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>
49 #include <linux/kernel_stat.h>
53 /* Data structures. */
55 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
57 #define RCU_STATE_INITIALIZER(name) { \
58 .level = { &name.node[0] }, \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
63 NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \
65 .signaled = RCU_GP_IDLE, \
68 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
69 .orphan_cbs_list = NULL, \
70 .orphan_cbs_tail = &name.orphan_cbs_list, \
72 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
74 .n_force_qs_ngp = 0, \
77 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
78 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
80 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
81 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
83 static int rcu_scheduler_active __read_mostly;
87 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
88 * permit this function to be invoked without holding the root rcu_node
89 * structure's ->lock, but of course results can be subject to change.
91 static int rcu_gp_in_progress(struct rcu_state *rsp)
93 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
97 * Note a quiescent state. Because we do not need to know
98 * how many quiescent states passed, just if there was at least
99 * one since the start of the grace period, this just sets a flag.
101 void rcu_sched_qs(int cpu)
103 struct rcu_data *rdp;
105 rdp = &per_cpu(rcu_sched_data, cpu);
106 rdp->passed_quiesc_completed = rdp->gpnum - 1;
108 rdp->passed_quiesc = 1;
109 rcu_preempt_note_context_switch(cpu);
112 void rcu_bh_qs(int cpu)
114 struct rcu_data *rdp;
116 rdp = &per_cpu(rcu_bh_data, cpu);
117 rdp->passed_quiesc_completed = rdp->gpnum - 1;
119 rdp->passed_quiesc = 1;
123 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
124 .dynticks_nesting = 1,
127 #endif /* #ifdef CONFIG_NO_HZ */
129 static int blimit = 10; /* Maximum callbacks per softirq. */
130 static int qhimark = 10000; /* If this many pending, ignore blimit. */
131 static int qlowmark = 100; /* Once only this many pending, use blimit. */
133 module_param(blimit, int, 0);
134 module_param(qhimark, int, 0);
135 module_param(qlowmark, int, 0);
137 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
138 static int rcu_pending(int cpu);
141 * Return the number of RCU-sched batches processed thus far for debug & stats.
143 long rcu_batches_completed_sched(void)
145 return rcu_sched_state.completed;
147 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
150 * Return the number of RCU BH batches processed thus far for debug & stats.
152 long rcu_batches_completed_bh(void)
154 return rcu_bh_state.completed;
156 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
159 * Does the CPU have callbacks ready to be invoked?
162 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
164 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
168 * Does the current CPU require a yet-as-unscheduled grace period?
171 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
173 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
177 * Return the root node of the specified rcu_state structure.
179 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
181 return &rsp->node[0];
187 * If the specified CPU is offline, tell the caller that it is in
188 * a quiescent state. Otherwise, whack it with a reschedule IPI.
189 * Grace periods can end up waiting on an offline CPU when that
190 * CPU is in the process of coming online -- it will be added to the
191 * rcu_node bitmasks before it actually makes it online. The same thing
192 * can happen while a CPU is in the process of coming online. Because this
193 * race is quite rare, we check for it after detecting that the grace
194 * period has been delayed rather than checking each and every CPU
195 * each and every time we start a new grace period.
197 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
200 * If the CPU is offline, it is in a quiescent state. We can
201 * trust its state not to change because interrupts are disabled.
203 if (cpu_is_offline(rdp->cpu)) {
208 /* If preemptable RCU, no point in sending reschedule IPI. */
209 if (rdp->preemptable)
212 /* The CPU is online, so send it a reschedule IPI. */
213 if (rdp->cpu != smp_processor_id())
214 smp_send_reschedule(rdp->cpu);
221 #endif /* #ifdef CONFIG_SMP */
226 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
228 * Enter nohz mode, in other words, -leave- the mode in which RCU
229 * read-side critical sections can occur. (Though RCU read-side
230 * critical sections can occur in irq handlers in nohz mode, a possibility
231 * handled by rcu_irq_enter() and rcu_irq_exit()).
233 void rcu_enter_nohz(void)
236 struct rcu_dynticks *rdtp;
238 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
239 local_irq_save(flags);
240 rdtp = &__get_cpu_var(rcu_dynticks);
242 rdtp->dynticks_nesting--;
243 WARN_ON_ONCE(rdtp->dynticks & 0x1);
244 local_irq_restore(flags);
248 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
250 * Exit nohz mode, in other words, -enter- the mode in which RCU
251 * read-side critical sections normally occur.
253 void rcu_exit_nohz(void)
256 struct rcu_dynticks *rdtp;
258 local_irq_save(flags);
259 rdtp = &__get_cpu_var(rcu_dynticks);
261 rdtp->dynticks_nesting++;
262 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
263 local_irq_restore(flags);
264 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
268 * rcu_nmi_enter - inform RCU of entry to NMI context
270 * If the CPU was idle with dynamic ticks active, and there is no
271 * irq handler running, this updates rdtp->dynticks_nmi to let the
272 * RCU grace-period handling know that the CPU is active.
274 void rcu_nmi_enter(void)
276 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
278 if (rdtp->dynticks & 0x1)
280 rdtp->dynticks_nmi++;
281 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
282 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
286 * rcu_nmi_exit - inform RCU of exit from NMI context
288 * If the CPU was idle with dynamic ticks active, and there is no
289 * irq handler running, this updates rdtp->dynticks_nmi to let the
290 * RCU grace-period handling know that the CPU is no longer active.
292 void rcu_nmi_exit(void)
294 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
296 if (rdtp->dynticks & 0x1)
298 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
299 rdtp->dynticks_nmi++;
300 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
304 * rcu_irq_enter - inform RCU of entry to hard irq context
306 * If the CPU was idle with dynamic ticks active, this updates the
307 * rdtp->dynticks to let the RCU handling know that the CPU is active.
309 void rcu_irq_enter(void)
311 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
313 if (rdtp->dynticks_nesting++)
316 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
317 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
321 * rcu_irq_exit - inform RCU of exit from hard irq context
323 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
324 * to put let the RCU handling be aware that the CPU is going back to idle
327 void rcu_irq_exit(void)
329 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
331 if (--rdtp->dynticks_nesting)
333 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
335 WARN_ON_ONCE(rdtp->dynticks & 0x1);
337 /* If the interrupt queued a callback, get out of dyntick mode. */
338 if (__get_cpu_var(rcu_sched_data).nxtlist ||
339 __get_cpu_var(rcu_bh_data).nxtlist)
346 * Snapshot the specified CPU's dynticks counter so that we can later
347 * credit them with an implicit quiescent state. Return 1 if this CPU
348 * is in dynticks idle mode, which is an extended quiescent state.
350 static int dyntick_save_progress_counter(struct rcu_data *rdp)
356 snap = rdp->dynticks->dynticks;
357 snap_nmi = rdp->dynticks->dynticks_nmi;
358 smp_mb(); /* Order sampling of snap with end of grace period. */
359 rdp->dynticks_snap = snap;
360 rdp->dynticks_nmi_snap = snap_nmi;
361 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
368 * Return true if the specified CPU has passed through a quiescent
369 * state by virtue of being in or having passed through an dynticks
370 * idle state since the last call to dyntick_save_progress_counter()
373 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
380 curr = rdp->dynticks->dynticks;
381 snap = rdp->dynticks_snap;
382 curr_nmi = rdp->dynticks->dynticks_nmi;
383 snap_nmi = rdp->dynticks_nmi_snap;
384 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
387 * If the CPU passed through or entered a dynticks idle phase with
388 * no active irq/NMI handlers, then we can safely pretend that the CPU
389 * already acknowledged the request to pass through a quiescent
390 * state. Either way, that CPU cannot possibly be in an RCU
391 * read-side critical section that started before the beginning
392 * of the current RCU grace period.
394 if ((curr != snap || (curr & 0x1) == 0) &&
395 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
400 /* Go check for the CPU being offline. */
401 return rcu_implicit_offline_qs(rdp);
404 #endif /* #ifdef CONFIG_SMP */
406 #else /* #ifdef CONFIG_NO_HZ */
410 static int dyntick_save_progress_counter(struct rcu_data *rdp)
415 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
417 return rcu_implicit_offline_qs(rdp);
420 #endif /* #ifdef CONFIG_SMP */
422 #endif /* #else #ifdef CONFIG_NO_HZ */
424 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
426 static void record_gp_stall_check_time(struct rcu_state *rsp)
428 rsp->gp_start = jiffies;
429 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
432 static void print_other_cpu_stall(struct rcu_state *rsp)
437 struct rcu_node *rnp = rcu_get_root(rsp);
439 /* Only let one CPU complain about others per time interval. */
441 spin_lock_irqsave(&rnp->lock, flags);
442 delta = jiffies - rsp->jiffies_stall;
443 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
444 spin_unlock_irqrestore(&rnp->lock, flags);
447 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
450 * Now rat on any tasks that got kicked up to the root rcu_node
451 * due to CPU offlining.
453 rcu_print_task_stall(rnp);
454 spin_unlock_irqrestore(&rnp->lock, flags);
456 /* OK, time to rat on our buddy... */
458 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
459 rcu_for_each_leaf_node(rsp, rnp) {
460 rcu_print_task_stall(rnp);
461 if (rnp->qsmask == 0)
463 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
464 if (rnp->qsmask & (1UL << cpu))
465 printk(" %d", rnp->grplo + cpu);
467 printk(" (detected by %d, t=%ld jiffies)\n",
468 smp_processor_id(), (long)(jiffies - rsp->gp_start));
469 trigger_all_cpu_backtrace();
471 force_quiescent_state(rsp, 0); /* Kick them all. */
474 static void print_cpu_stall(struct rcu_state *rsp)
477 struct rcu_node *rnp = rcu_get_root(rsp);
479 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
480 smp_processor_id(), jiffies - rsp->gp_start);
481 trigger_all_cpu_backtrace();
483 spin_lock_irqsave(&rnp->lock, flags);
484 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
486 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
487 spin_unlock_irqrestore(&rnp->lock, flags);
489 set_need_resched(); /* kick ourselves to get things going. */
492 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
495 struct rcu_node *rnp;
497 delta = jiffies - rsp->jiffies_stall;
499 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
501 /* We haven't checked in, so go dump stack. */
502 print_cpu_stall(rsp);
504 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
506 /* They had two time units to dump stack, so complain. */
507 print_other_cpu_stall(rsp);
511 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
513 static void record_gp_stall_check_time(struct rcu_state *rsp)
517 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
521 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
524 * Update CPU-local rcu_data state to record the newly noticed grace period.
525 * This is used both when we started the grace period and when we notice
526 * that someone else started the grace period. The caller must hold the
527 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
528 * and must have irqs disabled.
530 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
532 if (rdp->gpnum != rnp->gpnum) {
534 rdp->passed_quiesc = 0;
535 rdp->gpnum = rnp->gpnum;
539 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
542 struct rcu_node *rnp;
544 local_irq_save(flags);
546 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
547 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
548 local_irq_restore(flags);
551 __note_new_gpnum(rsp, rnp, rdp);
552 spin_unlock_irqrestore(&rnp->lock, flags);
556 * Did someone else start a new RCU grace period start since we last
557 * checked? Update local state appropriately if so. Must be called
558 * on the CPU corresponding to rdp.
561 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
566 local_irq_save(flags);
567 if (rdp->gpnum != rsp->gpnum) {
568 note_new_gpnum(rsp, rdp);
571 local_irq_restore(flags);
576 * Advance this CPU's callbacks, but only if the current grace period
577 * has ended. This may be called only from the CPU to whom the rdp
578 * belongs. In addition, the corresponding leaf rcu_node structure's
579 * ->lock must be held by the caller, with irqs disabled.
582 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
584 /* Did another grace period end? */
585 if (rdp->completed != rnp->completed) {
587 /* Advance callbacks. No harm if list empty. */
588 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
589 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
590 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
592 /* Remember that we saw this grace-period completion. */
593 rdp->completed = rnp->completed;
598 * Advance this CPU's callbacks, but only if the current grace period
599 * has ended. This may be called only from the CPU to whom the rdp
603 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
606 struct rcu_node *rnp;
608 local_irq_save(flags);
610 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
611 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
612 local_irq_restore(flags);
615 __rcu_process_gp_end(rsp, rnp, rdp);
616 spin_unlock_irqrestore(&rnp->lock, flags);
620 * Do per-CPU grace-period initialization for running CPU. The caller
621 * must hold the lock of the leaf rcu_node structure corresponding to
625 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
627 /* Prior grace period ended, so advance callbacks for current CPU. */
628 __rcu_process_gp_end(rsp, rnp, rdp);
631 * Because this CPU just now started the new grace period, we know
632 * that all of its callbacks will be covered by this upcoming grace
633 * period, even the ones that were registered arbitrarily recently.
634 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
636 * Other CPUs cannot be sure exactly when the grace period started.
637 * Therefore, their recently registered callbacks must pass through
638 * an additional RCU_NEXT_READY stage, so that they will be handled
639 * by the next RCU grace period.
641 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
642 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
644 /* Set state so that this CPU will detect the next quiescent state. */
645 __note_new_gpnum(rsp, rnp, rdp);
649 * Start a new RCU grace period if warranted, re-initializing the hierarchy
650 * in preparation for detecting the next grace period. The caller must hold
651 * the root node's ->lock, which is released before return. Hard irqs must
655 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
656 __releases(rcu_get_root(rsp)->lock)
658 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
659 struct rcu_node *rnp = rcu_get_root(rsp);
661 if (!cpu_needs_another_gp(rsp, rdp)) {
662 if (rnp->completed == rsp->completed) {
663 spin_unlock_irqrestore(&rnp->lock, flags);
666 spin_unlock(&rnp->lock); /* irqs remain disabled. */
669 * Propagate new ->completed value to rcu_node structures
670 * so that other CPUs don't have to wait until the start
671 * of the next grace period to process their callbacks.
673 rcu_for_each_node_breadth_first(rsp, rnp) {
674 spin_lock(&rnp->lock); /* irqs already disabled. */
675 rnp->completed = rsp->completed;
676 spin_unlock(&rnp->lock); /* irqs remain disabled. */
678 local_irq_restore(flags);
682 /* Advance to a new grace period and initialize state. */
684 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
685 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
686 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
687 record_gp_stall_check_time(rsp);
689 /* Special-case the common single-level case. */
690 if (NUM_RCU_NODES == 1) {
691 rcu_preempt_check_blocked_tasks(rnp);
692 rnp->qsmask = rnp->qsmaskinit;
693 rnp->gpnum = rsp->gpnum;
694 rnp->completed = rsp->completed;
695 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
696 rcu_start_gp_per_cpu(rsp, rnp, rdp);
697 spin_unlock_irqrestore(&rnp->lock, flags);
701 spin_unlock(&rnp->lock); /* leave irqs disabled. */
704 /* Exclude any concurrent CPU-hotplug operations. */
705 spin_lock(&rsp->onofflock); /* irqs already disabled. */
708 * Set the quiescent-state-needed bits in all the rcu_node
709 * structures for all currently online CPUs in breadth-first
710 * order, starting from the root rcu_node structure. This
711 * operation relies on the layout of the hierarchy within the
712 * rsp->node[] array. Note that other CPUs will access only
713 * the leaves of the hierarchy, which still indicate that no
714 * grace period is in progress, at least until the corresponding
715 * leaf node has been initialized. In addition, we have excluded
716 * CPU-hotplug operations.
718 * Note that the grace period cannot complete until we finish
719 * the initialization process, as there will be at least one
720 * qsmask bit set in the root node until that time, namely the
721 * one corresponding to this CPU, due to the fact that we have
724 rcu_for_each_node_breadth_first(rsp, rnp) {
725 spin_lock(&rnp->lock); /* irqs already disabled. */
726 rcu_preempt_check_blocked_tasks(rnp);
727 rnp->qsmask = rnp->qsmaskinit;
728 rnp->gpnum = rsp->gpnum;
729 rnp->completed = rsp->completed;
730 if (rnp == rdp->mynode)
731 rcu_start_gp_per_cpu(rsp, rnp, rdp);
732 spin_unlock(&rnp->lock); /* irqs remain disabled. */
735 rnp = rcu_get_root(rsp);
736 spin_lock(&rnp->lock); /* irqs already disabled. */
737 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
738 spin_unlock(&rnp->lock); /* irqs remain disabled. */
739 spin_unlock_irqrestore(&rsp->onofflock, flags);
743 * Report a full set of quiescent states to the specified rcu_state
744 * data structure. This involves cleaning up after the prior grace
745 * period and letting rcu_start_gp() start up the next grace period
746 * if one is needed. Note that the caller must hold rnp->lock, as
747 * required by rcu_start_gp(), which will release it.
749 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
750 __releases(rcu_get_root(rsp)->lock)
752 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
753 rsp->completed = rsp->gpnum;
754 rsp->signaled = RCU_GP_IDLE;
755 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
759 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
760 * Allows quiescent states for a group of CPUs to be reported at one go
761 * to the specified rcu_node structure, though all the CPUs in the group
762 * must be represented by the same rcu_node structure (which need not be
763 * a leaf rcu_node structure, though it often will be). That structure's
764 * lock must be held upon entry, and it is released before return.
