[PATCH v2 1/3] ALSA: pcm: add snd_pcm_period_elapsed() variant without acquiring lock of PCM substream

[Takashi Iwai]

On Thu, 10 Jun 2021 12:12:43 +0200,
Takashi Sakamoto wrote:
> 
> On Thu, Jun 10, 2021 at 10:36:57AM +0200, Takashi Iwai wrote:
> > On Thu, 10 Jun 2021 10:26:22 +0200,
> > Takashi Sakamoto wrote:
> > > 
> > > On Thu, Jun 10, 2021 at 10:08:39AM +0200, Takashi Iwai wrote:
> > > > On Thu, 10 Jun 2021 10:05:21 +0200,
> > > > Takashi Sakamoto wrote:
> > > > > 
> > > > > On Thu, Jun 10, 2021 at 09:39:37AM +0200, Takashi Iwai wrote:
> > > > > > On Thu, 10 Jun 2021 01:16:23 +0200,
> > > > > > Takashi Sakamoto wrote:
> > > > > > > 
> > > > > > > On Wed, Jun 09, 2021 at 05:27:29PM +0200, Takashi Iwai wrote:
> > > > > > > > On Wed, 09 Jun 2021 16:31:43 +0200,
> > > > > > > > Takashi Sakamoto wrote:
> > > > > > > > > diff --git a/sound/core/pcm_lib.c b/sound/core/pcm_lib.c
> > > > > > > > > index b7e3d8f44511..3488ec1e3674 100644
> > > > > > > > > --- a/sound/core/pcm_lib.c
> > > > > > > > > +++ b/sound/core/pcm_lib.c
> > > > > > > > > @@ -1778,27 +1778,41 @@ int snd_pcm_lib_ioctl(struct snd_pcm_substream *substream,
> > > > > > > > >  EXPORT_SYMBOL(snd_pcm_lib_ioctl);
> > > > > > > > >  
> > > > > > > > >  /**
> > > > > > > > > - * snd_pcm_period_elapsed - update the pcm status for the next period
> > > > > > > > > - * @substream: the pcm substream instance
> > > > > > > > > + * snd_pcm_period_elapsed_under_stream_lock() - update the status of runtime for the next period
> > > > > > > > > + *						under acquired lock of PCM substream.
> > > > > > > > > + * @substream: the instance of pcm substream.
> > > > > > > > > + *
> > > > > > > > > + * This function is called when the batch of audio data frames as the same size as the period of
> > > > > > > > > + * buffer is already processed in audio data transmission.
> > > > > > > > > + *
> > > > > > > > > + * The call of function updates the status of runtime with the latest position of audio data
> > > > > > > > > + * transmission, checks overrun and underrun over buffer, awaken user processes from waiting for
> > > > > > > > > + * available audio data frames, sampling audio timestamp, and performs stop or drain the PCM
> > > > > > > > > + * substream according to configured threshold.
> > > > > > > > > + *
> > > > > > > > > + * The function is intended to use for the case that PCM driver operates audio data frames under
> > > > > > > > > + * acquired lock of PCM substream; e.g. in callback of any operation of &snd_pcm_ops in process
> > > > > > > > > + * context. In any interrupt context, it's preferrable to use ``snd_pcm_period_elapsed()`` instead
> > > > > > > > > + * since lock of PCM substream should be acquired in advance.
> > > > > > > > >   *
> > > > > > > > > - * This function is called from the interrupt handler when the
> > > > > > > > > - * PCM has processed the period size.  It will update the current
> > > > > > > > > - * pointer, wake up sleepers, etc.
> > > > > > > > > + * Developer should pay enough attention that some callbacks in &snd_pcm_ops are done by the call of
> > > > > > > > > + * function:
> > > > > > > > >   *
> > > > > > > > > - * Even if more than one periods have elapsed since the last call, you
> > > > > > > > > - * have to call this only once.
> > > > > > > > > + * - .pointer - to retrieve current position of audio data transmission by frame count or XRUN state.
> > > > > > > > > + * - .trigger - with SNDRV_PCM_TRIGGER_STOP at XRUN or DRAINING state.
> > > > > > > > > + * - .get_time_info - to retrieve audio time stamp if needed.
> > > > > > > > > + *
> > > > > > > > > + * Even if more than one periods have elapsed since the last call, you have to call this only once.
