Add call_rcu_sched()

Hello!

Second cut of patch to provide the call_rcu_sched() needed for Mathieu's markers implementation. This is again to synchronize_sched() as call_rcu() is to synchronize_rcu().

Should be fine for experimental use, but not ready for inclusion.

Passes short rcutorture sessions, but should be treated with some caution given that very little of it is more than 24 hours old. Fixes since the first version include a bug that could result in indefinite blocking (spotted by Gautham Shenoy), better resiliency against CPU-hotplug operations, and other minor fixes.

Known/suspected shortcomings:

o Only moderately tested -- only short rcutorture sessions.

o Need to add call_rcu_sched() testing to rcutorture.

o If I remember correctly, an rcu_barrier_sched() is required

	(Mathieu?).

o Interaction of this patch with CPU hotplug should be viewed
with great suspicion.

o If there are no synchronize_sched() calls for more than two minutes, one can see messages of the form "INFO: task rcu_sched_grace:924 blocked for more than 120 seconds." Any thoughts on how to avoid this message? Should I be using something other than __wait_event() and wake_up(), which sleep uninterruptibly, thus triggering this message?

One other thing -- this patch also fixes a long-standing bug in the earlier preemptable-RCU implementation of synchronize_rcu() that could result in loss of concurrent external changes to a task's CPU affinity mask. I have lost track of who reported this...

Signed-off-by Paul E. McKenney <[email protected]>

---

 include/linux/rcuclassic.h |    3 
 include/linux/rcupdate.h   |   22 +++
 include/linux/rcupreempt.h |   15 ++
 init/main.c                |    1 
 kernel/rcupdate.c          |   20 ---
 kernel/rcupreempt.c        |  276 +++++++++++++++++++++++++++++++++++++++++++--
 6 files changed, 308 insertions(+), 29 deletions(-)

diff -urpNa -X dontdiff linux-2.6.25-rc6/include/linux/rcuclassic.h linux-2.6.25-rc6-C1-call_rcu_sched/include/linux/rcuclassic.h

include/linux/rcuclassic.h +3 -0

@@ -153,7 +153,10 @@ extern struct lockdep_map rcu_lock_map;
 
 #define __synchronize_sched() synchronize_rcu()
 
+#define call_rcu_sched(head, func) call_rcu(head, func)
+
 extern void __rcu_init(void);
+#define rcu_init_sched()	do { } while (0)
 extern void rcu_check_callbacks(int cpu, int user);
 extern void rcu_restart_cpu(int cpu);
 

diff -urpNa -X dontdiff linux-2.6.25-rc6/include/linux/rcupdate.h linux-2.6.25-rc6-C1-call_rcu_sched/include/linux/rcupdate.h

include/linux/rcupdate.h +22 -0

@@ -42,6 +42,7 @@
 #include <linux/cpumask.h>
 #include <linux/seqlock.h>
 #include <linux/lockdep.h>
+#include <linux/completion.h>
 
 /**
  * struct rcu_head - callback structure for use with RCU

@@ -182,6 +183,27 @@ struct rcu_head {
 		(p) = (v); \
 	})
 
+/* Infrastructure to implement the synchronize_() primitives. */
+
+struct rcu_synchronize {
+	struct rcu_head head;
+	struct completion completion;
+};
+
+extern void wakeme_after_rcu(struct rcu_head  *head);
+
+#define synchronize_rcu_xxx(name, func) \
+void name(void) \
+{ \
+	struct rcu_synchronize rcu; \
+	\
+	init_completion(&rcu.completion); \
+	/* Will wake me after RCU finished. */ \
+	func(&rcu.head, wakeme_after_rcu); \
+	/* Wait for it. */ \
+	wait_for_completion(&rcu.completion); \
+}
+
 /**
  * synchronize_sched - block until all CPUs have exited any non-preemptive
  * kernel code sequences.

diff -urpNa -X dontdiff linux-2.6.25-rc6/include/linux/rcupreempt.h linux-2.6.25-rc6-C1-call_rcu_sched/include/linux/rcupreempt.h

include/linux/rcupreempt.h +15 -0

@@ -46,6 +46,20 @@
 #define rcu_bh_qsctr_inc(cpu)
 #define call_rcu_bh(head, rcu) call_rcu(head, rcu)
 
+/**
+ * call_rcu_sched - Queue RCU callback for invocation after sched grace period.
+ * @head: structure to be used for queueing the RCU updates.
+ * @func: actual update function to be invoked after the grace period
+ *
+ * The update function will be invoked some time after a full
+ * synchronize_sched()-style grace period elapses, in other words after
+ * all currently executing preempt-disabled sections of code (including
+ * hardirq handlers, NMI handlers, and local_irq_save() blocks) have
+ * completed.
+ */
+extern void call_rcu_sched(struct rcu_head *head,
+			   void (*func)(struct rcu_head *head));
+
 extern void __rcu_read_lock(void)	__acquires(RCU);
 extern void __rcu_read_unlock(void)	__releases(RCU);
 extern int rcu_pending(int cpu);

@@ -57,6 +71,7 @@ extern int rcu_needs_cpu(int cpu);
 extern void __synchronize_sched(void);
 
 extern void __rcu_init(void);
+extern void rcu_init_sched(void);
 extern void rcu_check_callbacks(int cpu, int user);
 extern void rcu_restart_cpu(int cpu);
 extern long rcu_batches_completed(void);

diff -urpNa -X dontdiff linux-2.6.25-rc6/init/main.c linux-2.6.25-rc6-C1-call_rcu_sched/init/main.c

init/main.c +1 -0

@@ -736,6 +736,7 @@ static void __init do_basic_setup(void)
 	driver_init();
 	init_irq_proc();
 	do_initcalls();
+	rcu_init_sched();
 }
 
 static int __initdata nosoftlockup;

diff -urpNa -X dontdiff linux-2.6.25-rc6/kernel/rcupdate.c linux-2.6.25-rc6-C1-call_rcu_sched/kernel/rcupdate.c

kernel/rcupdate.c +2 -18

@@ -39,18 +39,12 @@
 #include <linux/sched.h>
 #include <asm/atomic.h>
 #include <linux/bitops.h>
-#include <linux/completion.h>
 #include <linux/percpu.h>
 #include <linux/notifier.h>
 #include <linux/cpu.h>
 #include <linux/mutex.h>
 #include <linux/module.h>
 
-struct rcu_synchronize {
-	struct rcu_head head;
-	struct completion completion;
-};
-
 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
 static atomic_t rcu_barrier_cpu_count;
 static DEFINE_MUTEX(rcu_barrier_mutex);

@@ -60,7 +54,7 @@ static struct completion rcu_barrier_com
  * Awaken the corresponding synchronize_rcu() instance now that a
  * grace period has elapsed.
  */
-static void wakeme_after_rcu(struct rcu_head  *head)
+void wakeme_after_rcu(struct rcu_head  *head)
 {
 	struct rcu_synchronize *rcu;
 

@@ -77,17 +71,7 @@ static void wakeme_after_rcu(struct rcu_
  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
  * and may be nested.
  */
-void synchronize_rcu(void)
-{
-	struct rcu_synchronize rcu;
-
-	init_completion(&rcu.completion);
-	/* Will wake me after RCU finished */
-	call_rcu(&rcu.head, wakeme_after_rcu);
-
-	/* Wait for it */
-	wait_for_completion(&rcu.completion);
-}
+synchronize_rcu_xxx(synchronize_rcu, call_rcu)
 EXPORT_SYMBOL_GPL(synchronize_rcu);
 
 static void rcu_barrier_callback(struct rcu_head *notused)

diff -urpNa -X dontdiff linux-2.6.25-rc6/kernel/rcupreempt.c linux-2.6.25-rc6-C1-call_rcu_sched/kernel/rcupreempt.c

kernel/rcupreempt.c +265 -11

@@ -46,6 +46,7 @@
 #include <asm/atomic.h>
 #include <linux/bitops.h>
 #include <linux/module.h>
+#include <linux/kthread.h>
 #include <linux/completion.h>
 #include <linux/moduleparam.h>
 #include <linux/percpu.h>

@@ -87,9 +88,14 @@ struct rcu_data {
 	struct rcu_head **nexttail;
 	struct rcu_head *waitlist[GP_STAGES];
 	struct rcu_head **waittail[GP_STAGES];
-	struct rcu_head *donelist;
+	struct rcu_head *donelist;	/* from waitlist & waitschedlist */
 	struct rcu_head **donetail;
 	long rcu_flipctr[2];
+	struct rcu_head *nextschedlist;
+	struct rcu_head **nextschedtail;
+	struct rcu_head *waitschedlist;
+	struct rcu_head **waitschedtail;
+	int rcu_sched_sleeping;
 #ifdef CONFIG_RCU_TRACE
 	struct rcupreempt_trace trace;
 #endif /* #ifdef CONFIG_RCU_TRACE */

@@ -131,11 +137,24 @@ enum rcu_try_flip_states {
 	rcu_try_flip_waitmb_state,
 };
 
+/*
+ * States for rcu_ctrlblk.rcu_sched_sleep.
+ */
+
+enum rcu_sched_sleep_states {
+	rcu_sched_not_sleeping,	/* Not sleeping, callbacks need GP.  */
+	rcu_sched_sleep_prep,	/* Thinking of sleeping, rechecking. */
+	rcu_sched_sleeping,	/* Sleeping, awaken if GP needed. */
+};
+
 struct rcu_ctrlblk {
 	spinlock_t	fliplock;	/* Protect state-machine transitions. */
 	long		completed;	/* Number of last completed batch. */
 	enum rcu_try_flip_states rcu_try_flip_state; /* The current state of
 							the rcu state machine */
+	spinlock_t	schedlock;	/* Protect rcu_sched sleep state. */
+	enum rcu_sched_sleep_states sched_sleep; /* rcu_sched state. */
+	wait_queue_head_t sched_wq;	/* Place for rcu_sched to sleep. */
 };
 
 static DEFINE_PER_CPU(struct rcu_data, rcu_data);

@@ -143,8 +162,12 @@ static struct rcu_ctrlblk rcu_ctrlblk = 
 	.fliplock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.fliplock),
 	.completed = 0,
 	.rcu_try_flip_state = rcu_try_flip_idle_state,
+	.schedlock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.schedlock),
+	.sched_sleep = rcu_sched_not_sleeping,
+	.sched_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rcu_ctrlblk.sched_wq),
 };
 
+static struct task_struct *rcu_sched_grace_period_task;
 
 #ifdef CONFIG_RCU_TRACE
 static char *rcu_try_flip_state_names[] =

@@ -871,6 +894,8 @@ void rcu_offline_cpu(int cpu)
 	struct rcu_head *list = NULL;
 	unsigned long flags;
 	struct rcu_data *rdp = RCU_DATA_CPU(cpu);
+	struct rcu_head *schedlist = NULL;
+	struct rcu_head **schedtail = &schedlist;
 	struct rcu_head **tail = &list;
 
