Skip to contentSkip to conversation

perf/core: Save raw sample data conditionally based on sample type

PATCH Superseded Yabin Cui 2 participants 4 messages May 2, 2024 – May 9, 2024

Subsystems (5)
BPF perf events Tracing BPF tracing +1 more subsystems

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Yabin Cui

author
#0 lore raw

PATCH v2 1/1 perf/core: Save raw sample data conditionally based on sample type

Currently, space for raw sample data is always allocated within sample records for both BPF output and tracepoint events. This leads to unused space in sample records when raw sample data is not requested.

This patch checks sample type of an event before saving raw sample data in both BPF output and tracepoint event handling logic. Raw sample data will only be saved if explicitly requested, reducing overhead when it is not needed.

Fixes 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")

---

Changes since v1:
- Check event->attr.sample_type & PERF_SAMPLE_RAW before calling perf_sample_save_raw_data().
- Subject has been changed to reflect the change of solution.

Original commit message from v1:
perf/core: Trim dyn_size if raw data is absent

 kernel/events/core.c     | 37 ++++++++++++++++++++-----------------
 kernel/trace/bpf_trace.c | 12 +++++++-----
 2 files changed, 27 insertions(+), 22 deletions(-)
kernel/events/core.c +20 -17
index 724e6d7e128f..dc5f3147feef 100644
@@ -10120,9 +10120,9 @@ static struct pmu perf_tracepoint = {
 };
 
 static int perf_tp_filter_match(struct perf_event *event,
-				struct perf_sample_data *data)
+				struct perf_raw_record *raw)
 {
-	void *record = data->raw->frag.data;
+	void *record = raw->frag.data;
 
 	/* only top level events have filters set */
 	if (event->parent)
@@ -10134,7 +10134,7 @@ static int perf_tp_filter_match(struct perf_event *event,
 }
 
 static int perf_tp_event_match(struct perf_event *event,
-				struct perf_sample_data *data,
+				struct perf_raw_record *raw,
 				struct pt_regs *regs)
 {
 	if (event->hw.state & PERF_HES_STOPPED)
@@ -10145,7 +10145,7 @@ static int perf_tp_event_match(struct perf_event *event,
 	if (event->attr.exclude_kernel && !user_mode(regs))
 		return 0;
 
-	if (!perf_tp_filter_match(event, data))
+	if (!perf_tp_filter_match(event, raw))
 		return 0;
 
 	return 1;
@@ -10171,6 +10171,7 @@ EXPORT_SYMBOL_GPL(perf_trace_run_bpf_submit);
 static void __perf_tp_event_target_task(u64 count, void *record,
 					struct pt_regs *regs,
 					struct perf_sample_data *data,
+					struct perf_raw_record *raw,
 					struct perf_event *event)
 {
 	struct trace_entry *entry = record;
@@ -10180,13 +10181,18 @@ static void __perf_tp_event_target_task(u64 count, void *record,
 	/* Cannot deliver synchronous signal to other task. */
 	if (event->attr.sigtrap)
 		return;
-	if (perf_tp_event_match(event, data, regs))
+	if (perf_tp_event_match(event, raw, regs)) {
+		perf_sample_data_init(data, 0, 0);
+		if (event->attr.sample_type & PERF_SAMPLE_RAW)
+			perf_sample_save_raw_data(data, raw);
 		perf_swevent_event(event, count, data, regs);
+	}
 }
 
 static void perf_tp_event_target_task(u64 count, void *record,
 				      struct pt_regs *regs,
 				      struct perf_sample_data *data,
+				      struct perf_raw_record *raw,
 				      struct perf_event_context *ctx)
 {
 	unsigned int cpu = smp_processor_id();
@@ -10194,15 +10200,15 @@ static void perf_tp_event_target_task(u64 count, void *record,
 	struct perf_event *event, *sibling;
 
 	perf_event_groups_for_cpu_pmu(event, &ctx->pinned_groups, cpu, pmu) {
-		__perf_tp_event_target_task(count, record, regs, data, event);
+		__perf_tp_event_target_task(count, record, regs, data, raw, event);
 		for_each_sibling_event(sibling, event)
-			__perf_tp_event_target_task(count, record, regs, data, sibling);
+			__perf_tp_event_target_task(count, record, regs, data, raw, sibling);
 	}
 
