Re: [PATCH v3] dmaengine: tegra-apb: Support per-burst residue granularity

[Dmitry Osipenko]

02.07.2019 14:37, Dmitry Osipenko ÐÐÑÐÑ:
> 02.07.2019 14:20, Jon Hunter ÐÐÑÐÑ:
>>
>> On 27/06/2019 20:47, Dmitry Osipenko wrote:
>>> Tegra's APB DMA engine updates words counter after each transferred burst
>>> of data, hence it can report transfer's residual with more fidelity which
>>> may be required in cases like audio playback. In particular this fixes
>>> audio stuttering during playback in a chromium web browser. The patch is
>>> based on the original work that was made by Ben Dooks and a patch from
>>> downstream kernel. It was tested on Tegra20 and Tegra30 devices.
>>>
>>> Link: https://lore.kernel.org/lkml/20190424162348.23692-1-ben.dooks@xxxxxxxxxxxxxxx/
>>> Link: https://nv-tegra.nvidia.com/gitweb/?p=linux-4.4.git;a=commit;h=c7bba40c6846fbf3eaad35c4472dcc7d8bbc02e5
>>> Inspired-by: Ben Dooks <ben.dooks@xxxxxxxxxxxxxxx>
>>> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx>
>>> ---
>>>
>>> Changelog:
>>>
>>> v3:  Added workaround for a hardware design shortcoming that results
>>>      in a words counter wraparound before end-of-transfer bit is set
>>>      in a cyclic mode.
>>>
>>> v2:  Addressed review comments made by Jon Hunter to v1. We won't try
>>>      to get words count if dma_desc is on free list as it will result
>>>      in a NULL dereference because this case wasn't handled properly.
>>>
>>>      The residual value is now updated properly, avoiding potential
>>>      integer overflow by adding the "bytes" to the "bytes_transferred"
>>>      instead of the subtraction.
>>>
>>>  drivers/dma/tegra20-apb-dma.c | 69 +++++++++++++++++++++++++++++++----
>>>  1 file changed, 62 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/drivers/dma/tegra20-apb-dma.c b/drivers/dma/tegra20-apb-dma.c
>>> index 79e9593815f1..71473eda28ee 100644
>>> --- a/drivers/dma/tegra20-apb-dma.c
>>> +++ b/drivers/dma/tegra20-apb-dma.c
>>> @@ -152,6 +152,7 @@ struct tegra_dma_sg_req {
>>>  	bool				last_sg;
>>>  	struct list_head		node;
>>>  	struct tegra_dma_desc		*dma_desc;
>>> +	unsigned int			words_xferred;
>>>  };
>>>  
>>>  /*
>>> @@ -496,6 +497,7 @@ static void tegra_dma_configure_for_next(struct tegra_dma_channel *tdc,
>>>  	tdc_write(tdc, TEGRA_APBDMA_CHAN_CSR,
>>>  				nsg_req->ch_regs.csr | TEGRA_APBDMA_CSR_ENB);
>>>  	nsg_req->configured = true;
>>> +	nsg_req->words_xferred = 0;
>>>  
>>>  	tegra_dma_resume(tdc);
>>>  }
>>> @@ -511,6 +513,7 @@ static void tdc_start_head_req(struct tegra_dma_channel *tdc)
>>>  					typeof(*sg_req), node);
>>>  	tegra_dma_start(tdc, sg_req);
>>>  	sg_req->configured = true;
>>> +	sg_req->words_xferred = 0;
>>>  	tdc->busy = true;
>>>  }
>>>  
>>> @@ -797,6 +800,61 @@ static int tegra_dma_terminate_all(struct dma_chan *dc)
>>>  	return 0;
>>>  }
>>>  
>>> +static unsigned int tegra_dma_sg_bytes_xferred(struct tegra_dma_channel *tdc,
>>> +					       struct tegra_dma_sg_req *sg_req)
>>> +{
>>> +	unsigned long status, wcount = 0;
>>> +
>>> +	if (!list_is_first(&sg_req->node, &tdc->pending_sg_req))
>>> +		return 0;
>>> +
>>> +	if (tdc->tdma->chip_data->support_separate_wcount_reg)
>>> +		wcount = tdc_read(tdc, TEGRA_APBDMA_CHAN_WORD_TRANSFER);
>>> +
>>> +	status = tdc_read(tdc, TEGRA_APBDMA_CHAN_STATUS);
>>> +
>>> +	if (!tdc->tdma->chip_data->support_separate_wcount_reg)
>>> +		wcount = status;
>>> +
>>> +	if (status & TEGRA_APBDMA_STATUS_ISE_EOC)
>>> +		return sg_req->req_len;
>>> +
>>> +	wcount = get_current_xferred_count(tdc, sg_req, wcount);
>>> +
>>> +	if (!wcount) {
>>> +		/*
>>> +		 * If wcount wasn't ever polled for this SG before, then
>>> +		 * simply assume that transfer hasn't started yet.
>>> +		 *
>>> +		 * Otherwise it's the end of the transfer.
>>> +		 *
>>> +		 * The alternative would be to poll the status register
>>> +		 * until EOC bit is set or wcount goes UP. That's so
>>> +		 * because EOC bit is getting set only after the last
>>> +		 * burst's completion and counter is less than the actual
>>> +		 * transfer size by 4 bytes. The counter value wraps around
>>> +		 * in a cyclic mode before EOC is set(!), so we can't easily
>>> +		 * distinguish start of transfer from its end.
>>> +		 */
>>> +		if (sg_req->words_xferred)
>>> +			wcount = sg_req->req_len - 4;
>>> +
>>> +	} else if (wcount < sg_req->words_xferred) {
>>> +		/*
>>> +		 * This case shall not ever happen because EOC bit
>>> +		 * must be set once next cyclic transfer is started.
>>
>> I am not sure I follow this and why this condition cannot happen for
>> cyclic transfers. What about non-cyclic transfers?
> 
> It cannot happen because the EOC bit will be set in that case. The counter wraps
> around when the transfer of a last burst happens, EOC bit is guaranteed to be set
> after completion of the last burst. That's my observation after a thorough testing,
> it will be very odd if EOC setting happened completely asynchronously.
> 
> For a non-cyclic transfers it doesn't matter.. because they are not cyclic and thus
> counter will be stopped by itself. It will be a disaster if all of sudden a
> non-cyclic transfer becomes cyclic, don't you think so? :)
> 

Ah, probably I was too focused on audio playback use-case. If it's a free-running
transfer, then that case of a wraparound seems should be legit.. hmm.