rcu: merge TREE_PREEPT_RCU blocked_tasks[] lists
[linux-2.6.git] / kernel / rcutree_plugin.h
index 1cee04f..774f010 100644 (file)
  *        Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  */
 
+#include <linux/delay.h>
+#include <linux/stop_machine.h>
+
+/*
+ * Check the RCU kernel configuration parameters and print informative
+ * messages about anything out of the ordinary.  If you like #ifdef, you
+ * will love this function.
+ */
+static void __init rcu_bootup_announce_oddness(void)
+{
+#ifdef CONFIG_RCU_TRACE
+       printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n");
+#endif
+#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
+       printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
+              CONFIG_RCU_FANOUT);
+#endif
+#ifdef CONFIG_RCU_FANOUT_EXACT
+       printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n");
+#endif
+#ifdef CONFIG_RCU_FAST_NO_HZ
+       printk(KERN_INFO
+              "\tRCU dyntick-idle grace-period acceleration is enabled.\n");
+#endif
+#ifdef CONFIG_PROVE_RCU
+       printk(KERN_INFO "\tRCU lockdep checking is enabled.\n");
+#endif
+#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
+       printk(KERN_INFO "\tRCU torture testing starts during boot.\n");
+#endif
+#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
+       printk(KERN_INFO "\tVerbose stalled-CPUs detection is disabled.\n");
+#endif
+#if NUM_RCU_LVL_4 != 0
+       printk(KERN_INFO "\tExperimental four-level hierarchy is enabled.\n");
+#endif
+}
 
 #ifdef CONFIG_TREE_PREEMPT_RCU
 
 struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state);
 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
 
+static int rcu_preempted_readers_exp(struct rcu_node *rnp);
+
 /*
  * Tell them what RCU they are running.
  */
-static inline void rcu_bootup_announce(void)
+static void __init rcu_bootup_announce(void)
 {
-       printk(KERN_INFO
-              "Experimental preemptable hierarchical RCU implementation.\n");
+       printk(KERN_INFO "Preemptable hierarchical RCU implementation.\n");
+       rcu_bootup_announce_oddness();
 }
 
 /*
@@ -59,29 +98,44 @@ long rcu_batches_completed(void)
 EXPORT_SYMBOL_GPL(rcu_batches_completed);
 
 /*
+ * Force a quiescent state for preemptible RCU.
+ */
+void rcu_force_quiescent_state(void)
+{
+       force_quiescent_state(&rcu_preempt_state, 0);
+}
+EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
+
+/*
  * Record a preemptable-RCU quiescent state for the specified CPU.  Note
  * that this just means that the task currently running on the CPU is
  * not in a quiescent state.  There might be any number of tasks blocked
  * while in an RCU read-side critical section.
+ *
+ * Unlike the other rcu_*_qs() functions, callers to this function
+ * must disable irqs in order to protect the assignment to
+ * ->rcu_read_unlock_special.
  */
 static void rcu_preempt_qs(int cpu)
 {
        struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
-       rdp->passed_quiesc_completed = rdp->completed;
+
+       rdp->passed_quiesc_completed = rdp->gpnum - 1;
        barrier();
        rdp->passed_quiesc = 1;
+       current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
 }
 
 /*
  * We have entered the scheduler, and the current task might soon be
  * context-switched away from.  If this task is in an RCU read-side
  * critical section, we will no longer be able to rely on the CPU to
- * record that fact, so we enqueue the task on the appropriate entry
- * of the blocked_tasks[] array.  The task will dequeue itself when
- * it exits the outermost enclosing RCU read-side critical section.
- * Therefore, the current grace period cannot be permitted to complete
- * until the blocked_tasks[] entry indexed by the low-order bit of
- * rnp->gpnum empties.
+ * record that fact, so we enqueue the task on the blkd_tasks list.
+ * The task will dequeue itself when it exits the outermost enclosing
+ * RCU read-side critical section.  Therefore, the current grace period
+ * cannot be permitted to complete until the blkd_tasks list entries
+ * predating the current grace period drain, in other words, until
+ * rnp->gp_tasks becomes NULL.
  *
  * Caller must disable preemption.
  */
@@ -89,7 +143,6 @@ static void rcu_preempt_note_context_switch(int cpu)
 {
        struct task_struct *t = current;
        unsigned long flags;
-       int phase;
        struct rcu_data *rdp;
        struct rcu_node *rnp;
 
@@ -97,9 +150,9 @@ static void rcu_preempt_note_context_switch(int cpu)
            (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
 
                /* Possibly blocking in an RCU read-side critical section. */
-               rdp = rcu_preempt_state.rda[cpu];
+               rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
                rnp = rdp->mynode;
-               spin_lock_irqsave(&rnp->lock, flags);
+               raw_spin_lock_irqsave(&rnp->lock, flags);
                t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
                t->rcu_blocked_node = rnp;
 
