Re: [PATCH] fix SEM_UNDO with namespaces, take 2

[Pavel Emelyanov <xemul@openvz.org>]

Manfred Spraul wrote:
> Hi,
> 
> below is the second attempt to fix SEM_UNDO + unshare():
> lookup_undo (in ipc/sem.c) is not namespace-aware, thus all entries in 
> sysvsem.undo_list must be from the same namespace.
> The patch enforces that by detaching the current thread from 
> sysvsem.undo_list in switch_task_namespaces() if the ipc namespace is 
> changed.
> 
> The patch boots and passes simple sysvsem+unshare tests.
> 
> Signed-Off-By: Manfred Spraul <manfred@colorfullife.com>
> 
> @@ -211,6 +211,18 @@ void switch_task_namespaces(struct task_struct *p, struct nsproxy *new)
>  
>  	might_sleep();
>  
> +	if ((p->nsproxy == NULL && new != NULL) ||
> +		(p->nsproxy != NULL && new == NULL) ||
> +		(p->nsproxy != NULL && new != NULL && p->nsproxy->ipc_ns != new->ipc_ns)) {
> +		/* switching the IPC namespace is considered equivalent to sys_exit() wrt.
> +		 * to outstanding SEM_UNDO undos: After switching to the new IPC namespace,
> +		 * the semaphore arrays from the old namespace are not accessible anymore.
> +		 * 
> +		 * Additionally, an implicit sys_unshare(CLONE_SYSVSEM) is performed.
> +		 */
> +		exit_sem(p);
> +	}
> +

No, switch_task_namespaces is the wrong place to do this. It is to be 
done in copy_ipc_ns. If you need a task for which a new namespace is 
being prepared, then pass one into it.

>  	ns = p->nsproxy;
>  
>  	rcu_assign_pointer(p->nsproxy, new);