Skip to contentSkip to conversation

perf tools: Add support to accumulate hist periods (v7)

PATCH In review Namhyung Kim 2 participants 29 messages Jan 23, 2014 – Jan 27, 2014 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

cover perf tools: Add support to accumulate hist periods (v7)

Hello,

This is a new attempt to implement cumulative hist period report. This work begins from Arun's SORT_INCLUSIVE patch [1] but I completely rewrote it from scratch.

This patchset is based on my previous patchset [2] but I think it's almost independent so that it can be applied separately.

Please see the patch 03/24. I refactored functions that add hist entries with struct hist_entry_iter. While I converted all functions carefully, it'd be better anyone can test and confirm that I didn't mess up something - especially for branch stack and mem stuff.

This patchset basically adds period in a sample to every node in the callchain. A hist_entry now has an additional fields to keep the cumulative period if --children option is given on perf report.

I changed the option as a separate --children and added a new "Children" column (and renamed the default "Overhead" column into "Self"). The output will be sorted by children (cumulative) overhead for now. The reason I changed to the --children is that I still think it's much different from other --call-graph options. The --call-graph option will take care of it even with --children option.

I know that the UI should be changed also to be more flexible as Ingo requested, but I'd like to do this first and then move to work on the next. I also added a new config option to enable it by default.

* chagnes in v7:
- add Tested-by tags from Arun
- rebase onto current acme/perf/core

* changes in v6:
- separate struct hist_iter_ops (Jiri)
- check iter->he before calling ->add_entry_cb (Jiri)
- fix locking issue on perf top (Jiri)

* changes in v5:
- support both of --children and --call-graph (Arun)
- refactor hist_entry_iter to share with perf top (Jiri)
- various cleanups and fixes (Jiri)
- add ack's from Jiri

* changes in v4:
- change to --children option (Ingo)
- rebased on new annotation change (Arnaldo)
- support perf top also
- enable --children option by default (Ingo)

* changes in v3:
- change to --cumulate option
- fix a couple of bugs (Jiri, Rodrigo)
- rename some help functions (Arnaldo)
- cache previous hist entries rathen than just symbol and dso
- add some preparatory cleanups
- add report.cumulate config option

Let me show you an example:

  $ cat abc.c
  #define barrier() asm volatile("" ::: "memory")
  void a(void)
  {
  	int i;
  	for (i = 0; i < 1000000; i++)
  		barrier();
  }
  void b(void)
  {
  	a();
  }
  void c(void)
  {
  	b();
  }
  int main(void)
  {
  	c();
  	return 0;
  }

With this simple program I ran perf record and report:

  $ perf record -g -e cycles:u ./abc

Case 1.

  $ perf report --stdio --no-call-graph --no-children

# Overhead Command Shared Object Symbol

  # ........  .......  .................  ..............
  #
      91.50%      abc  abc                [.] a         
       8.18%      abc  ld-2.17.so         [.] strlen    
       0.31%      abc  [kernel.kallsyms]  [k] page_fault
       0.01%      abc  ld-2.17.so         [.] _start    

Case 2. (current default behavior)

  $ perf report --stdio --call-graph --no-children

# Overhead Command Shared Object Symbol

  # ........  .......  .................  ..............
  #
      91.50%      abc  abc                [.] a         
                  |
                  --- a
                      b
                      c
                      main
                      __libc_start_main
       8.18%      abc  ld-2.17.so         [.] strlen    
                  |
                  --- strlen
                      _dl_sysdep_start

       0.31%      abc  [kernel.kallsyms]  [k] page_fault
                  |
                  --- page_fault
                      _start

       0.01%      abc  ld-2.17.so         [.] _start    
                  |
                  --- _start

Case 3.

  $ perf report --no-call-graph --children --stdio
  #     Self  Children  Command      Shared Object                 Symbol
  # ........  ........  .......  .................  .....................
  #
       0.00%    91.50%      abc  libc-2.17.so       [.] __libc_start_main
       0.00%    91.50%      abc  abc                [.] main             
       0.00%    91.50%      abc  abc                [.] c                
       0.00%    91.50%      abc  abc                [.] b                
      91.50%    91.50%      abc  abc                [.] a                
       0.00%     8.18%      abc  ld-2.17.so         [.] _dl_sysdep_start 
       8.18%     8.18%      abc  ld-2.17.so         [.] strlen           
       0.01%     0.33%      abc  ld-2.17.so         [.] _start           
       0.31%     0.31%      abc  [kernel.kallsyms]  [k] page_fault       

As you can see __libc_start_main -> main -> c -> b -> a callchain show up in the output.

Finally, it looks like below with both option enabled:

Case 4. (default behavior?)

  $ perf report --call-graph --children --stdio
  #     Self  Children  Command      Shared Object                 Symbol
  # ........  ........  .......  .................  .....................
  #
       0.00%    91.50%      abc  libc-2.17.so       [.] __libc_start_main
                  |
                  --- __libc_start_main

       0.00%    91.50%      abc  abc                [.] main             
                  |
                  --- main
                      __libc_start_main

       0.00%    91.50%      abc  abc                [.] c                
                  |
                  --- c
                      main
                      __libc_start_main

       0.00%    91.50%      abc  abc                [.] b                
                  |
                  --- b
                      c
                      main
                      __libc_start_main

      91.50%    91.50%      abc  abc                [.] a                
                  |
                  --- a
                      b
                      c
                      main
                      __libc_start_main

...

Currently the perf enables both of --call-graph and --children when it finds callchains in the samples. While this is useful for TUI or GTK, I'm not sure for stdio as it'd consume so much lines.

It does not handle all kind of cases like event groups and annotations yet, but I really want to release it and get reviews.

You can also get this series on 'perf/cumulate-v7' branch in my tree at:

git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Any comments are welcome, thanks.
Namhyung

Cc Arun Sharma <[email protected]>
Cc Frederic Weisbecker <[email protected]>

[1] https://lkml.org/lkml/2012/3/31/6
[2] https://lkml.org/lkml/2014/1/22/514

Namhyung Kim (21):
perf tools: Introduce struct hist_entry_iter
perf hists: Add support for accumulated stat of hist entry perf hists: Check if accumulated when adding a hist entry perf hists: Accumulate hist entry stat based on the callchain perf tools: Update cpumode for each cumulative entry perf report: Cache cumulative callchains
perf callchain: Add callchain_cursor_snapshot() perf tools: Save callchain info for each cumulative entry perf hists: Sort hist entries by accumulated period perf ui/hist: Add support to accumulated hist stat perf ui/browser: Add support to accumulated hist stat perf ui/gtk: Add support to accumulated hist stat perf tools: Apply percent-limit to cumulative percentage perf tools: Add more hpp helper functions
perf report: Add --children option
perf report: Add report.children config option
perf tools: Add callback function to hist_entry_iter perf top: Convert to hist_entry_iter
perf top: Add --children option
perf top: Add top.children config option
perf tools: Enable --children option by default

 tools/perf/Documentation/perf-report.txt |   5 +
 tools/perf/Documentation/perf-top.txt    |   6 +
 tools/perf/builtin-annotate.c            |   3 +-
 tools/perf/builtin-diff.c                |   2 +-
 tools/perf/builtin-report.c              | 184 +++--------
 tools/perf/builtin-top.c                 |  97 +++---
 tools/perf/tests/hists_link.c            |   4 +-
 tools/perf/ui/browsers/hists.c           |  50 +--
 tools/perf/ui/gtk/hists.c                |  20 +-
 tools/perf/ui/hist.c                     |  62 ++++
 tools/perf/ui/stdio/hist.c               |   4 +-
 tools/perf/util/callchain.c              |  42 +++
 tools/perf/util/callchain.h              |  11 +
 tools/perf/util/hist.c                   | 506 ++++++++++++++++++++++++++++++-
 tools/perf/util/hist.h                   |  47 ++-
 tools/perf/util/sort.h                   |  18 +-
 tools/perf/util/symbol.c                 |  11 +-
 tools/perf/util/symbol.h                 |   1 +
 18 files changed, 850 insertions(+), 223 deletions(-)
Signature · 1 line
1.7.11.7
author maintainer
#1 lore raw

PATCH 1/21 perf tools: Introduce struct hist_entry_iter

There're some duplicate code when adding hist entries. They are different in that some have branch info or mem info but generally do same thing. So introduce new struct hist_entry_iter and add callbacks to customize each case in general way.

The new perf_evsel__add_entry() function will look like:

  iter->prepare_entry();
  iter->add_single_entry();

  while (iter->next_entry())
    iter->add_next_entry();

  iter->finish_entry();

This will help further work like the cumulative callchain patchset.

Cc Jiri Olsa <[email protected]>
Cc David Ahern <[email protected]>
Cc Stephane Eranian <[email protected]>
Cc Frederic Weisbecker <[email protected]>
---
 tools/perf/builtin-report.c | 172 +++----------------------
 tools/perf/util/hist.c      | 304 ++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/hist.h      |  34 +++++
 3 files changed, 356 insertions(+), 154 deletions(-)
tools/perf/builtin-report.c +18 -154
index 7ddea46594ae..b6618ecb474a 100644
@@ -85,144 +85,6 @@ static int report__config(const char *var, const char *value, void *cb)
 	return perf_default_config(var, value, cb);
 }
 
