Messages in this thread | | | Date | Wed, 27 Jun 2018 08:53:15 -0700 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH tip/core/rcu 13/22] rcu: Fix grace-period hangs due to race with CPU offline |
| |
On Wed, Jun 27, 2018 at 11:11:06AM +0200, Peter Zijlstra wrote: > On Tue, Jun 26, 2018 at 04:40:04PM -0700, Paul E. McKenney wrote: > > The options I have considered are as follows: > > > 2. Stick with the no-failsafe approach, but rely on RCU's grace-period > > kthread to wake up later due to its timed wait during the > > force-quiescent-state process. This would be a bit obnoxious, > > as it requires passing a don't-wake flag (or some such) up the > > quiescent-state reporting mechanism. It would also needlessly > > delay grace-period ends, especially on large systems (RCU scales > > up the FQS delay on larger systems to maintain limited CPU > > consumption per unit time). > > > > 3. Stick with the no-failsafe approach, but have the quiescent-state > > reporting code hand back a value indicating that a wakeup is needed. > > Also a bit obnoxious, as this value would need to be threaded up > > the reporting code's return path. Simple in theory, but a bit > > of an ugly change, especially for the many places in the code that > > currently expect quiescent-state reporting to be an unconditional > > fire-and-forget operation. > > Here's a variant on 2+3, instead of propagating the state back, we > completely ignore if we needed a wakeup or not, and then unconditionally > wake the GP kthread on the managing CPU's rcutree_migrate_callbacks() > invocation. > > Hotplug is rare (or should damn well be), doing a spurious wake of the > GP thread shouldn't matter here.
Agreed. I decided that the extra lock acquisition was OK on similar grounds. Though that could be improved...
> The extra argument isn't really pretty but not nearly as bad as feared.
The patch is indeed quite a bit larger.
And note that the penalty for a typo in one of those rcu_report_qs_rnp() arguments is an intermittent grace-period hang that can be quite unpretty to track down...
Thanx, Paul
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c > index 7832dd556490..d4c38d8d3621 100644 > --- a/kernel/rcu/tree.c > +++ b/kernel/rcu/tree.c > @@ -104,7 +104,6 @@ struct rcu_state sname##_state = { \ > .abbr = sabbr, \ > .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \ > .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \ > - .ofl_lock = __SPIN_LOCK_UNLOCKED(sname##_state.ofl_lock), \ > } > > RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); > @@ -160,7 +159,8 @@ static int rcu_scheduler_fully_active __read_mostly; > > static void > rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, > - struct rcu_node *rnp, unsigned long gps, unsigned long flags); > + struct rcu_node *rnp, unsigned long gps, > + unsigned long flags, bool no_wakeup); > static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); > static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); > static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); > @@ -1928,13 +1928,11 @@ static bool rcu_gp_init(struct rcu_state *rsp) > */ > rsp->gp_state = RCU_GP_ONOFF; > rcu_for_each_leaf_node(rsp, rnp) { > - spin_lock(&rsp->ofl_lock); > raw_spin_lock_irq_rcu_node(rnp); > if (rnp->qsmaskinit == rnp->qsmaskinitnext && > !rnp->wait_blkd_tasks) { > /* Nothing to do on this leaf rcu_node structure. */ > raw_spin_unlock_irq_rcu_node(rnp); > - spin_unlock(&rsp->ofl_lock); > continue; > } > > @@ -1970,7 +1968,6 @@ static bool rcu_gp_init(struct rcu_state *rsp) > } > > raw_spin_unlock_irq_rcu_node(rnp); > - spin_unlock(&rsp->ofl_lock); > } > rcu_gp_slow(rsp, gp_preinit_delay); /* Races with CPU hotplug. */ > > @@ -2004,7 +2001,7 @@ static bool rcu_gp_init(struct rcu_state *rsp) > mask = rnp->qsmask & ~rnp->qsmaskinitnext; > rnp->rcu_gp_init_mask = mask; > if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp)) > - rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags); > + rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags, false); > else > raw_spin_unlock_irq_rcu_node(rnp); > cond_resched_tasks_rcu_qs(); > @@ -2247,14 +2244,17 @@ static int __noreturn rcu_gp_kthread(void *arg) > * just-completed grace period. Note that the caller must hold rnp->lock, > * which is released before return. > */ > -static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) > +static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags, > + bool