Radeon KMS bug(?) after Linux kernel 2.6.32 causes a crash of X server 1.8.0+ on termination

[Nix <nix@esperi.org.uk>]

On 18 Apr 2010, nix@esperi.org.uk said:

> So far, every time I've quit X 1.8.0 (1.8-stable tip of tree), it's
> coredumped and left my console unusable until I restart. (I'm using the
> tip of the xf86-video-ati tree, and KMS, both of which worked fine with
> 1.7.5. Obviously I've recompiled all the drivers I'm usingt, or X
> wouldn't work at all...)
>
> The backtrace differs depending on whether auditing is enabled or not.
>
> With auditing on, we are hit with a segfault here:
>
> #0  0x00007f7e06148985 in _xstat () from /lib/libc.so.6
> #1  0x00007f7e061198d0 in __tzfile_read () from /lib/libc.so.6
> #2  0x00007f7e06118c8a in tzset_internal () from /lib/libc.so.6
> #3  0x00007f7e06118df9 in __tz_convert () from /lib/libc.so.6
> #4  0x00007f7e06117439 in ctime () from /lib/libc.so.6
> #5  0x00000000004533c8 in AuditPrefix ()
> #6  0x0000000000453956 in VAuditF ()
> #7  0x0000000000453add in AuditF ()
> #8  0x000000000043e5c6 in CloseDownClient ()
> #9  0x0000000000443af8 in Dispatch ()
> #10 0x0000000000420dc5 in main ()
>
> With it off, I see this instead:
>
> Program received signal SIGTERM, Terminated.
> 0x000000000042904c in FreeClientResources ()
> (gdb) bt
> #0  0x000000000042904c in FreeClientResources ()
> #1  0x000000000043e4c2 in CloseDownClient ()
> #2  0x0000000000443af8 in Dispatch ()
> #3  0x0000000000420dc5 in main ()
>
> which might look like normal termination, except that
> FreeClientResources() of course does not contain an exit(), and the
> console is still unusable.
>
> I suspect a double-free() somewhere, and/or heap corruption.

Nope. This bug only appears with KMS enabled; it does not appear in
2.6.32.10 (when termination happens normally, with only

[   71.267834] Unpin not necessary for ffff88033d66dc00 !

in the kernel log) but does appear with current tip; seen with 2.6.34rc3
and later (i.e. as soon as I upgraded the X server).

So this is either a KMS bug corrupting something in userspace, or a bug
in the X server triggered by the presence of KMS and some feature only
available in 2.6.33+. It's nearly midnight here so I'm going to leave it
at that tonight. I'll do a bisection of the kernel tomorrow, and perhaps
a bisection of the X server, to see where it started crashing between
1.7.5 and 1.8.0.