 	perf_event_groups_for_cpu_pmu(event, &ctx->flexible_groups, cpu, pmu) {
-		__perf_tp_event_target_task(count, record, regs, data, event);
+		__perf_tp_event_target_task(count, record, regs, data, raw, event);
 		for_each_sibling_event(sibling, event)
-			__perf_tp_event_target_task(count, record, regs, data, sibling);
+			__perf_tp_event_target_task(count, record, regs, data, raw, sibling);
 	}
 }
 
@@ -10220,15 +10226,10 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
 		},
 	};
 
-	perf_sample_data_init(&data, 0, 0);
-	perf_sample_save_raw_data(&data, &raw);
-
 	perf_trace_buf_update(record, event_type);
 
 	hlist_for_each_entry_rcu(event, head, hlist_entry) {
-		if (perf_tp_event_match(event, &data, regs)) {
-			perf_swevent_event(event, count, &data, regs);
-
+		if (perf_tp_event_match(event, &raw, regs)) {
 			/*
 			 * Here use the same on-stack perf_sample_data,
 			 * some members in data are event-specific and
@@ -10238,7 +10239,9 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
 			 * because data->sample_flags is set.
 			 */
 			perf_sample_data_init(&data, 0, 0);
-			perf_sample_save_raw_data(&data, &raw);
+			if (event->attr.sample_type & PERF_SAMPLE_RAW)
+				perf_sample_save_raw_data(&data, &raw);
+			perf_swevent_event(event, count, &data, regs);
 		}
 	}
 
@@ -10255,7 +10258,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
 			goto unlock;
 
 		raw_spin_lock(&ctx->lock);
-		perf_tp_event_target_task(count, record, regs, &data, ctx);
+		perf_tp_event_target_task(count, record, regs, &data, &raw, ctx);
 		raw_spin_unlock(&ctx->lock);
 unlock:
 		rcu_read_unlock();
kernel/trace/bpf_trace.c +7 -5
index 9dc605f08a23..4b3ff71b4c0a 100644
@@ -620,7 +620,8 @@ static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 
 static __always_inline u64
 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
-			u64 flags, struct perf_sample_data *sd)
+			u64 flags, struct perf_raw_record *raw,
+			struct perf_sample_data *sd)
 {
 	struct bpf_array *array = container_of(map, struct bpf_array, map);
 	unsigned int cpu = smp_processor_id();
@@ -645,6 +646,9 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 	if (unlikely(event->oncpu != cpu))
 		return -EOPNOTSUPP;
 
+	if (event->attr.sample_type & PERF_SAMPLE_RAW)
+		perf_sample_save_raw_data(sd, raw);
+
 	return perf_event_output(event, sd, regs);
 }
 
@@ -688,9 +692,8 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
 	}
 
 	perf_sample_data_init(sd, 0, 0);
-	perf_sample_save_raw_data(sd, &raw);
 
-	err = __bpf_perf_event_output(regs, map, flags, sd);
+	err = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
 	this_cpu_dec(bpf_trace_nest_level);
 	preempt_enable();
@@ -749,9 +752,8 @@ u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 
 	perf_fetch_caller_regs(regs);
 	perf_sample_data_init(sd, 0, 0);
-	perf_sample_save_raw_data(sd, &raw);
 
-	ret = __bpf_perf_event_output(regs, map, flags, sd);
+	ret = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
 	this_cpu_dec(bpf_event_output_nest_level);
 	preempt_enable();
Signature · 1 line
2.45.0.rc0.197.gbae5840b3b-goog
maintainer Acked-by on v2 1/1
#1 lore raw

Hello,

Quoted 4 lines · Yabin Cui

Currently, space for raw sample data is always allocated within sample records for both BPF output and tracepoint events. This leads to unused space in sample records when raw sample data is not requested.

