On 05/24/2010 11:18 AM, Daniel P. Berrange wrote:
On Mon, May 24, 2010 at 10:56:39AM -0400, Cole Robinson wrote:
> On 05/24/2010 10:31 AM, Daniel P. Berrange wrote:
>> On Mon, May 24, 2010 at 10:26:43AM -0400, Cole Robinson wrote:
>>> On 05/24/2010 10:14 AM, Daniel P. Berrange wrote:
>>>> On Thu, May 20, 2010 at 12:04:04PM -0400, Cole Robinson wrote:
>>>>> Currently all host audio backends are disabled if a VM is using VNC,
in
>>>>> favor of the QEMU VNC audio extension. Unfortunately no released VNC
>>>>> client supports this extension, so users have no way of getting
audio
>>>>> to work if using VNC.
>>>>>
>>>>> Add a new config option in qemu.conf which allows changing
libvirt's
>>>>> behavior, but keep the default intact.
>>>>>
>>>>> Signed-off-by: Cole Robinson <crobinso(a)redhat.com>
>>>>> ---
>>>>> src/qemu/qemu.conf | 10 ++++++++++
>>>>> src/qemu/qemu_conf.c | 17 ++++++++++++-----
>>>>> src/qemu/qemu_conf.h | 2 ++
>>>>> 3 files changed, 24 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/src/qemu/qemu.conf b/src/qemu/qemu.conf
>>>>> index 3da332f..fec946d 100644
>>>>> --- a/src/qemu/qemu.conf
>>>>> +++ b/src/qemu/qemu.conf
>>>>> @@ -168,3 +168,13 @@
>>>>> # be assigned to guests.
>>>>> #
>>>>> # relaxed_acs_check = 1
>>>>> +
>>>>> +
>>>>> +# QEMU implements anextension for providing audio over a VNC
connection,
>>>>> +# though if your VNC client does not support it, your only chance
for getting
>>>>> +# sound output is through regular audio backends. By default,
libvirt will
>>>>> +# disable all QEMU sound backends if using VNC, since they can
cause
>>>>> +# permissions issues. Enabling this option will make libvirtd honor
the
>>>>> +# QEMU_AUDIO_DRV environment variable when using VNC.
>>>>> +#
>>>>> +# vnc_enable_audio_backend = 0
>>>>
>>>> I think this would be better named as 'vnc_allow_host_audio'
>>>> since its not toggling VNC audio, just whether it is allowed
>>>> to use the host audio drivers.
>>>>
>>>
>>> Okay, I'll change and resend.
>>>
>>>> NB, even with this toggled things are unlikely to magically
>>>> 'just work'. If QEMU is running as 'qemu' user ID, that
user
>>>> won't have any permissions to use /dev/snd/* devices. And
>>>> if running as 'root' then pulseaudio won't autospawn so will
>>>> need to be run manually.
>>>>
>>>
>>> Right. Currently getting sound to work in Fedora is a big pain, and
>>> there have been quite a few user complaints. I'm going to write up a
>>> 'known bug' page, documenting a workaround of qemu.conf user/group =
>>> $USER to fix the issue, but this patch will still need to be backported.
>>
>> I'd really recommend against telling people to configure libvirt to run
>> the guests as their own $USER because that'll open a huge can of worms.
>>
>
> What exactly are the downsides, besides the security implications? It
> should exercise all the same code as user/group = 'qemu', but solves
> issues like:
>
> - Making sure emulator user can access $HOME for install media
> - Audio
> - PTY access
> - SDL (if someone insists on it)
In the contex of desktop usage, these problems are all the result
of using the qemu:///system driver instead of the session driver.
We've not historically used the session driver, because it can't
setup a TAP device for guests out of the box. If we're going down
the route of manually post-install config steps as root, then instead
of changing libvirtd config, we'd be better of doing the manual
config step to allow TAP device access to $USER. This approach is
aligned with our driver design for qemu://session being the per
user bus, as opposed to hacking qemu://system todo something it
was never intended to support.
qemu:///session is the way forward, but its a lot more work then just
telling people to manually setup a tap device: better tool support in
virt-manager/virt-install, MUCH better/more visible docs covering the
differences/pros/cons, not to mention preparation for the onslaught of
'I'm running as root and can't see my VMs' bugs. Maybe docs about moving
your qemu:///system machines to qemu:///session. Until this work is
done, we should document workarounds, even if it uses qemu:///system in
a way it was never intended to support (we already do that since it is
used for 99% of all desktop virt libvirt usage).
Independantly of this, those 4 problems above can also be addressed
for the system driver while maintaining the proper privilege
separation. Audio access by fixing GTK-VNC. PTY access by using the
libvirt streams API. SDL access by pointing it to the users'
xauth file (and setting an ACL on xauth to allow qemu to read it).
Install media by adding libvirt APIs to upload a kernel+initrd or
boot.iso the appropriate location. This is again aligned with our
design of the system instance being highly privileged, but separated
from any other accounts on the host.
That SDL trick doesn't work for me on F13. Even running guests as root
doesn't work for SDL anymore, used to work on F11 at least.
That aside, I completely agree that we can and should properly solve
these issues with qemu:///system. But F12 and F13 users will likely
never see this work, so it's important to document all this: bug reports
tracking the proper way forward (we have), with possible workarounds in
the interim (we don't have). Setting user/group to $USER is a fairly
simple change that makes many of these issues disappear, and for a
single user system it doesn't seem too outrageous a concept.
>> Better to tell them to add an ACL to /dev/snd/* using setfacl
to add the
>> qemu user to the ACL for the sound card, or drop in a udev rule todo the
>> same.
>>
>
> I'll try that out, but I wonder will my local user actually hear that
> sound? Will the audio go to my pulseaudio session?
This will cause another pulseaudio daemon to open the sound device and play
directly, it shouldn't need to connect to the user's own pulse daemon. If
PA isn't in use, then QEMU will just try to access the sound card directly.
Hmm, couldn't get this to work with F13, qemu startup log is filled with
'Connection Refused' pa errors. Could be user error or require more
configuration, but it's already entering the realm of more difficulty
than user/group=$USER.
- Cole