Skip to contentSkip to conversation

perf x86: Add off-core event constraints for Sandy/IvyBridge micro architecture

RFC/PATCH Jiri Olsa 3 participants 4 messages Jan 27, 2013 – Jan 31, 2013

Subsystems (2)
lore
Share & tools
lore

Jiri Olsa

author
#0 lore raw

hi,
I was looking at the offcore stuff and it looks like we might be missing some constraints for offcore response events on Sandy/IvyBridge.

The table 18.8.5 (Off-core Response Performance Monitoring) in Intel SDM states PMC0 for 0xb7 and PMC3 for 0xbb, but there's no other explanation or related description.

I can't say/ack if the counters looks bad or right with or without the patch so far.. so just curious ;-)

thanks,
jirka

---
The Intel SDM (18.8.5 Off-core Response Performance Monitoring) states the off-core events MSR_OFFCORE_RSP_0/MSR_OFFCORE_RSP_03 to be defined only for PMC0/PMC3 respectively.

Adding related constraints.

Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 arch/x86/kernel/cpu/perf_event_intel.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/kernel/cpu/perf_event_intel.c +2 -0
index 5dc54fc..d1f240f 100644
@@ -105,6 +105,8 @@ static struct event_constraint intel_snb_event_constraints[] __read_mostly =
 	INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
 	INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
 	INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
+	INTEL_EVENT_CONSTRAINT(0xb7, 0x1), /* MSR_OFFCORE_RSP_0 - PMC0 only*/
+	INTEL_EVENT_CONSTRAINT(0xbb, 0x8), /* MSR_OFFCORE_RSP_1 - PMC3 only*/
 	EVENT_CONSTRAINT_END
 };
 
Signature · 1 line
1.7.11.7
#1 lore raw
Quoted 5 lines · Jiri Olsa

hi,
I was looking at the offcore stuff and it looks like we might be missing some constraints for offcore response events on Sandy/IvyBridge.

Quoted 7 lines

The table 18.8.5 (Off-core Response Performance Monitoring) in Intel SDM states PMC0 for 0xb7 and PMC3 for 0xbb, but there's no other explanation or related description.

I can't say/ack if the counters looks bad or right with or without the patch so far.. so just curious ;-)

Those are artificial constraints which should not be there. Remember that offcore_rsp uses an extra MSR which has to be shared by all the counters on the PMU. So a way to handle the sharing of that extra MSR is to impose an artificial constraint on the event itself. If it can only run on one counter, then you get the management of the extra MSR for free, i.e., only one event gets it.

In perf_events, we use a more sophisticated dynamic scheme which does not use this artificial constraint. We can measure the event multiple times and share the extra MSR if possible (same value). Why multiple times you might ask? For instance, with different priv levels.

Hope this helps.

Quoted 37 lines

---
The Intel SDM (18.8.5 Off-core Response Performance Monitoring) states the off-core events MSR_OFFCORE_RSP_0/MSR_OFFCORE_RSP_03 to be defined only for PMC0/PMC3 respectively.

Adding related constraints.

Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Stephane Eranian <[email protected]>
---
 arch/x86/kernel/cpu/perf_event_intel.c | 2 ++
 1 file changed, 2 insertions(+)
arch/x86/kernel/cpu/perf_event_intel.c +2 -1
index 5dc54fc..d1f240f 100644
@@ -105,6 +105,8 @@ static struct event_constraint intel_snb_event_constraints[] __read_mostly =
        INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
        INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
        INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
+       INTEL_EVENT_CONSTRAINT(0xb7, 0x1), /* MSR_OFFCORE_RSP_0 - PMC0 only*/
+       INTEL_EVENT_CONSTRAINT(0xbb, 0x8), /* MSR_OFFCORE_RSP_1 - PMC3 only*/
        EVENT_CONSTRAINT_END
 };
 
--

1.7.11.7

author
#2 lore raw
Quoted 21 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa

hi,
I was looking at the offcore stuff and it looks like we might be missing some constraints for offcore response events on Sandy/IvyBridge.

Quoted 7 lines

The table 18.8.5 (Off-core Response Performance Monitoring) in Intel SDM states PMC0 for 0xb7 and PMC3 for 0xbb, but there's no other explanation or related description.

I can't say/ack if the counters looks bad or right with or without the patch so far.. so just curious ;-)

Those are artificial constraints which should not be there. Remember that offcore_rsp uses an extra MSR which has to be shared by all the counters on the PMU. So a way to handle the sharing of that extra MSR is to impose an artificial constraint on the event itself. If it can only run on one counter, then you get the management of the extra MSR for free, i.e., only one event gets it.

Quoted 8 lines

In perf_events, we use a more sophisticated dynamic scheme which does not use this artificial constraint. We can measure the event multiple times and share the extra MSR if possible (same value). Why multiple times you might ask? For instance, with different priv levels.

Hope this helps.

nice, thanks a lot for explanation

jirka

#3 lore raw
Quoted 29 lines · Jiri Olsa
Quoted 28 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa

hi,
I was looking at the offcore stuff and it looks like we might be missing some constraints for offcore response events on Sandy/IvyBridge.

Quoted 7 lines

The table 18.8.5 (Off-core Response Performance Monitoring) in Intel SDM states PMC0 for 0xb7 and PMC3 for 0xbb, but there's no other explanation or related description.

I can't say/ack if the counters looks bad or right with or without the patch so far.. so just curious ;-)

Those are artificial constraints which should not be there. Remember that offcore_rsp uses an extra MSR which has to be shared by all the counters on the PMU. So a way to handle the sharing of that extra MSR is to impose an artificial constraint on the event itself. If it can only run on one counter, then you get the management of the extra MSR for free, i.e., only one event gets it.

In perf_events, we use a more sophisticated dynamic scheme which does not use this artificial constraint. We can measure the event multiple times and share the extra MSR if possible (same value). Why multiple times you might ask? For instance, with different priv levels.

Quoted 3 lines
Quoted 1 line

Hope this helps.

nice, thanks a lot for explanation

Would be nice to stick this into the code somewhere appropriate, AFAICS this information only lives in older commit logs atm.

Thanks,

Ingo