 	/*

@@ -884,6 +909,11 @@ void rcu_offline_cpu(int cpu)
 		rcu_offline_cpu_enqueue(rdp->waitlist[i], rdp->waittail[i],
 						list, tail);
 	rcu_offline_cpu_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
+	rcu_offline_cpu_enqueue(rdp->waitschedlist, rdp->waitschedtail,
+				schedlist, schedtail);
+	rcu_offline_cpu_enqueue(rdp->nextschedlist, rdp->nextschedtail,
+				schedlist, schedtail);
+	rdp->rcu_sched_sleeping = 0;
 	spin_unlock_irqrestore(&rdp->lock, flags);
 	rdp->waitlistcount = 0;
 

@@ -924,16 +954,35 @@ void rcu_offline_cpu(int cpu)
 	*rdp->nexttail = list;
 	if (list)
 		rdp->nexttail = tail;
+	*rdp->nextschedtail = schedlist;
+	if (schedlist)
+		rdp->nextschedtail = schedtail;
 	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 void __devinit rcu_online_cpu(int cpu)
 {
 	unsigned long flags;
+	struct rcu_data *rdp;
 
 	spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
 	cpu_set(cpu, rcu_cpu_online_map);
 	spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
+
+	/*
+	 * The rcu_sched grace-period processing might have bypassed
+	 * this CPU, given that it was not in the rcu_cpu_online_map
+	 * when the grace-period scan started.  This means that the
+	 * grace-period task might sleep.  So make sure that if this
+	 * should happen, the first callback posted to this CPU will
+	 * wake up the grace-period task if need be.
+	 */
+
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	rdp->rcu_sched_sleeping = 1;
+	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */

@@ -993,26 +1042,214 @@ void call_rcu(struct rcu_head *head, voi
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+{
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int wake_gp = 0;
+
+	head->func = func;
+	head->next = NULL;
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	*rdp->nextschedtail = head;
+	rdp->nextschedtail = &head->next;
+	if (rdp->rcu_sched_sleeping) {
+
+		/* Grace-period processing might be sleeping... */
+
+		rdp->rcu_sched_sleeping = 0;
+		wake_gp = 1;
+	}
+	spin_unlock(&rdp->lock);
+	local_irq_restore(flags);
+	if (wake_gp) {
+
+		/* Wake up grace-period processing, unless someone beat us. */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
+			wake_gp = 0;
+		rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		if (wake_gp)
+			wake_up(&rcu_ctrlblk.sched_wq);
+	}
+}
+EXPORT_SYMBOL_GPL(call_rcu_sched);
+
 /*
  * Wait until all currently running preempt_disable() code segments
  * (including hardware-irq-disable segments) complete.  Note that
  * in -rt this does -not- necessarily result in all currently executing
  * interrupt -handlers- having completed.
  */
-void __synchronize_sched(void)
+synchronize_rcu_xxx(__synchronize_sched, call_rcu_sched)
+EXPORT_SYMBOL_GPL(__synchronize_sched);
+
+/*
+ * kthread function that manages call_rcu_sched grace periods.
+ */
+static int
+rcu_sched_grace_period(void *arg)
 {
-	cpumask_t oldmask;
+	int couldsleep;		/* might sleep after current pass. */
+	int couldsleepnext = 0; /* might sleep after next pass. */
 	int cpu;
+	long err;
+	unsigned long flags;
+	int needsoftirq;
+	struct rcu_data *rdp;
 
-	if (sched_getaffinity(0, &oldmask) < 0)
-		oldmask = cpu_possible_map;
-	for_each_online_cpu(cpu) {
-		sched_setaffinity(0, cpumask_of_cpu(cpu));
-		schedule();
-	}
-	sched_setaffinity(0, oldmask);
+	/*
+	 * Each pass through the following loop handles one
+	 * rcu_sched grace period cycle.
+	 */
+
+	do {
+
+		/*
+		 * Sleep for about an RCU grace-period's worth to
+		 * allow better batching and to consume less CPU.
+		 */
+
+		schedule_timeout_interruptible(HZ / 20);
+
+		/*
+		 * If there was nothing to do last time, prepare to
+		 * sleep at the end of the current grace period cycle.
+		 */
+
+		couldsleep = couldsleepnext;
+		couldsleepnext = 1;
+		if (couldsleep) {
+			spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+			rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		}
+
+		/*
+		 * Schedule on each CPU in turn, advancing callbacks
+		 * as we go.  We will have visited each CPU between
+		 * the time we move a callback from the nextsched
+		 * list and the time we move that callback to the
+		 * done list.  Now, a given CPU might come online
+		 * during that interval, but that means that it
+		 * was offline when we started, so we can safely
+		 * ignore it.
+		 */
+
+		for_each_online_cpu(cpu) {
+
+retry:
+
+			/* Initialize and schedule onto current CPU. */
+
+			needsoftirq = 0;
+			err = sched_setaffinity(0, cpumask_of_cpu(cpu));
+			if (err < 0) {
+				printk(KERN_WARNING "sched_setaffinity(%d) error: %ld, cpu_is_offline: %ld\n", cpu, err, cpu_is_offline(cpu));
+				schedule_timeout_interruptible(HZ);
+				continue;
+			}
+
+			/*
+			 * Get a reference to this CPU's rcu_data
+			 * structure, lock it, and verify that this
+			 * CPU is still online (skip it otherwise).
+			 */
+
+			rdp = RCU_DATA_CPU(cpu);
+			spin_lock_irqsave(&rdp->lock, flags);
+			if (cpu_is_offline(cpu)) {
+				spin_unlock_irqrestore(&rdp->lock, flags);
+				continue;
+			}
+
+			/*
+			 * If we didn't end up on the CPU we expected
+			 * to, try again.  This can happen if a CPU
+			 * goes offline before we attempt to schedule
+			 * on it, but comes back online before we get
+			 * to this check.
+			 */
+
+			if (smp_processor_id() != cpu) {
+				spin_unlock_irqrestore(&rdp->lock, flags);
+				goto retry;
+			}
+
+			/*
+			 * We are running on the CPU irq-disabled, so it
+			 * cannot go offline until we re-enable irqs.
+			 *
+			 * Advance the callbacks!  We share normal RCU's
+			 * donelist, since callbacks are invoked the
+			 * same way in either case.
+			 */
+
+			if (rdp->waitschedlist != NULL) {
+				*rdp->donetail = rdp->waitschedlist;
+				rdp->donetail = rdp->waitschedtail;
+				needsoftirq = 1;
+			}
+			if (rdp->nextschedlist != NULL) {
+				rdp->waitschedlist = rdp->nextschedlist;
+				rdp->waitschedtail = rdp->nextschedtail;
+				couldsleep = 0;
+				couldsleepnext = 0;
+			} else {
+				rdp->waitschedlist = NULL;
+				rdp->waitschedtail = &rdp->waitschedlist;
+			}
+			rdp->nextschedlist = NULL;
+			rdp->nextschedtail = &rdp->nextschedlist;
+
+			/* Mark sleep intention. */
+
+			rdp->rcu_sched_sleeping = couldsleep;
+
+			spin_unlock_irqrestore(&rdp->lock, flags);
+
+			/* If we added callbacks to donelist, process. */
+
+			if (needsoftirq)
+				raise_softirq(RCU_SOFTIRQ);
+		}
+
+		/* If we saw callbacks on the last scan, go deal with them. */
+
+		if (!couldsleep)
+			continue;
+
+		/* Attempt to block... */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
+
+			/*
+			 * Someone posted a callback after we scanned.
+			 * Go take care of it.
+			 */
+
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+			couldsleepnext = 0;
+			continue;
+		}
+
+		/* Block until the next person posts a callback. */
+
+		rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		__wait_event(rcu_ctrlblk.sched_wq,
+			     rcu_ctrlblk.sched_sleep != rcu_sched_sleeping);
+		couldsleepnext = 0;
+
+	} while (!kthread_should_stop());
+
+	return (0);
 }
-EXPORT_SYMBOL_GPL(__synchronize_sched);
 
 /*
  * Check to see if any future RCU-related work will need to be done

@@ -1107,6 +1344,11 @@ void __init __rcu_init(void)
 		rdp->donetail = &rdp->donelist;
 		rdp->rcu_flipctr[0] = 0;
 		rdp->rcu_flipctr[1] = 0;
+		rdp->nextschedlist = NULL;
+		rdp->nextschedtail = &rdp->nextschedlist;
+		rdp->waitschedlist = NULL;
+		rdp->waitschedtail = &rdp->waitschedlist;
+		rdp->rcu_sched_sleeping = 0;
 	}
 	register_cpu_notifier(&rcu_nb);
 

@@ -1129,6 +1371,18 @@ void __init __rcu_init(void)
 }
 
 /*
+ * Late-boot-time RCU initialization that must wait until after scheduler
+ * has been initialized.
+ */
+void __init rcu_init_sched(void)
+{
+	rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
+						  NULL,
+						  "rcu_sched_grace_period");
+	WARN_ON(IS_ERR(rcu_sched_grace_period_task));
+}
+
+/*
  * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
  */
 void synchronize_kernel(void)

Quoted 14 lines · * Paul E. McKenney ([email protected]) wrote:

Hello!

Second cut of patch to provide the call_rcu_sched() needed for Mathieu's markers implementation. This is again to synchronize_sched() as call_rcu() is to synchronize_rcu().

Should be fine for experimental use, but not ready for inclusion.

Passes short rcutorture sessions, but should be treated with some caution given that very little of it is more than 24 hours old. Fixes since the first version include a bug that could result in indefinite blocking (spotted by Gautham Shenoy), better resiliency against CPU-hotplug operations, and other minor fixes.

Quoted 9 lines

Known/suspected shortcomings:

o Only moderately tested -- only short rcutorture sessions.

o Need to add call_rcu_sched() testing to rcutorture.

o If I remember correctly, an rcu_barrier_sched() is required

	(Mathieu?).

Hi Paul,

Thanks for this work, I'll give it a try (I'm just back from a weekend away from the city). Yes, my code needs a rcu_barrier_sched() so it can wait for call_rcu_sched completion before it tries to re-use the data structures at the next modification of the same marker.

I think rcu_barrier_sched should be quite straightforward to implement if we derive it from kernel/rcupdate.c:rcu_barrier. Actually, couldn't we just rename rcu_barrier into something else made static (_rcu_barrier) and call it with a different parameter telling which of call_rcu or call_rcu_sched to use ?

Something like this :

Add rcu_barrier_sched

Adds rcu_barrier_sched, which uses call_rcu_sched. It wait for each in flight call_rcu_sched to be completed before it returns.

