perf_events: fix BTS interrupt handling to avoid being dazed by NMI
Subsystems (2)
PATCH 1/1 perf_events: fix BTS interrupt handling to avoid being dazed by NMI
Fix a bug introduced with commit de725de and the change in the meaning of the return value of intel_pmu_handle_irq(). With the current code, when you are using the BTS, you get 'dazed by NMI' each time the BTS buffer fills up.
BTS does interrupt on the PMU vector, thus NMI. You need to take this into account in the return value of the function.
--
arch/x86/kernel/cpu/perf_event_intel.c +3 -3
@@ -713,18 +713,18 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)struct cpu_hw_events *cpuc;int bit, loops;u64 status; -int handled =0 ; +int handled;perf_sample_data_init (&data,0 ); cpuc = &__get_cpu_var (cpu_hw_events);intel_pmu_disable_all (); -intel_pmu_drain_bts_buffer (); + handled =intel_pmu_drain_bts_buffer (); status =intel_pmu_get_status ();if (!status) {intel_pmu_enable_all (0 ); -return 0 ; +return handled; } loops =0 ;
Sorry, my patch is missing one piece. I will repost.
Quoted 37 lines · Stephane Eranian
Fix a bug introduced with commit de725de and the change in the meaning of the return value of
intel_pmu_handle_irq(). With the current code, when you are using the BTS, you get 'dazed by NMI' each time the BTS buffer fills up.BTS does interrupt on the PMU vector, thus NMI. You need to take this into account in the return value of the function.
Signed-off-by Stephane Eranian <[email protected]>--
arch/x86/kernel/cpu/perf_event_intel.c +3
-3@@ -713,18 +713,18 @@ static int intel_pmu_handle_irq(struct pt_regs *regs) struct cpu_hw_events *cpuc; int bit, loops; u64 status; - int handled = 0; + int handled; perf_sample_data_init(&data, 0); cpuc = &__get_cpu_var(cpu_hw_events); intel_pmu_disable_all(); - intel_pmu_drain_bts_buffer(); + handled = intel_pmu_drain_bts_buffer(); status = intel_pmu_get_status(); if (!status) { intel_pmu_enable_all(0); - return 0; + return handled; } loops = 0;