Re: deadlock in scheduler enabling HRTICK feature

[Andy Lutomirski <luto@amacapital.net>]

On 06/27/2013 03:53 AM, Peter Zijlstra wrote:
> On Thu, Jun 27, 2013 at 12:43:09PM +0200, Peter Zijlstra wrote:
>> On Wed, Jun 26, 2013 at 10:46:33AM -0600, David Ahern wrote:
>>> On 6/26/13 1:05 AM, Peter Zijlstra wrote:
>>>>> What is the expectation that the feature provides? not a whole lot of
>>>>> documentation on it. I walked down the path wondering if it solved an odd
>>>>> problem we are seeing with the CFS in 2.6.27 kernel.
>>>>
>>>> Its supposed to use hrtimers for slice expiry instead of the regular tick.
>>>
>>> So theoretically CPU bound tasks would get preempted sooner? That was my
>>> guess/hope anyways.
>>
>> Doth the below worketh?
>>
> 
> Related to all this; the reason its not enabled by default is that mucking
> about with hrtimers all the while is god awful expensive.
> 
> I've had ideas about making this a special purpose 'hard-coded' timer in the
> hrtimer guts that's only ever re-programmed when the new value is sooner.
> 
> By making it a 'special' timer we can avoid the whole rb-tree song and dance;
> and by taking 'spurious' short interrupts we can avoid prodding the hardware
> too often.

Supposedly, on really new x86 systems, the TSC deadline timer is so fast
to reprogram that this isn't worth it.

I wonder if the general timing code should have a way to indicate that a
given clockevent is (a) very fast and (b) is actually locked to a
clocksource as opposed to just having a vaguely accurately calibrated
frequency.

--Andy