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(structname) { \
58 .level = { &structname.node[0] }, \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
64 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
66 .signaled = RCU_GP_IDLE, \
69 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
70 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
72 .n_force_qs_ngp = 0, \
73 .name = #structname, \
76 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
77 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
79 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
80 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
82 int rcu_scheduler_active __read_mostly;
83 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
86 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
87 * permit this function to be invoked without holding the root rcu_node
88 * structure's ->lock, but of course results can be subject to change.
90 static int rcu_gp_in_progress(struct rcu_state *rsp)
92 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
96 * Note a quiescent state. Because we do not need to know
97 * how many quiescent states passed, just if there was at least
98 * one since the start of the grace period, this just sets a flag.
100 void rcu_sched_qs(int cpu)
102 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
104 rdp->passed_quiesc_completed = rdp->gpnum - 1;
106 rdp->passed_quiesc = 1;
109 void rcu_bh_qs(int cpu)
111 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
113 rdp->passed_quiesc_completed = rdp->gpnum - 1;
115 rdp->passed_quiesc = 1;
119 * Note a context switch. This is a quiescent state for RCU-sched,
120 * and requires special handling for preemptible RCU.
122 void rcu_note_context_switch(int cpu)
125 rcu_preempt_note_context_switch(cpu);
129 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
130 .dynticks_nesting = 1,
133 #endif /* #ifdef CONFIG_NO_HZ */
135 static int blimit = 10; /* Maximum callbacks per softirq. */
136 static int qhimark = 10000; /* If this many pending, ignore blimit. */
137 static int qlowmark = 100; /* Once only this many pending, use blimit. */
139 module_param(blimit, int, 0);
140 module_param(qhimark, int, 0);
141 module_param(qlowmark, int, 0);
143 int rcu_cpu_stall_suppress __read_mostly;
144 module_param(rcu_cpu_stall_suppress, int, 0644);
146 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
147 static int rcu_pending(int cpu);
150 * Return the number of RCU-sched batches processed thus far for debug & stats.
152 long rcu_batches_completed_sched(void)
154 return rcu_sched_state.completed;
156 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
159 * Return the number of RCU BH batches processed thus far for debug & stats.
161 long rcu_batches_completed_bh(void)
163 return rcu_bh_state.completed;
165 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
168 * Force a quiescent state for RCU BH.
170 void rcu_bh_force_quiescent_state(void)
172 force_quiescent_state(&rcu_bh_state, 0);
174 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
177 * Force a quiescent state for RCU-sched.
179 void rcu_sched_force_quiescent_state(void)
181 force_quiescent_state(&rcu_sched_state, 0);
183 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
186 * Does the CPU have callbacks ready to be invoked?
189 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
191 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
195 * Does the current CPU require a yet-as-unscheduled grace period?
198 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
200 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
204 * Return the root node of the specified rcu_state structure.
206 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
208 return &rsp->node[0];
214 * If the specified CPU is offline, tell the caller that it is in
215 * a quiescent state. Otherwise, whack it with a reschedule IPI.
216 * Grace periods can end up waiting on an offline CPU when that
217 * CPU is in the process of coming online -- it will be added to the
218 * rcu_node bitmasks before it actually makes it online. The same thing
219 * can happen while a CPU is in the process of coming online. Because this
220 * race is quite rare, we check for it after detecting that the grace
221 * period has been delayed rather than checking each and every CPU
222 * each and every time we start a new grace period.
224 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
227 * If the CPU is offline, it is in a quiescent state. We can
228 * trust its state not to change because interrupts are disabled.
230 if (cpu_is_offline(rdp->cpu)) {
235 /* If preemptable RCU, no point in sending reschedule IPI. */
236 if (rdp->preemptable)
239 /* The CPU is online, so send it a reschedule IPI. */
240 if (rdp->cpu != smp_processor_id())
241 smp_send_reschedule(rdp->cpu);
248 #endif /* #ifdef CONFIG_SMP */
253 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
255 * Enter nohz mode, in other words, -leave- the mode in which RCU
256 * read-side critical sections can occur. (Though RCU read-side
257 * critical sections can occur in irq handlers in nohz mode, a possibility
258 * handled by rcu_irq_enter() and rcu_irq_exit()).
260 void rcu_enter_nohz(void)
263 struct rcu_dynticks *rdtp;
265 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
266 local_irq_save(flags);
267 rdtp = &__get_cpu_var(rcu_dynticks);
269 rdtp->dynticks_nesting--;
270 WARN_ON_ONCE(rdtp->dynticks & 0x1);
271 local_irq_restore(flags);
275 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
277 * Exit nohz mode, in other words, -enter- the mode in which RCU
278 * read-side critical sections normally occur.
280 void rcu_exit_nohz(void)
283 struct rcu_dynticks *rdtp;
285 local_irq_save(flags);
286 rdtp = &__get_cpu_var(rcu_dynticks);
288 rdtp->dynticks_nesting++;
289 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
290 local_irq_restore(flags);
291 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
295 * rcu_nmi_enter - inform RCU of entry to NMI context
297 * If the CPU was idle with dynamic ticks active, and there is no
298 * irq handler running, this updates rdtp->dynticks_nmi to let the
299 * RCU grace-period handling know that the CPU is active.
301 void rcu_nmi_enter(void)
303 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
305 if (rdtp->dynticks & 0x1)
307 rdtp->dynticks_nmi++;
308 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
309 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
313 * rcu_nmi_exit - inform RCU of exit from NMI context
315 * If the CPU was idle with dynamic ticks active, and there is no
316 * irq handler running, this updates rdtp->dynticks_nmi to let the
317 * RCU grace-period handling know that the CPU is no longer active.
319 void rcu_nmi_exit(void)
321 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
323 if (rdtp->dynticks & 0x1)
325 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
326 rdtp->dynticks_nmi++;
327 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
331 * rcu_irq_enter - inform RCU of entry to hard irq context
333 * If the CPU was idle with dynamic ticks active, this updates the
334 * rdtp->dynticks to let the RCU handling know that the CPU is active.
