Merge branch 'linus' into cpus4096
[pandora-kernel.git] / kernel / rcupreempt.c
index 41d275a..2782793 100644 (file)
 #include <asm/atomic.h>
 #include <linux/bitops.h>
 #include <linux/module.h>
+#include <linux/kthread.h>
 #include <linux/completion.h>
 #include <linux/moduleparam.h>
 #include <linux/percpu.h>
 #include <linux/notifier.h>
-#include <linux/rcupdate.h>
 #include <linux/cpu.h>
 #include <linux/random.h>
 #include <linux/delay.h>
@@ -82,14 +82,18 @@ struct rcu_data {
        spinlock_t      lock;           /* Protect rcu_data fields. */
        long            completed;      /* Number of last completed batch. */
        int             waitlistcount;
-       struct tasklet_struct rcu_tasklet;
        struct rcu_head *nextlist;
        struct rcu_head **nexttail;
        struct rcu_head *waitlist[GP_STAGES];
        struct rcu_head **waittail[GP_STAGES];
-       struct rcu_head *donelist;
+       struct rcu_head *donelist;      /* from waitlist & waitschedlist */
        struct rcu_head **donetail;
        long rcu_flipctr[2];
+       struct rcu_head *nextschedlist;
+       struct rcu_head **nextschedtail;
+       struct rcu_head *waitschedlist;
+       struct rcu_head **waitschedtail;
+       int rcu_sched_sleeping;
 #ifdef CONFIG_RCU_TRACE
        struct rcupreempt_trace trace;
 #endif /* #ifdef CONFIG_RCU_TRACE */
@@ -131,11 +135,24 @@ enum rcu_try_flip_states {
        rcu_try_flip_waitmb_state,
 };
 
+/*
+ * States for rcu_ctrlblk.rcu_sched_sleep.
+ */
+
+enum rcu_sched_sleep_states {
+       rcu_sched_not_sleeping, /* Not sleeping, callbacks need GP.  */
+       rcu_sched_sleep_prep,   /* Thinking of sleeping, rechecking. */
+       rcu_sched_sleeping,     /* Sleeping, awaken if GP needed. */
+};
+
 struct rcu_ctrlblk {
        spinlock_t      fliplock;       /* Protect state-machine transitions. */
        long            completed;      /* Number of last completed batch. */
        enum rcu_try_flip_states rcu_try_flip_state; /* The current state of
                                                        the rcu state machine */
+       spinlock_t      schedlock;      /* Protect rcu_sched sleep state. */
+       enum rcu_sched_sleep_states sched_sleep; /* rcu_sched state. */
+       wait_queue_head_t sched_wq;     /* Place for rcu_sched to sleep. */
 };
 
 static DEFINE_PER_CPU(struct rcu_data, rcu_data);
@@ -143,8 +160,12 @@ static struct rcu_ctrlblk rcu_ctrlblk = {
        .fliplock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.fliplock),
        .completed = 0,
        .rcu_try_flip_state = rcu_try_flip_idle_state,
+       .schedlock = __SPIN_LOCK_UNLOCKED(rcu_ctrlblk.schedlock),
+       .sched_sleep = rcu_sched_not_sleeping,
+       .sched_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rcu_ctrlblk.sched_wq),
 };
 
+static struct task_struct *rcu_sched_grace_period_task;
 
 #ifdef CONFIG_RCU_TRACE
 static char *rcu_try_flip_state_names[] =
@@ -207,6 +228,8 @@ static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_mb_flag_values, rcu_mb_flag)
  */
 #define RCU_TRACE_RDP(f, rdp) RCU_TRACE(f, &((rdp)->trace));
 
+#define RCU_SCHED_BATCH_TIME (HZ / 50)
+
 /*
  * Return the number of RCU batches processed thus far.  Useful
  * for debug and statistics.
@@ -411,32 +434,34 @@ static void __rcu_advance_callbacks(struct rcu_data *rdp)
        }
 }
 
-#ifdef CONFIG_NO_HZ
+DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_dyntick_sched, rcu_dyntick_sched) = {
+       .dynticks = 1,
+};
 
