[RFC][PATCH] ALSA: pcm_dmaengine: Use more shorter periods for SEGMENT granularity PCM DMA

[RFC][PATCH] ALSA: pcm_dmaengine: Use more shorter periods for SEGMENT granularity PCM DMA

[Marek Vasut]

The segment granularity PCM DMA updates PCM pointer at the end of every
segment. In case the segments are too long, which happens with pipewire,
it may happen that pcm_lib wait_for_avail() will time out while waiting
for PCM pointer update and playback will fail with IO error.

In case of SEGMENT granularity, use 8 shorter segments over the default
2 long segments set in dmaengine_pcm_set_runtime_hwparams(). That gives
the DMA engine some chance to update the PCM pointer every once in a
while and prevent the timeout.

Signed-off-by: Marek Vasut <marex@nabladev.com>
---
Cc: Jaroslav Kysela <perex@perex.cz>
Cc: Lars-Peter Clausen <lars@metafoo.de>
Cc: Mark Brown <broonie@kernel.org>
Cc: Michal Simek <monstr@monstr.eu>
Cc: Takashi Iwai <tiwai@suse.com>
Cc: linux-sound@vger.kernel.org
---
 sound/core/pcm_dmaengine.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/sound/core/pcm_dmaengine.c b/sound/core/pcm_dmaengine.c
index 1306b04be171a..092ce958e1aee 100644
--- a/sound/core/pcm_dmaengine.c
+++ b/sound/core/pcm_dmaengine.c
@@ -424,6 +424,10 @@ int snd_dmaengine_pcm_refine_runtime_hwparams(
 			hw->info |= SNDRV_PCM_INFO_PAUSE | SNDRV_PCM_INFO_RESUME;
 		if (dma_caps.residue_granularity <= DMA_RESIDUE_GRANULARITY_SEGMENT)
 			hw->info |= SNDRV_PCM_INFO_BATCH;
+		if (dma_caps.residue_granularity == DMA_RESIDUE_GRANULARITY_SEGMENT) {
+			hw->periods_min = 8;
+			hw->period_bytes_max /= 8;
+		}
 
 		if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK)
 			addr_widths = dma_caps.dst_addr_widths;
-- 
2.53.0

Re: [RFC][PATCH] ALSA: pcm_dmaengine: Use more shorter periods for SEGMENT granularity PCM DMA

[Takashi Iwai]

On Fri, 27 Mar 2026 15:29:43 +0100,
Marek Vasut wrote:
> 
> The segment granularity PCM DMA updates PCM pointer at the end of every
> segment. In case the segments are too long, which happens with pipewire,
> it may happen that pcm_lib wait_for_avail() will time out while waiting
> for PCM pointer update and playback will fail with IO error.
> 
> In case of SEGMENT granularity, use 8 shorter segments over the default
> 2 long segments set in dmaengine_pcm_set_runtime_hwparams(). That gives
> the DMA engine some chance to update the PCM pointer every once in a
> while and prevent the timeout.
> 
> Signed-off-by: Marek Vasut <marex@nabladev.com>
> ---
> Cc: Jaroslav Kysela <perex@perex.cz>
> Cc: Lars-Peter Clausen <lars@metafoo.de>
> Cc: Mark Brown <broonie@kernel.org>
> Cc: Michal Simek <monstr@monstr.eu>
> Cc: Takashi Iwai <tiwai@suse.com>
> Cc: linux-sound@vger.kernel.org
> ---
>  sound/core/pcm_dmaengine.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/sound/core/pcm_dmaengine.c b/sound/core/pcm_dmaengine.c
> index 1306b04be171a..092ce958e1aee 100644
> --- a/sound/core/pcm_dmaengine.c
> +++ b/sound/core/pcm_dmaengine.c
> @@ -424,6 +424,10 @@ int snd_dmaengine_pcm_refine_runtime_hwparams(
>  			hw->info |= SNDRV_PCM_INFO_PAUSE | SNDRV_PCM_INFO_RESUME;
>  		if (dma_caps.residue_granularity <= DMA_RESIDUE_GRANULARITY_SEGMENT)
>  			hw->info |= SNDRV_PCM_INFO_BATCH;
> +		if (dma_caps.residue_granularity == DMA_RESIDUE_GRANULARITY_SEGMENT) {
> +			hw->periods_min = 8;
> +			hw->period_bytes_max /= 8;
> +		}

Overriding with the hard-coded values at hw_params refine doesn't look
like a right way.  Those values are set and dealt by the drivers in
general, as they are pretty hardware-specific.  You can't guarantee
this magic number 8 working on all available hardware.


thanks,

Takashi

Re: [RFC][PATCH] ALSA: pcm_dmaengine: Use more shorter periods for SEGMENT granularity PCM DMA

[Marek Vasut]

On 3/30/26 9:41 AM, Takashi Iwai wrote:
> On Fri, 27 Mar 2026 15:29:43 +0100,
> Marek Vasut wrote:
>>
>> The segment granularity PCM DMA updates PCM pointer at the end of every
>> segment. In case the segments are too long, which happens with pipewire,
>> it may happen that pcm_lib wait_for_avail() will time out while waiting
>> for PCM pointer update and playback will fail with IO error.
>>
>> In case of SEGMENT granularity, use 8 shorter segments over the default
>> 2 long segments set in dmaengine_pcm_set_runtime_hwparams(). That gives
>> the DMA engine some chance to update the PCM pointer every once in a
>> while and prevent the timeout.
>>
>> Signed-off-by: Marek Vasut <marex@nabladev.com>
>> ---
>> Cc: Jaroslav Kysela <perex@perex.cz>
>> Cc: Lars-Peter Clausen <lars@metafoo.de>
>> Cc: Mark Brown <broonie@kernel.org>
>> Cc: Michal Simek <monstr@monstr.eu>
>> Cc: Takashi Iwai <tiwai@suse.com>
>> Cc: linux-sound@vger.kernel.org
>> ---
>>   sound/core/pcm_dmaengine.c | 4 ++++
>>   1 file changed, 4 insertions(+)
>>
>> diff --git a/sound/core/pcm_dmaengine.c b/sound/core/pcm_dmaengine.c
>> index 1306b04be171a..092ce958e1aee 100644
>> --- a/sound/core/pcm_dmaengine.c
>> +++ b/sound/core/pcm_dmaengine.c
>> @@ -424,6 +424,10 @@ int snd_dmaengine_pcm_refine_runtime_hwparams(
>>   			hw->info |= SNDRV_PCM_INFO_PAUSE | SNDRV_PCM_INFO_RESUME;
>>   		if (dma_caps.residue_granularity <= DMA_RESIDUE_GRANULARITY_SEGMENT)
>>   			hw->info |= SNDRV_PCM_INFO_BATCH;
>> +		if (dma_caps.residue_granularity == DMA_RESIDUE_GRANULARITY_SEGMENT) {
>> +			hw->periods_min = 8;
>> +			hw->period_bytes_max /= 8;
>> +		}
> 
> Overriding with the hard-coded values at hw_params refine doesn't look
> like a right way.  Those values are set and dealt by the drivers in
> general, as they are pretty hardware-specific.  You can't guarantee
> this magic number 8 working on all available hardware.
This is very much why I sent this patch as an RFC .

If these limits should be set per DMA driver, do you happen to have a 
suggestion where to plumb it in ?