LIBGL_ALWAYS_INDIRECT causes firefox crashes

[Alkis Georgopoulos]

In some cases, the default LIBGL_ALWAYS_INDIRECT=true is causing the following problems on thin clients:
 * When starting firefox, xorg crashes, without any relevant messages in .xsession-errors or in Xorg.log.
 * When *closing* chromium-browser, xorg crashes.

The first of the setups where this was observed had the following graphics card on both the server and the clients, and had no proprietary drivers installed:

# lspci -nn -k | grep -A 2 VGA
01:00.0 VGA compatible controller [0300]: NVIDIA Corporation NV18 [GeForce4 MX 440 AGP 8x] [10de:0181] (rev c1)
         Kernel driver in use: nouveau
         Kernel modules: nouveau, nvidiafb

That's on Ubuntu Precise, fully updated.

[Alkis Georgopoulos (alkisg)]

In another occurance of the problem, the client had:

# lspci -nn -k |grep -A 2 VGA
01:00.0 VGA compatible controller [0300]: NVIDIA Corporation NV5M64 [RIVA TNT2 Model 64/Model 64 Pro] [10de:002d] (rev 15)
Kernel driver in use: nouveau
Kernel modules: nouveau, nvidiafb, rivafb

...while the server did not have nvidia in that case:
# lspci -nn -k | grep -A 2 VGA
00:02.0 VGA compatible controller [0300]: Intel Corporation 4 Series Chipset Integrated Graphics Controller [8086:2e22] (rev 03)
 Subsystem: Hewlett-Packard Company Device [103c:2a94]
 Kernel driver in use: i915

[Alkis Georgopoulos (alkisg)]

Third case:

# lspci -nn -k | grep -A 2 VGA
01:00.0 VGA compatible controller [0300]: NVIDIA Corporation NV5M64 [RIVA TNT2 Model 64/Model 64 Pro] [10de:002d] (rev 15)
        Kernel driver in use: nouveau
        Kernel modules: nouveau, nvidiafb, rivafb

We should probably start by disabling LIBGL_ALWAYS_INDIRECT for all nvidia cards, and then try to hunt down the actual bug.

[Alkis Georgopoulos (alkisg)]

Attaching Xorg.log where the segfault in nouveau_vieux_dri.so can be seen.

An easy way to reproduce the problem without LTSP:
(from a client with an nvidia card):
$ ssh -X server
$ LIBGL_ALWAYS_INDIRECT=true firefox

One workaround mentioned in #nouveau:
(08:34:05 πμ) airlied: got all the accel you are going to get you might just mv the nouveau_vieux_dri.so aside

Tested it and it works fine, maybe we can do that on in /usr/share/ldm/rc.d/X51-opengl (inside the "if" so that it affects thin clients only).

[Alkis Georgopoulos (alkisg)]

In this case, although with Precise not updated for a couple of months, the problem did NOT happen:

# lspci -nn -k | grep -A 2 VGA
01:00.0 VGA compatible controller [0300]: NVIDIA Corporation NV11 [GeForce2 MX/MX 400] [10de:0110] (rev b2)
        Kernel driver in use: nouveau
        Kernel modules: nouveau, nvidiafb, rivafb

[Alkis Georgopoulos (alkisg)]

Some people are reporting that this doesn't happen in a fresh 12.04.1 installation, but it happens after they `apt-get update` their server.

[Alkis Georgopoulos (alkisg)]

Another teacher reports the same problem in these clients:

root@ltsp135:~# lspci -nn -k | grep -A 2 VGA
01:00.0 VGA compatible controller [0300]: NVIDIA Corporation NV18
[GeForce4 MX 4000] [10de:0185] (rev c1)
   Subsystem: Micro-Star International Co., Ltd. Device [1462:9363]
   Kernel driver in use: nouveau

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Christian (c-pradelli)]

I having this same problem in all my ltsp clients that have nvidia video card.
I'm using Ubuntu 12.04.1 64 bits.
Is there any workarround?

[Vagrant Cascadian (vagrantc)]

On Fri, Nov 30, 2012 at 09:25:38PM -0000, Christian wrote:
> I having this same problem in all my ltsp clients that have nvidia video card.
> I'm using Ubuntu 12.04.1 64 bits.
> Is there any workarround?

You can create an empty file to override it:

  sudo mkdir -p /opt/ltsp/i386/etc/ldm/rc.d/
  sudo touch /opt/ltsp/i386/etc/ldm/rc.d/X51-opengl
  sudo ltsp-update-image

Adjust i386 for amd64 if you're using amd64 clients.

Note that if it is ever fixed properly, you'll want to remove this empty file.

live well,
  vagrant

[Alkis Georgopoulos (alkisg)]

Fix committed in http://bazaar.launchpad.net/~ltsp-upstream/ltsp/ldm-trunk/revision/1453.
It is just a temporary workaround, it should be removed in the future once the nouveau bug is resolved.

It'd be nice if someone could SRU this, as many LTSP users are affected.

[Alkis Georgopoulos (alkisg)]

The committed fix is applied too late, after xorg is loaded. So it's not working on the first login, but only on subsequent logins.
The `rm -f /usr/lib/i386-linux-gnu/dri/nouveau_vieux_dri.so` command should be run from some init-ltsp.d script instead.

A workaround that doesn't involve updating the chroot is to put this in lts.conf:
[Default]
RCFILE_01="rm -f /usr/lib/i386-linux-gnu/dri/nouveau_vieux_dri.so"

[Alkis Georgopoulos (alkisg)]

Fix committed in http://bazaar.launchpad.net/~ltsp-upstream/ltsp/ltsp-trunk/revision/2447.

[Alkis Georgopoulos (alkisg)]

Fix released in LTSP 5.4.5.