[PATCH] dmaengine: delete redundant chan_id and chancnt initialization in dma drivers

Jassi Brar jassisinghbrar at gmail.com
Thu Sep 22 01:41:47 EDT 2011

On Thu, Sep 22, 2011 at 5:38 AM, Barry Song <21cnbao at gmail.com> wrote:
> 2011/9/21 Jassi Brar <jassisinghbrar at gmail.com>:
>> On Fri, Sep 16, 2011 at 3:13 PM, Barry Song <Baohua.Song at csr.com> wrote:
>>> dma_async_device_register will re-init chan_id and chancnt,
>>> so whatever chan_id and chancnt are set in drivers, they will
>>> be re-written by dma_async_device_register.
>>> Cc: Nicolas Ferre <nicolas.ferre at atmel.com>
>>> Cc: Viresh Kumar <viresh.kumar at st.com>
>>> Cc: Vinod Koul <vinod.koul at intel.com>
>>> Cc: Piotr Ziecik <kosmo at semihalf.com>
>>> Cc: Yong Wang <yong.y.wang at intel.com>
>>> Cc: Jaswinder Singh <jassi.brar at samsung.com>
>>> Cc: Pelagicore AB <info at pelagicore.com>
>>> Signed-off-by: Barry Song <Baohua.Song at csr.com>
>>> ---
>>>  drivers/dma/at_hdmac.c      |    5 ++---
>>>  drivers/dma/dw_dmac.c       |    5 ++---
>>>  drivers/dma/intel_mid_dma.c |    2 --
>>>  drivers/dma/mpc512x_dma.c   |    1 -
>>>  drivers/dma/pch_dma.c       |    2 --
>>>  drivers/dma/pl330.c         |    2 --
>>>  drivers/dma/timb_dma.c      |    3 +--
>> Apparently ....
>> drivers/dma/ppc4xx/adma.c
>> drivers/dma/ipu/ipu_idmac.c
>>       still write to chan_id.
>> drivers/dma/amba-pl08x.c
>> drivers/dma/fsldma.c
>> drivers/dma/ioat/dma_v2.c
>> drivers/dma/ioat/dma.c
>> drivers/dma/mpc512x_dma.c
>> drivers/dma/shdma.c
>>       still write to chancnt
My question was if you really intended to convert only some of
the drivers and not others?

>> Most of them are simply a matter of removal, but some seem
>> like really depending upon setting them(?)
>> Anyways, even after you modify those as well, chan_id and chancnt
>> are rendered overstaying guests in dmaengine core. Because chan_id
>> of each channel would be _precisely_ the order in which the
>> _dmac-driver_ added the channel to the 'channels' list.
>> So if their values are _always_ gonna be just contiguously incrementing
>> why need variables for that in the dmaengine api?
>> Dmac drivers could use local variables for that.
>> OTOH, why not chan_id be left solely for use by dmac drivers read by the
>> dmaengine only to create sysfs entries ?
> actually i don't like the way dmaengine core handles chan_id. in my
> opinion, every dmac only needs to tell dmaengine core the chan_base.
> then dmaengine core set the id in a global scale.
> for example, if you have two dmac in system:
> dmac0 chan_base 0
> dmac1 chan_base 16
> then core will have chan 0~31 by chan_id from 0 to 31 but not
> dmac0chan0~15, dmac1chan0~15.
But note the numbering is still decided by the dmac driver - directly
or indirectly.
That is, if the dmac driver reorder registering of dmac or adding
channels to the 'channels' list, the assigned chan_id's would change.
So why not rather get rid of chan_id assignment from dmaengine
and do explicitly in dmac drivers ?
Which I already submitted a patch for https://lkml.org/lkml/2011/7/21/1

> client drivers can request a specific chan_id just like it can request a gpio.
No dear. That's not gonna fly if we are to have common client drivers
across SoCs.

> now dmac drivers and dmaengine core have repeated chan_id and chancnt,
> anyway, it is a problem that should be fixed at first.
Correct, chan_id assignment should be completely either in dmaengine
or in dmac drivers. But this patch does neither.

More information about the linux-arm-kernel mailing list