Re: [PATCH V3 3/5] drm/vkms: Decouple crc operations from composer

[Daniel Vetter <daniel@ffwll.ch>]

On Tue, Jun 25, 2019 at 10:37:58PM -0300, Rodrigo Siqueira wrote:
> In the vkms_composer.c, some of the functions related to CRC and compose
> have interdependence between each other. This patch reworks some
> functions inside vkms_composer to make crc and composer computation
> decoupled.
> 
> This patch is preparation work for making vkms able to support new
> features.
> 
> Signed-off-by: Rodrigo Siqueira <rodrigosiqueiramelo@gmail.com>
> ---
>  drivers/gpu/drm/vkms/vkms_composer.c | 49 ++++++++++++++++------------
>  1 file changed, 29 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/gpu/drm/vkms/vkms_composer.c b/drivers/gpu/drm/vkms/vkms_composer.c
> index eb7ea8be1f98..51a270514219 100644
> --- a/drivers/gpu/drm/vkms/vkms_composer.c
> +++ b/drivers/gpu/drm/vkms/vkms_composer.c
> @@ -105,35 +105,31 @@ static void compose_cursor(struct vkms_composer *cursor_composer,
>  	      primary_composer, cursor_composer);
>  }
>  
> -static uint32_t _vkms_get_crc(struct vkms_composer *primary_composer,
> -			      struct vkms_composer *cursor_composer)
> +static int compose_planes(void **vaddr_out,
> +			  struct vkms_composer *primary_composer,
> +			  struct vkms_composer *cursor_composer)
>  {
>  	struct drm_framebuffer *fb = &primary_composer->fb;
>  	struct drm_gem_object *gem_obj = drm_gem_fb_get_obj(fb, 0);
>  	struct vkms_gem_object *vkms_obj = drm_gem_to_vkms_gem(gem_obj);
> -	void *vaddr_out = kzalloc(vkms_obj->gem.size, GFP_KERNEL);
> -	u32 crc = 0;
>  
> -	if (!vaddr_out) {
> -		DRM_ERROR("Failed to allocate memory for output frame.");
> -		return 0;
> +	if (!*vaddr_out) {
> +		*vaddr_out = kzalloc(vkms_obj->gem.size, GFP_KERNEL);

Uh allocating memory here isn't great, since you effectily can't handle
the error at all. Also for big resolutions kzalloc will fall back to
kvmalloc I think, which is rather expensive to set up.

But I guess this is a preexisting issue, so welp.

What I would do is pull out the allocation at least, so that
compose_planes really only dos composing, can never fail because it
doesn't need to allocate memory.

> +		if (!*vaddr_out) {
> +			DRM_ERROR("Cannot allocate memory for output frame.");
> +			return -ENOMEM;
> +		}
>  	}
>  
> -	if (WARN_ON(!vkms_obj->vaddr)) {
> -		kfree(vaddr_out);
> -		return crc;
> -	}
> +	if (WARN_ON(!vkms_obj->vaddr))
> +		return -EINVAL;
>  
> -	memcpy(vaddr_out, vkms_obj->vaddr, vkms_obj->gem.size);
> +	memcpy(*vaddr_out, vkms_obj->vaddr, vkms_obj->gem.size);
>  
>  	if (cursor_composer)
> -		compose_cursor(cursor_composer, primary_composer, vaddr_out);
> +		compose_cursor(cursor_composer, primary_composer, *vaddr_out);
>  
> -	crc = compute_crc(vaddr_out, primary_composer);
> -
> -	kfree(vaddr_out);
> -
> -	return crc;
> +	return 0;
>  }
>  
>  /**
> @@ -154,9 +150,11 @@ void vkms_composer_worker(struct work_struct *work)
>  	struct vkms_output *out = drm_crtc_to_vkms_output(crtc);
>  	struct vkms_composer *primary_composer = NULL;
>  	struct vkms_composer *cursor_composer = NULL;
> +	void *vaddr_out = NULL;
>  	u32 crc32 = 0;
>  	u64 frame_start, frame_end;
>  	bool crc_pending;
> +	int ret;
>  
>  	spin_lock_irq(&out->composer_lock);
>  	frame_start = crtc_state->frame_start;
> @@ -180,14 +178,25 @@ void vkms_composer_worker(struct work_struct *work)
>  	if (crtc_state->num_active_planes == 2)
>  		cursor_composer = crtc_state->active_planes[1]->composer;
>  
> -	if (primary_composer)
> -		crc32 = _vkms_get_crc(primary_composer, cursor_composer);
> +	if (!primary_composer)
> +		return;
> +
> +	ret = compose_planes(&vaddr_out, primary_composer, cursor_composer);
> +	if (ret) {
> +		if (ret == -EINVAL)
> +			kfree(vaddr_out);
> +		return;
> +	}
> +
> +	crc32 = compute_crc(vaddr_out, primary_composer);
>  
>  	/*
>  	 * The worker can fall behind the vblank hrtimer, make sure we catch up.
>  	 */
>  	while (frame_start <= frame_end)
>  		drm_crtc_add_crc_entry(crtc, true, frame_start++, &crc32);
> +
> +	kfree(vaddr_out);

Especially since you're freeing the memory _outside_ of compose_planes.

Aside: This all kinda doesn't go in the right direction for
high-performance composing, so I guess I need to get started with typing
up what that should look like.
-Daniel

>  }
>  
>  static const char * const pipe_crc_sources[] = {"auto"};
> -- 
> 2.21.0

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch