Skip to contentSkip to conversation

perf/x86/intel/uncore: customized event_read for client IMC uncore

PATCH Superseded [email protected] 2 participants 20 messages Nov 2, 2017 – Jan 15, 2018

Subsystems (2)

v4 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/8 perf/x86/intel/uncore: customized event_read for client IMC uncore

From: Kan Liang <[email protected]>

There are two free running counters for client IMC uncore. The custom event_init() function hardcode their index to 'UNCORE_PMC_IDX_FIXED' and 'UNCORE_PMC_IDX_FIXED + 1'. To support the 'UNCORE_PMC_IDX_FIXED + 1' case, the generic uncore_perf_event_update is obscurely hacked. The code quality issue will bring problem when new counter index is introduced into generic code. For example, free running counter index.

Introduce customized event_read function for client IMC uncore. The customized function is exactly copied from previous generic uncore_pmu_event_read.
The 'UNCORE_PMC_IDX_FIXED + 1' case will be isolated for client IMC uncore only.

---

Change since V3:
- Use the customized read function to replace uncore_perf_event_update.
- Move generic code change to patch 3/8.

 arch/x86/events/intel/uncore_snb.c | 33 +++++++++++++++++++++++++++++++--
 1 file changed, 31 insertions(+), 2 deletions(-)
arch/x86/events/intel/uncore_snb.c +31 -2
index db1127c..b6d0d72 100644
@@ -449,6 +449,35 @@ static void snb_uncore_imc_event_start(struct perf_event *event, int flags)
 		uncore_pmu_start_hrtimer(box);
 }
 
+static void snb_uncore_imc_event_read(struct perf_event *event)
+{
+	struct intel_uncore_box *box = uncore_event_to_box(event);
+	u64 prev_count, new_count, delta;
+	int shift;
+
+	/*
+	 * There are two free running counters in IMC.
+	 * The index for the second one is hardcoded to
+	 * UNCORE_PMC_IDX_FIXED + 1.
+	 */
+	if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
+		shift = 64 - uncore_fixed_ctr_bits(box);
+	else
+		shift = 64 - uncore_perf_ctr_bits(box);
+
+	/* the hrtimer might modify the previous event value */
+again:
+	prev_count = local64_read(&event->hw.prev_count);
+	new_count = uncore_read_counter(box, event);
+	if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
+		goto again;
+
+	delta = (new_count << shift) - (prev_count << shift);
+	delta >>= shift;
+
+	local64_add(delta, &event->count);
+}
+
 static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)
 {
 	struct intel_uncore_box *box = uncore_event_to_box(event);
@@ -471,7 +500,7 @@ static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)
 		 * Drain the remaining delta count out of a event
 		 * that we are disabling:
 		 */
-		uncore_perf_event_update(box, event);
+		snb_uncore_imc_event_read(event);
 		hwc->state |= PERF_HES_UPTODATE;
 	}
 }
@@ -533,7 +562,7 @@ static struct pmu snb_uncore_imc_pmu = {
 	.del		= snb_uncore_imc_event_del,
 	.start		= snb_uncore_imc_event_start,
 	.stop		= snb_uncore_imc_event_stop,
-	.read		= uncore_pmu_event_read,
+	.read		= snb_uncore_imc_event_read,
 };
 
 static struct intel_uncore_ops snb_uncore_imc_ops = {
Signature · 1 line
2.7.4
#1 lore raw

PATCH v4 2/8 perf/x86/intel/uncore: correct fixed counter index check for NHM

From: Kan Liang <[email protected]>

For Nehalem and Westmere, there is only one fixed counter for W-Box. There is no index which is bigger than UNCORE_PMC_IDX_FIXED. It is not correct to use >= to check fixed counter.
The code quality issue will bring problem when new counter index is introduced.

---

Changes since V3:
- New patch for similar '>=' issue on NHM.

 arch/x86/events/intel/uncore_nhmex.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/events/intel/uncore_nhmex.c +1 -1
index 6a5cbe9..6f0d0da 100644
@@ -245,7 +245,7 @@ static void nhmex_uncore_msr_enable_event(struct intel_uncore_box *box, struct p
 {
 	struct hw_perf_event *hwc = &event->hw;
 
-	if (hwc->idx >= UNCORE_PMC_IDX_FIXED)
+	if (hwc->idx == UNCORE_PMC_IDX_FIXED)
 		wrmsrl(hwc->config_base, NHMEX_PMON_CTL_EN_BIT0);
 	else if (box->pmu->type->event_mask & NHMEX_PMON_CTL_EN_BIT0)
 		wrmsrl(hwc->config_base, hwc->config | NHMEX_PMON_CTL_EN_BIT22);
Signature · 1 line
2.7.4
Reviewed-by on v4 2/8
#2 lore raw

Re: [PATCH V4 2/8] perf/x86/intel/uncore: correct fixed counter index check for NHM

Quoted 9 lines · On Thu, 2 Nov 2017, [email protected] wrote:

From: Kan Liang <[email protected]>

For Nehalem and Westmere, there is only one fixed counter for W-Box. There is no index which is bigger than UNCORE_PMC_IDX_FIXED. It is not correct to use >= to check fixed counter.
The code quality issue will bring problem when new counter index is introduced.

#3 lore raw

PATCH v4 3/8 perf/x86/intel/uncore: correct fixed counter index check in generic code

In reply to [email protected]

From: Kan Liang <[email protected]>

There is no index which is bigger than UNCORE_PMC_IDX_FIXED. The only exception is client IMC uncore. It has customized function to deal with the 'UNCORE_PMC_IDX_FIXED + 1' case. It does not touch the generic code. For generic code, it is not correct to use >= to check fixed counter. The code quality issue will bring problem when new counter index is introduced.

---

Changes since V3:
- Split the patch for generic change

 arch/x86/events/intel/uncore.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/events/intel/uncore.c +1 -1
index 1c5390f..3b8cd88 100644
@@ -218,7 +218,7 @@ void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *e
 	u64 prev_count, new_count, delta;
 	int shift;
 
-	if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
+	if (event->hw.idx == UNCORE_PMC_IDX_FIXED)
 		shift = 64 - uncore_fixed_ctr_bits(box);
 	else
 		shift = 64 - uncore_perf_ctr_bits(box);
Signature · 1 line
2.7.4
Reviewed-by on v4 3/8
#4 lore raw

Re: [PATCH V4 3/8] perf/x86/intel/uncore: correct fixed counter index check in generic code

Quoted 10 lines · On Thu, 2 Nov 2017, [email protected] wrote:

From: Kan Liang <[email protected]>

There is no index which is bigger than UNCORE_PMC_IDX_FIXED. The only exception is client IMC uncore. It has customized function to deal with the 'UNCORE_PMC_IDX_FIXED + 1' case. It does not touch the generic code. For generic code, it is not correct to use >= to check fixed counter. The code quality issue will bring problem when new counter index is introduced.

