[PATCH V10 1/4] perf/x86/intel: Apply static call for drain_pebs

[PATCH V10 1/4] perf/x86/intel: Apply static call for drain_pebs

[kan.liang]

From: "Peter Zijlstra (Intel)" <peterz@infradead.org>

The x86_pmu_drain_pebs static call was introduced in commit 7c9903c9bf71
("x86/perf, static_call: Optimize x86_pmu methods"), but it's not really
used to replace the old method.

Apply the static call for drain_pebs.

Fixes: 7c9903c9bf71 ("x86/perf, static_call: Optimize x86_pmu methods")
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
---

New for V10

 arch/x86/events/intel/core.c | 2 +-
 arch/x86/events/intel/ds.c   | 2 +-
 arch/x86/events/perf_event.h | 1 +
 3 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index 2a2824e9c50d..4daa45ae9bd2 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -3066,7 +3066,7 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
 
 		handled++;
 		x86_pmu_handle_guest_pebs(regs, &data);
-		x86_pmu.drain_pebs(regs, &data);
+		static_call(x86_pmu_drain_pebs)(regs, &data);
 		status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
 
 		/*
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index ba74e1198328..322963b02a91 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -957,7 +957,7 @@ static inline void intel_pmu_drain_pebs_buffer(void)
 {
 	struct perf_sample_data data;
 
-	x86_pmu.drain_pebs(NULL, &data);
+	static_call(x86_pmu_drain_pebs)(NULL, &data);
 }
 
 /*
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index 31c2771545a6..084e9196b458 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -1107,6 +1107,7 @@ extern struct x86_pmu x86_pmu __read_mostly;
 
 DECLARE_STATIC_CALL(x86_pmu_set_period, *x86_pmu.set_period);
 DECLARE_STATIC_CALL(x86_pmu_update,     *x86_pmu.update);
+DECLARE_STATIC_CALL(x86_pmu_drain_pebs,	*x86_pmu.drain_pebs);
 
 static __always_inline struct x86_perf_task_context_opt *task_context_opt(void *ctx)
 {
-- 
2.38.1

[PATCH V10 2/4] perf/x86/intel: Avoid disable PMU if !cpuc->enabled in sample read

[kan.liang]

From: Kan Liang <kan.liang@linux.intel.com>

The WARN_ON(this_cpu_read(cpu_hw_events.enabled)) in the
intel_pmu_save_and_restart_reload() is triggered, when sampling read
topdown events.

In a NMI handler, the cpu_hw_events.enabled is set and used to indicate
the status of core PMU. The generic pmu->pmu_disable_count, updated in
the perf_pmu_disable/enable pair, is not touched.
However, the perf_pmu_disable/enable pair is invoked when sampling read
in a NMI handler. The cpuc->enabled is mistakenly set by the
perf_pmu_enable().

Avoid disabling PMU if the core PMU is already disabled.
Merge the logic together.

Fixes: 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: stable@vger.kernel.org
---

Changes since V9:
- Merge the read_event related codes together

 arch/x86/events/intel/core.c | 41 ++++++++++++++++++++----------------
 arch/x86/events/intel/ds.c   | 11 +---------
 arch/x86/events/perf_event.h |  2 +-
 3 files changed, 25 insertions(+), 29 deletions(-)

diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index 4daa45ae9bd2..762b140c4953 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -2775,28 +2775,33 @@ static u64 icl_update_topdown_event(struct perf_event *event)
 
 DEFINE_STATIC_CALL(intel_pmu_update_topdown_event, x86_perf_event_update);
 
-static void intel_pmu_read_topdown_event(struct perf_event *event)
+static void intel_pmu_read_event(struct perf_event *event)
 {
-	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	if (event->hw.flags & (PERF_X86_EVENT_AUTO_RELOAD | PERF_X86_EVENT_TOPDOWN)) {
+		struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+		bool pmu_enabled = cpuc->enabled;
 
-	/* Only need to call update_topdown_event() once for group read. */
-	if ((cpuc->txn_flags & PERF_PMU_TXN_READ) &&
-	    !is_slots_event(event))
-		return;
+		/* Only need to call update_topdown_event() once for group read. */
+		if (is_metric_event(event) && (cpuc->txn_flags & PERF_PMU_TXN_READ))
+			return;
 
-	perf_pmu_disable(event->pmu);
-	static_call(intel_pmu_update_topdown_event)(event);
-	perf_pmu_enable(event->pmu);
-}
+		cpuc->enabled = 0;
+		if (pmu_enabled)
+			intel_pmu_disable_all();
 
-static void intel_pmu_read_event(struct perf_event *event)
-{
-	if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
-		intel_pmu_auto_reload_read(event);
-	else if (is_topdown_count(event))
-		intel_pmu_read_topdown_event(event);
-	else
-		x86_perf_event_update(event);
+		if (is_topdown_event(event))
+			static_call(intel_pmu_update_topdown_event)(event);
+		else
+			intel_pmu_drain_pebs_buffer();
+
+		cpuc->enabled = pmu_enabled;
+		if (pmu_enabled)
+			intel_pmu_enable_all(0);
+
+		return;
+	}
+
+	x86_perf_event_update(event);
 }
 
 static void intel_pmu_enable_fixed(struct perf_event *event)
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 322963b02a91..eb14b46423e5 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -953,7 +953,7 @@ int intel_pmu_drain_bts_buffer(void)
 	return 1;
 }
 
-static inline void intel_pmu_drain_pebs_buffer(void)
+void intel_pmu_drain_pebs_buffer(void)
 {
 	struct perf_sample_data data;
 
@@ -2094,15 +2094,6 @@ get_next_pebs_record_by_bit(void *base, void *top, int bit)
 	return NULL;
 }
 
-void intel_pmu_auto_reload_read(struct perf_event *event)
-{
-	WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
-
-	perf_pmu_disable(event->pmu);
-	intel_pmu_drain_pebs_buffer();
-	perf_pmu_enable(event->pmu);
-}
-
 /*
  * Special variant of intel_pmu_save_and_restart() for auto-reload.
  */
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index 084e9196b458..536a112f6353 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -1644,7 +1644,7 @@ void intel_pmu_pebs_disable_all(void);
 
 void intel_pmu_pebs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in);
 
-void intel_pmu_auto_reload_read(struct perf_event *event);
+void intel_pmu_drain_pebs_buffer(void);
 
 void intel_pmu_store_pebs_lbrs(struct lbr_entry *lbr);
 
-- 
2.38.1

[PATCH V10 3/4] perf: Avoid the read if the count is already updated

[kan.liang]

From: "Peter Zijlstra (Intel)" <peterz@infradead.org>

The event may have been updated in the PMU-specific implementation,
e.g., Intel PEBS counters snapshotting. The common code should not
read and overwrite the value.

The PERF_SAMPLE_READ in the data->sample_type can be used to detect
whether the PMU-specific value is available. If yes, avoid the
pmu->read() in the common code. Add a new flag, skip_read, to track the
case.

Factor out a perf_pmu_read() to clean up the code.

Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---

No changes since V9

 include/linux/perf_event.h  |  8 +++++++-
 kernel/events/core.c        | 33 ++++++++++++++++-----------------
 kernel/events/ring_buffer.c |  1 +
 3 files changed, 24 insertions(+), 18 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 8333f132f4a9..2d07bc1193f3 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -1062,7 +1062,13 @@ struct perf_output_handle {
 	struct perf_buffer		*rb;
 	unsigned long			wakeup;
 	unsigned long			size;
-	u64				aux_flags;
+	union {
+		u64			flags;		/* perf_output*() */
+		u64			aux_flags;	/* perf_aux_output*() */
+		struct {
+			u64		skip_read : 1;
+		};
+	};
 	union {
 		void			*addr;
 		unsigned long		head;
diff --git a/kernel/events/core.c b/kernel/events/core.c
index b2bc67791f84..f91ba29048ce 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -1191,6 +1191,12 @@ static void perf_assert_pmu_disabled(struct pmu *pmu)
 	WARN_ON_ONCE(*this_cpu_ptr(pmu->pmu_disable_count) == 0);
 }
 
+static inline void perf_pmu_read(struct perf_event *event)
+{
+	if (event->state == PERF_EVENT_STATE_ACTIVE)
+		event->pmu->read(event);
+}
+
 static void get_ctx(struct perf_event_context *ctx)
 {
 	refcount_inc(&ctx->refcount);
@@ -3473,8 +3479,7 @@ static void __perf_event_sync_stat(struct perf_event *event,
 	 * we know the event must be on the current CPU, therefore we
 	 * don't need to use it.
 	 */
-	if (event->state == PERF_EVENT_STATE_ACTIVE)
-		event->pmu->read(event);
+	perf_pmu_read(event);
 
 	perf_event_update_time(event);
 
@@ -4618,15 +4623,8 @@ static void __perf_event_read(void *info)
 
 	pmu->read(event);
 
-	for_each_sibling_event(sub, event) {
-		if (sub->state == PERF_EVENT_STATE_ACTIVE) {
-			/*
-			 * Use sibling's PMU rather than @event's since
-			 * sibling could be on different (eg: software) PMU.
-			 */
-			sub->pmu->read(sub);
-		}
-	}
+	for_each_sibling_event(sub, event)
+		perf_pmu_read(sub);
 
 	data->ret = pmu->commit_txn(pmu);
 
@@ -7400,9 +7398,8 @@ static void perf_output_read_group(struct perf_output_handle *handle,
 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 		values[n++] = running;
 
