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

[Arthur Marsh <arthur.marsh@internode.on.net>]

Mathieu Desnoyers wrote on 08/07/15 02:03:
> ----- On Jul 7, 2015, at 3:29 AM, Peter Zijlstra peterz@infradead.org wrote:
>
>> On Tue, Jul 07, 2015 at 02:59:06PM +0930, Arthur Marsh wrote:
>>> I had a single, non-reproducible case of the same lock-up happening on my
>>> other machine running the Linus git head kernel in 64-bit mode.
>>
>> Hmm, disturbing.. I've had my machines run this stuff for weeks and not
>> had anything like this :/
>>
>> Do you have a serial cable between those machines? serial console output
>> will allow capturing more complete traces than these pictures can and
>> might also aid in capturing some extra debug info.
>>
>> In any case, I'll go try and build some debug code.
>
> Arthur: can you double-check if you load any module with --force ?
> This could cause a module header layout mismatch, which can be an
> issue with the changes done by the identified commit: the module
> header layout changes there.
>
> Also, I'm attaching a small patch which serializes both updates and
> reads of the module rbree. Can you try it out ? If the problem
> still shows with the spinlocks in place, that would mean the issue
> is *not* a race between latched rbtree updates and traversals.
>
> Thanks!
>
> Mathieu
>

I'm not aware of any modules being loaded with --force .

I've applied the patch, thanks!

The resultant kernel locked up as follows:

http://www.users.on.net/~arthur.marsh/20150708469.jpg
http://www.users.on.net/~arthur.marsh/20150708470.jpg

Sorry that the first image isn't as clear as the second - it only 
appears for a few seconds.

Hopefully these will provide some clue as to what is happening.

Arthur.