[PATCH] perf tools: Synthesize anon MMAP records on the heap

[Gaurav Jain]

Anon records usually do not have the 'execname' entry. However if they are on
the heap, the execname shows up as '[heap]'. The fix considers any executable
entries in the map that do not have a name or are on the heap as anon records
and sets the name to '//anon'.

This fixes JIT profiling for records on the heap.

Signed-off-by: Gaurav Jain <gj...@fb.com>
Cc: Ingo Molnar <mi...@kernel.org>
Cc: Jiri Olsa <jo...@redhat.com>
Cc: Paul Mackerras <pau...@samba.org>
Cc: Peter Zijlstra <a.p.zi...@chello.nl>
Cc: Don Zickus <dzi...@redhat.com>
Cc: Arun Sharma <ash...@fb.com>
---
tools/perf/util/event.c | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/tools/perf/util/event.c b/tools/perf/util/event.c
index bb788c1..ae9c55b 100644
--- a/tools/perf/util/event.c
+++ b/tools/perf/util/event.c
@@ -224,10 +224,9 @@ static int perf_event__synthesize_mmap_events(struct perf_tool *tool,
continue;

event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
- }
-
- if (!strcmp(execname, ""))
+ } if (!strcmp(execname, "") || !strcmp(execname, "[heap]")) {
strcpy(execname, anonstr);
+ }

size = strlen(execname) + 1;
memcpy(event->mmap.filename, execname, size);
--
1.8.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majo...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

[Namhyung Kim]

Hi Gaurav,

On Sat, 11 Jan 2014 20:32:14 -0800, Gaurav Jain wrote:
> Anon records usually do not have the 'execname' entry. However if they are on
> the heap, the execname shows up as '[heap]'. The fix considers any executable
> entries in the map that do not have a name or are on the heap as anon records
> and sets the name to '//anon'.
>
> This fixes JIT profiling for records on the heap.

Acked-by: Namhyung Kim <namh...@kernel.org>

Thanks,
Namhyung

[Don Zickus]

On Sat, Jan 11, 2014 at 08:32:14PM -0800, Gaurav Jain wrote:
> Anon records usually do not have the 'execname' entry. However if they are on
> the heap, the execname shows up as '[heap]'. The fix considers any executable
> entries in the map that do not have a name or are on the heap as anon records
> and sets the name to '//anon'.
>
> This fixes JIT profiling for records on the heap.

I guess I don't understand the need for this fix. It seems breaking out
//anon vs. [heap] would be useful. Your patch is saying otherwise. Can
give a description of the problem you are trying to solve?

Also style issue below..

>
> Signed-off-by: Gaurav Jain <gj...@fb.com>
> Cc: Ingo Molnar <mi...@kernel.org>
> Cc: Jiri Olsa <jo...@redhat.com>
> Cc: Paul Mackerras <pau...@samba.org>
> Cc: Peter Zijlstra <a.p.zi...@chello.nl>
> Cc: Don Zickus <dzi...@redhat.com>
> Cc: Arun Sharma <ash...@fb.com>
> ---
> tools/perf/util/event.c | 5 ++---
> 1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/tools/perf/util/event.c b/tools/perf/util/event.c
> index bb788c1..ae9c55b 100644
> --- a/tools/perf/util/event.c
> +++ b/tools/perf/util/event.c
> @@ -224,10 +224,9 @@ static int perf_event__synthesize_mmap_events(struct perf_tool *tool,
> continue;
>
> event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
> - }
> -
> - if (!strcmp(execname, ""))
> + } if (!strcmp(execname, "") || !strcmp(execname, "[heap]")) {

The '} if' part should seperate the 'if' on to its own line (unless you
meant an 'else' in there).

Cheers,
Don

[Gaurav Jain]

On 1/13/14, 11:54 AM, "Don Zickus" <dzi...@redhat.com> wrote:

>On Sat, Jan 11, 2014 at 08:32:14PM -0800, Gaurav Jain wrote:
>> Anon records usually do not have the 'execname' entry. However if they
>>are on
>> the heap, the execname shows up as '[heap]'. The fix considers any
>>executable
>> entries in the map that do not have a name or are on the heap as anon
>>records
>> and sets the name to '//anon'.
>>
>> This fixes JIT profiling for records on the heap.
>
>I guess I don't understand the need for this fix. It seems breaking out
>//anon vs. [heap] would be useful. Your patch is saying otherwise. Can
>give a description of the problem you are trying to solve?

Thank you for looking at the patch.

