rcu: priority boosting for TREE_PREEMPT_RCU
authorPaul E. McKenney <paul.mckenney@linaro.org>
Mon, 7 Feb 2011 20:47:15 +0000 (12:47 -0800)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Fri, 6 May 2011 06:16:55 +0000 (23:16 -0700)
Add priority boosting for TREE_PREEMPT_RCU, similar to that for
TINY_PREEMPT_RCU.  This is enabled by the default-off RCU_BOOST
kernel parameter.  The priority to which to boost preempted
RCU readers is controlled by the RCU_BOOST_PRIO kernel parameter
(defaulting to real-time priority 1) and the time to wait before
boosting the readers who are blocking a given grace period is
controlled by the RCU_BOOST_DELAY kernel parameter (defaulting to
500 milliseconds).

Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
init/Kconfig
kernel/rcutree.c
kernel/rcutree.h
kernel/rcutree_plugin.h

index d886b1e..2d964fa 100644 (file)
@@ -485,7 +485,7 @@ config TREE_RCU_TRACE
 
 config RCU_BOOST
        bool "Enable RCU priority boosting"
-       depends on RT_MUTEXES && TINY_PREEMPT_RCU
+       depends on RT_MUTEXES && PREEMPT_RCU
        default n
        help
          This option boosts the priority of preempted RCU readers that
index 18e3331..28fd92a 100644 (file)
@@ -81,6 +81,8 @@ DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
 
+static struct rcu_state *rcu_state;
+
 int rcu_scheduler_active __read_mostly;
 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 
@@ -94,7 +96,7 @@ static DEFINE_PER_CPU(char, rcu_cpu_has_work);
 static char rcu_kthreads_spawnable;
 
 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp);
-static void invoke_rcu_kthread(void);
+static void invoke_rcu_cpu_kthread(void);
 
 #define RCU_KTHREAD_PRIO 1     /* RT priority for per-CPU kthreads. */
 
@@ -791,6 +793,7 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
                rnp->completed = rsp->completed;
                rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
                rcu_start_gp_per_cpu(rsp, rnp, rdp);
+               rcu_preempt_boost_start_gp(rnp);
                raw_spin_unlock_irqrestore(&rnp->lock, flags);
                return;
        }
@@ -826,6 +829,7 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
                rnp->completed = rsp->completed;
                if (rnp == rdp->mynode)
                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
+               rcu_preempt_boost_start_gp(rnp);
                raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
        }
 
@@ -882,7 +886,7 @@ rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
                        return;
                }
                rnp->qsmask &= ~mask;
-               if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
+               if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
 
                        /* Other bits still set at this level, so done. */
                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
@@ -1089,8 +1093,11 @@ static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
        t = rnp->node_kthread_task;
        if (t != NULL &&
            rnp->qsmaskinit == 0) {
-               kthread_stop(t);
+               raw_spin_lock_irqsave(&rnp->lock, flags);
                rnp->node_kthread_task = NULL;
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+               kthread_stop(t);
+               rcu_stop_boost_kthread(rnp);
        } else
                rcu_node_kthread_setaffinity(rnp);
 }
@@ -1190,7 +1197,7 @@ static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
 
        /* Re-raise the RCU softirq if there are callbacks remaining. */
        if (cpu_has_callbacks_ready_to_invoke(rdp))
-               invoke_rcu_kthread();
+               invoke_rcu_cpu_kthread();
 }
 
 /*
@@ -1236,7 +1243,7 @@ void rcu_check_callbacks(int cpu, int user)
        }
        rcu_preempt_check_callbacks(cpu);
        if (rcu_pending(cpu))
-               invoke_rcu_kthread();
+               invoke_rcu_cpu_kthread();
 }
 
