Re: Confused about hlist_unhashed_lockless()

From: Will Deacon
Date: Fri Jan 31 2020 - 12:21:06 EST


On Fri, Jan 31, 2020 at 09:06:27AM -0800, Eric Dumazet wrote:
> On Fri, Jan 31, 2020 at 8:57 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > On Fri, Jan 31, 2020 at 08:48:05AM -0800, Eric Dumazet wrote:
> > > On Fri, Jan 31, 2020 at 8:43 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > > > Then running these two concurrently on the same node means that
> > > > hlist_unhashed_lockless() doesn't really tell you anything about whether
> > > > or not the node is reachable in the list (i.e. there is another node
> > > > with a next pointer pointing to it). In other words, I think all of
> > > > these outcomes are permitted:
> > > >
> > > > hlist_unhashed_lockless(n) n reachable in list
> > > > 0 0 (No reordering)
> > > > 0 1 (No reordering)
> > > > 1 0 (No reordering)
> > > > 1 1 (Reorder first and last WRITE_ONCEs)
> > > >
> > > > So I must be missing some details about the use-case here. Please could
> > > > you enlighten me? The RCU implementation permits only the first three
> > > > outcomes afaict, why not use that and leave non-RCU hlist as it was?
> > > >
> > >
> > > I guess the following has been lost :
> >
> > Thanks, although...
> >
> > > Author: Eric Dumazet <edumazet@xxxxxxxxxx>
> > > Date: Thu Nov 7 11:23:14 2019 -0800
> > >
> > > timer: use hlist_unhashed_lockless() in timer_pending()
> > >
> > > timer_pending() is mostly used in lockless contexts.
> >
> > ... my point above still stands: the value returned by
> > hlist_unhashed_lockless() doesn't tell you anything about whether or
> > not the timer is reachable in the hlist or not. The comment above
> > timer_pending() also states that:
> >
> > | Callers must ensure serialization wrt. other operations done to
> > | this timer, e.g. interrupt contexts, or other CPUs on SMP.
> >
> > If that is intended to preclude list operations, shouldn't we use an
> > RCU hlist instead of throwing {READ,WRITE}_ONCE() at the problem to
> > shut the sanitiser up without actually fixing anything? :(
>
>
> Sorry, but timer_pending() requires no serialization.

Then we should update the comment!

Without serialisation, timer_pending() as currently implemented does
not reliably tell you whether the timer is in the hlist. Is that not a
problem? Using an RCU hlist does not introduce serialisation, but does
at least rule out the case where timer_pending() returns false for a
timer that /is/ reachable in the list by another CPU.

> The only thing we need is a READ_ONCE() so that compiler is not allowed
> to optimize out stuff like
>
> loop() {
> if (timer_pending())
> something;

If that was the case, then you wouldn't need to touch hlist_add_before()
at all so there's got to be more to it than that or we can revert that
part of the patch.

Will