Re: [KVM PATCH v3 2/2] kvm: add support for irqfd via eventfd-notification interface

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

On Sun, May 03, 2009 at 07:59:40PM +0300, Avi Kivity wrote:
> Michael S. Tsirkin wrote:
>> On Mon, Apr 27, 2009 at 02:33:34PM -0400, Gregory Haskins wrote:
>>   
>>> This allows an eventfd to be registered as an irq source with a guest.  Any
>>> signaling operation on the eventfd (via userspace or kernel) will inject
>>> the registered GSI at the next available window.
>>>
>>> Signed-off-by: Gregory Haskins <ghaskins@novell.com>
>>>     
>>
>> If we ever want to use this with e.g. MSI-X emulation in guest, and want
>> to be stricly compliant to MSI-X, we'll need a way for guest to mask
>> interrupts, and for host to report that a masked interrupt is pending.
>> Ideally, all this will be doable with a couple of mmapped pages to avoid
>> vmexits/system calls.
>>
>>   
>
> We could do this in two ways:
>
> - move msix entry emulation into the kernel

It's not too bad IMO: MSIX is just a table with a list
of vectors, you check the mask bit on each interrupt,
if masked mark vector pending and poll until unmasked.

> - require the device to support replacing its irqfd, and juggle it like so:
>   - guest disables msi
>       - replace device model fd with eventfd belonging to us
>       - when the device fires its eventfd, set the irq pending bit
>   - guest enables msi
>        - if the pending bit is set, fire the interrupt?
>        - replace device model fd with the real irqfd

Looks like a lot of code. No?

> I'm leaning towards the latter, though it's not an easy call.

Actually there's a third option: add KVM_MASK_IRQ, KVM_UNMASK_IRQ ioctls
which will block/unblock guest from getting interrupt on this irq,
whatever the source.  Interrupts are queued in kernel while masked. A
third ioctl KVM_PENDING_IRQS will return the status for a set if IRQs.
qemu would call these ioctls when guest edits the MSIX vector control or
reads the pending bit array.

>>> +static void
>>> +irqfd_inject(struct work_struct *work)
>>> +{
>>> +	struct _irqfd *irqfd = container_of(work, struct _irqfd, work);
>>> +	struct kvm *kvm = irqfd->kvm;
>>> +
>>> +	mutex_lock(&kvm->lock);
>>> +	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
>>> +	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
>>> +	mutex_unlock(&kvm->lock);
>>>     
>>
>> This will do weird stuff (deliver the irq twice) if the irq is
>> MSI/MSI-X. I know this was discussed already and is a temporary
>> shortcut, but maybe add a comment that we really want kvm_toggle_irq,
>> so that we won't forget?
>>   
>
> If so, that's a bug.  MSI should ignore kvm_set_irq(..., 0).

-- 
MST