Re: [PATCH] ALSA: usb-audio: Fix recursive locking on XRUN

[John Keeping]

On Sun, Mar 19, 2023 at 10:15:55AM +0100, Takashi Iwai wrote:
> On Sun, 19 Mar 2023 08:57:03 +0100,
> Takashi Iwai wrote:
> > 
> > On Sun, 19 Mar 2023 04:28:53 +0100,
> > Takashi Sakamoto wrote:
> > > 
> > > Hi,
> > > 
> > > On Sat, Mar 18, 2023 at 09:20:05AM +0900, Takashi Sakamoto wrote:
> > > > On Fri, Mar 17, 2023 at 07:51:27PM +0000, John Keeping wrote:
> > > > > snd_usb_queue_pending_output_urbs() may be called from
> > > > > snd_pcm_ops::ack() which means the PCM stream is locked.
> > > > > 
> > > > > For the normal case where the call back into the PCM core is via
> > > > > prepare_output_urb() the "_under_stream_lock" variant of
> > > > > snd_pcm_period_elapsed() is called, but when an error occurs and the
> > > > > stream is stopped as XRUN then snd_pcm_xrun() tries to recursively lock
> > > > > the stream which results in deadlock.
> > > > > 
> > > > > Follow the example of snd_pcm_period_elapsed() by adding
> > > > > snd_pcm_xrun_under_stream_lock() and use this when the PCM substream
> > > > > lock is already held.
> > > > > 
> > > > > Signed-off-by: John Keeping <john@xxxxxxxxxxxx>
> > > > > ---
> > > > >  include/sound/pcm.h     |  1 +
> > > > >  sound/core/pcm_native.c | 28 ++++++++++++++++++++++++----
> > > > >  sound/usb/endpoint.c    | 18 +++++++++++-------
> > > > >  3 files changed, 36 insertions(+), 11 deletions(-)
> > > >  
> > > > The name of added kernel API implies me that you refer to existent
> > > > 'snd_pcm_period_elapsed_under_stream_lock()' which I added to Linux
> > > > v5.14.
> > > > 
> > > > In my opinion, unlike the version of period elapsed API, the version of
> > > > XRUN API seems not to be necessarily required to ALSA PCM core, since PCM
> > > > device drivers can implement .pointer callback in the part of PCM operation.
> > > > When the callback returns SNDRV_PCM_POS_XRUN, ALSA PCM application get
> > > > occurence of XRUN as a result of any operation relevant to hwptr movement
> > > > (e.g. SNDRV_PCM_IOCTL_HWSYNC).
> > > > 
> > > > Therefore I think it possible to fix the issue without the proposed
> > > > kernel API. I can assume some scenario:
> > > > 
> > > > 1. Failure at tasklet for URB completion
> > > > 
> > > > It is softIRQ context. The stream lock is not acquired. It doesn't
> > > > matter to call current XRUN API.
> > > > 
> > > > 2. Failure at PCM operation called by ALSA PCM application
> > > > 
> > > > It is process context. The stream lock is acquired before calling driver
> > > > code. When detecting any type of failure, driver code stores the state.
> > > > Then .pointer callback should return SNDRV_PCM_POS_XRUNrefering to
> > > > the state.
> > > 
> > > Although being inexperienced to hack driver for USB audio device class,
> > > I attempt to post the patch to fix the issue of recursive stream lock.
> > > I apologies in advance since the patch is not tested yet...
> > > 
> > > The 'in_xrun' member is newly added to 'struct snd_usb_substream'. When
> > > detecting any failure, false is assigned to the member. The assignment
> > > is expected to be done in both softIRQ context, and process context with
> > > stream lock, thus no need to take care of cocurrent access (e.g. by usage
> > > of WRITE_ONCE/READ_ONCE).
> > > 
> > > Typical ALSA PCM application periodically calls PCM operation which calls
> > > .pointer in driver code. As I described, returning SNDRV_PCM_POS_XRUN
> > > takes ALSA PCM core to handle XRUN state of PCM substream in the timing.
> > > 
> > > The negative point of the patch is the delay of XRUN notification to user
> > > space application. In the point, I think the new kernel API introduced by
> > > your patch has advantage.
> > > 
> > > The in_xrun member can be replaced with a kind of EP_STATE_
> > > enumerations; i.e. EP_STATE_XRUN. In the case, we need some care so that
> > > the state should be referred from pcm.c.
> > 
> > Thanks for the patch.  That would work, but the shortcoming side of
> > this implementation is that it misses stopping / reporting the error
> > immediately but waiting for the next pointer update.
> > 
> > It might be simpler if we perform the xrun handling in the caller
> > side, i.e. a change like below:
> > 
> > --- a/sound/core/pcm_lib.c
> > +++ b/sound/core/pcm_lib.c
> > @@ -2155,6 +2155,8 @@ int pcm_lib_apply_appl_ptr(struct snd_pcm_substream *substream,
> >  		ret = substream->ops->ack(substream);
> >  		if (ret < 0) {
> >  			runtime->control->appl_ptr = old_appl_ptr;
> > +			if (ret == -EPIPE)
> > +				__snd_pcm_xrun(substream);
> >  			return ret;
> >  		}
> >  	}
> > 
> > ... and let the caller returning -EPIPE for XRUN:
> 
> and that misses the XRUN in the case of non-stream-lock.
> A revised version is below.

