Skip to contentSkip to conversation

perf data convert json: Fix trace_seq memory leak in process_sample_event()

PATCH Applied Tanushree Shah 4 participants 11 messages Jun 4 – Jul 8 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf data convert json: Fix trace_seq memory leak in process_sample_event()

Unlike the in-kernel trace_seq which uses a statically allocated buffer, the userspace traceevent library's trace_seq uses a dynamically allocated one. Therefore, every trace_seq_init() call must be paired with a trace_seq_destroy(), otherwise it produces a memory leak.

In process_sample_event(), a trace_seq is initialized for each field when formatting tracepoint raw_data, but the matching trace_seq_destroy() is never called, leaking memory for every field of every sample processed.

Add the missing trace_seq_destroy() after using the trace_seq buffer to properly free the allocated memory.

Detected with Valgrind on a perf.data file with 2,729 tracepoint samples:
Before: definitely lost: 55,537,664 bytes in 13,559 blocks
After: definitely lost: 0 bytes in 0 blocks

Fixes 9d895e468429 ("perf data: Add tracepoint fields when converting to JSON")
---
 tools/perf/util/data-convert-json.c | 1 +
 1 file changed, 1 insertion(+)
tools/perf/util/data-convert-json.c +1 -0
index d526c91312ed..6dd9349ab10f 100644
@@ -257,6 +257,7 @@ static int process_sample_event(const struct perf_tool *tool,
 				trace_seq_init(&s);
 				tep_print_field(&s, sample->raw_data, fields[i]);
 				output_json_key_string(out, true, 3, fields[i]->name, s.buffer);
+				trace_seq_destroy(&s);
 
 				i++;
 			}
Signature · 1 line
2.47.3
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] The process_sample_event() function leaks reference counts and memory associated with struct addr_location when a sample is skipped due to time filtering. --

commit de6e4710063f2942dd58ebe436d7d992872132b3
Author: Tanushree Shah <[email protected]>

perf data convert json: Fix trace_seq memory leak in process_sample_event()

This commit fixes a memory leak by ensuring trace_seq_destroy() is called to pair with trace_seq_init() when formatting tracepoint raw data.

tools/perf/util/data-convert-json.c static int process_sample_event(const struct perf_tool *tool, · line 263

Quoted 12 lines
tools/perf/util/data-convert-json.c +1 -0
index d526c91312ed8..6dd9349ab10f1 100644
@@ -257,6 +257,7 @@ static int process_sample_event(const struct perf_tool *tool,
 				trace_seq_init(&s);
 				tep_print_field(&s, sample->raw_data, fields[i]);
 				output_json_key_string(out, true, 3, fields[i]->name, s.buffer);
+				trace_seq_destroy(&s);
 
 				i++;
 			}

[Severity: High]
This is a pre-existing issue, but does the early return path for skipped samples higher up in process_sample_event() leak the addr_location reference handles?

tools/perf/util/data-convert-json.c:process_sample_event() {
        addr_location__init(&al);
        if (machine__resolve(machine, &al, sample) < 0) {
                pr_err("Sample resolution failed!\n");
                addr_location__exit(&al);
                return -1;
        }

        if (perf_time__ranges_skip_sample(c->ptime_range, c->range_num, sample->time)) {
                ++c->skipped;
                return 0;
        }
        ...
}

When machine__resolve() populates the addr_location structure, it acquires reference handles for objects like thread and map. The early return for time filtering bypasses the addr_location__exit() cleanup.

Could this cause reference counts and memory to leak when dropping skipped samples?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw

Hi Sashiko,

Thank you for the review and for spotting this!

Regarding the addr_location reference leak:

yes its correct. The early return path for time-filtered samples does leak addr_location references.

When machine__resolve() succeeds, it calls thread__find_map(), which acquires reference counts for thread, map, and maps objects and stores them in the addr_location structure. The early return bypasses addr_location__exit(), preventing these references from being released.