@@ -111,16 +164,27 @@ static void rcu_preempt_note_context_switch(int cpu)
                 * (i.e., this CPU has not yet passed through a quiescent
                 * state for the current grace period), then as long
                 * as that task remains queued, the current grace period
-                * cannot end.
+                * cannot end.  Note that there is some uncertainty as
+                * to exactly when the current grace period started.
+                * We take a conservative approach, which can result
+                * in unnecessarily waiting on tasks that started very
+                * slightly after the current grace period began.  C'est
+                * la vie!!!
                 *
                 * But first, note that the current CPU must still be
                 * on line!
                 */
                WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
                WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
-               phase = (rnp->gpnum + !(rnp->qsmask & rdp->grpmask)) & 0x1;
-               list_add(&t->rcu_node_entry, &rnp->blocked_tasks[phase]);
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
+                       list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
+                       rnp->gp_tasks = &t->rcu_node_entry;
+               } else {
+                       list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
+                       if (rnp->qsmask & rdp->grpmask)
+                               rnp->gp_tasks = &t->rcu_node_entry;
+               }
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
        }
 
        /*
@@ -132,9 +196,8 @@ static void rcu_preempt_note_context_switch(int cpu)
         * grace period, then the fact that the task has been enqueued
         * means that we continue to block the current grace period.
         */
-       rcu_preempt_qs(cpu);
        local_irq_save(flags);
-       t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
+       rcu_preempt_qs(cpu);
        local_irq_restore(flags);
 }
 
@@ -145,16 +208,83 @@ static void rcu_preempt_note_context_switch(int cpu)
  */
 void __rcu_read_lock(void)
 {
-       ACCESS_ONCE(current->rcu_read_lock_nesting)++;
+       current->rcu_read_lock_nesting++;
        barrier();  /* needed if we ever invoke rcu_read_lock in rcutree.c */
 }
 EXPORT_SYMBOL_GPL(__rcu_read_lock);
 
+/*
+ * Check for preempted RCU readers blocking the current grace period
+ * for the specified rcu_node structure.  If the caller needs a reliable
+ * answer, it must hold the rcu_node's ->lock.
+ */
+static int rcu_preempted_readers(struct rcu_node *rnp)
+{
+       return rnp->gp_tasks != NULL;
+}
+
+/*
+ * Record a quiescent state for all tasks that were previously queued
+ * on the specified rcu_node structure and that were blocking the current
+ * RCU grace period.  The caller must hold the specified rnp->lock with
+ * irqs disabled, and this lock is released upon return, but irqs remain
+ * disabled.
+ */
+static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
+       __releases(rnp->lock)
+{
+       unsigned long mask;
+       struct rcu_node *rnp_p;
+
+       if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+               return;  /* Still need more quiescent states! */
+       }
+
+       rnp_p = rnp->parent;
+       if (rnp_p == NULL) {
+               /*
+                * Either there is only one rcu_node in the tree,
+                * or tasks were kicked up to root rcu_node due to
+                * CPUs going offline.
+                */
+               rcu_report_qs_rsp(&rcu_preempt_state, flags);
+               return;
+       }
+
+       /* Report up the rest of the hierarchy. */
+       mask = rnp->grpmask;
+       raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
+       raw_spin_lock(&rnp_p->lock);    /* irqs already disabled. */
+       rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
+}
+
+/*
+ * Advance a ->blkd_tasks-list pointer to the next entry, instead
+ * returning NULL if at the end of the list.
+ */
+static struct list_head *rcu_next_node_entry(struct task_struct *t,
+                                            struct rcu_node *rnp)
+{
+       struct list_head *np;
+
+       np = t->rcu_node_entry.next;
+       if (np == &rnp->blkd_tasks)
+               np = NULL;
+       return np;
+}
+
+/*
+ * Handle special cases during rcu_read_unlock(), such as needing to
+ * notify RCU core processing or task having blocked during the RCU
+ * read-side critical section.
+ */
 static void rcu_read_unlock_special(struct task_struct *t)
 {
        int empty;
+       int empty_exp;
        unsigned long flags;
-       unsigned long mask;
+       struct list_head *np;
        struct rcu_node *rnp;
        int special;
 
@@ -170,7 +300,6 @@ static void rcu_read_unlock_special(struct task_struct *t)
         */
        special = t->rcu_read_unlock_special;
        if (special & RCU_READ_UNLOCK_NEED_QS) {
-               t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
                rcu_preempt_qs(smp_processor_id());
        }
 
@@ -191,42 +320,41 @@ static void rcu_read_unlock_special(struct task_struct *t)
                 */
                for (;;) {
                        rnp = t->rcu_blocked_node;
-                       spin_lock(&rnp->lock);  /* irqs already disabled. */
+                       raw_spin_lock(&rnp->lock);  /* irqs already disabled. */
                        if (rnp == t->rcu_blocked_node)
                                break;
-                       spin_unlock(&rnp->lock);  /* irqs remain disabled. */
+                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
                }
-               empty = list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]);
+               empty = !rcu_preempted_readers(rnp);
+               empty_exp = !rcu_preempted_readers_exp(rnp);
+               smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
+               np = rcu_next_node_entry(t, rnp);
                list_del_init(&t->rcu_node_entry);
+               if (&t->rcu_node_entry == rnp->gp_tasks)
+                       rnp->gp_tasks = np;
+               if (&t->rcu_node_entry == rnp->exp_tasks)
+                       rnp->exp_tasks = np;
                t->rcu_blocked_node = NULL;
 
