Re: Solid freezes with 2.6.25

[Bernhard Walle <bwalle@suse.de>]

* Gabor Gombas [2008-07-25 10:52]:
>
> On Sun, Jun 15, 2008 at 07:11:41PM +0200, Gabor Gombas wrote:
> > On Tue, May 13, 2008 at 04:39:29PM +0200, Gabor Gombas wrote:
> > > On Mon, May 12, 2008 at 02:21:46PM +0200, Bernhard Walle wrote:
> > > 
> > > > Well, can you try if that really not happens without
> > > > CONFIG_HPET_EMULATE_RTC? The question is if that 64 interrupts per second really
> > > > cause a machine hang ...
> > > 
> > > I'll try when I have a little time...
> > 
> > Ok, it took much more time than I'd have liked, but now I'm having:
> > 
> > $ uname -a
> > Linux boogie 2.6.25 #3 SMP Thu Jun 12 13:32:35 CEST 2008 x86_64 GNU/Linux
> > $ grep EMULATE_RTC /boot/config-2.6.25
> > # CONFIG_HPET_EMULATE_RTC is not set
> > $ uptime
> >  19:07:17 up 3 days,  4:07, 12 users,  load average: 0.14, 0.31, 0.20
> > 
> > So I'd say it works.
> 
> FYI, the bug is still present in 2.6.26. I also got the same hang on a
> Dell PowerEdge 850 using a Debian-provided kernel (based on 2.6.25.6);
> that is the 3rd distinct HW configuration that hangs.

Did you try that patch?

-------------- 8< ---------------------------------------------------
From: David Brownell <dbrownell@users.sourceforge.net>

Resolve http://bugzilla.kernel.org/show_bug.cgi?id=11051 and other bugs
related to the way the HPET glue code in rtc-cmos was incomplete and
inconsistent:

 * Switch the approach so that the basic driver code flow isn't
   changed by having HPET ... instead, just have HPET shadow the
   RTC_CONTROL irq enables and RTC_FREQ_SELECT data.  It's only
   coping with IRQ thievery, after all.

 * Do that consistently (!!) to avoid problems when the HPET code
   is out of sync with the real RTC intent.  Examples include:

   - cmos_procfs(), which now reports correct data

   - cmos_irq_set_state() ... also removing the previous PIE_{ON,OFF}
     ioctl support so only one code path manages "periodic" IRQs

   - cmos_do_shutdown() ... currently a "just in case" change.

   - cmos_suspend() and cmos_resume() ... also handling a bug that
     was specific to HPET's IRQ thievery, where the alarm wasn't
     disabled after waking the system

 * Always call that HPET code under the RTC spinlock (it doesn't do
   its own locking)
-------------- >8 ---------------------------------------------------


Bernhard
-- 
Bernhard Walle, SUSE LINUX Products GmbH, Architecture Development