Oh, I thought it was ok because even if it sets _RAW bit in the data->sample_flags unconditionally, perf_prepare_sample() and perf_output_sample() checks the original event's attr->sample_type so the raw data won't be recorded.

But I've realized that it increased data->dyn_size already. :( Which means the sample would have garbage at the end.

I need to check if there are other places that made the same mistake for other sample types.

Quoted 8 lines

This patch checks sample type of an event before saving raw sample data in both BPF output and tracepoint event handling logic. Raw sample data will only be saved if explicitly requested, reducing overhead when it is not needed.

Fixes 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")

Thanks,
Namhyung

Quoted 182 lines

---

Changes since v1:
- Check event->attr.sample_type & PERF_SAMPLE_RAW before calling perf_sample_save_raw_data().
- Subject has been changed to reflect the change of solution.

Original commit message from v1:
perf/core: Trim dyn_size if raw data is absent

 kernel/events/core.c     | 37 ++++++++++++++++++++-----------------
 kernel/trace/bpf_trace.c | 12 +++++++-----
 2 files changed, 27 insertions(+), 22 deletions(-)
kernel/events/core.c +20 -17
index 724e6d7e128f..dc5f3147feef 100644
@@ -10120,9 +10120,9 @@ static struct pmu perf_tracepoint = {
 };
 
 static int perf_tp_filter_match(struct perf_event *event,
-                               struct perf_sample_data *data)
+                               struct perf_raw_record *raw)
 {
-       void *record = data->raw->frag.data;
+       void *record = raw->frag.data;
 
        /* only top level events have filters set */
        if (event->parent)
@@ -10134,7 +10134,7 @@ static int perf_tp_filter_match(struct perf_event *event,
 }
 
 static int perf_tp_event_match(struct perf_event *event,
-                               struct perf_sample_data *data,
+                               struct perf_raw_record *raw,
                                struct pt_regs *regs)
 {
        if (event->hw.state & PERF_HES_STOPPED)
@@ -10145,7 +10145,7 @@ static int perf_tp_event_match(struct perf_event *event,
        if (event->attr.exclude_kernel && !user_mode(regs))
                return 0;
 
-       if (!perf_tp_filter_match(event, data))
+       if (!perf_tp_filter_match(event, raw))
                return 0;
 
        return 1;
@@ -10171,6 +10171,7 @@ EXPORT_SYMBOL_GPL(perf_trace_run_bpf_submit);
 static void __perf_tp_event_target_task(u64 count, void *record,
                                        struct pt_regs *regs,
                                        struct perf_sample_data *data,
+                                       struct perf_raw_record *raw,
                                        struct perf_event *event)
 {
        struct trace_entry *entry = record;
@@ -10180,13 +10181,18 @@ static void __perf_tp_event_target_task(u64 count, void *record,
        /* Cannot deliver synchronous signal to other task. */
        if (event->attr.sigtrap)
                return;
-       if (perf_tp_event_match(event, data, regs))
+       if (perf_tp_event_match(event, raw, regs)) {
+               perf_sample_data_init(data, 0, 0);
+               if (event->attr.sample_type & PERF_SAMPLE_RAW)
+                       perf_sample_save_raw_data(data, raw);
                perf_swevent_event(event, count, data, regs);
+       }
 }
 
 static void perf_tp_event_target_task(u64 count, void *record,
                                      struct pt_regs *regs,
                                      struct perf_sample_data *data,
+                                     struct perf_raw_record *raw,
                                      struct perf_event_context *ctx)
 {
        unsigned int cpu = smp_processor_id();
@@ -10194,15 +10200,15 @@ static void perf_tp_event_target_task(u64 count, void *record,
        struct perf_event *event, *sibling;
 
        perf_event_groups_for_cpu_pmu(event, &ctx->pinned_groups, cpu, pmu) {
-               __perf_tp_event_target_task(count, record, regs, data, event);
+               __perf_tp_event_target_task(count, record, regs, data, raw, event);
                for_each_sibling_event(sibling, event)
-                       __perf_tp_event_target_task(count, record, regs, data, sibling);
+                       __perf_tp_event_target_task(count, record, regs, data, raw, sibling);
        }
 
        perf_event_groups_for_cpu_pmu(event, &ctx->flexible_groups, cpu, pmu) {
-               __perf_tp_event_target_task(count, record, regs, data, event);
+               __perf_tp_event_target_task(count, record, regs, data, raw, event);
                for_each_sibling_event(sibling, event)
-                       __perf_tp_event_target_task(count, record, regs, data, sibling);
+                       __perf_tp_event_target_task(count, record, regs, data, raw, sibling);
        }
 }
 