336 void rcu_irq_enter(void)
338 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
340 if (rdtp->dynticks_nesting++)
343 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
344 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
348 * rcu_irq_exit - inform RCU of exit from hard irq context
350 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
351 * to put let the RCU handling be aware that the CPU is going back to idle
354 void rcu_irq_exit(void)
356 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
358 if (--rdtp->dynticks_nesting)
360 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
362 WARN_ON_ONCE(rdtp->dynticks & 0x1);
364 /* If the interrupt queued a callback, get out of dyntick mode. */
365 if (__this_cpu_read(rcu_sched_data.nxtlist) ||
366 __this_cpu_read(rcu_bh_data.nxtlist))
373 * Snapshot the specified CPU's dynticks counter so that we can later
374 * credit them with an implicit quiescent state. Return 1 if this CPU
375 * is in dynticks idle mode, which is an extended quiescent state.
377 static int dyntick_save_progress_counter(struct rcu_data *rdp)
383 snap = rdp->dynticks->dynticks;
384 snap_nmi = rdp->dynticks->dynticks_nmi;
385 smp_mb(); /* Order sampling of snap with end of grace period. */
386 rdp->dynticks_snap = snap;
387 rdp->dynticks_nmi_snap = snap_nmi;
388 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
395 * Return true if the specified CPU has passed through a quiescent
396 * state by virtue of being in or having passed through an dynticks
397 * idle state since the last call to dyntick_save_progress_counter()
400 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
407 curr = rdp->dynticks->dynticks;
408 snap = rdp->dynticks_snap;
409 curr_nmi = rdp->dynticks->dynticks_nmi;
410 snap_nmi = rdp->dynticks_nmi_snap;
411 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
414 * If the CPU passed through or entered a dynticks idle phase with
415 * no active irq/NMI handlers, then we can safely pretend that the CPU
416 * already acknowledged the request to pass through a quiescent
417 * state. Either way, that CPU cannot possibly be in an RCU
418 * read-side critical section that started before the beginning
419 * of the current RCU grace period.
421 if ((curr != snap || (curr & 0x1) == 0) &&
422 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
427 /* Go check for the CPU being offline. */
428 return rcu_implicit_offline_qs(rdp);
431 #endif /* #ifdef CONFIG_SMP */
433 #else /* #ifdef CONFIG_NO_HZ */
437 static int dyntick_save_progress_counter(struct rcu_data *rdp)
442 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
444 return rcu_implicit_offline_qs(rdp);
447 #endif /* #ifdef CONFIG_SMP */
449 #endif /* #else #ifdef CONFIG_NO_HZ */
451 int rcu_cpu_stall_suppress __read_mostly;
453 static void record_gp_stall_check_time(struct rcu_state *rsp)
455 rsp->gp_start = jiffies;
456 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
459 static void print_other_cpu_stall(struct rcu_state *rsp)
464 struct rcu_node *rnp = rcu_get_root(rsp);
466 /* Only let one CPU complain about others per time interval. */
468 raw_spin_lock_irqsave(&rnp->lock, flags);
469 delta = jiffies - rsp->jiffies_stall;
470 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
471 raw_spin_unlock_irqrestore(&rnp->lock, flags);
474 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
477 * Now rat on any tasks that got kicked up to the root rcu_node
478 * due to CPU offlining.
480 rcu_print_task_stall(rnp);
481 raw_spin_unlock_irqrestore(&rnp->lock, flags);
484 * OK, time to rat on our buddy...
485 * See Documentation/RCU/stallwarn.txt for info on how to debug
486 * RCU CPU stall warnings.
488 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
490 rcu_for_each_leaf_node(rsp, rnp) {
491 raw_spin_lock_irqsave(&rnp->lock, flags);
492 rcu_print_task_stall(rnp);
493 raw_spin_unlock_irqrestore(&rnp->lock, flags);
494 if (rnp->qsmask == 0)
496 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
497 if (rnp->qsmask & (1UL << cpu))
498 printk(" %d", rnp->grplo + cpu);
500 printk("} (detected by %d, t=%ld jiffies)\n",
501 smp_processor_id(), (long)(jiffies - rsp->gp_start));
502 trigger_all_cpu_backtrace();
504 /* If so configured, complain about tasks blocking the grace period. */
506 rcu_print_detail_task_stall(rsp);
508 force_quiescent_state(rsp, 0); /* Kick them all. */
511 static void print_cpu_stall(struct rcu_state *rsp)
514 struct rcu_node *rnp = rcu_get_root(rsp);
517 * OK, time to rat on ourselves...
518 * See Documentation/RCU/stallwarn.txt for info on how to debug
519 * RCU CPU stall warnings.
521 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
522 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
523 trigger_all_cpu_backtrace();
525 raw_spin_lock_irqsave(&rnp->lock, flags);
526 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
528 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
529 raw_spin_unlock_irqrestore(&rnp->lock, flags);
531 set_need_resched(); /* kick ourselves to get things going. */
534 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
537 struct rcu_node *rnp;
539 if (rcu_cpu_stall_suppress)
541 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
543 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && delta >= 0) {
545 /* We haven't checked in, so go dump stack. */
546 print_cpu_stall(rsp);
548 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
550 /* They had two time units to dump stack, so complain. */
551 print_other_cpu_stall(rsp);
555 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
557 rcu_cpu_stall_suppress = 1;
562 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
564 * Set the stall-warning timeout way off into the future, thus preventing
565 * any RCU CPU stall-warning messages from appearing in the current set of
568 * The caller must disable hard irqs.
