Re: Requirement for at91-sama5d2_adc timestamp buffer

[<Eugen.Hristev@microchip.com>]

On 01.11.2020 22:21, Jonathan Cameron wrote:
> On Sun, 1 Nov 2020 11:28:53 +0200
> Alexandru Ardelean <ardeleanalex@gmail.com> wrote:
> 
>> On Fri, Oct 30, 2020 at 2:29 PM <Eugen.Hristev@microchip.com> wrote:
>>>
>>> Hello Jonathan,
>>>
>>> I found an issue with at91-sama5d2_adc driver, namely, when using DMA
>>> and timestamp in the same time, the hardware provides the sample in the
>>> fashion (2 bytes per channel) * (number of channels) - as copied by the
>>> DMA master to memory.
>>> I compute a software timestamp , and then push to buffers with timestamp.
>>> However your push code will try to write this timestamp inside my buffer
>>> ! and overwrite my samples... I have multiple samples in the buffer
>>> (watermark number) and there is no space between them because the
>>> hardware copies the conversion data directly in this buffer.
>>>
>>> Do you have any suggestion on how to solve this, except 1) giving up the
>>> timestamp in this mode or 2) copy to another buffer with more space for
>>> timestamp storage ?
>>
>> I'm assuming the issue is here:
>>                  iio_push_to_buffers_with_timestamp(indio_dev,
>>                                  (st->dma_st.rx_buf + st->dma_st.buf_idx),
>>                                  (st->dma_st.dma_ts + interval * sample_index));
>>
>> Can the DMA be configured to add some padding in-between the samples?
>> It looks like the way this is currently working, timestamps cannot
>> work with the DMA buffers and multiple consecutive samples.
>> But, it may be that this case is a bit unrealistic; or shouldn't be supported.
>> DMA is used to provide really-fast transfers; computing timestamps in
>> SW for each sample would slow things down to the point where the
>> transfers aren't fast anymore.
>>
>> What would [possibly] be an alternative, is to do a
>> "iio_push_multiple_samples_to_buffers_with_timestamp(indio_dev, buffer
>> for 1 sample-set, n_samples, )".
>> That would basically mean, the DMA gets 10, 100, 1000 samples, and
>> adds a timestamp at the end.
>> Now, the only thing that I don't know here: is how userspace would be
>> able to determine the number of samples until the next timestamp.
>> I guess some mechanism could be extended inside IIO to accommodate for
>> this; N_samples_till_timestamp counter, with a default value of 1.
>> Not sure if it makes sense though.
> 
> There is a very obscure and little used facility to repeats of a single
> channel, but that's as close as we get.  Describing arbitrary patterns
> of data is hard to do without adding the overhead of a header and not
> doing tagged data has been one of the fundamental design decisions in
> IIO.
> 
> We've discussed adding a separate meta data buffer, but that would
> need extra storage anyway to flag when there was meta data so wouldn't
> help you here.
> 
> As Alexandru asked, can you persuade the dma engine to leave a gap?

This might work for a single channel. But if we use arbitrary number of 
channels, only the iio driver knows how many channels are enabled and 
how to separate the data into samples, by dividing total dma received 
data by the size of one sample
So the DMA channel is just copying data from the ADC as soon as it's 
ready, and then in the iio driver I am just passing the buffer to the 
subsystem. Thus it's pretty unaware to leave a gap when the channels are 
done for one sample.

> If not, you are going to have to keep copying the data.
> I theory you might be able to persuade the kfifo to take that
> data in different form and then introduce a different path for
> buffer management (not iio_push_to_buffers_*), but you'd need to
> figure out how to do that.

Currently, I hold two timestamps (one at DMA start and one at DMA 
finish), and then spread the time between them by the number of samples 
(Jonathan's idea when I initially added the DMA support). I could make 
another primitive that would receive a timestamp for DMA start , and one 
for DMA end, but this would mean that the memcopy would be done inside 
the iio subsystem (which is again a performance blocker)
Is there a way to have a channel enabled, but this channel would provide 
data not every time so to say, for example, once every X samples ?
I could alter the scan_mask to remove the timestamp channel and then 
reenable the timestamp and provide timestamp information at the end of 
one DMA chunk ?
Does this sound as a good solution ?

Thanks,
Eugen
>>
>>>
>>> Thanks,
>>> Eugen
>