Skip to contentSkip to conversation

perf stat: Support --all-kernel and --all-user

PATCH In review Jin Yao 4 participants 14 messages Sep 25, 2019 – Oct 12, 2019 perf events

lore
Share & tools
lore

Jin Yao

author
#0 lore raw

cover perf stat: Support --all-kernel and --all-user

This patch series supports the new options "--all-kernel" and "--all-user" in perf-stat.

For example,

root@kbl:~# perf stat -e cycles,instructions --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

        19,156,665      cycles:k
         7,265,342      instructions:k            #    0.38  insn per cycle
     4,511,186,293      cycles:u
       121,881,436      instructions:u            #    0.03  insn per cycle
       1.001153540 seconds time elapsed
 root@kbl:~# perf stat -a --topdown --all-kernel -- sleep 1

Performance counter stats for 'system wide':

retiring:k bad speculation:k frontend bound:k backend bound:k

S0-D0-C0           2                 7.6%                 1.8%                40.5%                50.0%
S0-D0-C1           2                15.4%                 3.4%                14.4%                66.8%
S0-D0-C2           2                15.8%                 5.1%                26.9%                52.2%
S0-D0-C3           2                 5.7%                 5.7%                46.2%                42.4%
       1.000771709 seconds time elapsed

More detail information are in the patch descriptions.

Jin Yao (2):
perf stat: Support --all-kernel and --all-user options perf stat: Support topdown with --all-kernel/--all-user

 tools/perf/Documentation/perf-record.txt |   3 +-
 tools/perf/Documentation/perf-stat.txt   |   7 +
 tools/perf/builtin-stat.c                | 200 ++++++++++++++++++++++-
 tools/perf/util/stat-shadow.c            | 167 ++++++++++++++-----
 tools/perf/util/stat.h                   |  23 +++
 5 files changed, 353 insertions(+), 47 deletions(-)
Signature · 1 line
2.17.1
author
#1 lore raw

PATCH 1/2 perf stat: Support --all-kernel and --all-user options

perf record has supported --all-kernel / --all-user to configure all used events to run in kernel space or in user space. But perf stat doesn't support that. It would be useful to support these options so that we can collect metrics for e.g. user space only without having to type ":u" in the events manually.

Also it would be useful if --all-user / --all-kernel could be specified together, and the tool would automatically add two copies of the events, so that we get a break down between user and kernel.

Since we need to support specifying both --all-user and --all-kernel together, we can't do as what the perf record does for supporting --all-user / --all-kernel (it sets attr->exclude_kernel and attr->exclude_user).

This patch uses another solution which appends the modifiers ":u"/":k" to event string.

For example,
perf stat -e cycles,instructions --all-user --all-kernel

It's automatically expanded to:
perf stat -e cycles:k,instructions:k,cycles:u,instructions:u

More examples,

 root@kbl:~# perf stat -e cycles --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

        20,884,637      cycles:k
     4,511,494,722      cycles:u

       1.000891147 seconds time elapsed
 root@kbl:~# perf stat -e cycles,instructions --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

        19,156,665      cycles:k
         7,265,342      instructions:k            #    0.38  insn per cycle
     4,511,186,293      cycles:u
       121,881,436      instructions:u            #    0.03  insn per cycle
       1.001153540 seconds time elapsed
 root@kbl:~#  perf stat -e "{cycles,instructions}" --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

        16,230,472      cycles:k
         5,357,549      instructions:k            #    0.33  insn per cycle
     4,510,695,030      cycles:u
       122,097,780      instructions:u            #    0.03  insn per cycle
       1.000933419 seconds time elapsed
 root@kbl:~# perf stat -e "{cycles,instructions},{cache-misses,cache-references}" --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

       111,688,302      cycles:k                                                      (74.81%)
        24,322,238      instructions:k            #    0.22  insn per cycle           (74.81%)
         1,115,414      cache-misses:k            #   21.292 % of all cache refs      (75.02%)
         5,238,665      cache-references:k                                            (75.02%)
     4,506,792,681      cycles:u                                                      (75.22%)
       124,199,635      instructions:u            #    0.03  insn per cycle           (75.22%)
        43,846,543      cache-misses:u            #   62.616 % of all cache refs      (74.97%)
        70,024,231      cache-references:u                                            (74.97%)
       1.001186804 seconds time elapsed

Note:

1. This patch can only configure the specified events (-e xxx) to run in user space or in kernel space. For supporting other options, such as --topdown, need follow-up patches.