@@ -10220,15 +10226,10 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                },
        };
 
-       perf_sample_data_init(&data, 0, 0);
-       perf_sample_save_raw_data(&data, &raw);
-
        perf_trace_buf_update(record, event_type);
 
        hlist_for_each_entry_rcu(event, head, hlist_entry) {
-               if (perf_tp_event_match(event, &data, regs)) {
-                       perf_swevent_event(event, count, &data, regs);
-
+               if (perf_tp_event_match(event, &raw, regs)) {
                        /*
                         * Here use the same on-stack perf_sample_data,
                         * some members in data are event-specific and
@@ -10238,7 +10239,9 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                         * because data->sample_flags is set.
                         */
                        perf_sample_data_init(&data, 0, 0);
-                       perf_sample_save_raw_data(&data, &raw);
+                       if (event->attr.sample_type & PERF_SAMPLE_RAW)
+                               perf_sample_save_raw_data(&data, &raw);
+                       perf_swevent_event(event, count, &data, regs);
                }
        }
 
@@ -10255,7 +10258,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                        goto unlock;
 
                raw_spin_lock(&ctx->lock);
-               perf_tp_event_target_task(count, record, regs, &data, ctx);
+               perf_tp_event_target_task(count, record, regs, &data, &raw, ctx);
                raw_spin_unlock(&ctx->lock);
 unlock:
                rcu_read_unlock();
kernel/trace/bpf_trace.c +7 -6
index 9dc605f08a23..4b3ff71b4c0a 100644
@@ -620,7 +620,8 @@ static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 
 static __always_inline u64
 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
-                       u64 flags, struct perf_sample_data *sd)
+                       u64 flags, struct perf_raw_record *raw,
+                       struct perf_sample_data *sd)
 {
        struct bpf_array *array = container_of(map, struct bpf_array, map);
        unsigned int cpu = smp_processor_id();
@@ -645,6 +646,9 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
        if (unlikely(event->oncpu != cpu))
                return -EOPNOTSUPP;
 
+       if (event->attr.sample_type & PERF_SAMPLE_RAW)
+               perf_sample_save_raw_data(sd, raw);
+
        return perf_event_output(event, sd, regs);
 }
 
@@ -688,9 +692,8 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
        }
 
        perf_sample_data_init(sd, 0, 0);
-       perf_sample_save_raw_data(sd, &raw);
 
-       err = __bpf_perf_event_output(regs, map, flags, sd);
+       err = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
        this_cpu_dec(bpf_trace_nest_level);
        preempt_enable();
@@ -749,9 +752,8 @@ u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 
        perf_fetch_caller_regs(regs);
        perf_sample_data_init(sd, 0, 0);
-       perf_sample_save_raw_data(sd, &raw);
 
-       ret = __bpf_perf_event_output(regs, map, flags, sd);
+       ret = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
        this_cpu_dec(bpf_event_output_nest_level);
        preempt_enable();
--

2.45.0.rc0.197.gbae5840b3b-goog

maintainer
#2 lore raw

Another thought.

Quoted 35 lines · Namhyung Kim

Hello,

Quoted 4 lines · Yabin Cui

Currently, space for raw sample data is always allocated within sample records for both BPF output and tracepoint events. This leads to unused space in sample records when raw sample data is not requested.

Oh, I thought it was ok because even if it sets _RAW bit in the data->sample_flags unconditionally, perf_prepare_sample() and perf_output_sample() checks the original event's attr->sample_type so the raw data won't be recorded.

But I've realized that it increased data->dyn_size already. :( Which means the sample would have garbage at the end.

I need to check if there are other places that made the same mistake for other sample types.

Quoted 8 lines

This patch checks sample type of an event before saving raw sample data in both BPF output and tracepoint event handling logic. Raw sample data will only be saved if explicitly requested, reducing overhead when it is not needed.

Fixes 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")

Thanks,
Namhyung

Quoted 1 line

---

[SNIP]

Quoted 9 lines

kernel/events/core.c static void __perf_tp_event_target_task(u64 count, void *record, · line 10187

Quoted 9 lines
kernel/events/core.c +4 -1
@@ -10180,13 +10181,18 @@ static void __perf_tp_event_target_task(u64 count, void *record,
        /* Cannot deliver synchronous signal to other task. */
        if (event->attr.sigtrap)
                return;
-       if (perf_tp_event_match(event, data, regs))
+       if (perf_tp_event_match(event, raw, regs)) {
+               perf_sample_data_init(data, 0, 0);
+               if (event->attr.sample_type & PERF_SAMPLE_RAW)
+                       perf_sample_save_raw_data(data, raw);

Hmm.. to prevent future mistakes, maybe we can move the check into the perf_sample_save_raw_data() itself.

And it'd be great if you could do the same for callchain and brstack too. :)

Thanks,
Namhyung

Quoted 117 lines
Quoted 117 lines
kernel/events/core.c +11 -12
                perf_swevent_event(event, count, data, regs);
+       }
 }
 
 static void perf_tp_event_target_task(u64 count, void *record,
                                      struct pt_regs *regs,
                                      struct perf_sample_data *data,
+                                     struct perf_raw_record *raw,
                                      struct perf_event_context *ctx)
 {
        unsigned int cpu = smp_processor_id();
@@ -10194,15 +10200,15 @@ static void perf_tp_event_target_task(u64 count, void *record,
        struct perf_event *event, *sibling;
 
        perf_event_groups_for_cpu_pmu(event, &ctx->pinned_groups, cpu, pmu) {
-               __perf_tp_event_target_task(count, record, regs, data, event);
+               __perf_tp_event_target_task(count, record, regs, data, raw, event);
                for_each_sibling_event(sibling, event)
-                       __perf_tp_event_target_task(count, record, regs, data, sibling);
+                       __perf_tp_event_target_task(count, record, regs, data, raw, sibling);
        }
 
        perf_event_groups_for_cpu_pmu(event, &ctx->flexible_groups, cpu, pmu) {
-               __perf_tp_event_target_task(count, record, regs, data, event);
+               __perf_tp_event_target_task(count, record, regs, data, raw, event);
                for_each_sibling_event(sibling, event)
-                       __perf_tp_event_target_task(count, record, regs, data, sibling);
+                       __perf_tp_event_target_task(count, record, regs, data, raw, sibling);
        }
 }
 
@@ -10220,15 +10226,10 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                },
        };
 
-       perf_sample_data_init(&data, 0, 0);
-       perf_sample_save_raw_data(&data, &raw);
-
        perf_trace_buf_update(record, event_type);
 
        hlist_for_each_entry_rcu(event, head, hlist_entry) {
-               if (perf_tp_event_match(event, &data, regs)) {
-                       perf_swevent_event(event, count, &data, regs);
-
+               if (perf_tp_event_match(event, &raw, regs)) {
                        /*
                         * Here use the same on-stack perf_sample_data,
                         * some members in data are event-specific and
@@ -10238,7 +10239,9 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                         * because data->sample_flags is set.
                         */
                        perf_sample_data_init(&data, 0, 0);
-                       perf_sample_save_raw_data(&data, &raw);
+                       if (event->attr.sample_type & PERF_SAMPLE_RAW)
+                               perf_sample_save_raw_data(&data, &raw);
+                       perf_swevent_event(event, count, &data, regs);
                }
        }
 
