Bug 14121

happens to me as well.
In my case it does recover after the crash but the keyboard map is acting weird until I restart dbus/hal.

I use xserver 1.4.0.90 with dbus 1.1.4 and hal 0.5.10 
And I'm also on gentoo.

Created attachment 13915 [details] [review]
Doublecheck all connections before DBus kills us with an assert.

Please try the attached patched, they stop the errors for me.

Problem is simply the libdbus insists on using assert to kill us all if something isn't quite right. So if we try to shutdown hal when the connection is already down, libdbus kills us.
If we try to get the dispatch status of the current message and the connection is already down, libdbus kills us.

Thanks Peter! They mostly look good, but the first one needs to also have an if (info->connection) guard _inside_ the do { } while block, else you get:
info->connection is NULL
do:
   dbus_blah(info->connection);
     ha ha, assert!

I'd actually fixed this in my Nokia tree, but forgot to cherry-pick the fix.  Please feel free to merge, and thanks again for picking up after me while I'm sans connectivity. :)

(In reply to comment #3)
> Thanks Peter! They mostly look good, but the first one needs to also have an if
> (info->connection) guard _inside_ the do { } while block, else you get:
> info->connection is NULL
> do:
>    dbus_blah(info->connection);
>      ha ha, assert!

I don't think so. Look at the context

if (info-connection != NULL)
   do
       dbus_oh_god_please_dont_kill_me_please_please_please()
   while(info->connection)

the first run is guarded and from then on the loop protects us.

Oh, right, I'm just being daft.  Must've just meant to cherry-pick the while (info->connection) fix then, 'cause you can get disconnected after a dispatch call.  Thanks!

The patches fix it only half way for me. Stopping dbus+hald is no problem anymore, but when starting dbus again, X shuts down after about 1 second. I find this message in the log file:

[config/hal] couldn't initialise context: (null) ((null))

(In reply to comment #6)
> The patches fix it only half way for me. Stopping dbus+hald is no problem
> anymore, but when starting dbus again, X shuts down after about 1 second. I
> find this message in the log file:
> 
> [config/hal] couldn't initialise context: (null) ((null))
> 

Yeah, I found that yesterday too but ran out of time to fix it. the dbus API doesn't init properly again either. I'll get the patch out ASAP.

Created attachment 13951 [details] [review]
patches to stop dbus from bringing down the server.

(In reply to comment #6)
> The patches fix it only half way for me. Stopping dbus+hald is no problem
> anymore, but when starting dbus again, X shuts down after about 1 second. I
> find this message in the log file:
> 
> [config/hal] couldn't initialise context: (null) ((null))

Ok, I had a look but I couldn't reproduce your bug. The errors I got was from re-initialising the dbus-api, they are fixed with the new patch.

I think you're hitting a race condition. If hal isn't up yet but dbus is running, then initialising the hal layer will fail. 

This shouldn't bring down the server though, so anything reproducable (or more extensive log messages) would help greatly. if you define DEBUG when compiling xserver/config/ then there should me more spew to the console.

Thanks, attachment 13951 [details] [review] seems to fix it. I tried a lot with restarting dbus and hald and I can't reproduce the bug anymore.

If dbus is up but hald is not, it seems not to connect to hal anymore, even if it comes up later. So restarting dbus (which itself stops and starts hald on gentoo) can cause a tiny race condition, if you care. But it's not a crasher.

(In reply to comment #10)
> Thanks, attachment 13951 [details] [review] seems to fix it. I tried a lot with restarting dbus
> and hald and I can't reproduce the bug anymore.
> 
> If dbus is up but hald is not, it seems not to connect to hal anymore, even if
> it comes up later. So restarting dbus (which itself stops and starts hald on
> gentoo) can cause a tiny race condition, if you care. But it's not a crasher.
> 

pushed as commits 
975ab11799c819a81da1dfe83505194410dbcb95
7dde5a694a06efed0a9186f05d33f5be6f5dba71
f0ba7707161b8866e6fde32d6f25be6afcdecb48
2cb0ebec2b85d96289c23c17cfdcdf97ef6877d2