[Spice-devel] Spice agent and LXC

[Charles Ricketts]

Despite my problem with building within my VM environment, I went ahead
and attempted a Fedora LXC+Spice configuration. I'm happy to report it
was successful as well. The only change that had to be made was adding
the -f argument to the Xspice script to keep spice-vdagentd from
attempting ioctl ops on the uinput pipe created by the QXL driver. I'm
not sure why I didn't see this before, looks like I had mistakenly
forgotten to use --prefix=/usr on the vdagent source, so I was using the
distribution-provided version which doesn't have the -f option at all
and apparently doesn't print the "improper ioctl for device" error.
However, upon ensuring I was using the git versions of the vdagents and
the -f flag was added in Xspice, I was again successful!

Hooray!

Well, if nothing else, I hope this discussion helps somebody else
looking to accomplish the same task. And, even if it wasn't intended,
props to the developers for having the foresight in making the QXL
driver/Xspice and the agents flexible enough to use non-character
devices and therefore eliminating the reliance on a virtualized
environment. You guys rock!

Chuck R.

On 11/27/2014 11:27 PM, Charles Ricketts wrote:
> Finally got it set up in an Ubuntu VM, but I get the same segfault
> error sans-LXC. Just a note, in order to actually get it compiled
> within Ubuntu I had to make a change to a Makefile:
>
> spice/server/tests/Makefile: add -lm to LDFLAGS (looks like the LIBM
> line below doesn't get used?)
>
> (hand typed diff, prone to error)
>
> --- Makefile    2014-11-27 22:00:04.138824336 -0600
> +++ Makefile.mod        2014-11-27 21:59:37.802824336 -0600
> @@ -334,7 +334,7 @@
>  INSTALL_STRIP_PROGRAM = $(install_sh) -c -s
>  JPEG_LIBS = -ljpeg
>  LD = /usr/bin/ld -m elf_x86_64
> -LDFLAGS = 
> +LDFLAGS = -lm
> LIBM = -lm
> LIBOBJS =
> LIBRT = -lrt
>
> But, even after getting it successfully built, I get the same segfault
> as in LXC (note, no X is currently running and again this is hand-typed):
>
> $ sudo Xspice --vdagent --vdagent-exec $(which spice-vdagent)
> --vdagentd-exec $(which spice-vdagentd) --disable-ticketing :0
> ...
> Loading extension GLX
> resizing surface0 to 1677216
> memory space from 0x7f10bfba0010 to 0x7f10c6b9d010
> memory space from 0x7f10b6b9f010 to 0x7f10beb9f010
> changing surface0 to 1677216
> memory spice from 0x7f10bfba0010 to 0x7f10c6b9d010
> memory space from 0x7f10b6b9f010 to 0x7f10beb9f010
> (EE)
> (EE) Backtrace:
> (EE) 0: /usr/bin/Xorg (xorg_backtrace+0x48) [0x7f10ce389c78]
> (EE) 1: /usr/bin/Xorg (0x7f10ce1e1000+0x1ac969) [0x7f10ce38d969]
> (EE) 2: /lib/x86_64-linux-gnu/libpthread.so.0 (0x7f10cd2de000+0x10340)
> [0x7f10cd2ee340]
> (EE) 3: /usr/lib/x86_64-linux-gnu/libspice-server.so.1
> (0x7f10c7b5d000+0x20d3c) [0x7f10c7b7dd3c]
> (EE) 4: /usr/lib/x86_64-linux-gnu/libspice-server.so.1
> (0x7f10c7b5d000+0x210dc) [0x7f10c7b7e0dc]
> (EE) 5: /usr/lib/xorg/modules/drivers/spiceqxl_drv.so
> (0x7f10c7e7a000+0x9885) [0x7f10c7e83885]
> (EE) 6: /usr/lib/x86_64-linux-gnu/libspice-server.so.1
> (0x7f10c7b5d000+0x22dde) [0x7f10c7b7fdde]
> (EE) 7: /usr/lib/x86_64-linux-gnu/libspice-server.so.1
> (spice_server_add_interface+0x3a6) [0x7f10c7ba6206]
> (EE) 8: /usr/lib/xorg/modules/drivers/spiceqxl_drv.so
> (0x7f10c7e7a000+0xc2a3) [0x7f10c7e862a3]
> (EE) 9: /usr/bin/Xorg (AddScreen+0x71) [0x7f10ce236d41]
> (EE) 10: /usr/bin/Xorg (InitOutput+0x3c8) [0x7f10ce277ce8]
> (EE) 11: /usr/bin/Xorg (0x7f10ce1e1000+0x5975b) [0x7f10ce23a75b]
> (EE) 12: /lib/x86_64-linux-gnu/libc.so.6 (__libc_start_main+0xf5)
> [0x7f10cbd1dec5]
