Skip to contentSkip to conversation

perf/lock: Fix non-atomic max/time and min_time updates in contention_data

PATCH RESEND In review Suchit Karunakaran 3 participants 4 messages May 4 – May 6

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

PATCH RESEND 1/1 perf/lock: Fix non-atomic max/time and min_time updates in contention_data

The update_contention_data() had a FIXME noting that max_time and min_time updates lacked atomicity. Two CPUs could simultaneously read a stale value, pass the comparison check and race on the write-back, with the smaller value potentially overwriting the larger one and silently corrupting the statistics.

Fix this by replacing the bare conditional assignments with a bpf_loop()-based CAS retry loop. Each field tracks its own convergence independently via max_done/min_done flags in cas_ctx, so a successful CAS on one field is never retried even if the other field needs more attempts.

---
 .../perf/util/bpf_skel/lock_contention.bpf.c  | 50 +++++++++++++++++--
 1 file changed, 45 insertions(+), 5 deletions(-)
tools/perf/util/bpf_skel/lock_contention.bpf.c +45 -5
index 96e7d853b9ed..5c8431be674a 100644
@@ -175,6 +175,13 @@ struct mm_struct___new {
 	struct rw_semaphore mmap_lock;
 } __attribute__((preserve_access_index));
 
+struct cas_ctx {
+	struct contention_data *data;
+	u64 duration;
+	int max_done;
+	int min_done;
+};
+
 extern struct kmem_cache *bpf_get_kmem_cache(u64 addr) __ksym __weak;
 
 /* control flags */
@@ -486,16 +493,49 @@ static inline s32 get_owner_stack_id(u64 *stacktrace)
 	return -1;
 }
 
+static long cas_min_max_cb(u64 idx, void *arg)
+{
+	struct cas_ctx *ctx = arg;
+
+	if (!ctx->max_done) {
+		u64 old_max = ctx->data->max_time;
+		if (old_max >= ctx->duration) {
+			ctx->max_done = 1;
+		} else {
+			u64 r = __sync_val_compare_and_swap(
+				&ctx->data->max_time, old_max, ctx->duration);
+			if (r == old_max)
+				ctx->max_done = 1;
+		}
+	}
+
+	if (!ctx->min_done) {
+		u64 old_min = ctx->data->min_time;
+		if (old_min <= ctx->duration) {
+			ctx->min_done = 1;
+		} else {
+			u64 r = __sync_val_compare_and_swap(
+				&ctx->data->min_time, old_min, ctx->duration);
+			if (r == old_min)
+				ctx->min_done = 1;
+		}
+	}
+
+	return (ctx->max_done && ctx->min_done) ? 1 : 0;
+}
+
 static inline void update_contention_data(struct contention_data *data, u64 duration, u32 count)
 {
 	__sync_fetch_and_add(&data->total_time, duration);
 	__sync_fetch_and_add(&data->count, count);
 
-	/* FIXME: need atomic operations */
-	if (data->max_time < duration)
-		data->max_time = duration;
-	if (data->min_time > duration)
-		data->min_time = duration;
+	struct cas_ctx ctx = {
+		.data     = data,
+		.duration = duration,
+		.max_done = 0,
+		.min_done = 0,
+	};
+	bpf_loop(64, cas_min_max_cb, &ctx, 0);
 }
 
 static inline void update_owner_stat(u32 id, u64 duration, u32 flags)
Signature · 1 line
2.53.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Low] The use of `__sync_val_compare_and_swap` introduces a compiler dependency on BPF ISA v3, breaking…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Low] The use of __sync_val_compare_and_swap introduces a compiler dependency on BPF ISA v3, breaking the perf tool build on older toolchains. --

commit 30820839a2cfcfa987fbc2e0f79fa0e2f2cf0dc3
Author: Suchit Karunakaran <[email protected]>

perf/lock: Fix non-atomic max/time and min_time updates in contention_data

This commit replaces bare conditional assignments with a bpf_loop()-based CAS retry loop to fix non-atomic updates of max_time and min_time.

Quoted 4 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +0 -0
index 96e7d853b9edf..5c8431be674ac 100644

[ ... ]

Quoted 5 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +1 -0
@@ -486,16 +493,49 @@ static inline s32 get_owner_stack_id(u64 *stacktrace)
 	return -1;
 }
 
+static long cas_min_max_cb(u64 idx, void *arg)

tools/perf/util/bpf_skel/lock_contention.bpf.c static inline s32 get_owner_stack_id(u64 *stacktrace) · line 506

Quoted 10 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +10 -0
+{
+	struct cas_ctx *ctx = arg;
+
+	if (!ctx->max_done) {
+		u64 old_max = ctx->data->max_time;
+		if (old_max >= ctx->duration) {
+			ctx->max_done = 1;
+		} else {
+			u64 r = __sync_val_compare_and_swap(
+				&ctx->data->max_time, old_max, ctx->duration);

Does this cause build failures on older toolchains?

The built-in __sync_val_compare_and_swap lowers to the BPF_CMPXCHG instruction, which requires BPF CPU v3 (-mcpu=v3).

The perf build system compiles BPF skeletons using the host compiler's default target features, without explicitly passing an -mcpu flag. On toolchains where the default BPF target is v1 or v2 (such as Clang < 18), the compiler may fail to generate the instruction and throw a fatal error (e.g., 'hardware compare and swap not supported' or an unresolved libcall to __sync_val_compare_and_swap_8).

Quoted 4 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +4 -0
+			if (r == old_max)
+				ctx->max_done = 1;
+		}
+	}

[ ... ]

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

Hello,

Quoted 35 lines · On Mon, May 04, 2026 at 05:23:37PM +0000, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Low] The use of __sync_val_compare_and_swap introduces a compiler dependency on BPF ISA v3, breaking the perf tool build on older toolchains. --

commit 30820839a2cfcfa987fbc2e0f79fa0e2f2cf0dc3
Author: Suchit Karunakaran <[email protected]>

perf/lock: Fix non-atomic max/time and min_time updates in contention_data

This commit replaces bare conditional assignments with a bpf_loop()-based CAS retry loop to fix non-atomic updates of max_time and min_time.

Quoted 4 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +0 -0
index 96e7d853b9edf..5c8431be674ac 100644

[ ... ]

Quoted 5 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +1 -0
@@ -486,16 +493,49 @@ static inline s32 get_owner_stack_id(u64 *stacktrace)
 	return -1;
 }
 
+static long cas_min_max_cb(u64 idx, void *arg)

tools/perf/util/bpf_skel/lock_contention.bpf.c static inline s32 get_owner_stack_id(u64 *stacktrace) · line 506

Quoted 10 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +10 -0
+{
+	struct cas_ctx *ctx = arg;
+
+	if (!ctx->max_done) {
+		u64 old_max = ctx->data->max_time;
+		if (old_max >= ctx->duration) {
+			ctx->max_done = 1;
+		} else {
+			u64 r = __sync_val_compare_and_swap(
+				&ctx->data->max_time, old_max, ctx->duration);

Does this cause build failures on older toolchains?

Quoted 2 lines

The built-in __sync_val_compare_and_swap lowers to the BPF_CMPXCHG instruction, which requires BPF CPU v3 (-mcpu=v3).

Hmm.. this looks like a real concern. We could add -mcpu=v3 to the compiler option, but then there would be compatibility issues with old kernels. It seems it's added in v5.1 kernel and the oldest longterm support kernel version is 5.10. So I think it's fine to add it.

Could you please update the build flag as well? It can be a separate commit.

Thanks,
Namhyung

Quoted 16 lines

The perf build system compiles BPF skeletons using the host compiler's default target features, without explicitly passing an -mcpu flag. On toolchains where the default BPF target is v1 or v2 (such as Clang < 18), the compiler may fail to generate the instruction and throw a fatal error (e.g., 'hardware compare and swap not supported' or an unresolved libcall to __sync_val_compare_and_swap_8).

Quoted 4 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +4 -0
+			if (r == old_max)
+				ctx->max_done = 1;
+		}
+	}

[ ... ]

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

Hi Namhyung,

Quoted 47 lines · Namhyung Kim

Hello,

Quoted 35 lines · On Mon, May 04, 2026 at 05:23:37PM +0000, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Low] The use of __sync_val_compare_and_swap introduces a compiler dependency on BPF ISA v3, breaking the perf tool build on older toolchains. --

commit 30820839a2cfcfa987fbc2e0f79fa0e2f2cf0dc3
Author: Suchit Karunakaran <[email protected]>

perf/lock: Fix non-atomic max/time and min_time updates in contention_data

This commit replaces bare conditional assignments with a bpf_loop()-based CAS retry loop to fix non-atomic updates of max_time and min_time.

Quoted 4 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +0 -0
index 96e7d853b9edf..5c8431be674ac 100644

[ ... ]

Quoted 5 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +1 -0
@@ -486,16 +493,49 @@ static inline s32 get_owner_stack_id(u64 *stacktrace)
    return -1;
 }
 
+static long cas_min_max_cb(u64 idx, void *arg)

tools/perf/util/bpf_skel/lock_contention.bpf.c static inline s32 get_owner_stack_id(u64 *stacktrace) · line 506

Quoted 10 lines
tools/perf/util/bpf_skel/lock_contention.bpf.c +10 -0
+{
+   struct cas_ctx *ctx = arg;
+
+   if (!ctx->max_done) {
+           u64 old_max = ctx->data->max_time;
+           if (old_max >= ctx->duration) {
+                   ctx->max_done = 1;
+           } else {
+                   u64 r = __sync_val_compare_and_swap(
+                           &ctx->data->max_time, old_max, ctx->duration);

Does this cause build failures on older toolchains?

Quoted 2 lines

The built-in __sync_val_compare_and_swap lowers to the BPF_CMPXCHG instruction, which requires BPF CPU v3 (-mcpu=v3).

Hmm.. this looks like a real concern. We could add -mcpu=v3 to the compiler option, but then there would be compatibility issues with old kernels. It seems it's added in v5.1 kernel and the oldest longterm support kernel version is 5.10. So I think it's fine to add it.

Quoted 7 lines

Could you please update the build flag as well? It can be a separate commit.

Thanks,
Namhyung

Quoted 1 line

Sure, I'll send a patch for that as well.