2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
52 /* Data structures. */
54 static struct lock_class_key rcu_root_class;
56 #define RCU_STATE_INITIALIZER(name) { \
57 .level = { &name.node[0] }, \
59 NUM_RCU_LVL_0, /* root of hierarchy. */ \
62 NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \
64 .signaled = RCU_GP_IDLE, \
67 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
68 .orphan_cbs_list = NULL, \
69 .orphan_cbs_tail = &name.orphan_cbs_list, \
71 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
73 .n_force_qs_ngp = 0, \
76 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
77 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
79 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
80 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
85 * permit this function to be invoked without holding the root rcu_node
86 * structure's ->lock, but of course results can be subject to change.
88 static int rcu_gp_in_progress(struct rcu_state *rsp)
90 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
94 * Note a quiescent state. Because we do not need to know
95 * how many quiescent states passed, just if there was at least
96 * one since the start of the grace period, this just sets a flag.
98 void rcu_sched_qs(int cpu)
100 struct rcu_data *rdp;
102 rdp = &per_cpu(rcu_sched_data, cpu);
103 rdp->passed_quiesc_completed = rdp->completed;
105 rdp->passed_quiesc = 1;
106 rcu_preempt_note_context_switch(cpu);
109 void rcu_bh_qs(int cpu)
111 struct rcu_data *rdp;
113 rdp = &per_cpu(rcu_bh_data, cpu);
114 rdp->passed_quiesc_completed = rdp->completed;
116 rdp->passed_quiesc = 1;
120 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
121 .dynticks_nesting = 1,
124 #endif /* #ifdef CONFIG_NO_HZ */
126 static int blimit = 10; /* Maximum callbacks per softirq. */
127 static int qhimark = 10000; /* If this many pending, ignore blimit. */
128 static int qlowmark = 100; /* Once only this many pending, use blimit. */
130 module_param(blimit, int, 0);
131 module_param(qhimark, int, 0);
132 module_param(qlowmark, int, 0);
134 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
135 static int rcu_pending(int cpu);
138 * Return the number of RCU-sched batches processed thus far for debug & stats.
140 long rcu_batches_completed_sched(void)
142 return rcu_sched_state.completed;
144 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
147 * Return the number of RCU BH batches processed thus far for debug & stats.
149 long rcu_batches_completed_bh(void)
151 return rcu_bh_state.completed;
153 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
156 * Does the CPU have callbacks ready to be invoked?
159 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
161 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
165 * Does the current CPU require a yet-as-unscheduled grace period?
168 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
170 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
174 * Return the root node of the specified rcu_state structure.
176 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
178 return &rsp->node[0];
182 * Record the specified "completed" value, which is later used to validate
183 * dynticks counter manipulations and CPU-offline checks. Specify
184 * "rsp->completed - 1" to unconditionally invalidate any future dynticks
185 * manipulations and CPU-offline checks. Such invalidation is useful at
186 * the beginning of a grace period.
188 static void dyntick_record_completed(struct rcu_state *rsp, long comp)
190 rsp->dynticks_completed = comp;
196 * Recall the previously recorded value of the completion for dynticks.
198 static long dyntick_recall_completed(struct rcu_state *rsp)
200 return rsp->dynticks_completed;
204 * If the specified CPU is offline, tell the caller that it is in
205 * a quiescent state. Otherwise, whack it with a reschedule IPI.
206 * Grace periods can end up waiting on an offline CPU when that
207 * CPU is in the process of coming online -- it will be added to the
208 * rcu_node bitmasks before it actually makes it online. The same thing
209 * can happen while a CPU is in the process of coming online. Because this
210 * race is quite rare, we check for it after detecting that the grace
211 * period has been delayed rather than checking each and every CPU
212 * each and every time we start a new grace period.
214 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
217 * If the CPU is offline, it is in a quiescent state. We can
218 * trust its state not to change because interrupts are disabled.
220 if (cpu_is_offline(rdp->cpu)) {
225 /* If preemptable RCU, no point in sending reschedule IPI. */
226 if (rdp->preemptable)
229 /* The CPU is online, so send it a reschedule IPI. */
230 if (rdp->cpu != smp_processor_id())
231 smp_send_reschedule(rdp->cpu);
238 #endif /* #ifdef CONFIG_SMP */
243 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
245 * Enter nohz mode, in other words, -leave- the mode in which RCU
246 * read-side critical sections can occur. (Though RCU read-side
247 * critical sections can occur in irq handlers in nohz mode, a possibility
248 * handled by rcu_irq_enter() and rcu_irq_exit()).
250 void rcu_enter_nohz(void)
253 struct rcu_dynticks *rdtp;
255 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
256 local_irq_save(flags);
257 rdtp = &__get_cpu_var(rcu_dynticks);
259 rdtp->dynticks_nesting--;
260 WARN_ON_ONCE(rdtp->dynticks & 0x1);
261 local_irq_restore(flags);
265 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
267 * Exit nohz mode, in other words, -enter- the mode in which RCU
268 * read-side critical sections normally occur.