-static int report__add_mem_hist_entry(struct perf_tool *tool, struct addr_location *al,
-				      struct perf_sample *sample, struct perf_evsel *evsel,
-				      union perf_event *event)
-{
-	struct report *rep = container_of(tool, struct report, tool);
-	struct symbol *parent = NULL;
-	u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
-	struct hist_entry *he;
-	struct mem_info *mi, *mx;
-	uint64_t cost;
-	int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);
-
-	if (err)
-		return err;
-
-	mi = machine__resolve_mem(al->machine, al->thread, sample, cpumode);
-	if (!mi)
-		return -ENOMEM;
-
-	if (rep->hide_unresolved && !al->sym)
-		return 0;
-
-	cost = sample->weight;
-	if (!cost)
-		cost = 1;
-
-	/*
-	 * must pass period=weight in order to get the correct
-	 * sorting from hists__collapse_resort() which is solely
-	 * based on periods. We want sorting be done on nr_events * weight
-	 * and this is indirectly achieved by passing period=weight here
-	 * and the he_stat__add_period() function.
-	 */
-	he = __hists__add_entry(&evsel->hists, al, parent, NULL, mi,
-				cost, cost, 0);
-	if (!he)
-		return -ENOMEM;
-
-	err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
-	if (err)
-		goto out;
-
-	mx = he->mem_info;
-	err = addr_map_symbol__inc_samples(&mx->daddr, evsel->idx);
-	if (err)
-		goto out;
-
-	evsel->hists.stats.total_period += cost;
-	hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
-	err = hist_entry__append_callchain(he, sample);
-out:
-	return err;
-}
-
-static int report__add_branch_hist_entry(struct perf_tool *tool, struct addr_location *al,
-					 struct perf_sample *sample, struct perf_evsel *evsel)
-{
-	struct report *rep = container_of(tool, struct report, tool);
-	struct symbol *parent = NULL;
-	unsigned i;
-	struct hist_entry *he;
-	struct branch_info *bi, *bx;
-	int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);
-
-	if (err)
-		return err;
-
-	bi = machine__resolve_bstack(al->machine, al->thread,
-				     sample->branch_stack);
-	if (!bi)
-		return -ENOMEM;
-
-	for (i = 0; i < sample->branch_stack->nr; i++) {
-		if (rep->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
-			continue;
-
-		err = -ENOMEM;
-
-		/* overwrite the 'al' to branch-to info */
-		al->map = bi[i].to.map;
-		al->sym = bi[i].to.sym;
-		al->addr = bi[i].to.addr;
-		/*
-		 * The report shows the percentage of total branches captured
-		 * and not events sampled. Thus we use a pseudo period of 1.
-		 */
-		he = __hists__add_entry(&evsel->hists, al, parent, &bi[i], NULL,
-					1, 1, 0);
-		if (he) {
-			bx = he->branch_info;
-			err = addr_map_symbol__inc_samples(&bx->from, evsel->idx);
-			if (err)
-				goto out;
-
-			err = addr_map_symbol__inc_samples(&bx->to, evsel->idx);
-			if (err)
-				goto out;
-
-			evsel->hists.stats.total_period += 1;
-			hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
-		} else
-			goto out;
-	}
-	err = 0;
-out:
-	free(bi);
-	return err;
-}
-
-static int report__add_hist_entry(struct perf_tool *tool, struct perf_evsel *evsel,
-				  struct addr_location *al, struct perf_sample *sample)
-{
-	struct report *rep = container_of(tool, struct report, tool);
-	struct symbol *parent = NULL;
-	struct hist_entry *he;
-	int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);
-
-	if (err)
-		return err;
-
-	he = __hists__add_entry(&evsel->hists, al, parent, NULL, NULL,
-				sample->period, sample->weight,
-				sample->transaction);
-	if (he == NULL)
-		return -ENOMEM;
-
-	err = hist_entry__append_callchain(he, sample);
-	if (err)
-		goto out;
-
-	err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
-	evsel->hists.stats.total_period += sample->period;
-	hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
-out:
-	return err;
-}
-
-
 static int process_sample_event(struct perf_tool *tool,
 				union perf_event *event,
 				struct perf_sample *sample,
@@ -231,6 +93,9 @@ static int process_sample_event(struct perf_tool *tool,
 {
 	struct report *rep = container_of(tool, struct report, tool);
 	struct addr_location al;
+	struct hist_entry_iter iter = {
+		.hide_unresolved = rep->hide_unresolved,
+	};
 	int ret;
 
 	if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
@@ -245,22 +110,21 @@ static int process_sample_event(struct perf_tool *tool,
 	if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
 		return 0;
 
-	if (sort__mode == SORT_MODE__BRANCH) {
-		ret = report__add_branch_hist_entry(tool, &al, sample, evsel);
-		if (ret < 0)
-			pr_debug("problem adding lbr entry, skipping event\n");
-	} else if (rep->mem_mode == 1) {
-		ret = report__add_mem_hist_entry(tool, &al, sample, evsel, event);
-		if (ret < 0)
-			pr_debug("problem adding mem entry, skipping event\n");
-	} else {
-		if (al.map != NULL)
-			al.map->dso->hit = 1;
-
-		ret = report__add_hist_entry(tool, evsel, &al, sample);
-		if (ret < 0)
-			pr_debug("problem incrementing symbol period, skipping event\n");
-	}
+	if (sort__mode == SORT_MODE__BRANCH)
+		iter.ops = &hist_iter_branch;
+	else if (rep->mem_mode == 1)
+		iter.ops = &hist_iter_mem;
+	else
+		iter.ops = &hist_iter_normal;
+
+	if (al.map != NULL)
+		al.map->dso->hit = 1;
+
+	ret = hist_entry_iter__add(&iter, &al, evsel, event, sample,
+				   rep->max_stack);
+	if (ret < 0)
+		pr_debug("problem adding hist entry, skipping event\n");
+
 	return ret;
 }
 
tools/perf/util/hist.c +304 -0

316 lines not on this page · raw message

tools/perf/util/hist.h +34 -0
index 7d1d973d9a39..8182cb51d218 100644
@@ -91,12 +91,46 @@ struct hists {
 	u16			col_len[HISTC_NR_COLS];
 };
 
+struct hist_entry_iter;
+
+struct hist_iter_ops {
+	int (*prepare_entry)(struct hist_entry_iter *, struct addr_location *);
+	int (*add_single_entry)(struct hist_entry_iter *, struct addr_location *);
+	int (*next_entry)(struct hist_entry_iter *, struct addr_location *);
+	int (*add_next_entry)(struct hist_entry_iter *, struct addr_location *);
+	int (*finish_entry)(struct hist_entry_iter *, struct addr_location *);
+};
+
+struct hist_entry_iter {
+	int total;
+	int curr;
+
+	bool hide_unresolved;
+
+	const union perf_event *event;
+	struct perf_evsel *evsel;
+	struct perf_sample *sample;
+	struct hist_entry *he;
+	struct symbol *parent;
+	void *priv;
+
+	const struct hist_iter_ops *ops;
+};
+
+extern const struct hist_iter_ops hist_iter_normal;
+extern const struct hist_iter_ops hist_iter_branch;
+extern const struct hist_iter_ops hist_iter_mem;
+
 struct hist_entry *__hists__add_entry(struct hists *hists,
 				      struct addr_location *al,
 				      struct symbol *parent,
 				      struct branch_info *bi,
 				      struct mem_info *mi, u64 period,
 				      u64 weight, u64 transaction);
+int hist_entry_iter__add(struct hist_entry_iter *iter, struct addr_location *al,
+			 struct perf_evsel *evsel, const union perf_event *event,
+			 struct perf_sample *sample, int max_stack_depth);
+
 int64_t hist_entry__cmp(struct hist_entry *left, struct hist_entry *right);
 int64_t hist_entry__collapse(struct hist_entry *left, struct hist_entry *right);
 int hist_entry__transaction_len(void);
Signature · 1 line
1.7.11.7
#2 lore raw

Re: [PATCH 01/21] perf tools: Introduce struct hist_entry_iter

Quoted 48 lines · Namhyung Kim

There're some duplicate code when adding hist entries. They are different in that some have branch info or mem info but generally do same thing. So introduce new struct hist_entry_iter and add callbacks to customize each case in general way.

The new perf_evsel__add_entry() function will look like:

  iter->prepare_entry();
  iter->add_single_entry();

  while (iter->next_entry())
    iter->add_next_entry();

  iter->finish_entry();

This will help further work like the cumulative callchain patchset.

Cc Jiri Olsa <[email protected]>
Cc David Ahern <[email protected]>
Cc Stephane Eranian <[email protected]>
Cc Frederic Weisbecker <[email protected]>
---
 tools/perf/builtin-report.c | 172 +++----------------------
 tools/perf/util/hist.c      | 304 ++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/hist.h      |  34 +++++
 3 files changed, 356 insertions(+), 154 deletions(-)
tools/perf/builtin-report.c +0 -11
index 7ddea46594ae..b6618ecb474a 100644
@@ -85,144 +85,6 @@ static int report__config(const char *var, const char *value, void *cb)
 	return perf_default_config(var, value, cb);
 }
 
-static int report__add_mem_hist_entry(struct perf_tool *tool, struct addr_location *al,
-				      struct perf_sample *sample, struct perf_evsel *evsel,
-				      union perf_event *event)
-{
-	struct report *rep = container_of(tool, struct report, tool);
-	struct symbol *parent = NULL;
-	u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
-	struct hist_entry *he;
-	struct mem_info *mi, *mx;
-	uint64_t cost;
-	int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);

tools/perf/builtin-report.c static int report__config(const char *var, const char *value, void *cb) · line 108

Quoted 10 lines
tools/perf/builtin-report.c +0 -10
-
-	if (err)
-		return err;
-
-	mi = machine__resolve_mem(al->machine, al->thread, sample, cpumode);
-	if (!mi)
-		return -ENOMEM;
-
-	if (rep->hide_unresolved && !al->sym)
-		return 0;

this check seems to be missing in iter_add_single_mem_entry

jirka

author maintainer
#3 lore raw

Re: [PATCH 01/21] perf tools: Introduce struct hist_entry_iter

Quoted 17 lines · Jiri Olsa
Quoted 16 lines · Namhyung Kim

There're some duplicate code when adding hist entries. They are different in that some have branch info or mem info but generally do same thing. So introduce new struct hist_entry_iter and add callbacks to customize each case in general way.

The new perf_evsel__add_entry() function will look like:

  iter->prepare_entry();
  iter->add_single_entry();

  while (iter->next_entry())
    iter->add_next_entry();

  iter->finish_entry();

This will help further work like the cumulative callchain patchset.

[SNIP]

Quoted 4 lines

tools/perf/builtin-report.c static int report__config(const char *var, const char *value, void *cb) · line 108

Quoted 2 lines
tools/perf/builtin-report.c +0 -2
-	if (rep->hide_unresolved && !al->sym)
-		return 0;

this check seems to be missing in iter_add_single_mem_entry

Right. But afaics it's duplicate as we already check it in process_sample_event().

Thanks,
Namhyung

#4 lore raw

Re: [PATCH 01/21] perf tools: Introduce struct hist_entry_iter

Quoted 19 lines · Namhyung Kim
Quoted 11 lines · Jiri Olsa
Quoted 10 lines · Namhyung Kim

There're some duplicate code when adding hist entries. They are different in that some have branch info or mem info but generally do same thing. So introduce new struct hist_entry_iter and add callbacks to customize each case in general way.

The new perf_evsel__add_entry() function will look like:

  iter->prepare_entry();
  iter->add_single_entry();
Quoted 6 lines
Quoted 6 lines
  while (iter->next_entry())
    iter->add_next_entry();

  iter->finish_entry();

This will help further work like the cumulative callchain patchset.

Quoted 8 lines

[SNIP]

Quoted 4 lines

tools/perf/builtin-report.c static int report__config(const char *var, const char *value, void *cb) · line 108

Quoted 2 lines
tools/perf/builtin-report.c +0 -2
-	if (rep->hide_unresolved && !al->sym)
-		return 0;

this check seems to be missing in iter_add_single_mem_entry

Right. But afaics it's duplicate as we already check it in process_sample_event().

ok.. missed that

jirka

#5 lore raw

Re: [PATCH 01/21] perf tools: Introduce struct hist_entry_iter

In reply to Namhyung Kim
Quoted 5 lines · Namhyung Kim

There're some duplicate code when adding hist entries. They are different in that some have branch info or mem info but generally do same thing. So introduce new struct hist_entry_iter and add callbacks to customize each case in general way.

SNIP

Quoted 10 lines
tools/perf/builtin-report.c +0 -9
-	if (sort__mode == SORT_MODE__BRANCH) {
-		ret = report__add_branch_hist_entry(tool, &al, sample, evsel);
-		if (ret < 0)
-			pr_debug("problem adding lbr entry, skipping event\n");
-	} else if (rep->mem_mode == 1) {
-		ret = report__add_mem_hist_entry(tool, &al, sample, evsel, event);
-		if (ret < 0)
-			pr_debug("problem adding mem entry, skipping event\n");
-	} else {

tools/perf/builtin-report.c static int process_sample_event(struct perf_tool *tool, · line 115

Quoted 10 lines
tools/perf/builtin-report.c +3 -7
-		if (al.map != NULL)
-			al.map->dso->hit = 1;
-
-		ret = report__add_hist_entry(tool, evsel, &al, sample);
-		if (ret < 0)
-			pr_debug("problem incrementing symbol period, skipping event\n");
-	}
+	if (sort__mode == SORT_MODE__BRANCH)
+		iter.ops = &hist_iter_branch;
+	else if (rep->mem_mode == 1)

nit.. mem_mode is bool

jirka

Quoted 1 line

+

author maintainer
#6 lore raw

Re: [PATCH 01/21] perf tools: Introduce struct hist_entry_iter

Quoted 23 lines · Jiri Olsa
Quoted 5 lines · Namhyung Kim

There're some duplicate code when adding hist entries. They are different in that some have branch info or mem info but generally do same thing. So introduce new struct hist_entry_iter and add callbacks to customize each case in general way.