767 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
768 struct rcu_node *rnp, unsigned long flags)
769 __releases(rnp->lock)
771 struct rcu_node *rnp_c;
773 /* Walk up the rcu_node hierarchy. */
775 if (!(rnp->qsmask & mask)) {
777 /* Our bit has already been cleared, so done. */
778 spin_unlock_irqrestore(&rnp->lock, flags);
781 rnp->qsmask &= ~mask;
782 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
784 /* Other bits still set at this level, so done. */
785 spin_unlock_irqrestore(&rnp->lock, flags);
789 if (rnp->parent == NULL) {
791 /* No more levels. Exit loop holding root lock. */
795 spin_unlock_irqrestore(&rnp->lock, flags);
798 spin_lock_irqsave(&rnp->lock, flags);
799 WARN_ON_ONCE(rnp_c->qsmask);
803 * Get here if we are the last CPU to pass through a quiescent
804 * state for this grace period. Invoke rcu_report_qs_rsp()
805 * to clean up and start the next grace period if one is needed.
807 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
811 * Record a quiescent state for the specified CPU to that CPU's rcu_data
812 * structure. This must be either called from the specified CPU, or
813 * called when the specified CPU is known to be offline (and when it is
814 * also known that no other CPU is concurrently trying to help the offline
815 * CPU). The lastcomp argument is used to make sure we are still in the
816 * grace period of interest. We don't want to end the current grace period
817 * based on quiescent states detected in an earlier grace period!
820 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
824 struct rcu_node *rnp;
827 spin_lock_irqsave(&rnp->lock, flags);
828 if (lastcomp != rnp->completed) {
831 * Someone beat us to it for this grace period, so leave.
832 * The race with GP start is resolved by the fact that we
833 * hold the leaf rcu_node lock, so that the per-CPU bits
834 * cannot yet be initialized -- so we would simply find our
835 * CPU's bit already cleared in rcu_report_qs_rnp() if this
838 rdp->passed_quiesc = 0; /* try again later! */
839 spin_unlock_irqrestore(&rnp->lock, flags);
843 if ((rnp->qsmask & mask) == 0) {
844 spin_unlock_irqrestore(&rnp->lock, flags);
849 * This GP can't end until cpu checks in, so all of our
850 * callbacks can be processed during the next GP.
852 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
854 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
859 * Check to see if there is a new grace period of which this CPU
860 * is not yet aware, and if so, set up local rcu_data state for it.
861 * Otherwise, see if this CPU has just passed through its first
862 * quiescent state for this grace period, and record that fact if so.
865 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
867 /* If there is now a new grace period, record and return. */
868 if (check_for_new_grace_period(rsp, rdp))
872 * Does this CPU still need to do its part for current grace period?
873 * If no, return and let the other CPUs do their part as well.
875 if (!rdp->qs_pending)
879 * Was there a quiescent state since the beginning of the grace
880 * period? If no, then exit and wait for the next call.
882 if (!rdp->passed_quiesc)
886 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
889 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
892 #ifdef CONFIG_HOTPLUG_CPU
895 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
896 * specified flavor of RCU. The callbacks will be adopted by the next
897 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
898 * comes first. Because this is invoked from the CPU_DYING notifier,
899 * irqs are already disabled.
901 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
904 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
906 if (rdp->nxtlist == NULL)
907 return; /* irqs disabled, so comparison is stable. */
908 spin_lock(&rsp->onofflock); /* irqs already disabled. */
909 *rsp->orphan_cbs_tail = rdp->nxtlist;
910 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
912 for (i = 0; i < RCU_NEXT_SIZE; i++)
913 rdp->nxttail[i] = &rdp->nxtlist;
914 rsp->orphan_qlen += rdp->qlen;
916 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
920 * Adopt previously orphaned RCU callbacks.
922 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
925 struct rcu_data *rdp;
927 spin_lock_irqsave(&rsp->onofflock, flags);
928 rdp = rsp->rda[smp_processor_id()];
929 if (rsp->orphan_cbs_list == NULL) {
930 spin_unlock_irqrestore(&rsp->onofflock, flags);
933 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
934 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
935 rdp->qlen += rsp->orphan_qlen;
936 rsp->orphan_cbs_list = NULL;
937 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
938 rsp->orphan_qlen = 0;
939 spin_unlock_irqrestore(&rsp->onofflock, flags);
943 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
944 * and move all callbacks from the outgoing CPU to the current one.
946 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
951 struct rcu_data *rdp = rsp->rda[cpu];
952 struct rcu_node *rnp;
954 /* Exclude any attempts to start a new grace period. */
955 spin_lock_irqsave(&rsp->onofflock, flags);
957 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
958 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
959 mask = rdp->grpmask; /* rnp->grplo is constant. */
961 spin_lock(&rnp->lock); /* irqs already disabled. */
962 rnp->qsmaskinit &= ~mask;
963 if (rnp->qsmaskinit != 0) {
964 if (rnp != rdp->mynode)
965 spin_unlock(&rnp->lock); /* irqs remain disabled. */
968 if (rnp == rdp->mynode)
969 need_quiet = rcu_preempt_offline_tasks(rsp, rnp, rdp);
971 spin_unlock(&rnp->lock); /* irqs remain disabled. */
974 } while (rnp != NULL);
977 * We still hold the leaf rcu_node structure lock here, and
978 * irqs are still disabled. The reason for this subterfuge is
979 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
980 * held leads to deadlock.
982 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
985 rcu_report_unblock_qs_rnp(rnp, flags);
987 spin_unlock_irqrestore(&rnp->lock, flags);
989 rcu_adopt_orphan_cbs(rsp);
993 * Remove the specified CPU from the RCU hierarchy and move any pending
994 * callbacks that it might have to the current CPU. This code assumes
995 * that at least one CPU in the system will remain running at all times.