570 void rcu_cpu_stall_reset(void)
572 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
573 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
574 rcu_preempt_stall_reset();
577 static struct notifier_block rcu_panic_block = {
578 .notifier_call = rcu_panic,
581 static void __init check_cpu_stall_init(void)
583 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
587 * Update CPU-local rcu_data state to record the newly noticed grace period.
588 * This is used both when we started the grace period and when we notice
589 * that someone else started the grace period. The caller must hold the
590 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
591 * and must have irqs disabled.
593 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
595 if (rdp->gpnum != rnp->gpnum) {
597 * If the current grace period is waiting for this CPU,
598 * set up to detect a quiescent state, otherwise don't
599 * go looking for one.
601 rdp->gpnum = rnp->gpnum;
602 if (rnp->qsmask & rdp->grpmask) {
604 rdp->passed_quiesc = 0;
610 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
613 struct rcu_node *rnp;
615 local_irq_save(flags);
617 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
618 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
619 local_irq_restore(flags);
622 __note_new_gpnum(rsp, rnp, rdp);
623 raw_spin_unlock_irqrestore(&rnp->lock, flags);
627 * Did someone else start a new RCU grace period start since we last
628 * checked? Update local state appropriately if so. Must be called
629 * on the CPU corresponding to rdp.
632 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
637 local_irq_save(flags);
638 if (rdp->gpnum != rsp->gpnum) {
639 note_new_gpnum(rsp, rdp);
642 local_irq_restore(flags);
647 * Advance this CPU's callbacks, but only if the current grace period
648 * has ended. This may be called only from the CPU to whom the rdp
649 * belongs. In addition, the corresponding leaf rcu_node structure's
650 * ->lock must be held by the caller, with irqs disabled.
653 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
655 /* Did another grace period end? */
656 if (rdp->completed != rnp->completed) {
658 /* Advance callbacks. No harm if list empty. */
659 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
660 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
661 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
663 /* Remember that we saw this grace-period completion. */
664 rdp->completed = rnp->completed;
667 * If we were in an extended quiescent state, we may have
668 * missed some grace periods that others CPUs handled on
669 * our behalf. Catch up with this state to avoid noting
670 * spurious new grace periods. If another grace period
671 * has started, then rnp->gpnum will have advanced, so
672 * we will detect this later on.
674 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
675 rdp->gpnum = rdp->completed;
678 * If RCU does not need a quiescent state from this CPU,
679 * then make sure that this CPU doesn't go looking for one.
681 if ((rnp->qsmask & rdp->grpmask) == 0)
687 * Advance this CPU's callbacks, but only if the current grace period
688 * has ended. This may be called only from the CPU to whom the rdp
692 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
695 struct rcu_node *rnp;
697 local_irq_save(flags);
699 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
700 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
701 local_irq_restore(flags);
704 __rcu_process_gp_end(rsp, rnp, rdp);
705 raw_spin_unlock_irqrestore(&rnp->lock, flags);
709 * Do per-CPU grace-period initialization for running CPU. The caller
710 * must hold the lock of the leaf rcu_node structure corresponding to
714 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
716 /* Prior grace period ended, so advance callbacks for current CPU. */
717 __rcu_process_gp_end(rsp, rnp, rdp);
720 * Because this CPU just now started the new grace period, we know
721 * that all of its callbacks will be covered by this upcoming grace
722 * period, even the ones that were registered arbitrarily recently.
723 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
725 * Other CPUs cannot be sure exactly when the grace period started.
726 * Therefore, their recently registered callbacks must pass through
727 * an additional RCU_NEXT_READY stage, so that they will be handled
728 * by the next RCU grace period.
730 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
731 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
733 /* Set state so that this CPU will detect the next quiescent state. */
734 __note_new_gpnum(rsp, rnp, rdp);
738 * Start a new RCU grace period if warranted, re-initializing the hierarchy
739 * in preparation for detecting the next grace period. The caller must hold
740 * the root node's ->lock, which is released before return. Hard irqs must
744 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
745 __releases(rcu_get_root(rsp)->lock)
747 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
748 struct rcu_node *rnp = rcu_get_root(rsp);
750 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
751 if (cpu_needs_another_gp(rsp, rdp))
752 rsp->fqs_need_gp = 1;
753 if (rnp->completed == rsp->completed) {
754 raw_spin_unlock_irqrestore(&rnp->lock, flags);
757 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
760 * Propagate new ->completed value to rcu_node structures
761 * so that other CPUs don't have to wait until the start
762 * of the next grace period to process their callbacks.
764 rcu_for_each_node_breadth_first(rsp, rnp) {
765 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
766 rnp->completed = rsp->completed;
767 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
769 local_irq_restore(flags);
773 /* Advance to a new grace period and initialize state. */
775 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
776 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
777 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
778 record_gp_stall_check_time(rsp);
780 /* Special-case the common single-level case. */
781 if (NUM_RCU_NODES == 1) {
782 rcu_preempt_check_blocked_tasks(rnp);
783 rnp->qsmask = rnp->qsmaskinit;
784 rnp->gpnum = rsp->gpnum;
785 rnp->completed = rsp->completed;
786 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
787 rcu_start_gp_per_cpu(rsp, rnp, rdp);
788 raw_spin_unlock_irqrestore(&rnp->lock, flags);
792 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
795 /* Exclude any concurrent CPU-hotplug operations. */
796 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
799 * Set the quiescent-state-needed bits in all the rcu_node
800 * structures for all currently online CPUs in breadth-first
801 * order, starting from the root rcu_node structure. This
802 * operation relies on the layout of the hierarchy within the
803 * rsp->node[] array. Note that other CPUs will access only
804 * the leaves of the hierarchy, which still indicate that no
805 * grace period is in progress, at least until the corresponding
806 * leaf node has been initialized. In addition, we have excluded
807 * CPU-hotplug operations.
