On Wed, 29 Jun 2016 14:23:42 +0200
Marc-André Lureau <marcandre.lur...@gmail.com> wrote:
> Hi
>
> On Wed, Jun 29, 2016 at 2:17 PM, Cornelia Huck <cornelia.h...@de.ibm.com>
> wrote:
> > When setting up host notifiers, virtio_bus_set_host_notifier()
> > simply switches the handler. This will only work, however, if
> > the ioeventfd has already been setup; this is true for dataplane,
> > but not for vhost.
> >
> > Fix this by starting the ioeventfd if that has not happened
> > before.
> >
> > While we're at it, also fixup the unsetting path of
> > set_host_notifier_internal().
> >
> > Fixes: 6798e245a3 ("virtio-bus: common ioeventfd infrastructure")
> > Reported-by: Jason Wang <jasow...@redhat.com>
> > Reported-by: Marc-André Lureau <marcandre.lur...@gmail.com>
> > Signed-off-by: Cornelia Huck <cornelia.h...@de.ibm.com>
> > ---
> >
> > This fixes the vhost regression for me, while dataplane continues
> > to work.
> >
>
> That doesn't work here,
> QTEST_QEMU_BINARY=x86_64-softmmu/qemu-system-x86_64 tests/vhost-user-test
>
> hangs in /x86_64/vhost-user/migrate
I can reproduce it, but I have zero ideas on how to proceed.
I can see that one of the qemus sits on event_notifier_test_and_clear
when vhost tries to shut down. (I am thoroughly confused by all of
that qtest setup, so I have no idea which qemu instance this is...)
Looking at the code path, we really should switch the handler around,
but virtio_queue_set_host_notifier_fd_handler always unsets the handler
unless both assign and set_handler are true. Is that really what we
want?
I fear I have stared at this for so long that I have now lost myself
between all these flags, so I hope one of the folks on cc: has a good
idea...
