[PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[Peter Ujfalusi]

When based on the CCR_ENABLE bit the channel is stopped we should not call
omap_dma_callback(), only change the return value to DMA_COMPLETE. Client
drivers will do the right thing to clean up the channel after the transfer
has been completed.
Check the CCR_ENABLE only if the channel is not paused since pause in sDMA
means that the channel is stopped.
This will fix one hard to reproduce race condition when the channel is
terminated during transfer (affecting cyclic operation).

Fixes: 1a7cf7b26f25 ("dmaengine: omap-dma: Handle cases when the channel is polled for completion")

Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
---
 drivers/dma/omap-dma.c | 16 ++++++++++------
 1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/drivers/dma/omap-dma.c b/drivers/dma/omap-dma.c
index f6bef0d93998..a6b189fdbbe6 100644
--- a/drivers/dma/omap-dma.c
+++ b/drivers/dma/omap-dma.c
@@ -671,18 +671,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
 	struct omap_chan *c = to_omap_dma_chan(chan);
 	struct virt_dma_desc *vd;
 	enum dma_status ret;
-	uint32_t ccr;
 	unsigned long flags;
 
-	ccr = omap_dma_chan_read(c, CCR);
-	/* The channel is no longer active, handle the completion right away */
-	if (!(ccr & CCR_ENABLE))
-		omap_dma_callback(c->dma_ch, 0, c);
-
 	ret = dma_cookie_status(chan, cookie, txstate);
 	if (ret == DMA_COMPLETE || !txstate)
 		return ret;
 
+	if (!c->paused) {
+		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+		 * The channel is no longer active, set the return value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;
+	}
+
 	spin_lock_irqsave(&c->vc.lock, flags);
 	vd = vchan_find_desc(&c->vc, cookie);
 	if (vd) {
-- 
2.7.1

Re: [PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[Russell King - ARM Linux]

On Thu, Feb 25, 2016 at 10:28:59AM +0200, Peter Ujfalusi wrote:
> When based on the CCR_ENABLE bit the channel is stopped we should not call
> omap_dma_callback(), only change the return value to DMA_COMPLETE. Client
> drivers will do the right thing to clean up the channel after the transfer
> has been completed.
> Check the CCR_ENABLE only if the channel is not paused since pause in sDMA
> means that the channel is stopped.
> This will fix one hard to reproduce race condition when the channel is
> terminated during transfer (affecting cyclic operation).
> 
> Fixes: 1a7cf7b26f25 ("dmaengine: omap-dma: Handle cases when the channel is polled for completion")
> 
> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
> ---
>  drivers/dma/omap-dma.c | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/dma/omap-dma.c b/drivers/dma/omap-dma.c
> index f6bef0d93998..a6b189fdbbe6 100644
> --- a/drivers/dma/omap-dma.c
> +++ b/drivers/dma/omap-dma.c
> @@ -671,18 +671,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
>  	struct omap_chan *c = to_omap_dma_chan(chan);
>  	struct virt_dma_desc *vd;
>  	enum dma_status ret;
> -	uint32_t ccr;
>  	unsigned long flags;
>  
> -	ccr = omap_dma_chan_read(c, CCR);
> -	/* The channel is no longer active, handle the completion right away */
> -	if (!(ccr & CCR_ENABLE))
> -		omap_dma_callback(c->dma_ch, 0, c);
> -
>  	ret = dma_cookie_status(chan, cookie, txstate);
>  	if (ret == DMA_COMPLETE || !txstate)
>  		return ret;
>  
> +	if (!c->paused) {
> +		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;
> +	}
> +

This looks very much like a hack, and surely opens a race condition
up: what happens when a request submitted and pending but not yet
started?  If the channel is idle, requesting status will report
that the request has completed.

It's also wrong for another reason.  If txstate is NULL...

Your original commit adding the original hack that you're now removing
above says that this is to support polled operation: I'm not aware of
DMA engine supporting such a mode.  DMA_PREP_INTERRUPT is a mechanism
where requests can be queued without an interrupt to allow batching.
See the raid5/async_tx code, which queues a set of operations without
DMA_PREP_INTERRUPT, with the final operation with DMA_PREP_INTERRUPT
set.