> (EE) 13: /usr/bin/Xorg (0x7f10ce1e1000+0x44e7e) [0x7f10ce225e77e]
> (EE)
> (EE) Segmentation fault at address 0xd8
> (EE)
> Fatal server error:
> (EE) Caught signal 11 (Segmentation fault). Server aborting
> ....
>
> Kind of glad I hand typed this, because I may have found the error....
> it's using the distribution-provided library from
> /usr/lib/x86_64-linux-gnu/libspice-server.so.1 instead of
> /usr/lib/libspice-server.so.1, that /would/ cause some weird
> problems.. Reordering via ldconfig... success!
>
> I actually got it fully working within Ubuntu. Using `Xspice
> --disable-ticketing --vdagent --vdagentd-exec $(which spice-vdagentd)
> --vdagent-exec $(which spice-vdagent)`. I had to make a modification
> to the Xspice script in order to get it working properly:
>
> --- /usr/bin/Xspice     2014-11-27 22:58:26.114824336 -0600
> +++ Xspice      2014-11-27 22:58:07.722824336 -0600
>  else:
>      if args.vdagent_enabled and args.vdagent_launch:
>          # XXX use systemd --user for this?
> -        vdagentd = launch(args=[args.vdagentd_exec, '-x', '-S',
> vdagentd_uds,
> +        vdagentd = launch(args=[args.vdagentd_exec, '-f', '-x', '-S',
> vdagentd_uds,
>                            '-s', args.vdagent_virtio_path, '-u',
> args.vdagent_uinput_path])
>          time.sleep(1)
>          # TODO wait for uinput pipe open for write
>
> This simply adds the -f parameter to spice-vdagentd since
> /tmp/xspice-uinput is a pipe and not a character devices and,
> therefore, "fake." Without the parameter, spice-vdagend (pretty sure
> it was vdagentd) exits complaining about a bad ioctl. Using the
> modified Xspice allows me to copy/paste and automatic resolution
> switching.
>
> Getting it working with Xspice is definitely an accomplishment and I
> now know the caveats I need to fix within my LXC container. But first,
> my LXC setup depends on X starting with the qxlspice settings --
> moving spiceqxl.xorg.conf to xorg.conf. In order to accomplish this I
> added the undocumented xorg.conf options: "SpiceVdagentEnabled" "0",
> "SpiceVdagentVirtioPath" "/tmp/xspice-virtio",
> "SpiceVdagentUinputPath" "/tmp/xspice-uinput" (taken from
> xf86-video-qxl/src/qxl_driver.c). Now I simply need to figure out how
> to get spice-vdagent and spice-vdagentd to start up alongside it, but
> that's another problem for another day. Overall, a huge success.
>
> But now, for Fedora. I'm unable to even get the QXL driver compiled
> for Fedora, unfortunately. I hit the following error when running `make`:
>
> ./configure: line 18133: syntax error near unexpected token `RANDR,'
> ./configure: line 18133: `XORG_DRIVER_CHECK_EXT(RANDR, randrproto)''
>
> Keep in mind, this is my first experience with Fedora, but I've
> installed every x11 devel and xrandr devel package I can find. I even
> used advice I found on another forum to use `yum groupinstall "X11
> Development Files"` (or some similar group name) to no avail. Oddly
> enough, I didn't have any issues with with compiling within my LXC
> container, but I'm not sure all the steps I took to set that up quite
> honestly. I even compared my installed packages with those from my LXC
> container, and they seem the same.
>
> # yum list '*x11*devel*'
> Installed Packages
> xorg-x11-proto-devel.noarch
> xorg-x11-server-devel.x86_64
> xorg-x11-xkb-utils-devel.x86_64
> xorg-x11-xtrans-devel.noarch
> ...
>
> #yum list '*randr*devel*'
> Installed Packages
> libXrandr-devel.x86_64
> ...
>
> If you guys could point me in the right direction here, it'd be
> greatly appreciated.
>
> Chuck
>
> On 11/27/2014 06:02 AM, Charles Ricketts wrote:
