Skip to contentSkip to conversation

perf record: Save build-id of DSO in callchains

RFC/PATCH Namhyung Kim 4 participants 6 messages Mar 23, 2015 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

Currently the build-id only recorded for sampled location, but in order to correctly view/annotate callchains it might need the build-id's of callchains too. I guess this choice was due to a performance impact on the post-processing at perf record time.

Add a new option --buildid-callchain to record this info.

$ perf record -o xxx -g -- perf > /dev/null
$ perf buildid-list -i xxx
6ebcee76c4b04895598b4df86ec445c49fc137a2 /lib/modules/3.18.6-1-ARCH/build/vmlinux
fc0759b71584d2513b04f1b530965a61b5e499e2 /usr/lib/ld-2.21.so
$ perf record -o yyy -g --buildid-callchain -- perf > /dev/null
$ perf buildid-list -i yyy
6ebcee76c4b04895598b4df86ec445c49fc137a2 /lib/modules/3.18.6-1-ARCH/build/vmlinux
4235ae90856f98d2746529648a3339ebfa6ede43 /home/namhyung/project/linux/tools/perf/perf
fc0759b71584d2513b04f1b530965a61b5e499e2 /usr/lib/ld-2.21.so
9ac81172d5ff96f40d984fe7c10073a98f1a6b2e /usr/lib/libc-2.21.so
864583bb881db1d260660a30a51b3f5022cd538b /usr/lib/libnuma.so.1.0.0
---
 tools/perf/builtin-record.c |  2 ++
 tools/perf/util/build-id.c  | 24 +++++++++++++++++++++++-
 2 files changed, 25 insertions(+), 1 deletion(-)
tools/perf/builtin-record.c +2 -0
index 5a2ff510b75b..a43b1d96714b 100644
@@ -841,6 +841,8 @@ struct option __record_options[] = {
 		    "Sample machine registers on interrupt"),
 	OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
 		    "Record running/enabled time of read (:S) events"),
+	OPT_BOOLEAN(0, "buildid-callchain", &symbol_conf.use_callchain,
+		    "Record build-id of DSOs in callchain"),
 	OPT_END()
 };
 
tools/perf/util/build-id.c +23 -1
index f7fb2587df69..401357f3425a 100644
@@ -17,6 +17,7 @@
 #include "tool.h"
 #include "header.h"
 #include "vdso.h"
+#include "callchain.h"
 
 
 static bool no_buildid_cache;
@@ -24,13 +25,15 @@ static bool no_buildid_cache;
 int build_id__mark_dso_hit(struct perf_tool *tool __maybe_unused,
 			   union perf_event *event,
 			   struct perf_sample *sample,
-			   struct perf_evsel *evsel __maybe_unused,
+			   struct perf_evsel *evsel,
 			   struct machine *machine)
 {
 	struct addr_location al;
 	u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
 							sample->tid);
+	int err;
+	struct callchain_cursor_node *node;
 
 	if (thread == NULL) {
 		pr_err("problem processing %d event, skipping it.\n",
@@ -43,6 +46,25 @@ int build_id__mark_dso_hit(struct perf_tool *tool __maybe_unused,
 	if (al.map != NULL)
 		al.map->dso->hit = 1;
 
+	if (!symbol_conf.use_callchain)
+		return 0;
+
+	err = sample__resolve_callchain(sample, NULL, evsel, &al,
+					PERF_MAX_STACK_DEPTH);
+	if (err)
+		return err;
+
+	callchain_cursor_commit(&callchain_cursor);
+	node = callchain_cursor_current(&callchain_cursor);
+
+	while (node) {
+		if (node->map)
+			node->map->dso->hit = 1;
+
+		callchain_cursor_advance(&callchain_cursor);
+		node = callchain_cursor_current(&callchain_cursor);
+	}
+
 	return 0;
 }
 
Signature · 1 line
2.3.3
#1 lore raw
Quoted 12 lines · Namhyung Kim

Currently the build-id only recorded for sampled location, but in order to correctly view/annotate callchains it might need the build-id's of callchains too. I guess this choice was due to a performance impact on the post-processing at perf record time.

Add a new option --buildid-callchain to record this info.

$ perf record -o xxx -g -- perf > /dev/null
$ perf buildid-list -i xxx
6ebcee76c4b04895598b4df86ec445c49fc137a2 /lib/modules/3.18.6-1-ARCH/build/vmlinux
fc0759b71584d2513b04f1b530965a61b5e499e2 /usr/lib/ld-2.21.so
Quoted 7 lines
$ perf record -o yyy -g --buildid-callchain -- perf > /dev/null
$ perf buildid-list -i yyy
6ebcee76c4b04895598b4df86ec445c49fc137a2 /lib/modules/3.18.6-1-ARCH/build/vmlinux
4235ae90856f98d2746529648a3339ebfa6ede43 /home/namhyung/project/linux/tools/perf/perf
fc0759b71584d2513b04f1b530965a61b5e499e2 /usr/lib/ld-2.21.so
9ac81172d5ff96f40d984fe7c10073a98f1a6b2e /usr/lib/libc-2.21.so
864583bb881db1d260660a30a51b3f5022cd538b /usr/lib/libnuma.so.1.0.0

Just curious, could you try to measure the performance impact of this change?

Also, unless the performance (or file size) effect is horrible, I think this flag should be implicitly set by -g.

99.999% of users won't know about this flag, and will see broken annotations!

