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

Russell, Kent Kent.Russell at amd.com
Mon Sep 28 10:54:39 UTC 2020


[AMD Public Use]

Some minor things inline

> -----Original Message-----
> From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> On Behalf Of Ramesh Errabolu
> Sent: Friday, September 25, 2020 6:03 PM
> To: amd-gfx at lists.freedesktop.org
> Cc: Errabolu, Ramesh <Ramesh.Errabolu at amd.com>
> Subject: [PATCH 3/3] drm/amd/amdkfd: Surface files in Sysfs to allow users to get number
> of compute units that are in use.
> 
> [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    | 25 +++++++++
>  drivers/gpu/drm/amd/amdkfd/kfd_process.c | 68 +++++++++++++++++++++++-
>  2 files changed, 92 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> index 56f92cfff591..3df2b9936458 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
> @@ -705,6 +705,31 @@ struct kfd_process_device {
> 
>  	struct kobject *kobj_stats;
>  	unsigned int doorbell_index;
> +
> +	/*
> +	 * @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.

is a snapshot

> +	 *
> +	 * 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

when they do use scratch memory

> +	 *  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;
>  };
> 
>  #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 17d909c86f50..26b716b5eb23 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
> @@ -249,6 +249,52 @@ 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 kfd_process_device *pdd = NULL;
> +
> +	pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
> +	dev = pdd->dev;
> +	if (dev->kfd2kgd->get_cu_occupancy == NULL)
> +		return -EINVAL;
> +
> +	cu_cnt = 0;
> +	proc = pdd->process;
> +	if (pdd->qpd.queue_count == 0) {
> +		pr_debug("Gpu-Id: %d has no active queues for process %d\n",
> +			 dev->id, proc->pasid);
> +		return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
> +	}
> +
> +	/* Collect wave count from device if it supports */
> +	wave_cnt = 0;
> +	max_waves_per_cu = 0;
> +	dev->kfd2kgd->get_cu_occupancy(dev->kgd, proc->pasid, &wave_cnt,
> +			&max_waves_per_cu);
> +
> +	/* Translate wave count to number of compute units */
> +	cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / 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)
>  {
> @@ -344,6 +390,7 @@ static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
> 
>  	return 0;
>  }
> +
>  static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
>  				     struct attribute *attr, char *buffer)
>  {
> @@ -359,6 +406,10 @@ static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
>  				PAGE_SIZE,
>  				"%llu\n",
>  				jiffies64_to_msecs(evict_jiffies));
> +
> +	/* Sysfs handle that gets CU occupancy is per device */
> +	} else if (strcmp(attr->name, "cu_occupancy") == 0) {
> +		return kfd_get_cu_occupancy(attr, buffer);
>  	} else
You should use } else { here, even for a single line, see https://www.kernel.org/doc/html/v4.10/process/coding-style.html right above section 3.1. Annoyingly, checkpatch doesn't catch this.
>  		pr_err("Invalid attribute");
> 
> @@ -466,6 +517,7 @@ static int kfd_procfs_add_sysfs_stats(struct kfd_process *p)
>  	 * Create sysfs files for each GPU:
>  	 * - proc/<pid>/stats_<gpuid>/
>  	 * - proc/<pid>/stats_<gpuid>/evicted_ms
> +	 * - proc/<pid>/stats_<gpuid>/cu_occupancy
>  	 */
>  	list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
>  		struct kobject *kobj_stats;
> @@ -496,6 +548,19 @@ static int kfd_procfs_add_sysfs_stats(struct kfd_process *p)
>  		if (ret)
>  			pr_warn("Creating eviction stats for gpuid %d failed",
>  					(int)pdd->dev->id);
> +
> +		/* Add sysfs file to report compute unit occupancy */
> +		if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL) {
> +			pdd->attr_cu_occupancy.name = "cu_occupancy";
> +			pdd->attr_cu_occupancy.mode = KFD_SYSFS_FILE_MODE;
> +			sysfs_attr_init(&pdd->attr_cu_occupancy);
> +			ret = sysfs_create_file(kobj_stats,
> +						&pdd->attr_cu_occupancy);
> +			if (ret)
> +				pr_warn("Creating %s failed for gpuid: %d",
> +					pdd->attr_cu_occupancy.name,
> +					(int)pdd->dev->id);
> +		}
>  	}
>  err:
>  	return ret;
> @@ -537,7 +602,6 @@ static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
>  	return ret;
>  }
> 
> -
>  void kfd_procfs_del_queue(struct queue *q)
>  {
>  	if (!q)
> @@ -909,6 +973,8 @@ static void kfd_process_wq_release(struct work_struct *work)
>  			sysfs_remove_file(p->kobj, &pdd->attr_vram);
>  			sysfs_remove_file(p->kobj, &pdd->attr_sdma);
>  			sysfs_remove_file(p->kobj, &pdd->attr_evict);
> +			if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL)
> +				sysfs_remove_file(p->kobj, &pdd->attr_cu_occupancy);
>  			kobject_del(pdd->kobj_stats);
>  			kobject_put(pdd->kobj_stats);
>  			pdd->kobj_stats = NULL;
> --
> 2.27.0
> 
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.or
> g%2Fmailman%2Flistinfo%2Famd-
> gfx&data=02%7C01%7Ckent.russell%40amd.com%7C4ed773ec69b04a9d2ee308d8619
> ee024%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637366682305777071&amp
> ;sdata=1NPpCMCB8dfWGwIokNotPm1pJU4AMkeurhcIe0y%2FIaw%3D&reserved=0


More information about the amd-gfx mailing list