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 .orphan_cbs_list = NULL, \
71 .orphan_cbs_tail = &structname.orphan_cbs_list, \
73 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
75 .n_force_qs_ngp = 0, \
76 .name = #structname, \
79 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
80 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
82 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
83 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
85 int rcu_scheduler_active __read_mostly;
86 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
89 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
90 * permit this function to be invoked without holding the root rcu_node
91 * structure's ->lock, but of course results can be subject to change.
93 static int rcu_gp_in_progress(struct rcu_state *rsp)
95 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
99 * Note a quiescent state. Because we do not need to know
100 * how many quiescent states passed, just if there was at least
101 * one since the start of the grace period, this just sets a flag.
103 void rcu_sched_qs(int cpu)
105 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
107 rdp->passed_quiesc_completed = rdp->gpnum - 1;
109 rdp->passed_quiesc = 1;
112 void rcu_bh_qs(int cpu)
114 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
116 rdp->passed_quiesc_completed = rdp->gpnum - 1;
118 rdp->passed_quiesc = 1;
122 * Note a context switch. This is a quiescent state for RCU-sched,
123 * and requires special handling for preemptible RCU.
125 void rcu_note_context_switch(int cpu)
128 rcu_preempt_note_context_switch(cpu);
132 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
133 .dynticks_nesting = 1,
136 #endif /* #ifdef CONFIG_NO_HZ */
138 static int blimit = 10; /* Maximum callbacks per softirq. */
139 static int qhimark = 10000; /* If this many pending, ignore blimit. */
140 static int qlowmark = 100; /* Once only this many pending, use blimit. */
142 module_param(blimit, int, 0);
143 module_param(qhimark, int, 0);
144 module_param(qlowmark, int, 0);
146 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
147 int rcu_cpu_stall_suppress __read_mostly;
148 module_param(rcu_cpu_stall_suppress, int, 0);
149 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
151 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
152 static int rcu_pending(int cpu);
155 * Return the number of RCU-sched batches processed thus far for debug & stats.
157 long rcu_batches_completed_sched(void)
159 return rcu_sched_state.completed;
161 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
164 * Return the number of RCU BH batches processed thus far for debug & stats.
166 long rcu_batches_completed_bh(void)
168 return rcu_bh_state.completed;
170 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
173 * Force a quiescent state for RCU BH.
175 void rcu_bh_force_quiescent_state(void)
177 force_quiescent_state(&rcu_bh_state, 0);
179 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
182 * Force a quiescent state for RCU-sched.
184 void rcu_sched_force_quiescent_state(void)
186 force_quiescent_state(&rcu_sched_state, 0);
188 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
191 * Does the CPU have callbacks ready to be invoked?
194 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
196 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
200 * Does the current CPU require a yet-as-unscheduled grace period?
203 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
205 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
209 * Return the root node of the specified rcu_state structure.
211 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
213 return &rsp->node[0];
219 * If the specified CPU is offline, tell the caller that it is in
220 * a quiescent state. Otherwise, whack it with a reschedule IPI.
221 * Grace periods can end up waiting on an offline CPU when that
222 * CPU is in the process of coming online -- it will be added to the
223 * rcu_node bitmasks before it actually makes it online. The same thing
224 * can happen while a CPU is in the process of coming online. Because this
225 * race is quite rare, we check for it after detecting that the grace
226 * period has been delayed rather than checking each and every CPU
227 * each and every time we start a new grace period.
229 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
232 * If the CPU is offline, it is in a quiescent state. We can
233 * trust its state not to change because interrupts are disabled.
235 if (cpu_is_offline(rdp->cpu)) {
240 /* If preemptable RCU, no point in sending reschedule IPI. */
241 if (rdp->preemptable)
244 /* The CPU is online, so send it a reschedule IPI. */
245 if (rdp->cpu != smp_processor_id())
246 smp_send_reschedule(rdp->cpu);
253 #endif /* #ifdef CONFIG_SMP */
258 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
260 * Enter nohz mode, in other words, -leave- the mode in which RCU
261 * read-side critical sections can occur. (Though RCU read-side
262 * critical sections can occur in irq handlers in nohz mode, a possibility
263 * handled by rcu_irq_enter() and rcu_irq_exit()).
265 void rcu_enter_nohz(void)
268 struct rcu_dynticks *rdtp;
270 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
271 local_irq_save(flags);
272 rdtp = &__get_cpu_var(rcu_dynticks);
274 rdtp->dynticks_nesting--;
275 WARN_ON_ONCE(rdtp->dynticks & 0x1);
276 local_irq_restore(flags);
280 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
282 * Exit nohz mode, in other words, -enter- the mode in which RCU
283 * read-side critical sections normally occur.
285 void rcu_exit_nohz(void)
288 struct rcu_dynticks *rdtp;
290 local_irq_save(flags);
291 rdtp = &__get_cpu_var(rcu_dynticks);
293 rdtp->dynticks_nesting++;
294 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
295 local_irq_restore(flags);
296 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
300 * rcu_nmi_enter - inform RCU of entry to NMI context
302 * If the CPU was idle with dynamic ticks active, and there is no
303 * irq handler running, this updates rdtp->dynticks_nmi to let the
304 * RCU grace-period handling know that the CPU is active.
306 void rcu_nmi_enter(void)
308 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
310 if (rdtp->dynticks & 0x1)
312 rdtp->dynticks_nmi++;
313 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
314 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
318 * rcu_nmi_exit - inform RCU of exit from NMI context
320 * If the CPU was idle with dynamic ticks active, and there is no
321 * irq handler running, this updates rdtp->dynticks_nmi to let the
322 * RCU grace-period handling know that the CPU is no longer active.
324 void rcu_nmi_exit(void)
326 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
328 if (rdtp->dynticks & 0x1)
330 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
331 rdtp->dynticks_nmi++;
332 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
336 * rcu_irq_enter - inform RCU of entry to hard irq context
338 * If the CPU was idle with dynamic ticks active, this updates the
339 * rdtp->dynticks to let the RCU handling know that the CPU is active.
341 void rcu_irq_enter(void)
343 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
345 if (rdtp->dynticks_nesting++)
348 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
349 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
353 * rcu_irq_exit - inform RCU of exit from hard irq context
355 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
356 * to put let the RCU handling be aware that the CPU is going back to idle
359 void rcu_irq_exit(void)
361 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
363 if (--rdtp->dynticks_nesting)
365 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
367 WARN_ON_ONCE(rdtp->dynticks & 0x1);
369 /* If the interrupt queued a callback, get out of dyntick mode. */
370 if (__get_cpu_var(rcu_sched_data).nxtlist ||
371 __get_cpu_var(rcu_bh_data).nxtlist)
378 * Snapshot the specified CPU's dynticks counter so that we can later
379 * credit them with an implicit quiescent state. Return 1 if this CPU
380 * is in dynticks idle mode, which is an extended quiescent state.
382 static int dyntick_save_progress_counter(struct rcu_data *rdp)
388 snap = rdp->dynticks->dynticks;
389 snap_nmi = rdp->dynticks->dynticks_nmi;
390 smp_mb(); /* Order sampling of snap with end of grace period. */
391 rdp->dynticks_snap = snap;
392 rdp->dynticks_nmi_snap = snap_nmi;
393 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
400 * Return true if the specified CPU has passed through a quiescent
401 * state by virtue of being in or having passed through an dynticks
402 * idle state since the last call to dyntick_save_progress_counter()
405 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
412 curr = rdp->dynticks->dynticks;
413 snap = rdp->dynticks_snap;
414 curr_nmi = rdp->dynticks->dynticks_nmi;
415 snap_nmi = rdp->dynticks_nmi_snap;
416 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
419 * If the CPU passed through or entered a dynticks idle phase with
420 * no active irq/NMI handlers, then we can safely pretend that the CPU
421 * already acknowledged the request to pass through a quiescent
422 * state. Either way, that CPU cannot possibly be in an RCU
423 * read-side critical section that started before the beginning
424 * of the current RCU grace period.