                /*
                 * If this was the last task on the current list, and if
                 * we aren't waiting on any CPUs, report the quiescent state.
-                * Note that both cpu_quiet_msk_finish() and cpu_quiet_msk()
-                * drop rnp->lock and restore irq.
+                * Note that rcu_report_unblock_qs_rnp() releases rnp->lock.
                 */
-               if (!empty && rnp->qsmask == 0 &&
-                   list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1])) {
-                       struct rcu_node *rnp_p;
-
-                       if (rnp->parent == NULL) {
-                               /* Only one rcu_node in the tree. */
-                               cpu_quiet_msk_finish(&rcu_preempt_state, flags);
-                               return;
-                       }
-                       /* Report up the rest of the hierarchy. */
-                       mask = rnp->grpmask;
-                       spin_unlock_irqrestore(&rnp->lock, flags);
-                       rnp_p = rnp->parent;
-                       spin_lock_irqsave(&rnp_p->lock, flags);
-                       WARN_ON_ONCE(rnp->qsmask);
-                       cpu_quiet_msk(mask, &rcu_preempt_state, rnp_p, flags);
-                       return;
-               }
-               spin_unlock(&rnp->lock);
+               if (empty)
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+               else
+                       rcu_report_unblock_qs_rnp(rnp, flags);
+
+               /*
+                * If this was the last task on the expedited lists,
+                * then we need to report up the rcu_node hierarchy.
+                */
+               if (!empty_exp && !rcu_preempted_readers_exp(rnp))
+                       rcu_report_exp_rnp(&rcu_preempt_state, rnp);
+       } else {
+               local_irq_restore(flags);
        }
-       local_irq_restore(flags);
 }
 
 /*
@@ -241,13 +369,58 @@ void __rcu_read_unlock(void)
        struct task_struct *t = current;
 
        barrier();  /* needed if we ever invoke rcu_read_unlock in rcutree.c */
-       if (--ACCESS_ONCE(t->rcu_read_lock_nesting) == 0 &&
+       --t->rcu_read_lock_nesting;
+       barrier();  /* decrement before load of ->rcu_read_unlock_special */
+       if (t->rcu_read_lock_nesting == 0 &&
            unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
                rcu_read_unlock_special(t);
+#ifdef CONFIG_PROVE_LOCKING
+       WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0);
+#endif /* #ifdef CONFIG_PROVE_LOCKING */
 }
 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 
-#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
+#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
+
+/*
+ * Dump detailed information for all tasks blocking the current RCU
+ * grace period on the specified rcu_node structure.
+ */
+static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
+{
+       unsigned long flags;
+       struct task_struct *t;
+
+       if (!rcu_preempted_readers(rnp))
+               return;
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       t = list_entry(rnp->gp_tasks,
+                      struct task_struct, rcu_node_entry);
+       list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+               sched_show_task(t);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+/*
+ * Dump detailed information for all tasks blocking the current RCU
+ * grace period.
+ */
+static void rcu_print_detail_task_stall(struct rcu_state *rsp)
+{
+       struct rcu_node *rnp = rcu_get_root(rsp);
+
+       rcu_print_detail_task_stall_rnp(rnp);
+       rcu_for_each_leaf_node(rsp, rnp)
+               rcu_print_detail_task_stall_rnp(rnp);
+}
+
+#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
+
+static void rcu_print_detail_task_stall(struct rcu_state *rsp)
+{
+}
+
+#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
 
 /*
  * Scan the current list of tasks blocked within RCU read-side critical
@@ -255,22 +428,25 @@ EXPORT_SYMBOL_GPL(__rcu_read_unlock);
  */
 static void rcu_print_task_stall(struct rcu_node *rnp)
 {
-       unsigned long flags;
-       struct list_head *lp;
-       int phase = rnp->gpnum & 0x1;
        struct task_struct *t;
 
-       if (!list_empty(&rnp->blocked_tasks[phase])) {
-               spin_lock_irqsave(&rnp->lock, flags);
-               phase = rnp->gpnum & 0x1; /* re-read under lock. */
-               lp = &rnp->blocked_tasks[phase];
-               list_for_each_entry(t, lp, rcu_node_entry)
-                       printk(" P%d", t->pid);
-               spin_unlock_irqrestore(&rnp->lock, flags);
-       }
+       if (!rcu_preempted_readers(rnp))
+               return;
+       t = list_entry(rnp->gp_tasks,
+                      struct task_struct, rcu_node_entry);
+       list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
+               printk(" P%d", t->pid);
 }
 
-#endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
+/*
+ * Suppress preemptible RCU's CPU stall warnings by pushing the
+ * time of the next stall-warning message comfortably far into the
+ * future.
+ */
+static void rcu_preempt_stall_reset(void)
+{
+       rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2;
+}
 
 /*
  * Check that the list of blocked tasks for the newly completed grace
@@ -278,23 +454,18 @@ static void rcu_print_task_stall(struct rcu_node *rnp)
  * period that still has RCU readers blocked!  This function must be
  * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
  * must be held by the caller.
+ *
+ * Also, if there are blocked tasks on the list, they automatically
+ * block the newly created grace period, so set up ->gp_tasks accordingly.
  */
 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 {
-       WARN_ON_ONCE(!list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]));
+       WARN_ON_ONCE(rcu_preempted_readers(rnp));
+       if (!list_empty(&rnp->blkd_tasks))
+               rnp->gp_tasks = rnp->blkd_tasks.next;
        WARN_ON_ONCE(rnp->qsmask);
 }
 
