* [PATCH 2/4] perf report: Enable data-type profiling with -F option too
2025-12-10 2:33 [PATCH 1/4] perf record: Split --data-mmap option Namhyung Kim
@ 2025-12-10 2:33 ` Namhyung Kim
2025-12-10 22:05 ` Ian Rogers
2025-12-10 2:33 ` [PATCH 3/4] perf report: Fix histogram entry collapsing for -F option Namhyung Kim
` (2 subsequent siblings)
3 siblings, 1 reply; 9+ messages in thread
From: Namhyung Kim @ 2025-12-10 2:33 UTC (permalink / raw)
To: Arnaldo Carvalho de Melo, Ian Rogers, James Clark
Cc: Jiri Olsa, Adrian Hunter, Peter Zijlstra, Ingo Molnar, LKML,
linux-perf-users
It checked -s/--sort options only. As the sort keys can be setup using
the -F/--fields option as well, it should enable data-type profiling
with it too.
The following two commands should have the same output.
$ perf report -s type
$ perf report -F overhead,type
But there's another problem on this. I'll handle it in the next commit.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
tools/perf/builtin-report.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index add6b1c2aaf04270..6c2b4f93ec78e579 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -1727,7 +1727,8 @@ int cmd_report(int argc, const char **argv)
sort_order = NULL;
}
- if (sort_order && strstr(sort_order, "type")) {
+ if ((sort_order && strstr(sort_order, "type")) ||
+ (field_order && strstr(field_order, "type"))) {
report.data_type = true;
annotate_opts.annotate_src = false;
--
2.52.0.223.gf5cc29aaa4-goog
^ permalink raw reply related [flat|nested] 9+ messages in thread* Re: [PATCH 2/4] perf report: Enable data-type profiling with -F option too
2025-12-10 2:33 ` [PATCH 2/4] perf report: Enable data-type profiling with -F option too Namhyung Kim
@ 2025-12-10 22:05 ` Ian Rogers
0 siblings, 0 replies; 9+ messages in thread
From: Ian Rogers @ 2025-12-10 22:05 UTC (permalink / raw)
To: Namhyung Kim
Cc: Arnaldo Carvalho de Melo, James Clark, Jiri Olsa, Adrian Hunter,
Peter Zijlstra, Ingo Molnar, LKML, linux-perf-users
On Tue, Dec 9, 2025 at 6:33 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> It checked -s/--sort options only. As the sort keys can be setup using
> the -F/--fields option as well, it should enable data-type profiling
> with it too.
>
> The following two commands should have the same output.
>
> $ perf report -s type
>
> $ perf report -F overhead,type
>
> But there's another problem on this. I'll handle it in the next commit.
>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Thanks,
Ian
> ---
> tools/perf/builtin-report.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
> index add6b1c2aaf04270..6c2b4f93ec78e579 100644
> --- a/tools/perf/builtin-report.c
> +++ b/tools/perf/builtin-report.c
> @@ -1727,7 +1727,8 @@ int cmd_report(int argc, const char **argv)
> sort_order = NULL;
> }
>
> - if (sort_order && strstr(sort_order, "type")) {
> + if ((sort_order && strstr(sort_order, "type")) ||
> + (field_order && strstr(field_order, "type"))) {
> report.data_type = true;
> annotate_opts.annotate_src = false;
>
> --
> 2.52.0.223.gf5cc29aaa4-goog
>
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 3/4] perf report: Fix histogram entry collapsing for -F option
2025-12-10 2:33 [PATCH 1/4] perf record: Split --data-mmap option Namhyung Kim
2025-12-10 2:33 ` [PATCH 2/4] perf report: Enable data-type profiling with -F option too Namhyung Kim
@ 2025-12-10 2:33 ` Namhyung Kim
2025-12-10 22:06 ` Ian Rogers
2025-12-10 2:33 ` [PATCH 4/4] perf report: Update sort key state from " Namhyung Kim
2025-12-10 22:05 ` [PATCH 1/4] perf record: Split --data-mmap option Ian Rogers
3 siblings, 1 reply; 9+ messages in thread
From: Namhyung Kim @ 2025-12-10 2:33 UTC (permalink / raw)
To: Arnaldo Carvalho de Melo, Ian Rogers, James Clark
Cc: Jiri Olsa, Adrian Hunter, Peter Zijlstra, Ingo Molnar, LKML,
linux-perf-users
Users can use -F/--fields option to set output fields and sort keys
together. But it missed to set perf_hpp_list->need_collapse for sort
entries that have se_collapse callbacks. So it ends up with having
duplicated entries separately.
