[PATCH] powerpc/perf: Never program book3s PMCs with values >= 0x80000000

[Anton Blanchard <anton@samba.org>]

[-- Attachment #1: Type: text/plain, Size: 1746 bytes --]

We are seeing a lot of PMU warnings on POWER8:

    Can't find PMC that caused IRQ

Looking closer, the active PMC is 0 at this point and we took a PMU
exception on the transition from negative to 0. Some versions of POWER8
have an issue where they edge detect and not level detect PMC overflows.

A number of places program the PMC with (0x80000000 - period_left),
where period_left can be negative. We can either fix all of these or
just ensure that period_left is always >= 1.

This patch takes the second option.

Cc: <stable@vger.kernel.org>
Signed-off-by: Anton Blanchard <anton@samba.org>
---

Attached is a graph that shows how this bug also causes problems in the
initial period calculation for samples. Once fixed we converge on
the correct period very quickly.

Index: b/arch/powerpc/perf/core-book3s.c
===================================================================
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c
@@ -996,7 +996,22 @@ static void power_pmu_read(struct perf_e
 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
 
 	local64_add(delta, &event->count);
-	local64_sub(delta, &event->hw.period_left);
+
+	/*
+	 * A number of places program the PMC with (0x80000000 - period_left).
+	 * We never want period_left to be less than 1 because we will program
+	 * the PMC with a value >= 0x800000000 and an edge detected PMC will
+	 * roll around to 0 before taking an exception. We have seen this
+	 * on POWER8.
+	 *
+	 * To fix this, clamp the minimum value of period_left to 1.
+	 */
+	do {
+		prev = local64_read(&event->hw.period_left);
+		val = prev - delta;
+		if (val < 1)
+			val = 1;
+	} while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
 }
 
 /*

[-- Attachment #2: perf_record_sampling_period.png --]
[-- Type: image/png, Size: 30902 bytes --]