As the driver is reliant on interrupts to move to the next transfer,
the patch which causes DMA_PREP_INTERRUPT to influence whether
interrupts are sent is actually buggy, and will prevent several
queued DMA operations to fail.

-- 
RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

Re: [PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[Peter Ujfalusi]

On 2016-02-26 03:06, Russell King - ARM Linux wrote:
> On Thu, Feb 25, 2016 at 10:28:59AM +0200, Peter Ujfalusi wrote:
>> When based on the CCR_ENABLE bit the channel is stopped we should not call
>> omap_dma_callback(), only change the return value to DMA_COMPLETE. Client
>> drivers will do the right thing to clean up the channel after the transfer
>> has been completed.
>> Check the CCR_ENABLE only if the channel is not paused since pause in sDMA
>> means that the channel is stopped.
>> This will fix one hard to reproduce race condition when the channel is
>> terminated during transfer (affecting cyclic operation).
>>
>> Fixes: 1a7cf7b26f25 ("dmaengine: omap-dma: Handle cases when the channel is polled for completion")
>>
>> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
>> ---
>>  drivers/dma/omap-dma.c | 16 ++++++++++------
>>  1 file changed, 10 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/dma/omap-dma.c b/drivers/dma/omap-dma.c
>> index f6bef0d93998..a6b189fdbbe6 100644
>> --- a/drivers/dma/omap-dma.c
>> +++ b/drivers/dma/omap-dma.c
>> @@ -671,18 +671,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
>>  	struct omap_chan *c = to_omap_dma_chan(chan);
>>  	struct virt_dma_desc *vd;
>>  	enum dma_status ret;
>> -	uint32_t ccr;
>>  	unsigned long flags;
>>  
>> -	ccr = omap_dma_chan_read(c, CCR);
>> -	/* The channel is no longer active, handle the completion right away */
>> -	if (!(ccr & CCR_ENABLE))
>> -		omap_dma_callback(c->dma_ch, 0, c);
>> -
>>  	ret = dma_cookie_status(chan, cookie, txstate);
>>  	if (ret == DMA_COMPLETE || !txstate)
>>  		return ret;
>>  
>> +	if (!c->paused) {
>> +		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;
>> +	}
>> +
> 
> This looks very much like a hack, and surely opens a race condition
> up: what happens when a request submitted and pending but not yet
> started?  If the channel is idle, requesting status will report
> that the request has completed.
> 
> It's also wrong for another reason.  If txstate is NULL...

True, I have fixed these up.

> Your original commit adding the original hack that you're now removing
> above says that this is to support polled operation: I'm not aware of
> DMA engine supporting such a mode.  DMA_PREP_INTERRUPT is a mechanism
> where requests can be queued without an interrupt to allow batching.

Also it is used to suppress DMA interrupts during audio playback for example.
In this case we will run w/o interrupts and the position is polled.

> See the raid5/async_tx code, which queues a set of operations without
> DMA_PREP_INTERRUPT, with the final operation with DMA_PREP_INTERRUPT
> set.

We only allow the interrupts to be disabled in cyclic or memcpy mode. With
slave_sg we have interrupts as it is needed to move to the next SG.

> As the driver is reliant on interrupts to move to the next transfer,
> the patch which causes DMA_PREP_INTERRUPT to influence whether
> interrupts are sent is actually buggy, and will prevent several
> queued DMA operations to fail.

Yes, the omap-dma only allows the interrupts to be actually disabled when it
is save to do so. slave_sg can not work w/o interrupts so there we don't
disable them.

> 


-- 
Péter

Re: [PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[Russell King - ARM Linux]

