[PATCH 3/3] drm/amd/amdkfd: Surface files in Sysfs to allow users to get number of compute units that are in use.

Ramesh Errabolu Ramesh.Errabolu at amd.com
Thu Sep 17 17:11:18 UTC 2020


[Why]
Allow user to know how many compute units (CU) are in use at any given
moment.

[How]
Surface files in Sysfs that allow user to determine the number of compute
units that are in use for a given process. One Sysfs file is used per
device.

Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu at amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h    |  28 ++++++-
 drivers/gpu/drm/amd/amdkfd/kfd_process.c | 101 +++++++++++++++++++----
 2 files changed, 114 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 023629f28495..2ce03586d0dc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -631,7 +631,7 @@ enum kfd_pdd_bound {
 	PDD_BOUND_SUSPENDED,
 };
 
-#define MAX_SYSFS_FILENAME_LEN 11
+#define MAX_SYSFS_FILENAME_LEN 32
 
 /*
  * SDMA counter runs at 100MHz frequency.
@@ -692,6 +692,32 @@ struct kfd_process_device {
 	uint64_t sdma_past_activity_counter;
 	struct attribute attr_sdma;
 	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
+
+	/*
+	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
+	 * that is associated with device encoded by "this" struct instance. The
+	 * value reflects CU usage by all of the waves launched by this process
+	 * on this device. A very important property of occupancy parameter is
+	 * that its value is a a snapshot of current use.
+	 *
+	 * Following is to be noted regarding how this parameter is reported:
+	 *
+	 *  The number of waves that a CU can launch is limited by couple of
+	 *  parameters. These are encoded by struct amdgpu_cu_info instance
+	 *  that is part of every device definition. For GFX9 devices this
+	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
+	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
+	 *  when they use. This could change for future devices and therefore
+	 *  this example should be considered as a guide.
+	 *
+	 *  All CU's of a device are available for the process. This may not be true
+	 *  under certain conditions - e.g. CU masking.
+	 *
+	 *  Finally number of CU's that are occupied by a process is affected by both
+	 *  number of CU's a device has along with number of other competing processes
+	 */
+	struct attribute attr_cu_occupancy;
+	char cu_occupancy_filename[MAX_SYSFS_FILENAME_LEN];
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index a0e12a79ab7d..3baffbc828b2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -249,6 +249,63 @@ static void kfd_sdma_activity_worker(struct work_struct *work)
 	}
 }
 
+/**
+ * @kfd_get_cu_occupancy() - Collect number of waves in-flight on this device
+ * by current process. Translates acquired wave count into number of compute units
+ * that are occupied.
+ *
+ * @atr: Handle of attribute that allows reporting of wave count. The attribute
+ * handle encapsulates GPU device it is associated with, thereby allowing collection
+ * of waves in flight, etc
+ *
+ * @buffer: Handle of user provided buffer updated with wave count
+ *
+ * Return: Number of bytes written to user buffer or an error value
+ */
+static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
+{
+	int cu_cnt;
+	int wave_cnt;
+	int max_waves_per_cu;
+	struct kfd_dev *dev = NULL;
+	struct kfd_process *proc = NULL;
+	struct amdgpu_device *adev = NULL;
+	struct kfd_process_device *pdd = NULL;
+
+	/* Acquire handle of Process-Device-Data associated with attribute */
+	pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
+
+	/*
+	 * Acquire handle of Gpu Device associated with attribute. Determine
+	 * if ability to collect CU occupancy is defined for this device. If
+	 * so acquire the handle of process which encapsulates the PASID of
+	 * process. It is possible that the process might have zero work on
+	 * device. This is determined by checking if process has any queues
+	 * @note: Ability to collect is defined only for Gfx9 devices
+	 */
+	cu_cnt = 0;
+	dev = pdd->dev;
+	proc = pdd->process;
+	if (pdd->qpd.queue_count == 0) {
+		pr_info("%s: Gpu-Id: %d has no active queues for process %d\n",
+			__func__, dev->id, proc->pasid);
+		return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
+	}
+
+	/* Collect wave count from device by reading relevant registers */
+	wave_cnt = 0;
+	if (dev->kfd2kgd->get_cu_occupancy != NULL)
+		dev->kfd2kgd->get_cu_occupancy(dev->kgd, proc->pasid,
+					       &wave_cnt);
+
+	/* Translate wave count to number of compute units */
+	adev = (struct amdgpu_device *)dev->kgd;
+	max_waves_per_cu = adev->gfx.cu_info.simd_per_cu *
+		adev->gfx.cu_info.max_waves_per_simd;
+	cu_cnt = wave_cnt / max_waves_per_cu;
+	return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
+}
+
 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
 			       char *buffer)
 {
@@ -279,6 +336,10 @@ static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
 		return snprintf(buffer, PAGE_SIZE, "%llu\n",
 				(sdma_activity_work_handler.sdma_activity_counter)/
 				 SDMA_ACTIVITY_DIVISOR);
+
+	/* Sysfs handle that gets CU occupancy is per device */
+	} else if (strncmp(attr->name, "cu_occupancy_", 13) == 0) {
+		return kfd_get_cu_occupancy(attr, buffer);
 	} else {
 		pr_err("Invalid attribute");
 		return -EINVAL;
@@ -432,27 +493,37 @@ static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
 	 * Create sysfs files for each GPU:
 	 * - proc/<pid>/vram_<gpuid>
 	 * - proc/<pid>/sdma_<gpuid>
+	 * - proc/<pid>/cu_occupancy_<gpuid>
 	 */
 	list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
 		snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
 			 pdd->dev->id);
-		ret = kfd_sysfs_create_file(p, &pdd->attr_vram, pdd->vram_filename);
+		ret = kfd_sysfs_create_file(p, &pdd->attr_vram,
+					    pdd->vram_filename);
 		if (ret)
 			pr_warn("Creating vram usage for gpu id %d failed",
 				(int)pdd->dev->id);
-
-		snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
-			 pdd->dev->id);
-		ret = kfd_sysfs_create_file(p, &pdd->attr_sdma, pdd->sdma_filename);
-		if (ret)
-			pr_warn("Creating sdma usage for gpu id %d failed",
-				(int)pdd->dev->id);
-	}
-
+		if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL) {
+			snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN,
+				 "sdma_%u", pdd->dev->id);
+			ret = kfd_sysfs_create_file(p, &pdd->attr_sdma,
+						    pdd->sdma_filename);
+			if (ret)
+				pr_warn("Creating sdma usage for gpu id %d failed",
+					(int)pdd->dev->id);
+			snprintf(pdd->cu_occupancy_filename,
+				 MAX_SYSFS_FILENAME_LEN, "cu_occupancy_%u",
+				 pdd->dev->id);
+			ret = kfd_sysfs_create_file(p, &pdd->attr_cu_occupancy,
+						    pdd->cu_occupancy_filename);
+			if (ret)
+				pr_warn("Creating CU occupancy file for gpu id %d failed",
+					(int)pdd->dev->id);
+			}
+		}
 	return ret;
 }
 
-
 void kfd_procfs_del_queue(struct queue *q)
 {
 	if (!q)
@@ -767,8 +838,7 @@ static void kfd_process_destroy_pdds(struct kfd_process *p)
 			amdgpu_amdkfd_gpuvm_release_process_vm(
 					pdd->dev->kgd, pdd->vm);
 			fput(pdd->drm_file);
-		}
-		else if (pdd->vm)
+		} else if (pdd->vm)
 			amdgpu_amdkfd_gpuvm_destroy_process_vm(
 				pdd->dev->kgd, pdd->vm);
 
@@ -815,7 +885,10 @@ static void kfd_process_wq_release(struct work_struct *work)
 
 		list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
 			sysfs_remove_file(p->kobj, &pdd->attr_vram);
-			sysfs_remove_file(p->kobj, &pdd->attr_sdma);
+			if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL) {
+				sysfs_remove_file(p->kobj, &pdd->attr_sdma);
+				sysfs_remove_file(p->kobj, &pdd->attr_cu_occupancy);
+			}
 		}
 
 		kobject_del(p->kobj);
-- 
2.27.0



More information about the amd-gfx mailing list