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>

6 files changed, 188 insertions, 8 deletions

diff --git a/drivers/misc/habanalabs/command_submission.c b/drivers/misc/habanalabs/command_submission.c
index e4dd3e83df8b..4777ec4c2b55 100644
--- a/drivers/misc/habanalabs/command_submission.c
+++ b/drivers/misc/habanalabs/command_submission.c
@@ -178,11 +178,23 @@ static void cs_do_release(struct kref *ref)
 
 	/* We also need to update CI for internal queues */
 	if (cs->submitted) {
-		int cs_cnt = atomic_dec_return(&hdev->cs_active_cnt);
+		hdev->asic_funcs->hw_queues_lock(hdev);
 
-		WARN_ONCE((cs_cnt < 0),
-			"hl%d: error in CS active cnt %d\n",
-			hdev->id, cs_cnt);
+		hdev->cs_active_cnt--;
+		if (!hdev->cs_active_cnt) {
+			struct hl_device_idle_busy_ts *ts;
+
+			ts = &hdev->idle_busy_ts_arr[hdev->idle_busy_ts_idx++];
+			ts->busy_to_idle_ts = ktime_get();
+
+			if (hdev->idle_busy_ts_idx == HL_IDLE_BUSY_TS_ARR_SIZE)
+				hdev->idle_busy_ts_idx = 0;
+		} else if (hdev->cs_active_cnt < 0) {
+			dev_crit(hdev->dev, "CS active cnt %d is negative\n",
+				hdev->cs_active_cnt);
+		}
+
+		hdev->asic_funcs->hw_queues_unlock(hdev);
 
 		hl_int_hw_queue_update_ci(cs);
 
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 */
diff --git a/drivers/misc/habanalabs/habanalabs.h b/drivers/misc/habanalabs/habanalabs.h
index a4d929f5bad8..23b86b7f9732 100644
--- a/drivers/misc/habanalabs/habanalabs.h
+++ b/drivers/misc/habanalabs/habanalabs.h
@@ -45,6 +45,8 @@
 /* MUST BE POWER OF 2 and larger than 1 */
 #define HL_MAX_PENDING_CS		64
 
+#define HL_IDLE_BUSY_TS_ARR_SIZE	4096
+
 /* Memory */
 #define MEM_HASH_TABLE_BITS		7 /* 1 << 7 buckets */
 
@@ -1157,6 +1159,16 @@ struct hl_device_reset_work {
 };
 
 /**
+ * struct hl_device_idle_busy_ts - used for calculating device utilization rate.
+ * @idle_to_busy_ts: timestamp where device changed from idle to busy.
+ * @busy_to_idle_ts: timestamp where device changed from busy to idle.
+ */
+struct hl_device_idle_busy_ts {
+	ktime_t				idle_to_busy_ts;
+	ktime_t				busy_to_idle_ts;
+};
+
+/**
  * struct hl_device - habanalabs device structure.
  * @pdev: pointer to PCI device, can be NULL in case of simulator device.
  * @pcie_bar: array of available PCIe bars.
@@ -1203,19 +1215,22 @@ struct hl_device_reset_work {
  *              when a user opens the device
  * @fpriv_list_lock: protects the fpriv_list
  * @compute_ctx: current compute context executing.
+ * @idle_busy_ts_arr: array to hold time stamps of transitions from idle to busy
+ *                    and vice-versa
  * @dram_used_mem: current DRAM memory consumption.
  * @timeout_jiffies: device CS timeout value.
  * @max_power: the max power of the device, as configured by the sysadmin. This
  *             value is saved so in case of hard-reset, KMD will restore this
  *             value and update the F/W after the re-initialization
  * @in_reset: is device in reset flow.
+ * @curr_pll_profile: current PLL profile.
  * @cs_active_cnt: number of active command submissions on this device (active
  *                 means already in H/W queues)
- * @curr_pll_profile: current PLL profile.
  * @major: habanalabs KMD major.
  * @high_pll: high PLL profile frequency.
  * @soft_reset_cnt: number of soft reset since KMD loading.
  * @hard_reset_cnt: number of hard reset since KMD loading.
+ * @idle_busy_ts_idx: index of current entry in idle_busy_ts_arr
  * @id: device minor.
  * @id_control: minor of the control device
  * @disabled: is device disabled.
@@ -1285,16 +1300,19 @@ struct hl_device {
 
 	struct hl_ctx			*compute_ctx;
 