For example, let's run this command first.
$ perf mem record -t load -U -- perf test -w datasym
This will record samples for memory access (load) to struct 'buf' and a
loop condition ('sig_atomic_t') types. So the following two commands
should have identical output.
$ perf report -s type --stdio --percent-limit=1 -q
87.80% perf buf
12.17% perf sig_atomic_t
But using -F option didn't collapse the entries based on types so the
result looked like below:
$ perf report -F overhead,type --stdio --percent-limit=1 -q
23.31% perf buf
22.84% perf buf
21.26% perf buf
20.39% perf buf
12.17% perf sig_atomic_t
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
tools/perf/util/sort.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/tools/perf/util/sort.c b/tools/perf/util/sort.c
index a0a5c1c40af0c318..c51604eaae0a4b90 100644
--- a/tools/perf/util/sort.c
+++ b/tools/perf/util/sort.c
@@ -3578,6 +3578,9 @@ static int __sort_dimension__add_output(struct perf_hpp_list *list,
if (__sort_dimension__add_hpp_output(sd, list, level) < 0)
return -1;
+ if (sd->entry->se_collapse)
+ list->need_collapse = 1;
+
sd->taken = 1;
return 0;
}
--
2.52.0.223.gf5cc29aaa4-goog
^ permalink raw reply related [flat|nested] 9+ messages in thread* Re: [PATCH 3/4] perf report: Fix histogram entry collapsing for -F option
2025-12-10 2:33 ` [PATCH 3/4] perf report: Fix histogram entry collapsing for -F option Namhyung Kim
@ 2025-12-10 22:06 ` Ian Rogers
0 siblings, 0 replies; 9+ messages in thread
From: Ian Rogers @ 2025-12-10 22:06 UTC (permalink / raw)
To: Namhyung Kim
Cc: Arnaldo Carvalho de Melo, James Clark, Jiri Olsa, Adrian Hunter,
Peter Zijlstra, Ingo Molnar, LKML, linux-perf-users
On Tue, Dec 9, 2025 at 6:33 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Users can use -F/--fields option to set output fields and sort keys
> together. But it missed to set perf_hpp_list->need_collapse for sort
> entries that have se_collapse callbacks. So it ends up with having
> duplicated entries separately.
>
> For example, let's run this command first.
>
> $ perf mem record -t load -U -- perf test -w datasym
>
> This will record samples for memory access (load) to struct 'buf' and a
> loop condition ('sig_atomic_t') types. So the following two commands
> should have identical output.
>
> $ perf report -s type --stdio --percent-limit=1 -q
> 87.80% perf buf
> 12.17% perf sig_atomic_t
>
> But using -F option didn't collapse the entries based on types so the
> result looked like below:
>
> $ perf report -F overhead,type --stdio --percent-limit=1 -q
> 23.31% perf buf
> 22.84% perf buf
> 21.26% perf buf
> 20.39% perf buf
> 12.17% perf sig_atomic_t
>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Thanks,
Ian
> ---
> tools/perf/util/sort.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/tools/perf/util/sort.c b/tools/perf/util/sort.c
> index a0a5c1c40af0c318..c51604eaae0a4b90 100644
> --- a/tools/perf/util/sort.c
> +++ b/tools/perf/util/sort.c
> @@ -3578,6 +3578,9 @@ static int __sort_dimension__add_output(struct perf_hpp_list *list,
> if (__sort_dimension__add_hpp_output(sd, list, level) < 0)
> return -1;
>
> + if (sd->entry->se_collapse)
> + list->need_collapse = 1;
> +
> sd->taken = 1;
> return 0;
> }
> --
> 2.52.0.223.gf5cc29aaa4-goog
>
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 4/4] perf report: Update sort key state from -F option
2025-12-10 2:33 [PATCH 1/4] perf record: Split --data-mmap option Namhyung Kim
2025-12-10 2:33 ` [PATCH 2/4] perf report: Enable data-type profiling with -F option too Namhyung Kim
2025-12-10 2:33 ` [PATCH 3/4] perf report: Fix histogram entry collapsing for -F option Namhyung Kim
@ 2025-12-10 2:33 ` Namhyung Kim
2025-12-10 22:06 ` Ian Rogers
2025-12-10 22:05 ` [PATCH 1/4] perf record: Split --data-mmap option Ian Rogers
3 siblings, 1 reply; 9+ messages in thread
From: Namhyung Kim @ 2025-12-10 2:33 UTC (permalink / raw)
To: Arnaldo Carvalho de Melo, Ian Rogers, James Clark
Cc: Jiri Olsa, Adrian Hunter, Peter Zijlstra, Ingo Molnar, LKML,
linux-perf-users
Factor out __sort_dimension__update() so that it can be called from -s
and -F option parsing logics. Otherwise the following command cannot go
into the annotation mode.
