diff mbox

alsa-lib: Provide a CLOCK_MONOTONIC_RAW timestamp type

Message ID 1404831152-8064-1-git-send-email-broonie@kernel.org
State New
Headers show

Commit Message

Mark Brown July 8, 2014, 2:52 p.m. UTC
From: Mark Brown <broonie@linaro.org>

For applications which need to synchronise with external timebases such
as broadcast TV applications the kernel monotonic time is not optimal as
it includes adjustments from NTP and so may still include discontinuities
due to that. A raw monotonic time which does not include any adjustments
is available in the kernel from getrawmonotonic() so provide userspace with
a new timestamp type SNDRV_PCM_TSTAMP_TYPE_MONOTONIC_RAW which provides
timestamps based on this as an option.

Reported-by: Daniel Thompson <daniel.thompson@linaro.org>
Signed-off-by: Mark Brown <broonie@linaro.org>
---
 include/sound/asound.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

Takashi Iwai July 9, 2014, 1:38 p.m. UTC | #1
At Tue,  8 Jul 2014 16:52:32 +0200,
Mark Brown wrote:
> 
> From: Mark Brown <broonie@linaro.org>
> 
> For applications which need to synchronise with external timebases such
> as broadcast TV applications the kernel monotonic time is not optimal as
> it includes adjustments from NTP and so may still include discontinuities
> due to that. A raw monotonic time which does not include any adjustments
> is available in the kernel from getrawmonotonic() so provide userspace with
> a new timestamp type SNDRV_PCM_TSTAMP_TYPE_MONOTONIC_RAW which provides
> timestamps based on this as an option.
> 
> Reported-by: Daniel Thompson <daniel.thompson@linaro.org>
> Signed-off-by: Mark Brown <broonie@linaro.org>

Currently, the timestamp mode is set implicitly in alsa-lib pcm_hw.c:
- When kernel PCM protocol version is high enough, alsa-lib hw prefers
  the monotonic always (if available), then set pcm->monotonic = 1.
- Application can ask whether the current timestamp is monotonic or
  not via snd_pcm_hw_params_is_monotonic().
So, only adding the flag above doesn't suffice.  If we need to add a
new mode, the API has to be extended as well.

But how?  The current API assumes that the monotonic mode was already
determined before hw_params.  We may add a set of new hw_params get
and set calls for tstamp mode while keeping the old API.  This would
be one option.  Another option would be to add a new PCM open flag
SND_PCM_TSTAMP_MONOTONIC_RAW, and snd_pcm_hw_params_is_monotonic_raw()
function.  The latter is easier (a simpler addition), while the former
is more extensible to newer formats in future.

Comments?

(I guess tinyalsa has a different story, but let's align genuine
 alsa-lib first.)


Takashi


> ---
>  include/sound/asound.h | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/include/sound/asound.h b/include/sound/asound.h
> index 1774a5c..9061cdd 100644
> --- a/include/sound/asound.h
> +++ b/include/sound/asound.h
> @@ -457,7 +457,8 @@ struct snd_xfern {
>  enum {
>  	SNDRV_PCM_TSTAMP_TYPE_GETTIMEOFDAY = 0,	/* gettimeofday equivalent */
>  	SNDRV_PCM_TSTAMP_TYPE_MONOTONIC,	/* posix_clock_monotonic equivalent */
> -	SNDRV_PCM_TSTAMP_TYPE_LAST = SNDRV_PCM_TSTAMP_TYPE_MONOTONIC,
> +	SNDRV_PCM_TSTAMP_TYPE_MONOTONIC_RAW,	/* monotonic_raw (no NTP) */
> +	SNDRV_PCM_TSTAMP_TYPE_LAST = SNDRV_PCM_TSTAMP_TYPE_MONOTONIC_RAW,
>  };
>  
>  /* channel positions */
> -- 
> 2.0.0
>
Clemens Ladisch July 9, 2014, 7:40 p.m. UTC | #2
Takashi Iwai wrote:
> Currently, the timestamp mode is set implicitly in alsa-lib pcm_hw.c:
> - When kernel PCM protocol version is high enough, alsa-lib hw prefers
>   the monotonic always (if available), then set pcm->monotonic = 1.
> - Application can ask whether the current timestamp is monotonic or
>   not via snd_pcm_hw_params_is_monotonic().
> So, only adding the flag above doesn't suffice.  If we need to add a
> new mode, the API has to be extended as well.
>
> But how?  The current API assumes that the monotonic mode was already
> determined before hw_params.  We may add a set of new hw_params get
> and set calls for tstamp mode while keeping the old API.  This would
> be one option.  Another option would be to add a new PCM open flag
> SND_PCM_TSTAMP_MONOTONIC_RAW, and snd_pcm_hw_params_is_monotonic_raw()
> function.  The latter is easier (a simpler addition), while the former
> is more extensible to newer formats in future.

