Query about TX BD Reclaim in Napi poll path (was Re: [PATCH v3] ethernet/arc/arc_emac - Add new driver)

[Vineet Gupta <Vineet.Gupta1@synopsys.com>]

Hi Francois,

Resurrecting an old thread.

On 06/14/2013 03:49 AM, Francois Romieu wrote:
>> +static irqreturn_t arc_emac_intr(int irq, void *dev_instance)
>> > +{
>> > +	struct net_device *ndev = dev_instance;
>> > +	struct arc_emac_priv *priv = netdev_priv(ndev);
>> > +	struct net_device_stats *stats = &priv->stats;
>> > +	unsigned int status;
>> > +
>> > +	status = arc_reg_get(priv, R_STATUS);
>> > +	status &= ~MDIO_MASK;
>> > +
>> > +	/* Reset all flags except "MDIO complete"*/
>> > +	arc_reg_set(priv, R_STATUS, status);
>> > +
>> > +	if (status & RXINT_MASK) {
>> > +		if (likely(napi_schedule_prep(&priv->napi))) {
>> > +			arc_reg_clr(priv, R_ENABLE, RXINT_MASK);
>> > +			__napi_schedule(&priv->napi);
>> > +		}
>> > +	}
>> > +
>> > +	if (status & TXINT_MASK) {
> You may consider moving everything into the napi poll handler.

I has to revisit this now-mainlined driver recently for fixing a bug. Per your
suggestion above, the TX BD reclaim was moved from interrupt context to NAPI
context. I was wondering if that is the right thing to do (I'm not a networking
expert but have worked on this driver heavily before it was mainlined by Alexey).

In case of large burst transfers by networking stack (say a large file copy over
NFS) will it not delay the TX BD reclaim possibly dropping more packets. Ofcourse
doing this requires enabling Tx interrupts which adds to overall cost from a
system perspective, but assuming the controller can coalesce the Tx interrupts,
will it not be better.

I did a quick hack to move the TX reclaim in intr path and it seems to be doing
slightly better than the current code - so the advantages are not sky high, but I
want to understand the implications nevertheless.

TIA,
-Vineet