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 #define RCU_STATE_INITIALIZER(name) { \
55 .level = { &name.node[0] }, \
57 NUM_RCU_LVL_0, /* root of hierarchy. */ \
60 NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \
62 .signaled = RCU_GP_IDLE, \
65 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
66 .orphan_cbs_list = NULL, \
67 .orphan_cbs_tail = &name.orphan_cbs_list, \
69 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
71 .n_force_qs_ngp = 0, \
74 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
75 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
77 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
78 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
82 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
83 * permit this function to be invoked without holding the root rcu_node
84 * structure's ->lock, but of course results can be subject to change.
86 static int rcu_gp_in_progress(struct rcu_state *rsp)
88 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
92 * Note a quiescent state. Because we do not need to know
93 * how many quiescent states passed, just if there was at least
94 * one since the start of the grace period, this just sets a flag.
96 void rcu_sched_qs(int cpu)
100 rdp = &per_cpu(rcu_sched_data, cpu);
101 rdp->passed_quiesc_completed = rdp->completed;
103 rdp->passed_quiesc = 1;
104 rcu_preempt_note_context_switch(cpu);
107 void rcu_bh_qs(int cpu)
109 struct rcu_data *rdp;
111 rdp = &per_cpu(rcu_bh_data, cpu);
112 rdp->passed_quiesc_completed = rdp->completed;
114 rdp->passed_quiesc = 1;
118 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
119 .dynticks_nesting = 1,
122 #endif /* #ifdef CONFIG_NO_HZ */
124 static int blimit = 10; /* Maximum callbacks per softirq. */
125 static int qhimark = 10000; /* If this many pending, ignore blimit. */
126 static int qlowmark = 100; /* Once only this many pending, use blimit. */
128 module_param(blimit, int, 0);
129 module_param(qhimark, int, 0);
130 module_param(qlowmark, int, 0);
132 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
133 static int rcu_pending(int cpu);
136 * Return the number of RCU-sched batches processed thus far for debug & stats.
138 long rcu_batches_completed_sched(void)
140 return rcu_sched_state.completed;
142 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
145 * Return the number of RCU BH batches processed thus far for debug & stats.
147 long rcu_batches_completed_bh(void)
149 return rcu_bh_state.completed;
151 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
154 * Does the CPU have callbacks ready to be invoked?
157 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
159 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
163 * Does the current CPU require a yet-as-unscheduled grace period?
166 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
168 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
172 * Return the root node of the specified rcu_state structure.
174 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
176 return &rsp->node[0];
180 * Record the specified "completed" value, which is later used to validate
181 * dynticks counter manipulations and CPU-offline checks. Specify
182 * "rsp->completed - 1" to unconditionally invalidate any future dynticks
183 * manipulations and CPU-offline checks. Such invalidation is useful at
184 * the beginning of a grace period.
186 static void dyntick_record_completed(struct rcu_state *rsp, long comp)
188 rsp->dynticks_completed = comp;
194 * Recall the previously recorded value of the completion for dynticks.
196 static long dyntick_recall_completed(struct rcu_state *rsp)
198 return rsp->dynticks_completed;
202 * If the specified CPU is offline, tell the caller that it is in
203 * a quiescent state. Otherwise, whack it with a reschedule IPI.
204 * Grace periods can end up waiting on an offline CPU when that
205 * CPU is in the process of coming online -- it will be added to the
206 * rcu_node bitmasks before it actually makes it online. The same thing
207 * can happen while a CPU is in the process of coming online. Because this
208 * race is quite rare, we check for it after detecting that the grace
209 * period has been delayed rather than checking each and every CPU
210 * each and every time we start a new grace period.
212 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
215 * If the CPU is offline, it is in a quiescent state. We can
216 * trust its state not to change because interrupts are disabled.
218 if (cpu_is_offline(rdp->cpu)) {
223 /* If preemptable RCU, no point in sending reschedule IPI. */
224 if (rdp->preemptable)
227 /* The CPU is online, so send it a reschedule IPI. */
228 if (rdp->cpu != smp_processor_id())
229 smp_send_reschedule(rdp->cpu);
236 #endif /* #ifdef CONFIG_SMP */
241 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
243 * Enter nohz mode, in other words, -leave- the mode in which RCU
244 * read-side critical sections can occur. (Though RCU read-side
245 * critical sections can occur in irq handlers in nohz mode, a possibility
246 * handled by rcu_irq_enter() and rcu_irq_exit()).
248 void rcu_enter_nohz(void)
251 struct rcu_dynticks *rdtp;
253 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
254 local_irq_save(flags);
255 rdtp = &__get_cpu_var(rcu_dynticks);
257 rdtp->dynticks_nesting--;
258 WARN_ON_ONCE(rdtp->dynticks & 0x1);
259 local_irq_restore(flags);
263 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
265 * Exit nohz mode, in other words, -enter- the mode in which RCU
266 * read-side critical sections normally occur.
268 void rcu_exit_nohz(void)
271 struct rcu_dynticks *rdtp;
273 local_irq_save(flags);
274 rdtp = &__get_cpu_var(rcu_dynticks);
276 rdtp->dynticks_nesting++;
277 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
278 local_irq_restore(flags);
279 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
283 * rcu_nmi_enter - inform RCU of entry to NMI context
285 * If the CPU was idle with dynamic ticks active, and there is no
286 * irq handler running, this updates rdtp->dynticks_nmi to let the
287 * RCU grace-period handling know that the CPU is active.
289 void rcu_nmi_enter(void)
291 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
293 if (rdtp->dynticks & 0x1)
295 rdtp->dynticks_nmi++;
296 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
297 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
301 * rcu_nmi_exit - inform RCU of exit from NMI context
303 * If the CPU was idle with dynamic ticks active, and there is no
304 * irq handler running, this updates rdtp->dynticks_nmi to let the
305 * RCU grace-period handling know that the CPU is no longer active.
307 void rcu_nmi_exit(void)
309 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
311 if (rdtp->dynticks & 0x1)
313 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
314 rdtp->dynticks_nmi++;
315 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
319 * rcu_irq_enter - inform RCU of entry to hard irq context
321 * If the CPU was idle with dynamic ticks active, this updates the
322 * rdtp->dynticks to let the RCU handling know that the CPU is active.
324 void rcu_irq_enter(void)
326 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
328 if (rdtp->dynticks_nesting++)
331 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
332 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
336 * rcu_irq_exit - inform RCU of exit from hard irq context
338 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
339 * to put let the RCU handling be aware that the CPU is going back to idle
342 void rcu_irq_exit(void)
344 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
346 if (--rdtp->dynticks_nesting)
348 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
350 WARN_ON_ONCE(rdtp->dynticks & 0x1);
352 /* If the interrupt queued a callback, get out of dyntick mode. */
353 if (__get_cpu_var(rcu_sched_data).nxtlist ||
354 __get_cpu_var(rcu_bh_data).nxtlist)
361 * Snapshot the specified CPU's dynticks counter so that we can later
362 * credit them with an implicit quiescent state. Return 1 if this CPU
363 * is in dynticks idle mode, which is an extended quiescent state.
365 static int dyntick_save_progress_counter(struct rcu_data *rdp)
371 snap = rdp->dynticks->dynticks;
372 snap_nmi = rdp->dynticks->dynticks_nmi;
373 smp_mb(); /* Order sampling of snap with end of grace period. */
374 rdp->dynticks_snap = snap;
375 rdp->dynticks_nmi_snap = snap_nmi;
376 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
383 * Return true if the specified CPU has passed through a quiescent
384 * state by virtue of being in or having passed through an dynticks
385 * idle state since the last call to dyntick_save_progress_counter()
388 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
395 curr = rdp->dynticks->dynticks;
396 snap = rdp->dynticks_snap;
397 curr_nmi = rdp->dynticks->dynticks_nmi;
398 snap_nmi = rdp->dynticks_nmi_snap;
399 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
402 * If the CPU passed through or entered a dynticks idle phase with
403 * no active irq/NMI handlers, then we can safely pretend that the CPU
404 * already acknowledged the request to pass through a quiescent
405 * state. Either way, that CPU cannot possibly be in an RCU
406 * read-side critical section that started before the beginning
407 * of the current RCU grace period.
409 if ((curr != snap || (curr & 0x1) == 0) &&
410 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
415 /* Go check for the CPU being offline. */
416 return rcu_implicit_offline_qs(rdp);
419 #endif /* #ifdef CONFIG_SMP */
421 #else /* #ifdef CONFIG_NO_HZ */
425 static int dyntick_save_progress_counter(struct rcu_data *rdp)
430 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
432 return rcu_implicit_offline_qs(rdp);
435 #endif /* #ifdef CONFIG_SMP */
437 #endif /* #else #ifdef CONFIG_NO_HZ */
439 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
441 static void record_gp_stall_check_time(struct rcu_state *rsp)
443 rsp->gp_start = jiffies;
444 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
447 static void print_other_cpu_stall(struct rcu_state *rsp)
452 struct rcu_node *rnp = rcu_get_root(rsp);
454 /* Only let one CPU complain about others per time interval. */
456 spin_lock_irqsave(&rnp->lock, flags);
457 delta = jiffies - rsp->jiffies_stall;
458 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
459 spin_unlock_irqrestore(&rnp->lock, flags);
462 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
465 * Now rat on any tasks that got kicked up to the root rcu_node
466 * due to CPU offlining.
468 rcu_print_task_stall(rnp);
469 spin_unlock_irqrestore(&rnp->lock, flags);
471 /* OK, time to rat on our buddy... */
473 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
474 rcu_for_each_leaf_node(rsp, rnp) {
475 rcu_print_task_stall(rnp);
476 if (rnp->qsmask == 0)
478 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
479 if (rnp->qsmask & (1UL << cpu))
480 printk(" %d", rnp->grplo + cpu);
482 printk(" (detected by %d, t=%ld jiffies)\n",
483 smp_processor_id(), (long)(jiffies - rsp->gp_start));
484 trigger_all_cpu_backtrace();
486 force_quiescent_state(rsp, 0); /* Kick them all. */
489 static void print_cpu_stall(struct rcu_state *rsp)
492 struct rcu_node *rnp = rcu_get_root(rsp);
494 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
495 smp_processor_id(), jiffies - rsp->gp_start);
496 trigger_all_cpu_backtrace();
498 spin_lock_irqsave(&rnp->lock, flags);
499 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
501 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
502 spin_unlock_irqrestore(&rnp->lock, flags);
504 set_need_resched(); /* kick ourselves to get things going. */
507 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
510 struct rcu_node *rnp;
512 delta = jiffies - rsp->jiffies_stall;
514 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
516 /* We haven't checked in, so go dump stack. */
517 print_cpu_stall(rsp);
519 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
521 /* They had two time units to dump stack, so complain. */
522 print_other_cpu_stall(rsp);
526 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
528 static void record_gp_stall_check_time(struct rcu_state *rsp)
532 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
536 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
539 * Update CPU-local rcu_data state to record the newly noticed grace period.
