Re: [PATCH] sched: Move task_mm_cid_work to mm delayed work

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

On 2024-12-09 10:33, Mathieu Desnoyers wrote:
> On 2024-12-09 08:45, Gabriele Monaco wrote:
>>
>>> Thinking back on this, you'll want a program that does the following
>>> on a system with N CPUs:
>>>
>>> - Phase 1: run one thread per cpu, pinned on each cpu. Print the
>>>     mm_cid from each thread with the cpu number every second or so.
>>>
>>> - Exit all threads except the main thread, join them from the main
>>>     thread,
>>>
>>> - Phase 2: the program is now single-threaded. We'd expect the
>>>     mm_cid value to converge towards 0 as the periodic task clears
>>>     unused CIDs.
>>>
>>> So I think in phase 2 we can have an actual automated test: If after
>>> an order of magnitude more time than the 100ms delay between periodic
>>> tasks we still observe mm_cid > 0 in phase 2, then something is
>>> wrong.
>>
>> Been thinking about this and came up with a simple draft, I'll probably
>> send it as a separate patch.
>>
>> Doing this can lead to false positives: the main thread may be assigned
>> the mm_cid 0 and keep it till the end, in this scenario the other
>> threads (CPUs) would get different mm_cids and exit, the main thread
>> will still have 0 and pass the test regardless.
>>
>> I have an idea to make it a bit more robust: we can run threads as you
>> described in phase 1, stop all but one (let's say the one running on
>> the last core), make sure the main thread doesn't accidentally run on
>> the same core by pinning to core 0 and wait until we see the 2
>> remaining threads holding 0 and 1, in any order.
>> Besides a special case if we have only 1 available core, this should
>> work fine, sure we could get false positives but it seems to me much
>> less likely.
>>
>> Does it make sense to you?
> 
> A small tweak on your proposed approach: in phase 1, get each thread
> to publish which mm_cid they observe, and select one thread which
> has observed mm_cid > 1 (possibly the largest mm_cid) as the thread
> that will keep running in phase 2 (in addition to the main thread).
> 
> All threads other than the main thread and that selected thread exit
> and are joined before phase 2.
> 
> So you end up in phase 2 with:
> 
> - main (observed any mm_cid)
> - selected thread (observed mm_cid > 1, possibly largest)
> 
> Then after a while, the selected thread should observe a
> mm_cid <= 1.
> 
> This test should be skipped if there are less than 3 CPUs in
> allowed cpumask (sched_getaffinity).

Even better:

For a sched_getaffinity with N cpus:

- If N == 1 -> skip (we cannot validate anything)

Phase 1: create N - 1 pthreads, each pinned to a CPU. main thread
also pinned to a cpu.

Publish the mm_cids observed by each thread, including main thread.

Select a new leader for phase 2: a thread which has observed nonzero
mm_cid. Each other thread including possibly main thread issue
pthread_exit, and the new leader does pthread join on each other.

Then check that the new leader eventually observe mm_cid == 0.

And it works with an allowed cpu mask that has only 2 cpus.

Thanks,

Mathieu

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