 #ifdef CONFIG_SMP
@@ -1244,6 +1251,8 @@ void rcu_check_callbacks(int cpu, int user)
 /*
  * Scan the leaf rcu_node structures, processing dyntick state for any that
  * have not yet encountered a quiescent state, using the function specified.
+ * Also initiate boosting for any threads blocked on the root rcu_node.
+ *
  * The caller must have suppressed start of new grace periods.
  */
 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
@@ -1262,6 +1271,7 @@ static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
                        return;
                }
                if (rnp->qsmask == 0) {
+                       rcu_initiate_boost(rnp);
                        raw_spin_unlock_irqrestore(&rnp->lock, flags);
                        continue;
                }
@@ -1280,6 +1290,11 @@ static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
                }
                raw_spin_unlock_irqrestore(&rnp->lock, flags);
        }
+       rnp = rcu_get_root(rsp);
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       if (rnp->qsmask == 0)
+               rcu_initiate_boost(rnp);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
 /*
@@ -1417,7 +1432,7 @@ static void rcu_process_callbacks(void)
  * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
  * cannot disappear out from under us.
  */
-static void invoke_rcu_kthread(void)
+static void invoke_rcu_cpu_kthread(void)
 {
        unsigned long flags;
        wait_queue_head_t *q;
@@ -1436,24 +1451,33 @@ static void invoke_rcu_kthread(void)
 }
 
 /*
+ * Wake up the specified per-rcu_node-structure kthread.
+ * The caller must hold ->lock.
+ */
+static void invoke_rcu_node_kthread(struct rcu_node *rnp)
+{
+       struct task_struct *t;
+
+       t = rnp->node_kthread_task;
+       if (t != NULL)
+               wake_up_process(t);
+}
+
+/*
  * Timer handler to initiate the waking up of per-CPU kthreads that
  * have yielded the CPU due to excess numbers of RCU callbacks.
+ * We wake up the per-rcu_node kthread, which in turn will wake up
+ * the booster kthread.
  */
 static void rcu_cpu_kthread_timer(unsigned long arg)
 {
        unsigned long flags;
-       struct rcu_data *rdp = (struct rcu_data *)arg;
+       struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
        struct rcu_node *rnp = rdp->mynode;
-       struct task_struct *t;
 
        raw_spin_lock_irqsave(&rnp->lock, flags);
        rnp->wakemask |= rdp->grpmask;
-       t = rnp->node_kthread_task;
-       if (t == NULL) {
-               raw_spin_unlock_irqrestore(&rnp->lock, flags);
-               return;
-       }
-       wake_up_process(t);
+       invoke_rcu_node_kthread(rnp);
        raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
@@ -1463,13 +1487,12 @@ static void rcu_cpu_kthread_timer(unsigned long arg)
  * remain preempted.  Either way, we restore our real-time priority
  * before returning.
  */
-static void rcu_yield(int cpu)
+static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
 {
-       struct rcu_data *rdp = per_cpu_ptr(rcu_sched_state.rda, cpu);
        struct sched_param sp;
        struct timer_list yield_timer;
 
-       setup_timer_on_stack(&yield_timer, rcu_cpu_kthread_timer, (unsigned long)rdp);
+       setup_timer_on_stack(&yield_timer, f, arg);
        mod_timer(&yield_timer, jiffies + 2);
        sp.sched_priority = 0;
        sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
@@ -1540,7 +1563,7 @@ static int rcu_cpu_kthread(void *arg)
                else
                        spincnt = 0;
                if (spincnt > 10) {
-                       rcu_yield(cpu);
+                       rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
                        spincnt = 0;
                }
        }
@@ -1597,6 +1620,7 @@ static int rcu_node_kthread(void *arg)
                raw_spin_lock_irqsave(&rnp->lock, flags);
                mask = rnp->wakemask;
                rnp->wakemask = 0;
+               rcu_initiate_boost(rnp);
                raw_spin_unlock_irqrestore(&rnp->lock, flags);
                for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
                        if ((mask & 0x1) == 0)
@@ -1618,7 +1642,8 @@ static int rcu_node_kthread(void *arg)
 
 /*
  * Set the per-rcu_node kthread's affinity to cover all CPUs that are
- * served by the rcu_node in question.
+ * served by the rcu_node in question.  The CPU hotplug lock is still
+ * held, so the value of rnp->qsmaskinit will be stable.
  */
 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp)
 {
@@ -1626,8 +1651,7 @@ static void rcu_node_kthread_setaffinity(struct rcu_node *rnp)
        int cpu;
        unsigned long mask = rnp->qsmaskinit;
 