996 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
998 static void rcu_offline_cpu(int cpu)
1000 __rcu_offline_cpu(cpu, &rcu_sched_state);
1001 __rcu_offline_cpu(cpu, &rcu_bh_state);
1002 rcu_preempt_offline_cpu(cpu);
1005 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1007 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1011 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1015 static void rcu_offline_cpu(int cpu)
1019 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1022 * Invoke any RCU callbacks that have made it to the end of their grace
1023 * period. Thottle as specified by rdp->blimit.
1025 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1027 unsigned long flags;
1028 struct rcu_head *next, *list, **tail;
1031 /* If no callbacks are ready, just return.*/
1032 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1036 * Extract the list of ready callbacks, disabling to prevent
1037 * races with call_rcu() from interrupt handlers.
1039 local_irq_save(flags);
1040 list = rdp->nxtlist;
1041 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1042 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1043 tail = rdp->nxttail[RCU_DONE_TAIL];
1044 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1045 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1046 rdp->nxttail[count] = &rdp->nxtlist;
1047 local_irq_restore(flags);
1049 /* Invoke callbacks. */
1056 if (++count >= rdp->blimit)
1060 local_irq_save(flags);
1062 /* Update count, and requeue any remaining callbacks. */
1065 *tail = rdp->nxtlist;
1066 rdp->nxtlist = list;
1067 for (count = 0; count < RCU_NEXT_SIZE; count++)
1068 if (&rdp->nxtlist == rdp->nxttail[count])
1069 rdp->nxttail[count] = tail;
1074 /* Reinstate batch limit if we have worked down the excess. */
1075 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1076 rdp->blimit = blimit;
1078 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1079 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1080 rdp->qlen_last_fqs_check = 0;
1081 rdp->n_force_qs_snap = rsp->n_force_qs;
1082 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1083 rdp->qlen_last_fqs_check = rdp->qlen;
1085 local_irq_restore(flags);
1087 /* Re-raise the RCU softirq if there are callbacks remaining. */
1088 if (cpu_has_callbacks_ready_to_invoke(rdp))
1089 raise_softirq(RCU_SOFTIRQ);
1093 * Check to see if this CPU is in a non-context-switch quiescent state
1094 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1095 * Also schedule the RCU softirq handler.
1097 * This function must be called with hardirqs disabled. It is normally
1098 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1099 * false, there is no point in invoking rcu_check_callbacks().
1101 void rcu_check_callbacks(int cpu, int user)
1103 if (!rcu_pending(cpu))
1104 return; /* if nothing for RCU to do. */
1106 (idle_cpu(cpu) && rcu_scheduler_active &&
1107 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1110 * Get here if this CPU took its interrupt from user
1111 * mode or from the idle loop, and if this is not a
1112 * nested interrupt. In this case, the CPU is in
1113 * a quiescent state, so note it.
1115 * No memory barrier is required here because both
1116 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1117 * variables that other CPUs neither access nor modify,
1118 * at least not while the corresponding CPU is online.
1124 } else if (!in_softirq()) {
1127 * Get here if this CPU did not take its interrupt from
1128 * softirq, in other words, if it is not interrupting
1129 * a rcu_bh read-side critical section. This is an _bh
1130 * critical section, so note it.
1135 rcu_preempt_check_callbacks(cpu);
1136 raise_softirq(RCU_SOFTIRQ);
1142 * Scan the leaf rcu_node structures, processing dyntick state for any that
1143 * have not yet encountered a quiescent state, using the function specified.
1144 * Returns 1 if the current grace period ends while scanning (possibly
1145 * because we made it end).
1147 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1148 int (*f)(struct rcu_data *))
1152 unsigned long flags;
1154 struct rcu_node *rnp;
1156 rcu_for_each_leaf_node(rsp, rnp) {
1158 spin_lock_irqsave(&rnp->lock, flags);
1159 if (rnp->completed != lastcomp) {
1160 spin_unlock_irqrestore(&rnp->lock, flags);
1163 if (rnp->qsmask == 0) {
1164 spin_unlock_irqrestore(&rnp->lock, flags);
1169 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1170 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1173 if (mask != 0 && rnp->completed == lastcomp) {
1175 /* rcu_report_qs_rnp() releases rnp->lock. */
1176 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1179 spin_unlock_irqrestore(&rnp->lock, flags);
1185 * Force quiescent states on reluctant CPUs, and also detect which
1186 * CPUs are in dyntick-idle mode.
1188 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1190 unsigned long flags;
1192 struct rcu_node *rnp = rcu_get_root(rsp);
1196 if (!rcu_gp_in_progress(rsp))
1197 return; /* No grace period in progress, nothing to force. */
1198 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1199 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1200 return; /* Someone else is already on the job. */
1203 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1204 goto unlock_ret; /* no emergency and done recently. */
1206 spin_lock(&rnp->lock);
1207 lastcomp = rsp->gpnum - 1;
1208 signaled = rsp->signaled;
1209 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1210 if(!rcu_gp_in_progress(rsp)) {
1211 rsp->n_force_qs_ngp++;
1212 spin_unlock(&rnp->lock);
1213 goto unlock_ret; /* no GP in progress, time updated. */
1215 spin_unlock(&rnp->lock);
1220 break; /* grace period idle or initializing, ignore. */
1222 case RCU_SAVE_DYNTICK:
1224 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1225 break; /* So gcc recognizes the dead code. */
1227 /* Record dyntick-idle state. */
1228 if (rcu_process_dyntick(rsp, lastcomp,
1229 dyntick_save_progress_counter))
1231 /* fall into next case. */
1233 case RCU_SAVE_COMPLETED:
1235 /* Update state, record completion counter. */
1237 spin_lock(&rnp->lock);
1238 if (lastcomp + 1 == rsp->gpnum &&
1239 lastcomp == rsp->completed &&
1240 rsp->signaled == signaled) {
1241 rsp->signaled = RCU_FORCE_QS;
1242 rsp->completed_fqs = lastcomp;
1243 forcenow = signaled == RCU_SAVE_COMPLETED;
1245 spin_unlock(&rnp->lock);
1248 /* fall into next case. */
1252 /* Check dyntick-idle state, send IPI to laggarts. */
1253 if (rcu_process_dyntick(rsp, rsp->completed_fqs,
1254 rcu_implicit_dynticks_qs))
1257 /* Leave state in case more forcing is required. */
1262 spin_unlock_irqrestore(&rsp->fqslock, flags);
1265 #else /* #ifdef CONFIG_SMP */
1267 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1272 #endif /* #else #ifdef CONFIG_SMP */
1275 * This does the RCU processing work from softirq context for the
1276 * specified rcu_state and rcu_data structures. This may be called
1277 * only from the CPU to whom the rdp belongs.
