[PATCH v2 3/4] perf/marvell: cn10k DDR perfmon event overflow handling

[Bharat Bhushan]

CN10k DSS h/w perfmon does not support event overflow
interrupt, so periodic timer is being used. Each
event counter is 48bit, which in worst case scenario
can increment at maximum 5.6 GT/s. At this rate it may
take many hours to overflow these counters. Therefore
polling period for overflow is set to 100 sec, which
can be changed using sysfs parameter.

Two fixed event counters starts counting from zero on
overflow, so overflow condition is when new count less
than previous count. While eight programmable event
counters freezes at maximum value. Also individual
counter cannot be restarted, so need to restart all
eight counters.

Signed-off-by: Bharat Bhushan <bbhushan2@xxxxxxxxxxx>
---
v1->v2:
 - No change

 drivers/perf/marvell_cn10k_ddr_pmu.c | 111 +++++++++++++++++++++++++++
 1 file changed, 111 insertions(+)

diff --git a/drivers/perf/marvell_cn10k_ddr_pmu.c b/drivers/perf/marvell_cn10k_ddr_pmu.c
index 8f9e3d1fcd8d..80f1441961d0 100644
--- a/drivers/perf/marvell_cn10k_ddr_pmu.c
+++ b/drivers/perf/marvell_cn10k_ddr_pmu.c
@@ -11,6 +11,7 @@
 #include <linux/of_address.h>
 #include <linux/of_device.h>
 #include <linux/perf_event.h>
+#include <linux/hrtimer.h>
 
 /* Performance Counters Operating Mode Control Registers */
 #define DDRC_PERF_CNT_OP_MODE_CTRL	0x8020
@@ -128,6 +129,7 @@ struct cn10k_ddr_pmu {
 	struct	device *dev;
 	int active_events;
 	struct perf_event *events[DDRC_PERF_NUM_COUNTERS];
+	struct hrtimer hrtimer;
 };
 
 #define to_cn10k_ddr_pmu(p)	container_of(p, struct cn10k_ddr_pmu, pmu)
@@ -251,6 +253,18 @@ static const struct attribute_group *cn10k_attr_groups[] = {
 	NULL,
 };
 
+/* Default poll timeout is 100 sec, which is very sufficient for
+ * 48 bit counter incremented max at 5.6 GT/s, which may take many
+ * hours to overflow.
+ */
+static unsigned long cn10k_ddr_pmu_poll_period_sec = 100;
+module_param_named(poll_period_sec, cn10k_ddr_pmu_poll_period_sec, ulong, 0644);
+
+static ktime_t cn10k_ddr_pmu_timer_period(void)
+{
+	return ms_to_ktime((u64)cn10k_ddr_pmu_poll_period_sec * USEC_PER_SEC);
+}
+
 static uint64_t ddr_perf_get_event_bitmap(int eventid)
 {
 	uint64_t event_bitmap = 0;
@@ -439,6 +453,10 @@ static int cn10k_ddr_perf_event_add(struct perf_event *event, int flags)
 	pmu->active_events++;
 	hwc->idx = counter;
 
+	if (pmu->active_events == 1)
+		hrtimer_start(&pmu->hrtimer, cn10k_ddr_pmu_timer_period(),
+			      HRTIMER_MODE_REL_PINNED);
+
 	if (counter < DDRC_PERF_NUM_GEN_COUNTERS) {
 		/* Generic counters, configure event id */
 		reg_offset = DDRC_PERF_CFG(counter);
@@ -487,6 +505,10 @@ static void cn10k_ddr_perf_event_del(struct perf_event *event, int flags)
 	cn10k_ddr_perf_free_counter(pmu, counter);
 	pmu->active_events--;
 	hwc->idx = -1;
+
+	/* Cancel timer when no events to capture */
+	if (pmu->active_events == 0)
+		hrtimer_cancel(&pmu->hrtimer);
 }
 
 static void cn10k_ddr_perf_pmu_enable(struct pmu *pmu)
@@ -505,6 +527,92 @@ static void cn10k_ddr_perf_pmu_disable(struct pmu *pmu)
 		       DDRC_PERF_CNT_END_OP_CTRL);
 }
 
