Re: [RFC, PATCH] fix SEM_UNDO with namespaces

["Serge E. Hallyn" <serue@us.ibm.com>]

Quoting Serge E. Hallyn (serue@us.ibm.com):
> Quoting Eric W. Biederman (ebiederm@xmission.com):
> > Pavel Emelyanov <xemul@openvz.org> writes:
> > >>> I agree, that we should probably destroy this one when the task calls 
> > >>> unshare, but trying to keep this list relevant is useless.
> > >>>   
> > >> A very tricky question: Let's assume we have a process with two threads.
> > >> The undo structure is shared, as per opengroup standard.
> > >> Now one thread calls unshare(CLONE_NEWIPC). What should happen? We 
> > >> cannot destroy the undo structure, the other thread might be still 
> > >> interested in it.
> > >> If we allow sys_unshare() for multithreaded processes with CLONE_NEWIPC 
> > >> and without CLONE_SYSVSEM, then we must handle this case.
> > >
> > > Hm... I'd simply disable creating any new namespaces for threads.
> > > I think other namespaces developers agree with me. Serge, Suka, Eric
> > > what do you think?
> > 
> > I almost agree.  sys_unshare() in a multithreaded process breaks
> > all kinds of user space libs.  So you can only reasonably look at
> > the problem as what we do with linux tasks that share some things
> > and not others.  The posix/opengroup notion of processes and threads
> > are a distraction.
> > 
> > In this case requiring it appears that to require unsharing both
> > CLONE_SYSVSEM and CLONE_NEWIPC at the same time.  (i.e. unshare
> > of CLONE_SYSVSEM should fail if CLONE_NEWIPC is not also specified).
> > 
> > Then to make it work we make unshare of SYSVSEM succeed when it is
> > not shared.
> > 
> > This looks like about a 5 line patch or two.
> > 
> > The effect is because we don't support unsharing of SYSVSEM currently
> > we don't support a threaded process unsharing the ipc namespace.
> > 
> > Eric
> 
> Eric, does the following patch correctly interpret your recommendation?
> 
> Pavel does it make sense to you?
> 
> thanks,
> -serge
> 
> >From 9c85fb3cb80cea1d888c3c253a9fb6e9bc173649 Mon Sep 17 00:00:00 2001
> From: Serge E. Hallyn <serue@us.ibm.com>
> Date: Thu, 3 Apr 2008 12:43:23 -0700
> Subject: [PATCH 1/1] ipc namespaces: fix svsem unsharing issue
> 
> Refuse to unshare an ipcns if the semundo is shared and we
> are not requesting a new SYSVSEM
> 
> Signed-off-by: Serge E. Hallyn <serue@us.ibm.com>
> ---
>  ipc/namespace.c |   10 ++++++++++
>  1 files changed, 10 insertions(+), 0 deletions(-)
> 
> 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?

> +		if (!current->sysvsem.undo_list)
> +			goto ok;
> +		if (atomic_read(&current->sysvsem.undo_list->refcnt) == 1)

And the refcnt check really isn't needed, right?  Either undo_list
exists and both parent and child have it, or it doesn't and we're fine.
So I suspect I can remove that.

Manfred, I'm trying to test this, but can't get an error without this
patch.  Do you have a testcase?  (I've tried some combinations of
creating a semarray in the new ipcns, creating one with fewer elements
than the index on which the undo operation was done, and not creating
a semarray in the new ipcsns at all, but the current code seems to
do fine.

Or maybe I fundamentally misunderstand the bug.

> +			goto ok;
> +		put_ipc_ns(ns);
> +		return ERR_PTR(-EINVAL);
> +	}
> +
> +ok:
>  	new_ns = clone_ipc_ns(ns);
> 
>  	put_ipc_ns(ns);
> -- 
> 1.5.3.6