On Thu, Nov 07, 2019 at 08:39:42AM -0800, Eric Dumazet wrote:
> On Thu, Nov 7, 2019 at 8:35 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> >
> > On Thu, Nov 7, 2019 at 8:11 AM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > >
> > > OK, so this is due to timer_pending() lockless access to ->entry.pprev
> > > to determine whether or not the timer is on the list. New one on me!
> > >
> > > Given that use case, I don't have an objection to your patch to list.h.
> > >
> > > Except...
> > >
> > > Would it make sense to add a READ_ONCE() to hlist_unhashed()
> > > and to then make timer_pending() invoke hlist_unhashed()? That
> > > would better confine the needed uses of READ_ONCE().
> >
> > Sounds good to me, I had the same idea but was too lazy to look at the
> > history of timer_pending()
> > to check if the pprev pointer check was really the same underlying idea.
>
> Note that forcing READ_ONCE() in hlist_unhashed() might force the compiler
> to read the pprev pointer twice in some cases.
>
> This was one of the reason for me to add skb_queue_empty_lockless()
> variant in include/linux/skbuff.h
Ouch!
> /**
> * skb_queue_empty_lockless - check if a queue is empty
> * @list: queue head
> *
> * Returns true if the queue is empty, false otherwise.
> * This variant can be used in lockless contexts.
> */
> static inline bool skb_queue_empty_lockless(const struct sk_buff_head *list)
> {
> return READ_ONCE(list->next) == (const struct sk_buff *) list;
> }
>
> So maybe add a hlist_unhashed_lockless() to clearly document why
> callers are using the lockless variant ?
That sounds like a reasonable approach to me. There aren't all that
many uses of hlist_unhashed(), so a name change should not be a problem.
Thanx, Paul
|