Yes, it looks like this also solves the problem.  If you roll this into
a proper patch feel free to add:

Tested-by: John Keeping <john@xxxxxxxxxxxx>

> 
> -- 8< --
> --- a/sound/core/pcm_lib.c
> +++ b/sound/core/pcm_lib.c
> @@ -2155,6 +2155,8 @@ int pcm_lib_apply_appl_ptr(struct snd_pcm_substream *substream,
>  		ret = substream->ops->ack(substream);
>  		if (ret < 0) {
>  			runtime->control->appl_ptr = old_appl_ptr;
> +			if (ret == -EPIPE)
> +				__snd_pcm_xrun(substream);
>  			return ret;
>  		}
>  	}
> --- a/sound/usb/endpoint.c
> +++ b/sound/usb/endpoint.c
> @@ -455,8 +455,8 @@ static void push_back_to_ready_list(struct snd_usb_endpoint *ep,
>   * This function is used both for implicit feedback endpoints and in low-
>   * latency playback mode.
>   */
> -void snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
> -				       bool in_stream_lock)
> +int snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
> +				      bool in_stream_lock)
>  {
>  	bool implicit_fb = snd_usb_endpoint_implicit_feedback_sink(ep);
>  
> @@ -480,7 +480,7 @@ void snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
>  		spin_unlock_irqrestore(&ep->lock, flags);
>  
>  		if (ctx == NULL)
> -			return;
> +			break;
>  
>  		/* copy over the length information */
>  		if (implicit_fb) {
> @@ -495,11 +495,14 @@ void snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
>  			break;
>  		if (err < 0) {
>  			/* push back to ready list again for -EAGAIN */
> -			if (err == -EAGAIN)
> +			if (err == -EAGAIN) {
>  				push_back_to_ready_list(ep, ctx);
> -			else
> +				break;
> +			}
> +
> +			if (!in_stream_lock)
>  				notify_xrun(ep);
> -			return;
> +			return -EPIPE;
>  		}
>  
>  		err = usb_submit_urb(ctx->urb, GFP_ATOMIC);
> @@ -507,13 +510,16 @@ void snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
>  			usb_audio_err(ep->chip,
>  				      "Unable to submit urb #%d: %d at %s\n",
>  				      ctx->index, err, __func__);
> -			notify_xrun(ep);
> -			return;
> +			if (!in_stream_lock)
> +				notify_xrun(ep);
> +			return -EPIPE;
>  		}
>  
>  		set_bit(ctx->index, &ep->active_mask);
>  		atomic_inc(&ep->submitted_urbs);
>  	}
> +
> +	return 0;
>  }
>  
>  /*
> --- a/sound/usb/endpoint.h
> +++ b/sound/usb/endpoint.h
> @@ -52,7 +52,7 @@ int snd_usb_endpoint_implicit_feedback_sink(struct snd_usb_endpoint *ep);
>  int snd_usb_endpoint_next_packet_size(struct snd_usb_endpoint *ep,
>  				      struct snd_urb_ctx *ctx, int idx,
>  				      unsigned int avail);
> -void snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
> -				       bool in_stream_lock);
> +int snd_usb_queue_pending_output_urbs(struct snd_usb_endpoint *ep,
> +				      bool in_stream_lock);
>  
>  #endif /* __USBAUDIO_ENDPOINT_H */
> --- a/sound/usb/pcm.c
> +++ b/sound/usb/pcm.c
> @@ -1639,7 +1639,7 @@ static int snd_usb_pcm_playback_ack(struct snd_pcm_substream *substream)
>  	 * outputs here
>  	 */
>  	if (!ep->active_mask)
> -		snd_usb_queue_pending_output_urbs(ep, true);
> +		return snd_usb_queue_pending_output_urbs(ep, true);
>  	return 0;
>  }
>