Re: [PATCH 0/2] livepatch: Move tests from lib/livepatch to selftests/livepatch

[Shuah Khan <skhan@linuxfoundation.org>]

On 6/10/22 8:48 AM, Petr Mladek wrote:
> On Fri 2022-06-10 09:06:16, Joe Lawrence wrote:
>> On 6/9/22 4:16 PM, Shuah Khan wrote:
>>> On 6/3/22 8:32 AM, Marcos Paulo de Souza wrote:
>>>> Hi there,
>>>>
>>>> The first patch moves the current livepatch tests to selftests,
>>>> allowing it
>>>> be better suited to contain more complex tests, like using userspace C
>>>> code
>>>> to use the livepatched kernel code. As a bonus it allows to use
>>>> "gen_tar" to export the livepatch selftests, rebuild the modules by
>>>> running make in selftests/livepatch directory and simplifies the process
>>>> of creating and debugging new selftests.
>>>>
>>>
>>> In general selftests don't include modules. We keep test modules under lib.
>>> One of the reasons is that modules have dependencies on the kernel and
>>> should
>>> be built when kernel is built.
>>>
>>> I don't fully buy the argument that moving modules under selftest would
>>> simplify
>>> the process.
>>>
>>
>> Hi Shuah,
>>
>> I see that there is tools/testing/selftests/bpf/bpf_testmod/ which
>> claims to be a "conceptually out-of-tree module".  Would similarly
>> moving livepatch test modules under tools/ give us flexibility to write
>> them build for multiple kernel versions?  Then one could theoretically
>> build and run the latest, greatest selftests against older kernels
>> (assuming the associate script/module/kernel supports the idea)?
> 
> +1
> 
> Another motivation is that the new selftest also needs
> an executable binary. It would be nice to handle both modules
> and binaries the same way.
> 
> Honestly, lib/* is a mess. It mixes real functionality and test
> modules. The relation between the modules and tools/testing/*
> is far from clear. IMHO, it would be more clean to have the related
> stuff together.
> 
> Of course, we could not move all test modules from lib/* easily.
> Some of them might be used on its own or even as built-in
> tests. But preventing the move looks like a step in
> the wrong direction to me.
> 

As such bpf_testmod is the only one that is currently under kselftests.
I don't have an objection to it from technical stand point. My concern
is more from the standpoint of people writing modules that can't be built
out of tree. We would add another requirement to kselftest that the out
of tree modules should build successfully.

As long as that concern is addressed and also test gracefully fails if the
module fails to build, we can move on that direction. I would hesitate to
extend this to modules dependent on hardware and architecture features
such as cpufreq test drivers for example.

thanks,
-- Shuah