[PATCH] drm/amdgpu: enable separate timeout setting for every ring type

Koenig, Christian Christian.Koenig at amd.com
Mon Apr 29 09:33:59 UTC 2019


Am 29.04.19 um 10:57 schrieb Evan Quan:
> Every ring type can have its own timeout setting.
>
> Change-Id: I992f224f36bb33acd560162bffd2c3e987840a7e
> Signed-off-by: Evan Quan <evan.quan at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  7 +-
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 17 +++--
>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    | 82 ++++++++++++++++++++--
>   drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c  | 35 +++++++--
>   drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c      |  2 +-
>   5 files changed, 124 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index f6965b9403eb..9d7bf9fbf2ba 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -118,7 +118,6 @@ extern int amdgpu_disp_priority;
>   extern int amdgpu_hw_i2c;
>   extern int amdgpu_pcie_gen2;
>   extern int amdgpu_msi;
> -extern int amdgpu_lockup_timeout;
>   extern int amdgpu_dpm;
>   extern int amdgpu_fw_load_type;
>   extern int amdgpu_aspm;
> @@ -428,6 +427,7 @@ struct amdgpu_fpriv {
>   };
>   
>   int amdgpu_file_to_fpriv(struct file *filp, struct amdgpu_fpriv **fpriv);
> +int amdgpu_device_get_job_timeout(struct amdgpu_device *adev);
>   
>   int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>   		  unsigned size, struct amdgpu_ib *ib);
> @@ -1001,6 +1001,11 @@ struct amdgpu_device {
>   	struct work_struct		xgmi_reset_work;
>   
>   	bool                            in_baco_reset;
> +
> +	long				gfx_timeout;
> +	long				sdma_timeout;
> +	long				video_timeout;
> +	long				compute_timeout;
>   };
>   
>   static inline struct amdgpu_device *amdgpu_ttm_adev(struct ttm_bo_device *bdev)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 80bf604019b1..bb111c05949d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -912,8 +912,10 @@ static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
>    * Validates certain module parameters and updates
>    * the associated values used by the driver (all asics).
>    */
> -static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
> +static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
>   {
> +	int ret = 0;
> +
>   	if (amdgpu_sched_jobs < 4) {
>   		dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
>   			 amdgpu_sched_jobs);
> @@ -958,13 +960,16 @@ static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
>   		amdgpu_vram_page_split = 1024;
>   	}
>   
> -	if (amdgpu_lockup_timeout == 0) {
> -		dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
> -		amdgpu_lockup_timeout = 10000;
> +	ret = amdgpu_device_get_job_timeout(adev);
> +	if (ret) {
> +		dev_err(adev->dev, "invalid job timeout setting\n");
> +		return ret;
>   	}
>   
>   	adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
>   	amdgpu_direct_gma_size = min(amdgpu_direct_gma_size, 96);
> +
> +	return ret;
>   }
>   
>   /**
> @@ -2468,7 +2473,9 @@ int amdgpu_device_init(struct amdgpu_device *adev,
>   	mutex_init(&adev->lock_reset);
>   	mutex_init(&adev->virt.dpm_mutex);
>   
> -	amdgpu_device_check_arguments(adev);
> +	ret = amdgpu_device_check_arguments(adev);
> +	if (ret)
> +		return ret;
>   
>   	spin_lock_init(&adev->mmio_idx_lock);
>   	spin_lock_init(&adev->smc_idx_lock);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index 71df27cd03de..4694e36d1bf1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -83,6 +83,8 @@
>   
>   #define AMDGPU_VERSION		"19.10.9.418"
>   
> +#define MAX_PARAM_LENGTH	256

This needs a better name, especially but an AMDGPU_ prefix in front of it.