Signed-off-by Mathieu Desnoyers <[email protected]>

---
 include/linux/rcupdate.h |    1 +
 kernel/rcupdate.c        |   40 +++++++++++++++++++++++++++++++++-------
 2 files changed, 34 insertions(+), 7 deletions(-)

Index: linux-2.6-lttng/include/linux/rcupdate.h
===================================================================

include/linux/rcupdate.h +1 -0

@@ -260,6 +260,7 @@ extern void call_rcu_bh(struct rcu_head 
 /* Exported common interfaces */
 extern void synchronize_rcu(void);
 extern void rcu_barrier(void);
+extern void rcu_barrier_sched(void);
 extern long rcu_batches_completed(void);
 extern long rcu_batches_completed_bh(void);
 

Index: linux-2.6-lttng/kernel/rcupdate.c
===================================================================

kernel/rcupdate.c +33 -7

@@ -45,6 +45,11 @@
 #include <linux/mutex.h>
 #include <linux/module.h>
 
+enum rcu_barrier {
+	RCU_BARRIER_STD,
+	RCU_BARRIER_SCHED,
+};
+
 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
 static atomic_t rcu_barrier_cpu_count;
 static DEFINE_MUTEX(rcu_barrier_mutex);

@@ -83,19 +88,23 @@ static void rcu_barrier_callback(struct 
 /*
  * Called with preemption disabled, and from cross-cpu IRQ context.
  */
-static void rcu_barrier_func(void *notused)
+static void rcu_barrier_func(void *type)
 {
 	int cpu = smp_processor_id();
 	struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
 
 	atomic_inc(&rcu_barrier_cpu_count);
-	call_rcu(head, rcu_barrier_callback);
+	switch((enum rcu_barrier)type) {
+	case RCU_BARRIER_STD:
+		call_rcu(head, rcu_barrier_callback);
+		break;
+	case RCU_BARRIER_SCHED:
+		call_rcu_sched(head, rcu_barrier_callback);
+		break;
+	}
 }
 
-/**
- * rcu_barrier - Wait until all the in-flight RCUs are complete.
- */
-void rcu_barrier(void)
+static void _rcu_barrier(enum rcu_barrier type)
 {
 	BUG_ON(in_interrupt());
 	/* Take cpucontrol mutex to protect against CPU hotplug */

@@ -111,13 +120,30 @@ void rcu_barrier(void)
 	 * until all the callbacks are queued.
 	 */
 	rcu_read_lock();
-	on_each_cpu(rcu_barrier_func, NULL, 0, 1);
+	on_each_cpu(rcu_barrier_func, (void *)type, 0, 1);
 	rcu_read_unlock();
 	wait_for_completion(&rcu_barrier_completion);
 	mutex_unlock(&rcu_barrier_mutex);
 }
+
+/**
+ * rcu_barrier - Wait until all the in-flight RCUs are complete.
+ */
+void rcu_barrier(void)
+{
+	_rcu_barrier(RCU_BARRIER_STD);
+}
 EXPORT_SYMBOL_GPL(rcu_barrier);
 
+/**
+ * rcu_barrier_sched - Wait until all the in-flight call_rcu_sched are complete.
+ */
+void rcu_barrier_sched(void)
+{
+	_rcu_barrier(RCU_BARRIER_SCHED);
+}
+EXPORT_SYMBOL_GPL(rcu_barrier_sched);
+
 void __init rcu_init(void)
 {
 	__rcu_init();

Quoted 2 lines

o Interaction of this patch with CPU hotplug should be viewed
with great suspicion.

Fix call_rcu_sched wait

Quoted 7 lines

o If there are no synchronize_sched() calls for more than two minutes, one can see messages of the form "INFO: task rcu_sched_grace:924 blocked for more than 120 seconds." Any thoughts on how to avoid this message? Should I be using something other than __wait_event() and wake_up(), which sleep uninterruptibly, thus triggering this message?

Could you use __wait_event_interruptible and wake_up_interruptible instead ? softlockup.c only seems to complain when uninterruptible tasks are not scheduled for 2 minutes. I guess that when we receive a signal we could simply go through another loop.

Signed-off-by Mathieu Desnoyers <[email protected]>

---
 kernel/rcupreempt.c |    9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

Index: linux-2.6-lttng/kernel/rcupreempt.c
===================================================================

kernel/rcupreempt.c +2 -1

@@ -1074,7 +1074,7 @@ void call_rcu_sched(struct rcu_head *hea
 		rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
 		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
 		if (wake_gp)
-			wake_up(&rcu_ctrlblk.sched_wq);
+			wake_up_interruptible(&rcu_ctrlblk.sched_wq);
 	}
 }
 EXPORT_SYMBOL_GPL(call_rcu_sched);

@@ -1097,6 +1097,7 @@ rcu_sched_grace_period(void *arg)
 	int couldsleep;		/* might sleep after current pass. */
 	int couldsleepnext = 0; /* might sleep after next pass. */
 	int cpu;
+	int ret;
 	long err;
 	unsigned long flags;
 	int needsoftirq;

kernel/rcupreempt.c +0 -0

 		couldsleepnext = 0;
 
 	} while (!kthread_should_stop());

Quoted 5 lines

One other thing -- this patch also fixes a long-standing bug in the earlier preemptable-RCU implementation of synchronize_rcu() that could result in loss of concurrent external changes to a task's CPU affinity mask. I have lost track of who reported this...

That's always good :)

Mathieu

Quoted 535 lines

Signed-off-by Paul E. McKenney <[email protected]>

---

 include/linux/rcuclassic.h |    3 
 include/linux/rcupdate.h   |   22 +++
 include/linux/rcupreempt.h |   15 ++
 init/main.c                |    1 
 kernel/rcupdate.c          |   20 ---
 kernel/rcupreempt.c        |  276 +++++++++++++++++++++++++++++++++++++++++++--
 6 files changed, 308 insertions(+), 29 deletions(-)

diff -urpNa -X dontdiff linux-2.6.25-rc6/include/linux/rcuclassic.h linux-2.6.25-rc6-C1-call_rcu_sched/include/linux/rcuclassic.h

include/linux/rcuclassic.h +3 -0

@@ -153,7 +153,10 @@ extern struct lockdep_map rcu_lock_map;
 
 #define __synchronize_sched() synchronize_rcu()
 
+#define call_rcu_sched(head, func) call_rcu(head, func)
+
 extern void __rcu_init(void);
+#define rcu_init_sched()	do { } while (0)
 extern void rcu_check_callbacks(int cpu, int user);
 extern void rcu_restart_cpu(int cpu);
 

diff -urpNa -X dontdiff linux-2.6.25-rc6/include/linux/rcupdate.h linux-2.6.25-rc6-C1-call_rcu_sched/include/linux/rcupdate.h

include/linux/rcupdate.h +22 -0

@@ -42,6 +42,7 @@
 #include <linux/cpumask.h>
 #include <linux/seqlock.h>
 #include <linux/lockdep.h>
+#include <linux/completion.h>
 
 /**
  * struct rcu_head - callback structure for use with RCU

@@ -182,6 +183,27 @@ struct rcu_head {
 		(p) = (v); \
 	})
 
+/* Infrastructure to implement the synchronize_() primitives. */
+
+struct rcu_synchronize {
+	struct rcu_head head;
+	struct completion completion;
+};
+
+extern void wakeme_after_rcu(struct rcu_head  *head);
+
+#define synchronize_rcu_xxx(name, func) \
+void name(void) \
+{ \
+	struct rcu_synchronize rcu; \
+	\
+	init_completion(&rcu.completion); \
+	/* Will wake me after RCU finished. */ \
+	func(&rcu.head, wakeme_after_rcu); \
+	/* Wait for it. */ \
+	wait_for_completion(&rcu.completion); \
+}
+
 /**
  * synchronize_sched - block until all CPUs have exited any non-preemptive
  * kernel code sequences.

diff -urpNa -X dontdiff linux-2.6.25-rc6/include/linux/rcupreempt.h linux-2.6.25-rc6-C1-call_rcu_sched/include/linux/rcupreempt.h

include/linux/rcupreempt.h +15 -0

@@ -46,6 +46,20 @@
 #define rcu_bh_qsctr_inc(cpu)
 #define call_rcu_bh(head, rcu) call_rcu(head, rcu)
 
+/**
+ * call_rcu_sched - Queue RCU callback for invocation after sched grace period.
+ * @head: structure to be used for queueing the RCU updates.
+ * @func: actual update function to be invoked after the grace period
+ *
+ * The update function will be invoked some time after a full
+ * synchronize_sched()-style grace period elapses, in other words after
+ * all currently executing preempt-disabled sections of code (including
+ * hardirq handlers, NMI handlers, and local_irq_save() blocks) have
+ * completed.
+ */
+extern void call_rcu_sched(struct rcu_head *head,
+			   void (*func)(struct rcu_head *head));
+
 extern void __rcu_read_lock(void)	__acquires(RCU);
 extern void __rcu_read_unlock(void)	__releases(RCU);
 extern int rcu_pending(int cpu);

@@ -57,6 +71,7 @@ extern int rcu_needs_cpu(int cpu);
 extern void __synchronize_sched(void);
 
 extern void __rcu_init(void);
+extern void rcu_init_sched(void);
 extern void rcu_check_callbacks(int cpu, int user);
 extern void rcu_restart_cpu(int cpu);
 extern long rcu_batches_completed(void);

diff -urpNa -X dontdiff linux-2.6.25-rc6/init/main.c linux-2.6.25-rc6-C1-call_rcu_sched/init/main.c

init/main.c +1 -0

@@ -736,6 +736,7 @@ static void __init do_basic_setup(void)
 	driver_init();
 	init_irq_proc();
 	do_initcalls();
+	rcu_init_sched();
 }
 
 static int __initdata nosoftlockup;

diff -urpNa -X dontdiff linux-2.6.25-rc6/kernel/rcupdate.c linux-2.6.25-rc6-C1-call_rcu_sched/kernel/rcupdate.c

kernel/rcupdate.c +2 -18

@@ -39,18 +39,12 @@
 #include <linux/sched.h>
 #include <asm/atomic.h>
 #include <linux/bitops.h>
-#include <linux/completion.h>
 #include <linux/percpu.h>
 #include <linux/notifier.h>
 #include <linux/cpu.h>
 #include <linux/mutex.h>
 #include <linux/module.h>
 
-struct rcu_synchronize {
-	struct rcu_head head;
-	struct completion completion;
-};
-
 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
 static atomic_t rcu_barrier_cpu_count;
 static DEFINE_MUTEX(rcu_barrier_mutex);

@@ -60,7 +54,7 @@ static struct completion rcu_barrier_com
  * Awaken the corresponding synchronize_rcu() instance now that a
  * grace period has elapsed.
  */
-static void wakeme_after_rcu(struct rcu_head  *head)
+void wakeme_after_rcu(struct rcu_head  *head)
 {
 	struct rcu_synchronize *rcu;
 

@@ -77,17 +71,7 @@ static void wakeme_after_rcu(struct rcu_
  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
  * and may be nested.
  */
-void synchronize_rcu(void)
-{
-	struct rcu_synchronize rcu;
-
-	init_completion(&rcu.completion);
-	/* Will wake me after RCU finished */
-	call_rcu(&rcu.head, wakeme_after_rcu);
-
-	/* Wait for it */
-	wait_for_completion(&rcu.completion);
-}
+synchronize_rcu_xxx(synchronize_rcu, call_rcu)
 EXPORT_SYMBOL_GPL(synchronize_rcu);
 
 static void rcu_barrier_callback(struct rcu_head *notused)

diff -urpNa -X dontdiff linux-2.6.25-rc6/kernel/rcupreempt.c linux-2.6.25-rc6-C1-call_rcu_sched/kernel/rcupreempt.c

kernel/rcupreempt.c +265 -11

@@ -46,6 +46,7 @@
 #include <asm/atomic.h>
 #include <linux/bitops.h>
 #include <linux/module.h>
+#include <linux/kthread.h>
 #include <linux/completion.h>
 #include <linux/moduleparam.h>
 #include <linux/percpu.h>