+static void cn10k_ddr_perf_event_update_all(struct cn10k_ddr_pmu *pmu)
+{
+	struct hw_perf_event *hwc;
+	int i;
+
+	for (i = 0; i < DDRC_PERF_NUM_GEN_COUNTERS; i++) {
+		if (pmu->events[i] == NULL)
+			continue;
+
+		cn10k_ddr_perf_event_update(pmu->events[i]);
+	}
+
+	/* Reset previous count as h/w counter are reset */
+	for (i = 0; i < DDRC_PERF_NUM_GEN_COUNTERS; i++) {
+		if (pmu->events[i] == NULL)
+			continue;
+
+		hwc = &pmu->events[i]->hw;
+		local64_set(&hwc->prev_count, 0);
+	}
+}
+
+static irqreturn_t cn10k_ddr_pmu_overflow_handler(struct cn10k_ddr_pmu *pmu)
+{
+	struct perf_event *event;
+	struct hw_perf_event *hwc;
+	uint64_t prev_count, new_count;
+	uint64_t value;
+	int i;
+
+	event = pmu->events[DDRC_PERF_READ_COUNTER_IDX];
+	if (event) {
+		hwc = &event->hw;
+		prev_count = local64_read(&hwc->prev_count);
+		new_count = cn10k_ddr_perf_read_counter(pmu, hwc->idx);
+
+		/* Overflow condition is when new count less than
+		 * previous count
+		 */
+		if (new_count < prev_count)
+			cn10k_ddr_perf_event_update(event);
+	}
+
+	event = pmu->events[DDRC_PERF_WRITE_COUNTER_IDX];
+	if (event) {
+		hwc = &event->hw;
+		prev_count = local64_read(&hwc->prev_count);
+		new_count = cn10k_ddr_perf_read_counter(pmu, hwc->idx);
+
+		/* Overflow condition is when new count less than
+		 * previous count
+		 */
+		if (new_count < prev_count)
+			cn10k_ddr_perf_event_update(event);
+	}
+
+	for (i = 0; i < DDRC_PERF_NUM_GEN_COUNTERS; i++) {
+		if (pmu->events[i] == NULL)
+			continue;
+
+		value = cn10k_ddr_perf_read_counter(pmu, i);
+		if (value == DDRC_PERF_CNT_MAX_VALUE) {
+			pr_info("Counter-(%d) reached max value\n", i);
+			cn10k_ddr_perf_event_update_all(pmu);
+			cn10k_ddr_perf_pmu_disable(&pmu->pmu);
+			cn10k_ddr_perf_pmu_enable(&pmu->pmu);
+		}
+	}
+
+	return IRQ_HANDLED;
+}
+
+static enum hrtimer_restart cn10k_ddr_pmu_timer_handler(struct hrtimer *hrtimer)
+{
+	struct cn10k_ddr_pmu *pmu = container_of(hrtimer, struct cn10k_ddr_pmu,
+						 hrtimer);
+	unsigned long flags;
+
+	local_irq_save(flags);
+	cn10k_ddr_pmu_overflow_handler(pmu);
+	local_irq_restore(flags);
+
+	hrtimer_forward_now(hrtimer, cn10k_ddr_pmu_timer_period());
+	return HRTIMER_RESTART;
+}
+
 static int cn10k_ddr_perf_probe(struct platform_device *pdev)
 {
 	struct cn10k_ddr_pmu *ddr_pmu;
@@ -555,6 +663,9 @@ static int cn10k_ddr_perf_probe(struct platform_device *pdev)
 	if (!name)
 		return -ENOMEM;
 
+	hrtimer_init(&ddr_pmu->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+	ddr_pmu->hrtimer.function = cn10k_ddr_pmu_timer_handler;
+
 	ret = perf_pmu_register(&ddr_pmu->pmu, name, -1);
 	if (ret)
 		return ret;
-- 
2.17.1