[PATCH] drm/v3d: clean caches at the end of render jobs on request from user space

[Eric Anholt]

Iago Toral Quiroga <itoral at igalia.com> writes:

> Extends the user space ioctl for CL submissions so it can include a request
> to flush the cache once the CL execution has completed. Fixes memory
> write violation messages reported by the kernel in workloads involving
> shader memory writes (SSBOs, shader images, scratch, etc) which sometimes
> also lead to GPU resets during Piglit and CTS workloads.

Some context for any other reviewers: This patch is the interface change
necessary to expose GLES 3.1 on V3D.  It turns out the HW packets for
flushing the caches were broken in multiple ways.

> Signed-off-by: Iago Toral Quiroga <itoral at igalia.com>
> ---
>  drivers/gpu/drm/v3d/v3d_gem.c | 51 +++++++++++++++++++++++++++++------
>  include/uapi/drm/v3d_drm.h    |  7 ++---
>  2 files changed, 47 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/gpu/drm/v3d/v3d_gem.c b/drivers/gpu/drm/v3d/v3d_gem.c
> index 5d80507b539b..530fe9d9d5bd 100644
> --- a/drivers/gpu/drm/v3d/v3d_gem.c
> +++ b/drivers/gpu/drm/v3d/v3d_gem.c
> @@ -530,13 +530,16 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
>  	struct drm_v3d_submit_cl *args = data;
>  	struct v3d_bin_job *bin = NULL;
>  	struct v3d_render_job *render;
> +	struct v3d_job *clean_job = NULL;
> +	struct v3d_job *last_job;
>  	struct ww_acquire_ctx acquire_ctx;
>  	int ret = 0;
>  
>  	trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
>  
> -	if (args->pad != 0) {
> -		DRM_INFO("pad must be zero: %d\n", args->pad);
> +	if (args->flags != 0 &&
> +	    args->flags != DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG) {
> +		DRM_INFO("invalid flags: %d\n", args->flags);
>  		return -EINVAL;
>  	}
>  
> @@ -575,12 +578,28 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
>  		bin->render = render;
>  	}
>  
> -	ret = v3d_lookup_bos(dev, file_priv, &render->base,
> +	if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG) {
> +		clean_job = kcalloc(1, sizeof(*clean_job), GFP_KERNEL);
> +		if (!clean_job) {
> +			ret = -ENOMEM;
> +			goto fail;
> +		}
> +
> +		ret = v3d_job_init(v3d, file_priv, clean_job, v3d_job_free, 0);
> +		if (ret)
> +			goto fail;

Only issue I see: If v3d_job_init() fails, we need to not v3d_job_put()
it.  I'm fine with either kfree() it and NULL the ptr before jumping to
fail, or open code the bin/render puts.

With that,

Reviewed-by: Eric Anholt <eric at anholt.net>

> +
> +		last_job = clean_job;
> +	} else {
> +		last_job = &render->base;
> +	}
> +
> +	ret = v3d_lookup_bos(dev, file_priv, last_job,
>  			     args->bo_handles, args->bo_handle_count);
>  	if (ret)
>  		goto fail;
>  
> -	ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx);
> +	ret = v3d_lock_bo_reservations(last_job, &acquire_ctx);
>  	if (ret)
>  		goto fail;
>  
> @@ -599,28 +618,44 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
>  	ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
>  	if (ret)
>  		goto fail_unreserve;
> +
> +	if (clean_job) {
> +		struct dma_fence *render_fence =
> +			dma_fence_get(render->base.done_fence);
> +		ret = drm_gem_fence_array_add(&clean_job->deps, render_fence);
> +		if (ret)
> +			goto fail_unreserve;
> +		ret = v3d_push_job(v3d_priv, clean_job, V3D_CACHE_CLEAN);
> +		if (ret)
> +			goto fail_unreserve;
> +	}
> +
>  	mutex_unlock(&v3d->sched_lock);
>  
>  	v3d_attach_fences_and_unlock_reservation(file_priv,
> -						 &render->base,
> +						 last_job,
>  						 &acquire_ctx,
>  						 args->out_sync,
> -						 render->base.done_fence);
> +						 last_job->done_fence);
>  
>  	if (bin)
>  		v3d_job_put(&bin->base);
>  	v3d_job_put(&render->base);
> +	if (clean_job)
> +		v3d_job_put(clean_job);
>  
>  	return 0;
>  
>  fail_unreserve:
>  	mutex_unlock(&v3d->sched_lock);
> -	drm_gem_unlock_reservations(render->base.bo,
> -				    render->base.bo_count, &acquire_ctx);
> +	drm_gem_unlock_reservations(last_job->bo,
> +				    last_job->bo_count, &acquire_ctx);
>  fail:
>  	if (bin)
>  		v3d_job_put(&bin->base);
>  	v3d_job_put(&render->base);
> +	if (clean_job)
> +		v3d_job_put(clean_job);
>  
>  	return ret;
>  }
> diff --git a/include/uapi/drm/v3d_drm.h b/include/uapi/drm/v3d_drm.h
> index 58fbe48c91e9..58d2040ea48c 100644
> --- a/include/uapi/drm/v3d_drm.h
> +++ b/include/uapi/drm/v3d_drm.h
> @@ -48,6 +48,8 @@ extern "C" {
>  #define DRM_IOCTL_V3D_SUBMIT_TFU          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_TFU, struct drm_v3d_submit_tfu)
>  #define DRM_IOCTL_V3D_SUBMIT_CSD          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CSD, struct drm_v3d_submit_csd)
>  
> +#define DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG        0x01
> +
>  /**
>   * struct drm_v3d_submit_cl - ioctl argument for submitting commands to the 3D
>   * engine.
> @@ -61,7 +63,7 @@ extern "C" {
>   * flushed by the time the render done IRQ happens, which is the
>   * trigger for out_sync.  Any dirtying of cachelines by the job (only
>   * possible using TMU writes) must be flushed by the caller using the
> - * CL's cache flush commands.
> + * DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG flag.
>   */
>  struct drm_v3d_submit_cl {
>  	/* Pointer to the binner command list.
> @@ -124,8 +126,7 @@ struct drm_v3d_submit_cl {
>  	/* Number of BO handles passed in (size is that times 4). */
>  	__u32 bo_handle_count;
>  
> -	/* Pad, must be zero-filled. */
> -	__u32 pad;
> +	__u32 flags;
>  };
>  
>  /**
> -- 
> 2.17.1
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/dri-devel/attachments/20190912/a35f01dd/attachment.sig>