Skip to contentSkip to conversation

perf: fix symbol processing bug and greatly improve performance

PATCH Superseded Waiman Long 4 participants 13 messages May 6, 2013 – May 9, 2013 perf events

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf: fix symbol processing bug and greatly improve performance

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 96.90    822.10   822.10   192156     0.00     0.00  dsos__find
  0.81    828.96     6.86 172089958     0.00     0.00  rb_next
  0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
              822.10    0.00  192156/192156      dsos__find [8]
                0.01    0.00  192156/192156      dsos__add [62]
                0.01    0.00  192156/192366      dso__new [61]
                0.00    0.00       1/45282525     memdup [31]
                0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
              822.10    0.00  192156/192156      vdso__dso_findnew [7]
[8]     96.9  822.10    0.00  192156         dsos__find [8]

-----------------------------------------------

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

To resolve this problem, the dsos__find() function interface was modified to enable searching either the long name or the short name. The vdso__dso_findnew() will now search only the short name while the other call sites search for the long name as before.

With this change, the cpu time of perf was reduced from 848.38s to 15.77s and dsos__find() only accounted for 0.06% of the total time.

0.06 15.73 0.01 192151 0.00 0.00 dsos__find

---
 tools/perf/util/dso.c  |   10 ++++++++--
 tools/perf/util/dso.h  |    2 +-
 tools/perf/util/vdso.c |    2 +-
 3 files changed, 10 insertions(+), 4 deletions(-)
tools/perf/util/dso.c +8 -2
index 6f7d5a9..3d80f92 100644
@@ -513,10 +513,16 @@ void dsos__add(struct list_head *head, struct dso *dso)
 	list_add_tail(&dso->node, head);
 }
 
-struct dso *dsos__find(struct list_head *head, const char *name)
+struct dso *dsos__find(struct list_head *head, const char *name, bool cmp_short)
 {
 	struct dso *pos;
 
+	if (cmp_short) {
+		list_for_each_entry(pos, head, node)
+			if (strcmp(pos->short_name, name) == 0)
+				return pos;
+		return NULL;
+	}
 	list_for_each_entry(pos, head, node)
 		if (strcmp(pos->long_name, name) == 0)
 			return pos;
@@ -525,7 +531,7 @@ struct dso *dsos__find(struct list_head *head, const char *name)
 
 struct dso *__dsos__findnew(struct list_head *head, const char *name)
 {
-	struct dso *dso = dsos__find(head, name);
+	struct dso *dso = dsos__find(head, name, FALSE);
 
 	if (!dso) {
 		dso = dso__new(name);
tools/perf/util/dso.h +1 -1
index 450199a..44c9fdd 100644
@@ -133,7 +133,7 @@ struct dso *dso__kernel_findnew(struct machine *machine, const char *name,
 				const char *short_name, int dso_type);
 
 void dsos__add(struct list_head *head, struct dso *dso);
-struct dso *dsos__find(struct list_head *head, const char *name);
+struct dso *dsos__find(struct list_head *head, const char *name, bool cmp_short);
 struct dso *__dsos__findnew(struct list_head *head, const char *name);
 bool __dsos__read_build_ids(struct list_head *head, bool with_hits);
 
tools/perf/util/vdso.c +1 -1
index e60951f..a8fd73d 100644
@@ -91,7 +91,7 @@ void vdso__exit(void)
 
 struct dso *vdso__dso_findnew(struct list_head *head)
 {
-	struct dso *dso = dsos__find(head, VDSO__MAP_NAME);
+	struct dso *dso = dsos__find(head, VDSO__MAP_NAME, TRUE);
 
 	if (!dso) {
 		char *file;
Signature · 1 line
1.7.1
#1 lore raw
Quoted 50 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 96.90    822.10   822.10   192156     0.00     0.00  dsos__find
  0.81    828.96     6.86 172089958     0.00     0.00  rb_next
  0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
              822.10    0.00  192156/192156      dsos__find [8]
                0.01    0.00  192156/192156      dsos__add [62]
                0.01    0.00  192156/192366      dso__new [61]
                0.00    0.00       1/45282525     memdup [31]
                0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
              822.10    0.00  192156/192156      vdso__dso_findnew [7]
[8]     96.9  822.10    0.00  192156         dsos__find [8]

-----------------------------------------------

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

To resolve this problem, the dsos__find() function interface was modified to enable searching either the long name or the short name. The vdso__dso_findnew() will now search only the short name while the other call sites search for the long name as before.