no_wakeup) > __releases(rcu_get_root(rsp)->lock) > { > raw_lockdep_assert_held_rcu_node(rcu_get_root(rsp)); > WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); > WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); > raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags); > - rcu_gp_kthread_wake(rsp); > + > + if (!no_wakeup) > + rcu_gp_kthread_wake(rsp); > } > > /* > @@ -2273,7 +2273,8 @@ static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) > */ > static void > rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, > - struct rcu_node *rnp, unsigned long gps, unsigned long flags) > + struct rcu_node *rnp, unsigned long gps, > + unsigned long flags, bool no_wakeup) > __releases(rnp->lock) > { > unsigned long oldmask = 0; > @@ -2326,7 +2327,7 @@ rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, > * state for this grace period. Invoke rcu_report_qs_rsp() > * to clean up and start the next grace period if one is needed. > */ > - rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ > + rcu_report_qs_rsp(rsp, flags, no_wakeup); /* releases rnp->lock. */ > } > > /* > @@ -2361,7 +2362,7 @@ rcu_report_unblock_qs_rnp(struct rcu_state *rsp, > * Only one rcu_node structure in the tree, so don't > * try to report up to its nonexistent parent! > */ > - rcu_report_qs_rsp(rsp, flags); > + rcu_report_qs_rsp(rsp, flags, false); > return; > } > > @@ -2370,7 +2371,7 @@ rcu_report_unblock_qs_rnp(struct rcu_state *rsp, > mask = rnp->grpmask; > raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ > raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */ > - rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); > + rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags, false); > } > > /* > @@ -2413,7 +2414,7 @@ rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) > */ > needwake = rcu_accelerate_cbs(rsp, rnp, rdp); > > - rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags); > + rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags, false); > /* ^^^ Released rnp->lock */ > if (needwake) > rcu_gp_kthread_wake(rsp); > @@ -2711,7 +2712,7 @@ static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp)) > } > if (mask != 0) { > /* Idle/offline CPUs, report (releases rnp->lock). */ > - rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags); > + rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags, false); > } else { > /* Nothing to do here, so just drop the lock. */ > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > @@ -3745,7 +3746,7 @@ void rcu_cpu_starting(unsigned int cpu) > rdp->rcu_onl_gp_flags = READ_ONCE(rsp->gp_flags); > if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */ > /* Report QS -after- changing ->qsmaskinitnext! */ > - rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags); > + rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags, false); > } else { > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > } > @@ -3768,18 +3769,15 @@ static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) > > /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ > mask = rdp->grpmask; > - spin_lock(&rsp->ofl_lock); > raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */ > rdp->rcu_ofl_gp_seq = READ_ONCE(rsp->gp_seq); > rdp->rcu_ofl_gp_flags = READ_ONCE(rsp->gp_flags); > + rnp->qsmaskinitnext &= ~mask; > if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */ > - /* Report quiescent state -before- changing ->qsmaskinitnext! */ > - rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags); > + rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags, true); > raw_spin_lock_irqsave_rcu_node(rnp, flags); > } > - rnp->qsmaskinitnext &= ~mask; > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > - spin_unlock(&rsp->ofl_lock); > } > > /* > @@ -3849,6 +3847,12 @@ void rcutree_migrate_callbacks(int cpu) > { > struct rcu_state *rsp; > > + /* > + * Just in case the outgoing CPU needed to wake the GP kthread > + * do so here. > + */ > + rcu_gp_kthread_wake(rsp); > + > for_each_rcu_flavor(rsp) > rcu_migrate_callbacks(cpu, rsp); > } > diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h > index 4e74df768c57..8dab71838141 100644 > --- a/kernel/rcu/tree.h > +++ b/kernel/rcu/tree.h > @@ -367,10 +367,6 @@ struct rcu_state { > const char *name; /* Name of structure. */ > char abbr; /* Abbreviated name. */ > struct list_head flavors; /* List of RCU flavors. */ > - > - spinlock_t ofl_lock ____cacheline_internodealigned_in_smp; > - /* Synchronize offline with */ > - /* GP pre-initialization. */ > }; > > /* Values for rcu_state structure's gp_flags field. */ >
| |