[gdm-list] screen locks up after a gdm greeter login

[Eyal Lebedinsky <gdm eyal emu id au>]

I found many references to similar problems but none are suitable in my case.
I know that I am using an old video card but it is understood to be supported.
I think that the problem is in the way the server starts and may be easy to
work around. Details below.

The symptom is this:
- I start the computer (cold start)
  Waking up from a hibernation does not have a problem
- I am presented with a login screen, and use it
- the screen freezes, no keyboard or mouse         <<<<< the problem
- I kill Xorg (ssh from another machine)           <<<<< the current ugly fix
- I am presented with a login screen again, and use it
- my session starts and all is well

I investigated the matter and now think that I see a problem.

First, /var/log/Xorg.1.log (from the first attempt) ends with
    [    35.865] (II) NVIDIA(0): NVIDIA GPU GeForce 8600 GT (G84) at PCI:1:0:0 (GPU-0)
    [    35.865] (--) NVIDIA(0): Memory: 524288 kBytes
    [    35.865] (--) NVIDIA(0): VideoBIOS: 60.84.59.00.00
    [    35.865] (II) NVIDIA(0): Detected PCI Express Link width: 16X
    [    35.869] (EE) NVIDIA(GPU-0): EVO Push buffer channel allocation failed
    [    35.872] (EE) NVIDIA(GPU-0): Failed to allocate EVO core DMA push buffer
    [    35.873] (EE) NVIDIA(0): Failing initialization of X screen 0
    ...
    [    35.874] (EE) Server terminated with error (1). Closing log file.
(and I think that this is where the freeze comes from).

BTW, It may be a problem with the nvidia driver that cannot handle multiple
concurrent sessions, but in this case there should not be multiple sessions.

As described above, the second login attempt shows no error.
Looking at the timestamps one sees this:

       Xorg.0.log         Xorg.1.log
1st     19.310- 35.891     35.764- 35.874
               [35.869] (EE) NVIDIA(GPU-0): EVO Push buffer channel allocation failed
2nd    502.160-516.936    517.023-519.530

It is clear that in the first attempt the second session starts (and fails) before the first one exits.
In the second attempt there is no overlap.

Is it possible that the overlap causes the error? Where can I put a (small) delay to test this
proposition and avoid the overlap? Or is there another way to avoid it?
[later] I tried "another way" by changing from GDM to LXDM and this resolves the problem.
Is this a gdm problem or something in the distro (f26) or im setup?

My system
=========

I am on f26, using xfce wm (not that this should matter?). The driver:
  (II) NVIDIA GLX Module  340.102  Mon Jan 16 12:37:38 PST 2017
  (II) NVIDIA dlloader X Driver  340.102  Mon Jan 16 12:17:09 PST 2017

$ dnf list installed '*nvidia*'
akmod-nvidia-340xx.x86_64                                  1:340.102-4.fc26            @@commandline
kmod-nvidia-340xx-4.11.10-100.fc24.x86_64.x86_64           1:340.102-4.fc26            @@commandline
kmod-nvidia-340xx-4.11.12-100.fc24.x86_64.x86_64           1:340.102-4.fc26            @@commandline
kmod-nvidia-340xx-4.12.5-300.fc26.x86_64.x86_64            1:340.102-4.fc26            @@commandline
xorg-x11-drv-nvidia-340xx.x86_64                           1:340.102-1.fc26            @@commandline
xorg-x11-drv-nvidia-340xx-kmodsrc.x86_64                   1:340.102-1.fc26            @@commandline
xorg-x11-drv-nvidia-340xx-libs.x86_64                      1:340.102-1.fc26            @@commandline
xorg-x11-drv-nvidia-kmodsrc.x86_64                         2:375.66-7.fc26             @@commandline

I think that I had this problem on f24 too, but instead of a frozen screen
I had a blank (monitor soon went into power saving) screen. Same workaround.

TIA

--
Eyal Lebedinsky (gdm eyal emu id au)