SNIP

Quoted 14 lines
tools/perf/builtin-report.c +0 -13
-	if (sort__mode == SORT_MODE__BRANCH) {
-		ret = report__add_branch_hist_entry(tool, &al, sample, evsel);
-		if (ret < 0)
-			pr_debug("problem adding lbr entry, skipping event\n");
-	} else if (rep->mem_mode == 1) {
-		ret = report__add_mem_hist_entry(tool, &al, sample, evsel, event);
-		if (ret < 0)
-			pr_debug("problem adding mem entry, skipping event\n");
-	} else {
-		if (al.map != NULL)
-			al.map->dso->hit = 1;
-
-		ret = report__add_hist_entry(tool, evsel, &al, sample);
Quoted 8 lines

tools/perf/builtin-report.c static int process_sample_event(struct perf_tool *tool, · line 115

Quoted 6 lines
tools/perf/builtin-report.c +3 -3
-		if (ret < 0)
-			pr_debug("problem incrementing symbol period, skipping event\n");
-	}
+	if (sort__mode == SORT_MODE__BRANCH)
+		iter.ops = &hist_iter_branch;
+	else if (rep->mem_mode == 1)

nit.. mem_mode is bool

Ah, right! As you can see the old code already did it. ;-p

Thanks,
Namhyung

author maintainer
#7 lore raw

PATCH 2/21 perf hists: Add support for accumulated stat of hist entry

In reply to Namhyung Kim

Maintain accumulated stat information in hist_entry->stat_acc if symbol_conf.cumulate_callchain is set. Fields in ->stat_acc have same vaules initially, and will be updated as callchain is processed later.

---
 tools/perf/util/hist.c   | 28 ++++++++++++++++++++++++++--
 tools/perf/util/sort.h   |  1 +
 tools/perf/util/symbol.h |  1 +
 3 files changed, 28 insertions(+), 2 deletions(-)
tools/perf/util/hist.c +26 -2
index d60ea8da342c..45a962f40cea 100644
@@ -231,6 +231,8 @@ static bool hists__decay_entry(struct hists *hists, struct hist_entry *he)
 		return true;
 
 	he_stat__decay(&he->stat);
+	if (symbol_conf.cumulate_callchain)
+		he_stat__decay(he->stat_acc);
 
 	if (!he->filtered)
 		hists->stats.total_period -= prev_period - he->stat.period;
@@ -272,12 +274,26 @@ void hists__decay_entries(struct hists *hists, bool zap_user, bool zap_kernel)
 
 static struct hist_entry *hist_entry__new(struct hist_entry *template)
 {
-	size_t callchain_size = symbol_conf.use_callchain ? sizeof(struct callchain_root) : 0;
-	struct hist_entry *he = zalloc(sizeof(*he) + callchain_size);
+	size_t callchain_size = 0;
+	struct hist_entry *he;
+
+	if (symbol_conf.use_callchain || symbol_conf.cumulate_callchain)
+		callchain_size = sizeof(struct callchain_root);
+
+	he = zalloc(sizeof(*he) + callchain_size);
 
 	if (he != NULL) {
 		*he = *template;
 
+		if (symbol_conf.cumulate_callchain) {
+			he->stat_acc = malloc(sizeof(he->stat));
+			if (he->stat_acc == NULL) {
+				free(he);
+				return NULL;
+			}
+			memcpy(he->stat_acc, &he->stat, sizeof(he->stat));
+		}
+
 		if (he->ms.map)
 			he->ms.map->referenced = true;
 
@@ -289,6 +305,7 @@ static struct hist_entry *hist_entry__new(struct hist_entry *template)
 			 */
 			he->branch_info = malloc(sizeof(*he->branch_info));
 			if (he->branch_info == NULL) {
+				free(he->stat_acc);
 				free(he);
 				return NULL;
 			}
@@ -362,6 +379,8 @@ static struct hist_entry *add_hist_entry(struct hists *hists,
 
 		if (!cmp) {
 			he_stat__add_period(&he->stat, period, weight);
+			if (symbol_conf.cumulate_callchain)
+				he_stat__add_period(he->stat_acc, period, weight);
 
 			/*
 			 * This mem info was allocated from machine__resolve_mem
@@ -398,6 +417,8 @@ static struct hist_entry *add_hist_entry(struct hists *hists,
 	rb_insert_color(&he->rb_node_in, hists->entries_in);
 out:
 	he_stat__add_cpumode_period(&he->stat, al->cpumode, period);
+	if (symbol_conf.cumulate_callchain)
+		he_stat__add_cpumode_period(he->stat_acc, al->cpumode, period);
 	return he;
 }
 
@@ -775,6 +796,7 @@ void hist_entry__free(struct hist_entry *he)
 {
 	zfree(&he->branch_info);
 	zfree(&he->mem_info);
+	zfree(&he->stat_acc);
 	free_srcline(he->srcline);
 	free(he);
 }
@@ -800,6 +822,8 @@ static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused,
 
 		if (!cmp) {
 			he_stat__add_stat(&iter->stat, &he->stat);
+			if (symbol_conf.cumulate_callchain)
+				he_stat__add_stat(iter->stat_acc, he->stat_acc);
 
 			if (symbol_conf.use_callchain) {
 				callchain_cursor_reset(&callchain_cursor);
tools/perf/util/sort.h +1 -0
index 43e5ff42a609..309f2838a1b4 100644
@@ -81,6 +81,7 @@ struct hist_entry {
 		struct list_head head;
 	} pairs;
 	struct he_stat		stat;
+	struct he_stat		*stat_acc;
 	struct map_symbol	ms;
 	struct thread		*thread;
 	struct comm		*comm;
tools/perf/util/symbol.h +1 -0
index da779b23469b..0657e724d6a8 100644
@@ -97,6 +97,7 @@ struct symbol_conf {
 			show_nr_samples,
 			show_total_period,
 			use_callchain,
+			cumulate_callchain,
 			exclude_other,
 			show_cpu_utilization,
 			initialized,
Signature · 1 line
1.7.11.7
author maintainer
#8 lore raw

PATCH 3/21 perf hists: Check if accumulated when adding a hist entry

In reply to Namhyung Kim

To support callchain accumulation, @entry should be recognized if it's accumulated or not when add_hist_entry() called. The period of an accumulated entry should be added to ->stat_acc but not ->stat. Add @sample_self arg for that.

---
 tools/perf/builtin-annotate.c |  3 ++-
 tools/perf/builtin-diff.c     |  2 +-
 tools/perf/builtin-top.c      |  2 +-
 tools/perf/tests/hists_link.c |  4 ++--
 tools/perf/util/hist.c        | 29 ++++++++++++++++++-----------
 tools/perf/util/hist.h        |  3 ++-
 6 files changed, 26 insertions(+), 17 deletions(-)
tools/perf/builtin-annotate.c +2 -1
index 0da603b79b61..70b2d52c3b2e 100644
@@ -65,7 +65,8 @@ static int perf_evsel__add_sample(struct perf_evsel *evsel,
 		return 0;
 	}
 
-	he = __hists__add_entry(&evsel->hists, al, NULL, NULL, NULL, 1, 1, 0);
+	he = __hists__add_entry(&evsel->hists, al, NULL, NULL, NULL, 1, 1, 0,
+				true);
 	if (he == NULL)
 		return -ENOMEM;
 
tools/perf/builtin-diff.c +1 -1
index a77e31246c00..93912add75b5 100644
@@ -308,7 +308,7 @@ static int hists__add_entry(struct hists *hists,
 			    u64 weight, u64 transaction)
 {
 	if (__hists__add_entry(hists, al, NULL, NULL, NULL, period, weight,
-			       transaction) != NULL)
+			       transaction, true) != NULL)
 		return 0;
 	return -ENOMEM;
 }
tools/perf/builtin-top.c +1 -1
index 76cd510d34d0..c574c291383c 100644
@@ -245,7 +245,7 @@ static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 	pthread_mutex_lock(&evsel->hists.lock);
 	he = __hists__add_entry(&evsel->hists, al, NULL, NULL, NULL,
 				sample->period, sample->weight,
-				sample->transaction);
+				sample->transaction, true);
 	pthread_mutex_unlock(&evsel->hists.lock);
 	if (he == NULL)
 		return NULL;
tools/perf/tests/hists_link.c +2 -2
index 2b6519e0e36f..e4e931ec1dbb 100644
@@ -223,7 +223,7 @@ static int add_hist_entries(struct perf_evlist *evlist, struct machine *machine)
 				goto out;
 
 			he = __hists__add_entry(&evsel->hists, &al, NULL,
-						NULL, NULL, 1, 1, 0);
+						NULL, NULL, 1, 1, 0, true);
 			if (he == NULL)
 				goto out;
 
@@ -246,7 +246,7 @@ static int add_hist_entries(struct perf_evlist *evlist, struct machine *machine)
 				goto out;
 
 			he = __hists__add_entry(&evsel->hists, &al, NULL,
-						NULL, NULL, 1, 1, 0);
+						NULL, NULL, 1, 1, 0, true);
 			if (he == NULL)
 				goto out;
 
tools/perf/util/hist.c +18 -11
index 45a962f40cea..2e9dd5d4ca1d 100644
@@ -272,7 +272,8 @@ void hists__decay_entries(struct hists *hists, bool zap_user, bool zap_kernel)
  * histogram, sorted on item, collects periods
  */
 
-static struct hist_entry *hist_entry__new(struct hist_entry *template)
+static struct hist_entry *hist_entry__new(struct hist_entry *template,
+					  bool sample_self)
 {
 	size_t callchain_size = 0;
 	struct hist_entry *he;
@@ -292,6 +293,8 @@ static struct hist_entry *hist_entry__new(struct hist_entry *template)
 				return NULL;
 			}
 			memcpy(he->stat_acc, &he->stat, sizeof(he->stat));
+			if (!sample_self)
+				memset(&he->stat, 0, sizeof(he->stat));
 		}
 
 		if (he->ms.map)
@@ -354,7 +357,8 @@ static u8 symbol__parent_filter(const struct symbol *parent)
 
 static struct hist_entry *add_hist_entry(struct hists *hists,
 					 struct hist_entry *entry,
-					 struct addr_location *al)
+					 struct addr_location *al,
+					 bool sample_self)
 {
 	struct rb_node **p;
 	struct rb_node *parent = NULL;
@@ -378,7 +382,8 @@ static struct hist_entry *add_hist_entry(struct hists *hists,
 		cmp = hist_entry__cmp(he, entry);
 
 		if (!cmp) {
-			he_stat__add_period(&he->stat, period, weight);
+			if (sample_self)
+				he_stat__add_period(&he->stat, period, weight);
 			if (symbol_conf.cumulate_callchain)
 				he_stat__add_period(he->stat_acc, period, weight);
 
@@ -408,7 +413,7 @@ static struct hist_entry *add_hist_entry(struct hists *hists,
 			p = &(*p)->rb_right;
 	}
 
