perf, x86: Provide a PEBS capable cycle event
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Tue, 14 Dec 2010 20:26:40 +0000 (21:26 +0100)
committerIngo Molnar <mingo@elte.hu>
Thu, 16 Dec 2010 10:36:44 +0000 (11:36 +0100)
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/kernel/cpu/perf_event_intel.c

index c8f5c088cad11ae3f245e1e7374bb43c915170d6..24e390e40f2e0b484d4b2b09084deb9d120d59b4 100644 (file)
@@ -816,6 +816,32 @@ static int intel_pmu_hw_config(struct perf_event *event)
        if (ret)
                return ret;
 
+       if (event->attr.precise_ip &&
+           (event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
+               /*
+                * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
+                * (0x003c) so that we can use it with PEBS.
+                *
+                * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
+                * PEBS capable. However we can use INST_RETIRED.ANY_P
+                * (0x00c0), which is a PEBS capable event, to get the same
+                * count.
+                *
+                * INST_RETIRED.ANY_P counts the number of cycles that retires
+                * CNTMASK instructions. By setting CNTMASK to a value (16)
+                * larger than the maximum number of instructions that can be
+                * retired per cycle (4) and then inverting the condition, we
+                * count all cycles that retire 16 or less instructions, which
+                * is every cycle.
+                *
+                * Thereby we gain a PEBS capable cycle counter.
+                */
+               u64 alt_config = 0x108000c0; /* INST_RETIRED.TOTAL_CYCLES */
+
+               alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
+               event->hw.config = alt_config;
+       }
+
        if (event->attr.type != PERF_TYPE_RAW)
                return 0;