Don,

On Wed, May 10, 2006 at 10:59:53AM -0400, Don Zickus wrote:
> > 
> > This means you can at runtime enable/disbale nmi_watchdog, i.e., reserve
> > some performance counters on the fly. This gets complicated because now
> > the perfmon subsystem (and probably oprofile) cannot check register
> > availability when they are first initialized. Basically each time,
> > the /sys entry is modified, they would have to scan the list of available
> > performance counters. I don't know exactly when Oprofile does this checking.
> > For perfmon, this is done only once, when the PMU description table is 
> > loaded.

> How often did you plan on enabling/disabling the nmi_watchdog?  My
> understanding was you disable nmi_watchdog, run oprofile/perfmon,
> re-enable nmi_watchdog.  I guess I don't understand what type of funky
> scenarios you are dealing with.  
> 
I thought that part of the exercise was to allow oprofile/perfmon and 
nmi_watchdog
to work simultaneously, i.e., don't disable watchdog when you monitor. Leave 
watchdog
on and have oprofile/perfmon run with fewer counters. The alternative would be 
to do
what you say: disbale nmi, monitor, re-enable nmi.

> > 
> > Also something that I did not see in this code is the error detection in
> > case enable_lapic_nmi_watchdog() fails. Oprofile runs on all CPUs or none.
> > Perfmon lets you monitor on subsets on CPUs. In case NMI was disabled and
> > a monitoring session was active on some CPUs. The 
> > enable_lapic_nmi_watchdog()
> > will fail on some CPUs. How is that handled?
> 
> It's not.  In fact I wouldn't know what to do in such situations.  Is it
> really wrong to only have a subset of cpus being monitored by the
> nmi_watchdog?  This seems to be wandering into the area where the user is
> looking to do something complicated (profiling a subset of cpus) and as
> such might be expected to make sure the nmi_watchdog is properly enabled
> on all cpus when they are done.
> 
On larger machines (think NUMA-style), it does not always make sense to monitor 
all CPUs.

-- 

-Stephane
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/

Reply via email to