From 93472aff802fd7b61f2209335207e9bd793012f7 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 24 Jun 2015 16:47:50 +0200 Subject: [PATCH] perf/x86: Fix 'active_events' imbalance Commit 1b7b938f1817 ("perf/x86/intel: Fix PMI handling for Intel PT") conditionally increments active_events in x86_add_exclusive() but unconditionally decrements in x86_del_exclusive(). These extra decrements can lead to the situation where active_events is zero and thus the PMI handler is 'disabled' while we have active events on the PMU generating PMIs. This leads to a truckload of: Uhhuh. NMI received for unknown reason 21 on CPU 28. Do you have a strange power saving mode enabled? Dazed and confused, but trying to continue messages and generally messes up perf. Remove the condition on the increment, double increment balanced by a double decrement is perfectly fine. Restructure the code a little bit to make the unconditional inc a bit more natural. Signed-off-by: Peter Zijlstra (Intel) Cc: Andrew Morton Cc: Borislav Petkov Cc: H. Peter Anvin Cc: Linus Torvalds Cc: Peter Zijlstra Cc: Thomas Gleixner Cc: alexander.shishkin@linux.intel.com Cc: brgerst@gmail.com Cc: dvlasenk@redhat.com Cc: luto@amacapital.net Cc: oleg@redhat.com Fixes: 1b7b938f1817 ("perf/x86/intel: Fix PMI handling for Intel PT") Link: http://lkml.kernel.org/r/20150624144750.GJ18673@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event.c | 36 +++++++++++++----------------------- 1 file changed, 13 insertions(+), 23 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c index 5801a14f7524..3658de47900f 100644 --- a/arch/x86/kernel/cpu/perf_event.c +++ b/arch/x86/kernel/cpu/perf_event.c @@ -357,34 +357,24 @@ void x86_release_hardware(void) */ int x86_add_exclusive(unsigned int what) { - int ret = -EBUSY, i; - - if (atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) - return 0; + int i; - mutex_lock(&pmc_reserve_mutex); - for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) { - if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i])) - goto out; + if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) { + mutex_lock(&pmc_reserve_mutex); + for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) { + if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i])) + goto fail_unlock; + } + atomic_inc(&x86_pmu.lbr_exclusive[what]); + mutex_unlock(&pmc_reserve_mutex); } - atomic_inc(&x86_pmu.lbr_exclusive[what]); - ret = 0; + atomic_inc(&active_events); + return 0; -out: +fail_unlock: mutex_unlock(&pmc_reserve_mutex); - - /* - * Assuming that all exclusive events will share the PMI handler - * (which checks active_events for whether there is work to do), - * we can bump active_events counter right here, except for - * x86_lbr_exclusive_lbr events that go through x86_pmu_event_init() - * path, which already bumps active_events for them. - */ - if (!ret && what != x86_lbr_exclusive_lbr) - atomic_inc(&active_events); - - return ret; + return -EBUSY; } void x86_del_exclusive(unsigned int what) -- 2.11.0