On Tue, Jan 28, 2014 at 10:16:53AM +0100, Arnd Bergmann wrote:

On Tuesday 28 January 2014 10:05:35 Lars-Peter Clausen wrote:

quoted

quoted

+
+Clients must use the format described in the dma.txt file, using a three cell
+specifier for each channel.
+
+The three cells in order are:
+  1. A phandle pointing to the DMA controller
+  2. The channel number
+  3. Direction of the fixed unidirectional channel
+     0 - Memory to Device
+     1 - Device to Memory
+     2 - Device to Device
+

Why does the direction needs to be specified in specifier? I see two
options, either the direction per is fixed in hardware. In that case the DMA
controller node should describe which channel is which direction. Or the
direction is not fixed in hardware and can be changed at runtime in which
case it should be set on a per descriptor basis.

Normally the direction is implied by dmaengine_slave_config().
Note that neither the dma slave API nor the generic DT binding
can actually support device-to-device transfers, since this
normally implies using two dma-request lines rather than one.

There might be a case where the direction is required in order
to allocate a channel, because the engine has specialized channels
per direction, and might connect any of them to any dma request
line. This does not seem to be the case for "bam", because
the DMA specifier already contains a specific channel number, not
a request line or slave ID number.

After some deliberation, I think the best solution is removing the direction
from the DT for now.  It doesn't add anything except some verification
of direction.

As for the device to device:
As I mentioned before, each bam dma node is attached to a specific peripheral
(with one exception, but lets skip over that).  The peripherals allow for more
than one execution environment to access the peripheral and attached bam.  2 bam
channels can be connected to form a unidirectional pipe from one execution
environment to another.  Once the pipe is configured, the actually transfer
resembles a cyclical dma transfer and continues until you explicitly stop it.

That functionality will come later.

-- 
sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
hosted by The Linux Foundation