[PATCH] perf/x86/intel: Fix PMI handling for Intel PT

[Alexander Shishkin]

Since Intel PT is a separate pmu and is not using any of the x86_pmu
code paths, which means in particular that active_events counter remains
intact when new PT events are created. However, PT uses x86_pmu PMI
handler for its PMI handling needs. The problem here is that the latter
checks active_events and in case of it being zero, exits without calling
the actual x86_pmu.handle_nmi(), which results in unknown NMI errors and
massive data loss for PT.

The effect is not visible if there are other perf events in the system
at the same time that keep active_events counter non-zero, for instance
if the NMI watchdog is running, so one needs to disable it to reproduce
the problem.

At the same time, the active_events counter besides doing what the name
suggests also implicitly serves as a pmc hardware and DS area reference
counter.

This patch adds a separate reference counter for the pmc hardware, leaving
active_events for actually counting the events and makes sure it also
counts PT and BTS events.

Signed-off-by: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
---
 arch/x86/kernel/cpu/perf_event.c | 27 +++++++++++++++++++++++----
 1 file changed, 23 insertions(+), 4 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
index 8487714da2..3b0257e222 100644
--- a/arch/x86/kernel/cpu/perf_event.c
+++ b/arch/x86/kernel/cpu/perf_event.c
@@ -135,6 +135,7 @@ static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
 }
 
 static atomic_t active_events;
+static atomic_t pmc_refcount;
 static DEFINE_MUTEX(pmc_reserve_mutex);
 
 #ifdef CONFIG_X86_LOCAL_APIC
@@ -271,6 +272,7 @@ msr_fail:
 static void hw_perf_event_destroy(struct perf_event *event)
 {
 	x86_release_hardware();
+	atomic_dec(&active_events);
 }
 
 void hw_perf_lbr_event_destroy(struct perf_event *event)
@@ -324,16 +326,16 @@ int x86_reserve_hardware(void)
 {
 	int err = 0;
 
-	if (!atomic_inc_not_zero(&active_events)) {
+	if (!atomic_inc_not_zero(&pmc_refcount)) {
 		mutex_lock(&pmc_reserve_mutex);
-		if (atomic_read(&active_events) == 0) {
+		if (atomic_read(&pmc_refcount) == 0) {
 			if (!reserve_pmc_hardware())
 				err = -EBUSY;
 			else
 				reserve_ds_buffers();
 		}
 		if (!err)
-			atomic_inc(&active_events);
+			atomic_inc(&pmc_refcount);
 		mutex_unlock(&pmc_reserve_mutex);
 	}
 
@@ -342,7 +344,7 @@ int x86_reserve_hardware(void)
 
 void x86_release_hardware(void)
 {
-	if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
+	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
 		release_pmc_hardware();
 		release_ds_buffers();
 		mutex_unlock(&pmc_reserve_mutex);
@@ -371,12 +373,24 @@ int x86_add_exclusive(unsigned int what)
 
 out:
 	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;
 }
 
 void x86_del_exclusive(unsigned int what)
 {
 	atomic_dec(&x86_pmu.lbr_exclusive[what]);
+	atomic_dec(&active_events);
 }
 
 int x86_setup_perfctr(struct perf_event *event)
@@ -557,6 +571,7 @@ static int __x86_pmu_event_init(struct perf_event *event)
 	if (err)
 		return err;
 
+	atomic_inc(&active_events);
 	event->destroy = hw_perf_event_destroy;
 
 	event->hw.idx = -1;
@@ -1426,6 +1441,10 @@ perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
 	u64 finish_clock;
 	int ret;
 
+	/*
+	 * All PMUs/events that share this PMI handler should make sure to
+	 * increment active_events for their events.
+	 */
 	if (!atomic_read(&active_events))
 		return NMI_DONE;
 
-- 
2.1.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/