-	if ((leader != event) &&
-	    (leader->state == PERF_EVENT_STATE_ACTIVE))
-		leader->pmu->read(leader);
+	if ((leader != event) && !handle->skip_read)
+		perf_pmu_read(leader);
 
 	values[n++] = perf_event_count(leader, self);
 	if (read_format & PERF_FORMAT_ID)
@@ -7415,9 +7412,8 @@ static void perf_output_read_group(struct perf_output_handle *handle,
 	for_each_sibling_event(sub, leader) {
 		n = 0;
 
-		if ((sub != event) &&
-		    (sub->state == PERF_EVENT_STATE_ACTIVE))
-			sub->pmu->read(sub);
+		if ((sub != event) && !handle->skip_read)
+			perf_pmu_read(sub);
 
 		values[n++] = perf_event_count(sub, self);
 		if (read_format & PERF_FORMAT_ID)
@@ -7476,6 +7472,9 @@ void perf_output_sample(struct perf_output_handle *handle,
 {
 	u64 sample_type = data->type;
 
+	if (data->sample_flags & PERF_SAMPLE_READ)
+		handle->skip_read = 1;
+
 	perf_output_put(handle, *header);
 
 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 4f46f688d0d4..9b49ecca693e 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -185,6 +185,7 @@ __perf_output_begin(struct perf_output_handle *handle,
 
 	handle->rb    = rb;
 	handle->event = event;
+	handle->flags = 0;
 
 	have_lost = local_read(&rb->lost);
 	if (unlikely(have_lost)) {
-- 
2.38.1

[PATCH V10 4/4] perf/x86/intel: Support PEBS counters snapshotting

[kan.liang]

From: Kan Liang <kan.liang@linux.intel.com>

The counters snapshotting is a new adaptive PEBS extension, which can
capture programmable counters, fixed-function counters, and performance
metrics in a PEBS record. The feature is available in the PEBS format
V6.

The target counters can be configured in the new fields of MSR_PEBS_CFG.
Then the PEBS HW will generate the bit mask of counters (Counters Group
Header) followed by the content of all the requested counters into a
PEBS record.

The current Linux perf sample read feature can read all events in the
group when any event in the group is overflowed. But the rdpmc in the
NMI/overflow handler has a small gap from overflow. Also, there is some
overhead for each rdpmc read. The counters snapshotting feature can be
used as an accurate and low-overhead replacement.

Extend intel_update_topdown_event() to accept the value from PEBS
records.

Add a new PEBS_CNTR flag to indicate a sample read group that utilizes
the counters snapshotting feature. When the group is scheduled, the
PEBS configure can be updated accordingly.

To prevent the case that a PEBS record value might be in the past
relative to what is already in the event, perf always stops the PMU and
drains the PEBS buffer before updating the corresponding event->count.

Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
---

Changes since V9:
- Remove the stale Reviewed-by tag
- Add late_setup static call in struct x86_pmu
- Rename and move the late_setup to the begin of x86_pmu_enable()
- Update the comment in intel_perf_event_update_pmc() to explain
  the case of HES_STOPPED counter.
- Move INTEL_PMC_IDX_FIXED out of PEBS_DATACFG_FIX_BIT
- Rename intel_perf_event_pmc_to_count() intel_perf_event_update_pmc()
- Add a helper function __setup_pebs_counter_group() to avoid goto

 arch/x86/events/core.c             |  13 ++
 arch/x86/events/intel/core.c       |  75 ++++++++---
 arch/x86/events/intel/ds.c         | 191 +++++++++++++++++++++++++++--
 arch/x86/events/perf_event.h       |  13 ++
 arch/x86/events/perf_event_flags.h |   2 +-
 arch/x86/include/asm/perf_event.h  |  15 +++
 6 files changed, 284 insertions(+), 25 deletions(-)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 8f218ac0d445..7b6430e5a77b 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -94,6 +94,8 @@ DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
 
 DEFINE_STATIC_CALL_NULL(x86_pmu_filter, *x86_pmu.filter);
 
+DEFINE_STATIC_CALL_NULL(x86_pmu_late_setup, *x86_pmu.late_setup);
+
 /*
  * This one is magic, it will get called even when PMU init fails (because
  * there is no PMU), in which case it should simply return NULL.
@@ -1298,6 +1300,15 @@ static void x86_pmu_enable(struct pmu *pmu)
 
 	if (cpuc->n_added) {
 		int n_running = cpuc->n_events - cpuc->n_added;
+
+		/*
+		 * The late setup (after counters are scheduled)
+		 * is required for some cases, e.g., PEBS counters
+		 * snapshotting. Because an accurate counter index
+		 * is needed.
+		 */
+		static_call_cond(x86_pmu_late_setup)();
+
 		/*
 		 * apply assignment obtained either from
 		 * hw_perf_group_sched_in() or x86_pmu_enable()
@@ -2035,6 +2046,8 @@ static void x86_pmu_static_call_update(void)
 