#5 lore raw

PATCH v4 4/8 perf/x86/intel/uncore: add new data structures for free running counters

In reply to [email protected]

From: Kan Liang <[email protected]>

There are a number of free running counters introduced for uncore, which provide highly valuable information to a wide array of customers. For example, Skylake Server has IIO free running counters to collect Input/Output x BW/Utilization.
The precious generic counters could be saved to collect other customer interested data.

The free running counter is read-only and always active. Current generic uncore code does not support this kind of counters.

Introduce a new index to indicate the free running counters. Only one index is enough for all free running counters. Because the free running countes are always active, and the event and free running counter are always 1:1 mapped. It does not need extra index to indicate the assigned counter.

Introduce some rules to encode the event for free running counters.
- The event for free running counter has the same event code 0xff as the event for fixed counter.
- The umask of the event starts from 0x10. The umask which is less than 0x10 is reserved for the event of fixed counter.
- The free running counters can be divided into different types according to the MSR location, bit width or definition. The start point of the umask for different type has 0x10 offset.
For example, there are three types of IIO free running counters on Skylake server, IO CLOCKS counters, BANDWIDTH counters and UTILIZATION counters.
The event code for all free running counters is 0xff. 'ioclk' is the first counter of IO CLOCKS. IO CLOCKS is the first type of free running counters, which umask starts from 0x10.
So 'ioclk' is encoded as event=0xff,umask=0x10
'bw_in_port2' is the third counter of BANDWIDTH counters. BANDWIDTH is the second type which umask starts from 0x20.
So 'bw_in_port2' is encoded as event=0xff,umask=0x22.

Introduce a new data structure to store free running counters related information for each type. It includes the number of counters, bit width, base address, offset between counters and offset between boxes.

Introduce several inline helpers to check index for fixed counter and free running counter, validate free running counter event, and retrieve the free running counter information according to box and event.

---

Changes since V3:
- Split the patch to add the new data structures and struct members along with the inline helpers.

 arch/x86/events/intel/uncore.h | 118 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 117 insertions(+), 1 deletion(-)
arch/x86/events/intel/uncore.h +117 -1
index df5989f..a508bc5 100644
@@ -11,8 +11,13 @@
 
 #define UNCORE_FIXED_EVENT		0xff
 #define UNCORE_PMC_IDX_MAX_GENERIC	8
+#define UNCORE_PMC_IDX_MAX_FIXED	1
+#define UNCORE_PMC_IDX_MAX_FREERUNNING	1
 #define UNCORE_PMC_IDX_FIXED		UNCORE_PMC_IDX_MAX_GENERIC
-#define UNCORE_PMC_IDX_MAX		(UNCORE_PMC_IDX_FIXED + 1)
+#define UNCORE_PMC_IDX_FREERUNNING	(UNCORE_PMC_IDX_FIXED + \
+					UNCORE_PMC_IDX_MAX_FIXED)
+#define UNCORE_PMC_IDX_MAX		(UNCORE_PMC_IDX_FREERUNNING + \
+					UNCORE_PMC_IDX_MAX_FREERUNNING)
 
 #define UNCORE_PCI_DEV_FULL_DATA(dev, func, type, idx)	\
 		((dev << 24) | (func << 16) | (type << 8) | idx)
@@ -34,6 +39,7 @@ struct intel_uncore_ops;
 struct intel_uncore_pmu;
 struct intel_uncore_box;
 struct uncore_event_desc;
+struct freerunning_counters;
 
 struct intel_uncore_type {
 	const char *name;
@@ -41,6 +47,7 @@ struct intel_uncore_type {
 	int num_boxes;
 	int perf_ctr_bits;
 	int fixed_ctr_bits;
+	int num_freerunning_types;
 	unsigned perf_ctr;
 	unsigned event_ctl;
 	unsigned event_mask;
@@ -58,6 +65,7 @@ struct intel_uncore_type {
 	struct intel_uncore_pmu *pmus;
 	struct intel_uncore_ops *ops;
 	struct uncore_event_desc *event_descs;
+	struct freerunning_counters *freerunning;
 	const struct attribute_group *attr_groups[4];
 	struct pmu *pmu; /* for custom pmu ops */
 };
@@ -128,6 +136,14 @@ struct uncore_event_desc {
 	const char *config;
 };
 
+struct freerunning_counters {
+	unsigned int counter_base;
+	unsigned int counter_offset;
+	unsigned int box_offset;
+	unsigned int num_counters;
+	unsigned int bits;
+};
+
 struct pci2phy_map {
 	struct list_head list;
 	int segment;
@@ -156,6 +172,16 @@ static ssize_t __uncore_##_var##_show(struct kobject *kobj,		\
 static struct kobj_attribute format_attr_##_var =			\
 	__ATTR(_name, 0444, __uncore_##_var##_show, NULL)
 
+static inline bool uncore_pmc_fixed(int idx)
+{
+	return idx == UNCORE_PMC_IDX_FIXED;
+}
+
+static inline bool uncore_pmc_freerunning(int idx)
+{
+	return idx == UNCORE_PMC_IDX_FREERUNNING;
+}
+
 static inline unsigned uncore_pci_box_ctl(struct intel_uncore_box *box)
 {
 	return box->pmu->type->box_ctl;
@@ -213,6 +239,55 @@ static inline unsigned uncore_msr_fixed_ctr(struct intel_uncore_box *box)
 	return box->pmu->type->fixed_ctr + uncore_msr_box_offset(box);
 }
 
+
+/*
+ * Free running counter is similar as fixed counter, except it is read-only
+ * and always active when the uncore box is powered up.
+ *
+ * Here are the rules which are used to encode the event for free running
+ * counter.
+ * - The event for free running counter has the same event code 0xff as
+ *   the event for fixed counter.
+ * - The umask of the event starts from 0x10. The umask which is less
+ *   than 0x10 is reserved for the event of fixed counter.
+ * - The free running counters can be divided into different types according
+ *   to the MSR location, bit width or definition. The start point of the
+ *   umask for different type has 0x10 offset.
+ *
+ * For example, there are three types of IIO free running counters on Skylake
+ * server, IO CLOCKS counters, BANDWIDTH counters and UTILIZATION counters.
+ * The event code for all the free running counters is 0xff.
+ * 'ioclk' is the first counter of IO CLOCKS. IO CLOCKS is the first type,
+ * which umask starts from 0x10.
+ * So 'ioclk' is encoded as event=0xff,umask=0x10
+ * 'bw_in_port2' is the third counter of BANDWIDTH counters. BANDWIDTH is
+ * the second type, which umask starts from 0x20.
+ * So 'bw_in_port2' is encoded as event=0xff,umask=0x22
+ */
+static inline unsigned int uncore_freerunning_idx(u64 config)
+{
+	return ((config >> 8) & 0xf);
+}
+
+#define UNCORE_FREERUNNING_UMASK_START		0x10
+static inline unsigned int uncore_freerunning_type(u64 config)
+{
+	return ((((config >> 8) - UNCORE_FREERUNNING_UMASK_START) >> 4) & 0xf);
+}
+
+static inline
+unsigned int uncore_freerunning_counter(struct intel_uncore_box *box,
+					struct perf_event *event)
+{
+	unsigned int type = uncore_freerunning_type(event->attr.config);
+	unsigned int idx = uncore_freerunning_idx(event->attr.config);
+	struct intel_uncore_pmu *pmu = box->pmu;
+
+	return pmu->type->freerunning[type].counter_base +
+	       pmu->type->freerunning[type].counter_offset * idx +
+	       pmu->type->freerunning[type].box_offset * pmu->pmu_idx;
+}
+
 static inline
 unsigned uncore_msr_event_ctl(struct intel_uncore_box *box, int idx)
 {
@@ -275,11 +350,52 @@ static inline int uncore_fixed_ctr_bits(struct intel_uncore_box *box)
 	return box->pmu->type->fixed_ctr_bits;
 }
 
+static inline
+unsigned int uncore_freerunning_bits(struct intel_uncore_box *box,
+				     struct perf_event *event)
+{
+	unsigned int type = uncore_freerunning_type(event->attr.config);
+
+	return box->pmu->type->freerunning[type].bits;
+}
+
+static inline int uncore_num_freerunning(struct intel_uncore_box *box,
+					 struct perf_event *event)
+{
+	unsigned int type = uncore_freerunning_type(event->attr.config);
+
+	return box->pmu->type->freerunning[type].num_counters;
+}
+
+static inline int uncore_num_freerunning_types(struct intel_uncore_box *box,
+					       struct perf_event *event)
+{
+	return box->pmu->type->num_freerunning_types;
+}
+
+static inline bool check_valid_freerunning_event(struct intel_uncore_box *box,
+						 struct perf_event *event)
+{
+	unsigned int type = uncore_freerunning_type(event->attr.config);
+	unsigned int idx = uncore_freerunning_idx(event->attr.config);
+
+	return (type < uncore_num_freerunning_types(box, event)) &&
+	       (idx < uncore_num_freerunning(box, event));
+}
+
 static inline int uncore_num_counters(struct intel_uncore_box *box)
 {
 	return box->pmu->type->num_counters;
 }
 
+static inline bool is_freerunning_event(struct perf_event *event)
+{
+	u64 cfg = event->attr.config;
+
+	return ((cfg & UNCORE_FIXED_EVENT) == UNCORE_FIXED_EVENT) &&
+	       (((cfg >> 8) & 0xff) >= UNCORE_FREERUNNING_UMASK_START);
+}
+
 static inline void uncore_disable_box(struct intel_uncore_box *box)
 {
 	if (box->pmu->type->ops->disable_box)
Signature · 1 line
2.7.4
#6 lore raw

Re: [PATCH V4 4/8] perf/x86/intel/uncore: add new data structures for free running counters

Quoted 21 lines · On Thu, 2 Nov 2017, [email protected] wrote:
arch/x86/events/intel/uncore.h +21 -0
+/*
+ * Free running counter is similar as fixed counter, except it is read-only
+ * and always active when the uncore box is powered up.
+ *
+ * Here are the rules which are used to encode the event for free running
+ * counter.
+ * - The event for free running counter has the same event code 0xff as
+ *   the event for fixed counter.
+ * - The umask of the event starts from 0x10. The umask which is less
+ *   than 0x10 is reserved for the event of fixed counter.
+ * - The free running counters can be divided into different types according
+ *   to the MSR location, bit width or definition. The start point of the
+ *   umask for different type has 0x10 offset.
+ *
+ * For example, there are three types of IIO free running counters on Skylake
+ * server, IO CLOCKS counters, BANDWIDTH counters and UTILIZATION counters.
+ * The event code for all the free running counters is 0xff.
+ * 'ioclk' is the first counter of IO CLOCKS. IO CLOCKS is the first type,
+ * which umask starts from 0x10.
+ * So 'ioclk' is encoded as event=0xff,umask=0x10
+ * 'bw_in_port2' is the third counter of BANDWIDTH counters. BANDWIDTH is

arch/x86/events/intel/uncore.h static inline unsigned uncore_msr_fixed_ctr(struct intel_uncore_box *box) · line 273

Quoted 10 lines
arch/x86/events/intel/uncore.h +10 -0
+ * the second type, which umask starts from 0x20.
+ * So 'bw_in_port2' is encoded as event=0xff,umask=0x22
+ */
+static inline unsigned int uncore_freerunning_idx(u64 config)
+{
+	return ((config >> 8) & 0xf);
+}
+
+#define UNCORE_FREERUNNING_UMASK_START		0x10
+static inline unsigned int uncore_freerunning_type(u64 config)

Groan. I asked that before that you please stop glueing defines right in front of a function declaration w/o any visible space. Is it that hard to read and address _ALL_ review comments?

arch/x86/events/intel/uncore.h static inline unsigned uncore_msr_fixed_ctr(struct intel_uncore_box *box) · line 276

Quoted 3 lines
arch/x86/events/intel/uncore.h +3 -0
+{
+	return ((((config >> 8) - UNCORE_FREERUNNING_UMASK_START) >> 4) & 0xf);
+}

Other than that this looks reasonable.

