Re: [PATCH V3 6/9] virtio-ccw: implement synchronize_cbs()

[Michael S. Tsirkin]

On Thu, Apr 28, 2022 at 04:43:15AM +0200, Halil Pasic wrote:
> On Wed, 27 Apr 2022 11:27:03 +0200
> Cornelia Huck <cohuck@xxxxxxxxxx> wrote:
> 
> > On Tue, Apr 26 2022, "Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:
> > 
> > > On Tue, Apr 26, 2022 at 05:47:17PM +0200, Cornelia Huck wrote:  
> > >> On Mon, Apr 25 2022, "Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:
> > >>   
> > >> > On Mon, Apr 25, 2022 at 11:53:24PM -0400, Michael S. Tsirkin wrote:  
> > >> >> On Tue, Apr 26, 2022 at 11:42:45AM +0800, Jason Wang wrote:  
> > >> >> > 
> > >> >> > 在 2022/4/26 11:38, Michael S. Tsirkin 写道:  
> > >> >> > > On Mon, Apr 25, 2022 at 11:35:41PM -0400, Michael S. Tsirkin wrote:  
> > >> >> > > > On Tue, Apr 26, 2022 at 04:29:11AM +0200, Halil Pasic wrote:  
> > >> >> > > > > On Mon, 25 Apr 2022 09:59:55 -0400
> > >> >> > > > > "Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:
> > >> >> > > > >   
> > >> >> > > > > > On Mon, Apr 25, 2022 at 10:54:24AM +0200, Cornelia Huck wrote:  
> > >> >> > > > > > > On Mon, Apr 25 2022, "Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:  
> > >> >> > > > > > > > On Mon, Apr 25, 2022 at 10:44:15AM +0800, Jason Wang wrote:  
> > >> >> > > > > > > > > This patch tries to implement the synchronize_cbs() for ccw. For the
> > >> >> > > > > > > > > vring_interrupt() that is called via virtio_airq_handler(), the
> > >> >> > > > > > > > > synchronization is simply done via the airq_info's lock. For the
> > >> >> > > > > > > > > vring_interrupt() that is called via virtio_ccw_int_handler(), a per
> > >> >> > > > > > > > > device spinlock for irq is introduced ans used in the synchronization
> > >> >> > > > > > > > > method.
> > >> >> > > > > > > > > 
> > >> >> > > > > > > > > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > >> >> > > > > > > > > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > >> >> > > > > > > > > Cc: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
> > >> >> > > > > > > > > Cc: Marc Zyngier <maz@xxxxxxxxxx>
> > >> >> > > > > > > > > Cc: Halil Pasic <pasic@xxxxxxxxxxxxx>
> > >> >> > > > > > > > > Cc: Cornelia Huck <cohuck@xxxxxxxxxx>
> > >> >> > > > > > > > > Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>  
> > >> >> > > > > > > > 
> > >> >> > > > > > > > This is the only one that is giving me pause. Halil, Cornelia,
> > >> >> > > > > > > > should we be concerned about the performance impact here?
> > >> >> > > > > > > > Any chance it can be tested?  
> > >> >> > > > > > > We can have a bunch of devices using the same airq structure, and the
> > >> >> > > > > > > sync cb creates a choke point, same as registering/unregistering.  
> > >> >> > > > > > BTW can callbacks for multiple VQs run on multiple CPUs at the moment?  
> > >> >> > > > > I'm not sure I understand the question.
> > >> >> > > > > 
> > >> >> > > > > I do think we can have multiple CPUs that are executing some portion of
> > >> >> > > > > virtio_ccw_int_handler(). So I guess the answer is yes. Connie what do you think?
> > >> >> > > > > 
> > >> >> > > > > On the other hand we could also end up serializing synchronize_cbs()
> > >> >> > > > > calls for different devices if they happen to use the same airq_info. But
> > >> >> > > > > this probably was not your question  
> > >> >> > > > 
> > >> >> > > > I am less concerned about  synchronize_cbs being slow and more about
> > >> >> > > > the slowdown in interrupt processing itself.
> > >> >> > > >   
> > >> >> > > > > > this patch serializes them on a spinlock.
> > >> >> > > > > >   
> > >> >> > > > > Those could then pile up on the newly introduced spinlock.  
> > >> 
> > >> How bad would that be in practice? IIUC, we hit on the spinlock when
> > >> - doing synchronize_cbs (should be rare)
> > >> - processing queue interrupts for devices using per-device indicators
> > >>   (which is the non-preferred path, which I would basically only expect
> > >>   when running on an ancient or non-standard hypervisor)  
> > >
> > > this one is my concern. I am worried serializing everything on a single lock
> > > will drastically regress performance here.  
> > 
> > Yeah, that case could get much worse. OTOH, how likely is it that any
> > setup that runs a recent kernel will actually end up with devices using
> > per-device indicators? Anything running under a QEMU released in the
> > last couple of years is unlikely to not use airqs, I think. Halil, do
> > you think that the classic indicator setup would be more common on any
> > non-QEMU hypervisors?
> > 
> 
> I really don't know. My opinion is that, two stages indicators are kind
> of recommended for anybody who cares about notifications performance.
> 
> > IOW, how much effort is it worth spending on optimizing this case? We
> > certainly should explore any simple solutions, but I don't think we need
> > to twist ourselves into pretzels to solve it.
> > 
> 
> Frankly, I would be fine with an rwlock based solution as proposed by
> Jason. My rationale is: we recommend two stage indicators, and the two
> stage indicators are already encumbered by an rwlock on the interrupt
> path.

Which lock is it? Can we take that lock to synchronize?

> Yes, the coalescence of adapter interrupts is architecturally
> different, and so it is with GISA (without GISA, I'm not even sure), so
> this rwlock end up being worse than the one for 2 stage. But my feeling
> is, that it should be fine. On the other hand, I don't feel comfortable
> with plain spinlock, and I am curious about a more advanced solution.
> But my guess is that rwlock + some testing for the legacy indicator case
> just to double check if there is a heavy regression despite of our
> expectations to see none should do the trick.
> 
> Regards,
> Halil

Makes sense.

> > >
> > >  
> > >> - configuration change interrupts (should be rare)
> > >> - during setup, reset, etc. (should not be a concern)  
> >