Le Sat, Feb 15, 2025 at 02:38:04AM -0800, Paul E. McKenney a écrit : > On Fri, Feb 14, 2025 at 01:10:52PM +0100, Frederic Weisbecker wrote: > > Le Fri, Feb 14, 2025 at 01:01:56AM -0800, Paul E. McKenney a écrit : > > > On Fri, Feb 14, 2025 at 12:25:59AM +0100, Frederic Weisbecker wrote: > > > > A CPU coming online checks for an ongoing grace period and reports > > > > a quiescent state accordingly if needed. This special treatment that > > > > shortcuts the expedited IPI finds its origin as an optimization purpose > > > > on the following commit: > > > > > > > > 338b0f760e84 (rcu: Better hotplug handling for synchronize_sched_expedited() > > > > > > > > The point is to avoid an IPI while waiting for a CPU to become online > > > > or failing to become offline. > > > > > > > > However this is pointless and even error prone for several reasons: > > > > > > > > * If the CPU has been seen offline in the first round scanning offline > > > > and idle CPUs, no IPI is even tried and the quiescent state is > > > > reported on behalf of the CPU. > > > > > > > > * This means that if the IPI fails, the CPU just became offline. So > > > > it's unlikely to become online right away, unless the cpu hotplug > > > > operation failed and rolled back, which is a rare event that can > > > > wait a jiffy for a new IPI to be issued. > > But the expedited grace period might be preempted for an arbitrarily > long period, especially if a hypervisor is in play. And we do drop > that lock midway through... Well, then that delays the expedited grace period as a whole anyway... > > > > For all those reasons, remove this optimization that doesn't look worthy > > > > to keep around. > > > > > > Thank you for digging into this! > > > > > > When I ran tests that removed the call to sync_sched_exp_online_cleanup() > > > a few months ago, I got grace-period hangs [1]. Has something changed > > > to make this safe? > > > > Hmm, but was it before or after "rcu: Fix get_state_synchronize_rcu_full() > > GP-start detection" ? > > Before. There was also some buggy debug code in play. Also, to get the > failure, it was necessary to make TREE03 disable preemption, as stock > TREE03 has an empty sync_sched_exp_online_cleanup() function. > > I am rerunning the test with a WARN_ON_ONCE() after the early exit from > the sync_sched_exp_online_cleanup(). Of course, lack of a failure does > not necessairly indicate Cool, thanks! > > > And if after do we know why? > > Here are some (possibly bogus) possibilities that came to mind: > > 1. There is some coming-online race that deprives the incoming > CPU of an IPI, but nevertheless marks that CPU as blocking the > current grace period. Arguably there is a tiny window between rcutree_report_cpu_starting() and set_cpu_online() that could make ->qsmaskinitnext visible before cpu_online() and therefore delay the IPI a bit. But I don't expect more than a jiffy to fill up the gap. And if that's relevant, note that only !PREEMPT_RCU is then "fixed" by sync_sched_exp_online_cleanup() here. > > 2. Some strange scenario involves the CPU going offline for just a > little bit, so that the IPI gets wasted on the outgoing due to > neither of the "if" conditions in rcu_exp_handler() being true. > The outgoing CPU just says "I need a QS", then leaves and > comes back. (The expedited grace period doesn't retry because > it believes that it already sent that IPI.) I don't think this is possible. Once the CPU enters CPUHP_TEARDOWN_CPU with stop_machine, no more IPIs can be issued. The remaining ones are executed at CPUHP_AP_SMPCFD_DYING, still in stop_machine. So this is the last call for rcu_exp_handler() execution. And this last call has to be followed by rcu_note_context_switch() between stop_machine and the final schedule to idle. And that rcu_note_context_switch() must report the rdp exp context switch. One easy way to assert that is: diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c index 86935fe00397..40d6090a33f5 100644 --- a/kernel/rcu/tree.c +++ b/kernel/rcu/tree.c @@ -4347,6 +4347,12 @@ void rcutree_report_cpu_dead(void) * may introduce a new READ-side while it is actually off the QS masks. */ lockdep_assert_irqs_disabled(); + /* + * CPUHP_AP_SMPCFD_DYING was the last call for rcu_exp_handler() execution. + * The requested QS must have been reported on the last context switch + * from stop machine to idle. + */ + WARN_ON_ONCE(rdp->cpu_no_qs.b.exp); // Do any dangling deferred wakeups. do_nocb_deferred_wakeup(rdp); > > 3. Your ideas here! ;-) :-) > > Thanx, Paul