Reviewed-by
#7 lore raw

Re: [PATCH V4 4/8] perf/x86/intel/uncore: add new data structures for free running counters

Quoted 37 lines · Thomas Gleixner
Quoted 21 lines · On Thu, 2 Nov 2017, [email protected] wrote:
arch/x86/events/intel/uncore.h +21 -0
+/*
+ * Free running counter is similar as fixed counter, except it is read-only
+ * and always active when the uncore box is powered up.
+ *
+ * Here are the rules which are used to encode the event for free running
+ * counter.
+ * - The event for free running counter has the same event code 0xff as
+ *   the event for fixed counter.
+ * - The umask of the event starts from 0x10. The umask which is less
+ *   than 0x10 is reserved for the event of fixed counter.
+ * - The free running counters can be divided into different types according
+ *   to the MSR location, bit width or definition. The start point of the
+ *   umask for different type has 0x10 offset.
+ *
+ * For example, there are three types of IIO free running counters on Skylake
+ * server, IO CLOCKS counters, BANDWIDTH counters and UTILIZATION counters.
+ * The event code for all the free running counters is 0xff.
+ * 'ioclk' is the first counter of IO CLOCKS. IO CLOCKS is the first type,
+ * which umask starts from 0x10.
+ * So 'ioclk' is encoded as event=0xff,umask=0x10
+ * 'bw_in_port2' is the third counter of BANDWIDTH counters. BANDWIDTH is

arch/x86/events/intel/uncore.h static inline unsigned uncore_msr_fixed_ctr(struct intel_uncore_box *box) · line 273

Quoted 10 lines
arch/x86/events/intel/uncore.h +10 -0
+ * the second type, which umask starts from 0x20.
+ * So 'bw_in_port2' is encoded as event=0xff,umask=0x22
+ */
+static inline unsigned int uncore_freerunning_idx(u64 config)
+{
+	return ((config >> 8) & 0xf);
+}
+
+#define UNCORE_FREERUNNING_UMASK_START		0x10
+static inline unsigned int uncore_freerunning_type(u64 config)

Groan. I asked that before that you please stop glueing defines right in front of a function declaration w/o any visible space. Is it that hard to read and address _ALL_ review comments?

Quoted 5 lines

arch/x86/events/intel/uncore.h static inline unsigned uncore_msr_fixed_ctr(struct intel_uncore_box *box) · line 276

Quoted 3 lines
arch/x86/events/intel/uncore.h +3 -0
+{
+	return ((((config >> 8) - UNCORE_FREERUNNING_UMASK_START) >> 4) & 0xf);
+}

Other than that this looks reasonable.

Hit send to early:

#8 lore raw

PATCH v4 5/8 perf/x86/intel/uncore: add infrastructure for free running counter

In reply to [email protected]

From: Kan Liang <[email protected]>

The free running counter is read-only and always active. Current generic uncore code does not support this kind of counters.

The free running counter is read-only. It cannot be enable/disable in event_start/stop.
The free running counter event and free running counter are 1:1 mapped. It doesn't need to be assigned nor tracked by event_list. The free running counters are always active. It doesn't need to check the free running counters' availability.

Use inline helpers to check fixed counter index and free running counter index.

---

Changes since V3:
- Split the patch to use the data structures and inline helpers which are introduced from previous patch.

 arch/x86/events/intel/uncore.c | 68 +++++++++++++++++++++++++++++++++++++++---
 1 file changed, 64 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +64 -4
index 3b8cd88..9849ba1 100644
@@ -203,7 +203,7 @@ static void uncore_assign_hw_event(struct intel_uncore_box *box,
 	hwc->idx = idx;
 	hwc->last_tag = ++box->tags[idx];
 
-	if (hwc->idx == UNCORE_PMC_IDX_FIXED) {
+	if (uncore_pmc_fixed(hwc->idx)) {
 		hwc->event_base = uncore_fixed_ctr(box);
 		hwc->config_base = uncore_fixed_ctl(box);
 		return;
@@ -218,7 +218,9 @@ void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *e
 	u64 prev_count, new_count, delta;
 	int shift;
 
-	if (event->hw.idx == UNCORE_PMC_IDX_FIXED)
+	if (uncore_pmc_freerunning(event->hw.idx))
+		shift = 64 - uncore_freerunning_bits(box, event);
+	else if (uncore_pmc_fixed(event->hw.idx))
 		shift = 64 - uncore_fixed_ctr_bits(box);
 	else
 		shift = 64 - uncore_perf_ctr_bits(box);
@@ -454,10 +456,25 @@ static void uncore_pmu_event_start(struct perf_event *event, int flags)
 	struct intel_uncore_box *box = uncore_event_to_box(event);
 	int idx = event->hw.idx;
 
-	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
+	if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX))
 		return;
 
-	if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX))
+	/*
+	 * Free running counter is read-only and always active.
+	 * Use the current counter value as start point.
+	 * There is no overflow interrupt for free running counter.
+	 * Use hrtimer to periodically poll the counter to avoid overflow.
+	 */
+	if (uncore_pmc_freerunning(event->hw.idx)) {
+		list_add_tail(&event->active_entry, &box->active_list);
+		local64_set(&event->hw.prev_count,
+			    uncore_read_counter(box, event));
+		if (box->n_active++ == 0)
+			uncore_pmu_start_hrtimer(box);
+		return;
+	}
+
+	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
 		return;
 
 	event->hw.state = 0;
@@ -479,6 +496,15 @@ static void uncore_pmu_event_stop(struct perf_event *event, int flags)
 	struct intel_uncore_box *box = uncore_event_to_box(event);
 	struct hw_perf_event *hwc = &event->hw;
 
+	/* Cannot disable free running counter which is read-only */
+	if (uncore_pmc_freerunning(hwc->idx)) {
+		list_del(&event->active_entry);
+		if (--box->n_active == 0)
+			uncore_pmu_cancel_hrtimer(box);
+		uncore_perf_event_update(box, event);
+		return;
+	}
+
 	if (__test_and_clear_bit(hwc->idx, box->active_mask)) {
 		uncore_disable_event(box, event);
 		box->n_active--;
@@ -512,6 +538,17 @@ static int uncore_pmu_event_add(struct perf_event *event, int flags)
 	if (!box)
 		return -ENODEV;
 
+	/*
+	 * The free funning counter is assigned in event_init().
+	 * The free running counter event and free running counter
+	 * are 1:1 mapped. It doesn't need to be tracked in event_list.
+	 */
+	if (uncore_pmc_freerunning(hwc->idx)) {
+		if (flags & PERF_EF_START)
+			uncore_pmu_event_start(event, 0);
+		return 0;
+	}
+
 	ret = n = uncore_collect_events(box, event, false);
 	if (ret < 0)
 		return ret;
@@ -570,6 +607,14 @@ static void uncore_pmu_event_del(struct perf_event *event, int flags)
 
 	uncore_pmu_event_stop(event, PERF_EF_UPDATE);
 
