Re: [PATCH for 7.2-rc3 v1 0/2] virtio fixes

[Alex Bennée]

"Michael S. Tsirkin" <mst@redhat.com> writes:

> On Thu, Nov 24, 2022 at 09:21:15AM +0000, Alex Bennée wrote:
>> 
>> "Michael S. Tsirkin" <mst@redhat.com> writes:
>> 
>> > On Wed, Nov 23, 2022 at 04:03:49PM +0000, Alex Bennée wrote:
>> >> 
>> >> "Michael S. Tsirkin" <mst@redhat.com> writes:
>> >> 
>> >> > On Wed, Nov 23, 2022 at 03:21:32PM +0000, Alex Bennée wrote:
>> >> >> Hi,
>> >> >> 
>> >> >> This hopefully fixes the problems with VirtIO migration caused by the
>> >> >> previous refactoring of virtio_device_started(). That introduced a
>> >> >> different order of checking which didn't give the VM state primacy but
>> >> >> wasn't noticed as we don't properly exercise VirtIO device migration
>> >> >> and caused issues when dev->started wasn't checked in the core code.
>> >> >> The introduction of virtio_device_should_start() split the overloaded
>> >> >> function up but the broken order still remained. The series finally
>> >> >> fixes that by restoring the original semantics but with the cleaned up
>> >> >> functions.
>> >> >> 
>> >> >> I've added more documentation to the various structures involved as
>> >> >> well as the functions. There is still some inconsistencies in the
>> >> >> VirtIO code between different devices but I think that can be looked
>> >> >> at over the 8.0 cycle.
>> >> >
>> >> >
>> >> > Thanks a lot! Did you try this with gitlab CI? A patch similar to your
>> >> > 2/2 broke it previously ...
>> >> 
>> >> Looking into it now - so far hasn't broken locally but I guess there is
>> >> something different about the CI.
>> >
>> >
>> > yes - pls push to gitlab, create pipeline e.g. with QEMU_CI set to 2
>> >
>> > Or with QEMU_CI set to 1 and then run fedora container and then
>> > clang-system manually.
>> 
>> I'm having trouble re-creating the failures in CI locally on my boxen. I
>> have triggered a bug on s390 but that looks like a pre-existing problem
>> with VRING_SET_ENDIAN being triggered for the vhost-user-gpio tests. I
>> think that is a limitation of the test harness.
>> 
>> Will keep looking.
>
> Why not just trigger it on gitlab CI - it's very repeatable there?

I've got a fix for gpio and am running it through CI now:

  https://gitlab.com/stsquad/qemu/-/pipelines/704285944

My main concern is I had to do something no other vhost-user device does
and I'm not sure if thats down to misunderstanding or the other devices
just getting lucky.

-- 
Alex Bennée