809 * Note that the grace period cannot complete until we finish
810 * the initialization process, as there will be at least one
811 * qsmask bit set in the root node until that time, namely the
812 * one corresponding to this CPU, due to the fact that we have
815 rcu_for_each_node_breadth_first(rsp, rnp) {
816 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
817 rcu_preempt_check_blocked_tasks(rnp);
818 rnp->qsmask = rnp->qsmaskinit;
819 rnp->gpnum = rsp->gpnum;
820 rnp->completed = rsp->completed;
821 if (rnp == rdp->mynode)
822 rcu_start_gp_per_cpu(rsp, rnp, rdp);
823 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
826 rnp = rcu_get_root(rsp);
827 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
828 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
829 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
830 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
834 * Report a full set of quiescent states to the specified rcu_state
835 * data structure. This involves cleaning up after the prior grace
836 * period and letting rcu_start_gp() start up the next grace period
837 * if one is needed. Note that the caller must hold rnp->lock, as
838 * required by rcu_start_gp(), which will release it.
840 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
841 __releases(rcu_get_root(rsp)->lock)
843 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
844 rsp->completed = rsp->gpnum;
845 rsp->signaled = RCU_GP_IDLE;
846 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
850 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
851 * Allows quiescent states for a group of CPUs to be reported at one go
852 * to the specified rcu_node structure, though all the CPUs in the group
853 * must be represented by the same rcu_node structure (which need not be
854 * a leaf rcu_node structure, though it often will be). That structure's
855 * lock must be held upon entry, and it is released before return.
858 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
859 struct rcu_node *rnp, unsigned long flags)
860 __releases(rnp->lock)
862 struct rcu_node *rnp_c;
864 /* Walk up the rcu_node hierarchy. */
866 if (!(rnp->qsmask & mask)) {
868 /* Our bit has already been cleared, so done. */
869 raw_spin_unlock_irqrestore(&rnp->lock, flags);
872 rnp->qsmask &= ~mask;
873 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
875 /* Other bits still set at this level, so done. */
876 raw_spin_unlock_irqrestore(&rnp->lock, flags);
880 if (rnp->parent == NULL) {
882 /* No more levels. Exit loop holding root lock. */
886 raw_spin_unlock_irqrestore(&rnp->lock, flags);
889 raw_spin_lock_irqsave(&rnp->lock, flags);
890 WARN_ON_ONCE(rnp_c->qsmask);
894 * Get here if we are the last CPU to pass through a quiescent
895 * state for this grace period. Invoke rcu_report_qs_rsp()
896 * to clean up and start the next grace period if one is needed.
898 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
902 * Record a quiescent state for the specified CPU to that CPU's rcu_data
903 * structure. This must be either called from the specified CPU, or
904 * called when the specified CPU is known to be offline (and when it is
905 * also known that no other CPU is concurrently trying to help the offline
906 * CPU). The lastcomp argument is used to make sure we are still in the
907 * grace period of interest. We don't want to end the current grace period
908 * based on quiescent states detected in an earlier grace period!
911 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
915 struct rcu_node *rnp;
918 raw_spin_lock_irqsave(&rnp->lock, flags);
919 if (lastcomp != rnp->completed) {
922 * Someone beat us to it for this grace period, so leave.
923 * The race with GP start is resolved by the fact that we
924 * hold the leaf rcu_node lock, so that the per-CPU bits
925 * cannot yet be initialized -- so we would simply find our
926 * CPU's bit already cleared in rcu_report_qs_rnp() if this
929 rdp->passed_quiesc = 0; /* try again later! */
930 raw_spin_unlock_irqrestore(&rnp->lock, flags);
934 if ((rnp->qsmask & mask) == 0) {
935 raw_spin_unlock_irqrestore(&rnp->lock, flags);
940 * This GP can't end until cpu checks in, so all of our
941 * callbacks can be processed during the next GP.
943 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
945 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
950 * Check to see if there is a new grace period of which this CPU
951 * is not yet aware, and if so, set up local rcu_data state for it.
952 * Otherwise, see if this CPU has just passed through its first
953 * quiescent state for this grace period, and record that fact if so.
956 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
958 /* If there is now a new grace period, record and return. */
959 if (check_for_new_grace_period(rsp, rdp))
963 * Does this CPU still need to do its part for current grace period?
964 * If no, return and let the other CPUs do their part as well.
966 if (!rdp->qs_pending)
970 * Was there a quiescent state since the beginning of the grace
971 * period? If no, then exit and wait for the next call.
973 if (!rdp->passed_quiesc)
977 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
980 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
983 #ifdef CONFIG_HOTPLUG_CPU
986 * Move a dying CPU's RCU callbacks to online CPU's callback list.
987 * Synchronization is not required because this function executes
988 * in stop_machine() context.
990 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
993 /* current DYING CPU is cleared in the cpu_online_mask */
994 int receive_cpu = cpumask_any(cpu_online_mask);
995 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
996 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
998 if (rdp->nxtlist == NULL)
999 return; /* irqs disabled, so comparison is stable. */
1001 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1002 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1003 receive_rdp->qlen += rdp->qlen;
1004 receive_rdp->n_cbs_adopted += rdp->qlen;
1005 rdp->n_cbs_orphaned += rdp->qlen;
1007 rdp->nxtlist = NULL;
1008 for (i = 0; i < RCU_NEXT_SIZE; i++)
1009 rdp->nxttail[i] = &rdp->nxtlist;
1014 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1015 * and move all callbacks from the outgoing CPU to the current one.
1017 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1019 unsigned long flags;
1021 int need_report = 0;
1022 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1023 struct rcu_node *rnp;
1025 /* Exclude any attempts to start a new grace period. */
1026 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1028 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1029 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1030 mask = rdp->grpmask; /* rnp->grplo is constant. */
1032 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1033 rnp->qsmaskinit &= ~mask;
1034 if (rnp->qsmaskinit != 0) {
1035 if (rnp != rdp->mynode)
1036 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1039 if (rnp == rdp->mynode)
1040 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1042 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1043 mask = rnp->grpmask;
1045 } while (rnp != NULL);
1048 * We still hold the leaf rcu_node structure lock here, and
1049 * irqs are still disabled. The reason for this subterfuge is
1050 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1051 * held leads to deadlock.
1053 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1055 if (need_report & RCU_OFL_TASKS_NORM_GP)
1056 rcu_report_unblock_qs_rnp(rnp, flags);
1058 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1059 if (need_report & RCU_OFL_TASKS_EXP_GP)
1060 rcu_report_exp_rnp(rsp, rnp);
1064 * Remove the specified CPU from the RCU hierarchy and move any pending
1065 * callbacks that it might have to the current CPU. This code assumes
1066 * that at least one CPU in the system will remain running at all times.
1067 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1069 static void rcu_offline_cpu(int cpu)
1071 __rcu_offline_cpu(cpu, &rcu_sched_state);
1072 __rcu_offline_cpu(cpu, &rcu_bh_state);
1073 rcu_preempt_offline_cpu(cpu);
1076 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1078 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1082 static void rcu_offline_cpu(int cpu)
1086 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1089 * Invoke any RCU callbacks that have made it to the end of their grace
1090 * period. Thottle as specified by rdp->blimit.
1092 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1094 unsigned long flags;
1095 struct rcu_head *next, *list, **tail;
1098 /* If no callbacks are ready, just return.*/
1099 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1103 * Extract the list of ready callbacks, disabling to prevent
1104 * races with call_rcu() from interrupt handlers.
1106 local_irq_save(flags);
1107 list = rdp->nxtlist;
1108 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1109 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1110 tail = rdp->nxttail[RCU_DONE_TAIL];
1111 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1112 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1113 rdp->nxttail[count] = &rdp->nxtlist;
1114 local_irq_restore(flags);
1116 /* Invoke callbacks. */
1121 debug_rcu_head_unqueue(list);
1124 if (++count >= rdp->blimit)
1128 local_irq_save(flags);
1130 /* Update count, and requeue any remaining callbacks. */
1132 rdp->n_cbs_invoked += count;
1134 *tail = rdp->nxtlist;
1135 rdp->nxtlist = list;
1136 for (count = 0; count < RCU_NEXT_SIZE; count++)
1137 if (&rdp->nxtlist == rdp->nxttail[count])
1138 rdp->nxttail[count] = tail;
1143 /* Reinstate batch limit if we have worked down the excess. */
1144 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1145 rdp->blimit = blimit;
1147 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1148 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1149 rdp->qlen_last_fqs_check = 0;
1150 rdp->n_force_qs_snap = rsp->n_force_qs;
1151 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1152 rdp->qlen_last_fqs_check = rdp->qlen;
1154 local_irq_restore(flags);
1156 /* Re-raise the RCU softirq if there are callbacks remaining. */
1157 if (cpu_has_callbacks_ready_to_invoke(rdp))
1158 raise_softirq(RCU_SOFTIRQ);
1162 * Check to see if this CPU is in a non-context-switch quiescent state
1163 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1164 * Also schedule the RCU softirq handler.
1166 * This function must be called with hardirqs disabled. It is normally
1167 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1168 * false, there is no point in invoking rcu_check_callbacks().
1170 void rcu_check_callbacks(int cpu, int user)
1173 (idle_cpu(cpu) && rcu_scheduler_active &&
1174 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1177 * Get here if this CPU took its interrupt from user
1178 * mode or from the idle loop, and if this is not a
1179 * nested interrupt. In this case, the CPU is in
1180 * a quiescent state, so note it.
1182 * No memory barrier is required here because both
1183 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1184 * variables that other CPUs neither access nor modify,
1185 * at least not while the corresponding CPU is online.
1191 } else if (!in_softirq()) {
1194 * Get here if this CPU did not take its interrupt from
1195 * softirq, in other words, if it is not interrupting
1196 * a rcu_bh read-side critical section. This is an _bh
1197 * critical section, so note it.
1202 rcu_preempt_check_callbacks(cpu);
1203 if (rcu_pending(cpu))
1204 raise_softirq(RCU_SOFTIRQ);
1210 * Scan the leaf rcu_node structures, processing dyntick state for any that
1211 * have not yet encountered a quiescent state, using the function specified.
1212 * The caller must have suppressed start of new grace periods.
1214 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1218 unsigned long flags;
1220 struct rcu_node *rnp;
1222 rcu_for_each_leaf_node(rsp, rnp) {
1224 raw_spin_lock_irqsave(&rnp->lock, flags);
1225 if (!rcu_gp_in_progress(rsp)) {
1226 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1229 if (rnp->qsmask == 0) {
1230 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1235 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1236 if ((rnp->qsmask & bit) != 0 &&
1237 f(per_cpu_ptr(rsp->rda, cpu)))
1242 /* rcu_report_qs_rnp() releases rnp->lock. */
1243 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1246 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1251 * Force quiescent states on reluctant CPUs, and also detect which
1252 * CPUs are in dyntick-idle mode.
1254 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1256 unsigned long flags;
1257 struct rcu_node *rnp = rcu_get_root(rsp);
1259 if (!rcu_gp_in_progress(rsp))
1260 return; /* No grace period in progress, nothing to force. */
1261 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1262 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1263 return; /* Someone else is already on the job. */
1265 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1266 goto unlock_fqs_ret; /* no emergency and done recently. */
1268 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1269 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1270 if(!rcu_gp_in_progress(rsp)) {
1271 rsp->n_force_qs_ngp++;
1272 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1273 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1275 rsp->fqs_active = 1;
1276 switch (rsp->signaled) {
1280 break; /* grace period idle or initializing, ignore. */
1282 case RCU_SAVE_DYNTICK:
1283 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1284 break; /* So gcc recognizes the dead code. */
1286 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1288 /* Record dyntick-idle state. */
1289 force_qs_rnp(rsp, dyntick_save_progress_counter);
1290 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1291 if (rcu_gp_in_progress(rsp))
1292 rsp->signaled = RCU_FORCE_QS;
1297 /* Check dyntick-idle state, send IPI to laggarts. */
1298 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1299 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1301 /* Leave state in case more forcing is required. */
1303 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1306 rsp->fqs_active = 0;
1307 if (rsp->fqs_need_gp) {
1308 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1309 rsp->fqs_need_gp = 0;
1310 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1313 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1315 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1318 #else /* #ifdef CONFIG_SMP */
1320 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1325 #endif /* #else #ifdef CONFIG_SMP */
1328 * This does the RCU processing work from softirq context for the
1329 * specified rcu_state and rcu_data structures. This may be called
1330 * only from the CPU to whom the rdp belongs.