Validation with Valgrind:

I've validated the issue with Valgrind using a test case with 2,385 total samples (150 converted, 2,235 skipped by time filter):

Without the fix:

definitely lost:  8,336 bytes in 32 blocks
indirectly lost:  1,539,361 bytes in 21,307 blocks
possibly lost:    10,325,465 bytes in 146,283 blocks
Total leaked:     ~11.9 MB

With the fix:

definitely lost:  0 bytes in 0 blocks
indirectly lost:  0 bytes in 0 blocks
possibly lost:    0 bytes in 0 blocks
Total leaked:     0 bytes

I will send a v2 patch series that addresses both memory leaks:

The trace_seq_destroy() fix (original issue) The addr_location__exit() fix for the early return path

Thanks
Tanushree Shah

Quoted 46 lines · On 04/06/26 12:36, [email protected] wrote:

perf data convert json: Fix trace_seq memory leak in process_sample_event()

This commit fixes a memory leak by ensuring trace_seq_destroy() is called to pair with trace_seq_init() when formatting tracepoint raw data.

tools/perf/util/data-convert-json.c static int process_sample_event(const struct perf_tool *tool, · line 263

Quoted 12 lines
tools/perf/util/data-convert-json.c +1 -0
index d526c91312ed8..6dd9349ab10f1 100644
@@ -257,6 +257,7 @@ static int process_sample_event(const struct perf_tool *tool,
  				trace_seq_init(&s);
  				tep_print_field(&s, sample->raw_data, fields[i]);
  				output_json_key_string(out, true, 3, fields[i]->name, s.buffer);
+				trace_seq_destroy(&s);
 
  				i++;
  			}

[Severity: High]
This is a pre-existing issue, but does the early return path for skipped samples higher up in process_sample_event() leak the addr_location reference handles?

tools/perf/util/data-convert-json.c:process_sample_event() {
         addr_location__init(&al);
         if (machine__resolve(machine, &al, sample) < 0) {
                 pr_err("Sample resolution failed!\n");
                 addr_location__exit(&al);
                 return -1;
         }

         if (perf_time__ranges_skip_sample(c->ptime_range, c->range_num, sample->time)) {
                 ++c->skipped;
                 return 0;
         }
         ...
}

When machine__resolve() populates the addr_location structure, it acquires reference handles for objects like thread and map. The early return for time filtering bypasses the addr_location__exit() cleanup.

Could this cause reference counts and memory to leak when dropping skipped samples?

-- Sashiko AI review · https://sashiko.dev/#/
patchset/[email protected]?part=1

maintainer Applied
#3 lore raw
In reply to Tanushree Shah
Quoted 9 lines · Tanushree Shah

Unlike the in-kernel trace_seq which uses a statically allocated buffer, the userspace traceevent library's trace_seq uses a dynamically allocated one. Therefore, every trace_seq_init() call must be paired with a trace_seq_destroy(), otherwise it produces a memory leak.

In process_sample_event(), a trace_seq is initialized for each field when formatting tracepoint raw_data, but the matching trace_seq_destroy() is never called, leaking memory for every field of every sample processed.

Quoted 6 lines

Add the missing trace_seq_destroy() after using the trace_seq buffer to properly free the allocated memory.

Detected with Valgrind on a perf.data file with 2,729 tracepoint samples:
Before: definitely lost: 55,537,664 bytes in 13,559 blocks
After: definitely lost: 0 bytes in 0 blocks

Thanks, applied to perf-tools-next, for v7.2.

- Arnaldo

Quoted 20 lines
Fixes 9d895e468429 ("perf data: Add tracepoint fields when converting to JSON")
---
 tools/perf/util/data-convert-json.c | 1 +
 1 file changed, 1 insertion(+)
tools/perf/util/data-convert-json.c +1 -0
index d526c91312ed..6dd9349ab10f 100644
@@ -257,6 +257,7 @@ static int process_sample_event(const struct perf_tool *tool,
 				trace_seq_init(&s);
 				tep_print_field(&s, sample->raw_data, fields[i]);
 				output_json_key_string(out, true, 3, fields[i]->name, s.buffer);
+				trace_seq_destroy(&s);
 
 				i++;
 			}
Signature · 1 line
2.47.3
#4 lore raw

Hello Arnaldo

Thanks for checking the patch. I cannot see this patch applied on perf-tools-next yet.

https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/log/?h=perf-tools-next

Thanks.

Quoted 41 lines · Arnaldo Carvalho de Melo
Quoted 9 lines · Tanushree Shah

Unlike the in-kernel trace_seq which uses a statically allocated buffer, the userspace traceevent library's trace_seq uses a dynamically allocated one. Therefore, every trace_seq_init() call must be paired with a trace_seq_destroy(), otherwise it produces a memory leak.

In process_sample_event(), a trace_seq is initialized for each field when formatting tracepoint raw_data, but the matching trace_seq_destroy() is never called, leaking memory for every field of every sample processed.

Quoted 6 lines

Add the missing trace_seq_destroy() after using the trace_seq buffer to properly free the allocated memory.

