lkml.org 
[lkml]   [2009]   [Jun]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] tick: add check for the existence of broadcast clock event device
On Sun, 7 Jun 2009 00:18:55 +0800
Thomas Gleixner <tglx@linutronix.de> wrote:

> On Sat, 6 Jun 2009, Thomas Gleixner wrote:
> > Feng,
> >
> > On Sat, 6 Jun 2009, Feng Tang wrote:
> > > > If your percpu devices are always on (not affected by C3 stop)
> > > > then you never dereference bc. So why do we need an extra check
> > > > for !bc ?
> > >
> > > Hi tglx,
> >
> > > Thanks for the explanation. But we really ran into the NULL
> > > pointer case, in our platform, there are 2 X86 CPUs which have
> > > lapic, also it has 2 external timers which are pretty similar
> > > with HPET timers, those 2 external timers will be used as per-cpu
> > > timers (higher rating than lapic timer). In system's power cycle
> > > of suspend and resume, disable_nontboot_cpus will be called
> > > before goto suspend state,and enable_nonboot_cpus will be called
> > > for the resume process, so lapic timer of cpu1 will be first
> > > registered as per-cpu timer, and our external timer will be
> > > registered later after get a CPU_ONLINE notifier (similar with
> > > HPET), right in this time slot that lapic is the per-cpu timer,
> > > when system get the CLOCK_EVT_BROADCAST_ENTER/EXIT msg,
> > > tick_do_broadcast_on_off() is called and hit the NULL pointer
> > > case.
> >
> > Ok, I can understand now why we need it. I'll apply your patch and
> > add some more info into the commit msg so we do not look at it in a
> > year and scratch our heads. :)
>
> Hmm, thought more about it.
>
> 1) How do you calibrate the local APIC timer if you do not have some
> initial timer device ?
Yes, we have external timer device with the name "apbt"
>
> 2) If you have some initial timer device (PIT/HPET) why isn't it
> registered as broadcast device.
We only have 2 available apbt, apbt0 for cpu0 and apbt1 for cpu1. And based on my understanding, there is no explicit API to register a timer as bc, and bc is setup during clockevents_register_device() --> tick_check_new_device() -->tick_check_broadcast_device(), and HPET case can just meet the condition: lapic's rating is 110, timer "hpet"is 50 and set to be the broadcast device, and other hpet(2/3/4/5..)'s are 150 and set to be the per-cpu timer

>
> 3) When the CPU uses the local APIC before the external timer is
> initialized what happens if the system goes into a deeper c-state ?
Good point, for cpu1, this is not an issue, as cpu0 will register apbt1 for cpu1. there is a window cpu1 is in deep c-idle state and no timer wake it up, but window is small as cpu0 will send IPI soon.
>
> Thanks,
>
> tglx


\
 
 \ /
  Last update: 2009-06-08 04:01    [from the cache]
©2003-2011 Jasper Spaans