Re: [PATCH 1/2] KVM: selftests: Make rseq compatible with glibc-2.35

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Aug 9, 2022, at 8:37 PM, Gavin Shan gshan@redhat.com wrote:

> Hi Mathieu and Sean,
> 
> On 8/10/22 7:38 AM, Sean Christopherson wrote:
>> On Tue, Aug 09, 2022, Mathieu Desnoyers wrote:
>>> ----- On Aug 9, 2022, at 8:21 AM, Mathieu Desnoyers
>>> mathieu.desnoyers@efficios.com wrote:
>>>> ----- Gavin Shan <gshan@redhat.com> wrote:
>>>>> On 8/9/22 5:16 PM, Florian Weimer wrote:
>>>>>>>> __builtin_thread_pointer doesn't work on all architectures/GCC
>>>>>>>> versions.
>>>>>>>> Is this a problem for selftests?
>>>>>>>>
>>>>>>>
>>>>>>> It's a problem as the test case is running on all architectures. I think I
>>>>>>> need introduce our own __builtin_thread_pointer() for where it's not
>>>>>>> supported: (1) PowerPC  (2) x86 without GCC 11
>>>>>>>
>>>>>>> Please let me know if I still have missed cases where
>>>>>>> __buitin_thread_pointer() isn't supported?
>>>>>>
>>>>>> As far as I know, these are the two outliers that also have rseq
>>>>>> support.  The list is a bit longer if we also consider non-rseq
>>>>>> architectures (csky, hppa, ia64, m68k, microblaze, sparc, don't know
>>>>>> about the Linux architectures without glibc support).
>>>>>>
>>>>>
>>>>> For kvm/selftests, there are 3 architectures involved actually. So we
>>>>> just need consider 4 cases: aarch64, x86, s390 and other. For other
>>>>> case, we just use __builtin_thread_pointer() to maintain code's
>>>>> integrity, but it's not called at all.
>>>>>
>>>>> I think kvm/selftest is always relying on glibc if I'm correct.
>>>>
>>>> All those are handled in the rseq selftests and in librseq. Why duplicate all
>>>> that logic again?
>>>
>>> More to the point, considering that we have all the relevant rseq registration
>>> code in tools/testing/selftests/rseq/rseq.c already, and the relevant thread
>>> pointer getter code in tools/testing/selftests/rseq/rseq-*thread-pointer.h,
>>> is there an easy way to get test applications in tools/testing/selftests/kvm
>>> and in tools/testing/selftests/rseq to share that common code ?
>>>
>>> Keeping duplicated compatibility code is bad for long-term maintainability.
>> 
>> Any reason not to simply add tools/lib/rseq.c and then expose a helper to get
>> the
>> registered rseq struct?
>> 
> 
> There are couple of reasons, not to share
> tools/testing/selftests/rseq/librseq.so
> or add tools/lib/librseq.so. Please let me know if the arguments making sense
> to you?
> 
> - By design, selftests/rseq and selftests/kvm are parallel. It's going to
> introduce
>   unnecessary dependency for selftests/kvm to use selftests/rseq/librseq.so. To
>   me,
>   it makes the maintainability even harder.

In terms of build system, yes, selftests/rseq and selftests/kvm are side-by-side,
and I agree it is odd to have a cross-dependency.

That's where moving rseq.c to tools/lib/ makes sense.

> 
> - What selftests/kvm needs is rseq-thread-pointer.h, which accounts for ~5% of
>   functionalities, provided by selftests/rseq/librseq.so.

I've never seen this type of argument used to prevent using a library before, except
on extremely memory-constrained devices, which is not our target here.

Even if you would only use 1% of the features of a library, it does not justify
reimplementing that 1% if that code already sits within the same project (kernel
selftests).

> 
> - I'm not too much familiar with selftests/rseq, but it seems it need heavy
>   rework before it can become tools/lib/librseq.so. However, I'm not sure if
>   the effort is worthwhile. The newly added library is fully used by
>   testtests/rseq. ~5% of that is going to be used by selftests/kvm.
>   In this case, we still have cross-dependency issue.

No, it's just moving files around and a bit of Makefile modifications. That's
the simple part.

> 
> I personally prefer not to use selftests/rseq/librseq.so or add
> tools/lib/librseq.so,
> but I need your feedback. Please share your thoughts.

I strongly favor that we use a two steps approach:

1) immediate fix: include ../rseq/rseq.c into your test code and use the headers,
   as proposed by Paolo.

2) I'll move librseq code into tools/lib/ and tools/include/rseq/, and adapt the
   users accordingly. (after the end of my vacation)

Thanks,

Mathieu

> Thanks,
> Gavin

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com