$ perf report -F overhead,type,sym
Warning: Annotation is only available for symbolic views, include "sym*" in --sort to use it.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
tools/perf/util/sort.c | 100 ++++++++++++++++++++++-------------------
1 file changed, 54 insertions(+), 46 deletions(-)
diff --git a/tools/perf/util/sort.c b/tools/perf/util/sort.c
index c51604eaae0a4b90..5e27a66e3ccb0bbe 100644
--- a/tools/perf/util/sort.c
+++ b/tools/perf/util/sort.c
@@ -3531,6 +3531,56 @@ static int add_dynamic_entry(struct evlist *evlist, const char *tok,
return ret;
}
+static int __sort_dimension__update(struct sort_dimension *sd,
+ struct perf_hpp_list *list)
+{
+ if (sd->entry == &sort_parent && parent_pattern) {
+ int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
+ if (ret) {
+ char err[BUFSIZ];
+
+ regerror(ret, &parent_regex, err, sizeof(err));
+ pr_err("Invalid regex: %s\n%s", parent_pattern, err);
+ return -EINVAL;
+ }
+ list->parent = 1;
+ } else if (sd->entry == &sort_sym) {
+ list->sym = 1;
+ /*
+ * perf diff displays the performance difference amongst
+ * two or more perf.data files. Those files could come
+ * from different binaries. So we should not compare
+ * their ips, but the name of symbol.
+ */
+ if (sort__mode == SORT_MODE__DIFF)
+ sd->entry->se_collapse = sort__sym_sort;
+
+ } else if (sd->entry == &sort_sym_offset) {
+ list->sym = 1;
+ } else if (sd->entry == &sort_dso) {
+ list->dso = 1;
+ } else if (sd->entry == &sort_socket) {
+ list->socket = 1;
+ } else if (sd->entry == &sort_thread) {
+ list->thread = 1;
+ } else if (sd->entry == &sort_comm) {
+ list->comm = 1;
+ } else if (sd->entry == &sort_type_offset) {
+ symbol_conf.annotate_data_member = true;
+ } else if (sd->entry == &sort_sym_from || sd->entry == &sort_sym_to) {
+ list->sym = 1;
+ } else if (sd->entry == &sort_mem_dcacheline && cacheline_size() == 0) {
+ return -EINVAL;
+ } else if (sd->entry == &sort_mem_daddr_sym) {
+ list->sym = 1;
+ }
+
+ if (sd->entry->se_collapse)
+ list->need_collapse = 1;
+
+ return 0;
+}
+
static int __sort_dimension__add(struct sort_dimension *sd,
struct perf_hpp_list *list,
int level)
@@ -3541,8 +3591,8 @@ static int __sort_dimension__add(struct sort_dimension *sd,
if (__sort_dimension__add_hpp_sort(sd, list, level) < 0)
return -1;
- if (sd->entry->se_collapse)
- list->need_collapse = 1;
+ if (__sort_dimension__update(sd, list) < 0)
+ return -1;
sd->taken = 1;
@@ -3578,8 +3628,8 @@ static int __sort_dimension__add_output(struct perf_hpp_list *list,
if (__sort_dimension__add_hpp_output(sd, list, level) < 0)
return -1;
- if (sd->entry->se_collapse)
- list->need_collapse = 1;
+ if (__sort_dimension__update(sd, list) < 0)
+ return -1;
sd->taken = 1;
return 0;
@@ -3644,39 +3694,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
sort_dimension_add_dynamic_header(sd, env);
}
- if (sd->entry == &sort_parent && parent_pattern) {
- int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
- if (ret) {
- char err[BUFSIZ];
-
- regerror(ret, &parent_regex, err, sizeof(err));
- pr_err("Invalid regex: %s\n%s", parent_pattern, err);
- return -EINVAL;
- }
- list->parent = 1;
- } else if (sd->entry == &sort_sym) {
- list->sym = 1;
- /*
- * perf diff displays the performance difference amongst
- * two or more perf.data files. Those files could come
- * from different binaries. So we should not compare
- * their ips, but the name of symbol.
- */
- if (sort__mode == SORT_MODE__DIFF)
- sd->entry->se_collapse = sort__sym_sort;
-
- } else if (sd->entry == &sort_dso) {
- list->dso = 1;
- } else if (sd->entry == &sort_socket) {
- list->socket = 1;
- } else if (sd->entry == &sort_thread) {
- list->thread = 1;
- } else if (sd->entry == &sort_comm) {
- list->comm = 1;
- } else if (sd->entry == &sort_type_offset) {
- symbol_conf.annotate_data_member = true;
- }
-
return __sort_dimension__add(sd, list, level);
}
@@ -3695,9 +3712,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
strlen(tok)))
return -EINVAL;
- if (sd->entry == &sort_sym_from || sd->entry == &sort_sym_to)
- list->sym = 1;
-
__sort_dimension__add(sd, list, level);
return 0;
}
@@ -3711,12 +3725,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
if (sort__mode != SORT_MODE__MEMORY)
return -EINVAL;
- if (sd->entry == &sort_mem_dcacheline && cacheline_size() == 0)
- return -EINVAL;
-
- if (sd->entry == &sort_mem_daddr_sym)
- list->sym = 1;
-
__sort_dimension__add(sd, list, level);
return 0;
}
--
2.52.0.223.gf5cc29aaa4-goog
^ permalink raw reply related [flat|nested] 9+ messages in thread* Re: [PATCH 4/4] perf report: Update sort key state from -F option
2025-12-10 2:33 ` [PATCH 4/4] perf report: Update sort key state from " Namhyung Kim
@ 2025-12-10 22:06 ` Ian Rogers
0 siblings, 0 replies; 9+ messages in thread
From: Ian Rogers @ 2025-12-10 22:06 UTC (permalink / raw)
To: Namhyung Kim
Cc: Arnaldo Carvalho de Melo, James Clark, Jiri Olsa, Adrian Hunter,
Peter Zijlstra, Ingo Molnar, LKML, linux-perf-users
On Tue, Dec 9, 2025 at 6:33 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Factor out __sort_dimension__update() so that it can be called from -s
> and -F option parsing logics. Otherwise the following command cannot go
> into the annotation mode.
>
> $ perf report -F overhead,type,sym
>
> Warning: Annotation is only available for symbolic views, include "sym*" in --sort to use it.
>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Thanks,
Ian
> ---
> tools/perf/util/sort.c | 100 ++++++++++++++++++++++-------------------
> 1 file changed, 54 insertions(+), 46 deletions(-)
>
> diff --git a/tools/perf/util/sort.c b/tools/perf/util/sort.c
> index c51604eaae0a4b90..5e27a66e3ccb0bbe 100644
> --- a/tools/perf/util/sort.c
> +++ b/tools/perf/util/sort.c
> @@ -3531,6 +3531,56 @@ static int add_dynamic_entry(struct evlist *evlist, const char *tok,
> return ret;
> }
>
> +static int __sort_dimension__update(struct sort_dimension *sd,
> + struct perf_hpp_list *list)
> +{
> + if (sd->entry == &sort_parent && parent_pattern) {
> + int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
> + if (ret) {
> + char err[BUFSIZ];
> +
> + regerror(ret, &parent_regex, err, sizeof(err));
> + pr_err("Invalid regex: %s\n%s", parent_pattern, err);
> + return -EINVAL;
> + }
> + list->parent = 1;
> + } else if (sd->entry == &sort_sym) {
> + list->sym = 1;
> + /*
> + * perf diff displays the performance difference amongst
> + * two or more perf.data files. Those files could come
> + * from different binaries. So we should not compare
> + * their ips, but the name of symbol.
