Whenever a single or cyclic transaction is prepared, the driver could eventually split it over several SG descriptors in order to deal with the HW maximum transfer length. This could end up in DMA operations starting from a misaligned address. This seems fatal for the HW if DRE (Data Realignment Engine) is not enabled. This patch eventually adjusts the transfer size in order to make sure all operations start from an aligned address. Cc: Radhey Shyam Pandey <radhey.shyam.pandey@xxxxxxxxxx> Signed-off-by: Andrea Merello <andrea.merello@xxxxxxxxx> Reviewed-by: Radhey Shyam Pandey <radhey.shyam.pandey@xxxxxxxxxx> --- Changes in v2: - don't introduce copy_mask field, rather rely on already-esistent copy_align field. Suggested by Radhey Shyam Pandey - reword title Changes in v3: - fix bug introduced in v2: wrong copy size when DRE is enabled - use implementation suggested by Radhey Shyam Pandey Changes in v4: - rework on the top of 1/6 Changes in v5: - fix typo in commit title - add hint about "DRE" meaning in commit message Changes in v6: None --- drivers/dma/xilinx/xilinx_dma.c | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/drivers/dma/xilinx/xilinx_dma.c b/drivers/dma/xilinx/xilinx_dma.c index 2c1db500284f..cbf34dd5e966 100644 --- a/drivers/dma/xilinx/xilinx_dma.c +++ b/drivers/dma/xilinx/xilinx_dma.c @@ -975,6 +975,15 @@ static int xilinx_dma_calc_copysize(struct xilinx_dma_chan *chan, copy = min_t(size_t, size - done, chan->xdev->max_buffer_len); + if ((copy + done < size) && + chan->xdev->common.copy_align) { + /* + * If this is not the last descriptor, make sure + * the next one will be properly aligned + */ + copy = rounddown(copy, + (1 << chan->xdev->common.copy_align)); + } return copy; } -- 2.17.1