hung_task: fix false positive during vfork
[linux-2.6.git] / kernel / rcutree.c
index d920285..ba06207 100644 (file)
@@ -36,7 +36,7 @@
 #include <linux/interrupt.h>
 #include <linux/sched.h>
 #include <linux/nmi.h>
-#include <asm/atomic.h>
+#include <linux/atomic.h>
 #include <linux/bitops.h>
 #include <linux/module.h>
 #include <linux/completion.h>
@@ -47,6 +47,9 @@
 #include <linux/mutex.h>
 #include <linux/time.h>
 #include <linux/kernel_stat.h>
+#include <linux/wait.h>
+#include <linux/kthread.h>
+#include <linux/prefetch.h>
 
 #include "rcutree.h"
 
@@ -54,8 +57,8 @@
 
 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
 
-#define RCU_STATE_INITIALIZER(name) { \
-       .level = { &name.node[0] }, \
+#define RCU_STATE_INITIALIZER(structname) { \
+       .level = { &structname.node[0] }, \
        .levelcnt = { \
                NUM_RCU_LVL_0,  /* root of hierarchy. */ \
                NUM_RCU_LVL_1, \
@@ -66,13 +69,11 @@ static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
        .signaled = RCU_GP_IDLE, \
        .gpnum = -300, \
        .completed = -300, \
-       .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
-       .orphan_cbs_list = NULL, \
-       .orphan_cbs_tail = &name.orphan_cbs_list, \
-       .orphan_qlen = 0, \
-       .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
+       .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
+       .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
        .n_force_qs = 0, \
        .n_force_qs_ngp = 0, \
+       .name = #structname, \
 }
 
 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
@@ -81,8 +82,65 @@ 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 int rcu_scheduler_active __read_mostly;
+static struct rcu_state *rcu_state;
 
+/*
+ * The rcu_scheduler_active variable transitions from zero to one just
+ * before the first task is spawned.  So when this variable is zero, RCU
+ * can assume that there is but one task, allowing RCU to (for example)
+ * optimized synchronize_sched() to a simple barrier().  When this variable
+ * is one, RCU must actually do all the hard work required to detect real
+ * grace periods.  This variable is also used to suppress boot-time false
+ * positives from lockdep-RCU error checking.
+ */
+int rcu_scheduler_active __read_mostly;
+EXPORT_SYMBOL_GPL(rcu_scheduler_active);
+
+/*
+ * The rcu_scheduler_fully_active variable transitions from zero to one
+ * during the early_initcall() processing, which is after the scheduler
+ * is capable of creating new tasks.  So RCU processing (for example,
+ * creating tasks for RCU priority boosting) must be delayed until after
+ * rcu_scheduler_fully_active transitions from zero to one.  We also
+ * currently delay invocation of any RCU callbacks until after this point.
+ *
+ * It might later prove better for people registering RCU callbacks during
+ * early boot to take responsibility for these callbacks, but one step at
+ * a time.
+ */
+static int rcu_scheduler_fully_active __read_mostly;
+
+#ifdef CONFIG_RCU_BOOST
+
+/*
+ * Control variables for per-CPU and per-rcu_node kthreads.  These
+ * handle all flavors of RCU.
+ */
+static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
+DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
+DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
+DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
+DEFINE_PER_CPU(char, rcu_cpu_has_work);
+
+#endif /* #ifdef CONFIG_RCU_BOOST */
+
+static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
+static void invoke_rcu_core(void);
+static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
+
+#define RCU_KTHREAD_PRIO 1     /* RT priority for per-CPU kthreads. */
+
+/*
+ * Track the rcutorture test sequence number and the update version
+ * number within a given test.  The rcutorture_testseq is incremented
+ * on every rcutorture module load and unload, so has an odd value
+ * when a test is running.  The rcutorture_vernum is set to zero
+ * when rcutorture starts and is incremented on each rcutorture update.
+ * These variables enable correlating rcutorture output with the
+ * RCU tracing information.
+ */
+unsigned long rcutorture_testseq;
+unsigned long rcutorture_vernum;
 
 /*
  * Return true if an RCU grace period is in progress.  The ACCESS_ONCE()s
@@ -101,29 +159,37 @@ static int rcu_gp_in_progress(struct rcu_state *rsp)
  */
 void rcu_sched_qs(int cpu)
 {
-       struct rcu_data *rdp;
+       struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
 
-       rdp = &per_cpu(rcu_sched_data, cpu);
        rdp->passed_quiesc_completed = rdp->gpnum - 1;
        barrier();
        rdp->passed_quiesc = 1;
-       rcu_preempt_note_context_switch(cpu);
 }
 
 void rcu_bh_qs(int cpu)
 {
-       struct rcu_data *rdp;
+       struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
 
-       rdp = &per_cpu(rcu_bh_data, cpu);
        rdp->passed_quiesc_completed = rdp->gpnum - 1;
        barrier();
        rdp->passed_quiesc = 1;
 }
 
+/*
+ * Note a context switch.  This is a quiescent state for RCU-sched,
+ * and requires special handling for preemptible RCU.
+ */
+void rcu_note_context_switch(int cpu)
+{
+       rcu_sched_qs(cpu);
+       rcu_preempt_note_context_switch(cpu);
+}
+EXPORT_SYMBOL_GPL(rcu_note_context_switch);
+
 #ifdef CONFIG_NO_HZ
 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
        .dynticks_nesting = 1,
-       .dynticks = 1,
+       .dynticks = ATOMIC_INIT(1),
 };
 #endif /* #ifdef CONFIG_NO_HZ */
 
@@ -135,6 +201,9 @@ module_param(blimit, int, 0);
 module_param(qhimark, int, 0);
 module_param(qlowmark, int, 0);
 
+int rcu_cpu_stall_suppress __read_mostly;
+module_param(rcu_cpu_stall_suppress, int, 0644);
+
 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
 static int rcu_pending(int cpu);
 
@@ -157,6 +226,49 @@ long rcu_batches_completed_bh(void)
 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 
 /*
+ * Force a quiescent state for RCU BH.
+ */
+void rcu_bh_force_quiescent_state(void)
+{
+       force_quiescent_state(&rcu_bh_state, 0);
+}
+EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
+
+/*
+ * Record the number of times rcutorture tests have been initiated and
+ * terminated.  This information allows the debugfs tracing stats to be
+ * correlated to the rcutorture messages, even when the rcutorture module
+ * is being repeatedly loaded and unloaded.  In other words, we cannot
+ * store this state in rcutorture itself.
+ */
+void rcutorture_record_test_transition(void)
+{
+       rcutorture_testseq++;
+       rcutorture_vernum = 0;
+}
+EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
+
+/*
+ * Record the number of writer passes through the current rcutorture test.
+ * This is also used to correlate debugfs tracing stats with the rcutorture
+ * messages.
+ */
+void rcutorture_record_progress(unsigned long vernum)
+{
+       rcutorture_vernum++;
+}
+EXPORT_SYMBOL_GPL(rcutorture_record_progress);
+
+/*
+ * Force a quiescent state for RCU-sched.
+ */
+void rcu_sched_force_quiescent_state(void)
+{
+       force_quiescent_state(&rcu_sched_state, 0);
+}
+EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
+
+/*
  * Does the CPU have callbacks ready to be invoked?
  */
 static int
@@ -206,8 +318,8 @@ static int rcu_implicit_offline_qs(struct rcu_data *rdp)
                return 1;
        }
 