1280 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1282 unsigned long flags;
1284 WARN_ON_ONCE(rdp->beenonline == 0);
1287 * If an RCU GP has gone long enough, go check for dyntick
1288 * idle CPUs and, if needed, send resched IPIs.
1290 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1291 force_quiescent_state(rsp, 1);
1294 * Advance callbacks in response to end of earlier grace
1295 * period that some other CPU ended.
1297 rcu_process_gp_end(rsp, rdp);
1299 /* Update RCU state based on any recent quiescent states. */
1300 rcu_check_quiescent_state(rsp, rdp);
1302 /* Does this CPU require a not-yet-started grace period? */
1303 if (cpu_needs_another_gp(rsp, rdp)) {
1304 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1305 rcu_start_gp(rsp, flags); /* releases above lock */
1308 /* If there are callbacks ready, invoke them. */
1309 rcu_do_batch(rsp, rdp);
1313 * Do softirq processing for the current CPU.
1315 static void rcu_process_callbacks(struct softirq_action *unused)
1318 * Memory references from any prior RCU read-side critical sections
1319 * executed by the interrupted code must be seen before any RCU
1320 * grace-period manipulations below.
1322 smp_mb(); /* See above block comment. */
1324 __rcu_process_callbacks(&rcu_sched_state,
1325 &__get_cpu_var(rcu_sched_data));
1326 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1327 rcu_preempt_process_callbacks();
1330 * Memory references from any later RCU read-side critical sections
1331 * executed by the interrupted code must be seen after any RCU
1332 * grace-period manipulations above.
1334 smp_mb(); /* See above block comment. */
1338 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1339 struct rcu_state *rsp)
1341 unsigned long flags;
1342 struct rcu_data *rdp;
1347 smp_mb(); /* Ensure RCU update seen before callback registry. */
1350 * Opportunistically note grace-period endings and beginnings.
1351 * Note that we might see a beginning right after we see an
1352 * end, but never vice versa, since this CPU has to pass through
1353 * a quiescent state betweentimes.
1355 local_irq_save(flags);
1356 rdp = rsp->rda[smp_processor_id()];
1357 rcu_process_gp_end(rsp, rdp);
1358 check_for_new_grace_period(rsp, rdp);
1360 /* Add the callback to our list. */
1361 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1362 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1364 /* Start a new grace period if one not already started. */
1365 if (!rcu_gp_in_progress(rsp)) {
1366 unsigned long nestflag;
1367 struct rcu_node *rnp_root = rcu_get_root(rsp);
1369 spin_lock_irqsave(&rnp_root->lock, nestflag);
1370 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1374 * Force the grace period if too many callbacks or too long waiting.
1375 * Enforce hysteresis, and don't invoke force_quiescent_state()
1376 * if some other CPU has recently done so. Also, don't bother
1377 * invoking force_quiescent_state() if the newly enqueued callback
1378 * is the only one waiting for a grace period to complete.
1380 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1381 rdp->blimit = LONG_MAX;
1382 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1383 *rdp->nxttail[RCU_DONE_TAIL] != head)
1384 force_quiescent_state(rsp, 0);
1385 rdp->n_force_qs_snap = rsp->n_force_qs;
1386 rdp->qlen_last_fqs_check = rdp->qlen;
1387 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1388 force_quiescent_state(rsp, 1);
1389 local_irq_restore(flags);
1393 * Queue an RCU-sched callback for invocation after a grace period.
1395 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1397 __call_rcu(head, func, &rcu_sched_state);
1399 EXPORT_SYMBOL_GPL(call_rcu_sched);
1402 * Queue an RCU for invocation after a quicker grace period.
1404 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1406 __call_rcu(head, func, &rcu_bh_state);
1408 EXPORT_SYMBOL_GPL(call_rcu_bh);
1411 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1413 * Control will return to the caller some time after a full rcu-sched
1414 * grace period has elapsed, in other words after all currently executing
1415 * rcu-sched read-side critical sections have completed. These read-side
1416 * critical sections are delimited by rcu_read_lock_sched() and
1417 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1418 * local_irq_disable(), and so on may be used in place of
1419 * rcu_read_lock_sched().
1421 * This means that all preempt_disable code sequences, including NMI and
1422 * hardware-interrupt handlers, in progress on entry will have completed
1423 * before this primitive returns. However, this does not guarantee that
1424 * softirq handlers will have completed, since in some kernels, these
1425 * handlers can run in process context, and can block.
1427 * This primitive provides the guarantees made by the (now removed)
1428 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1429 * guarantees that rcu_read_lock() sections will have completed.
1430 * In "classic RCU", these two guarantees happen to be one and
1431 * the same, but can differ in realtime RCU implementations.
1433 void synchronize_sched(void)
1435 struct rcu_synchronize rcu;
1437 if (rcu_blocking_is_gp())
1440 init_completion(&rcu.completion);
1441 /* Will wake me after RCU finished. */
1442 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1444 wait_for_completion(&rcu.completion);
1446 EXPORT_SYMBOL_GPL(synchronize_sched);
1449 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1451 * Control will return to the caller some time after a full rcu_bh grace
1452 * period has elapsed, in other words after all currently executing rcu_bh
1453 * read-side critical sections have completed. RCU read-side critical
1454 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1455 * and may be nested.
1457 void synchronize_rcu_bh(void)
1459 struct rcu_synchronize rcu;
1461 if (rcu_blocking_is_gp())
1464 init_completion(&rcu.completion);
1465 /* Will wake me after RCU finished. */
1466 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1468 wait_for_completion(&rcu.completion);
1470 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1473 * Check to see if there is any immediate RCU-related work to be done
1474 * by the current CPU, for the specified type of RCU, returning 1 if so.
1475 * The checks are in order of increasing expense: checks that can be
1476 * carried out against CPU-local state are performed first. However,
1477 * we must check for CPU stalls first, else we might not get a chance.
1479 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1481 struct rcu_node *rnp = rdp->mynode;
1483 rdp->n_rcu_pending++;
1485 /* Check for CPU stalls, if enabled. */
1486 check_cpu_stall(rsp, rdp);
1488 /* Is the RCU core waiting for a quiescent state from this CPU? */
1489 if (rdp->qs_pending) {
1490 rdp->n_rp_qs_pending++;
1494 /* Does this CPU have callbacks ready to invoke? */
1495 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1496 rdp->n_rp_cb_ready++;
1500 /* Has RCU gone idle with this CPU needing another grace period? */
1501 if (cpu_needs_another_gp(rsp, rdp)) {
1502 rdp->n_rp_cpu_needs_gp++;
1506 /* Has another RCU grace period completed? */
1507 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1508 rdp->n_rp_gp_completed++;
1512 /* Has a new RCU grace period started? */
1513 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1514 rdp->n_rp_gp_started++;
1518 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1519 if (rcu_gp_in_progress(rsp) &&
1520 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1521 rdp->n_rp_need_fqs++;
1526 rdp->n_rp_need_nothing++;
1531 * Check to see if there is any immediate RCU-related work to be done
1532 * by the current CPU, returning 1 if so. This function is part of the
1533 * RCU implementation; it is -not- an exported member of the RCU API.
1535 static int rcu_pending(int cpu)
1537 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1538 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1539 rcu_preempt_pending(cpu);
1543 * Check to see if any future RCU-related work will need to be done
1544 * by the current CPU, even if none need be done immediately, returning
1545 * 1 if so. This function is part of the RCU implementation; it is -not-
1546 * an exported member of the RCU API.
1548 int rcu_needs_cpu(int cpu)
1550 /* RCU callbacks either ready or pending? */
1551 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1552 per_cpu(rcu_bh_data, cpu).nxtlist ||
1553 rcu_preempt_needs_cpu(cpu);
1557 * This function is invoked towards the end of the scheduler's initialization
1558 * process. Before this is called, the idle task might contain
1559 * RCU read-side critical sections (during which time, this idle
1560 * task is booting the system). After this function is called, the
1561 * idle tasks are prohibited from containing RCU read-side critical
1564 void rcu_scheduler_starting(void)
1566 WARN_ON(num_online_cpus() != 1);
1567 WARN_ON(nr_context_switches() > 0);
1568 rcu_scheduler_active = 1;
1571 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1572 static atomic_t rcu_barrier_cpu_count;
1573 static DEFINE_MUTEX(rcu_barrier_mutex);
1574 static struct completion rcu_barrier_completion;
1576 static void rcu_barrier_callback(struct rcu_head *notused)
1578 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1579 complete(&rcu_barrier_completion);
1583 * Called with preemption disabled, and from cross-cpu IRQ context.
1585 static void rcu_barrier_func(void *type)
1587 int cpu = smp_processor_id();
1588 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1589 void (*call_rcu_func)(struct rcu_head *head,
1590 void (*func)(struct rcu_head *head));
1592 atomic_inc(&rcu_barrier_cpu_count);
1593 call_rcu_func = type;
1594 call_rcu_func(head, rcu_barrier_callback);
1598 * Orchestrate the specified type of RCU barrier, waiting for all
1599 * RCU callbacks of the specified type to complete.
1601 static void _rcu_barrier(struct rcu_state *rsp,
1602 void (*call_rcu_func)(struct rcu_head *head,
1603 void (*func)(struct rcu_head *head)))
1605 BUG_ON(in_interrupt());
1606 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1607 mutex_lock(&rcu_barrier_mutex);
1608 init_completion(&rcu_barrier_completion);
1610 * Initialize rcu_barrier_cpu_count to 1, then invoke
1611 * rcu_barrier_func() on each CPU, so that each CPU also has
1612 * incremented rcu_barrier_cpu_count. Only then is it safe to
1613 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1614 * might complete its grace period before all of the other CPUs
1615 * did their increment, causing this function to return too
1618 atomic_set(&rcu_barrier_cpu_count, 1);
1619 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1620 rcu_adopt_orphan_cbs(rsp);
1621 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1622 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1623 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1624 complete(&rcu_barrier_completion);
1625 wait_for_completion(&rcu_barrier_completion);
1626 mutex_unlock(&rcu_barrier_mutex);
1630 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1632 void rcu_barrier_bh(void)
1634 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1636 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1639 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1641 void rcu_barrier_sched(void)
1643 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1645 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1648 * Do boot-time initialization of a CPU's per-CPU RCU data.