 	static_call_update(x86_pmu_guest_get_msrs, x86_pmu.guest_get_msrs);
 	static_call_update(x86_pmu_filter, x86_pmu.filter);
+
+	static_call_update(x86_pmu_late_setup, x86_pmu.late_setup);
 }
 
 static void _x86_pmu_read(struct perf_event *event)
diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
index 762b140c4953..7f1b6b90a5fb 100644
--- a/arch/x86/events/intel/core.c
+++ b/arch/x86/events/intel/core.c
@@ -2710,7 +2710,7 @@ static void update_saved_topdown_regs(struct perf_event *event, u64 slots,
  * modify by a NMI. PMU has to be disabled before calling this function.
  */
 
-static u64 intel_update_topdown_event(struct perf_event *event, int metric_end)
+static u64 intel_update_topdown_event(struct perf_event *event, int metric_end, u64 *val)
 {
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 	struct perf_event *other;
@@ -2718,13 +2718,24 @@ static u64 intel_update_topdown_event(struct perf_event *event, int metric_end)
 	bool reset = true;
 	int idx;
 
-	/* read Fixed counter 3 */
-	rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots);
-	if (!slots)
-		return 0;
+	if (!val) {
+		/* read Fixed counter 3 */
+		rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots);
+		if (!slots)
+			return 0;
 
-	/* read PERF_METRICS */
-	rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics);
+		/* read PERF_METRICS */
+		rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics);
+	} else {
+		slots = val[0];
+		metrics = val[1];
+		/*
+		 * Don't reset the PERF_METRICS and Fixed counter 3
+		 * for each PEBS record read. Utilize the RDPMC metrics
+		 * clear mode.
+		 */
+		reset = false;
+	}
 
 	for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) {
 		if (!is_topdown_idx(idx))
@@ -2767,17 +2778,19 @@ static u64 intel_update_topdown_event(struct perf_event *event, int metric_end)
 	return slots;
 }
 
-static u64 icl_update_topdown_event(struct perf_event *event)
+static u64 icl_update_topdown_event(struct perf_event *event, u64 *val)
 {
 	return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE +
-						 x86_pmu.num_topdown_events - 1);
+						 x86_pmu.num_topdown_events - 1,
+					  val);
 }
 
-DEFINE_STATIC_CALL(intel_pmu_update_topdown_event, x86_perf_event_update);
+DEFINE_STATIC_CALL(intel_pmu_update_topdown_event, intel_pmu_topdown_event_update);
 
 static void intel_pmu_read_event(struct perf_event *event)
 {
-	if (event->hw.flags & (PERF_X86_EVENT_AUTO_RELOAD | PERF_X86_EVENT_TOPDOWN)) {
+	if (event->hw.flags & (PERF_X86_EVENT_AUTO_RELOAD | PERF_X86_EVENT_TOPDOWN) ||
+	    is_pebs_counter_event_group(event)) {
 		struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
 		bool pmu_enabled = cpuc->enabled;
 
@@ -2789,8 +2802,12 @@ static void intel_pmu_read_event(struct perf_event *event)
 		if (pmu_enabled)
 			intel_pmu_disable_all();
 
-		if (is_topdown_event(event))
-			static_call(intel_pmu_update_topdown_event)(event);
+		/*
+		 * If the PEBS counters snapshotting is enabled,
+		 * the topdown event is available in PEBS records.
+		 */
+		if (is_topdown_event(event) && !is_pebs_counter_event_group(event))
+			static_call(intel_pmu_update_topdown_event)(event, NULL);
 		else
 			intel_pmu_drain_pebs_buffer();
 
@@ -2933,7 +2950,7 @@ static int intel_pmu_set_period(struct perf_event *event)
 static u64 intel_pmu_update(struct perf_event *event)
 {
 	if (unlikely(is_topdown_count(event)))
-		return static_call(intel_pmu_update_topdown_event)(event);
+		return static_call(intel_pmu_update_topdown_event)(event, NULL);
 
 	return x86_perf_event_update(event);
 }
@@ -3099,7 +3116,7 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
 	 */
 	if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) {
 		handled++;
-		static_call(intel_pmu_update_topdown_event)(NULL);
+		static_call(intel_pmu_update_topdown_event)(NULL, NULL);
 	}
 
 	/*
@@ -3117,6 +3134,27 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
 		if (!test_bit(bit, cpuc->active_mask))
 			continue;
 