2. In perf-record, it has already supported the --all-kernel and --all-user, but they can't be combined. We should keep the behavior consistent among all perf subtools. So if this patch can be accepted, will post follow-up patches for supporting other subtools for the same behavior.

---
 tools/perf/Documentation/perf-record.txt |   3 +-
 tools/perf/Documentation/perf-stat.txt   |   7 +
 tools/perf/builtin-stat.c                | 163 ++++++++++++++++++++++-
 tools/perf/util/stat.h                   |  11 ++
 4 files changed, 182 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-record.txt +2 -1
index c6f9f31b6039..739e29905184 100644
@@ -495,7 +495,8 @@ Produce compressed trace using specified level n (default: 1 - fastest compressi
 Configure all used events to run in kernel space.
 
 --all-user::
-Configure all used events to run in user space.
+Configure all used events to run in user space. The --all-kernel
+and --all-user can't be combined.
 
 --kernel-callchains::
 Collect callchains only from kernel space. I.e. this option sets
tools/perf/Documentation/perf-stat.txt +7 -0
index 930c51c01201..3f630e2f4144 100644
@@ -323,6 +323,13 @@ The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
 
 Users who wants to get the actual value can apply --no-metric-only.
 
+--all-kernel::
+Configure all specified events to run in kernel space.
+
+--all-user::
+Configure all specified events to run in user space. The --all-kernel
+and --all-user can be combined.
+
 EXAMPLES
 --------
 
tools/perf/builtin-stat.c +162 -1

198 lines not on this page · raw message

tools/perf/util/stat.h +11 -0
index edbeb2f63e8d..8154e07ced64 100644
@@ -93,6 +93,13 @@ struct runtime_stat {
 typedef int (*aggr_get_id_t)(struct perf_stat_config *config,
 			     struct perf_cpu_map *m, int cpu);
 
+struct perf_stat_event_opt {
+	const struct option	*opt;
+	const char		*event_str;
+};
+
+#define PERF_STAT_EVENT_OPT_MAX        128
+
 struct perf_stat_config {
 	enum aggr_mode		 aggr_mode;
 	bool			 scale;
@@ -106,6 +113,8 @@ struct perf_stat_config {
 	bool			 big_num;
 	bool			 no_merge;
 	bool			 walltime_run_table;
+	bool			 all_kernel;
+	bool			 all_user;
 	FILE			*output;
 	unsigned int		 interval;
 	unsigned int		 timeout;
@@ -126,6 +135,8 @@ struct perf_stat_config {
 	struct perf_cpu_map		*cpus_aggr_map;
 	u64			*walltime_run;
 	struct rblist		 metric_events;
+	struct perf_stat_event_opt	event_opt[PERF_STAT_EVENT_OPT_MAX];
+	int			opt_num;
 };
 
 void update_stats(struct stats *stats, u64 val);
Signature · 1 line
2.17.1
author
#2 lore raw

PATCH 2/2 perf stat: Support topdown with --all-kernel/--all-user

In reply to Jin Yao

When perf stat --topdown is enabled, the internal event list is expanded to:
"{topdown-total-slots,topdown-slots-retired,topdown-recovery-bubbles,topdown-fetch-bubbles,topdown-slots-issued}".

With this patch,

1. When --all-user is enabled, it's expanded to:
"{topdown-total-slots:u,topdown-slots-retired:u,topdown-recovery-bubbles:u,topdown-fetch-bubbles:u,topdown-slots-issued:u}"

2. When --all-kernel is enabled, it's expanded to:
"{topdown-total-slots:k,topdown-slots-retired:k,topdown-recovery-bubbles:k,topdown-fetch-bubbles:k,topdown-slots-issued:k}"

3. Both are enabled, it's expanded to:
"{topdown-total-slots:k,topdown-slots-retired:k,topdown-recovery-bubbles:k,topdown-fetch-bubbles:k,topdown-slots-issued:k},{topdown-total-slots:u,topdown-slots-retired:u,topdown-recovery-bubbles:u,topdown-fetch-bubbles:u,topdown-slots-issued:u}"

This patch creates new topdown stat type (STAT_TOPDOWN_XXX_K / STAT_TOPDOWN_XXX_U), and save the event counting value to type related entry in runtime_stat rblist.

