[Spice-devel] Spice agent for XSpice

[Alon Levy]

On Wed, 2013-07-31 at 17:53 +0200, Hans de Goede wrote:
> Hi,
> 
> On 07/31/2013 05:43 PM, Alon Levy wrote:
> >> Hi,
> >>
> >> On 03/09/2012 11:57 AM, Alon Levy wrote:
> >>> On Fri, Mar 09, 2012 at 10:46:50AM +0100, Hans de Goede wrote:
> >>>> Hi Eike, Alon,
> >>>>
> >>>> On 03/08/2012 08:33 PM, Eike Hein wrote:
> >>>>>
> >>>>> Hi,
> >>>>>
> >>>>> I recently sent a mail to Marc-André Lureau, inquiring about
> >>>>> clipboard sharing support in his virt-viewer builds for Win-
> >>>>> dows. It turned out that the reason clipboard sharing doesn't
> >>>>> work is because Xspice does not yet spawn an agent.
> >>>>>
> >>>>> Marc-André subsequently got Alon Levy into the discussion,
> >>>>> who had this to say on how this might be pulled off:
> >>>>>
> >>>>> "I'm really glad to hear someone is actually using this. To implement
> >>>>> clipboard sharing is indeed just an Xspice issue. You'll need to have an
> >>>>> agent talking to spice server not via a virtio device and qemu. Looking
> >>>>> at vdagent-linux I guess there are a few questions:
> >>>>> * do we run vdagentd and vdagent as subprocesses of Xspice
> >>>>> (actually Xorg)
> >>>>> * is there a way to emulate uinput (not related to clipboard)
> >>>>> * more a statement - I think the clipboard part is relatively easy,
> >>>>> you can replace the hardcoded /dev path for the virtio-serial port
> >>>>> with a pipe.
> >>>>>
> >>>>> I guess I would try to split vdagent to a library and app, and then link
> >>>>> the library into spiceqxl_drv.so (i.e. xf86-video-qxl)."
> >>>>>
> >>>>> I'm still pretty keen on getting this to work in my Xspice-
> >>>>> based setup, and since I was encouraged to bring this topic
> >>>>> to the list here goes. Please chime in :).
> >>>>
> >>>> Eike, in case you don't know, I'm the main author of the Linux
> >>>> spice-vdagent. I've been thinking a bit about re-using that
> >>>> for Xspice (after Alon asked me this a couple of days ago) and
> >>>> here is my 2 cents:
> >>>>
> >>>> The linux agent consists of 2 parts, a system level daemon and
> >>>> a per user (X) session agent process.
> >>>>
> >>>> The system level daemon (spice-vdagentd):
> >>>> -talks to the agent virtio serial port
> >>>> -handles client mouse mode through uinput
> >>>> -dispatches other messages to session process for the
> >>>>    currently active session (it uses consolekit to find out
> >>>>    which is the currently active session, think fast
> >>>>    user switching and having multiple X-session open
> >>>>    on different VTs)
> >>>>
> >>>> The user session process (spice-vdagent):
> >>>> -sends the resolution X is currently running at to
> >>>>    spice-vdagentd, which needs this info to interpret
> >>>>    the mouse events it gets from the client and feeds to
> >>>>    the uinput device
> >>>> -receives the desired (client native) resolution from the
> >>>>    client, and uses Xrandr to switch to this
> >>>> -handles copy and paste
> >>>>
> >>>> If we want to re-use this all for X-spice, then the choosen
> >>>> split actually comes in quite handy, with X-spice we don't
> >>>> have a agent virtio serial port, and we don't need to send
> >>>> mouse event through a uinput device either. So I suggest that
> >>>> for X-spice we simply re-use the user session agent process
> >>>> as is (making it connect to the running X-spice server), and
> >>>> "throw away" the system level daemon, replacing it with some
> >>>> functionality inside X-spice.
> >>>>
> >>>> The user session agent process talks to the (to be removed
> >>>> in the X-spice scenario) system level agent daemon through
> >>>> the following unix domain socket:
> >>>> /var/run/spice-vdagentd/spice-vdagent-sock
> >>>>
> >>>> So I think that if we modify X-spice to register a chardev
> >>>> with spice-server for the agent channel, and create a
> >>>> /var/run/spice-vdagentd/spice-vdagent-sock unix socket and
> >>>> forward all client agent messages between the 2 we are almost
> >>>> done. All that then needs to be done is filter out mouse
> >>>> event messages and inject those directly into the X-server.
> >>>>
> >>>> Note that the protocol on the unix socket != the protocol
> >>>> on the virtio serial port, so you will need to lift some code
> >>>> from spice-vdagentd which does the parsing of the one and
> >>>> wrapping of the other.
> >>>>
> >>>> Later on we should make the path of the unix domain socket
> >>>> a cmdline option for both X-spice and spice-vdagent so
> >>>> that we can run multiple X-spice sessions on the same machine,
> >>>> with each using there own socket.
> >>>>
> >>>> Something else to worry about later is sharing the necessary
> >>>> code from spice-vdagent with X-spice, for now I simply suggest
> >>>> copying over the necessary files.
> >>>>
> >>>
> >>> Sounds like a plan. My only comment was the hardcoded /var/run.. and
> >>> multiple servers, but you forsaw that :)
> >>>
> >>> I could start from the end part - spliting the system level daemon to a
> >>> library and daemon, and reusing the library for Xspice (is X-spice
> >>> really more readable?)
> >>
> >> I don't think there is much in there to split into a library. The only
> >> re-usable part is udscs.c / udscs.h which stands for unix domain socket
> >> client server. Which does the low-level uds handling. Everything else
> >> is not really suitable for re-use (ie assumes a virtio serial port,
> >> rather then spice-servers chardev interface).
> >>
> >> I think the only thing which you should copy 1 on 1 is vdagentd-proto*.h
> >> which defines the protocol over the unix domainsocket.h and maybe udscs.?
> >
> > Hi,
> >
> >   I've finally sat down to do this. It turns out in practice that incorporating udscs is not such a good idea:
> > 1. udscs code is GPL, driver is BSD (solvable, but see the rest)
> > 2. you get a lot of duplication with vdagentd - parsing chunks and messages, basically reading all messages, passing through some, some with changes (clipboard). (~700 lines just for monitors config + mouse + untested xfer, lacking clipboard)
> >
> >   I have a working implementation but Uri suggested a simpler idea and I wanted your opinion:
> > 1. use a named pipe provided by Xspice to vdagentd instead of virtio serial com.redhat.spice.0
> > 2. use a named pipe / unix domain socket to implement a pseudo uinput device. pseudo since I can't implement any of the ioctls, so this will require a small change to vdagentd-uinput.c and a command line argument.
> >
> > Overall the changes to vdagentd/vdagent will be small:
> > 1. vdagent option to pick a different path to udscs UDS.
> > 2. same for vdagentd
> > 3. vdagentd pseudio-uinput support (same packets, no ioctls - presumed device layout, i.e. 2 axis 5 buttons 2 of which are a wheel).
> >
> > Any comments?
> 
> Sounds like a good idea to me. Note you don't need pseudo uinput support you can just make Xspice use the uinput device created by the
> agent using the xorg-x11-drv-evdev driver :)

The reason for not using uinput directly was to avoid having the agent
have the permissions required to do that. But as you suggested in real
life I can just skip agent mouse and use usb tablet for the client mouse
goodness.

> 
> Regards,
> 
> Hans