Skip to contentSkip to conversation

perf: Update perf tool to monitor uncore events

RFC PATCH Superseded Lin Ming 4 participants 10 messages Nov 21, 2010 – Nov 23, 2010 perf events

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Lin Ming

author
#0 lore raw

RFC PATCH v2 3/3 perf: Update perf tool to monitor uncore events

Uncore events are monitored with raw events with "ru" prefix("u" for uncore).

$ ./perf stat -C 0 -e ru0101 -- ls

Performance counter stats for 'ls':

            1033791  raw 0x101               

        0.003027583  seconds time elapsed
$ sudo ./perf top -e ru0101

--------------------------------------------------------------------------------------------------------------------
PerfTop: 22 irqs/sec kernel:68.2% exact: 0.0% [1000Hz raw 0x101], (all, 8 CPUs)
--------------------------------------------------------------------------------------------------------------------

             samples  pcnt function               DSO
             _______ _____ ______________________ ____________________________________

                8.00 18.6% kallsyms_expand_symbol [kernel.kallsyms]                   
                5.00 11.6% rb_next                /home/mlin/linux-2.6/tools/perf/perf
                5.00 11.6% clear_page_c           [kernel.kallsyms]
---
 tools/perf/util/parse-events.c |   14 ++++++++++----
 1 files changed, 10 insertions(+), 4 deletions(-)
tools/perf/util/parse-events.c +10 -4
index 898aa94..b03457a 100644
@@ -278,7 +278,7 @@ const char *__event_name(int type, u64 config)
 {
 	static char buf[32];
 
-	if (type == PERF_TYPE_RAW) {
+	if (type == PERF_TYPE_RAW || type == PERF_TYPE_UNCORE) {
 		sprintf(buf, "raw 0x%llx", config);
 		return buf;
 	}
@@ -655,14 +655,20 @@ parse_raw_event(const char **strp, struct perf_event_attr *attr)
 {
 	const char *str = *strp;
 	u64 config;
+	int uncore = 0;
 	int n;
 
 	if (*str != 'r')
 		return EVT_FAILED;
-	n = hex2u64(str + 1, &config);
+	if (*(str+1) == 'u')
+		uncore = 1;
+	n = hex2u64(str + 1 + uncore, &config);
 	if (n > 0) {
-		*strp = str + n + 1;
-		attr->type = PERF_TYPE_RAW;
+		*strp = str + n + 1 + uncore;
+		if (!uncore)
+			attr->type = PERF_TYPE_RAW;
+		else
+			attr->type = PERF_TYPE_UNCORE;
 		attr->config = config;
 		return EVT_HANDLED;
 	}
Signature · 1 line
1.7.2.3
#1 lore raw
Quoted 7 lines

samples pcnt function DSO
_______ _____ ______________________
____________________________________

                8.00 18.6% kallsyms_expand_symbol [kernel.kallsyms]

Reporting a symbol for an uncore event seems highly misleading. After all the uncore counter has no idea for which core the event was, so there isn't really any instruction pointer to report. The event could be event caused by a PCI device or similar.

For per function monitoring of uncore events one has to use OFFCORE_RESPONSE, like I implemented recently.

I would suggest to not report any symbol names for uncore events. Doing so just will confuse users.

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

-Andi

#2 lore raw
Quoted 16 lines · Andi Kleen
Quoted 7 lines

             samples  pcnt function               DSO _______ _____ ______________________
____________________________________

                8.00 18.6% kallsyms_expand_symbol [kernel.kallsyms]

Reporting a symbol for an uncore event seems highly misleading. After all the uncore counter has no idea for which core the event was, so there isn't really any instruction pointer to report. The event could be event caused by a PCI device or similar.

For per function monitoring of uncore events one has to use OFFCORE_RESPONSE, like I implemented recently.

Quoted 7 lines

I would suggest to not report any symbol names for uncore events. Doing so just will confuse users.

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

I agree, uncore should only be used for counting on a per-cpu basis. You can leave the perf tool as is, but that opens up the risk of misinterpretation by many users, or you restrict this in the tool directly which is the better solution in my mind.

author
#9 lore raw
In reply to Andi Kleen
Quoted 7 lines · Andi Kleen
Quoted 6 lines

samples pcnt function DSO
_______ _____ ______________________
____________________________________

                8.00 18.6% kallsyms_expand_symbol [kernel.kallsyms]
Quoted 7 lines

Reporting a symbol for an uncore event seems highly misleading. After all the uncore counter has no idea for which core the event was, so there isn't really any instruction pointer to report. The event could be event caused by a PCI device or similar.

For per function monitoring of uncore events one has to use OFFCORE_RESPONSE, like I implemented recently.

Let me look at this and then get back to you.