@@ -87,9 +88,14 @@ struct rcu_data {
 	struct rcu_head **nexttail;
 	struct rcu_head *waitlist[GP_STAGES];
 	struct rcu_head **waittail[GP_STAGES];
-	struct rcu_head *donelist;
+	struct rcu_head *donelist;	/* from waitlist & waitschedlist */
 	struct rcu_head **donetail;
 	long rcu_flipctr[2];
+	struct rcu_head *nextschedlist;
+	struct rcu_head **nextschedtail;
+	struct rcu_head *waitschedlist;
+	struct rcu_head **waitschedtail;
+	int rcu_sched_sleeping;
 #ifdef CONFIG_RCU_TRACE
 	struct rcupreempt_trace trace;
 #endif /* #ifdef CONFIG_RCU_TRACE */

@@ -131,11 +137,24 @@ enum rcu_try_flip_states {
 	rcu_try_flip_waitmb_state,
 };
 
+/*
+ * States for rcu_ctrlblk.rcu_sched_sleep.
+ */
+
+enum rcu_sched_sleep_states {
+	rcu_sched_not_sleeping,	/* Not sleeping, callbacks need GP.  */
+	rcu_sched_sleep_prep,	/* Thinking of sleeping, rechecking. */
+	rcu_sched_sleeping,	/* Sleeping, awaken if GP needed. */
+};
+
 struct rcu_ctrlblk {
 	spinlock_t	fliplock;	/* Protect state-machine transitions. */
 	long		completed;	/* Number of last completed batch. */
 	enum rcu_try_flip_states rcu_try_flip_state; /* The current state of
 							the rcu state machine */
+	spinlock_t	schedlock;	/* Protect rcu_sched sleep state. */
+	enum rcu_sched_sleep_states sched_sleep; /* rcu_sched state. */
+	wait_queue_head_t sched_wq;	/* Place for rcu_sched to sleep. */
 };
 
 static DEFINE_PER_CPU(struct rcu_data, rcu_data);

@@ -143,8 +162,12 @@ static struct rcu_ctrlblk rcu_ctrlblk = 
 	.fliplock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.fliplock),
 	.completed = 0,
 	.rcu_try_flip_state = rcu_try_flip_idle_state,
+	.schedlock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.schedlock),
+	.sched_sleep = rcu_sched_not_sleeping,
+	.sched_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rcu_ctrlblk.sched_wq),
 };
 
+static struct task_struct *rcu_sched_grace_period_task;
 
 #ifdef CONFIG_RCU_TRACE
 static char *rcu_try_flip_state_names[] =

@@ -871,6 +894,8 @@ void rcu_offline_cpu(int cpu)
 	struct rcu_head *list = NULL;
 	unsigned long flags;
 	struct rcu_data *rdp = RCU_DATA_CPU(cpu);
+	struct rcu_head *schedlist = NULL;
+	struct rcu_head **schedtail = &schedlist;
 	struct rcu_head **tail = &list;
 
 	/*

@@ -884,6 +909,11 @@ void rcu_offline_cpu(int cpu)
 		rcu_offline_cpu_enqueue(rdp->waitlist[i], rdp->waittail[i],
 						list, tail);
 	rcu_offline_cpu_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
+	rcu_offline_cpu_enqueue(rdp->waitschedlist, rdp->waitschedtail,
+				schedlist, schedtail);
+	rcu_offline_cpu_enqueue(rdp->nextschedlist, rdp->nextschedtail,
+				schedlist, schedtail);
+	rdp->rcu_sched_sleeping = 0;
 	spin_unlock_irqrestore(&rdp->lock, flags);
 	rdp->waitlistcount = 0;
 

@@ -924,16 +954,35 @@ void rcu_offline_cpu(int cpu)
 	*rdp->nexttail = list;
 	if (list)
 		rdp->nexttail = tail;
+	*rdp->nextschedtail = schedlist;
+	if (schedlist)
+		rdp->nextschedtail = schedtail;
 	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 void __devinit rcu_online_cpu(int cpu)
 {
 	unsigned long flags;
+	struct rcu_data *rdp;
 
 	spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
 	cpu_set(cpu, rcu_cpu_online_map);
 	spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
+
+	/*
+	 * The rcu_sched grace-period processing might have bypassed
+	 * this CPU, given that it was not in the rcu_cpu_online_map
+	 * when the grace-period scan started.  This means that the
+	 * grace-period task might sleep.  So make sure that if this
+	 * should happen, the first callback posted to this CPU will
+	 * wake up the grace-period task if need be.
+	 */
+
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	rdp->rcu_sched_sleeping = 1;
+	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */

@@ -993,26 +1042,214 @@ void call_rcu(struct rcu_head *head, voi
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+{
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int wake_gp = 0;
+
+	head->func = func;
+	head->next = NULL;
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	*rdp->nextschedtail = head;
+	rdp->nextschedtail = &head->next;
+	if (rdp->rcu_sched_sleeping) {
+
+		/* Grace-period processing might be sleeping... */
+
+		rdp->rcu_sched_sleeping = 0;
+		wake_gp = 1;
+	}
+	spin_unlock(&rdp->lock);
+	local_irq_restore(flags);
+	if (wake_gp) {
+
+		/* Wake up grace-period processing, unless someone beat us. */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
+			wake_gp = 0;
+		rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		if (wake_gp)
+			wake_up(&rcu_ctrlblk.sched_wq);
+	}
+}
+EXPORT_SYMBOL_GPL(call_rcu_sched);
+
 /*
  * Wait until all currently running preempt_disable() code segments
  * (including hardware-irq-disable segments) complete.  Note that
  * in -rt this does -not- necessarily result in all currently executing
  * interrupt -handlers- having completed.
  */
-void __synchronize_sched(void)
+synchronize_rcu_xxx(__synchronize_sched, call_rcu_sched)
+EXPORT_SYMBOL_GPL(__synchronize_sched);
+
+/*
+ * kthread function that manages call_rcu_sched grace periods.
+ */
+static int
+rcu_sched_grace_period(void *arg)
 {
-	cpumask_t oldmask;
+	int couldsleep;		/* might sleep after current pass. */
+	int couldsleepnext = 0; /* might sleep after next pass. */
 	int cpu;
+	long err;
+	unsigned long flags;
+	int needsoftirq;
+	struct rcu_data *rdp;
 
-	if (sched_getaffinity(0, &oldmask) < 0)
-		oldmask = cpu_possible_map;
-	for_each_online_cpu(cpu) {
-		sched_setaffinity(0, cpumask_of_cpu(cpu));
-		schedule();
-	}
-	sched_setaffinity(0, oldmask);
+	/*
+	 * Each pass through the following loop handles one
+	 * rcu_sched grace period cycle.
+	 */
+
+	do {
+
+		/*
+		 * Sleep for about an RCU grace-period's worth to
+		 * allow better batching and to consume less CPU.
+		 */
+
+		schedule_timeout_interruptible(HZ / 20);
+
+		/*
+		 * If there was nothing to do last time, prepare to
+		 * sleep at the end of the current grace period cycle.
+		 */
+
+		couldsleep = couldsleepnext;
+		couldsleepnext = 1;
+		if (couldsleep) {
+			spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+			rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		}
+
+		/*
+		 * Schedule on each CPU in turn, advancing callbacks
+		 * as we go.  We will have visited each CPU between
+		 * the time we move a callback from the nextsched
+		 * list and the time we move that callback to the
+		 * done list.  Now, a given CPU might come online
+		 * during that interval, but that means that it
+		 * was offline when we started, so we can safely
+		 * ignore it.
+		 */
+
+		for_each_online_cpu(cpu) {
+
+retry:
+
+			/* Initialize and schedule onto current CPU. */
+
+			needsoftirq = 0;
+			err = sched_setaffinity(0, cpumask_of_cpu(cpu));
+			if (err < 0) {
+				printk(KERN_WARNING "sched_setaffinity(%d) error: %ld, cpu_is_offline: %ld\n", cpu, err, cpu_is_offline(cpu));
+				schedule_timeout_interruptible(HZ);
+				continue;
+			}
+
+			/*
+			 * Get a reference to this CPU's rcu_data
+			 * structure, lock it, and verify that this
+			 * CPU is still online (skip it otherwise).
+			 */
+
+			rdp = RCU_DATA_CPU(cpu);
+			spin_lock_irqsave(&rdp->lock, flags);
+			if (cpu_is_offline(cpu)) {
+				spin_unlock_irqrestore(&rdp->lock, flags);
+				continue;
+			}
+
+			/*
+			 * If we didn't end up on the CPU we expected
+			 * to, try again.  This can happen if a CPU
+			 * goes offline before we attempt to schedule
+			 * on it, but comes back online before we get
+			 * to this check.
+			 */
+
+			if (smp_processor_id() != cpu) {
+				spin_unlock_irqrestore(&rdp->lock, flags);
+				goto retry;
+			}
+
+			/*
+			 * We are running on the CPU irq-disabled, so it
+			 * cannot go offline until we re-enable irqs.
+			 *
+			 * Advance the callbacks!  We share normal RCU's
+			 * donelist, since callbacks are invoked the
+			 * same way in either case.
+			 */
+
+			if (rdp->waitschedlist != NULL) {
+				*rdp->donetail = rdp->waitschedlist;
+				rdp->donetail = rdp->waitschedtail;
+				needsoftirq = 1;
+			}
+			if (rdp->nextschedlist != NULL) {
+				rdp->waitschedlist = rdp->nextschedlist;
+				rdp->waitschedtail = rdp->nextschedtail;
+				couldsleep = 0;
+				couldsleepnext = 0;
+			} else {
+				rdp->waitschedlist = NULL;
+				rdp->waitschedtail = &rdp->waitschedlist;
+			}
+			rdp->nextschedlist = NULL;
+			rdp->nextschedtail = &rdp->nextschedlist;
+
+			/* Mark sleep intention. */
+
+			rdp->rcu_sched_sleeping = couldsleep;
+
+			spin_unlock_irqrestore(&rdp->lock, flags);
+
+			/* If we added callbacks to donelist, process. */
+
+			if (needsoftirq)
+				raise_softirq(RCU_SOFTIRQ);
+		}
+
+		/* If we saw callbacks on the last scan, go deal with them. */
+
+		if (!couldsleep)
+			continue;
+
+		/* Attempt to block... */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
+
+			/*
+			 * Someone posted a callback after we scanned.
+			 * Go take care of it.
+			 */
+
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+			couldsleepnext = 0;
+			continue;
+		}
+
+		/* Block until the next person posts a callback. */
+
+		rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		__wait_event(rcu_ctrlblk.sched_wq,
+			     rcu_ctrlblk.sched_sleep != rcu_sched_sleeping);
+		couldsleepnext = 0;
+
+	} while (!kthread_should_stop());
+
+	return (0);
 }
-EXPORT_SYMBOL_GPL(__synchronize_sched);
 
 /*
  * Check to see if any future RCU-related work will need to be done

@@ -1107,6 +1344,11 @@ void __init __rcu_init(void)
 		rdp->donetail = &rdp->donelist;
 		rdp->rcu_flipctr[0] = 0;
 		rdp->rcu_flipctr[1] = 0;
+		rdp->nextschedlist = NULL;
+		rdp->nextschedtail = &rdp->nextschedlist;
+		rdp->waitschedlist = NULL;
+		rdp->waitschedtail = &rdp->waitschedlist;
+		rdp->rcu_sched_sleeping = 0;
 	}
 	register_cpu_notifier(&rcu_nb);
 

@@ -1129,6 +1371,18 @@ void __init __rcu_init(void)
 }
 
 /*
+ * Late-boot-time RCU initialization that must wait until after scheduler
+ * has been initialized.
+ */
+void __init rcu_init_sched(void)
+{
+	rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
+						  NULL,
+						  "rcu_sched_grace_period");
+	WARN_ON(IS_ERR(rcu_sched_grace_period_task));
+}
+
+/*
  * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
  */
 void synchronize_kernel(void)

Signature · 3 lines

Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68

Quoted 1 line · Mathieu Desnoyers

* Paul E. McKenney ([email protected]) wrote:

Quoted 4 lines

Quoted 2 lines

o Interaction of this patch with CPU hotplug should be viewed
with great suspicion.

Fix call_rcu_sched wait

There are definitely some problems here... Though I am seeing them in the sched_setaffinity() call rather than in the wait processing.

Quoted 8 lines

Quoted 7 lines

o If there are no synchronize_sched() calls for more than two minutes, one can see messages of the form "INFO: task rcu_sched_grace:924 blocked for more than 120 seconds." Any thoughts on how to avoid this message? Should I be using something other than __wait_event() and wake_up(), which sleep uninterruptibly, thus triggering this message?

Quoted 4 lines

Could you use __wait_event_interruptible and wake_up_interruptible instead ? softlockup.c only seems to complain when uninterruptible tasks are not scheduled for 2 minutes. I guess that when we receive a signal we could simply go through another loop.

I will give these a try.

Review

Quoted 4 lines

(quoted diff) +4 -0

+		ret = 0;
+		__wait_event_interruptible(rcu_ctrlblk.sched_wq,
+			     rcu_ctrlblk.sched_sleep != rcu_sched_sleeping,
+			     ret);

Don't we have to do something here to clear signal state if we are ever to block again? Maybe something like the following?

flush_signals(current):

Or am I missing something?

Quoted 2 lines

That's always good :)

Fixing the bug or losing track? ;-)

o If there are no synchronize_sched() calls for more than two minutes, one can see messages of the form "INFO: task rcu_sched_grace:924 blocked for more than 120 seconds." Any thoughts on how to avoid this message? Should I be using something other than __wait_event() and wake_up(), which sleep uninterruptibly, thus triggering this message?

Initialize call_rcu_sched sooner so it can be used in module_init stage.

(needed for LTTng)

Signed-off-by Mathieu Desnoyers <[email protected]>

---
 init/main.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Index: linux-2.6-lttng/init/main.c
===================================================================

init/main.c +1 -1

@@ -738,13 +738,13 @@ static void __init do_initcalls(void)
  */
 static void __init do_basic_setup(void)
 {
+	rcu_init_sched();
 	/* drivers will send hotplug events */
 	init_workqueues();
 	usermodehelper_init();
 	driver_init();
 	init_irq_proc();
 	do_initcalls();
-	rcu_init_sched();
 }
 
 static int __initdata nosoftlockup;

Signature · 3 lines

Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68

Hello!

Third cut of patch to provide the call_rcu_sched(). This is again to synchronize_sched() as call_rcu() is to synchronize_rcu().

Should be fine for experimental use, but not ready for inclusion.

Passes multi-hour rcutorture sessions with concurrent CPU hotplugging.

Fixes since the first version include a bug that could result in indefinite blocking (spotted by Gautham Shenoy), better resiliency against CPU-hotplug operations, and other minor fixes.

Fixes since the second version include reworking grace-period detection to avoid deadlocks that could happen when running concurrently with CPU hotplug, adding Mathieu's fix to avoid the softlockup messages, as well as Mathieu's fix to allow use earlier in boot.

Known/suspected shortcomings:

o Only moderately tested on x86-64 and POWER -- a few hours of
rcutorture with concurrent CPU hotplugging. In particular, I still do not trust the sleep/wakeup logic between call_rcu_sched() and rcu_sched_grace_period().

o Need to add call_rcu_sched() testing to rcutorture.

o Still needs rcu_barrier_sched() -- intending to incorporate
the version Mathieu provided.

This patch also fixes a long-standing bug in the earlier preemptable-RCU implementation of synchronize_rcu() that could result in loss of concurrent external changes to a task's CPU affinity mask. I still cannot remember who reported this...

...

kernel/rcupreempt.c +24 -0

+{
+	long curr;
+	long snap;
+	struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+	curr = rdssp->dynticks;
+	snap = rdssp->sched_dynticks_snap;
+	smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
+
+	/*
+	 * If the CPU remained in dynticks mode for the entire time
+	 * and didn't take any interrupts, NMIs, SMIs, or whatever,
+	 * then it cannot be in the middle of an rcu_read_lock(), so
+	 * the next rcu_read_lock() it executes must use the new value
+	 * of the counter.  Therefore, this CPU has been in a quiescent
+	 * state the entire time, and we don't need to wait for it.
+	 */
+
+	if ((curr == snap) && ((curr & 0x1) == 0))
+		return 0;
+
+	/*
+	 * If the CPU passed through or entered a dynticks idle phase with
+	 * no active irq handlers, then, as above, this CPU has already

(quoted diff) +7 -0

+		smp_mb();	/* Guard against aggressive schedule(). */
+	     	rcu_qsctr_inc(cpu);
+	}
+
 	rcu_check_mb(cpu);
 	if (rcu_ctrlblk.completed == rdp->completed)
 		rcu_try_flip();
 
...
 
+
+	/*
+	 * The rcu_sched grace-period processing might have bypassed

kernel/rcupreempt.c +14 -0

+	rdp->rcu_sched_sleeping = 1;
+	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */

@@ -993,26 +1129,194 @@ void call_rcu(struct rcu_head *head, voi
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+{
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int wake_gp = 0;
+
+	head->func = func;
+	head->next = NULL;
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	*rdp->nextschedtail = head;

kernel/rcupreempt.c +4 -0

+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
+			wake_gp = 0;
+		rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);

{

kernel/rcupreempt.c +5 -8

-	cpumask_t oldmask;
+	int couldsleep;		/* might sleep after current pass. */
+	int couldsleepnext = 0; /* might sleep after next pass. */
 	int cpu;
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int ret;
 
-	if (sched_getaffinity(0, &oldmask) < 0)
-		oldmask = cpu_possible_map;
-	for_each_online_cpu(cpu) {
-		sched_setaffinity(0, cpumask_of_cpu(cpu));
-		schedule();
-	}
-	sched_setaffinity(0, oldmask);

kernel/rcupreempt.c +8 -0

+		/*
+		 * Wait on each CPU in turn to have either visited
+		 * a quiescent state or been in dynticks-idle mode.
+		 */
+
+		for_each_online_cpu(cpu) {
+			while (rcu_qsctr_inc_needed(cpu) &&
+			       rcu_qsctr_inc_needed_dyntick(cpu)) {

kernel/rcupreempt.c +53 -0

+			 * Advance the callbacks!  We share normal RCU's
+			 * donelist, since callbacks are invoked the
+			 * same way in either case.
+			 */
+
+			if (rdp->waitschedlist != NULL) {
+				*rdp->donetail = rdp->waitschedlist;
+				rdp->donetail = rdp->waitschedtail;
+
+				/*
+				 * Next rcu_check_callbacks() will
+				 * do the required raise_softirq().
+				 */
+			}
+			if (rdp->nextschedlist != NULL) {
+				rdp->waitschedlist = rdp->nextschedlist;
+				rdp->waitschedtail = rdp->nextschedtail;
+				couldsleep = 0;
+				couldsleepnext = 0;
+			} else {
+				rdp->waitschedlist = NULL;
+				rdp->waitschedtail = &rdp->waitschedlist;
+			}
+			rdp->nextschedlist = NULL;
+			rdp->nextschedtail = &rdp->nextschedlist;
+
+			/* Mark sleep intention. */
+
+			rdp->rcu_sched_sleeping = couldsleep;
+
+			spin_unlock_irqrestore(&rdp->lock, flags);
+		}
+
+		/* If we saw callbacks on the last scan, go deal with them. */
+
+		if (!couldsleep)
+			continue;
+
+		/* Attempt to block... */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
+
+			/*
+			 * Someone posted a callback after we scanned.
+			 * Go take care of it.
+			 */
+
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+			couldsleepnext = 0;
+			continue;
+		}
+

kernel/rcupreempt.c +21 -3

+		couldsleepnext = 0;
+
+	} while (!kthread_should_stop());
+
+	return (0);
 }
-EXPORT_SYMBOL_GPL(__synchronize_sched);
 
 /*
  * Check to see if any future RCU-related work will need to be done

@@ -1029,7 +1333,9 @@ int rcu_needs_cpu(int cpu)
 
 	return (rdp->donelist != NULL ||
 		!!rdp->waitlistcount ||
-		rdp->nextlist != NULL);
+		rdp->nextlist != NULL ||
+		rdp->nextschedlist != NULL ||
+		rdp->waitschedlist != NULL);
 }
 
 int rcu_pending(int cpu)

@@ -1040,7 +1346,9 @@ int rcu_pending(int cpu)
 
 	if (rdp->donelist != NULL ||
 	    !!rdp->waitlistcount ||
-	    rdp->nextlist != NULL)
+	    rdp->nextlist != NULL ||
+	    rdp->nextschedlist != NULL ||
+	    rdp->waitschedlist != NULL)
 		return 1;
 
 	/* The RCU core needs an acknowledgement from this CPU. */

@@ -1107,6 +1415,11 @@ void __init __rcu_init(void)
 		rdp->donetail = &rdp->donelist;
 		rdp->rcu_flipctr[0] = 0;
 		rdp->rcu_flipctr[1] = 0;
+		rdp->nextschedlist = NULL;
+		rdp->nextschedtail = &rdp->nextschedlist;
+		rdp->waitschedlist = NULL;
+		rdp->waitschedtail = &rdp->waitschedlist;
+		rdp->rcu_sched_sleeping = 0;
 	}
 	register_cpu_notifier(&rcu_nb);
 

@@ -1129,6 +1442,18 @@ void __init __rcu_init(void)
 }
 
 /*
+ * Late-boot-time RCU initialization that must wait until after scheduler
+ * has been initialized.
+ */
+void __init rcu_init_sched(void)
+{

Quoted 6 lines · On Sun, 6 Apr 2008 14:37:19 -0700 "Paul E. McKenney"…

Hello!

