Re: [tip: timers/core] hrtimer: Annotate lockless access to timer->state

[Paul E. McKenney]

On Wed, Nov 06, 2019 at 02:59:36PM -0800, Eric Dumazet wrote:
> On Wed, Nov 6, 2019 at 2:53 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> >
> > On Wed, Nov 6, 2019 at 2:24 PM tip-bot2 for Eric Dumazet
> > <tip-bot2@xxxxxxxxxxxxx> wrote:
> > >
> > > The following commit has been merged into the timers/core branch of tip:
> > >
> > > Commit-ID:     56144737e67329c9aaed15f942d46a6302e2e3d8
> > > Gitweb:        https://git.kernel.org/tip/56144737e67329c9aaed15f942d46a6302e2e3d8
> > > Author:        Eric Dumazet <edumazet@xxxxxxxxxx>
> > > AuthorDate:    Wed, 06 Nov 2019 09:48:04 -08:00
> > > Committer:     Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > > CommitterDate: Wed, 06 Nov 2019 23:18:31 +01:00
> > >
> > > hrtimer: Annotate lockless access to timer->state
> > >
> >
> > I guess we also need to fix timer_pending(), since timer->entry.pprev
> > could change while we read it.
> 
> It is interesting seeing hlist_add_head() has a WRITE_ONCE(h->first, n);,
> but no WRITE_ONCE() for the pprev change.
> 
> The WRITE_ONCE() was added in commit 1c97be677f72b3c338312aecd36d8fff20322f32
> ("list: Use WRITE_ONCE() when adding to lists and hlists")

The theory is that while the ->next pointer is concurrently accessed by
RCU readers, the ->pprev pointer is accessed only by updaters, who need
to supply sufficient synchronization.

But what is this theory missing in practice?

							Thanx, Paul