Hi Martin,
Based on the dmaengine document
https://kernel.org/doc/html/v6.3-rc2/driver-api/dmaengine/client.html
The sg_len used for dmaengine_prep_slave_sg should be returned by
dma_map_sg.
This implies there should not be zero segment for the first sg_len segments.
I think that is why we need to provide 'sg_len' for
dmaengine_prep_slave_sg().
With 'sg_len', we do not need to worry about zero length sg in the
driver callback.
Thanks,
Lizhi
On 3/16/23 10:03, Martin Tůma wrote:
Hi,
The Xilinx XDMA driver crashes when the scatterlist provided to
xdma_prep_device_sg() contains an empty entry, i.e. sg_dma_len()
returns zero. As I do get such sgl from v4l2 I suppose that this
is a valid scenario and not a bug in our parent mgb4 driver. Also
the documentation for sg_dma_len() suggests that there may be
zero-length entries.
The following trivial patch fixes the crash:
diff --git a/drivers/dma/xilinx/xdma.c b/drivers/dma/xilinx/xdma.c
index 462109c61653..cd5fcd911c50 100644
--- a/drivers/dma/xilinx/xdma.c
+++ b/drivers/dma/xilinx/xdma.c
@@ -487,6 +487,8 @@ xdma_prep_device_sg(struct dma_chan *chan, struct
scatterlist *sgl,
for_each_sg(sgl, sg, sg_len, i) {
addr = sg_dma_address(sg);
rest = sg_dma_len(sg);
+ if (!rest)
+ break;
do {
len = min_t(u32, rest, XDMA_DESC_BLEN_MAX);
M.