@@ -10255,7 +10258,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                        goto unlock;
 
                raw_spin_lock(&ctx->lock);
-               perf_tp_event_target_task(count, record, regs, &data, ctx);
+               perf_tp_event_target_task(count, record, regs, &data, &raw, ctx);
                raw_spin_unlock(&ctx->lock);
 unlock:
                rcu_read_unlock();
kernel/trace/bpf_trace.c +7 -6
index 9dc605f08a23..4b3ff71b4c0a 100644
@@ -620,7 +620,8 @@ static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 
 static __always_inline u64
 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
-                       u64 flags, struct perf_sample_data *sd)
+                       u64 flags, struct perf_raw_record *raw,
+                       struct perf_sample_data *sd)
 {
        struct bpf_array *array = container_of(map, struct bpf_array, map);
        unsigned int cpu = smp_processor_id();
@@ -645,6 +646,9 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
        if (unlikely(event->oncpu != cpu))
                return -EOPNOTSUPP;
 
+       if (event->attr.sample_type & PERF_SAMPLE_RAW)
+               perf_sample_save_raw_data(sd, raw);
+
        return perf_event_output(event, sd, regs);
 }
 
@@ -688,9 +692,8 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
        }
 
        perf_sample_data_init(sd, 0, 0);
-       perf_sample_save_raw_data(sd, &raw);
 
-       err = __bpf_perf_event_output(regs, map, flags, sd);
+       err = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
        this_cpu_dec(bpf_trace_nest_level);
        preempt_enable();
@@ -749,9 +752,8 @@ u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 
        perf_fetch_caller_regs(regs);
        perf_sample_data_init(sd, 0, 0);
-       perf_sample_save_raw_data(sd, &raw);
 
-       ret = __bpf_perf_event_output(regs, map, flags, sd);
+       ret = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
        this_cpu_dec(bpf_event_output_nest_level);
        preempt_enable();
--

2.45.0.rc0.197.gbae5840b3b-goog

author
#3 lore raw

Thanks for reviewing the patch! I will update a v3 patch based on the comments.

Quoted 177 lines · Namhyung Kim

Another thought.

Quoted 35 lines · Namhyung Kim

Hello,

Quoted 4 lines · Yabin Cui

Currently, space for raw sample data is always allocated within sample records for both BPF output and tracepoint events. This leads to unused space in sample records when raw sample data is not requested.

Oh, I thought it was ok because even if it sets _RAW bit in the data->sample_flags unconditionally, perf_prepare_sample() and perf_output_sample() checks the original event's attr->sample_type so the raw data won't be recorded.

But I've realized that it increased data->dyn_size already. :( Which means the sample would have garbage at the end.

I need to check if there are other places that made the same mistake for other sample types.

Quoted 8 lines

This patch checks sample type of an event before saving raw sample data in both BPF output and tracepoint event handling logic. Raw sample data will only be saved if explicitly requested, reducing overhead when it is not needed.

Fixes 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")

Thanks,
Namhyung

Quoted 1 line

---

[SNIP]

Quoted 9 lines

kernel/events/core.c static void __perf_tp_event_target_task(u64 count, void *record, · line 10187

Quoted 9 lines
kernel/events/core.c +4 -1
@@ -10180,13 +10181,18 @@ static void __perf_tp_event_target_task(u64 count, void *record,
        /* Cannot deliver synchronous signal to other task. */
        if (event->attr.sigtrap)
                return;
-       if (perf_tp_event_match(event, data, regs))
+       if (perf_tp_event_match(event, raw, regs)) {
+               perf_sample_data_init(data, 0, 0);
+               if (event->attr.sample_type & PERF_SAMPLE_RAW)
+                       perf_sample_save_raw_data(data, raw);

Hmm.. to prevent future mistakes, maybe we can move the check into the perf_sample_save_raw_data() itself.