1333 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1335 unsigned long flags;
1337 WARN_ON_ONCE(rdp->beenonline == 0);
1340 * If an RCU GP has gone long enough, go check for dyntick
1341 * idle CPUs and, if needed, send resched IPIs.
1343 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1344 force_quiescent_state(rsp, 1);
1347 * Advance callbacks in response to end of earlier grace
1348 * period that some other CPU ended.
1350 rcu_process_gp_end(rsp, rdp);
1352 /* Update RCU state based on any recent quiescent states. */
1353 rcu_check_quiescent_state(rsp, rdp);
1355 /* Does this CPU require a not-yet-started grace period? */
1356 if (cpu_needs_another_gp(rsp, rdp)) {
1357 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1358 rcu_start_gp(rsp, flags); /* releases above lock */
1361 /* If there are callbacks ready, invoke them. */
1362 rcu_do_batch(rsp, rdp);
1366 * Do softirq processing for the current CPU.
1368 static void rcu_process_callbacks(struct softirq_action *unused)
1371 * Memory references from any prior RCU read-side critical sections
1372 * executed by the interrupted code must be seen before any RCU
1373 * grace-period manipulations below.
1375 smp_mb(); /* See above block comment. */
1377 __rcu_process_callbacks(&rcu_sched_state,
1378 &__get_cpu_var(rcu_sched_data));
1379 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1380 rcu_preempt_process_callbacks();
1383 * Memory references from any later RCU read-side critical sections
1384 * executed by the interrupted code must be seen after any RCU
1385 * grace-period manipulations above.
1387 smp_mb(); /* See above block comment. */
1389 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1390 rcu_needs_cpu_flush();
1394 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1395 struct rcu_state *rsp)
1397 unsigned long flags;
1398 struct rcu_data *rdp;
1400 debug_rcu_head_queue(head);
1404 smp_mb(); /* Ensure RCU update seen before callback registry. */
1407 * Opportunistically note grace-period endings and beginnings.
1408 * Note that we might see a beginning right after we see an
1409 * end, but never vice versa, since this CPU has to pass through
1410 * a quiescent state betweentimes.
1412 local_irq_save(flags);
1413 rdp = this_cpu_ptr(rsp->rda);
1415 /* Add the callback to our list. */
1416 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1417 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1420 * Force the grace period if too many callbacks or too long waiting.
1421 * Enforce hysteresis, and don't invoke force_quiescent_state()
1422 * if some other CPU has recently done so. Also, don't bother
1423 * invoking force_quiescent_state() if the newly enqueued callback
1424 * is the only one waiting for a grace period to complete.
1426 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1428 /* Are we ignoring a completed grace period? */
1429 rcu_process_gp_end(rsp, rdp);
1430 check_for_new_grace_period(rsp, rdp);
1432 /* Start a new grace period if one not already started. */
1433 if (!rcu_gp_in_progress(rsp)) {
1434 unsigned long nestflag;
1435 struct rcu_node *rnp_root = rcu_get_root(rsp);
1437 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1438 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1440 /* Give the grace period a kick. */
1441 rdp->blimit = LONG_MAX;
1442 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1443 *rdp->nxttail[RCU_DONE_TAIL] != head)
1444 force_quiescent_state(rsp, 0);
1445 rdp->n_force_qs_snap = rsp->n_force_qs;
1446 rdp->qlen_last_fqs_check = rdp->qlen;
1448 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1449 force_quiescent_state(rsp, 1);
1450 local_irq_restore(flags);
1454 * Queue an RCU-sched callback for invocation after a grace period.
1456 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1458 __call_rcu(head, func, &rcu_sched_state);
1460 EXPORT_SYMBOL_GPL(call_rcu_sched);
1463 * Queue an RCU for invocation after a quicker grace period.
1465 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1467 __call_rcu(head, func, &rcu_bh_state);
1469 EXPORT_SYMBOL_GPL(call_rcu_bh);
1472 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1474 * Control will return to the caller some time after a full rcu-sched
1475 * grace period has elapsed, in other words after all currently executing
1476 * rcu-sched read-side critical sections have completed. These read-side
1477 * critical sections are delimited by rcu_read_lock_sched() and
1478 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1479 * local_irq_disable(), and so on may be used in place of
1480 * rcu_read_lock_sched().
1482 * This means that all preempt_disable code sequences, including NMI and
1483 * hardware-interrupt handlers, in progress on entry will have completed
1484 * before this primitive returns. However, this does not guarantee that
1485 * softirq handlers will have completed, since in some kernels, these
1486 * handlers can run in process context, and can block.
1488 * This primitive provides the guarantees made by the (now removed)
1489 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1490 * guarantees that rcu_read_lock() sections will have completed.
1491 * In "classic RCU", these two guarantees happen to be one and
1492 * the same, but can differ in realtime RCU implementations.
1494 void synchronize_sched(void)
1496 struct rcu_synchronize rcu;
1498 if (rcu_blocking_is_gp())
1501 init_rcu_head_on_stack(&rcu.head);
1502 init_completion(&rcu.completion);
1503 /* Will wake me after RCU finished. */
1504 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1506 wait_for_completion(&rcu.completion);
1507 destroy_rcu_head_on_stack(&rcu.head);
1509 EXPORT_SYMBOL_GPL(synchronize_sched);
1512 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1514 * Control will return to the caller some time after a full rcu_bh grace
1515 * period has elapsed, in other words after all currently executing rcu_bh
1516 * read-side critical sections have completed. RCU read-side critical
1517 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1518 * and may be nested.
1520 void synchronize_rcu_bh(void)
1522 struct rcu_synchronize rcu;
1524 if (rcu_blocking_is_gp())
1527 init_rcu_head_on_stack(&rcu.head);
1528 init_completion(&rcu.completion);
1529 /* Will wake me after RCU finished. */
1530 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1532 wait_for_completion(&rcu.completion);
1533 destroy_rcu_head_on_stack(&rcu.head);
1535 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1538 * Check to see if there is any immediate RCU-related work to be done
1539 * by the current CPU, for the specified type of RCU, returning 1 if so.
1540 * The checks are in order of increasing expense: checks that can be
1541 * carried out against CPU-local state are performed first. However,
1542 * we must check for CPU stalls first, else we might not get a chance.
1544 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1546 struct rcu_node *rnp = rdp->mynode;
1548 rdp->n_rcu_pending++;
1550 /* Check for CPU stalls, if enabled. */
1551 check_cpu_stall(rsp, rdp);
1553 /* Is the RCU core waiting for a quiescent state from this CPU? */
1554 if (rdp->qs_pending && !rdp->passed_quiesc) {
1557 * If force_quiescent_state() coming soon and this CPU
1558 * needs a quiescent state, and this is either RCU-sched
1559 * or RCU-bh, force a local reschedule.
1561 rdp->n_rp_qs_pending++;
1562 if (!rdp->preemptable &&
1563 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1566 } else if (rdp->qs_pending && rdp->passed_quiesc) {
1567 rdp->n_rp_report_qs++;
1571 /* Does this CPU have callbacks ready to invoke? */
1572 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1573 rdp->n_rp_cb_ready++;
1577 /* Has RCU gone idle with this CPU needing another grace period? */
1578 if (cpu_needs_another_gp(rsp, rdp)) {
1579 rdp->n_rp_cpu_needs_gp++;
1583 /* Has another RCU grace period completed? */
1584 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1585 rdp->n_rp_gp_completed++;
1589 /* Has a new RCU grace period started? */
1590 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1591 rdp->n_rp_gp_started++;
1595 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1596 if (rcu_gp_in_progress(rsp) &&
1597 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1598 rdp->n_rp_need_fqs++;
1603 rdp->n_rp_need_nothing++;
1608 * Check to see if there is any immediate RCU-related work to be done
1609 * by the current CPU, returning 1 if so. This function is part of the
1610 * RCU implementation; it is -not- an exported member of the RCU API.
1612 static int rcu_pending(int cpu)
1614 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1615 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1616 rcu_preempt_pending(cpu);
1620 * Check to see if any future RCU-related work will need to be done
1621 * by the current CPU, even if none need be done immediately, returning
1624 static int rcu_needs_cpu_quick_check(int cpu)
1626 /* RCU callbacks either ready or pending? */
1627 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1628 per_cpu(rcu_bh_data, cpu).nxtlist ||
1629 rcu_preempt_needs_cpu(cpu);
1632 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1633 static atomic_t rcu_barrier_cpu_count;
1634 static DEFINE_MUTEX(rcu_barrier_mutex);
1635 static struct completion rcu_barrier_completion;
1637 static void rcu_barrier_callback(struct rcu_head *notused)
1639 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1640 complete(&rcu_barrier_completion);
1644 * Called with preemption disabled, and from cross-cpu IRQ context.
1646 static void rcu_barrier_func(void *type)
1648 int cpu = smp_processor_id();
1649 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1650 void (*call_rcu_func)(struct rcu_head *head,
1651 void (*func)(struct rcu_head *head));
1653 atomic_inc(&rcu_barrier_cpu_count);
1654 call_rcu_func = type;
1655 call_rcu_func(head, rcu_barrier_callback);
1659 * Orchestrate the specified type of RCU barrier, waiting for all
1660 * RCU callbacks of the specified type to complete.
1662 static void _rcu_barrier(struct rcu_state *rsp,
1663 void (*call_rcu_func)(struct rcu_head *head,
1664 void (*func)(struct rcu_head *head)))
1666 BUG_ON(in_interrupt());
1667 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1668 mutex_lock(&rcu_barrier_mutex);
1669 init_completion(&rcu_barrier_completion);
1671 * Initialize rcu_barrier_cpu_count to 1, then invoke
1672 * rcu_barrier_func() on each CPU, so that each CPU also has
1673 * incremented rcu_barrier_cpu_count. Only then is it safe to
1674 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1675 * might complete its grace period before all of the other CPUs
1676 * did their increment, causing this function to return too
1677 * early. Note that on_each_cpu() disables irqs, which prevents
1678 * any CPUs from coming online or going offline until each online
1679 * CPU has queued its RCU-barrier callback.
1681 atomic_set(&rcu_barrier_cpu_count, 1);
1682 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1683 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1684 complete(&rcu_barrier_completion);
1685 wait_for_completion(&rcu_barrier_completion);
1686 mutex_unlock(&rcu_barrier_mutex);
1690 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1692 void rcu_barrier_bh(void)
1694 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1696 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1699 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1701 void rcu_barrier_sched(void)
1703 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1705 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1708 * Do boot-time initialization of a CPU's per-CPU RCU data.