270 void rcu_exit_nohz(void)
273 struct rcu_dynticks *rdtp;
275 local_irq_save(flags);
276 rdtp = &__get_cpu_var(rcu_dynticks);
278 rdtp->dynticks_nesting++;
279 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
280 local_irq_restore(flags);
281 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
285 * rcu_nmi_enter - inform RCU of entry to NMI context
287 * If the CPU was idle with dynamic ticks active, and there is no
288 * irq handler running, this updates rdtp->dynticks_nmi to let the
289 * RCU grace-period handling know that the CPU is active.
291 void rcu_nmi_enter(void)
293 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
295 if (rdtp->dynticks & 0x1)
297 rdtp->dynticks_nmi++;
298 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
299 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
303 * rcu_nmi_exit - inform RCU of exit from NMI context
305 * If the CPU was idle with dynamic ticks active, and there is no
306 * irq handler running, this updates rdtp->dynticks_nmi to let the
307 * RCU grace-period handling know that the CPU is no longer active.
309 void rcu_nmi_exit(void)
311 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
313 if (rdtp->dynticks & 0x1)
315 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
316 rdtp->dynticks_nmi++;
317 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
321 * rcu_irq_enter - inform RCU of entry to hard irq context
323 * If the CPU was idle with dynamic ticks active, this updates the
324 * rdtp->dynticks to let the RCU handling know that the CPU is active.
326 void rcu_irq_enter(void)
328 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
330 if (rdtp->dynticks_nesting++)
333 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
334 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
338 * rcu_irq_exit - inform RCU of exit from hard irq context
340 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
341 * to put let the RCU handling be aware that the CPU is going back to idle
344 void rcu_irq_exit(void)
346 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
348 if (--rdtp->dynticks_nesting)
350 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
352 WARN_ON_ONCE(rdtp->dynticks & 0x1);
354 /* If the interrupt queued a callback, get out of dyntick mode. */
355 if (__get_cpu_var(rcu_sched_data).nxtlist ||
356 __get_cpu_var(rcu_bh_data).nxtlist)
363 * Snapshot the specified CPU's dynticks counter so that we can later
364 * credit them with an implicit quiescent state. Return 1 if this CPU
365 * is in dynticks idle mode, which is an extended quiescent state.
367 static int dyntick_save_progress_counter(struct rcu_data *rdp)
373 snap = rdp->dynticks->dynticks;
374 snap_nmi = rdp->dynticks->dynticks_nmi;
375 smp_mb(); /* Order sampling of snap with end of grace period. */
376 rdp->dynticks_snap = snap;
377 rdp->dynticks_nmi_snap = snap_nmi;
378 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
385 * Return true if the specified CPU has passed through a quiescent
386 * state by virtue of being in or having passed through an dynticks
387 * idle state since the last call to dyntick_save_progress_counter()
390 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
397 curr = rdp->dynticks->dynticks;
398 snap = rdp->dynticks_snap;
399 curr_nmi = rdp->dynticks->dynticks_nmi;
400 snap_nmi = rdp->dynticks_nmi_snap;
401 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
404 * If the CPU passed through or entered a dynticks idle phase with
405 * no active irq/NMI handlers, then we can safely pretend that the CPU
406 * already acknowledged the request to pass through a quiescent
407 * state. Either way, that CPU cannot possibly be in an RCU
408 * read-side critical section that started before the beginning
409 * of the current RCU grace period.
411 if ((curr != snap || (curr & 0x1) == 0) &&
412 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
417 /* Go check for the CPU being offline. */
418 return rcu_implicit_offline_qs(rdp);
421 #endif /* #ifdef CONFIG_SMP */
423 #else /* #ifdef CONFIG_NO_HZ */
427 static int dyntick_save_progress_counter(struct rcu_data *rdp)
432 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
434 return rcu_implicit_offline_qs(rdp);
437 #endif /* #ifdef CONFIG_SMP */
439 #endif /* #else #ifdef CONFIG_NO_HZ */
441 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
443 static void record_gp_stall_check_time(struct rcu_state *rsp)
445 rsp->gp_start = jiffies;
446 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
449 static void print_other_cpu_stall(struct rcu_state *rsp)
454 struct rcu_node *rnp = rcu_get_root(rsp);
456 /* Only let one CPU complain about others per time interval. */
458 spin_lock_irqsave(&rnp->lock, flags);
459 delta = jiffies - rsp->jiffies_stall;
460 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
461 spin_unlock_irqrestore(&rnp->lock, flags);
464 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
467 * Now rat on any tasks that got kicked up to the root rcu_node
468 * due to CPU offlining.
470 rcu_print_task_stall(rnp);
471 spin_unlock_irqrestore(&rnp->lock, flags);
473 /* OK, time to rat on our buddy... */
475 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
476 rcu_for_each_leaf_node(rsp, rnp) {
477 rcu_print_task_stall(rnp);
478 if (rnp->qsmask == 0)
480 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
481 if (rnp->qsmask & (1UL << cpu))
482 printk(" %d", rnp->grplo + cpu);
484 printk(" (detected by %d, t=%ld jiffies)\n",
485 smp_processor_id(), (long)(jiffies - rsp->gp_start));
486 trigger_all_cpu_backtrace();
488 force_quiescent_state(rsp, 0); /* Kick them all. */
491 static void print_cpu_stall(struct rcu_state *rsp)
494 struct rcu_node *rnp = rcu_get_root(rsp);
496 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
497 smp_processor_id(), jiffies - rsp->gp_start);
498 trigger_all_cpu_backtrace();
500 spin_lock_irqsave(&rnp->lock, flags);
501 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
503 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
504 spin_unlock_irqrestore(&rnp->lock, flags);
506 set_need_resched(); /* kick ourselves to get things going. */
509 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
512 struct rcu_node *rnp;
514 delta = jiffies - rsp->jiffies_stall;
516 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
518 /* We haven't checked in, so go dump stack. */
519 print_cpu_stall(rsp);
521 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
523 /* They had two time units to dump stack, so complain. */
524 print_other_cpu_stall(rsp);
528 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
530 static void record_gp_stall_check_time(struct rcu_state *rsp)
534 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
538 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
541 * Update CPU-local rcu_data state to record the newly noticed grace period.
542 * This is used both when we started the grace period and when we notice
543 * that someone else started the grace period.
545 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
548 rdp->passed_quiesc = 0;
549 rdp->gpnum = rsp->gpnum;
553 * Did someone else start a new RCU grace period start since we last
554 * checked? Update local state appropriately if so. Must be called
555 * on the CPU corresponding to rdp.
558 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
563 local_irq_save(flags);
564 if (rdp->gpnum != rsp->gpnum) {
565 note_new_gpnum(rsp, rdp);
568 local_irq_restore(flags);
573 * Start a new RCU grace period if warranted, re-initializing the hierarchy
574 * in preparation for detecting the next grace period. The caller must hold
575 * the root node's ->lock, which is released before return. Hard irqs must
579 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
580 __releases(rcu_get_root(rsp)->lock)
582 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
583 struct rcu_node *rnp = rcu_get_root(rsp);
585 if (!cpu_needs_another_gp(rsp, rdp)) {
586 spin_unlock_irqrestore(&rnp->lock, flags);
590 /* Advance to a new grace period and initialize state. */
592 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
593 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
594 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
595 record_gp_stall_check_time(rsp);
596 dyntick_record_completed(rsp, rsp->completed - 1);
597 note_new_gpnum(rsp, rdp);
600 * Because this CPU just now started the new grace period, we know
601 * that all of its callbacks will be covered by this upcoming grace
602 * period, even the ones that were registered arbitrarily recently.
603 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
605 * Other CPUs cannot be sure exactly when the grace period started.
606 * Therefore, their recently registered callbacks must pass through
607 * an additional RCU_NEXT_READY stage, so that they will be handled
608 * by the next RCU grace period.
610 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
611 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
613 /* Special-case the common single-level case. */
614 if (NUM_RCU_NODES == 1) {
615 rcu_preempt_check_blocked_tasks(rnp);
616 rnp->qsmask = rnp->qsmaskinit;
617 rnp->gpnum = rsp->gpnum;
618 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
619 spin_unlock_irqrestore(&rnp->lock, flags);
623 spin_unlock(&rnp->lock); /* leave irqs disabled. */
626 /* Exclude any concurrent CPU-hotplug operations. */
627 spin_lock(&rsp->onofflock); /* irqs already disabled. */
630 * Set the quiescent-state-needed bits in all the rcu_node
631 * structures for all currently online CPUs in breadth-first
632 * order, starting from the root rcu_node structure. This
633 * operation relies on the layout of the hierarchy within the
634 * rsp->node[] array. Note that other CPUs will access only
635 * the leaves of the hierarchy, which still indicate that no
636 * grace period is in progress, at least until the corresponding
637 * leaf node has been initialized. In addition, we have excluded
638 * CPU-hotplug operations.
640 * Note that the grace period cannot complete until we finish
641 * the initialization process, as there will be at least one
642 * qsmask bit set in the root node until that time, namely the
643 * one corresponding to this CPU, due to the fact that we have
646 rcu_for_each_node_breadth_first(rsp, rnp) {
647 spin_lock(&rnp->lock); /* irqs already disabled. */
648 rcu_preempt_check_blocked_tasks(rnp);
649 rnp->qsmask = rnp->qsmaskinit;
650 rnp->gpnum = rsp->gpnum;
651 spin_unlock(&rnp->lock); /* irqs remain disabled. */
654 rnp = rcu_get_root(rsp);
655 spin_lock(&rnp->lock); /* irqs already disabled. */
656 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
657 spin_unlock(&rnp->lock); /* irqs remain disabled. */
658 spin_unlock_irqrestore(&rsp->onofflock, flags);
662 * Advance this CPU's callbacks, but only if the current grace period
663 * has ended. This may be called only from the CPU to whom the rdp
667 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
672 local_irq_save(flags);
673 completed_snap = ACCESS_ONCE(rsp->completed); /* outside of lock. */
675 /* Did another grace period end? */
676 if (rdp->completed != completed_snap) {
678 /* Advance callbacks. No harm if list empty. */
679 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
680 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
681 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
683 /* Remember that we saw this grace-period completion. */
684 rdp->completed = completed_snap;
686 local_irq_restore(flags);
690 * Clean up after the prior grace period and let rcu_start_gp() start up
691 * the next grace period if one is needed. Note that the caller must
692 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
694 static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
695 __releases(rcu_get_root(rsp)->lock)
697 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
698 rsp->completed = rsp->gpnum;
699 rsp->signaled = RCU_GP_IDLE;
700 rcu_process_gp_end(rsp, rsp->rda[smp_processor_id()]);
701 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
705 * Similar to cpu_quiet(), for which it is a helper function. Allows
706 * a group of CPUs to be quieted at one go, though all the CPUs in the
707 * group must be represented by the same leaf rcu_node structure.
