Re: [PATCH RFC] kvm: enable irq injection from interrupt context

[Michael S. Tsirkin]

On Thu, Sep 16, 2010 at 01:17:52PM +0200, Gleb Natapov wrote:
> On Thu, Sep 16, 2010 at 12:53:52PM +0200, Michael S. Tsirkin wrote:
> > On Thu, Sep 16, 2010 at 12:54:03PM +0200, Gleb Natapov wrote:
> > > On Thu, Sep 16, 2010 at 12:44:55PM +0200, Michael S. Tsirkin wrote:
> > > > On Thu, Sep 16, 2010 at 12:20:47PM +0200, Gleb Natapov wrote:
> > > > > On Thu, Sep 16, 2010 at 12:13:39PM +0200, Michael S. Tsirkin wrote:
> > > > > > On Thu, Sep 16, 2010 at 12:13:32PM +0200, Gleb Natapov wrote:
> > > > > > > On Thu, Sep 16, 2010 at 11:53:10AM +0200, Michael S. Tsirkin wrote:
> > > > > > > > On Thu, Sep 16, 2010 at 11:46:03AM +0200, Avi Kivity wrote:
> > > > > > > > >  On 09/16/2010 11:25 AM, Gleb Natapov wrote:
> > > > > > > > > >>
> > > > > > > > > >>  MSI only appeared in rhel6, older guests still use level interrupts.
> > > > > > > > > >So they are already slow for other reasons.
> > > > > > > > > >
> > > > > > > > > 
> > > > > > > > > Exactly, for example they need to exit to userspace to ack the
> > > > > > > > > interrupt.  That's far slower than the workqueue.
> > > > > > > > 
> > > > > > > > Well, this is not exactly comparable: you might get
> > > > > > > > same irq asserted multiple times and only deasserted once.
> > > > > > > > 
> > > > > > > Are we talking about level interrupts? Why would you assert level
> > > > > > > triggered interrupt multiple times before deasserting it?
> > > > > > 
> > > > > > User of irqfd has no way to know what current interrupt level is.
> > > > > > So it has to keep asserting.
> > > > > > 
> > > > > Why can't it keep track of current level?
> > > > 
> > > > This breaks the model: eventfd user is unaware of PCI, levels and such:
> > > > it just signals the event.  Remember that asserts are done from e.g. vhost-net,
> > > > deasserts need to be handled by qemu.
> > > > 
> > > eventfd user implements HW and it knows exactly what type of interrupt
> > > this HW generates.
> > 
> > We haver two users: qemu does deasserts, vhost-net does asserts.
> Well this is broken. You want KVM to track level for you and this is
> wrong. KVM does this anyway because it can't relay on devise model
> to behave correctly [0], but in your case it is designed to behave
> incorrectly.
> 
> Interrupt type is a device property. PCI devices just happen to be level
> triggered according to PCI spec. What if you want to use vhost-net to
> implement network device which has active-low interrupt line? [1]

The polarity would have to be reversed in gsi (irq line can be shared,
all devices must be active high or low consistently).

> If you want to split parts that asserts irq and de-asserts it then we
> should have irqfd that tracks line status and knows interrupt line
> polarity.

Yes, it can know about polarity even though I think it's cleaner to do this
per gsi. But it can not track line status as line is shared with
other devices.

> > Another application is out of process virtio (sandboxing!).
> It will still assert and de-assert irq at the same code, so it will be
> able to track irq line status.
> 
> > Again, pci stuff needs to stay in qemu.
> > 
> 
> Nothing to do with PCI whatsoever.
> 
> [0] most qemu devices behave incorrectly and trigger level irq more then
>     needed.

Which devices?
pci core tracks line status and will never assert the same
line multiple times.

> [1] this is how correct PCI device should behave but we override
>     polarity in ACPI, but now incorrect behaviour is deeply designed
>     into vhost-net.

Not really, vhost net signals an eventfd. What happens then is
up to kvm.

> --
> 			Gleb.



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/