Re: [PATCH v2 1/3] media/videobuf2-dma-sg: Fix handling of sg_table structure

[Hans Verkuil <hverkuil@xs4all.nl>]

Hi Ricardo, Marek,

I have a few questions, mostly to improve my own understanding.

First of all, is this solving an actual bug for you, or did you just find
it while reviewing code? And if it solves a bug, then which architecture
are you using? ARM? Intel?

On 02/11/2015 11:33 AM, Ricardo Ribalda Delgado wrote:
> When sg_alloc_table_from_pages() does not fail it returns a sg_table
> structure with nents and nents_orig initialized to the same value.
> 
> dma_map_sg returns the number of areas mapped by the hardware,
> which could be different than the areas given as an input.
> The output must be saved to nent.
> 
> The output of dma_map, should be used to transverse the scatter list.
> 
> dma_unmap_sg needs the value passed to dma_map_sg (nents_orig).

I noticed that few dma_unmap_sg calls actually use orig_nents. It makes
me wonder if the dma_unmap_sg documentation is actually correct. It does
clearly state that orig_nents should be used, and it might well be that
the only reason this hasn't led to problems is that very few architectures
actually seem to return nents < orig_nents.

> 
> sg_free_tables uses also orig_nent.
> 
> This patch fix the file to follow this paradigm.
> 
> Signed-off-by: Ricardo Ribalda Delgado <ricardo.ribalda@gmail.com>
> ---
>  drivers/media/v4l2-core/videobuf2-dma-sg.c | 22 +++++++++++++---------
>  1 file changed, 13 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/media/v4l2-core/videobuf2-dma-sg.c b/drivers/media/v4l2-core/videobuf2-dma-sg.c
> index b1838ab..40c330f 100644
> --- a/drivers/media/v4l2-core/videobuf2-dma-sg.c
> +++ b/drivers/media/v4l2-core/videobuf2-dma-sg.c
> @@ -147,8 +147,9 @@ static void *vb2_dma_sg_alloc(void *alloc_ctx, unsigned long size,
>  	 * No need to sync to the device, this will happen later when the
>  	 * prepare() memop is called.
>  	 */
> -	if (dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->nents,
> -			     buf->dma_dir, &attrs) == 0)
> +	sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
> +				      buf->dma_dir, &attrs);

Is a driver free to change sgt->nents? It's unclear from the documentation
or code that that is actually the purpose of sgt->nents. Most drivers seem
to store the result of dma_map_sg into a driver-specific struct.

Regards,

	Hans

> +	if (!sgt->nents)
>  		goto fail_map;
>  
>  	buf->handler.refcount = &buf->refcount;
> @@ -187,7 +188,7 @@ static void vb2_dma_sg_put(void *buf_priv)
>  		dma_set_attr(DMA_ATTR_SKIP_CPU_SYNC, &attrs);
>  		dprintk(1, "%s: Freeing buffer of %d pages\n", __func__,
>  			buf->num_pages);
> -		dma_unmap_sg_attrs(buf->dev, sgt->sgl, sgt->nents,
> +		dma_unmap_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
>  				   buf->dma_dir, &attrs);
>  		if (buf->vaddr)
>  			vm_unmap_ram(buf->vaddr, buf->num_pages);
> @@ -314,9 +315,11 @@ static void *vb2_dma_sg_get_userptr(void *alloc_ctx, unsigned long vaddr,
>  	 * No need to sync to the device, this will happen later when the
>  	 * prepare() memop is called.
>  	 */
> -	if (dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->nents,
> -			     buf->dma_dir, &attrs) == 0)
> +	sgt->nents = dma_map_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents,
> +				      buf->dma_dir, &attrs);
> +	if (!sgt->nents)
>  		goto userptr_fail_map;
> +
>  	return buf;
>  
>  userptr_fail_map:
> @@ -351,7 +354,8 @@ static void vb2_dma_sg_put_userptr(void *buf_priv)
>  
>  	dprintk(1, "%s: Releasing userspace buffer of %d pages\n",
>  	       __func__, buf->num_pages);
> -	dma_unmap_sg_attrs(buf->dev, sgt->sgl, sgt->nents, buf->dma_dir, &attrs);
> +	dma_unmap_sg_attrs(buf->dev, sgt->sgl, sgt->orig_nents, buf->dma_dir,
> +			   &attrs);
>  	if (buf->vaddr)
>  		vm_unmap_ram(buf->vaddr, buf->num_pages);
>  	sg_free_table(buf->dma_sgt);
> @@ -502,7 +506,6 @@ static struct sg_table *vb2_dma_sg_dmabuf_ops_map(
>  	/* stealing dmabuf mutex to serialize map/unmap operations */
>  	struct mutex *lock = &db_attach->dmabuf->lock;
>  	struct sg_table *sgt;
> -	int ret;
>  
>  	mutex_lock(lock);
>  
> @@ -521,8 +524,9 @@ static struct sg_table *vb2_dma_sg_dmabuf_ops_map(
>  	}
>  
>  	/* mapping to the client with new direction */
> -	ret = dma_map_sg(db_attach->dev, sgt->sgl, sgt->orig_nents, dma_dir);
> -	if (ret <= 0) {
> +	sgt->nents = dma_map_sg(db_attach->dev, sgt->sgl, sgt->orig_nents,
> +				dma_dir);
> +	if (!sgt->nents) {
>  		pr_err("failed to map scatterlist\n");
>  		mutex_unlock(lock);
>  		return ERR_PTR(-EIO);
>