708 * That structure's lock must be held upon entry, and it is released
712 cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp,
714 __releases(rnp->lock)
716 struct rcu_node *rnp_c;
718 /* Walk up the rcu_node hierarchy. */
720 if (!(rnp->qsmask & mask)) {
722 /* Our bit has already been cleared, so done. */
723 spin_unlock_irqrestore(&rnp->lock, flags);
726 rnp->qsmask &= ~mask;
727 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
729 /* Other bits still set at this level, so done. */
730 spin_unlock_irqrestore(&rnp->lock, flags);
734 if (rnp->parent == NULL) {
736 /* No more levels. Exit loop holding root lock. */
740 spin_unlock_irqrestore(&rnp->lock, flags);
743 spin_lock_irqsave(&rnp->lock, flags);
744 WARN_ON_ONCE(rnp_c->qsmask);
748 * Get here if we are the last CPU to pass through a quiescent
749 * state for this grace period. Invoke cpu_quiet_msk_finish()
750 * to clean up and start the next grace period if one is needed.
752 cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */
756 * Record a quiescent state for the specified CPU, which must either be
757 * the current CPU. The lastcomp argument is used to make sure we are
758 * still in the grace period of interest. We don't want to end the current
759 * grace period based on quiescent states detected in an earlier grace
763 cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
767 struct rcu_node *rnp;
770 spin_lock_irqsave(&rnp->lock, flags);
771 if (lastcomp != ACCESS_ONCE(rsp->completed)) {
774 * Someone beat us to it for this grace period, so leave.
775 * The race with GP start is resolved by the fact that we
776 * hold the leaf rcu_node lock, so that the per-CPU bits
777 * cannot yet be initialized -- so we would simply find our
778 * CPU's bit already cleared in cpu_quiet_msk() if this race
781 rdp->passed_quiesc = 0; /* try again later! */
782 spin_unlock_irqrestore(&rnp->lock, flags);
786 if ((rnp->qsmask & mask) == 0) {
787 spin_unlock_irqrestore(&rnp->lock, flags);
792 * This GP can't end until cpu checks in, so all of our
793 * callbacks can be processed during the next GP.
795 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
797 cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */
802 * Check to see if there is a new grace period of which this CPU
803 * is not yet aware, and if so, set up local rcu_data state for it.
804 * Otherwise, see if this CPU has just passed through its first
805 * quiescent state for this grace period, and record that fact if so.
808 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
810 /* If there is now a new grace period, record and return. */
811 if (check_for_new_grace_period(rsp, rdp))
815 * Does this CPU still need to do its part for current grace period?
816 * If no, return and let the other CPUs do their part as well.
818 if (!rdp->qs_pending)
822 * Was there a quiescent state since the beginning of the grace
823 * period? If no, then exit and wait for the next call.
825 if (!rdp->passed_quiesc)
828 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
829 cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
832 #ifdef CONFIG_HOTPLUG_CPU
835 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
836 * specified flavor of RCU. The callbacks will be adopted by the next
837 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
838 * comes first. Because this is invoked from the CPU_DYING notifier,
839 * irqs are already disabled.
841 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
844 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
846 if (rdp->nxtlist == NULL)
847 return; /* irqs disabled, so comparison is stable. */
848 spin_lock(&rsp->onofflock); /* irqs already disabled. */
849 *rsp->orphan_cbs_tail = rdp->nxtlist;
850 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
852 for (i = 0; i < RCU_NEXT_SIZE; i++)
853 rdp->nxttail[i] = &rdp->nxtlist;
854 rsp->orphan_qlen += rdp->qlen;
856 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
860 * Adopt previously orphaned RCU callbacks.
862 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
865 struct rcu_data *rdp;
867 spin_lock_irqsave(&rsp->onofflock, flags);
868 rdp = rsp->rda[smp_processor_id()];
869 if (rsp->orphan_cbs_list == NULL) {
870 spin_unlock_irqrestore(&rsp->onofflock, flags);
873 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
874 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
875 rdp->qlen += rsp->orphan_qlen;
876 rsp->orphan_cbs_list = NULL;
877 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
878 rsp->orphan_qlen = 0;
879 spin_unlock_irqrestore(&rsp->onofflock, flags);
883 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
884 * and move all callbacks from the outgoing CPU to the current one.
886 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
891 struct rcu_data *rdp = rsp->rda[cpu];
892 struct rcu_node *rnp;
894 /* Exclude any attempts to start a new grace period. */
895 spin_lock_irqsave(&rsp->onofflock, flags);
897 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
898 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
899 mask = rdp->grpmask; /* rnp->grplo is constant. */
901 spin_lock(&rnp->lock); /* irqs already disabled. */
902 rnp->qsmaskinit &= ~mask;
903 if (rnp->qsmaskinit != 0) {
904 spin_unlock(&rnp->lock); /* irqs remain disabled. */
909 * If there was a task blocking the current grace period,
910 * and if all CPUs have checked in, we need to propagate
911 * the quiescent state up the rcu_node hierarchy. But that
912 * is inconvenient at the moment due to deadlock issues if
913 * this should end the current grace period. So set the
914 * offlined CPU's bit in ->qsmask in order to force the
915 * next force_quiescent_state() invocation to clean up this
916 * mess in a deadlock-free manner.