Third cut of patch to provide the call_rcu_sched(). This is again to synchronize_sched() as call_rcu() is to synchronize_rcu().

Should be fine for experimental use, but not ready for inclusion.

Let me know when to come out of hiding ;)

Quoted 30 lines

Passes multi-hour rcutorture sessions with concurrent CPU hotplugging.

Fixes since the first version include a bug that could result in indefinite blocking (spotted by Gautham Shenoy), better resiliency against CPU-hotplug operations, and other minor fixes.

Fixes since the second version include reworking grace-period detection to avoid deadlocks that could happen when running concurrently with CPU hotplug, adding Mathieu's fix to avoid the softlockup messages, as well as Mathieu's fix to allow use earlier in boot.

Known/suspected shortcomings:

o Only moderately tested on x86-64 and POWER -- a few hours of
rcutorture with concurrent CPU hotplugging. In particular, I still do not trust the sleep/wakeup logic between call_rcu_sched() and rcu_sched_grace_period().

o Need to add call_rcu_sched() testing to rcutorture.

o Still needs rcu_barrier_sched() -- intending to incorporate
the version Mathieu provided.

This patch also fixes a long-standing bug in the earlier preemptable-RCU implementation of synchronize_rcu() that could result in loss of concurrent external changes to a task's CPU affinity mask. I still cannot remember who reported this...

...

include/linux/rcuclassic.h extern struct lockdep_map rcu_lock_map; · line 159

Quoted 4 lines

include/linux/rcuclassic.h +3 -0

+#define call_rcu_sched(head, func) call_rcu(head, func)
+
 extern void __rcu_init(void);
+#define rcu_init_sched()	do { } while (0)

There are lots of creepy macros-which-probably-dont-need-to-be-macros in here.

kernel/rcupreempt.c rcu_try_flip_waitmb_needed(int cpu) · line 651

Quoted 3 lines

kernel/rcupreempt.c +3 -0

+
+static inline int
+rcu_qsctr_inc_needed_dyntick(int cpu)

Unneeded newline.

Quoted 24 lines

kernel/rcupreempt.c +24 -0

+{
+	long curr;
+	long snap;
+	struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+	curr = rdssp->dynticks;
+	snap = rdssp->sched_dynticks_snap;
+	smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
+
+	/*
+	 * If the CPU remained in dynticks mode for the entire time
+	 * and didn't take any interrupts, NMIs, SMIs, or whatever,
+	 * then it cannot be in the middle of an rcu_read_lock(), so
+	 * the next rcu_read_lock() it executes must use the new value
+	 * of the counter.  Therefore, this CPU has been in a quiescent
+	 * state the entire time, and we don't need to wait for it.
+	 */
+
+	if ((curr == snap) && ((curr & 0x1) == 0))
+		return 0;
+
+	/*
+	 * If the CPU passed through or entered a dynticks idle phase with
+	 * no active irq handlers, then, as above, this CPU has already

kernel/rcupreempt.c rcu_try_flip_waitmb_needed(int cpu) · line 685

Quoted 10 lines

kernel/rcupreempt.c +10 -0

+	 * passed through a quiescent state.
+	 */
+
+	if ((curr - snap) > 2 || (snap & 0x1) == 0)
+		return 0;
+
+	/* We need this CPU to go through a quiescent state. */
+
+	return 1;
+}

That's a pretty big inline. It only has a single callsite so the compiler should inline it for us. And if it grows a second callsite, the inlining is probably wrong.

Review

Quoted 2 lines

(quoted diff) +2 -0

+static inline int
+rcu_qsctr_inc_needed(int cpu)

Unneeded newline.

kernel/rcupreempt.c void rcu_check_callbacks(int cpu, int us · line 929

Quoted 10 lines

/*
* Get here when RCU is idle. Decide whether we need to
* move out of idle state, and return non-zero if so.

kernel/rcupreempt.c +3 -0

@@ -821,6 +924,13 @@ void rcu_check_callbacks(int cpu, int us
 	unsigned long flags;
 	struct rcu_data *rdp = RCU_DATA_CPU(cpu);
 
+	if (user ||
+	    (idle_cpu(cpu) && !in_softirq() &&
+	     hardirq_count() <= (1 << HARDIRQ_SHIFT))) {

I think this test could do with a bigfatcomment explaining what it is doing.

Quoted 10 lines

kernel/rcupreempt.c +4 -0

+		smp_mb();	/* Guard against aggressive schedule(). */
+	     	rcu_qsctr_inc(cpu);
+	}
+
 	rcu_check_mb(cpu);
 	if (rcu_ctrlblk.completed == rdp->completed)
 		rcu_try_flip();
 
...

kernel/rcupreempt.c void rcu_offline_cpu(int cpu) · line 1070

Quoted 13 lines

kernel/rcupreempt.c +13 -0

+
+	/*
+	 * The rcu_sched grace-period processing might have bypassed
+	 * this CPU, given that it was not in the rcu_cpu_online_map
+	 * when the grace-period scan started.  This means that the
+	 * grace-period task might sleep.  So make sure that if this
+	 * should happen, the first callback posted to this CPU will
+	 * wake up the grace-period task if need be.
+	 */
+
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);

I assume that splitting the irq-disable from the spin_lock is a little latency optimisation?

Quoted 22 lines

kernel/rcupreempt.c +14 -0

+	rdp->rcu_sched_sleeping = 1;
+	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */

@@ -993,26 +1129,194 @@ void call_rcu(struct rcu_head *head, voi
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+{
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int wake_gp = 0;
+
+	head->func = func;
+	head->next = NULL;
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	*rdp->nextschedtail = head;

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1153

Quoted 10 lines

kernel/rcupreempt.c +10 -0

+	rdp->nextschedtail = &head->next;
+	if (rdp->rcu_sched_sleeping) {
+
+		/* Grace-period processing might be sleeping... */
+
+		rdp->rcu_sched_sleeping = 0;
+		wake_gp = 1;
+	}
+	spin_unlock(&rdp->lock);
+	local_irq_restore(flags);

spin_unlock_irqrestore() here would be consistent with the above.

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1158

Quoted 5 lines

kernel/rcupreempt.c +5 -0

+	if (wake_gp) {
+
+		/* Wake up grace-period processing, unless someone beat us. */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);

If wake_gp!=0 is common then we could microoptimise straight-line performance here by retaining the irq-offness from above.

Quoted 4 lines

kernel/rcupreempt.c +4 -0

+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
+			wake_gp = 0;
+		rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi

Quoted 10 lines

kernel/rcupreempt.c +7 -0

+		if (wake_gp)
+			wake_up_interruptible(&rcu_ctrlblk.sched_wq);
+	}
+}
+EXPORT_SYMBOL_GPL(call_rcu_sched);
 
...
 
+static int
+rcu_sched_grace_period(void *arg)

Unneeded newline.

Quoted 16 lines

{

kernel/rcupreempt.c +5 -8

-	cpumask_t oldmask;
+	int couldsleep;		/* might sleep after current pass. */
+	int couldsleepnext = 0; /* might sleep after next pass. */
 	int cpu;
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int ret;
 
-	if (sched_getaffinity(0, &oldmask) < 0)
-		oldmask = cpu_possible_map;
-	for_each_online_cpu(cpu) {
-		sched_setaffinity(0, cpumask_of_cpu(cpu));
-		schedule();
-	}
-	sched_setaffinity(0, oldmask);

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1200

Quoted 10 lines

kernel/rcupreempt.c +10 -0

+	/*
+	 * Each pass through the following loop handles one
+	 * rcu_sched grace period cycle.
+	 */
+
+	do {
+		
+		/* Save each CPU's current state. */
+
+		for_each_online_cpu(cpu) {

Numerous unneeded newline ;)

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1210

Quoted 10 lines

kernel/rcupreempt.c +10 -0

+			dyntick_save_progress_counter_sched(cpu);
+			save_qsctr_sched(cpu);
+		}
+
+		/*
+		 * Sleep for about an RCU grace-period's worth to
+		 * allow better batching and to consume less CPU.
+		 */
+
+		schedule_timeout_interruptible(HZ / 20);

eek, a magic number.

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1223

Quoted 12 lines

kernel/rcupreempt.c +12 -0

+		/*
+		 * If there was nothing to do last time, prepare to
+		 * sleep at the end of the current grace period cycle.
+		 */
+
+		couldsleep = couldsleepnext;
+		couldsleepnext = 1;
+		if (couldsleep) {
+			spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+			rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		}

If the above locking actually correct and needed? The write to rcu_ctrlblk.sched_sleep is a single word...

Quoted 8 lines

kernel/rcupreempt.c +8 -0

+		/*
+		 * Wait on each CPU in turn to have either visited
+		 * a quiescent state or been in dynticks-idle mode.
+		 */
+
+		for_each_online_cpu(cpu) {
+			while (rcu_qsctr_inc_needed(cpu) &&
+			       rcu_qsctr_inc_needed_dyntick(cpu)) {

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1242

Quoted 10 lines

kernel/rcupreempt.c +10 -0

+				/* resched_cpu(cpu); */
+				schedule_timeout_interruptible(1);
+			}
+		}
+
+		/*
+		 * Advance callbacks for each CPU.
+		 */
+
+		for_each_online_cpu(cpu) {

It's more conventional to omit the blank line after the above form of comment block.

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1249

Quoted 6 lines

kernel/rcupreempt.c +6 -0

+			rdp = RCU_DATA_CPU(cpu);
+			spin_lock_irqsave(&rdp->lock, flags);
+
+			/*
+			 * We are running on this CPU irq-disabled, so no
+			 * CPU can go offline until we re-enable irqs.

but, but, but. The cpu at `cpu' could have gone offline just before we disabled local interrupts.

Quoted 54 lines

kernel/rcupreempt.c +53 -0

+			 * Advance the callbacks!  We share normal RCU's
+			 * donelist, since callbacks are invoked the
+			 * same way in either case.
+			 */
+
+			if (rdp->waitschedlist != NULL) {
+				*rdp->donetail = rdp->waitschedlist;
+				rdp->donetail = rdp->waitschedtail;
+
+				/*
+				 * Next rcu_check_callbacks() will
+				 * do the required raise_softirq().
+				 */
+			}
+			if (rdp->nextschedlist != NULL) {
+				rdp->waitschedlist = rdp->nextschedlist;
+				rdp->waitschedtail = rdp->nextschedtail;
+				couldsleep = 0;
+				couldsleepnext = 0;
+			} else {
+				rdp->waitschedlist = NULL;
+				rdp->waitschedtail = &rdp->waitschedlist;
+			}
+			rdp->nextschedlist = NULL;
+			rdp->nextschedtail = &rdp->nextschedlist;
+
+			/* Mark sleep intention. */
+
+			rdp->rcu_sched_sleeping = couldsleep;
+
+			spin_unlock_irqrestore(&rdp->lock, flags);
+		}
+
+		/* If we saw callbacks on the last scan, go deal with them. */
+
+		if (!couldsleep)
+			continue;
+
+		/* Attempt to block... */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
+
+			/*
+			 * Someone posted a callback after we scanned.
+			 * Go take care of it.
+			 */
+
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+			couldsleepnext = 0;
+			continue;
+		}
+

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1313

Quoted 10 lines

kernel/rcupreempt.c +10 -0

+		/* Block until the next person posts a callback. */
+
+		rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		ret = 0;
+		__wait_event_interruptible(rcu_ctrlblk.sched_wq,
+			rcu_ctrlblk.sched_sleep != rcu_sched_sleeping,
+			ret);
+		if (ret)
+			flush_signals(current);

That flush_signals() was a surprise. A desurprising comment would be nice.