+		/*
+		 * There may be unprocessed PEBS records in the PEBS buffer,
+		 * which still stores the previous values.
+		 * Process those records first before handling the latest value.
+		 * For example,
+		 * A is a regular counter
+		 * B is a PEBS event which reads A
+		 * C is a PEBS event
+		 *
+		 * The following can happen:
+		 * B-assist			A=1
+		 * C				A=2
+		 * B-assist			A=3
+		 * A-overflow-PMI		A=4
+		 * C-assist-PMI (PEBS buffer)	A=5
+		 *
+		 * The PEBS buffer has to be drained before handling the A-PMI
+		 */
+		if (is_pebs_counter_event_group(event))
+			x86_pmu.drain_pebs(regs, &data);
+
 		if (!intel_pmu_save_and_restart(event))
 			continue;
 
@@ -4064,6 +4102,13 @@ static int intel_pmu_hw_config(struct perf_event *event)
 		event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT;
 	}
 
+	if ((event->attr.sample_type & PERF_SAMPLE_READ) &&
+	    (x86_pmu.intel_cap.pebs_format >= 6) &&
+	    x86_pmu.intel_cap.pebs_baseline &&
+	    is_sampling_event(event) &&
+	    event->attr.precise_ip)
+		event->group_leader->hw.flags |= PERF_X86_EVENT_PEBS_CNTR;
+
 	if ((event->attr.type == PERF_TYPE_HARDWARE) ||
 	    (event->attr.type == PERF_TYPE_HW_CACHE))
 		return 0;
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index eb14b46423e5..fad4352e6931 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1294,6 +1294,19 @@ static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
 	ds->pebs_interrupt_threshold = threshold;
 }
 
+#define PEBS_DATACFG_CNTRS(x)						\
+	((x >> PEBS_DATACFG_CNTR_SHIFT) & PEBS_DATACFG_CNTR_MASK)
+
+#define PEBS_DATACFG_CNTR_BIT(x)					\
+	(((1ULL << x) & PEBS_DATACFG_CNTR_MASK) << PEBS_DATACFG_CNTR_SHIFT)
+
+#define PEBS_DATACFG_FIX(x)						\
+	((x >> PEBS_DATACFG_FIX_SHIFT) & PEBS_DATACFG_FIX_MASK)
+
+#define PEBS_DATACFG_FIX_BIT(x)						\
+	(((1ULL << (x)) & PEBS_DATACFG_FIX_MASK)			\
+	 << PEBS_DATACFG_FIX_SHIFT)
+
 static void adaptive_pebs_record_size_update(void)
 {
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
@@ -1308,10 +1321,58 @@ static void adaptive_pebs_record_size_update(void)
 		sz += sizeof(struct pebs_xmm);
 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
 		sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
+	if (pebs_data_cfg & (PEBS_DATACFG_METRICS | PEBS_DATACFG_CNTR)) {
+		sz += sizeof(struct pebs_cntr_header);
+
+		/* Metrics base and Metrics Data */
+		if (pebs_data_cfg & PEBS_DATACFG_METRICS)
+			sz += 2 * sizeof(u64);
+
+		if (pebs_data_cfg & PEBS_DATACFG_CNTR) {
+			sz += (hweight64(PEBS_DATACFG_CNTRS(pebs_data_cfg)) +
+			       hweight64(PEBS_DATACFG_FIX(pebs_data_cfg))) *
+			      sizeof(u64);
+		}
+	}
 
 	cpuc->pebs_record_size = sz;
 }
 
+static void __intel_pmu_pebs_update_cfg(struct perf_event *event,
+					int idx, u64 *pebs_data_cfg)
+{
+	if (is_metric_event(event)) {
+		*pebs_data_cfg |= PEBS_DATACFG_METRICS;
+		return;
+	}
+
+	*pebs_data_cfg |= PEBS_DATACFG_CNTR;
+
+	if (idx >= INTEL_PMC_IDX_FIXED)
+		*pebs_data_cfg |= PEBS_DATACFG_FIX_BIT(idx - INTEL_PMC_IDX_FIXED);
+	else
+		*pebs_data_cfg |= PEBS_DATACFG_CNTR_BIT(idx);
+}
+
+
+static void intel_pmu_late_setup(void)
+{
+	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	struct perf_event *event;
+	u64 pebs_data_cfg = 0;
+	int i;
+
+	for (i = 0; i < cpuc->n_events; i++) {
+		event = cpuc->event_list[i];
+		if (!is_pebs_counter_event_group(event))
+			continue;
+		__intel_pmu_pebs_update_cfg(event, cpuc->assign[i], &pebs_data_cfg);
+	}
+
+	if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
+		cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
+}
+
 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
 				PERF_SAMPLE_PHYS_ADDR |			     \
 				PERF_SAMPLE_WEIGHT_TYPE |		     \
@@ -1914,12 +1975,89 @@ static void adaptive_pebs_save_regs(struct pt_regs *regs,
 #endif
 }
 
