Re: [RFC PATCH] Regression in linux 2.6.32 virtio_net seen with vhost-net

[Rusty Russell <rusty@rustcorp.com.au>]

On Thu, 17 Dec 2009 07:55:31 pm Michael S. Tsirkin wrote:
> > +static int virtnet_xmit_poll(struct napi_struct *xmit_napi, int budget)
> > +{
> > +	struct virtnet_info *vi =
> > +		container_of(xmit_napi, struct virtnet_info, xmit_napi);
> > +
> > +	/* Don't access vq/capacity at same time as start_xmit. */
> > +	__netif_tx_lock(netdev_get_tx_queue(vi->dev, 0), smp_processor_id());
> 
> So now that we are locking, we could build a variant of this
> without NAPI (maybe with trylock: we can't spin on xmit lock from
> from hard irq context, can we?)? Possibly, if we do, that would be
> a small enough change to be applicable in 2.6.32.

We'd need a separate lock and irq disable; it's not obvious to me that
trylock would never cause us to miss a required wakeup.

I'd rather get this right and backport, than introduce YA random change.
The NAPI change is no more complex that this.

>From what Sridhar has said, this doesn't even fix the issue.  I'm confused...

Rusty.