On Fri, Jan 20, 2023 at 12:33:00PM -0800, Paul E. McKenney wrote:
> On Fri, Jan 20, 2023 at 08:27:03AM -0500, Joel Fernandes wrote:
> >
> >
> > > On Jan 20, 2023, at 3:19 AM, Zhang, Qiang1 <qiang1.zhang@xxxxxxxxx> wrote:
> > >
> > >
> > >>
> > >>
> > >>>> On Wed, Jan 18, 2023 at 03:30:14PM +0800, Zqiang wrote:
> > >>>>> When inovke rcu_report_qs_rdp(), if current CPU's rcu_data structure's ->
> > >>>>> grpmask has not been cleared from the corresponding rcu_node structure's
> > >>>>> ->qsmask, after that will clear and report quiescent state, but in this
> > >>>>> time, this also means that current grace period is not end, the current
> > >>>>> grace period is ongoing, because the rcu_gp_in_progress() currently return
> > >>>>> true, so for non-offloaded rdp, invoke rcu_accelerate_cbs() is impossible
> > >>>>> to return true.
> > >>>>>
> > >>>>> This commit therefore remove impossible rcu_gp_kthread_wake() calling.
> > >>>>>
> > >>>>> Signed-off-by: Zqiang <qiang1.zhang@xxxxxxxxx>
> > >>>>> Reviewed-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > >>>
> > >>> Queued (wordsmithed as shown below, as always, please check) for further
> > >>> testing and review, thank you both!
> > >>>
> > >>> Thanx, Paul
> > >>>
> > >>> ------------------------------------------------------------------------
> > >>>
> > >>> commit fbe3e300ec8b3edd2b8f84dab4dc98947cf71eb8
> > >>> Author: Zqiang <qiang1.zhang@xxxxxxxxx>
> > >>> Date: Wed Jan 18 15:30:14 2023 +0800
> > >>>
> > >>> rcu: Remove never-set needwake assignment from rcu_report_qs_rdp()
> > >>>
> > >>> The rcu_accelerate_cbs() function is invoked by rcu_report_qs_rdp()
> > >>> only if there is a grace period in progress that is still blocked
> > >>> by at least one CPU on this rcu_node structure. This means that
> > >>> rcu_accelerate_cbs() should never return the value true, and thus that
> > >>> this function should never set the needwake variable and in turn never
> > >>> invoke rcu_gp_kthread_wake().
> > >>>
> > >>> This commit therefore removes the needwake variable and the invocation
> > >>> of rcu_gp_kthread_wake() in favor of a WARN_ON_ONCE() on the call to
> > >>> rcu_accelerate_cbs(). The purpose of this new WARN_ON_ONCE() is to
> > >>> detect situations where the system's opinion differs from ours.
> > >>>
> > >>> Signed-off-by: Zqiang <qiang1.zhang@xxxxxxxxx>
> > >>> Reviewed-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > >>> Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > >>>
> > >>> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > >>> index b2c2045294780..7a3085ad0a7df 100644
> > >>> --- a/kernel/rcu/tree.c
> > >>> +++ b/kernel/rcu/tree.c
> > >>> @@ -1956,7 +1956,6 @@ rcu_report_qs_rdp(struct rcu_data *rdp)
> > >>> {
> > >>> unsigned long flags;
> > >>> unsigned long mask;
> > >>> - bool needwake = false;
> > >>> bool needacc = false;
> > >>> struct rcu_node *rnp;
> > >>>
> > >>> @@ -1988,7 +1987,12 @@ rcu_report_qs_rdp(struct rcu_data *rdp)
> > >>> * NOCB kthreads have their own way to deal with that...
> > >>> */
> > >>> if (!rcu_rdp_is_offloaded(rdp)) {
> > >>> - needwake = rcu_accelerate_cbs(rnp, rdp);
> > >>> + /*
> > >>> + * The current GP has not yet ended, so it
> > >>> + * should not be possible for rcu_accelerate_cbs()
> > >>> + * to return true. So complain, but don't awaken.
