Re: [PATCH 01/11] dmaengine: add context parameter toprep_slave_sgand prep_dma_cyclic

From: Vinod Koul
Date: Mon Feb 06 2012 - 13:02:33 EST

On Mon, 2012-02-06 at 15:53 +0000, Russell King wrote:
> On Mon, Feb 06, 2012 at 08:58:54PM +0530, Vinod Koul wrote:
> > On Mon, 2012-02-06 at 07:04 -0800, Bounine, Alexandre wrote:
> > > I was thinking about keeping the original scatterlist for dmac unchanged,
> > > but allocating another scatterlist in rio_dma_prep_slave_sg() and chaining
> > > two lists together. After it passed to device specific function, it parses
> > > first section of the chain for additional transfer parameters and use
> > > following scatterlist as intended for dmac.
> > hmmm, but that wouldn't make it generic for other systems like proposed
> > MSM box mode..., right?
> > >
> > > But Russell's idea (See ) seems to be
> > > a better way without added complexity and I am ready to take that path and
> > > make new patches if you and Dan agree with it.
> > but it still doesn't fix his concerns for the using void pointer.
> It helps because it makes it easier to find those drivers who are not
> using the generic interface (and so would be tied to their particular
> DMA engine.)
> Let's take a step back.
> One of the fundamental points for having a DMA engine API is to separate
> the DMA users from the DMA engines, and make the two independent. This
> allows DMA users (like MMC, UARTs, etc) to be written in a way that they
> are totally independent of the DMA engine.
> Why is this important? We're seeing an increasing number of SoCs with
> the same peripheral IP integrated onto them but with different DMA
> controllers. Not only does that happen within an architecture, but you
> can bet it'll start happening outside. For example, AMBA Primecell
> peripherals are not only found on ARM but also on SoCs with differing
> CPUs. Some PXA peripherals are now being found on x86 hardware.
Yes this is the basic premise we are working on.
> So, we need a basic DMA engine API which ensures that users of the API
> do not have to know any details about the DMA engine itself for them to
> be able to get on with their business.
For memcpy I believe that is mostly the case.
But problem arises for slave-dma partly due to
a) controllers need specific channel as only specific channel(s) can
talk to their peripheral, and today we don't have a good way to do this
across arch. (x86, arm, RIO.....)
b) some dmacs which support a generic transfer pattern but need more
information in order to get a transfer done, so some specific parameters
have to be passed. For most of drivers today we have reached a point
where we need to remove any specific params introduced (chan->private).
But for these controllers (RIO, MSM), it seems they would need few more
non generic parameters!!!
So we can
a) say no, don't use dmaengine APIs and invent your own stuff
b) try to accommodate while keeping dmaengine neutral.
> Now, if we start allowing a 'void *' per-transfer random pointer, then
> what will happen is that we'll end up with people abusing it for passing
> other stuff, maybe such as DMA request lines because they don't want to
> bother with virtual channels in their DMA engine. At that point, the
> peripheral drivers gain knowledge about the DMA engine they're attached
> to, and it becomes impossible to reuse them without resorting to ifdeffery
> in the peripheral drivers.
Yes that was my concern too which was listed when we discussed original
patch from Alexandre and this as possible approach.
> This brings up a fundamental question: if you have a DMA engine which
> has some unique specific feature, and needs extra data to use that funky
> feature passed through the generic API, is it appropriate to use the
> generic API, or does it make sense to augment the API in some way?
> It might be that the peripherals of this funky feature are soo tied to
> the DMA engine that it wouldn't ever be appropriate for them to live
> with a different DMA engine. In that case, does the DMA engine API
> actually benefit them, or does it result in an erosion of the API
> boundary, and therefore the separation of DMA engine from DMA user?
> If the net result will be that the DMA engine API boundaries are blured,
> then we've lost, because we need a replacement API to do the job that we
> desperately need our existing API to be doing for us.
> And that's my biggest concern about nebulous undefined 'void *'
> arguments across this API.
Okay, so now how about _defining_ these arch specific parameters in
dmaengine_arch (or some better place) and clients which are not on those
will use generic dmaengine_prep_slave_sg() as you outlined, and the ones
like RIO use

struct rio_specfic_data {

static inline struct dma_async_tx_descriptor *dmaengine_prep_slave_sg_rio(
struct dma_chan *chan, struct scatterlist *sg, unsigned sg_len,
enum dma_transfer_direction direction, unsigned long flags,
struct rio_specfic_data *data)
return chan->device->device_prep_slave_sg(chan, sg, sglen, dir, flags, (void *)data);

That way no client needs to pass a loose parameter. And we ensure dmacs
are using the last argument only when defined for particular arch.


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at