Re: [patch] CFS scheduler, -v19

[Bill Davidsen <davidsen@tmr.com>]

Linus Torvalds wrote:
> On Tue, 17 Jul 2007, Ingo Molnar wrote:
>   
>> * Ian Kent <raven@themaw.net> wrote:
>>     
>>> In several places I have code similar to:
>>>
>>> wait.tv_sec = time(NULL) + 1;
>>> wait.tv_nsec = 0;
>>>       
>
> Ok, that definitely should work.
>
> Does the patch below help?
>
>   
Spectacularly no! With this patch the "glitch1" script with multiple 
scrolling windows has all xterms and glxgears stop totally dead for 
~200ms once per second. I didn't properly test anything else after that. 
Since the automount issue doesn't seem to start until something kicks it 
off, I didn't see it but that doesn't mean it's fixed.
>> ah! It passes in a low-res time source into a high-res time interface 
>> (pthread_cond_timedwait()). Could you change the time(NULL) + 1 to 
>> time(NULL) + 2, or change it to:
>>
>> 	gettimeofday(&wait, NULL);
>> 	wait.tv_sec++;
>>     
>
> This is wrong. It's wrong for two reasons:
>
>  - it really shouldn't be needed. I don't think "time()" has to be 
>    *exactly* in sync, but I don't think it can be off by a third of a 
>    second or whatever (as the "30% CPU load" would seem to imply)
>
>  - gettimeofday works on a timeval, pthread_cond_timedwait() works on a 
>    timespec.
>
> So if it actually makes a difference, it makes a difference for the 
> *wrong* reason: the time is still totally nonsensical in the tv_nsec field 
> (because it actually got filled in with msecs!), but now the tv_sec field 
> is in sync, so it hides the bug.
>
> Anyway, hopefully the patch below might help. But we probably should make 
> this whole thing a much more generic routine (ie we have our internal 
> "getnstimeofday()" that still is missing the second-overflow logic, and 
> that is quite possibly the one that triggers the "30% off" behaviour).
>
>   
Hope that info helps.

> Ingo, I'd suggest:
>  - ger rid of "timespec_add_ns()", or at least make it return a return 
>    value for when it overflows.
>  - make all the people who overflow into tv_sec call a "fix_up_seconds()" 
>    thing that does the xtime overflow handling.
>
> 		Linus
>