> > >>> + */
> > >>> + WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp));
> > >>> } else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
> > >>> /*
> > >>> * ...but NOCB kthreads may miss or delay callbacks acceleration
> > >>> @@ -2000,8 +2004,6 @@ rcu_report_qs_rdp(struct rcu_data *rdp)
> > >>> rcu_disable_urgency_upon_qs(rdp);
> > >>> rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
> > >>> /* ^^^ Released rnp->lock */
> > >>> - if (needwake)
> > >>> - rcu_gp_kthread_wake();
> > >>>
> > >>> AFAICS, there is almost no compiler benefit of doing this, and zero runtime
> > >>> benefit of doing this. The WARN_ON_ONCE() also involves a runtime condition
> > >>> check of the return value of rcu_accelerate_cbs(), so you still have a
> > >>> branch. Yes, maybe slightly smaller code without the wake call, but I'm not
> > >>> sure that is worth it.
> > >>>
> > >>> And, if the opinion of system differs, its a bug anyway, so more added risk.
> > >>>
> > >>>
> > >>>
> > >>> if (needacc) {
> > >>> rcu_nocb_lock_irqsave(rdp, flags);
> > >>>
> > >>> And when needacc = true, rcu_accelerate_cbs_unlocked() tries to do a wake up
> > >>> anyway, so it is consistent with nocb vs !nocb.
> > >>
> > >> For !nocb, we invoked rcu_accelerate_cbs() before report qs, so this GP is impossible to end
> > >> and we also not set RCU_GP_FLAG_INIT to start new GP in rcu_accelerate_cbs().
> > >> but for nocb, when needacc = true, we invoke rcu_accelerate_cbs_unlocked() after current CPU
> > >> has reported qs, if all CPU have been reported qs, we will wakeup gp kthread to end this GP in
> > >> rcu_report_qs_rnp(). after that, the rcu_accelerate_cbs_unlocked() is possible to try to wake up
> > >> gp kthread if this GP has ended at this time. so nocb vs !nocb is likely to be inconsistent.
> > >>
> > >>
> > >> That is a fair point. But after gp ends, rcu_check_quiescent_state()
> > >> -> note_gp_changes() which will do a accel + GP thread wake up at that
> > >> point anyway, once it notices that a GP has come to an end. That
> > >> should happen for both the nocb and !nocb cases right?
> > >
> > > For nocb rdp, we won't invoke rcu_accelerate_cbs() and rcu_advance_cbs() in
> > > note_gp_changes(). so also not wakeup gp kthread in note_gp_changes().
> >
> > Yes correct, ok but…
> > >
> > >>
> > >> I am wondering if rcu_report_qs_rdp() needs to be rethought to make
> > >> both cases consistent.
> > >>
> > >> Why does the nocb case need an accel + GP thread wakeup in the
> > >> rcu_report_qs_rdp() function, but the !nocb case does not?
> > >
> > > For nocb accel + GP kthread wakeup only happen in the middle of a (de-)offloading process.
> > > this is an intermediate state.
> >
> > Sure, I know what the code currently does, I am asking why and it feels wrong.
> >
> > I suggest you slightly change your approach to not assuming the code should be bonafide correct and then fixing it (which is ok once in a while), and asking higher level questions to why things are the way they are in the first place (that is just my suggestion and I am not in a place to provide advice, far from it, but I am just telling you my approach — I care more about the code than increasing my patch count :P).
> >
> > If you are in an intermediate state, part way to a !nocb state — you may have missed a nocb-related accel and wake, correct? Why does that matter? Once we transition to a !nocb state, we do not do a post-qs-report accel+wake anyway as we clearly know from the discussion. So why do we need to do it if we missed it for the intermediate stage? So, I am not fully sure yet what that needac is doing and why it is needed.
> >
> > Do not get me wrong, stellar work here. But I suggest challenge the assumptions and the design, not always just the code that was already written :), apologies for any misplaced or noisy advice.
>
> To add to Joel's point, an extra unnecessary check on a slow path can
> be OK, but missing a necessary check is of course very bad.
>
> Just to make sure that I am following along, here are the options I see:
>
> 1. Status quo.
>
> 2. Zqiang's current patch, as in remove the wakeup and
> add the WARN_ON_ONCE().
>
> 3. Status quo, and only add the WARN_ON_ONCE(), but still
> keep the needless check for the wakeup.
>
> Are there other options that I have missed?
I'm personally in favour of keeping 2.
Removing an imaginary path and consolidating an expectation from such
a complicated codebase always makes me able to sleep a few more minutes
everyday :)
Thanks.
>
> Thanx, Paul
