Skip to contentSkip to conversation

perf/core: Prepare sample data for BPF

PATCH In review Namhyung Kim 4 participants 23 messages Jan 12, 2023 – Jan 17, 2023

Subsystems (8)
x86 PowerPC BPF perf events +4 more subsystems
lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

cover perf/core: Prepare sample data for BPF

Hello,

The perf_prepare_sample() is to fill the perf sample data and update the header info before sending it to the ring buffer. But we want to use it for BPF overflow handler so that it can access the sample data to filter relevant ones.

Changes in v2)
* the layout change is merged
* reduce branches using __cond_set (PeterZ)
* add helpers to set dynamic sample data (PeterZ)
* introduce perf_prepare_header() (PeterZ)
* call perf_prepare_sample() before bpf_overflow_handler unconditionally

This means the perf_prepare_handler() can be called more than once. To avoid duplicate work, use the data->sample_flags and save the data size.

I also added a few of helpers to set those information accordingly. But it looks some fields like REGS_USER, STACK_USER and AUX are saved in the perf_prepare_sample() so I didn't add the helpers for them.

After than we can just check the filtered_sample_type flags begin zero to determine if it has more work. In that case, it needs to update the data->type since it's possible to miss when PMU driver sets all required sample flags before calling perf_prepare_sample().

The code is also available at 'perf/prepare-sample-v2' branch in

git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Thanks,
Namhyung

Namhyung Kim (8):
perf/core: Save the dynamic parts of sample data size perf/core: Add perf_sample_save_callchain() helper perf/core: Add perf_sample_save_raw_data() helper perf/core: Add perf_sample_save_brstack() helper perf/core: Set data->sample_flags in perf_prepare_sample() perf/core: Do not pass header for sample id init perf/core: Introduce perf_prepare_header() perf/core: Call perf_prepare_sample() before running BPF

 arch/powerpc/perf/core-book3s.c    |   3 +-
 arch/s390/kernel/perf_cpum_cf.c    |   4 +-
 arch/s390/kernel/perf_cpum_sf.c    |   3 +-
 arch/s390/kernel/perf_pai_crypto.c |   4 +-
 arch/s390/kernel/perf_pai_ext.c    |   4 +-
 arch/x86/events/amd/core.c         |   6 +-
 arch/x86/events/amd/ibs.c          |   9 +-
 arch/x86/events/intel/core.c       |   6 +-
 arch/x86/events/intel/ds.c         |  24 ++--
 include/linux/perf_event.h         | 133 +++++++++++++-----
 kernel/events/core.c               | 207 ++++++++++++++++-------------
 kernel/trace/bpf_trace.c           |   6 +-
 12 files changed, 236 insertions(+), 173 deletions(-)

base-commit: 9fcad995c6c52cc9791f7ee9f1386a5684055f9c

Signature · 1 line
2.39.0.314.g84b9a713c41-goog
author maintainer
#1 lore raw

PATCH 1/8 perf/core: Save the dynamic parts of sample data size

The perf sample data can be divided into parts. The event->header_size and event->id_header_size keep the static part of the sample data which is determined by the sample_type flags.

But other parts like CALLCHAIN and BRANCH_STACK are changing dynamically so it needs to see the actual data. In preparation of handling repeated calls for perf_prepare_sample(), it can save the dynamic size to the perf sample data to avoid the duplicate work.

---
 include/linux/perf_event.h |  2 ++
 kernel/events/core.c       | 17 ++++++++++-------
 2 files changed, 12 insertions(+), 7 deletions(-)
include/linux/perf_event.h +2 -0
index 03949d017ac9..16b980014449 100644
@@ -1103,6 +1103,7 @@ struct perf_sample_data {
 	 */
 	u64				sample_flags;
 	u64				period;
+	u64				dyn_size;
 
 	/*
 	 * Fields commonly set by __perf_event_header__init_id(),
@@ -1158,6 +1159,7 @@ static inline void perf_sample_data_init(struct perf_sample_data *data,
 	/* remaining struct members initialized in perf_prepare_sample() */
 	data->sample_flags = PERF_SAMPLE_PERIOD;
 	data->period = period;
+	data->dyn_size = 0;
 
 	if (addr) {
 		data->addr = addr;
kernel/events/core.c +10 -7
index eacc3702654d..8c8de26f04ab 100644
@@ -7593,7 +7593,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 
 		size += data->callchain->nr;
 
-		header->size += size * sizeof(u64);
+		data->dyn_size += size * sizeof(u64);
 	}
 
 	if (sample_type & PERF_SAMPLE_RAW) {
@@ -7619,7 +7619,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 			data->raw = NULL;
 		}
 
-		header->size += size;
+		data->dyn_size += size;
 	}
 
 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
@@ -7631,7 +7631,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 			size += data->br_stack->nr
 			      * sizeof(struct perf_branch_entry);
 		}
-		header->size += size;
+		data->dyn_size += size;
 	}
 
 	if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
@@ -7646,7 +7646,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 			size += hweight64(mask) * sizeof(u64);
 		}
 
-		header->size += size;
+		data->dyn_size += size;
 	}
 
 	if (sample_type & PERF_SAMPLE_STACK_USER) {
@@ -7671,7 +7671,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 			size += sizeof(u64) + stack_size;
 
 		data->stack_user_size = stack_size;
-		header->size += size;
+		data->dyn_size += size;
 	}
 
 	if (filtered_sample_type & PERF_SAMPLE_WEIGHT_TYPE)
@@ -7700,7 +7700,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 			size += hweight64(mask) * sizeof(u64);
 		}
 
-		header->size += size;
+		data->dyn_size += size;
 	}
 
 	if (sample_type & PERF_SAMPLE_PHYS_ADDR &&
@@ -7745,8 +7745,11 @@ void perf_prepare_sample(struct perf_event_header *header,
 		size = perf_prepare_sample_aux(event, data, size);
 
 		WARN_ON_ONCE(size + header->size > U16_MAX);
-		header->size += size;
+		data->dyn_size += size + sizeof(u64); /* size above */
 	}
+
+	header->size += data->dyn_size;
+
 	/*
 	 * If you're adding more sample types here, you likely need to do
 	 * something about the overflowing header::size, like repurpose the
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer Acked-by on v1 1/8
#2 lore raw

Re: [PATCH 1/8] perf/core: Save the dynamic parts of sample data size

Quoted 11 lines · Namhyung Kim

The perf sample data can be divided into parts. The event->header_size and event->id_header_size keep the static part of the sample data which is determined by the sample_type flags.

But other parts like CALLCHAIN and BRANCH_STACK are changing dynamically so it needs to see the actual data. In preparation of handling repeated calls for perf_prepare_sample(), it can save the dynamic size to the perf sample data to avoid the duplicate work.

author maintainer
#3 lore raw

PATCH 2/8 perf/core: Add perf_sample_save_callchain() helper

In reply to Namhyung Kim

When it saves the callchain to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_callchain() helper and convert all call sites.