+static void intel_perf_event_update_pmc(struct perf_event *event, u64 pmc)
+{
+	int shift = 64 - x86_pmu.cntval_bits;
+	struct hw_perf_event *hwc;
+	u64 delta, prev_pmc;
+
+	/*
+	 * A recorded counter may not have an assigned event in the
+	 * following cases. The value should be dropped.
+	 * - An event is deleted. There is still an active PEBS event.
+	 *   The PEBS record doesn't shrink on pmu::del().
+	 *   If the counter of the deleted event once occurred in a PEBS
+	 *   record, PEBS still records the counter until the counter is
+	 *   reassigned.
+	 * - An event is stopped for some reason, e.g., throttled.
+	 *   During this period, another event is added and takes the
+	 *   counter of the stopped event. The stopped event is assigned
+	 *   to another new and uninitialized counter, since the
+	 *   x86_pmu_start(RELOAD) is not invoked for a stopped event.
+	 *   The PEBS__DATA_CFG is updated regardless of the event state.
+	 *   The uninitialized counter can be recorded in a PEBS record.
+	 *   But the cpuc->events[uninitialized_counter] is always NULL,
+	 *   because the event is stopped. The uninitialized value is
+	 *   safely dropped.
+	 */
+	if (!event)
+		return;
+
+	hwc = &event->hw;
+	prev_pmc = local64_read(&hwc->prev_count);
+
+	/* Only update the count when the PMU is disabled */
+	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
+	local64_set(&hwc->prev_count, pmc);
+
+	delta = (pmc << shift) - (prev_pmc << shift);
+	delta >>= shift;
+
+	local64_add(delta, &event->count);
+	local64_sub(delta, &hwc->period_left);
+}
+
+static inline void __setup_pebs_counter_group(struct cpu_hw_events *cpuc,
+					      struct perf_event *event,
+					      struct pebs_cntr_header *cntr,
+					      void *next_record)
+{
+	int bit;
+
+	for_each_set_bit(bit, (unsigned long *)&cntr->cntr, INTEL_PMC_MAX_GENERIC) {
+		intel_perf_event_update_pmc(cpuc->events[bit], *(u64 *)next_record);
+		next_record += sizeof(u64);
+	}
+
+	for_each_set_bit(bit, (unsigned long *)&cntr->fixed, INTEL_PMC_MAX_FIXED) {
+		/* The slots event will be handled with perf_metric later */
+		if ((cntr->metrics == INTEL_CNTR_METRICS) &&
+		    (bit + INTEL_PMC_IDX_FIXED == INTEL_PMC_IDX_FIXED_SLOTS)) {
+			next_record += sizeof(u64);
+			continue;
+		}
+		intel_perf_event_update_pmc(cpuc->events[bit + INTEL_PMC_IDX_FIXED],
+					    *(u64 *)next_record);
+		next_record += sizeof(u64);
+	}
+
+	/* HW will reload the value right after the overflow. */
+	if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
+		local64_set(&event->hw.prev_count, (u64)-event->hw.sample_period);
+
+	if (cntr->metrics == INTEL_CNTR_METRICS) {
+		static_call(intel_pmu_update_topdown_event)
+			   (cpuc->events[INTEL_PMC_IDX_FIXED_SLOTS],
+			    (u64 *)next_record);
+		next_record += 2 * sizeof(u64);
+	}
+}
+
 #define PEBS_LATENCY_MASK			0xffff
 
 /*
  * With adaptive PEBS the layout depends on what fields are configured.
  */
-
 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
 					    struct pt_regs *iregs, void *__pebs,
 					    struct perf_sample_data *data,
@@ -2049,6 +2187,28 @@ static void setup_pebs_adaptive_sample_data(struct perf_event *event,
 		}
 	}
 
+	if (format_group & (PEBS_DATACFG_CNTR | PEBS_DATACFG_METRICS)) {
+		struct pebs_cntr_header *cntr = next_record;
+		unsigned int nr;
+
+		next_record += sizeof(struct pebs_cntr_header);
+		/*
+		 * The PEBS_DATA_CFG is a global register, which is the
+		 * superset configuration for all PEBS events.
+		 * For the PEBS record of non-sample-read group, ignore
+		 * the counter snapshot fields.
+		 */
+		if (is_pebs_counter_event_group(event)) {
+			__setup_pebs_counter_group(cpuc, event, cntr, next_record);
+			data->sample_flags |= PERF_SAMPLE_READ;
+		}
+
+		nr = hweight32(cntr->cntr) + hweight32(cntr->fixed);
+		if (cntr->metrics == INTEL_CNTR_METRICS)
+			nr += 2;
+		next_record += nr * sizeof(u64);
+	}
+
 	WARN_ONCE(next_record != __pebs + basic->format_size,
 			"PEBS record size %u, expected %llu, config %llx\n",
 			basic->format_size,
@@ -2202,13 +2362,21 @@ __intel_pmu_pebs_last_event(struct perf_event *event,
 	}
 