918 if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
922 spin_unlock(&rnp->lock); /* irqs remain disabled. */
924 } while (rnp != NULL);
925 lastcomp = rsp->completed;
927 spin_unlock_irqrestore(&rsp->onofflock, flags);
929 rcu_adopt_orphan_cbs(rsp);
933 * Remove the specified CPU from the RCU hierarchy and move any pending
934 * callbacks that it might have to the current CPU. This code assumes
935 * that at least one CPU in the system will remain running at all times.
936 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
938 static void rcu_offline_cpu(int cpu)
940 __rcu_offline_cpu(cpu, &rcu_sched_state);
941 __rcu_offline_cpu(cpu, &rcu_bh_state);
942 rcu_preempt_offline_cpu(cpu);
945 #else /* #ifdef CONFIG_HOTPLUG_CPU */
947 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
951 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
955 static void rcu_offline_cpu(int cpu)
959 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
962 * Invoke any RCU callbacks that have made it to the end of their grace
963 * period. Thottle as specified by rdp->blimit.
965 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
968 struct rcu_head *next, *list, **tail;
971 /* If no callbacks are ready, just return.*/
972 if (!cpu_has_callbacks_ready_to_invoke(rdp))
976 * Extract the list of ready callbacks, disabling to prevent
977 * races with call_rcu() from interrupt handlers.
979 local_irq_save(flags);
981 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
982 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
983 tail = rdp->nxttail[RCU_DONE_TAIL];
984 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
985 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
986 rdp->nxttail[count] = &rdp->nxtlist;
987 local_irq_restore(flags);
989 /* Invoke callbacks. */
996 if (++count >= rdp->blimit)
1000 local_irq_save(flags);
1002 /* Update count, and requeue any remaining callbacks. */
1005 *tail = rdp->nxtlist;
1006 rdp->nxtlist = list;
1007 for (count = 0; count < RCU_NEXT_SIZE; count++)
1008 if (&rdp->nxtlist == rdp->nxttail[count])
1009 rdp->nxttail[count] = tail;
1014 /* Reinstate batch limit if we have worked down the excess. */
1015 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1016 rdp->blimit = blimit;
1018 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1019 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1020 rdp->qlen_last_fqs_check = 0;
1021 rdp->n_force_qs_snap = rsp->n_force_qs;
1022 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1023 rdp->qlen_last_fqs_check = rdp->qlen;
1025 local_irq_restore(flags);
1027 /* Re-raise the RCU softirq if there are callbacks remaining. */
1028 if (cpu_has_callbacks_ready_to_invoke(rdp))
1029 raise_softirq(RCU_SOFTIRQ);
1033 * Check to see if this CPU is in a non-context-switch quiescent state
1034 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1035 * Also schedule the RCU softirq handler.
1037 * This function must be called with hardirqs disabled. It is normally
1038 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1039 * false, there is no point in invoking rcu_check_callbacks().
1041 void rcu_check_callbacks(int cpu, int user)
1043 if (!rcu_pending(cpu))
1044 return; /* if nothing for RCU to do. */
1046 (idle_cpu(cpu) && rcu_scheduler_active &&
1047 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1050 * Get here if this CPU took its interrupt from user
1051 * mode or from the idle loop, and if this is not a
1052 * nested interrupt. In this case, the CPU is in
1053 * a quiescent state, so note it.
1055 * No memory barrier is required here because both
1056 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1057 * variables that other CPUs neither access nor modify,
1058 * at least not while the corresponding CPU is online.
1064 } else if (!in_softirq()) {
1067 * Get here if this CPU did not take its interrupt from
1068 * softirq, in other words, if it is not interrupting
1069 * a rcu_bh read-side critical section. This is an _bh
1070 * critical section, so note it.
1075 rcu_preempt_check_callbacks(cpu);
1076 raise_softirq(RCU_SOFTIRQ);
1082 * Scan the leaf rcu_node structures, processing dyntick state for any that
1083 * have not yet encountered a quiescent state, using the function specified.
1084 * Returns 1 if the current grace period ends while scanning (possibly
1085 * because we made it end).
1087 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1088 int (*f)(struct rcu_data *))
1092 unsigned long flags;
1094 struct rcu_node *rnp;
1096 rcu_for_each_leaf_node(rsp, rnp) {
1098 spin_lock_irqsave(&rnp->lock, flags);
1099 if (rsp->completed != lastcomp) {
1100 spin_unlock_irqrestore(&rnp->lock, flags);
1103 if (rnp->qsmask == 0) {
1104 spin_unlock_irqrestore(&rnp->lock, flags);
1109 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1110 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1113 if (mask != 0 && rsp->completed == lastcomp) {
1115 /* cpu_quiet_msk() releases rnp->lock. */
1116 cpu_quiet_msk(mask, rsp, rnp, flags);
1119 spin_unlock_irqrestore(&rnp->lock, flags);
1125 * Force quiescent states on reluctant CPUs, and also detect which
1126 * CPUs are in dyntick-idle mode.