-/*
- * Check for preempted RCU readers for the specified rcu_node structure.
- * If the caller needs a reliable answer, it must hold the rcu_node's
- * >lock.
- */
-static int rcu_preempted_readers(struct rcu_node *rnp)
-{
-       return !list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]);
-}
-
 #ifdef CONFIG_HOTPLUG_CPU
 
 /*
@@ -303,45 +474,62 @@ static int rcu_preempted_readers(struct rcu_node *rnp)
  * rcu_node.  The reason for not just moving them to the immediate
  * parent is to remove the need for rcu_read_unlock_special() to
  * make more than two attempts to acquire the target rcu_node's lock.
+ * Returns true if there were tasks blocking the current RCU grace
+ * period.
+ *
+ * Returns 1 if there was previously a task blocking the current grace
+ * period on the specified rcu_node structure.
  *
  * The caller must hold rnp->lock with irqs disabled.
  */
-static void rcu_preempt_offline_tasks(struct rcu_state *rsp,
-                                     struct rcu_node *rnp,
-                                     struct rcu_data *rdp)
+static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
+                                    struct rcu_node *rnp,
+                                    struct rcu_data *rdp)
 {
-       int i;
        struct list_head *lp;
        struct list_head *lp_root;
+       int retval = 0;
        struct rcu_node *rnp_root = rcu_get_root(rsp);
-       struct task_struct *tp;
+       struct task_struct *t;
 
        if (rnp == rnp_root) {
                WARN_ONCE(1, "Last CPU thought to be offlined?");
-               return;  /* Shouldn't happen: at least one CPU online. */
+               return 0;  /* Shouldn't happen: at least one CPU online. */
        }
-       WARN_ON_ONCE(rnp != rdp->mynode &&
-                    (!list_empty(&rnp->blocked_tasks[0]) ||
-                     !list_empty(&rnp->blocked_tasks[1])));
+
+       /* If we are on an internal node, complain bitterly. */
+       WARN_ON_ONCE(rnp != rdp->mynode);
 
        /*
-        * Move tasks up to root rcu_node.  Rely on the fact that the
-        * root rcu_node can be at most one ahead of the rest of the
-        * rcu_nodes in terms of gp_num value.  This fact allows us to
-        * move the blocked_tasks[] array directly, element by element.
+        * Move tasks up to root rcu_node.  Don't try to get fancy for
+        * this corner-case operation -- just put this node's tasks
+        * at the head of the root node's list, and update the root node's
+        * ->gp_tasks and ->exp_tasks pointers to those of this node's,
+        * if non-NULL.  This might result in waiting for more tasks than
+        * absolutely necessary, but this is a good performance/complexity
+        * tradeoff.
         */
-       for (i = 0; i < 2; i++) {
-               lp = &rnp->blocked_tasks[i];
-               lp_root = &rnp_root->blocked_tasks[i];
-               while (!list_empty(lp)) {
-                       tp = list_entry(lp->next, typeof(*tp), rcu_node_entry);
-                       spin_lock(&rnp_root->lock); /* irqs already disabled */
-                       list_del(&tp->rcu_node_entry);
-                       tp->rcu_blocked_node = rnp_root;
-                       list_add(&tp->rcu_node_entry, lp_root);
-                       spin_unlock(&rnp_root->lock); /* irqs remain disabled */
-               }
+       if (rcu_preempted_readers(rnp))
+               retval |= RCU_OFL_TASKS_NORM_GP;
+       if (rcu_preempted_readers_exp(rnp))
+               retval |= RCU_OFL_TASKS_EXP_GP;
+       lp = &rnp->blkd_tasks;
+       lp_root = &rnp_root->blkd_tasks;
+       while (!list_empty(lp)) {
+               t = list_entry(lp->next, typeof(*t), rcu_node_entry);
+               raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
+               list_del(&t->rcu_node_entry);
+               t->rcu_blocked_node = rnp_root;
+               list_add(&t->rcu_node_entry, lp_root);
+               if (&t->rcu_node_entry == rnp->gp_tasks)
+                       rnp_root->gp_tasks = rnp->gp_tasks;
+               if (&t->rcu_node_entry == rnp->exp_tasks)
+                       rnp_root->exp_tasks = rnp->exp_tasks;
+               raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
        }
+       rnp->gp_tasks = NULL;
+       rnp->exp_tasks = NULL;
+       return retval;
 }
 
 /*
@@ -366,7 +554,6 @@ static void rcu_preempt_check_callbacks(int cpu)
        struct task_struct *t = current;
 
        if (t->rcu_read_lock_nesting == 0) {
-               t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
                rcu_preempt_qs(cpu);
                return;
        }
@@ -392,6 +579,188 @@ void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+/**
+ * synchronize_rcu - wait until a grace period has elapsed.
+ *
+ * Control will return to the caller some time after a full grace
+ * period has elapsed, in other words after all currently executing RCU
+ * read-side critical sections have completed.  Note, however, that
+ * upon return from synchronize_rcu(), the caller might well be executing
+ * concurrently with new RCU read-side critical sections that began while
+ * synchronize_rcu() was waiting.  RCU read-side critical sections are
+ * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
+ */
+void synchronize_rcu(void)
+{
+       struct rcu_synchronize rcu;
+
+       if (!rcu_scheduler_active)
+               return;
+
+       init_rcu_head_on_stack(&rcu.head);
+       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);
+       destroy_rcu_head_on_stack(&rcu.head);
+}
+EXPORT_SYMBOL_GPL(synchronize_rcu);
+
+static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
+static long sync_rcu_preempt_exp_count;
+static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
+
+/*
+ * Return non-zero if there are any tasks in RCU read-side critical
+ * sections blocking the current preemptible-RCU expedited grace period.
+ * If there is no preemptible-RCU expedited grace period currently in
+ * progress, returns zero unconditionally.
+ */
+static int rcu_preempted_readers_exp(struct rcu_node *rnp)
+{
+       return rnp->exp_tasks != NULL;
+}
+
+/*
+ * return non-zero if there is no RCU expedited grace period in progress
+ * for the specified rcu_node structure, in other words, if all CPUs and
+ * tasks covered by the specified rcu_node structure have done their bit
+ * for the current expedited grace period.  Works only for preemptible
+ * RCU -- other RCU implementation use other means.
+ *
+ * Caller must hold sync_rcu_preempt_exp_mutex.
+ */
+static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
+{
+       return !rcu_preempted_readers_exp(rnp) &&
+              ACCESS_ONCE(rnp->expmask) == 0;
+}
+
+/*
+ * Report the exit from RCU read-side critical section for the last task
+ * that queued itself during or before the current expedited preemptible-RCU
+ * grace period.  This event is reported either to the rcu_node structure on
+ * which the task was queued or to one of that rcu_node structure's ancestors,
+ * recursively up the tree.  (Calm down, calm down, we do the recursion
+ * iteratively!)
+ *
+ * Caller must hold sync_rcu_preempt_exp_mutex.
+ */
+static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
+{
+       unsigned long flags;
+       unsigned long mask;
+
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       for (;;) {
+               if (!sync_rcu_preempt_exp_done(rnp))
+                       break;
+               if (rnp->parent == NULL) {
+                       wake_up(&sync_rcu_preempt_exp_wq);
+                       break;
+               }
+               mask = rnp->grpmask;
+               raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
+               rnp = rnp->parent;
+               raw_spin_lock(&rnp->lock); /* irqs already disabled */
+               rnp->expmask &= ~mask;
+       }
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+/*
+ * Snapshot the tasks blocking the newly started preemptible-RCU expedited
+ * grace period for the specified rcu_node structure.  If there are no such
+ * tasks, report it up the rcu_node hierarchy.
+ *
+ * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock.
+ */
+static void
+sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
+{
+       int must_wait = 0;
+
+       raw_spin_lock(&rnp->lock); /* irqs already disabled */
+       if (!list_empty(&rnp->blkd_tasks)) {
+               rnp->exp_tasks = rnp->blkd_tasks.next;
+               must_wait = 1;
+       }
+       raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
+       if (!must_wait)
+               rcu_report_exp_rnp(rsp, rnp);
+}
+
+/*
+ * Wait for an rcu-preempt grace period, but expedite it.  The basic idea
+ * is to invoke synchronize_sched_expedited() to push all the tasks to
+ * the ->blkd_tasks lists and wait for this list to drain.
+ */
+void synchronize_rcu_expedited(void)
+{
+       unsigned long flags;
+       struct rcu_node *rnp;
+       struct rcu_state *rsp = &rcu_preempt_state;
+       long snap;
+       int trycount = 0;
+
+       smp_mb(); /* Caller's modifications seen first by other CPUs. */
+       snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
+       smp_mb(); /* Above access cannot bleed into critical section. */
+
+       /*
+        * Acquire lock, falling back to synchronize_rcu() if too many
+        * lock-acquisition failures.  Of course, if someone does the
+        * expedited grace period for us, just leave.
+        */
+       while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
+               if (trycount++ < 10)
+                       udelay(trycount * num_online_cpus());
+               else {
+                       synchronize_rcu();
+                       return;
+               }
+               if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
+                       goto mb_ret; /* Others did our work for us. */
+       }
+       if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0)
+               goto unlock_mb_ret; /* Others did our work for us. */
+
+       /* force all RCU readers onto ->blkd_tasks lists. */
+       synchronize_sched_expedited();
+
+       raw_spin_lock_irqsave(&rsp->onofflock, flags);
+
+       /* Initialize ->expmask for all non-leaf rcu_node structures. */
+       rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
+               raw_spin_lock(&rnp->lock); /* irqs already disabled. */
+               rnp->expmask = rnp->qsmaskinit;
+               raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
+       }
+
+       /* Snapshot current state of ->blkd_tasks lists. */
+       rcu_for_each_leaf_node(rsp, rnp)
+               sync_rcu_preempt_exp_init(rsp, rnp);
+       if (NUM_RCU_NODES > 1)
+               sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
+
+       raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
+
+       /* Wait for snapshotted ->blkd_tasks lists to drain. */
+       rnp = rcu_get_root(rsp);
+       wait_event(sync_rcu_preempt_exp_wq,
+                  sync_rcu_preempt_exp_done(rnp));
+
+       /* Clean up and exit. */
+       smp_mb(); /* ensure expedited GP seen before counter increment. */
+       ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
+unlock_mb_ret:
+       mutex_unlock(&sync_rcu_preempt_exp_mutex);
+mb_ret:
+       smp_mb(); /* ensure subsequent action seen after grace period. */
+}
+EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
+
 /*
  * Check to see if there is any immediate preemptable-RCU-related work
  * to be done.
@@ -410,6 +779,15 @@ static int rcu_preempt_needs_cpu(int cpu)
        return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
 }
 