-DEFINE_PER_CPU(long, dynticks_progress_counter) = 1;
-static DEFINE_PER_CPU(long, rcu_dyntick_snapshot);
+#ifdef CONFIG_NO_HZ
 static DEFINE_PER_CPU(int, rcu_update_flag);
 
 /**
  * rcu_irq_enter - Called from Hard irq handlers and NMI/SMI.
  *
  * If the CPU was idle with dynamic ticks active, this updates the
- * dynticks_progress_counter to let the RCU handling know that the
+ * rcu_dyntick_sched.dynticks to let the RCU handling know that the
  * CPU is active.
  */
 void rcu_irq_enter(void)
 {
        int cpu = smp_processor_id();
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
 
        if (per_cpu(rcu_update_flag, cpu))
                per_cpu(rcu_update_flag, cpu)++;
 
        /*
         * Only update if we are coming from a stopped ticks mode
-        * (dynticks_progress_counter is even).
+        * (rcu_dyntick_sched.dynticks is even).
         */
        if (!in_interrupt() &&
-           (per_cpu(dynticks_progress_counter, cpu) & 0x1) == 0) {
+           (rdssp->dynticks & 0x1) == 0) {
                /*
                 * The following might seem like we could have a race
                 * with NMI/SMIs. But this really isn't a problem.
@@ -459,12 +484,12 @@ void rcu_irq_enter(void)
                 * RCU read-side critical sections on this CPU would
                 * have already completed.
                 */
-               per_cpu(dynticks_progress_counter, cpu)++;
+               rdssp->dynticks++;
                /*
                 * The following memory barrier ensures that any
                 * rcu_read_lock() primitives in the irq handler
                 * are seen by other CPUs to follow the above
-                * increment to dynticks_progress_counter. This is
+                * increment to rcu_dyntick_sched.dynticks. This is
                 * required in order for other CPUs to correctly
                 * determine when it is safe to advance the RCU
                 * grace-period state machine.
@@ -472,7 +497,7 @@ void rcu_irq_enter(void)
                smp_mb(); /* see above block comment. */
                /*
                 * Since we can't determine the dynamic tick mode from
-                * the dynticks_progress_counter after this routine,
+                * the rcu_dyntick_sched.dynticks after this routine,
                 * we use a second flag to acknowledge that we came
                 * from an idle state with ticks stopped.
                 */
@@ -480,7 +505,7 @@ void rcu_irq_enter(void)
                /*
                 * If we take an NMI/SMI now, they will also increment
                 * the rcu_update_flag, and will not update the
-                * dynticks_progress_counter on exit. That is for
+                * rcu_dyntick_sched.dynticks on exit. That is for
                 * this IRQ to do.
                 */
        }
@@ -490,12 +515,13 @@ void rcu_irq_enter(void)
  * rcu_irq_exit - Called from exiting Hard irq context.
  *
  * If the CPU was idle with dynamic ticks active, update the
- * dynticks_progress_counter to put let the RCU handling be
+ * rcu_dyntick_sched.dynticks to put let the RCU handling be
  * aware that the CPU is going back to idle with no ticks.
  */
 void rcu_irq_exit(void)
 {
        int cpu = smp_processor_id();
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
 
        /*
         * rcu_update_flag is set if we interrupted the CPU
@@ -503,7 +529,7 @@ void rcu_irq_exit(void)
         * Once this occurs, we keep track of interrupt nesting
         * because a NMI/SMI could also come in, and we still
         * only want the IRQ that started the increment of the
-        * dynticks_progress_counter to be the one that modifies
+        * rcu_dyntick_sched.dynticks to be the one that modifies
         * it on exit.
         */
        if (per_cpu(rcu_update_flag, cpu)) {
@@ -515,28 +541,29 @@ void rcu_irq_exit(void)
 
                /*
                 * If an NMI/SMI happens now we are still
-                * protected by the dynticks_progress_counter being odd.
+                * protected by the rcu_dyntick_sched.dynticks being odd.
                 */
 
                /*
                 * The following memory barrier ensures that any
                 * rcu_read_unlock() primitives in the irq handler
                 * are seen by other CPUs to preceed the following
-                * increment to dynticks_progress_counter. This
+                * increment to rcu_dyntick_sched.dynticks. This
                 * is required in order for other CPUs to determine
                 * when it is safe to advance the RCU grace-period
                 * state machine.
                 */
                smp_mb(); /* see above block comment. */
-               per_cpu(dynticks_progress_counter, cpu)++;
-               WARN_ON(per_cpu(dynticks_progress_counter, cpu) & 0x1);
+               rdssp->dynticks++;
+               WARN_ON(rdssp->dynticks & 0x1);
        }
 }
 
 static void dyntick_save_progress_counter(int cpu)
 {
-       per_cpu(rcu_dyntick_snapshot, cpu) =
-               per_cpu(dynticks_progress_counter, cpu);
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+       rdssp->dynticks_snap = rdssp->dynticks;
 }
 
 static inline int
@@ -544,9 +571,10 @@ rcu_try_flip_waitack_needed(int cpu)
 {
        long curr;
        long snap;
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
 
-       curr = per_cpu(dynticks_progress_counter, cpu);
-       snap = per_cpu(rcu_dyntick_snapshot, cpu);
+       curr = rdssp->dynticks;
+       snap = rdssp->dynticks_snap;
        smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
 
        /*
@@ -567,7 +595,7 @@ rcu_try_flip_waitack_needed(int cpu)
         * that this CPU already acknowledged the counter.
         */
 
-       if ((curr - snap) > 2 || (snap & 0x1) == 0)
+       if ((curr - snap) > 2 || (curr & 0x1) == 0)
                return 0;
 
        /* We need this CPU to explicitly acknowledge the counter flip. */
@@ -580,9 +608,10 @@ rcu_try_flip_waitmb_needed(int cpu)
 {
        long curr;
        long snap;
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
 
-       curr = per_cpu(dynticks_progress_counter, cpu);
-       snap = per_cpu(rcu_dyntick_snapshot, cpu);
+       curr = rdssp->dynticks;
+       snap = rdssp->dynticks_snap;
        smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
 
        /*
@@ -609,14 +638,86 @@ rcu_try_flip_waitmb_needed(int cpu)
        return 1;
 }
 
+static void dyntick_save_progress_counter_sched(int cpu)
+{
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+       rdssp->sched_dynticks_snap = rdssp->dynticks;
+}
+
+static int rcu_qsctr_inc_needed_dyntick(int cpu)
+{
+       long curr;
+       long snap;
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+       curr = rdssp->dynticks;
+       snap = rdssp->sched_dynticks_snap;
+       smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
+
+       /*
+        * If the CPU remained in dynticks mode for the entire time
+        * and didn't take any interrupts, NMIs, SMIs, or whatever,
+        * then it cannot be in the middle of an rcu_read_lock(), so
+        * the next rcu_read_lock() it executes must use the new value
+        * of the counter.  Therefore, this CPU has been in a quiescent
+        * state the entire time, and we don't need to wait for it.
+        */
+
+       if ((curr == snap) && ((curr & 0x1) == 0))
+               return 0;
+
+       /*
+        * If the CPU passed through or entered a dynticks idle phase with
+        * no active irq handlers, then, as above, this CPU has already
+        * passed through a quiescent state.
+        */
+
+       if ((curr - snap) > 2 || (snap & 0x1) == 0)
+               return 0;
+
+       /* We need this CPU to go through a quiescent state. */
+
+       return 1;
+}
+
 #else /* !CONFIG_NO_HZ */
 
-# define dyntick_save_progress_counter(cpu)    do { } while (0)
-# define rcu_try_flip_waitack_needed(cpu)      (1)
-# define rcu_try_flip_waitmb_needed(cpu)       (1)
+# define dyntick_save_progress_counter(cpu)            do { } while (0)
+# define rcu_try_flip_waitack_needed(cpu)              (1)
+# define rcu_try_flip_waitmb_needed(cpu)               (1)
+
+# define dyntick_save_progress_counter_sched(cpu)      do { } while (0)
+# define rcu_qsctr_inc_needed_dyntick(cpu)             (1)
 
