Re: [tip: sched/core] sched/tracing: Don't re-read p->state when emitting sched_switch event

["Holger Hoffstätte" <holger@applied-asynchrony.com>]

On 2022-04-11 09:28, Greg KH wrote:
> On Mon, Apr 11, 2022 at 09:18:19AM +0200, Holger Hoffstätte wrote:
>> On 2022-04-11 01:22, Holger Hoffstätte wrote:
>>> On 2022-04-11 00:06, Qais Yousef wrote:
>>>> On 04/10/22 00:38, Qais Yousef wrote:
>>>>> On 03/08/22 18:51, Qais Yousef wrote:
>>>>>> On 03/08/22 19:10, Greg KH wrote:
>>>>>>> On Tue, Mar 08, 2022 at 06:02:40PM +0000, Qais Yousef wrote:
>>>>>>>> +CC stable
>>>>>>>>
>>>>>>>> On 03/01/22 15:24, tip-bot2 for Valentin Schneider wrote:
>>>>>>>>> The following commit has been merged into the sched/core branch of tip:
>>>>>>>>>
>>>>>>>>> Commit-ID:     fa2c3254d7cfff5f7a916ab928a562d1165f17bb
>>>>>>>>> Gitweb:        https://git.kernel.org/tip/fa2c3254d7cfff5f7a916ab928a562d1165f17bb
>>>>>>>>> Author:        Valentin Schneider <valentin.schneider@arm.com>
>>>>>>>>> AuthorDate:    Thu, 20 Jan 2022 16:25:19
>>>>>>>>> Committer:     Peter Zijlstra <peterz@infradead.org>
>>>>>>>>> CommitterDate: Tue, 01 Mar 2022 16:18:39 +01:00
>>>>>>>>>
>>>>>>>>> sched/tracing: Don't re-read p->state when emitting sched_switch event
>>>>>>>>>
>>>>>>>>> As of commit
>>>>>>>>>
>>>>>>>>>     c6e7bd7afaeb ("sched/core: Optimize ttwu() spinning on p->on_cpu")
>>>>>>>>>
>>>>>>>>> the following sequence becomes possible:
>>>>>>>>>
>>>>>>>>>                p->__state = TASK_INTERRUPTIBLE;
>>>>>>>>>                __schedule()
>>>>>>>>>              deactivate_task(p);
>>>>>>>>>     ttwu()
>>>>>>>>>       READ !p->on_rq
>>>>>>>>>       p->__state=TASK_WAKING
>>>>>>>>>              trace_sched_switch()
>>>>>>>>>                __trace_sched_switch_state()
>>>>>>>>>                  task_state_index()
>>>>>>>>>                    return 0;
>>>>>>>>>
>>>>>>>>> TASK_WAKING isn't in TASK_REPORT, so the task appears as TASK_RUNNING in
>>>>>>>>> the trace event.
>>>>>>>>>
>>>>>>>>> Prevent this by pushing the value read from __schedule() down the trace
>>>>>>>>> event.
>>>>>>>>>
>>>>>>>>> Reported-by: Abhijeet Dharmapurikar <adharmap@quicinc.com>
>>>>>>>>> Signed-off-by: Valentin Schneider <valentin.schneider@arm.com>
>>>>>>>>> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
>>>>>>>>> Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
>>>>>>>>> Link: https://lore.kernel.org/r/20220120162520.570782-2-valentin.schneider@arm.com
>>>>>>>>
>>>>>>>> Any objection to picking this for stable? I'm interested in this one for some
>>>>>>>> Android users but prefer if it can be taken by stable rather than backport it
>>>>>>>> individually.
>>>>>>>>
>>>>>>>> I think it makes sense to pick the next one in the series too.
>>>>>>>
>>>>>>> What commit does this fix in Linus's tree?
>>>>>>
>>>>>> It should be this one: c6e7bd7afaeb ("sched/core: Optimize ttwu() spinning on p->on_cpu")
>>>>>
>>>>> Should this be okay to be picked up by stable now? I can see AUTOSEL has picked
>>>>> it up for v5.15+, but it impacts v5.10 too.
>>>>
>>>> commit: fa2c3254d7cfff5f7a916ab928a562d1165f17bb
>>>> subject: sched/tracing: Don't re-read p->state when emitting sched_switch event
>>>>
>>>> This patch has an impact on Android 5.10 users who experience tooling breakage.
>>>> Is it possible to include in 5.10 LTS please?
>>>>
>>>> It was already picked up for 5.15+ by AUTOSEL and only 5.10 is missing.
>>>>
>>>
>>> https://lore.kernel.org/stable/Yk2PQzynOVOzJdPo@kroah.com/
>>>
>>> However, since then further investigation (still in progress) has shown that this
>>> may have been the fault of the tool in question, so if you can verify that tracing
>>> sched still works for you with this patch in 5.15.x then by all means
>>> let's merge it.
>>
>> So it turns out the lockup is indeed the fault of the tool, which contains multiple
>> kernel-version dependent tracepoint definitions and now fails with this
>> patch.
> 
> What tools is this?

sysdig - which uses a helper kernel module which accesses tracepoints, but of course
(as I just found) with copypasta'd TP definitions, which broke with this patch due to
the additional parameter in the function signature. It's been prone to breakage forever
because of a lack of a stable kernel ABI.

Took me a while to find/figure out, but IMHO better safe than sorry. We've had
autoselected scheduler patches before that looked fine but really were not.

> 
>> Greg, please re-enqueue this patch where necessary (5.10, 5.15+)
> 
> If I queue it up again, will the tools keep breaking?

Yes, but that's their problem with an out-of-tree module; a few more #ifdefs
are not going to make a big difference.

thanks
Holger