Skip to contentSkip to conversation

perf/lock: enable end-timestamp accounting for cgroup aggregation

PATCH RESEND Applied Suchit Karunakaran 4 participants 5 messages May 30 – Jun 4

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH RESEND v2 1/1 perf/lock: enable end-timestamp accounting for cgroup aggregation

update_lock_stat() handles lock contentions that start but never reach a contention_end event (e.g., locks still held when profiling stops), but previously treated LOCK_AGGR_CGROUP as a no-op due to missing cgroup context in userspace.

Fix this by adding a cgroup_id field to struct tstamp_data, recording it at contention_begin using get_current_cgroup_id() when aggr_mode is LOCK_AGGR_CGROUP. Capturing it at contention_begin is semantically correct, the contention cost is incurred by the task that had to wait, not by whatever task happens to be running at contention_end. It is also preferable from a performance standpoint, as contention_end runs just before the task enters the critical section.

Update contention_end to use pelem->cgroup_id instead of calling get_current_cgroup_id() dynamically, ensuring both complete and incomplete contention events attribute the wait time to the cgroup at wait-start time consistently.

---
 tools/perf/util/bpf_lock_contention.c          | 4 ++--
 tools/perf/util/bpf_skel/lock_contention.bpf.c | 4 +++-
 tools/perf/util/bpf_skel/lock_data.h           | 1 +
 3 files changed, 6 insertions(+), 3 deletions(-)
tools/perf/util/bpf_lock_contention.c +2 -2
index cbd7435579fe..1a5bd2ff8ee4 100644
@@ -463,8 +463,8 @@ static void update_lock_stat(int map_fd, int pid, u64 end_ts,
 		stat_key.lock_addr_or_cgroup = ts_data->lock;
 		break;
 	case LOCK_AGGR_CGROUP:
-		/* TODO */
-		return;
+		stat_key.lock_addr_or_cgroup = ts_data->cgroup_id;
+		break;
 	default:
 		return;
 	}
tools/perf/util/bpf_skel/lock_contention.bpf.c +3 -1
index 96e7d853b9ed..139199811020 100644
@@ -536,6 +536,8 @@ int contention_begin(u64 *ctx)
 	pelem->timestamp = bpf_ktime_get_ns();
 	pelem->lock = (__u64)ctx[0];
 	pelem->flags = (__u32)ctx[1];
+	if (aggr_mode == LOCK_AGGR_CGROUP)
+		pelem->cgroup_id = get_current_cgroup_id();
 
 	if (needs_callstack) {
 		u32 i = 0;
@@ -771,7 +773,7 @@ int contention_end(u64 *ctx)
 			key.stack_id = pelem->stack_id;
 		break;
 	case LOCK_AGGR_CGROUP:
-		key.lock_addr_or_cgroup = get_current_cgroup_id();
+		key.lock_addr_or_cgroup = pelem->cgroup_id;
 		break;
 	default:
 		/* should not happen */
tools/perf/util/bpf_skel/lock_data.h +1 -0
index 28c5e5aced7f..652e114e6b87 100644
@@ -13,6 +13,7 @@ struct owner_tracing_data {
 struct tstamp_data {
 	u64 timestamp;
 	u64 lock;
+	u64 cgroup_id;
 	u32 flags;
 	s32 stack_id;
 };
Signature · 1 line
2.54.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Incomplete lock contentions for per-CPU locks (spinlocks/rwlocks) are silently dropped when…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Incomplete lock contentions for per-CPU locks (spinlocks/rwlocks) are silently dropped when using cgroup aggregation due to an unmodified early return in account_end_timestamp(). --

commit 9be2409410efd2d4d44f23abbecf590bbbc25d34
Author: Suchit Karunakaran <[email protected]>

perf/lock: enable end-timestamp accounting for cgroup aggregation

This commit enables end-timestamp accounting for cgroup aggregation by tracking the cgroup_id in the BPF program during contention_begin. It updates update_lock_stat() in userspace and contention_end() in BPF to use this recorded cgroup context instead of fetching it dynamically.

Quoted 5 lines
tools/perf/util/bpf_lock_contention.c +0 -0
index cbd7435579fea..1a5bd2ff8ee4c 100644

tools/perf/util/bpf_lock_contention.c static void update_lock_stat(int map_fd, int pid, u64 end_ts, · line 470

Quoted 10 lines
tools/perf/util/bpf_lock_contention.c +2 -2
@@ -463,8 +463,8 @@ static void update_lock_stat(int map_fd, int pid, u64 end_ts,
 		stat_key.lock_addr_or_cgroup = ts_data->lock;
 		break;
 	case LOCK_AGGR_CGROUP:
-		/* TODO */
-		return;
+		stat_key.lock_addr_or_cgroup = ts_data->cgroup_id;
+		break;
 	default:
 		return;
 	}

[Severity: Medium]
Since this patch enables end-timestamp accounting for cgroup aggregation, should the early return for LOCK_AGGR_CGROUP in account_end_timestamp() also be removed?

tools/perf/util/bpf_lock_contention.c:account_end_timestamp() {
        ...
        /* Now it'll check per-cpu tstamp map which doesn't have TID. */
        if (aggr_mode == LOCK_AGGR_TASK || aggr_mode == LOCK_AGGR_CGROUP)
                return;
        ...
}

Without removing this check, incomplete lock contentions for per-CPU locks (like spinlocks and rwlocks) appear to be silently skipped when using cgroup aggregation.

[ ... ]

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Suchit Karunakaran
Quoted 13 lines · Suchit Karunakaran

update_lock_stat() handles lock contentions that start but never reach a contention_end event (e.g., locks still held when profiling stops), but previously treated LOCK_AGGR_CGROUP as a no-op due to missing cgroup context in userspace.

Fix this by adding a cgroup_id field to struct tstamp_data, recording it at contention_begin using get_current_cgroup_id() when aggr_mode is LOCK_AGGR_CGROUP. Capturing it at contention_begin is semantically correct, the contention cost is incurred by the task that had to wait, not by whatever task happens to be running at contention_end. It is also preferable from a performance standpoint, as contention_end runs just before the task enters the critical section.

Quoted 4 lines

Update contention_end to use pelem->cgroup_id instead of calling get_current_cgroup_id() dynamically, ensuring both complete and incomplete contention events attribute the wait time to the cgroup at wait-start time consistently.

Namhyung, can you provide an Acked-by or Reviewed-by?

Thanks,

- Arnaldo

Quoted 58 lines
---
 tools/perf/util/bpf_lock_contention.c          | 4 ++--
 tools/perf/util/bpf_skel/lock_contention.bpf.c | 4 +++-
 tools/perf/util/bpf_skel/lock_data.h           | 1 +
 3 files changed, 6 insertions(+), 3 deletions(-)
tools/perf/util/bpf_lock_contention.c +2 -2
index cbd7435579fe..1a5bd2ff8ee4 100644
@@ -463,8 +463,8 @@ static void update_lock_stat(int map_fd, int pid, u64 end_ts,
 		stat_key.lock_addr_or_cgroup = ts_data->lock;
 		break;
 	case LOCK_AGGR_CGROUP:
-		/* TODO */
-		return;
+		stat_key.lock_addr_or_cgroup = ts_data->cgroup_id;
+		break;
 	default:
 		return;
 	}
tools/perf/util/bpf_skel/lock_contention.bpf.c +3 -1
index 96e7d853b9ed..139199811020 100644
@@ -536,6 +536,8 @@ int contention_begin(u64 *ctx)
 	pelem->timestamp = bpf_ktime_get_ns();
 	pelem->lock = (__u64)ctx[0];
 	pelem->flags = (__u32)ctx[1];
+	if (aggr_mode == LOCK_AGGR_CGROUP)
+		pelem->cgroup_id = get_current_cgroup_id();
 
 	if (needs_callstack) {
 		u32 i = 0;
@@ -771,7 +773,7 @@ int contention_end(u64 *ctx)
 			key.stack_id = pelem->stack_id;
 		break;
 	case LOCK_AGGR_CGROUP:
-		key.lock_addr_or_cgroup = get_current_cgroup_id();
+		key.lock_addr_or_cgroup = pelem->cgroup_id;
 		break;
 	default:
 		/* should not happen */
tools/perf/util/bpf_skel/lock_data.h +1 -0
index 28c5e5aced7f..652e114e6b87 100644
@@ -13,6 +13,7 @@ struct owner_tracing_data {
 struct tstamp_data {
 	u64 timestamp;
 	u64 lock;
+	u64 cgroup_id;
 	u32 flags;
 	s32 stack_id;
 };
Signature · 1 line
2.54.0
maintainer Reviewed-by
#3 lore raw
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 13 lines · Suchit Karunakaran

update_lock_stat() handles lock contentions that start but never reach a contention_end event (e.g., locks still held when profiling stops), but previously treated LOCK_AGGR_CGROUP as a no-op due to missing cgroup context in userspace.

Fix this by adding a cgroup_id field to struct tstamp_data, recording it at contention_begin using get_current_cgroup_id() when aggr_mode is LOCK_AGGR_CGROUP. Capturing it at contention_begin is semantically correct, the contention cost is incurred by the task that had to wait, not by whatever task happens to be running at contention_end. It is also preferable from a performance standpoint, as contention_end runs just before the task enters the critical section.

Quoted 6 lines
Quoted 4 lines

Update contention_end to use pelem->cgroup_id instead of calling get_current_cgroup_id() dynamically, ensuring both complete and incomplete contention events attribute the wait time to the cgroup at wait-start time consistently.

Namhyung, can you provide an Acked-by or Reviewed-by?

Thanks,
Namhyung

maintainer Applied
#4 lore raw
Quoted 15 lines · Namhyung Kim
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 13 lines · Suchit Karunakaran

update_lock_stat() handles lock contentions that start but never reach a contention_end event (e.g., locks still held when profiling stops), but previously treated LOCK_AGGR_CGROUP as a no-op due to missing cgroup context in userspace.

Fix this by adding a cgroup_id field to struct tstamp_data, recording it at contention_begin using get_current_cgroup_id() when aggr_mode is LOCK_AGGR_CGROUP. Capturing it at contention_begin is semantically correct, the contention cost is incurred by the task that had to wait, not by whatever task happens to be running at contention_end. It is also preferable from a performance standpoint, as contention_end runs just before the task enters the critical section.

Quoted 8 lines
Quoted 6 lines
Quoted 4 lines

Update contention_end to use pelem->cgroup_id instead of calling get_current_cgroup_id() dynamically, ensuring both complete and incomplete contention events attribute the wait time to the cgroup at wait-start time consistently.

Namhyung, can you provide an Acked-by or Reviewed-by?

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

- Arnaldo