bcm_sysport_priv *priv = ring->priv;
>>  	unsigned int work_done = 0;
>>  
>> -	work_done = bcm_sysport_tx_reclaim(ring->priv, ring);
>> +	spin_lock(&ring->lock);
>> +	work_done = __bcm_sysport_tx_reclaim(priv, ring);
>> +	if (work_done)
>> +		netif_tx_wake_queue(netdev_get_tx_queue(priv->netdev,
>> +							ring->index));
>> +	spin_unlock(&ring->lock);
>>  
>>  	if (work_done == 0) {
>>  		napi_complete(napi);
>> @@ -984,7 +970,7 @@ static void bcm_sysport_tx_reclaim_all(struct bcm_sysport_priv *priv)
>>  	unsigned int q;
>>  
>>  	for (q = 0; q < priv->netdev->num_tx_queues; q++)
>> -		bcm_sysport_tx_reclaim(priv, &priv->tx_rings[q]);
>> +		bcm_sysport_tx_clean(priv, &priv->tx_rings[q]);
>>  }
>>  
>>  static int bcm_sysport_poll(struct napi_struct *napi, int budget)
>>
> 
> I am a bit confused by this patch, the changelog mixes hard and soft irqs.
> 
> This driver seems to call bcm_sysport_tx_reclaim_all() from hard irq handler 
> (INTRL2_0_TX_RING_FULL condition)
> 
> So it looks you need to acquire ring->lock with some _irqsave() variant when
> bcm_sysport_tx_poll() is running (from BH context)

You are right, I completely missed that path and the very reason why
this is spin_lock_irqsave() in the first place... time to get some sleep.

Thanks!
-- 
Florian