 #endif /* CONFIG_NO_HZ */
 
+static void save_qsctr_sched(int cpu)
+{
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+       rdssp->sched_qs_snap = rdssp->sched_qs;
+}
+
+static inline int rcu_qsctr_inc_needed(int cpu)
+{
+       struct rcu_dyntick_sched *rdssp = &per_cpu(rcu_dyntick_sched, cpu);
+
+       /*
+        * If there has been a quiescent state, no more need to wait
+        * on this CPU.
+        */
+
+       if (rdssp->sched_qs != rdssp->sched_qs_snap) {
+               smp_mb(); /* force ordering with cpu entering schedule(). */
+               return 0;
+       }
+
+       /* We need this CPU to go through a quiescent state. */
+
+       return 1;
+}
+
 /*
  * Get here when RCU is idle.  Decide whether we need to
  * move out of idle state, and return non-zero if so.
@@ -655,7 +756,7 @@ rcu_try_flip_idle(void)
 
        /* Now ask each CPU for acknowledgement of the flip. */
 
-       for_each_cpu_mask(cpu, rcu_cpu_online_map) {
+       for_each_cpu_mask_nr(cpu, rcu_cpu_online_map) {
                per_cpu(rcu_flip_flag, cpu) = rcu_flipped;
                dyntick_save_progress_counter(cpu);
        }
@@ -673,7 +774,7 @@ rcu_try_flip_waitack(void)
        int cpu;
 
        RCU_TRACE_ME(rcupreempt_trace_try_flip_a1);
-       for_each_cpu_mask(cpu, rcu_cpu_online_map)
+       for_each_cpu_mask_nr(cpu, rcu_cpu_online_map)
                if (rcu_try_flip_waitack_needed(cpu) &&
                    per_cpu(rcu_flip_flag, cpu) != rcu_flip_seen) {
                        RCU_TRACE_ME(rcupreempt_trace_try_flip_ae1);
@@ -705,7 +806,7 @@ rcu_try_flip_waitzero(void)
        /* Check to see if the sum of the "last" counters is zero. */
 
        RCU_TRACE_ME(rcupreempt_trace_try_flip_z1);
-       for_each_cpu_mask(cpu, rcu_cpu_online_map)
+       for_each_cpu_mask_nr(cpu, rcu_cpu_online_map)
                sum += RCU_DATA_CPU(cpu)->rcu_flipctr[lastidx];
        if (sum != 0) {
                RCU_TRACE_ME(rcupreempt_trace_try_flip_ze1);
@@ -720,7 +821,7 @@ rcu_try_flip_waitzero(void)
        smp_mb();  /*  ^^^^^^^^^^^^ */
 
        /* Call for a memory barrier from each CPU. */
-       for_each_cpu_mask(cpu, rcu_cpu_online_map) {
+       for_each_cpu_mask_nr(cpu, rcu_cpu_online_map) {
                per_cpu(rcu_mb_flag, cpu) = rcu_mb_needed;
                dyntick_save_progress_counter(cpu);
        }
@@ -740,7 +841,7 @@ rcu_try_flip_waitmb(void)
        int cpu;
 
        RCU_TRACE_ME(rcupreempt_trace_try_flip_m1);
-       for_each_cpu_mask(cpu, rcu_cpu_online_map)
+       for_each_cpu_mask_nr(cpu, rcu_cpu_online_map)
                if (rcu_try_flip_waitmb_needed(cpu) &&
                    per_cpu(rcu_mb_flag, cpu) != rcu_mb_done) {
                        RCU_TRACE_ME(rcupreempt_trace_try_flip_me1);
@@ -819,6 +920,26 @@ void rcu_check_callbacks(int cpu, int user)
        unsigned long flags;
        struct rcu_data *rdp = RCU_DATA_CPU(cpu);
 