Quoted 4 lines

With this change, the cpu time of perf was reduced from 848.38s to 15.77s and dsos__find() only accounted for 0.06% of the total time.

0.06 15.73 0.01 192151 0.00 0.00 dsos__find

Very nice!

I noticed that you used gprof to instrument perf itself on a call graph level.

Does this method of profiling perf via perf:

perf record -g perf report
perf report

... produce similarly useful call-graph instrumentation for you?

If not or not quite then could you describe the differences? We could use that to further improve perf call-graph profiling.

Thanks,

Ingo

#2 lore raw
Quoted 63 lines · Ingo Molnar
Quoted 50 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

   %   cumulative   self              self     total
  time   seconds   seconds    calls   s/call   s/call  name
  96.90    822.10   822.10   192156     0.00     0.00  dsos__find
   0.81    828.96     6.86 172089958     0.00     0.00  rb_next
   0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                 0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
               822.10    0.00  192156/192156      dsos__find [8]
                 0.01    0.00  192156/192156      dsos__add [62]
                 0.01    0.00  192156/192366      dso__new [61]
                 0.00    0.00       1/45282525     memdup [31]
                 0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
               822.10    0.00  192156/192156      vdso__dso_findnew [7]

[8] 96.9 822.10 0.00 192156 dsos__find [8]
-----------------------------------------------

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

To resolve this problem, the dsos__find() function interface was modified to enable searching either the long name or the short name. The vdso__dso_findnew() will now search only the short name while the other call sites search for the long name as before.

Quoted 4 lines

With this change, the cpu time of perf was reduced from 848.38s to 15.77s and dsos__find() only accounted for 0.06% of the total time.

   0.06     15.73     0.01   192151     0.00     0.00  dsos__find

Very nice!

I noticed that you used gprof to instrument perf itself on a call graph level.

Does this method of profiling perf via perf:

Quoted 7 lines

perf record -g perf report
perf report

... produce similarly useful call-graph instrumentation for you?

If not or not quite then could you describe the differences? We could use that to further improve perf call-graph profiling.

Thank for the comment.

The slowdown that I was trying to fix was in the "perf record" part of the profiling process, not the "perf report" part. I didn't try perf-record on perf-record as the performance counters are limited resources and I don't want resource conflicts to affect the results. As the slow-down was entirely in the user space, I decided to use gprof to do the profiling.

One thing I noticed about the gprof instrumentation and perf call-graph was that inlined functions were properly identified by gprof, but not perf. This makes sense as the compiler produces additional instrumentation code and data into the binary that can be used by gprof. That is not the case for perf.

I also looked into possible performance bottleneck in the perf-report part of the process. Most of the time was spent in sorting and merging the call-chain data and there didn't seem to have low hanging fruit for performance tuning opportunity.

Please let me know what other tests or data would you like to me to gather.

Best regards,
Longman

#3 lore raw
Quoted 1 line · Waiman Long

The slowdown that I was trying to fix was in the "perf record" part of the profiling process, not the "perf report" part. I didn't try perf-record on perf-record as the performance counters are limited resources and I don't want resource conflicts to affect the results. As the slow-down was entirely in the user space, I decided to use gprof to do the profiling.

Just to clarify that the post-processing time that I referred to in the patch is the time spent after the workload finishes in the perf-record session and the perf command itself quits. I could take a really long time for perf to quit without the patch.

Regards,
Longman

#4 lore raw
In reply to Waiman Long
Quoted 72 lines · Waiman Long
Quoted 63 lines · Ingo Molnar
Quoted 50 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 96.90    822.10   822.10   192156     0.00     0.00  dsos__find
  0.81    828.96     6.86 172089958     0.00     0.00  rb_next
  0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
              822.10    0.00  192156/192156      dsos__find [8]
                0.01    0.00  192156/192156      dsos__add [62]
                0.01    0.00  192156/192366      dso__new [61]
                0.00    0.00       1/45282525     memdup [31]
                0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
              822.10    0.00  192156/192156      vdso__dso_findnew [7]
[8]     96.9  822.10    0.00  192156         dsos__find [8]

-----------------------------------------------

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

To resolve this problem, the dsos__find() function interface was modified to enable searching either the long name or the short name. The vdso__dso_findnew() will now search only the short name while the other call sites search for the long name as before.

Quoted 4 lines

With this change, the cpu time of perf was reduced from 848.38s to 15.77s and dsos__find() only accounted for 0.06% of the total time.

0.06 15.73 0.01 192151 0.00 0.00 dsos__find

Very nice!

I noticed that you used gprof to instrument perf itself on a call graph level.

Does this method of profiling perf via perf:

Quoted 7 lines

perf record -g perf report
perf report

... produce similarly useful call-graph instrumentation for you?

If not or not quite then could you describe the differences? We could use that to further improve perf call-graph profiling.

Quoted 6 lines

Thank for the comment.

The slowdown that I was trying to fix was in the "perf record" part of the profiling process, not the "perf report" part. I didn't try perf-record on perf-record as the performance counters are limited resources and I don't want resource conflicts to affect the results.

ah, ok. In general two instances of cycles-profiling should work just fine without any resource conflicts.

But yeah, with "perf record" that's a valid worry and I can see how you wanted to not worry about that.

Thanks,

Ingo

#5 lore raw
In reply to Waiman Long
Quoted 34 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 96.90    822.10   822.10   192156     0.00     0.00  dsos__find
  0.81    828.96     6.86 172089958     0.00     0.00  rb_next
  0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
              822.10    0.00  192156/192156      dsos__find [8]
                0.01    0.00  192156/192156      dsos__add [62]
                0.01    0.00  192156/192366      dso__new [61]
                0.00    0.00       1/45282525     memdup [31]
                0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
              822.10    0.00  192156/192156      vdso__dso_findnew [7]
[8]     96.9  822.10    0.00  192156         dsos__find [8]

-----------------------------------------------

Quoted 10 lines

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

hi,
the issue is there and fix looks ok, thanks!

though I'm not able to get vdso callchains to pop out even by investigating report with vdso heavy workload.

I'll have a closer look..

thanks,
jirka

#6 lore raw
Quoted 45 lines · Jiri Olsa
Quoted 34 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

   %   cumulative   self              self     total
  time   seconds   seconds    calls   s/call   s/call  name
  96.90    822.10   822.10   192156     0.00     0.00  dsos__find
   0.81    828.96     6.86 172089958     0.00     0.00  rb_next
   0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                 0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
               822.10    0.00  192156/192156      dsos__find [8]
                 0.01    0.00  192156/192156      dsos__add [62]
                 0.01    0.00  192156/192366      dso__new [61]
                 0.00    0.00       1/45282525     memdup [31]
                 0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
               822.10    0.00  192156/192156      vdso__dso_findnew [7]

[8] 96.9 822.10 0.00 192156 dsos__find [8]
-----------------------------------------------

Quoted 10 lines

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

Quoted 7 lines

hi,
the issue is there and fix looks ok, thanks!

though I'm not able to get vdso callchains to pop out even by investigating report with vdso heavy workload.

I'll have a closer look..

The test machine that I used have RHEL 6.4 installed in it with a upstream 3.9 kernel layered on top. The kernel config is based on the 6.4 configuration file with modification to enable the X2APIC option needed by the machine. Other than that, I didn't make too much modification to the base configuration. I used the "-a -s" option when running perf-record.

I don't think the vdso callchains were major part of the workload that I tested. I think it is the high number of CPU cores plus the high number of users (1500) that cause the performance bottleneck to surface. In a smaller machine, those bottlenecks may be much less noticeable. The vdso call-chain dominates the post-processsing time because of the need to search through the while DSO list for the vdso library which can grow to 2M+ in my test case.

Regards,
Longman

#7 lore raw
In reply to Jiri Olsa
Quoted 45 lines · Jiri Olsa
Quoted 34 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

   %   cumulative   self              self     total
  time   seconds   seconds    calls   s/call   s/call  name
  96.90    822.10   822.10   192156     0.00     0.00  dsos__find
   0.81    828.96     6.86 172089958     0.00     0.00  rb_next
   0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                 0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
               822.10    0.00  192156/192156      dsos__find [8]
                 0.01    0.00  192156/192156      dsos__add [62]
                 0.01    0.00  192156/192366      dso__new [61]
                 0.00    0.00       1/45282525     memdup [31]
                 0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
               822.10    0.00  192156/192156      vdso__dso_findnew [7]

[8] 96.9 822.10 0.00 192156 dsos__find [8]
-----------------------------------------------

Quoted 10 lines

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

Quoted 7 lines

hi,
the issue is there and fix looks ok, thanks!

though I'm not able to get vdso callchains to pop out even by investigating report with vdso heavy workload.

I'll have a closer look..

Is there a chance that the fix will go to v3.10 or have to wait for v3.11?

Regards,
Longman

Acked-by
#8 lore raw
Quoted 49 lines · Waiman Long
Quoted 48 lines · Jiri Olsa
Quoted 34 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 96.90    822.10   822.10   192156     0.00     0.00  dsos__find
  0.81    828.96     6.86 172089958     0.00     0.00  rb_next
  0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
              822.10    0.00  192156/192156      dsos__find [8]
                0.01    0.00  192156/192156      dsos__add [62]
                0.01    0.00  192156/192366      dso__new [61]
                0.00    0.00       1/45282525     memdup [31]
                0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
              822.10    0.00  192156/192156      vdso__dso_findnew [7]
[8]     96.9  822.10    0.00  192156         dsos__find [8]

-----------------------------------------------

Quoted 10 lines

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

hi,
the issue is there and fix looks ok, thanks!

Quoted 6 lines
Quoted 4 lines

though I'm not able to get vdso callchains to pop out even by investigating report with vdso heavy workload.

I'll have a closer look..

Is there a chance that the fix will go to v3.10 or have to wait for v3.11?

I got this from scripts/checkpatch.pl:

WARNING: line over 80 characters
#104: FILE: tools/perf/util/dso.h:136:
+struct dso *dsos__find(struct list_head *head, const char *name, bool cmp_short);

otherwise it looks ok:

thanks,
jirka

#9 lore raw
Quoted 55 lines · Jiri Olsa
Quoted 54 lines · Waiman Long
Quoted 52 lines · Jiri Olsa
Quoted 34 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

  %   cumulative   self              self     total
 time   seconds   seconds    calls   s/call   s/call  name
 96.90    822.10   822.10   192156     0.00     0.00  dsos__find
  0.81    828.96     6.86 172089958     0.00     0.00  rb_next
  0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
              822.10    0.00  192156/192156      dsos__find [8]
                0.01    0.00  192156/192156      dsos__add [62]
                0.01    0.00  192156/192366      dso__new [61]
                0.00    0.00       1/45282525     memdup [31]
                0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
              822.10    0.00  192156/192156      vdso__dso_findnew [7]
[8]     96.9  822.10    0.00  192156         dsos__find [8]

-----------------------------------------------

Quoted 10 lines

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

hi,
the issue is there and fix looks ok, thanks!

though I'm not able to get vdso callchains to pop out even by investigating report with vdso heavy workload.

I'll have a closer look..

Quoted 8 lines
Quoted 1 line

Is there a chance that the fix will go to v3.10 or have to wait for v3.11?

I got this from scripts/checkpatch.pl:

WARNING: line over 80 characters
#104: FILE: tools/perf/util/dso.h:136:
+struct dso *dsos__find(struct list_head *head, const char *name, bool cmp_short);

For a prototype line that's OK I think.

Thanks,

Ingo

#10 lore raw
In reply to Jiri Olsa
Quoted 57 lines · Jiri Olsa
Quoted 46 lines · Waiman Long
Quoted 45 lines · Jiri Olsa
Quoted 44 lines · Waiman Long

When "perf record" was used on a large machine with a lot of CPUs, the perf post-processing time could take a lot of minutes and even hours depending on how large the resulting perf.data file was.

While running AIM7 1500-user high_systime workload on a 80-core x86-64 system with a 3.9 kernel, the workload itself took about 2 minutes to run and the perf.data file had a size of 1108.746 MB. However, the post-processing step took more than 10 minutes.

With a gprof-profiled perf binary, the time spent by perf was as follows:

   %   cumulative   self              self     total
  time   seconds   seconds    calls   s/call   s/call  name
  96.90    822.10   822.10   192156     0.00     0.00  dsos__find
   0.81    828.96     6.86 172089958     0.00     0.00  rb_next
   0.41    832.44     3.48 48539289     0.00     0.00  rb_erase

So 97% (822 seconds) of the time was spent in a single dsos_find() function. After analyzing the call-graph data below:

-----------------------------------------------

                 0.00  822.12  192156/192156      map__new [6]
[7]     96.9    0.00  822.12  192156         vdso__dso_findnew [7]
               822.10    0.00  192156/192156      dsos__find [8]
                 0.01    0.00  192156/192156      dsos__add [62]
                 0.01    0.00  192156/192366      dso__new [61]
                 0.00    0.00       1/45282525     memdup [31]
                 0.00    0.00  192156/192230      dso__set_long_name [91]
-----------------------------------------------
               822.10    0.00  192156/192156      vdso__dso_findnew [7]

[8] 96.9 822.10 0.00 192156 dsos__find [8]
-----------------------------------------------

It was found that the vdso__dso_findnew() function failed to locate VDSO__MAP_NAME ("[vdso]") in the dso list and have to insert a new entry at the end for 192156 times. This problem is due to the fact that there are 2 types of name in the dso entry - short name and long name. The initial dso__new() adds "[vdso]" to both the short and long names. After that, vdso__dso_findnew() modifies the long name to something like /tmp/perf-vdso.so-NoXkDj. The dsos__find() function only compares the long name. As a result, the same vdso entry is duplicated many time in the dso list. This bug increases memory consumption as well as slows the symbol processing time to a crawl.

Quoted 8 lines
Quoted 7 lines

hi,
the issue is there and fix looks ok, thanks!

though I'm not able to get vdso callchains to pop out even by investigating report with vdso heavy workload.

I'll have a closer look..

Is there a chance that the fix will go to v3.10 or have to wait for v3.11?

I got this from scripts/checkpatch.pl:

Quoted 8 lines

WARNING: line over 80 characters
#104: FILE: tools/perf/util/dso.h:136:
+struct dso *dsos__find(struct list_head *head, const char *name, bool cmp_short);

otherwise it looks ok:

I am sorry that I forgot to rerun checkpatch.pl again after changing the argument type from int to bool. So I missed this warning.

I had resent an updated patch with the warning fixed. I also made some update to the patch description.

Regards,
Longman

#11 lore raw
Quoted 2 lines · Waiman Long

I am sorry that I forgot to rerun checkpatch.pl again after changing the argument type from int to bool.

You can edit/create .git/hooks/pre-commit and have it run checkpatch.pl so you don't forget:

#!/bin/bash
exec git diff --cached | scripts/checkpatch.pl --no-signoff -q -

If it spits out warnings that are ok, add -n to git-commit to bypass.

David

#12 lore raw
Quoted 10 lines · David Ahern
Quoted 2 lines · Waiman Long

I am sorry that I forgot to rerun checkpatch.pl again after changing the argument type from int to bool.

You can edit/create .git/hooks/pre-commit and have it run checkpatch.pl so you don't forget:

#!/bin/bash
exec git diff --cached | scripts/checkpatch.pl --no-signoff -q -

If it spits out warnings that are ok, add -n to git-commit to bypass.

Thanks for the tip! I will implement that in my development environment.

Regards,
Longman