Re: [PATCH ipsec-next 01/10] xfrm: state: fix sparse warnings on xfrm_state_hold_rcu

[Sabrina Dubroca <sd@queasysnail.net>]

(adding Simon: maybe you can help with the "comment wording" issue here :))

2026-03-10, 22:10:21 +0200, Leon Romanovsky wrote:
> On Tue, Mar 10, 2026 at 08:49:44PM +0100, Florian Westphal wrote:
> > Leon Romanovsky <leon@kernel.org> wrote:
> > > > You can only use xfrm_state_hold() if the refcount is already > 0.
> > > > xfrm_state_hold uses refcount_inc(), so you get a UaF warn splat
> > > > if this assuption doesn't hold true.
> > > 
> > > I know it, the thing that bothers me is that it is unclear how
> > > xfrm_state_hold_rcu() can have refcount equal to 0.
> > > 
> > > xfrm_state_put() decreases refcount and when it is zero, it calls
> > > to __xfrm_state_destroy(). The latter assumes that the state was
> > > already removed from various hlists.
> > 
> > Yes, insertion in the table means refcount is 1, but userspace
> > can zap states at any time, e.g.:
> > 
> > xfrm_del_sa -> xfrm_state_delete -> __xfrm_state_delete (which
> > unlinks from hash lists).
> > 
> > The last xfrm_state_put() in that function may cause 1 -> 0
> > transition.  Parallel lookup can still observe that state,
> > so it has to pretend it wasn't there to begin with.

Thanks Florian.


> Yes, this is possible scenario and this is what is worth to document.


We could add something like:

/* Take a reference to @x, when we know the state has a refcount >= 1.
 * In this case, we can avoid refcount_inc_not_zero and the error
 * handling it requires.
 * In contexts where concurrent state deletion is possible and we
 * don't already hold a reference to that state, xfrm_state_hold_rcu
 * must be used.
 */

Though it may not make much sense to refer to xfrm_state_hold_rcu
(implemented in net/xfrm/xfrm_state.c) from include/net/xfrm.h.

And if we consider the hashtables to be private to
net/xfrm/xfrm_state.c, nothing outside of it should ever see a state
with refcount=0, since they will only ever see states that already
have one reference held by whatever gave them the pointer.

So maybe it's more xfrm_state_hold_rcu that needs a mention of
"concurrent state deletion could bring the refcount to 0 while we're
doing the lookup"? I don't know, for me it's pretty obvious with the
_rcu suffix that RCU -> unlocked -> could be deleted in parallel.


We also have __xfrm_state_put, the commit message that introduced it
states:

    We often just do an atomic_dec(&x->refcnt) on an xfrm_state object
    because we know there is more than 1 reference remaining and thus
    we can elide the heavier xfrm_state_put() call.

so we could add:

/* Drop a reference to @x, when we know there is more than 1 reference remaining.
 * In this case, we can avoid refcount_dec_and_test and just decrement refcnt.
 */

but maybe someone has a better suggestion.


-- 
Sabrina