Quoted 53 lines

kernel/rcupreempt.c +21 -3

+		couldsleepnext = 0;
+
+	} while (!kthread_should_stop());
+
+	return (0);
 }
-EXPORT_SYMBOL_GPL(__synchronize_sched);
 
 /*
  * Check to see if any future RCU-related work will need to be done

@@ -1029,7 +1333,9 @@ int rcu_needs_cpu(int cpu)
 
 	return (rdp->donelist != NULL ||
 		!!rdp->waitlistcount ||
-		rdp->nextlist != NULL);
+		rdp->nextlist != NULL ||
+		rdp->nextschedlist != NULL ||
+		rdp->waitschedlist != NULL);
 }
 
 int rcu_pending(int cpu)

@@ -1040,7 +1346,9 @@ int rcu_pending(int cpu)
 
 	if (rdp->donelist != NULL ||
 	    !!rdp->waitlistcount ||
-	    rdp->nextlist != NULL)
+	    rdp->nextlist != NULL ||
+	    rdp->nextschedlist != NULL ||
+	    rdp->waitschedlist != NULL)
 		return 1;
 
 	/* The RCU core needs an acknowledgement from this CPU. */

@@ -1107,6 +1415,11 @@ void __init __rcu_init(void)
 		rdp->donetail = &rdp->donelist;
 		rdp->rcu_flipctr[0] = 0;
 		rdp->rcu_flipctr[1] = 0;
+		rdp->nextschedlist = NULL;
+		rdp->nextschedtail = &rdp->nextschedlist;
+		rdp->waitschedlist = NULL;
+		rdp->waitschedtail = &rdp->waitschedlist;
+		rdp->rcu_sched_sleeping = 0;
 	}
 	register_cpu_notifier(&rcu_nb);
 

@@ -1129,6 +1442,18 @@ void __init __rcu_init(void)
 }
 
 /*
+ * Late-boot-time RCU initialization that must wait until after scheduler
+ * has been initialized.
+ */
+void __init rcu_init_sched(void)
+{

kernel/rcupreempt.c void __init __rcu_init(void) · line 1459

Quoted 10 lines

kernel/rcupreempt.c +7 -0

+	rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
+						  NULL,
+						  "rcu_sched_grace_period");
+	WARN_ON(IS_ERR(rcu_sched_grace_period_task));
+}
+
+/*
  * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
  */
 void synchronize_kernel(void)

I suspect I don't understand any of the RCU code any more.

Quoted 1 line

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1244

Quoted 1 line

kernel/rcupreempt.c +1 -0

+			rdp = RCU_DATA_CPU(cpu);

                                                               <-- here ------\

Quoted 8 lines

Review

Quoted 5 lines

(quoted diff) +5 -0

+			spin_lock_irqsave(&rdp->lock, flags);                   |
+                                                                         |
+			/*                                                      |
+			 * We are running on this CPU irq-disabled, so no       |
+			 * CPU can go offline until we re-enable irqs.          |

                                                                            |
but, but, but.  The cpu at `cpu' could have gone offline just before we     |
disabled local interrupts.                                                  |

                                                                              |
In that case the CPU_DEAD callback should have migrated the rcu-lists to      |
a cpu which is online.                                                        |

Quoted 10 lines · On Tue, 8 Apr 2008 13:40:48 +0530 Gautham R Shenoy <[email protected]> wrote:

Quoted 1 line

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1244

Quoted 1 line

kernel/rcupreempt.c +1 -0

+			rdp = RCU_DATA_CPU(cpu);

                                                               <-- here ------\

Quoted 8 lines

Review

Quoted 5 lines

(quoted diff) +5 -0

+			spin_lock_irqsave(&rdp->lock, flags);                   |
+                                                                         |
+			/*                                                      |
+			 * We are running on this CPU irq-disabled, so no       |
+			 * CPU can go offline until we re-enable irqs.          |

                                                                            |
but, but, but.  The cpu at `cpu' could have gone offline just before we     |
disabled local interrupts.                                                  |

Quoted 3 lines

                                                                              |
In that case the CPU_DEAD callback should have migrated the rcu-lists to      |
a cpu which is online.                                                        |

                                                                                |

But local variable rdp might be pointing at the now-offlined CPU's data? -------/

Quoted 15 lines · Andrew Morton

Quoted 13 lines · On Tue, 8 Apr 2008 13:40:48 +0530 Gautham R Shenoy <[email protected]> wrote:

Quoted 1 line

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1244

Quoted 1 line

kernel/rcupreempt.c +1 -0

+			rdp = RCU_DATA_CPU(cpu);

                                                               <-- here ------\

Quoted 8 lines

Review

Quoted 5 lines

(quoted diff) +5 -0

+			spin_lock_irqsave(&rdp->lock, flags);                   |
+                                                                         |
+			/*                                                      |
+			 * We are running on this CPU irq-disabled, so no       |
+			 * CPU can go offline until we re-enable irqs.          |

                                                                            |
but, but, but.  The cpu at `cpu' could have gone offline just before we     |
disabled local interrupts.                                                  |

                                                                              |
In that case the CPU_DEAD callback should have migrated the rcu-lists to      |
a cpu which is online.                                                        |

Quoted 2 lines

                                                                                |

But local variable rdp might be pointing at the now-offlined CPU's data? -------/

Right. But then rdp wouldn't contain anything useful at this point. So, we may only end up taking the rdp->lock, observe that there's nothing to do, and move on.

Is there something else that I am missing?

Quoted 23 lines · On Tue, 8 Apr 2008 14:26:37 +0530 Gautham R Shenoy <[email protected]> wrote:

Quoted 12 lines · Andrew Morton

Quoted 10 lines · On Tue, 8 Apr 2008 13:40:48 +0530 Gautham R Shenoy <[email protected]> wrote:

Quoted 1 line

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1244

Quoted 1 line

kernel/rcupreempt.c +1 -0

+			rdp = RCU_DATA_CPU(cpu);

                                                               <-- here ------\

Quoted 8 lines

Review

Quoted 5 lines

(quoted diff) +5 -0

+			spin_lock_irqsave(&rdp->lock, flags);                   |
+                                                                         |
+			/*                                                      |
+			 * We are running on this CPU irq-disabled, so no       |
+			 * CPU can go offline until we re-enable irqs.          |

                                                                            |
but, but, but.  The cpu at `cpu' could have gone offline just before we     |
disabled local interrupts.                                                  |

Quoted 5 lines

Quoted 3 lines

                                                                              |
In that case the CPU_DEAD callback should have migrated the rcu-lists to      |
a cpu which is online.                                                        |

                                                                                |

But local variable rdp might be pointing at the now-offlined CPU's data? -------/

Right. But then rdp wouldn't contain anything useful at this point. So, we may only end up taking the rdp->lock, observe that there's nothing to do, and move on.

Quoted 1 line

Is there something else that I am missing?

erm, I guess that'll work OK.

There were intentions to release the per-cpu memory during unplug, but nobody has threatened to do that for a while.

Quoted 6 lines · On Sun, 6 Apr 2008 14:37:19 -0700 "Paul E. McKenney"…

Hello!

Third cut of patch to provide the call_rcu_sched(). This is again to synchronize_sched() as call_rcu() is to synchronize_rcu().

Should be fine for experimental use, but not ready for inclusion.

Let me know when to come out of hiding ;)

Quoted 30 lines

Passes multi-hour rcutorture sessions with concurrent CPU hotplugging.

Fixes since the first version include a bug that could result in indefinite blocking (spotted by Gautham Shenoy), better resiliency against CPU-hotplug operations, and other minor fixes.

Fixes since the second version include reworking grace-period detection to avoid deadlocks that could happen when running concurrently with CPU hotplug, adding Mathieu's fix to avoid the softlockup messages, as well as Mathieu's fix to allow use earlier in boot.

Known/suspected shortcomings:

o Only moderately tested on x86-64 and POWER -- a few hours of
rcutorture with concurrent CPU hotplugging. In particular, I still do not trust the sleep/wakeup logic between call_rcu_sched() and rcu_sched_grace_period().

o Need to add call_rcu_sched() testing to rcutorture.

o Still needs rcu_barrier_sched() -- intending to incorporate
the version Mathieu provided.

This patch also fixes a long-standing bug in the earlier preemptable-RCU implementation of synchronize_rcu() that could result in loss of concurrent external changes to a task's CPU affinity mask. I still cannot remember who reported this...

...

include/linux/rcuclassic.h extern struct lockdep_map rcu_lock_map; · line 159

Quoted 4 lines

include/linux/rcuclassic.h +3 -0

+#define call_rcu_sched(head, func) call_rcu(head, func)
+
 extern void __rcu_init(void);
+#define rcu_init_sched()	do { } while (0)

There are lots of creepy macros-which-probably-dont-need-to-be-macros in here.

kernel/rcupreempt.c rcu_try_flip_waitmb_needed(int cpu) · line 651

Quoted 3 lines

kernel/rcupreempt.c +3 -0

+
+static inline int
+rcu_qsctr_inc_needed_dyntick(int cpu)

Unneeded newline.

Quoted 30 lines

kernel/rcupreempt.c +30 -0

+{
+	long curr;
+	long snap;
+	struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+	curr = rdssp->dynticks;
+	snap = rdssp->sched_dynticks_snap;
+	smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
+
+	/*
+	 * If the CPU remained in dynticks mode for the entire time
+	 * and didn't take any interrupts, NMIs, SMIs, or whatever,
+	 * then it cannot be in the middle of an rcu_read_lock(), so
+	 * the next rcu_read_lock() it executes must use the new value
+	 * of the counter.  Therefore, this CPU has been in a quiescent
+	 * state the entire time, and we don't need to wait for it.
+	 */
+
+	if ((curr == snap) && ((curr & 0x1) == 0))
+		return 0;
+
+	/*
+	 * If the CPU passed through or entered a dynticks idle phase with
+	 * no active irq handlers, then, as above, this CPU has already
+	 * passed through a quiescent state.
+	 */
+
+	if ((curr - snap) > 2 || (snap & 0x1) == 0)
+		return 0;
+

kernel/rcupreempt.c rcu_try_flip_waitmb_needed(int cpu) · line 685

Quoted 4 lines

kernel/rcupreempt.c +4 -0

+	/* We need this CPU to go through a quiescent state. */
+
+	return 1;
+}

That's a pretty big inline. It only has a single callsite so the compiler should inline it for us. And if it grows a second callsite, the inlining is probably wrong.