On Fri, Feb 26, 2016 at 12:23:27PM +0200, Peter Ujfalusi wrote:
> On 2016-02-26 03:06, Russell King - ARM Linux wrote:
> > On Thu, Feb 25, 2016 at 10:28:59AM +0200, Peter Ujfalusi wrote:
> >> When based on the CCR_ENABLE bit the channel is stopped we should not call
> >> omap_dma_callback(), only change the return value to DMA_COMPLETE. Client
> >> drivers will do the right thing to clean up the channel after the transfer
> >> has been completed.
> >> Check the CCR_ENABLE only if the channel is not paused since pause in sDMA
> >> means that the channel is stopped.
> >> This will fix one hard to reproduce race condition when the channel is
> >> terminated during transfer (affecting cyclic operation).
> >>
> >> Fixes: 1a7cf7b26f25 ("dmaengine: omap-dma: Handle cases when the channel is polled for completion")
> >>
> >> Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com>
> >> ---
> >>  drivers/dma/omap-dma.c | 16 ++++++++++------
> >>  1 file changed, 10 insertions(+), 6 deletions(-)
> >>
> >> diff --git a/drivers/dma/omap-dma.c b/drivers/dma/omap-dma.c
> >> index f6bef0d93998..a6b189fdbbe6 100644
> >> --- a/drivers/dma/omap-dma.c
> >> +++ b/drivers/dma/omap-dma.c
> >> @@ -671,18 +671,22 @@ static enum dma_status omap_dma_tx_status(struct dma_chan *chan,
> >>  	struct omap_chan *c = to_omap_dma_chan(chan);
> >>  	struct virt_dma_desc *vd;
> >>  	enum dma_status ret;
> >> -	uint32_t ccr;
> >>  	unsigned long flags;
> >>  
> >> -	ccr = omap_dma_chan_read(c, CCR);
> >> -	/* The channel is no longer active, handle the completion right away */
> >> -	if (!(ccr & CCR_ENABLE))
> >> -		omap_dma_callback(c->dma_ch, 0, c);
> >> -
> >>  	ret = dma_cookie_status(chan, cookie, txstate);
> >>  	if (ret == DMA_COMPLETE || !txstate)
> >>  		return ret;
> >>  
> >> +	if (!c->paused) {
> >> +		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;
> >> +	}
> >> +
> > 
> > This looks very much like a hack, and surely opens a race condition
> > up: what happens when a request submitted and pending but not yet
> > started?  If the channel is idle, requesting status will report
> > that the request has completed.
> > 
> > It's also wrong for another reason.  If txstate is NULL...
> 
> True, I have fixed these up.
> 
> > Your original commit adding the original hack that you're now removing
> > above says that this is to support polled operation: I'm not aware of
> > DMA engine supporting such a mode.  DMA_PREP_INTERRUPT is a mechanism
> > where requests can be queued without an interrupt to allow batching.
> 
> Also it is used to suppress DMA interrupts during audio playback for example.
> In this case we will run w/o interrupts and the position is polled.
> 
> > See the raid5/async_tx code, which queues a set of operations without
> > DMA_PREP_INTERRUPT, with the final operation with DMA_PREP_INTERRUPT
> > set.
> 
> We only allow the interrupts to be disabled in cyclic or memcpy mode. With
> slave_sg we have interrupts as it is needed to move to the next SG.
> 
> > As the driver is reliant on interrupts to move to the next transfer,
> > the patch which causes DMA_PREP_INTERRUPT to influence whether
> > interrupts are sent is actually buggy, and will prevent several
> > queued DMA operations to fail.
> 
> Yes, the omap-dma only allows the interrupts to be actually disabled when it
> is save to do so. slave_sg can not work w/o interrupts so there we don't
> disable them.

I get the impression that you haven't taken in what I've said, because
each fragment of your reply is just repeating what the previous fragment
said.

Let me state that I don't believe you need any hacks here, and this patch
is not necessary: the final operation in a set of chained memcpy()s should
have DMA_PREP_INTERRUPT set.

For the periodic stuff for audio, it's irrelevant anyway: periodic
transfers never complete - you can start them, and they continue running
until terminated.  There's no completion.

-- 
RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

Re: [PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[Peter Ujfalusi]