540 * This is used both when we started the grace period and when we notice
541 * that someone else started the grace period.
543 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
546 rdp->passed_quiesc = 0;
547 rdp->gpnum = rsp->gpnum;
551 * Did someone else start a new RCU grace period start since we last
552 * checked? Update local state appropriately if so. Must be called
553 * on the CPU corresponding to rdp.
556 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
561 local_irq_save(flags);
562 if (rdp->gpnum != rsp->gpnum) {
563 note_new_gpnum(rsp, rdp);
566 local_irq_restore(flags);
571 * Advance this CPU's callbacks, but only if the current grace period
572 * has ended. This may be called only from the CPU to whom the rdp
573 * belongs. In addition, the corresponding leaf rcu_node structure's
574 * ->lock must be held by the caller, with irqs disabled.
577 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
579 /* Did another grace period end? */
580 if (rdp->completed != rnp->completed) {
582 /* Advance callbacks. No harm if list empty. */
583 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
584 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
585 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
587 /* Remember that we saw this grace-period completion. */
588 rdp->completed = rnp->completed;
593 * Advance this CPU's callbacks, but only if the current grace period
594 * has ended. This may be called only from the CPU to whom the rdp
598 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
601 struct rcu_node *rnp;
603 local_irq_save(flags);
605 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
606 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
607 local_irq_restore(flags);
610 __rcu_process_gp_end(rsp, rnp, rdp);
611 spin_unlock_irqrestore(&rnp->lock, flags);
615 * Do per-CPU grace-period initialization for running CPU. The caller
616 * must hold the lock of the leaf rcu_node structure corresponding to
620 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
622 /* Prior grace period ended, so advance callbacks for current CPU. */
623 __rcu_process_gp_end(rsp, rnp, rdp);
626 * Because this CPU just now started the new grace period, we know
627 * that all of its callbacks will be covered by this upcoming grace
628 * period, even the ones that were registered arbitrarily recently.
629 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
631 * Other CPUs cannot be sure exactly when the grace period started.
632 * Therefore, their recently registered callbacks must pass through
633 * an additional RCU_NEXT_READY stage, so that they will be handled
634 * by the next RCU grace period.
636 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
637 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
641 * Start a new RCU grace period if warranted, re-initializing the hierarchy
642 * in preparation for detecting the next grace period. The caller must hold
643 * the root node's ->lock, which is released before return. Hard irqs must
647 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
648 __releases(rcu_get_root(rsp)->lock)
650 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
651 struct rcu_node *rnp = rcu_get_root(rsp);
653 if (!cpu_needs_another_gp(rsp, rdp)) {
654 spin_unlock_irqrestore(&rnp->lock, flags);
658 /* Advance to a new grace period and initialize state. */
660 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
661 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
662 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
663 record_gp_stall_check_time(rsp);
664 dyntick_record_completed(rsp, rsp->completed - 1);
665 note_new_gpnum(rsp, rdp);
667 /* Special-case the common single-level case. */
668 if (NUM_RCU_NODES == 1) {
669 rcu_preempt_check_blocked_tasks(rnp);
670 rnp->qsmask = rnp->qsmaskinit;
671 rnp->gpnum = rsp->gpnum;
672 rnp->completed = rsp->completed;
673 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
674 rcu_start_gp_per_cpu(rsp, rnp, rdp);
675 spin_unlock_irqrestore(&rnp->lock, flags);
679 spin_unlock(&rnp->lock); /* leave irqs disabled. */
682 /* Exclude any concurrent CPU-hotplug operations. */
683 spin_lock(&rsp->onofflock); /* irqs already disabled. */
686 * Set the quiescent-state-needed bits in all the rcu_node
687 * structures for all currently online CPUs in breadth-first
688 * order, starting from the root rcu_node structure. This
689 * operation relies on the layout of the hierarchy within the
690 * rsp->node[] array. Note that other CPUs will access only
691 * the leaves of the hierarchy, which still indicate that no
692 * grace period is in progress, at least until the corresponding
693 * leaf node has been initialized. In addition, we have excluded
694 * CPU-hotplug operations.
696 * Note that the grace period cannot complete until we finish
697 * the initialization process, as there will be at least one
698 * qsmask bit set in the root node until that time, namely the
699 * one corresponding to this CPU, due to the fact that we have
702 rcu_for_each_node_breadth_first(rsp, rnp) {
703 spin_lock(&rnp->lock); /* irqs already disabled. */
704 rcu_preempt_check_blocked_tasks(rnp);
705 rnp->qsmask = rnp->qsmaskinit;
706 rnp->gpnum = rsp->gpnum;
707 rnp->completed = rsp->completed;
708 if (rnp == rdp->mynode)
709 rcu_start_gp_per_cpu(rsp, rnp, rdp);
710 spin_unlock(&rnp->lock); /* irqs remain disabled. */
713 rnp = rcu_get_root(rsp);
714 spin_lock(&rnp->lock); /* irqs already disabled. */
715 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
716 spin_unlock(&rnp->lock); /* irqs remain disabled. */
717 spin_unlock_irqrestore(&rsp->onofflock, flags);
721 * Clean up after the prior grace period and let rcu_start_gp() start up
722 * the next grace period if one is needed. Note that the caller must
723 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
725 static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
726 __releases(rcu_get_root(rsp)->lock)
728 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
729 rsp->completed = rsp->gpnum;
730 rsp->signaled = RCU_GP_IDLE;
731 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
735 * Similar to cpu_quiet(), for which it is a helper function. Allows
736 * a group of CPUs to be quieted at one go, though all the CPUs in the
737 * group must be represented by the same leaf rcu_node structure.
