Hi Mark, On Wed, 2015-06-10 at 16:06 +0800, Eddie Huang wrote: > On Tue, 2015-06-09 at 11:39 +0100, Mark Brown wrote: > > On Tue, Jun 09, 2015 at 06:05:21PM +0800, Eddie Huang wrote: > > > > > --- a/drivers/spi/spi.c > > > +++ b/drivers/spi/spi.c > > > @@ -539,8 +539,8 @@ static int __spi_map_msg(struct spi_master *master, > > > struct spi_message *msg) > > > if (!master->can_dma) > > > return 0; > > > > > - tx_dev = master->dma_tx->device->dev; > > > - rx_dev = master->dma_rx->device->dev; > > > + tx_dev = master->dma_tx ? master->dma_tx->device->dev : > > > master->dev; > > > + rx_dev = master->dma_rx ? master->dma_rx->device->dev : master- > > > > > Is this what you want ? Actually, I don't like first one at all. > > > > Not quite what I'd been thinking of - we can't just pick the device in > > the core safely, the device might be a MFD or have some other > > restriction that needs us to use a separate struct device. However most > > of those cases are likely to point towards implementing a dmaengine > > device so probably the above will work for most cases and is fine. Can > > you send a proper patch please? > After further investigate, we found problem when use can_dma() to implement spi driver. If can_dma() return true in __spi_map_msg() function, it will map data buffer to sg_table, then pass this sg_table to transfer function. In transfer function, spi hardware can do tx and rx at the same time (full duplex), and the data size must be the same. Here comes the problem, although total length of tx, rx is the same, each entry in rx and tx scatterlist may not be the same (in the case data buffer allocate from vmalloc). Other vendor have dmaengine driver to send entry-by-entry automatically, so it is ok due to total length is the same.But mediatek hw can only send tx and rx scatterlist entry one by on, which means in full duplex, mediatek SPI hardware need send each entry separately, will cause full duplex fail because tx/rx entry size or entry number may not be the same. The problem is not dma map discuss earlier, it is mediatek spi hardware limitation that can't support scatterlist dma transfer in full-duplex mode. We can only support both tx and rx has the same size continuous memory data in full-duplex mode. I don't know whether should modify spi.c to support mediatek spi, or we just return can_dma() false and do transfer one continuous data in transfer function. By the way, I think maybe it is better to change can_dma() to can_dma_sg(). Eddie Thanks -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html