+	/*
+	 * The event for free running counter is not tracked by event_list.
+	 * It doesn't need to force event->hw.idx = -1 to reassign the counter.
+	 * Because the event and the free running counter are 1:1 mapped.
+	 */
+	if (uncore_pmc_freerunning(event->hw.idx))
+		return;
+
 	for (i = 0; i < box->n_events; i++) {
 		if (event == box->event_list[i]) {
 			uncore_put_event_constraint(box, event);
@@ -603,6 +648,10 @@ static int uncore_validate_group(struct intel_uncore_pmu *pmu,
 	struct intel_uncore_box *fake_box;
 	int ret = -EINVAL, n;
 
+	/* The free running counter is always active. */
+	if (uncore_pmc_freerunning(event->hw.idx))
+		return 0;
+
 	fake_box = uncore_alloc_box(pmu->type, NUMA_NO_NODE);
 	if (!fake_box)
 		return -ENOMEM;
@@ -690,6 +739,17 @@ static int uncore_pmu_event_init(struct perf_event *event)
 
 		/* fixed counters have event field hardcoded to zero */
 		hwc->config = 0ULL;
+	} else if (is_freerunning_event(event)) {
+		if (!check_valid_freerunning_event(box, event))
+			return -EINVAL;
+		event->hw.idx = UNCORE_PMC_IDX_FREERUNNING;
+		/*
+		 * The free running counter event and free running counter
+		 * are always 1:1 mapped.
+		 * The free running counter is always active.
+		 * Assign the free running counter here.
+		 */
+		event->hw.event_base = uncore_freerunning_counter(box, event);
 	} else {
 		hwc->config = event->attr.config &
 			      (pmu->type->event_mask | ((u64)pmu->type->event_mask_ext << 32));
Signature · 1 line
2.7.4
Reviewed-by on v4 5/8
#9 lore raw

Re: [PATCH V4 5/8] perf/x86/intel/uncore: add infrastructure for free running counter

Quoted 12 lines · On Thu, 2 Nov 2017, [email protected] wrote:

From: Kan Liang <[email protected]>

The free running counter is read-only and always active. Current generic uncore code does not support this kind of counters.

The free running counter is read-only. It cannot be enable/disable in event_start/stop.
The free running counter event and free running counter are 1:1 mapped. It doesn't need to be assigned nor tracked by event_list. The free running counters are always active. It doesn't need to check the free running counters' availability.

Quoted 3 lines

Use inline helpers to check fixed counter index and free running counter index.

That changelog is horrible to read. Other than that:

author
#10 lore raw

RE: [PATCH V4 5/8] perf/x86/intel/uncore: add infrastructure for free running counter

Quoted 14 lines · On Thu, 2 Nov 2017, [email protected] wrote:
Quoted 13 lines

From: Kan Liang <[email protected]>

The free running counter is read-only and always active. Current generic uncore code does not support this kind of counters.

The free running counter is read-only. It cannot be enable/disable in event_start/stop.
The free running counter event and free running counter are 1:1 mapped. It doesn't need to be assigned nor tracked by event_list. The free running counters are always active. It doesn't need to check the free running counters' availability.

Quoted 6 lines
Quoted 3 lines

Use inline helpers to check fixed counter index and free running counter index.

That changelog is horrible to read. Other than that:

Hi Thomas,

Thanks for your review.

Sorry about the changelog.
I will modify the changelog and resend all the patches.

Thanks,
Kan

Quoted 1 line
#11 lore raw

PATCH v4 6/8 perf/x86/intel/uncore: SKX support for IIO free running counters

In reply to [email protected]

From: Kan Liang <[email protected]>

As of Skylake Server, there are a number of free-running counters in each IIO Box that collect counts for per box IO clocks and per Port Input/Output x BW/Utilization.

The free running counter is read-only and always active. Counting will be suspended only when the IIO Box is powered down.

There are three types of IIO free running counters on Skylake server, IO CLOCKS counter, BANDWIDTH counters and UTILIZATION counters. IO CLOCKS counter is to count IO clocks.
BANDWIDTH counters are to count inbound(PCIe->CPU)/outbound(CPU->PCIe) bandwidth.
UTILIZATION counters are to count input/output utilization.

The bit width of the free running counters is 36-bits.

---

No changes since V3

 arch/x86/events/intel/uncore_snbep.c | 58 ++++++++++++++++++++++++++++++++++++
 1 file changed, 58 insertions(+)
arch/x86/events/intel/uncore_snbep.c +58 -0
index db1fe37..4ee1e86 100644
@@ -3459,6 +3459,61 @@ static struct intel_uncore_ops skx_uncore_iio_ops = {
 	.read_counter		= uncore_msr_read_counter,
 };
 
+enum perf_uncore_iio_freerunning_type_id {
+	SKX_IIO_MSR_IOCLK			= 0,
+	SKX_IIO_MSR_BW				= 1,
+	SKX_IIO_MSR_UTIL			= 2,
+
+	SKX_IIO_FREERUNNING_TYPE_MAX,
+};
+
+
+static struct freerunning_counters skx_iio_freerunning[] = {
+	[SKX_IIO_MSR_IOCLK]	= { 0xa45, 0x1, 0x20, 1, 36 },
+	[SKX_IIO_MSR_BW]	= { 0xb00, 0x1, 0x10, 8, 36 },
+	[SKX_IIO_MSR_UTIL]	= { 0xb08, 0x1, 0x10, 8, 36 },
+};
+
+static struct uncore_event_desc skx_uncore_iio_events[] = {
+	/* Free-Running IO CLOCKS Counter */
+	INTEL_UNCORE_EVENT_DESC(ioclk,			"event=0xff,umask=0x10"),
+	/* Free-Running IIO BANDWIDTH Counters */
+	INTEL_UNCORE_EVENT_DESC(bw_in_port0,		"event=0xff,umask=0x20"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port0.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port0.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port1,		"event=0xff,umask=0x21"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port1.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port1.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port2,		"event=0xff,umask=0x22"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port2.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port2.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port3,		"event=0xff,umask=0x23"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port3.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_in_port3.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port0,		"event=0xff,umask=0x24"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port0.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port0.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port1,		"event=0xff,umask=0x25"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port1.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port1.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port2,		"event=0xff,umask=0x26"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port2.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port2.unit,	"MiB"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port3,		"event=0xff,umask=0x27"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port3.scale,	"3.814697266e-6"),
+	INTEL_UNCORE_EVENT_DESC(bw_out_port3.unit,	"MiB"),
+	/* Free-running IIO UTILIZATION Counters */
+	INTEL_UNCORE_EVENT_DESC(util_in_port0,		"event=0xff,umask=0x30"),
+	INTEL_UNCORE_EVENT_DESC(util_out_port0,		"event=0xff,umask=0x31"),
+	INTEL_UNCORE_EVENT_DESC(util_in_port1,		"event=0xff,umask=0x32"),
+	INTEL_UNCORE_EVENT_DESC(util_out_port1,		"event=0xff,umask=0x33"),
+	INTEL_UNCORE_EVENT_DESC(util_in_port2,		"event=0xff,umask=0x34"),
+	INTEL_UNCORE_EVENT_DESC(util_out_port2,		"event=0xff,umask=0x35"),
+	INTEL_UNCORE_EVENT_DESC(util_in_port3,		"event=0xff,umask=0x36"),
+	INTEL_UNCORE_EVENT_DESC(util_out_port3,		"event=0xff,umask=0x37"),
+	{ /* end: all zeroes */ },
+};
+
 static struct intel_uncore_type skx_uncore_iio = {
 	.name			= "iio",
 	.num_counters		= 4,
@@ -3470,8 +3525,11 @@ static struct intel_uncore_type skx_uncore_iio = {
 	.event_mask_ext		= SKX_IIO_PMON_RAW_EVENT_MASK_EXT,
 	.box_ctl		= SKX_IIO0_MSR_PMON_BOX_CTL,
 	.msr_offset		= SKX_IIO_MSR_OFFSET,
+	.num_freerunning_types	= SKX_IIO_FREERUNNING_TYPE_MAX,
+	.freerunning		= skx_iio_freerunning,
 	.constraints		= skx_uncore_iio_constraints,
 	.ops			= &skx_uncore_iio_ops,
+	.event_descs		= skx_uncore_iio_events,
 	.format_group		= &skx_uncore_iio_format_group,
 };
 
Signature · 1 line
2.7.4
Reviewed-by on v4 6/8
#12 lore raw

Re: [PATCH V4 6/8] perf/x86/intel/uncore: SKX support for IIO free running counters

Quoted 16 lines · On Thu, 2 Nov 2017, [email protected] wrote:

From: Kan Liang <[email protected]>

As of Skylake Server, there are a number of free-running counters in each IIO Box that collect counts for per box IO clocks and per Port Input/Output x BW/Utilization.

The free running counter is read-only and always active. Counting will be suspended only when the IIO Box is powered down.

There are three types of IIO free running counters on Skylake server, IO CLOCKS counter, BANDWIDTH counters and UTILIZATION counters. IO CLOCKS counter is to count IO clocks.
BANDWIDTH counters are to count inbound(PCIe->CPU)/outbound(CPU->PCIe) bandwidth.
UTILIZATION counters are to count input/output utilization.

Quoted 3 lines

The bit width of the free running counters is 36-bits.

Assumed that all the magic numbers are correct:

#13 lore raw

PATCH v4 7/8 perf/x86/intel/uncore: expose uncore_pmu_event functions

In reply to [email protected]

From: Kan Liang <[email protected]>

Some uncore has custom pmu. For custom pmu, it does not need to customize everything. For example, it only needs to customize init() function for client IMC uncore. Other functions like add()/del()/start()/stop()/read() can use generic code.

Expose the uncore_pmu_event_add/del/start/stop functions.

---

No changes since V3

 arch/x86/events/intel/uncore.c | 8 ++++----
 arch/x86/events/intel/uncore.h | 4 ++++
 2 files changed, 8 insertions(+), 4 deletions(-)
arch/x86/events/intel/uncore.c +4 -4
index 9849ba1..77cc975 100644
@@ -451,7 +451,7 @@ static int uncore_assign_events(struct intel_uncore_box *box, int assign[], int
 	return ret ? -EINVAL : 0;
 }
 
-static void uncore_pmu_event_start(struct perf_event *event, int flags)
+void uncore_pmu_event_start(struct perf_event *event, int flags)
 {
 	struct intel_uncore_box *box = uncore_event_to_box(event);
 	int idx = event->hw.idx;
@@ -491,7 +491,7 @@ static void uncore_pmu_event_start(struct perf_event *event, int flags)
 	}
 }
 
-static void uncore_pmu_event_stop(struct perf_event *event, int flags)
+void uncore_pmu_event_stop(struct perf_event *event, int flags)
 {
 	struct intel_uncore_box *box = uncore_event_to_box(event);
 	struct hw_perf_event *hwc = &event->hw;
@@ -528,7 +528,7 @@ static void uncore_pmu_event_stop(struct perf_event *event, int flags)
 	}
 }
 
-static int uncore_pmu_event_add(struct perf_event *event, int flags)
+int uncore_pmu_event_add(struct perf_event *event, int flags)
 {
 	struct intel_uncore_box *box = uncore_event_to_box(event);
 	struct hw_perf_event *hwc = &event->hw;
@@ -600,7 +600,7 @@ static int uncore_pmu_event_add(struct perf_event *event, int flags)
 	return 0;
 }
 
-static void uncore_pmu_event_del(struct perf_event *event, int flags)
+void uncore_pmu_event_del(struct perf_event *event, int flags)
 {
 	struct intel_uncore_box *box = uncore_event_to_box(event);
 	int i;
arch/x86/events/intel/uncore.h +4 -0
index a508bc5..abb6e9a 100644
@@ -461,6 +461,10 @@ struct intel_uncore_box *uncore_pmu_to_box(struct intel_uncore_pmu *pmu, int cpu
 u64 uncore_msr_read_counter(struct intel_uncore_box *box, struct perf_event *event);
 void uncore_pmu_start_hrtimer(struct intel_uncore_box *box);
 void uncore_pmu_cancel_hrtimer(struct intel_uncore_box *box);
+void uncore_pmu_event_start(struct perf_event *event, int flags);
+void uncore_pmu_event_stop(struct perf_event *event, int flags);
+int uncore_pmu_event_add(struct perf_event *event, int flags);
+void uncore_pmu_event_del(struct perf_event *event, int flags);
 void uncore_pmu_event_read(struct perf_event *event);
 void uncore_perf_event_update(struct intel_uncore_box *box, struct perf_event *event);
 struct event_constraint *
Signature · 1 line
2.7.4
Reviewed-by on v4 7/8
#14 lore raw

Re: [PATCH V4 7/8] perf/x86/intel/uncore: expose uncore_pmu_event functions

Quoted 10 lines · On Thu, 2 Nov 2017, [email protected] wrote:

From: Kan Liang <[email protected]>

Some uncore has custom pmu. For custom pmu, it does not need to customize everything. For example, it only needs to customize init() function for client IMC uncore. Other functions like add()/del()/start()/stop()/read() can use generic code.

