Re: [PATCH 2/2] fix sys_unshare()+SEM_UNDO: perform an implicit CLONE_SYSVSEM in CLONE_NEWIPC

[Manfred Spraul <manfred@colorfullife.com>]

Serge E. Hallyn wrote:
> Quoting Manfred Spraul (manfred@colorfullife.com):
>   
>> sys_unshare(CLONE_NEWIPC) doesn't handle the undo lists properly, this can
>> cause a kernel memory corruption. CLONE_NEWIPC must detach from the existing
>> undo lists.
>> Fix, part 2: perform an implicit CLONE_SYSVSEM in CLONE_NEWIPC.
>> CLONE_NEWIPC creates a new IPC namespace, the task cannot access the
>> existing semaphore arrays after the unshare syscall. Thus the task
>> can/must detach from the existing undo list entries, too.
>>
>> This fixes the kernel corruption, because it makes it impossible that
>> undo records from two different namespaces are in sysvsem.undo_list.
>>     
>
> So this was never an issue with clone(CLONE_NEWIPC|CLONE_SYSVSEM), which
> should have had the same result as unshare(CLONE_NEWIPC&~CLONE_SYSVSEM)?
>
>   
Actually, the story is slightly different:
unshare(CLONE_NEWIPC|CLONE_SYSVSEM) returns -EINVAL right now.

Thus all apps right now call unshare(CLONE_NEWIPC|&~CLONE_SYSVSEM).
This combination doesn't make much sense. Even worse - it easily causes 
a kernel oops.
Thus my fix is twofold:
- add support for unshare(CLONE_SYSVSEM).
- implicitely add CLONE_SYSVSEM to all calls that set CLONE_NEWIPC.

It's not really pretty: If a pivot_namespace syscall is ever added, then 
CLONE_NEWIPC&~CLONE_SYSVSEM would make sense again.
What do you think? Can we break backward compatibility and add
    if ( (unshare_flags & CLONE_NEWIPC) && !(unshare_flags & 
CLONE_SYSVSEM) )
       return -EINVAL;
into sys_unshare()?
I have decided against that, it breaks the current ABI.
And we gain virtually nothing - most if not all unshare users will be 
single threaded apps that do not use sysvsem at all, and even most 
sysvsem users do not use SEM_UNDO.

--
    Manfred