Re: 0.9.3 fails on amd64 2.6.19

[Philippe E. <ph...@wa...>]

On Sat, 21 Jul 2007 at 16:27 +0000, John Levon wrote:

> On Sat, Jul 21, 2007 at 11:21:13AM -0400, William Cohen wrote:
> 
> > That patch is the patch mentioned november 17, "Multiple users of the 
> > performance counters". It adjusts the allocation mechanism to only use the 
> > registers that are actually available. This became an issue with the watchdog 
> > timer code grabbing one of the performance counters. I am not sure why it wasn't 
> > actually checked in.
> 
> They broke userspace compatibility. Haven't they fixed that yet?

It's fixed and I'm happy with the used solution, I understand now where
the confusion come from. CONFIG_WATCHDOG=no is no longer sufficient, nowaday
it means don't build driver for hardware watchdog, the watchdog through
local/io apic is always built in if an apic is present so always true with
SMP. I did not look why boot params no longer works, quite possible it has
been broken only in some kernel version and work now, anyway the right way is:

echo 0 > /proc/sys/kernel/nmi_watchdog

which is lot better than the old solution.

People getting this problem must also do an opcontrol --deinit, because
the driver is incorrectly initialized in such case.

Will I did that http://oprofile.cvs.sourceforge.net/oprofile/oprofile/utils/opcontrol?r1=1.146&r2=1.147 in opcontrol, it breaks your patch for RH. If you
want to continue to use your patch in RH you'll need to turn it as a warning
ala "if you want to use all counter do echo .... " and do not error out.

John, do we want to save /proc/sys/kernel/nmi_watchdog and restore it
at opcontrol --deinit time and hide this thing ?

-- 
regards,
Philippe Elie