Expose the uncore_pmu_event_add/del/start/stop functions.

#15 lore raw

PATCH v4 8/8 perf/x86/intel/uncore: clean up client IMC uncore

In reply to [email protected]

From: Kan Liang <[email protected]>

The counters in client IMC uncore are free running counters, not fixed counters. It should be corrected. The new infrastructure for free running counter should be applied.

Introduce free running counter type SNB_PCI_UNCORE_IMC_DATA for data read and data write counters.

Keep the custom event_init() function compatible with old event encoding.

Clean up other custom event_* functions.

---

No changes since V3

 arch/x86/events/intel/uncore_snb.c | 132 ++++++-------------------------------
 1 file changed, 20 insertions(+), 112 deletions(-)
arch/x86/events/intel/uncore_snb.c +20 -112
index b6d0d72..29473e9 100644
@@ -284,6 +284,15 @@ static struct uncore_event_desc snb_uncore_imc_events[] = {
 #define SNB_UNCORE_PCI_IMC_DATA_WRITES_BASE	0x5054
 #define SNB_UNCORE_PCI_IMC_CTR_BASE		SNB_UNCORE_PCI_IMC_DATA_READS_BASE
 
+enum perf_snb_uncore_imc_freerunning_types {
+	SNB_PCI_UNCORE_IMC_DATA		= 0,
+	SNB_PCI_UNCORE_IMC_FREERUNNING_TYPE_MAX,
+};
+
+static struct freerunning_counters snb_uncore_imc_freerunning[] = {
+	[SNB_PCI_UNCORE_IMC_DATA]     = { SNB_UNCORE_PCI_IMC_DATA_READS_BASE, 0x4, 0x0, 2, 32 },
+};
+
 static struct attribute *snb_uncore_imc_formats_attr[] = {
 	&format_attr_event.attr,
 	NULL,
@@ -340,9 +349,8 @@ static u64 snb_uncore_imc_read_counter(struct intel_uncore_box *box, struct perf
 }
 
 /*
- * custom event_init() function because we define our own fixed, free
- * running counters, so we do not want to conflict with generic uncore
- * logic. Also simplifies processing
+ * Keep the custom event_init() function compatible with old event
+ * encoding for free running counters.
  */
 static int snb_uncore_imc_event_init(struct perf_event *event)
 {
@@ -404,11 +412,11 @@ static int snb_uncore_imc_event_init(struct perf_event *event)
 	switch (cfg) {
 	case SNB_UNCORE_PCI_IMC_DATA_READS:
 		base = SNB_UNCORE_PCI_IMC_DATA_READS_BASE;
-		idx = UNCORE_PMC_IDX_FIXED;
+		idx = UNCORE_PMC_IDX_FREERUNNING;
 		break;
 	case SNB_UNCORE_PCI_IMC_DATA_WRITES:
 		base = SNB_UNCORE_PCI_IMC_DATA_WRITES_BASE;
-		idx = UNCORE_PMC_IDX_FIXED + 1;
+		idx = UNCORE_PMC_IDX_FREERUNNING;
 		break;
 	default:
 		return -EINVAL;
@@ -429,104 +437,6 @@ static int snb_uncore_imc_hw_config(struct intel_uncore_box *box, struct perf_ev
 	return 0;
 }
 
-static void snb_uncore_imc_event_start(struct perf_event *event, int flags)
-{
-	struct intel_uncore_box *box = uncore_event_to_box(event);
-	u64 count;
-
-	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
-		return;
-
-	event->hw.state = 0;
-	box->n_active++;
-
-	list_add_tail(&event->active_entry, &box->active_list);
-
-	count = snb_uncore_imc_read_counter(box, event);
-	local64_set(&event->hw.prev_count, count);
-
-	if (box->n_active == 1)
-		uncore_pmu_start_hrtimer(box);
-}
-
-static void snb_uncore_imc_event_read(struct perf_event *event)
-{
-	struct intel_uncore_box *box = uncore_event_to_box(event);
-	u64 prev_count, new_count, delta;
-	int shift;
-
-	/*
-	 * There are two free running counters in IMC.
-	 * The index for the second one is hardcoded to
-	 * UNCORE_PMC_IDX_FIXED + 1.
-	 */
-	if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
-		shift = 64 - uncore_fixed_ctr_bits(box);
-	else
-		shift = 64 - uncore_perf_ctr_bits(box);
-
-	/* the hrtimer might modify the previous event value */
-again:
-	prev_count = local64_read(&event->hw.prev_count);
-	new_count = uncore_read_counter(box, event);
-	if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
-		goto again;
-
-	delta = (new_count << shift) - (prev_count << shift);
-	delta >>= shift;
-
-	local64_add(delta, &event->count);
-}
-
-static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)
-{
-	struct intel_uncore_box *box = uncore_event_to_box(event);
-	struct hw_perf_event *hwc = &event->hw;
-
-	if (!(hwc->state & PERF_HES_STOPPED)) {
-		box->n_active--;
-
-		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
-		hwc->state |= PERF_HES_STOPPED;
-
-		list_del(&event->active_entry);
-
-		if (box->n_active == 0)
-			uncore_pmu_cancel_hrtimer(box);
-	}
-
-	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
-		/*
-		 * Drain the remaining delta count out of a event
-		 * that we are disabling:
-		 */
-		snb_uncore_imc_event_read(event);
-		hwc->state |= PERF_HES_UPTODATE;
-	}
-}
-
-static int snb_uncore_imc_event_add(struct perf_event *event, int flags)
-{
-	struct intel_uncore_box *box = uncore_event_to_box(event);
-	struct hw_perf_event *hwc = &event->hw;
-
-	if (!box)
-		return -ENODEV;
-
-	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
-	if (!(flags & PERF_EF_START))
-		hwc->state |= PERF_HES_ARCH;
-
-	snb_uncore_imc_event_start(event, 0);
-
-	return 0;
-}
-
-static void snb_uncore_imc_event_del(struct perf_event *event, int flags)
-{
-	snb_uncore_imc_event_stop(event, PERF_EF_UPDATE);
-}
-
 int snb_pci2phy_map_init(int devid)
 {
 	struct pci_dev *dev = NULL;
@@ -558,11 +468,11 @@ int snb_pci2phy_map_init(int devid)
 static struct pmu snb_uncore_imc_pmu = {
 	.task_ctx_nr	= perf_invalid_context,
 	.event_init	= snb_uncore_imc_event_init,
-	.add		= snb_uncore_imc_event_add,
-	.del		= snb_uncore_imc_event_del,
-	.start		= snb_uncore_imc_event_start,
-	.stop		= snb_uncore_imc_event_stop,
-	.read		= snb_uncore_imc_event_read,
+	.add		= uncore_pmu_event_add,
+	.del		= uncore_pmu_event_del,
+	.start		= uncore_pmu_event_start,
+	.stop		= uncore_pmu_event_stop,
+	.read		= uncore_pmu_event_read,
 };
 
 static struct intel_uncore_ops snb_uncore_imc_ops = {
@@ -580,12 +490,10 @@ static struct intel_uncore_type snb_uncore_imc = {
 	.name		= "imc",
 	.num_counters   = 2,
 	.num_boxes	= 1,
-	.fixed_ctr_bits	= 32,
-	.fixed_ctr	= SNB_UNCORE_PCI_IMC_CTR_BASE,
+	.num_freerunning_types	= SNB_PCI_UNCORE_IMC_FREERUNNING_TYPE_MAX,
+	.freerunning	= snb_uncore_imc_freerunning,
 	.event_descs	= snb_uncore_imc_events,
 	.format_group	= &snb_uncore_imc_format_group,
-	.perf_ctr	= SNB_UNCORE_PCI_IMC_DATA_READS_BASE,
-	.event_mask	= SNB_UNCORE_PCI_IMC_EVENT_MASK,
 	.ops		= &snb_uncore_imc_ops,
 	.pmu		= &snb_uncore_imc_pmu,
 };
Signature · 1 line
2.7.4
Reviewed-by on v4 8/8
#16 lore raw

Re: [PATCH V4 8/8] perf/x86/intel/uncore: clean up client IMC uncore

Quoted 9 lines · On Thu, 2 Nov 2017, [email protected] wrote:

From: Kan Liang <[email protected]>

The counters in client IMC uncore are free running counters, not fixed counters. It should be corrected. The new infrastructure for free running counter should be applied.

Introduce free running counter type SNB_PCI_UNCORE_IMC_DATA for data read and data write counters.

Quoted 6 lines

Keep the custom event_init() function compatible with old event encoding.

Clean up other custom event_* functions.

author
#17 lore raw
In reply to [email protected]

Hi Thomas,

Any comments for this patch series?

Thanks,
Kan

Quoted 89 lines

From: Kan Liang <[email protected]>

There are two free running counters for client IMC uncore. The custom event_init() function hardcode their index to 'UNCORE_PMC_IDX_FIXED' and 'UNCORE_PMC_IDX_FIXED + 1'. To support the 'UNCORE_PMC_IDX_FIXED + 1'
case, the generic uncore_perf_event_update is obscurely hacked. The code quality issue will bring problem when new counter index is introduced into generic code. For example, free running counter index.

Introduce customized event_read function for client IMC uncore. The customized function is exactly copied from previous generic uncore_pmu_event_read.
The 'UNCORE_PMC_IDX_FIXED + 1' case will be isolated for client IMC uncore only.

---

Change since V3:
- Use the customized read function to replace uncore_perf_event_update.
- Move generic code change to patch 3/8.

 arch/x86/events/intel/uncore_snb.c | 33
+++++++++++++++++++++++++++++++--
 1 file changed, 31 insertions(+), 2 deletions(-)
arch/x86/events/intel/uncore_snb.c +28 -0
index db1127c..b6d0d72 100644
@@ -449,6 +449,35 @@ static void snb_uncore_imc_event_start(struct
 perf_event *event, int flags)
 		uncore_pmu_start_hrtimer(box);
 }
 
+static void snb_uncore_imc_event_read(struct perf_event *event) {
+	struct intel_uncore_box *box = uncore_event_to_box(event);
+	u64 prev_count, new_count, delta;
+	int shift;
+
+	/*
+	 * There are two free running counters in IMC.
+	 * The index for the second one is hardcoded to
+	 * UNCORE_PMC_IDX_FIXED + 1.
+	 */
+	if (event->hw.idx >= UNCORE_PMC_IDX_FIXED)
+		shift = 64 - uncore_fixed_ctr_bits(box);
+	else
+		shift = 64 - uncore_perf_ctr_bits(box);
+
+	/* the hrtimer might modify the previous event value */
+again:
+	prev_count = local64_read(&event->hw.prev_count);
+	new_count = uncore_read_counter(box, event);
+	if (local64_xchg(&event->hw.prev_count, new_count) != prev_count)
+		goto again;
+
+	delta = (new_count << shift) - (prev_count << shift);
+	delta >>= shift;
+
+	local64_add(delta, &event->count);
+}
+
 static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)  {
 	struct intel_uncore_box *box = uncore_event_to_box(event); @@ -

471,7 +500,7 @@ static void snb_uncore_imc_event_stop(struct perf_event *event, int flags)

arch/x86/events/intel/uncore_snb.c +2 -3
 		 * Drain the remaining delta count out of a event
 		 * that we are disabling:
 		 */
-		uncore_perf_event_update(box, event);
+		snb_uncore_imc_event_read(event);
 		hwc->state |= PERF_HES_UPTODATE;
 	}
 }
@@ -533,7 +562,7 @@ static struct pmu snb_uncore_imc_pmu = {
 	.del		= snb_uncore_imc_event_del,
 	.start		= snb_uncore_imc_event_start,
 	.stop		= snb_uncore_imc_event_stop,
-	.read		= uncore_pmu_event_read,
+	.read		= snb_uncore_imc_event_read,
 };
 
 static struct intel_uncore_ops snb_uncore_imc_ops = {
--

2.7.4

#18 lore raw
Quoted 3 lines · Kan

Hi Thomas,

Any comments for this patch series?

it's on my todo list.

author
#19 lore raw

Hi Thomas,

Did you get a chance to review the patch series?

Thanks,
Kan

Quoted 8 lines · Kan
Quoted 4 lines

Hi Thomas,

Any comments for this patch series?

it's on my todo list.