---
 arch/x86/events/amd/ibs.c  |  6 ++----
 arch/x86/events/intel/ds.c | 12 ++++--------
 include/linux/perf_event.h | 16 +++++++++++++++-
 kernel/events/core.c       | 12 ++----------
 4 files changed, 23 insertions(+), 23 deletions(-)
arch/x86/events/amd/ibs.c +2 -4
index da3f5ebac4e1..417c80bd3274 100644
@@ -1122,10 +1122,8 @@ static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
 	 * recorded as part of interrupt regs. Thus we need to use rip from
 	 * interrupt regs while unwinding call stack.
 	 */
-	if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
-		data.callchain = perf_callchain(event, iregs);
-		data.sample_flags |= PERF_SAMPLE_CALLCHAIN;
-	}
+	if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
+		perf_sample_save_callchain(&data, event, iregs);
 
 	throttle = perf_event_overflow(event, &data, &regs);
 out:
arch/x86/events/intel/ds.c +4 -8
index 6ec326b47e2e..158cf845fc80 100644
@@ -1617,10 +1617,8 @@ static void setup_pebs_fixed_sample_data(struct perf_event *event,
 	 * previous PMI context or an (I)RET happened between the record and
 	 * PMI.
 	 */
-	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
-		data->callchain = perf_callchain(event, iregs);
-		data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
-	}
+	if (sample_type & PERF_SAMPLE_CALLCHAIN)
+		perf_sample_save_callchain(data, event, iregs);
 
 	/*
 	 * We use the interrupt regs as a base because the PEBS record does not
@@ -1795,10 +1793,8 @@ static void setup_pebs_adaptive_sample_data(struct perf_event *event,
 	 * previous PMI context or an (I)RET happened between the record and
 	 * PMI.
 	 */
-	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
-		data->callchain = perf_callchain(event, iregs);
-		data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
-	}
+	if (sample_type & PERF_SAMPLE_CALLCHAIN)
+		perf_sample_save_callchain(data, event, iregs);
 
 	*regs = *iregs;
 	/* The ip in basic is EventingIP */
include/linux/perf_event.h +15 -1
index 16b980014449..a9419608402b 100644
@@ -1095,6 +1095,8 @@ int perf_event_read_local(struct perf_event *event, u64 *value,
 extern u64 perf_event_read_value(struct perf_event *event,
 				 u64 *enabled, u64 *running);
 
+extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
+
 
 struct perf_sample_data {
 	/*
@@ -1167,6 +1169,19 @@ static inline void perf_sample_data_init(struct perf_sample_data *data,
 	}
 }
 
+static inline void perf_sample_save_callchain(struct perf_sample_data *data,
+					      struct perf_event *event,
+					      struct pt_regs *regs)
+{
+	int size = 1;
+
+	data->callchain = perf_callchain(event, regs);
+	size += data->callchain->nr;
+
+	data->dyn_size += size * sizeof(u64);
+	data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
+}
+
 /*
  * Clear all bitfields in the perf_branch_entry.
  * The to and from fields are not cleared because they are
@@ -1408,7 +1423,6 @@ extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct
 extern struct perf_callchain_entry *
 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
 		   u32 max_stack, bool crosstask, bool add_mark);
-extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
 extern int get_callchain_buffers(int max_stack);
 extern void put_callchain_buffers(void);
 extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
kernel/events/core.c +2 -10
index 8c8de26f04ab..0fba98b9cd65 100644
@@ -7585,16 +7585,8 @@ void perf_prepare_sample(struct perf_event_header *header,
 	if (sample_type & (PERF_SAMPLE_IP | PERF_SAMPLE_CODE_PAGE_SIZE))
 		data->ip = perf_instruction_pointer(regs);
 
-	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
-		int size = 1;
-
-		if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
-			data->callchain = perf_callchain(event, regs);
-
-		size += data->callchain->nr;
-
-		data->dyn_size += size * sizeof(u64);
-	}
+	if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
+		perf_sample_save_callchain(data, event, regs);
 
 	if (sample_type & PERF_SAMPLE_RAW) {
 		struct perf_raw_record *raw = data->raw;
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer Acked-by on v1 2/8
#4 lore raw

Re: [PATCH 2/8] perf/core: Add perf_sample_save_callchain() helper

Quoted 8 lines · Namhyung Kim

When it saves the callchain to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_callchain() helper and convert all call sites.

author maintainer
#5 lore raw

PATCH 3/8 perf/core: Add perf_sample_save_raw_data() helper

In reply to Namhyung Kim

When it saves the raw_data to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_raw_data() helper and convert all call sites.

---
 arch/s390/kernel/perf_cpum_cf.c    |  4 +---
 arch/s390/kernel/perf_pai_crypto.c |  4 +---
 arch/s390/kernel/perf_pai_ext.c    |  4 +---
 arch/x86/events/amd/ibs.c          |  3 +--
 include/linux/perf_event.h         | 33 +++++++++++++++++++++++++-----
 kernel/events/core.c               | 31 +++++-----------------------
 kernel/trace/bpf_trace.c           |  6 ++----
 7 files changed, 39 insertions(+), 46 deletions(-)
arch/s390/kernel/perf_cpum_cf.c +1 -3
index f043a7ff220b..aa38649c7c27 100644
@@ -662,9 +662,7 @@ static int cfdiag_push_sample(struct perf_event *event,
 	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
 		raw.frag.size = cpuhw->usedss;
 		raw.frag.data = cpuhw->stop;
-		raw.size = raw.frag.size;
-		data.raw = &raw;
-		data.sample_flags |= PERF_SAMPLE_RAW;
+		perf_sample_save_raw_data(&data, &raw);
 	}
 
 	overflow = perf_event_overflow(event, &data, &regs);
arch/s390/kernel/perf_pai_crypto.c +1 -3
index 985e243a2ed8..a7b339c4fd7c 100644
@@ -362,9 +362,7 @@ static int paicrypt_push_sample(void)
 	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
 		raw.frag.size = rawsize;
 		raw.frag.data = cpump->save;
-		raw.size = raw.frag.size;
-		data.raw = &raw;
-		data.sample_flags |= PERF_SAMPLE_RAW;
+		perf_sample_save_raw_data(&data, &raw);
 	}
 
 	overflow = perf_event_overflow(event, &data, &regs);
arch/s390/kernel/perf_pai_ext.c +1 -3
index 1138f57baae3..555597222bad 100644
@@ -451,9 +451,7 @@ static int paiext_push_sample(void)
 	if (event->attr.sample_type & PERF_SAMPLE_RAW) {
 		raw.frag.size = rawsize;
 		raw.frag.data = cpump->save;
-		raw.size = raw.frag.size;
-		data.raw = &raw;
-		data.sample_flags |= PERF_SAMPLE_RAW;
+		perf_sample_save_raw_data(&data, &raw);
 	}
 
 	overflow = perf_event_overflow(event, &data, &regs);
arch/x86/events/amd/ibs.c +1 -2
index 417c80bd3274..64582954b5f6 100644
@@ -1110,8 +1110,7 @@ static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
 				.data = ibs_data.data,
 			},
 		};
-		data.raw = &raw;
-		data.sample_flags |= PERF_SAMPLE_RAW;
+		perf_sample_save_raw_data(&data, &raw);
 	}
 
 	if (perf_ibs == &perf_ibs_op)
include/linux/perf_event.h +28 -5
index a9419608402b..569dfac5887f 100644
@@ -95,6 +95,11 @@ struct perf_raw_record {
 	u32				size;
 };
 
+static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
+{
+	return frag->pad < sizeof(u64);
+}
+
 /*
  * branch stack layout:
  *  nr: number of taken branches stored in entries[]
@@ -1182,6 +1187,29 @@ static inline void perf_sample_save_callchain(struct perf_sample_data *data,
 	data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
 }
 
+static inline void perf_sample_save_raw_data(struct perf_sample_data *data,
+					     struct perf_raw_record *raw)
+{
+	struct perf_raw_frag *frag = &raw->frag;
+	u32 sum = 0;
+	int size;
+
+	do {
+		sum += frag->size;
+		if (perf_raw_frag_last(frag))
+			break;
+		frag = frag->next;
+	} while (1);
+
+	size = round_up(sum + sizeof(u32), sizeof(u64));
+	raw->size = size - sizeof(u32);
+	frag->pad = raw->size - sum;
+
+	data->raw = raw;
+	data->dyn_size += size;
+	data->sample_flags |= PERF_SAMPLE_RAW;
+}
+
 /*
  * Clear all bitfields in the perf_branch_entry.
  * The to and from fields are not cleared because they are
@@ -1690,11 +1718,6 @@ extern void perf_restore_debug_store(void);
 static inline void perf_restore_debug_store(void)			{ }
 #endif
 
-static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
-{
-	return frag->pad < sizeof(u64);
-}
-
 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 
 struct perf_pmu_events_attr {
kernel/events/core.c +5 -26
index 0fba98b9cd65..133894ae5e30 100644
@@ -7588,30 +7588,10 @@ void perf_prepare_sample(struct perf_event_header *header,
 	if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
 		perf_sample_save_callchain(data, event, regs);
 