For example,

 root@kbl:~# perf stat -a --topdown --all-kernel -- sleep 1

Performance counter stats for 'system wide':

retiring:k bad speculation:k frontend bound:k backend bound:k

S0-D0-C0           2                 7.6%                 1.8%                40.5%                50.0%
S0-D0-C1           2                15.4%                 3.4%                14.4%                66.8%
S0-D0-C2           2                15.8%                 5.1%                26.9%                52.2%
S0-D0-C3           2                 5.7%                 5.7%                46.2%                42.4%
       1.000771709 seconds time elapsed
 root@kbl:~# perf stat -a --topdown --all-user -- sleep 1

Performance counter stats for 'system wide':

retiring:u bad speculation:u frontend bound:u backend bound:u

S0-D0-C0           2                 0.5%                 0.0%                 0.0%                99.4%
S0-D0-C1           2                 5.7%                 5.8%                77.7%                10.7%
S0-D0-C2           2                15.5%                20.5%                35.8%                28.2%
S0-D0-C3           2                14.1%                 0.5%                 1.5%                83.9%
       1.000773028 seconds time elapsed
---
 tools/perf/builtin-stat.c     |  37 +++++++-
 tools/perf/util/stat-shadow.c | 167 +++++++++++++++++++++++++---------
 tools/perf/util/stat.h        |  12 +++
 3 files changed, 171 insertions(+), 45 deletions(-)
tools/perf/builtin-stat.c +36 -1
index 7f4d22b00d04..b766293b9a15 100644
@@ -1436,7 +1436,8 @@ static int add_default_attributes(void)
 
 	if (topdown_run) {
 		char *str = NULL;
-		bool warn = false;
+		bool warn = false, append_uk = false;
+		struct strbuf new_str;
 
 		if (stat_config.aggr_mode != AGGR_GLOBAL &&
 		    stat_config.aggr_mode != AGGR_CORE) {
@@ -1457,6 +1458,21 @@ static int add_default_attributes(void)
 			return -1;
 		}
 		if (topdown_attrs[0] && str) {
+			int ret;
+
+			if (stat_config.all_kernel || stat_config.all_user) {
+				ret = append_modifier(&new_str, str,
+						stat_config.all_kernel,
+						stat_config.all_user);
+				if (ret)
+					return ret;
+
+				free(str);
+				str = strbuf_detach(&new_str, NULL);
+				strbuf_release(&new_str);
+				append_uk = true;
+			}
+
 			if (warn)
 				arch_topdown_group_warn();
 			err = parse_events(evsel_list, str, &errinfo);
@@ -1468,6 +1484,25 @@ static int add_default_attributes(void)
 				free(str);
 				return -1;
 			}
+
+			if (append_uk) {
+				struct evsel *evsel;
+				char *p;
+
+				evlist__for_each_entry(evsel_list, evsel) {
+					/*
+					 * We appended the modifiers ":u"/":k"
+					 * to evsel->name. Since the events have
+					 * been parsed, remove the appended
+					 * modifiers from event name here.
+					 */
+					if (evsel->name) {
+						p = strchr(evsel->name, ':');
+						if (p)
+							*p = 0;
+					}
+				}
+			}
 		} else {
 			fprintf(stderr, "System does not support topdown\n");
 			return -1;
tools/perf/util/stat-shadow.c +123 -44

256 lines not on this page · raw message

tools/perf/util/stat.h +12 -0
index 8154e07ced64..1bae80ed5543 100644
@@ -60,6 +60,8 @@ enum {
 
 #define NUM_CTX CTX_BIT_MAX
 
+#define STAT_TOPDOWN_NUM	5
+
 enum stat_type {
 	STAT_NONE = 0,
 	STAT_NSECS,
@@ -81,6 +83,16 @@ enum stat_type {
 	STAT_TOPDOWN_SLOTS_RETIRED,
 	STAT_TOPDOWN_FETCH_BUBBLES,
 	STAT_TOPDOWN_RECOVERY_BUBBLES,
+	STAT_TOPDOWN_TOTAL_SLOTS_K,
+	STAT_TOPDOWN_SLOTS_ISSUED_K,
+	STAT_TOPDOWN_SLOTS_RETIRED_K,
+	STAT_TOPDOWN_FETCH_BUBBLES_K,
+	STAT_TOPDOWN_RECOVERY_BUBBLES_K,
+	STAT_TOPDOWN_TOTAL_SLOTS_U,
+	STAT_TOPDOWN_SLOTS_ISSUED_U,
+	STAT_TOPDOWN_SLOTS_RETIRED_U,
+	STAT_TOPDOWN_FETCH_BUBBLES_U,
+	STAT_TOPDOWN_RECOVERY_BUBBLES_U,
 	STAT_SMI_NUM,
 	STAT_APERF,
 	STAT_MAX
Signature · 1 line
2.17.1
#3 lore raw
In reply to Jin Yao
Quoted 5 lines · Jin Yao

This patch series supports the new options "--all-kernel" and "--all-user" in perf-stat.

For example,

Quoted 8 lines
root@kbl:~# perf stat -e cycles,instructions --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

        19,156,665      cycles:k
         7,265,342      instructions:k            #    0.38  insn per cycle
     4,511,186,293      cycles:u
       121,881,436      instructions:u            #    0.03  insn per cycle

hi,
I think we should follow --all-kernel/--all-user behaviour from record command, adding extra events seems like unnecesary complexity to me

jirka

Quoted 32 lines
       1.001153540 seconds time elapsed
 root@kbl:~# perf stat -a --topdown --all-kernel -- sleep 1

Performance counter stats for 'system wide':

retiring:k bad speculation:k frontend bound:k backend bound:k

S0-D0-C0           2                 7.6%                 1.8%                40.5%                50.0%
S0-D0-C1           2                15.4%                 3.4%                14.4%                66.8%
S0-D0-C2           2                15.8%                 5.1%                26.9%                52.2%
S0-D0-C3           2                 5.7%                 5.7%                46.2%                42.4%
       1.000771709 seconds time elapsed

More detail information are in the patch descriptions.

Jin Yao (2):
perf stat: Support --all-kernel and --all-user options perf stat: Support topdown with --all-kernel/--all-user

 tools/perf/Documentation/perf-record.txt |   3 +-
 tools/perf/Documentation/perf-stat.txt   |   7 +
 tools/perf/builtin-stat.c                | 200 ++++++++++++++++++++++-
 tools/perf/util/stat-shadow.c            | 167 ++++++++++++++-----
 tools/perf/util/stat.h                   |  23 +++
 5 files changed, 353 insertions(+), 47 deletions(-)
Signature · 1 line
2.17.1
#4 lore raw
Quoted 10 lines · Jiri Olsa
Quoted 9 lines · Jin Yao

This patch series supports the new options "--all-kernel" and "--all-user" in perf-stat.

For example,

root@kbl:~# perf stat -e cycles,instructions --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

Quoted 8 lines
Quoted 4 lines
        19,156,665      cycles:k
         7,265,342      instructions:k            #    0.38  insn per cycle
     4,511,186,293      cycles:u
       121,881,436      instructions:u            #    0.03  insn per cycle

hi,
I think we should follow --all-kernel/--all-user behaviour from record command, adding extra events seems like unnecesary complexity to me

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

-Andi

#5 lore raw

Em Mon, Sep 30, 2019 at 11:21:36AM -0700, Andi Kleen escreveu:

Quoted 16 lines · Jiri Olsa
Quoted 14 lines · Jin Yao
Quoted 13 lines

This patch series supports the new options "--all-kernel" and "--all-user" in perf-stat.