We generate a perf map file which includes certain JIT�ed functions that
show up as [heap] entries. As a result, I included the executable heap
entries as anon pages so that it would be handled in
util/map.c:map__new(). The alternative would be to handle heap entries in
map__new() directly, however I wasn�t sure if this would break something
as it seems that heap and stack entries are expected to fail all
map__find_* functions. Thus I considered executable heap entries as
//anon, but perhaps there is a better way.

Bah yes, I intended 'else if'. I apologize for that. Does the fact that I
filtered anon entries to only those marked as executable break the
existing behavior?

Thanks,

Gaurav

[Namhyung Kim]

Hi Gaurav,

I'd like to take my ack back - it seems I missed some points.

On Tue, 14 Jan 2014 20:48:23 +0000, Gaurav Jain wrote:
> On 1/13/14, 11:54 AM, "Don Zickus" <dzi...@redhat.com> wrote:
>
>>On Sat, Jan 11, 2014 at 08:32:14PM -0800, Gaurav Jain wrote:
>>> Anon records usually do not have the 'execname' entry. However if they
>>>are on
>>> the heap, the execname shows up as '[heap]'. The fix considers any
>>>executable
>>> entries in the map that do not have a name or are on the heap as anon
>>>records
>>> and sets the name to '//anon'.
>>>
>>> This fixes JIT profiling for records on the heap.
>>
>>I guess I don't understand the need for this fix. It seems breaking out
>>//anon vs. [heap] would be useful. Your patch is saying otherwise. Can
>>give a description of the problem you are trying to solve?
>
> Thank you for looking at the patch.
>
> We generate a perf map file which includes certain JIT¹ed functions that
> show up as [heap] entries. As a result, I included the executable heap
> entries as anon pages so that it would be handled in
> util/map.c:map__new(). The alternative would be to handle heap entries in
> map__new() directly, however I wasn¹t sure if this would break something
> as it seems that heap and stack entries are expected to fail all
> map__find_* functions. Thus I considered executable heap entries as
> //anon, but perhaps there is a better way.

Hmm.. so the point is that an executable heap mapping should have
/tmp/perf-XXX.map as a file name, right? If so, does something like
below work well for you?

Thanks,
Namhyung


diff --git a/tools/perf/util/map.c b/tools/perf/util/map.c
index 9b9bd719aa19..d52387fe83f1 100644
--- a/tools/perf/util/map.c
+++ b/tools/perf/util/map.c
@@ -69,7 +69,7 @@ struct map *map__new(struct list_head *dsos__list, u64 start, u64 len,
map->ino = ino;
map->ino_generation = ino_gen;

- if (anon) {
+ if (anon || (no_dso && type == MAP__FUNCTION)) {
snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", pid);
filename = newfilename;
}
@@ -93,7 +93,7 @@ struct map *map__new(struct list_head *dsos__list, u64 start, u64 len,
* functions still return NULL, and we avoid the
* unnecessary map__load warning.
*/
- if (no_dso)
+ if (no_dso && type != MAP__FUNCTION)
dso__set_loaded(dso, map->type);

[Gaurav Jain]

Hi Namhyung,

On 1/15/14, 12:46 AM, "Namhyung Kim" <namh...@kernel.org> wrote:

>I'd like to take my ack back - it seems I missed some points.

No worries, looks like the patch wasn’t well thought out.

Just gave it a try and it fixed the issue perfectly! Thanks for the help.
This looks like a much better solution than treating the heap mapping as
an anon record.

Gaurav

[Don Zickus]

On Tue, Jan 14, 2014 at 08:48:23PM +0000, Gaurav Jain wrote:
> On 1/13/14, 11:54 AM, "Don Zickus" <dzi...@redhat.com> wrote:
>
> >On Sat, Jan 11, 2014 at 08:32:14PM -0800, Gaurav Jain wrote:
> >> Anon records usually do not have the 'execname' entry. However if they
> >>are on
> >> the heap, the execname shows up as '[heap]'. The fix considers any
> >>executable
> >> entries in the map that do not have a name or are on the heap as anon
> >>records
> >> and sets the name to '//anon'.
> >>
> >> This fixes JIT profiling for records on the heap.
> >
> >I guess I don't understand the need for this fix. It seems breaking out
> >//anon vs. [heap] would be useful. Your patch is saying otherwise. Can
> >give a description of the problem you are trying to solve?
>
> Thank you for looking at the patch.
>
> We generate a perf map file which includes certain JIT�ed functions that
> show up as [heap] entries. As a result, I included the executable heap
> entries as anon pages so that it would be handled in
> util/map.c:map__new(). The alternative would be to handle heap entries in
> map__new() directly, however I wasn�t sure if this would break something
> as it seems that heap and stack entries are expected to fail all
> map__find_* functions. Thus I considered executable heap entries as
> //anon, but perhaps there is a better way.

