On Fri, Jan 30 2026, Eric Farman <[email protected]> wrote:

> On Mon, 2026-01-26 at 23:00 +0100, Halil Pasic wrote:
>> On Mon, 26 Jan 2026 19:40:09 +0300
>> Vladimir Sementsov-Ogievskiy <[email protected]> wrote:
>> 
>> > > > +/**
>> > > > + * virtio_queue_set_guest_notifier - set/unset queue or config guest
>> > > > + *     notifier
>> > > > + *
>> > > > + * @vdev: the VirtIO device
>> > > > + * @n: queue number, or VIRTIO_CONFIG_IRQ_IDX to set config notifer  
>> > > 
>> > > Semantically are mixing queue numbers and interrupt indexes here.  
>> > 
>> > That's preexisting logic in virtio_pci_set_guest_notifier_fd_handler().
>> > 
>> > But, you are right, that's doesn't answer the question, is it good to share
>> > the logic with mmio and ccw.
>> > 
>> > Is config notifier only pci specific, or similar feature may appear in 
>> > future
>> > for mmio and/or ccw?
>> 
>> AFAIR there is no  notifier for ccw. A possible explanation for
>> that is, that AFAIK that config space notifications (device -> driver)
>> are always classic I/O interrupts while queue indicators can be so called
>> adapter I/O interrupts and irqfd is only available for adapter I/O
>> interrupts. 
>> 
>> If you like you can have a look at 
>> "4.3.2.6 Setting Up Indicators" in
>> https://docs.oasis-open.org/virtio/virtio/v1.3/csd01/virtio-v1.3-csd01.html#x1-1940003
>> 
>> I see no reason why virtio_config_guest_notifier_read() based, i.e
>> !irqfd virtio config notifier should not exist or would not work. And I
>> guess it is needed for certain vhost things. But nevertheless, I can't
>> find the place where a virtio-ccw config notifier is set up.
>> 
>> Eric, what is your opinion on this? My current thinking is that
>> virtio-ccw not having a config notifier is basically a bug.
>
> (Sorry that I missed this; thanks for the ping, Halil)
>
> Yeah, we don't have have such a config notifier and I don't recall the 
> history behind it.

Adapter interrupts were not a thing for virtio-ccw in the beginning (we
only added the airq infrastructure later IIRC, so all interrupts were
"classic" subchannel-based interrupts.)

> But I don't know that it's a bug; spec points out config notifications for 
> ccw come as unsolicited
> interrupts versus config notifier, so not having one is fine by me as far as 
> this refactoring goes.
> I'll add it to a list of things to play around with, though.

I think I agree.

(Although my brain is already more than halfway into the weekend...)


Reply via email to