+	struct hl_device_idle_busy_ts	*idle_busy_ts_arr;
+
 	atomic64_t			dram_used_mem;
 	u64				timeout_jiffies;
 	u64				max_power;
 	atomic_t			in_reset;
-	atomic_t			cs_active_cnt;
 	enum hl_pll_frequency		curr_pll_profile;
+	int				cs_active_cnt;
 	u32				major;
 	u32				high_pll;
 	u32				soft_reset_cnt;
 	u32				hard_reset_cnt;
+	u32				idle_busy_ts_idx;
 	u16				id;
 	u16				id_control;
 	u8				disabled;
@@ -1457,6 +1475,7 @@ int hl_device_reset(struct hl_device *hdev, bool hard_reset,
 void hl_hpriv_get(struct hl_fpriv *hpriv);
 void hl_hpriv_put(struct hl_fpriv *hpriv);
 int hl_device_set_frequency(struct hl_device *hdev, enum hl_pll_frequency freq);
+uint32_t hl_device_utilization(struct hl_device *hdev, uint32_t period_ms);
 
 int hl_build_hwmon_channel_info(struct hl_device *hdev,
 		struct armcp_sensor *sensors_arr);
diff --git a/drivers/misc/habanalabs/habanalabs_ioctl.c b/drivers/misc/habanalabs/habanalabs_ioctl.c
index 8fbca2c0f44b..f958568f7996 100644
--- a/drivers/misc/habanalabs/habanalabs_ioctl.c
+++ b/drivers/misc/habanalabs/habanalabs_ioctl.c
@@ -197,6 +197,29 @@ out:
 	return rc;
 }
 
+static int device_utilization(struct hl_device *hdev, struct hl_info_args *args)
+{
+	struct hl_info_device_utilization device_util = {0};
+	u32 max_size = args->return_size;
+	void __user *out = (void __user *) (uintptr_t) args->return_pointer;
+
+	if ((!max_size) || (!out))
+		return -EINVAL;
+
+	if ((args->period_ms < 100) || (args->period_ms > 1000) ||
+		(args->period_ms % 100)) {
+		dev_err(hdev->dev,
+			"period %u must be between 100 - 1000 and must be divisible by 100\n",
+			args->period_ms);
+		return -EINVAL;
+	}
+
+	device_util.utilization = hl_device_utilization(hdev, args->period_ms);
+
+	return copy_to_user(out, &device_util,
+		min((size_t) max_size, sizeof(device_util))) ? -EFAULT : 0;
+}
+
 static int _hl_info_ioctl(struct hl_fpriv *hpriv, void *data,
 				struct device *dev)
 {
@@ -239,6 +262,10 @@ static int _hl_info_ioctl(struct hl_fpriv *hpriv, void *data,
 		rc = hw_idle(hdev, args);
 		break;
 
+	case HL_INFO_DEVICE_UTILIZATION:
+		rc = device_utilization(hdev, args);
+		break;
+
 	default:
 		dev_err(dev, "Invalid request %d\n", args->op);
 		rc = -ENOTTY;
diff --git a/drivers/misc/habanalabs/hw_queue.c b/drivers/misc/habanalabs/hw_queue.c
index 696cf7d206c6..55b383b2a116 100644
--- a/drivers/misc/habanalabs/hw_queue.c
+++ b/drivers/misc/habanalabs/hw_queue.c
@@ -364,7 +364,13 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
 		spin_unlock(&hdev->hw_queues_mirror_lock);
 	}
 
-	atomic_inc(&hdev->cs_active_cnt);
+	if (!hdev->cs_active_cnt++) {
+		struct hl_device_idle_busy_ts *ts;
+
+		ts = &hdev->idle_busy_ts_arr[hdev->idle_busy_ts_idx];
+		ts->busy_to_idle_ts = ktime_set(0, 0);
+		ts->idle_to_busy_ts = ktime_get();
+	}
 
 	list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
 		if (job->ext_queue)
diff --git a/drivers/misc/habanalabs/include/goya/goya.h b/drivers/misc/habanalabs/include/goya/goya.h
index 3f02a52ba4ce..43d241891e45 100644
--- a/drivers/misc/habanalabs/include/goya/goya.h
+++ b/drivers/misc/habanalabs/include/goya/goya.h
@@ -38,4 +38,6 @@
 
 #define TPC_MAX_NUM		8
 
+#define MME_MAX_NUM		1
+
 #endif /* GOYA_H */