On 2016-02-26 13:25, Russell King - ARM Linux wrote:
>>> Your original commit adding the original hack that you're now removing
>>> above says that this is to support polled operation: I'm not aware of
>>> DMA engine supporting such a mode.  DMA_PREP_INTERRUPT is a mechanism
>>> where requests can be queued without an interrupt to allow batching.
>>
>> Also it is used to suppress DMA interrupts during audio playback for example.
>> In this case we will run w/o interrupts and the position is polled.
>>
>>> See the raid5/async_tx code, which queues a set of operations without
>>> DMA_PREP_INTERRUPT, with the final operation with DMA_PREP_INTERRUPT
>>> set.
>>
>> We only allow the interrupts to be disabled in cyclic or memcpy mode. With
>> slave_sg we have interrupts as it is needed to move to the next SG.
>>
>>> As the driver is reliant on interrupts to move to the next transfer,
>>> the patch which causes DMA_PREP_INTERRUPT to influence whether
>>> interrupts are sent is actually buggy, and will prevent several
>>> queued DMA operations to fail.
>>
>> Yes, the omap-dma only allows the interrupts to be actually disabled when it
>> is save to do so. slave_sg can not work w/o interrupts so there we don't
>> disable them.
> 
> I get the impression that you haven't taken in what I've said, because
> each fragment of your reply is just repeating what the previous fragment
> said.

Sorry about that. I got it.

> Let me state that I don't believe you need any hacks here, and this patch
> is not necessary: the final operation in a set of chained memcpy()s should
> have DMA_PREP_INTERRUPT set.

calling the omap_dma_callback() from the tx_status() was a bad idea, this is
why I'm fixing it.
As for the DMA_PREP_INTERRUPT not enabling interrupt was introduced by:
4ce98c0a20bef (dmaengine: omap-dma: Add support for memcpy)

While we do not have users submitting multiple descriptors I agree that this
is a possibility and the driver should not fail in such a case.
I can send a followup patch to fix the omap_dma_prep_dma_memcpy()

-- 
Péter

Re: [PATCH] dmaengine: omap-dma: Do not call omap_dma_callback() from tx_status()

[Peter Ujfalusi]

On 2016-02-26 14:45, Peter Ujfalusi wrote:
> On 2016-02-26 13:25, Russell King - ARM Linux wrote:
>>>> Your original commit adding the original hack that you're now removing
>>>> above says that this is to support polled operation: I'm not aware of
>>>> DMA engine supporting such a mode.  DMA_PREP_INTERRUPT is a mechanism
>>>> where requests can be queued without an interrupt to allow batching.
>>>
>>> Also it is used to suppress DMA interrupts during audio playback for example.
>>> In this case we will run w/o interrupts and the position is polled.
>>>
>>>> See the raid5/async_tx code, which queues a set of operations without
>>>> DMA_PREP_INTERRUPT, with the final operation with DMA_PREP_INTERRUPT
>>>> set.
>>>
>>> We only allow the interrupts to be disabled in cyclic or memcpy mode. With
>>> slave_sg we have interrupts as it is needed to move to the next SG.
>>>
>>>> As the driver is reliant on interrupts to move to the next transfer,
>>>> the patch which causes DMA_PREP_INTERRUPT to influence whether
>>>> interrupts are sent is actually buggy, and will prevent several
>>>> queued DMA operations to fail.
>>>
>>> Yes, the omap-dma only allows the interrupts to be actually disabled when it
>>> is save to do so. slave_sg can not work w/o interrupts so there we don't
>>> disable them.
>>
>> I get the impression that you haven't taken in what I've said, because
>> each fragment of your reply is just repeating what the previous fragment
>> said.
> 
> Sorry about that. I got it.
> 
>> Let me state that I don't believe you need any hacks here, and this patch
>> is not necessary: the final operation in a set of chained memcpy()s should
>> have DMA_PREP_INTERRUPT set.
> 
> calling the omap_dma_callback() from the tx_status() was a bad idea, this is
> why I'm fixing it.

The reason that we need to be able to determine if the channel is done with
the copy by polling the tx_status is: we need to read and write tiler
registers with DMA on dra7 as per one ERRATA document. This read/write can be
executed in interrupt context so we can not rely on DMA callback to be
notified about the completion of the read/write.

> As for the DMA_PREP_INTERRUPT not enabling interrupt was introduced by:
> 4ce98c0a20bef (dmaengine: omap-dma: Add support for memcpy)
> 
> While we do not have users submitting multiple descriptors I agree that this
> is a possibility and the driver should not fail in such a case.
> I can send a followup patch to fix the omap_dma_prep_dma_memcpy()

With the updated patch based on your comments I can fix the polling in the
tx_status callback and with a separate patch I can address your concern
regarding to multiple transfers queued with only the last one having
DMA_PREP_INTERRUPT (and callback provided).

-- 
Péter