-	if (sample_type & PERF_SAMPLE_RAW) {
-		struct perf_raw_record *raw = data->raw;
-		int size;
-
-		if (raw && (data->sample_flags & PERF_SAMPLE_RAW)) {
-			struct perf_raw_frag *frag = &raw->frag;
-			u32 sum = 0;
-
-			do {
-				sum += frag->size;
-				if (perf_raw_frag_last(frag))
-					break;
-				frag = frag->next;
-			} while (1);
-
-			size = round_up(sum + sizeof(u32), sizeof(u64));
-			raw->size = size - sizeof(u32);
-			frag->pad = raw->size - sum;
-		} else {
-			size = sizeof(u64);
-			data->raw = NULL;
-		}
-
-		data->dyn_size += size;
+	if (filtered_sample_type & PERF_SAMPLE_RAW) {
+		data->raw = NULL;
+		data->dyn_size += sizeof(u64);
+		data->sample_flags |= PERF_SAMPLE_RAW;
 	}
 
 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
@@ -10127,8 +10107,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
 	};
 
 	perf_sample_data_init(&data, 0, 0);
-	data.raw = &raw;
-	data.sample_flags |= PERF_SAMPLE_RAW;
+	perf_sample_save_raw_data(&data, &raw);
 
 	perf_trace_buf_update(record, event_type);
 
kernel/trace/bpf_trace.c +2 -4
index 3bbd3f0c810c..ad37608afc35 100644
@@ -687,8 +687,7 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
 	}
 
 	perf_sample_data_init(sd, 0, 0);
-	sd->raw = &raw;
-	sd->sample_flags |= PERF_SAMPLE_RAW;
+	perf_sample_save_raw_data(sd, &raw);
 
 	err = __bpf_perf_event_output(regs, map, flags, sd);
 
@@ -746,8 +745,7 @@ 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);
-	sd->raw = &raw;
-	sd->sample_flags |= PERF_SAMPLE_RAW;
+	perf_sample_save_raw_data(sd, &raw);
 
 	ret = __bpf_perf_event_output(regs, map, flags, sd);
 out:
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer
#6 lore raw

Re: [PATCH 3/8] perf/core: Add perf_sample_save_raw_data() helper

Quoted 16 lines · Namhyung Kim
include/linux/perf_event.h +12 -0
@@ -1182,6 +1187,29 @@ static inline void perf_sample_save_callchain(struct perf_sample_data *data,
 	data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
 }
 
+static inline void perf_sample_save_raw_data(struct perf_sample_data *data,
+					     struct perf_raw_record *raw)
+{
+	struct perf_raw_frag *frag = &raw->frag;
+	u32 sum = 0;
+	int size;
+
+	do {
+		sum += frag->size;
+		if (perf_raw_frag_last(frag))
+			break;
+		frag = frag->next;

include/linux/perf_event.h static inline void perf_sample_save_callchain(struct perf_sample_data *data, · line 1211

Quoted 10 lines
include/linux/perf_event.h +10 -0
+	} while (1);
+
+	size = round_up(sum + sizeof(u32), sizeof(u64));
+	raw->size = size - sizeof(u32);
+	frag->pad = raw->size - sum;
+
+	data->raw = raw;
+	data->dyn_size += size;
+	data->sample_flags |= PERF_SAMPLE_RAW;
+}

This might be a wee big for inline, but I suppose it doesn't matter too much.

maintainer
#7 lore raw

Re: [PATCH 3/8] perf/core: Add perf_sample_save_raw_data() helper

In reply to Namhyung Kim
Quoted 164 lines · Namhyung Kim

When it saves the raw_data to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_raw_data() helper and convert all call sites.

---
 arch/s390/kernel/perf_cpum_cf.c    |  4 +---
 arch/s390/kernel/perf_pai_crypto.c |  4 +---
 arch/s390/kernel/perf_pai_ext.c    |  4 +---
 arch/x86/events/amd/ibs.c          |  3 +--
 include/linux/perf_event.h         | 33 +++++++++++++++++++++++++-----
 kernel/events/core.c               | 31 +++++-----------------------
 kernel/trace/bpf_trace.c           |  6 ++----
 7 files changed, 39 insertions(+), 46 deletions(-)
arch/s390/kernel/perf_cpum_cf.c +1 -3
index f043a7ff220b..aa38649c7c27 100644
@@ -662,9 +662,7 @@ static int cfdiag_push_sample(struct perf_event *event,
        if (event->attr.sample_type & PERF_SAMPLE_RAW) {
                raw.frag.size = cpuhw->usedss;
                raw.frag.data = cpuhw->stop;
-               raw.size = raw.frag.size;
-               data.raw = &raw;
-               data.sample_flags |= PERF_SAMPLE_RAW;
+               perf_sample_save_raw_data(&data, &raw);
        }
 
        overflow = perf_event_overflow(event, &data, &regs);
arch/s390/kernel/perf_pai_crypto.c +1 -3
index 985e243a2ed8..a7b339c4fd7c 100644
@@ -362,9 +362,7 @@ static int paicrypt_push_sample(void)
        if (event->attr.sample_type & PERF_SAMPLE_RAW) {
                raw.frag.size = rawsize;
                raw.frag.data = cpump->save;
-               raw.size = raw.frag.size;
-               data.raw = &raw;
-               data.sample_flags |= PERF_SAMPLE_RAW;
+               perf_sample_save_raw_data(&data, &raw);
        }
 
        overflow = perf_event_overflow(event, &data, &regs);
arch/s390/kernel/perf_pai_ext.c +1 -3
index 1138f57baae3..555597222bad 100644
@@ -451,9 +451,7 @@ static int paiext_push_sample(void)
        if (event->attr.sample_type & PERF_SAMPLE_RAW) {
                raw.frag.size = rawsize;
                raw.frag.data = cpump->save;
-               raw.size = raw.frag.size;
-               data.raw = &raw;
-               data.sample_flags |= PERF_SAMPLE_RAW;
+               perf_sample_save_raw_data(&data, &raw);
        }
 
        overflow = perf_event_overflow(event, &data, &regs);
arch/x86/events/amd/ibs.c +1 -2
index 417c80bd3274..64582954b5f6 100644
@@ -1110,8 +1110,7 @@ static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
                                .data = ibs_data.data,
                        },
                };
-               data.raw = &raw;
-               data.sample_flags |= PERF_SAMPLE_RAW;
+               perf_sample_save_raw_data(&data, &raw);
        }
 
        if (perf_ibs == &perf_ibs_op)
include/linux/perf_event.h +28 -5
index a9419608402b..569dfac5887f 100644
@@ -95,6 +95,11 @@ struct perf_raw_record {
        u32                             size;
 };
 
+static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
+{
+       return frag->pad < sizeof(u64);
+}
+
 /*
  * branch stack layout:
  *  nr: number of taken branches stored in entries[]
@@ -1182,6 +1187,29 @@ static inline void perf_sample_save_callchain(struct perf_sample_data *data,
        data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
 }
 
+static inline void perf_sample_save_raw_data(struct perf_sample_data *data,
+                                            struct perf_raw_record *raw)
+{
+       struct perf_raw_frag *frag = &raw->frag;
+       u32 sum = 0;
+       int size;
+
+       do {
+               sum += frag->size;
+               if (perf_raw_frag_last(frag))
+                       break;
+               frag = frag->next;
+       } while (1);
+
+       size = round_up(sum + sizeof(u32), sizeof(u64));
+       raw->size = size - sizeof(u32);
+       frag->pad = raw->size - sum;
+
+       data->raw = raw;
+       data->dyn_size += size;
+       data->sample_flags |= PERF_SAMPLE_RAW;
+}
+
 /*
  * Clear all bitfields in the perf_branch_entry.
  * The to and from fields are not cleared because they are
@@ -1690,11 +1718,6 @@ extern void perf_restore_debug_store(void);
 static inline void perf_restore_debug_store(void)                      { }
 #endif
 
-static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
-{
-       return frag->pad < sizeof(u64);
-}
-
 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 
 struct perf_pmu_events_attr {
kernel/events/core.c +0 -19
index 0fba98b9cd65..133894ae5e30 100644
@@ -7588,30 +7588,10 @@ void perf_prepare_sample(struct perf_event_header *header,
        if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
                perf_sample_save_callchain(data, event, regs);
 
