V7 test-in-container patch

[Florian Weimer]

On 08/16/2018 01:57 AM, DJ Delorie wrote:

>> Hurd doesn't have unshare, so you need to avoid building the container
>> framework there.
> 
> Is it OK to return UNSUPPORTED for those?  I don't think it makes sense
> to try to work around the whole containerized-test setup if containers
> are missing.

UNSUPPORTED would work there.  If the Hurd maintainers do not like it, 
we could probably filter out test-containers tests there.

>> The Red Hat Enterprise Linux 7 kernel doesn't seem to like this
>> combination of flags, even as root:
>>
>> unshare(CLONE_NEWNS|CLONE_NEWUSER|CLONE_NEWPID) = -1 EINVAL (Invalid
>> argument)
>>
>> Is CLONE_NEWPID really required?
> 
> Yes.  You can't mount /proc without it.  Don't ask me why ;-)

Can you perform a bind mount of the existing /proc instead?  Maybe you 
can drop the CLONE_NEWPID this way.

If the bind mount doesn't work, I suggest to ask internally what our 
kernel needs to get this working.

>> My concern about the over-use of FAIL_UNSUPPORTED and the UNSUPPORTED
>> test status in the container framework remains.  Sure, there are some
>> things that can fail due to missing host support, but e.g. a fork file
>> shouldn't lead to UNSUPPORTED, but FAIL.
> 
> I replaced them all with FAILs.  I was trying to keep "failure of the
> test" separate from "failure of the test harness" but I'm OK with it
> either way.

It's hard to draw a line.  Fork failures in particular tend to leak 
between tests because of kernel bugs (still unfixed upstream, 
unfortunately).

Thanks,
Florian