 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
-		/*
-		 * Now, auto-reload is only enabled in fixed period mode.
-		 * The reload value is always hwc->sample_period.
-		 * May need to change it, if auto-reload is enabled in
-		 * freq mode later.
-		 */
-		intel_pmu_save_and_restart_reload(event, count);
+		if ((is_pebs_counter_event_group(event))) {
+			/*
+			 * The value of each sample has been updated when setup
+			 * the corresponding sample data.
+			 */
+			perf_event_update_userpage(event);
+		} else {
+			/*
+			 * Now, auto-reload is only enabled in fixed period mode.
+			 * The reload value is always hwc->sample_period.
+			 * May need to change it, if auto-reload is enabled in
+			 * freq mode later.
+			 */
+			intel_pmu_save_and_restart_reload(event, count);
+		}
 	} else
 		intel_pmu_save_and_restart(event);
 }
@@ -2543,6 +2711,11 @@ void __init intel_ds_init(void)
 			break;
 
 		case 6:
+			if (x86_pmu.intel_cap.pebs_baseline) {
+				x86_pmu.large_pebs_flags |= PERF_SAMPLE_READ;
+				x86_pmu.late_setup = intel_pmu_late_setup;
+			}
+			fallthrough;
 		case 5:
 			x86_pmu.pebs_ept = 1;
 			fallthrough;
@@ -2567,7 +2740,7 @@ void __init intel_ds_init(void)
 					  PERF_SAMPLE_REGS_USER |
 					  PERF_SAMPLE_REGS_INTR);
 			}
-			pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
+			pr_cont("PEBS fmt%d%c%s, ", format, pebs_type, pebs_qual);
 
 			if (!is_hybrid() && x86_pmu.intel_cap.pebs_output_pt_available) {
 				pr_cont("PEBS-via-PT, ");
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index 536a112f6353..a698e6484b3b 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -115,6 +115,11 @@ static inline bool is_branch_counters_group(struct perf_event *event)
 	return event->group_leader->hw.flags & PERF_X86_EVENT_BRANCH_COUNTERS;
 }
 
+static inline bool is_pebs_counter_event_group(struct perf_event *event)
+{
+	return event->group_leader->hw.flags & PERF_X86_EVENT_PEBS_CNTR;
+}
+
 struct amd_nb {
 	int nb_id;  /* NorthBridge id */
 	int refcnt; /* reference count */
@@ -800,6 +805,7 @@ struct x86_pmu {
 	u64		(*update)(struct perf_event *event);
 	int		(*hw_config)(struct perf_event *event);
 	int		(*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
+	void		(*late_setup)(void);
 	unsigned	eventsel;
 	unsigned	perfctr;
 	unsigned	fixedctr;
@@ -1108,6 +1114,7 @@ extern struct x86_pmu x86_pmu __read_mostly;
 DECLARE_STATIC_CALL(x86_pmu_set_period, *x86_pmu.set_period);
 DECLARE_STATIC_CALL(x86_pmu_update,     *x86_pmu.update);
 DECLARE_STATIC_CALL(x86_pmu_drain_pebs,	*x86_pmu.drain_pebs);
+DECLARE_STATIC_CALL(x86_pmu_late_setup,	*x86_pmu.late_setup);
 
 static __always_inline struct x86_perf_task_context_opt *task_context_opt(void *ctx)
 {
@@ -1149,6 +1156,12 @@ extern u64 __read_mostly hw_cache_extra_regs
 
 u64 x86_perf_event_update(struct perf_event *event);
 
+static inline u64 intel_pmu_topdown_event_update(struct perf_event *event, u64 *val)
+{
+	return x86_perf_event_update(event);
+}
+DECLARE_STATIC_CALL(intel_pmu_update_topdown_event, intel_pmu_topdown_event_update);
+
 static inline unsigned int x86_pmu_config_addr(int index)
 {
 	return x86_pmu.eventsel + (x86_pmu.addr_offset ?
diff --git a/arch/x86/events/perf_event_flags.h b/arch/x86/events/perf_event_flags.h
index 6c977c19f2cd..1d9e385649b5 100644
--- a/arch/x86/events/perf_event_flags.h
+++ b/arch/x86/events/perf_event_flags.h
@@ -9,7 +9,7 @@ PERF_ARCH(PEBS_LD_HSW,		0x00008) /* haswell style datala, load */
 PERF_ARCH(PEBS_NA_HSW,		0x00010) /* haswell style datala, unknown */
 PERF_ARCH(EXCL,			0x00020) /* HT exclusivity on counter */
 PERF_ARCH(DYNAMIC,		0x00040) /* dynamic alloc'd constraint */
-			/*	0x00080	*/
+PERF_ARCH(PEBS_CNTR,		0x00080) /* PEBS counters snapshot */
 PERF_ARCH(EXCL_ACCT,		0x00100) /* accounted EXCL event */
 PERF_ARCH(AUTO_RELOAD,		0x00200) /* use PEBS auto-reload */
 PERF_ARCH(LARGE_PEBS,		0x00400) /* use large PEBS */
diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
index 1ac79f361645..adaeb8ca3a8a 100644
--- a/arch/x86/include/asm/perf_event.h
+++ b/arch/x86/include/asm/perf_event.h
@@ -141,6 +141,12 @@
 #define PEBS_DATACFG_XMMS	BIT_ULL(2)
 #define PEBS_DATACFG_LBRS	BIT_ULL(3)
 #define PEBS_DATACFG_LBR_SHIFT	24
+#define PEBS_DATACFG_CNTR	BIT_ULL(4)
+#define PEBS_DATACFG_CNTR_SHIFT	32
+#define PEBS_DATACFG_CNTR_MASK	GENMASK_ULL(15, 0)
+#define PEBS_DATACFG_FIX_SHIFT	48
+#define PEBS_DATACFG_FIX_MASK	GENMASK_ULL(7, 0)
+#define PEBS_DATACFG_METRICS	BIT_ULL(5)
 