426 if ((curr != snap || (curr & 0x1) == 0) &&
427 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
432 /* Go check for the CPU being offline. */
433 return rcu_implicit_offline_qs(rdp);
436 #endif /* #ifdef CONFIG_SMP */
438 #else /* #ifdef CONFIG_NO_HZ */
442 static int dyntick_save_progress_counter(struct rcu_data *rdp)
447 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
449 return rcu_implicit_offline_qs(rdp);
452 #endif /* #ifdef CONFIG_SMP */
454 #endif /* #else #ifdef CONFIG_NO_HZ */
456 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
458 int rcu_cpu_stall_suppress __read_mostly;
460 static void record_gp_stall_check_time(struct rcu_state *rsp)
462 rsp->gp_start = jiffies;
463 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
466 static void print_other_cpu_stall(struct rcu_state *rsp)
471 struct rcu_node *rnp = rcu_get_root(rsp);
473 /* Only let one CPU complain about others per time interval. */
475 raw_spin_lock_irqsave(&rnp->lock, flags);
476 delta = jiffies - rsp->jiffies_stall;
477 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
478 raw_spin_unlock_irqrestore(&rnp->lock, flags);
481 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
484 * Now rat on any tasks that got kicked up to the root rcu_node
485 * due to CPU offlining.
487 rcu_print_task_stall(rnp);
488 raw_spin_unlock_irqrestore(&rnp->lock, flags);
490 /* OK, time to rat on our buddy... */
492 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
494 rcu_for_each_leaf_node(rsp, rnp) {
495 raw_spin_lock_irqsave(&rnp->lock, flags);
496 rcu_print_task_stall(rnp);
497 raw_spin_unlock_irqrestore(&rnp->lock, flags);
498 if (rnp->qsmask == 0)
500 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
501 if (rnp->qsmask & (1UL << cpu))
502 printk(" %d", rnp->grplo + cpu);
504 printk("} (detected by %d, t=%ld jiffies)\n",
505 smp_processor_id(), (long)(jiffies - rsp->gp_start));
506 trigger_all_cpu_backtrace();
508 /* If so configured, complain about tasks blocking the grace period. */
510 rcu_print_detail_task_stall(rsp);
512 force_quiescent_state(rsp, 0); /* Kick them all. */
515 static void print_cpu_stall(struct rcu_state *rsp)
518 struct rcu_node *rnp = rcu_get_root(rsp);
520 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
521 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
522 trigger_all_cpu_backtrace();
524 raw_spin_lock_irqsave(&rnp->lock, flags);
525 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
527 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
528 raw_spin_unlock_irqrestore(&rnp->lock, flags);
530 set_need_resched(); /* kick ourselves to get things going. */
533 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
536 struct rcu_node *rnp;
538 if (rcu_cpu_stall_suppress)
540 delta = jiffies - rsp->jiffies_stall;
542 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
544 /* We haven't checked in, so go dump stack. */
545 print_cpu_stall(rsp);
547 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
549 /* They had two time units to dump stack, so complain. */
550 print_other_cpu_stall(rsp);
554 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
556 rcu_cpu_stall_suppress = 1;
560 static struct notifier_block rcu_panic_block = {
561 .notifier_call = rcu_panic,
564 static void __init check_cpu_stall_init(void)
566 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
569 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
571 static void record_gp_stall_check_time(struct rcu_state *rsp)
575 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
579 static void __init check_cpu_stall_init(void)
583 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
586 * Update CPU-local rcu_data state to record the newly noticed grace period.
587 * This is used both when we started the grace period and when we notice
588 * that someone else started the grace period. The caller must hold the
589 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
590 * and must have irqs disabled.
592 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
594 if (rdp->gpnum != rnp->gpnum) {
596 rdp->passed_quiesc = 0;
597 rdp->gpnum = rnp->gpnum;
601 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
604 struct rcu_node *rnp;
606 local_irq_save(flags);
608 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
609 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
610 local_irq_restore(flags);
613 __note_new_gpnum(rsp, rnp, rdp);
614 raw_spin_unlock_irqrestore(&rnp->lock, flags);
618 * Did someone else start a new RCU grace period start since we last
619 * checked? Update local state appropriately if so. Must be called
620 * on the CPU corresponding to rdp.
623 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
628 local_irq_save(flags);
629 if (rdp->gpnum != rsp->gpnum) {
630 note_new_gpnum(rsp, rdp);
633 local_irq_restore(flags);
638 * Advance this CPU's callbacks, but only if the current grace period
639 * has ended. This may be called only from the CPU to whom the rdp
640 * belongs. In addition, the corresponding leaf rcu_node structure's
641 * ->lock must be held by the caller, with irqs disabled.
644 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
646 /* Did another grace period end? */
647 if (rdp->completed != rnp->completed) {
649 /* Advance callbacks. No harm if list empty. */
650 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
651 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
652 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
654 /* Remember that we saw this grace-period completion. */
655 rdp->completed = rnp->completed;
660 * Advance this CPU's callbacks, but only if the current grace period
661 * has ended. This may be called only from the CPU to whom the rdp
665 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
668 struct rcu_node *rnp;
670 local_irq_save(flags);
672 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
673 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
674 local_irq_restore(flags);
677 __rcu_process_gp_end(rsp, rnp, rdp);
678 raw_spin_unlock_irqrestore(&rnp->lock, flags);
682 * Do per-CPU grace-period initialization for running CPU. The caller
683 * must hold the lock of the leaf rcu_node structure corresponding to
687 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
689 /* Prior grace period ended, so advance callbacks for current CPU. */
690 __rcu_process_gp_end(rsp, rnp, rdp);
693 * Because this CPU just now started the new grace period, we know
694 * that all of its callbacks will be covered by this upcoming grace
695 * period, even the ones that were registered arbitrarily recently.
696 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
698 * Other CPUs cannot be sure exactly when the grace period started.
699 * Therefore, their recently registered callbacks must pass through
700 * an additional RCU_NEXT_READY stage, so that they will be handled
701 * by the next RCU grace period.
703 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
704 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
706 /* Set state so that this CPU will detect the next quiescent state. */
707 __note_new_gpnum(rsp, rnp, rdp);
711 * Start a new RCU grace period if warranted, re-initializing the hierarchy
712 * in preparation for detecting the next grace period. The caller must hold
713 * the root node's ->lock, which is released before return. Hard irqs must
717 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
718 __releases(rcu_get_root(rsp)->lock)
720 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
721 struct rcu_node *rnp = rcu_get_root(rsp);
723 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
724 if (cpu_needs_another_gp(rsp, rdp))
725 rsp->fqs_need_gp = 1;
726 if (rnp->completed == rsp->completed) {
727 raw_spin_unlock_irqrestore(&rnp->lock, flags);
730 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
733 * Propagate new ->completed value to rcu_node structures
734 * so that other CPUs don't have to wait until the start
735 * of the next grace period to process their callbacks.
737 rcu_for_each_node_breadth_first(rsp, rnp) {
738 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
739 rnp->completed = rsp->completed;
740 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
742 local_irq_restore(flags);
746 /* Advance to a new grace period and initialize state. */
748 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
749 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
750 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
751 record_gp_stall_check_time(rsp);
753 /* Special-case the common single-level case. */
754 if (NUM_RCU_NODES == 1) {
755 rcu_preempt_check_blocked_tasks(rnp);
756 rnp->qsmask = rnp->qsmaskinit;
757 rnp->gpnum = rsp->gpnum;
758 rnp->completed = rsp->completed;
759 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
760 rcu_start_gp_per_cpu(rsp, rnp, rdp);
761 raw_spin_unlock_irqrestore(&rnp->lock, flags);
765 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
768 /* Exclude any concurrent CPU-hotplug operations. */
769 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
772 * Set the quiescent-state-needed bits in all the rcu_node
773 * structures for all currently online CPUs in breadth-first
774 * order, starting from the root rcu_node structure. This
775 * operation relies on the layout of the hierarchy within the
776 * rsp->node[] array. Note that other CPUs will access only
777 * the leaves of the hierarchy, which still indicate that no
778 * grace period is in progress, at least until the corresponding
779 * leaf node has been initialized. In addition, we have excluded
780 * CPU-hotplug operations.
782 * Note that the grace period cannot complete until we finish
783 * the initialization process, as there will be at least one
784 * qsmask bit set in the root node until that time, namely the
785 * one corresponding to this CPU, due to the fact that we have
788 rcu_for_each_node_breadth_first(rsp, rnp) {
789 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
790 rcu_preempt_check_blocked_tasks(rnp);
791 rnp->qsmask = rnp->qsmaskinit;
792 rnp->gpnum = rsp->gpnum;
793 rnp->completed = rsp->completed;
794 if (rnp == rdp->mynode)
795 rcu_start_gp_per_cpu(rsp, rnp, rdp);
796 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
799 rnp = rcu_get_root(rsp);
800 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
801 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
802 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
803 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
807 * Report a full set of quiescent states to the specified rcu_state
808 * data structure. This involves cleaning up after the prior grace
809 * period and letting rcu_start_gp() start up the next grace period
810 * if one is needed. Note that the caller must hold rnp->lock, as
811 * required by rcu_start_gp(), which will release it.