>> Well, I have Spice working perfectly fine in a Windows install. However,
>> seeing as that's not pertinent to the Linux side of things I went ahead
>> and installed Ubuntu 14.04 in Qemu and, as expected, everything worked.
>> I didn't bother with the git sources in this install, because I was 99%
>> sure it was going to work anyway. I don't have a Fedora ISO lying around
>> to test it with, but I imagine that the results would be the same.
>>
>> However, I don't think that even this is pertinent to the problem. The
>> reason I think this is because Qemu acts as the Spice server if I am
>> correct. Qemu relays information from a network socket assigned on the
>> command line to the virtualized serial port and vice versa. Since an LXC
>> installation is sans-Qemu server then I must use Xspice in order to take
>> the place of Qemu and act as a Spice server in order to relay
>> information between the agents/QXL driver and the Spice client. So,
>> testing it within Qemu doesn't really reflect the problem at all. Beyond
>> Qemu, there's really no way to test it sans-LXC. Actually, now that I
>> think about it I may be able to run Xspice directly within a VM and then
>> attempt to connect to it... I'll try that out later on and let you know
>> how/if that works out. I may have to get that Fedora ISO after all just
>> to broaden the test cases.
>>
>> I realize that I'm effectively attempting to use Spice outside of normal
>> circumstances. However, the way that Xspice behaves -- such as creating
>> its own versions of the virtio port (as a socket rather than a character
>> device) and uinput (as a pipe) and attempting to destroy any existing
>> versions of those files -- leads me to believe that Xspice was almost
>> built for the purpose even if not intentionally. And, as I had said
>> before, I got it mostly working in a Fedora LXC container (only lacking
>> client functionality, which is why I asked for input in the first place ;).
>>
>> I would like to be able to provide a stack trace of the segfault I get
>> in the Ubuntu LXC, but I'm unsure how to compile the sources with
>> debugging symbols. Any help in that respect?
>>
>> On 11/27/2014 03:50 AM, Christophe Fergeau wrote:
>>> On Wed, Nov 26, 2014 at 05:16:52PM -0600, Charles Ricketts wrote:
>>>> Yes, I had seen those options. That was part of why I was asking about the
>>>> ucds socket. I found now that the ucds socket is used to talk to multiple
>>>> agents. I have tried both setting each argument to specify the paths of
>>>> each piece (ucds socket, uinput, and virtio port) and letting Xspice set
>>>> them up automatically. Xspice by default re-creates these devices as
>>>> sockets and pipes, so it seems that Xspice is actually ideal for this
>>>> purpose; there is no need to create the devices by hand. However, any way
>>>> it's done even with the newest sources results in the same thing. In the
>>>> Fedora LXC under Ubuntu 14.04, I get a display but no agent functionality
>>>> for some reason even though the same ucds port is used by both agentd and
>>>> agent. In an Ubuntu 14.04 LXC container I can't even get X to start via the
>>>> Xspice script or a direct call to X because it segfaults when using the
>>>> spiceqxl driver.
>>>>
>>>> There also appears to be no difference between letting the Xspice script
>>>> start the agents or starting the agents by hand, which isn't unexpected.
>>>>
>>> Have you tried all of this without lxc to see how well/bad it works
>>> without adding lxc to the mix?
>>>
>>> Christophe
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/spice-devel/attachments/20141128/3c7ded3a/attachment-0001.html>