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

[Dmitry Osipenko]

06.06.2019 17:25, Jon Hunter ÐÐÑÐÑ:
> 
> 
> On 06/06/2019 14:45, Dmitry Osipenko wrote:
>> 06.06.2019 15:37, Jon Hunter ÐÐÑÐÑ:
>>>
>>> On 06/06/2019 12:54, Peter Ujfalusi wrote:
>>>>
>>>>
>>>> On 06/06/2019 13.49, Jon Hunter wrote:
>>>>>
>>>>> On 06/06/2019 11:22, Peter Ujfalusi wrote:
>>>>>
>>>>> ...
>>>>>
>>>>>>>>> It does sounds like that FIFO_SIZE == src/dst_maxburst in your case as
>>>>>>>>> well.
>>>>>>>> Not exactly equal.
>>>>>>>> ADMA burst_size can range from 1(WORD) to 16(WORDS)
>>>>>>>> FIFO_SIZE can be adjusted from 16(WORDS) to 1024(WORDS) [can vary in
>>>>>>>> multiples of 16]
>>>>>>>
>>>>>>> So I think that the key thing to highlight here, is that the as Sameer
>>>>>>> highlighted above for the Tegra ADMA there are two values that need to
>>>>>>> be programmed; the DMA client FIFO size and the max burst size. The ADMA
>>>>>>> has register fields for both of these.
>>>>>>
>>>>>> How does the ADMA uses the 'client FIFO size' and 'max burst size'
>>>>>> values and what is the relation of these values to the peripheral side
>>>>>> (ADMAIF)?
>>>>>
>>>>> Per Sameer's previous comment, the FIFO size is used by the ADMA to
>>>>> determine how much space is available in the FIFO. I assume the burst
>>>>> size just limits how much data is transferred per transaction.
>>>>>
>>>>>>> As you can see from the above the FIFO size can be much greater than the
>>>>>>> burst size and so ideally both of these values would be passed to the DMA.
>>>>>>>
>>>>>>> We could get by with just passing the FIFO size (as the max burst size)
>>>>>>> and then have the DMA driver set the max burst size depending on this,
>>>>>>> but this does feel quite correct for this DMA. Hence, ideally, we would
>>>>>>> like to pass both.
>>>>>>>
>>>>>>> We are also open to other ideas.
>>>>>>
>>>>>> I can not find public documentation (I think they are walled off by
>>>>>> registration), but correct me if I'm wrong:
>>>>>
>>>>> No unfortunately, you are not wrong here :-(
>>>>>
>>>>>> ADMAIF - peripheral side
>>>>>>  - kind of a small DMA for audio preipheral(s)?
>>>>>
>>>>> Yes this is the interface to the APE (audio processing engine) and data
>>>>> sent to the ADMAIF is then sent across a crossbar to one of many
>>>>> devices/interfaces (I2S, DMIC, etc). Basically a large mux that is user
>>>>> configurable depending on the use-case.
>>>>>
>>>>>>  - Variable FIFO size
>>>>>
>>>>> Yes.
>>>>>
>>>>>>  - sends DMA request to ADMA per words
>>>>>
>>>>> From Sameer's notes it says the ADMAIF send a signal to the ADMA per
>>>>> word, yes.
>>>>>
>>>>>> ADMA - system DMA
>>>>>>  - receives the DMA requests from ADMAIF
>>>>>>  - counts the requests
>>>>>>  - based on some threshold of the counter it will send/read from ADMAIF?
>>>>>>   - maxburst number of words probably?
>>>>>
>>>>> Sounds about right to me.
>>>>>
>>>>>> ADMA needs to know the ADMAIF's FIFO size because, it is the one who is
>>>>>> managing that FIFO from the outside, making sure that it does not over
>>>>>> or underrun?
>>>>>
>>>>> Yes.
>>>>>
>>>>>> And it is the one who sets the pace (in effect the DMA burst size - how
>>>>>> many bytes the DMA jumps between refills) of refills to the ADMAIF's FIFO?
>>>>>
>>>>> Yes.
>>>>>
>>>>> So currently, if you look at the ADMA driver
>>>>> (drivers/dma/tegra210-adma.c) you will see we use the src/dst_maxburst
>>>>> for the burst, but the FIFO size is hard-coded (see the
>>>>> TEGRA210_FIFO_CTRL_DEFAULT and TEGRA186_FIFO_CTRL_DEFAULT definitions).
>>>>> Ideally, we should not hard-code this but pass it.
>>>>
>>>> Sure, hardcoding is never good ;)
>>>>
>>>>> Given that there are no current users of the ADMA upstream, we could
>>>>> change the usage of the src/dst_maxburst, but being able to set the FIFO
>>>>> size as well would be ideal.
>>>>
>>>> Looking at the drivers/dma/tegra210-adma.c for the
>>>> TEGRA*_FIFO_CTRL_DEFAULT definition it is still not clear where the
>>>> remote FIFO size would fit.
>>>> There are fields for overflow and starvation(?) thresholds and TX/RX
>>>> size (assuming word length, 3 == 32bits?).
>>>
>>> The TX/RX size are the FIFO size. So 3 equates to a FIFO size of 3 * 64
>>> bytes.
>>>
>>>> Both threshold is set to one, so I assume currently ADMA is
>>>> pushing/pulling data word by word.
>>>
>>> That's different. That indicates thresholds when transfers start.
>>>
>>>> Not sure what the burst size is used for, my guess would be that it is
>>>> used on the memory (DDR) side for optimized, more efficient accesses?
>>>
>>> That is the actual burst size.
>>>
>>>> My guess is that the threshold values are the counter limits, if the DMA
>>>> request counter reaches it then ADMA would do a threshold limit worth of
>>>> push/pull to ADMAIF.
>>>> Or there is another register where the remote FIFO size can be written
>>>> and ADMA is counting back from there until it reaches the threshold (and
>>>> pushes/pulling again threshold amount of data) so it keeps the FIFO
>>>> filled with at least threshold amount of data?
>>>>
>>>> I think in both cases the threshold would be the maxburst.
>>>>
>>>> I suppose you have the patch for adma on how to use the fifo_size
>>>> parameter? That would help understand what you are trying to achieve better.
>>>
>>> Its quite simple, we would just use the FIFO size to set the fields
>>> TEGRAXXX_ADMA_CH_FIFO_CTRL_TXSIZE/RXSIZE in the
>>> TEGRAXXX_ADMA_CH_FIFO_CTRL register. That's all.
>>>
>>> Jon
>>>
>>
>> Hi,
>>
>> If I understood everything correctly, the FIFO buffer is shared among
>> all of the ADMA clients and hence it should be up to the ADMA driver to
>> manage the quotas of the clients. So if there is only one client that
>> uses ADMA at a time, then this client will get a whole FIFO buffer, but
>> once another client starts to use ADMA, then the ADMA driver will have
>> to reconfigure hardware to split the quotas.
> 
> The FIFO quotas are managed by the ADMAIF driver (does not exist in
> mainline currently but we are working to upstream this) because it is
> this device that owns and needs to configure the FIFOs. So it is really
> a means to pass the information from the ADMAIF to the ADMA.

So you'd want to reserve a larger FIFO for an audio channel that has a
higher audio rate since it will perform reads more often. You could also
prioritize one channel over the others, like in a case of audio call for
example.

Is the shared buffer smaller than may be needed by clients in a worst
case scenario? If you could split the quotas statically such that each
client won't ever starve, then seems there is no much need in the
dynamic configuration.