-       if (rnp->node_kthread_task == NULL ||
-           rnp->qsmaskinit == 0)
+       if (rnp->node_kthread_task == NULL || mask == 0)
                return;
        if (!alloc_cpumask_var(&cm, GFP_KERNEL))
                return;
@@ -1636,31 +1660,40 @@ static void rcu_node_kthread_setaffinity(struct rcu_node *rnp)
                if (mask & 0x1)
                        cpumask_set_cpu(cpu, cm);
        set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
+       rcu_boost_kthread_setaffinity(rnp, cm);
        free_cpumask_var(cm);
 }
 
 /*
  * Spawn a per-rcu_node kthread, setting priority and affinity.
+ * Called during boot before online/offline can happen, or, if
+ * during runtime, with the main CPU-hotplug locks held.  So only
+ * one of these can be executing at a time.
  */
 static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
                                                struct rcu_node *rnp)
 {
+       unsigned long flags;
        int rnp_index = rnp - &rsp->node[0];
        struct sched_param sp;
        struct task_struct *t;
 
        if (!rcu_kthreads_spawnable ||
-           rnp->qsmaskinit == 0 ||
-           rnp->node_kthread_task != NULL)
+           rnp->qsmaskinit == 0)
                return 0;
-       t = kthread_create(rcu_node_kthread, (void *)rnp, "rcun%d", rnp_index);
-       if (IS_ERR(t))
-               return PTR_ERR(t);
-       rnp->node_kthread_task = t;
-       wake_up_process(t);
-       sp.sched_priority = 99;
-       sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
-       return 0;
+       if (rnp->node_kthread_task == NULL) {
+               t = kthread_create(rcu_node_kthread, (void *)rnp,
+                                  "rcun%d", rnp_index);
+               if (IS_ERR(t))
+                       return PTR_ERR(t);
+               raw_spin_lock_irqsave(&rnp->lock, flags);
+               rnp->node_kthread_task = t;
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+               wake_up_process(t);
+               sp.sched_priority = 99;
+               sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
+       }
+       return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
 }
 
 /*
@@ -1678,10 +1711,16 @@ static int __init rcu_spawn_kthreads(void)
                if (cpu_online(cpu))
                        (void)rcu_spawn_one_cpu_kthread(cpu);
        }
-       rcu_for_each_leaf_node(&rcu_sched_state, rnp) {
-               init_waitqueue_head(&rnp->node_wq);
-               (void)rcu_spawn_one_node_kthread(&rcu_sched_state, rnp);
-       }
+       rnp = rcu_get_root(rcu_state);
+       init_waitqueue_head(&rnp->node_wq);
+       rcu_init_boost_waitqueue(rnp);
+       (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
+       if (NUM_RCU_NODES > 1)
+               rcu_for_each_leaf_node(rcu_state, rnp) {
+                       init_waitqueue_head(&rnp->node_wq);
+                       rcu_init_boost_waitqueue(rnp);
+                       (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
+               }
        return 0;
 }
 early_initcall(rcu_spawn_kthreads);
@@ -2087,14 +2126,14 @@ static void __cpuinit rcu_online_cpu(int cpu)
 
 static void __cpuinit rcu_online_kthreads(int cpu)
 {
-       struct rcu_data *rdp = per_cpu_ptr(rcu_sched_state.rda, cpu);
+       struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
        struct rcu_node *rnp = rdp->mynode;
 
        /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
        if (rcu_kthreads_spawnable) {
                (void)rcu_spawn_one_cpu_kthread(cpu);
                if (rnp->node_kthread_task == NULL)
-                       (void)rcu_spawn_one_node_kthread(&rcu_sched_state, rnp);
+                       (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
        }
 }
 
@@ -2105,7 +2144,7 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
                                    unsigned long action, void *hcpu)
 {
        long cpu = (long)hcpu;
-       struct rcu_data *rdp = per_cpu_ptr(rcu_sched_state.rda, cpu);
+       struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
        struct rcu_node *rnp = rdp->mynode;
 
        switch (action) {
index c021380..8db0cdc 100644 (file)
@@ -135,6 +135,24 @@ struct rcu_node {
                                /*  if there is no such task.  If there */
                                /*  is no current expedited grace period, */
                                /*  then there can cannot be any such task. */
+#ifdef CONFIG_RCU_BOOST
+       struct list_head *boost_tasks;
+                               /* Pointer to first task that needs to be */
+                               /*  priority boosted, or NULL if no priority */
+                               /*  boosting is needed for this rcu_node */
+                               /*  structure.  If there are no tasks */
+                               /*  queued on this rcu_node structure that */
+                               /*  are blocking the current grace period, */
+                               /*  there can be no such task. */
+       unsigned long boost_time;
+                               /* When to start boosting (jiffies). */
+       struct task_struct *boost_kthread_task;
+                               /* kthread that takes care of priority */
+                               /*  boosting for this rcu_node structure. */
+       wait_queue_head_t boost_wq;
+                               /* Wait queue on which to park the boost */
+                               /*  kthread. */
+#endif /* #ifdef CONFIG_RCU_BOOST */
        struct task_struct *node_kthread_task;
                                /* kthread that takes care of this rcu_node */
                                /*  structure, for example, awakening the */
@@ -365,7 +383,7 @@ DECLARE_PER_CPU(struct rcu_data, rcu_preempt_data);
 static void rcu_bootup_announce(void);
 long rcu_batches_completed(void);
 static void rcu_preempt_note_context_switch(int cpu);
-static int rcu_preempted_readers(struct rcu_node *rnp);
+static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
 #ifdef CONFIG_HOTPLUG_CPU
 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp,
                                      unsigned long flags);
@@ -392,5 +410,16 @@ static void __cpuinit rcu_preempt_init_percpu_data(int cpu);
 static void rcu_preempt_send_cbs_to_online(void);
 static void __init __rcu_init_preempt(void);
 static void rcu_needs_cpu_flush(void);
+static void __init rcu_init_boost_waitqueue(struct rcu_node *rnp);
+static void rcu_initiate_boost(struct rcu_node *rnp);
+static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
+                                         cpumask_var_t cm);
+static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
+static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
+                                                struct rcu_node *rnp,
+                                                int rnp_index);
+#ifdef CONFIG_HOTPLUG_CPU
+static void rcu_stop_boost_kthread(struct rcu_node *rnp);
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
 
 #endif /* #ifndef RCU_TREE_NONCORE */
index b9bd69a..5964f82 100644 (file)
@@ -66,6 +66,7 @@ static void __init rcu_bootup_announce_oddness(void)
 
 struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state);
 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
+static struct rcu_state *rcu_state = &rcu_preempt_state;
 
 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
 
@@ -179,6 +180,10 @@ static void rcu_preempt_note_context_switch(int cpu)
                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;
+#ifdef CONFIG_RCU_BOOST
+                       if (rnp->boost_tasks != NULL)
+                               rnp->boost_tasks = rnp->gp_tasks;
+#endif /* #ifdef CONFIG_RCU_BOOST */
                } else {
                        list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
                        if (rnp->qsmask & rdp->grpmask)
@@ -218,7 +223,7 @@ EXPORT_SYMBOL_GPL(__rcu_read_lock);
  * 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)
+static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 {
        return rnp->gp_tasks != NULL;
 }
@@ -236,7 +241,7 @@ static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
        unsigned long mask;
        struct rcu_node *rnp_p;
 
-       if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
+       if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
                raw_spin_unlock_irqrestore(&rnp->lock, flags);
                return;  /* Still need more quiescent states! */
        }
@@ -325,7 +330,7 @@ static void rcu_read_unlock_special(struct task_struct *t)
                                break;
                        raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
                }
-               empty = !rcu_preempted_readers(rnp);
+               empty = !rcu_preempt_blocked_readers_cgp(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);
@@ -334,6 +339,10 @@ static void rcu_read_unlock_special(struct task_struct *t)
                        rnp->gp_tasks = np;
                if (&t->rcu_node_entry == rnp->exp_tasks)
                        rnp->exp_tasks = np;
+#ifdef CONFIG_RCU_BOOST
+               if (&t->rcu_node_entry == rnp->boost_tasks)
+                       rnp->boost_tasks = np;
+#endif /* #ifdef CONFIG_RCU_BOOST */
                t->rcu_blocked_node = NULL;
 