These timestamps cannot be handled by hardware drivers; they are always
read by the ALSA framework, in software.  Furthermore, switching between
MONOTONIC and MONOTONIC_RAW is possible at any time.  Therefore, the
timestamp mode should be made a part of sw_params; just add a field
named tstamp_mode ...


Regards,
Clemens
Takashi Iwai July 10, 2014, 10:22 a.m. UTC | #3
At Wed, 09 Jul 2014 21:40:02 +0200,
Clemens Ladisch wrote:
> 
> Takashi Iwai wrote:
> > Currently, the timestamp mode is set implicitly in alsa-lib pcm_hw.c:
> > - When kernel PCM protocol version is high enough, alsa-lib hw prefers
> >   the monotonic always (if available), then set pcm->monotonic = 1.
> > - Application can ask whether the current timestamp is monotonic or
> >   not via snd_pcm_hw_params_is_monotonic().
> > So, only adding the flag above doesn't suffice.  If we need to add a
> > new mode, the API has to be extended as well.
> >
> > But how?  The current API assumes that the monotonic mode was already
> > determined before hw_params.  We may add a set of new hw_params get
> > and set calls for tstamp mode while keeping the old API.  This would
> > be one option.  Another option would be to add a new PCM open flag
> > SND_PCM_TSTAMP_MONOTONIC_RAW, and snd_pcm_hw_params_is_monotonic_raw()
> > function.  The latter is easier (a simpler addition), while the former
> > is more extensible to newer formats in future.
> 
> These timestamps cannot be handled by hardware drivers; they are always
> read by the ALSA framework, in software.  Furthermore, switching between
> MONOTONIC and MONOTONIC_RAW is possible at any time.  Therefore, the
> timestamp mode should be made a part of sw_params; just add a field
> named tstamp_mode ...

Sounds reasonable.  I'll respin the kernel patches as well.


Takashi
Pierre-Louis Bossart July 10, 2014, 3:10 p.m. UTC | #4
On 7/9/14, 2:40 PM, Clemens Ladisch wrote:
> Takashi Iwai wrote:
>> Currently, the timestamp mode is set implicitly in alsa-lib pcm_hw.c:
>> - When kernel PCM protocol version is high enough, alsa-lib hw prefers
>>    the monotonic always (if available), then set pcm->monotonic = 1.
>> - Application can ask whether the current timestamp is monotonic or
>>    not via snd_pcm_hw_params_is_monotonic().
>> So, only adding the flag above doesn't suffice.  If we need to add a
>> new mode, the API has to be extended as well.
>>
>> But how?  The current API assumes that the monotonic mode was already
>> determined before hw_params.  We may add a set of new hw_params get
>> and set calls for tstamp mode while keeping the old API.  This would
>> be one option.  Another option would be to add a new PCM open flag
>> SND_PCM_TSTAMP_MONOTONIC_RAW, and snd_pcm_hw_params_is_monotonic_raw()
>> function.  The latter is easier (a simpler addition), while the former
>> is more extensible to newer formats in future.
>
> These timestamps cannot be handled by hardware drivers; they are always
> read by the ALSA framework, in software.  Furthermore, switching between
> MONOTONIC and MONOTONIC_RAW is possible at any time.  Therefore, the
> timestamp mode should be made a part of sw_params; just add a field
> named tstamp_mode ...

Humm... why would anyone switch modes at run time during 
playback/capture? I have seen timestamps being used mainly to track that 
playback/capture is happening as predicted, improve A/V sync or sleep 
for a predictable time with interrupts disabled (PulseAudio, Android, 
ChromeOS, etc). NTP adjustments are really adding noise more than 
information for those usages. I have yet to see a case where the use of 
those NTP adjustments would matter for userspace, and for now I really 
don't see the point of making this dynamically configurable as a 
software parameter. I would personally make this new mode the default.
-Pierre
Takashi Iwai July 10, 2014, 3:33 p.m. UTC | #5
At Thu, 10 Jul 2014 10:10:56 -0500,
Pierre-Louis Bossart wrote:
> 
> On 7/9/14, 2:40 PM, Clemens Ladisch wrote:
> > Takashi Iwai wrote:
> >> Currently, the timestamp mode is set implicitly in alsa-lib pcm_hw.c:
> >> - When kernel PCM protocol version is high enough, alsa-lib hw prefers
> >>    the monotonic always (if available), then set pcm->monotonic = 1.
> >> - Application can ask whether the current timestamp is monotonic or
> >>    not via snd_pcm_hw_params_is_monotonic().
> >> So, only adding the flag above doesn't suffice.  If we need to add a
> >> new mode, the API has to be extended as well.
> >>
> >> But how?  The current API assumes that the monotonic mode was already
> >> determined before hw_params.  We may add a set of new hw_params get
> >> and set calls for tstamp mode while keeping the old API.  This would
> >> be one option.  Another option would be to add a new PCM open flag
> >> SND_PCM_TSTAMP_MONOTONIC_RAW, and snd_pcm_hw_params_is_monotonic_raw()
> >> function.  The latter is easier (a simpler addition), while the former
> >> is more extensible to newer formats in future.
> >
> > These timestamps cannot be handled by hardware drivers; they are always
> > read by the ALSA framework, in software.  Furthermore, switching between
> > MONOTONIC and MONOTONIC_RAW is possible at any time.  Therefore, the
> > timestamp mode should be made a part of sw_params; just add a field
> > named tstamp_mode ...
> 
> Humm... why would anyone switch modes at run time during 
> playback/capture? I have seen timestamps being used mainly to track that 
> playback/capture is happening as predicted, improve A/V sync or sleep 
> for a predictable time with interrupts disabled (PulseAudio, Android, 
> ChromeOS, etc). NTP adjustments are really adding noise more than 
> information for those usages. I have yet to see a case where the use of 
> those NTP adjustments would matter for userspace, and for now I really 
> don't see the point of making this dynamically configurable as a 
> software parameter. I would personally make this new mode the default.

