Re: [PATCH 1/2] [RT] hrtimers stuck in waitqueue

[Gregory Haskins <ghaskins@novell.com>]

[-- Attachment #1: Type: text/plain, Size: 1649 bytes --]

Hi Gilles,

Gilles Carry wrote:
>
> Le 19 août 08 à 16:10, Gregory Haskins a écrit :
>
>> Hi Gilles,
>>
>> Gilles Carry wrote:
>>> This patch makes hrtimers initialized with hrtimer_init_sleeper
>>> to use another mode and then not be stuck in waitqueues when
>>> hrtimer_interrupt is very busy.
>>>  
>>
>> Not sure if this is related, but I just posted a patch to fix a 
>> condition where hrtimer_interrupt could get artificially busy for 
>> *very* long periods of time (I observed 500ms).
>>
>> http://article.gmane.org/gmane.linux.rt.user/3357
>>
>> This bug was specific to 26-rt1.  Did you investigate the cause of 
>> the busy hrtimer_interrupt?  If not, its possible that your issue 
>> would be indirectly fixed now that the HRT subsystem doesn't 
>> spike-out like that.  Or do you believe these are orthogonal issues 
>> and we should still protect against non-pathological cases of 
>> hrtimer_interrupt running hot?
>
> Hello Greg,
> The cause of busy hrtimer interrupt is the test that stresses the 
> system. Many timers expire and are treated in a single interrupt run.
>
> Anyway looking at __run_hrtimer we can see:
> spin_unlock (&cpu_base->lock);
> restart = fn(timer);
> spin_lock(&cpu_base->lock);
>
> I guess It is the point where awaken threads might migrate to another cpu and reach hrtimer_cancel *before* __run_hrtimer has changed the timer state. (at the bottom of the function)
> This is why a new state machine is needed.

It sounds reasonable that both fixes are needed, then.  Thanks.

-Greg

>
> Regards,
> Gilles.
>
>>
>>
>> Regards,
>> -Greg
>>
>>
>>
>



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 257 bytes --]