1651 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1653 unsigned long flags;
1655 struct rcu_data *rdp = rsp->rda[cpu];
1656 struct rcu_node *rnp = rcu_get_root(rsp);
1658 /* Set up local state, ensuring consistent view of global state. */
1659 spin_lock_irqsave(&rnp->lock, flags);
1660 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1661 rdp->nxtlist = NULL;
1662 for (i = 0; i < RCU_NEXT_SIZE; i++)
1663 rdp->nxttail[i] = &rdp->nxtlist;
1666 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1667 #endif /* #ifdef CONFIG_NO_HZ */
1669 spin_unlock_irqrestore(&rnp->lock, flags);
1673 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1674 * offline event can be happening at a given time. Note also that we
1675 * can accept some slop in the rsp->completed access due to the fact
1676 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1678 static void __cpuinit
1679 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1681 unsigned long flags;
1683 struct rcu_data *rdp = rsp->rda[cpu];
1684 struct rcu_node *rnp = rcu_get_root(rsp);
1686 /* Set up local state, ensuring consistent view of global state. */
1687 spin_lock_irqsave(&rnp->lock, flags);
1688 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1689 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1690 rdp->beenonline = 1; /* We have now been online. */
1691 rdp->preemptable = preemptable;
1692 rdp->qlen_last_fqs_check = 0;
1693 rdp->n_force_qs_snap = rsp->n_force_qs;
1694 rdp->blimit = blimit;
1695 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1698 * A new grace period might start here. If so, we won't be part
1699 * of it, but that is OK, as we are currently in a quiescent state.
1702 /* Exclude any attempts to start a new GP on large systems. */
1703 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1705 /* Add CPU to rcu_node bitmasks. */
1707 mask = rdp->grpmask;
1709 /* Exclude any attempts to start a new GP on small systems. */
1710 spin_lock(&rnp->lock); /* irqs already disabled. */
1711 rnp->qsmaskinit |= mask;
1712 mask = rnp->grpmask;
1713 if (rnp == rdp->mynode) {
1714 rdp->gpnum = rnp->completed; /* if GP in progress... */
1715 rdp->completed = rnp->completed;
1716 rdp->passed_quiesc_completed = rnp->completed - 1;
1718 spin_unlock(&rnp->lock); /* irqs already disabled. */
1720 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1722 spin_unlock_irqrestore(&rsp->onofflock, flags);
1725 static void __cpuinit rcu_online_cpu(int cpu)
1727 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1728 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1729 rcu_preempt_init_percpu_data(cpu);
1733 * Handle CPU online/offline notification events.
1735 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1736 unsigned long action, void *hcpu)
1738 long cpu = (long)hcpu;
1741 case CPU_UP_PREPARE:
1742 case CPU_UP_PREPARE_FROZEN:
1743 rcu_online_cpu(cpu);
1746 case CPU_DYING_FROZEN:
1748 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1749 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1750 * returns, all online cpus have queued rcu_barrier_func().
1751 * The dying CPU clears its cpu_online_mask bit and
1752 * moves all of its RCU callbacks to ->orphan_cbs_list
1753 * in the context of stop_machine(), so subsequent calls
1754 * to _rcu_barrier() will adopt these callbacks and only
1755 * then queue rcu_barrier_func() on all remaining CPUs.
1757 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1758 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1759 rcu_preempt_send_cbs_to_orphanage();
1762 case CPU_DEAD_FROZEN:
1763 case CPU_UP_CANCELED:
1764 case CPU_UP_CANCELED_FROZEN:
1765 rcu_offline_cpu(cpu);
1774 * Compute the per-level fanout, either using the exact fanout specified
1775 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1777 #ifdef CONFIG_RCU_FANOUT_EXACT
1778 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1782 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1783 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1785 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1786 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1793 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1794 ccur = rsp->levelcnt[i];
1795 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1799 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1802 * Helper function for rcu_init() that initializes one rcu_state structure.
1804 static void __init rcu_init_one(struct rcu_state *rsp)
1809 struct rcu_node *rnp;
1811 /* Initialize the level-tracking arrays. */
1813 for (i = 1; i < NUM_RCU_LVLS; i++)
1814 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1815 rcu_init_levelspread(rsp);
1817 /* Initialize the elements themselves, starting from the leaves. */
1819 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1820 cpustride *= rsp->levelspread[i];
1821 rnp = rsp->level[i];
1822 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1823 spin_lock_init(&rnp->lock);
1824 lockdep_set_class(&rnp->lock, &rcu_node_class[i]);
1827 rnp->qsmaskinit = 0;
1828 rnp->grplo = j * cpustride;
1829 rnp->grphi = (j + 1) * cpustride - 1;
1830 if (rnp->grphi >= NR_CPUS)
1831 rnp->grphi = NR_CPUS - 1;
1837 rnp->grpnum = j % rsp->levelspread[i - 1];
1838 rnp->grpmask = 1UL << rnp->grpnum;
1839 rnp->parent = rsp->level[i - 1] +
1840 j / rsp->levelspread[i - 1];
1843 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1844 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1850 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1851 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1854 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1858 struct rcu_node *rnp; \
1860 rcu_init_one(rsp); \
1861 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1863 for_each_possible_cpu(i) { \
1864 if (i > rnp[j].grphi) \
1866 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1867 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1868 rcu_boot_init_percpu_data(i, rsp); \
1872 void __init rcu_init(void)
1876 rcu_bootup_announce();
1877 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1878 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1879 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1880 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1881 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1882 __rcu_init_preempt();
1883 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1886 * We don't need protection against CPU-hotplug here because
1887 * this is called early in boot, before either interrupts
1888 * or the scheduler are operational.
1890 cpu_notifier(rcu_cpu_notify, 0);
1891 for_each_online_cpu(i)
1892 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)i);
1895 #include "rcutree_plugin.h"