-       /* If preemptable RCU, no point in sending reschedule IPI. */
-       if (rdp->preemptable)
+       /* If preemptible RCU, no point in sending reschedule IPI. */
+       if (rdp->preemptible)
                return 0;
 
        /* The CPU is online, so send it a reschedule IPI. */
@@ -236,13 +348,25 @@ void rcu_enter_nohz(void)
        unsigned long flags;
        struct rcu_dynticks *rdtp;
 
-       smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
        local_irq_save(flags);
        rdtp = &__get_cpu_var(rcu_dynticks);
-       rdtp->dynticks++;
-       rdtp->dynticks_nesting--;
-       WARN_ON_ONCE(rdtp->dynticks & 0x1);
+       if (--rdtp->dynticks_nesting) {
+               local_irq_restore(flags);
+               return;
+       }
+       /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
+       smp_mb__before_atomic_inc();  /* See above. */
+       atomic_inc(&rdtp->dynticks);
+       smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
+       WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
        local_irq_restore(flags);
+
+       /* If the interrupt queued a callback, get out of dyntick mode. */
+       if (in_irq() &&
+           (__get_cpu_var(rcu_sched_data).nxtlist ||
+            __get_cpu_var(rcu_bh_data).nxtlist ||
+            rcu_preempt_needs_cpu(smp_processor_id())))
+               set_need_resched();
 }
 
 /*
@@ -258,11 +382,16 @@ void rcu_exit_nohz(void)
 
        local_irq_save(flags);
        rdtp = &__get_cpu_var(rcu_dynticks);
-       rdtp->dynticks++;
-       rdtp->dynticks_nesting++;
-       WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
+       if (rdtp->dynticks_nesting++) {
+               local_irq_restore(flags);
+               return;
+       }
+       smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
+       atomic_inc(&rdtp->dynticks);
+       /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
+       smp_mb__after_atomic_inc();  /* See above. */
+       WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
        local_irq_restore(flags);
-       smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
 }
 
 /**
@@ -276,11 +405,15 @@ void rcu_nmi_enter(void)
 {
        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 
-       if (rdtp->dynticks & 0x1)
+       if (rdtp->dynticks_nmi_nesting == 0 &&
+           (atomic_read(&rdtp->dynticks) & 0x1))
                return;
-       rdtp->dynticks_nmi++;
-       WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
-       smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
+       rdtp->dynticks_nmi_nesting++;
+       smp_mb__before_atomic_inc();  /* Force delay from prior write. */
+       atomic_inc(&rdtp->dynticks);
+       /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
+       smp_mb__after_atomic_inc();  /* See above. */
+       WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
 }
 
 /**
@@ -294,11 +427,14 @@ void rcu_nmi_exit(void)
 {
        struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
 
-       if (rdtp->dynticks & 0x1)
+       if (rdtp->dynticks_nmi_nesting == 0 ||
+           --rdtp->dynticks_nmi_nesting != 0)
                return;
-       smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
-       rdtp->dynticks_nmi++;
-       WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
+       /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
+       smp_mb__before_atomic_inc();  /* See above. */
+       atomic_inc(&rdtp->dynticks);
+       smp_mb__after_atomic_inc();  /* Force delay to next write. */
+       WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
 }
 
 /**
@@ -309,13 +445,7 @@ void rcu_nmi_exit(void)
  */
 void rcu_irq_enter(void)
 {
-       struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
-
-       if (rdtp->dynticks_nesting++)
-               return;
-       rdtp->dynticks++;
-       WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
-       smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
+       rcu_exit_nohz();
 }
 
 /**
@@ -327,18 +457,7 @@ void rcu_irq_enter(void)
  */
 void rcu_irq_exit(void)
 {
-       struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
-
-       if (--rdtp->dynticks_nesting)
-               return;
-       smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
-       rdtp->dynticks++;
-       WARN_ON_ONCE(rdtp->dynticks & 0x1);
-
-       /* If the interrupt queued a callback, get out of dyntick mode. */
-       if (__get_cpu_var(rcu_sched_data).nxtlist ||
-           __get_cpu_var(rcu_bh_data).nxtlist)
-               set_need_resched();
+       rcu_enter_nohz();
 }
 
 #ifdef CONFIG_SMP
@@ -350,19 +469,8 @@ void rcu_irq_exit(void)
  */
 static int dyntick_save_progress_counter(struct rcu_data *rdp)
 {
-       int ret;
-       int snap;
-       int snap_nmi;
-
-       snap = rdp->dynticks->dynticks;
-       snap_nmi = rdp->dynticks->dynticks_nmi;
-       smp_mb();       /* Order sampling of snap with end of grace period. */
-       rdp->dynticks_snap = snap;
-       rdp->dynticks_nmi_snap = snap_nmi;
-       ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
-       if (ret)
-               rdp->dynticks_fqs++;
-       return ret;
+       rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
+       return 0;
 }
 
 /*
@@ -373,16 +481,11 @@ static int dyntick_save_progress_counter(struct rcu_data *rdp)
  */
 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 {
-       long curr;
-       long curr_nmi;
-       long snap;
-       long snap_nmi;
+       unsigned long curr;
+       unsigned long snap;
 
-       curr = rdp->dynticks->dynticks;
-       snap = rdp->dynticks_snap;
-       curr_nmi = rdp->dynticks->dynticks_nmi;
-       snap_nmi = rdp->dynticks_nmi_snap;
-       smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
+       curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
+       snap = (unsigned long)rdp->dynticks_snap;
 
        /*
         * If the CPU passed through or entered a dynticks idle phase with
@@ -392,8 +495,7 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
         * read-side critical section that started before the beginning
         * of the current RCU grace period.
         */
-       if ((curr != snap || (curr & 0x1) == 0) &&
-           (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
+       if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
                rdp->dynticks_fqs++;
                return 1;
        }
@@ -422,7 +524,7 @@ static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
 
 #endif /* #else #ifdef CONFIG_NO_HZ */
 
-#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
+int rcu_cpu_stall_suppress __read_mostly;
 
 static void record_gp_stall_check_time(struct rcu_state *rsp)
 {
@@ -439,10 +541,10 @@ static void print_other_cpu_stall(struct rcu_state *rsp)
 