+/**
+ * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
+ */
+void rcu_barrier(void)
+{
+       _rcu_barrier(&rcu_preempt_state, call_rcu);
+}
+EXPORT_SYMBOL_GPL(rcu_barrier);
+
 /*
  * Initialize preemptable RCU's per-CPU data.
  */
@@ -419,6 +797,22 @@ static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
 }
 
 /*
+ * Move preemptable RCU's callbacks from dying CPU to other online CPU.
+ */
+static void rcu_preempt_send_cbs_to_online(void)
+{
+       rcu_send_cbs_to_online(&rcu_preempt_state);
+}
+
+/*
+ * Initialize preemptable RCU's state structures.
+ */
+static void __init __rcu_init_preempt(void)
+{
+       rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
+}
+
+/*
  * Check for a task exiting while in a preemptable-RCU read-side
  * critical section, clean up if so.  No need to issue warnings,
  * as debug_check_no_locks_held() already does this if lockdep
@@ -439,9 +833,10 @@ void exit_rcu(void)
 /*
  * Tell them what RCU they are running.
  */
-static inline void rcu_bootup_announce(void)
+static void __init rcu_bootup_announce(void)
 {
        printk(KERN_INFO "Hierarchical RCU implementation.\n");
+       rcu_bootup_announce_oddness();
 }
 
 /*
@@ -454,6 +849,16 @@ long rcu_batches_completed(void)
 EXPORT_SYMBOL_GPL(rcu_batches_completed);
 
 /*
+ * Force a quiescent state for RCU, which, because there is no preemptible
+ * RCU, becomes the same as rcu-sched.
+ */
+void rcu_force_quiescent_state(void)
+{
+       rcu_sched_force_quiescent_state();
+}
+EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
+
+/*
  * Because preemptable RCU does not exist, we never have to check for
  * CPUs being in quiescent states.
  */
@@ -461,7 +866,32 @@ static void rcu_preempt_note_context_switch(int cpu)
 {
 }
 
-#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
+/*
+ * Because preemptable RCU does not exist, there are never any preempted
+ * RCU readers.
+ */
+static int rcu_preempted_readers(struct rcu_node *rnp)
+{
+       return 0;
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+
+/* Because preemptible RCU does not exist, no quieting of tasks. */
+static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
+{
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
+
+/*
+ * Because preemptable RCU does not exist, we never have to check for
+ * tasks blocked within RCU read-side critical sections.
+ */
+static void rcu_print_detail_task_stall(struct rcu_state *rsp)
+{
+}
 
 /*
  * Because preemptable RCU does not exist, we never have to check for
@@ -471,7 +901,13 @@ static void rcu_print_task_stall(struct rcu_node *rnp)
 {
 }
 
-#endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
+/*
+ * Because preemptible RCU does not exist, there is no need to suppress
+ * its CPU stall warnings.
+ */
+static void rcu_preempt_stall_reset(void)
+{
+}
 
 /*
  * Because there is no preemptable RCU, there can be no readers blocked,
@@ -483,25 +919,19 @@ static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
        WARN_ON_ONCE(rnp->qsmask);
 }
 
-/*
- * Because preemptable RCU does not exist, there are never any preempted
- * RCU readers.
- */
-static int rcu_preempted_readers(struct rcu_node *rnp)
-{
-       return 0;
-}
-
 #ifdef CONFIG_HOTPLUG_CPU
 
 /*
  * Because preemptable RCU does not exist, it never needs to migrate
- * tasks that were blocked within RCU read-side critical sections.
+ * tasks that were blocked within RCU read-side critical sections, and
+ * such non-existent tasks cannot possibly have been blocking the current
+ * grace period.
  */
-static void rcu_preempt_offline_tasks(struct rcu_state *rsp,
-                                     struct rcu_node *rnp,
-                                     struct rcu_data *rdp)
+static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
+                                    struct rcu_node *rnp,
+                                    struct rcu_data *rdp)
 {
+       return 0;
 }
 
 /*
@@ -518,7 +948,7 @@ static void rcu_preempt_offline_cpu(int cpu)
  * Because preemptable RCU does not exist, it never has any callbacks
  * to check.
  */
-void rcu_preempt_check_callbacks(int cpu)
+static void rcu_preempt_check_callbacks(int cpu)
 {
 }
 
@@ -526,18 +956,33 @@ void rcu_preempt_check_callbacks(int cpu)
  * Because preemptable RCU does not exist, it never has any callbacks
  * to process.
  */
-void rcu_preempt_process_callbacks(void)
+static void rcu_preempt_process_callbacks(void)
 {
 }
 
