[v1,1/2] at_hdmac: check and return DMA_PAUSED status when suitable

Message ID 20221108074938.48853-1-andriy.shevchenko@linux.intel.com
State New
Headers
Series [v1,1/2] at_hdmac: check and return DMA_PAUSED status when suitable |

Commit Message

Andy Shevchenko Nov. 8, 2022, 7:49 a.m. UTC
  device_tx_status() may return DMA_PAUSED status when driver supports it.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
---
 drivers/dma/at_hdmac.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)
  

Comments

Tudor Ambarus Nov. 8, 2022, 11:50 a.m. UTC | #1
Hi, Andy,

On 11/8/22 09:49, Andy Shevchenko wrote:
> EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> 
> device_tx_status() may return DMA_PAUSED status when driver supports it.

Yeah, but you haven't described the why. And that's because dma_cookie_status()
currently considers just DMA_COMPLETE and DMA_IN_PROGRESS for the status of the
cookie, so the controller drivers are forced to query the DMA_PAUSED state
themselves.

Also, I noticed that Vinod prefers that you use the full paragraph in the commit
message, and not a continuation of the patch title, so you may want to reword
the commit message.

> 
> Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
> ---
>  drivers/dma/at_hdmac.c | 15 ++++++++++++++-
>  1 file changed, 14 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/dma/at_hdmac.c b/drivers/dma/at_hdmac.c
> index 8858470246e1..a9d8dd990d6e 100644
> --- a/drivers/dma/at_hdmac.c
> +++ b/drivers/dma/at_hdmac.c
> @@ -1669,9 +1669,19 @@ atc_tx_status(struct dma_chan *chan,
>         int ret;
> 
>         dma_status = dma_cookie_status(chan, cookie, txstate);
> -       if (dma_status == DMA_COMPLETE || !txstate)
> +       if (dma_status == DMA_COMPLETE)
>                 return dma_status;
> 
> +       /*
> +        * There's no point in calculating the residue if there's
> +        * no txstate to store the value.
> +        */
> +       if (!txstate) {
> +               if (test_bit(ATC_IS_PAUSED, &atchan->status))

there's a helper function that you can use instead: atc_chan_is_paused()

> +                       return DMA_PAUSED;
> +               return DMA_ERROR;

return dma_status please

> +       }
> +
>         spin_lock_irqsave(&atchan->vc.lock, flags);
>         /*  Get number of bytes left in the active transactions */
>         ret = atc_get_residue(chan, cookie, &residue);
> @@ -1684,6 +1694,9 @@ atc_tx_status(struct dma_chan *chan,
>                 dma_set_residue(txstate, residue);
>         }
> 
> +       if (test_bit(ATC_IS_PAUSED, &atchan->status))

The status may change after spin_unlock_irqrestore(). Should the residue
be in sync with the dma status? If yes, you should check the status while
holding the lock.

> +               dma_status = DMA_PAUSED;
> +
>         dev_vdbg(chan2dev(chan), "tx_status %d: cookie = %d residue = %u\n",
>                  dma_status, cookie, residue);
> 
> --
> 2.35.1
> 

-- 
Cheers,
ta
  
Andy Shevchenko Nov. 8, 2022, 12:18 p.m. UTC | #2
On Tue, Nov 08, 2022 at 11:50:19AM +0000, Tudor.Ambarus@microchip.com wrote:
> On 11/8/22 09:49, Andy Shevchenko wrote:
> > EXTERNAL EMAIL: Do not click links or open attachments unless you know the content is safe
> > 
> > device_tx_status() may return DMA_PAUSED status when driver supports it.
> 
> Yeah, but you haven't described the why. And that's because dma_cookie_status()
> currently considers just DMA_COMPLETE and DMA_IN_PROGRESS for the status of the
> cookie, so the controller drivers are forced to query the DMA_PAUSED state
> themselves.

At last without this change it's inconvenient and requires a lot of additional
(unneeded) code to be written by the caller. Moreover, it's racy. If you query
status twice in a raw, it may be well changed (for example from PAUSED to
IN_PROGRESS or COMPETE).

I will add a word summarizing this.

> Also, I noticed that Vinod prefers that you use the full paragraph in the commit
> message, and not a continuation of the patch title, so you may want to reword
> the commit message.

Yes, I will fix that.

...

> > +       if (!txstate) {
> > +               if (test_bit(ATC_IS_PAUSED, &atchan->status))
> 
> there's a helper function that you can use instead: atc_chan_is_paused()

Will use it.

> > +                       return DMA_PAUSED;
> > +               return DMA_ERROR;
> 
> return dma_status please

Will squash the patch.

> > +       }

...

> > +       if (test_bit(ATC_IS_PAUSED, &atchan->status))
> 
> The status may change after spin_unlock_irqrestore(). Should the residue
> be in sync with the dma status? If yes, you should check the status while
> holding the lock.

You should tell me actually. Because I'm a bit puzzled why we need a spin lock
_and_ atomic bit operations together.

> > +               dma_status = DMA_PAUSED;

...

Thank you for the review.
  

Patch

diff --git a/drivers/dma/at_hdmac.c b/drivers/dma/at_hdmac.c
index 8858470246e1..a9d8dd990d6e 100644
--- a/drivers/dma/at_hdmac.c
+++ b/drivers/dma/at_hdmac.c
@@ -1669,9 +1669,19 @@  atc_tx_status(struct dma_chan *chan,
 	int ret;
 
 	dma_status = dma_cookie_status(chan, cookie, txstate);
-	if (dma_status == DMA_COMPLETE || !txstate)
+	if (dma_status == DMA_COMPLETE)
 		return dma_status;
 
+	/*
+	 * There's no point in calculating the residue if there's
+	 * no txstate to store the value.
+	 */
+	if (!txstate) {
+		if (test_bit(ATC_IS_PAUSED, &atchan->status))
+			return DMA_PAUSED;
+		return DMA_ERROR;
+	}
+
 	spin_lock_irqsave(&atchan->vc.lock, flags);
 	/*  Get number of bytes left in the active transactions */
 	ret = atc_get_residue(chan, cookie, &residue);
@@ -1684,6 +1694,9 @@  atc_tx_status(struct dma_chan *chan,
 		dma_set_residue(txstate, residue);
 	}
 
+	if (test_bit(ATC_IS_PAUSED, &atchan->status))
+		dma_status = DMA_PAUSED;
+
 	dev_vdbg(chan2dev(chan), "tx_status %d: cookie = %d residue = %u\n",
 		 dma_status, cookie, residue);