Review

Quoted 2 lines

(quoted diff) +2 -0

+static inline int
+rcu_qsctr_inc_needed(int cpu)

Unneeded newline.

Quoted 7 lines

/*
* Get here when RCU is idle. Decide whether we need to
* move out of idle state, and return non-zero if so.

kernel/rcupreempt.c +0 -0

@@ -821,6 +924,13 @@ void rcu_check_callbacks(int cpu, int us
 	unsigned long flags;
 	struct rcu_data *rdp = RCU_DATA_CPU(cpu);
 

kernel/rcupreempt.c void rcu_check_callbacks(int cpu, int us · line 929

Quoted 3 lines

kernel/rcupreempt.c +3 -0

+	if (user ||
+	    (idle_cpu(cpu) && !in_softirq() &&
+	     hardirq_count() <= (1 << HARDIRQ_SHIFT))) {

I think this test could do with a bigfatcomment explaining what it is doing.

Quoted 10 lines

kernel/rcupreempt.c +4 -0

+		smp_mb();	/* Guard against aggressive schedule(). */
+	     	rcu_qsctr_inc(cpu);
+	}
+
 	rcu_check_mb(cpu);
 	if (rcu_ctrlblk.completed == rdp->completed)
 		rcu_try_flip();
 
...

kernel/rcupreempt.c void rcu_offline_cpu(int cpu) · line 1070

Quoted 13 lines

kernel/rcupreempt.c +13 -0

+
+	/*
+	 * The rcu_sched grace-period processing might have bypassed
+	 * this CPU, given that it was not in the rcu_cpu_online_map
+	 * when the grace-period scan started.  This means that the
+	 * grace-period task might sleep.  So make sure that if this
+	 * should happen, the first callback posted to this CPU will
+	 * wake up the grace-period task if need be.
+	 */
+
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);

I assume that splitting the irq-disable from the spin_lock is a little latency optimisation?

Quoted 25 lines

kernel/rcupreempt.c +18 -0

+	rdp->rcu_sched_sleeping = 1;
+	spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */

@@ -993,26 +1129,194 @@ void call_rcu(struct rcu_head *head, voi
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+{
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int wake_gp = 0;
+
+	head->func = func;
+	head->next = NULL;
+	local_irq_save(flags);
+	rdp = RCU_DATA_ME();
+	spin_lock(&rdp->lock);
+	*rdp->nextschedtail = head;
+	rdp->nextschedtail = &head->next;
+	if (rdp->rcu_sched_sleeping) {
+
+		/* Grace-period processing might be sleeping... */

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1153

Quoted 6 lines

kernel/rcupreempt.c +6 -0

+
+		rdp->rcu_sched_sleeping = 0;
+		wake_gp = 1;
+	}
+	spin_unlock(&rdp->lock);
+	local_irq_restore(flags);

spin_unlock_irqrestore() here would be consistent with the above.

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1158

Quoted 5 lines

kernel/rcupreempt.c +5 -0

+	if (wake_gp) {
+
+		/* Wake up grace-period processing, unless someone beat us. */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);

If wake_gp!=0 is common then we could microoptimise straight-line performance here by retaining the irq-offness from above.

Quoted 10 lines

kernel/rcupreempt.c +9 -0

+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
+			wake_gp = 0;
+		rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		if (wake_gp)
+			wake_up_interruptible(&rcu_ctrlblk.sched_wq);
+	}
+}
+EXPORT_SYMBOL_GPL(call_rcu_sched);
 

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1182

Quoted 4 lines

...

kernel/rcupreempt.c +2 -0

+static int
+rcu_sched_grace_period(void *arg)

Unneeded newline.

Quoted 20 lines

{

kernel/rcupreempt.c +9 -8

-	cpumask_t oldmask;
+	int couldsleep;		/* might sleep after current pass. */
+	int couldsleepnext = 0; /* might sleep after next pass. */
 	int cpu;
+	unsigned long flags;
+	struct rcu_data *rdp;
+	int ret;
 
-	if (sched_getaffinity(0, &oldmask) < 0)
-		oldmask = cpu_possible_map;
-	for_each_online_cpu(cpu) {
-		sched_setaffinity(0, cpumask_of_cpu(cpu));
-		schedule();
-	}
-	sched_setaffinity(0, oldmask);
+	/*
+	 * Each pass through the following loop handles one
+	 * rcu_sched grace period cycle.
+	 */

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1200

Quoted 6 lines

kernel/rcupreempt.c +6 -0

+
+	do {
+		
+		/* Save each CPU's current state. */
+
+		for_each_online_cpu(cpu) {

Numerous unneeded newline ;)

Quoted 4 lines

kernel/rcupreempt.c +4 -0

+			dyntick_save_progress_counter_sched(cpu);
+			save_qsctr_sched(cpu);
+		}
+

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1210

Quoted 6 lines

kernel/rcupreempt.c +6 -0

+		/*
+		 * Sleep for about an RCU grace-period's worth to
+		 * allow better batching and to consume less CPU.
+		 */
+
+		schedule_timeout_interruptible(HZ / 20);

eek, a magic number.

Quoted 7 lines

kernel/rcupreempt.c +7 -0

+		/*
+		 * If there was nothing to do last time, prepare to
+		 * sleep at the end of the current grace period cycle.
+		 */
+
+		couldsleep = couldsleepnext;
+		couldsleepnext = 1;

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1223

Quoted 5 lines

kernel/rcupreempt.c +5 -0

+		if (couldsleep) {
+			spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+			rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+		}

If the above locking actually correct and needed? The write to rcu_ctrlblk.sched_sleep is a single word...

Quoted 13 lines

kernel/rcupreempt.c +13 -0

+		/*
+		 * Wait on each CPU in turn to have either visited
+		 * a quiescent state or been in dynticks-idle mode.
+		 */
+
+		for_each_online_cpu(cpu) {
+			while (rcu_qsctr_inc_needed(cpu) &&
+			       rcu_qsctr_inc_needed_dyntick(cpu)) {
+				/* resched_cpu(cpu); */
+				schedule_timeout_interruptible(1);
+			}
+		}
+

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1242

Quoted 5 lines

kernel/rcupreempt.c +5 -0

+		/*
+		 * Advance callbacks for each CPU.
+		 */
+
+		for_each_online_cpu(cpu) {

It's more conventional to omit the blank line after the above form of comment block.

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1249

Quoted 6 lines

kernel/rcupreempt.c +6 -0

+			rdp = RCU_DATA_CPU(cpu);
+			spin_lock_irqsave(&rdp->lock, flags);
+
+			/*
+			 * We are running on this CPU irq-disabled, so no
+			 * CPU can go offline until we re-enable irqs.

but, but, but. The cpu at `cpu' could have gone offline just before we disabled local interrupts.

Quoted 57 lines

kernel/rcupreempt.c +57 -0

+			 * Advance the callbacks!  We share normal RCU's
+			 * donelist, since callbacks are invoked the
+			 * same way in either case.
+			 */
+
+			if (rdp->waitschedlist != NULL) {
+				*rdp->donetail = rdp->waitschedlist;
+				rdp->donetail = rdp->waitschedtail;
+
+				/*
+				 * Next rcu_check_callbacks() will
+				 * do the required raise_softirq().
+				 */
+			}
+			if (rdp->nextschedlist != NULL) {
+				rdp->waitschedlist = rdp->nextschedlist;
+				rdp->waitschedtail = rdp->nextschedtail;
+				couldsleep = 0;
+				couldsleepnext = 0;
+			} else {
+				rdp->waitschedlist = NULL;
+				rdp->waitschedtail = &rdp->waitschedlist;
+			}
+			rdp->nextschedlist = NULL;
+			rdp->nextschedtail = &rdp->nextschedlist;
+
+			/* Mark sleep intention. */
+
+			rdp->rcu_sched_sleeping = couldsleep;
+
+			spin_unlock_irqrestore(&rdp->lock, flags);
+		}
+
+		/* If we saw callbacks on the last scan, go deal with them. */
+
+		if (!couldsleep)
+			continue;
+
+		/* Attempt to block... */
+
+		spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+		if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
+
+			/*
+			 * Someone posted a callback after we scanned.
+			 * Go take care of it.
+			 */
+
+			spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+			couldsleepnext = 0;
+			continue;
+		}
+
+		/* Block until the next person posts a callback. */
+
+		rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
+		spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);

kernel/rcupreempt.c void call_rcu(struct rcu_head *head, voi · line 1313

Quoted 6 lines

kernel/rcupreempt.c +6 -0

+		ret = 0;
+		__wait_event_interruptible(rcu_ctrlblk.sched_wq,
+			rcu_ctrlblk.sched_sleep != rcu_sched_sleeping,
+			ret);
+		if (ret)
+			flush_signals(current);

That flush_signals() was a surprise. A desurprising comment would be nice.

Quoted 57 lines

kernel/rcupreempt.c +25 -3

+		couldsleepnext = 0;
+
+	} while (!kthread_should_stop());
+
+	return (0);
 }
-EXPORT_SYMBOL_GPL(__synchronize_sched);
 
 /*
  * Check to see if any future RCU-related work will need to be done

@@ -1029,7 +1333,9 @@ int rcu_needs_cpu(int cpu)
 
 	return (rdp->donelist != NULL ||
 		!!rdp->waitlistcount ||
-		rdp->nextlist != NULL);
+		rdp->nextlist != NULL ||
+		rdp->nextschedlist != NULL ||
+		rdp->waitschedlist != NULL);
 }
 
 int rcu_pending(int cpu)

@@ -1040,7 +1346,9 @@ int rcu_pending(int cpu)
 
 	if (rdp->donelist != NULL ||
 	    !!rdp->waitlistcount ||
-	    rdp->nextlist != NULL)
+	    rdp->nextlist != NULL ||
+	    rdp->nextschedlist != NULL ||
+	    rdp->waitschedlist != NULL)
 		return 1;
 
 	/* The RCU core needs an acknowledgement from this CPU. */

@@ -1107,6 +1415,11 @@ void __init __rcu_init(void)
 		rdp->donetail = &rdp->donelist;
 		rdp->rcu_flipctr[0] = 0;
 		rdp->rcu_flipctr[1] = 0;
+		rdp->nextschedlist = NULL;
+		rdp->nextschedtail = &rdp->nextschedlist;
+		rdp->waitschedlist = NULL;
+		rdp->waitschedtail = &rdp->waitschedlist;
+		rdp->rcu_sched_sleeping = 0;
 	}
 	register_cpu_notifier(&rcu_nb);
 

@@ -1129,6 +1442,18 @@ void __init __rcu_init(void)
 }
 
 /*
+ * Late-boot-time RCU initialization that must wait until after scheduler
+ * has been initialized.
+ */
+void __init rcu_init_sched(void)
+{
+	rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
+						  NULL,
+						  "rcu_sched_grace_period");
+	WARN_ON(IS_ERR(rcu_sched_grace_period_task));

kernel/rcupreempt.c void __init __rcu_init(void) · line 1459

Quoted 6 lines

kernel/rcupreempt.c +3 -0

+}
+
+/*
  * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
  */
 void synchronize_kernel(void)

I suspect I don't understand any of the RCU code any more.