 /* Steal the highest bit of pebs_data_cfg for SW usage */
 #define PEBS_UPDATE_DS_SW	BIT_ULL(63)
@@ -471,6 +477,15 @@ struct pebs_xmm {
 
 #define IBS_CPUID_FEATURES		0x8000001b
 
+struct pebs_cntr_header {
+	u32 cntr;
+	u32 fixed;
+	u32 metrics;
+	u32 reserved;
+};
+
+#define INTEL_CNTR_METRICS		0x3
+
 /*
  * Same bit mask as for IBS cpuid feature flags (Fn8000_001B_EAX), but
  * bit 0 is used to indicate the existence of IBS.
-- 
2.38.1

Re: [PATCH V10 4/4] perf/x86/intel: Support PEBS counters snapshotting

[Peter Zijlstra]

On Tue, Jan 21, 2025 at 07:23:03AM -0800, kan.liang@linux.intel.com wrote:

> @@ -471,6 +477,15 @@ struct pebs_xmm {
>  
>  #define IBS_CPUID_FEATURES		0x8000001b
>  
> +struct pebs_cntr_header {
> +	u32 cntr;
> +	u32 fixed;
> +	u32 metrics;
> +	u32 reserved;
> +};
> +
> +#define INTEL_CNTR_METRICS		0x3

I've moved this up a little, so that all the adaptive pebs crud is
together.

Other than that, I've picked up these patches and will merge the lot
past -rc1.

Re: [PATCH V10 4/4] perf/x86/intel: Support PEBS counters snapshotting

[Liang, Kan]

On 2025-01-27 10:51 a.m., Peter Zijlstra wrote:
> On Tue, Jan 21, 2025 at 07:23:03AM -0800, kan.liang@linux.intel.com wrote:
> 
>> @@ -471,6 +477,15 @@ struct pebs_xmm {
>>  
>>  #define IBS_CPUID_FEATURES		0x8000001b
>>  
>> +struct pebs_cntr_header {
>> +	u32 cntr;
>> +	u32 fixed;
>> +	u32 metrics;
>> +	u32 reserved;
>> +};
>> +
>> +#define INTEL_CNTR_METRICS		0x3
> 
> I've moved this up a little, so that all the adaptive pebs crud is
> together.
> 

Sure. Thanks!

> Other than that, I've picked up these patches and will merge the lot
> past -rc1.
> 

Thanks!

Kan

Re: [PATCH V10 4/4] perf/x86/intel: Support PEBS counters snapshotting

[Liang, Kan]

Hi Peter,

On 2025-01-27 10:51 a.m., Peter Zijlstra wrote:
> On Tue, Jan 21, 2025 at 07:23:03AM -0800, kan.liang@linux.intel.com wrote:
> 
>> @@ -471,6 +477,15 @@ struct pebs_xmm {
>>  
>>  #define IBS_CPUID_FEATURES		0x8000001b
>>  
>> +struct pebs_cntr_header {
>> +	u32 cntr;
>> +	u32 fixed;
>> +	u32 metrics;
>> +	u32 reserved;
>> +};
>> +
>> +#define INTEL_CNTR_METRICS		0x3
> 
> I've moved this up a little, so that all the adaptive pebs crud is
> together.
> 
> Other than that, I've picked up these patches and will merge the lot
> past -rc1.
> 

We found a bug when testing with some non-precise events in freq mode.
I've submitted a patch to fix it.
https://lore.kernel.org/lkml/20250204210514.4089680-1-kan.liang@linux.intel.com/

Could you please fold it into the patch ("perf/x86/intel: Support PEBS
counters snapshotting")?

Sorry for any inconvenience.

Thanks,
Kan