813 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
814 __releases(rcu_get_root(rsp)->lock)
816 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
817 rsp->completed = rsp->gpnum;
818 rsp->signaled = RCU_GP_IDLE;
819 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
823 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
824 * Allows quiescent states for a group of CPUs to be reported at one go
825 * to the specified rcu_node structure, though all the CPUs in the group
826 * must be represented by the same rcu_node structure (which need not be
827 * a leaf rcu_node structure, though it often will be). That structure's
828 * lock must be held upon entry, and it is released before return.
831 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
832 struct rcu_node *rnp, unsigned long flags)
833 __releases(rnp->lock)
835 struct rcu_node *rnp_c;
837 /* Walk up the rcu_node hierarchy. */
839 if (!(rnp->qsmask & mask)) {
841 /* Our bit has already been cleared, so done. */
842 raw_spin_unlock_irqrestore(&rnp->lock, flags);
845 rnp->qsmask &= ~mask;
846 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
848 /* Other bits still set at this level, so done. */
849 raw_spin_unlock_irqrestore(&rnp->lock, flags);
853 if (rnp->parent == NULL) {
855 /* No more levels. Exit loop holding root lock. */
859 raw_spin_unlock_irqrestore(&rnp->lock, flags);
862 raw_spin_lock_irqsave(&rnp->lock, flags);
863 WARN_ON_ONCE(rnp_c->qsmask);
867 * Get here if we are the last CPU to pass through a quiescent
868 * state for this grace period. Invoke rcu_report_qs_rsp()
869 * to clean up and start the next grace period if one is needed.
871 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
875 * Record a quiescent state for the specified CPU to that CPU's rcu_data
876 * structure. This must be either called from the specified CPU, or
877 * called when the specified CPU is known to be offline (and when it is
878 * also known that no other CPU is concurrently trying to help the offline
879 * CPU). The lastcomp argument is used to make sure we are still in the
880 * grace period of interest. We don't want to end the current grace period
881 * based on quiescent states detected in an earlier grace period!
884 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
888 struct rcu_node *rnp;
891 raw_spin_lock_irqsave(&rnp->lock, flags);
892 if (lastcomp != rnp->completed) {
895 * Someone beat us to it for this grace period, so leave.
896 * The race with GP start is resolved by the fact that we
897 * hold the leaf rcu_node lock, so that the per-CPU bits
898 * cannot yet be initialized -- so we would simply find our
899 * CPU's bit already cleared in rcu_report_qs_rnp() if this
902 rdp->passed_quiesc = 0; /* try again later! */
903 raw_spin_unlock_irqrestore(&rnp->lock, flags);
907 if ((rnp->qsmask & mask) == 0) {
908 raw_spin_unlock_irqrestore(&rnp->lock, flags);
913 * This GP can't end until cpu checks in, so all of our
914 * callbacks can be processed during the next GP.
916 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
918 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
923 * Check to see if there is a new grace period of which this CPU
924 * is not yet aware, and if so, set up local rcu_data state for it.
925 * Otherwise, see if this CPU has just passed through its first
926 * quiescent state for this grace period, and record that fact if so.
929 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
931 /* If there is now a new grace period, record and return. */
932 if (check_for_new_grace_period(rsp, rdp))
936 * Does this CPU still need to do its part for current grace period?
937 * If no, return and let the other CPUs do their part as well.
939 if (!rdp->qs_pending)
943 * Was there a quiescent state since the beginning of the grace
944 * period? If no, then exit and wait for the next call.
946 if (!rdp->passed_quiesc)
950 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
953 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
956 #ifdef CONFIG_HOTPLUG_CPU
959 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
960 * specified flavor of RCU. The callbacks will be adopted by the next
961 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
962 * comes first. Because this is invoked from the CPU_DYING notifier,
963 * irqs are already disabled.
965 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
968 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
970 if (rdp->nxtlist == NULL)
971 return; /* irqs disabled, so comparison is stable. */
972 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
973 *rsp->orphan_cbs_tail = rdp->nxtlist;
974 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
976 for (i = 0; i < RCU_NEXT_SIZE; i++)
977 rdp->nxttail[i] = &rdp->nxtlist;
978 rsp->orphan_qlen += rdp->qlen;
980 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
984 * Adopt previously orphaned RCU callbacks.