738 * That structure's lock must be held upon entry, and it is released
742 cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp,
744 __releases(rnp->lock)
746 struct rcu_node *rnp_c;
748 /* Walk up the rcu_node hierarchy. */
750 if (!(rnp->qsmask & mask)) {
752 /* Our bit has already been cleared, so done. */
753 spin_unlock_irqrestore(&rnp->lock, flags);
756 rnp->qsmask &= ~mask;
757 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
759 /* Other bits still set at this level, so done. */
760 spin_unlock_irqrestore(&rnp->lock, flags);
764 if (rnp->parent == NULL) {
766 /* No more levels. Exit loop holding root lock. */
770 spin_unlock_irqrestore(&rnp->lock, flags);
773 spin_lock_irqsave(&rnp->lock, flags);
774 WARN_ON_ONCE(rnp_c->qsmask);
778 * Get here if we are the last CPU to pass through a quiescent
779 * state for this grace period. Invoke cpu_quiet_msk_finish()
780 * to clean up and start the next grace period if one is needed.
782 cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */
786 * Record a quiescent state for the specified CPU, which must either be
787 * the current CPU. The lastcomp argument is used to make sure we are
788 * still in the grace period of interest. We don't want to end the current
789 * grace period based on quiescent states detected in an earlier grace
793 cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
797 struct rcu_node *rnp;
800 spin_lock_irqsave(&rnp->lock, flags);
801 if (lastcomp != ACCESS_ONCE(rsp->completed)) {
804 * Someone beat us to it for this grace period, so leave.
805 * The race with GP start is resolved by the fact that we
806 * hold the leaf rcu_node lock, so that the per-CPU bits
807 * cannot yet be initialized -- so we would simply find our
808 * CPU's bit already cleared in cpu_quiet_msk() if this race
811 rdp->passed_quiesc = 0; /* try again later! */
812 spin_unlock_irqrestore(&rnp->lock, flags);
816 if ((rnp->qsmask & mask) == 0) {
817 spin_unlock_irqrestore(&rnp->lock, flags);
822 * This GP can't end until cpu checks in, so all of our
823 * callbacks can be processed during the next GP.
825 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
827 cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */
832 * Check to see if there is a new grace period of which this CPU
833 * is not yet aware, and if so, set up local rcu_data state for it.
834 * Otherwise, see if this CPU has just passed through its first
835 * quiescent state for this grace period, and record that fact if so.
838 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
840 /* If there is now a new grace period, record and return. */
841 if (check_for_new_grace_period(rsp, rdp))
845 * Does this CPU still need to do its part for current grace period?
846 * If no, return and let the other CPUs do their part as well.
848 if (!rdp->qs_pending)
852 * Was there a quiescent state since the beginning of the grace
853 * period? If no, then exit and wait for the next call.
855 if (!rdp->passed_quiesc)
858 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
859 cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
862 #ifdef CONFIG_HOTPLUG_CPU
865 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
866 * specified flavor of RCU. The callbacks will be adopted by the next
867 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
868 * comes first. Because this is invoked from the CPU_DYING notifier,
869 * irqs are already disabled.
871 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
874 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
876 if (rdp->nxtlist == NULL)
877 return; /* irqs disabled, so comparison is stable. */
878 spin_lock(&rsp->onofflock); /* irqs already disabled. */
879 *rsp->orphan_cbs_tail = rdp->nxtlist;
880 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
882 for (i = 0; i < RCU_NEXT_SIZE; i++)
883 rdp->nxttail[i] = &rdp->nxtlist;
884 rsp->orphan_qlen += rdp->qlen;
886 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
890 * Adopt previously orphaned RCU callbacks.
892 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
895 struct rcu_data *rdp;
897 spin_lock_irqsave(&rsp->onofflock, flags);
898 rdp = rsp->rda[smp_processor_id()];
899 if (rsp->orphan_cbs_list == NULL) {
900 spin_unlock_irqrestore(&rsp->onofflock, flags);
903 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
904 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
905 rdp->qlen += rsp->orphan_qlen;
906 rsp->orphan_cbs_list = NULL;
907 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
908 rsp->orphan_qlen = 0;
909 spin_unlock_irqrestore(&rsp->onofflock, flags);
913 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
914 * and move all callbacks from the outgoing CPU to the current one.
916 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
921 struct rcu_data *rdp = rsp->rda[cpu];
922 struct rcu_node *rnp;
924 /* Exclude any attempts to start a new grace period. */
925 spin_lock_irqsave(&rsp->onofflock, flags);
927 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
928 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
929 mask = rdp->grpmask; /* rnp->grplo is constant. */
931 spin_lock(&rnp->lock); /* irqs already disabled. */
932 rnp->qsmaskinit &= ~mask;
933 if (rnp->qsmaskinit != 0) {
934 spin_unlock(&rnp->lock); /* irqs remain disabled. */
939 * If there was a task blocking the current grace period,
940 * and if all CPUs have checked in, we need to propagate
941 * the quiescent state up the rcu_node hierarchy. But that
942 * is inconvenient at the moment due to deadlock issues if
943 * this should end the current grace period. So set the
944 * offlined CPU's bit in ->qsmask in order to force the
945 * next force_quiescent_state() invocation to clean up this
946 * mess in a deadlock-free manner.
948 if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
952 spin_unlock(&rnp->lock); /* irqs remain disabled. */
954 } while (rnp != NULL);
955 lastcomp = rsp->completed;
957 spin_unlock_irqrestore(&rsp->onofflock, flags);
959 rcu_adopt_orphan_cbs(rsp);
963 * Remove the specified CPU from the RCU hierarchy and move any pending
964 * callbacks that it might have to the current CPU. This code assumes
965 * that at least one CPU in the system will remain running at all times.
966 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
968 static void rcu_offline_cpu(int cpu)
970 __rcu_offline_cpu(cpu, &rcu_sched_state);
971 __rcu_offline_cpu(cpu, &rcu_bh_state);
972 rcu_preempt_offline_cpu(cpu);
975 #else /* #ifdef CONFIG_HOTPLUG_CPU */
977 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
981 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
985 static void rcu_offline_cpu(int cpu)
989 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
992 * Invoke any RCU callbacks that have made it to the end of their grace
993 * period. Thottle as specified by rdp->blimit.
995 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
998 struct rcu_head *next, *list, **tail;
1001 /* If no callbacks are ready, just return.*/
1002 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1006 * Extract the list of ready callbacks, disabling to prevent
1007 * races with call_rcu() from interrupt handlers.
1009 local_irq_save(flags);
1010 list = rdp->nxtlist;
1011 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1012 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1013 tail = rdp->nxttail[RCU_DONE_TAIL];
1014 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1015 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1016 rdp->nxttail[count] = &rdp->nxtlist;
1017 local_irq_restore(flags);
1019 /* Invoke callbacks. */
1026 if (++count >= rdp->blimit)
1030 local_irq_save(flags);
1032 /* Update count, and requeue any remaining callbacks. */
1035 *tail = rdp->nxtlist;
1036 rdp->nxtlist = list;
1037 for (count = 0; count < RCU_NEXT_SIZE; count++)
1038 if (&rdp->nxtlist == rdp->nxttail[count])
1039 rdp->nxttail[count] = tail;
1044 /* Reinstate batch limit if we have worked down the excess. */
1045 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1046 rdp->blimit = blimit;
1048 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1049 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1050 rdp->qlen_last_fqs_check = 0;
1051 rdp->n_force_qs_snap = rsp->n_force_qs;
1052 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1053 rdp->qlen_last_fqs_check = rdp->qlen;
1055 local_irq_restore(flags);
1057 /* Re-raise the RCU softirq if there are callbacks remaining. */
1058 if (cpu_has_callbacks_ready_to_invoke(rdp))
1059 raise_softirq(RCU_SOFTIRQ);
1063 * Check to see if this CPU is in a non-context-switch quiescent state
1064 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1065 * Also schedule the RCU softirq handler.
1067 * This function must be called with hardirqs disabled. It is normally
1068 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1069 * false, there is no point in invoking rcu_check_callbacks().
1071 void rcu_check_callbacks(int cpu, int user)
1073 if (!rcu_pending(cpu))
1074 return; /* if nothing for RCU to do. */
1076 (idle_cpu(cpu) && rcu_scheduler_active &&
1077 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1080 * Get here if this CPU took its interrupt from user
1081 * mode or from the idle loop, and if this is not a
1082 * nested interrupt. In this case, the CPU is in
1083 * a quiescent state, so note it.
1085 * No memory barrier is required here because both
1086 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1087 * variables that other CPUs neither access nor modify,
1088 * at least not while the corresponding CPU is online.
1094 } else if (!in_softirq()) {
1097 * Get here if this CPU did not take its interrupt from
1098 * softirq, in other words, if it is not interrupting
1099 * a rcu_bh read-side critical section. This is an _bh
1100 * critical section, so note it.
1105 rcu_preempt_check_callbacks(cpu);
1106 raise_softirq(RCU_SOFTIRQ);
1112 * Scan the leaf rcu_node structures, processing dyntick state for any that
1113 * have not yet encountered a quiescent state, using the function specified.
1114 * Returns 1 if the current grace period ends while scanning (possibly
1115 * because we made it end).
1117 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1118 int (*f)(struct rcu_data *))
1122 unsigned long flags;
1124 struct rcu_node *rnp;
1126 rcu_for_each_leaf_node(rsp, rnp) {
1128 spin_lock_irqsave(&rnp->lock, flags);
1129 if (rsp->completed != lastcomp) {
1130 spin_unlock_irqrestore(&rnp->lock, flags);
1133 if (rnp->qsmask == 0) {
1134 spin_unlock_irqrestore(&rnp->lock, flags);
1139 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1140 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1143 if (mask != 0 && rsp->completed == lastcomp) {
1145 /* cpu_quiet_msk() releases rnp->lock. */
1146 cpu_quiet_msk(mask, rsp, rnp, flags);
1149 spin_unlock_irqrestore(&rnp->lock, flags);
1155 * Force quiescent states on reluctant CPUs, and also detect which
1156 * CPUs are in dyntick-idle mode.
