Re: [Qemu-devel] [PATCH 0/6] Misc PCI cleanups

[Michael S. Tsirkin]

On Wed, Oct 10, 2012 at 01:31:52PM -0600, Alex Williamson wrote:
> On Tue, 2012-10-09 at 09:09 +0200, Jan Kiszka wrote:
> > On 2012-10-08 23:11, Alex Williamson wrote:
> > > On Mon, 2012-10-08 at 23:40 +0200, Michael S. Tsirkin wrote:
> > >> On Mon, Oct 08, 2012 at 01:27:33PM -0600, Alex Williamson wrote:
> > >>> On Mon, 2012-10-08 at 22:15 +0200, Michael S. Tsirkin wrote:
> > >>>> On Mon, Oct 08, 2012 at 09:58:32AM -0600, Alex Williamson wrote:
> > >>>>> Michael, Jan,
> > >>>>>
> > >>>>> Any comments on these?  I'd like to make the PCI changes before I 
> > >>>>> update
> > >>>>> vfio-pci to make use of the new resampling irqfd in kvm.  We don't 
> > >>>>> have
> > >>>>> anyone officially listed as maintainer of pci-assign since it's been
> > >>>>> moved to qemu.  I could include the pci-assign patches in my tree if 
> > >>>>> you
> > >>>>> prefer.  Thanks,
> > >>>>>
> > >>>>> Alex
> > >>>>
> > >>>> Patches themselves look fine, but I'd like to
> > >>>> better understand why do we want the INTx fallback.
> > >>>> Isn't it easier to add intx routing support?
> > >>>
> > >>> vfio-pci can work with or without intx routing support.  Its presence is
> > >>> just one requirement to enable kvm accelerated intx support.  Regardless
> > >>> of whether it's easy or hard to implement intx routing in a given
> > >>> chipset, I currently can't probe for it and make useful decisions about
> > >>> whether or not to enable kvm support without potentially hitting an
> > >>> assert.  It's arguable how important intx acceleration is for specific
> > >>> applications, so while I'd like all chipsets to implement it, I don't
> > >>> know that it should be a gating factor to chipset integration.  Thanks,
> > >>>
> > >>> Alex
> > >>
> > >> Yes but there's nothing kvm specific in the routing API,
> > >> and IIRC it actually works fine without kvm.
> > > 
> > > Correct, but intx routing isn't very useful without kvm.
> > 
> > Right now: yes. Long-term: no. The concept in general is also required
> > for decoupling I/O paths lock-wise from our main thread. We need to
> > explore the IRQ path and cache it in order to avoid taking lots of locks
> > on each delivery, possibly even the BQL. But we will likely need
> > something smarter at that point, i.e. something PCI-independent.
> 
> That sounds great long term, but in the interim I think this trivial
> extension to the API is more than justified.  I hope that it can go in
> soon so we can get vfio-pci kvm intx acceleration in before freeze
> deadlines get much closer.  Thanks,
> 
> Alex

Simply reorder the patches:
1. add vfio acceleration with no fallback
2. add way for intx routing to fail
3. add vfio fallback if intx routing fails

Then we can apply 1 and argue about the need for 2/3
afterwards.

-- 
MST