-       if (sample_type & PERF_SAMPLE_RAW) {
-               struct perf_raw_record *raw = data->raw;
-               int size;
-
-               if (raw && (data->sample_flags & PERF_SAMPLE_RAW)) {
-                       struct perf_raw_frag *frag = &raw->frag;
-                       u32 sum = 0;
-
-                       do {
-                               sum += frag->size;
-                               if (perf_raw_frag_last(frag))
-                                       break;
-                               frag = frag->next;
-                       } while (1);
-
-                       size = round_up(sum + sizeof(u32), sizeof(u64));
-                       raw->size = size - sizeof(u32);
-                       frag->pad = raw->size - sum;
-               } else {

kernel/events/core.c void perf_prepare_sample(struct perf_event_header *header, · line 7595

Quoted 10 lines
kernel/events/core.c +4 -5
-                       size = sizeof(u64);
-                       data->raw = NULL;
-               }
-
-               data->dyn_size += size;
+       if (filtered_sample_type & PERF_SAMPLE_RAW) {
+               data->raw = NULL;
+               data->dyn_size += sizeof(u64);
+               data->sample_flags |= PERF_SAMPLE_RAW;
        }

I don't quite follow this change, and the commit log doesn't seem to cover this part.

Quoted 39 lines
kernel/events/core.c +1 -2
        if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
@@ -10127,8 +10107,7 @@ void perf_tp_event(u16 event_type, u64 count, void *record, int entry_size,
        };
 
        perf_sample_data_init(&data, 0, 0);
-       data.raw = &raw;
-       data.sample_flags |= PERF_SAMPLE_RAW;
+       perf_sample_save_raw_data(&data, &raw);
 
        perf_trace_buf_update(record, event_type);
 
kernel/trace/bpf_trace.c +2 -5
index 3bbd3f0c810c..ad37608afc35 100644
@@ -687,8 +687,7 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
        }
 
        perf_sample_data_init(sd, 0, 0);
-       sd->raw = &raw;
-       sd->sample_flags |= PERF_SAMPLE_RAW;
+       perf_sample_save_raw_data(sd, &raw);
 
        err = __bpf_perf_event_output(regs, map, flags, sd);
 
@@ -746,8 +745,7 @@ 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);
-       sd->raw = &raw;
-       sd->sample_flags |= PERF_SAMPLE_RAW;
+       perf_sample_save_raw_data(sd, &raw);
 
        ret = __bpf_perf_event_output(regs, map, flags, sd);
 out:
--

2.39.0.314.g84b9a713c41-goog

author maintainer
#8 lore raw

Re: [PATCH 3/8] perf/core: Add perf_sample_save_raw_data() helper

Hi Song,

Quoted 13 lines · Song Liu
Quoted 11 lines · Namhyung Kim

When it saves the raw_data to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_raw_data() helper and convert all call sites.

---

[SNIP]

Quoted 32 lines
Quoted 32 lines
kernel/events/core.c +0 -24
index 0fba98b9cd65..133894ae5e30 100644
@@ -7588,30 +7588,10 @@ void perf_prepare_sample(struct perf_event_header *header,
        if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
                perf_sample_save_callchain(data, event, regs);
 
-       if (sample_type & PERF_SAMPLE_RAW) {
-               struct perf_raw_record *raw = data->raw;
-               int size;
-
-               if (raw && (data->sample_flags & PERF_SAMPLE_RAW)) {
-                       struct perf_raw_frag *frag = &raw->frag;
-                       u32 sum = 0;
-
-                       do {
-                               sum += frag->size;
-                               if (perf_raw_frag_last(frag))
-                                       break;
-                               frag = frag->next;
-                       } while (1);
-
-                       size = round_up(sum + sizeof(u32), sizeof(u64));
-                       raw->size = size - sizeof(u32);
-                       frag->pad = raw->size - sum;
-               } else {
-                       size = sizeof(u64);
-                       data->raw = NULL;
-               }
-
-               data->dyn_size += size;
Quoted 8 lines

kernel/events/core.c void perf_prepare_sample(struct perf_event_header *header, · line 7595

Quoted 5 lines
kernel/events/core.c +4 -0
+       if (filtered_sample_type & PERF_SAMPLE_RAW) {
+               data->raw = NULL;
+               data->dyn_size += sizeof(u64);
+               data->sample_flags |= PERF_SAMPLE_RAW;
        }

I don't quite follow this change, and the commit log doesn't seem to cover this part.

It's for when the user requested RAW but no actual data. It assumes PMU drivers call perf_sample_save_raw_data() before perf_prepare_sample() if there's RAW data. So we can only handle the 'else' part in the original code.

Thanks,
Namhyung

maintainer
#9 lore raw

Re: [PATCH 3/8] perf/core: Add perf_sample_save_raw_data() helper

Quoted 60 lines · Namhyung Kim

Hi Song,

Quoted 13 lines · Song Liu
Quoted 11 lines · Namhyung Kim

When it saves the raw_data to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_raw_data() helper and convert all call sites.

---

[SNIP]

Quoted 32 lines
Quoted 32 lines
kernel/events/core.c +0 -24
index 0fba98b9cd65..133894ae5e30 100644
@@ -7588,30 +7588,10 @@ void perf_prepare_sample(struct perf_event_header *header,
        if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
                perf_sample_save_callchain(data, event, regs);
 
-       if (sample_type & PERF_SAMPLE_RAW) {
-               struct perf_raw_record *raw = data->raw;
-               int size;
-
-               if (raw && (data->sample_flags & PERF_SAMPLE_RAW)) {
-                       struct perf_raw_frag *frag = &raw->frag;
-                       u32 sum = 0;
-
-                       do {
-                               sum += frag->size;
-                               if (perf_raw_frag_last(frag))
-                                       break;
-                               frag = frag->next;
-                       } while (1);
-
-                       size = round_up(sum + sizeof(u32), sizeof(u64));
-                       raw->size = size - sizeof(u32);
-                       frag->pad = raw->size - sum;
-               } else {
-                       size = sizeof(u64);
-                       data->raw = NULL;
-               }
-
-               data->dyn_size += size;
Quoted 8 lines

kernel/events/core.c void perf_prepare_sample(struct perf_event_header *header, · line 7595

Quoted 5 lines
kernel/events/core.c +4 -0
+       if (filtered_sample_type & PERF_SAMPLE_RAW) {
+               data->raw = NULL;
+               data->dyn_size += sizeof(u64);
+               data->sample_flags |= PERF_SAMPLE_RAW;
        }

I don't quite follow this change, and the commit log doesn't seem to cover this part.

Quoted 4 lines

It's for when the user requested RAW but no actual data. It assumes PMU drivers call perf_sample_save_raw_data() before perf_prepare_sample() if there's RAW data. So we can only handle the 'else' part in the original code.

Got it. Thanks for the explanation.

Song

author maintainer
#10 lore raw

PATCH 4/8 perf/core: Add perf_sample_save_brstack() helper

In reply to Namhyung Kim

When it saves the branch stack to the perf sample data, it needs to update the sample flags and the dynamic size. To make sure this, add the perf_sample_save_brstack() helper and convert all call sites.

