[alsa-devel] [PATCH] ASoC: fsl-ssi: fix SDMA starvation
From: Juergen Beisert jbe@pengutronix.de
In cases when capturing is already running and someone enables also playback, the SDMA unit of the i.MX SoC does not see an edge on its TX DMA request line. This is due to the TX FIFOs are empty, and this request line is active all the time, when the CCSR_SSI_SIER_TDMAE bit is set. In this case also the TX FIFO underrun interrupt will flood the system forever.
To ensure the first edge happens after enabling the TX side of the SSI unit and to trigger the SDMA unit successfully enable the CCSR_SSI_SIER_TDMAE on demand only. To be sure, this patch does the same for the CCSR_SSI_SIER_RDMAE (RX side) bit.
Signed-off-by: Juergen Beisert jbe@pengutronix.de Cc: stable@vger.kernel.org Signed-off-by: Markus Pargmann mpa@pengutronix.de --- Hi,
The patch is based on Mark's "fix/fsl" branch.
Regards,
Markus Pargmann
sound/soc/fsl/fsl_ssi.c | 24 ++++++++++++++++-------- 1 file changed, 16 insertions(+), 8 deletions(-)
diff --git a/sound/soc/fsl/fsl_ssi.c b/sound/soc/fsl/fsl_ssi.c index 6b81d0c..8c94632 100644 --- a/sound/soc/fsl/fsl_ssi.c +++ b/sound/soc/fsl/fsl_ssi.c @@ -107,10 +107,9 @@ static inline void write_ssi_mask(u32 __iomem *addr, u32 clear, u32 set) #endif
/* SIER bitflag of interrupts to enable */ -#define SIER_FLAGS (CCSR_SSI_SIER_TFRC_EN | CCSR_SSI_SIER_TDMAE | \ - CCSR_SSI_SIER_TIE | CCSR_SSI_SIER_TUE0_EN | \ - CCSR_SSI_SIER_TUE1_EN | CCSR_SSI_SIER_RFRC_EN | \ - CCSR_SSI_SIER_RDMAE | CCSR_SSI_SIER_RIE | \ +#define SIER_FLAGS (CCSR_SSI_SIER_TFRC_EN | CCSR_SSI_SIER_TIE | \ + CCSR_SSI_SIER_TUE0_EN | CCSR_SSI_SIER_TUE1_EN | \ + CCSR_SSI_SIER_RFRC_EN | CCSR_SSI_SIER_RIE | \ CCSR_SSI_SIER_ROE0_EN | CCSR_SSI_SIER_ROE1_EN)
/** @@ -548,6 +547,9 @@ static int fsl_ssi_hw_params(struct snd_pcm_substream *substream, * * The DMA channel is in external master start and pause mode, which * means the SSI completely controls the flow of data. + * + * To use the i.MX's SDMA unit it is important to create an edge on the + * corresponding request line. Otherwise the request will be ignored! */ static int fsl_ssi_trigger(struct snd_pcm_substream *substream, int cmd, struct snd_soc_dai *dai) @@ -580,20 +582,26 @@ static int fsl_ssi_trigger(struct snd_pcm_substream *substream, int cmd, switch (cmd) { case SNDRV_PCM_TRIGGER_START: case SNDRV_PCM_TRIGGER_PAUSE_RELEASE: - if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) + if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) { write_ssi_mask(&ssi->scr, 0, CCSR_SSI_SCR_SSIEN | CCSR_SSI_SCR_TE); - else + write_ssi_mask(&ssi->sier, 0, CCSR_SSI_SIER_TDMAE); + } else { write_ssi_mask(&ssi->scr, 0, CCSR_SSI_SCR_SSIEN | CCSR_SSI_SCR_RE); + write_ssi_mask(&ssi->sier, 0, CCSR_SSI_SIER_RDMAE); + } break;
case SNDRV_PCM_TRIGGER_STOP: case SNDRV_PCM_TRIGGER_PAUSE_PUSH: - if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) + if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK) { write_ssi_mask(&ssi->scr, CCSR_SSI_SCR_TE, 0); - else + write_ssi_mask(&ssi->sier, CCSR_SSI_SIER_TDMAE, 0); + } else { write_ssi_mask(&ssi->scr, CCSR_SSI_SCR_RE, 0); + write_ssi_mask(&ssi->sier, CCSR_SSI_SIER_RDMAE, 0); + }
if (!ssi_private->imx_ac97 && (read_ssi(&ssi->scr) & (CCSR_SSI_SCR_TE | CCSR_SSI_SCR_RE)) == 0)
Markus Pargmann wrote:
To ensure the first edge happens after enabling the TX side of the SSI unit and to trigger the SDMA unit successfully enable the CCSR_SSI_SIER_TDMAE on demand only. To be sure, this patch does the same for the CCSR_SSI_SIER_RDMAE (RX side) bit.