                /*
@@ -346,6 +355,15 @@ static void rcu_read_unlock_special(struct task_struct *t)
                else
                        rcu_report_unblock_qs_rnp(rnp, flags);
 
+#ifdef CONFIG_RCU_BOOST
+               /* Unboost if we were boosted. */
+               if (special & RCU_READ_UNLOCK_BOOSTED) {
+                       t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BOOSTED;
+                       rt_mutex_unlock(t->rcu_boost_mutex);
+                       t->rcu_boost_mutex = NULL;
+               }
+#endif /* #ifdef CONFIG_RCU_BOOST */
+
                /*
                 * If this was the last task on the expedited lists,
                 * then we need to report up the rcu_node hierarchy.
@@ -391,7 +409,7 @@ static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
        unsigned long flags;
        struct task_struct *t;
 
-       if (!rcu_preempted_readers(rnp))
+       if (!rcu_preempt_blocked_readers_cgp(rnp))
                return;
        raw_spin_lock_irqsave(&rnp->lock, flags);
        t = list_entry(rnp->gp_tasks,
@@ -430,7 +448,7 @@ static void rcu_print_task_stall(struct rcu_node *rnp)
 {
        struct task_struct *t;
 
-       if (!rcu_preempted_readers(rnp))
+       if (!rcu_preempt_blocked_readers_cgp(rnp))
                return;
        t = list_entry(rnp->gp_tasks,
                       struct task_struct, rcu_node_entry);
@@ -460,7 +478,7 @@ static void rcu_preempt_stall_reset(void)
  */
 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
 {
-       WARN_ON_ONCE(rcu_preempted_readers(rnp));
+       WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
        if (!list_empty(&rnp->blkd_tasks))
                rnp->gp_tasks = rnp->blkd_tasks.next;
        WARN_ON_ONCE(rnp->qsmask);
@@ -509,7 +527,7 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
         * absolutely necessary, but this is a good performance/complexity
         * tradeoff.
         */
-       if (rcu_preempted_readers(rnp))
+       if (rcu_preempt_blocked_readers_cgp(rnp))
                retval |= RCU_OFL_TASKS_NORM_GP;
        if (rcu_preempted_readers_exp(rnp))
                retval |= RCU_OFL_TASKS_EXP_GP;
@@ -525,8 +543,22 @@ static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
                        rnp_root->gp_tasks = rnp->gp_tasks;
                if (&t->rcu_node_entry == rnp->exp_tasks)
                        rnp_root->exp_tasks = rnp->exp_tasks;
+#ifdef CONFIG_RCU_BOOST
+               if (&t->rcu_node_entry == rnp->boost_tasks)
+                       rnp_root->boost_tasks = rnp->boost_tasks;
+#endif /* #ifdef CONFIG_RCU_BOOST */
                raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
        }
+
+#ifdef CONFIG_RCU_BOOST
+       /* In case root is being boosted and leaf is not. */
+       raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
+       if (rnp_root->boost_tasks != NULL &&
+           rnp_root->boost_tasks != rnp_root->gp_tasks)
+               rnp_root->boost_tasks = rnp_root->gp_tasks;
+       raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
+#endif /* #ifdef CONFIG_RCU_BOOST */
+
        rnp->gp_tasks = NULL;
        rnp->exp_tasks = NULL;
        return retval;
@@ -684,6 +716,7 @@ sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
        raw_spin_lock(&rnp->lock); /* irqs already disabled */
        if (!list_empty(&rnp->blkd_tasks)) {
                rnp->exp_tasks = rnp->blkd_tasks.next;
+               rcu_initiate_boost(rnp);
                must_wait = 1;
        }
        raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
@@ -830,6 +863,8 @@ void exit_rcu(void)
 
 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
 
+static struct rcu_state *rcu_state = &rcu_sched_state;
+
 /*
  * Tell them what RCU they are running.
  */
@@ -870,7 +905,7 @@ static void rcu_preempt_note_context_switch(int cpu)
  * Because preemptable RCU does not exist, there are never any preempted
  * RCU readers.
  */
-static int rcu_preempted_readers(struct rcu_node *rnp)
+static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
 {
        return 0;
 }
@@ -1034,6 +1069,263 @@ static void __init __rcu_init_preempt(void)
 