 /*
- * In classic RCU, call_rcu() is just call_rcu_sched().
+ * Wait for an rcu-preempt grace period, but make it happen quickly.
+ * But because preemptable RCU does not exist, map to rcu-sched.
  */
-void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+void synchronize_rcu_expedited(void)
 {
-       call_rcu_sched(head, func);
+       synchronize_sched_expedited();
 }
-EXPORT_SYMBOL_GPL(call_rcu);
+EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
+
+#ifdef CONFIG_HOTPLUG_CPU
+
+/*
+ * Because preemptable RCU does not exist, there is never any need to
+ * report on tasks preempted in RCU read-side critical sections during
+ * expedited RCU grace periods.
+ */
+static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp)
+{
+       return;
+}
+
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
 
 /*
  * Because preemptable RCU does not exist, it never has any work to do.
@@ -556,6 +1001,16 @@ static int rcu_preempt_needs_cpu(int cpu)
 }
 
 /*
+ * Because preemptable RCU does not exist, rcu_barrier() is just
+ * another name for rcu_barrier_sched().
+ */
+void rcu_barrier(void)
+{
+       rcu_barrier_sched();
+}
+EXPORT_SYMBOL_GPL(rcu_barrier);
+
+/*
  * Because preemptable RCU does not exist, there is no per-CPU
  * data to initialize.
  */
@@ -563,4 +1018,263 @@ static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
 {
 }
 