Thanks.

Quoted 11 lines

I would suggest to not report any symbol names for uncore events. Doing so just will confuse users.

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

-Andi

author
#3 lore raw
In reply to Stephane Eranian
Quoted 24 lines · Stephane Eranian
Quoted 16 lines · Andi Kleen
Quoted 7 lines

samples pcnt function DSO
_______ _____ ______________________
____________________________________

                8.00 18.6% kallsyms_expand_symbol [kernel.kallsyms]

Reporting a symbol for an uncore event seems highly misleading. After all the uncore counter has no idea for which core the event was, so there isn't really any instruction pointer to report. The event could be event caused by a PCI device or similar.

For per function monitoring of uncore events one has to use OFFCORE_RESPONSE, like I implemented recently.

Quoted 7 lines

I would suggest to not report any symbol names for uncore events. Doing so just will confuse users.

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

Quoted 5 lines

I agree, uncore should only be used for counting on a per-cpu basis. You can leave the perf tool as is, but that opens up the risk of misinterpretation by many users, or you restrict this in the tool directly which is the better solution in my mind.

I agree restricting the tool is a better idea. How about below?

#one cpu stat is allowed
perf stat -e ruNNNN -C <cpu>
perf stat -e ruNNNN -C <cpu> -p <pid> perf stat -e ruNNNN -C <cpu> -- <command>

#per thread is NOT allowed
perf stat -e ruNNNN -p <pid> perf stat -e ruNNNN -- <command>

#all cpus stat is NOT allowed, because it will mess the results perf stat -e ruNNNN -a

#4 lore raw
In reply to Stephane Eranian
Quoted 8 lines
Quoted 6 lines

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

I agree, uncore should only be used for counting on a per-cpu basis. You can leave the perf tool as is, but

I think you mean per socket base, right? per cpu (= cpu core thread) doesn't make sense for the uncore either.

Per socket would need a new command line mode. The new mode should never report any symbols or processes, just events per socket.

-Andi

#6 lore raw
In reply to Stephane Eranian
Quoted 24 lines · Stephane Eranian
Quoted 16 lines · Andi Kleen
Quoted 7 lines

samples pcnt function DSO
_______ _____ ______________________
____________________________________

                8.00 18.6% kallsyms_expand_symbol [kernel.kallsyms]

Reporting a symbol for an uncore event seems highly misleading. After all the uncore counter has no idea for which core the event was, so there isn't really any instruction pointer to report. The event could be event caused by a PCI device or similar.

For per function monitoring of uncore events one has to use OFFCORE_RESPONSE, like I implemented recently.

Quoted 7 lines

I would suggest to not report any symbol names for uncore events. Doing so just will confuse users.

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

Quoted 5 lines

I agree, uncore should only be used for counting on a per-cpu basis. You can leave the perf tool as is, but that opens up the risk of misinterpretation by many users, or you restrict this in the tool directly which is the better solution in my mind.

I would argue against restricting the tool, print a warning perhaps. I mean, give the user all the rope he needs and tell him how to tie the knot is the unix way, right? :-)

#7 lore raw
Quoted 3 lines

I would argue against restricting the tool, print a warning perhaps. I mean, give the user all the rope he needs and tell him how to tie the knot is the unix way, right? :-)

WARNING: the following information is completely useless and wrong

Basically reporting rip here is lying and don't think lying is the "Unix way".

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only.
#8 lore raw
Quoted 8 lines · Andi Kleen
Quoted 3 lines

I would argue against restricting the tool, print a warning perhaps. I mean, give the user all the rope he needs and tell him how to tie the knot is the unix way, right? :-)

WARNING: the following information is completely useless and wrong

Basically reporting rip here is lying and don't think lying is the "Unix way".

Yeah, unless there's only one task running on that socket.. Dude, really, lighten up.

#5 lore raw
In reply to Andi Kleen
Quoted 9 lines · Andi Kleen
Quoted 8 lines
Quoted 6 lines

In fact I suspect uncore events are only really useful with "stat", but not with "top", or if they are used in top then the symbol reporting should be disabled.

I agree, uncore should only be used for counting on a per-cpu basis. You can leave the perf tool as is, but

I meant one cpu per socket: perf stat -a -C 0 -e unc_clk_unhalted. One could envision the tool picking one CPU per socket. It can figure this out from the topology information in sysfs.

If you do perf stat -a -e unc_*, then you'll get into multiplexing very quickly AND it is useless especially because auto aggregation will yield bogus counts.

Quoted 9 lines

I think you mean per socket base, right? per cpu (= cpu core thread) doesn't make sense for the uncore either.

Per socket would need a new command line mode. The new mode should never report any symbols or processes, just events per socket.

-Andi