Le Fri, Sep 08, 2023 at 09:48:44PM -0400, Joel Fernandes a écrit :
> On Fri, Sep 8, 2023 at 4:36 PM Frederic Weisbecker <frederic@xxxxxxxxxx> wrote:
> >
> > The LOAD-ACQUIRE access performed on rdp->nocb_cb_sleep advertizes
> > ordering callback execution against grace period completion. However
> > this is contradicted by the following:
> >
> > * This LOAD-ACQUIRE doesn't pair with anything. The only counterpart
> > barrier that can be found is the smp_mb() placed after callbacks
> > advancing in nocb_gp_wait(). However the barrier is placed _after_
> > ->nocb_cb_sleep write.
>
> Hmm, on one side you have:
>
> WRITE_ONCE(rdp->nocb_cb_sleep, false);
> smp_mb();
> swake_up_one(&rdp->nocb_cb_wq); /* wakeup -- consider this to be a STORE */
>
> And on another side you have:
> swait_event_interruptible_exclusive(rdp->nocb_cb_wq, ..cond..) /*
> consider this to be a LOAD */
> smp_load_acquire(&rdp->nocb_cb_sleep)
> /* exec CBs (LOAD operations) */
>
> So there seems to be pairing AFAICS.
I must be confused, that would give such pattern:
WRITE X LOAD Y
smp_mb()
WRITE Y smp_load_acquire(X)
How does this pair?
>
> But maybe you are referring to pairing between advancing the callbacks
> and storing to nocb_cb_sleep. In this case, the RELEASE of the nocb
> unlock operation just after advancing should be providing the
> ordering
Right.
> but we still need the acquire this patch deletes.
Why?
>
> > * Callbacks can be concurrently advanced between the LOAD-ACQUIRE on
> > ->nocb_cb_sleep and the call to rcu_segcblist_extract_done_cbs() in
> > rcu_do_batch(), making any ordering based on ->nocb_cb_sleep broken.
>
> If you don't mind, could you elaborate more?
So imagine:
1) Some callbacks are pending
2) A grace period completes, nocb_gp_wait() advance some callbacks to DONE and
some callbacks to WAIT, another grace period starts to handle the latter.
3) Because some callbacks are ready to invoke, nocb_gp_wait() sets
rdp->nocb_cb_sleep to false and wakes up nocb_cb_wait()
4) nocb_cb_wait() does smp_load_acquire(rdp->nocb_cb_sleep) and proceeds
with rcu_do_batch() but it gets preempted right before.
5) The new grace period completes.
6) nocb_gp_wait() does one more round and advances the WAIT callbacks to the
non-empty DONE segment. Also it doesn't need to wake up nocb_cb_wait()
since it's pending and ->nocb_cb_sleep is still false but it force writes
again ->nocb_cb_sleep to false.
7) nocb_cb_wait() resumes and calls rcu_do_batch() without doing a new
load-acquire on ->nocb_cb_sleep, this means the ordering only applies to the
callbacks that were moved to DONE on step 2) but not to those moved to DONE
on step 6).
>
> > * Both rcu_segcblist_extract_done_cbs() and rcu_advance_cbs() are called
> > under the nocb_lock, the latter hereby providing already the desired
> > ACQUIRE semantics.
>
> The acquire orders loads to nocb_cb_sleep with all later loads/stores.
> I am not sure how nocb_lock gives that same behavior since that's
> doing ACQUIRE on the lock access itself and not on nocb_cb_sleep
> access, I'd appreciate it if we can debate this out.
Well, the nocb_lock releases not only the write to nocb_cb_sleep but also
everything that precedes it. So it plays the same role and, most importantly,
it's acquired before calling rcu_segcblist_extract_done_cbs().
>
> Every few months I need a memory-ordering workout so this can be that.
> ;-) You could be onto something.
No worries, I have some more headaches upcoming for all of us on the plate ;-)
Thanks.
