fsl_ssi: set fifo watermark to more reliable value

Commit Message

Caleb Crome Jan. 3, 2017, 6:22 p.m. UTC

From: Caleb Crome <[e-mail]>

The fsl_ssi fifo watermark is by default set to 2 free spaces (i.e.
activate DMA on FIFO when only 2 spaces are left.)  This means the
DMA must service the fifo within 2 audio samples, which is just not
enough time  for many use cases with high data rate.  In many
configurations the audio channel slips (causing l/r swap in stereo
configurations, or channel slipping in multi-channel configurations).

This patch gives more breathing room and allows the SSI to operate
reliably by changing the fifio refill watermark to 8.

There is no change in behavior for older chips (with an 8-deep fifo).
Only the newer chips with a 15-deep fifo get the new behavior. I
suspect a new fifo depth setting could be optimized on the older
chips too, but I have not tested.

Signed-off-by: Caleb Crome <[e-mail]>
---
 sound/soc/fsl/fsl_ssi.c | 74 +++++++++++++++++++++++++++++++++++--------------
 1 file changed, 53 insertions(+), 21 deletions(-)

Comments

Fabio Estevam Jan. 3, 2017, 9:08 p.m. UTC | #1

On Tue, Jan 3, 2017 at 4:22 PM, Caleb Crome <[e-mail]> wrote:
> From: Caleb Crome <[e-mail]>
>
> The fsl_ssi fifo watermark is by default set to 2 free spaces (i.e.
> activate DMA on FIFO when only 2 spaces are left.)  This means the
> DMA must service the fifo within 2 audio samples, which is just not
> enough time  for many use cases with high data rate.  In many
> configurations the audio channel slips (causing l/r swap in stereo
> configurations, or channel slipping in multi-channel configurations).
>
> This patch gives more breathing room and allows the SSI to operate
> reliably by changing the fifio refill watermark to 8.
>
> There is no change in behavior for older chips (with an 8-deep fifo).
> Only the newer chips with a 15-deep fifo get the new behavior. I
> suspect a new fifo depth setting could be optimized on the older
> chips too, but I have not tested.
>
> Signed-off-by: Caleb Crome <[e-mail]>

Reviewed-by: Fabio Estevam <[e-mail]>

Mark Brown Jan. 4, 2017, 6:27 p.m. UTC | #2

On Tue, Jan 03, 2017 at 10:22:57AM -0800, Caleb Crome wrote:
> From: Caleb Crome <[e-mail]>
> 
> The fsl_ssi fifo watermark is by default set to 2 free spaces (i.e.
> activate DMA on FIFO when only 2 spaces are left.)  This means the

Please submit patches using subject lines reflecting the style for the
subsystem.  This makes it easier for people to identify relevant
patches.  Look at what existing commits in the area you're changing are
doing and make sure your subject lines visually resemble what they're
doing.

Caleb Crome Jan. 4, 2017, 6:31 p.m. UTC | #3

On Wed, Jan 4, 2017 at 10:27 AM, Mark Brown <[e-mail]> wrote:
>
> On Tue, Jan 03, 2017 at 10:22:57AM -0800, Caleb Crome wrote:
> > From: Caleb Crome <[e-mail]>
> >
> > The fsl_ssi fifo watermark is by default set to 2 free spaces (i.e.
> > activate DMA on FIFO when only 2 spaces are left.)  This means the
>
> Please submit patches using subject lines reflecting the style for the
> subsystem.  This makes it easier for people to identify relevant
> patches.  Look at what existing commits in the area you're changing are
> doing and make sure your subject lines visually resemble what they're
> doing.

Shoot, yes of course.  I forgot to check the subject format against
extant patches.

The subject should be

[PATCH] ASoC: fsl_ssi: set fifo watermark to more reliable value

Correct?

Should I re-submit?

Thanks,
 -Caleb

Mark Brown Jan. 4, 2017, 6:35 p.m. UTC | #4

On Wed, Jan 04, 2017 at 10:31:42AM -0800, ccrome wrote:

> The subject should be
> 
> [PATCH] ASoC: fsl_ssi: set fifo watermark to more reliable value
> 
> Correct?

Yes.

> Should I re-submit?

No, it's fine - if I'm sending one of those mails it'll be fine as I've
actually seen the mail to reply to it, the problem is if I don't see the
mail.

Patch

diff --git a/sound/soc/fsl/fsl_ssi.c b/sound/soc/fsl/fsl_ssi.c
index bedec4a..56245e6 100644
--- a/sound/soc/fsl/fsl_ssi.c
+++ b/sound/soc/fsl/fsl_ssi.c
@@ -224,6 +224,12 @@  struct fsl_ssi_soc_data {
  * @dbg_stats: Debugging statistics
  *
  * @soc: SoC specific data
+ *
+ * @fifo_watermark: the FIFO watermark setting.  Notifies DMA when
+ *             there are @fifo_watermark or fewer words in TX fifo or
+ *             @fifo_watermark or more empty words in RX fifo.
+ * @dma_maxburst: max number of words to transfer in one go.  So far,
+ *             this is always the same as fifo_watermark.
  */
 struct fsl_ssi_private {
 	struct regmap *regs;
@@ -263,6 +269,9 @@  struct fsl_ssi_private {
 
 	const struct fsl_ssi_soc_data *soc;
 	struct device *dev;
+
+	u32 fifo_watermark;
+	u32 dma_maxburst;
 };
 
 /*
@@ -1051,21 +1060,7 @@  static int _fsl_ssi_set_dai_fmt(struct device *dev,
 	regmap_write(regs, CCSR_SSI_SRCR, srcr);
 	regmap_write(regs, CCSR_SSI_SCR, scr);
 
-	/*
-	 * Set the watermark for transmit FIFI 0 and receive FIFO 0. We don't
-	 * use FIFO 1. We program the transmit water to signal a DMA transfer
-	 * if there are only two (or fewer) elements left in the FIFO. Two
-	 * elements equals one frame (left channel, right channel). This value,
-	 * however, depends on the depth of the transmit buffer.
-	 *
-	 * We set the watermark on the same level as the DMA burstsize.  For
-	 * fiq it is probably better to use the biggest possible watermark
-	 * size.
-	 */
-	if (ssi_private->use_dma)
-		wm = ssi_private->fifo_depth - 2;
-	else
-		wm = ssi_private->fifo_depth;
+	wm = ssi_private->fifo_watermark;
 
 	regmap_write(regs, CCSR_SSI_SFCSR,
 			CCSR_SSI_SFCSR_TFWM0(wm) | CCSR_SSI_SFCSR_RFWM0(wm) |
@@ -1373,12 +1368,8 @@  static int fsl_ssi_imx_probe(struct platform_device *pdev,
 		dev_dbg(&pdev->dev, "could not get baud clock: %ld\n",
 			 PTR_ERR(ssi_private->baudclk));
 
-	/*
-	 * We have burstsize be "fifo_depth - 2" to match the SSI
-	 * watermark setting in fsl_ssi_startup().
-	 */
-	ssi_private->dma_params_tx.maxburst = ssi_private->fifo_depth - 2;
-	ssi_private->dma_params_rx.maxburst = ssi_private->fifo_depth - 2;
+	ssi_private->dma_params_tx.maxburst = ssi_private->dma_maxburst;
+	ssi_private->dma_params_rx.maxburst = ssi_private->dma_maxburst;
 	ssi_private->dma_params_tx.addr = ssi_private->ssi_phys + CCSR_SSI_STX0;
 	ssi_private->dma_params_rx.addr = ssi_private->ssi_phys + CCSR_SSI_SRX0;
 
@@ -1543,6 +1534,47 @@  static int fsl_ssi_probe(struct platform_device *pdev)
                 /* Older 8610 DTs didn't have the fifo-depth property */
 		ssi_private->fifo_depth = 8;
 
+	/*
+	 * Set the watermark for transmit FIFO 0 and receive FIFO 0. We don't
+	 * use FIFO 1 but set the watermark appropriately nontheless.
+	 * We program the transmit water to signal a DMA transfer
+	 * if there are N elements left in the FIFO. For chips with 15-deep
+	 * FIFOs, set watermark to 8.  This allows the SSI to operate at a
+	 * high data rate without channel slipping. Behavior is unchanged
+	 * for the older chips with a fifo depth of only 8.  A value of 4
+	 * might be appropriate for the older chips, but is left at
+	 * fifo_depth-2 until sombody has a chance to test.
+	 *
+	 * We set the watermark on the same level as the DMA burstsize.  For
+	 * fiq it is probably better to use the biggest possible watermark
+	 * size.
+	 */
+	switch (ssi_private->fifo_depth) {
+	case 15:
+		/*
+		 * 2 samples is not enough when running at high data
+		 * rates (like 48kHz @ 16 bits/channel, 16 channels)
+		 * 8 seems to split things evenly and leave enough time
+		 * for the DMA to fill the FIFO before it's over/under
+		 * run.
+		 */
+		ssi_private->fifo_watermark = 8;
+		ssi_private->dma_maxburst = 8;
+		break;
+	case 8:
+	default:
+		/*
+		 * maintain old behavior for older chips.
+		 * Keeping it the same because I don't have an older
+		 * board to test with.
+		 * I suspect this could be changed to be something to
+		 * leave some more space in the fifo.
+		 */
+		ssi_private->fifo_watermark = ssi_private->fifo_depth - 2;
+		ssi_private->dma_maxburst = ssi_private->fifo_depth - 2;
+		break;
+	}
+
 	dev_set_drvdata(&pdev->dev, ssi_private);
 
 	if (ssi_private->soc->imx) {