Re: [PATCH net-next] virtio_net: Prevent misidentified spurious interrupts from killing the irq

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Aug 06, 2024 at 11:18:14AM +0800, Jason Wang wrote:
> On Mon, Aug 5, 2024 at 2:29 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Mon, Aug 05, 2024 at 11:26:56AM +0800, Jason Wang wrote:
> > > On Fri, Aug 2, 2024 at 9:11 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > >
> > > > On Fri, Aug 02, 2024 at 11:41:57AM +0800, Jason Wang wrote:
> > > > > On Thu, Aug 1, 2024 at 9:56 PM Heng Qi <hengqi@linux.alibaba.com> wrote:
> > > > > >
> > > > > > Michael has effectively reduced the number of spurious interrupts in
> > > > > > commit a7766ef18b33 ("virtio_net: disable cb aggressively") by disabling
> > > > > > irq callbacks before cleaning old buffers.
> > > > > >
> > > > > > But it is still possible that the irq is killed by mistake:
> > > > > >
> > > > > >   When a delayed tx interrupt arrives, old buffers has been cleaned in
> > > > > >   other paths (start_xmit and virtnet_poll_cleantx), then the interrupt is
> > > > > >   mistakenly identified as a spurious interrupt in vring_interrupt.
> > > > > >
> > > > > >   We should refrain from labeling it as a spurious interrupt; otherwise,
> > > > > >   note_interrupt may inadvertently kill the legitimate irq.
> > > > >
> > > > > I think the evil came from where we do free_old_xmit() in
> > > > > start_xmit(). I know it is for performance, but we may need to make
> > > > > the code work correctly instead of adding endless hacks. Personally, I
> > > > > think the virtio-net TX path is over-complicated. We probably pay too
> > > > > much (e.g there's netif_tx_lock in TX NAPI path) to try to "optimize"
> > > > > the performance.
> > > > >
> > > > > How about just don't do free_old_xmit and do that solely in the TX NAPI?
> > > >
> > > > Not getting interrupts is always better than getting interrupts.
> > >
> > > Not sure. For example letting 1 cpu to do the transmission without the
> > > dealing of xmit skbs should give us better performance.
> >
> > Hmm. It's a subtle thing. I suspect until certain limit
> > (e.g. ping pong test) free_old_xmit will win anyway.
> 
> Not sure I understand here.

If you transmit 1 packet and then wait for another one anyway,
you are better off just handling the tx interrupt.


> >
> > > > This is not new code, there are no plans to erase it all and start
> > > > anew "to make it work correctly" - it's widely deployed,
> > > > you will cause performance regressions and they are hard
> > > > to debug.
> > >
> > > I actually meant the TX NAPI mode, we tried to hold the TX lock in the
> > > TX NAPI, which turns out to slow down both the transmission and the
> > > NAPI itself.
> > >
> > > Thanks
> >
> > We do need to synchronize anyway though, virtio expects drivers to do
> > their own serialization of vq operations.
> 
> Right, but currently add and get needs to be serialized which is a
> bottleneck. I don't see any issue to parallelize that.

Do you see this in traces?

> > You could try to instead move
> > skbs to some kind of array under the tx lock, then free them all up
> > later after unlocking tx.
> >
> > Can be helpful for batching as well?
> 
> It's worth a try and see.

Why not.

> >
> >
> > I also always wondered whether it is an issue that free_old_xmit
> > just polls vq until it is empty, without a limit.
> 
> Did you mean schedule a NAPI if free_old_xmit() exceeds the NAPI quota?

yes

> > napi is supposed to poll until a limit is reached.
> > I guess not many people have very deep vqs.
> 
> Current NAPI weight is 64, so I think we can meet it in stressful workload.
> 
> Thanks

yes, but it's just a random number.  since we hold the tx lock,
we get at most vq size bufs, so it's limited.

> >
> > --
> > MST
> >