pcm_meter.c issue at s16_update

[Pavel Hofman]

Dne 09. 08. 20 v 22:29 Jaroslav Kysela napsal(a):
> Dne 09. 08. 20 v 9:05 Pavel Hofman napsal(a):
>> Dne 03. 08. 20 v 12:48 Pavel Hofman napsal(a):
>>>
>>>
>>> Dne 03. 08. 20 v 9:22 Jaroslav Kysela napsal(a):
>>>> Dne 03. 08. 20 v 8:17 Takashi Iwai napsal(a):
>>>>> On Sun, 02 Aug 2020 19:50:44 +0200,
>>>>>>>
>>>>>>> Optionally the second case could be handled just like the first
>>>>>>> case by
>>>>>>> resetting s16->old, assuming the boundary wrap occurs very
>>>>>>> infrequently.
>>>>>>
>>>>>> The following patch is tested to work OK, no CPU peaks and no meter
>>>>>> output glitches when the size < 0 condition occurs:
>>>>>>
>>>>>> diff --git a/src/pcm/pcm_meter.c b/src/pcm/pcm_meter.c
>>>>>> index 20b41876..48df5945 100644
>>>>>> --- a/src/pcm/pcm_meter.c
>>>>>> +++ b/src/pcm/pcm_meter.c
>>>>>> @@ -1098,8 +1098,15 @@ static void s16_update(snd_pcm_scope_t *scope)
>>>>>>           snd_pcm_sframes_t size;
>>>>>>           snd_pcm_uframes_t offset;
>>>>>>           size = meter->now - s16->old;
>>>>>> -       if (size < 0)
>>>>>> -               size += spcm->boundary;
>>>>>> +       if (size < 0) {
>>>>>> +               /**
>>>>>> +                * Application pointer adjusted for delay (meter->now)
>>>>>> has dropped compared
>>>>>> +                * to the previous update cycle. Either spcm->boundary
>>>>>> wraparound, pcm rewinding,
>>>>>> +                * or pcm restart without s16->old properly reset.
>>>>>> +                * In any case the safest solution is skipping this
>>>>>> conversion cycle.
>>>>>> +                */
>>>>>> +               size = 0;
>>>>>> +       }
>>>>>>           offset = s16->old % meter->buf_size;
>>>>>>           while (size > 0) {
>>>>>>                   snd_pcm_uframes_t frames = size;
>>>>>>
>>>>>>
>>>>>>
>>>>>> Please will you accept this (workaround) bugfix? If so, I would send a
>>>>>> proper patch.
>>>>>
>>>>> It looks OK, at least this must be safe.
>>>>> So yes, I'll happily apply if you submit a proper patch.
>>>>
>>>> It would be probably better to check against the boundary / 2 value to
>>>> check
>>>> correctly the boundary wrap instead to drop all negative size values:
>>>>
>>>>     if (size < 0) {
>>>>        if (size < -(spcm->boundary / 2))
>>>>           size += spcm->boundary;
>>>>        else
>>>>           size = 0;
>>>>     }
>>>
>>> Is there a reliable way to detect the boundary wraparound, at best using
>>> some dedicated API? I could find any, IMO the wraparound does not create
>>> any notification. The check is OK for a rewind, half of boundary is
>>> usually a very large number too. I am not sure what would happen at
>>> reset when application pointer was already past the boundary half - see
>>> below.
> 
> Yes, it's a good argument. In this case, the s16->old value is not properly
> synced during the reset operation, otherwise the boundary / 2 threshold
> (change limit) is sufficient to detect the boundary wrap.
> 
>>>> The "hidden" pcm restart referred in the comment should not occur,
>>>> otherwise
>>>> it's another bug somewhere.
>>>
>>> I do not know the exact moments when plugin API methods are called. The
>>> fact is Takashi's suggestion to call s16 reset explicitely in
>>> snd_pcm_meter_reset created this order:
>>>
>>> snd_pcm_meter_reset -> s16->reset
>>> s16_update: meter->now 22751, s16->old 22751, size 0
>>> s16_update: meter->now 839, s16->old 22751, size -21912
>>>
>>> I.e. AFTER resetting meter/s16 the variable meter->now was still at the
>>> original large 22751 (with s16->old equal to its value due to
>>> s16->reset). The value of meter->now was reset to 839 (= app pointer -
>>> delay) only in the next call of s16_update (when s16->old was still the
>>> previous old value => size < 0 => huge size => high CPU load).  From
>>> this I kind of conclude that the reset is buggy. Maybe the reset code
>>> should re-calculate meter->now = appl.pointer - delay before aligning
>>> s16->old = meter->now.
>>>
>>> Nevertheless all this (except for the boundary wraparound) would result
>>> in the same size = 0, thus skipping samples from the last cycle, just
>>> like what the proposed patch does.
>>>
>>>
>>
>> Please can we reach a decision and close the problem so that affected
>> use cases do not have to be patched with the next the alsa-lib version?
> 
> I think that this problem should be fixed for reset and rewind separately. The
> meter->reset should be set in snd_pcm_meter_reset() inside the running_mutex
> lock to serialize correctly the update operations in the
> snd_pcm_meter_thread(). And perhaps, we can follow this logic for the rewind.
> 
> I mean, we should ensure to call the s16->reset at the proper time to avoid
> broken old/now combinations inside the scope "clients".
> 
> Your proposed solution is just a workaround.

I am well aware of that. The main cause of the problem is that the 
existing code assumes that a drop in the meter->now value is caused by 
the pcm->boundary wraparound. Only for that particular case the existing 
  size += spcm->boundary code is correct, for all the other cases it is 
grossly wrong, locking the thread for many tens of seconds and jamming 
CPU. If there was a callback or some other way to signal the boundary 
wraparound that "dangerous" code would be called only for that special 
case (which is extremely rare in usual setups).

I do not know all cases when the meter->now can drop. Reset, rewind, any 
other (xrun)? If a single case is omitted, the same problem prevails.

No problem with resetting where appropriate, but still I would suggest 
to not keep size += spcm->boundary in the s16_update as is now.

Regards,

Pavel.