 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
 
+#ifdef CONFIG_RCU_BOOST
+
+#include "rtmutex_common.h"
+
+/*
+ * Carry out RCU priority boosting on the task indicated by ->exp_tasks
+ * or ->boost_tasks, advancing the pointer to the next task in the
+ * ->blkd_tasks list.
+ *
+ * Note that irqs must be enabled: boosting the task can block.
+ * Returns 1 if there are more tasks needing to be boosted.
+ */
+static int rcu_boost(struct rcu_node *rnp)
+{
+       unsigned long flags;
+       struct rt_mutex mtx;
+       struct task_struct *t;
+       struct list_head *tb;
+
+       if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
+               return 0;  /* Nothing left to boost. */
+
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+
+       /*
+        * Recheck under the lock: all tasks in need of boosting
+        * might exit their RCU read-side critical sections on their own.
+        */
+       if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+               return 0;
+       }
+
+       /*
+        * Preferentially boost tasks blocking expedited grace periods.
+        * This cannot starve the normal grace periods because a second
+        * expedited grace period must boost all blocked tasks, including
+        * those blocking the pre-existing normal grace period.
+        */
+       if (rnp->exp_tasks != NULL)
+               tb = rnp->exp_tasks;
+       else
+               tb = rnp->boost_tasks;
+
+       /*
+        * We boost task t by manufacturing an rt_mutex that appears to
+        * be held by task t.  We leave a pointer to that rt_mutex where
+        * task t can find it, and task t will release the mutex when it
+        * exits its outermost RCU read-side critical section.  Then
+        * simply acquiring this artificial rt_mutex will boost task
+        * t's priority.  (Thanks to tglx for suggesting this approach!)
+        *
+        * Note that task t must acquire rnp->lock to remove itself from
+        * the ->blkd_tasks list, which it will do from exit() if from
+        * nowhere else.  We therefore are guaranteed that task t will
+        * stay around at least until we drop rnp->lock.  Note that
+        * rnp->lock also resolves races between our priority boosting
+        * and task t's exiting its outermost RCU read-side critical
+        * section.
+        */
+       t = container_of(tb, struct task_struct, rcu_node_entry);
+       rt_mutex_init_proxy_locked(&mtx, t);
+       t->rcu_boost_mutex = &mtx;
+       t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BOOSTED;
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+       rt_mutex_lock(&mtx);  /* Side effect: boosts task t's priority. */
+       rt_mutex_unlock(&mtx);  /* Keep lockdep happy. */
+
+       return rnp->exp_tasks != NULL || rnp->boost_tasks != NULL;
+}
+
+/*
+ * Timer handler to initiate waking up of boost kthreads that
+ * have yielded the CPU due to excessive numbers of tasks to
+ * boost.  We wake up the per-rcu_node kthread, which in turn
+ * will wake up the booster kthread.
+ */
+static void rcu_boost_kthread_timer(unsigned long arg)
+{
+       unsigned long flags;
+       struct rcu_node *rnp = (struct rcu_node *)arg;
+
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       invoke_rcu_node_kthread(rnp);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+/*
+ * Priority-boosting kthread.  One per leaf rcu_node and one for the
+ * root rcu_node.
+ */
+static int rcu_boost_kthread(void *arg)
+{
+       struct rcu_node *rnp = (struct rcu_node *)arg;
+       int spincnt = 0;
+       int more2boost;
+
+       for (;;) {
+               wait_event_interruptible(rnp->boost_wq, rnp->boost_tasks ||
+                                                       rnp->exp_tasks ||
+                                                       kthread_should_stop());
+               if (kthread_should_stop())
+                       break;
+               more2boost = rcu_boost(rnp);
+               if (more2boost)
+                       spincnt++;
+               else
+                       spincnt = 0;
+               if (spincnt > 10) {
+                       rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp);
+                       spincnt = 0;
+               }
+       }
+       return 0;
+}
+
+/*
+ * Check to see if it is time to start boosting RCU readers that are
+ * blocking the current grace period, and, if so, tell the per-rcu_node
+ * kthread to start boosting them.  If there is an expedited grace
+ * period in progress, it is always time to boost.
+ *
