Re: [PATCH 07/28] ALSA: Support nonatomic dmaengine PCMs

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] alsa: missing workqueue cancellation on stream stop in dmaengine PCM
- [Medium] alsa: workqueue coalescing drops DMA completions for nonatomic PCMs
--

commit 377967383b868c91dc9fcc127ab2f8b8e50a24e6
Author: Martin Povišer <povik+lin@cutebit.org>

ALSA: Support nonatomic dmaengine PCMs

Add support for nonatomic dmaengine PCMs.

> diff --git a/sound/core/pcm_dmaengine.c b/sound/core/pcm_dmaengine.c
> index e59b561b849d2..839ff2711f5e1 100644
> --- a/sound/core/pcm_dmaengine.c
> +++ b/sound/core/pcm_dmaengine.c
[ ... ]
> @@ -147,6 +149,24 @@ static void dmaengine_pcm_dma_complete(void *arg)
>  	snd_pcm_period_elapsed(substream);
>  }
>  
> +static void dmaengine_pcm_dma_complete_nonatomic(struct work_struct *wq)
> +{
> +	struct dmaengine_pcm_runtime_data *prtd = container_of(wq,
> +							       struct dmaengine_pcm_runtime_data,
> +							       complete_wq);
> +	struct snd_pcm_substream *substream = prtd->substream;
> +
> +	dmaengine_pcm_dma_complete(substream);
> +}
> +
> +static void dmaengine_pcm_dma_complete_nonatomic_callback(void *arg)
> +{
> +	struct snd_pcm_substream *substream = arg;
> +	struct dmaengine_pcm_runtime_data *prtd = substream_to_prtd(substream);
> +
> +	schedule_work(&prtd->complete_wq);

[Severity: Medium]
Will this workqueue coalescing cause dropped DMA completions?

If multiple DMA period interrupts fire in quick succession before the system
workqueue thread can run, schedule_work() will coalesce the requests so the
work item only runs once.

dmaengine_pcm_dma_complete() blindly increments prtd->pos by exactly one
period:

sound/core/pcm_dmaengine.c:dmaengine_pcm_dma_complete() {
	new_pos = prtd->pos + snd_pcm_lib_period_bytes(substream);
	if (new_pos >= snd_pcm_lib_buffer_bytes(substream))
		new_pos = 0;
	prtd->pos = new_pos;
}

Could this cause prtd->pos to permanently lag behind the actual hardware
position if events are coalesced?

> +}
> +
>  static int dmaengine_pcm_prepare_and_submit(struct snd_pcm_substream *substream)
>  {
>  	struct dmaengine_pcm_runtime_data *prtd = substream_to_prtd(substream);
> @@ -169,7 +189,11 @@ static int dmaengine_pcm_prepare_and_submit(struct snd_pcm_substream *substream)
>  	if (!desc)
>  		return -ENOMEM;
>  
> -	desc->callback = dmaengine_pcm_dma_complete;
> +	if (substream->pcm->nonatomic)
> +		desc->callback = dmaengine_pcm_dma_complete_nonatomic_callback;
> +	else
> +		desc->callback = dmaengine_pcm_dma_complete;
> +

[Severity: High]
Is a workqueue cancellation missing when the stream stops?

If a DMA completion interrupt fires just before SNDRV_PCM_TRIGGER_STOP,
it will schedule complete_wq.

snd_dmaengine_pcm_sync_stop() waits for DMA callbacks via
dmaengine_synchronize(), but doesn't cancel the work item:

sound/core/pcm_dmaengine.c:snd_dmaengine_pcm_sync_stop() {
	status = dmaengine_tx_status(prtd->dma_chan, prtd->cookie, &state);
	if (status != DMA_PAUSED)
		dmaengine_synchronize(prtd->dma_chan);

	return 0;
}

Could this leaked work execute after the stream is prepared or restarted,
erroneously advancing prtd->pos and permanently corrupting the position
state for the new stream cycle?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260920-macaudio-v1-0-741cc20a74e5@gmail.com?part=7