Hi Alex,

On 26/02/15 20:04, Alex Smith wrote:

Hi Zubair,

On 26 February 2015 at 12:43, Zubair Lutfullah Kakakhel
[off-list ref] wrote:

quoted

From: Alex Smith <redacted>

Add device tree bindings for the DMA controller on JZ4780 SoCs, used by
the dma-jz4780 driver.

Signed-off-by: Alex Smith <redacted>
Signed-off-by: Zubair Lutfullah Kakakhel <redacted>

---
V3 -> V2
Changed binding.
Used to be 3 DMA cells required. <&dma TX_type RX_type Reserved>
Now 2 DMA cells are required. <&dma Transfer_type Reserved>

This is more common in DMA bindings.
And I couldn't figure any reason that 3 cells were used.

There are different request type numbers for transfers to/from the
same device (see the JZ4780 programmers manual, page 505). While only
having the option to specify one transfer type is OK when the driver
is using separate channels for read/writes, I recall seeing/writing
some other drivers which use a single channel for both reads and
writes. This would not be possible if we can only specify one transfer
type, you'd have to have them separate.

I know. I looked at the drivers and did this on purpose.

We'd like to keep the same bindings/code for jz4740/jz4780 peripheral drivers and dma code.

Our jz47xx-mmc driver we had was the main culprit I found. As well as the jz4740-i2s one.

However, jz4740-mmc upstream driver has improved already for dma and takes two dma channels.
And the jz4740-i2s also takes two channels. One for Tx and one for Rx.

If we move to 3 cells for jz4780-dma. Then 'ideally' jz4740-dma would need 3 cells too.
Or we'd have a binding nightmare everywhere.

But when we use jz4740-mmc and jz4740-i2s, we still have to change the driver to share one channel 
or simply pass them
<&dma Tx 0 Reserved>
<&dma 0 Rx Reserved>

Which makes the binding redundant.

There isn't any particular reason a driver would need to share one channel only.
The 'special' nand/nemc channels don't have a request type.

Thanks,
ZubairLK

Thanks,
Alex