Thanks for the improved problem description. I see it led to a better
patch. :-) That is why it is generally a good idea to describe the
problem you are trying to solve to see if others have a better solution.

Cheers,
Don

[Namhyung Kim]

Hi Don,

Yes, thank you very much for pointing it out and helping to resolve this
issue!

Thanks,
Namhyung

[Namhyung Kim]

Hi Gaurav,

On Wed, 15 Jan 2014 06:44:39 +0000, Gaurav Jain wrote:
> Hi Namhyung,
>
>
> On 1/15/14, 12:46 AM, "Namhyung Kim" <namh...@kernel.org> wrote:
>>Hmm.. so the point is that an executable heap mapping should have
>>/tmp/perf-XXX.map as a file name, right? If so, does something like
>>below work well for you?
>
> Just gave it a try and it fixed the issue perfectly! Thanks for the help.
> This looks like a much better solution than treating the heap mapping as
> an anon record.

Thanks for testing! It's good to hear it solved your problem. :)

>
> Gaurav
>
>>diff --git a/tools/perf/util/map.c b/tools/perf/util/map.c
>>index 9b9bd719aa19..d52387fe83f1 100644
>>--- a/tools/perf/util/map.c
>>+++ b/tools/perf/util/map.c
>>@@ -69,7 +69,7 @@ struct map *map__new(struct list_head *dsos__list, u64
>>start, u64 len,
>> map->ino = ino;
>> map->ino_generation = ino_gen;
>>
>>- if (anon) {
>>+ if (anon || (no_dso && type == MAP__FUNCTION)) {

Hmm.. I think it should check type of anon mapping too (assuming JIT
interface only provides function symbols, no?):

if ((anon || no_dso) && type == MAP__FUNCTION)) {

>> snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", pid);
>> filename = newfilename;
>> }
>>@@ -93,7 +93,7 @@ struct map *map__new(struct list_head *dsos__list, u64
>>start, u64 len,
>> * functions still return NULL, and we avoid the
>> * unnecessary map__load warning.
>> */
>>- if (no_dso)
>>+ if (no_dso && type != MAP__FUNCTION)

And it should be simply:

if (type != MAP__FUNCTION)


>> dso__set_loaded(dso, map->type);
>> }
>> }

I'll update this and send a formal patch soon. Could you please test it with
the new patch then?

Thanks,
Namhyung

[Gaurav Jain]

Hi Namhyung,

> On Jan 15, 2014, at 8:01 PM, "Namhyung Kim" <namh...@kernel.org> wrote:
>
> Hi Gaurav,
>
>> On Wed, 15 Jan 2014 06:44:39 +0000, Gaurav Jain wrote:
>> Hi Namhyung,
>>
>>
>>> On 1/15/14, 12:46 AM, "Namhyung Kim" <namh...@kernel.org> wrote:
>>> diff --git a/tools/perf/util/map.c b/tools/perf/util/map.c
>>> index 9b9bd719aa19..d52387fe83f1 100644
>>> --- a/tools/perf/util/map.c
>>> +++ b/tools/perf/util/map.c
>>> @@ -69,7 +69,7 @@ struct map *map__new(struct list_head *dsos__list, u64
>>> start, u64 len,
>>> map->ino = ino;
>>> map->ino_generation = ino_gen;
>>>
>>> - if (anon) {
>>> + if (anon || (no_dso && type == MAP__FUNCTION)) {
>
> Hmm.. I think it should check type of anon mapping too (assuming JIT
> interface only provides function symbols, no?):
>
> if ((anon || no_dso) && type == MAP__FUNCTION)) {

In that case, shouldn't we filter out all mappings that are not executable as I had intended to do in my original patch?

>>> snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", pid);
>>> filename = newfilename;
>>> }
>>> @@ -93,7 +93,7 @@ struct map *map__new(struct list_head *dsos__list, u64
>>> start, u64 len,
>>> * functions still return NULL, and we avoid the
>>> * unnecessary map__load warning.
>>> */
>>> - if (no_dso)
>>> + if (no_dso && type != MAP__FUNCTION)
>
> And it should be simply:
>
> if (type != MAP__FUNCTION)
>
>
>>> dso__set_loaded(dso, map->type);
>>> }
>>> }
>
> I'll update this and send a formal patch soon. Could you please test it with
> the new patch then?

Sure I'll test the new patch, though please note my previous suggestion.

Gaurav--