On Sat, Apr 13, 2019 at 08:24:14PM +0200, Noralf Trønnes wrote: > @@ -844,6 +816,17 @@ static int bcm2835_spi_prepare_message(struct spi_master *master, > struct spi_device *spi = msg->spi; > struct bcm2835_spi *bs = spi_master_get_devdata(master); > u32 cs = bcm2835_rd(bs, BCM2835_SPI_CS); > + int ret; > + > + /* > + * DMA transfers are limited to 16 bit (0 to 65535 bytes) by the SPI HW > + * due to DLEN. Split up transfers (32-bit FIFO aligned) if the limit is > + * exceeded. > + */ > + ret = spi_split_transfers_maxsize(master, msg, 65532, > + GFP_KERNEL | GFP_DMA); > + if (ret) > + return ret; This looks much better than the previous version because spi_split_transfers_maxsize() is now used as a library function by an individual driver, rather than something that is inflicted on *all* drivers (midlayer fallacy). Of course the performance is suboptimal compared to an approach which transmits the sglist in portions which are < 65535 each, but we can move to that later. In case it helps, Reviewed-by: Lukas Wunner <lukas@xxxxxxxxx> Thanks, Lukas