Re: [PATCH] dmaengine: bcm2835: Add slave dma support

["Noralf Trønnes" <noralf@tronnes.org>]

Den 16.04.2015 21:06, skrev Alexander Stein:
> Hi Stefan,
>
> On Wednesday 15 April 2015, 21:00:26 wrote Stefan Wahren:
>> Am 15.04.2015 um 11:56 schrieb Noralf Trønnes:
>>> Add slave transfer capability to BCM2835 dmaengine driver.
>>> This patch is pulled from the bcm2708-dmaengine driver in the
>>> Raspberry Pi repo. The work was done by Gellert Weisz.
>>>
>>> Tested with the bcm2835-mmc driver from the same repo.
>> why not with the upstream kernel?
> I also looked at slave dma support, especially for use in mmc. It turns our that bcm2835-mmc is written more or less completly new.
> Mainline linux uses sdhci "framework" which internally uses the SDMA and/or ADMA (both internal, to SD/MMC controller, DMA units) which can be supported by an SDHCI compatible controller.
> AFAIK the SD/MMC controller in bcm2835 lacks both that is why the driver only uses PIO. I dunno if external DMA usage can so easily be integrated into the sdhci, I have my doubts.

I asked Jonathan Bell (Raspberry Pi) about why a new driver was made
instead of extending sdhci-bcm2835.

On 10.04.2015 20:02, Jonathan Bell wrote:
 > Basically, it's impossible to integrate platform DMA channel support
 > within the SDHCI framework. The Arasan controller (and the Broadcom
 > MMCI controller) both use platform DMA channels to pump data to/from
 > the host FIFO. Our old "sdhci-bcm2708" driver basically hacked sdhci.c
 > to allow platform DMA support in a way that was guaranteed to cause
 > merge conflicts with every new kernel branch. The reasoning behind
 > creating an MMC-level driver was to minimise disruption of incorporating
 > platform DMA and to have additional control e.g. on sequencing of 
commands
 > that are known to have bugs/problems. There are drivers in the source
 > tree that are "SDHCI compliant" but have their own various idiosyncrasies
 > - e.g. : 
http://lxr.free-electrons.com/source/drivers/mmc/host/davinci_mmc.c
 > Implementing an MMC-level driver was the easiest way to incorporate all
 > our various bits of baggage (random necessary delays here, busy-wait 
there)
 > without disrupting the rest of the codebase. I agree that some functions
 > could just substitute the sdhci.c equivalents and deduplicate some of 
the code.


Stephen Warren made this comment on a previous attempt to upstream the
bcm2835-mmc driver:

On Tue, 28 Oct 2014 19:55:20 -0600, Stephen Warren wrote:
 > On 10/28/2014 06:00 PM, Piotr Król wrote:
 > > This is driver for Arasan External Mass Media Controller provided in
 > > Raspberry Pi single board computer.
 >
 > We should not have multiple drivers for the same HW. The correct
 > approach would be to enhance the existing sdhci-bcm2835.c to support any
 > new features or bug-fixes embodied within this driver. Presumably that
 > way, you'd also end up with a lot of small feature patches, which would
 > make patch review easier. Consequently I haven't reviewed this patch 
much.