1128 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1130 unsigned long flags;
1132 struct rcu_node *rnp = rcu_get_root(rsp);
1136 if (!rcu_gp_in_progress(rsp))
1137 return; /* No grace period in progress, nothing to force. */
1138 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1139 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1140 return; /* Someone else is already on the job. */
1143 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1144 goto unlock_ret; /* no emergency and done recently. */
1146 spin_lock(&rnp->lock);
1147 lastcomp = rsp->completed;
1148 signaled = rsp->signaled;
1149 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1150 if (lastcomp == rsp->gpnum) {
1151 rsp->n_force_qs_ngp++;
1152 spin_unlock(&rnp->lock);
1153 goto unlock_ret; /* no GP in progress, time updated. */
1155 spin_unlock(&rnp->lock);
1160 break; /* grace period idle or initializing, ignore. */
1162 case RCU_SAVE_DYNTICK:
1164 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1165 break; /* So gcc recognizes the dead code. */
1167 /* Record dyntick-idle state. */
1168 if (rcu_process_dyntick(rsp, lastcomp,
1169 dyntick_save_progress_counter))
1171 /* fall into next case. */
1173 case RCU_SAVE_COMPLETED:
1175 /* Update state, record completion counter. */
1177 spin_lock(&rnp->lock);
1178 if (lastcomp == rsp->completed &&
1179 rsp->signaled == signaled) {
1180 rsp->signaled = RCU_FORCE_QS;
1181 dyntick_record_completed(rsp, lastcomp);
1182 forcenow = signaled == RCU_SAVE_COMPLETED;
1184 spin_unlock(&rnp->lock);
1187 /* fall into next case. */
1191 /* Check dyntick-idle state, send IPI to laggarts. */
1192 if (rcu_process_dyntick(rsp, dyntick_recall_completed(rsp),
1193 rcu_implicit_dynticks_qs))
1196 /* Leave state in case more forcing is required. */
1201 spin_unlock_irqrestore(&rsp->fqslock, flags);
1204 #else /* #ifdef CONFIG_SMP */
1206 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1211 #endif /* #else #ifdef CONFIG_SMP */
1214 * This does the RCU processing work from softirq context for the
1215 * specified rcu_state and rcu_data structures. This may be called
1216 * only from the CPU to whom the rdp belongs.
1219 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1221 unsigned long flags;
1223 WARN_ON_ONCE(rdp->beenonline == 0);
1226 * If an RCU GP has gone long enough, go check for dyntick
1227 * idle CPUs and, if needed, send resched IPIs.
1229 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1230 force_quiescent_state(rsp, 1);
1233 * Advance callbacks in response to end of earlier grace
1234 * period that some other CPU ended.
1236 rcu_process_gp_end(rsp, rdp);
1238 /* Update RCU state based on any recent quiescent states. */
1239 rcu_check_quiescent_state(rsp, rdp);
1241 /* Does this CPU require a not-yet-started grace period? */
1242 if (cpu_needs_another_gp(rsp, rdp)) {
1243 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1244 rcu_start_gp(rsp, flags); /* releases above lock */
1247 /* If there are callbacks ready, invoke them. */
1248 rcu_do_batch(rsp, rdp);
1252 * Do softirq processing for the current CPU.
1254 static void rcu_process_callbacks(struct softirq_action *unused)
1257 * Memory references from any prior RCU read-side critical sections
1258 * executed by the interrupted code must be seen before any RCU
1259 * grace-period manipulations below.
1261 smp_mb(); /* See above block comment. */
1263 __rcu_process_callbacks(&rcu_sched_state,
1264 &__get_cpu_var(rcu_sched_data));
1265 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1266 rcu_preempt_process_callbacks();
1269 * Memory references from any later RCU read-side critical sections
1270 * executed by the interrupted code must be seen after any RCU
1271 * grace-period manipulations above.
1273 smp_mb(); /* See above block comment. */
1277 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1278 struct rcu_state *rsp)
1280 unsigned long flags;
1281 struct rcu_data *rdp;
1286 smp_mb(); /* Ensure RCU update seen before callback registry. */
1289 * Opportunistically note grace-period endings and beginnings.
1290 * Note that we might see a beginning right after we see an
1291 * end, but never vice versa, since this CPU has to pass through
1292 * a quiescent state betweentimes.
1294 local_irq_save(flags);
1295 rdp = rsp->rda[smp_processor_id()];
1296 rcu_process_gp_end(rsp, rdp);
1297 check_for_new_grace_period(rsp, rdp);
1299 /* Add the callback to our list. */
1300 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1301 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1303 /* Start a new grace period if one not already started. */
1304 if (!rcu_gp_in_progress(rsp)) {
1305 unsigned long nestflag;
1306 struct rcu_node *rnp_root = rcu_get_root(rsp);
1308 spin_lock_irqsave(&rnp_root->lock, nestflag);
1309 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1313 * Force the grace period if too many callbacks or too long waiting.
1314 * Enforce hysteresis, and don't invoke force_quiescent_state()
1315 * if some other CPU has recently done so. Also, don't bother
1316 * invoking force_quiescent_state() if the newly enqueued callback
1317 * is the only one waiting for a grace period to complete.
