Re: [RFC 0/3] aio-posix: call ->poll_end() when removing AioHandler

[Paolo Bonzini <pbonzini@redhat.com>]

On 1/3/24 12:40, Fiona Ebner wrote:
> I'm happy to report that I cannot reproduce the CPU-usage-spike issue
> with the patch, but I did run into an assertion failure when trying to
> verify that it fixes my original stuck-guest-IO issue. See below for the
> backtrace [0]. Hanna wrote in https://issues.redhat.com/browse/RHEL-3934
> 
>> I think it’s sufficient to simply call virtio_queue_notify_vq(vq)
>> after the virtio_queue_aio_attach_host_notifier(vq, ctx) call, because
>> both virtio-scsi’s and virtio-blk’s .handle_output() implementations
>> acquire the device’s context, so this should be directly callable from
>> any context.
> 
> I guess this is not true anymore now that the AioContext locking was
> removed?

Good point and, in fact, even before it was much safer to use 
virtio_queue_notify() instead.  Not only does it use the event notifier 
handler, but it also calls it in the right thread/AioContext just by 
doing event_notifier_set().

The call to virtio_queue_set_notification(..., 1) is safe; not sure 
about the call to virtio_queue_set_notification(..., 0) though.  I'd 
rather have that executed synchronously in the AioContext using 
aio_wait_bh_oneshot().  This is consistent with the pattern used by 
virtio_scsi_dataplane_stop() and virtio_blk_data_plane_stop().

Paolo