Re: [PATCH v2 2/3] dmaengine: imx-sdma: add memcpy interface

[Sascha Hauer]

On Fri, Jul 20, 2018 at 09:40:51AM +0000, Robin Gong wrote:
> > -----Original Message-----
> > From: Sascha Hauer [mailto:s.hauer@xxxxxxxxxxxxxx]
> > Sent: 2018å7æ13æ 14:16
> > To: Robin Gong <yibin.gong@xxxxxxx>
> > Cc: vkoul@xxxxxxxxxx; dan.j.williams@xxxxxxxxx; shawnguo@xxxxxxxxxx; Fabio
> > Estevam <fabio.estevam@xxxxxxx>; linux@xxxxxxxxxxxxxxx;
> > linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; kernel@xxxxxxxxxxxxxx;
> > dmaengine@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; dl-linux-imx
> > <linux-imx@xxxxxxx>
> > Subject: Re: [PATCH v2 2/3] dmaengine: imx-sdma: add memcpy interface
> > 
> > On Fri, Jul 13, 2018 at 09:08:46PM +0800, Robin Gong wrote:
> > > Add MEMCPY capability for imx-sdma driver.
> > >
> > > Signed-off-by: Robin Gong <yibin.gong@xxxxxxx>
> > > ---
> > >  drivers/dma/imx-sdma.c | 95
> > > ++++++++++++++++++++++++++++++++++++++++++++++++--
> > >  1 file changed, 92 insertions(+), 3 deletions(-)
> > >
> > > @@ -1318,6 +1347,63 @@ static struct sdma_desc
> > *sdma_transfer_init(struct sdma_channel *sdmac,
> > >  	return NULL;
> > >  }
> > >
> > > +static struct dma_async_tx_descriptor *sdma_prep_memcpy(
> > > +		struct dma_chan *chan, dma_addr_t dma_dst,
> > > +		dma_addr_t dma_src, size_t len, unsigned long flags) {
> > > +	struct sdma_channel *sdmac = to_sdma_chan(chan);
> > > +	struct sdma_engine *sdma = sdmac->sdma;
> > > +	int channel = sdmac->channel;
> > > +	size_t count;
> > > +	int i = 0, param;
> > > +	struct sdma_buffer_descriptor *bd;
> > > +	struct sdma_desc *desc;
> > > +
> > > +	if (!chan || !len)
> > > +		return NULL;
> > > +
> > > +	dev_dbg(sdma->dev, "memcpy: %pad->%pad, len=%zu, channel=%d.\n",
> > > +		&dma_src, &dma_dst, len, channel);
> > > +
> > > +	desc = sdma_transfer_init(sdmac, DMA_MEM_TO_MEM,
> > > +					len / SDMA_BD_MAX_CNT + 1);
> > > +	if (!desc)
> > > +		return NULL;
> > > +
> > > +	do {
> > > +		count = min_t(size_t, len, SDMA_BD_MAX_CNT);
> > > +		bd = &desc->bd[i];
> > > +		bd->buffer_addr = dma_src;
> > > +		bd->ext_buffer_addr = dma_dst;
> > > +		bd->mode.count = count;
> > > +		desc->chn_count += count;
> > > +		/* align with sdma->dma_device.copy_align: 4bytes */
> > > +		bd->mode.command = 0;
> > > +
> > > +		dma_src += count;
> > > +		dma_dst += count;
> > > +		len -= count;
> > > +		i++;
> > 
> > NACK.
> > 
> > Please actually look at your code and find out where you do unaligned DMA
> > accesses. Hint: What happens when this loop body is executed more than
> > once?
> > 
> > Sascha
> Actually internal sdma script has already take such 'align' matter like below:
> Burst with words if source/dest address and data length are all aligned with words, burst with
> Byte if not.

If that's the case why did you add the DMAENGINE_ALIGN_4_BYTES flag in the
first place?

Also, if the SDMA script falls back to byte copy when necessary, doing
so probably has performance impacts, so I wonder why you don't just cut
'count' to something that is four byte aligned.

Sascha

-- 
Pengutronix e.K.                           |                             |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |
Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |