Re: Possible Oprofile crash/race when stopping

[Robert Richter <robert.richter@amd.com>]

On 22.07.10 01:14:40, Benjamin Herrenschmidt wrote:
> Hi folks !
> 
> We've hit a strange crash internally, that we -think- we have tracked
> down to an oprofile bug. It's hard to hit, so I can't guarantee yet that
> we have fully smashed it but I'd like to share our findings in case you
> guys have a better idea.
> 
> So the initial observation is a spinlock bad magic followed by a crash
> in the spinlock debug code:

Benjamin,

thanks for reporting this. I was trying to reproduce this with various
loads and scenarios, but without success so far. Can you give me a
hint of the load you have (number of processes running, cpu load, do
you switch off oprofile while many processes are still running)? Are
you able to regularly trigger it?

> I think the right sequence however requires breaking up end_sync. Ie, we
> need to do in that order:
> 
>   - cancel the workqueues
>   - unregister the notifier
>   - process the mortuary
> 
> What do you think ?

This could potentially fix it, I will have to look deeper into the
code. Try to do this next week.

Thanks,

-Robert

-- 
Advanced Micro Devices, Inc.
Operating System Research Center