+       /*
+        * If this CPU took its interrupt from user mode or from the
+        * idle loop, and this is not a nested interrupt, then
+        * this CPU has to have exited all prior preept-disable
+        * sections of code.  So increment the counter to note this.
+        *
+        * The memory barrier is needed to handle the case where
+        * writes from a preempt-disable section of code get reordered
+        * into schedule() by this CPU's write buffer.  So the memory
+        * barrier makes sure that the rcu_qsctr_inc() is seen by other
+        * CPUs to happen after any such write.
+        */
+
+       if (user ||
+           (idle_cpu(cpu) && !in_softirq() &&
+            hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
+               smp_mb();       /* Guard against aggressive schedule(). */
+               rcu_qsctr_inc(cpu);
+       }
+
        rcu_check_mb(cpu);
        if (rcu_ctrlblk.completed == rdp->completed)
                rcu_try_flip();
@@ -869,6 +990,8 @@ void rcu_offline_cpu(int cpu)
        struct rcu_head *list = NULL;
        unsigned long flags;
        struct rcu_data *rdp = RCU_DATA_CPU(cpu);
+       struct rcu_head *schedlist = NULL;
+       struct rcu_head **schedtail = &schedlist;
        struct rcu_head **tail = &list;
 
        /*
@@ -882,6 +1005,11 @@ void rcu_offline_cpu(int cpu)
                rcu_offline_cpu_enqueue(rdp->waitlist[i], rdp->waittail[i],
                                                list, tail);
        rcu_offline_cpu_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
+       rcu_offline_cpu_enqueue(rdp->waitschedlist, rdp->waitschedtail,
+                               schedlist, schedtail);
+       rcu_offline_cpu_enqueue(rdp->nextschedlist, rdp->nextschedtail,
+                               schedlist, schedtail);
+       rdp->rcu_sched_sleeping = 0;
        spin_unlock_irqrestore(&rdp->lock, flags);
        rdp->waitlistcount = 0;
 
@@ -916,12 +1044,15 @@ void rcu_offline_cpu(int cpu)
         * fix.
         */
 
-       local_irq_save(flags);
+       local_irq_save(flags);  /* disable preempt till we know what lock. */
        rdp = RCU_DATA_ME();
        spin_lock(&rdp->lock);
        *rdp->nexttail = list;
        if (list)
                rdp->nexttail = tail;
+       *rdp->nextschedtail = schedlist;
+       if (schedlist)
+               rdp->nextschedtail = schedtail;
        spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
@@ -936,10 +1067,25 @@ void rcu_offline_cpu(int cpu)
 void __cpuinit rcu_online_cpu(int cpu)
 {
        unsigned long flags;
+       struct rcu_data *rdp;
 
        spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
        cpu_set(cpu, rcu_cpu_online_map);
        spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
+
+       /*
+        * The rcu_sched grace-period processing might have bypassed
+        * this CPU, given that it was not in the rcu_cpu_online_map
+        * when the grace-period scan started.  This means that the
+        * grace-period task might sleep.  So make sure that if this
+        * should happen, the first callback posted to this CPU will
+        * wake up the grace-period task if need be.
+        */
+
+       rdp = RCU_DATA_CPU(cpu);
+       spin_lock_irqsave(&rdp->lock, flags);
+       rdp->rcu_sched_sleeping = 1;
+       spin_unlock_irqrestore(&rdp->lock, flags);
 }
 
 static void rcu_process_callbacks(struct softirq_action *unused)
@@ -982,31 +1128,196 @@ void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
        *rdp->nexttail = head;
        rdp->nexttail = &head->next;
        RCU_TRACE_RDP(rcupreempt_trace_next_add, rdp);
-       spin_unlock(&rdp->lock);
-       local_irq_restore(flags);
+       spin_unlock_irqrestore(&rdp->lock, flags);
 }
 EXPORT_SYMBOL_GPL(call_rcu);
 
+void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+{
+       unsigned long flags;
+       struct rcu_data *rdp;
+       int wake_gp = 0;
+
+       head->func = func;
+       head->next = NULL;
+       local_irq_save(flags);
+       rdp = RCU_DATA_ME();
+       spin_lock(&rdp->lock);
+       *rdp->nextschedtail = head;
+       rdp->nextschedtail = &head->next;
+       if (rdp->rcu_sched_sleeping) {
+
+               /* Grace-period processing might be sleeping... */
+
+               rdp->rcu_sched_sleeping = 0;
+               wake_gp = 1;
+       }
+       spin_unlock_irqrestore(&rdp->lock, flags);
+       if (wake_gp) {
+
+               /* Wake up grace-period processing, unless someone beat us. */
+
+               spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+               if (rcu_ctrlblk.sched_sleep != rcu_sched_sleeping)
+                       wake_gp = 0;
+               rcu_ctrlblk.sched_sleep = rcu_sched_not_sleeping;
+               spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+               if (wake_gp)
+                       wake_up_interruptible(&rcu_ctrlblk.sched_wq);
+       }
+}
+EXPORT_SYMBOL_GPL(call_rcu_sched);
+
 /*
  * Wait until all currently running preempt_disable() code segments
  * (including hardware-irq-disable segments) complete.  Note that
  * in -rt this does -not- necessarily result in all currently executing
  * interrupt -handlers- having completed.
  */
-void __synchronize_sched(void)
+synchronize_rcu_xxx(__synchronize_sched, call_rcu_sched)
+EXPORT_SYMBOL_GPL(__synchronize_sched);
+
+/*
+ * kthread function that manages call_rcu_sched grace periods.
+ */
+static int rcu_sched_grace_period(void *arg)
 {
-       cpumask_t oldmask;
+       int couldsleep;         /* might sleep after current pass. */
+       int couldsleepnext = 0; /* might sleep after next pass. */
        int cpu;
+       unsigned long flags;
+       struct rcu_data *rdp;
+       int ret;
 
-       if (sched_getaffinity(0, &oldmask) < 0)
-               oldmask = cpu_possible_map;
-       for_each_online_cpu(cpu) {
-               sched_setaffinity(0, &cpumask_of_cpu(cpu));
-               schedule();
-       }
-       sched_setaffinity(0, &oldmask);
+       /*
+        * Each pass through the following loop handles one
+        * rcu_sched grace period cycle.
+        */
+       do {
+               /* Save each CPU's current state. */
+
+               for_each_online_cpu(cpu) {
+                       dyntick_save_progress_counter_sched(cpu);
+                       save_qsctr_sched(cpu);
+               }
+
+               /*
+                * Sleep for about an RCU grace-period's worth to
+                * allow better batching and to consume less CPU.
+                */
+               schedule_timeout_interruptible(RCU_SCHED_BATCH_TIME);
+
+               /*
+                * If there was nothing to do last time, prepare to
+                * sleep at the end of the current grace period cycle.
+                */
+               couldsleep = couldsleepnext;
+               couldsleepnext = 1;
+               if (couldsleep) {
+                       spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+                       rcu_ctrlblk.sched_sleep = rcu_sched_sleep_prep;
+                       spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+               }
+
+               /*
+                * Wait on each CPU in turn to have either visited
+                * a quiescent state or been in dynticks-idle mode.
+                */
+               for_each_online_cpu(cpu) {
+                       while (rcu_qsctr_inc_needed(cpu) &&
+                              rcu_qsctr_inc_needed_dyntick(cpu)) {
+                               /* resched_cpu(cpu); @@@ */
+                               schedule_timeout_interruptible(1);
+                       }
+               }
+
+               /* Advance callbacks for each CPU.  */
+
+               for_each_online_cpu(cpu) {
+
+                       rdp = RCU_DATA_CPU(cpu);
+                       spin_lock_irqsave(&rdp->lock, flags);
+
+                       /*
+                        * We are running on this CPU irq-disabled, so no
+                        * CPU can go offline until we re-enable irqs.
+                        * The current CPU might have already gone
+                        * offline (between the for_each_offline_cpu and
+                        * the spin_lock_irqsave), but in that case all its
+                        * callback lists will be empty, so no harm done.
+                        *
+                        * Advance the callbacks!  We share normal RCU's
+                        * donelist, since callbacks are invoked the
+                        * same way in either case.
+                        */
+                       if (rdp->waitschedlist != NULL) {
+                               *rdp->donetail = rdp->waitschedlist;
+                               rdp->donetail = rdp->waitschedtail;
+
+                               /*
+                                * Next rcu_check_callbacks() will
+                                * do the required raise_softirq().
+                                */
+                       }
+                       if (rdp->nextschedlist != NULL) {
+                               rdp->waitschedlist = rdp->nextschedlist;
+                               rdp->waitschedtail = rdp->nextschedtail;
+                               couldsleep = 0;
+                               couldsleepnext = 0;
+                       } else {
+                               rdp->waitschedlist = NULL;
+                               rdp->waitschedtail = &rdp->waitschedlist;
+                       }
+                       rdp->nextschedlist = NULL;
+                       rdp->nextschedtail = &rdp->nextschedlist;
+
+                       /* Mark sleep intention. */
+
+                       rdp->rcu_sched_sleeping = couldsleep;
+
+                       spin_unlock_irqrestore(&rdp->lock, flags);
+               }
+
+               /* If we saw callbacks on the last scan, go deal with them. */
+
+               if (!couldsleep)
+                       continue;
+
+               /* Attempt to block... */
+
+               spin_lock_irqsave(&rcu_ctrlblk.schedlock, flags);
+               if (rcu_ctrlblk.sched_sleep != rcu_sched_sleep_prep) {
+
+                       /*
+                        * Someone posted a callback after we scanned.
+                        * Go take care of it.
+                        */
+                       spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+                       couldsleepnext = 0;
+                       continue;
+               }
+
+               /* Block until the next person posts a callback. */
+
+               rcu_ctrlblk.sched_sleep = rcu_sched_sleeping;
+               spin_unlock_irqrestore(&rcu_ctrlblk.schedlock, flags);
+               ret = 0;
+               __wait_event_interruptible(rcu_ctrlblk.sched_wq,
+                       rcu_ctrlblk.sched_sleep != rcu_sched_sleeping,
+                       ret);
+
+               /*
+                * Signals would prevent us from sleeping, and we cannot
+                * do much with them in any case.  So flush them.
+                */
+               if (ret)
+                       flush_signals(current);
+               couldsleepnext = 0;
+
+       } while (!kthread_should_stop());
+
+       return (0);
 }
-EXPORT_SYMBOL_GPL(__synchronize_sched);
 
