Skip to contentSkip to conversation

perf_counter, x86: Make NMI lockups more robust

tip:perfcounters/core tip-bot for Ingo Molnar 1 participant 1 message May 26, 2009 x86

lore
Share & tools
lore
#0 lore raw
Commit-ID:  aaba98018b8295dfa2119345d17f833d74448cd0
Gitweb:     http://git.kernel.org/tip/aaba98018b8295dfa2119345d17f833d74448cd0
Author:     Ingo Molnar <[email protected]>
AuthorDate: Tue, 26 May 2009 08:10:00 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Tue, 26 May 2009 09:52:03 +0200

perf_counter, x86: Make NMI lockups more robust

We have a debug check that detects stuck NMIs and returns with the PMU disabled in the global ctrl MSR - but i managed to trigger a situation where this was not enough to deassert the NMI.

So clear/reset the full PMU and keep the disable count balanced when exiting from here. This way the box produces a debug warning but stays up and is more debuggable.

[ Impact: in case of PMU related bugs, recover more gracefully ]

Cc: Peter Zijlstra <[email protected]>
Cc: Mike Galbraith <[email protected]>
Cc: Paul Mackerras <[email protected]>
Cc: Corey Ashford <[email protected]>
Cc: Marcelo Tosatti <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: John Kacur <[email protected]>
LKML-Reference: <new-submission>

---
 arch/x86/kernel/cpu/perf_counter.c |   26 ++++++++++++++++++++++++++
 1 files changed, 26 insertions(+), 0 deletions(-)
arch/x86/kernel/cpu/perf_counter.c +26 -0
index ece3813..2eeaa99 100644
@@ -724,6 +724,30 @@ static void intel_pmu_save_and_restart(struct perf_counter *counter)
 		intel_pmu_enable_counter(hwc, idx);
 }
 
+static void intel_pmu_reset(void)
+{
+	unsigned long flags;
+	int idx;
+
+	if (!x86_pmu.num_counters)
+		return;
+
+	local_irq_save(flags);
+
+	printk("clearing PMU state on CPU#%d\n", smp_processor_id());
+
+	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
+		checking_wrmsrl(x86_pmu.eventsel + idx, 0ull);
+		checking_wrmsrl(x86_pmu.perfctr  + idx, 0ull);
+	}
+	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
+		checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
+	}
+
+	local_irq_restore(flags);
+}
+
+
 /*
  * This handler is triggered by the local APIC, so the APIC IRQ handling
  * rules apply:
@@ -750,6 +774,8 @@ again:
 	if (++loops > 100) {
 		WARN_ONCE(1, "perfcounters: irq loop stuck!\n");
 		perf_counter_print_debug();
+		intel_pmu_reset();
+		perf_enable();
 		return 1;
 	}