clockevent: Don't remove broadcast device when cpu is dead
[linux-2.6.git] / kernel / rcutree_plugin.h
index ef2a58c..37fbccd 100644 (file)
  *        Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  */
 
+#include <linux/delay.h>
 
 #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");
@@ -67,7 +70,7 @@ EXPORT_SYMBOL_GPL(rcu_batches_completed);
 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;
 }
@@ -157,14 +160,58 @@ EXPORT_SYMBOL_GPL(__rcu_read_lock);
  */
 static int rcu_preempted_readers(struct rcu_node *rnp)
 {
-       return !list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]);
+       int phase = rnp->gpnum & 0x1;
+
+       return !list_empty(&rnp->blocked_tasks[phase]) ||
+              !list_empty(&rnp->blocked_tasks[phase + 2]);
+}
+
+/*
+ * 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)) {
+               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;
+       spin_unlock(&rnp->lock);        /* irqs remain disabled. */
+       spin_lock(&rnp_p->lock);        /* irqs already disabled. */
+       rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
 }
 
+/*
+ * 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 rcu_node *rnp;
        int special;
 
@@ -207,36 +254,30 @@ static void rcu_read_unlock_special(struct task_struct *t)
                        spin_unlock(&rnp->lock);  /* irqs remain disabled. */
                }
                empty = !rcu_preempted_readers(rnp);
+               empty_exp = !rcu_preempted_readers_exp(rnp);
+               smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
                list_del_init(&t->rcu_node_entry);
                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 &&
-                   !rcu_preempted_readers(rnp)) {
-                       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;
+               if (empty)
                        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);
+               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);
 }
 
 /*
@@ -303,6 +344,8 @@ static void rcu_preempt_check_blocked_tasks(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.
@@ -316,7 +359,7 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
        int i;
        struct list_head *lp;
        struct list_head *lp_root;
-       int retval = rcu_preempted_readers(rnp);
+       int retval = 0;
        struct rcu_node *rnp_root = rcu_get_root(rsp);
        struct task_struct *tp;
 
@@ -326,7 +369,9 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
        }
        WARN_ON_ONCE(rnp != rdp->mynode &&
                     (!list_empty(&rnp->blocked_tasks[0]) ||
-                     !list_empty(&rnp->blocked_tasks[1])));
+                     !list_empty(&rnp->blocked_tasks[1]) ||
+                     !list_empty(&rnp->blocked_tasks[2]) ||
+                     !list_empty(&rnp->blocked_tasks[3])));
 
        /*
         * Move tasks up to root rcu_node.  Rely on the fact that the
@@ -334,7 +379,11 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
         * rcu_nodes in terms of gp_num value.  This fact allows us to
         * move the blocked_tasks[] array directly, element by element.
         */
-       for (i = 0; i < 2; i++) {
+       if (rcu_preempted_readers(rnp))
+               retval |= RCU_OFL_TASKS_NORM_GP;
+       if (rcu_preempted_readers_exp(rnp))
+               retval |= RCU_OFL_TASKS_EXP_GP;
+       for (i = 0; i < 4; i++) {
                lp = &rnp->blocked_tasks[i];
                lp_root = &rnp_root->blocked_tasks[i];
                while (!list_empty(lp)) {
@@ -346,7 +395,6 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
                        spin_unlock(&rnp_root->lock); /* irqs remain disabled */
                }
        }
-
        return retval;
 }
 
@@ -398,14 +446,183 @@ 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.  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_completion(&rcu.completion);
+       /* Will wake me after RCU finished. */
+       call_rcu(&rcu.head, wakeme_after_rcu);
+       /* Wait for it. */
+       wait_for_completion(&rcu.completion);
+}
+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);
+
 /*
- * Wait for an rcu-preempt grace period.  We are supposed to expedite the
- * grace period, but this is the crude slow compatability hack, so just
- * invoke synchronize_rcu().
+ * 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 !list_empty(&rnp->blocked_tasks[2]) ||
+              !list_empty(&rnp->blocked_tasks[3]);
+}
+
+/*
+ * 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;
+
+       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;
+               spin_unlock(&rnp->lock); /* irqs remain disabled */
+               rnp = rnp->parent;
+               spin_lock(&rnp->lock); /* irqs already disabled */
+               rnp->expmask &= ~mask;
+       }
+       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;
+
+       spin_lock(&rnp->lock); /* irqs already disabled */
+       list_splice_init(&rnp->blocked_tasks[0], &rnp->blocked_tasks[2]);
+       list_splice_init(&rnp->blocked_tasks[1], &rnp->blocked_tasks[3]);
+       must_wait = rcu_preempted_readers_exp(rnp);
+       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 ->blocked_tasks[] lists, move all entries from the first set of
+ * ->blocked_tasks[] lists to the second set, and finally wait for this
+ * second set to drain.
  */
 void synchronize_rcu_expedited(void)
 {
-       synchronize_rcu();
+       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 blocked_tasks[]. */
+       synchronize_sched_expedited();
+
+       spin_lock_irqsave(&rsp->onofflock, flags);
+
+       /* Initialize ->expmask for all non-leaf rcu_node structures. */
+       rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
+               spin_lock(&rnp->lock); /* irqs already disabled. */
+               rnp->expmask = rnp->qsmaskinit;
+               spin_unlock(&rnp->lock); /* irqs remain disabled. */
+       }
+
+       /* Snapshot current state of ->blocked_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));
+
+       spin_unlock_irqrestore(&rsp->onofflock, flags);
+
+       /* Wait for snapshotted ->blocked_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);
 
@@ -481,7 +698,7 @@ 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");
 }
@@ -512,6 +729,16 @@ 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)
+{
+       spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
+
 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
 
 /*
@@ -594,6 +821,20 @@ void synchronize_rcu_expedited(void)
 }
 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.
  */