Re: [PATCH] mm/kmemleak: Fix sleeping function called from invalid context in kmemleak_seq_show

[Catalin Marinas <catalin.marinas@arm.com>]

On Fri, Nov 22, 2024 at 09:14:37AM +0100, Sebastian Andrzej Siewior wrote:
> On 2024-11-21 19:19:05 [+0000], Catalin Marinas wrote:
> …
> > > Maybe SELinux locks should be converted to raw? I don't know how long that
> > > lock is held. There are some loops though :-/
> > > 
> > > avc_insert():
> > > 
> > > 	spin_lock_irqsave(lock, flag);
> > > 	hlist_for_each_entry(pos, head, list) {
> > > 		if (pos->ae.ssid == ssid &&
> > > 			pos->ae.tsid == tsid &&
> > > 			pos->ae.tclass == tclass) {
> > > 			avc_node_replace(node, pos);
> > > 			goto found;
> > > 		}
> > > 	}
> > > 	hlist_add_head_rcu(&node->list, head);
> > > found:
> > > 	spin_unlock_irqrestore(lock, flag);
> > > 
> > > Perhaps that could be converted to simple RCU?
> > > 
> > > As I'm sure there's other places that call vsprintf() under a raw_spin_lock
> > > or non-preemptable context, perhaps this should be the fix we do.
> > 
> > My preference would also be to convert SELinux rather than avoiding the
> > issue in kmemleak (and other similar places).
> 
> No. kmemleak has been made use a raw_spinlock_t because most of what it
> does is something that is not used in production on a PREEMPT_RT system
> and falls in the same category as lockdep for instance. And that code
> calls into LSM/ selinux.
> Before making the lock in selinux a raw_spinlock_t you have to think
> about the consequences in general and audit the code. From a quick
> look, there is also avc_insert() invoked in that callchain which
> allocates memory and this is a no no.
> Also, if you make the solution here in selinux to use a raw_spinlock_t
> you would have to do it also in every LSM as they might be used instead
> of selinux.

Good point, thanks. Kmemleak is indeed a debug tool not supposed to be
used in production. Modifying SELinux has wider implications for
PREEMPT_RT.

> Therefore, I still prefer adding PREEMPT_RT to the restricted_pointer()
> category for atomic invocations.

This should work. If one wants the actual (hashed) pointers with
kmemleak, I guess they can disable kptr_restrict.

-- 
Catalin