986 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
989 struct rcu_data *rdp;
991 raw_spin_lock_irqsave(&rsp->onofflock, flags);
992 rdp = this_cpu_ptr(rsp->rda);
993 if (rsp->orphan_cbs_list == NULL) {
994 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
997 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
998 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
999 rdp->qlen += rsp->orphan_qlen;
1000 rsp->orphan_cbs_list = NULL;
1001 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
1002 rsp->orphan_qlen = 0;
1003 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1007 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1008 * and move all callbacks from the outgoing CPU to the current one.
1010 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1012 unsigned long flags;
1014 int need_report = 0;
1015 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1016 struct rcu_node *rnp;
1018 /* Exclude any attempts to start a new grace period. */
1019 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1021 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1022 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1023 mask = rdp->grpmask; /* rnp->grplo is constant. */
1025 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1026 rnp->qsmaskinit &= ~mask;
1027 if (rnp->qsmaskinit != 0) {
1028 if (rnp != rdp->mynode)
1029 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1032 if (rnp == rdp->mynode)
1033 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1035 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1036 mask = rnp->grpmask;
1038 } while (rnp != NULL);
1041 * We still hold the leaf rcu_node structure lock here, and
1042 * irqs are still disabled. The reason for this subterfuge is
1043 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1044 * held leads to deadlock.
1046 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1048 if (need_report & RCU_OFL_TASKS_NORM_GP)
1049 rcu_report_unblock_qs_rnp(rnp, flags);
1051 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1052 if (need_report & RCU_OFL_TASKS_EXP_GP)
1053 rcu_report_exp_rnp(rsp, rnp);
1055 rcu_adopt_orphan_cbs(rsp);
1059 * Remove the specified CPU from the RCU hierarchy and move any pending
1060 * callbacks that it might have to the current CPU. This code assumes
1061 * that at least one CPU in the system will remain running at all times.
1062 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1064 static void rcu_offline_cpu(int cpu)
1066 __rcu_offline_cpu(cpu, &rcu_sched_state);
1067 __rcu_offline_cpu(cpu, &rcu_bh_state);
1068 rcu_preempt_offline_cpu(cpu);
1071 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1073 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1077 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1081 static void rcu_offline_cpu(int cpu)
1085 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1088 * Invoke any RCU callbacks that have made it to the end of their grace
1089 * period. Thottle as specified by rdp->blimit.
1091 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1093 unsigned long flags;
1094 struct rcu_head *next, *list, **tail;
1097 /* If no callbacks are ready, just return.*/
1098 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1102 * Extract the list of ready callbacks, disabling to prevent
1103 * races with call_rcu() from interrupt handlers.
1105 local_irq_save(flags);
1106 list = rdp->nxtlist;
1107 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1108 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1109 tail = rdp->nxttail[RCU_DONE_TAIL];
1110 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1111 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1112 rdp->nxttail[count] = &rdp->nxtlist;
1113 local_irq_restore(flags);
1115 /* Invoke callbacks. */
1120 debug_rcu_head_unqueue(list);
1123 if (++count >= rdp->blimit)
1127 local_irq_save(flags);
1129 /* Update count, and requeue any remaining callbacks. */
1132 *tail = rdp->nxtlist;
1133 rdp->nxtlist = list;
1134 for (count = 0; count < RCU_NEXT_SIZE; count++)
1135 if (&rdp->nxtlist == rdp->nxttail[count])
1136 rdp->nxttail[count] = tail;
1141 /* Reinstate batch limit if we have worked down the excess. */
1142 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1143 rdp->blimit = blimit;
1145 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1146 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1147 rdp->qlen_last_fqs_check = 0;
1148 rdp->n_force_qs_snap = rsp->n_force_qs;
1149 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1150 rdp->qlen_last_fqs_check = rdp->qlen;
1152 local_irq_restore(flags);
1154 /* Re-raise the RCU softirq if there are callbacks remaining. */
1155 if (cpu_has_callbacks_ready_to_invoke(rdp))
1156 raise_softirq(RCU_SOFTIRQ);
1160 * Check to see if this CPU is in a non-context-switch quiescent state
1161 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1162 * Also schedule the RCU softirq handler.
1164 * This function must be called with hardirqs disabled. It is normally
1165 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1166 * false, there is no point in invoking rcu_check_callbacks().
1168 void rcu_check_callbacks(int cpu, int user)
1171 (idle_cpu(cpu) && rcu_scheduler_active &&
1172 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1175 * Get here if this CPU took its interrupt from user
1176 * mode or from the idle loop, and if this is not a
1177 * nested interrupt. In this case, the CPU is in
1178 * a quiescent state, so note it.
