Skip to contentSkip to conversation

perf tools: Synthesize anon MMAP records on the heap

PATCH In review Gaurav Jain 3 participants 10 messages Jan 12, 2014 – Jan 16, 2014 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf tools: Synthesize anon MMAP records on the heap

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.

Cc Ingo Molnar <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Don Zickus <[email protected]>
Cc Arun Sharma <[email protected]>
---
 tools/perf/util/event.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
tools/perf/util/event.c +2 -3
index bb788c1..ae9c55b 100644
@@ -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);
Signature · 1 line
1.8.1
maintainer Acked-by on v1 1/1
#1 lore raw

Hi Gaurav,

Quoted 6 lines · 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.

Thanks,
Namhyung

#2 lore raw
In reply to Gaurav Jain
Quoted 6 lines · 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.

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..

Quoted 17 lines
Cc Ingo Molnar <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Don Zickus <[email protected]>
Cc Arun Sharma <[email protected]>
---
 tools/perf/util/event.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
tools/perf/util/event.c +0 -0
index bb788c1..ae9c55b 100644

tools/perf/util/event.c static int perf_event__synthesize_mmap_events(struct perf_tool *tool, · line 227

Quoted 7 lines
tools/perf/util/event.c +1 -3
@@ -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

#3 lore raw
Quoted 9 lines · Don Zickus
Quoted 8 lines · 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'.

Quoted 5 lines
Quoted 1 line

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.

Quoted 22 lines

Also style issue below..

Quoted 20 lines
Cc Ingo Molnar <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Don Zickus <[email protected]>
Cc Arun Sharma <[email protected]>
---
 tools/perf/util/event.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
tools/perf/util/event.c +0 -0
index bb788c1..ae9c55b 100644
@@ -224,10 +224,9 @@ static int
 perf_event__synthesize_mmap_events(struct perf_tool *tool,
 				continue;
 
Quoted 8 lines

tools/perf/util/event.c static int perf_event__synthesize_mmap_events(struct perf_tool *tool, · line 227

Quoted 5 lines
tools/perf/util/event.c +1 -3
 			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).

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

maintainer
#4 lore raw

Hi Gaurav,

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

Quoted 19 lines · Gaurav Jain
Quoted 9 lines · Don Zickus
Quoted 8 lines · 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'.

Quoted 5 lines
Quoted 1 line

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.

Quoted 8 lines

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

tools/perf/util/map.c +2 -2
index 9b9bd719aa19..d52387fe83f1 100644
@@ -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);
 		}
 	}
#5 lore raw

Hi Namhyung,

Quoted 1 line · Namhyung Kim

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.

Quoted 30 lines · Gaurav Jain
Quoted 19 lines · Don Zickus
Quoted 10 lines · Gaurav Jain
Quoted 8 lines

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'.

Quoted 5 lines
Quoted 1 line

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.

Quoted 9 lines

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.

Quoted 3 lines

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.

Gaurav

Quoted 23 lines
tools/perf/util/map.c +1 -1
index 9b9bd719aa19..d52387fe83f1 100644
@@ -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;
		}
tools/perf/util/map.c +1 -1
@@ -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);
 		}
Quoted 1 line
	}

�淸�{.n�+돴윯돪�+%듚�깁負쪐w��{.n�+돴쪐{깸G쳺��{ay����,j�f"톒쉹��묎�z_溫(�쉸듶줷"얎�m����G쳺�◀?솳鈺�&x�~뤳턨O뺞촾퇉v�^�m��� �띠�◀?뻀�

maintainer
#6 lore raw

Hi Gaurav,

Quoted 11 lines · Gaurav Jain

Hi Namhyung,

Quoted 3 lines · Namhyung Kim

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. :)

Quoted 12 lines

Gaurav

Quoted 9 lines
tools/perf/util/map.c +0 -0
index 9b9bd719aa19..d52387fe83f1 100644
@@ -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;
 
Quoted 2 lines

tools/perf/util/map.c struct map *map__new(struct list_head *dsos__list, u64 start, u64 len, · line 72

Quoted 2 lines
tools/perf/util/map.c +1 -1
-		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)) {
Quoted 10 lines

tools/perf/util/map.c struct map *map__new(struct list_head *dsos__list, u64 · line 97

Quoted 10 lines
tools/perf/util/map.c +0 -0
 			snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", pid);
 			filename = newfilename;
		}
tools/perf/util/map.c +1 -1
@@ -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)
Quoted 3 lines
Quoted 3 lines
tools/perf/util/map.c +0 -0
 				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

#7 lore raw

Hi Namhyung,

Quoted 18 lines · Namhyung Kim

Hi Gaurav,

Quoted 14 lines · Gaurav Jain

Hi Namhyung,

Quoted 10 lines · Namhyung Kim
tools/perf/util/map.c +0 -0
index 9b9bd719aa19..d52387fe83f1 100644
@@ -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;
 
Quoted 7 lines
Quoted 2 lines

tools/perf/util/map.c struct map *map__new(struct list_head *dsos__list, u64 start, u64 len, · line 72

Quoted 2 lines
tools/perf/util/map.c +1 -1
-        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?

Quoted 16 lines
Quoted 10 lines

tools/perf/util/map.c struct map *map__new(struct list_head *dsos__list, u64 · line 97

Quoted 10 lines
tools/perf/util/map.c +0 -0
           snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", pid);
           filename = newfilename;
       }
tools/perf/util/map.c +1 -1
@@ -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)
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
tools/perf/util/map.c +0 -0
               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

#8 lore raw
In reply to Gaurav Jain
Quoted 19 lines · Gaurav Jain
Quoted 9 lines · Don Zickus
Quoted 8 lines · 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'.

Quoted 5 lines
Quoted 1 line

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.

Quoted 8 lines

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.

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

Quoted 40 lines
Quoted 23 lines

Also style issue below..

Quoted 20 lines
Cc Ingo Molnar <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Paul Mackerras <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Don Zickus <[email protected]>
Cc Arun Sharma <[email protected]>
---
 tools/perf/util/event.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
tools/perf/util/event.c +0 -0
index bb788c1..ae9c55b 100644
@@ -224,10 +224,9 @@ static int
 perf_event__synthesize_mmap_events(struct perf_tool *tool,
 				continue;
 
Quoted 8 lines

tools/perf/util/event.c static int perf_event__synthesize_mmap_events(struct perf_tool *tool, · line 227

Quoted 5 lines
tools/perf/util/event.c +1 -3
 			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).

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

maintainer
#9 lore raw

Hi Don,

Quoted 29 lines · Don Zickus
Quoted 19 lines · Gaurav Jain
Quoted 9 lines · Don Zickus
Quoted 8 lines · 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'.

Quoted 5 lines
Quoted 1 line

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.

Quoted 8 lines

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.

Quoted 3 lines

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.

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

Thanks,
Namhyung