Re: [PATCH 0/4 v9] MSI-X MMIO support for KVM

[Sasha Levin <levinsasha928@gmail.com>]

On Tue, 2011-09-06 at 11:05 +0200, Jan Kiszka wrote:
> On 2011-09-06 11:00, Sasha Levin wrote:
> > On Tue, 2011-09-06 at 10:49 +0200, Jan Kiszka wrote:
> >> On 2011-09-06 10:46, Sasha Levin wrote:
> >>> On Tue, 2011-09-06 at 10:36 +0200, Jan Kiszka wrote:
> >>>> On 2011-09-06 10:12, Michael S. Tsirkin wrote:
> >>>>> On Tue, Sep 06, 2011 at 10:52:41AM +0300, Avi Kivity wrote:
> >>>>>>>
> >>>>>>> BTW, the same is also true for that optional per-vector masking of
> >>>>>>> legacy MSI. Are there devices in the field that actually support this? I
> >>>>>>> haven't found one so far and tend to consider this feature not worth
> >>>>>>> implementing.
> >>>>>>
> >>>>>> Don't know.  I don't like implementing features on the basis of bug
> >>>>>> reports, though.  On the other hand we can't really test it without
> >>>>>> a real device.
> >>>>>
> >>>>> Linux will use this capability if present. So
> >>>>> we could add support for an emulated device (e.g. e1000),
> >>>>> then test with nested virt once iommu emulation lands :)
> >>>>
> >>>> Yeah, would be kind of cool. Still, I would feel better having it tested
> >>>> against a real silicon as well. Also to prove that there is a real need.
> >>>>
> >>>> So, in case someone stumbles for such a device (bit 8 set in MSI control
> >>>> word), please let us know!
> >>>
> >>> Doesn't any device that supports MSI-X supports per-vector masking?
> >>>
> >>> >From the spec:
> >>>
> >>> "MSI and MSI-X each support per-vector masking. Per-vector masking is an
> >>> optional extension to MSI, and a standard feature with MSI-X"
> >>
> >> Right, but the cap flag still has to be set.
> >>
> > 
> > Are you sure? Take a look at the table in section 6.8.2.3 in the spec
> > (pci v3). Unlike the message control for MSI, this table doesn't mention
> > anything about bit 8 or the per-vector masking capability for MSI-X, it
> > just assumes it's there.
> 
> [Err, I should stop doing n things in parallel.]
> 
> Of course, MSI-X implies per-vector masking, but in a totally different
> way with different data structures etc. That's not interesting for the
> case in question: per-vector masking for legacy MSI.
> 

Ah, Okay.

> Back to square #1: We need a device with MSI support and cap bit 8 set
> in its _MSI_ control word.

Alright, so I've looked at some of my servers, and one of them has both
a bunch of MSI-X devices, and some MSI devices which show this:

	Capabilities: [60] MSI: Enable+ Count=1/2 Maskable+ 64bit-
		Address: 00000020  Data: 0000
		Masking: 00000000  Pending: 00000000

Which would suggest that the support per-vector masking, right?

-- 

Sasha.