[PATCH V3 2/5] perf/x86/intel/uncore: add infrastructure for free running counter

[kan . liang]

From: Kan Liang <Kan.liang@xxxxxxxxx>

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 freerunning counters to collect
Input/Output x BW/Utilization.
The precious generic counters could be used 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 freerunning counters. Because the free running
counter is always active, and the event and free running counter are
always 1:1 mapped. It doesnot need extra index to indicate the assigned
counter.
The free running counter is read-only. So it cannot be disabled/enabled,
which is specially handled in generic code. It doesnot need to be
tracked in the events list.
There is no overflow interrupt for free running counter. Use hrtimer to
periodically poll the counter to avoid overflow.

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
The event codes for IIO free running counters on Skylake server can be
found in next patch.

Signed-off-by: Kan Liang <Kan.liang@xxxxxxxxx>
---

Changes since V2:
 - Modify the changelog and comments.
 - The free running counter can be in MSR space or PCI space.
   Use generic name counter_base/offset to replace msr_base/off.
   The related function name is also changed accordingly
 - Use hrtimer to periodically poll the counter to avoid overflow.

 arch/x86/events/intel/uncore.c |  65 +++++++++++++++++++++--
 arch/x86/events/intel/uncore.h | 118 ++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 178 insertions(+), 5 deletions(-)

diff --git a/arch/x86/events/intel/uncore.c b/arch/x86/events/intel/uncore.c
index 3b8cd88..c05971b 100644
--- a/arch/x86/events/intel/uncore.c
+++ b/arch/x86/events/intel/uncore.c
@@ -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;
 
+	/*
+	 * Doesn't need to assign free running counter for event.
+	 * Doesn't need to track the event in event_list.
+	 * They are 1:1 mapped. The free running counter is always active.
+	 */
+	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);
 
+	/*
+	 * Event for free running counter is not tracked by event_list.
+	 * Event and free running counter are 1:1 mapped.
+	 * Doesn't need to force event->hw.idx = -1 to reassign the counter.
+	 */
+	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 +648,13 @@ static int uncore_validate_group(struct intel_uncore_pmu *pmu,
 	struct intel_uncore_box *fake_box;
 	int ret = -EINVAL, n;
 
+	/*
+	 * Event and free running counter are 1:1 mapped
+	 * which is always available.
+	 */
+	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 +742,11 @@ 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;
+		event->hw.event_base = uncore_freerunning_event(box, event);
 	} else {
 		hwc->config = event->attr.config &
 			      (pmu->type->event_mask | ((u64)pmu->type->event_mask_ext << 32));
diff --git a/arch/x86/events/intel/uncore.h b/arch/x86/events/intel/uncore.h
index df5989f..0ed0e54 100644
--- a/arch/x86/events/intel/uncore.h
+++ b/arch/x86/events/intel/uncore.h
@@ -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_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);
+	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)
-- 
2.7.4