---
 arch/powerpc/perf/core-book3s.c |  3 +-
 arch/x86/events/amd/core.c      |  6 +--
 arch/x86/events/intel/core.c    |  6 +--
 arch/x86/events/intel/ds.c      |  9 ++---
 include/linux/perf_event.h      | 66 ++++++++++++++++++++-------------
 kernel/events/core.c            | 16 +++-----
 6 files changed, 53 insertions(+), 53 deletions(-)
arch/powerpc/perf/core-book3s.c +1 -2
index bf318dd9b709..8c1f7def596e 100644
@@ -2313,8 +2313,7 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
 			struct cpu_hw_events *cpuhw;
 			cpuhw = this_cpu_ptr(&cpu_hw_events);
 			power_pmu_bhrb_read(event, cpuhw);
-			data.br_stack = &cpuhw->bhrb_stack;
-			data.sample_flags |= PERF_SAMPLE_BRANCH_STACK;
+			perf_sample_save_brstack(&data, event, &cpuhw->bhrb_stack);
 		}
 
 		if (event->attr.sample_type & PERF_SAMPLE_DATA_SRC &&
arch/x86/events/amd/core.c +2 -4
index d6f3703e4119..463f3eb8bbd7 100644
@@ -928,10 +928,8 @@ static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
 		if (!x86_perf_event_set_period(event))
 			continue;
 
-		if (has_branch_stack(event)) {
-			data.br_stack = &cpuc->lbr_stack;
-			data.sample_flags |= PERF_SAMPLE_BRANCH_STACK;
-		}
+		if (has_branch_stack(event))
+			perf_sample_save_brstack(&data, event, &cpuc->lbr_stack);
 
 		if (perf_event_overflow(event, &data, regs))
 			x86_pmu_stop(event, 0);
arch/x86/events/intel/core.c +2 -4
index 29d2d0411caf..14f0a746257d 100644
@@ -3036,10 +3036,8 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
 
 		perf_sample_data_init(&data, 0, event->hw.last_period);
 
-		if (has_branch_stack(event)) {
-			data.br_stack = &cpuc->lbr_stack;
-			data.sample_flags |= PERF_SAMPLE_BRANCH_STACK;
-		}
+		if (has_branch_stack(event))
+			perf_sample_save_brstack(&data, event, &cpuc->lbr_stack);
 
 		if (perf_event_overflow(event, &data, regs))
 			x86_pmu_stop(event, 0);
arch/x86/events/intel/ds.c +3 -6
index 158cf845fc80..07c8a2cdc3ee 100644
@@ -1720,10 +1720,8 @@ static void setup_pebs_fixed_sample_data(struct perf_event *event,
 		data->sample_flags |= PERF_SAMPLE_TIME;
 	}
 
-	if (has_branch_stack(event)) {
-		data->br_stack = &cpuc->lbr_stack;
-		data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
-	}
+	if (has_branch_stack(event))
+		perf_sample_save_brstack(data, event, &cpuc->lbr_stack);
 }
 
 static void adaptive_pebs_save_regs(struct pt_regs *regs,
@@ -1883,8 +1881,7 @@ static void setup_pebs_adaptive_sample_data(struct perf_event *event,
 
 		if (has_branch_stack(event)) {
 			intel_pmu_store_pebs_lbrs(lbr);
-			data->br_stack = &cpuc->lbr_stack;
-			data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
+			perf_sample_save_brstack(data, event, &cpuc->lbr_stack);
 		}
 	}
 
include/linux/perf_event.h +40 -26

82 lines not on this page · raw message

kernel/events/core.c +5 -11
index 133894ae5e30..0218b6ffaf36 100644
@@ -7317,7 +7317,7 @@ void perf_output_sample(struct perf_output_handle *handle,
 	}
 
 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
-		if (data->sample_flags & PERF_SAMPLE_BRANCH_STACK) {
+		if (data->br_stack) {
 			size_t size;
 
 			size = data->br_stack->nr
@@ -7594,16 +7594,10 @@ void perf_prepare_sample(struct perf_event_header *header,
 		data->sample_flags |= PERF_SAMPLE_RAW;
 	}
 
-	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
-		int size = sizeof(u64); /* nr */
-		if (data->sample_flags & PERF_SAMPLE_BRANCH_STACK) {
-			if (branch_sample_hw_index(event))
-				size += sizeof(u64);
-
-			size += data->br_stack->nr
-			      * sizeof(struct perf_branch_entry);
-		}
-		data->dyn_size += size;
+	if (filtered_sample_type & PERF_SAMPLE_BRANCH_STACK) {
+		data->br_stack = NULL;
+		data->dyn_size += sizeof(u64);
+		data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
 	}
 
 	if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
author maintainer
#11 lore raw

PATCH 5/8 perf/core: Set data->sample_flags in perf_prepare_sample()

In reply to Namhyung Kim

The perf_prepare_sample() sets the perf_sample_data according to the attr->sample_type before copying it to the ring buffer. But BPF also wants to access the sample data so it needs to prepare the sample even before the regular path.

That means the perf_prepare_sample() can be called more than once. Set the data->sample_flags consistently so that it can indicate which fields are set already and skip them if sets.

Also update the filtered_sample_type to have the dependent flags to reduce the number of branches.

---
 kernel/events/core.c | 85 +++++++++++++++++++++++++++++++++-----------
 1 file changed, 65 insertions(+), 20 deletions(-)
kernel/events/core.c +65 -20
index 0218b6ffaf36..7c5555051662 100644
@@ -7053,6 +7053,14 @@ static void perf_aux_sample_output(struct perf_event *event,
 	ring_buffer_put(rb);
 }
 
+/*
+ * A set of common sample data types saved even for non-sample records
+ * when event->attr.sample_id_all is set.
+ */
+#define PERF_SAMPLE_ID_ALL  (PERF_SAMPLE_TID | PERF_SAMPLE_TIME |	\
+			     PERF_SAMPLE_CPU | PERF_SAMPLE_ID |		\
+			     PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
+
 static void __perf_event_header__init_id(struct perf_event_header *header,
 					 struct perf_sample_data *data,
 					 struct perf_event *event,
@@ -7561,6 +7569,11 @@ perf_callchain(struct perf_event *event, struct pt_regs *regs)
 	return callchain ?: &__empty_callchain;
 }
 
+static __always_inline u64 __cond_set(u64 flags, u64 s, u64 d)
+{
+	return d * !!(flags & s);
+}
+
 void perf_prepare_sample(struct perf_event_header *header,
 			 struct perf_sample_data *data,
 			 struct perf_event *event,
@@ -7576,14 +7589,25 @@ void perf_prepare_sample(struct perf_event_header *header,
 	header->misc |= perf_misc_flags(regs);
 
 	/*
-	 * Clear the sample flags that have already been done by the
-	 * PMU driver.
+	 * Add the sample flags that are dependent to others.  And clear the
+	 * sample flags that have already been done by the PMU driver.
 	 */
-	filtered_sample_type = sample_type & ~data->sample_flags;
+	filtered_sample_type = sample_type;
+	filtered_sample_type |= __cond_set(sample_type, PERF_SAMPLE_CODE_PAGE_SIZE,
+					   PERF_SAMPLE_IP);
+	filtered_sample_type |= __cond_set(sample_type, PERF_SAMPLE_DATA_PAGE_SIZE |
+					   PERF_SAMPLE_PHYS_ADDR, PERF_SAMPLE_ADDR);
+	filtered_sample_type |= __cond_set(sample_type, PERF_SAMPLE_STACK_USER,
+					   PERF_SAMPLE_REGS_USER);
+	filtered_sample_type &= ~data->sample_flags;
+
 	__perf_event_header__init_id(header, data, event, filtered_sample_type);
+	data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;
 
-	if (sample_type & (PERF_SAMPLE_IP | PERF_SAMPLE_CODE_PAGE_SIZE))
+	if (filtered_sample_type & PERF_SAMPLE_IP) {
 		data->ip = perf_instruction_pointer(regs);
+		data->sample_flags |= PERF_SAMPLE_IP;
+	}
 
 	if (filtered_sample_type & PERF_SAMPLE_CALLCHAIN)
 		perf_sample_save_callchain(data, event, regs);
@@ -7600,10 +7624,15 @@ void perf_prepare_sample(struct perf_event_header *header,
 		data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
 	}
 
-	if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
+	if (filtered_sample_type & PERF_SAMPLE_REGS_USER)
 		perf_sample_regs_user(&data->regs_user, regs);
 
-	if (sample_type & PERF_SAMPLE_REGS_USER) {
+	/*
+	 * It cannot use the filtered_sample_type here as REGS_USER can be set
+	 * by STACK_USER (using __cond_set() above) and we don't want to update
+	 * the dyn_size if it's not requested by users.
+	 */
+	if ((sample_type & ~data->sample_flags) & PERF_SAMPLE_REGS_USER) {
 		/* regs dump ABI info */
 		int size = sizeof(u64);
 