Detected with Valgrind on a perf.data file with 2,729 tracepoint samples:
Before: definitely lost: 55,537,664 bytes in 13,559 blocks
After: definitely lost: 0 bytes in 0 blocks

Thanks, applied to perf-tools-next, for v7.2.

- Arnaldo

Quoted 20 lines
Fixes 9d895e468429 ("perf data: Add tracepoint fields when converting to JSON")
---
  tools/perf/util/data-convert-json.c | 1 +
  1 file changed, 1 insertion(+)
tools/perf/util/data-convert-json.c +1 -0
index d526c91312ed..6dd9349ab10f 100644
@@ -257,6 +257,7 @@ static int process_sample_event(const struct perf_tool *tool,
  				trace_seq_init(&s);
  				tep_print_field(&s, sample->raw_data, fields[i]);
  				output_json_key_string(out, true, 3, fields[i]->name, s.buffer);
+				trace_seq_destroy(&s);
 
  				i++;
  			}
Signature · 1 line
2.47.3
maintainer
#5 lore raw

Hello,

Quoted 4 lines · Tanushree Shah

Hello Arnaldo

Thanks for checking the patch. I cannot see this patch applied on perf-tools-next yet.

According to the other comment, I guess you're supposed to send v2. No?

Thanks,
Namhyung

#6 lore raw

Hi Namhyung,

Yes, I was planning to send a v2 based on the review comments to fix the pre-existing issue.

However, Arnaldo mentioned that the original patch had already been merged before I got to send a v2, so I addressed the pre-existing issue in a separate patch instead. The patch is available here:

https://lore.kernel.org/linux-perf-users/[email protected]/

That said, although Arnaldo mentioned that this patch had been merged, I don't see it in perf-tools-next yet either. I'll wait for him to clarify its current status.

Thanks,
Tanushree Shah

Quoted 13 lines · Namhyung Kim

Hello,

Quoted 4 lines · Tanushree Shah

Hello Arnaldo

Thanks for checking the patch. I cannot see this patch applied on perf-tools-next yet.

According to the other comment, I guess you're supposed to send v2. No?

Thanks,
Namhyung

maintainer
#7 lore raw
Quoted 9 lines · Tanushree Shah

Hi Namhyung,

Yes, I was planning to send a v2 based on the review comments to fix the pre-existing issue.

However, Arnaldo mentioned that the original patch had already been merged before I got to send a v2, so I addressed the pre-existing issue in a separate patch instead. The patch is available here:

Quoted 5 lines

https://lore.kernel.org/linux-perf-users/[email protected]/

That said, although Arnaldo mentioned that this patch had been merged, I don't see it in perf-tools-next yet either. I'll wait for him to clarify its current status.

It's not in the perf-tools-next, and no plan to add it. Please send v2.

Thanks,
Namhyung

#8 lore raw

Hi Namhyung,
The only change I had planned for v2 was the fix for the pre-existing issue, which has now already been merged separately upstream.

https://github.com/torvalds/linux/commit/2857a5dca750ea989c6fb70b4c14e801e4b7b4ad

Please let me know if there is any additional change you would like me to include in v2.

Quoted 12 lines · Namhyung Kim
Quoted 11 lines · Tanushree Shah

Hi Namhyung,

Yes, I was planning to send a v2 based on the review comments to fix the pre-existing issue.

However, Arnaldo mentioned that the original patch had already been merged before I got to send a v2, so I addressed the pre-existing issue in a separate patch instead. The patch is available here:

https://lore.kernel.org/linux-perf-users/[email protected]/

Quoted 10 lines
Quoted 3 lines

That said, although Arnaldo mentioned that this patch had been merged, I don't see it in perf-tools-next yet either. I'll wait for him to clarify its current status.

It's not in the perf-tools-next, and no plan to add it. Please send v2.

Thanks,
Namhyung

Thanks
Tanushree Shah

maintainer Applied
#10 lore raw
In reply to Tanushree Shah
Quoted 10 lines · Tanushree Shah

Unlike the in-kernel trace_seq which uses a statically allocated buffer, the userspace traceevent library's trace_seq uses a dynamically allocated one. Therefore, every trace_seq_init() call must be paired with a trace_seq_destroy(), otherwise it produces a memory leak.

In process_sample_event(), a trace_seq is initialized for each field when formatting tracepoint raw_data, but the matching trace_seq_destroy() is never called, leaking memory for every field of every sample processed.

[...]

Applied to perf-tools-next, thanks!

Best regards,
Namhyung