-	he = hist_entry__new(entry);
+	he = hist_entry__new(entry, sample_self);
 	if (!he)
 		return NULL;
 
@@ -416,7 +421,8 @@ static struct hist_entry *add_hist_entry(struct hists *hists,
 	rb_link_node(&he->rb_node_in, parent, p);
 	rb_insert_color(&he->rb_node_in, hists->entries_in);
 out:
-	he_stat__add_cpumode_period(&he->stat, al->cpumode, period);
+	if (sample_self)
+		he_stat__add_cpumode_period(&he->stat, al->cpumode, period);
 	if (symbol_conf.cumulate_callchain)
 		he_stat__add_cpumode_period(he->stat_acc, al->cpumode, period);
 	return he;
@@ -427,7 +433,8 @@ struct hist_entry *__hists__add_entry(struct hists *hists,
 				      struct symbol *sym_parent,
 				      struct branch_info *bi,
 				      struct mem_info *mi,
-				      u64 period, u64 weight, u64 transaction)
+				      u64 period, u64 weight, u64 transaction,
+				      bool sample_self)
 {
 	struct hist_entry entry = {
 		.thread	= al->thread,
@@ -452,7 +459,7 @@ struct hist_entry *__hists__add_entry(struct hists *hists,
 		.transaction = transaction,
 	};
 
-	return add_hist_entry(hists, &entry, al);
+	return add_hist_entry(hists, &entry, al, sample_self);
 }
 
 static int
@@ -509,7 +516,7 @@ iter_add_single_mem_entry(struct hist_entry_iter *iter, struct addr_location *al
 	 * and the he_stat__add_period() function.
 	 */
 	he = __hists__add_entry(&iter->evsel->hists, al, iter->parent, NULL, mi,
-				cost, cost, 0);
+				cost, cost, 0, true);
 	if (!he)
 		return -ENOMEM;
 
@@ -621,7 +628,7 @@ iter_add_next_branch_entry(struct hist_entry_iter *iter, struct addr_location *a
 	 * and not events sampled. Thus we use a pseudo period of 1.
 	 */
 	he = __hists__add_entry(&evsel->hists, al, iter->parent, &bi[i], NULL,
-				1, 1, 0);
+				1, 1, 0, true);
 	if (he == NULL)
 		return -ENOMEM;
 
@@ -667,7 +674,7 @@ iter_add_single_normal_entry(struct hist_entry_iter *iter, struct addr_location
 
 	he = __hists__add_entry(&evsel->hists, al, iter->parent, NULL, NULL,
 				sample->period, sample->weight,
-				sample->transaction);
+				sample->transaction, true);
 	if (he == NULL)
 		return -ENOMEM;
 
@@ -1177,7 +1184,7 @@ static struct hist_entry *hists__add_dummy_entry(struct hists *hists,
 			p = &(*p)->rb_right;
 	}
 
-	he = hist_entry__new(pair);
+	he = hist_entry__new(pair, true);
 	if (he) {
 		memset(&he->stat, 0, sizeof(he->stat));
 		he->hists = hists;
tools/perf/util/hist.h +2 -1
index 8182cb51d218..d482e673ecf5 100644
@@ -126,7 +126,8 @@ struct hist_entry *__hists__add_entry(struct hists *hists,
 				      struct symbol *parent,
 				      struct branch_info *bi,
 				      struct mem_info *mi, u64 period,
-				      u64 weight, u64 transaction);
+				      u64 weight, u64 transaction,
+				      bool sample_self);
 int hist_entry_iter__add(struct hist_entry_iter *iter, struct addr_location *al,
 			 struct perf_evsel *evsel, const union perf_event *event,
 			 struct perf_sample *sample, int max_stack_depth);
Signature · 1 line
1.7.11.7
author maintainer
#9 lore raw

PATCH 4/21 perf hists: Accumulate hist entry stat based on the callchain

In reply to Namhyung Kim

Call __hists__add_entry() for each callchain node to get an accumulated stat for an entry. Introduce new cumulative_iter ops to process them properly.

---
 tools/perf/builtin-report.c |  2 ++
 tools/perf/util/hist.c      | 87 +++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/hist.h      |  1 +
 3 files changed, 90 insertions(+)
tools/perf/builtin-report.c +2 -0
index b6618ecb474a..3ed0669d7620 100644
@@ -114,6 +114,8 @@ static int process_sample_event(struct perf_tool *tool,
 		iter.ops = &hist_iter_branch;
 	else if (rep->mem_mode == 1)
 		iter.ops = &hist_iter_mem;
+	else if (symbol_conf.cumulate_callchain)
+		iter.ops = &hist_iter_cumulative;
 	else
 		iter.ops = &hist_iter_normal;
 
tools/perf/util/hist.c +87 -0

99 lines not on this page · raw message

tools/perf/util/hist.h +1 -0
index d482e673ecf5..091bf81df8c3 100644
@@ -120,6 +120,7 @@ struct hist_entry_iter {
 extern const struct hist_iter_ops hist_iter_normal;
 extern const struct hist_iter_ops hist_iter_branch;
 extern const struct hist_iter_ops hist_iter_mem;
+extern const struct hist_iter_ops hist_iter_cumulative;
 
 struct hist_entry *__hists__add_entry(struct hists *hists,
 				      struct addr_location *al,
Signature · 1 line
1.7.11.7
author maintainer
#10 lore raw

PATCH 5/21 perf tools: Update cpumode for each cumulative entry

In reply to Namhyung Kim

The cpumode and level in struct addr_localtion was set for a sample and but updated as cumulative callchains were added. This led to have non-matching symbol and cpumode in the output.

Update it accordingly based on the fact whether the map is a part of the kernel or not. This is a reverse of what thread__find_addr_map() does.

---
 tools/perf/util/callchain.c | 42 ++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/callchain.h |  2 ++
 tools/perf/util/hist.c      | 13 ++-----------
 3 files changed, 46 insertions(+), 11 deletions(-)
tools/perf/util/callchain.c +42 -0
index 8d9db454f1a9..ac658135079f 100644
@@ -551,3 +551,45 @@ int hist_entry__append_callchain(struct hist_entry *he, struct perf_sample *samp
 		return 0;
 	return callchain_append(he->callchain, &callchain_cursor, sample->period);
 }
+
+int fill_callchain_info(struct addr_location *al, struct callchain_cursor_node *node,
+			bool hide_unresolved)
+{
+	al->map = node->map;
+	al->sym = node->sym;
+	if (node->map)
+		al->addr = node->map->map_ip(node->map, node->ip);
+	else
+		al->addr = node->ip;
+
+	if (al->sym == NULL) {
+		if (hide_unresolved)
+			return 0;
+		if (al->map == NULL)
+			goto out;
+	}
+
+	if (al->map->groups == &al->machine->kmaps) {
+		if (machine__is_host(al->machine)) {
+			al->cpumode = PERF_RECORD_MISC_KERNEL;
+			al->level = 'k';
+		} else {
+			al->cpumode = PERF_RECORD_MISC_GUEST_KERNEL;
+			al->level = 'g';
+		}
+	} else {
+		if (machine__is_host(al->machine)) {
+			al->cpumode = PERF_RECORD_MISC_USER;
+			al->level = '.';
+		} else if (perf_guest) {
+			al->cpumode = PERF_RECORD_MISC_GUEST_USER;
+			al->level = 'u';
+		} else {
+			al->cpumode = PERF_RECORD_MISC_HYPERVISOR;
+			al->level = 'H';
+		}
+	}
+
+out:
+	return 1;
+}
tools/perf/util/callchain.h +2 -0
index 8ad97e9b119f..66faae21370d 100644
@@ -155,6 +155,8 @@ int sample__resolve_callchain(struct perf_sample *sample, struct symbol **parent
 			      struct perf_evsel *evsel, struct addr_location *al,
 			      int max_stack);
 int hist_entry__append_callchain(struct hist_entry *he, struct perf_sample *sample);
+int fill_callchain_info(struct addr_location *al, struct callchain_cursor_node *node,
+			bool hide_unresolved);
 
 extern const char record_callchain_help[];
 #endif	/* __PERF_CALLCHAIN_H */
tools/perf/util/hist.c +2 -11
index 46402fbf4c0e..beb9f96e4e4f 100644
@@ -739,18 +739,9 @@ iter_next_cumulative_entry(struct hist_entry_iter *iter,
 	if (node == NULL)
 		return 0;
 
-	al->map = node->map;
-	al->sym = node->sym;
-	if (node->map)
-		al->addr = node->map->map_ip(node->map, node->ip);
-	else
-		al->addr = node->ip;
-
-	if (iter->hide_unresolved && al->sym == NULL)
-		return 0;
-
 	callchain_cursor_advance(&callchain_cursor);
-	return 1;
+
+	return fill_callchain_info(al, node, iter->hide_unresolved);
 }
 
 static int
Signature · 1 line
1.7.11.7
author maintainer
#11 lore raw

PATCH 6/21 perf report: Cache cumulative callchains

In reply to Namhyung Kim

It is possble that a callchain has cycles or recursive calls. In that case it'll end up having entries more than 100% overhead in the output. In order to prevent such entries, cache each callchain node and skip if same entry already cumulated.

---
 tools/perf/util/hist.c | 43 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 43 insertions(+)
tools/perf/util/hist.c +43 -0
index beb9f96e4e4f..1b477f8d2d17 100644
@@ -708,7 +708,22 @@ static int
 iter_prepare_cumulative_entry(struct hist_entry_iter *iter __maybe_unused,
 			      struct addr_location *al __maybe_unused)
 {
+	struct hist_entry **he_cache;
+
 	callchain_cursor_commit(&callchain_cursor);
+
+	/*
+	 * This is for detecting cycles or recursions so that they're
+	 * cumulated only one time to prevent entries more than 100%
+	 * overhead.
+	 */
+	he_cache = malloc(sizeof(*he_cache) * (PERF_MAX_STACK_DEPTH + 1));
+	if (he_cache == NULL)
+		return -ENOMEM;
+
+	iter->priv = he_cache;
+	iter->curr = 0;
+
 	return 0;
 }
 
@@ -718,6 +733,7 @@ iter_add_single_cumulative_entry(struct hist_entry_iter *iter,
 {
 	struct perf_evsel *evsel = iter->evsel;
 	struct perf_sample *sample = iter->sample;
+	struct hist_entry **he_cache = iter->priv;
 	struct hist_entry *he;
 
 	he = __hists__add_entry(&evsel->hists, al, iter->parent, NULL, NULL,
@@ -726,6 +742,8 @@ iter_add_single_cumulative_entry(struct hist_entry_iter *iter,
 	if (he == NULL)
 		return -ENOMEM;
 
+	he_cache[iter->curr++] = he;
+
 	return hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 }
 
@@ -750,7 +768,29 @@ iter_add_next_cumulative_entry(struct hist_entry_iter *iter,
 {
 	struct perf_evsel *evsel = iter->evsel;
 	struct perf_sample *sample = iter->sample;
+	struct hist_entry **he_cache = iter->priv;
 	struct hist_entry *he;
+	struct hist_entry he_tmp = {
+		.cpu = al->cpu,
+		.thread = al->thread,
+		.comm = thread__comm(al->thread),
+		.ip = al->addr,
+		.ms = {
+			.map = al->map,
+			.sym = al->sym,
+		},
+		.parent = iter->parent,
+	};
+	int i;
+
+	/*
+	 * Check if there's duplicate entries in the callchain.
+	 * It's possible that it has cycles or recursive calls.
+	 */
+	for (i = 0; i < iter->curr; i++) {
+		if (hist_entry__cmp(he_cache[i], &he_tmp) == 0)
+			return 0;
+	}
 
 	he = __hists__add_entry(&evsel->hists, al, iter->parent, NULL, NULL,
 				sample->period, sample->weight,
@@ -758,6 +798,8 @@ iter_add_next_cumulative_entry(struct hist_entry_iter *iter,
 	if (he == NULL)
 		return -ENOMEM;
 
+	he_cache[iter->curr++] = he;
+
 	return hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 }
 
@@ -771,6 +813,7 @@ iter_finish_cumulative_entry(struct hist_entry_iter *iter,
 	evsel->hists.stats.total_period += sample->period;
 	hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 
+	zfree(&iter->priv);
 	return 0;
 }
 
Signature · 1 line
1.7.11.7
author maintainer
#12 lore raw

PATCH 7/21 perf callchain: Add callchain_cursor_snapshot()

In reply to Namhyung Kim

The callchain_cursor_snapshot() is for saving current status of the callchain. It'll be used to accumulate callchain information for each node.

---
 tools/perf/util/callchain.h | 9 +++++++++
 1 file changed, 9 insertions(+)
tools/perf/util/callchain.h +9 -0
index 66faae21370d..bbd63dfbe112 100644
@@ -159,4 +159,13 @@ int fill_callchain_info(struct addr_location *al, struct callchain_cursor_node *
 			bool hide_unresolved);
 
 extern const char record_callchain_help[];
+
+static inline void callchain_cursor_snapshot(struct callchain_cursor *dest,
+					     struct callchain_cursor *src)
+{
+	*dest = *src;
+
+	dest->first = src->curr;
+	dest->nr -= src->pos;
+}
 #endif	/* __PERF_CALLCHAIN_H */
Signature · 1 line
1.7.11.7
author maintainer
#13 lore raw

PATCH 8/21 perf tools: Save callchain info for each cumulative entry

In reply to Namhyung Kim

When accumulating callchain entry, also save current snapshot of the chain so that it can show the rest of the chain.

---
 tools/perf/util/hist.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)
tools/perf/util/hist.c +15 -2
index 1b477f8d2d17..765ac5953201 100644
@@ -744,6 +744,14 @@ iter_add_single_cumulative_entry(struct hist_entry_iter *iter,
 
 	he_cache[iter->curr++] = he;
 
+	callchain_append(he->callchain, &callchain_cursor, sample->period);
+
+	/*
+	 * We need to re-initialize the cursor since callchain_append()
+	 * advanced the cursor to the end.
+	 */
+	callchain_cursor_commit(&callchain_cursor);
+
 	return hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 }
 
@@ -757,8 +765,6 @@ iter_next_cumulative_entry(struct hist_entry_iter *iter,
 	if (node == NULL)
 		return 0;
 
-	callchain_cursor_advance(&callchain_cursor);
-
 	return fill_callchain_info(al, node, iter->hide_unresolved);
 }
 
@@ -782,6 +788,11 @@ iter_add_next_cumulative_entry(struct hist_entry_iter *iter,
 		.parent = iter->parent,
 	};
 	int i;
+	struct callchain_cursor cursor;
+
+	callchain_cursor_snapshot(&cursor, &callchain_cursor);
+
+	callchain_cursor_advance(&callchain_cursor);
 
 	/*
 	 * Check if there's duplicate entries in the callchain.
@@ -800,6 +811,8 @@ iter_add_next_cumulative_entry(struct hist_entry_iter *iter,
 
 	he_cache[iter->curr++] = he;
 
+	callchain_append(he->callchain, &cursor, sample->period);
+
 	return hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 }
 
Signature · 1 line
1.7.11.7
author maintainer
#14 lore raw

PATCH 9/21 perf hists: Sort hist entries by accumulated period

In reply to Namhyung Kim

When callchain accumulation is requested, we need to sort the entries by accumulated period value. When accumulated periods of two entries are same (i.e. single path callchain) put the caller above since accumulation tends to put callers on higher position for obvious reason.

---
 tools/perf/util/hist.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
tools/perf/util/hist.c +12 -0
index 765ac5953201..da4fbec3d0d8 100644
@@ -1063,6 +1063,18 @@ static int hist_entry__sort_on_period(struct hist_entry *a,
 	struct hist_entry *pair;
 	u64 *periods_a, *periods_b;
 
+	if (symbol_conf.cumulate_callchain) {
+		/*
+		 * Put caller above callee when they have equal period.
+		 */
+		if (a->stat_acc->period != b->stat_acc->period)
+			return a->stat_acc->period > b->stat_acc->period ? 1 : -1;
+
+		if (a->callchain->max_depth != b->callchain->max_depth)
+			return a->callchain->max_depth < b->callchain->max_depth ?
+				1 : -1;
+	}
+
 	ret = period_cmp(a->stat.period, b->stat.period);
 	if (ret || !symbol_conf.event_group)
 		return ret;
Signature · 1 line
1.7.11.7
author maintainer
#15 lore raw

PATCH 10/21 perf ui/hist: Add support to accumulated hist stat

In reply to Namhyung Kim

Print accumulated stat of a hist entry if requested.

---
 tools/perf/ui/hist.c   | 45 +++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/hist.h |  1 +
 2 files changed, 46 insertions(+)
tools/perf/ui/hist.c +45 -0
index 78f4c92e9b73..eb9c07bcdb01 100644
@@ -129,6 +129,28 @@ static int hpp__entry_##_type(struct perf_hpp_fmt *_fmt __maybe_unused,		\
 			  scnprintf, true);					\
 }
 
+#define __HPP_COLOR_ACC_PERCENT_FN(_type, _field)				\
+static u64 he_get_acc_##_field(struct hist_entry *he)				\
+{										\
+	return he->stat_acc->_field;						\
+}										\
+										\
+static int hpp__color_acc_##_type(struct perf_hpp_fmt *fmt __maybe_unused,	\
+			      struct perf_hpp *hpp, struct hist_entry *he) 	\
+{										\
+	return __hpp__fmt(hpp, he, he_get_acc_##_field, " %6.2f%%",		\
+			  (hpp_snprint_fn)percent_color_snprintf, true);	\
+}
+
+#define __HPP_ENTRY_ACC_PERCENT_FN(_type, _field)				\
+static int hpp__entry_acc_##_type(struct perf_hpp_fmt *_fmt __maybe_unused,	\
+			      struct perf_hpp *hpp, struct hist_entry *he) 	\
+{										\
+	const char *fmt = symbol_conf.field_sep ? " %.2f" : " %6.2f%%";		\
+	return __hpp__fmt(hpp, he, he_get_acc_##_field, fmt,			\
+			  scnprintf, true);					\
+}
+
 #define __HPP_ENTRY_RAW_FN(_type, _field)					\
 static u64 he_get_raw_##_field(struct hist_entry *he)				\
 {										\
@@ -148,17 +170,25 @@ __HPP_WIDTH_FN(_type, _min_width, _unit_width)				\
 __HPP_COLOR_PERCENT_FN(_type, _field)					\
 __HPP_ENTRY_PERCENT_FN(_type, _field)
 
+#define HPP_PERCENT_ACC_FNS(_type, _str, _field, _min_width, _unit_width)\
+__HPP_HEADER_FN(_type, _str, _min_width, _unit_width)			\
+__HPP_WIDTH_FN(_type, _min_width, _unit_width)				\
+__HPP_COLOR_ACC_PERCENT_FN(_type, _field)				\
+__HPP_ENTRY_ACC_PERCENT_FN(_type, _field)
+
 #define HPP_RAW_FNS(_type, _str, _field, _min_width, _unit_width)	\
 __HPP_HEADER_FN(_type, _str, _min_width, _unit_width)			\
 __HPP_WIDTH_FN(_type, _min_width, _unit_width)				\
 __HPP_ENTRY_RAW_FN(_type, _field)
 
+__HPP_HEADER_FN(overhead_self, "Self", 8, 8)
 
 HPP_PERCENT_FNS(overhead, "Overhead", period, 8, 8)
 HPP_PERCENT_FNS(overhead_sys, "sys", period_sys, 8, 8)
 HPP_PERCENT_FNS(overhead_us, "usr", period_us, 8, 8)
 HPP_PERCENT_FNS(overhead_guest_sys, "guest sys", period_guest_sys, 9, 8)
 HPP_PERCENT_FNS(overhead_guest_us, "guest usr", period_guest_us, 9, 8)
+HPP_PERCENT_ACC_FNS(overhead_acc, "Children", period, 8, 8)
 
 HPP_RAW_FNS(samples, "Samples", nr_events, 12, 12)
 HPP_RAW_FNS(period, "Period", period, 12, 12)
@@ -171,6 +201,14 @@ HPP_RAW_FNS(period, "Period", period, 12, 12)
 		.entry	= hpp__entry_ ## _name		\
 	}
 
+#define HPP__COLOR_ACC_PRINT_FNS(_name)			\
+	{						\
+		.header	= hpp__header_ ## _name,	\
+		.width	= hpp__width_ ## _name,		\
+		.color	= hpp__color_acc_ ## _name,	\
+		.entry	= hpp__entry_acc_ ## _name	\
+	}
+
 #define HPP__PRINT_FNS(_name)				\
 	{						\
 		.header	= hpp__header_ ## _name,	\
@@ -184,6 +222,7 @@ struct perf_hpp_fmt perf_hpp__format[] = {
 	HPP__COLOR_PRINT_FNS(overhead_us),
 	HPP__COLOR_PRINT_FNS(overhead_guest_sys),
 	HPP__COLOR_PRINT_FNS(overhead_guest_us),
+	HPP__COLOR_ACC_PRINT_FNS(overhead_acc),
 	HPP__PRINT_FNS(samples),
 	HPP__PRINT_FNS(period)
 };
@@ -208,6 +247,12 @@ void perf_hpp__init(void)
 {
 	perf_hpp__column_enable(PERF_HPP__OVERHEAD);
 
+	if (symbol_conf.cumulate_callchain) {
+		perf_hpp__format[PERF_HPP__OVERHEAD].header =
+						hpp__header_overhead_self;
+		perf_hpp__column_enable(PERF_HPP__OVERHEAD_ACC);
+	}
+
 	if (symbol_conf.show_cpu_utilization) {
 		perf_hpp__column_enable(PERF_HPP__OVERHEAD_SYS);
 		perf_hpp__column_enable(PERF_HPP__OVERHEAD_US);
tools/perf/util/hist.h +1 -0
index 091bf81df8c3..59d8ba4665a4 100644
@@ -199,6 +199,7 @@ enum {
 	PERF_HPP__OVERHEAD_US,
 	PERF_HPP__OVERHEAD_GUEST_SYS,
 	PERF_HPP__OVERHEAD_GUEST_US,
+	PERF_HPP__OVERHEAD_ACC,
 	PERF_HPP__SAMPLES,
 	PERF_HPP__PERIOD,
 
Signature · 1 line
1.7.11.7
author maintainer
#16 lore raw

PATCH 11/21 perf ui/browser: Add support to accumulated hist stat

In reply to Namhyung Kim

Print accumulated stat of a hist entry if requested.