@@ -7613,9 +7642,10 @@ void perf_prepare_sample(struct perf_event_header *header,
 		}
 
 		data->dyn_size += size;
+		data->sample_flags |= PERF_SAMPLE_REGS_USER;
 	}
 
-	if (sample_type & PERF_SAMPLE_STACK_USER) {
+	if (filtered_sample_type & PERF_SAMPLE_STACK_USER) {
 		/*
 		 * Either we need PERF_SAMPLE_STACK_USER bit to be always
 		 * processed as the last one or have additional check added
@@ -7638,23 +7668,30 @@ void perf_prepare_sample(struct perf_event_header *header,
 
 		data->stack_user_size = stack_size;
 		data->dyn_size += size;
+		data->sample_flags |= PERF_SAMPLE_STACK_USER;
 	}
 
-	if (filtered_sample_type & PERF_SAMPLE_WEIGHT_TYPE)
+	if (filtered_sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
 		data->weight.full = 0;
+		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
+	}
 
-	if (filtered_sample_type & PERF_SAMPLE_DATA_SRC)
+	if (filtered_sample_type & PERF_SAMPLE_DATA_SRC) {
 		data->data_src.val = PERF_MEM_NA;
+		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
+	}
 
-	if (filtered_sample_type & PERF_SAMPLE_TRANSACTION)
+	if (filtered_sample_type & PERF_SAMPLE_TRANSACTION) {
 		data->txn = 0;
+		data->sample_flags |= PERF_SAMPLE_TRANSACTION;
+	}
 
-	if (sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR | PERF_SAMPLE_DATA_PAGE_SIZE)) {
-		if (filtered_sample_type & PERF_SAMPLE_ADDR)
-			data->addr = 0;
+	if (filtered_sample_type & PERF_SAMPLE_ADDR) {
+		data->addr = 0;
+		data->sample_flags |= PERF_SAMPLE_ADDR;
 	}
 
-	if (sample_type & PERF_SAMPLE_REGS_INTR) {
+	if (filtered_sample_type & PERF_SAMPLE_REGS_INTR) {
 		/* regs dump ABI info */
 		int size = sizeof(u64);
 
@@ -7667,19 +7704,22 @@ void perf_prepare_sample(struct perf_event_header *header,
 		}
 
 		data->dyn_size += size;
+		data->sample_flags |= PERF_SAMPLE_REGS_INTR;
 	}
 
-	if (sample_type & PERF_SAMPLE_PHYS_ADDR &&
-	    filtered_sample_type & PERF_SAMPLE_PHYS_ADDR)
+	if (filtered_sample_type & PERF_SAMPLE_PHYS_ADDR) {
 		data->phys_addr = perf_virt_to_phys(data->addr);
+		data->sample_flags |= PERF_SAMPLE_PHYS_ADDR;
+	}
 
 #ifdef CONFIG_CGROUP_PERF
-	if (sample_type & PERF_SAMPLE_CGROUP) {
+	if (filtered_sample_type & PERF_SAMPLE_CGROUP) {
 		struct cgroup *cgrp;
 
 		/* protected by RCU */
 		cgrp = task_css_check(current, perf_event_cgrp_id, 1)->cgroup;
 		data->cgroup = cgroup_id(cgrp);
+		data->sample_flags |= PERF_SAMPLE_CGROUP;
 	}
 #endif
 
@@ -7688,13 +7728,17 @@ void perf_prepare_sample(struct perf_event_header *header,
 	 * require PERF_SAMPLE_ADDR, kernel implicitly retrieve the data->addr,
 	 * but the value will not dump to the userspace.
 	 */
-	if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)
+	if (filtered_sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) {
 		data->data_page_size = perf_get_page_size(data->addr);
+		data->sample_flags |= PERF_SAMPLE_DATA_PAGE_SIZE;
+	}
 
-	if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)
+	if (filtered_sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) {
 		data->code_page_size = perf_get_page_size(data->ip);
+		data->sample_flags |= PERF_SAMPLE_CODE_PAGE_SIZE;
+	}
 
-	if (sample_type & PERF_SAMPLE_AUX) {
+	if (filtered_sample_type & PERF_SAMPLE_AUX) {
 		u64 size;
 
 		header->size += sizeof(u64); /* size */
@@ -7712,6 +7756,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 
 		WARN_ON_ONCE(size + header->size > U16_MAX);
 		data->dyn_size += size + sizeof(u64); /* size above */
+		data->sample_flags |= PERF_SAMPLE_AUX;
 	}
 
 	header->size += data->dyn_size;
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer
#12 lore raw

Re: [PATCH 5/8] perf/core: Set data->sample_flags in perf_prepare_sample()

Quoted 8 lines · Namhyung Kim
kernel/events/core.c +2 -0
@@ -7053,6 +7053,14 @@ static void perf_aux_sample_output(struct perf_event *event,
 	ring_buffer_put(rb);
 }
 
+/*
+ * A set of common sample data types saved even for non-sample records

kernel/events/core.c static void perf_aux_sample_output(struct perf_event *event,

Quoted 10 lines
kernel/events/core.c +7 -0
+ * when event->attr.sample_id_all is set.
+ */
+#define PERF_SAMPLE_ID_ALL  (PERF_SAMPLE_TID | PERF_SAMPLE_TIME |	\
+			     PERF_SAMPLE_CPU | PERF_SAMPLE_ID |		\
+			     PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
+
 
 
 	__perf_event_header__init_id(header, data, event, filtered_sample_type);
+	data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;

Why not update __perf_event_header__init_id() ?

author maintainer
#13 lore raw

Re: [PATCH 5/8] perf/core: Set data->sample_flags in perf_prepare_sample()

Quoted 19 lines · Peter Zijlstra
Quoted 14 lines · Namhyung Kim
kernel/events/core.c +8 -0
@@ -7053,6 +7053,14 @@ static void perf_aux_sample_output(struct perf_event *event,
      ring_buffer_put(rb);
 }
 
+/*
+ * A set of common sample data types saved even for non-sample records
+ * when event->attr.sample_id_all is set.
+ */
+#define PERF_SAMPLE_ID_ALL  (PERF_SAMPLE_TID | PERF_SAMPLE_TIME |    \
+                          PERF_SAMPLE_CPU | PERF_SAMPLE_ID |         \
+                          PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
+
Quoted 4 lines

kernel/events/core.c void perf_prepare_sample(struct perf_event_header *header, · line 7605

Quoted 2 lines
kernel/events/core.c +1 -0
      __perf_event_header__init_id(header, data, event, filtered_sample_type);
+     data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;

Why not update __perf_event_header__init_id() ?

My concern was the non-sample records which don't check the sample flags at all. So I'd like to skip the write for them.

But that's not a big deal. I can move it if you want.

Thanks,
Namhyung

author maintainer
#14 lore raw

PATCH 6/8 perf/core: Do not pass header for sample id init

In reply to Namhyung Kim

The only thing it does for header in __perf_event_header__init_id() is to update the header size with event->id_header_size. We can do this outside and get rid of the argument for the later change.

