Re: [PATCH net] udp: drop secpath before storing an skb in a receive queue

[Eric Dumazet <edumazet@google.com>]

On Tue, Oct 14, 2025 at 1:28 AM Sabrina Dubroca <sd@queasysnail.net> wrote:
>
> 2025-10-14, 01:06:04 -0700, Eric Dumazet wrote:
> > On Tue, Oct 14, 2025 at 1:01 AM Eric Dumazet <edumazet@google.com> wrote:
> > >
> > > On Tue, Oct 14, 2025 at 12:43 AM Eric Dumazet <edumazet@google.com> wrote:
> > > >
> > > > On Tue, Oct 14, 2025 at 12:32 AM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > >
> > > > >
> > > > >
> > > > > On 10/14/25 8:37 AM, Sabrina Dubroca wrote:
> > > > > > 2025-10-14, 06:04:54 +0000, Eric Dumazet wrote:
> > > > > >> Michal reported and bisected an issue after recent adoption
> > > > > >> of skb_attempt_defer_free() in UDP.
> > > > > >>
> > > > > >> We had the same issue for TCP, that Sabrina fixed in commit 9b6412e6979f
> > > > > >> ("tcp: drop secpath at the same time as we currently drop dst")
> > > > > >
> > > > > > I'm not convinced this is the same bug. The TCP one was a "leaked"
> > > > > > reference (delayed put). This looks more like a double put/missing
> > > > > > hold to me (we get to the destroy path without having done the proper
> > > > > > delete, which would set XFRM_STATE_DEAD).
> > > > > >
> > > > > > And this shouldn't be an issue after b441cf3f8c4b ("xfrm: delete
> > > > > > x->tunnel as we delete x").
> > > > >
> > > > > I think Sabrina is right. If the skb carries a secpath,
> > > > > UDP_SKB_IS_STATELESS is not set, and skb_release_head_state() will be
> > > > > called by skb_consume_udp().
> > > > >
> > > > > skb_ext_put() does not clear skb->extensions nor ext->refcnt, if
> > > > > skb_attempt_defer_free() enters the slow path (kfree_skb_napi_cache()),
> > > > > the skb will go through again skb_release_head_state(), with a double free.
> > > > >
> > > > > I think something alike the following (completely untested) should work:
> > > > > ---
> > > > > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> > > > > index 95241093b7f0..4a308fd6aa6c 100644
> > > > > --- a/net/ipv4/udp.c
> > > > > +++ b/net/ipv4/udp.c
> > > > > @@ -1851,8 +1851,10 @@ void skb_consume_udp(struct sock *sk, struct
> > > > > sk_buff *skb, int len)
> > > > >                 sk_peek_offset_bwd(sk, len);
> > > > >
> > > > >         if (!skb_shared(skb)) {
> > > > > -               if (unlikely(udp_skb_has_head_state(skb)))
> > > > > +               if (unlikely(udp_skb_has_head_state(skb))) {
> > > > >                         skb_release_head_state(skb);
> > > > > +                       skb->active_extensions = 0;
> > >
> > > We probably also want to clear CONNTRACK state as well.
> >
> > Perhaps not use skb_release_head_state() ?
> >
> > We know there is no dst, and no destructor.
>
> Then, do we need to do anything before calling skb_attempt_defer_free()?
> skb_attempt_defer_free() only wants no dst and no destructor, and the
> secpath issue that we dealt with in TCP is not a problem anymore.
>
> Can we just drop the udp_skb_has_head_state() special handling and
> simply call skb_attempt_defer_free()?

Good point ! I need a second cup of coffee !