[Qemu-devel] Re: KVM: High CPU load with audio enabled

[Qemu-devel] Re: KVM: High CPU load with audio enabled

[Jan Kiszka]

[-- Attachment #1: Type: text/plain, Size: 972 bytes --]

Am 09.11.2010 00:48, Michael Tokarev wrote:
> 08.11.2010 15:40, Stefan Pietsch wrote:
>> Hi,
>>
>> I'm running a Windows XP guest in KVM. After enabling the audio support
>> (-soundhw es1370) the CPU load of the KVM process in the host system
>> goes up to 100%.
>>
>> Tested with: qemu-kvm-0.13.0 Debian 0.13.0+dfsg-1
>>
>> KVM start command: kvm -m 512 -hda xp.img -soundhw es1370
>>
>> Did you experience similar problems?
> 
> Yes, sound in 0.13 appears to be broken.  It is looping in
> select/read loop forever, read returns EAGAIN and select
> says the filedescriptor is ready.

Try to set QEMU_AUDIO_DAC_TRY_POLL=0 for the environment of the qemu
process. That's long required to get sound out of the Musicpal emulation
as well (malc, you may recall this). It even got worse with current
qemu.git: previously, the sounds was just quirky. Now the system is
booting way slower in polling mode and is partially not reactive anymore.

Jan



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 259 bytes --]

[Qemu-devel] Re: KVM: High CPU load with audio enabled

[malc]

On Tue, 9 Nov 2010, Jan Kiszka wrote:

> Am 09.11.2010 00:48, Michael Tokarev wrote:
> > 08.11.2010 15:40, Stefan Pietsch wrote:
> >> Hi,
> >>
> >> I'm running a Windows XP guest in KVM. After enabling the audio support
> >> (-soundhw es1370) the CPU load of the KVM process in the host system
> >> goes up to 100%.
> >>
> >> Tested with: qemu-kvm-0.13.0 Debian 0.13.0+dfsg-1
> >>
> >> KVM start command: kvm -m 512 -hda xp.img -soundhw es1370
> >>
> >> Did you experience similar problems?
> > 
> > Yes, sound in 0.13 appears to be broken.  It is looping in
> > select/read loop forever, read returns EAGAIN and select
> > says the filedescriptor is ready.
> 
> Try to set QEMU_AUDIO_DAC_TRY_POLL=0 for the environment of the qemu
> process. That's long required to get sound out of the Musicpal emulation
> as well (malc, you may recall this). It even got worse with current
> qemu.git: previously, the sounds was just quirky. Now the system is
> booting way slower in polling mode and is partially not reactive anymore.

Wich driver is being used? ALSA stuff should have been fixed by Jindrich
Makovicka in:
38cc9b607f85017b095793cab6c129bc9844f441 and
22d948a2d97434192018bdabaf0a50cda7f994be


-- 
mailto:av1474@comtv.ru

[Qemu-devel] Re: KVM: High CPU load with audio enabled

[Michael Tokarev]

09.11.2010 16:48, malc wrote:
> On Tue, 9 Nov 2010, Jan Kiszka wrote:
> 
>> Am 09.11.2010 00:48, Michael Tokarev wrote:
[]
>>> Yes, sound in 0.13 appears to be broken.  It is looping in
>>> select/read loop forever, read returns EAGAIN and select
>>> says the filedescriptor is ready.
>>
>> Try to set QEMU_AUDIO_DAC_TRY_POLL=0 for the environment of the qemu
>> process. That's long required to get sound out of the Musicpal emulation
>> as well (malc, you may recall this). It even got worse with current
>> qemu.git: previously, the sounds was just quirky. Now the system is
>> booting way slower in polling mode and is partially not reactive anymore.
> 
> Wich driver is being used? ALSA stuff should have been fixed by Jindrich
> Makovicka in:
> 38cc9b607f85017b095793cab6c129bc9844f441 and
> 22d948a2d97434192018bdabaf0a50cda7f994be

As far as I can see, both are in 0.13.0 release, which is
the one I'm running currently.

This is with QEMU_AUDIO_DRV=alsa (default on Debian):

...
gettimeofday({1289311843, 426751}, NULL) = 0
select(24, [0 5 13 18 20 23], [], [], {1, 0}) = 1 (in [13], left {0, 999998})
poll([{fd=13, events=POLLIN|POLLERR|POLLNVAL}], 1, 0) = 1 ([{fd=13, revents=POLLIN}])
ioctl(14, 0xc0844123, 0x8e75200)        = 0
ioctl(14, 0xc0844123, 0x8e75200)        = 0
clock_gettime(CLOCK_MONOTONIC, {654227, 9502215}) = 0
clock_gettime(CLOCK_MONOTONIC, {654227, 9521568}) = 0
gettimeofday({1289311843, 426911}, NULL) = 0
select(24, [0 5 13 18 20 23], [], [], {1, 0}) = 1 (in [13], left {0, 999998})
poll([{fd=13, events=POLLIN|POLLERR|POLLNVAL}], 1, 0) = 1 ([{fd=13, revents=POLLIN}])
ioctl(14, 0xc0844123, 0x8e75200)        = 0
ioctl(14, 0xc0844123, 0x8e75200)        = 0
clock_gettime(CLOCK_MONOTONIC, {654227, 9660799}) = 0
clock_gettime(CLOCK_MONOTONIC, {654227, 9680088}) = 0
gettimeofday({1289311843, 427070}, NULL) = 0
...

Fd#13 is /dev/snd/timer.

For this test, I tried -soundhw ac97.  The same happens with
es1370 too.

With oss or pulseaudio it works just fine.

Note that the problem does not occur always.  I just
booted a win7 guest with ac97, installed a driver, and it
whole thing appears to work without this 100% cpu loop, at
least after settling down.  But several attempts to run a
linux guest with es1370 or ac97 results in a reproduceable
tight loop - right after bootup.

And I just discovered another issue with QUEMU_AUDIO_DRV=pa:
when I tried to close sdl window, kvm wents into tight loop

waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
waitpid(30478, 0xffa84ccc, 0)           = -1 ECHILD (No child processes)
...

obviously consuming 100% cpu, and the only way to stop this is to
kill it.  Unrelated, but also buggy ;)

Thanks!

/mjt

[Qemu-devel] Re: KVM: High CPU load with audio enabled

[Michael Tokarev]

09.11.2010 17:33, Michael Tokarev wrote:
> 09.11.2010 16:48, malc wrote:
>> On Tue, 9 Nov 2010, Jan Kiszka wrote:
>>
>>> Am 09.11.2010 00:48, Michael Tokarev wrote:
> []
>>>> Yes, sound in 0.13 appears to be broken.  It is looping in
>>>> select/read loop forever, read returns EAGAIN and select
>>>> says the filedescriptor is ready.
>>>
>>> Try to set QEMU_AUDIO_DAC_TRY_POLL=0 for the environment of the qemu
>>> process. That's long required to get sound out of the Musicpal emulation
>>> as well (malc, you may recall this). It even got worse with current
>>> qemu.git: previously, the sounds was just quirky. Now the system is
>>> booting way slower in polling mode and is partially not reactive anymore.
>>
>> Wich driver is being used? ALSA stuff should have been fixed by Jindrich
>> Makovicka in:
>> 38cc9b607f85017b095793cab6c129bc9844f441 and
>> 22d948a2d97434192018bdabaf0a50cda7f994be
> 
> As far as I can see, both are in 0.13.0 release, which is
> the one I'm running currently.

Ok. This is wrong: neither of the two changes went into any
stable release.  Hence we've bugs which are long fixed but
no one knows about this.  As usual for qemu[-kvm] development.

/mjt

[Qemu-devel] Re: KVM: High CPU load with audio enabled

[Stefan Pietsch]

* Michael Tokarev <mjt@tls.msk.ru> [2010-11-09 15:41]:

> Ok. This is wrong: neither of the two changes went into any
> stable release.  Hence we've bugs which are long fixed but
> no one knows about this.  As usual for qemu[-kvm] development.

I can reproduce the problem with the Debian release 0.12.5+dfsg-4.
So this is not related to the 0.13.0 release.


Regards,
Stefan

[Qemu-devel] Re: KVM: High CPU load with audio enabled

[Michael Tokarev]

10.11.2010 12:39, Stefan Pietsch wrote:
> * Michael Tokarev <mjt@tls.msk.ru> [2010-11-09 15:41]:
> 
>> Ok. This is wrong: neither of the two changes went into any
>> stable release.  Hence we've bugs which are long fixed but
>> no one knows about this.  As usual for qemu[-kvm] development.
> 
> I can reproduce the problem with the Debian release 0.12.5+dfsg-4.
> So this is not related to the 0.13.0 release.

It is related to supporting stable releases, or lack thereof.

As you just discovered, the same bug exists in 0.12 stable
version too, also not fixed there.  Earlier today when I
prepared new 0.13 release for debian I also checked if the
same fix should be applied to 0.12, and indeed it should -
and I updated 0.12 debian branch too.

/mjt