> +
>   int amdgpu_vram_limit = 0;
>   int amdgpu_vis_vram_limit = 0;
>   int amdgpu_gart_size = -1; /* auto */
> @@ -95,7 +97,7 @@ int amdgpu_disp_priority = 0;
>   int amdgpu_hw_i2c = 0;
>   int amdgpu_pcie_gen2 = -1;
>   int amdgpu_msi = -1;
> -int amdgpu_lockup_timeout = 10000;
> +char amdgpu_lockup_timeout[MAX_PARAM_LENGTH];
>   int amdgpu_dpm = -1;
>   int amdgpu_fw_load_type = -1;
>   int amdgpu_aspm = -1;
> @@ -232,12 +234,20 @@ MODULE_PARM_DESC(msi, "MSI support (1 = enable, 0 = disable, -1 = auto)");
>   module_param_named(msi, amdgpu_msi, int, 0444);
>   
>   /**
> - * DOC: lockup_timeout (int)
> - * Set GPU scheduler timeout value in ms. Value 0 is invalidated, will be adjusted to 10000.
> - * Negative values mean 'infinite timeout' (MAX_JIFFY_OFFSET). The default is 10000.
> + * DOC: lockup_timeout (string)
> + * Set GPU scheduler timeout value in ms.
> + *
> + * The format is non-compute[.gfx.sdma.video][:compute].
> + * With no "non-compute[.gfx.sdma.video]" timeout specified, the default timeout for non-compute_job is 10000.
> + * The "non-compute" timeout setting applys to all non compute IPs(gfx, sdma and video). Unless
> + * the timeout for this IP is specified separately(by "[.gfx.sdma.video]").
> + *
> + * With no "[:compute]" timeout specified, there will be no timeout enforced on compute jobs.
> + * Otherwise, the timeout for compute jobs is set as specified.
>    */
> -MODULE_PARM_DESC(lockup_timeout, "GPU lockup timeout in ms > 0 (default 10000)");
> -module_param_named(lockup_timeout, amdgpu_lockup_timeout, int, 0444);
> +MODULE_PARM_DESC(lockup_timeout, "GPU lockup timeout in ms (default: 10000 for non-compute jobs and no timeout for compute jobs), "
> +		"format is non-compute[.gfx.sdma.video][:compute]");
> +module_param_string(lockup_timeout, amdgpu_lockup_timeout, sizeof(amdgpu_lockup_timeout), 0444);
>   
>   /**
>    * DOC: dpm (int)
> @@ -1307,6 +1317,66 @@ int amdgpu_file_to_fpriv(struct file *filp, struct amdgpu_fpriv **fpriv)
>   	return 0;
>   }
>   
> +int amdgpu_device_get_job_timeout(struct amdgpu_device *adev)
> +{
> +	char *input = amdgpu_lockup_timeout;
> +	char *non_compute_setting = NULL;
> +	char *timeout_setting = NULL;
> +	int index = 0;
> +	int ret = 0;
> +
> +	/* Default timeout for non compute job is 10000 */
> +	adev->gfx_timeout =
> +		adev->sdma_timeout =
> +			adev->video_timeout = 10000;
> +	/* Enforce no timeout on compute job at default */
> +	adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
> +
> +	if (strnlen(input, MAX_PARAM_LENGTH)) {
> +		non_compute_setting = strsep(&input, ":");
> +		if (input) {
> +			ret = kstrtol(input, 0, &(adev->compute_timeout));
> +			if (ret)
> +				return ret;
> +		}
> +
> +		while ((timeout_setting = strsep(&non_compute_setting, ".")) &&
> +				strnlen(timeout_setting, MAX_PARAM_LENGTH)) {
> +			switch (index) {
> +			case 0:
> +				ret = kstrtol(timeout_setting, 0, &adev->gfx_timeout);

Maybe move the kstrtol call outside of the switch.

That would de-duplicate error handling quite a bit.

Christian.

> +				if (ret)
> +					return ret;
> +
> +				adev->sdma_timeout =
> +					adev->video_timeout =
> +						adev->gfx_timeout;
> +				break;
> +			case 1:
> +				ret = kstrtol(timeout_setting, 0, &adev->gfx_timeout);
> +				if (ret)
> +					return ret;
> +				break;
> +			case 2:
> +				ret = kstrtol(timeout_setting, 0, &adev->sdma_timeout);
> +				if (ret)
> +					return ret;
> +				break;
> +			case 3:
> +				ret = kstrtol(timeout_setting, 0, &adev->video_timeout);
> +				if (ret)
> +					return ret;
> +				break;
> +			default:
> +				break;
> +			}
> +			index++;
> +		}
> +	}
> +
> +	return ret;
> +}
> +
>   static bool
>   amdgpu_get_crtc_scanout_position(struct drm_device *dev, unsigned int pipe,
>   				 bool in_vblank_irq, int *vpos, int *hpos,
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> index d19ad34bef75..bb40e8946bc2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> @@ -436,9 +436,13 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
>   int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
>   				  unsigned num_hw_submission)
>   {
> +	struct amdgpu_device *adev = ring->adev;
>   	long timeout;
>   	int r;
>   
> +	if (!adev)
> +		return -EINVAL;
> +
>   	/* Check that num_hw_submission is a power of two */
>   	if ((num_hw_submission & (num_hw_submission - 1)) != 0)
>   		return -EINVAL;
> @@ -465,12 +469,31 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
>   
>   	/* No need to setup the GPU scheduler for KIQ ring */
>   	if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
> -		/* for non-sriov case, no timeout enforce on compute ring */
> -		if ((ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
> -				&& !amdgpu_sriov_vf(ring->adev))
> -			timeout = MAX_SCHEDULE_TIMEOUT;
> -		else
> -			timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
> +		switch (ring->funcs->type) {
> +		case AMDGPU_RING_TYPE_GFX:
> +			timeout = adev->gfx_timeout;
> +			break;
> +		case AMDGPU_RING_TYPE_COMPUTE:
> +			/*
> +			 * For non-sriov case, no timeout enforce
> +			 * on compute ring at default. Unless user
> +			 * specifies a timeout for compute ring.
> +			 *
> +			 * For sriov case, always use the timeout
> +			 * as gfx ring
> +			 */
> +			if (!amdgpu_sriov_vf(ring->adev))
> +				timeout = adev->compute_timeout;
> +			else
> +				timeout = adev->gfx_timeout;
> +			break;
> +		case AMDGPU_RING_TYPE_SDMA:
> +			timeout = adev->sdma_timeout;
> +			break;
> +		default:
> +			timeout = adev->video_timeout;
> +			break;
> +		}
>   
>   		r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
>   				   num_hw_submission, amdgpu_job_hang_limit,
> diff --git a/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c b/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
> index 8dbad496b29f..089952a1e6b0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
> +++ b/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
> @@ -343,7 +343,7 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
>   
>   	/* Trigger recovery for world switch failure if no TDR */
>   	if (amdgpu_device_should_recover_gpu(adev)
> -		&& amdgpu_lockup_timeout == MAX_SCHEDULE_TIMEOUT)
> +		&& adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT)
>   		amdgpu_device_gpu_recover(adev, NULL);
>   }
>   



More information about the amd-gfx mailing list