1158 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1160 unsigned long flags;
1162 struct rcu_node *rnp = rcu_get_root(rsp);
1166 if (!rcu_gp_in_progress(rsp))
1167 return; /* No grace period in progress, nothing to force. */
1168 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1169 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1170 return; /* Someone else is already on the job. */
1173 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1174 goto unlock_ret; /* no emergency and done recently. */
1176 spin_lock(&rnp->lock);
1177 lastcomp = rsp->completed;
1178 signaled = rsp->signaled;
1179 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1180 if (lastcomp == rsp->gpnum) {
1181 rsp->n_force_qs_ngp++;
1182 spin_unlock(&rnp->lock);
1183 goto unlock_ret; /* no GP in progress, time updated. */
1185 spin_unlock(&rnp->lock);
1190 break; /* grace period idle or initializing, ignore. */
1192 case RCU_SAVE_DYNTICK:
1194 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1195 break; /* So gcc recognizes the dead code. */
1197 /* Record dyntick-idle state. */
1198 if (rcu_process_dyntick(rsp, lastcomp,
1199 dyntick_save_progress_counter))
1201 /* fall into next case. */
1203 case RCU_SAVE_COMPLETED:
1205 /* Update state, record completion counter. */
1207 spin_lock(&rnp->lock);
1208 if (lastcomp == rsp->completed &&
1209 rsp->signaled == signaled) {
1210 rsp->signaled = RCU_FORCE_QS;
1211 dyntick_record_completed(rsp, lastcomp);
1212 forcenow = signaled == RCU_SAVE_COMPLETED;
1214 spin_unlock(&rnp->lock);
1217 /* fall into next case. */
1221 /* Check dyntick-idle state, send IPI to laggarts. */
1222 if (rcu_process_dyntick(rsp, dyntick_recall_completed(rsp),
1223 rcu_implicit_dynticks_qs))
1226 /* Leave state in case more forcing is required. */
1231 spin_unlock_irqrestore(&rsp->fqslock, flags);
1234 #else /* #ifdef CONFIG_SMP */
1236 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1241 #endif /* #else #ifdef CONFIG_SMP */
1244 * This does the RCU processing work from softirq context for the
1245 * specified rcu_state and rcu_data structures. This may be called
1246 * only from the CPU to whom the rdp belongs.
1249 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1251 unsigned long flags;
1253 WARN_ON_ONCE(rdp->beenonline == 0);
1256 * If an RCU GP has gone long enough, go check for dyntick
1257 * idle CPUs and, if needed, send resched IPIs.
1259 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1260 force_quiescent_state(rsp, 1);
1263 * Advance callbacks in response to end of earlier grace
1264 * period that some other CPU ended.
1266 rcu_process_gp_end(rsp, rdp);
1268 /* Update RCU state based on any recent quiescent states. */
1269 rcu_check_quiescent_state(rsp, rdp);
1271 /* Does this CPU require a not-yet-started grace period? */
1272 if (cpu_needs_another_gp(rsp, rdp)) {
1273 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1274 rcu_start_gp(rsp, flags); /* releases above lock */
1277 /* If there are callbacks ready, invoke them. */
1278 rcu_do_batch(rsp, rdp);
1282 * Do softirq processing for the current CPU.
1284 static void rcu_process_callbacks(struct softirq_action *unused)
1287 * Memory references from any prior RCU read-side critical sections
1288 * executed by the interrupted code must be seen before any RCU
1289 * grace-period manipulations below.
1291 smp_mb(); /* See above block comment. */
1293 __rcu_process_callbacks(&rcu_sched_state,
1294 &__get_cpu_var(rcu_sched_data));
1295 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1296 rcu_preempt_process_callbacks();
1299 * Memory references from any later RCU read-side critical sections
1300 * executed by the interrupted code must be seen after any RCU
1301 * grace-period manipulations above.
1303 smp_mb(); /* See above block comment. */
1307 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1308 struct rcu_state *rsp)
1310 unsigned long flags;
1311 struct rcu_data *rdp;
1316 smp_mb(); /* Ensure RCU update seen before callback registry. */
1319 * Opportunistically note grace-period endings and beginnings.
1320 * Note that we might see a beginning right after we see an
1321 * end, but never vice versa, since this CPU has to pass through
1322 * a quiescent state betweentimes.
1324 local_irq_save(flags);
1325 rdp = rsp->rda[smp_processor_id()];
1326 rcu_process_gp_end(rsp, rdp);
1327 check_for_new_grace_period(rsp, rdp);
1329 /* Add the callback to our list. */
1330 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1331 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1333 /* Start a new grace period if one not already started. */
1334 if (!rcu_gp_in_progress(rsp)) {
1335 unsigned long nestflag;
1336 struct rcu_node *rnp_root = rcu_get_root(rsp);
1338 spin_lock_irqsave(&rnp_root->lock, nestflag);
1339 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1343 * Force the grace period if too many callbacks or too long waiting.
1344 * Enforce hysteresis, and don't invoke force_quiescent_state()
1345 * if some other CPU has recently done so. Also, don't bother
1346 * invoking force_quiescent_state() if the newly enqueued callback
1347 * is the only one waiting for a grace period to complete.