> > > > > > > > > + *
> > > > > > > > > + * Context: Any context in which lock of PCM substream is already acquired. This function may not
> > > > > > > > > + * sleep.
> > > > > > > > 
> > > > > > > > Hm, this text still remains here.  Overlooked?
> > > > > > > 
> > > > > > > It's my intension for documentation of
> > > > > > > snd_pcm_period_elapsed_under_stream_lock() since it's expected to call
> > > > > > > it under acquired lock. Its implementation doesn't yield processor
> > > > > > > voluntarily by itself. If it yielded, it would depend on implementation
> > > > > > > of each driver for struct snd_pcm_ops.{pointer, trigger, get_time_info},
> > > > > > > but it's not preferable implementation of driver, in my opinion.
> > > > > > 
> > > > > > My point is again about the sleep.  This function may sleep in the
> > > > > > nonatomic mode.  The type of the PCM stream lock depends on it.
> > > > > 
> > > > > Would I simply request you to show how the added function yields except
> > > > > for the driver implementation? The lock of stream is expected to be
> > > > > acquired already.
> > > > 
> > > > In the nonatomic mode, the PCM stream lock is a mutex (no
> > > > spin_lock_irqsave), hence it can sleep -- which contradicts with the
> > > > added description above.
> > > > 
> > > > Or do I misunderstand your question...? 
> > > 
> > > Thanks to clarify the role of PCM stream lock, and I'm ease that we have
> > > the same understanding about the lock.
> > > 
> > > Here, let us see deleted/added line again.
> > > 
> > > > diff --git a/sound/core/pcm_lib.c b/sound/core/pcm_lib.c
> > > > index b7e3d8f44511..3488ec1e3674 100644
> > > > --- a/sound/core/pcm_lib.c
> > > > +++ b/sound/core/pcm_lib.c
> > > > @@ -1778,27 +1778,41 @@ int snd_pcm_lib_ioctl(struct snd_pcm_substream *substream,
> > > >  EXPORT_SYMBOL(snd_pcm_lib_ioctl);
> > > >  
> > > >  /**
> > > > - * snd_pcm_period_elapsed - update the pcm status for the next period
> > > > - * @substream: the pcm substream instance
> > > > + * snd_pcm_period_elapsed_under_stream_lock() - update the status of runtime for the next period
> > > > + *						under acquired lock of PCM substream.
> > > > + ...
> > > > + * Context: Any context in which lock of PCM substream is already acquired. This function may not
> > > > + * sleep.
> > > 
> > > The issued documentation is for the new function. Inner the function, the
> > > lock of PCM substream is not acquired again since it causes dead lock
> > > (it's not nest-able lock) regardless of usage of mutex or spin_lock.
> > > 
> > > The well-known function, snd_pcm_period_elapsed(), is rewritten to call
> > > the new function between lock/unlock operations:
> > > 
> > > ->snd_pcm_period_elapsed()
> > >   ->snd_pcm_stream_lock_irqsave()
> > >   ->snd_pcm_period_elapsed_under_stream_lock()
> > >   ->snd_pcm_stream_unlock_irqrestore()
> > > 
> > > Or the new function can acquire the lock somewhere I overlook? However I
> > > think it is unlikely since it necessarily causes dead lock or corruption
> > > of irq context...
> > 
> > Again, my *only* point is about the sleep.  You addition was:
> > 
> > + * Context: Any context in which lock of PCM substream is already acquired. This function may not
> > + * sleep.
> > 
> > where "This function may not sleep" is stated incorrectly.
> 
> Hm. Would I request you to show the detail case that the call of function
> (snd_pcm_period_elapsed_under_stream_lock()) goes sleep except for
> driver-side implementation of snd_pcm_ops.{pointer, trigger,
> get_time_info}? At least, in callgraph I find no function call to
> yield...

True.  But the fact that those callbacks may sleep means that the
function would go sleeping after all.


Takashi