2009-12-09 19:24:30

by Wojciech Zabołotny

[permalink] [raw]
Subject: Driver snd-hrtimer - causing system freeze with Rosegarden

When I try to use the "HR timer" as a MIDI sequencer timer in the
Rosegarden (http://www.rosegardenmusic.com) application, the system
freezes. Another people also reported similar problems with other
sequencer - OpenOctave (
https://linuxmusicians.com/viewtopic.php?f=27&p=8917 )

I have tested the problem on two SMP machines with diffrent kernels
(2.6.30, 2.6.31.6, 2.6.32).
The problem seems to be associated with the "snd-hrtimer" module.
I've tried to run the application vie remote X session, while having the
text console active.
Before the machine froze, the following messages have been displayed in
the console:

System 1 (Intel(R) Pentium(R) 4 CPU 2.80GHz HT)

hrtimer: interrupt too slow, forcing clock min delta to 3568 ns

System 2 (Intel(R) Core(TM)2 CPU T5500 @ 1.66GHz)
hrtimer: interrupt too slow, forcing clock min delta to 11103 ns
CE: hpet increasing min_delta_ns to 15000 nsec
CE: hpet increasing min_delta_ns to 22500 nsec
CE: hpet increasing min_delta_ns to 33750 nsec

Analyzing the problem, I had an idea, that maybe the snd-hrtimer should
be rewritten to call
the callback via tasklet (I attach my patch only as an ilustration, my
code doesn't work either),
however then I got the message "BUG: scheduling while atomic" or
something like this
from the snd_hrtimer_stop function called from _snd_timer_stop.
Therefore I've tried to investigate the snd_hrtimer_stop more thoroughly.

Looking further into the code, I have found, that snd_hrtimer_stop may
be called
from the callback routine (e.g.:
http://lxr.linux.no/linux+v2.6.32/sound/core/timer.c#L709 or
http://lxr.linux.no/linux+v2.6.32/sound/core/timer.c#L719 )

It seems to me, that calling hrtimer_cancel (in the original code),
which waits for the handler to finish,
from that handler may lock the machine...
(snd_hrtimer_callback calls the snd_timer_interrupt:
http://lxr.linux.no/linux+v2.6.32/sound/core/hrtimer.c#L47
which in turn calls snd_hrtimer_stop (as above) which calls hrtimer_cancel).


Unfortunately currently I have no idea how to fix this problem. Maybe
snd_hrtimer_stop shouldn't try to cancel the timer,
instead setting a flag causing callback to return immediately without
rearming of the timer?

--
Best regards,
Wojtek Zabolotny
[email protected]


Attachments:
hrtimer.patch (1.92 kB)

2009-12-10 08:16:48

by Takashi Iwai

[permalink] [raw]
Subject: Re: Driver snd-hrtimer - causing system freeze with Rosegarden

At Wed, 09 Dec 2009 23:04:30 +0100,
[email protected] wrote:
>
> Dear Takashi,
>
> I've sent the attached message to the LKML, but I think that it may be
> also reasonable to contact you - the author of the snd-hrtimer.
> I've tried to find reasons for instability of system running the
> sequencer using HR timer.
> My results are summarized in the attached bug report.
> Could you please check if calling of hrtimer_cancel from
> xnd_hrtimer_stop, which sometimes is called from the callback may cause
> the obderved instability?

hrtimer_cancel() is very simple and doesn't sleep (but spins). So,
this shouldn't be the real issue.

> I have tested the problem on two SMP machines with diffrent kernels
> (2.6.30, 2.6.31.6, 2.6.32).
> The problem seems to be associated with the "snd-hrtimer" module.
> I've tried to run the application vie remote X session, while having the
> text console active.
> Before the machine froze, the following messages have been displayed in
> the console:
>
> System 1 (Intel(R) Pentium(R) 4 CPU 2.80GHz HT)
>
> hrtimer: interrupt too slow, forcing clock min delta to 3568 ns
>
> System 2 (Intel(R) Core(TM)2 CPU T5500 @ 1.66GHz)
> hrtimer: interrupt too slow, forcing clock min delta to 11103 ns
> CE: hpet increasing min_delta_ns to 15000 nsec
> CE: hpet increasing min_delta_ns to 22500 nsec
> CE: hpet increasing min_delta_ns to 33750 nsec

Hm, this sounds weird.
Does hrtimer work reliably on your machine?

> Analyzing the problem, I had an idea, that maybe the snd-hrtimer should
> be rewritten to call
> the callback via tasklet (I attach my patch only as an ilustration, my
> code doesn't work either),

Basically snd-hrtimer already uses tasklet (more exactly, snd-timer
handles callbacks in tasklet as default except for the objects with
SNDRV_TIMER_HW_TASKLET). So your patch is wrapping over a wrap.


Takashi