ARM: 7117/1: perf: fix HW_CACHE_* events on Cortex-A9
Will Deacon [Mon, 3 Oct 2011 17:30:53 +0000 (18:30 +0100)]
Using COHERENT_LINE_{MISS,HIT} for cache misses and references
respectively is completely wrong. Instead, use the L1D events which
are a better and more useful approximation despite ignoring instruction
traffic.

Reported-by: Alasdair Grant <alasdair.grant@arm.com>
Reported-by: Matt Horsnell <matt.horsnell@arm.com>
Reported-by: Michael Williams <michael.williams@arm.com>
Cc: stable@kernel.org
Cc: Jean Pihet <j-pihet@ti.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

arch/arm/kernel/perf_event_v7.c

index 4c85183..6be3e2e 100644 (file)
@@ -321,8 +321,8 @@ static const unsigned armv7_a9_perf_map[PERF_COUNT_HW_MAX] = {
        [PERF_COUNT_HW_CPU_CYCLES]          = ARMV7_PERFCTR_CPU_CYCLES,
        [PERF_COUNT_HW_INSTRUCTIONS]        =
                                        ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE,
-       [PERF_COUNT_HW_CACHE_REFERENCES]    = ARMV7_PERFCTR_COHERENT_LINE_HIT,
-       [PERF_COUNT_HW_CACHE_MISSES]        = ARMV7_PERFCTR_COHERENT_LINE_MISS,
+       [PERF_COUNT_HW_CACHE_REFERENCES]    = ARMV7_PERFCTR_DCACHE_ACCESS,
+       [PERF_COUNT_HW_CACHE_MISSES]        = ARMV7_PERFCTR_DCACHE_REFILL,
        [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
        [PERF_COUNT_HW_BRANCH_MISSES]       = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
        [PERF_COUNT_HW_BUS_CYCLES]          = ARMV7_PERFCTR_CLOCK_CYCLES,