Re: [PATCH v2 5/6] kvm: KVM_EOIFD, an eventfd for EOIs

[Gleb Natapov <gleb@redhat.com>]

On Thu, Jun 28, 2012 at 04:11:40PM +0300, Michael S. Tsirkin wrote:
> On Wed, Jun 27, 2012 at 09:55:44PM -0600, Alex Williamson wrote:
> > On Wed, 2012-06-27 at 17:51 +0300, Gleb Natapov wrote:
> > > On Wed, Jun 27, 2012 at 08:29:04AM -0600, Alex Williamson wrote:
> > > > On Wed, 2012-06-27 at 16:58 +0300, Gleb Natapov wrote:
> > > > > On Tue, Jun 26, 2012 at 11:10:08PM -0600, Alex Williamson wrote:
> > > > > > This new ioctl enables an eventfd to be triggered when an EOI is
> > > > > > written for a specified irqchip pin.  By default this is a simple
> > > > > > notification, but we can also tie the eoifd to a level irqfd, which
> > > > > > enables the irqchip pin to be automatically de-asserted on EOI.
> > > > > > This mode is particularly useful for device-assignment applications
> > > > > > where the unmask and notify triggers a hardware unmask.  The default
> > > > > > mode is most applicable to simple notify with no side-effects for
> > > > > > userspace usage, such as Qemu.
> > > > > > 
> > > > > > Here we make use of the reference counting of the _irq_source
> > > > > > object allowing us to share it with an irqfd and cleanup regardless
> > > > > > of the release order.
> > > > > > 
> > > > > > Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> > > > > > ---
> > > > > > 
> > > > > >  Documentation/virtual/kvm/api.txt |   24 +++++
> > > > > >  arch/x86/kvm/x86.c                |    1 
> > > > > >  include/linux/kvm.h               |   14 +++
> > > > > >  include/linux/kvm_host.h          |   13 +++
> > > > > >  virt/kvm/eventfd.c                |  189 +++++++++++++++++++++++++++++++++++++
> > > > > >  virt/kvm/kvm_main.c               |   11 ++
> > > > > >  6 files changed, 250 insertions(+), 2 deletions(-)
> > > > > > 
> > > > > > diff --git a/Documentation/virtual/kvm/api.txt b/Documentation/virtual/kvm/api.txt
> > > > > > index b216709..87a2558 100644
> > > > > > --- a/Documentation/virtual/kvm/api.txt
> > > > > > +++ b/Documentation/virtual/kvm/api.txt
> > > > > > @@ -1987,6 +1987,30 @@ interrupts with those injected through KVM_IRQ_LINE.  IRQFDs created
> > > > > >  with KVM_IRQFD_FLAG_LEVEL must also set this flag when de-assiging.
> > > > > >  KVM_IRQFD_FLAG_LEVEL support is indicated by KVM_CAP_IRQFD_LEVEL.
> > > > > >  
> > > > > > +4.77 KVM_EOIFD
> > > > > > +
> > > > > > +Capability: KVM_CAP_EOIFD
> > > > > > +Architectures: x86
> > > > > > +Type: vm ioctl
> > > > > > +Parameters: struct kvm_eoifd (in)
> > > > > > +Returns: 0 on success, -1 on error
> > > > > > +
> > > > > > +KVM_EOIFD allows userspace to receive EOI notification through an
> > > > > > +eventfd for level triggered irqchip interrupts.  Behavior for edge
> > > > > > +triggered interrupts is undefined.  kvm_eoifd.fd specifies the eventfd
> > > > > Lets make it defined. EOI notification can be used by userspace to fix
> > > > > time drift due to lost interrupts. But than userspace needs to know
> > > > > which vcpu did EOI.
> > > > 
> > > > Hmm, do we need an additional flag and field in kvm_eoifd to filter by
> > > > vCPU then?
> > > > 
> > > This will be enough for a use case I am aware of. Don't know if this
> > > interface is generic enough for all possible use cases.
> > 
> > That's generally a hard prediction to make ;)  We currently don't pass a
> > kvm_vcpu anywhere close to the irq ack notifier.  The ioapic path could
> > be relatively trivial, but the pic path is a bit further disconnected.
> > If we had that plumbing, a KVM_CAP plus vcpu filter flag and specifying
> > the vcpu using some of the padding space seems like it's sufficient.
> > I'll drop mention of level-only from the description, but the plumbing
> > and vcpu filtering can be a follow-on.  Thanks,
> > 
> > Alex
> 
> If we don't implement what's needed for timedrift to be fixed,
> then IMO it's better to simply require an IRQFD for EOIFD for now,
> and limit this to level. Otherwise when we actually try to implement
> we might find issues.
> 
> Another reason to explicitly say EOI is not supported for edge is that
> EOI might not get invoked at all with PV EOI.
> 
Good point, but easily addressable by disabling PV EOI for a GSI that
has EOI notifier registered.

--
			Gleb.