        /* Only let one CPU complain about others per time interval. */
 
-       spin_lock_irqsave(&rnp->lock, flags);
+       raw_spin_lock_irqsave(&rnp->lock, flags);
        delta = jiffies - rsp->jiffies_stall;
        if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
                return;
        }
        rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
@@ -452,23 +554,33 @@ static void print_other_cpu_stall(struct rcu_state *rsp)
         * due to CPU offlining.
         */
        rcu_print_task_stall(rnp);
-       spin_unlock_irqrestore(&rnp->lock, flags);
-
-       /* OK, time to rat on our buddy... */
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
 
-       printk(KERN_ERR "INFO: RCU detected CPU stalls:");
+       /*
+        * OK, time to rat on our buddy...
+        * See Documentation/RCU/stallwarn.txt for info on how to debug
+        * RCU CPU stall warnings.
+        */
+       printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
+              rsp->name);
        rcu_for_each_leaf_node(rsp, rnp) {
+               raw_spin_lock_irqsave(&rnp->lock, flags);
                rcu_print_task_stall(rnp);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
                if (rnp->qsmask == 0)
                        continue;
                for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
                        if (rnp->qsmask & (1UL << cpu))
                                printk(" %d", rnp->grplo + cpu);
        }
-       printk(" (detected by %d, t=%ld jiffies)\n",
+       printk("} (detected by %d, t=%ld jiffies)\n",
               smp_processor_id(), (long)(jiffies - rsp->gp_start));
        trigger_all_cpu_backtrace();
 
+       /* If so configured, complain about tasks blocking the grace period. */
+
+       rcu_print_detail_task_stall(rsp);
+
        force_quiescent_state(rsp, 0);  /* Kick them all. */
 }
 
@@ -477,49 +589,78 @@ static void print_cpu_stall(struct rcu_state *rsp)
        unsigned long flags;
        struct rcu_node *rnp = rcu_get_root(rsp);
 
-       printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
-                       smp_processor_id(), jiffies - rsp->gp_start);
+       /*
+        * OK, time to rat on ourselves...
+        * See Documentation/RCU/stallwarn.txt for info on how to debug
+        * RCU CPU stall warnings.
+        */
+       printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
+              rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
        trigger_all_cpu_backtrace();
 
-       spin_lock_irqsave(&rnp->lock, flags);
-       if ((long)(jiffies - rsp->jiffies_stall) >= 0)
+       raw_spin_lock_irqsave(&rnp->lock, flags);
+       if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
                rsp->jiffies_stall =
                        jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
-       spin_unlock_irqrestore(&rnp->lock, flags);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
 
        set_need_resched();  /* kick ourselves to get things going. */
 }
 
 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 {
-       long delta;
+       unsigned long j;
+       unsigned long js;
        struct rcu_node *rnp;
 
-       delta = jiffies - rsp->jiffies_stall;
+       if (rcu_cpu_stall_suppress)
+               return;
+       j = ACCESS_ONCE(jiffies);
+       js = ACCESS_ONCE(rsp->jiffies_stall);
        rnp = rdp->mynode;
-       if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
+       if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
 
                /* We haven't checked in, so go dump stack. */
                print_cpu_stall(rsp);
 
-       } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
+       } else if (rcu_gp_in_progress(rsp) &&
+                  ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
 
-               /* They had two time units to dump stack, so complain. */
+               /* They had a few time units to dump stack, so complain. */
                print_other_cpu_stall(rsp);
        }
 }
 
-#else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
-
-static void record_gp_stall_check_time(struct rcu_state *rsp)
+static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
 {
+       rcu_cpu_stall_suppress = 1;
+       return NOTIFY_DONE;
 }
 
-static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
+/**
+ * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
+ *
+ * Set the stall-warning timeout way off into the future, thus preventing
+ * any RCU CPU stall-warning messages from appearing in the current set of
+ * RCU grace periods.
+ *
+ * The caller must disable hard irqs.
+ */
+void rcu_cpu_stall_reset(void)
 {
+       rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
+       rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
+       rcu_preempt_stall_reset();
 }
 
-#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
+static struct notifier_block rcu_panic_block = {
+       .notifier_call = rcu_panic,
+};
+
+static void __init check_cpu_stall_init(void)
+{
+       atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
+}
 
 /*
  * Update CPU-local rcu_data state to record the newly noticed grace period.
@@ -531,9 +672,17 @@ static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
 {
        if (rdp->gpnum != rnp->gpnum) {
-               rdp->qs_pending = 1;
-               rdp->passed_quiesc = 0;
+               /*
+                * If the current grace period is waiting for this CPU,
+                * set up to detect a quiescent state, otherwise don't
+                * go looking for one.
+                */
                rdp->gpnum = rnp->gpnum;
+               if (rnp->qsmask & rdp->grpmask) {
+                       rdp->qs_pending = 1;
+                       rdp->passed_quiesc = 0;
+               } else
+                       rdp->qs_pending = 0;
        }
 }
 
@@ -545,12 +694,12 @@ static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
        local_irq_save(flags);
        rnp = rdp->mynode;
        if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
-           !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
+           !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
                local_irq_restore(flags);
                return;
        }
        __note_new_gpnum(rsp, rnp, rdp);
-       spin_unlock_irqrestore(&rnp->lock, flags);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
 /*
@@ -592,6 +741,24 @@ __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_dat
 
                /* Remember that we saw this grace-period completion. */
                rdp->completed = rnp->completed;
+
+               /*
+                * If we were in an extended quiescent state, we may have
+                * missed some grace periods that others CPUs handled on
+                * our behalf. Catch up with this state to avoid noting
+                * spurious new grace periods.  If another grace period
+                * has started, then rnp->gpnum will have advanced, so
+                * we will detect this later on.
+                */
+               if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
+                       rdp->gpnum = rdp->completed;
+
+               /*
+                * If RCU does not need a quiescent state from this CPU,
+                * then make sure that this CPU doesn't go looking for one.
+                */
+               if ((rnp->qsmask & rdp->grpmask) == 0)
+                       rdp->qs_pending = 0;
        }
 }
 
@@ -609,12 +776,12 @@ rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
        local_irq_save(flags);
        rnp = rdp->mynode;
        if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
-           !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
+           !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
                local_irq_restore(flags);
                return;
        }
        __rcu_process_gp_end(rsp, rnp, rdp);
-       spin_unlock_irqrestore(&rnp->lock, flags);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
 /*
@@ -656,15 +823,17 @@ static void
 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
        __releases(rcu_get_root(rsp)->lock)
 {
-       struct rcu_data *rdp = rsp->rda[smp_processor_id()];
+       struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
        struct rcu_node *rnp = rcu_get_root(rsp);
 
        if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
+               if (cpu_needs_another_gp(rsp, rdp))
+                       rsp->fqs_need_gp = 1;
                if (rnp->completed == rsp->completed) {
-                       spin_unlock_irqrestore(&rnp->lock, flags);
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
                        return;
                }
-               spin_unlock(&rnp->lock);         /* irqs remain disabled. */
+               raw_spin_unlock(&rnp->lock);     /* irqs remain disabled. */
 
