Serge E. Hallyn wrote:
>> diff --git a/ipc/namespace.c b/ipc/namespace.c
>> index 9171d94..9044505 100644
>> --- a/ipc/namespace.c
>> +++ b/ipc/namespace.c
>> @@ -48,6 +48,16 @@ struct ipc_namespace *copy_ipcs(unsigned long flags, struct ipc_namespace *ns)
>>  	if (!(flags & CLONE_NEWIPC))
>>  		return ns;
>>
>> +	if (!(flags & CLONE_SYSVSEM)) {
>>     
>
> Wait, this should be opposite, right?
>
> [snip]
> Manfred, I'm trying to test this, but can't get an error without this
> patch.  Do you have a testcase?

The patch is bogus, sorry that I didn't notice it immediately.
The problem is ipc/sem.c, function find_undo, lookup_undo:
lookup_undo doesn't check the namespace pointer, thus a simple single 
threaded app can trigger the problem.

Attached is a test app and a kernel patch that shows the problem.
Run the test app immediately after boot (or within a new ipc namespace), 
otherwise the ipc sequence counter will prevent the app from triggering 
the problem: The undo structure that was created before unshare() [i.e. 
with 1 semaphore in it] will be used after unshare() [i.e. semaphore 100 
will be accessed].

With kernel debugging (full slub debugging) enabled, I even got an oops 
when I tried to ipcrm the left over array after running undons, probably 
because the undo structure was freed at exit_sem() within the new 
namespace, but still used in the outer namespace.

--
    Manfred