[alsa-devel] [RFC PATCH] ALSA: hda - hdmi begin to support dynamic PCM assignment

[Takashi Iwai]

On Fri, 13 Nov 2015 16:33:22 +0100,
Yang, Libin wrote:
> 
> 
> > -----Original Message-----
> > From: Takashi Iwai [mailto:tiwai at suse.de]
> > Sent: Friday, November 13, 2015 5:21 PM
> > To: libin.yang at linux.intel.com
> > Cc: alsa-devel at alsa-project.org; mengdong.lin at linux.intel.com; Yang,
> > Libin
> > Subject: Re: [alsa-devel] [RFC PATCH] ALSA: hda - hdmi begin to support
> > dynamic PCM assignment
> > 
> > On Fri, 13 Nov 2015 09:56:30 +0100,
> > libin.yang at linux.intel.com wrote:
> > >
> > > From: Libin Yang <libin.yang at linux.intel.com>
> > >
> > > Begin to support dynamic PCM assignment to pin in
> > > hdmi audio driver.
> > >
> > > This means PCM will not be statically bound with pin.
> > > When there is a monitor connected, the corresponding pin
> > > will try to find a proper PCM to bind. When the monitor
> > > is disconnected, the corresponding pin will unbind
> > > the PCM. This helps to reduce the PCM number when there
> > > are many pins (device entries in DP MST mode) and only
> > > a few of them work at the same time.
> > >
> > > This patch adds the pcm member in struct hdmi_spec_per_pin.
> > > When PCM is dynamically bound to the pin, the member pcm
> > > will pointer to the corresponding pcm_rec[] in hdmi_spec,
> > > which means the hda_pcm is bound to the pin.
> > >
> > > Signed-off-by: Libin Yang <libin.yang at linux.intel.com>
> > > ---
> > >  sound/pci/hda/patch_hdmi.c | 26 +++++++++++++++++++++-----
> > >  1 file changed, 21 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/sound/pci/hda/patch_hdmi.c
> > b/sound/pci/hda/patch_hdmi.c
> > > index f503a88..f3363b8 100644
> > > --- a/sound/pci/hda/patch_hdmi.c
> > > +++ b/sound/pci/hda/patch_hdmi.c
> > > @@ -82,6 +82,7 @@ struct hdmi_spec_per_pin {
> > >  	struct mutex lock;
> > >  	struct delayed_work work;
> > >  	struct snd_kcontrol *eld_ctl;
> > > +	struct hda_pcm *pcm; /* pointer to spec->pcm_rec[n]
> > dynamically*/
> > >  	int repoll_count;
> > >  	bool setup; /* the stream has been set up by prepare callback */
> > >  	int channels; /* current number of channels */
> > > @@ -140,7 +141,8 @@ struct hdmi_spec {
> > >  	struct hdmi_ops ops;
> > >
> > >  	bool dyn_pin_out;
> > > -
> > > +	bool dyn_pcm_assign;
> > > +	struct mutex pcm_lock;
> > >  	/*
> > >  	 * Non-generic VIA/NVIDIA specific
> > >  	 */
> > > @@ -378,13 +380,26 @@ static int hinfo_to_pin_index(struct
> > hda_codec *codec,
> > >  			      struct hda_pcm_stream *hinfo)
> > >  {
> > >  	struct hdmi_spec *spec = codec->spec;
> > > +	struct hdmi_spec_per_pin *per_pin;
> > >  	int pin_idx;
> > >
> > > -	for (pin_idx = 0; pin_idx < spec->num_pins; pin_idx++)
> > > -		if (get_pcm_rec(spec, pin_idx)->stream == hinfo)
> > > -			return pin_idx;
> > > +	if (!spec->dyn_pcm_assign) {
> > > +		for (pin_idx = 0; pin_idx < spec->num_pins; pin_idx++)
> > > +			if (get_pcm_rec(spec, pin_idx)->stream == hinfo)
> > > +				return pin_idx;
> > > +	} else {
> > > +		mutex_lock(&spec->pcm_lock);
> > > +		for (pin_idx = 0; pin_idx < spec->num_pins; pin_idx++) {
> > > +			per_pin = get_pin(spec, pin_idx);
> > > +			if (per_pin->pcm && per_pin->pcm->stream ==
> > hinfo) {
> > > +				mutex_unlock(&spec->pcm_lock);
> > > +				return pin_idx;
> > > +			}
> > > +		}
> > > +		mutex_unlock(&spec->pcm_lock);
> > 
> > You can assign per_pin->pcm statically at init time when
> > dyn_pcm_assign=false.  Then the same code is used in the above.
> 
> Yes, I will change it.
> 
> > 
> > The need of spec->pcm_lock isn't clear at this moment.  I guess you're
> > trying to avoid the race of dynamic pin assignment and release.  But
> > then the code above doesn't protect anything because you return the
> > pin_idx as an available value after the unlock.  Then the pin release
> > can happen after this point while you're accessing with the pin_idx.
> > 
> > So, better to consider the implementation of lock or other mechanism
> > once when the whole picture is ready.
> 
> I add the lock to avoid:
>  if (per_pin->pcm && per_pin->pcm->stream == hinfo)
> per_pin->pcm is not null at first and when judge 
> "per_pin->pcm->stream == hinfo" per_pin->pcm has been null.
> 
> I found it's very hard to protect the pin_idx in the whole process.
> For example, pin is connected when open, but disconnected when
> prepare or close, or vise verse. Or open is called twice at the same
> time, but pin is assigned to 2 different PCMs (hotplug happens).
 
I think it's feasible.  We need to protect it only in each PCM
operation, not through the whole PCM open time, i.e. only in each
prepare(), cleanup() or whatever call.  For example, make a rw
semaphore and do read lock in each operation.  When the PCM assign or
release happens, it takes write lock.  So it waits until the PCM
operation finishes and blocks it while switching.

> Maybe we can keep the lock now, and will refine it later? This lock 
> can make sure when per_pin->pcm->stream == hinfo, it will never
> access invalid memory.

As I wrote, there is no reason to introduce the lock in *this* patch,
as there is no race at all here.  There will be a race in future, but
then let's think the best way at that point.

Also, the mapping between hinfo and per_pin can be done in a different
way instead of loop search.  We may still need some lock, but then the
required lock would be different, too.


Takashi