 /*
  * Check to see if any future RCU-related work will need to be done
@@ -1023,7 +1334,9 @@ int rcu_needs_cpu(int cpu)
 
        return (rdp->donelist != NULL ||
                !!rdp->waitlistcount ||
-               rdp->nextlist != NULL);
+               rdp->nextlist != NULL ||
+               rdp->nextschedlist != NULL ||
+               rdp->waitschedlist != NULL);
 }
 
 int rcu_pending(int cpu)
@@ -1034,7 +1347,9 @@ int rcu_pending(int cpu)
 
        if (rdp->donelist != NULL ||
            !!rdp->waitlistcount ||
-           rdp->nextlist != NULL)
+           rdp->nextlist != NULL ||
+           rdp->nextschedlist != NULL ||
+           rdp->waitschedlist != NULL)
                return 1;
 
        /* The RCU core needs an acknowledgement from this CPU. */
@@ -1101,6 +1416,11 @@ void __init __rcu_init(void)
                rdp->donetail = &rdp->donelist;
                rdp->rcu_flipctr[0] = 0;
                rdp->rcu_flipctr[1] = 0;
+               rdp->nextschedlist = NULL;
+               rdp->nextschedtail = &rdp->nextschedlist;
+               rdp->waitschedlist = NULL;
+               rdp->waitschedtail = &rdp->waitschedlist;
+               rdp->rcu_sched_sleeping = 0;
        }
        register_cpu_notifier(&rcu_nb);
 
@@ -1119,15 +1439,19 @@ void __init __rcu_init(void)
        for_each_online_cpu(cpu)
                rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE, (void *)(long) cpu);
 
-       open_softirq(RCU_SOFTIRQ, rcu_process_callbacks, NULL);
+       open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
 }
 
 /*
- * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
+ * Late-boot-time RCU initialization that must wait until after scheduler
+ * has been initialized.
  */
-void synchronize_kernel(void)
+void __init rcu_init_sched(void)
 {
-       synchronize_rcu();
+       rcu_sched_grace_period_task = kthread_run(rcu_sched_grace_period,
+                                                 NULL,
+                                                 "rcu_sched_grace_period");
+       WARN_ON(IS_ERR(rcu_sched_grace_period_task));
 }
 
 #ifdef CONFIG_RCU_TRACE