                /*
                 * Propagate new ->completed value to rcu_node structures
@@ -672,9 +841,9 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
                 * of the next grace period to process their callbacks.
                 */
                rcu_for_each_node_breadth_first(rsp, rnp) {
-                       spin_lock(&rnp->lock);   /* irqs already disabled. */
+                       raw_spin_lock(&rnp->lock); /* irqs already disabled. */
                        rnp->completed = rsp->completed;
-                       spin_unlock(&rnp->lock); /* irqs remain disabled. */
+                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
                }
                local_irq_restore(flags);
                return;
@@ -695,15 +864,16 @@ 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);
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               rcu_preempt_boost_start_gp(rnp);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
                return;
        }
 
-       spin_unlock(&rnp->lock);  /* leave irqs disabled. */
+       raw_spin_unlock(&rnp->lock);  /* leave irqs disabled. */
 
 
        /* Exclude any concurrent CPU-hotplug operations. */
-       spin_lock(&rsp->onofflock);  /* irqs already disabled. */
+       raw_spin_lock(&rsp->onofflock);  /* irqs already disabled. */
 
        /*
         * Set the quiescent-state-needed bits in all the rcu_node
@@ -723,21 +893,22 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
         * irqs disabled.
         */
        rcu_for_each_node_breadth_first(rsp, rnp) {
-               spin_lock(&rnp->lock);          /* irqs already disabled. */
+               raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
                rcu_preempt_check_blocked_tasks(rnp);
                rnp->qsmask = rnp->qsmaskinit;
                rnp->gpnum = rsp->gpnum;
                rnp->completed = rsp->completed;
                if (rnp == rdp->mynode)
                        rcu_start_gp_per_cpu(rsp, rnp, rdp);
-               spin_unlock(&rnp->lock);        /* irqs remain disabled. */
+               rcu_preempt_boost_start_gp(rnp);
+               raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
        }
 
        rnp = rcu_get_root(rsp);
-       spin_lock(&rnp->lock);                  /* irqs already disabled. */
+       raw_spin_lock(&rnp->lock);              /* irqs already disabled. */
        rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
-       spin_unlock(&rnp->lock);                /* irqs remain disabled. */
-       spin_unlock_irqrestore(&rsp->onofflock, flags);
+       raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
+       raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
 }
 
 /*
@@ -750,7 +921,18 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
        __releases(rcu_get_root(rsp)->lock)
 {
+       unsigned long gp_duration;
+
        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
+
+       /*
+        * Ensure that all grace-period and pre-grace-period activity
+        * is seen before the assignment to rsp->completed.
+        */
+       smp_mb(); /* See above block comment. */
+       gp_duration = jiffies - rsp->gp_start;
+       if (gp_duration > rsp->gp_max)
+               rsp->gp_max = gp_duration;
        rsp->completed = rsp->gpnum;
        rsp->signaled = RCU_GP_IDLE;
        rcu_start_gp(rsp, flags);  /* releases root node's rnp->lock. */
@@ -776,14 +958,14 @@ rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
                if (!(rnp->qsmask & mask)) {
 
                        /* Our bit has already been cleared, so done. */
-                       spin_unlock_irqrestore(&rnp->lock, flags);
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
                        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. */
-                       spin_unlock_irqrestore(&rnp->lock, flags);
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
                        return;
                }
                mask = rnp->grpmask;
@@ -793,10 +975,10 @@ rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
 
                        break;
                }
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
                rnp_c = rnp;
                rnp = rnp->parent;
-               spin_lock_irqsave(&rnp->lock, flags);
+               raw_spin_lock_irqsave(&rnp->lock, flags);
                WARN_ON_ONCE(rnp_c->qsmask);
        }
 
@@ -825,7 +1007,7 @@ rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long las
        struct rcu_node *rnp;
 
        rnp = rdp->mynode;
-       spin_lock_irqsave(&rnp->lock, flags);
+       raw_spin_lock_irqsave(&rnp->lock, flags);
        if (lastcomp != rnp->completed) {
 
                /*
@@ -837,12 +1019,12 @@ rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long las
                 * race occurred.
                 */
                rdp->passed_quiesc = 0; /* try again later! */
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
                return;
        }
        mask = rdp->grpmask;
        if ((rnp->qsmask & mask) == 0) {
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
        } else {
                rdp->qs_pending = 0;
 
@@ -893,83 +1075,67 @@ rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
 #ifdef CONFIG_HOTPLUG_CPU
 
 /*
- * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
- * specified flavor of RCU.  The callbacks will be adopted by the next
- * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
- * comes first.  Because this is invoked from the CPU_DYING notifier,
- * irqs are already disabled.
+ * Move a dying CPU's RCU callbacks to online CPU's callback list.
+ * Synchronization is not required because this function executes
+ * in stop_machine() context.
  */
-static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
+static void rcu_send_cbs_to_online(struct rcu_state *rsp)
 {
        int i;
-       struct rcu_data *rdp = rsp->rda[smp_processor_id()];
+       /* current DYING CPU is cleared in the cpu_online_mask */
+       int receive_cpu = cpumask_any(cpu_online_mask);
+       struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
+       struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
 
        if (rdp->nxtlist == NULL)
                return;  /* irqs disabled, so comparison is stable. */
-       spin_lock(&rsp->onofflock);  /* irqs already disabled. */
-       *rsp->orphan_cbs_tail = rdp->nxtlist;
-       rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
+
+       *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
+       receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+       receive_rdp->qlen += rdp->qlen;
+       receive_rdp->n_cbs_adopted += rdp->qlen;
+       rdp->n_cbs_orphaned += rdp->qlen;
+
        rdp->nxtlist = NULL;
        for (i = 0; i < RCU_NEXT_SIZE; i++)
                rdp->nxttail[i] = &rdp->nxtlist;
-       rsp->orphan_qlen += rdp->qlen;
        rdp->qlen = 0;
-       spin_unlock(&rsp->onofflock);  /* irqs remain disabled. */
-}
-
-/*
- * Adopt previously orphaned RCU callbacks.
- */
-static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
-{
-       unsigned long flags;
-       struct rcu_data *rdp;
-
-       spin_lock_irqsave(&rsp->onofflock, flags);
-       rdp = rsp->rda[smp_processor_id()];
-       if (rsp->orphan_cbs_list == NULL) {
-               spin_unlock_irqrestore(&rsp->onofflock, flags);
-               return;
-       }
-       *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
-       rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
-       rdp->qlen += rsp->orphan_qlen;
-       rsp->orphan_cbs_list = NULL;
-       rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
-       rsp->orphan_qlen = 0;
-       spin_unlock_irqrestore(&rsp->onofflock, flags);
 }
 
