Hi Hans,

On 21/09/15 14:13, Hans Verkuil wrote:

Hi Tiffany!

On 21-09-15 14:26, Tiffany Lin wrote:

quoted

vb2_dc_prepare use the number of SG entries dma_map_sg_attrs return.
But in dma_sync_sg_for_device, it use lengths of each SG entries
before dma_map_sg_attrs. dma_map_sg_attrs will concatenate
SGs until dma length > dma seg bundary. sgt->nents will less than
sgt->orig_nents. Using SG entries after dma_map_sg_attrs
in vb2_dc_prepare will make some SGs are not sync to device.
After add DMA_ATTR_SKIP_CPU_SYNC in vb2_dc_get_userptr to remove
sync data to device twice. Device randomly get incorrect data because
some SGs are not sync to device. Change to use number of SG entries
before dma_map_sg_attrs in vb2_dc_prepare to prevent this issue.

Signed-off-by: Tiffany Lin <tiffany.lin@mediatek.com>
---
  drivers/media/v4l2-core/videobuf2-dma-contig.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/media/v4l2-core/videobuf2-dma-contig.c b/drivers/media/v4l2-core/videobuf2-dma-contig.c
index 2397ceb..c5d00bd 100644
--- a/drivers/media/v4l2-core/videobuf2-dma-contig.c
+++ b/drivers/media/v4l2-core/videobuf2-dma-contig.c

@@ -100,7 +100,7 @@ static void vb2_dc_prepare(void *buf_priv)
  	if (!sgt || buf->db_attach)
  		return;

-	dma_sync_sg_for_device(buf->dev, sgt->sgl, sgt->nents, buf->dma_dir);
+	dma_sync_sg_for_device(buf->dev, sgt->sgl, sgt->orig_nents, buf->dma_dir);
  }

  static void vb2_dc_finish(void *buf_priv)

@@ -112,7 +112,7 @@ static void vb2_dc_finish(void *buf_priv)
  	if (!sgt || buf->db_attach)
  		return;

-	dma_sync_sg_for_cpu(buf->dev, sgt->sgl, sgt->nents, buf->dma_dir);
+	dma_sync_sg_for_cpu(buf->dev, sgt->sgl, sgt->orig_nents, buf->dma_dir);
  }

I don't really understand it. I am assuming that this happens on an arm and that
the dma_map_sg_attrs and dma_sync_sg_* functions used are arm_iommu_map_sg() and
arm_iommu_sync_sg_* as implemented in arch/arm/mm/dma-mapping.c.

Now, as I understand it (and my understanding may very well be flawed!) the map_sg
function concatenates SG entries if possible, so it may return fewer entries. But
the dma_sync_sg functions use those updated SG entries, so the full buffer should
be covered by this. Using orig_nents will actually sync parts of the buffer twice!
The first nents entries already cover the full buffer so any remaining entries up
to orig_nents will just duplicate parts of the buffer.

As Documentation/DMA-API.txt says, the parameters to dma_sync_sg_* must 
be the same as those originally passed into dma_map_sg. The segments are 
only merged *from the point of view of the device*: if I have a 
scatterlist of two discontiguous 4K segments, I can remap them with an 
IOMMU so the device sees them as a single 8K buffer, and tell it as 
such. If on the other hand I want to do maintenance from the CPU side 
(i.e. any DMA API call), then those DMA addresses mean nothing and I can 
only operate on the CPU addresses of the underlying pages, which are 
still very much discontiguous in the linear map; ergo I still need to 
iterate over the original entries.

Whilst I can't claim much familiarity with v4l itself, from a brief look 
over the existing code this patch does look to be doing the right thing.

Robin.