RFC: PV vfb's and the handling of global vnc options

RFC: PV vfb's and the handling of global vnc options

[George Dunlap]

So I had a user report that he couldn't get the vnclisten option to
work.  It turns out he was using a PV guest, and had the following in
his config file:

vfb=[ 'type=vnc' ]
vnclisten='0.0.0.0'

After digging around in the code, it turns out that the following is
accepted for PV guests:

vnc=1
vnclisten='0.0.0.0'

...on the condition that there be *no* vfb option.  If you set the vfb
option, then the top-level vnc entries are completely ignored.

The following also works (as one might expect):

vfb=[ 'type=vnc,vnclisten=0.0.0.0' ]

The man page is a bit unclear here; it says:

If "Emulated VGA Graphics Device" options are used in a PV guest
configuration, xl will pick up vnc, vnclisten, vncpasswd, vncdisplay,
vncunused, sdl, opengl and keymap to construct paravirtual framebuffer
device for the guest.

I see two options:

1. Make vfb take defaults from the top-level vnc options

2. Clarify the man page to make it clear that top-level vnc options
will only have an effect if no vfb is specified.

I'm of two minds which one is best.  At the moment I'm leaning towards
#2, possibly adding a warning if it finds both vfb and toplevel vnc
options, saying that it's ignoring the second.

Any opinions?

 -George

Re: RFC: PV vfb's and the handling of global vnc options

[Konrad Rzeszutek Wilk]

On Mon, Nov 09, 2015 at 03:03:17PM +0000, George Dunlap wrote:
> So I had a user report that he couldn't get the vnclisten option to
> work.  It turns out he was using a PV guest, and had the following in
> his config file:
> 
> vfb=[ 'type=vnc' ]
> vnclisten='0.0.0.0'
> 
> After digging around in the code, it turns out that the following is
> accepted for PV guests:
> 
> vnc=1
> vnclisten='0.0.0.0'
> 
> ...on the condition that there be *no* vfb option.  If you set the vfb
> option, then the top-level vnc entries are completely ignored.

Ha!
> 
> The following also works (as one might expect):
> 
> vfb=[ 'type=vnc,vnclisten=0.0.0.0' ]
> 
> The man page is a bit unclear here; it says:
> 
> If "Emulated VGA Graphics Device" options are used in a PV guest
> configuration, xl will pick up vnc, vnclisten, vncpasswd, vncdisplay,
> vncunused, sdl, opengl and keymap to construct paravirtual framebuffer
> device for the guest.
> 
> I see two options:
> 
> 1. Make vfb take defaults from the top-level vnc options

Keep in mind that for HVM guests we would want only the 'vnc' options
from the 'vfb'. That is because by default we boot with 'vga=cirrus'
and with both backends (VGA VNC and PV FB) the guest gets mighty
confused during bootup. That is - the initial 'text' console goes
to the PV FB and the KDM/GDM goes in the VGA VNC. But the 'PV FB'
is the first one and most tools except only one VNC backend.

> 
> 2. Clarify the man page to make it clear that top-level vnc options
> will only have an effect if no vfb is specified.

The manpage was probably constructed during the reign of Xend. In
which you could have both options (vfb= and vnc=) .. and Xend would
happily expose in Xenstore the pvfb keys with the old QEMU
(which would default to the VGA for HVM guests). Which meant that the
guest would spend 30 seconds trying to use the PVFB backend driver :-(

> 
> I'm of two minds which one is best.  At the moment I'm leaning towards
> #2, possibly adding a warning if it finds both vfb and toplevel vnc
> options, saying that it's ignoring the second.
> 
> Any opinions?

I like #1 as it would make it easy to use the old style Xend
configuration files doing the right thing.

With PVH and with 'qemu=none'..you would want the PVFB, if the
user used 'vnc' or 'vfb'.

> 
>  -George
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel

Re: RFC: PV vfb's and the handling of global vnc options

[Ian Campbell]

On Mon, 2015-11-09 at 15:03 +0000, George Dunlap wrote:

Adding other tools maintainers.

> So I had a user report that he couldn't get the vnclisten option to
> work.  It turns out he was using a PV guest, and had the following in
> his config file:
> 
> vfb=[ 'type=vnc' ]
> vnclisten='0.0.0.0'
> 
> After digging around in the code, it turns out that the following is
> accepted for PV guests:
> 
> vnc=1
> vnclisten='0.0.0.0'
> 
> ...on the condition that there be *no* vfb option.  If you set the vfb
> option, then the top-level vnc entries are completely ignored.
> 
> The following also works (as one might expect):
> 
> vfb=[ 'type=vnc,vnclisten=0.0.0.0' ]
> 
> The man page is a bit unclear here; it says:
> 
> If "Emulated VGA Graphics Device" options are used in a PV guest
> configuration, xl will pick up vnc, vnclisten, vncpasswd, vncdisplay,
> vncunused, sdl, opengl and keymap to construct paravirtual framebuffer
> device for the guest.
> 
> I see two options:
> 
> 1. Make vfb take defaults from the top-level vnc options
> 
> 2. Clarify the man page to make it clear that top-level vnc options
> will only have an effect if no vfb is specified.
> 
> I'm of two minds which one is best.  At the moment I'm leaning towards
> #2, possibly adding a warning if it finds both vfb and toplevel vnc
> options, saying that it's ignoring the second.

One problem with #1 is that for some options (vncdisplay) having a global
default for potentially multiple vfb=[] stanzas would be wrong and a bit
confusing (although perhaps no more so than the current behaviour).

I suppose you could limit some or (better) all of the globals to only being
used for the first entry in the array (but I bet the code gets quite funky
in that case).

You'd probably want the warning you mention in #2 even in the #1 case.

Like you I'm not really sure.

A third option just to muddy the waters:

3. If any global options are given then a new vfb is pushed onto the front
of any array given by vfb=[], i.e. those vfbs implicitly go from devices
0..N to 1..N+1.

Ian.



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel