perf, x86, Do not user perf_disable from NMI context
Peter Zijlstra [Mon, 8 Mar 2010 12:51:01 +0000 (13:51 +0100)]
Explicitly use intel_pmu_{disable,enable}_all() in intel_pmu_handle_irq()
to avoid the NMI race conditions in perf_{disable,enable}

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: paulus@samba.org
Cc: eranian@google.com
Cc: robert.richter@amd.com
Cc: fweisbec@gmail.com
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

arch/x86/kernel/cpu/perf_event_intel.c

index 12e811a..c582449 100644 (file)
@@ -745,11 +745,11 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
 
        cpuc = &__get_cpu_var(cpu_hw_events);
 
-       perf_disable();
+       intel_pmu_disable_all();
        intel_pmu_drain_bts_buffer();
        status = intel_pmu_get_status();
        if (!status) {
-               perf_enable();
+               intel_pmu_enable_all();
                return 0;
        }
 
@@ -759,8 +759,7 @@ again:
                WARN_ONCE(1, "perfevents: irq loop stuck!\n");
                perf_event_print_debug();
                intel_pmu_reset();
-               perf_enable();
-               return 1;
+               goto done;
        }
 
        inc_irq_stat(apic_perf_irqs);
@@ -790,8 +789,8 @@ again:
        if (status)
                goto again;
 
-       perf_enable();
-
+done:
+       intel_pmu_enable_all();
        return 1;
 }