On Monday 25 June 2012, Vinod Koul wrote:

On Mon, 2012-06-25 at 11:51 -0500, Jon Hunter wrote:

quoted

Hi Russell,

quoted

Dan, Vinod, in this thread we have been discussing the addition of a
generic device-tree binding for DMA controllers. In the below, we were
discussing the addition of a device-tree API, which would work as a
wrapper to the dma-engine dma_request_channel() API. I apologise for
adding you late into the discussion. If you have any questions/comments
let me know.

Looks like this a long discussion, I will try to go through archives.

But am still unsure about about dmaengine part. If we have DT binding
for dma controllers, why it they worry about dma_request_channel()
IMO, the problem of channel mapping is not DT specific, it need to be
solved at dmaengine and possibly the required mapping can come from DT
among other mechanisms for various platforms.

This is what google told me about this patch set:

dma_request_channel is called with some information about the channel
provided in its arguments, and the driver might get that from a number
of places.

In the case of having a fully populated device tree with this binding,
the driver calling (of_)dma_request_channel does not need to know about
any of that information because we should be able to encapsulate that
completely in device tree data. It does not replace the regular interface
but wraps around it to provide a higher abstraction level where possible.

Of course if you think we should not be doing that but instead
have of_dma_request_channel() live besides dma_request_channel()
rather than calling it, that should be absolutely fine too.

quoted

   In the case of DMA controllers that are using DMA Engine, requesting a
   channel is performed by calling the following function.

	struct dma_chan *dma_request_channel(dma_cap_mask_t mask,
			dma_filter_fn filter_fn,
			void *filter_param);

   The mask variable is used to identify the device controller in a list of
   controllers. The filter_fn and filter_param are used to identify the
   required dma channel and return a handle to the dma channel of type
   dma_chan. From the examples I have seen, the mask and filter_fn are constant
   for a given DMA controller. Therefore, when registering a DMA controller with
   device tree we can pass these parameters and store them so that a device can
   request them when requesting a channel. Hence, based upon this our register
   function for the DMA controller now looks like this.

	int of_dma_controller_register(struct device_node *np,
		dma_cap_mask_t *mask, dma_filter_fn fn);

IMO we should do away with filter functions.
If we solve the mapping problem, then we don't need a filer.

The channel data in the device tree is still in a format
that is specific to that dmaengine driver and interpreted
by it. Using the regular dma_filter_fn prototype is not
necessary, but it would be convenient because the dmaengine
code already knows how to deal with it. If we don't use this
method, how about adding another callback to struct dma_device
like

bool (*device_match)(struct dma_chan *chan, struct property *req);

quoted

2. Supporting legacy devices not using DMA Engine

   These devices present a problem, as there may not be a uniform way to easily
   support them with regard to device tree. However, _IF_ legacy devices that
   are not using DMA Engine, only have a single DMA controller, then this
   problem is a lot simpler. For example, if we look at the previously proposed
   API for registering a DMA controller (where we pass the mask and function
   pointer to the DMA Engine filter function) we can simply pass NULL and hence,
   a driver requesting the DMA channel information would receive NULL for the
   DMA Engine specific parameters. Then for legacy devices we simply need a
   means to return the channel information (more on this later). If there are
   legacy devices that do have multiple DMA controllers, then maybe they need to
   be converted to support DMA Engine. I am not sure if this is unreasonable???

Why should these be supported? They should be converted to use dmaengine
over a reasonable amount of time.

I agree, at least for the long run. However, that is a separate issue to work on.
Right now we need a generic way to represent dma requests independent of how
they are used in the kernel. The device tree binding is supposed to be
operating system independent so there should be nothing in it that requires
the use of the linux dmaengine code.

For drivers that do not use dmaengine, we have to make a decision whether
it's worth adding support for the DT binding first and converting the driver
and its users to dmaengine later, or whether it's better to use the dmaengine
API right away to avoid having to do changes twice.

quoted

3. Representing and requesting channel information

   From a hardware perspective, a DMA channel could be represented as ...

   i. channel index/number
   ii. channel transfer type (optional)
   iii. DMA interrupt mapping (optional)

  Please note that the transfer type is used to indicate if the transfer is to
  device from memory, to memory from device, to memory from memory, etc. This
  can be useful when there is a device such as an MMC device that uses two DMA
  channels, one for reading (RX) and one for writing (TX).
From a dma controller perspective, it can service both with single

channel.
I have dma controller which can talk to three peripherals on both
transmit and receive direction. The point is that 1:1 mapping of dma
channel does not exist. So any representation which tries to do this may
not work. 

In the device tree, we know both the dmaengine and its slave, so we also know
whether to put one or more requests in there.

quoted

	struct of_dma_channel_info {
		int		dma_channel;
		dma_cap_mask_t	dma_cap;
		dma_filter_fn	dma_filter_func;
	};

   Here dma_channel will always be valid and the other fields are optional
   depending on whether DMA Engine is used.

This implementation has been tested on OMAP4430 using Russell King's latest
DMA Engine series for OMAP [3] and with Benoit Cousson latest DT changes for
OMAP4 [4]. I have validated that MMC is working on the PANDA board with this
implementation. I have not included all the changes for PANDA board here but
just wished to share the implementation.

I am still unclear on how this attempts to solve mapping problem? Maybe
i need more coffee at midnight break!

I believe we have moved on from this proposal to a simpler one,
doing away with the of_dma_channel_info.

quoted

quoted

On Fri, Jun 22, 2012 at 05:52:08PM -0500, Jon Hunter wrote:

quoted

quoted

1. chan = of_dma_request_channel(dev->of_node, 0);
2. chan = of_dma_request_channel(dev->of_node, 0);
3. rxchan = of_dma_request_channel(dev->of_node, DMA_MEM_TO_DEV);
   txchan = of_dma_request_channel(dev->of_node, DMA_DEV_TO_MEM);
4. rxchan = of_dma_request_channel(dev->of_node, DMA_MEM_TO_DEV);
   txchan = of_dma_request_channel(dev->of_node, DMA_DEV_TO_MEM);
5. chan = of_dma_request_named_channel(dev->of_node, "rwdata", 0);
   auxchan = of_dma_request_named_channel(dev->of_node, "status", DMA_DEV_TO_MEM);
6. chan = of_dma_request_named_channel(dev->of_node, "rwdata", 0);
   auxchan = of_dma_request_named_channel(dev->of_node, "status", DMA_DEV_TO_MEM);

In the above examples, did you imply that the of_dma_request_channel()
function would return a type of "struct dma_chan" and so be calling
dma_request_channel() underneath?

I am been prototyping something, but wanted to make sure I am completely
aligned on this :-)

This is what I think we need to be heading to.

	Arnd