---
 tools/perf/ui/browsers/hists.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
tools/perf/ui/browsers/hists.c +17 -0
index b720b92eba6e..c8e3a2e7b5c0 100644
@@ -693,11 +693,26 @@ hist_browser__hpp_color_##_type(struct perf_hpp_fmt *fmt __maybe_unused,\
 	return __hpp__color_fmt(hpp, he, __hpp_get_##_field, _cb);	\
 }
 
+#define __HPP_COLOR_ACC_PERCENT_FN(_type, _field, _cb)			\
+static u64 __hpp_get_acc_##_field(struct hist_entry *he)		\
+{									\
+	return he->stat_acc->_field;					\
+}									\
+									\
+static int								\
+hist_browser__hpp_color_##_type(struct perf_hpp_fmt *fmt __maybe_unused,\
+				struct perf_hpp *hpp,			\
+				struct hist_entry *he)			\
+{									\
+	return __hpp__color_fmt(hpp, he, __hpp_get_acc_##_field, _cb);	\
+}
+
 __HPP_COLOR_PERCENT_FN(overhead, period, __hpp__color_callchain)
 __HPP_COLOR_PERCENT_FN(overhead_sys, period_sys, NULL)
 __HPP_COLOR_PERCENT_FN(overhead_us, period_us, NULL)
 __HPP_COLOR_PERCENT_FN(overhead_guest_sys, period_guest_sys, NULL)
 __HPP_COLOR_PERCENT_FN(overhead_guest_us, period_guest_us, NULL)
+__HPP_COLOR_ACC_PERCENT_FN(overhead_acc, period, NULL)
 
 #undef __HPP_COLOR_PERCENT_FN
 
@@ -715,6 +730,8 @@ void hist_browser__init_hpp(void)
 				hist_browser__hpp_color_overhead_guest_sys;
 	perf_hpp__format[PERF_HPP__OVERHEAD_GUEST_US].color =
 				hist_browser__hpp_color_overhead_guest_us;
+	perf_hpp__format[PERF_HPP__OVERHEAD_ACC].color =
+				hist_browser__hpp_color_overhead_acc;
 }
 
 static int hist_browser__show_entry(struct hist_browser *browser,
Signature · 1 line
1.7.11.7
author maintainer
#17 lore raw

PATCH 12/21 perf ui/gtk: Add support to accumulated hist stat

In reply to Namhyung Kim

Print accumulated stat of a hist entry if requested.

---
 tools/perf/ui/gtk/hists.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
tools/perf/ui/gtk/hists.c +16 -0
index 5b95c44f3435..49531c1c7121 100644
@@ -98,11 +98,25 @@ static int perf_gtk__hpp_color_##_type(struct perf_hpp_fmt *fmt __maybe_unused,
 	return __hpp__color_fmt(hpp, he, he_get_##_field);			\
 }
 
+#define __HPP_COLOR_ACC_PERCENT_FN(_type, _field)				\
+static u64 he_get_acc_##_field(struct hist_entry *he)				\
+{										\
+	return he->stat_acc->_field;						\
+}										\
+										\
+static int perf_gtk__hpp_color_##_type(struct perf_hpp_fmt *fmt __maybe_unused,	\
+				       struct perf_hpp *hpp,			\
+				       struct hist_entry *he)			\
+{										\
+	return __hpp__color_fmt(hpp, he, he_get_acc_##_field);			\
+}
+
 __HPP_COLOR_PERCENT_FN(overhead, period)
 __HPP_COLOR_PERCENT_FN(overhead_sys, period_sys)
 __HPP_COLOR_PERCENT_FN(overhead_us, period_us)
 __HPP_COLOR_PERCENT_FN(overhead_guest_sys, period_guest_sys)
 __HPP_COLOR_PERCENT_FN(overhead_guest_us, period_guest_us)
+__HPP_COLOR_ACC_PERCENT_FN(overhead_acc, period)
 
 #undef __HPP_COLOR_PERCENT_FN
 
@@ -121,6 +135,8 @@ void perf_gtk__init_hpp(void)
 				perf_gtk__hpp_color_overhead_guest_sys;
 	perf_hpp__format[PERF_HPP__OVERHEAD_GUEST_US].color =
 				perf_gtk__hpp_color_overhead_guest_us;
+	perf_hpp__format[PERF_HPP__OVERHEAD_ACC].color =
+				perf_gtk__hpp_color_overhead_acc;
 }
 
 static void callchain_list__sym_name(struct callchain_list *cl,
Signature · 1 line
1.7.11.7
author maintainer
#18 lore raw

PATCH 13/21 perf tools: Apply percent-limit to cumulative percentage

In reply to Namhyung Kim

If -g cumulative option is given, it needs to show entries which don't have self overhead. So apply percent-limit to accumulated overhead percentage in this case.

---
 tools/perf/ui/browsers/hists.c | 33 +++++++++++----------------------
 tools/perf/ui/gtk/hists.c      |  4 ++--
 tools/perf/ui/stdio/hist.c     |  4 ++--
 tools/perf/util/hist.h         |  1 +
 tools/perf/util/sort.h         | 17 ++++++++++++++++-
 5 files changed, 32 insertions(+), 27 deletions(-)
tools/perf/ui/browsers/hists.c +11 -22
index c8e3a2e7b5c0..0a088535ebac 100644
@@ -828,12 +828,12 @@ static unsigned int hist_browser__refresh(struct ui_browser *browser)
 
 	for (nd = browser->top; nd; nd = rb_next(nd)) {
 		struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
-		float percent = h->stat.period * 100.0 /
-					hb->hists->stats.total_period;
+		float percent;
 
 		if (h->filtered)
 			continue;
 
+		percent = hist_entry__get_percent_limit(h);
 		if (percent < hb->min_pcnt)
 			continue;
 
@@ -846,18 +846,13 @@ static unsigned int hist_browser__refresh(struct ui_browser *browser)
 }
 
 static struct rb_node *hists__filter_entries(struct rb_node *nd,
-					     struct hists *hists,
 					     float min_pcnt)
 {
 	while (nd != NULL) {
 		struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
-		float percent = h->stat.period * 100.0 /
-					hists->stats.total_period;
+		float percent = hist_entry__get_percent_limit(h);
 
-		if (percent < min_pcnt)
-			return NULL;
-
-		if (!h->filtered)
+		if (!h->filtered && percent >= min_pcnt)
 			return nd;
 
 		nd = rb_next(nd);
@@ -867,13 +862,11 @@ static struct rb_node *hists__filter_entries(struct rb_node *nd,
 }
 
 static struct rb_node *hists__filter_prev_entries(struct rb_node *nd,
-						  struct hists *hists,
 						  float min_pcnt)
 {
 	while (nd != NULL) {
 		struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
-		float percent = h->stat.period * 100.0 /
-					hists->stats.total_period;
+		float percent = hist_entry__get_percent_limit(h);
 
 		if (!h->filtered && percent >= min_pcnt)
 			return nd;
@@ -902,14 +895,14 @@ static void ui_browser__hists_seek(struct ui_browser *browser,
 	switch (whence) {
 	case SEEK_SET:
 		nd = hists__filter_entries(rb_first(browser->entries),
-					   hb->hists, hb->min_pcnt);
+					   hb->min_pcnt);
 		break;
 	case SEEK_CUR:
 		nd = browser->top;
 		goto do_offset;
 	case SEEK_END:
 		nd = hists__filter_prev_entries(rb_last(browser->entries),
-						hb->hists, hb->min_pcnt);
+						hb->min_pcnt);
 		first = false;
 		break;
 	default:
@@ -952,8 +945,7 @@ do_offset:
 					break;
 				}
 			}
-			nd = hists__filter_entries(rb_next(nd), hb->hists,
-						   hb->min_pcnt);
+			nd = hists__filter_entries(rb_next(nd), hb->min_pcnt);
 			if (nd == NULL)
 				break;
 			--offset;
@@ -986,7 +978,7 @@ do_offset:
 				}
 			}
 
-			nd = hists__filter_prev_entries(rb_prev(nd), hb->hists,
+			nd = hists__filter_prev_entries(rb_prev(nd),
 							hb->min_pcnt);
 			if (nd == NULL)
 				break;
@@ -1157,7 +1149,6 @@ static int hist_browser__fprintf_entry(struct hist_browser *browser,
 static int hist_browser__fprintf(struct hist_browser *browser, FILE *fp)
 {
 	struct rb_node *nd = hists__filter_entries(rb_first(browser->b.entries),
-						   browser->hists,
 						   browser->min_pcnt);
 	int printed = 0;
 
@@ -1165,8 +1156,7 @@ static int hist_browser__fprintf(struct hist_browser *browser, FILE *fp)
 		struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
 
 		printed += hist_browser__fprintf_entry(browser, h, fp);
-		nd = hists__filter_entries(rb_next(nd), browser->hists,
-					   browser->min_pcnt);
+		nd = hists__filter_entries(rb_next(nd), browser->min_pcnt);
 	}
 
 	return printed;
@@ -1390,8 +1380,7 @@ static void hist_browser__update_pcnt_entries(struct hist_browser *hb)
 
 	while (nd) {
 		nr_entries++;
-		nd = hists__filter_entries(rb_next(nd), hb->hists,
-					   hb->min_pcnt);
+		nd = hists__filter_entries(rb_next(nd), hb->min_pcnt);
 	}
 
 	hb->nr_pcnt_entries = nr_entries;
tools/perf/ui/gtk/hists.c +2 -2
index 49531c1c7121..ded22b4da677 100644
@@ -296,12 +296,12 @@ static void perf_gtk__show_hists(GtkWidget *window, struct hists *hists,
 	for (nd = rb_first(&hists->entries); nd; nd = rb_next(nd)) {
 		struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
 		GtkTreeIter iter;
-		float percent = h->stat.period * 100.0 /
-					hists->stats.total_period;
+		float percent;
 
 		if (h->filtered)
 			continue;
 
+		percent = hist_entry__get_percent_limit(h);
 		if (percent < min_pcnt)
 			continue;
 
tools/perf/ui/stdio/hist.c +2 -2
index 831fbb77d1ff..b48c7ba49a1c 100644
@@ -487,12 +487,12 @@ print_entries:
 
 	for (nd = rb_first(&hists->entries); nd; nd = rb_next(nd)) {
 		struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
-		float percent = h->stat.period * 100.0 /
-					hists->stats.total_period;
+		float percent;
 
 		if (h->filtered)
 			continue;
 
+		percent = hist_entry__get_percent_limit(h);
 		if (percent < min_pcnt)
 			continue;
 
tools/perf/util/hist.h +1 -0
index 59d8ba4665a4..421c2d809769 100644
@@ -140,6 +140,7 @@ int hist_entry__sort_snprintf(struct hist_entry *he, char *bf, size_t size,
 			      struct hists *hists);
 void hist_entry__free(struct hist_entry *);
 
+
 void hists__output_resort(struct hists *hists);
 void hists__collapse_resort(struct hists *hists, struct ui_progress *prog);
 
tools/perf/util/sort.h +16 -1
index 309f2838a1b4..103748ba25f2 100644
@@ -20,7 +20,7 @@
 
 #include "parse-options.h"
 #include "parse-events.h"
-
+#include "hist.h"
 #include "thread.h"
 
 extern regex_t parent_regex;
@@ -130,6 +130,21 @@ static inline void hist_entry__add_pair(struct hist_entry *pair,
 	list_add_tail(&pair->pairs.node, &he->pairs.head);
 }
 
+static inline float hist_entry__get_percent_limit(struct hist_entry *he)
+{
+	u64 period = he->stat.period;
+	u64 total_period = he->hists->stats.total_period;
+
+	if (unlikely(total_period == 0))
+		return 0;
+
+	if (symbol_conf.cumulate_callchain)
+		period = he->stat_acc->period;
+
+	return period * 100.0 / total_period;
+}
+
+
 enum sort_mode {
 	SORT_MODE__NORMAL,
 	SORT_MODE__BRANCH,
Signature · 1 line
1.7.11.7
author maintainer
#19 lore raw

PATCH 14/21 perf tools: Add more hpp helper functions

In reply to Namhyung Kim

Sometimes it needs to disable some columns at runtime. Add help functions to support that.