As Clemens already pointed, if the application itself refers to the
timestamp and compares with the own timestamp by CLOCK_MONOTONIC,
using CLOCK_MONOTONIC_RAW would break.  So we cannot replace it with
CLOCK_MONOTONIC_RAW silently, unfortunately.


Takashi
Pierre-Louis Bossart July 10, 2014, 3:38 p.m. UTC | #6
On 7/10/14, 10:33 AM, Takashi Iwai wrote:
> At Thu, 10 Jul 2014 10:10:56 -0500,
> Pierre-Louis Bossart wrote:
>>
>> On 7/9/14, 2:40 PM, Clemens Ladisch wrote:
>>> Takashi Iwai wrote:
>>>> Currently, the timestamp mode is set implicitly in alsa-lib pcm_hw.c:
>>>> - When kernel PCM protocol version is high enough, alsa-lib hw prefers
>>>>     the monotonic always (if available), then set pcm->monotonic = 1.
>>>> - Application can ask whether the current timestamp is monotonic or
>>>>     not via snd_pcm_hw_params_is_monotonic().
>>>> So, only adding the flag above doesn't suffice.  If we need to add a
>>>> new mode, the API has to be extended as well.
>>>>
>>>> But how?  The current API assumes that the monotonic mode was already
>>>> determined before hw_params.  We may add a set of new hw_params get
>>>> and set calls for tstamp mode while keeping the old API.  This would
>>>> be one option.  Another option would be to add a new PCM open flag
>>>> SND_PCM_TSTAMP_MONOTONIC_RAW, and snd_pcm_hw_params_is_monotonic_raw()
>>>> function.  The latter is easier (a simpler addition), while the former
>>>> is more extensible to newer formats in future.
>>>
>>> These timestamps cannot be handled by hardware drivers; they are always
>>> read by the ALSA framework, in software.  Furthermore, switching between
>>> MONOTONIC and MONOTONIC_RAW is possible at any time.  Therefore, the
>>> timestamp mode should be made a part of sw_params; just add a field
>>> named tstamp_mode ...
>>
>> Humm... why would anyone switch modes at run time during
>> playback/capture? I have seen timestamps being used mainly to track that
>> playback/capture is happening as predicted, improve A/V sync or sleep
>> for a predictable time with interrupts disabled (PulseAudio, Android,
>> ChromeOS, etc). NTP adjustments are really adding noise more than
>> information for those usages. I have yet to see a case where the use of
>> those NTP adjustments would matter for userspace, and for now I really
>> don't see the point of making this dynamically configurable as a
>> software parameter. I would personally make this new mode the default.
>
> As Clemens already pointed, if the application itself refers to the
> timestamp and compares with the own timestamp by CLOCK_MONOTONIC,
> using CLOCK_MONOTONIC_RAW would break.  So we cannot replace it with
> CLOCK_MONOTONIC_RAW silently, unfortunately.

ok, fine, it's a different mode then. That still doesn't clarify who 
would ever switch modes dynamically while streaming data.
diff mbox

Patch

diff --git a/include/sound/asound.h b/include/sound/asound.h
index 1774a5c..9061cdd 100644
--- a/include/sound/asound.h
+++ b/include/sound/asound.h
@@ -457,7 +457,8 @@  struct snd_xfern {
 enum {
 	SNDRV_PCM_TSTAMP_TYPE_GETTIMEOFDAY = 0,	/* gettimeofday equivalent */
 	SNDRV_PCM_TSTAMP_TYPE_MONOTONIC,	/* posix_clock_monotonic equivalent */
-	SNDRV_PCM_TSTAMP_TYPE_LAST = SNDRV_PCM_TSTAMP_TYPE_MONOTONIC,
+	SNDRV_PCM_TSTAMP_TYPE_MONOTONIC_RAW,	/* monotonic_raw (no NTP) */
+	SNDRV_PCM_TSTAMP_TYPE_LAST = SNDRV_PCM_TSTAMP_TYPE_MONOTONIC_RAW,
 };
 
 /* channel positions */