> + */
> + if (sort__mode == SORT_MODE__DIFF)
> + sd->entry->se_collapse = sort__sym_sort;
> +
> + } else if (sd->entry == &sort_sym_offset) {
> + list->sym = 1;
> + } else if (sd->entry == &sort_dso) {
> + list->dso = 1;
> + } else if (sd->entry == &sort_socket) {
> + list->socket = 1;
> + } else if (sd->entry == &sort_thread) {
> + list->thread = 1;
> + } else if (sd->entry == &sort_comm) {
> + list->comm = 1;
> + } else if (sd->entry == &sort_type_offset) {
> + symbol_conf.annotate_data_member = true;
> + } else if (sd->entry == &sort_sym_from || sd->entry == &sort_sym_to) {
> + list->sym = 1;
> + } else if (sd->entry == &sort_mem_dcacheline && cacheline_size() == 0) {
> + return -EINVAL;
> + } else if (sd->entry == &sort_mem_daddr_sym) {
> + list->sym = 1;
> + }
> +
> + if (sd->entry->se_collapse)
> + list->need_collapse = 1;
> +
> + return 0;
> +}
> +
> static int __sort_dimension__add(struct sort_dimension *sd,
> struct perf_hpp_list *list,
> int level)
> @@ -3541,8 +3591,8 @@ static int __sort_dimension__add(struct sort_dimension *sd,
> if (__sort_dimension__add_hpp_sort(sd, list, level) < 0)
> return -1;
>
> - if (sd->entry->se_collapse)
> - list->need_collapse = 1;
> + if (__sort_dimension__update(sd, list) < 0)
> + return -1;
>
> sd->taken = 1;
>
> @@ -3578,8 +3628,8 @@ static int __sort_dimension__add_output(struct perf_hpp_list *list,
> if (__sort_dimension__add_hpp_output(sd, list, level) < 0)
> return -1;
>
> - if (sd->entry->se_collapse)
> - list->need_collapse = 1;
> + if (__sort_dimension__update(sd, list) < 0)
> + return -1;
>
> sd->taken = 1;
> return 0;
> @@ -3644,39 +3694,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
> sort_dimension_add_dynamic_header(sd, env);
> }
>
> - if (sd->entry == &sort_parent && parent_pattern) {
> - int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
> - if (ret) {
> - char err[BUFSIZ];
> -
> - regerror(ret, &parent_regex, err, sizeof(err));
> - pr_err("Invalid regex: %s\n%s", parent_pattern, err);
> - return -EINVAL;
> - }
> - list->parent = 1;
> - } else if (sd->entry == &sort_sym) {
> - list->sym = 1;
> - /*
> - * perf diff displays the performance difference amongst
> - * two or more perf.data files. Those files could come
> - * from different binaries. So we should not compare
> - * their ips, but the name of symbol.
> - */
> - if (sort__mode == SORT_MODE__DIFF)
> - sd->entry->se_collapse = sort__sym_sort;
> -
> - } else if (sd->entry == &sort_dso) {
> - list->dso = 1;
> - } else if (sd->entry == &sort_socket) {
> - list->socket = 1;
> - } else if (sd->entry == &sort_thread) {
> - list->thread = 1;
> - } else if (sd->entry == &sort_comm) {
> - list->comm = 1;
> - } else if (sd->entry == &sort_type_offset) {
> - symbol_conf.annotate_data_member = true;
> - }
> -
> return __sort_dimension__add(sd, list, level);
> }
>
> @@ -3695,9 +3712,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
> strlen(tok)))
> return -EINVAL;
>
> - if (sd->entry == &sort_sym_from || sd->entry == &sort_sym_to)
> - list->sym = 1;
> -
> __sort_dimension__add(sd, list, level);
> return 0;
> }
> @@ -3711,12 +3725,6 @@ int sort_dimension__add(struct perf_hpp_list *list, const char *tok,
> if (sort__mode != SORT_MODE__MEMORY)
> return -EINVAL;
>
> - if (sd->entry == &sort_mem_dcacheline && cacheline_size() == 0)
> - return -EINVAL;
> -
> - if (sd->entry == &sort_mem_daddr_sym)
> - list->sym = 1;
> -
> __sort_dimension__add(sd, list, level);
> return 0;
> }
> --
> 2.52.0.223.gf5cc29aaa4-goog
>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 1/4] perf record: Split --data-mmap option
2025-12-10 2:33 [PATCH 1/4] perf record: Split --data-mmap option Namhyung Kim
` (2 preceding siblings ...)