1349 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1350 rdp->blimit = LONG_MAX;
1351 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1352 *rdp->nxttail[RCU_DONE_TAIL] != head)
1353 force_quiescent_state(rsp, 0);
1354 rdp->n_force_qs_snap = rsp->n_force_qs;
1355 rdp->qlen_last_fqs_check = rdp->qlen;
1356 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1357 force_quiescent_state(rsp, 1);
1358 local_irq_restore(flags);
1362 * Queue an RCU-sched callback for invocation after a grace period.
1364 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1366 __call_rcu(head, func, &rcu_sched_state);
1368 EXPORT_SYMBOL_GPL(call_rcu_sched);
1371 * Queue an RCU for invocation after a quicker grace period.
1373 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1375 __call_rcu(head, func, &rcu_bh_state);
1377 EXPORT_SYMBOL_GPL(call_rcu_bh);
1380 * Check to see if there is any immediate RCU-related work to be done
1381 * by the current CPU, for the specified type of RCU, returning 1 if so.
1382 * The checks are in order of increasing expense: checks that can be
1383 * carried out against CPU-local state are performed first. However,
1384 * we must check for CPU stalls first, else we might not get a chance.
1386 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1388 rdp->n_rcu_pending++;
1390 /* Check for CPU stalls, if enabled. */
1391 check_cpu_stall(rsp, rdp);
1393 /* Is the RCU core waiting for a quiescent state from this CPU? */
1394 if (rdp->qs_pending) {
1395 rdp->n_rp_qs_pending++;
1399 /* Does this CPU have callbacks ready to invoke? */
1400 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1401 rdp->n_rp_cb_ready++;
1405 /* Has RCU gone idle with this CPU needing another grace period? */
1406 if (cpu_needs_another_gp(rsp, rdp)) {
1407 rdp->n_rp_cpu_needs_gp++;
1411 /* Has another RCU grace period completed? */
1412 if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */
1413 rdp->n_rp_gp_completed++;
1417 /* Has a new RCU grace period started? */
1418 if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */
1419 rdp->n_rp_gp_started++;
1423 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1424 if (rcu_gp_in_progress(rsp) &&
1425 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1426 rdp->n_rp_need_fqs++;
1431 rdp->n_rp_need_nothing++;
1436 * Check to see if there is any immediate RCU-related work to be done
1437 * by the current CPU, returning 1 if so. This function is part of the
1438 * RCU implementation; it is -not- an exported member of the RCU API.
1440 static int rcu_pending(int cpu)
1442 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1443 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1444 rcu_preempt_pending(cpu);
1448 * Check to see if any future RCU-related work will need to be done
1449 * by the current CPU, even if none need be done immediately, returning
1450 * 1 if so. This function is part of the RCU implementation; it is -not-
1451 * an exported member of the RCU API.
1453 int rcu_needs_cpu(int cpu)
1455 /* RCU callbacks either ready or pending? */
1456 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1457 per_cpu(rcu_bh_data, cpu).nxtlist ||
1458 rcu_preempt_needs_cpu(cpu);
1461 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1462 static atomic_t rcu_barrier_cpu_count;
1463 static DEFINE_MUTEX(rcu_barrier_mutex);
1464 static struct completion rcu_barrier_completion;
1466 static void rcu_barrier_callback(struct rcu_head *notused)
1468 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1469 complete(&rcu_barrier_completion);
1473 * Called with preemption disabled, and from cross-cpu IRQ context.
1475 static void rcu_barrier_func(void *type)
1477 int cpu = smp_processor_id();
1478 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1479 void (*call_rcu_func)(struct rcu_head *head,
1480 void (*func)(struct rcu_head *head));
1482 atomic_inc(&rcu_barrier_cpu_count);
1483 call_rcu_func = type;
1484 call_rcu_func(head, rcu_barrier_callback);
1488 * Orchestrate the specified type of RCU barrier, waiting for all
1489 * RCU callbacks of the specified type to complete.
1491 static void _rcu_barrier(struct rcu_state *rsp,
1492 void (*call_rcu_func)(struct rcu_head *head,
1493 void (*func)(struct rcu_head *head)))
1495 BUG_ON(in_interrupt());
1496 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1497 mutex_lock(&rcu_barrier_mutex);
1498 init_completion(&rcu_barrier_completion);
1500 * Initialize rcu_barrier_cpu_count to 1, then invoke
1501 * rcu_barrier_func() on each CPU, so that each CPU also has
1502 * incremented rcu_barrier_cpu_count. Only then is it safe to
1503 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1504 * might complete its grace period before all of the other CPUs
1505 * did their increment, causing this function to return too
1508 atomic_set(&rcu_barrier_cpu_count, 1);
1509 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1510 rcu_adopt_orphan_cbs(rsp);
1511 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1512 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1513 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1514 complete(&rcu_barrier_completion);
1515 wait_for_completion(&rcu_barrier_completion);
1516 mutex_unlock(&rcu_barrier_mutex);
1520 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1522 void rcu_barrier_bh(void)
1524 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1526 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1529 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1531 void rcu_barrier_sched(void)
1533 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1535 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1538 * Do boot-time initialization of a CPU's per-CPU RCU data.