Thanks,

Ingo

author maintainer
#2 lore raw

Hi Ingo,

Quoted 23 lines · Ingo Molnar
Quoted 12 lines · Namhyung Kim

Currently the build-id only recorded for sampled location, but in order to correctly view/annotate callchains it might need the build-id's of callchains too. I guess this choice was due to a performance impact on the post-processing at perf record time.

Add a new option --buildid-callchain to record this info.

$ perf record -o xxx -g -- perf > /dev/null
$ perf buildid-list -i xxx
6ebcee76c4b04895598b4df86ec445c49fc137a2 /lib/modules/3.18.6-1-ARCH/build/vmlinux
fc0759b71584d2513b04f1b530965a61b5e499e2 /usr/lib/ld-2.21.so
Quoted 7 lines
$ perf record -o yyy -g --buildid-callchain -- perf > /dev/null
$ perf buildid-list -i yyy
6ebcee76c4b04895598b4df86ec445c49fc137a2 /lib/modules/3.18.6-1-ARCH/build/vmlinux
4235ae90856f98d2746529648a3339ebfa6ede43 /home/namhyung/project/linux/tools/perf/perf
fc0759b71584d2513b04f1b530965a61b5e499e2 /usr/lib/ld-2.21.so
9ac81172d5ff96f40d984fe7c10073a98f1a6b2e /usr/lib/libc-2.21.so
864583bb881db1d260660a30a51b3f5022cd538b /usr/lib/libnuma.so.1.0.0
Quoted 2 lines

Just curious, could you try to measure the performance impact of this change?

  $ time perf record --call-graph dwarf -o kbuild.xxx -a -- sleep 30
  [ perf record: Woken up 34674 times to write data ]
  [ perf record: Captured and wrote 9538.729 MB kbuild.xxx (1211997 samples) ]
  real   0m32.244s
  user   0m1.097s
  sys    0m8.997s
  $ time perf record --call-graph dwarf -o kbuild.yyy --buildid-callchain -a -- sleep 30
  [ perf record: Woken up 40902 times to write data ]
  [ perf record: Captured and wrote 11208.500 MB kbuild.yyy (1433922 samples) ]
  real   2m21.695s
  user   1m33.127s
  sys    0m22.077s
Quoted 3 lines

Also, unless the performance (or file size) effect is horrible, I think this flag should be implicitly set by -g.

Agreed. But unfortunately, it seems horrible. ;-p

Quoted 3 lines

99.999% of users won't know about this flag, and will see broken annotations!

Well, most of them will see correct one unless the binary is changed in the meantime. :)

Thanks,
Namhyung

#3 lore raw
Quoted 11 lines · Namhyung Kim
Quoted 2 lines

Just curious, could you try to measure the performance impact of this change?

   $ time perf record --call-graph dwarf -o kbuild.xxx -a -- sleep 30
   [ perf record: Woken up 34674 times to write data ]
   [ perf record: Captured and wrote 9538.729 MB kbuild.xxx (1211997 samples) ]
   real   0m32.244s
   user   0m1.097s
   sys    0m8.997s
Quoted 8 lines
   $ time perf record --call-graph dwarf -o kbuild.yyy --buildid-callchain -a -- sleep 30
   [ perf record: Woken up 40902 times to write data ]
   [ perf record: Captured and wrote 11208.500 MB kbuild.yyy (1433922 samples) ]
   real   2m21.695s
   user   1m33.127s
   sys    0m22.077s

How many CPUs and processes?

David

#4 lore raw

Em Mon, Mar 23, 2015 at 07:44:37AM -0600, David Ahern escreveu:

Quoted 16 lines · Namhyung Kim
Quoted 15 lines
Quoted 2 lines

Just curious, could you try to measure the performance impact of this change?

  $ time perf record --call-graph dwarf -o kbuild.xxx -a -- sleep 30
  [ perf record: Woken up 34674 times to write data ]
  [ perf record: Captured and wrote 9538.729 MB kbuild.xxx (1211997 samples) ]
  real   0m32.244s
  user   0m1.097s
  sys    0m8.997s
  $ time perf record --call-graph dwarf -o kbuild.yyy --buildid-callchain -a -- sleep 30
  [ perf record: Woken up 40902 times to write data ]
  [ perf record: Captured and wrote 11208.500 MB kbuild.yyy (1433922 samples) ]
Quoted 6 lines
Quoted 4 lines
  real   2m21.695s
  user   1m33.127s
  sys    0m22.077s

How many CPUs and processes?

I think I'll add those statistics to the default 'perf record' output when build-id processing is done...

- Arnaldo

author maintainer
#5 lore raw
In reply to David Ahern

Hi David,

Quoted 16 lines · David Ahern
Quoted 15 lines · Namhyung Kim
Quoted 2 lines

Just curious, could you try to measure the performance impact of this change?

  $ time perf record --call-graph dwarf -o kbuild.xxx -a -- sleep 30
  [ perf record: Woken up 34674 times to write data ]
  [ perf record: Captured and wrote 9538.729 MB kbuild.xxx (1211997 samples) ]
  real   0m32.244s
  user   0m1.097s
  sys    0m8.997s
  $ time perf record --call-graph dwarf -o kbuild.yyy --buildid-callchain -a -- sleep 30
  [ perf record: Woken up 40902 times to write data ]
  [ perf record: Captured and wrote 11208.500 MB kbuild.yyy (1433922 samples) ]
Quoted 6 lines
Quoted 4 lines
  real   2m21.695s
  user   1m33.127s
  sys    0m22.077s

How many CPUs and processes?

I was running kbuild with allyesconfig on 12-cpu box.

Thanks,
Namhyung