diff mbox

[1/1] dma: imx-sdma: add 1ms delay to ensure SDMA channel is stopped

Message ID 1486738005-4297-1-git-send-email-jiada_wang@mentor.com (mailing list archive)
State Changes Requested
Headers show

Commit Message

Wang, Jiada Feb. 10, 2017, 2:46 p.m. UTC 
From: Jiada Wang <jiada_wang@mentor.com>

sdma_disable_channel() cannot ensure dma is stopped to access
module's FIFOs. Maybe SDMA core is running and accessing BD when
disable of corresponding channel, this may cause sometimes even
after call of .sdma_disable_channel(), SDMA core still be running
and accessing module's FIFOs.

We should add delay of one BD SDMA cost time, the maximum is 1ms.
So that SDMA clients by calling .device_terminate_all can
ensure SDMA core has really been stopped.

Signed-off-by: Jiada Wang <jiada_wang@mentor.com>
---
 drivers/dma/imx-sdma.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

Comments

Vinod Koul Feb. 13, 2017, 2:05 a.m. UTC | #1 
On Fri, Feb 10, 2017 at 06:46:45AM -0800, jiada_wang@mentor.com wrote:
> From: Jiada Wang <jiada_wang@mentor.com>
> 
> sdma_disable_channel() cannot ensure dma is stopped to access
> module's FIFOs. Maybe SDMA core is running and accessing BD when
> disable of corresponding channel, this may cause sometimes even
> after call of .sdma_disable_channel(), SDMA core still be running
> and accessing module's FIFOs.
> 
> We should add delay of one BD SDMA cost time, the maximum is 1ms.
> So that SDMA clients by calling .device_terminate_all can
> ensure SDMA core has really been stopped.
> 
> Signed-off-by: Jiada Wang <jiada_wang@mentor.com>
> ---
>  drivers/dma/imx-sdma.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c
> index d1651a5..7332c40 100644
> --- a/drivers/dma/imx-sdma.c
> +++ b/drivers/dma/imx-sdma.c
> @@ -937,6 +937,14 @@ static int sdma_disable_channel(struct dma_chan *chan)
>  	return 0;
>  }
>  
> +static int sdma_disable_channel_with_delay(struct dma_chan *chan)
> +{
> +	sdma_disable_channel(chan);
> +	mdelay(1);

what is the gaurantee that 1ms is fine? Shouldn't you poll the bit to see
channel is disabled properly..

> +
> +	return 0;
> +}
> +
>  static void sdma_set_watermarklevel_for_p2p(struct sdma_channel *sdmac)
>  {
>  	struct sdma_engine *sdma = sdmac->sdma;
> @@ -1828,7 +1836,7 @@ static int sdma_probe(struct platform_device *pdev)
>  	sdma->dma_device.device_prep_slave_sg = sdma_prep_slave_sg;
>  	sdma->dma_device.device_prep_dma_cyclic = sdma_prep_dma_cyclic;
>  	sdma->dma_device.device_config = sdma_config;
> -	sdma->dma_device.device_terminate_all = sdma_disable_channel;
> +	sdma->dma_device.device_terminate_all = sdma_disable_channel_with_delay;
>  	sdma->dma_device.src_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
>  	sdma->dma_device.dst_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
>  	sdma->dma_device.directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
> -- 
> 2.7.4
> 
>
Wang, Jiada Feb. 13, 2017, 6:30 a.m. UTC | #2 
Hello Vinod

On 02/13/2017 11:05 AM, Vinod Koul wrote:
> On Fri, Feb 10, 2017 at 06:46:45AM -0800, jiada_wang@mentor.com wrote:
>> From: Jiada Wang <jiada_wang@mentor.com>
>>
>> sdma_disable_channel() cannot ensure dma is stopped to access
>> module's FIFOs. Maybe SDMA core is running and accessing BD when
>> disable of corresponding channel, this may cause sometimes even
>> after call of .sdma_disable_channel(), SDMA core still be running
>> and accessing module's FIFOs.
>>
>> We should add delay of one BD SDMA cost time, the maximum is 1ms.
>> So that SDMA clients by calling .device_terminate_all can
>> ensure SDMA core has really been stopped.
>>
>> Signed-off-by: Jiada Wang <jiada_wang@mentor.com>
>> ---
>>  drivers/dma/imx-sdma.c | 10 +++++++++-
>>  1 file changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c
>> index d1651a5..7332c40 100644
>> --- a/drivers/dma/imx-sdma.c
>> +++ b/drivers/dma/imx-sdma.c
>> @@ -937,6 +937,14 @@ static int sdma_disable_channel(struct dma_chan *chan)
>>  	return 0;
>>  }
>>
>> +static int sdma_disable_channel_with_delay(struct dma_chan *chan)
>> +{
>> +	sdma_disable_channel(chan);
>> +	mdelay(1);
>
> what is the gaurantee that 1ms is fine? Shouldn't you poll the bit to see
> channel is disabled properly..
>
I got the information from NXP (freescale) R&D team,
according to them, by write '1' to SDMA_H_STATSTOP, only disables
the related sdma channel (so poll HE bit will indicates the channel has 
been disabled),
but it cannot ensure SDMA core stop to access modules' FIFO,
SDMA core may still is running, this is a bug in HW.

regarding if the '1ms' is enough to ensure SDMA core has stopped,
NXP R&D team mentioned:
"we should add some delay of one BD SDMA cost time after disable the 
channel bit, the maximum is 1ms"
so I assume 1ms should work for all cases

Thanks,
Jiada
>> +
>> +	return 0;
>> +}
>> +
>>  static void sdma_set_watermarklevel_for_p2p(struct sdma_channel *sdmac)
>>  {
>>  	struct sdma_engine *sdma = sdmac->sdma;
>> @@ -1828,7 +1836,7 @@ static int sdma_probe(struct platform_device *pdev)
>>  	sdma->dma_device.device_prep_slave_sg = sdma_prep_slave_sg;
>>  	sdma->dma_device.device_prep_dma_cyclic = sdma_prep_dma_cyclic;
>>  	sdma->dma_device.device_config = sdma_config;
>> -	sdma->dma_device.device_terminate_all = sdma_disable_channel;
>> +	sdma->dma_device.device_terminate_all = sdma_disable_channel_with_delay;
>>  	sdma->dma_device.src_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
>>  	sdma->dma_device.dst_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
>>  	sdma->dma_device.directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
>> --
>> 2.7.4
>>
>>
>
--
To unsubscribe from this list: send the line "unsubscribe dmaengine" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Vinod Koul Feb. 13, 2017, 10:22 a.m. UTC | #3 
On Mon, Feb 13, 2017 at 03:30:19PM +0900, Jiada Wang wrote:
> >>+static int sdma_disable_channel_with_delay(struct dma_chan *chan)
> >>+{
> >>+	sdma_disable_channel(chan);
> >>+	mdelay(1);
> >
> >what is the gaurantee that 1ms is fine? Shouldn't you poll the bit to see
> >channel is disabled properly..
> >
> I got the information from NXP (freescale) R&D team,
> according to them, by write '1' to SDMA_H_STATSTOP, only disables
> the related sdma channel (so poll HE bit will indicates the channel
> has been disabled),
> but it cannot ensure SDMA core stop to access modules' FIFO,
> SDMA core may still is running, this is a bug in HW.

Okay b ut you are not doing the HE bit here..??
> 
> regarding if the '1ms' is enough to ensure SDMA core has stopped,
> NXP R&D team mentioned:
> "we should add some delay of one BD SDMA cost time after disable the
> channel bit, the maximum is 1ms"
> so I assume 1ms should work for all cases

At least please document this in changelog and comments in code.
Wang, Jiada March 1, 2017, 8:14 a.m. UTC | #4 
Hello Vinod

On 02/13/2017 07:22 PM, Vinod Koul wrote:
> On Mon, Feb 13, 2017 at 03:30:19PM +0900, Jiada Wang wrote:
>>>> +static int sdma_disable_channel_with_delay(struct dma_chan *chan)
>>>> +{
>>>> +	sdma_disable_channel(chan);
>>>> +	mdelay(1);
>>>
>>> what is the gaurantee that 1ms is fine? Shouldn't you poll the bit to see
>>> channel is disabled properly..
>>>
>> I got the information from NXP (freescale) R&D team,
>> according to them, by write '1' to SDMA_H_STATSTOP, only disables
>> the related sdma channel (so poll HE bit will indicates the channel
>> has been disabled),
>> but it cannot ensure SDMA core stop to access modules' FIFO,
>> SDMA core may still is running, this is a bug in HW.
>
> Okay b ut you are not doing the HE bit here..??
by calling sdma_disable_channel(chan) here, sdma driver clears
corresponding HE bit, thus disables the channel. But it can't ensure
SDMA core is stopped by this operation.

>>
>> regarding if the '1ms' is enough to ensure SDMA core has stopped,
>> NXP R&D team mentioned:
>> "we should add some delay of one BD SDMA cost time after disable the
>> channel bit, the maximum is 1ms"
>> so I assume 1ms should work for all cases
>
> At least please document this in changelog and comments in code.
>
I will update my patch to add comments once all concerns are addressed.

Thanks,
Jiada

--
To unsubscribe from this list: send the line "unsubscribe dmaengine" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/dma/imx-sdma.c b/drivers/dma/imx-sdma.c
index d1651a5..7332c40 100644
--- a/drivers/dma/imx-sdma.c
+++ b/drivers/dma/imx-sdma.c
@@ -937,6 +937,14 @@  static int sdma_disable_channel(struct dma_chan *chan)
 	return 0;
 }
 
+static int sdma_disable_channel_with_delay(struct dma_chan *chan)
+{
+	sdma_disable_channel(chan);
+	mdelay(1);
+
+	return 0;
+}
+
 static void sdma_set_watermarklevel_for_p2p(struct sdma_channel *sdmac)
 {
 	struct sdma_engine *sdma = sdmac->sdma;
@@ -1828,7 +1836,7 @@  static int sdma_probe(struct platform_device *pdev)
 	sdma->dma_device.device_prep_slave_sg = sdma_prep_slave_sg;
 	sdma->dma_device.device_prep_dma_cyclic = sdma_prep_dma_cyclic;
 	sdma->dma_device.device_config = sdma_config;
-	sdma->dma_device.device_terminate_all = sdma_disable_channel;
+	sdma->dma_device.device_terminate_all = sdma_disable_channel_with_delay;
 	sdma->dma_device.src_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
 	sdma->dma_device.dst_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
 	sdma->dma_device.directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);