+ * The caller must hold rnp->lock.
+ */
+static void rcu_initiate_boost(struct rcu_node *rnp)
+{
+       struct task_struct *t;
+
+       if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL)
+               return;
+       if (rnp->exp_tasks != NULL ||
+           (rnp->gp_tasks != NULL &&
+            rnp->boost_tasks == NULL &&
+            rnp->qsmask == 0 &&
+            ULONG_CMP_GE(jiffies, rnp->boost_time))) {
+               if (rnp->exp_tasks == NULL)
+                       rnp->boost_tasks = rnp->gp_tasks;
+               t = rnp->boost_kthread_task;
+               if (t != NULL)
+                       wake_up_process(t);
+       }
+}
+
+static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
+                                         cpumask_var_t cm)
+{
+       unsigned long flags;
+       struct task_struct *t;
+
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       t = rnp->boost_kthread_task;
+       if (t != NULL)
+               set_cpus_allowed_ptr(rnp->boost_kthread_task, cm);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
+
+/*
+ * Do priority-boost accounting for the start of a new grace period.
+ */
+static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
+{
+       rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
+}
+
+/*
+ * Initialize the RCU-boost waitqueue.
+ */
+static void __init rcu_init_boost_waitqueue(struct rcu_node *rnp)
+{
+       init_waitqueue_head(&rnp->boost_wq);
+}
+
+/*
+ * Create an RCU-boost kthread for the specified node if one does not
+ * already exist.  We only create this kthread for preemptible RCU.
+ * Returns zero if all is well, a negated errno otherwise.
+ */
+static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
+                                                struct rcu_node *rnp,
+                                                int rnp_index)
+{
+       unsigned long flags;
+       struct sched_param sp;
+       struct task_struct *t;
+
+       if (&rcu_preempt_state != rsp)
+               return 0;
+       if (rnp->boost_kthread_task != NULL)
+               return 0;
+       t = kthread_create(rcu_boost_kthread, (void *)rnp,
+                          "rcub%d", rnp_index);
+       if (IS_ERR(t))
+               return PTR_ERR(t);
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       rnp->boost_kthread_task = t;
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+       wake_up_process(t);
+       sp.sched_priority = RCU_KTHREAD_PRIO;
+       sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
+       return 0;
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+
+static void rcu_stop_boost_kthread(struct rcu_node *rnp)
+{
+       unsigned long flags;
+       struct task_struct *t;
+
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       t = rnp->boost_kthread_task;
+       rnp->boost_kthread_task = NULL;
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+       if (t != NULL)
+               kthread_stop(t);
+}
+
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
+
+#else /* #ifdef CONFIG_RCU_BOOST */
+
+static void rcu_initiate_boost(struct rcu_node *rnp)
+{
+}
+
+static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
+                                         cpumask_var_t cm)
+{
+}
+
+static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
+{
+}
+
+static void __init rcu_init_boost_waitqueue(struct rcu_node *rnp)
+{
+}
+
+static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
+                                                struct rcu_node *rnp,
+                                                int rnp_index)
+{
+       return 0;
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+
+static void rcu_stop_boost_kthread(struct rcu_node *rnp)
+{
+}
+
+#endif /* #ifdef CONFIG_HOTPLUG_CPU */
+
+#endif /* #else #ifdef CONFIG_RCU_BOOST */
+
 #ifndef CONFIG_SMP
 
 void synchronize_sched_expedited(void)
@@ -1206,8 +1498,8 @@ static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
  *
  * Because it is not legal to invoke rcu_process_callbacks() with irqs
  * disabled, we do one pass of force_quiescent_state(), then do a
- * invoke_rcu_kthread() to cause rcu_process_callbacks() to be invoked later.
- * The per-cpu rcu_dyntick_drain variable controls the sequencing.
+ * invoke_rcu_cpu_kthread() 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)
 {
@@ -1257,7 +1549,7 @@ int rcu_needs_cpu(int cpu)
 
        /* If RCU callbacks are still pending, RCU still needs this CPU. */
        if (c)
-               invoke_rcu_kthread();
+               invoke_rcu_cpu_kthread();
        return c;
 }