2025-12-10 2:33 ` [PATCH 4/4] perf report: Update sort key state from " Namhyung Kim
@ 2025-12-10 22:05 ` Ian Rogers
2025-12-17 12:37 ` Arnaldo Carvalho de Melo
3 siblings, 1 reply; 9+ messages in thread
From: Ian Rogers @ 2025-12-10 22:05 UTC (permalink / raw)
To: Namhyung Kim
Cc: Arnaldo Carvalho de Melo, James Clark, Jiri Olsa, Adrian Hunter,
Peter Zijlstra, Ingo Molnar, LKML, linux-perf-users
On Tue, Dec 9, 2025 at 6:33 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Currently -d/--data option controls both PERF_SAMPLE_ADDR bit and
> perf_event_attr.mmap_data flag. Separate them using new --data-mmap
> option to support recording only one of them.
>
> For data-type profiling, data MMAP is unnecessary but it wastes a lot
> of space in the ring buffer and data file.
>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Thanks,
Ian
> ---
> tools/perf/Documentation/perf-record.txt | 8 +++++++-
> tools/perf/builtin-record.c | 19 +++++++++++++------
> tools/perf/util/evsel.c | 5 +++--
> tools/perf/util/record.h | 2 ++
> 4 files changed, 25 insertions(+), 9 deletions(-)
>
> diff --git a/tools/perf/Documentation/perf-record.txt b/tools/perf/Documentation/perf-record.txt
> index e8b9aadbbfa50574..c402e74172f6a22d 100644
> --- a/tools/perf/Documentation/perf-record.txt
> +++ b/tools/perf/Documentation/perf-record.txt
> @@ -344,7 +344,8 @@ OPTIONS
>
> -d::
> --data::
> - Record the sample virtual addresses. Implies --sample-mem-info.
> + Record the sample virtual addresses. Implies --sample-mem-info and
> + --data-mmap.
>
> --phys-data::
> Record the sample physical addresses.
> @@ -861,6 +862,11 @@ filtered through the mask provided by -C option.
> Prepare BPF filter to be used by regular users. The action should be
> either "pin" or "unpin". The filter can be used after it's pinned.
>
> +--data-mmap::
> + Enable recording MMAP events for non-executable mappings. Basically
> + perf only records executable mappings but data mmaping can be useful
> + when you analyze data access with sample addresses. So using -d option
> + would enable this unless you specify --no-data-mmap manually.