This patch impacts PowerPC as well. Have you tested it on that platform?
On Tue, Nov 12, 2013 at 06:55:05AM -0600, Timur Tabi wrote:
Markus Pargmann wrote:
To ensure the first edge happens after enabling the TX side of the SSI unit and to trigger the SDMA unit successfully enable the CCSR_SSI_SIER_TDMAE on demand only. To be sure, this patch does the same for the CCSR_SSI_SIER_RDMAE (RX side) bit.
This patch impacts PowerPC as well. Have you tested it on that platform?
No, I don't have access on PowerPC hardware, so I didn't test it on PowerPC.
Regards,
Markus Pargmann
On Tue, Nov 12, 2013 at 7:24 AM, Markus Pargmann mpa@pengutronix.de wrote:
No, I don't have access on PowerPC hardware, so I didn't test it on PowerPC.
Can you change the code so that the interrupts are enabled only on i.MX? On PowerPC, the interrupts are used only for debugging.
Hi Timur,
On Tuesday 12 November 2013 21:17:34 Timur Tabi wrote:
On Tue, Nov 12, 2013 at 7:24 AM, Markus Pargmann mpa@pengutronix.de wrote:
No, I don't have access on PowerPC hardware, so I didn't test it on PowerPC.
Can you change the code so that the interrupts are enabled only on i.MX? On PowerPC, the interrupts are used only for debugging.
Hmm, you are right. My patch is partially bogus. It enables the CCSR_SSI_SIER_TDMAE and CCSR_SSI_SIER_RDMAE bits on demand, but this setting will be overwritten by the "write_ssi(sier_bits, &ssi->sier);" just a few lines below. :(
But this patch works, because it removes the CCSR_SSI_SIER_TDMAE and CCSR_SSI_SIER_RDMAE from the SIER_FLAGS macro. Its content is set _unconditionally_ in lines 570 and 575 and is written to the register in line 607. This should hit PowerPC as well.
So, the real (and small) fix is to remove these bits from the SIER_FLAGS macro:
diff --git a/sound/soc/fsl/fsl_ssi.c b/sound/soc/fsl/fsl_ssi.c index c6b7439..f176e34 100644 --- a/sound/soc/fsl/fsl_ssi.c +++ b/sound/soc/fsl/fsl_ssi.c @@ -107,10 +107,10 @@ static inline void write_ssi_mask(u32 __iomem *addr, u32 clear, u32 set) #endif
/* SIER bitflag of interrupts to enable */ -#define SIER_FLAGS (CCSR_SSI_SIER_TFRC_EN | CCSR_SSI_SIER_TDMAE | \ +#define SIER_FLAGS (CCSR_SSI_SIER_TFRC_EN | \ CCSR_SSI_SIER_TIE | CCSR_SSI_SIER_TUE0_EN | \ CCSR_SSI_SIER_TUE1_EN | CCSR_SSI_SIER_RFRC_EN | \ - CCSR_SSI_SIER_RDMAE | CCSR_SSI_SIER_RIE | \ + CCSR_SSI_SIER_RIE | \ CCSR_SSI_SIER_ROE0_EN | CCSR_SSI_SIER_ROE1_EN)
/**
@Markus: can you please test it on your hardware?