---
 tools/perf/ui/hist.c   | 17 +++++++++++++++++
 tools/perf/util/hist.h |  3 +++
 2 files changed, 20 insertions(+)
tools/perf/ui/hist.c +17 -0
index eb9c07bcdb01..1f9e25211da2 100644
@@ -275,12 +275,29 @@ void perf_hpp__column_register(struct perf_hpp_fmt *format)
 	list_add_tail(&format->list, &perf_hpp__list);
 }
 
+void perf_hpp__column_unregister(struct perf_hpp_fmt *format)
+{
+	list_del(&format->list);
+}
+
 void perf_hpp__column_enable(unsigned col)
 {
 	BUG_ON(col >= PERF_HPP__MAX_INDEX);
 	perf_hpp__column_register(&perf_hpp__format[col]);
 }
 
+void perf_hpp__column_disable(unsigned col)
+{
+	BUG_ON(col >= PERF_HPP__MAX_INDEX);
+	perf_hpp__column_unregister(&perf_hpp__format[col]);
+}
+
+void perf_hpp__cancel_cumulate(void)
+{
+	perf_hpp__column_disable(PERF_HPP__OVERHEAD_ACC);
+	perf_hpp__format[PERF_HPP__OVERHEAD].header = hpp__header_overhead;
+}
+
 int hist_entry__sort_snprintf(struct hist_entry *he, char *s, size_t size,
 			      struct hists *hists)
 {
tools/perf/util/hist.h +3 -0
index 421c2d809769..1a91d536e443 100644
@@ -209,7 +209,10 @@ enum {
 
 void perf_hpp__init(void);
 void perf_hpp__column_register(struct perf_hpp_fmt *format);
+void perf_hpp__column_unregister(struct perf_hpp_fmt *format);
 void perf_hpp__column_enable(unsigned col);
+void perf_hpp__column_disable(unsigned col);
+void perf_hpp__cancel_cumulate(void);
 
 static inline size_t perf_hpp__use_color(void)
 {
Signature · 1 line
1.7.11.7
author maintainer
#20 lore raw

PATCH 15/21 perf report: Add --children option

In reply to Namhyung Kim

The --children option is for showing accumulated overhead (period) value as well as self overhead.

---
 tools/perf/Documentation/perf-report.txt |  5 +++++
 tools/perf/builtin-report.c              | 10 ++++++++++
 2 files changed, 15 insertions(+)
tools/perf/Documentation/perf-report.txt +5 -0
index 09af66298564..36eb1877563f 100644
@@ -144,6 +144,11 @@ OPTIONS
 
 	Default: fractal,0.5,callee,function.
 
+--children::
+	Accumulate callchain of children to parent entry so that then can
+	show up in the output.  The output will have a new "Children" column
+	and will be sorted on the data.  It requires callchains are recorded.
+
 --max-stack::
 	Set the stack depth limit when parsing the callchain, anything
 	beyond the specified depth will be ignored. This is a trade-off
tools/perf/builtin-report.c +10 -0
index 3ed0669d7620..56fe8c00b1be 100644
@@ -183,6 +183,14 @@ static int report__setup_sample_type(struct report *rep)
 			}
 	}
 
+	if (symbol_conf.cumulate_callchain) {
+		/* Silently ignore if callchain is missing */
+		if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
+			symbol_conf.cumulate_callchain = false;
+			perf_hpp__cancel_cumulate();
+		}
+	}
+
 	if (sort__mode == SORT_MODE__BRANCH) {
 		if (!is_pipe &&
 		    !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
@@ -661,6 +669,8 @@ int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
 	OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
 		     "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). "
 		     "Default: fractal,0.5,callee,function", &parse_callchain_opt, callchain_default_opt),
+	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
+		    "Accumulate callchains of children and show total overhead as well"),
 	OPT_INTEGER(0, "max-stack", &report.max_stack,
 		    "Set the maximum stack depth when parsing the callchain, "
 		    "anything beyond the specified depth will be ignored. "
Signature · 1 line
1.7.11.7
author maintainer
#21 lore raw

PATCH 16/21 perf report: Add report.children config option

In reply to Namhyung Kim

Add report.children config option for setting default value of callchain accumulation. It affects the report output only if perf.data contains callchain info.

A user can write .perfconfig file like below to enable accumulation by default:

  $ cat ~/.perfconfig
  [report]
  children = true

And it can be disabled through command line:

  $ perf report --no-children
---
 tools/perf/builtin-report.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/perf/builtin-report.c +3 -1
index 56fe8c00b1be..b8541ec61a00 100644
@@ -78,7 +78,9 @@ static int report__config(const char *var, const char *value, void *cb)
 			symbol_conf.filter_relative = false;
 		else
 			return -1;
-
+	}
+	if (!strcmp(var, "report.children")) {
+		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
 		return 0;
 	}
 
Signature · 1 line
1.7.11.7
author maintainer
#22 lore raw

PATCH 17/21 perf tools: Add callback function to hist_entry_iter

In reply to Namhyung Kim

The new ->add_entry_cb() will be called after an entry was added to the histogram. It's used for code sharing between perf report and perf top. Note that ops->add_*_entry() should set iter->he properly in order to call the ->add_entry_cb.

Also pass @arg to the callback function. It'll be used by perf top later.

---
 tools/perf/builtin-report.c | 30 +++++++++++++++++++++++++++---
 tools/perf/util/hist.c      | 45 +++++++++++++++++++++++++--------------------
 tools/perf/util/hist.h      |  6 +++++-
 3 files changed, 57 insertions(+), 24 deletions(-)
tools/perf/builtin-report.c +27 -3
index b8541ec61a00..3dbf2b83fc5e 100644
@@ -87,6 +87,27 @@ static int report__config(const char *var, const char *value, void *cb)
 	return perf_default_config(var, value, cb);
 }
 
+static int hist_iter_cb(struct hist_entry_iter *iter,
+			struct addr_location *al, bool single,
+			void *arg __maybe_unused)
+{
+	int err;
+	struct hist_entry *he = iter->he;
+	struct perf_evsel *evsel = iter->evsel;
+	struct perf_sample *sample = iter->sample;
+
+	err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
+
+	if (!single)
+		goto out;
+
+	evsel->hists.stats.total_period += sample->period;
+	hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
+
+out:
+	return err;
+}
+
 static int process_sample_event(struct perf_tool *tool,
 				union perf_event *event,
 				struct perf_sample *sample,
@@ -116,16 +137,19 @@ static int process_sample_event(struct perf_tool *tool,
 		iter.ops = &hist_iter_branch;
 	else if (rep->mem_mode == 1)
 		iter.ops = &hist_iter_mem;
-	else if (symbol_conf.cumulate_callchain)
+	else if (symbol_conf.cumulate_callchain) {
 		iter.ops = &hist_iter_cumulative;
-	else
+		iter.add_entry_cb = hist_iter_cb;
+	} else {
 		iter.ops = &hist_iter_normal;
+		iter.add_entry_cb = hist_iter_cb;
+	}
 
 	if (al.map != NULL)
 		al.map->dso->hit = 1;
 
 	ret = hist_entry_iter__add(&iter, &al, evsel, event, sample,
-				   rep->max_stack);
+				   rep->max_stack, NULL);
 	if (ret < 0)
 		pr_debug("problem adding hist entry, skipping event\n");
 
tools/perf/util/hist.c +25 -20

111 lines not on this page · raw message

tools/perf/util/hist.h +5 -1
index 1a91d536e443..2f52c4dd61de 100644
@@ -115,6 +115,9 @@ struct hist_entry_iter {
 	void *priv;
 
 	const struct hist_iter_ops *ops;
+	/* user-defined callback function (optional) */
+	int (*add_entry_cb)(struct hist_entry_iter *iter,
+			    struct addr_location *al, bool single, void *arg);
 };
 
 extern const struct hist_iter_ops hist_iter_normal;
@@ -131,7 +134,8 @@ struct hist_entry *__hists__add_entry(struct hists *hists,
 				      bool sample_self);
 int hist_entry_iter__add(struct hist_entry_iter *iter, struct addr_location *al,
 			 struct perf_evsel *evsel, const union perf_event *event,
-			 struct perf_sample *sample, int max_stack_depth);
+			 struct perf_sample *sample, int max_stack_depth,
+			 void *arg);
 
 int64_t hist_entry__cmp(struct hist_entry *left, struct hist_entry *right);
 int64_t hist_entry__collapse(struct hist_entry *left, struct hist_entry *right);
Signature · 1 line
1.7.11.7
author maintainer
#23 lore raw

PATCH 18/21 perf top: Convert to hist_entry_iter

In reply to Namhyung Kim

Reuse hist_entry_iter__add() function to share the similar code with perf report. Note that it needs to be called with hists.lock so tweak some internal functions not to deadlock or hold the lock too long.

---
 tools/perf/builtin-top.c | 78 ++++++++++++++++++++++++++----------------------
 1 file changed, 43 insertions(+), 35 deletions(-)
tools/perf/builtin-top.c +43 -35
index c574c291383c..9de14f71c196 100644
@@ -194,6 +194,12 @@ static void perf_top__record_precise_ip(struct perf_top *top,
 
 	pthread_mutex_unlock(&notes->lock);
 
+	/*
+	 * This function is now called with he->hists->lock held.
+	 * Release it before going to sleep.
+	 */
+	pthread_mutex_unlock(&he->hists->lock);
+
 	if (err == -ERANGE && !he->ms.map->erange_warned)
 		ui__warn_map_erange(he->ms.map, sym, ip);
 	else if (err == -ENOMEM) {
@@ -201,6 +207,8 @@ static void perf_top__record_precise_ip(struct perf_top *top,
 		       sym->name);
 		sleep(1);
 	}
+
+	pthread_mutex_lock(&he->hists->lock);
 }
 
 static void perf_top__show_details(struct perf_top *top)
@@ -236,24 +244,6 @@ out_unlock:
 	pthread_mutex_unlock(&notes->lock);
 }
 
-static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
-						     struct addr_location *al,
-						     struct perf_sample *sample)
-{
-	struct hist_entry *he;
-
-	pthread_mutex_lock(&evsel->hists.lock);
-	he = __hists__add_entry(&evsel->hists, al, NULL, NULL, NULL,
-				sample->period, sample->weight,
-				sample->transaction, true);
-	pthread_mutex_unlock(&evsel->hists.lock);
-	if (he == NULL)
-		return NULL;
-
-	hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
-	return he;
-}
-
 static void perf_top__print_sym_table(struct perf_top *top)
 {
 	char bf[160];
@@ -657,6 +647,28 @@ static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 	return 0;
 }
 
+static int hist_iter_cb(struct hist_entry_iter *iter, struct addr_location *al,
+			bool single, void *arg)
+{
+	struct perf_top *top = arg;
+	struct hist_entry *he = iter->he;
+	struct perf_evsel *evsel = iter->evsel;
+
+	if (single)
+		hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
+
+	if (sort__has_sym) {
+		u64 ip = al->addr;
+
+		if (al->map)
+			ip = al->map->unmap_ip(al->map, ip);
+
+		perf_top__record_precise_ip(top, he, evsel->idx, ip);
+	}
+
+	return 0;
+}
+
 static void perf_event__process_sample(struct perf_tool *tool,
 				       const union perf_event *event,
 				       struct perf_evsel *evsel,
@@ -664,8 +676,6 @@ static void perf_event__process_sample(struct perf_tool *tool,
 				       struct machine *machine)
 {
 	struct perf_top *top = container_of(tool, struct perf_top, tool);
-	struct symbol *parent = NULL;
-	u64 ip = sample->ip;
 	struct addr_location al;
 	int err;
 
@@ -741,25 +751,23 @@ static void perf_event__process_sample(struct perf_tool *tool,
 	}
 
 	if (al.sym == NULL || !al.sym->ignore) {
-		struct hist_entry *he;
+		struct hist_entry_iter iter = {
+			.add_entry_cb = hist_iter_cb,
+		};
 
-		err = sample__resolve_callchain(sample, &parent, evsel, &al,
-						top->max_stack);
-		if (err)
-			return;
+		if (symbol_conf.cumulate_callchain)
+			iter.ops = &hist_iter_cumulative;
+		else
+			iter.ops = &hist_iter_normal;
 
-		he = perf_evsel__add_hist_entry(evsel, &al, sample);
-		if (he == NULL) {
-			pr_err("Problem incrementing symbol period, skipping event\n");
-			return;
-		}
+		pthread_mutex_lock(&evsel->hists.lock);
 
-		err = hist_entry__append_callchain(he, sample);
-		if (err)
-			return;
+		err = hist_entry_iter__add(&iter, &al, evsel, event, sample,
+					   top->max_stack, top);
+		if (err < 0)
+			pr_err("Problem incrementing symbol period, skipping event\n");
 
-		if (sort__has_sym)
-			perf_top__record_precise_ip(top, he, evsel->idx, ip);
+		pthread_mutex_unlock(&evsel->hists.lock);
 	}
 
 	return;
Signature · 1 line
1.7.11.7
author maintainer
#24 lore raw

PATCH 19/21 perf top: Add --children option

In reply to Namhyung Kim

The --children option is for showing accumulated overhead (period) value as well as self overhead. It should be used with one of -g or --call-graph option.

---
 tools/perf/Documentation/perf-top.txt | 6 ++++++
 tools/perf/builtin-top.c              | 7 +++++++
 2 files changed, 13 insertions(+)
tools/perf/Documentation/perf-top.txt +6 -0
index cdd8d4946dba..01b6fd1a4428 100644
@@ -149,6 +149,12 @@ Default is to monitor all CPUS.
 	Setup and enable call-graph (stack chain/backtrace) recording,
 	implies -g.
 
+--children::
+	Accumulate callchain of children to parent entry so that then can
+	show up in the output.  The output will have a new "Children" column
+	and will be sorted on the data.  It requires -g/--call-graph option
+	enabled.
+
 --max-stack::
 	Set the stack depth limit when parsing the callchain, anything
 	beyond the specified depth will be ignored. This is a trade-off
tools/perf/builtin-top.c +7 -0
index 9de14f71c196..439e000b0f39 100644
@@ -1087,6 +1087,8 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
 	OPT_CALLBACK(0, "call-graph", &top.record_opts,
 		     "mode[,dump_size]", record_callchain_help,
 		     &parse_callchain_opt),
+	OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
+		    "Accumulate callchains of children and show total overhead as well"),
 	OPT_INTEGER(0, "max-stack", &top.max_stack,
 		    "Set the maximum stack depth when parsing the callchain. "
 		    "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
@@ -1186,6 +1188,11 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
 
 	top.sym_evsel = perf_evlist__first(top.evlist);
 
+	if (!symbol_conf.use_callchain) {
+		symbol_conf.cumulate_callchain = false;
+		perf_hpp__cancel_cumulate();
+	}
+
 	symbol_conf.priv_size = sizeof(struct annotation);
 
 	symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
Signature · 1 line
1.7.11.7
author maintainer
#25 lore raw

PATCH 20/21 perf top: Add top.children config option

In reply to Namhyung Kim

Add top.children config option for setting default value of callchain accumulation. It affects the output only if one of -g or --call-graph option is given as well.

A user can write .perfconfig file like below to enable accumulation by default:

  $ cat ~/.perfconfig
  [top]
  children = true

And it can be disabled through command line:

  $ perf top --no-children
---
 tools/perf/builtin-top.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
tools/perf/builtin-top.c +12 -0
index 439e000b0f39..2e521d00c0a4 100644
@@ -87,6 +87,16 @@ static void perf_top__sig_winch(int sig __maybe_unused,
 	perf_top__update_print_entries(top);
 }
 
+static int top__config(const char *var, const char *value, void *cb)
+{
+	if (!strcmp(var, "top.children")) {
+		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
+		return 0;
+	}
+
+	return perf_default_config(var, value, cb);
+}
+
 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 {
 	struct symbol *sym;
@@ -1125,6 +1135,8 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
 	if (top.evlist == NULL)
 		return -ENOMEM;
 
+	perf_config(top__config, &top);
+
 	argc = parse_options(argc, argv, options, top_usage, 0);
 	if (argc)
 		usage_with_options(top_usage, options);
Signature · 1 line
1.7.11.7
author maintainer
#26 lore raw

PATCH 21/21 perf tools: Enable --children option by default

In reply to Namhyung Kim

Now perf top and perf report will show children column by default if it has callchain information.

---
 tools/perf/util/symbol.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
tools/perf/util/symbol.c +6 -5
index 39ce9adbaaf0..186dca9df434 100644
@@ -29,11 +29,12 @@ int vmlinux_path__nr_entries;
 char **vmlinux_path;
 
 struct symbol_conf symbol_conf = {
-	.use_modules	  = true,
-	.try_vmlinux_path = true,
-	.annotate_src	  = true,
-	.demangle	  = true,
-	.symfs            = "",
+	.use_modules		= true,
+	.try_vmlinux_path	= true,
+	.annotate_src		= true,
+	.demangle		= true,
+	.cumulate_callchain	= true,
+	.symfs			= "",
 };
 
 static enum dso_binary_type binary_type_symtab[] = {
Signature · 1 line
1.7.11.7
#27 lore raw
In reply to Namhyung Kim
Quoted 28 lines · Namhyung Kim

Hello,

This is a new attempt to implement cumulative hist period report. This work begins from Arun's SORT_INCLUSIVE patch [1] but I completely rewrote it from scratch.

This patchset is based on my previous patchset [2] but I think it's almost independent so that it can be applied separately.

Please see the patch 03/24. I refactored functions that add hist entries with struct hist_entry_iter. While I converted all functions carefully, it'd be better anyone can test and confirm that I didn't mess up something - especially for branch stack and mem stuff.

This patchset basically adds period in a sample to every node in the callchain. A hist_entry now has an additional fields to keep the cumulative period if --children option is given on perf report.

I changed the option as a separate --children and added a new "Children" column (and renamed the default "Overhead" column into "Self"). The output will be sorted by children (cumulative) overhead for now. The reason I changed to the --children is that I still think it's much different from other --call-graph options. The --call-graph option will take care of it even with --children option.

I know that the UI should be changed also to be more flexible as Ingo requested, but I'd like to do this first and then move to work on the next. I also added a new config option to enable it by default.

SNIP

Quoted 7 lines
                  --- _start

Case 3.

  $ perf report --no-call-graph --children --stdio
Quoted 12 lines
  #     Self  Children  Command      Shared Object                 Symbol
  # ........  ........  .......  .................  .....................
  #
       0.00%    91.50%      abc  libc-2.17.so       [.] __libc_start_main
       0.00%    91.50%      abc  abc                [.] main             
       0.00%    91.50%      abc  abc                [.] c                
       0.00%    91.50%      abc  abc                [.] b                
      91.50%    91.50%      abc  abc                [.] a                
       0.00%     8.18%      abc  ld-2.17.so         [.] _dl_sysdep_start 
       8.18%     8.18%      abc  ld-2.17.so         [.] strlen           
       0.01%     0.33%      abc  ld-2.17.so         [.] _start           
       0.31%     0.31%      abc  [kernel.kallsyms]  [k] page_fault       

I couldnt get this one working:

[jolsa@krava perf]$ ./perf report --no-call-graph --children --stdio

#

#     Self  Children  Command  Shared Object                      Symbol
# ........  ........  .......  .............  ..........................
#
    94.47%    94.47%      abc  abc            [.] a                     
     4.74%     4.74%      abc  ld-2.17.so     [.] _dl_map_object_from_fd
     0.75%     0.75%      abc  ld-2.17.so     [.] _dl_start             
     0.04%     0.04%      abc  ld-2.17.so     [.] _start                

I was looking for cumulated stdio output without callchain unwinded, TUI shows it properly as default

jirka

author maintainer
#28 lore raw
Quoted 30 lines · Jiri Olsa
Quoted 15 lines · Namhyung Kim
  $ perf report --no-call-graph --children --stdio
  #     Self  Children  Command      Shared Object                 Symbol
  # ........  ........  .......  .................  .....................
  #
       0.00%    91.50%      abc  libc-2.17.so       [.] __libc_start_main
       0.00%    91.50%      abc  abc                [.] main             
       0.00%    91.50%      abc  abc                [.] c                
       0.00%    91.50%      abc  abc                [.] b                
      91.50%    91.50%      abc  abc                [.] a                
       0.00%     8.18%      abc  ld-2.17.so         [.] _dl_sysdep_start 
       8.18%     8.18%      abc  ld-2.17.so         [.] strlen           
       0.01%     0.33%      abc  ld-2.17.so         [.] _start           
       0.31%     0.31%      abc  [kernel.kallsyms]  [k] page_fault       

I couldnt get this one working:

[jolsa@krava perf]$ ./perf report --no-call-graph --children --stdio

#

#     Self  Children  Command  Shared Object                      Symbol
# ........  ........  .......  .............  ..........................
#
    94.47%    94.47%      abc  abc            [.] a                     
     4.74%     4.74%      abc  ld-2.17.so     [.] _dl_map_object_from_fd
     0.75%     0.75%      abc  ld-2.17.so     [.] _dl_start             
     0.04%     0.04%      abc  ld-2.17.so     [.] _start                
Quoted 2 lines

I was looking for cumulated stdio output without callchain unwinded, TUI shows it properly as default

Oops, I missed a change below, sorry.

tools/perf/util/callchain.c +2 -1
index ac658135079f..c1d4ccd755ed 100644
@@ -538,7 +538,8 @@ int sample__resolve_callchain(struct perf_sample *sample, struct symbol **parent
 	if (sample->callchain == NULL)
 		return 0;
 
-	if (symbol_conf.use_callchain || sort__has_parent) {
+	if (symbol_conf.use_callchain || symbol_conf.cumulate_callchain ||
+	    sort__has_parent) {
 		return machine__resolve_callchain(al->machine, evsel, al->thread,
 						  sample, parent, al, max_stack);
 	}
Signature · 1 line
1.7.11.7