Re: [PATCH] [RFC] dmaengine: add fifo_size member

[Dmitry Osipenko]

07.06.2019 16:35, Peter Ujfalusi ÐÐÑÐÑ:
> 
> 
> On 07/06/2019 15.58, Jon Hunter wrote:
>>> Imho if you can explain it without using 'HACK' in the sentences it
>>> might be OK, but it does not feel right.
>>
>> I don't perceive this as a hack. Although from looking at the
>> description of the src/dst_maxburst these are burst size with regard to
>> the device, so maybe it is a stretch.
>>
>>> However since your ADMA and ADMIF is highly coupled and it does needs
>>> special maxburst information (burst and allocated FIFO depth) I would
>>> rather use src_maxburst/dst_maxburst alone for DEV_TO_MEM/MEM_TO_DEV:
>>>
>>> ADMA_BURST_SIZE(maxburst)	((maxburst) & 0xff)
>>> ADMA_FIFO_SIZE(maxburst)	(((maxburst) >> 8) & 0xffffff)
>>>
>>> So lower 1 byte is the burst value you want from ADMA
>>> the other 3 bytes are the allocated FIFO size for the given ADMAIF channel.
>>>
>>> Sure, you need a header for this to make sure there is no
>>> misunderstanding between the two sides.
>>
>> I don't like this because as I mentioned to Dmitry, the ADMA can perform
>> memory-to-memory transfers where such encoding would not be applicable.
> 
> mem2mem does not really use dma_slave_config, it is for used by
> is_slave_direction() == true type of transfers.
> 
> But true, if you use ADMA against anything other than ADMAIF then this
> might be not right for non cyclic transfers.
> 
>> That does not align with the description in the
>> include/linux/dmaengine.h either.
> 
> True.
> 
>>> Or pass the allocated FIFO size via maxburst and then the ADMA driver
>>> will pick a 'good/safe' burst value for it.
>>>
>>> Or new member, but do you need two of them for src/dst? Probably
>>> fifo_depth is better word for it, or allocated_fifo_depth.
>>
>> Right, so looking at the struct dma_slave_config we have ...
>>
>> u32 src_maxburst;
>> u32 dst_maxburst;
>> u32 src_port_window_size;
>> u32 dst_port_window_size;
>>
>> Now if we could make these window sizes a union like the following this
>> could work ...
>>
>> diff --git a/include/linux/dmaengine.h b/include/linux/dmaengine.h
>> index 8fcdee1c0cf9..851251263527 100644
>> --- a/include/linux/dmaengine.h
>> +++ b/include/linux/dmaengine.h
>> @@ -360,8 +360,14 @@ struct dma_slave_config {
>>         enum dma_slave_buswidth dst_addr_width;
>>         u32 src_maxburst;
>>         u32 dst_maxburst;
>> -       u32 src_port_window_size;
>> -       u32 dst_port_window_size;
>> +       union {
>> +               u32 port_window_size;
>> +               u32 port_fifo_size;
>> +       } src;
>> +       union {
>> +               u32 port_window_size;
>> +               u32 port_fifo_size;
>> +       } dst;
> 
> What if in the future someone will have a setup where they would need both?
> 
> So not sure. Your problems are coming from a split DMA setup where the
> two are highly coupled, but sits in a different place and need to be
> configured as one device.
> 
> I think xilinx_dma is facing with similar issues and they have a custom
> API to set parameters which does not fit or is peripheral specific:
> include/linux/dma/xilinx_dma.h
> 
> Not sure if that is an acceptable solution.

If there are no other drivers with the exactly same requirement, then
the custom API is an a good variant given that there is a precedent
already. It is always possible to convert to a common thing later on
since that's all internal to kernel.

Jon / Sameer, you should check all the other drivers thoroughly to find
anyone who is doing the same thing as you need in order to achieve
something that is really common. I'm also wondering if it will be
possible to make dma_slave_config more flexible in order to start
accepting vendor specific properties in a somewhat common fashion, maybe
Vinod and Dan already have some thoughts on it? Apparently there is
already a need for the customization and people are just starting to
invent their own thing, but maybe that's fine too. That's really up to
subsys maintainer to decide in what direction to steer.