>
> include::intel-hybrid.txt[]
>
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index 2584d0d8bc820676..cbfbd9bb10634093 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -1881,7 +1881,7 @@ static int record__synthesize_workload(struct record *rec, bool tail)
> process_synthesized_event,
> &rec->session->machines.host,
> needs_mmap,
> - rec->opts.sample_address);
> + rec->opts.record_data_mmap);
> perf_thread_map__put(thread_map);
> return err;
> }
> @@ -2191,7 +2191,7 @@ static int record__synthesize(struct record *rec, bool tail)
>
> err = __machine__synthesize_threads(machine, tool, &opts->target,
> rec->evlist->core.threads,
> - f, needs_mmap, opts->sample_address,
> + f, needs_mmap, opts->record_data_mmap,
> rec->opts.nr_threads_synthesize);
> }
>
> @@ -3006,8 +3006,9 @@ int record_opts__parse_callchain(struct record_opts *record,
> ret = parse_callchain_record_opt(arg, callchain);
> if (!ret) {
> /* Enable data address sampling for DWARF unwind. */
> - if (callchain->record_mode == CALLCHAIN_DWARF)
> - record->sample_address = true;
> + if (callchain->record_mode == CALLCHAIN_DWARF &&
> + !record->record_data_mmap_set)
> + record->record_data_mmap = true;
> callchain_debug(callchain);
> }
>
> @@ -3686,6 +3687,9 @@ static struct option __record_options[] = {
> OPT_CALLBACK(0, "off-cpu-thresh", &record.opts, "ms",
> "Dump off-cpu samples if off-cpu time exceeds this threshold (in milliseconds). (Default: 500ms)",
> record__parse_off_cpu_thresh),
> + OPT_BOOLEAN_SET(0, "data-mmap", &record.opts.record_data_mmap,
> + &record.opts.record_data_mmap_set,
> + "Record mmap events for non-executable mappings"),
> OPT_END()
> };
>
> @@ -4249,9 +4253,12 @@ int cmd_record(int argc, const char **argv)
> goto out_opts;
> }
>
> - /* For backward compatibility, -d implies --mem-info */
> - if (rec->opts.sample_address)
> + /* For backward compatibility, -d implies --mem-info and --data-mmap */
> + if (rec->opts.sample_address) {
> rec->opts.sample_data_src = true;
> + if (!rec->opts.record_data_mmap_set)
> + rec->opts.record_data_mmap = true;
> + }
>
> /*
> * Allow aliases to facilitate the lookup of symbols for address
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index 9cd706f6279313c2..ec6552a6f667fec6 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -1445,10 +1445,11 @@ void evsel__config(struct evsel *evsel, struct record_opts *opts,
> attr->inherit_stat = 1;
> }
>
> - if (opts->sample_address) {
> + if (opts->sample_address)
> evsel__set_sample_bit(evsel, ADDR);
> +
> + if (opts->record_data_mmap)
> attr->mmap_data = track;
> - }
>
> /*
> * We don't allow user space callchains for function trace
> diff --git a/tools/perf/util/record.h b/tools/perf/util/record.h
> index ea3a6c4657eefb74..93627c9a73387ddd 100644
> --- a/tools/perf/util/record.h
> +++ b/tools/perf/util/record.h
> @@ -40,6 +40,8 @@ struct record_opts {
> bool record_cgroup;
> bool record_switch_events;
> bool record_switch_events_set;
> + bool record_data_mmap;
> + bool record_data_mmap_set;
> bool all_kernel;
> bool all_user;
> bool kernel_callchains;
> --
> 2.52.0.223.gf5cc29aaa4-goog
>
^ permalink raw reply [flat|nested] 9+ messages in thread* Re: [PATCH 1/4] perf record: Split --data-mmap option
2025-12-10 22:05 ` [PATCH 1/4] perf record: Split --data-mmap option Ian Rogers
@ 2025-12-17 12:37 ` Arnaldo Carvalho de Melo
0 siblings, 0 replies; 9+ messages in thread
From: Arnaldo Carvalho de Melo @ 2025-12-17 12:37 UTC (permalink / raw)
To: Ian Rogers
Cc: Namhyung Kim, James Clark, Jiri Olsa, Adrian Hunter,
Peter Zijlstra, Ingo Molnar, LKML, linux-perf-users
On Wed, Dec 10, 2025 at 02:05:23PM -0800, Ian Rogers wrote:
> On Tue, Dec 9, 2025 at 6:33 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > Currently -d/--data option controls both PERF_SAMPLE_ADDR bit and
> > perf_event_attr.mmap_data flag. Separate them using new --data-mmap
> > option to support recording only one of them.
> >
> > For data-type profiling, data MMAP is unnecessary but it wastes a lot
> > of space in the ring buffer and data file.
> >
> > Signed-off-by: Namhyung Kim <namhyung@kernel.org>
>
> Reviewed-by: Ian Rogers <irogers@google.com>
Tested the series and applied to perf-tools-next (6.20),
Thanks,
- Arnaldo
^ permalink raw reply [flat|nested] 9+ messages in thread