---
 kernel/events/core.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
kernel/events/core.c +7 -7
index 7c5555051662..127dddd20f93 100644
@@ -7061,13 +7061,11 @@ static void perf_aux_sample_output(struct perf_event *event,
 			     PERF_SAMPLE_CPU | PERF_SAMPLE_ID |		\
 			     PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_STREAM_ID)
 
-static void __perf_event_header__init_id(struct perf_event_header *header,
-					 struct perf_sample_data *data,
+static void __perf_event_header__init_id(struct perf_sample_data *data,
 					 struct perf_event *event,
 					 u64 sample_type)
 {
 	data->type = event->attr.sample_type;
-	header->size += event->id_header_size;
 
 	if (sample_type & PERF_SAMPLE_TID) {
 		/* namespace issues */
@@ -7094,8 +7092,10 @@ void perf_event_header__init_id(struct perf_event_header *header,
 				struct perf_sample_data *data,
 				struct perf_event *event)
 {
-	if (event->attr.sample_id_all)
-		__perf_event_header__init_id(header, data, event, event->attr.sample_type);
+	if (event->attr.sample_id_all) {
+		header->size += event->id_header_size;
+		__perf_event_header__init_id(data, event, event->attr.sample_type);
+	}
 }
 
 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
@@ -7583,7 +7583,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 	u64 filtered_sample_type;
 
 	header->type = PERF_RECORD_SAMPLE;
-	header->size = sizeof(*header) + event->header_size;
+	header->size = sizeof(*header) + event->header_size + event->id_header_size;
 
 	header->misc = 0;
 	header->misc |= perf_misc_flags(regs);
@@ -7601,7 +7601,7 @@ void perf_prepare_sample(struct perf_event_header *header,
 					   PERF_SAMPLE_REGS_USER);
 	filtered_sample_type &= ~data->sample_flags;
 
-	__perf_event_header__init_id(header, data, event, filtered_sample_type);
+	__perf_event_header__init_id(data, event, filtered_sample_type);
 	data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;
 
 	if (filtered_sample_type & PERF_SAMPLE_IP) {
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer Acked-by on v1 6/8
#15 lore raw

Re: [PATCH 6/8] perf/core: Do not pass header for sample id init

Quoted 6 lines · Namhyung Kim

The only thing it does for header in __perf_event_header__init_id() is to update the header size with event->id_header_size. We can do this outside and get rid of the argument for the later change.

author maintainer
#16 lore raw

PATCH 7/8 perf/core: Introduce perf_prepare_header()

In reply to Namhyung Kim

Factor out perf_prepare_header() so that it can call perf_prepare_sample() without a header if not needed.

Also it checks the filtered_sample_type to avoid duplicate work when perf_prepare_sample() is called twice (or more).

---
 arch/s390/kernel/perf_cpum_sf.c |  3 ++-
 arch/x86/events/intel/ds.c      |  3 ++-
 include/linux/perf_event.h      | 16 +++++++++++++-
 kernel/events/core.c            | 38 +++++++++++++++++++++------------
 4 files changed, 43 insertions(+), 17 deletions(-)
arch/s390/kernel/perf_cpum_sf.c +2 -1
index 332a49965130..fd02f8423243 100644
@@ -671,7 +671,8 @@ static void cpumsf_output_event_pid(struct perf_event *event,
 	/* Protect callchain buffers, tasks */
 	rcu_read_lock();
 
-	perf_prepare_sample(&header, data, event, regs);
+	perf_prepare_sample(data, event, regs);
+	perf_prepare_header(&header, data, event, regs);
 	if (perf_output_begin(&handle, data, event, header.size))
 		goto out;
 
arch/x86/events/intel/ds.c +2 -1
index 07c8a2cdc3ee..183efa914b99 100644
@@ -807,7 +807,8 @@ int intel_pmu_drain_bts_buffer(void)
 	 * the sample.
 	 */
 	rcu_read_lock();
-	perf_prepare_sample(&header, &data, event, &regs);
+	perf_prepare_sample(&data, event, &regs);
+	perf_prepare_header(&header, &data, event, &regs);
 
 	if (perf_output_begin(&handle, &data, event,
 			      header.size * (top - base - skip)))
include/linux/perf_event.h +15 -1
index 7db0e9cc2682..d5628a7b5eaa 100644
@@ -1250,6 +1250,17 @@ static inline void perf_sample_save_brstack(struct perf_sample_data *data,
 	data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
 }
 
+static inline u32 perf_sample_data_size(struct perf_sample_data *data,
+					struct perf_event *event)
+{
+	u32 size = sizeof(struct perf_event_header);
+
+	size += event->header_size + event->id_header_size;
+	size += data->dyn_size;
+
+	return size;
+}
+
 /*
  * Clear all bitfields in the perf_branch_entry.
  * The to and from fields are not cleared because they are
@@ -1271,7 +1282,10 @@ extern void perf_output_sample(struct perf_output_handle *handle,
 			       struct perf_event_header *header,
 			       struct perf_sample_data *data,
 			       struct perf_event *event);
-extern void perf_prepare_sample(struct perf_event_header *header,
+extern void perf_prepare_sample(struct perf_sample_data *data,
+				struct perf_event *event,
+				struct pt_regs *regs);
+extern void perf_prepare_header(struct perf_event_header *header,
 				struct perf_sample_data *data,
 				struct perf_event *event,
 				struct pt_regs *regs);
kernel/events/core.c +24 -14
index 127dddd20f93..5c4f3fa3d2b7 100644
@@ -7574,20 +7574,13 @@ static __always_inline u64 __cond_set(u64 flags, u64 s, u64 d)
 	return d * !!(flags & s);
 }
 
-void perf_prepare_sample(struct perf_event_header *header,
-			 struct perf_sample_data *data,
+void perf_prepare_sample(struct perf_sample_data *data,
 			 struct perf_event *event,
 			 struct pt_regs *regs)
 {
 	u64 sample_type = event->attr.sample_type;
 	u64 filtered_sample_type;
 
-	header->type = PERF_RECORD_SAMPLE;
-	header->size = sizeof(*header) + event->header_size + event->id_header_size;
-
-	header->misc = 0;
-	header->misc |= perf_misc_flags(regs);
-
 	/*
 	 * Add the sample flags that are dependent to others.  And clear the
 	 * sample flags that have already been done by the PMU driver.
@@ -7601,6 +7594,12 @@ void perf_prepare_sample(struct perf_event_header *header,
 					   PERF_SAMPLE_REGS_USER);
 	filtered_sample_type &= ~data->sample_flags;
 
+	if (filtered_sample_type == 0) {
+		/* Make sure it has the correct data->type for output */
+		data->type = event->attr.sample_type;
+		return;
+	}
+
 	__perf_event_header__init_id(data, event, filtered_sample_type);
 	data->sample_flags |= sample_type & PERF_SAMPLE_ID_ALL;
 
@@ -7653,9 +7652,10 @@ void perf_prepare_sample(struct perf_event_header *header,
 		 * up the rest of the sample size.
 		 */
 		u16 stack_size = event->attr.sample_stack_user;
+		u16 header_size = perf_sample_data_size(data, event);
 		u16 size = sizeof(u64);
 
-		stack_size = perf_sample_ustack_size(stack_size, header->size,
+		stack_size = perf_sample_ustack_size(stack_size, header_size,
 						     data->regs_user.regs);
 
 		/*
@@ -7740,8 +7740,9 @@ void perf_prepare_sample(struct perf_event_header *header,
 
 	if (filtered_sample_type & PERF_SAMPLE_AUX) {
 		u64 size;
+		u16 header_size = perf_sample_data_size(data, event);
 
-		header->size += sizeof(u64); /* size */
+		header_size += sizeof(u64); /* size */
 
 		/*
 		 * Given the 16bit nature of header::size, an AUX sample can
@@ -7749,17 +7750,25 @@ void perf_prepare_sample(struct perf_event_header *header,
 		 * Make sure this doesn't happen by using up to U16_MAX bytes
 		 * per sample in total (rounded down to 8 byte boundary).
 		 */
-		size = min_t(size_t, U16_MAX - header->size,
+		size = min_t(size_t, U16_MAX - header_size,
 			     event->attr.aux_sample_size);
 		size = rounddown(size, 8);
 		size = perf_prepare_sample_aux(event, data, size);
 
-		WARN_ON_ONCE(size + header->size > U16_MAX);
+		WARN_ON_ONCE(size + header_size > U16_MAX);
 		data->dyn_size += size + sizeof(u64); /* size above */
 		data->sample_flags |= PERF_SAMPLE_AUX;
 	}
+}
 
-	header->size += data->dyn_size;
+void perf_prepare_header(struct perf_event_header *header,
+			 struct perf_sample_data *data,
+			 struct perf_event *event,
+			 struct pt_regs *regs)
+{
+	header->type = PERF_RECORD_SAMPLE;
+	header->size = perf_sample_data_size(data, event);
+	header->misc = perf_misc_flags(regs);
 
 	/*
 	 * If you're adding more sample types here, you likely need to do
@@ -7788,7 +7797,8 @@ __perf_event_output(struct perf_event *event,
 	/* protect the callchain buffers */
 	rcu_read_lock();
 
-	perf_prepare_sample(&header, data, event, regs);
+	perf_prepare_sample(data, event, regs);
+	perf_prepare_header(&header, data, event, regs);
 
 	err = output_begin(&handle, data, event, header.size);
 	if (err)
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer Acked-by on v1 7/8
#17 lore raw

Re: [PATCH 7/8] perf/core: Introduce perf_prepare_header()

Quoted 11 lines · Namhyung Kim

Factor out perf_prepare_header() so that it can call perf_prepare_sample() without a header if not needed.

