Re: [PATCH v2] dmaengine: bcm2835: Add slave dma support

[Jonathan Bell <jonathan@raspberrypi.org>]

On 06/05/2015 20:21, Noralf Trønnes wrote:
>
> Den 01.05.2015 23:07, skrev Martin Sperl:
>> Tests with the initial (and incomplete) version of the spi-bcm2835 
>> driver
>> with DMA transfer support show that the dma-engine works as expected 
>> with
>> this patch.
>>
>> There is one one observation:
>>
>>> On 18.04.2015, at 13:06, Noralf Trønnes <noralf@tronnes.org> wrote:
>>> +static struct dma_async_tx_descriptor *
>>> +bcm2835_dma_prep_slave_sg(struct dma_chan *chan,
>>> +              struct scatterlist *sgl,
>>> +              unsigned int sg_len,
>>> +              enum dma_transfer_direction direction,
>>> +              unsigned long flags, void *context)
>>> +{
>> ...
>>> +            /* Enable */
>>> +            if (i == sg_len - 1 && len - j <= max_size)
>>> +                control_block->info |= BCM2835_DMA_INT_EN;
>> The observation is that an interrupt is always triggered - even in 
>> the case
>> where flags does NOT have DMA_PREP_INTERRUPT set.
>> This may not be necessary and avoid interrupts.
>>
>> So maybe the above if clause should get extended by:
>>     && (flags & DMA_PREP_INTERRUPT)
>> to only trigger an interrupt when really requested.
>>
>> I am not sure if there are any side-effects because of this besides 
>> having the
>> requirement on the client to run dmaengine_terminate_all() on that 
>> specific dma
>> channel without interrupts when the transfer is finished.
>>
>> In the case of SPI we have TX feed the fifo - which finishes early - 
>> , but we
>> only need to the interrupt when RX finishes reading the fifo, which 
>> indicates
>> that the SPI-transfer is fully finished.
>> So having an interrupt on TX is not necessary for the process.
>>
>> The same observations may also apply to bcm2835_dma_prep_dma_cyclic 
>> (which is
>> outside of this patch provided by Noralf).
>
> Jonathan, can you comment on this?
>
>
> Noralf.
>
>
I agree that the interrupt generated would be spurious - in the case 
where it is not required.

However if you do && (flags & DMA_PREP_INTERRUPT) then all users of this 
driver need to explicitly set interrupt flags when doing a 
scatter-gather transfer. As I understand it, currently the only upstream 
client of this driver is the I2S driver which only uses cyclic anyway.

Not requiring an interrupt on completion is a bit of an edge case - the 
default among other dmaengine drivers appears to be to enable interrupts 
unconditionally.