Re: [PATCH tip/core/rcu 13/22] rcu: Fix grace-period hangs due to race with CPU offline

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

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. */
>