Skip to contentSkip to conversation

perf stat: Do not open cgroups for BPF counters

PATCH Applied Namhyung Kim 2 participants 4 messages Jul 8 – Jul 16 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

PATCH 1/1 perf stat: Do not open cgroups for BPF counters

The --bpf-counters and --for-each-cgroup options use a set of shared events among the given cgroups rather than adding events for each cgroup respectively. It only uses cgroup-ID to compare and calculate the result. So no need to open and keep FDs for cgroups in BPF mode.

---
 tools/perf/builtin-stat.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/builtin-stat.c +2 -1
index 92cdb2df7285befb..49e523058185b753 100644
@@ -2888,7 +2888,8 @@ int cmd_stat(int argc, const char **argv)
 			goto out;
 		}
 
-		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, true) < 0) {
+		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
+					  !target.use_bpf) < 0) {
 			parse_options_usage(stat_usage, stat_options,
 					    "for-each-cgroup", 0);
 			goto out;
Signature · 1 line
2.55.0.795.g602f6c329a-goog
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Fallback from BPF counter mode to standard cgroup mode leaves cgroup file descriptors…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Fallback from BPF counter mode to standard cgroup mode leaves cgroup file descriptors uninitialized (-1), causing standard perf event opening to fail with -EBADF. --

--- Patch [1]: [PATCH] perf stat: Do not open cgroups for BPF counters --- commit 510c63ce03b2362940d1f693b45d5faa86cf3bc1
Author: Namhyung Kim <[email protected]>

perf stat: Do not open cgroups for BPF counters

This commit prevents opening and keeping file descriptors for cgroups when using the --bpf-counters and --for-each-cgroup options. BPF mode uses a set of shared events and compares cgroup-IDs, so it doesn't need individual cgroup file descriptors.

tools/perf/builtin-stat.c int cmd_stat(int argc, const char **argv) · line 2878

Quoted 10 lines
tools/perf/builtin-stat.c +2 -1
@@ -2874,7 +2874,8 @@ int cmd_stat(int argc, const char **argv)
 			goto out;
 		}
 
-		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, true) < 0) {
+		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
+					  !target.use_bpf) < 0) {

[Severity: Medium]
Will this break the fallback path when we exceed BPERF_CGROUP__MAX_EVENTS?

Later in cmd_stat(), we have a fallback that disables BPF counters if there are too many events:

	if (target.use_bpf && nr_cgroups &&
	    (evsel_list->core.nr_entries / nr_cgroups) > BPERF_CGROUP__MAX_EVENTS) {
		pr_warning("Disabling BPF counters due to more events...");
		target.use_bpf = false;
	}

If this fallback is triggered, target.use_bpf becomes false, but the cgroups were already expanded with !target.use_bpf (which was false at the time), leaving the cgroup file descriptors uninitialized (-1).

When the standard event opening later calls evsel__open_cpu() in tools/perf/util/evsel.c:

	if (evsel->cgrp)
		pid = evsel->cgrp->fd;

Could it attempt to use the uninitialized cgroup file descriptor (-1) with the perf_event_open syscall, causing it to fail with -EBADF? Should the cgroup file descriptors be reopened if the fallback to standard mode is taken?

Quoted 3 lines
tools/perf/builtin-stat.c +0 -0
 			parse_options_usage(stat_usage, stat_options,
 					    "for-each-cgroup", 0);
 			goto out;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author maintainer
#2 lore raw
Quoted 38 lines · On Wed, Jul 08, 2026 at 08:14:33PM +0000, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Fallback from BPF counter mode to standard cgroup mode leaves cgroup file descriptors uninitialized (-1), causing standard perf event opening to fail with -EBADF. --

--- Patch [1]: [PATCH] perf stat: Do not open cgroups for BPF counters --- commit 510c63ce03b2362940d1f693b45d5faa86cf3bc1
Author: Namhyung Kim <[email protected]>

perf stat: Do not open cgroups for BPF counters

This commit prevents opening and keeping file descriptors for cgroups when using the --bpf-counters and --for-each-cgroup options. BPF mode uses a set of shared events and compares cgroup-IDs, so it doesn't need individual cgroup file descriptors.

tools/perf/builtin-stat.c int cmd_stat(int argc, const char **argv) · line 2878

Quoted 10 lines
tools/perf/builtin-stat.c +2 -1
@@ -2874,7 +2874,8 @@ int cmd_stat(int argc, const char **argv)
 			goto out;
 		}
 
-		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list, true) < 0) {
+		if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
+					  !target.use_bpf) < 0) {

[Severity: Medium]
Will this break the fallback path when we exceed BPERF_CGROUP__MAX_EVENTS?

Later in cmd_stat(), we have a fallback that disables BPF counters if there are too many events:

	if (target.use_bpf && nr_cgroups &&
	    (evsel_list->core.nr_entries / nr_cgroups) > BPERF_CGROUP__MAX_EVENTS) {
		pr_warning("Disabling BPF counters due to more events...");
		target.use_bpf = false;
	}
Quoted 3 lines

If this fallback is triggered, target.use_bpf becomes false, but the cgroups were already expanded with !target.use_bpf (which was false at the time), leaving the cgroup file descriptors uninitialized (-1).

I think it's a false alarm. The evlist__expand_cgroup() is only used if stat_config.cgroup_list is set and then it should not have nr_cgroups. There's a check for that right before the call to the function.

So in that case, it cannot pass the if condition below and never set to target.use_bpf to false.

Thanks,
Namhyung

author maintainer Applied
#3 lore raw
In reply to Namhyung Kim
Quoted 6 lines · Namhyung Kim

The --bpf-counters and --for-each-cgroup options use a set of shared events among the given cgroups rather than adding events for each cgroup respectively. It only uses cgroup-ID to compare and calculate the result. So no need to open and keep FDs for cgroups in BPF mode.

Applied to perf-tools-next, thanks!

Best regards,
Namhyung