For example,

root@kbl:~# perf stat -e cycles,instructions --all-kernel --all-user -a -- sleep 1

Performance counter stats for 'system wide':

        19,156,665      cycles:k
         7,265,342      instructions:k            #    0.38  insn per cycle
     4,511,186,293      cycles:u
       121,881,436      instructions:u            #    0.03  insn per cycle
Quoted 5 lines
Quoted 2 lines

I think we should follow --all-kernel/--all-user behaviour from record command, adding extra events seems like unnecesary complexity to me

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

Performance counter stats for 'system wide':

        69,937,445      cycles:u
        23,459,595      instructions:u            #    0.34  insn per cycle
        51,040,704      cycles:k
        11,368,152      instructions:k            #    0.22  insn per cycle
       1.002887417 seconds time elapsed

[root@quaco ~]# perf record -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1 [ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.340 MB perf.data (927 samples) ] [root@quaco ~]# perf evlist
cycles:u
instructions:u
cycles:k
instructions:k
[root@quaco ~]#

To make it shorter we could have:

tools/perf/util/parse-events.l +2 -2
index 7469497cd28e..7df28b0e9682 100644
@@ -313,11 +313,11 @@ aux-output		{ return term(yyscanner, PARSE_EVENTS__TERM_TYPE_AUX_OUTPUT); }
 <<EOF>>			{ BEGIN(INITIAL); }
 }
 
-cpu-cycles|cycles				{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES); }
+cpu-cycles|cycles|cyc				{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES); }
 stalled-cycles-frontend|idle-cycles-frontend	{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_STALLED_CYCLES_FRONTEND); }
 stalled-cycles-backend|idle-cycles-backend	{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_STALLED_CYCLES_BACKEND); }
-instructions					{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS); }
+instructions|insn|ins				{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS); }
 cache-misses					{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES); }
 branch-instructions|branches			{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_BRANCH_INSTRUCTIONS); }
 branch-misses					{ return sym(yyscanner, PERF_TYPE_HARDWARE, PERF_COUNT_HW_BRANCH_MISSES); }
 

And another thing that comes to mind is to make -M metrics be accepted as -e arg, as someone suggested recently (Andi?), and also make it set its events honouring the :k or :u modifiers:

[root@quaco ~]# perf stat -M ipc
^C
Performance counter stats for 'system wide':

        15,030,011      inst_retired.any          #      0.3 IPC
        54,449,797      cpu_clk_unhalted.thread

       1.186531715 seconds time elapsed

[root@quaco ~]# perf stat -M ipc:k,ipc:u
Cannot find metric or group `ipc:k'

Usage: perf stat [<options>] [<command>]

-M, --metrics <metric/metric group list>
monitor specified metrics or metric groups (separated by ,)
[root@quaco ~]#

- Arnaldo

#6 lore raw
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

author
#7 lore raw
Quoted 11 lines · Andi Kleen
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

Thanks
Jin Yao

#8 lore raw
Quoted 21 lines · Yao
Quoted 11 lines · Andi Kleen
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

Quoted 7 lines

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

I still need to think a bit more about this.. did you consider other options like cloning of the perf_evlist/perf_evsel and changing just the exclude* bits? might be event worse actualy ;-)

or maybe if we add modifier we could add extra events/groups within the parser.. like:

"{cycles,instructions}:A,{cache-misses,cache-references}:A,cycles:A"

but that might be still more complicated then what you did

also please add the perf record changes so we have same code and logic for both if we are going to change it

jirka

author
#9 lore raw
Quoted 26 lines · Jiri Olsa
Quoted 25 lines · Yao
Quoted 11 lines · Andi Kleen
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

Quoted 8 lines
Quoted 3 lines

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

I still need to think a bit more about this.. did you consider other options like cloning of the perf_evlist/perf_evsel and changing just the exclude* bits? might be event worse actualy ;-)

That should be another approach, but it might be a bit more complicated than just appending ":u"/":k" modifiers to the event name string.

Quoted 7 lines

or maybe if we add modifier we could add extra events/groups within the parser.. like:

"{cycles,instructions}:A,{cache-misses,cache-references}:A,cycles:A"

but that might be still more complicated then what you did

Yes agree.

Quoted 3 lines

also please add the perf record changes so we have same code and logic for both if we are going to change it

If this new semantics can be accepted, I'd like to add perf record supporting as well. :)

Another difficulty for the new semantics is we need to create user and kernel stat type in runtime_stat rblist (see patch "perf stat: Support topdown with --all-kernel/--all-user"). That has to bring extra complexity.

Thanks
Jin Yao

Quoted 2 lines

jirka

#10 lore raw

Em Thu, Oct 10, 2019 at 04:33:57PM +0800, Jin, Yao escreveu:

Quoted 49 lines · Jiri Olsa
Quoted 26 lines · Yao
Quoted 25 lines · Andi Kleen
Quoted 11 lines
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

Quoted 8 lines
Quoted 3 lines

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

I still need to think a bit more about this.. did you consider other options like cloning of the perf_evlist/perf_evsel and changing just the exclude* bits? might be event worse actualy ;-)

That should be another approach, but it might be a bit more complicated than just appending ":u"/":k" modifiers to the event name string.

Quoted 7 lines

or maybe if we add modifier we could add extra events/groups within the parser.. like:

"{cycles,instructions}:A,{cache-misses,cache-references}:A,cycles:A"

but that might be still more complicated then what you did

Quoted 7 lines

Yes agree.

Quoted 2 lines

also please add the perf record changes so we have same code and logic for both if we are going to change it

If this new semantics can be accepted, I'd like to add perf record supporting as well. :)

Changes in semantics should be avoided, when we add an option already present in some other tool, we should strive to keep the semantics, so that people can reuse their knowledge and just switch tools to go from sampling to counting, say.

So if at all possible, and without having really looked deep in this specific case, I would prefer that new semantics come with a new syntax, would that be possible?

- Arnaldo

Quoted 9 lines

Another difficulty for the new semantics is we need to create user and kernel stat type in runtime_stat rblist (see patch "perf stat: Support topdown with --all-kernel/--all-user"). That has to bring extra complexity.

Thanks
Jin Yao

Quoted 2 lines

jirka

Signature · 1 line
- Arnaldo
author
#11 lore raw
Quoted 55 lines · Arnaldo Carvalho de Melo

Em Thu, Oct 10, 2019 at 04:33:57PM +0800, Jin, Yao escreveu:

Quoted 53 lines · Jiri Olsa
Quoted 26 lines · Yao
Quoted 25 lines · Andi Kleen
Quoted 11 lines
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

Quoted 8 lines
Quoted 3 lines

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

I still need to think a bit more about this.. did you consider other options like cloning of the perf_evlist/perf_evsel and changing just the exclude* bits? might be event worse actualy ;-)

That should be another approach, but it might be a bit more complicated than just appending ":u"/":k" modifiers to the event name string.

Quoted 7 lines

or maybe if we add modifier we could add extra events/groups within the parser.. like:

"{cycles,instructions}:A,{cache-misses,cache-references}:A,cycles:A"

but that might be still more complicated then what you did

Yes agree.

Quoted 2 lines

also please add the perf record changes so we have same code and logic for both if we are going to change it

Quoted 8 lines
Quoted 2 lines

If this new semantics can be accepted, I'd like to add perf record supporting as well. :)

Changes in semantics should be avoided, when we add an option already present in some other tool, we should strive to keep the semantics, so that people can reuse their knowledge and just switch tools to go from sampling to counting, say.

Yes, that makes sense. We need to try our best to keep the original semantics. I will post a patch for perf-stat which just follows the semantics in perf-record.

Quoted 4 lines

So if at all possible, and without having really looked deep in this specific case, I would prefer that new semantics come with a new syntax, would that be possible?

Yes, that's possible. Maybe we can use a new option for automatically adding two copies of the events (one copy for user and the other copy for kernel). The option something like "--all-space"?

Thanks
Jin Yao

Quoted 15 lines

- Arnaldo

Quoted 9 lines

Another difficulty for the new semantics is we need to create user and kernel stat type in runtime_stat rblist (see patch "perf stat: Support topdown with --all-kernel/--all-user"). That has to bring extra complexity.

Thanks
Jin Yao

Quoted 2 lines

jirka

#12 lore raw
Quoted 70 lines · Yao
Quoted 54 lines · Arnaldo Carvalho de Melo

Em Thu, Oct 10, 2019 at 04:33:57PM +0800, Jin, Yao escreveu:

Quoted 53 lines · Jiri Olsa
Quoted 26 lines · Yao
Quoted 25 lines · Andi Kleen
Quoted 11 lines
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

Quoted 8 lines
Quoted 3 lines

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

I still need to think a bit more about this.. did you consider other options like cloning of the perf_evlist/perf_evsel and changing just the exclude* bits? might be event worse actualy ;-)

That should be another approach, but it might be a bit more complicated than just appending ":u"/":k" modifiers to the event name string.

