[PATCH 0/3] perf/arm_dsu: Support newer DSUs

[PATCH 0/3] perf/arm_dsu: Support newer DSUs

[Robin Murphy]

Hi all,

DSU-110 and 120 have been out for a while, and although they continue
to share the same basic system register interface, they have subtle
functional changes that require a little attention to support properly.
And while DSUs still aren't widely exposed in current systems, at least
some folks do apparently care.

Sanity-checked on DSU-120 in FPGA, and original DSU on Rockchip RK3566
with TF-A hacked to set ACTLR_EL3.CLUSTERPMUEN.

Cheers,
Robin.


Robin Murphy (3):
  perf/arm_dsu: Support DSU-110
  perf/arm_dsu: Support DSU-120
  perf/arm_dsu: Allow standard cycles events

 drivers/perf/arm_dsu_pmu.c | 37 +++++++++++++++++++++++--------------
 1 file changed, 23 insertions(+), 14 deletions(-)

-- 
2.39.2.101.g768bb238c484.dirty

[PATCH 1/3] perf/arm_dsu: Support DSU-110

[Robin Murphy]

DSU-110 sneakily made all the event counters 64-bit, perhaps related
to no longer having AArch32 EL1 to worry about. While the DSU version
itself is not easily discoverable, the size of a counter certainly is.

Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---
 drivers/perf/arm_dsu_pmu.c | 24 ++++++++++++++----------
 1 file changed, 14 insertions(+), 10 deletions(-)

diff --git a/drivers/perf/arm_dsu_pmu.c b/drivers/perf/arm_dsu_pmu.c
index cb4fb59fe04b..8663721ee018 100644
--- a/drivers/perf/arm_dsu_pmu.c
+++ b/drivers/perf/arm_dsu_pmu.c
@@ -66,13 +66,6 @@
  */
 #define DSU_PMU_IDX_CYCLE_COUNTER	31
 
-/* All event counters are 32bit, with a 64bit Cycle counter */
-#define DSU_PMU_COUNTER_WIDTH(idx)	\
-	(((idx) == DSU_PMU_IDX_CYCLE_COUNTER) ? 64 : 32)
-
-#define DSU_PMU_COUNTER_MASK(idx)	\
-	GENMASK_ULL((DSU_PMU_COUNTER_WIDTH((idx)) - 1), 0)
-
 #define DSU_EXT_ATTR(_name, _func, _config)		\
 	(&((struct dev_ext_attribute[]) {				\
 		{							\
@@ -107,6 +100,7 @@ struct dsu_hw_events {
  * @num_counters	: Number of event counters implemented by the PMU,
  *			  excluding the cycle counter.
  * @irq			: Interrupt line for counter overflow.
+ * @has_32b_pmevcntr	: Are the non-cycle counters only 32-bit?
  * @cpmceid_bitmap	: Bitmap for the availability of architected common
  *			  events (event_code < 0x40).
  */
@@ -120,6 +114,7 @@ struct dsu_pmu {
 	struct hlist_node		cpuhp_node;
 	s8				num_counters;
 	int				irq;
+	bool				has_32b_pmevcntr;
 	DECLARE_BITMAP(cpmceid_bitmap, DSU_PMU_MAX_COMMON_EVENTS);
 };
 
@@ -328,6 +323,11 @@ static inline void dsu_pmu_set_event(struct dsu_pmu *dsu_pmu,
 	raw_spin_unlock_irqrestore(&dsu_pmu->pmu_lock, flags);
 }
 
+static u64 dsu_pmu_counter_mask(struct hw_perf_event *hw)
+{
+	return (hw->flags && hw->idx != DSU_PMU_IDX_CYCLE_COUNTER) ? U32_MAX : U64_MAX;
+}
+
 static void dsu_pmu_event_update(struct perf_event *event)
 {
 	struct hw_perf_event *hwc = &event->hw;
@@ -339,7 +339,7 @@ static void dsu_pmu_event_update(struct perf_event *event)
 		new_count = dsu_pmu_read_counter(event);
 	} while (local64_cmpxchg(&hwc->prev_count, prev_count, new_count) !=
 			prev_count);
-	delta = (new_count - prev_count) & DSU_PMU_COUNTER_MASK(hwc->idx);
+	delta = (new_count - prev_count) & dsu_pmu_counter_mask(hwc);
 	local64_add(delta, &event->count);
 }
 
@@ -362,8 +362,7 @@ static inline u32 dsu_pmu_get_reset_overflow(void)
  */
 static void dsu_pmu_set_event_period(struct perf_event *event)
 {
-	int idx = event->hw.idx;
-	u64 val = DSU_PMU_COUNTER_MASK(idx) >> 1;
+	u64 val = dsu_pmu_counter_mask(&event->hw) >> 1;
 
 	local64_set(&event->hw.prev_count, val);
 	dsu_pmu_write_counter(event, val);
@@ -564,6 +563,7 @@ static int dsu_pmu_event_init(struct perf_event *event)
 		return -EINVAL;
 
 	event->hw.config_base = event->attr.config;
+	event->hw.flags = dsu_pmu->has_32b_pmevcntr;
 	return 0;
 }
 
@@ -664,6 +664,10 @@ static void dsu_pmu_probe_pmu(struct dsu_pmu *dsu_pmu)
 	cpmceid[1] = __dsu_pmu_read_pmceid(1);
 	bitmap_from_arr32(dsu_pmu->cpmceid_bitmap, cpmceid,
 			  DSU_PMU_MAX_COMMON_EVENTS);
+	/* Newer DSUs have 64-bit counters */
+	__dsu_pmu_write_counter(0, U64_MAX);
+	if (__dsu_pmu_read_counter(0) != U64_MAX)
+		dsu_pmu->has_32b_pmevcntr = true;
 }
 
 static void dsu_pmu_set_active_cpu(int cpu, struct dsu_pmu *dsu_pmu)
-- 
2.39.2.101.g768bb238c484.dirty

[PATCH 2/3] perf/arm_dsu: Support DSU-120

[Robin Murphy]

DSU-120 has the same system register interface as previous DSUs, but
no longer offers a dedicated cycle counter. While this is not directly
discoverable via PMCR, the PMCCNTR register is still defined to exist
with RAZ/WI behaviour, allowing for a straightforward heuristic.

Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---
 drivers/perf/arm_dsu_pmu.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/perf/arm_dsu_pmu.c b/drivers/perf/arm_dsu_pmu.c
index 8663721ee018..56c592f0dae3 100644
--- a/drivers/perf/arm_dsu_pmu.c
+++ b/drivers/perf/arm_dsu_pmu.c
@@ -101,6 +101,7 @@ struct dsu_hw_events {
  *			  excluding the cycle counter.
  * @irq			: Interrupt line for counter overflow.
  * @has_32b_pmevcntr	: Are the non-cycle counters only 32-bit?
+ * @has_pmccntr		: Do we even have a dedicated cycle counter?
  * @cpmceid_bitmap	: Bitmap for the availability of architected common
  *			  events (event_code < 0x40).
  */
@@ -115,6 +116,7 @@ struct dsu_pmu {
 	s8				num_counters;
 	int				irq;
 	bool				has_32b_pmevcntr;
+	bool				has_pmccntr;
 	DECLARE_BITMAP(cpmceid_bitmap, DSU_PMU_MAX_COMMON_EVENTS);
 };
 
@@ -281,7 +283,7 @@ static int dsu_pmu_get_event_idx(struct dsu_hw_events *hw_events,
 	struct dsu_pmu *dsu_pmu = to_dsu_pmu(event->pmu);
 	unsigned long *used_mask = hw_events->used_mask;
 
-	if (evtype == DSU_PMU_EVT_CYCLES) {
+	if (evtype == DSU_PMU_EVT_CYCLES && dsu_pmu->has_pmccntr) {
 		if (test_and_set_bit(DSU_PMU_IDX_CYCLE_COUNTER, used_mask))
 			return -EAGAIN;
 		return DSU_PMU_IDX_CYCLE_COUNTER;
@@ -668,6 +670,10 @@ static void dsu_pmu_probe_pmu(struct dsu_pmu *dsu_pmu)
 	__dsu_pmu_write_counter(0, U64_MAX);
 	if (__dsu_pmu_read_counter(0) != U64_MAX)
 		dsu_pmu->has_32b_pmevcntr = true;
+	/* On even newer DSUs, PMCCNTR is RAZ/WI */
+	__dsu_pmu_write_pmccntr(U64_MAX);
+	if (__dsu_pmu_read_pmccntr() == U64_MAX)
+		dsu_pmu->has_pmccntr = true;
 }
 
 static void dsu_pmu_set_active_cpu(int cpu, struct dsu_pmu *dsu_pmu)
-- 
2.39.2.101.g768bb238c484.dirty

[PATCH 3/3] perf/arm_dsu: Allow standard cycles events

[Robin Murphy]

Since we do not use the divide-by-64 option, there should be no
significant difference between the dedicated cycle counter and the
standard cycles event. Since using the latter on DSU-120 now has
the side-effect of allowing multiple cycles events to be scheduled
simultaneously (beneficial for multiple cycle-based metrics), there
seems little reason not to allow the same on older DSUs as well.

Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---
 drivers/perf/arm_dsu_pmu.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/drivers/perf/arm_dsu_pmu.c b/drivers/perf/arm_dsu_pmu.c
index 56c592f0dae3..32b0dd7c693b 100644
--- a/drivers/perf/arm_dsu_pmu.c
+++ b/drivers/perf/arm_dsu_pmu.c
@@ -284,9 +284,8 @@ static int dsu_pmu_get_event_idx(struct dsu_hw_events *hw_events,
 	unsigned long *used_mask = hw_events->used_mask;
 
 	if (evtype == DSU_PMU_EVT_CYCLES && dsu_pmu->has_pmccntr) {
-		if (test_and_set_bit(DSU_PMU_IDX_CYCLE_COUNTER, used_mask))
-			return -EAGAIN;
-		return DSU_PMU_IDX_CYCLE_COUNTER;
+		if (!test_and_set_bit(DSU_PMU_IDX_CYCLE_COUNTER, used_mask))
+			return DSU_PMU_IDX_CYCLE_COUNTER;
 	}
 
 	idx = find_first_zero_bit(used_mask, dsu_pmu->num_counters);
-- 
2.39.2.101.g768bb238c484.dirty

Re: [PATCH 0/3] perf/arm_dsu: Support newer DSUs

[Will Deacon]

On Mon, 15 Dec 2025 13:04:57 +0000, Robin Murphy wrote:
> DSU-110 and 120 have been out for a while, and although they continue
> to share the same basic system register interface, they have subtle
> functional changes that require a little attention to support properly.
> And while DSUs still aren't widely exposed in current systems, at least
> some folks do apparently care.
> 
> Sanity-checked on DSU-120 in FPGA, and original DSU on Rockchip RK3566
> with TF-A hacked to set ACTLR_EL3.CLUSTERPMUEN.
> 
> [...]

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

[1/3] perf/arm_dsu: Support DSU-110
      https://git.kernel.org/will/c/0113affc9101
[2/3] perf/arm_dsu: Support DSU-120
      https://git.kernel.org/will/c/85c0dbd8b6e2
[3/3] perf/arm_dsu: Allow standard cycles events
      https://git.kernel.org/will/c/79448fa1f495

Cheers,
-- 
Will

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