Re: 2.6.38-rc2: Uhhuh. NMI received for unknown reason 2d on CPU 0.

[Cyrill Gorcunov <gorcunov@gmail.com>]

On 02/01/2011 07:27 PM, George Spelvin wrote:
> Since upgrading to -rc2 (-rc3 is compiling right now), I've been getting
> complaints at irregular intervals.  This didn't used to happen with 2.6.37.
> 
...
> Should I bisect this, or does someone know what might be happening?
> 
> Thank you!
> 

 I fear it's known issue at moment, we're trying to resolve it. There is
an option -- to disable nmi_watchdog (nmi_watchdog=0 boot option).

 But if you have a will or would like to help debug the problem -- mind to
try the patch below? Note the patch is ugly at moment and must *not* be
running on non-P4 system (and I only compile-tested it so no guarantees
at all, and I've CC'ed a couple of people as well)

    Cyrill

---
 arch/x86/kernel/cpu/perf_event.c    |   12 +++++++++++-
 arch/x86/kernel/cpu/perf_event_p4.c |    8 +++++++-
 2 files changed, 18 insertions(+), 2 deletions(-)

Index: linux-2.6.git/arch/x86/kernel/cpu/perf_event.c
=====================================================================
--- linux-2.6.git.orig/arch/x86/kernel/cpu/perf_event.c
+++ linux-2.6.git/arch/x86/kernel/cpu/perf_event.c
@@ -1075,7 +1075,17 @@ static void x86_pmu_start(struct perf_ev

 	cpuc->events[idx] = event;
 	__set_bit(idx, cpuc->active_mask);
-	__set_bit(idx, cpuc->running);
+	if (1) {
+		/* running mask is shared across a core */
+		int leader_cpu;
+		struct cpu_hw_events *leader_cpuc;
+
+		leader_cpu	= cpumask_first(__get_cpu_var(cpu_sibling_map));
+		leader_cpuc	= &per_cpu(cpu_hw_events, leader_cpu);
+
+		__set_bit(idx, leader_cpuc->running);
+	} else
+		__set_bit(idx, cpuc->running);
 	x86_pmu.enable(event);
 	perf_event_update_userpage(event);
 }
Index: linux-2.6.git/arch/x86/kernel/cpu/perf_event_p4.c
=====================================================================
--- linux-2.6.git.orig/arch/x86/kernel/cpu/perf_event_p4.c
+++ linux-2.6.git/arch/x86/kernel/cpu/perf_event_p4.c
@@ -907,8 +907,14 @@ static int p4_pmu_handle_irq(struct pt_r
 		int overflow;

 		if (!test_bit(idx, cpuc->active_mask)) {
+			int leader_cpu;
+			struct cpu_hw_events *leader_cpuc;
+
+			leader_cpu	= cpumask_first(__get_cpu_var(cpu_sibling_map));
+			leader_cpuc	= &per_cpu(cpu_hw_events, leader_cpu);
+
 			/* catch in-flight IRQs */
-			if (__test_and_clear_bit(idx, cpuc->running))
+			if (__test_and_clear_bit(idx, leader_cpuc->running))
 				handled++;
 			continue;
 		}