1319 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1320 rdp->blimit = LONG_MAX;
1321 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1322 *rdp->nxttail[RCU_DONE_TAIL] != head)
1323 force_quiescent_state(rsp, 0);
1324 rdp->n_force_qs_snap = rsp->n_force_qs;
1325 rdp->qlen_last_fqs_check = rdp->qlen;
1326 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1327 force_quiescent_state(rsp, 1);
1328 local_irq_restore(flags);
1332 * Queue an RCU-sched callback for invocation after a grace period.
1334 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1336 __call_rcu(head, func, &rcu_sched_state);
1338 EXPORT_SYMBOL_GPL(call_rcu_sched);
1341 * Queue an RCU for invocation after a quicker grace period.
1343 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1345 __call_rcu(head, func, &rcu_bh_state);
1347 EXPORT_SYMBOL_GPL(call_rcu_bh);
1350 * Check to see if there is any immediate RCU-related work to be done
1351 * by the current CPU, for the specified type of RCU, returning 1 if so.
1352 * The checks are in order of increasing expense: checks that can be
1353 * carried out against CPU-local state are performed first. However,
1354 * we must check for CPU stalls first, else we might not get a chance.
1356 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1358 rdp->n_rcu_pending++;
1360 /* Check for CPU stalls, if enabled. */
1361 check_cpu_stall(rsp, rdp);
1363 /* Is the RCU core waiting for a quiescent state from this CPU? */
1364 if (rdp->qs_pending) {
1365 rdp->n_rp_qs_pending++;
1369 /* Does this CPU have callbacks ready to invoke? */
1370 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1371 rdp->n_rp_cb_ready++;
1375 /* Has RCU gone idle with this CPU needing another grace period? */
1376 if (cpu_needs_another_gp(rsp, rdp)) {
1377 rdp->n_rp_cpu_needs_gp++;
1381 /* Has another RCU grace period completed? */
1382 if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */
1383 rdp->n_rp_gp_completed++;
1387 /* Has a new RCU grace period started? */
1388 if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */
1389 rdp->n_rp_gp_started++;
1393 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1394 if (rcu_gp_in_progress(rsp) &&
1395 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1396 rdp->n_rp_need_fqs++;
1401 rdp->n_rp_need_nothing++;
1406 * Check to see if there is any immediate RCU-related work to be done
1407 * by the current CPU, returning 1 if so. This function is part of the
1408 * RCU implementation; it is -not- an exported member of the RCU API.
1410 static int rcu_pending(int cpu)
1412 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1413 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1414 rcu_preempt_pending(cpu);
1418 * Check to see if any future RCU-related work will need to be done
1419 * by the current CPU, even if none need be done immediately, returning
1420 * 1 if so. This function is part of the RCU implementation; it is -not-
1421 * an exported member of the RCU API.
1423 int rcu_needs_cpu(int cpu)
1425 /* RCU callbacks either ready or pending? */
1426 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1427 per_cpu(rcu_bh_data, cpu).nxtlist ||
1428 rcu_preempt_needs_cpu(cpu);
1431 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1432 static atomic_t rcu_barrier_cpu_count;
1433 static DEFINE_MUTEX(rcu_barrier_mutex);
1434 static struct completion rcu_barrier_completion;
1436 static void rcu_barrier_callback(struct rcu_head *notused)
1438 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1439 complete(&rcu_barrier_completion);
1443 * Called with preemption disabled, and from cross-cpu IRQ context.
1445 static void rcu_barrier_func(void *type)
1447 int cpu = smp_processor_id();
1448 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1449 void (*call_rcu_func)(struct rcu_head *head,
1450 void (*func)(struct rcu_head *head));
1452 atomic_inc(&rcu_barrier_cpu_count);
1453 call_rcu_func = type;
1454 call_rcu_func(head, rcu_barrier_callback);
1458 * Orchestrate the specified type of RCU barrier, waiting for all
1459 * RCU callbacks of the specified type to complete.
1461 static void _rcu_barrier(struct rcu_state *rsp,
1462 void (*call_rcu_func)(struct rcu_head *head,
1463 void (*func)(struct rcu_head *head)))
1465 BUG_ON(in_interrupt());
1466 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1467 mutex_lock(&rcu_barrier_mutex);
1468 init_completion(&rcu_barrier_completion);
1470 * Initialize rcu_barrier_cpu_count to 1, then invoke
1471 * rcu_barrier_func() on each CPU, so that each CPU also has
1472 * incremented rcu_barrier_cpu_count. Only then is it safe to
1473 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1474 * might complete its grace period before all of the other CPUs
1475 * did their increment, causing this function to return too
1478 atomic_set(&rcu_barrier_cpu_count, 1);
1479 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1480 rcu_adopt_orphan_cbs(rsp);
1481 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1482 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1483 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1484 complete(&rcu_barrier_completion);
1485 wait_for_completion(&rcu_barrier_completion);
1486 mutex_unlock(&rcu_barrier_mutex);
1490 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1492 void rcu_barrier_bh(void)
1494 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1496 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1499 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1501 void rcu_barrier_sched(void)
1503 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1505 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1508 * Do boot-time initialization of a CPU's per-CPU RCU data.
