[PATCH v2 0/2] dmaengine: ti: omap-dma: Improved polling support

[PATCH v2 0/2] dmaengine: ti: omap-dma: Improved polling support

[Peter Ujfalusi]

Hi,

changes since v1:
- New preparation cleanup patch
- Simplified code for the main patch to be easier to read

This series fine-tunes the omap-dma polled memcpy support to be inline with how
the EDMA driver is handling it.

The polled completion can be tested by applying:
https://patchwork.kernel.org/patch/10966499/

and run the dmatest with polled = 1 on boards where sDMA is used.

Or boot up any dra7 family device with display enabled. The workaround for DMM
errata i878 uses polled DMA memcpy.

Regards,
Peter
---
Peter Ujfalusi (2):
  dmaengine: ti: omap-dma: Readability cleanup in omap_dma_tx_status()
  dmaengine: ti: omap-dma: Improved memcpy polling support

 drivers/dma/ti/omap-dma.c | 52 +++++++++++++++++++++++++--------------
 1 file changed, 33 insertions(+), 19 deletions(-)

-- 
Peter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

[PATCH v2 1/2] dmaengine: ti: omap-dma: Readability cleanup in omap_dma_tx_status()

[Peter Ujfalusi]

The tx_status is most likely going to be asked for the current transfer, so
check that first then try to fall back to lookup of non started transfers.

In this way the code is a bit more readable and in most cases we will avoid
to run vchan_find_desc() all the time.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
---
 drivers/dma/ti/omap-dma.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/drivers/dma/ti/omap-dma.c b/drivers/dma/ti/omap-dma.c
index ba2489d4ea24..029c0bd550d5 100644
--- a/drivers/dma/ti/omap-dma.c
+++ b/drivers/dma/ti/omap-dma.c
@@ -832,10 +832,8 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
 		return ret;
 
 	spin_lock_irqsave(&c->vc.lock, flags);
-	vd = vchan_find_desc(&c->vc, cookie);
-	if (vd) {
-		txstate->residue = omap_dma_desc_size(to_omap_dma_desc(&vd->tx));
-	} else if (c->desc && c->desc->vd.tx.cookie == cookie) {
+
+	if (c->desc && c->desc->vd.tx.cookie == cookie) {
 		struct omap_desc *d = c->desc;
 		dma_addr_t pos;
 
@@ -847,11 +845,15 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
 			pos = 0;
 
 		txstate->residue = omap_dma_desc_size_pos(d, pos);
+	} else if ((vd = vchan_find_desc(&c->vc, cookie))) {
+		txstate->residue = omap_dma_desc_size(to_omap_dma_desc(&vd->tx));
 	} else {
 		txstate->residue = 0;
 	}
+
 	if (ret == DMA_IN_PROGRESS && c->paused)
 		ret = DMA_PAUSED;
+
 	spin_unlock_irqrestore(&c->vc.lock, flags);
 
 	return ret;
-- 
Peter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

[PATCH v2 2/2] dmaengine: ti: omap-dma: Improved memcpy polling support

[Peter Ujfalusi]

When a DMA client driver does not set the DMA_PREP_INTERRUPT because it
does not want to use interrupts for DMA completion or because it can not
rely on DMA interrupts due to executing the memcpy when interrupts are
disabled it will poll the status of the transfer.

If the interrupts are enabled then the cookie will be set completed in the
interrupt handler so only check in HW completion when the polling is really
needed.

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
---
 drivers/dma/ti/omap-dma.c | 44 +++++++++++++++++++++++++--------------
 1 file changed, 28 insertions(+), 16 deletions(-)

diff --git a/drivers/dma/ti/omap-dma.c b/drivers/dma/ti/omap-dma.c
index 029c0bd550d5..966d8f0323b5 100644
--- a/drivers/dma/ti/omap-dma.c
+++ b/drivers/dma/ti/omap-dma.c
@@ -91,6 +91,7 @@ struct omap_desc {
 	bool using_ll;
 	enum dma_transfer_direction dir;
 	dma_addr_t dev_addr;
+	bool polled;
 
 	int32_t fi;		/* for OMAP_DMA_SYNC_PACKET / double indexing */
 	int16_t ei;		/* for double indexing */
@@ -815,26 +816,20 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
 	struct virt_dma_desc *vd;
 	enum dma_status ret;
 	unsigned long flags;
+	struct omap_desc *d = NULL;
 
 	ret = dma_cookie_status(chan, cookie, txstate);
-
-	if (!c->paused && c->running) {
-		uint32_t ccr = omap_dma_chan_read(c, CCR);
-		/*
-		 * The channel is no longer active, set the return value
-		 * accordingly
-		 */
-		if (!(ccr & CCR_ENABLE))
-			ret = DMA_COMPLETE;
-	}
-
-	if (ret == DMA_COMPLETE || !txstate)
+	if (ret == DMA_COMPLETE)
 		return ret;
 
 	spin_lock_irqsave(&c->vc.lock, flags);
+	if (c->desc && c->desc->vd.tx.cookie == cookie)
+		d = c->desc;
+
+	if (!txstate)
+		goto out;
 
-	if (c->desc && c->desc->vd.tx.cookie == cookie) {
-		struct omap_desc *d = c->desc;
+	if (d) {
 		dma_addr_t pos;
 
 		if (d->dir == DMA_MEM_TO_DEV)
@@ -851,8 +846,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
 		txstate->residue = 0;
 	}
 
-	if (ret == DMA_IN_PROGRESS && c->paused)
+out:
+	if (ret == DMA_IN_PROGRESS && c->paused) {
 		ret = DMA_PAUSED;
+	} else if (d && d->polled && c->running) {
+		uint32_t ccr = omap_dma_chan_read(c, CCR);
+		/*
+		 * The channel is no longer active, set the return value
+		 * accordingly and mark it as completed
+		 */
+		if (!(ccr & CCR_ENABLE)) {
+			struct omap_desc *d = c->desc;
+			ret = DMA_COMPLETE;
+			omap_dma_start_desc(c);
+			vchan_cookie_complete(&d->vd);
+		}
+	}
 
 	spin_unlock_irqrestore(&c->vc.lock, flags);
 
@@ -1180,7 +1189,10 @@ static struct dma_async_tx_descriptor *omap_dma_prep_dma_memcpy(
 	d->ccr = c->ccr;
 	d->ccr |= CCR_DST_AMODE_POSTINC | CCR_SRC_AMODE_POSTINC;
 
-	d->cicr = CICR_DROP_IE | CICR_FRAME_IE;
+	if (tx_flags & DMA_PREP_INTERRUPT)
+		d->cicr |= CICR_FRAME_IE;
+	else
+		d->polled = true;
 
 	d->csdp = data_type;
 
-- 
Peter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

Re: [PATCH v2 2/2] dmaengine: ti: omap-dma: Improved memcpy polling support

[Vinod Koul]

On 16-07-19, 11:24, Peter Ujfalusi wrote:
> When a DMA client driver does not set the DMA_PREP_INTERRUPT because it
> does not want to use interrupts for DMA completion or because it can not
> rely on DMA interrupts due to executing the memcpy when interrupts are
> disabled it will poll the status of the transfer.
> 
> If the interrupts are enabled then the cookie will be set completed in the
> interrupt handler so only check in HW completion when the polling is really
> needed.
> 
> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
> ---
>  drivers/dma/ti/omap-dma.c | 44 +++++++++++++++++++++++++--------------
>  1 file changed, 28 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/dma/ti/omap-dma.c b/drivers/dma/ti/omap-dma.c
> index 029c0bd550d5..966d8f0323b5 100644
> --- a/drivers/dma/ti/omap-dma.c
> +++ b/drivers/dma/ti/omap-dma.c
> @@ -91,6 +91,7 @@ struct omap_desc {
>  	bool using_ll;
>  	enum dma_transfer_direction dir;
>  	dma_addr_t dev_addr;
> +	bool polled;
>  
>  	int32_t fi;		/* for OMAP_DMA_SYNC_PACKET / double indexing */
>  	int16_t ei;		/* for double indexing */
> @@ -815,26 +816,20 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
>  	struct virt_dma_desc *vd;
>  	enum dma_status ret;
>  	unsigned long flags;
> +	struct omap_desc *d = NULL;
>  
>  	ret = dma_cookie_status(chan, cookie, txstate);
> -
> -	if (!c->paused && c->running) {
> -		uint32_t ccr = omap_dma_chan_read(c, CCR);
> -		/*
> -		 * The channel is no longer active, set the return value
> -		 * accordingly
> -		 */
> -		if (!(ccr & CCR_ENABLE))
> -			ret = DMA_COMPLETE;
> -	}
> -
> -	if (ret == DMA_COMPLETE || !txstate)
> +	if (ret == DMA_COMPLETE)

why do you want to continue for txstate being null?
Also it would lead to NULL ptr deref for txstate

>  		return ret;
>  
>  	spin_lock_irqsave(&c->vc.lock, flags);
> +	if (c->desc && c->desc->vd.tx.cookie == cookie)
> +		d = c->desc;
> +
> +	if (!txstate)
> +		goto out;
>  
> -	if (c->desc && c->desc->vd.tx.cookie == cookie) {
> -		struct omap_desc *d = c->desc;
> +	if (d) {
>  		dma_addr_t pos;
>  
>  		if (d->dir == DMA_MEM_TO_DEV)
> @@ -851,8 +846,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
>  		txstate->residue = 0;
>  	}
>  
> -	if (ret == DMA_IN_PROGRESS && c->paused)
> +out:
> +	if (ret == DMA_IN_PROGRESS && c->paused) {
>  		ret = DMA_PAUSED;
> +	} else if (d && d->polled && c->running) {
> +		uint32_t ccr = omap_dma_chan_read(c, CCR);
> +		/*
> +		 * The channel is no longer active, set the return value
> +		 * accordingly and mark it as completed
> +		 */
> +		if (!(ccr & CCR_ENABLE)) {
> +			struct omap_desc *d = c->desc;
> +			ret = DMA_COMPLETE;
> +			omap_dma_start_desc(c);
> +			vchan_cookie_complete(&d->vd);
> +		}
> +	}
>  
>  	spin_unlock_irqrestore(&c->vc.lock, flags);
>  
> @@ -1180,7 +1189,10 @@ static struct dma_async_tx_descriptor *omap_dma_prep_dma_memcpy(
>  	d->ccr = c->ccr;
>  	d->ccr |= CCR_DST_AMODE_POSTINC | CCR_SRC_AMODE_POSTINC;
>  
> -	d->cicr = CICR_DROP_IE | CICR_FRAME_IE;
> +	if (tx_flags & DMA_PREP_INTERRUPT)
> +		d->cicr |= CICR_FRAME_IE;
> +	else
> +		d->polled = true;
>  
>  	d->csdp = data_type;
>  
> -- 
> Peter
> 
> Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
> Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

-- 
~Vinod

Re: [PATCH v2 2/2] dmaengine: ti: omap-dma: Improved memcpy polling support

[Peter Ujfalusi]

On 25/07/2019 16.37, Vinod Koul wrote:
> On 16-07-19, 11:24, Peter Ujfalusi wrote:
>> When a DMA client driver does not set the DMA_PREP_INTERRUPT because it
>> does not want to use interrupts for DMA completion or because it can not
>> rely on DMA interrupts due to executing the memcpy when interrupts are
>> disabled it will poll the status of the transfer.
>>
>> If the interrupts are enabled then the cookie will be set completed in the
>> interrupt handler so only check in HW completion when the polling is really
>> needed.
>>
>> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
>> ---
>>  drivers/dma/ti/omap-dma.c | 44 +++++++++++++++++++++++++--------------
>>  1 file changed, 28 insertions(+), 16 deletions(-)
>>
>> diff --git a/drivers/dma/ti/omap-dma.c b/drivers/dma/ti/omap-dma.c
>> index 029c0bd550d5..966d8f0323b5 100644
>> --- a/drivers/dma/ti/omap-dma.c
>> +++ b/drivers/dma/ti/omap-dma.c
>> @@ -91,6 +91,7 @@ struct omap_desc {
>>  	bool using_ll;
>>  	enum dma_transfer_direction dir;
>>  	dma_addr_t dev_addr;
>> +	bool polled;
>>  
>>  	int32_t fi;		/* for OMAP_DMA_SYNC_PACKET / double indexing */
>>  	int16_t ei;		/* for double indexing */
>> @@ -815,26 +816,20 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
>>  	struct virt_dma_desc *vd;
>>  	enum dma_status ret;
>>  	unsigned long flags;
>> +	struct omap_desc *d = NULL;
>>  
>>  	ret = dma_cookie_status(chan, cookie, txstate);
>> -
>> -	if (!c->paused && c->running) {
>> -		uint32_t ccr = omap_dma_chan_read(c, CCR);
>> -		/*
>> -		 * The channel is no longer active, set the return value
>> -		 * accordingly
>> -		 */
>> -		if (!(ccr & CCR_ENABLE))
>> -			ret = DMA_COMPLETE;
>> -	}
>> -
>> -	if (ret == DMA_COMPLETE || !txstate)
>> +	if (ret == DMA_COMPLETE)
> 
> why do you want to continue for txstate being null?

The caller could opt to not provide txstate and I still need to check if
the non completed transfer is actually done by the HW or not.

> Also it would lead to NULL ptr deref for txstate

There is a !txstate check to avoid that.

> 
>>  		return ret;
>>  
>>  	spin_lock_irqsave(&c->vc.lock, flags);
>> +	if (c->desc && c->desc->vd.tx.cookie == cookie)
>> +		d = c->desc;
>> +
>> +	if (!txstate)
>> +		goto out;
>>  
>> -	if (c->desc && c->desc->vd.tx.cookie == cookie) {
>> -		struct omap_desc *d = c->desc;
>> +	if (d) {
>>  		dma_addr_t pos;
>>  
>>  		if (d->dir == DMA_MEM_TO_DEV)
>> @@ -851,8 +846,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
>>  		txstate->residue = 0;
>>  	}
>>  
>> -	if (ret == DMA_IN_PROGRESS && c->paused)
>> +out:
>> +	if (ret == DMA_IN_PROGRESS && c->paused) {
>>  		ret = DMA_PAUSED;
>> +	} else if (d && d->polled && c->running) {
>> +		uint32_t ccr = omap_dma_chan_read(c, CCR);
>> +		/*
>> +		 * The channel is no longer active, set the return value
>> +		 * accordingly and mark it as completed
>> +		 */
>> +		if (!(ccr & CCR_ENABLE)) {
>> +			struct omap_desc *d = c->desc;
>> +			ret = DMA_COMPLETE;
>> +			omap_dma_start_desc(c);
>> +			vchan_cookie_complete(&d->vd);
>> +		}
>> +	}
>>  
>>  	spin_unlock_irqrestore(&c->vc.lock, flags);
>>  
>> @@ -1180,7 +1189,10 @@ static struct dma_async_tx_descriptor *omap_dma_prep_dma_memcpy(
>>  	d->ccr = c->ccr;
>>  	d->ccr |= CCR_DST_AMODE_POSTINC | CCR_SRC_AMODE_POSTINC;
>>  
>> -	d->cicr = CICR_DROP_IE | CICR_FRAME_IE;
>> +	if (tx_flags & DMA_PREP_INTERRUPT)
>> +		d->cicr |= CICR_FRAME_IE;
>> +	else
>> +		d->polled = true;
>>  
>>  	d->csdp = data_type;
>>  
>> -- 
>> Peter
>>
>> Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
>> Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki
> 

- Péter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

Re: [PATCH v2 2/2] dmaengine: ti: omap-dma: Improved memcpy polling support

[Vinod Koul]

On 25-07-19, 17:07, Peter Ujfalusi wrote:
> 
> 
> On 25/07/2019 16.37, Vinod Koul wrote:
> > On 16-07-19, 11:24, Peter Ujfalusi wrote:
> >> When a DMA client driver does not set the DMA_PREP_INTERRUPT because it
> >> does not want to use interrupts for DMA completion or because it can not
> >> rely on DMA interrupts due to executing the memcpy when interrupts are
> >> disabled it will poll the status of the transfer.
> >>
> >> If the interrupts are enabled then the cookie will be set completed in the
> >> interrupt handler so only check in HW completion when the polling is really
> >> needed.
> >>
> >> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
> >> ---
> >>  drivers/dma/ti/omap-dma.c | 44 +++++++++++++++++++++++++--------------
> >>  1 file changed, 28 insertions(+), 16 deletions(-)
> >>
> >> diff --git a/drivers/dma/ti/omap-dma.c b/drivers/dma/ti/omap-dma.c
> >> index 029c0bd550d5..966d8f0323b5 100644
> >> --- a/drivers/dma/ti/omap-dma.c
> >> +++ b/drivers/dma/ti/omap-dma.c
> >> @@ -91,6 +91,7 @@ struct omap_desc {
> >>  	bool using_ll;
> >>  	enum dma_transfer_direction dir;
> >>  	dma_addr_t dev_addr;
> >> +	bool polled;
> >>  
> >>  	int32_t fi;		/* for OMAP_DMA_SYNC_PACKET / double indexing */
> >>  	int16_t ei;		/* for double indexing */
> >> @@ -815,26 +816,20 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
> >>  	struct virt_dma_desc *vd;
> >>  	enum dma_status ret;
> >>  	unsigned long flags;
> >> +	struct omap_desc *d = NULL;
> >>  
> >>  	ret = dma_cookie_status(chan, cookie, txstate);
> >> -
> >> -	if (!c->paused && c->running) {
> >> -		uint32_t ccr = omap_dma_chan_read(c, CCR);
> >> -		/*
> >> -		 * The channel is no longer active, set the return value
> >> -		 * accordingly
> >> -		 */
> >> -		if (!(ccr & CCR_ENABLE))
> >> -			ret = DMA_COMPLETE;
> >> -	}
> >> -
> >> -	if (ret == DMA_COMPLETE || !txstate)
> >> +	if (ret == DMA_COMPLETE)
> > 
> > why do you want to continue for txstate being null?
> 
> The caller could opt to not provide txstate and I still need to check if
> the non completed transfer is actually done by the HW or not.
> 
> > Also it would lead to NULL ptr deref for txstate
> 
> There is a !txstate check to avoid that.
> 
> > 
> >>  		return ret;
> >>  
> >>  	spin_lock_irqsave(&c->vc.lock, flags);
> >> +	if (c->desc && c->desc->vd.tx.cookie == cookie)
> >> +		d = c->desc;
> >> +
> >> +	if (!txstate)
> >> +		goto out;

Oops missed that, let me check again and do the needful

-- 
~Vinod

Re: [PATCH v2 0/2] dmaengine: ti: omap-dma: Improved polling support

[Vinod Koul]

On 16-07-19, 11:24, Peter Ujfalusi wrote:
> Hi,
> 
> changes since v1:
> - New preparation cleanup patch
> - Simplified code for the main patch to be easier to read
> 
> This series fine-tunes the omap-dma polled memcpy support to be inline with how
> the EDMA driver is handling it.
> 
> The polled completion can be tested by applying:
> https://patchwork.kernel.org/patch/10966499/
> 
> and run the dmatest with polled = 1 on boards where sDMA is used.
> 
> Or boot up any dra7 family device with display enabled. The workaround for DMM
> errata i878 uses polled DMA memcpy.

Applied, thanks

-- 
~Vinod