Re: [PATCH] ALSA: compress: allow pause and resume during draining

[Takashi Iwai <tiwai@suse.de>]

On Wed, 30 Sep 2020 11:57:45 +0200,
Jaroslav Kysela wrote:
> 
> Dne 30. 09. 20 v 11:35 Takashi Iwai napsal(a):
> > On Tue, 29 Sep 2020 19:27:17 +0200,
> > Jaroslav Kysela wrote:
> >>
> >> Dne 29. 09. 20 v 9:12 Takashi Iwai napsal(a):
> >>> On Tue, 29 Sep 2020 03:51:35 +0200,
> >>> Gyeongtaek Lee wrote:
> >>>>
> >>>> On 9/28/20 11:35 PM, Pierre-Louis Bossart wrote:
> >>>>> On 9/28/20 6:13 AM, Jaroslav Kysela wrote:
> >>>>>> Dne 28. 09. 20 v 12:50 Gyeongtaek Lee napsal(a):
> >>>>>>> With a stream with low bitrate, user can't pause or resume the stream
> >>>>>>> near the end of the stream because current ALSA doesn't allow it.
> >>>>>>> If the stream has very low bitrate enough to store whole stream into
> >>>>>>> the buffer, user can't do anything except stop the stream and then
> >>>>>>> restart it from the first.
> >>>>>>> If pause and resume is allowed during draining, user experience can be
> >>>>>>> enhanced.
> >>>>>>
> >>>>>> It seems that we need a new state to handle the pause + drain condition for
> >>>>>> this case.
> >>>>>>
> >>>>>> With this proposed change, the pause state in drain is invisible.
> >>>>>
> >>>>> Indeed it's be much nicer to have a new state, e..g 
> >>>>> SNDRV_PCM_STATE_DRAINING_PAUSED.
> >>>> Ok. I will add the new state.
> >>>>>
> >>>>> One concern is that states are defined in uapi/sound/asoc.h, so wouldn't 
> >>>>> this have impacts on userspace as well? We'd change the value of 
> >>>>> SNDRV_PCM_STATE_LAST.
> >>>>>
> >>>> I also agree that adding new state and increase LAST value in the header of uapi
> >>>> could be dangerous. So, I added it to comress_offload.h for now.
> >>>> It could be merged into snd_pcm_state_t in someday with big changes.
> >>>> Could you review the fixed patch below?
> >>
> >> I don't see a big problem to improve the API, but don't forget to increase the
> >> SNDRV_COMPRESS_VERSION, so the user space apps can check for this new behaviour.
> >>
> >>> Hrm, this resulted in rather more complex changes than the original
> >>> patch.  It shows that introducing yet another state is no good idea
> >>> for this particular case.
> >>
> >> I don't think so. The states should be isolated and it's clearly a new state
> >> and the resulted code at least gives a commented idea, what's going on. It
> >> seems that the compress driver state is not exported to the user space at the
> >> moment, so I would consider this extension as harmless. We can add this state
> >> to asound.h so the user space can be updated. We may use this state for the
> >> standard PCM devices one day, too. It makes sense to reserve it sooner than later.
> > 
> > Well, adding a new state can be cumbersome sometimes. For example, the
> > code like below may hit a segfault out of sudden after the upgrade:
> > 
> > 	const char *states[SNDRV_PCM_STATE_LAST + 1] = {
> > 		[SNDRV_PCM_STATE_RUNNING] = "running",
> > 		....
> > 	};
> > 
> > 	printf("current state = %s\n", states[s]);
> > 
> > It's not much frequent breakage, but this can give certainly some
> > incompatibilities even in the source code level.
> 
> alsa-lib has already the correct protection for this case:
> 
> const char *snd_pcm_state_name(const snd_pcm_state_t state)
> {
>         if (state > SND_PCM_STATE_LAST)
>                 return NULL;
>         return snd_pcm_state_names[state];
> }
> 
> If there's no check, it's a clear bug.

That's not what I meant; the code I showed was just an example
implying that the addition of a new state may require the deep code
change that can't be caught by a compiler.  It may be silently
broken.

And imagine the user-space library code that contains handling of the
PCM state.  All those has to be updated as well to deal with a new
state, not only alsa-lib.

IOW, by adding a new item to an exposed attribute like PCM state, the
possibly needed change would be spread over all lib / application
code, and its influence shouldn't be underestimated.  If it were only
some internal change in alsa-lib, I won't be concerned at all.

> > That's the reason I'm reluctant to add a new state unless it's a must.
> > As mentioned, the expected application's behavior is just like the
> > normal pause state, either resuming pause or dropping.  The only case
> > where a new state would help for application is at most that they may
> > foresee beforehand which state it'll go after the resume, to drain or
> > to running.  If this is a must-to-have feature, we can reconsider.
> 
> I don't agree here. It's much better to not hide the state related transitions
> even in the kernel in my eyes. For example drivers may behave differently when
> they resume from running+pause or drain+pause states.

Yes, but that's basically the driver's business.  As mentioned, "if
this is a must-to-have feature" for applications, we'll need to
reconsider.  But it's not clear from the scenario yet.
(FWIW, if any, we may add another function to tell the after-resume
state, too; this might be even safer from the compatibility POV,
although it can be more complicated.)

> The correct SNDRV_PCM_STATE_LAST is just an implementation issue, which can be
> easily solved.

How easily solvable -- that's the question :)


thanks,

Takashi