And it'd be great if you could do the same for callchain and brstack too. :)

Thanks,
Namhyung

Quoted 117 lines
Quoted 117 lines
kernel/events/core.c +11 -12
                perf_swevent_event(event, count, data, regs);
+       }
 }
 
 static void perf_tp_event_target_task(u64 count, void *record,
                                      struct pt_regs *regs,
                                      struct perf_sample_data *data,
+                                     struct perf_raw_record *raw,
                                      struct perf_event_context *ctx)
 {
        unsigned int cpu = smp_processor_id();
@@ -10194,15 +10200,15 @@ static void perf_tp_event_target_task(u64 count, void *record,
        struct perf_event *event, *sibling;
 
        perf_event_groups_for_cpu_pmu(event, &ctx->pinned_groups, cpu, pmu) {
-               __perf_tp_event_target_task(count, record, regs, data, event);
+               __perf_tp_event_target_task(count, record, regs, data, raw, event);
                for_each_sibling_event(sibling, event)
-                       __perf_tp_event_target_task(count, record, regs, data, sibling);
+                       __perf_tp_event_target_task(count, record, regs, data, raw, sibling);
        }
 
        perf_event_groups_for_cpu_pmu(event, &ctx->flexible_groups, cpu, pmu) {
-               __perf_tp_event_target_task(count, record, regs, data, event);
+               __perf_tp_event_target_task(count, record, regs, data, raw, event);
                for_each_sibling_event(sibling, event)
-                       __perf_tp_event_target_task(count, record, regs, data, sibling);
+                       __perf_tp_event_target_task(count, record, regs, data, raw, sibling);
        }
 }
 
@@ -10220,15 +10226,10 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                },
        };
 
-       perf_sample_data_init(&data, 0, 0);
-       perf_sample_save_raw_data(&data, &raw);
-
        perf_trace_buf_update(record, event_type);
 
        hlist_for_each_entry_rcu(event, head, hlist_entry) {
-               if (perf_tp_event_match(event, &data, regs)) {
-                       perf_swevent_event(event, count, &data, regs);
-
+               if (perf_tp_event_match(event, &raw, regs)) {
                        /*
                         * Here use the same on-stack perf_sample_data,
                         * some members in data are event-specific and
@@ -10238,7 +10239,9 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                         * because data->sample_flags is set.
                         */
                        perf_sample_data_init(&data, 0, 0);
-                       perf_sample_save_raw_data(&data, &raw);
+                       if (event->attr.sample_type & PERF_SAMPLE_RAW)
+                               perf_sample_save_raw_data(&data, &raw);
+                       perf_swevent_event(event, count, &data, regs);
                }
        }
 
@@ -10255,7 +10258,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
                        goto unlock;
 
                raw_spin_lock(&ctx->lock);
-               perf_tp_event_target_task(count, record, regs, &data, ctx);
+               perf_tp_event_target_task(count, record, regs, &data, &raw, ctx);
                raw_spin_unlock(&ctx->lock);
 unlock:
                rcu_read_unlock();
kernel/trace/bpf_trace.c +7 -6
index 9dc605f08a23..4b3ff71b4c0a 100644
@@ -620,7 +620,8 @@ static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
 
 static __always_inline u64
 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
-                       u64 flags, struct perf_sample_data *sd)
+                       u64 flags, struct perf_raw_record *raw,
+                       struct perf_sample_data *sd)
 {
        struct bpf_array *array = container_of(map, struct bpf_array, map);
        unsigned int cpu = smp_processor_id();
@@ -645,6 +646,9 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
        if (unlikely(event->oncpu != cpu))
                return -EOPNOTSUPP;
 
+       if (event->attr.sample_type & PERF_SAMPLE_RAW)
+               perf_sample_save_raw_data(sd, raw);
+
        return perf_event_output(event, sd, regs);
 }
 
@@ -688,9 +692,8 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
        }
 
        perf_sample_data_init(sd, 0, 0);
-       perf_sample_save_raw_data(sd, &raw);
 
-       err = __bpf_perf_event_output(regs, map, flags, sd);
+       err = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
        this_cpu_dec(bpf_trace_nest_level);
        preempt_enable();
@@ -749,9 +752,8 @@ u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
 
        perf_fetch_caller_regs(regs);
        perf_sample_data_init(sd, 0, 0);
-       perf_sample_save_raw_data(sd, &raw);
 
-       ret = __bpf_perf_event_output(regs, map, flags, sd);
+       ret = __bpf_perf_event_output(regs, map, flags, &raw, sd);
 out:
        this_cpu_dec(bpf_event_output_nest_level);
        preempt_enable();
--

2.45.0.rc0.197.gbae5840b3b-goog