Re: [PATCH v7 01/10] ARM: davinci: move private EDMA API to arm/common

[Linus Walleij <linus.walleij@linaro.org>]

On Tue, Feb 5, 2013 at 6:14 PM, Russell King - ARM Linux
<linux@arm.linux.org.uk> wrote:
> On Tue, Feb 05, 2013 at 04:30:45PM +0100, Linus Walleij wrote:

>> So put them on a wait list? Surely you will have a list of pending
>> cookies and pick from the front of the queue if there isn't a hole on
>> queue position 0.
>
> Not quite.  The key is the cookie system DMA engine employs to indicate
> when a cookie is complete.
>
> Cookies between the "issued sequence" and "completed sequence" are defined
> to be in progress, everything else is defined to be completed.
>
> This means that if "completed sequence" is 1, and "issued sequence" is 5,
> then cookies with values 2, 3, 4, 5 are in progress.  You can't mark
> sequence 4 as being complete until 2 and 3 have completed.

Yes that is true. DMA transfers on a certain channel are defined
as progressing linearly per-cookie. I wonder if that is a problem
in this case though (actually it seems the reverse, this helps
in Cyril's case.)

> If we need out-of-order completion, then that's a problem for the DMA
> engine API, because you'd need to radically change the way "completion"
> is marked.

True. I wonder if this usecase is ever going to be applicable
however. It could maybe be useful in some instances of
memcpy() I could dream up, whereas for device transfers it
seems unlikely to me.

Yours,
Linus Walleij