Quoted 7 lines

or maybe if we add modifier we could add extra events/groups within the parser.. like:

"{cycles,instructions}:A,{cache-misses,cache-references}:A,cycles:A"

but that might be still more complicated then what you did

Yes agree.

Quoted 2 lines

also please add the perf record changes so we have same code and logic for both if we are going to change it

Quoted 8 lines
Quoted 2 lines

If this new semantics can be accepted, I'd like to add perf record supporting as well. :)

Changes in semantics should be avoided, when we add an option already present in some other tool, we should strive to keep the semantics, so that people can reuse their knowledge and just switch tools to go from sampling to counting, say.

Yes, that makes sense. We need to try our best to keep the original semantics. I will post a patch for perf-stat which just follows the semantics in perf-record.

Quoted 8 lines
Quoted 4 lines

So if at all possible, and without having really looked deep in this specific case, I would prefer that new semantics come with a new syntax, would that be possible?

Yes, that's possible. Maybe we can use a new option for automatically adding two copies of the events (one copy for user and the other copy for kernel). The option something like "--all-space"?

some other ideas:

	--all
	--uk
	--both
	--full
	-e {cycles,cache-misses}:A,cycles,instructions:A
	-e {cycles,cache-misses}:B,cycles,instructions:B
	--duplicate-every-event-or-group-of-events-for-each-address-space ;-)

jirka

author
#13 lore raw
Quoted 90 lines · Jiri Olsa
Quoted 70 lines · Yao
Quoted 54 lines · Arnaldo Carvalho de Melo

Em Thu, Oct 10, 2019 at 04:33:57PM +0800, Jin, Yao escreveu:

Quoted 53 lines · Jiri Olsa
Quoted 26 lines · Yao
Quoted 25 lines · Andi Kleen
Quoted 11 lines
Quoted 6 lines
Quoted 2 lines

I think it's useful. Makes it easy to do kernel/user break downs. perf record should support the same.

Don't we have this already with:

[root@quaco ~]# perf stat -e cycles:u,instructions:u,cycles:k,instructions:k -a -- sleep 1

This only works for simple cases. Try it for --topdown or multiple -M metrics.

-Andi

Hi Arnaldo, Jiri,

We think it should be very useful if --all-user / --all-kernel can be specified together, so that we can get a break down between user and kernel easily.

But yes, the patches for supporting this new semantics is much complicated than the patch which just follows original perf-record behavior. I fully understand this concern.

Quoted 8 lines
Quoted 3 lines

So if this new semantics can be accepted, that would be very good. But if you think the new semantics is too complicated, I'm also fine for posting a new patch which just follows the perf-record behavior.

I still need to think a bit more about this.. did you consider other options like cloning of the perf_evlist/perf_evsel and changing just the exclude* bits? might be event worse actualy ;-)

That should be another approach, but it might be a bit more complicated than just appending ":u"/":k" modifiers to the event name string.

Quoted 7 lines

or maybe if we add modifier we could add extra events/groups within the parser.. like:

"{cycles,instructions}:A,{cache-misses,cache-references}:A,cycles:A"

but that might be still more complicated then what you did

Yes agree.

Quoted 2 lines

also please add the perf record changes so we have same code and logic for both if we are going to change it

Quoted 8 lines
Quoted 2 lines

If this new semantics can be accepted, I'd like to add perf record supporting as well. :)

Changes in semantics should be avoided, when we add an option already present in some other tool, we should strive to keep the semantics, so that people can reuse their knowledge and just switch tools to go from sampling to counting, say.

Yes, that makes sense. We need to try our best to keep the original semantics. I will post a patch for perf-stat which just follows the semantics in perf-record.

Quoted 8 lines
Quoted 4 lines

So if at all possible, and without having really looked deep in this specific case, I would prefer that new semantics come with a new syntax, would that be possible?

Yes, that's possible. Maybe we can use a new option for automatically adding two copies of the events (one copy for user and the other copy for kernel). The option something like "--all-space"?

some other ideas:

	--all
	--uk
	--both
	--full
	-e {cycles,cache-misses}:A,cycles,instructions:A
	-e {cycles,cache-misses}:B,cycles,instructions:B
	--duplicate-every-event-or-group-of-events-for-each-address-space ;-)
Quoted 2 lines

jirka

I like '--uk'. :)

Thanks
Jin Yao