Re: [External] : Re: [REPORT] Null pointer deref in net/core/dev.c on PowerPC

[ALOK TIWARI]

On 12/17/2025 8:52 PM, Eric Dumazet wrote:
> > > I will send the following fix, thanks.
> > >
> > > diff --git a/net/core/dev.c b/net/core/dev.c
> > > index 9094c0fb8c68..36dc5199037e 100644
> > > --- a/net/core/dev.c
> > > +++ b/net/core/dev.c
> > > @@ -4241,9 +4241,11 @@ static inline int __dev_xmit_skb(struct sk_buff
> > > *skb, struct Qdisc *q,
> > >                   int count = 0;
> > >
> > >                   llist_for_each_entry_safe(skb, next, ll_list, ll_node) {
> > > -                       prefetch(next);
> > > -                       prefetch(&next->priority);
> > > -                       skb_mark_not_on_list(skb);
> > > +                       if (next) {
> > > +                               prefetch(next);
> > > +                               prefetch(&next->priority);
> > > +                               skb_mark_not_on_list(skb);
> > > +                       }
> > >                           rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
> > >                           count++;
> > >                   }
> > why not only ?
> > if (likely(next)) {
> >       prefetch(next);
> >       prefetch(&next->priority);
> > }
> Because we also can avoid clearing skb->next, we know it is already NULL.
>
> Since we pay the price of a conditional, let's amortize its cost :/

Thanks a lot for the explanation, I understand the goal of amortizing 
the cost and avoiding unnecessary writes to skb->next.
Would it make sense to add if (likely(next)) around the prefetch?

Thanks,
Alok