Also it checks the filtered_sample_type to avoid duplicate work when perf_prepare_sample() is called twice (or more).

author maintainer
#18 lore raw

PATCH 8/8 perf/core: Call perf_prepare_sample() before running BPF

In reply to Namhyung Kim

As BPF can access sample data, it needs to populate the data. Also remove the logic to get the callchain specifically as it's covered by the perf_prepare_sample() now.

---
 kernel/events/core.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)
kernel/events/core.c +1 -7
index 5c4f3fa3d2b7..af8365fb639a 100644
@@ -10363,13 +10363,7 @@ static void bpf_overflow_handler(struct perf_event *event,
 	rcu_read_lock();
 	prog = READ_ONCE(event->prog);
 	if (prog) {
-		if (prog->call_get_stack &&
-		    (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) &&
-		    !(data->sample_flags & PERF_SAMPLE_CALLCHAIN)) {
-			data->callchain = perf_callchain(event, regs);
-			data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
-		}
-
+		perf_prepare_sample(data, event, regs);
 		ret = bpf_prog_run(prog, &ctx);
 	}
 	rcu_read_unlock();
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
maintainer Acked-by on v1 8/8
#19 lore raw

Re: [PATCH 8/8] perf/core: Call perf_prepare_sample() before running BPF

Quoted 8 lines · Namhyung Kim

As BPF can access sample data, it needs to populate the data. Also remove the logic to get the callchain specifically as it's covered by the perf_prepare_sample() now.

Quoted 26 lines
---
 kernel/events/core.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)
kernel/events/core.c +1 -8
index 5c4f3fa3d2b7..af8365fb639a 100644
@@ -10363,13 +10363,7 @@ static void bpf_overflow_handler(struct perf_event *event,
        rcu_read_lock();
        prog = READ_ONCE(event->prog);
        if (prog) {
-               if (prog->call_get_stack &&
-                   (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) &&
-                   !(data->sample_flags & PERF_SAMPLE_CALLCHAIN)) {
-                       data->callchain = perf_callchain(event, regs);
-                       data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
-               }
-
+               perf_prepare_sample(data, event, regs);
                ret = bpf_prog_run(prog, &ctx);
        }
        rcu_read_unlock();
--

2.39.0.314.g84b9a713c41-goog

maintainer
#20 lore raw
In reply to Namhyung Kim
Quoted 9 lines · Namhyung Kim

Namhyung Kim (8):
perf/core: Save the dynamic parts of sample data size perf/core: Add perf_sample_save_callchain() helper perf/core: Add perf_sample_save_raw_data() helper perf/core: Add perf_sample_save_brstack() helper perf/core: Set data->sample_flags in perf_prepare_sample() perf/core: Do not pass header for sample id init perf/core: Introduce perf_prepare_header() perf/core: Call perf_prepare_sample() before running BPF

Aside from a few small niggles, this looks really good, Thanks!

#21 lore raw
In reply to Namhyung Kim
Quoted 38 lines · Namhyung Kim

Hello,

The perf_prepare_sample() is to fill the perf sample data and update the header info before sending it to the ring buffer. But we want to use it for BPF overflow handler so that it can access the sample data to filter relevant ones.

Changes in v2)
* the layout change is merged
* reduce branches using __cond_set (PeterZ)
* add helpers to set dynamic sample data (PeterZ)
* introduce perf_prepare_header() (PeterZ)
* call perf_prepare_sample() before bpf_overflow_handler unconditionally

This means the perf_prepare_handler() can be called more than once. To avoid duplicate work, use the data->sample_flags and save the data size.

I also added a few of helpers to set those information accordingly. But it looks some fields like REGS_USER, STACK_USER and AUX are saved in the perf_prepare_sample() so I didn't add the helpers for them.

After than we can just check the filtered_sample_type flags begin zero to determine if it has more work. In that case, it needs to update the data->type since it's possible to miss when PMU driver sets all required sample flags before calling perf_prepare_sample().

The code is also available at 'perf/prepare-sample-v2' branch in

git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Thanks,
Namhyung

Quoted 9 lines

Namhyung Kim (8):
perf/core: Save the dynamic parts of sample data size perf/core: Add perf_sample_save_callchain() helper perf/core: Add perf_sample_save_raw_data() helper perf/core: Add perf_sample_save_brstack() helper perf/core: Set data->sample_flags in perf_prepare_sample() perf/core: Do not pass header for sample id init perf/core: Introduce perf_prepare_header() perf/core: Call perf_prepare_sample() before running BPF

lgtm, I ran the bpf selftests on top of that and it's ok

jirka

Quoted 20 lines
 arch/powerpc/perf/core-book3s.c    |   3 +-
 arch/s390/kernel/perf_cpum_cf.c    |   4 +-
 arch/s390/kernel/perf_cpum_sf.c    |   3 +-
 arch/s390/kernel/perf_pai_crypto.c |   4 +-
 arch/s390/kernel/perf_pai_ext.c    |   4 +-
 arch/x86/events/amd/core.c         |   6 +-
 arch/x86/events/amd/ibs.c          |   9 +-
 arch/x86/events/intel/core.c       |   6 +-
 arch/x86/events/intel/ds.c         |  24 ++--
 include/linux/perf_event.h         | 133 +++++++++++++-----
 kernel/events/core.c               | 207 ++++++++++++++++-------------
 kernel/trace/bpf_trace.c           |   6 +-
 12 files changed, 236 insertions(+), 173 deletions(-)

base-commit: 9fcad995c6c52cc9791f7ee9f1386a5684055f9c

Signature · 1 line
2.39.0.314.g84b9a713c41-goog
author maintainer
#22 lore raw

Hi Jiri,

Quoted 50 lines · Jiri Olsa
Quoted 38 lines · Namhyung Kim

Hello,

The perf_prepare_sample() is to fill the perf sample data and update the header info before sending it to the ring buffer. But we want to use it for BPF overflow handler so that it can access the sample data to filter relevant ones.

Changes in v2)
* the layout change is merged
* reduce branches using __cond_set (PeterZ)
* add helpers to set dynamic sample data (PeterZ)
* introduce perf_prepare_header() (PeterZ)
* call perf_prepare_sample() before bpf_overflow_handler unconditionally

This means the perf_prepare_handler() can be called more than once. To avoid duplicate work, use the data->sample_flags and save the data size.

I also added a few of helpers to set those information accordingly. But it looks some fields like REGS_USER, STACK_USER and AUX are saved in the perf_prepare_sample() so I didn't add the helpers for them.

After than we can just check the filtered_sample_type flags begin zero to determine if it has more work. In that case, it needs to update the data->type since it's possible to miss when PMU driver sets all required sample flags before calling perf_prepare_sample().

The code is also available at 'perf/prepare-sample-v2' branch in

git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Thanks,
Namhyung

Quoted 9 lines

Namhyung Kim (8):
perf/core: Save the dynamic parts of sample data size perf/core: Add perf_sample_save_callchain() helper perf/core: Add perf_sample_save_raw_data() helper perf/core: Add perf_sample_save_brstack() helper perf/core: Set data->sample_flags in perf_prepare_sample() perf/core: Do not pass header for sample id init perf/core: Introduce perf_prepare_header() perf/core: Call perf_prepare_sample() before running BPF

Quoted 1 line

lgtm, I ran the bpf selftests on top of that and it's ok

Thanks for your review. I'll add your Acked-by and Tested-by in the v3. :)

Thanks,
Namhyung