1541 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1543 unsigned long flags;
1545 struct rcu_data *rdp = rsp->rda[cpu];
1546 struct rcu_node *rnp = rcu_get_root(rsp);
1548 /* Set up local state, ensuring consistent view of global state. */
1549 spin_lock_irqsave(&rnp->lock, flags);
1550 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1551 rdp->nxtlist = NULL;
1552 for (i = 0; i < RCU_NEXT_SIZE; i++)
1553 rdp->nxttail[i] = &rdp->nxtlist;
1556 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1557 #endif /* #ifdef CONFIG_NO_HZ */
1559 spin_unlock_irqrestore(&rnp->lock, flags);
1563 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1564 * offline event can be happening at a given time. Note also that we
1565 * can accept some slop in the rsp->completed access due to the fact
1566 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1568 static void __cpuinit
1569 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1571 unsigned long flags;
1573 struct rcu_data *rdp = rsp->rda[cpu];
1574 struct rcu_node *rnp = rcu_get_root(rsp);
1576 /* Set up local state, ensuring consistent view of global state. */
1577 spin_lock_irqsave(&rnp->lock, flags);
1578 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1579 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1580 rdp->beenonline = 1; /* We have now been online. */
1581 rdp->preemptable = preemptable;
1582 rdp->qlen_last_fqs_check = 0;
1583 rdp->n_force_qs_snap = rsp->n_force_qs;
1584 rdp->blimit = blimit;
1585 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1588 * A new grace period might start here. If so, we won't be part
1589 * of it, but that is OK, as we are currently in a quiescent state.
1592 /* Exclude any attempts to start a new GP on large systems. */
1593 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1595 /* Add CPU to rcu_node bitmasks. */
1597 mask = rdp->grpmask;
1599 /* Exclude any attempts to start a new GP on small systems. */
1600 spin_lock(&rnp->lock); /* irqs already disabled. */
1601 rnp->qsmaskinit |= mask;
1602 mask = rnp->grpmask;
1603 if (rnp == rdp->mynode) {
1604 rdp->gpnum = rnp->completed; /* if GP in progress... */
1605 rdp->completed = rnp->completed;
1606 rdp->passed_quiesc_completed = rnp->completed - 1;
1608 spin_unlock(&rnp->lock); /* irqs already disabled. */
1610 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1612 spin_unlock_irqrestore(&rsp->onofflock, flags);
1615 static void __cpuinit rcu_online_cpu(int cpu)
1617 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1618 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1619 rcu_preempt_init_percpu_data(cpu);
1623 * Handle CPU online/offline notification events.
1625 int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1626 unsigned long action, void *hcpu)
1628 long cpu = (long)hcpu;
1631 case CPU_UP_PREPARE:
1632 case CPU_UP_PREPARE_FROZEN:
1633 rcu_online_cpu(cpu);
1636 case CPU_DYING_FROZEN:
1638 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1639 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1640 * returns, all online cpus have queued rcu_barrier_func().
1641 * The dying CPU clears its cpu_online_mask bit and
1642 * moves all of its RCU callbacks to ->orphan_cbs_list
1643 * in the context of stop_machine(), so subsequent calls
1644 * to _rcu_barrier() will adopt these callbacks and only
1645 * then queue rcu_barrier_func() on all remaining CPUs.
1647 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1648 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1649 rcu_preempt_send_cbs_to_orphanage();
1652 case CPU_DEAD_FROZEN:
1653 case CPU_UP_CANCELED:
1654 case CPU_UP_CANCELED_FROZEN:
1655 rcu_offline_cpu(cpu);
1664 * Compute the per-level fanout, either using the exact fanout specified
1665 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1667 #ifdef CONFIG_RCU_FANOUT_EXACT
1668 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1672 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1673 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1675 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1676 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1683 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1684 ccur = rsp->levelcnt[i];
1685 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1689 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1692 * Helper function for rcu_init() that initializes one rcu_state structure.
1694 static void __init rcu_init_one(struct rcu_state *rsp)
1699 struct rcu_node *rnp;
1701 /* Initialize the level-tracking arrays. */
1703 for (i = 1; i < NUM_RCU_LVLS; i++)
1704 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1705 rcu_init_levelspread(rsp);
1707 /* Initialize the elements themselves, starting from the leaves. */
1709 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1710 cpustride *= rsp->levelspread[i];
1711 rnp = rsp->level[i];
1712 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1713 if (rnp != rcu_get_root(rsp))
1714 spin_lock_init(&rnp->lock);
1717 rnp->qsmaskinit = 0;
1718 rnp->grplo = j * cpustride;
1719 rnp->grphi = (j + 1) * cpustride - 1;
1720 if (rnp->grphi >= NR_CPUS)
1721 rnp->grphi = NR_CPUS - 1;
1727 rnp->grpnum = j % rsp->levelspread[i - 1];
1728 rnp->grpmask = 1UL << rnp->grpnum;
1729 rnp->parent = rsp->level[i - 1] +
1730 j / rsp->levelspread[i - 1];
1733 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1734 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1737 spin_lock_init(&rcu_get_root(rsp)->lock);
1741 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1742 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1745 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1749 struct rcu_node *rnp; \
1751 rcu_init_one(rsp); \
1752 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1754 for_each_possible_cpu(i) { \
1755 if (i > rnp[j].grphi) \
1757 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1758 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1759 rcu_boot_init_percpu_data(i, rsp); \
1763 void __init __rcu_init(void)
1765 rcu_bootup_announce();
1766 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1767 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1768 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1769 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1770 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1771 __rcu_init_preempt();
1772 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1775 #include "rcutree_plugin.h"