[PATCH v4 0/2] Mesh PMU: Add tertiary match group support

[PATCH v4 0/2] Mesh PMU: Add tertiary match group support

[Ilkka Koskinen]

v4:	* No changes, only rebased on top of v6.10-rc4

v3:
	* Replaced wp_cfg in arm_cmn_hw_event with wp_idx that keeps
	  track, whether the event uses 0 or 1 index for the given direction
	* Cleaned and simplified allocation/claiming of wp config
	* arm_cmn_val_add_event() can't and won't fail anymore
	* Separated wp_combine from wp[] in event validation phase
	* use memset()/sizeof() when clearing an event
	* Still kept wp config allocator in separate functions - at least
	  for now
	* https://lore.kernel.org/all/20240329013215.169345-1-ilkka@os.amperecomputing.com/

v2:
        * Wp config registers are allocated on node basis, instead
	  of using the same one globally
	* Use bitmap in the event structure to keep track on the
	  wp config registers assigned to the specific event.
	* The bitmap tracks only either UP (wp conf 0&1) or DOWN
	  (wp conf 2&3) registers.
	* Dropped the second patch
	  ("The patch set v2 is now in the internal mailing list")
	  as perf/sysfs doesn't really support items with the same
	  name even if visibility would handle them.
	* Addressed a bunch of other comments by the author
	* https://lore.kernel.org/all/20240307230929.6233-2-ilkka@os.amperecomputing.com/

v1:
	* https://lore.kernel.org/all/20240126221215.1537377-1-ilkka@os.amperecomputing.com/

Ilkka Koskinen (2):
  perf/arm-cmn: Decouple wp_config registers from filter group number
  perf/arm-cmn: Enable support for tertiary match group

 drivers/perf/arm-cmn.c | 116 +++++++++++++++++++++++++++++++++--------
 1 file changed, 93 insertions(+), 23 deletions(-)

-- 
2.40.1

[PATCH v4 1/2] perf/arm-cmn: Decouple wp_config registers from filter group number

[Ilkka Koskinen]

Previously, wp_config0/2 registers were used for primary match group and
wp_config1/3 registers for secondary match group. In order to support
tertiary match group, this patch decouples the registers and the groups.

Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
---
 drivers/perf/arm-cmn.c | 97 ++++++++++++++++++++++++++++++++++--------
 1 file changed, 80 insertions(+), 17 deletions(-)

diff --git a/drivers/perf/arm-cmn.c b/drivers/perf/arm-cmn.c
index e26ad1d3ed0b..f2349d23df57 100644
--- a/drivers/perf/arm-cmn.c
+++ b/drivers/perf/arm-cmn.c
@@ -590,6 +590,13 @@ struct arm_cmn_hw_event {
 	s8 dtc_idx[CMN_MAX_DTCS];
 	u8 num_dns;
 	u8 dtm_offset;
+
+	/*
+	 * WP config registers are divided to UP and DOWN events. We need to
+	 * keep to track only one of them.
+	 */
+	DECLARE_BITMAP(wp_idx, CMN_MAX_XPS);
+
 	bool wide_sel;
 	enum cmn_filter_select filter_sel;
 };
@@ -617,6 +624,17 @@ static unsigned int arm_cmn_get_index(u64 x[], unsigned int pos)
 	return (x[pos / 32] >> ((pos % 32) * 2)) & 3;
 }
 
+static void arm_cmn_set_wp_idx(unsigned long *wp_idx, unsigned int pos, bool val)
+{
+	if (val)
+		set_bit(pos, wp_idx);
+}
+
+static unsigned int arm_cmn_get_wp_idx(unsigned long *wp_idx, unsigned int pos)
+{
+	return test_bit(pos, wp_idx);
+}
+
 struct arm_cmn_event_attr {
 	struct device_attribute attr;
 	enum cmn_model model;
@@ -1336,9 +1354,34 @@ static const struct attribute_group *arm_cmn_attr_groups[] = {
 	NULL
 };
 