1711 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1713 unsigned long flags;
1715 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1716 struct rcu_node *rnp = rcu_get_root(rsp);
1718 /* Set up local state, ensuring consistent view of global state. */
1719 raw_spin_lock_irqsave(&rnp->lock, flags);
1720 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1721 rdp->nxtlist = NULL;
1722 for (i = 0; i < RCU_NEXT_SIZE; i++)
1723 rdp->nxttail[i] = &rdp->nxtlist;
1726 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1727 #endif /* #ifdef CONFIG_NO_HZ */
1729 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1733 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1734 * offline event can be happening at a given time. Note also that we
1735 * can accept some slop in the rsp->completed access due to the fact
1736 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1738 static void __cpuinit
1739 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1741 unsigned long flags;
1743 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1744 struct rcu_node *rnp = rcu_get_root(rsp);
1746 /* Set up local state, ensuring consistent view of global state. */
1747 raw_spin_lock_irqsave(&rnp->lock, flags);
1748 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1749 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1750 rdp->beenonline = 1; /* We have now been online. */
1751 rdp->preemptable = preemptable;
1752 rdp->qlen_last_fqs_check = 0;
1753 rdp->n_force_qs_snap = rsp->n_force_qs;
1754 rdp->blimit = blimit;
1755 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1758 * A new grace period might start here. If so, we won't be part
1759 * of it, but that is OK, as we are currently in a quiescent state.
1762 /* Exclude any attempts to start a new GP on large systems. */
1763 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1765 /* Add CPU to rcu_node bitmasks. */
1767 mask = rdp->grpmask;
1769 /* Exclude any attempts to start a new GP on small systems. */
1770 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1771 rnp->qsmaskinit |= mask;
1772 mask = rnp->grpmask;
1773 if (rnp == rdp->mynode) {
1774 rdp->gpnum = rnp->completed; /* if GP in progress... */
1775 rdp->completed = rnp->completed;
1776 rdp->passed_quiesc_completed = rnp->completed - 1;
1778 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1780 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1782 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1785 static void __cpuinit rcu_online_cpu(int cpu)
1787 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1788 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1789 rcu_preempt_init_percpu_data(cpu);
1793 * Handle CPU online/offline notification events.
1795 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1796 unsigned long action, void *hcpu)
1798 long cpu = (long)hcpu;
1801 case CPU_UP_PREPARE:
1802 case CPU_UP_PREPARE_FROZEN:
1803 rcu_online_cpu(cpu);
1806 case CPU_DYING_FROZEN:
1808 * The whole machine is "stopped" except this CPU, so we can
1809 * touch any data without introducing corruption. We send the
1810 * dying CPU's callbacks to an arbitrarily chosen online CPU.
1812 rcu_send_cbs_to_online(&rcu_bh_state);
1813 rcu_send_cbs_to_online(&rcu_sched_state);
1814 rcu_preempt_send_cbs_to_online();
1817 case CPU_DEAD_FROZEN:
1818 case CPU_UP_CANCELED:
1819 case CPU_UP_CANCELED_FROZEN:
1820 rcu_offline_cpu(cpu);
1829 * This function is invoked towards the end of the scheduler's initialization
1830 * process. Before this is called, the idle task might contain
1831 * RCU read-side critical sections (during which time, this idle
1832 * task is booting the system). After this function is called, the
1833 * idle tasks are prohibited from containing RCU read-side critical
1834 * sections. This function also enables RCU lockdep checking.
1836 void rcu_scheduler_starting(void)
1838 WARN_ON(num_online_cpus() != 1);
1839 WARN_ON(nr_context_switches() > 0);
1840 rcu_scheduler_active = 1;
1844 * Compute the per-level fanout, either using the exact fanout specified
1845 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1847 #ifdef CONFIG_RCU_FANOUT_EXACT
1848 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1852 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
1853 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1854 rsp->levelspread[0] = RCU_FANOUT_LEAF;
1856 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1857 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1864 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1865 ccur = rsp->levelcnt[i];
1866 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1870 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1873 * Helper function for rcu_init() that initializes one rcu_state structure.
1875 static void __init rcu_init_one(struct rcu_state *rsp,
1876 struct rcu_data __percpu *rda)
1878 static char *buf[] = { "rcu_node_level_0",
1881 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1885 struct rcu_node *rnp;
1887 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1889 /* Initialize the level-tracking arrays. */
1891 for (i = 1; i < NUM_RCU_LVLS; i++)
1892 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1893 rcu_init_levelspread(rsp);
1895 /* Initialize the elements themselves, starting from the leaves. */
1897 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1898 cpustride *= rsp->levelspread[i];
1899 rnp = rsp->level[i];
1900 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1901 raw_spin_lock_init(&rnp->lock);
1902 lockdep_set_class_and_name(&rnp->lock,
1903 &rcu_node_class[i], buf[i]);
1906 rnp->qsmaskinit = 0;
1907 rnp->grplo = j * cpustride;
1908 rnp->grphi = (j + 1) * cpustride - 1;
1909 if (rnp->grphi >= NR_CPUS)
1910 rnp->grphi = NR_CPUS - 1;
1916 rnp->grpnum = j % rsp->levelspread[i - 1];
1917 rnp->grpmask = 1UL << rnp->grpnum;
1918 rnp->parent = rsp->level[i - 1] +
1919 j / rsp->levelspread[i - 1];
1922 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1923 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1924 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1925 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1930 rnp = rsp->level[NUM_RCU_LVLS - 1];
1931 for_each_possible_cpu(i) {
1932 while (i > rnp->grphi)
1934 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
1935 rcu_boot_init_percpu_data(i, rsp);
1939 void __init rcu_init(void)
1943 rcu_bootup_announce();
1944 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
1945 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
1946 __rcu_init_preempt();
1947 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1950 * We don't need protection against CPU-hotplug here because
1951 * this is called early in boot, before either interrupts
1952 * or the scheduler are operational.
1954 cpu_notifier(rcu_cpu_notify, 0);
1955 for_each_online_cpu(cpu)
1956 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1957 check_cpu_stall_init();
1960 #include "rcutree_plugin.h"