Re: [PATCH RFC 1/4] dma: imx-sdma - reduce transfer latency for DMA cyclic clients

[Peter Senna Tschudin]

On Thu, Jun 09, 2016 at 03:16:30PM +0300, Nandor Han wrote:
> Having the SDMA driver use a tasklet for running the clients
> callback introduce some issues:
>   - probability to have desynchronized data because of the
>     race condition created since the DMA transaction status
>     is retrieved only when the callback is executed, leaving
>     plenty of time for transaction status to get altered.
>   - inter-transfer latency which can leave channels idle.
> 
> Move the callback execution, for cyclic channels, to SDMA
> interrupt (as advised in `Documentation/dmaengine/provider.txt`)
> to (a)reduce the inter-transfer latency and (b) eliminate the
> race condition possibility where DMA transaction status might
> be changed by the time is read.
> 
> The responsibility of the SDMA interrupt latency
> is moved to the SDMA clients which case by case should defer
> the work to bottom-halves when needed.
> 
> Signed-off-by: Nandor Han <nandor.han@xxxxxx>
For the whole series:
Tested-by: Peter Senna Tschudin <peter.senna@xxxxxxxxxxxxx>
Acked-by: Peter Senna Tschudin <peter.senna@xxxxxxxxxxxxx>
> ---
>  drivers/dma/imx-sdma.c | 36 ++++++++++++++++--------------------
>  1 file changed, 16 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c
> index 0f6fd42..e497847 100644
> --- a/drivers/dma/imx-sdma.c
> +++ b/drivers/dma/imx-sdma.c
> @@ -654,12 +654,6 @@ static void sdma_event_disable(struct sdma_channel *sdmac, unsigned int event)
>  	writel_relaxed(val, sdma->regs + chnenbl);
>  }
>  
> -static void sdma_handle_channel_loop(struct sdma_channel *sdmac)
> -{
> -	if (sdmac->desc.callback)
> -		sdmac->desc.callback(sdmac->desc.callback_param);
> -}
> -
>  static void sdma_update_channel_loop(struct sdma_channel *sdmac)
>  {
>  	struct sdma_buffer_descriptor *bd;
> @@ -678,13 +672,25 @@ static void sdma_update_channel_loop(struct sdma_channel *sdmac)
>  			sdmac->status = DMA_ERROR;
>  
>  		bd->mode.status |= BD_DONE;
> +
> +		/*
> +		 * The callback is called from the interrupt context in order
> +		 * to reduce latency and to avoid the risk of altering the
> +		 * SDMA transaction status by the time the client tasklet is
> +		 * executed.
> +		 */
> +
> +		if (sdmac->desc.callback)
> +			sdmac->desc.callback(sdmac->desc.callback_param);
> +
>  		sdmac->buf_tail++;
>  		sdmac->buf_tail %= sdmac->num_bd;
>  	}
>  }
>  
> -static void mxc_sdma_handle_channel_normal(struct sdma_channel *sdmac)
> +static void mxc_sdma_handle_channel_normal(unsigned long data)
>  {
> +	struct sdma_channel *sdmac = (struct sdma_channel *) data;
>  	struct sdma_buffer_descriptor *bd;
>  	int i, error = 0;
>  
> @@ -711,16 +717,6 @@ static void mxc_sdma_handle_channel_normal(struct sdma_channel *sdmac)
>  		sdmac->desc.callback(sdmac->desc.callback_param);
>  }
>  
> -static void sdma_tasklet(unsigned long data)
> -{
> -	struct sdma_channel *sdmac = (struct sdma_channel *) data;
> -
> -	if (sdmac->flags & IMX_DMA_SG_LOOP)
> -		sdma_handle_channel_loop(sdmac);
> -	else
> -		mxc_sdma_handle_channel_normal(sdmac);
> -}
> -
>  static irqreturn_t sdma_int_handler(int irq, void *dev_id)
>  {
>  	struct sdma_engine *sdma = dev_id;
> @@ -737,8 +733,8 @@ static irqreturn_t sdma_int_handler(int irq, void *dev_id)
>  
>  		if (sdmac->flags & IMX_DMA_SG_LOOP)
>  			sdma_update_channel_loop(sdmac);
> -
> -		tasklet_schedule(&sdmac->tasklet);
> +		else
> +			tasklet_schedule(&sdmac->tasklet);
>  
>  		__clear_bit(channel, &stat);
>  	}
> @@ -1739,7 +1735,7 @@ static int sdma_probe(struct platform_device *pdev)
>  		dma_cookie_init(&sdmac->chan);
>  		sdmac->channel = i;
>  
> -		tasklet_init(&sdmac->tasklet, sdma_tasklet,
> +		tasklet_init(&sdmac->tasklet, mxc_sdma_handle_channel_normal,
>  			     (unsigned long) sdmac);
>  		/*
>  		 * Add the channel to the DMAC list. Do not add channel 0 though
> -- 
> 2.8.3
> 
>