alsa timer slippage
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 11 lines · Chris Cannam
While trying to track down the source of some poor timing in sequencing, I've noticed that my ALSA sequencer queue timer has a tendency to fall suddenly behind.
I have a little test program (available on request) that just starts a queue and every second or so compares the queue timer against real time as returned by
gettimeofday(). It doesn't mind if the two don't quite match, but it does complain if the difference between the two timers changes dramatically between two sample points. When I run it, it never lasts for more than about a minute before the ALSA queue timer suddenly slips by anything from 10 to 60 milliseconds.
Please, send me your test program.
Quoted 15 lines
This is a non-low-latency kernel so I'm not surprised that there may be some occasional timing issues, but 60ms is a lot on an unloaded machine, and I am vaguely surprised that the timer doesn't notice it's fallen behind and recover -- instead all events on the queue continue to be delivered late forever. This obviously makes for some disconcerting audible effects.
The system is SuSE 9.0 on a dual 2GHz Athlon using SuSE's stock SMP kernel. I have tried both ALSA 0.9.6 (from SuSE) and 1.0.0rc2 drivers and libraries. I haven't managed to reproduce it using a PlanetCCRMA SMP kernel on the same machine, nor on my uniprocessor laptop. I've surveyed a few other people on rosegarden-devel and nobody's corroborated my findings, so I guess it might be related to using a dual-processor machine.
Quoted 5 lines
Any thoughts on this, anyone? I'm finding it a little depressing that I can't play even a minute of 4/4 beats from an ALSA test program without the timing slipping audibly at least once. I'm ready to delve cluelessly into the timer code to take a look, but (glancing at
alsa-kernel/core/timer.c) I'm not at all sure how far I'd get...
The ALSA MIDI sequencer can use several timer sources. By default, it uses the Linux system timer functions, see:
http://hegel.ittc.ukans.edu/topics/linux/man-pages/man9/init_timer.9.html
Other timer sources are PCM devices and the RTC timer. The module snd-rtctimer provides this functionality. To compile it for a 2.4.x kernel, you need to patch your rtc driver (and rebuild), see alsa-driver/utils/patches
You can set a system-wide alternate timer in your /etc/modules.conf, for instance:
options snd-seq seq_default_timer_device=1
alias snd-timer-1 snd-rtctimer
Or your program can select an alternate timer source for a single queue. See:
http://www.mail-archive.com/[email protected]/msg09048.html
Regards,
Pedro
Signature · 10 lines
ALSA Library Bindings for Pascal http://alsapas.alturl.com ------------------------------------------------------- This SF.net email is sponsored by: SF.net Giveback Program. Does SourceForge.net help you be more productive? Does it help you create better code? SHARE THE LOVE, and help us help YOU! Click Here: http://sourceforge.net/donate/
Quoted 3 lines · Pedro Lopez-Cabanillas
The ALSA MIDI sequencer can use several timer sources. By default, it uses the Linux system timer functions [...]
Other timer sources are PCM devices and the RTC timer.
Ah... so that puts Jan's comment into perspective. I could have been using the soundcard to sync from, but I wasn't. And indeed, switching the timer for this queue to the first PCM device gives me vastly better results.
This seems to suggest that the default system-based timer is doing something a bit naive, like assuming it was called back on time without actually checking.
Chris
-------------------------------------------------------
This SF.net email is sponsored by: SF.net Giveback Program. Does SourceForge.net help you be more productive? Does it help you create better code? SHARE THE LOVE, and help us help YOU! Click Here: http://sourceforge.net/donate/
Quoted 3 lines · Chris Cannam
This seems to suggest that the default system-based timer is doing something a bit naive, like assuming it was called back on time without actually checking.
Hmm yes, even this crappy little patch to alsa-kernel/core/timer.c (gross adjustment to each system timer tick according to how late the previous one was) improves things for me. With this the timer still slips, but never quite enough to be audible in a MIDI-only stream. The PCM clock is still much much better though.
@@ -940,9 +940,16 @@
return 1000000000L / HZ;
}
+static long lateness;
+
static void snd_timer_s_function(unsigned long data) {
+ struct timer_list *tlist;
snd_timer_t *timer = (snd_timer_t *)data;+
+ tlist = (struct timer_list *) timer->private_data;
+ lateness = (long)jiffies - (long)tlist->expires;
+
snd_timer_interrupt(timer, timer->sticks);
}
@@ -951,7 +958,7 @@
struct timer_list *tlist;
tlist = (struct timer_list *) timer->private_data;
- tlist->expires = jiffies + timer->sticks;
+ tlist->expires = jiffies + timer->sticks - lateness;
add_timer(tlist);
return 0;}
Chris
-------------------------------------------------------
This SF.net email is sponsored by: SF.net Giveback Program. Does SourceForge.net help you be more productive? Does it help you create better code? SHARE THE LOVE, and help us help YOU! Click Here: http://sourceforge.net/donate/
Quoted 29 lines · Chris Cannam
Quoted 3 lines · Chris Cannam
This seems to suggest that the default system-based timer is doing something a bit naive, like assuming it was called back on time without actually checking.