+/*
+ * Because there is no preemptable RCU, there are no callbacks to move.
+ */
+static void rcu_preempt_send_cbs_to_online(void)
+{
+}
+
+/*
+ * Because preemptable RCU does not exist, it need not be initialized.
+ */
+static void __init __rcu_init_preempt(void)
+{
+}
+
 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
+
+#ifndef CONFIG_SMP
+
+void synchronize_sched_expedited(void)
+{
+       cond_resched();
+}
+EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
+
+#else /* #ifndef CONFIG_SMP */
+
+static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
+static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
+
+static int synchronize_sched_expedited_cpu_stop(void *data)
+{
+       /*
+        * There must be a full memory barrier on each affected CPU
+        * between the time that try_stop_cpus() is called and the
+        * time that it returns.
+        *
+        * In the current initial implementation of cpu_stop, the
+        * above condition is already met when the control reaches
+        * this point and the following smp_mb() is not strictly
+        * necessary.  Do smp_mb() anyway for documentation and
+        * robustness against future implementation changes.
+        */
+       smp_mb(); /* See above comment block. */
+       return 0;
+}
+
+/*
+ * Wait for an rcu-sched grace period to elapse, but use "big hammer"
+ * approach to force grace period to end quickly.  This consumes
+ * significant time on all CPUs, and is thus not recommended for
+ * any sort of common-case code.
+ *
+ * Note that it is illegal to call this function while holding any
+ * lock that is acquired by a CPU-hotplug notifier.  Failing to
+ * observe this restriction will result in deadlock.
+ *
+ * This implementation can be thought of as an application of ticket
+ * locking to RCU, with sync_sched_expedited_started and
+ * sync_sched_expedited_done taking on the roles of the halves
+ * of the ticket-lock word.  Each task atomically increments
+ * sync_sched_expedited_started upon entry, snapshotting the old value,
+ * then attempts to stop all the CPUs.  If this succeeds, then each
+ * CPU will have executed a context switch, resulting in an RCU-sched
+ * grace period.  We are then done, so we use atomic_cmpxchg() to
+ * update sync_sched_expedited_done to match our snapshot -- but
+ * only if someone else has not already advanced past our snapshot.
+ *
+ * On the other hand, if try_stop_cpus() fails, we check the value
+ * of sync_sched_expedited_done.  If it has advanced past our
+ * initial snapshot, then someone else must have forced a grace period
+ * some time after we took our snapshot.  In this case, our work is
+ * done for us, and we can simply return.  Otherwise, we try again,
+ * but keep our initial snapshot for purposes of checking for someone
+ * doing our work for us.
+ *
+ * If we fail too many times in a row, we fall back to synchronize_sched().
+ */
+void synchronize_sched_expedited(void)
+{
+       int firstsnap, s, snap, trycount = 0;
+
+       /* Note that atomic_inc_return() implies full memory barrier. */
+       firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
+       get_online_cpus();
+
+       /*
+        * Each pass through the following loop attempts to force a
+        * context switch on each CPU.
+        */
+       while (try_stop_cpus(cpu_online_mask,
+                            synchronize_sched_expedited_cpu_stop,
+                            NULL) == -EAGAIN) {
+               put_online_cpus();
+
+               /* No joy, try again later.  Or just synchronize_sched(). */
+               if (trycount++ < 10)
+                       udelay(trycount * num_online_cpus());
+               else {
+                       synchronize_sched();
+                       return;
+               }
+
+               /* Check to see if someone else did our work for us. */
+               s = atomic_read(&sync_sched_expedited_done);
+               if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
+                       smp_mb(); /* ensure test happens before caller kfree */
+                       return;
+               }
+
+               /*
+                * Refetching sync_sched_expedited_started allows later
+                * callers to piggyback on our grace period.  We subtract
+                * 1 to get the same token that the last incrementer got.
+                * We retry after they started, so our grace period works
+                * for them, and they started after our first try, so their
+                * grace period works for us.
+                */
+               get_online_cpus();
+               snap = atomic_read(&sync_sched_expedited_started) - 1;
+               smp_mb(); /* ensure read is before try_stop_cpus(). */
+       }
+
+       /*
+        * Everyone up to our most recent fetch is covered by our grace
+        * period.  Update the counter, but only if our work is still
+        * relevant -- which it won't be if someone who started later
+        * than we did beat us to the punch.
+        */
+       do {
+               s = atomic_read(&sync_sched_expedited_done);
+               if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
+                       smp_mb(); /* ensure test happens before caller kfree */
+                       break;
+               }
+       } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
+
+       put_online_cpus();
+}
+EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
+
+#endif /* #else #ifndef CONFIG_SMP */
+
+#if !defined(CONFIG_RCU_FAST_NO_HZ)
+
+/*
+ * Check to see if any future RCU-related work will need to be done
+ * by the current CPU, even if none need be done immediately, returning
+ * 1 if so.  This function is part of the RCU implementation; it is -not-
+ * an exported member of the RCU API.
+ *
+ * Because we have preemptible RCU, just check whether this CPU needs
+ * any flavor of RCU.  Do not chew up lots of CPU cycles with preemption
+ * disabled in a most-likely vain attempt to cause RCU not to need this CPU.
+ */
+int rcu_needs_cpu(int cpu)
+{
+       return rcu_needs_cpu_quick_check(cpu);
+}
+
+/*
+ * Check to see if we need to continue a callback-flush operations to
+ * allow the last CPU to enter dyntick-idle mode.  But fast dyntick-idle
+ * entry is not configured, so we never do need to.
+ */
+static void rcu_needs_cpu_flush(void)
+{
+}
+
+#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
+
+#define RCU_NEEDS_CPU_FLUSHES 5
+static DEFINE_PER_CPU(int, rcu_dyntick_drain);
+static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
+
+/*
+ * Check to see if any future RCU-related work will need to be done
+ * by the current CPU, even if none need be done immediately, returning
+ * 1 if so.  This function is part of the RCU implementation; it is -not-
+ * an exported member of the RCU API.
+ *
+ * Because we are not supporting preemptible RCU, attempt to accelerate
+ * any current grace periods so that RCU no longer needs this CPU, but
+ * only if all other CPUs are already in dynticks-idle mode.  This will
+ * allow the CPU cores to be powered down immediately, as opposed to after
+ * waiting many milliseconds for grace periods to elapse.
+ *
+ * Because it is not legal to invoke rcu_process_callbacks() with irqs
+ * disabled, we do one pass of force_quiescent_state(), then do a
+ * raise_softirq() to cause rcu_process_callbacks() to be invoked later.
+ * The per-cpu rcu_dyntick_drain variable controls the sequencing.
+ */
+int rcu_needs_cpu(int cpu)
+{
+       int c = 0;
+       int snap;
+       int thatcpu;
+
+       /* Check for being in the holdoff period. */
+       if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies)
+               return rcu_needs_cpu_quick_check(cpu);
+
+       /* Don't bother unless we are the last non-dyntick-idle CPU. */
+       for_each_online_cpu(thatcpu) {
+               if (thatcpu == cpu)
+                       continue;
+               snap = atomic_add_return(0, &per_cpu(rcu_dynticks,
+                                                    thatcpu).dynticks);
+               smp_mb(); /* Order sampling of snap with end of grace period. */
+               if ((snap & 0x1) != 0) {
+                       per_cpu(rcu_dyntick_drain, cpu) = 0;
+                       per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1;
+                       return rcu_needs_cpu_quick_check(cpu);
+               }
+       }
+
+       /* Check and update the rcu_dyntick_drain sequencing. */
+       if (per_cpu(rcu_dyntick_drain, cpu) <= 0) {
+               /* First time through, initialize the counter. */
+               per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES;
+       } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) {
+               /* We have hit the limit, so time to give up. */
+               per_cpu(rcu_dyntick_holdoff, cpu) = jiffies;
+               return rcu_needs_cpu_quick_check(cpu);
+       }
+
+       /* Do one step pushing remaining RCU callbacks through. */
+       if (per_cpu(rcu_sched_data, cpu).nxtlist) {
+               rcu_sched_qs(cpu);
+               force_quiescent_state(&rcu_sched_state, 0);
+               c = c || per_cpu(rcu_sched_data, cpu).nxtlist;
+       }
+       if (per_cpu(rcu_bh_data, cpu).nxtlist) {
+               rcu_bh_qs(cpu);
+               force_quiescent_state(&rcu_bh_state, 0);
+               c = c || per_cpu(rcu_bh_data, cpu).nxtlist;
+       }
+
+       /* If RCU callbacks are still pending, RCU still needs this CPU. */
+       if (c)
+               raise_softirq(RCU_SOFTIRQ);
+       return c;
+}
+
+/*
+ * Check to see if we need to continue a callback-flush operations to
+ * allow the last CPU to enter dyntick-idle mode.
+ */
+static void rcu_needs_cpu_flush(void)
+{
+       int cpu = smp_processor_id();
+       unsigned long flags;
+
+       if (per_cpu(rcu_dyntick_drain, cpu) <= 0)
+               return;
+       local_irq_save(flags);
+       (void)rcu_needs_cpu(cpu);
+       local_irq_restore(flags);
+}
+
+#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */