Re: [PATCH] SYSVIPC - Fix the ipc structures initialization

[Manfred Spraul <manfred@colorfullife.com>]

Nadia Derbey wrote:
> That's what the patch I sent at the beginning of this thread
> (http://lkml.org/lkml/2008/10/28/177) fixes: the ipc structure is not
> completely initialized before being inserted in the tree.
>
> The patched kernel has been running for 4 days on my victim without any
> problem, but Clement is still facing a problem (that comes later).
>
> I think this patch should be included anyways, because ipc structures
> initialization is not correct the way it is done today.
>
>   
Definitively: Your patch belongs into 2.6.28 and the next 2.6.27.

> I'm still reviewing the code to see what else could be wrong.
>   
I started looking into the idr code: perhaps idr returns a stale entry? 
A race somewhere in the RCU code.
I.e. the new bug is not corruption due to access before 
spin_lock_init(), but due to access after kfree()?

Clement: Could you send us an ipcs output from before the test?
Do you have any sysv msg queues on your system?
And: Do you use PREEMPT, do you use PREEMPT_RCU?

If there are no message queues around, then every msgget()/IPC_RMID will 
create a new idr layer/tear the layer down again.

--
    Manfred