Hmm yes, even this crappy little patch to
alsa-kernel/core/timer.c(gross adjustment to each system timer tick according to how late the previous one was) improves things for me. With this the timer still slips, but never quite enough to be audible in a MIDI-only stream. The PCM clock is still much much better though.(quoted diff) +7
-0@@ -940,9 +940,16 @@ return 1000000000L / HZ; } +static long lateness; + static void snd_timer_s_function(unsigned long data) { + struct timer_list *tlist; snd_timer_t *timer = (snd_timer_t *)data; + + tlist = (struct timer_list *) timer->private_data; + lateness = (long)jiffies - (long)tlist->expires; + snd_timer_interrupt(timer, timer->sticks); }
Review line 964
Quoted 8 lines
(quoted diff) +1
-1@@ -951,7 +958,7 @@ struct timer_list *tlist; tlist = (struct timer_list *) timer->private_data; - tlist->expires = jiffies + timer->sticks; + tlist->expires = jiffies + timer->sticks - lateness; add_timer(tlist); return 0; }
Thanks for insigh to the problem. Can you try the attached patch? (It's also in CVS.)
Jaroslav
-----
Jaroslav Kysela <[email protected]>
Linux Kernel Sound Maintainer
ALSA Project, SuSE Labs
Index: timer.c
===================================================================
RCS file: /cvsroot/alsa/alsa-kernel/core/timer.c,v retrieving revision 1.52
retrieving revision 1.55
timer.c +42 -16
@@ -935,6 +935,14 @@* System timer */ +struct snd_timer_system_private { +struct timer_list tlist; +struct timer * timer; +unsigned long last_expires; +unsigned long last_jiffies; +unsigned long correction; +}; +unsigned int snd_timer_system_resolution (void ) {return 1000000000L /HZ ;
@@ -943,26 +951,44 @@static void snd_timer_s_function (unsigned long data) {snd_timer_t *timer = (snd_timer_t *)data; -snd_timer_interrupt (timer, timer->sticks); +struct snd_timer_system_private *priv = timer->private_data; +unsigned long jiff = jiffies; +if (time_after (jiff, priv->last_expires)) + priv->correction = (long )jiff - (long )priv->last_expires; +snd_timer_interrupt (timer, (long )jiff - (long )priv->last_jiffies); }static int snd_timer_s_start (snd_timer_t * timer) { -struct timer_list *tlist; +struct snd_timer_system_private *priv; +unsigned long njiff; - tlist = (struct timer_list *) timer->private_data; - tlist->expires = jiffies + timer->sticks; -add_timer (tlist); + priv = (struct snd_timer_system_private *) timer->private_data; + njiff = (priv->last_jiffies = jiffies); +if (priv->correction > timer->sticks -1 ) { + priv->correction -= timer->sticks -1 ; + njiff++; + }else { + njiff += timer->sticks - priv->correction; + priv->correction -= timer->sticks; + } + priv->last_expires = priv->tlist.expires = njiff; +add_timer (&priv->tlist);return 0 ; }static int snd_timer_s_stop (snd_timer_t * timer) { -struct timer_list *tlist; +struct snd_timer_system_private *priv; +unsigned long jiff; - tlist = (struct timer_list *) timer->private_data; -del_timer (tlist); - timer->sticks = tlist->expires - jiffies; + priv = (struct snd_timer_system_private *) timer->private_data; +del_timer (&priv->tlist); + jiff = jiffies; +if (time_before (jiff, priv->last_expires)) + timer->sticks = priv->last_expires - jiff; +else + timer->sticks =1 ;return 0 ; }
@@ -984,22 +1010,22 @@static int snd_timer_register_system (void ) {snd_timer_t *timer; -struct timer_list *tlist; +struct snd_timer_system_private *priv;int err;if ((err =snd_timer_global_new ("system" ,SNDRV_TIMER_GLOBAL_SYSTEM , &timer)) <0 )return err;strcpy (timer->name,"system timer" ); timer->hw = snd_timer_system; - tlist = (struct timer_list *)snd_kcalloc (sizeof (struct timer_list ),GFP_KERNEL ); -if (tlist ==NULL ) { + priv = (struct snd_timer_system_private *)snd_kcalloc (sizeof (struct snd_timer_system_private ),GFP_KERNEL ); +if (priv ==NULL ) {snd_timer_free (timer);return -ENOMEM ; } -init_timer (tlist); - tlist->function = snd_timer_s_function; - tlist->data = (unsigned long ) timer; - timer->private_data = tlist; +init_timer (&priv->tlist); + priv->tlist.function = snd_timer_s_function; + priv->tlist.data = (unsigned long ) timer; + timer->private_data = priv; timer->private_free = snd_timer_free_system;return snd_timer_global_register (timer); }
Quoted 1 line · Jaroslav Kysela
Thanks for insigh to the problem. Can you try the attached patch?
Yes, that's a success. Thank you. It's still not absolutely perfect, but it's significantly better than the original version. I've seen one slippage (of about 40ms, under load) in about twelve minutes of testing -- enough to know there's still a problem in there somewhere, but perhaps acceptable for a non-low-latency kernel and the like.
Chris
-------------------------------------------------------
This SF.net email is sponsored by: IBM Linux Tutorials. Become an expert in LINUX or just sharpen your skills. Sign up for IBM's Free Linux Tutorials. Learn everything from the bash shell to sys admin. Click now! http://ads.osdn.com/?ad_id=1278&alloc_id=3371&op=click