Re: [RFC 2/3] preempt_tracer: Disable IRQ while starting/stopping due to a preempt_counter change

[Daniel Bristot de Oliveira]

On 31/05/2019 09:47, Joel Fernandes wrote:
> On Wed, May 29, 2019 at 11:40:34AM +0200, Daniel Bristot de Oliveira wrote:
>> On 29/05/2019 10:33, Peter Zijlstra wrote:
>>> On Tue, May 28, 2019 at 05:16:23PM +0200, Daniel Bristot de Oliveira wrote:
>>>> The preempt_disable/enable tracepoint only traces in the disable <-> enable
>>>> case, which is correct. But think about this case:
>>>>
>>>> ---------------------------- %< ------------------------------
>>>> 	THREAD					IRQ
>>>> 	   |					 |
>>>> preempt_disable() {
>>>>     __preempt_count_add(1)
>>>> 	------->	    smp_apic_timer_interrupt() {
>>>> 				preempt_disable()
>>>> 				    do not trace (preempt count >= 1)
>>>> 				    ....
>>>> 				preempt_enable()
>>>> 				    do not trace (preempt count >= 1)
>>>> 			    }
>>>>     trace_preempt_disable();
>>>> }
>>>> ---------------------------- >% ------------------------------
>>>>
>>>> The tracepoint will be skipped.
>>>
>>> .... for the IRQ. But IRQs are not preemptible anyway, so what the
>>> problem?
>>
>>
>> right, they are.
>>
>> exposing my problem in a more specific way:
>>
>> To show in a model that an event always takes place with preemption disabled,
>> but not necessarily with IRQs disabled, it is worth having the preemption
>> disable events separated from IRQ disable ones.
>>
>> The main reason is that, although IRQs disabled postpone the execution of the
>> scheduler, it is more pessimistic, as it also delays IRQs. So the more precise
>> the model is, the less pessimistic the analysis will be.
>>
>> But there are other use-cases, for instance:
>>
>> (Steve, correct me if I am wrong)
>>
>> The preempt_tracer will not notice a "preempt disabled" section in an IRQ
>> handler if the problem above happens.
>>
>> (Yeah, I know these problems are very specific... but...)
> 
> I agree with the problem. I think Daniel does not want to miss the preemption
> disabled event caused by the IRQ disabling.

Hi Joel!

Correct, but ... look bellow.

>>>> To avoid skipping the trace, the change in the counter should be "atomic"
>>>> with the start/stop, w.r.t the interrupts.
>>>>
>>>> Disable interrupts while the adding/starting stopping/subtracting.
>>>
>>>> +static inline void preempt_add_start_latency(int val)
>>>> +{
>>>> +	unsigned long flags;
>>>> +
>>>> +	raw_local_irq_save(flags);
>>>> +	__preempt_count_add(val);
>>>> +	preempt_latency_start(val);
>>>> +	raw_local_irq_restore(flags);
>>>> +}
>>>
>>>> +static inline void preempt_sub_stop_latency(int val)
>>>> +{
>>>> +	unsigned long flags;
>>>> +
>>>> +	raw_local_irq_save(flags);
>>>> +	preempt_latency_stop(val);
>>>> +	__preempt_count_sub(val);
>>>> +	raw_local_irq_restore(flags);
>>>> +}
>>>
>>> That is hideously expensive :/
>>
>> Yeah... :-( Is there another way to provide such "atomicity"?
>>
>> Can I use the argument "if one has these tracepoints enabled, they are not
>> considering it as a hot-path?"
> 
> The only addition here seems to  the raw_local_irq_{save,restore} around the
> calls to increment the preempt counter and start the latency tracking.
> 
> Is there any performance data with the tracepoint enabled and with/without
> this patch? Like with hackbench?

I discussed this with Steve at the Summit on the Summit (the reason why I did
not reply this email earlier is because I was in the conf/traveling), and he
also agrees with peterz, disabling and (mainly) re-enabling IRQs costs too much.

We need to find another way to resolve this problem (or mitigate the cost).... :-(.

Ideas?

Thanks!!

-- Daniel