Re: [PATCH v9 1/2] kvm: Use a reserved IRQ source ID for irqfd

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

On Wed, Sep 05, 2012 at 06:22:57PM +0300, Avi Kivity wrote:
> On 09/05/2012 06:13 PM, Michael S. Tsirkin wrote:
> > On Wed, Sep 05, 2012 at 05:59:46PM +0300, Avi Kivity wrote:
> >> On 09/05/2012 05:51 PM, Michael S. Tsirkin wrote:
> >> > On Wed, Sep 05, 2012 at 05:35:43PM +0300, Avi Kivity wrote:
> >> >> On 08/22/2012 03:41 AM, Michael S. Tsirkin wrote:
> >> >> >> 
> >> >> >> I assumed you were pointing out the level vs edge interaction.  If we
> >> >> >> call that a userspace bug, I can just drop this.  Thanks,
> >> >> >> 
> >> >> >> Alex
> >> >> > 
> >> >> > level is userspace bug I think :)
> >> >> 
> >> >> I don't see how it's a bug.  Suppose we have a vfio device that shares a
> >> >> gsi with an emulated device.  The emulated device naturally uses
> >> >> KVM_IRQ_LINE (it has no need to re-sample on ADN), while vfio naturally
> >> >> has to use irqfd.
> >> > 
> >> > Absolutely. But vfio needs to use irqfd with the new flag.
> >> > Using existing irqfd for level is a bug.
> >> 
> >> I see we're not reusing this irq source id for level irqfd.  But I think
> >> we should, there's no need for per-gsi irq source id.
> > 
> > I agree. All resample irqfds are deasserted at the same time,
> > tracking them separately gets us nothing.
> 
> That's not the reason.  Separate irq source ids only have meanings
> within a gsi.  We could have two lines (gsi 3 isid 4) and (gsi 4 isid 4)
> that can be toggled independently with no effect on the other gsi.
> Within a gsi we do need a separate irq source id usually, but as 2/2
> recognizes, AODNs are a special case since we clear all inputs anyway.
> The end result is that all AODNs can share a single isid.
> 
> > 
> >> Plus I'd like to
> >> fix the theoretical bug even if it doesn't bite in practice.
> >> 
> > 
> > I'm not sure what the bug is, for edge, and how a separate ID fixes it.
> > Could you clarify?
> 
> gsi 3 is configured as edge in the ioapic.  It has (unusually) two
> inputs: one driven by userspace, the other by irqfd.
> 
> cpu 0                    cpu 1
> ------------------------ -------------------------
> irqfd: set to 1
> ioapic: recognize edge
> inject irq
> EOI
>                          KVM_IRQ_LINE: set to 1
>                          ioapic: ignore
>                          KVM_IRQ_LINE: set to 0
> irqfd: set to 0
> 
> We had two edges with an EOI between them, but injected just on interrupt.

I see. Makes sense, ACK this patch.


> -- 
> error compiling committee.c: too many arguments to function