Re: [PATCH net-next] configs/debug: run and debug PREEMPT

Re: [PATCH net-next] configs/debug: run and debug PREEMPT

[Jakub Kicinski]

On Tue, 8 Apr 2025 20:18:26 +0200 Matthieu Baerts wrote:
> On 02/04/2025 19:23, Stanislav Fomichev wrote:
> > Recent change [0] resulted in a "BUG: using __this_cpu_read() in
> > preemptible" splat [1]. PREEMPT kernels have additional requirements
> > on what can and can not run with/without preemption enabled.
> > Expose those constrains in the debug kernels.  
> 
> Good idea to suggest this to find more bugs!
> 
> I did some quick tests on my side with our CI, and the MPTCP selftests
> seem to take a bit more time, but without impacting the results.
> Hopefully, there will be no impact in slower/busy environments :)

What kind of slow down do you see? I think we get up to 50% more time
spent in the longer tests. Not sure how bad is too bad.. I'm leaning
towards applying this to net-next and we can see if people running
on linux-next complain?

Let me CC kselftests, patch in question:
https://lore.kernel.org/all/20250402172305.1775226-1-sdf@fomichev.me/

Re: [PATCH net-next] configs/debug: run and debug PREEMPT

[Matthieu Baerts]

Hi Jakub,

On 08/04/2025 21:03, Jakub Kicinski wrote:
> On Tue, 8 Apr 2025 20:18:26 +0200 Matthieu Baerts wrote:
>> On 02/04/2025 19:23, Stanislav Fomichev wrote:
>>> Recent change [0] resulted in a "BUG: using __this_cpu_read() in
>>> preemptible" splat [1]. PREEMPT kernels have additional requirements
>>> on what can and can not run with/without preemption enabled.
>>> Expose those constrains in the debug kernels.  
>>
>> Good idea to suggest this to find more bugs!
>>
>> I did some quick tests on my side with our CI, and the MPTCP selftests
>> seem to take a bit more time, but without impacting the results.
>> Hopefully, there will be no impact in slower/busy environments :)
> 
> What kind of slow down do you see? I think we get up to 50% more time
> spent in the longer tests.

That's difficult to measure in our CI because we have a majority of
tests either creating test envs with random parameters (latency, losses,
etc.), or waiting for a transfer at a limited speed to finish. Plus, we
don't control the host running our tests. But if we omit that, our
packetdrill tests take ~20% longer on the CI, and our 2 mini KUnit tests
took ~10% longer (275ms -> 305ms). Globally, our test suite took maybe
~10-20% longer, and that's acceptable.

So not 50%. Is this difference acceptable for NIPA? Even when some tests
are restarted automatically in case of instabilities?

One last thing, Stanislav's patch has been shared during Linus' merge
window: perhaps something else could also impact the time?

> Not sure how bad is too bad..

Did you observe more instabilities? Maybe the individual results should
be omitted, and only debug specific issues (calltraces, kmemleak, etc.)
should be looked at?

> I'm leaning
> towards applying this to net-next and we can see if people running
> on linux-next complain?

Good idea! But I do wonder how run **and monitor** the selftests in
linux-next with a debug kernel :)

Cheers,
Matt
-- 
Sponsored by the NGI0 Core fund.

Re: [PATCH net-next] configs/debug: run and debug PREEMPT

[Jakub Kicinski]

On Wed, 9 Apr 2025 13:58:13 +0200 Matthieu Baerts wrote:
> On 08/04/2025 21:03, Jakub Kicinski wrote:
> > On Tue, 8 Apr 2025 20:18:26 +0200 Matthieu Baerts wrote:  
> >> On 02/04/2025 19:23, Stanislav Fomichev wrote:  
> >>> Recent change [0] resulted in a "BUG: using __this_cpu_read() in
> >>> preemptible" splat [1]. PREEMPT kernels have additional requirements
> >>> on what can and can not run with/without preemption enabled.
> >>> Expose those constrains in the debug kernels.    
> >>
> >> Good idea to suggest this to find more bugs!
> >>
> >> I did some quick tests on my side with our CI, and the MPTCP selftests
> >> seem to take a bit more time, but without impacting the results.
> >> Hopefully, there will be no impact in slower/busy environments :)  
> > 
> > What kind of slow down do you see? I think we get up to 50% more time
> > spent in the longer tests.  
> 
> That's difficult to measure in our CI because we have a majority of
> tests either creating test envs with random parameters (latency, losses,
> etc.), or waiting for a transfer at a limited speed to finish. Plus, we
> don't control the host running our tests. But if we omit that, our
> packetdrill tests take ~20% longer on the CI, and our 2 mini KUnit tests
> took ~10% longer (275ms -> 305ms). Globally, our test suite took maybe
> ~10-20% longer, and that's acceptable.
> 
> So not 50%. Is this difference acceptable for NIPA? Even when some tests
> are restarted automatically in case of instabilities?

We also see 10%+ on most cases, the 50% was the worst one I glanced.
The worst offenders in terms of runtime only increased by 10% so still
within the guidelines.

> One last thing, Stanislav's patch has been shared during Linus' merge
> window: perhaps something else could also impact the time?
> 
> > Not sure how bad is too bad..  
> 
> Did you observe more instabilities? Maybe the individual results should
> be omitted, and only debug specific issues (calltraces, kmemleak, etc.)
> should be looked at?

A couple but unclear at this stage whether that was just the merge
window or enabling preempt debug. Now patchwork is super unstable
so again, hard to judge the source of the problems :(

> > I'm leaning
> > towards applying this to net-next and we can see if people running
> > on linux-next complain?  
> 
> Good idea! But I do wonder how run **and monitor** the selftests in
> linux-next with a debug kernel :)

One way to find out :)