 /*
  * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
  * and move all callbacks from the outgoing CPU to the current one.
+ * There can only be one CPU hotplug operation at a time, so no other
+ * CPU can be attempting to update rcu_cpu_kthread_task.
  */
 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
 {
        unsigned long flags;
        unsigned long mask;
        int need_report = 0;
-       struct rcu_data *rdp = rsp->rda[cpu];
+       struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
        struct rcu_node *rnp;
 
+       rcu_stop_cpu_kthread(cpu);
+
        /* Exclude any attempts to start a new grace period. */
-       spin_lock_irqsave(&rsp->onofflock, flags);
+       raw_spin_lock_irqsave(&rsp->onofflock, flags);
 
        /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
        rnp = rdp->mynode;      /* this is the outgoing CPU's rnp. */
        mask = rdp->grpmask;    /* rnp->grplo is constant. */
        do {
-               spin_lock(&rnp->lock);          /* irqs already disabled. */
+               raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
                rnp->qsmaskinit &= ~mask;
                if (rnp->qsmaskinit != 0) {
                        if (rnp != rdp->mynode)
-                               spin_unlock(&rnp->lock); /* irqs remain disabled. */
+                               raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
                        break;
                }
                if (rnp == rdp->mynode)
                        need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
                else
-                       spin_unlock(&rnp->lock); /* irqs remain disabled. */
+                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
                mask = rnp->grpmask;
                rnp = rnp->parent;
        } while (rnp != NULL);
@@ -980,16 +1146,15 @@ static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
         * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
         * held leads to deadlock.
         */
-       spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
+       raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
        rnp = rdp->mynode;
        if (need_report & RCU_OFL_TASKS_NORM_GP)
                rcu_report_unblock_qs_rnp(rnp, flags);
        else
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
        if (need_report & RCU_OFL_TASKS_EXP_GP)
                rcu_report_exp_rnp(rsp, rnp);
-
-       rcu_adopt_orphan_cbs(rsp);
+       rcu_node_kthread_setaffinity(rnp, -1);
 }
 
 /*
@@ -1007,11 +1172,7 @@ static void rcu_offline_cpu(int cpu)
 
 #else /* #ifdef CONFIG_HOTPLUG_CPU */
 
-static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
-{
-}
-
-static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
+static void rcu_send_cbs_to_online(struct rcu_state *rsp)
 {
 }
 
@@ -1054,7 +1215,8 @@ static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
        while (list) {
                next = list->next;
                prefetch(next);
-               list->func(list);
+               debug_rcu_head_unqueue(list);
+               __rcu_reclaim(list);
                list = next;
                if (++count >= rdp->blimit)
                        break;
@@ -1064,6 +1226,7 @@ static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
 
        /* Update count, and requeue any remaining callbacks. */
        rdp->qlen -= count;
+       rdp->n_cbs_invoked += count;
        if (list != NULL) {
                *tail = rdp->nxtlist;
                rdp->nxtlist = list;
@@ -1089,7 +1252,7 @@ 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))
-               raise_softirq(RCU_SOFTIRQ);
+               invoke_rcu_core();
 }
 
 /*
@@ -1103,8 +1266,6 @@ static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
  */
 void rcu_check_callbacks(int cpu, int user)
 {
-       if (!rcu_pending(cpu))
-               return; /* if nothing for RCU to do. */
        if (user ||
            (idle_cpu(cpu) && rcu_scheduler_active &&
             !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
@@ -1136,7 +1297,8 @@ void rcu_check_callbacks(int cpu, int user)
                rcu_bh_qs(cpu);
        }
        rcu_preempt_check_callbacks(cpu);
-       raise_softirq(RCU_SOFTIRQ);
+       if (rcu_pending(cpu))
+               invoke_rcu_core();
 }
 
 #ifdef CONFIG_SMP
@@ -1144,6 +1306,8 @@ 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 *))
@@ -1156,19 +1320,20 @@ static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
 
        rcu_for_each_leaf_node(rsp, rnp) {
                mask = 0;
-               spin_lock_irqsave(&rnp->lock, flags);
+               raw_spin_lock_irqsave(&rnp->lock, flags);
                if (!rcu_gp_in_progress(rsp)) {
-                       spin_unlock_irqrestore(&rnp->lock, flags);
+                       raw_spin_unlock_irqrestore(&rnp->lock, flags);
                        return;
                }
                if (rnp->qsmask == 0) {
-                       spin_unlock_irqrestore(&rnp->lock, flags);
+                       rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
                        continue;
                }
                cpu = rnp->grplo;
                bit = 1;
                for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
-                       if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
+                       if ((rnp->qsmask & bit) != 0 &&
+                           f(per_cpu_ptr(rsp->rda, cpu)))
                                mask |= bit;
                }
                if (mask != 0) {
@@ -1177,7 +1342,12 @@ static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
                        rcu_report_qs_rnp(mask, rsp, rnp, flags);
                        continue;
                }
-               spin_unlock_irqrestore(&rnp->lock, flags);
+               raw_spin_unlock_irqrestore(&rnp->lock, flags);
+       }
+       rnp = rcu_get_root(rsp);
+       if (rnp->qsmask == 0) {
+               raw_spin_lock_irqsave(&rnp->lock, flags);
+               rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
        }
 }
 
@@ -1192,19 +1362,18 @@ static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
 
        if (!rcu_gp_in_progress(rsp))
                return;  /* No grace period in progress, nothing to force. */
-       if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
+       if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
                rsp->n_force_qs_lh++; /* Inexact, can lose counts.  Tough! */
                return; /* Someone else is already on the job. */
        }
-       if (relaxed &&
-           (long)(rsp->jiffies_force_qs - jiffies) >= 0)
+       if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
                goto unlock_fqs_ret; /* no emergency and done recently. */
        rsp->n_force_qs++;
-       spin_lock(&rnp->lock);  /* irqs already disabled */
+       raw_spin_lock(&rnp->lock);  /* irqs already disabled */
        rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
        if(!rcu_gp_in_progress(rsp)) {
                rsp->n_force_qs_ngp++;
-               spin_unlock(&rnp->lock);  /* irqs remain disabled */
+               raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
                goto unlock_fqs_ret;  /* no GP in progress, time updated. */
        }
        rsp->fqs_active = 1;
@@ -1215,14 +1384,14 @@ static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
                break; /* grace period idle or initializing, ignore. */
 
        case RCU_SAVE_DYNTICK:
-
-               spin_unlock(&rnp->lock);  /* irqs remain disabled */
                if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
                        break; /* So gcc recognizes the dead code. */
 
+               raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
+
                /* Record dyntick-idle state. */
                force_qs_rnp(rsp, dyntick_save_progress_counter);
-               spin_lock(&rnp->lock);  /* irqs already disabled */
+               raw_spin_lock(&rnp->lock);  /* irqs already disabled */
                if (rcu_gp_in_progress(rsp))
                        rsp->signaled = RCU_FORCE_QS;
                break;
@@ -1230,18 +1399,24 @@ static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
        case RCU_FORCE_QS:
 
                /* Check dyntick-idle state, send IPI to laggarts. */
-               spin_unlock(&rnp->lock);  /* irqs remain disabled */
+               raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
                force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
 
                /* Leave state in case more forcing is required. */
 
-               spin_lock(&rnp->lock);  /* irqs already disabled */
+               raw_spin_lock(&rnp->lock);  /* irqs already disabled */
                break;
        }
        rsp->fqs_active = 0;
-       spin_unlock(&rnp->lock);  /* irqs remain disabled */
+       if (rsp->fqs_need_gp) {
+               raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
+               rsp->fqs_need_gp = 0;
+               rcu_start_gp(rsp, flags); /* releases rnp->lock */
+               return;
+       }
+       raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
 unlock_fqs_ret:
-       spin_unlock_irqrestore(&rsp->fqslock, flags);
+       raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
 }
 
 #else /* #ifdef CONFIG_SMP */
@@ -1269,7 +1444,7 @@ __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
         * If an RCU GP has gone long enough, go check for dyntick
         * idle CPUs and, if needed, send resched IPIs.
         */
-       if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
+       if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
                force_quiescent_state(rsp, 1);
 
        /*
@@ -1283,12 +1458,13 @@ __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
 
        /* Does this CPU require a not-yet-started grace period? */
        if (cpu_needs_another_gp(rsp, rdp)) {
-               spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
+               raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
                rcu_start_gp(rsp, flags);  /* releases above lock */
        }
 
        /* If there are callbacks ready, invoke them. */
-       rcu_do_batch(rsp, rdp);
+       if (cpu_has_callbacks_ready_to_invoke(rdp))
+               invoke_rcu_callbacks(rsp, rdp);
 }
 
 /*
@@ -1296,24 +1472,35 @@ __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
  */
 static void rcu_process_callbacks(struct softirq_action *unused)
 {
-       /*
-        * Memory references from any prior RCU read-side critical sections
-        * executed by the interrupted code must be seen before any RCU
-        * grace-period manipulations below.
-        */
-       smp_mb(); /* See above block comment. */
-
        __rcu_process_callbacks(&rcu_sched_state,
                                &__get_cpu_var(rcu_sched_data));
        __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
        rcu_preempt_process_callbacks();
 
-       /*
-        * Memory references from any later RCU read-side critical sections
-        * executed by the interrupted code must be seen after any RCU
-        * grace-period manipulations above.
-        */
-       smp_mb(); /* See above block comment. */
+       /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
+       rcu_needs_cpu_flush();
+}
+
+/*
+ * Wake up the current CPU's kthread.  This replaces raise_softirq()
+ * in earlier versions of RCU.  Note that because we are running on
+ * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
+ * cannot disappear out from under us.
+ */
+static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
+{
+       if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
+               return;
+       if (likely(!rsp->boost)) {
+               rcu_do_batch(rsp, rdp);
+               return;
+       }
+       invoke_rcu_callbacks_kthread();
+}
+
+static void invoke_rcu_core(void)
+{
+       raise_softirq(RCU_SOFTIRQ);
 }
 
 static void
@@ -1323,6 +1510,7 @@ __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
        unsigned long flags;
        struct rcu_data *rdp;
 
+       debug_rcu_head_queue(head);
        head->func = func;
        head->next = NULL;
 
@@ -1335,21 +1523,17 @@ __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
         * a quiescent state betweentimes.
         */
        local_irq_save(flags);
-       rdp = rsp->rda[smp_processor_id()];
-       rcu_process_gp_end(rsp, rdp);
-       check_for_new_grace_period(rsp, rdp);
+       rdp = this_cpu_ptr(rsp->rda);
 
        /* Add the callback to our list. */
        *rdp->nxttail[RCU_NEXT_TAIL] = head;
        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
+       rdp->qlen++;
 
-       /* Start a new grace period if one not already started. */
-       if (!rcu_gp_in_progress(rsp)) {
-               unsigned long nestflag;
-               struct rcu_node *rnp_root = rcu_get_root(rsp);
-
-               spin_lock_irqsave(&rnp_root->lock, nestflag);
-               rcu_start_gp(rsp, nestflag);  /* releases rnp_root->lock. */
+       /* If interrupts were disabled, don't dive into RCU core. */
+       if (irqs_disabled_flags(flags)) {
+               local_irq_restore(flags);
+               return;
        }
 
        /*
@@ -1359,14 +1543,29 @@ __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
         * invoking force_quiescent_state() if the newly enqueued callback
         * is the only one waiting for a grace period to complete.
         */
-       if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
-               rdp->blimit = LONG_MAX;
-               if (rsp->n_force_qs == rdp->n_force_qs_snap &&
-                   *rdp->nxttail[RCU_DONE_TAIL] != head)
-                       force_quiescent_state(rsp, 0);
-               rdp->n_force_qs_snap = rsp->n_force_qs;
-               rdp->qlen_last_fqs_check = rdp->qlen;
-       } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
+       if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
+
+               /* Are we ignoring a completed grace period? */
+               rcu_process_gp_end(rsp, rdp);
+               check_for_new_grace_period(rsp, rdp);
+
+               /* Start a new grace period if one not already started. */
+               if (!rcu_gp_in_progress(rsp)) {
+                       unsigned long nestflag;
+                       struct rcu_node *rnp_root = rcu_get_root(rsp);
+
+                       raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
+                       rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
+               } else {
+                       /* Give the grace period a kick. */
+                       rdp->blimit = LONG_MAX;
+                       if (rsp->n_force_qs == rdp->n_force_qs_snap &&
+                           *rdp->nxttail[RCU_DONE_TAIL] != head)
+                               force_quiescent_state(rsp, 0);
+                       rdp->n_force_qs_snap = rsp->n_force_qs;
+                       rdp->qlen_last_fqs_check = rdp->qlen;
+               }
+       } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
                force_quiescent_state(rsp, 1);
        local_irq_restore(flags);
 }
@@ -1419,11 +1618,13 @@ void synchronize_sched(void)
        if (rcu_blocking_is_gp())
                return;
 
+       init_rcu_head_on_stack(&rcu.head);
        init_completion(&rcu.completion);
        /* Will wake me after RCU finished. */
        call_rcu_sched(&rcu.head, wakeme_after_rcu);
        /* Wait for it. */
        wait_for_completion(&rcu.completion);
+       destroy_rcu_head_on_stack(&rcu.head);
 }
 EXPORT_SYMBOL_GPL(synchronize_sched);
 
@@ -1443,11 +1644,13 @@ void synchronize_rcu_bh(void)
        if (rcu_blocking_is_gp())
                return;
 
+       init_rcu_head_on_stack(&rcu.head);
        init_completion(&rcu.completion);
        /* Will wake me after RCU finished. */
        call_rcu_bh(&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_bh);
 
@@ -1468,8 +1671,20 @@ static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
        check_cpu_stall(rsp, rdp);
 
        /* Is the RCU core waiting for a quiescent state from this CPU? */
-       if (rdp->qs_pending) {
+       if (rdp->qs_pending && !rdp->passed_quiesc) {
+
+               /*
+                * If force_quiescent_state() coming soon and this CPU
+                * needs a quiescent state, and this is either RCU-sched
+                * or RCU-bh, force a local reschedule.
+                */
                rdp->n_rp_qs_pending++;
+               if (!rdp->preemptible &&
+                   ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
+                                jiffies))
+                       set_need_resched();
+       } else if (rdp->qs_pending && rdp->passed_quiesc) {
+               rdp->n_rp_report_qs++;
                return 1;
        }
 
@@ -1499,7 +1714,7 @@ static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
 
        /* Has an RCU GP gone long enough to send resched IPIs &c? */
        if (rcu_gp_in_progress(rsp) &&
-           ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
+           ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
                rdp->n_rp_need_fqs++;
                return 1;
        }
@@ -1524,10 +1739,9 @@ static int rcu_pending(int cpu)
 /*
  * 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.
+ * 1 if so.
  */
-int rcu_needs_cpu(int cpu)
+static int rcu_needs_cpu_quick_check(int cpu)
 {
        /* RCU callbacks either ready or pending? */
        return per_cpu(rcu_sched_data, cpu).nxtlist ||
@@ -1535,21 +1749,6 @@ int rcu_needs_cpu(int cpu)
               rcu_preempt_needs_cpu(cpu);
 }
 
-/*
- * This function is invoked towards the end of the scheduler's initialization
- * process.  Before this is called, the idle task might contain
- * RCU read-side critical sections (during which time, this idle
- * task is booting the system).  After this function is called, the
- * idle tasks are prohibited from containing RCU read-side critical
- * sections.
- */
-void rcu_scheduler_starting(void)
-{
-       WARN_ON(num_online_cpus() != 1);
-       WARN_ON(nr_context_switches() > 0);
-       rcu_scheduler_active = 1;
-}
-
 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
 static atomic_t rcu_barrier_cpu_count;
 static DEFINE_MUTEX(rcu_barrier_mutex);
@@ -1595,13 +1794,12 @@ static void _rcu_barrier(struct rcu_state *rsp,
         * decrement rcu_barrier_cpu_count -- otherwise the first CPU
         * might complete its grace period before all of the other CPUs
         * did their increment, causing this function to return too
-        * early.
+        * early.  Note that on_each_cpu() disables irqs, which prevents
+        * any CPUs from coming online or going offline until each online
+        * CPU has queued its RCU-barrier callback.
         */
        atomic_set(&rcu_barrier_cpu_count, 1);
-       preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
-       rcu_adopt_orphan_cbs(rsp);
        on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
-       preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
        if (atomic_dec_and_test(&rcu_barrier_cpu_count))
                complete(&rcu_barrier_completion);
        wait_for_completion(&rcu_barrier_completion);
@@ -1634,11 +1832,11 @@ rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
 {
        unsigned long flags;
        int i;
-       struct rcu_data *rdp = rsp->rda[cpu];
+       struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
        struct rcu_node *rnp = rcu_get_root(rsp);
 
        /* Set up local state, ensuring consistent view of global state. */
-       spin_lock_irqsave(&rnp->lock, flags);
+       raw_spin_lock_irqsave(&rnp->lock, flags);
        rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
        rdp->nxtlist = NULL;
        for (i = 0; i < RCU_NEXT_SIZE; i++)
@@ -1648,7 +1846,7 @@ rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
        rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
 #endif /* #ifdef CONFIG_NO_HZ */
        rdp->cpu = cpu;
-       spin_unlock_irqrestore(&rnp->lock, flags);
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
 }
 
 /*
@@ -1658,23 +1856,23 @@ rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
  * that this CPU cannot possibly have any RCU callbacks in flight yet.
  */
 static void __cpuinit
-rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
+rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
 {
        unsigned long flags;
        unsigned long mask;
-       struct rcu_data *rdp = rsp->rda[cpu];
+       struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
        struct rcu_node *rnp = rcu_get_root(rsp);
 
        /* Set up local state, ensuring consistent view of global state. */
-       spin_lock_irqsave(&rnp->lock, flags);
+       raw_spin_lock_irqsave(&rnp->lock, flags);
        rdp->passed_quiesc = 0;  /* We could be racing with new GP, */
        rdp->qs_pending = 1;     /*  so set up to respond to current GP. */
        rdp->beenonline = 1;     /* We have now been online. */
-       rdp->preemptable = preemptable;
+       rdp->preemptible = preemptible;
        rdp->qlen_last_fqs_check = 0;
        rdp->n_force_qs_snap = rsp->n_force_qs;
        rdp->blimit = blimit;
-       spin_unlock(&rnp->lock);                /* irqs remain disabled. */
+       raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
 
        /*
         * A new grace period might start here.  If so, we won't be part
@@ -1682,14 +1880,14 @@ rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
         */
 
        /* Exclude any attempts to start a new GP on large systems. */
-       spin_lock(&rsp->onofflock);             /* irqs already disabled. */
+       raw_spin_lock(&rsp->onofflock);         /* irqs already disabled. */
 
        /* Add CPU to rcu_node bitmasks. */
        rnp = rdp->mynode;
        mask = rdp->grpmask;
        do {
                /* Exclude any attempts to start a new GP on small systems. */
-               spin_lock(&rnp->lock);  /* irqs already disabled. */
+               raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
                rnp->qsmaskinit |= mask;
                mask = rnp->grpmask;
                if (rnp == rdp->mynode) {
@@ -1697,14 +1895,14 @@ rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
                        rdp->completed = rnp->completed;
                        rdp->passed_quiesc_completed = rnp->completed - 1;
                }
-               spin_unlock(&rnp->lock); /* irqs already disabled. */
+               raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
                rnp = rnp->parent;
        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
 
-       spin_unlock_irqrestore(&rsp->onofflock, flags);
+       raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
 }
 
-static void __cpuinit rcu_online_cpu(int cpu)
+static void __cpuinit rcu_prepare_cpu(int cpu)
 {
        rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
        rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
@@ -1718,27 +1916,34 @@ 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_state->rda, cpu);
+       struct rcu_node *rnp = rdp->mynode;
 
        switch (action) {
        case CPU_UP_PREPARE:
        case CPU_UP_PREPARE_FROZEN:
-               rcu_online_cpu(cpu);
+               rcu_prepare_cpu(cpu);
+               rcu_prepare_kthreads(cpu);
+               break;
+       case CPU_ONLINE:
+       case CPU_DOWN_FAILED:
+               rcu_node_kthread_setaffinity(rnp, -1);
+               rcu_cpu_kthread_setrt(cpu, 1);
+               break;
+       case CPU_DOWN_PREPARE:
+               rcu_node_kthread_setaffinity(rnp, cpu);
+               rcu_cpu_kthread_setrt(cpu, 0);
                break;
        case CPU_DYING:
        case CPU_DYING_FROZEN:
                /*
-                * preempt_disable() in _rcu_barrier() prevents stop_machine(),
-                * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
-                * returns, all online cpus have queued rcu_barrier_func().
-                * The dying CPU clears its cpu_online_mask bit and
-                * moves all of its RCU callbacks to ->orphan_cbs_list
-                * in the context of stop_machine(), so subsequent calls
-                * to _rcu_barrier() will adopt these callbacks and only
-                * then queue rcu_barrier_func() on all remaining CPUs.
+                * The whole machine is "stopped" except this CPU, so we can
+                * touch any data without introducing corruption. We send the
+                * dying CPU's callbacks to an arbitrarily chosen online CPU.
                 */
-               rcu_send_cbs_to_orphanage(&rcu_bh_state);
-               rcu_send_cbs_to_orphanage(&rcu_sched_state);
-               rcu_preempt_send_cbs_to_orphanage();
+               rcu_send_cbs_to_online(&rcu_bh_state);
+               rcu_send_cbs_to_online(&rcu_sched_state);
+               rcu_preempt_send_cbs_to_online();
                break;
        case CPU_DEAD:
        case CPU_DEAD_FROZEN:
@@ -1753,6 +1958,21 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 }
 
 /*
+ * This function is invoked towards the end of the scheduler's initialization
+ * process.  Before this is called, the idle task might contain
+ * RCU read-side critical sections (during which time, this idle
+ * task is booting the system).  After this function is called, the
+ * idle tasks are prohibited from containing RCU read-side critical
+ * sections.  This function also enables RCU lockdep checking.
+ */
+void rcu_scheduler_starting(void)
+{
+       WARN_ON(num_online_cpus() != 1);
+       WARN_ON(nr_context_switches() > 0);
+       rcu_scheduler_active = 1;
+}
+
+/*
  * Compute the per-level fanout, either using the exact fanout specified
  * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
  */
@@ -1761,8 +1981,9 @@ static void __init rcu_init_levelspread(struct rcu_state *rsp)
 {
        int i;
 
-       for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
+       for (i = NUM_RCU_LVLS - 1; i > 0; i--)
                rsp->levelspread[i] = CONFIG_RCU_FANOUT;
+       rsp->levelspread[0] = RCU_FANOUT_LEAF;
 }
 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
 static void __init rcu_init_levelspread(struct rcu_state *rsp)
@@ -1783,13 +2004,20 @@ static void __init rcu_init_levelspread(struct rcu_state *rsp)
 /*
  * Helper function for rcu_init() that initializes one rcu_state structure.
  */
-static void __init rcu_init_one(struct rcu_state *rsp)
+static void __init rcu_init_one(struct rcu_state *rsp,
+               struct rcu_data __percpu *rda)
 {
+       static char *buf[] = { "rcu_node_level_0",
+                              "rcu_node_level_1",
+                              "rcu_node_level_2",
+                              "rcu_node_level_3" };  /* Match MAX_RCU_LVLS */
        int cpustride = 1;
        int i;
        int j;
        struct rcu_node *rnp;
 
+       BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
+
        /* Initialize the level-tracking arrays. */
 
        for (i = 1; i < NUM_RCU_LVLS; i++)
@@ -1802,8 +2030,9 @@ static void __init rcu_init_one(struct rcu_state *rsp)
                cpustride *= rsp->levelspread[i];
                rnp = rsp->level[i];
                for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
-                       spin_lock_init(&rnp->lock);
-                       lockdep_set_class(&rnp->lock, &rcu_node_class[i]);
+                       raw_spin_lock_init(&rnp->lock);
+                       lockdep_set_class_and_name(&rnp->lock,
+                                                  &rcu_node_class[i], buf[i]);
                        rnp->gpnum = 0;
                        rnp->qsmask = 0;
                        rnp->qsmaskinit = 0;
@@ -1822,52 +2051,29 @@ static void __init rcu_init_one(struct rcu_state *rsp)
                                              j / rsp->levelspread[i - 1];
                        }
                        rnp->level = i;
-                       INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
-                       INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
-                       INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
-                       INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
+                       INIT_LIST_HEAD(&rnp->blkd_tasks);
                }
        }
-}
 
-/*
- * Helper macro for __rcu_init() and __rcu_init_preempt().  To be used
- * nowhere else!  Assigns leaf node pointers into each CPU's rcu_data
- * structure.
- */
-#define RCU_INIT_FLAVOR(rsp, rcu_data) \
-do { \
-       int i; \
-       int j; \
-       struct rcu_node *rnp; \
-       \
-       rcu_init_one(rsp); \
-       rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
-       j = 0; \
-       for_each_possible_cpu(i) { \
-               if (i > rnp[j].grphi) \
-                       j++; \
-               per_cpu(rcu_data, i).mynode = &rnp[j]; \
-               (rsp)->rda[i] = &per_cpu(rcu_data, i); \
-               rcu_boot_init_percpu_data(i, rsp); \
-       } \
-} while (0)
+       rsp->rda = rda;
+       rnp = rsp->level[NUM_RCU_LVLS - 1];
+       for_each_possible_cpu(i) {
+               while (i > rnp->grphi)
+                       rnp++;
+               per_cpu_ptr(rsp->rda, i)->mynode = rnp;
+               rcu_boot_init_percpu_data(i, rsp);
+       }
+}
 
 void __init rcu_init(void)
 {
-       int i;
+       int cpu;
 
        rcu_bootup_announce();
-#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
-       printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
-#endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
-#if NUM_RCU_LVL_4 != 0
-       printk(KERN_INFO "Experimental four-level hierarchy is enabled.\n");
-#endif /* #if NUM_RCU_LVL_4 != 0 */
-       RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
-       RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
+       rcu_init_one(&rcu_sched_state, &rcu_sched_data);
+       rcu_init_one(&rcu_bh_state, &rcu_bh_data);
        __rcu_init_preempt();
-       open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
+        open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
 
        /*
         * We don't need protection against CPU-hotplug here because
@@ -1875,8 +2081,9 @@ void __init rcu_init(void)
         * or the scheduler are operational.
         */
        cpu_notifier(rcu_cpu_notify, 0);
-       for_each_online_cpu(i)
-               rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)i);
+       for_each_online_cpu(cpu)
+               rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
+       check_cpu_stall_init();
 }
 
 #include "rcutree_plugin.h"