Re: perf failed with kernel 2.6.35-rc

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Tue, 2010-07-13 at 17:16 +0200, Stephane Eranian wrote:
> On Tue, Jul 13, 2010 at 10:14 AM, Zhang, Yanmin
> <yanmin_zhang@linux.intel.com> wrote:
> > Peter,
> >
> > perf doesn't work on my Nehalem EX machine.
> > 1) The 1st start of 'perf top' is ok;
> > 2) Kill the 1st perf and restart it. It doesn't work. No data is showed.
> >
> > I located below commit:
> > commit 1ac62cfff252fb668405ef3398a1fa7f4a0d6d15
> > Author: Peter Zijlstra <peterz@infradead.org>
> > Date:   Fri Mar 26 14:08:44 2010 +0100
> >
> >    perf, x86: Add Nehelem PMU programming errata workaround
> >
> >    workaround From: Peter Zijlstra <a.p.zijlstra@chello.nl>
> >    Date: Fri Mar 26 13:59:41 CET 2010
> >
> >    Implement the workaround for Intel Errata AAK100 and AAP53.
> >
> >    Also, remove the Core-i7 name for Nehalem events since there are
> >    also Westmere based i7 chips.
> >
> >
> > If I comment out the workaround in function intel_pmu_nhm_enable_all,
> > perf could work.
> >
> > A quick glance shows:
> > wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x3);
> > should be:
> > wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x7);
> >
> >
> > I triggered sysrq to dump PMU registers and found the last bit of
> > global status register is 1. I added a status reset operation like below patch:
> >
> What do you call the last bit? bit0 or bit63?
Sorry for confusing you. It's bit0, mapping to PERFMON_EVENTSEL0.

> 
> > --- linux-2.6.35-rc5/arch/x86/kernel/cpu/perf_event_intel.c     2010-07-14 09:38:11.000000000 +0800
> > +++ linux-2.6.35-rc5_fork/arch/x86/kernel/cpu/perf_event_intel.c        2010-07-14 14:41:42.000000000 +0800
> > @@ -505,8 +505,13 @@ static void intel_pmu_nhm_enable_all(int
> >                wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 1, 0x4300B1);
> >                wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + 2, 0x4300B5);
> >
> > -               wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x3);
> > +               wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x7);
> >                wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
> > +               /*
> > +                * Reset the last 3 bits of global status register in case
> > +                * previous enabling causes overflows.
> > +                */
> 
> The workaround cannot cause on overflow because the associated counters
> won't count anything given their umask value is 0 (which does not correspond
> to anything for event 0xB1, event 0xB5 is undocumented). This is for the events
> described in table A.2. If NHM-EX has a different definition for 0xB1, 0xB5,
> then that's another story.
I found the status bit is set by triggering sysrq to dump PMU registers.

If I start perf by gdb, sometimes, perf could work. I found one processor's 1st status
register is equal to 0 while other processors' are 1. If just starting perf, all 1st
status registers are equal to 1.

> 
> 
> > +               wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, 0x7);
> >
> >                for (i = 0; i < 3; i++) {
> >                        struct perf_event *event = cpuc->events[i];
> >
> >
> >
> > However, it still doesn't work. Current right way is to comment out
> > the workaround.