Re: [PATCH v6 11/16] dmaengine: tegra-apb: Keep clock enabled only during of DMA transfer

[Jon Hunter <jonathanh@nvidia.com>]

On 30/01/2020 16:11, Dmitry Osipenko wrote:
> 30.01.2020 17:09, Jon Hunter пишет:
>>
>> On 30/01/2020 04:37, Dmitry Osipenko wrote:
>>> It's a bit impractical to enable hardware's clock at the time of DMA
>>> channel's allocation because most of DMA client drivers allocate DMA
>>> channel at the time of the driver's probing, and thus, DMA clock is kept
>>> always-enabled in practice, defeating the whole purpose of runtime PM.
>>>
>>> Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
>>> ---
>>>  drivers/dma/tegra20-apb-dma.c | 47 ++++++++++++++++++++++++-----------
>>>  1 file changed, 32 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/drivers/dma/tegra20-apb-dma.c b/drivers/dma/tegra20-apb-dma.c
>>> index 22b88ccff05d..0ee28d8e3c96 100644
>>> --- a/drivers/dma/tegra20-apb-dma.c
>>> +++ b/drivers/dma/tegra20-apb-dma.c
>>> @@ -436,6 +436,8 @@ static void tegra_dma_stop(struct tegra_dma_channel *tdc)
>>>  		tdc_write(tdc, TEGRA_APBDMA_CHAN_STATUS, status);
>>>  	}
>>>  	tdc->busy = false;
>>> +
>>> +	pm_runtime_put(tdc->tdma->dev);
>>>  }
>>>  
>>>  static void tegra_dma_start(struct tegra_dma_channel *tdc,
>>> @@ -500,18 +502,25 @@ static void tegra_dma_configure_for_next(struct tegra_dma_channel *tdc,
>>>  	tegra_dma_resume(tdc);
>>>  }
>>>  
>>> -static void tdc_start_head_req(struct tegra_dma_channel *tdc)
>>> +static bool tdc_start_head_req(struct tegra_dma_channel *tdc)
>>>  {
>>>  	struct tegra_dma_sg_req *sg_req;
>>> +	int err;
>>>  
>>>  	if (list_empty(&tdc->pending_sg_req))
>>> -		return;
>>> +		return false;
>>> +
>>> +	err = pm_runtime_get_sync(tdc->tdma->dev);
>>> +	if (WARN_ON_ONCE(err < 0))
>>> +		return false;
>>>  
>>>  	sg_req = list_first_entry(&tdc->pending_sg_req, typeof(*sg_req), node);
>>>  	tegra_dma_start(tdc, sg_req);
>>>  	sg_req->configured = true;
>>>  	sg_req->words_xferred = 0;
>>>  	tdc->busy = true;
>>> +
>>> +	return true;
>>>  }
>>>  
>>>  static void tdc_configure_next_head_desc(struct tegra_dma_channel *tdc)
>>> @@ -615,6 +624,8 @@ static void handle_once_dma_done(struct tegra_dma_channel *tdc,
>>>  	}
>>>  	list_add_tail(&sgreq->node, &tdc->free_sg_req);
>>>  
>>> +	pm_runtime_put(tdc->tdma->dev);
>>> +
>>>  	/* Do not start DMA if it is going to be terminate */
>>>  	if (to_terminate || list_empty(&tdc->pending_sg_req))
>>>  		return;
>>> @@ -730,9 +741,7 @@ static void tegra_dma_issue_pending(struct dma_chan *dc)
>>>  		dev_err(tdc2dev(tdc), "No DMA request\n");
>>>  		goto end;
>>>  	}
>>> -	if (!tdc->busy) {
>>> -		tdc_start_head_req(tdc);
>>> -
>>> +	if (!tdc->busy && tdc_start_head_req(tdc)) {
>>>  		/* Continuous single mode: Configure next req */
>>>  		if (tdc->cyclic) {
>>>  			/*
>>> @@ -775,6 +784,13 @@ static int tegra_dma_terminate_all(struct dma_chan *dc)
>>>  	else
>>>  		wcount = status;
>>>  
>>> +	/*
>>> +	 * tegra_dma_stop() will drop the RPM's usage refcount, but
>>> +	 * tegra_dma_resume() touches hardware and thus we should keep
>>> +	 * the DMA clock active while it's needed.
>>> +	 */
>>> +	pm_runtime_get(tdc->tdma->dev);
>>> +
>>
>> Would it work and make it simpler to just enable in the issue_pending
>> and disable in the handle_once_dma_done or terminate_all?
>>
>> diff --git a/drivers/dma/tegra20-apb-dma.c b/drivers/dma/tegra20-apb-dma.c
>> index 3a45079d11ec..86bbb45da93d 100644
>> --- a/drivers/dma/tegra20-apb-dma.c
>> +++ b/drivers/dma/tegra20-apb-dma.c
>> @@ -616,9 +616,14 @@ static void handle_once_dma_done(struct
>> tegra_dma_channel *tdc,
>>         list_add_tail(&sgreq->node, &tdc->free_sg_req);
>>
>>         /* Do not start DMA if it is going to be terminate */
>> -       if (to_terminate || list_empty(&tdc->pending_sg_req))
>> +       if (to_terminate)
>>                 return;
>>
>> +       if (list_empty(&tdc->pending_sg_req)) {
>> +               pm_runtime_put(tdc->tdma->dev);
>> +               return;
>> +       }
>> +
>>         tdc_start_head_req(tdc);
>>  }
>>
>> @@ -729,6 +734,11 @@ static void tegra_dma_issue_pending(struct dma_chan
>> *dc)
>>                 goto end;
>>         }
>>         if (!tdc->busy) {
>> +               if (pm_runtime_get_sync(tdc->tdma->dev) < 0) {
>> +                       dev_err(tdc2dev(tdc), "Failed to enable DMA!\n");
>> +                       goto end;
>> +               }
>> +
>>                 tdc_start_head_req(tdc);
>>
>>                 /* Continuous single mode: Configure next req */
>> @@ -788,6 +798,7 @@ static int tegra_dma_terminate_all(struct dma_chan *dc)
>>                                 get_current_xferred_count(tdc, sgreq,
>> wcount);
>>         }
>>         tegra_dma_resume(tdc);
>> +       pm_runtime_put(tdc->tdma->dev);
>>
>>  skip_dma_stop:
>>         tegra_dma_abort_all(tdc);
>>
> 
> The tegra_dma_stop() should put RPM anyways, which is missed in yours
> sample. Please see handle_continuous_head_request().

Yes and that is deliberate. The cyclic transfers the transfers *should*
not stop until terminate_all is called. The tegra_dma_stop in
handle_continuous_head_request() is an error condition and so I am not
sure it is actually necessary to call pm_runtime_put() here.

> I'm also finding the explicit get/put a bit easier to follow in the
> code, don't you think so?

I can see that, but I was thinking that in the case of cyclic transfers,
it should only really be necessary to call the get/put at the beginning
and end. So in my mind there should only be two exit points which are
the ISR handler for SG and terminate_all for SG and cyclic.

Jon

-- 
nvpublic