Re: [PATCH net] virtio-net: add cond_resched() to the command waiting loop

[Jason Wang]

On Wed, Oct 12, 2022 at 11:19 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
>
> On Tue, Oct 11, 2022 at 1:11 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >
> > On Sun, Oct 09, 2022 at 01:58:53PM +0800, Jason Wang wrote:
> > >
> > > 在 2022/9/8 13:19, Michael S. Tsirkin 写道:
> > > > On Thu, Sep 08, 2022 at 10:21:45AM +0800, Jason Wang wrote:
> > > > > 在 2022/9/7 15:46, Michael S. Tsirkin 写道:
> > > > > > On Wed, Sep 07, 2022 at 09:07:20AM +0200, Paolo Abeni wrote:
> > > > > > > On Wed, 2022-09-07 at 10:09 +0800, Jason Wang wrote:
> > > > > > > > On Tue, Sep 6, 2022 at 6:56 PM Paolo Abeni <pabeni@xxxxxxxxxx> wrote:
> > > > > > > > > On Mon, 2022-09-05 at 15:49 +0800, Jason Wang wrote:
> > > > > > > > > > On Mon, Sep 5, 2022 at 3:15 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > > > > > > > > > On Mon, Sep 05, 2022 at 12:53:41PM +0800, Jason Wang wrote:
> > > > > > > > > > > > Adding cond_resched() to the command waiting loop for a better
> > > > > > > > > > > > co-operation with the scheduler. This allows to give CPU a breath to
> > > > > > > > > > > > run other task(workqueue) instead of busy looping when preemption is
> > > > > > > > > > > > not allowed.
> > > > > > > > > > > >
> > > > > > > > > > > > What's more important. This is a must for some vDPA parent to work
> > > > > > > > > > > > since control virtqueue is emulated via a workqueue for those parents.
> > > > > > > > > > > >
> > > > > > > > > > > > Fixes: bda324fd037a ("vdpasim: control virtqueue support")
> > > > > > > > > > > That's a weird commit to fix. so it fixes the simulator?
> > > > > > > > > > Yes, since the simulator is using a workqueue to handle control virtueue.
> > > > > > > > > Uhmm... touching a driver for a simulator's sake looks a little weird.
> > > > > > > > Simulator is not the only one that is using a workqueue (but should be
> > > > > > > > the first).
> > > > > > > >
> > > > > > > > I can see  that the mlx5 vDPA driver is using a workqueue as well (see
> > > > > > > > mlx5_vdpa_kick_vq()).
> > > > > > > >
> > > > > > > > And in the case of VDUSE, it needs to wait for the response from the
> > > > > > > > userspace, this means cond_resched() is probably a must for the case
> > > > > > > > like UP.
> > > > > > > >
> > > > > > > > > Additionally, if the bug is vdpasim, I think it's better to try to
> > > > > > > > > solve it there, if possible.
> > > > > > > > >
> > > > > > > > > Looking at vdpasim_net_work() and vdpasim_blk_work() it looks like
> > > > > > > > > neither needs a process context, so perhaps you could rework it to run
> > > > > > > > > the work_fn() directly from vdpasim_kick_vq(), at least for the control
> > > > > > > > > virtqueue?
> > > > > > > > It's possible (but require some rework on the simulator core). But
> > > > > > > > considering we have other similar use cases, it looks better to solve
> > > > > > > > it in the virtio-net driver.
> > > > > > > I see.
> > > > > > >
> > > > > > > > Additionally, this may have better behaviour when using for the buggy
> > > > > > > > hardware (e.g the control virtqueue takes too long to respond). We may
> > > > > > > > consider switching to use interrupt/sleep in the future (but not
> > > > > > > > suitable for -net).
> > > > > > > Agreed. Possibly a timeout could be useful, too.
> > > > > > >
> > > > > > > Cheers,
> > > > > > >
> > > > > > > Paolo
> > > > > > Hmm timeouts are kind of arbitrary.
> > > > > > regular drivers basically derive them from hardware
> > > > > > behaviour but with a generic driver like virtio it's harder.
> > > > > > I guess we could add timeout as a config field, have
> > > > > > device make a promise to the driver.
> > > > > >
> > > > > > Making the wait interruptible seems more reasonable.
> > > > >
> > > > > Yes, but I think we still need this patch for -net and -stable.
> > > > >
> > > > > Thanks
> > > > I was referring to Paolo's idea of having a timeout.
> > >
> > >
> > > Ok, I think we're fine with this patch. Any chance to merge this or do I
> > > need to resend?
> > >
> > > Thanks
> >
> > Last question: do we want cpu_relax here now? Or is cond_resched
> > sufficient?
>
> (Have answered in another thread)
>
> I think we need cpu_relax() since there could be no high priority task
> in the current cpu so we still need to relax.
>
> Thanks

Michael, does this answer make sense? If yes, would you like to ack the patch?

Thanks

>
> >
> > >
> > > >
> >