-static int arm_cmn_wp_idx(struct perf_event *event)
+static int arm_cmn_find_free_wp_idx(struct arm_cmn_dtm *dtm,
+				    struct perf_event *event)
+{
+	int wp_idx = CMN_EVENT_EVENTID(event);
+
+	if (dtm->wp_event[wp_idx] >= 0)
+		if (dtm->wp_event[++wp_idx] >= 0)
+			return -ENOSPC;
+
+	return wp_idx;
+}
+
+static int arm_cmn_get_assigned_wp_idx(struct perf_event *event,
+				       struct arm_cmn_hw_event *hw,
+				       unsigned int pos)
 {
-	return CMN_EVENT_EVENTID(event) + CMN_EVENT_WP_GRP(event);
+	return CMN_EVENT_EVENTID(event) + arm_cmn_get_wp_idx(hw->wp_idx, pos);
+}
+
+static void arm_cmn_claim_wp_idx(struct arm_cmn_dtm *dtm,
+				 struct perf_event *event,
+				 unsigned int dtc, int wp_idx,
+				 unsigned int pos)
+{
+	struct arm_cmn_hw_event *hw = to_cmn_hw(event);
+
+	dtm->wp_event[wp_idx] = hw->dtc_idx[dtc];
+	arm_cmn_set_wp_idx(hw->wp_idx, pos, wp_idx - CMN_EVENT_EVENTID(event));
 }
 
 static u32 arm_cmn_wp_config(struct perf_event *event)
@@ -1520,12 +1563,12 @@ static void arm_cmn_event_start(struct perf_event *event, int flags)
 		writeq_relaxed(CMN_CC_INIT, cmn->dtc[i].base + CMN_DT_PMCCNTR);
 		cmn->dtc[i].cc_active = true;
 	} else if (type == CMN_TYPE_WP) {
-		int wp_idx = arm_cmn_wp_idx(event);
 		u64 val = CMN_EVENT_WP_VAL(event);
 		u64 mask = CMN_EVENT_WP_MASK(event);
 
 		for_each_hw_dn(hw, dn, i) {
 			void __iomem *base = dn->pmu_base + CMN_DTM_OFFSET(hw->dtm_offset);
+			int wp_idx = arm_cmn_get_assigned_wp_idx(event, hw, i);
 
 			writeq_relaxed(val, base + CMN_DTM_WPn_VAL(wp_idx));
 			writeq_relaxed(mask, base + CMN_DTM_WPn_MASK(wp_idx));
@@ -1550,10 +1593,9 @@ static void arm_cmn_event_stop(struct perf_event *event, int flags)
 		i = hw->dtc_idx[0];
 		cmn->dtc[i].cc_active = false;
 	} else if (type == CMN_TYPE_WP) {
-		int wp_idx = arm_cmn_wp_idx(event);
-
 		for_each_hw_dn(hw, dn, i) {
 			void __iomem *base = dn->pmu_base + CMN_DTM_OFFSET(hw->dtm_offset);
+			int wp_idx = arm_cmn_get_assigned_wp_idx(event, hw, i);
 
 			writeq_relaxed(0, base + CMN_DTM_WPn_MASK(wp_idx));
 			writeq_relaxed(~0ULL, base + CMN_DTM_WPn_VAL(wp_idx));
@@ -1571,10 +1613,23 @@ struct arm_cmn_val {
 	u8 dtm_count[CMN_MAX_DTMS];
 	u8 occupid[CMN_MAX_DTMS][SEL_MAX];
 	u8 wp[CMN_MAX_DTMS][4];
+	u8 wp_combine[CMN_MAX_DTMS][2];
 	int dtc_count[CMN_MAX_DTCS];
 	bool cycles;
 };
 
+static int arm_cmn_val_find_free_wp_config(struct perf_event *event,
+					  struct arm_cmn_val *val, int dtm)
+{
+	int wp_idx = CMN_EVENT_EVENTID(event);
+
+	if (val->wp[dtm][wp_idx])
+		if (val->wp[dtm][++wp_idx])
+			return -ENOSPC;
+
+	return wp_idx;
+}
+
 static void arm_cmn_val_add_event(struct arm_cmn *cmn, struct arm_cmn_val *val,
 				  struct perf_event *event)
 {
@@ -1606,8 +1661,9 @@ static void arm_cmn_val_add_event(struct arm_cmn *cmn, struct arm_cmn_val *val,
 		if (type != CMN_TYPE_WP)
 			continue;
 
-		wp_idx = arm_cmn_wp_idx(event);
-		val->wp[dtm][wp_idx] = CMN_EVENT_WP_COMBINE(event) + 1;
+		wp_idx = arm_cmn_val_find_free_wp_config(event, val, dtm);
+		val->wp[dtm][wp_idx] = 1;
+		val->wp_combine[dtm][wp_idx >> 1] += !!CMN_EVENT_WP_COMBINE(event);
 	}
 }
 
@@ -1631,6 +1687,7 @@ static int arm_cmn_validate_group(struct arm_cmn *cmn, struct perf_event *event)
 		return -ENOMEM;
 
 	arm_cmn_val_add_event(cmn, val, leader);
+
 	for_each_sibling_event(sibling, leader)
 		arm_cmn_val_add_event(cmn, val, sibling);
 
@@ -1645,7 +1702,7 @@ static int arm_cmn_validate_group(struct arm_cmn *cmn, struct perf_event *event)
 			goto done;
 
 	for_each_hw_dn(hw, dn, i) {
-		int wp_idx, wp_cmb, dtm = dn->dtm, sel = hw->filter_sel;
+		int wp_idx, dtm = dn->dtm, sel = hw->filter_sel;
 
 		if (val->dtm_count[dtm] == CMN_DTM_NUM_COUNTERS)
 			goto done;
@@ -1657,12 +1714,12 @@ static int arm_cmn_validate_group(struct arm_cmn *cmn, struct perf_event *event)
 		if (type != CMN_TYPE_WP)
 			continue;
 
-		wp_idx = arm_cmn_wp_idx(event);
-		if (val->wp[dtm][wp_idx])
+		wp_idx = arm_cmn_val_find_free_wp_config(event, val, dtm);
+		if (wp_idx < 0)
 			goto done;
 
-		wp_cmb = val->wp[dtm][wp_idx ^ 1];
-		if (wp_cmb && wp_cmb != CMN_EVENT_WP_COMBINE(event) + 1)
+		if (wp_idx & 1 &&
+		    val->wp_combine[dtm][wp_idx >> 1] != !!CMN_EVENT_WP_COMBINE(event))
 			goto done;
 	}
 
@@ -1773,8 +1830,11 @@ static void arm_cmn_event_clear(struct arm_cmn *cmn, struct perf_event *event,
 		struct arm_cmn_dtm *dtm = &cmn->dtms[hw->dn[i].dtm] + hw->dtm_offset;
 		unsigned int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
 
-		if (type == CMN_TYPE_WP)
-			dtm->wp_event[arm_cmn_wp_idx(event)] = -1;
+		if (type == CMN_TYPE_WP) {
+			int wp_idx = arm_cmn_get_assigned_wp_idx(event, hw, i);
+
+			dtm->wp_event[wp_idx] = -1;
+		}
 
 		if (hw->filter_sel > SEL_NONE)
 			hw->dn[i].occupid[hw->filter_sel].count--;
@@ -1783,6 +1843,7 @@ static void arm_cmn_event_clear(struct arm_cmn *cmn, struct perf_event *event,
 		writel_relaxed(dtm->pmu_config_low, dtm->base + CMN_DTM_PMU_CONFIG);
 	}
 	memset(hw->dtm_idx, 0, sizeof(hw->dtm_idx));
+	memset(hw->wp_idx, 0, sizeof(hw->wp_idx));
 
 	for_each_hw_dtc_idx(hw, j, idx)
 		cmn->dtc[j].counters[idx] = NULL;
@@ -1836,10 +1897,11 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
 		if (type == CMN_TYPE_XP) {
 			input_sel = CMN__PMEVCNT0_INPUT_SEL_XP + dtm_idx;
 		} else if (type == CMN_TYPE_WP) {
-			int tmp, wp_idx = arm_cmn_wp_idx(event);
+			int tmp, wp_idx;
 			u32 cfg = arm_cmn_wp_config(event);
 
-			if (dtm->wp_event[wp_idx] >= 0)
+			wp_idx = arm_cmn_find_free_wp_idx(dtm, event);
+			if (wp_idx < 0)
 				goto free_dtms;
 
 			tmp = dtm->wp_event[wp_idx ^ 1];
@@ -1848,7 +1910,8 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
 				goto free_dtms;
 
 			input_sel = CMN__PMEVCNT0_INPUT_SEL_WP + wp_idx;
-			dtm->wp_event[wp_idx] = hw->dtc_idx[d];
+
+			arm_cmn_claim_wp_idx(dtm, event, d, wp_idx, i);
 			writel_relaxed(cfg, dtm->base + CMN_DTM_WPn_CONFIG(wp_idx));
 		} else {
 			struct arm_cmn_nodeid nid = arm_cmn_nid(cmn, dn->id);
-- 
2.40.1

[PATCH v4 2/2] perf/arm-cmn: Enable support for tertiary match group

[Ilkka Koskinen]

Add support for tertiary match group.

Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
---
 drivers/perf/arm-cmn.c | 19 +++++++++++++------
 1 file changed, 13 insertions(+), 6 deletions(-)

diff --git a/drivers/perf/arm-cmn.c b/drivers/perf/arm-cmn.c
index f2349d23df57..c932d9d355cf 100644
--- a/drivers/perf/arm-cmn.c
+++ b/drivers/perf/arm-cmn.c
@@ -174,9 +174,8 @@
 #define CMN_CONFIG_WP_COMBINE		GENMASK_ULL(30, 27)
 #define CMN_CONFIG_WP_DEV_SEL		GENMASK_ULL(50, 48)
 #define CMN_CONFIG_WP_CHN_SEL		GENMASK_ULL(55, 51)
-/* Note that we don't yet support the tertiary match group on newer IPs */
-#define CMN_CONFIG_WP_GRP		BIT_ULL(56)
-#define CMN_CONFIG_WP_EXCLUSIVE		BIT_ULL(57)
+#define CMN_CONFIG_WP_GRP		GENMASK_ULL(57, 56)
+#define CMN_CONFIG_WP_EXCLUSIVE		BIT_ULL(58)
 #define CMN_CONFIG1_WP_VAL		GENMASK_ULL(63, 0)
 #define CMN_CONFIG2_WP_MASK		GENMASK_ULL(63, 0)
 
@@ -1384,7 +1383,7 @@ static void arm_cmn_claim_wp_idx(struct arm_cmn_dtm *dtm,
 	arm_cmn_set_wp_idx(hw->wp_idx, pos, wp_idx - CMN_EVENT_EVENTID(event));
 }
 
-static u32 arm_cmn_wp_config(struct perf_event *event)
+static u32 arm_cmn_wp_config(struct perf_event *event, int wp_idx)
 {
 	u32 config;
 	u32 dev = CMN_EVENT_WP_DEV_SEL(event);
@@ -1394,6 +1393,10 @@ static u32 arm_cmn_wp_config(struct perf_event *event)
 	u32 combine = CMN_EVENT_WP_COMBINE(event);
 	bool is_cmn600 = to_cmn(event->pmu)->part == PART_CMN600;
 
+	/* CMN-600 supports only primary and secondary matching groups */
+	if (is_cmn600)
+		grp &= 1;
+
 	config = FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_DEV_SEL, dev) |
 		 FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_CHN_SEL, chn) |
 		 FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_GRP, grp) |
@@ -1401,7 +1404,9 @@ static u32 arm_cmn_wp_config(struct perf_event *event)
 	if (exc)
 		config |= is_cmn600 ? CMN600_WPn_CONFIG_WP_EXCLUSIVE :
 				      CMN_DTM_WPn_CONFIG_WP_EXCLUSIVE;
-	if (combine && !grp)
+
+	/*  wp_combine is available only on WP0 and WP2 */
+	if (combine && !(wp_idx & 0x1))
 		config |= is_cmn600 ? CMN600_WPn_CONFIG_WP_COMBINE :
 				      CMN_DTM_WPn_CONFIG_WP_COMBINE;
 	return config;
@@ -1898,12 +1903,14 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
 			input_sel = CMN__PMEVCNT0_INPUT_SEL_XP + dtm_idx;
 		} else if (type == CMN_TYPE_WP) {
 			int tmp, wp_idx;
-			u32 cfg = arm_cmn_wp_config(event);
+			u32 cfg;
 
 			wp_idx = arm_cmn_find_free_wp_idx(dtm, event);
 			if (wp_idx < 0)
 				goto free_dtms;
 
+			cfg = arm_cmn_wp_config(event, wp_idx);
+
 			tmp = dtm->wp_event[wp_idx ^ 1];
 			if (tmp >= 0 && CMN_EVENT_WP_COMBINE(event) !=
 					CMN_EVENT_WP_COMBINE(cmn->dtc[d].counters[tmp]))
-- 
2.40.1

Re: [PATCH v4 0/2] Mesh PMU: Add tertiary match group support

[Will Deacon]

On Mon, 17 Jun 2024 17:50:54 -0700, Ilkka Koskinen wrote:
> v4:	* No changes, only rebased on top of v6.10-rc4
> 
> v3:
> 	* Replaced wp_cfg in arm_cmn_hw_event with wp_idx that keeps
> 	  track, whether the event uses 0 or 1 index for the given direction
> 	* Cleaned and simplified allocation/claiming of wp config
> 	* arm_cmn_val_add_event() can't and won't fail anymore
> 	* Separated wp_combine from wp[] in event validation phase
> 	* use memset()/sizeof() when clearing an event
> 	* Still kept wp config allocator in separate functions - at least
> 	  for now
> 	* https://lore.kernel.org/all/20240329013215.169345-1-ilkka@os.amperecomputing.com/
> 
> [...]

Applied to will (for-next/perf), thanks!

[1/2] perf/arm-cmn: Decouple wp_config registers from filter group number
      https://git.kernel.org/will/c/4a112585ebe8
[2/2] perf/arm-cmn: Enable support for tertiary match group
      https://git.kernel.org/will/c/f9a7a91f6406

Cheers,
-- 
Will

https://fixes.arm64.dev
https://next.arm64.dev
https://will.arm64.dev