Skip to contentSkip to conversation

alsa-lib: Provide a CLOCK_MONOTONIC_RAW timestamp type

PATCH In review Mark Brown 4 participants 7 messages Jul 8, 2014 – Jul 10, 2014 Sound

lore
Share & tools
lore

Mark Brown

author
#0 lore raw

PATCH 1/1 alsa-lib: Provide a CLOCK_MONOTONIC_RAW timestamp type

From: Mark Brown <[email protected]>

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.

---
 include/sound/asound.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
include/sound/asound.h +2 -1
index 1774a5c..9061cdd 100644
@@ -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 */
Signature · 1 line
2.0.0
#1 lore raw

At Tue, 8 Jul 2014 16:52:32 +0200,

Quoted 11 lines · Mark Brown

From: Mark Brown <[email protected]>

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.

Quoted 2 lines

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

Quoted 21 lines
---
 include/sound/asound.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
include/sound/asound.h +2 -1
index 1774a5c..9061cdd 100644
@@ -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 */
Signature · 1 line
2.0.0
#2 lore raw
Quoted 8 lines · Takashi Iwai

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.

Quoted 7 lines

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

#3 lore raw

At Wed, 09 Jul 2014 21:40:02 +0200,

Quoted 18 lines · Clemens Ladisch
Quoted 8 lines · Takashi Iwai

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.

Quoted 7 lines

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.

Quoted 5 lines

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

#4 lore raw
In reply to Clemens Ladisch
Quoted 17 lines · Clemens Ladisch
Quoted 8 lines · Takashi Iwai

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.

Quoted 7 lines

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.

Quoted 5 lines

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

#5 lore raw

At Thu, 10 Jul 2014 10:10:56 -0500,

Quoted 25 lines · Pierre-Louis Bossart
Quoted 17 lines · Clemens Ladisch
Quoted 8 lines · Takashi Iwai

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.

Quoted 7 lines

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.

Quoted 5 lines

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 ...

Quoted 9 lines

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

#6 lore raw
Quoted 37 lines · Takashi Iwai

At Thu, 10 Jul 2014 10:10:56 -0500,

Quoted 25 lines · Pierre-Louis Bossart
Quoted 17 lines · Clemens Ladisch
Quoted 8 lines · Takashi Iwai

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.

Quoted 7 lines

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.

Quoted 5 lines

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 ...

Quoted 9 lines

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.

Quoted 4 lines

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.