1511 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1513 unsigned long flags;
1515 struct rcu_data *rdp = rsp->rda[cpu];
1516 struct rcu_node *rnp = rcu_get_root(rsp);
1518 /* Set up local state, ensuring consistent view of global state. */
1519 spin_lock_irqsave(&rnp->lock, flags);
1520 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1521 rdp->nxtlist = NULL;
1522 for (i = 0; i < RCU_NEXT_SIZE; i++)
1523 rdp->nxttail[i] = &rdp->nxtlist;
1526 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1527 #endif /* #ifdef CONFIG_NO_HZ */
1529 spin_unlock_irqrestore(&rnp->lock, flags);
1533 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1534 * offline event can be happening at a given time. Note also that we
1535 * can accept some slop in the rsp->completed access due to the fact
1536 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1538 static void __cpuinit
1539 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1541 unsigned long flags;
1544 struct rcu_data *rdp = rsp->rda[cpu];
1545 struct rcu_node *rnp = rcu_get_root(rsp);
1547 /* Set up local state, ensuring consistent view of global state. */
1548 spin_lock_irqsave(&rnp->lock, flags);
1549 lastcomp = rsp->completed;
1550 rdp->completed = lastcomp;
1551 rdp->gpnum = lastcomp;
1552 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1553 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1554 rdp->beenonline = 1; /* We have now been online. */
1555 rdp->preemptable = preemptable;
1556 rdp->passed_quiesc_completed = lastcomp - 1;
1557 rdp->qlen_last_fqs_check = 0;
1558 rdp->n_force_qs_snap = rsp->n_force_qs;
1559 rdp->blimit = blimit;
1560 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1563 * A new grace period might start here. If so, we won't be part
1564 * of it, but that is OK, as we are currently in a quiescent state.
1567 /* Exclude any attempts to start a new GP on large systems. */
1568 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1570 /* Add CPU to rcu_node bitmasks. */
1572 mask = rdp->grpmask;
1574 /* Exclude any attempts to start a new GP on small systems. */
1575 spin_lock(&rnp->lock); /* irqs already disabled. */
1576 rnp->qsmaskinit |= mask;
1577 mask = rnp->grpmask;
1578 spin_unlock(&rnp->lock); /* irqs already disabled. */
1580 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1582 spin_unlock_irqrestore(&rsp->onofflock, flags);
1585 static void __cpuinit rcu_online_cpu(int cpu)
1587 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1588 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1589 rcu_preempt_init_percpu_data(cpu);
1593 * Handle CPU online/offline notification events.
1595 int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1596 unsigned long action, void *hcpu)
1598 long cpu = (long)hcpu;
1601 case CPU_UP_PREPARE:
1602 case CPU_UP_PREPARE_FROZEN:
1603 rcu_online_cpu(cpu);
1606 case CPU_DYING_FROZEN:
1608 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1609 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1610 * returns, all online cpus have queued rcu_barrier_func().
1611 * The dying CPU clears its cpu_online_mask bit and
1612 * moves all of its RCU callbacks to ->orphan_cbs_list
1613 * in the context of stop_machine(), so subsequent calls
1614 * to _rcu_barrier() will adopt these callbacks and only
1615 * then queue rcu_barrier_func() on all remaining CPUs.
1617 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1618 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1619 rcu_preempt_send_cbs_to_orphanage();
1622 case CPU_DEAD_FROZEN:
1623 case CPU_UP_CANCELED:
1624 case CPU_UP_CANCELED_FROZEN:
1625 rcu_offline_cpu(cpu);
1634 * Compute the per-level fanout, either using the exact fanout specified
1635 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1637 #ifdef CONFIG_RCU_FANOUT_EXACT
1638 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1642 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1643 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1645 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1646 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1653 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1654 ccur = rsp->levelcnt[i];
1655 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1659 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1662 * Helper function for rcu_init() that initializes one rcu_state structure.
1664 static void __init rcu_init_one(struct rcu_state *rsp)
1669 struct rcu_node *rnp;
1671 /* Initialize the level-tracking arrays. */
1673 for (i = 1; i < NUM_RCU_LVLS; i++)
1674 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1675 rcu_init_levelspread(rsp);
1677 /* Initialize the elements themselves, starting from the leaves. */
1679 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1680 cpustride *= rsp->levelspread[i];
1681 rnp = rsp->level[i];
1682 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1683 spin_lock_init(&rnp->lock);
1686 rnp->qsmaskinit = 0;
1687 rnp->grplo = j * cpustride;
1688 rnp->grphi = (j + 1) * cpustride - 1;
1689 if (rnp->grphi >= NR_CPUS)
1690 rnp->grphi = NR_CPUS - 1;
1696 rnp->grpnum = j % rsp->levelspread[i - 1];
1697 rnp->grpmask = 1UL << rnp->grpnum;
1698 rnp->parent = rsp->level[i - 1] +
1699 j / rsp->levelspread[i - 1];
1702 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1703 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1706 lockdep_set_class(&rcu_get_root(rsp)->lock, &rcu_root_class);
1710 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1711 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1714 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1718 struct rcu_node *rnp; \
1720 rcu_init_one(rsp); \
1721 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1723 for_each_possible_cpu(i) { \
1724 if (i > rnp[j].grphi) \
1726 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1727 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1728 rcu_boot_init_percpu_data(i, rsp); \
1732 void __init __rcu_init(void)
1734 rcu_bootup_announce();
1735 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1736 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1737 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1738 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1739 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1740 __rcu_init_preempt();
1741 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1744 #include "rcutree_plugin.h"