Re: lock-up with module: Optimize __module_address() using a latched RB-tree

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Jul 06, 2015 at 04:03:45AM +0930, Arthur Marsh wrote:
> On this machine, a single core Athlon 64 with a 32 bit current Linus' git
> head kernel, I get a lock-up early in the boot process. (A dmesg output of a
> successful boot-up of kernel 4.1.0 up to and slightly passed the point where
> the git head kernel locks up is attached).
> 
> A photo of the lock-up appears at:
> 
> http://www.users.on.net/~arthur.marsh/20150706462.jpg

So building a kernel with your .config (and a similar GCC) I was able to
match up the Code: with actual compiler output.

The faulting instruction is a dereference of mod->module_core through:

__modules_address()
  mod_find()
    latch_tree_find()
      mod_tree_comp()
        __mod_tree_val()

    20e0:       8b 90 44 01 00 00       mov    0x144(%eax),%edx

Now eax is NULL, which will have given you the splat.

The curious thing is that the mod pointer is obtained from the
mod_tree_node structure:

    20d0:       8b 46 fc                mov    -0x4(%esi),%eax

And esi looks like a regular kernel pointer.

Furthermore, we explicitly set the mod pointers in mod_tree_insert()
before linking in the nodes.

This all has the smell of memory corruption, this gfx card you need to
pull, you're not accidentally using a binary driver for that?