Re: acceptance-system-fedora failures

[Pavel Dovgalyuk]

On 07.10.2020 12:57, Philippe Mathieu-Daudé wrote:
> On 10/7/20 10:51 AM, Pavel Dovgalyuk wrote:
> > On 07.10.2020 11:23, Thomas Huth wrote:
> > > On 07/10/2020 09.13, Philippe Mathieu-Daudé wrote:
> > > > On 10/7/20 7:20 AM, Philippe Mathieu-Daudé wrote:
> > > > > On 10/7/20 1:07 AM, John Snow wrote:
> > > > > > I'm seeing this gitlab test fail quite often in my Python work; I
> > > > > > don't
> > > > > > *think* this has anything to do with my patches, but maybe I need
> > > > > > to try
> > > > > > and bisect this more aggressively.
> > > [...]
> > > > > w.r.t. the error in your build, I told Thomas about the
> > > > > test_ppc_mac99/day15/invaders.elf timeouting but he said this is
> > > > > not his area. Richard has been looking yesterday to see if it is
> > > > > a TCG regression, and said the test either finished/crashed raising
> > > > > SIGCHLD, but Avocado parent is still waiting for a timeout, so the
> > > > > children become zombie and the test hang.
> > > >
> > > > Expected output:
> > > >
> > > > Quiescing Open Firmware ...
> > > > Booting Linux via __start() @ 0x01000000 ...
> > > >
> > > > But QEMU exits in replay_char_write_event_load():
> > > >
> > > > Quiescing Open Firmware ...
> > > > qemu-system-ppc: Missing character write event in the replay log
> > > > $ echo $?
> > > > 1
> > > >
> > > > Latest events are CHECKPOINT CHECKPOINT INTERRUPT INTERRUPT INTERRUPT.
> > > >
> > > > Replay file is ~22MiB. End of record using "system_powerdown + quit"
> > > > in HMP.
> > > >
> > > > I guess we have 2 bugs:
> > > > - replay log
> > > > - avocado doesn't catch children exit(1)
> > > >
> > > > Quick reproducer:
> > > >
> > > > $ make qemu-system-ppc check-venv
> > > > $ tests/venv/bin/python -m \
> > > >     avocado --show=app,console,replay \
> > > >     run --job-timeout 300 -t machine:mac99 \
> > > >     tests/acceptance/replay_kernel.py
> > >
> > > Thanks, that was helpful. ... and the winner is:
> > >
> > >       commit   55adb3c45620c31f29978f209e2a44a08d34e2da
> > >       Author:  John Snow <jsnow@redhat.com>
> > >       Date:    Fri Jul 24 01:23:00 2020 -0400
> > >       Subject: ide: cancel pending callbacks on SRST
> > >
> > > ... starting with this commit, the tests starts failing. John, any
> > > idea what
> > > might be causing this?
> >
> > This patch includes the following lines:
> >
> > +        aio_bh_schedule_oneshot(qemu_get_aio_context(),
> > +                                ide_bus_perform_srst, bus);
> >
> > replay_bh_schedule_oneshot_event should be used instead of this
> > function, because it synchronizes non-deterministic BHs.
>
> Why do we have 2 different functions? BH are already complex
> enough, and we need to also think about the replay API...

There is note about it in docs/devel/replay.txt
It's hard to protect the guest state from incorrect operations.
There was the similar problem with icount - everyone who modify 
translate modules, needs to take it info account.
But now we have record/replay tests that assert that patches do not 
break icount/rr.

> What about the other cases such vhost-user (blk/net), virtio-blk?

I do not know much about these modules.
The main idea is the following: if the code works with the guest state, 
it should deal with icount and rr functions.

Pavel Dovgalyuk