Re: make -j check failing on master, interesting valgrind errors on qos-test vhost-user-blk-test/basic

[Alex Bennée]

Dario Faggioli <dfaggioli@suse.com> writes:

> [[PGP Signed Part:Undecided]]
> On Fri, 2022-05-27 at 10:18 +0200, Claudio Fontana wrote:
>> On 5/27/22 9:26 AM, Dario Faggioli wrote:
>> > > 
>> > Yes, this kind of matches what I've also seen and reported about in
>> > <5bcb5ceb44dd830770d66330e27de6a4345fcb69.camel@suse.com>. If
>> > enable/run just one of:
>> > - reconnect
>> > - flags_mismatch
>> > - connect_fail
>> > 
>> > I see no issues.
>> 
>> On the countrary, for me just running a single one of those can fail.
>> 
> Well, but you said (or at least so I understood) that running the test
> for the first time, works.
>
> Then, when you run it multiple times, things start to fail.
>
> That was, in fact, my point... I was making the parallelism between the
> fact running only one of those tests works for me and the fact that
> running the test for the first time works for you too.

Hmm so the qos-test is a bit weird as it:

 - forks itself to run a single subtest (g_test_trap_subprocess)
 - forks itself again for provide the dummy vhost-user daemon
 - as well as the fork/execve for qemu itself

while all the paths used for communication should be unique I wouldn't
be surprised if there is a racey interaction or two in the whole thing.
We even see a bit of this is the fact we don't cleanly tear stuff down
so QEMU sees the vhost-user socket disappear under it's feet.

>
> And between the fact that running two tests, one after the other, fails
> for me and the fact that running the same tests multiple times fails
> for you too.
>
> :-)
>
>> > However, Claudio, AFAIUI, you're seeing this with an older GCC and
>> > without LTO, right?
>> 
>> Yes, to provide a different angle I tried on veteran OpenSUSE Leap
>> 15.2, so gcc is based on 7.5.0.
>> 
>> I don't think LTO is being used in any way.
>> 
> Yep, agreed. Now I don't think it's related to LTO specifically either.
>
> Although, it's at least a bit of an Heisenbug. I mean, we're seeing it
> (with two different setups), but for others, things work fine, I guess?
>
> Regards


-- 
Alex Bennée