rcu: Fix grace-period-stall bug on large systems with CPU hotplug

When the last CPU of a given leaf rcu_node structure goes
offline, all of the tasks queued on that leaf rcu_node structure
(due to having blocked in their current RCU read-side critical
sections) are requeued onto the root rcu_node structure.  This
requeuing is carried out by rcu_preempt_offline_tasks().
However, it is possible that these queued tasks are the only
thing preventing the leaf rcu_node structure from reporting a
quiescent state up the rcu_node hierarchy.  Unfortunately, the
old code would fail to do this reporting, resulting in a
grace-period stall given the following sequence of events:

1.	Kernel built for more than 32 CPUs on 32-bit systems or for more
	than 64 CPUs on 64-bit systems, so that there is more than one
	rcu_node structure.  (Or CONFIG_RCU_FANOUT is artificially set
	to a number smaller than CONFIG_NR_CPUS.)

2.	The kernel is built with CONFIG_TREE_PREEMPT_RCU.

3.	A task running on a CPU associated with a given leaf rcu_node
	structure blocks while in an RCU read-side critical section
	-and- that CPU has not yet passed through a quiescent state
	for the current RCU grace period.  This will cause the task
	to be queued on the leaf rcu_node's blocked_tasks[] array, in
	particular, on the element of this array corresponding to the
	current grace period.

4.	Each of the remaining CPUs corresponding to this same leaf rcu_node
	structure pass through a quiescent state.  However, the task is
	still in its RCU read-side critical section, so these quiescent
	states cannot be reported further up the rcu_node hierarchy.
	Nevertheless, all bits in the leaf rcu_node structure's ->qsmask
	field are now zero.

5.	Each of the remaining CPUs go offline.  (The events in step
	#4 and #5 can happen in any order as long as each CPU passes
	through a quiescent state before going offline.)

6.	When the last CPU goes offline, __rcu_offline_cpu() will invoke
	rcu_preempt_offline_tasks(), which will move the task to the
	root rcu_node structure, but without reporting a quiescent state
	up the rcu_node hierarchy (and this failure to report a quiescent
	state is the bug).

	But because this leaf rcu_node structure's ->qsmask field is
	already zero and its ->block_tasks[] entries are all empty,
	force_quiescent_state() will skip this rcu_node structure.

	Therefore, grace periods are now hung.

This patch abstracts some code out of rcu_read_unlock_special(),
calling the result task_quiet() by analogy with cpu_quiet(), and
invokes task_quiet() from both rcu_read_lock_special() and
__rcu_offline_cpu().  Invoking task_quiet() from
__rcu_offline_cpu() reports the quiescent state up the rcu_node
hierarchy, fixing the bug.  This ends up requiring a separate
lock_class_key per level of the rcu_node hierarchy, which this
patch also provides.

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: laijs@cn.fujitsu.com
Cc: dipankar@in.ibm.com
Cc: mathieu.desnoyers@polymtl.ca
Cc: josh@joshtriplett.org
Cc: dvhltc@us.ibm.com
Cc: niv@us.ibm.com
Cc: peterz@infradead.org
Cc: rostedt@goodmis.org
Cc: Valdis.Kletnieks@vt.edu
Cc: dhowells@redhat.com
LKML-Reference: <12589088301770-git-send-email->
Signed-off-by: Ingo Molnar <mingo@elte.hu>

1 files changed, 21 insertions, 19 deletions

diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index 9b36d6d..b79bfcd 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -51,7 +51,7 @@
 
 /* Data structures. */
 
-static struct lock_class_key rcu_root_class;
+static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
 
 #define RCU_STATE_INITIALIZER(name) { \
 	.level = { &name.node[0] }, \
@@ -936,6 +936,7 @@ static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
 {
 	unsigned long flags;
 	unsigned long mask;
+	int need_quiet = 0;
 	struct rcu_data *rdp = rsp->rda[cpu];
 	struct rcu_node *rnp;
 
@@ -949,29 +950,30 @@ static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
 		spin_lock(&rnp->lock);		/* irqs already disabled. */
 		rnp->qsmaskinit &= ~mask;
 		if (rnp->qsmaskinit != 0) {
-			spin_unlock(&rnp->lock); /* irqs remain disabled. */
+			if (rnp != rdp->mynode)
+				spin_unlock(&rnp->lock); /* irqs remain disabled. */
 			break;
 		}
-
-		/*
-		 * If there was a task blocking the current grace period,
-		 * and if all CPUs have checked in, we need to propagate
-		 * the quiescent state up the rcu_node hierarchy.  But that
-		 * is inconvenient at the moment due to deadlock issues if
-		 * this should end the current grace period.  So set the
-		 * offlined CPU's bit in ->qsmask in order to force the
-		 * next force_quiescent_state() invocation to clean up this
-		 * mess in a deadlock-free manner.
-		 */
-		if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
-			rnp->qsmask |= mask;
-
+		if (rnp == rdp->mynode)
+			need_quiet = rcu_preempt_offline_tasks(rsp, rnp, rdp);
+		else
+			spin_unlock(&rnp->lock); /* irqs remain disabled. */
 		mask = rnp->grpmask;
-		spin_unlock(&rnp->lock);	/* irqs remain disabled. */
 		rnp = rnp->parent;
 	} while (rnp != NULL);
 
-	spin_unlock_irqrestore(&rsp->onofflock, flags);
+	/*
+	 * We still hold the leaf rcu_node structure lock here, and
+	 * irqs are still disabled.  The reason for this subterfuge is
+	 * because invoking task_quiet() with ->onofflock held leads
+	 * to deadlock.
+	 */
+	spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
+	rnp = rdp->mynode;
+	if (need_quiet)
+		task_quiet(rnp, flags);
+	else
+		spin_unlock_irqrestore(&rnp->lock, flags);
 
 	rcu_adopt_orphan_cbs(rsp);
 }
@@ -1731,6 +1733,7 @@ static void __init rcu_init_one(struct rcu_state *rsp)
 		rnp = rsp->level[i];
 		for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
 			spin_lock_init(&rnp->lock);
+			lockdep_set_class(&rnp->lock, &rcu_node_class[i]);
 			rnp->gpnum = 0;
 			rnp->qsmask = 0;
 			rnp->qsmaskinit = 0;
@@ -1753,7 +1756,6 @@ static void __init rcu_init_one(struct rcu_state *rsp)
 			INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
 		}
 	}
-	lockdep_set_class(&rcu_get_root(rsp)->lock, &rcu_root_class);
 }
 
 /*