Re: [PATCH] dma: edma: add device_slave_caps() support

[Lars-Peter Clausen <lars@metafoo.de>]

On 07/24/2013 08:55 PM, Joel Fernandes wrote:
> On 07/24/2013 03:40 AM, Lars-Peter Clausen wrote:
>> On 07/24/2013 10:28 AM, Fernandes, Joel wrote:
>>>
>>> On Jul 24, 2013, at 3:23 AM, "Lars-Peter Clausen" <lars@metafoo.de> wrote:
>>>
>>>> On 07/24/2013 10:11 AM, Joel Fernandes wrote:
>>>>> On 07/24/2013 03:03 AM, Lars-Peter Clausen wrote:
>>>>>> On 07/23/2013 06:43 PM, Joel Fernandes wrote:
>>>>>>> Implement device_slave_caps(). EDMA has a limited number of slots.
>>>>>>> Slave drivers such as omap_hsmmc will query the driver to make
>>>>>>> sure they don't pass in more than these many scatter segments.
>>>>>>>
>>>>>>> Signed-off-by: Joel Fernandes <joelf@ti.com>
>>>>>>> ---
>>>>>>> Vinod, or Dan- If this patch looks ok, can you please merge in for
>>>>>>> -rc cycle? This patch is required to fix MMC support on AM33xx. This
>>>>>>> patch is blocking 3 other patches which fix various MMC things. Thanks!
>>>>>>>
>>>>>>> Notes:
>>>>>>> (1) this approach is temporary and only for -rc cycle to fix MMC on
>>>>>>> AM335x. It will be replace by the RFC series in future kernels:
>>>>>>> http://www.spinics.net/lists/arm-kernel/msg260094.html
>>>>>>>
>>>>>>> (2) Patch depends Vinod's patch at:
>>>>>>> http://permalink.gmane.org/gmane.linux.kernel/1525112
>>>>>>>
>>>>>>> drivers/dma/edma.c |    9 +++++++++
>>>>>>> 1 file changed, 9 insertions(+)
>>>>>>>
>>>>>>> diff --git a/drivers/dma/edma.c b/drivers/dma/edma.c
>>>>>>> index 7222cbe..81d5429 100644
>>>>>>> --- a/drivers/dma/edma.c
>>>>>>> +++ b/drivers/dma/edma.c
>>>>>>> @@ -517,6 +517,14 @@ static void edma_issue_pending(struct dma_chan *chan)
>>>>>>>     spin_unlock_irqrestore(&echan->vchan.lock, flags);
>>>>>>> }
>>>>>>>
>>>>>>> +static inline int edma_slave_caps(struct dma_chan *chan,
>>>>>>> +    struct dma_slave_caps *caps)
>>>>>>> +{
>>>>>>> +    caps->max_sg_nr = MAX_NR_SG;
>>>>>>
>>>>>> Hm, what about the other fields?
>>>>>
>>>>> Other fields are unused, the max segment size is supposed to be
>>>>> calculated "given" the address width and burst size. Since these
>>>>> can't be provided to get_caps, I have left it out for now.
>>>>> See: https://lkml.org/lkml/2013/3/6/464
>>>>
>>>> The PL330 driver is similar in this regard, the maximum segment size also
>>>> depends on address width and burst width. What I did for the get_slave_caps
>>>> implementation is to set it to the minimum maximum size. E.g. in you case
>>>> that should be SZ_64K - 1 (burstsize and addrwidth both set to 1).
>>>
>>> So you're setting max to minimum maximum size? Isn't that like telling the driver that its segments can't be bigger than this... Unless I'm missing something..
>>
>> Yes. This is a limitation of the current slave_caps API. The maximum needs
>> to be the maximum for all possible configurations. A specific configuration
>> may allow a larger maximum. So we maybe have to extend the API to be able to
>> query the limits for a certain configuration. Not sure what the best way
>> would be to do that, either adding a config parameter to get_slave_caps or
>> to break it into two functions like you proposed one for the static
>> capabilities and one for the sg limits.
>
> I am OK with either approach as long as a decision can be made quickly
> by maintainers. Right now lot of back and forth has happened and 3
> different versions of the same thing have been posted since January.
> Since this is such a trivial change, it doesn't make sense to spend so
> much time on it IMO.... The sad part is though this change is trivial,
> other drivers such as MMC are broken and cannot be enabled due to this.
> We cannot afford to leave them broken.

Well this is a new API, so it is kind of expected that there is some back and 
forth and that there will be a few revisions.

>
> If Vinod is not available, can Dan please respond on how to proceed on
> this? We really need this trivial change to go into this -rc cycle and
> not delay it by another kernel release. Thank you.

This is not something you'd merge for rc3 or even later. If the MMC driver does 
not work without this I guess it never worked, so strictly speaking there is no 
regression and it is just a new feature.

- Lars