Thanks jbe
Jürgen Beisert wrote:
-#define SIER_FLAGS (CCSR_SSI_SIER_TFRC_EN | CCSR_SSI_SIER_TDMAE | \ +#define SIER_FLAGS (CCSR_SSI_SIER_TFRC_EN | \ CCSR_SSI_SIER_TIE | CCSR_SSI_SIER_TUE0_EN | \ CCSR_SSI_SIER_TUE1_EN | CCSR_SSI_SIER_RFRC_EN | \
CCSR_SSI_SIER_RDMAE | CCSR_SSI_SIER_RIE | \
CCSR_SSI_SIER_RIE | \
CCSR_SSI_SIER_ROE0_EN | CCSR_SSI_SIER_ROE1_EN)
/**
NACK, this will break PowerPC.
Hi,
On Tue, Nov 12, 2013 at 02:17:34PM -0600, Timur Tabi wrote:
On Tue, Nov 12, 2013 at 7:24 AM, Markus Pargmann mpa@pengutronix.de wrote:
No, I don't have access on PowerPC hardware, so I didn't test it on PowerPC.
Can you change the code so that the interrupts are enabled only on i.MX? On PowerPC, the interrupts are used only for debugging.
Why is this necessary? The interrupts were enabled for all architectures before.
However, this patch is wrong as Jürgen mentioned, the SIER_FLAGS are written to the SIER register after we enable the DMA enable bits which leads to missing DMA enable bits at the end. I created a new patch to cleanup the whole interrupt flag setup. I will post it later.
Regards,
Markus
Markus Pargmann wrote:
Can you change the code so that the interrupts are enabled only on i.MX? On PowerPC, the interrupts are used only for debugging.
Why is this necessary? The interrupts were enabled for all architectures before.
My concern is with these lines:
write_ssi_mask(&ssi->sier, 0, CCSR_SSI_SIER_TDMAE); ... write_ssi_mask(&ssi->sier, CCSR_SSI_SIER_TDMAE, 0);
This toggles the TDMAE interrupt on PowerPC, which I believe is VITAL to making it work. The SSI and the DMA controller talk to each other automatically, and if I'm reading the documentation correctly, that communication depends on TDMAE and RDMAE.
On Wed, Nov 13, 2013 at 07:10:49AM -0600, Timur Tabi wrote:
Markus Pargmann wrote:
Can you change the code so that the interrupts are enabled only on i.MX? On PowerPC, the interrupts are used only for debugging.
Why is this necessary? The interrupts were enabled for all architectures before.
My concern is with these lines:
write_ssi_mask(&ssi->sier, 0, CCSR_SSI_SIER_TDMAE); ... write_ssi_mask(&ssi->sier, CCSR_SSI_SIER_TDMAE, 0);
This toggles the TDMAE interrupt on PowerPC, which I believe is VITAL to making it work. The SSI and the DMA controller talk to each other automatically, and if I'm reading the documentation correctly, that communication depends on TDMAE and RDMAE.
Is it necessary to keep TDMAE enabled the whole time? In the patch we enable it when playback starts and disable when it stops, so the communication between DMA and SSI shouldn't be disturbed.
Regards,
Markus
Markus Pargmann wrote:
Is it necessary to keep TDMAE enabled the whole time? In the patch we enable it when playback starts and disable when it stops, so the communication between DMA and SSI shouldn't be disturbed.
I will have to test it to be sure, but there's a risk that the bit won't be set early enough. I've seen weird problems like that on my board.
participants (3)
-
Jürgen Beisert
-
Markus Pargmann
-
Timur Tabi