Message-ID: <13f42e6a-e20b-490c-a75b-a32b5e6c5553@redhat.com>
Date: Wed, 9 Apr 2025 13:02:15 +0200
From: David Hildenbrand <david@...hat.com>
To: Daniel Verkamp <dverkamp@...omium.org>
Cc: "Michael S. Tsirkin" <mst@...hat.com>, Halil Pasic <pasic@...ux.ibm.com>,
 linux-kernel@...r.kernel.org, linux-s390@...r.kernel.org,
 virtualization@...ts.linux.dev, kvm@...r.kernel.org,
 Chandra Merla <cmerla@...hat.com>, Stable@...r.kernel.org,
 Cornelia Huck <cohuck@...hat.com>, Thomas Huth <thuth@...hat.com>,
 Eric Farman <farman@...ux.ibm.com>, Heiko Carstens <hca@...ux.ibm.com>,
 Vasily Gorbik <gor@...ux.ibm.com>, Alexander Gordeev
 <agordeev@...ux.ibm.com>, Christian Borntraeger <borntraeger@...ux.ibm.com>,
 Sven Schnelle <svens@...ux.ibm.com>, Wei Wang <wei.w.wang@...el.com>
Subject: Re: [PATCH v1] s390/virtio_ccw: don't allocate/assign airqs for
 non-existing queues

On 07.04.25 23:09, Daniel Verkamp wrote:
> On Mon, Apr 7, 2025 at 11:47 AM David Hildenbrand <david@...hat.com> wrote:
>>
>>>>> Heh, but that one said:
>>>>>
>>>>> +\item[ VIRTIO_BALLOON_F_WS_REPORTING(6) ] The device has support for
>>>>> Working Set
>>>>>
>>>>> Which does not seem to reflect reality ...
>>>
>>> Please feel free to disregard these features and reuse their bits and
>>> queue indexes; as far as I know, they are not actually enabled
>>> anywhere currently and the corresponding guest patches were only
>>> applied to some (no-longer-used) ChromeOS kernel trees, so the
>>> compatibility impact should be minimal. I will also try to clean up
>>> the leftover bits on the crosvm side just to clear things up.
>>
>> Thanks for your reply, and thanks for clarifying+cleaning it up.
>>
> [...]
>>>> IIRC, in that commit they switched to the "spec" behavior.
>>>>
>>>> That's when they started hard-coding the queue indexes.
>>>>
>>>> CCing Daniel. All Linux versions should be incompatible with cross-vmm regarding free page reporting.
>>>> How is that handled?
>>>
>>> In practice, it only works because nobody calls crosvm with
>>> --balloon-page-reporting (it's off by default), so the balloon device
>>> does not advertise the VIRTIO_BALLOON_F_PAGE_REPORTING feature.
>>>
>>> (I just went searching now, and it does seem like there is actually
>>> one user in Android that does try to enable page reporting[1], which
>>> I'll have to look into...)
>>>
>>> In my opinion, it makes the most sense to keep the spec as it is and
>>> change QEMU and the kernel to match, but obviously that's not trivial
>>> to do in a way that doesn't break existing devices and drivers.
>>
>> If only it would be limited to QEMU and Linux ... :)
>>
>> Out of curiosity, assuming we'd make the spec match the current
>> QEMU/Linux implementation at least for the 3 involved features only,
>> would there be a way to adjust crossvm without any disruption?
>>
>> I still have the feeling that it will be rather hard to get that all
>> implementations match the spec ... For new features+queues it will be
>> easy to force the usage of fixed virtqueue numbers, but for
>> free-page-hinting and reporting, it's a mess :(
> 
> If the spec is changed, we can certainly update crosvm to match it; I
> think this only really affects a few devices (balloon and technically
> filesystem, but see below), only affects features that are generally
> not turned on, and in many cases, the guest kernel is updated
> simultaneously with the crosvm binary. I'm not opposed to changing the
> spec to match reality, although that feels like a bad move from a
> spec-integrity perspective.

Right. We didn't pay attention that the spec would reflect reality, and 
the reality was a bad decision :)

> 
> Regardless of the chosen path, I think the spec should be clarified -
> the meaning of "queue only exists if <feature> is set" leaves the
> reader with too many questions:

Right, that's what we've been discussing.

> - What does "if <feature> is set" mean? If it's advertised by the
> device? If it's acked by the driver? (To me, "set" definitely hints at
> the latter, but it should be explicit.)

Currently it's "feature is offered by the device".

> - What does it mean for a virtqueue to "exist"? Does that queue index
> disappear from the numbering if it does not exist, sliding all of the
> later queues down?

Currently it's like that, yes.

> If so, the spec should really not have the static
> queue numbers listed for the later queues, since they are only correct
> if all previous feature-dependent queues were also "set", whatever
> that means.

Yes, that's also what we've been discussing. And that started restarted 
the whole "can we fix device/drivers instead" :)

> 
> The way crosvm interpreted this was:
> - "if <feature> is set" means "if the device advertised <feature>
> *and* driver acknowledged <feature>"
> - "queue only exists" means "if <feature> was not acked, the
> corresponding virtqueue cannot be enabled by the driver" (attempting
> to set queue_enable = 1 has no effect).
> - Any later virtqueues are unaffected and still have the same queue indexes.

Yes, that matches my understanding.

> 
> The way QEMU interpeted this (I think, just skimming the code and
> working from memory here):
> - "if <feature> is set" means "if the device advertised <feature>" (it
> is checking host_features, not guest_features)

Right, "offered features".

> - "queue only exists" means "if <feature> was not offered by the
> device, all later virtqueues are shifted down one index"

Exactly.

> 
> ---
> 
> The spec for the filesystem device has a similar issue to the balloon device:
> - Queue 0 (hiprio) is always available regardless of features.
> - Queue 1 (notification queue) has a note that "The notification queue
> only exists if VIRTIO_FS_F_NOTIFICATION is set."
> - Queue 2..n are supposed to be the request queues per the numbering
> in the spec.
> 
> This is how it has been specified since virtio 1.2 when the fs device
> was originally added. However, the Linux driver (and probably all
> existing device implementations - at least virtiofsd and crosvm's fs
> device) don't support VIRTIO_FS_F_NOTIFICATION and use queue 1 as a
> request queue, which matches the QEMU/Linux interpretation but means
> the spec doesn't match reality again.

Yes, these are the two known cases we have to sort out.

Thanks for the information on crossvm, and that whatever we decide to 
do, adjusting crossvm should not be a big problem (at least for now :) ).

-- 
Cheers,

David / dhildenb

Powered by blists - more mailing lists