Re: question about block layer and dma alignment issue.

[Jens Axboe <axboe@kernel.dk>]

On 10/07/2014 03:37 AM, Oleksij Rempel wrote:
> Am 04.10.2014 um 19:00 schrieb Jens Axboe:
>> On 2014-10-04 00:59, Oleksij Rempel wrote:
>>> Hello Jens,
>>>
>>> i hope it is OK to ask you directly, i didn't found better way to do
>>> this.
>>> I'm writing MMC driver based on reverse engineering for au6601 mmc
>>> controller. This driver uses MMC API, and last one trying to use BLK API
>>> as directly as possible. So far so gut, but i have a problem by using SG
>>> requests with DMA without conversation. This controller accepts only
>>> addresses with this mask 0xffff.f000 - also page aligned.
>>> Till now i tried different QUEUE_FLAG_ flags and blk_queue settings, but
>>> i never had absolute aligned quarantine.
>>> Please tell me if i do some thing wrong or go in the wrong direction.
>>
>> Unless you can accept a 4kb hardware block size (and I'm assuming you
>> cannot, since you need to support 512b file systems?), then you'd have
>> to bounce the requests that don't align properly. You can fix some of
>> them by setting blk_queue_dma_alignment(q, 4095).
> 
> I needed to include MMC list with Ulf Hansson at the beginning, so here
> we are.
> 
> Hi Ulf,
> 
> here is the list of drivers which do care about dma alignment . Some of
> them use bounce_buffer, some will use PIO instead of DMA, for not
> aligned requests:
> 
> dw_mmc.c: dw_mci_pre_dma_transfer
> wbsd.c
> usdhi6rol0.c
> tmio_mmc_dma.c
> sunxi-mmc.c
> sdhci.c
> pxamci.c
> mvsdio.c
> au6601.c(not upstream jet)
> 
> The problem of all this driver is that each of them use own solution. I
> think this code can be moved to some common place. For example, it can
> be done in drivers/mmc/card/queue.c and enabled by
> mmc_host_dma_alignment variable. Or provide a mmc_sg_align() function,
> which should be called explicatively by driver.
> 
> Which option is preferable?

I would roll a block version that they all can use, you can look at
block/bounce.c for inspiration. That file deals with bouncing higher
pages to lower memory, so it's not exactly what you want. But the metod
of allocating a new bio that meets criteria and copying data over, that
is the same that you need.

-- 
Jens Axboe