habanalabs: add uapi to retrieve device utilization

Users and sysadmins usually want to know what is the device utilization as
a level 0 indication if they are efficiently using the device.

Add a new opcode to the INFO IOCTL that will return the device utilization
over the last period of 100-1000ms. The return value is 0-100,
representing as percentage the total utilization rate.

Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Omer Shpigelman <oshpigelman@habana.ai>

1 files changed, 115 insertions, 1 deletions

diff --git a/drivers/misc/habanalabs/device.c b/drivers/misc/habanalabs/device.c
index 41c3ddbca351..cebdceb72298 100644
--- a/drivers/misc/habanalabs/device.c
+++ b/drivers/misc/habanalabs/device.c
@@ -293,6 +293,14 @@ static int device_early_init(struct hl_device *hdev)
 		goto free_eq_wq;
 	}
 
+	hdev->idle_busy_ts_arr = kmalloc_array(HL_IDLE_BUSY_TS_ARR_SIZE,
+					sizeof(struct hl_device_idle_busy_ts),
+					(GFP_KERNEL | __GFP_ZERO));
+	if (!hdev->idle_busy_ts_arr) {
+		rc = -ENOMEM;
+		goto free_chip_info;
+	}
+
 	hl_cb_mgr_init(&hdev->kernel_cb_mgr);
 
 	mutex_init(&hdev->send_cpu_message_lock);
@@ -303,10 +311,11 @@ static int device_early_init(struct hl_device *hdev)
 	INIT_LIST_HEAD(&hdev->fpriv_list);
 	mutex_init(&hdev->fpriv_list_lock);
 	atomic_set(&hdev->in_reset, 0);
-	atomic_set(&hdev->cs_active_cnt, 0);
 
 	return 0;
 
+free_chip_info:
+	kfree(hdev->hl_chip_info);
 free_eq_wq:
 	destroy_workqueue(hdev->eq_wq);
 free_cq_wq:
@@ -336,6 +345,7 @@ static void device_early_fini(struct hl_device *hdev)
 
 	hl_cb_mgr_fini(hdev, &hdev->kernel_cb_mgr);
 
+	kfree(hdev->idle_busy_ts_arr);
 	kfree(hdev->hl_chip_info);
 
 	destroy_workqueue(hdev->eq_wq);
@@ -451,6 +461,102 @@ static void device_late_fini(struct hl_device *hdev)
 	hdev->late_init_done = false;
 }
 
+uint32_t hl_device_utilization(struct hl_device *hdev, uint32_t period_ms)
+{
+	struct hl_device_idle_busy_ts *ts;
+	ktime_t zero_ktime, curr = ktime_get();
+	u32 overlap_cnt = 0, last_index = hdev->idle_busy_ts_idx;
+	s64 period_us, last_start_us, last_end_us, last_busy_time_us,
+		total_busy_time_us = 0, total_busy_time_ms;
+
+	zero_ktime = ktime_set(0, 0);
+	period_us = period_ms * USEC_PER_MSEC;
+	ts = &hdev->idle_busy_ts_arr[last_index];
+
+	/* check case that device is currently in idle */
+	if (!ktime_compare(ts->busy_to_idle_ts, zero_ktime) &&
+			!ktime_compare(ts->idle_to_busy_ts, zero_ktime)) {
+
+		last_index--;
+		/* Handle case idle_busy_ts_idx was 0 */
+		if (last_index > HL_IDLE_BUSY_TS_ARR_SIZE)
+			last_index = HL_IDLE_BUSY_TS_ARR_SIZE - 1;
+
+		ts = &hdev->idle_busy_ts_arr[last_index];
+	}
+
+	while (overlap_cnt < HL_IDLE_BUSY_TS_ARR_SIZE) {
+		/* Check if we are in last sample case. i.e. if the sample
+		 * begun before the sampling period. This could be a real
+		 * sample or 0 so need to handle both cases
+		 */
+		last_start_us = ktime_to_us(
+				ktime_sub(curr, ts->idle_to_busy_ts));
+
+		if (last_start_us > period_us) {
+
+			/* First check two cases:
+			 * 1. If the device is currently busy
+			 * 2. If the device was idle during the whole sampling
+			 *    period
+			 */
+
+			if (!ktime_compare(ts->busy_to_idle_ts, zero_ktime)) {
+				/* Check if the device is currently busy */
+				if (ktime_compare(ts->idle_to_busy_ts,
+						zero_ktime))
+					return 100;
+
+				/* We either didn't have any activity or we
+				 * reached an entry which is 0. Either way,
+				 * exit and return what was accumulated so far
+				 */
+				break;
+			}
+
+			/* If sample has finished, check it is relevant */
+			last_end_us = ktime_to_us(
+					ktime_sub(curr, ts->busy_to_idle_ts));
+
+			if (last_end_us > period_us)
+				break;
+
+			/* It is relevant so add it but with adjustment */
+			last_busy_time_us = ktime_to_us(
+						ktime_sub(ts->busy_to_idle_ts,
+						ts->idle_to_busy_ts));
+			total_busy_time_us += last_busy_time_us -
+					(last_start_us - period_us);
+			break;
+		}
+
+		/* Check if the sample is finished or still open */
+		if (ktime_compare(ts->busy_to_idle_ts, zero_ktime))
+			last_busy_time_us = ktime_to_us(
+						ktime_sub(ts->busy_to_idle_ts,
+						ts->idle_to_busy_ts));
+		else
+			last_busy_time_us = ktime_to_us(
+					ktime_sub(curr, ts->idle_to_busy_ts));
+
+		total_busy_time_us += last_busy_time_us;
+
+		last_index--;
+		/* Handle case idle_busy_ts_idx was 0 */
+		if (last_index > HL_IDLE_BUSY_TS_ARR_SIZE)
+			last_index = HL_IDLE_BUSY_TS_ARR_SIZE - 1;
+
+		ts = &hdev->idle_busy_ts_arr[last_index];
+
+		overlap_cnt++;
+	}
+
+	total_busy_time_ms = DIV_ROUND_UP_ULL(total_busy_time_us,
+						USEC_PER_MSEC);
+
+	return DIV_ROUND_UP_ULL(total_busy_time_ms * 100, period_ms);
+}
+
 /*
  * hl_device_set_frequency - set the frequency of the device
  *
@@ -808,6 +914,14 @@ again:
 	for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
 		hl_cq_reset(hdev, &hdev->completion_queue[i]);
 
+	hdev->idle_busy_ts_idx = 0;
+	hdev->idle_busy_ts_arr[0].busy_to_idle_ts = ktime_set(0, 0);
+	hdev->idle_busy_ts_arr[0].idle_to_busy_ts = ktime_set(0, 0);
+
+	if (hdev->cs_active_cnt)
+		dev_crit(hdev->dev, "CS active cnt %d is not 0 during reset\n",
+			hdev->cs_active_cnt);
+
 	mutex_lock(&hdev->fpriv_list_lock);
 
 	/* Make sure the context switch phase will run again */