1180 * No memory barrier is required here because both
1181 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1182 * variables that other CPUs neither access nor modify,
1183 * at least not while the corresponding CPU is online.
1189 } else if (!in_softirq()) {
1192 * Get here if this CPU did not take its interrupt from
1193 * softirq, in other words, if it is not interrupting
1194 * a rcu_bh read-side critical section. This is an _bh
1195 * critical section, so note it.
1200 rcu_preempt_check_callbacks(cpu);
1201 if (rcu_pending(cpu))
1202 raise_softirq(RCU_SOFTIRQ);
1208 * Scan the leaf rcu_node structures, processing dyntick state for any that
1209 * have not yet encountered a quiescent state, using the function specified.
1210 * The caller must have suppressed start of new grace periods.
1212 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1216 unsigned long flags;
1218 struct rcu_node *rnp;
1220 rcu_for_each_leaf_node(rsp, rnp) {
1222 raw_spin_lock_irqsave(&rnp->lock, flags);
1223 if (!rcu_gp_in_progress(rsp)) {
1224 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1227 if (rnp->qsmask == 0) {
1228 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1233 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1234 if ((rnp->qsmask & bit) != 0 &&
1235 f(per_cpu_ptr(rsp->rda, cpu)))
1240 /* rcu_report_qs_rnp() releases rnp->lock. */
1241 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1244 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1249 * Force quiescent states on reluctant CPUs, and also detect which
1250 * CPUs are in dyntick-idle mode.
1252 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1254 unsigned long flags;
1255 struct rcu_node *rnp = rcu_get_root(rsp);
1257 if (!rcu_gp_in_progress(rsp))
1258 return; /* No grace period in progress, nothing to force. */
1259 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1260 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1261 return; /* Someone else is already on the job. */
1263 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1264 goto unlock_fqs_ret; /* no emergency and done recently. */
1266 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1267 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1268 if(!rcu_gp_in_progress(rsp)) {
1269 rsp->n_force_qs_ngp++;
1270 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1271 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1273 rsp->fqs_active = 1;
1274 switch (rsp->signaled) {
1278 break; /* grace period idle or initializing, ignore. */
1280 case RCU_SAVE_DYNTICK:
1281 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1282 break; /* So gcc recognizes the dead code. */
1284 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1286 /* Record dyntick-idle state. */
1287 force_qs_rnp(rsp, dyntick_save_progress_counter);
1288 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1289 if (rcu_gp_in_progress(rsp))
1290 rsp->signaled = RCU_FORCE_QS;
1295 /* Check dyntick-idle state, send IPI to laggarts. */
1296 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1297 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1299 /* Leave state in case more forcing is required. */
1301 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1304 rsp->fqs_active = 0;
1305 if (rsp->fqs_need_gp) {
1306 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1307 rsp->fqs_need_gp = 0;
1308 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1311 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1313 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1316 #else /* #ifdef CONFIG_SMP */
1318 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1323 #endif /* #else #ifdef CONFIG_SMP */
1326 * This does the RCU processing work from softirq context for the
1327 * specified rcu_state and rcu_data structures. This may be called
1328 * only from the CPU to whom the rdp belongs.
1331 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1333 unsigned long flags;
1335 WARN_ON_ONCE(rdp->beenonline == 0);
1338 * If an RCU GP has gone long enough, go check for dyntick
1339 * idle CPUs and, if needed, send resched IPIs.
1341 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1342 force_quiescent_state(rsp, 1);
1345 * Advance callbacks in response to end of earlier grace
1346 * period that some other CPU ended.
1348 rcu_process_gp_end(rsp, rdp);
1350 /* Update RCU state based on any recent quiescent states. */
1351 rcu_check_quiescent_state(rsp, rdp);
1353 /* Does this CPU require a not-yet-started grace period? */
1354 if (cpu_needs_another_gp(rsp, rdp)) {
1355 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1356 rcu_start_gp(rsp, flags); /* releases above lock */
1359 /* If there are callbacks ready, invoke them. */
1360 rcu_do_batch(rsp, rdp);
1364 * Do softirq processing for the current CPU.
1366 static void rcu_process_callbacks(struct softirq_action *unused)
1369 * Memory references from any prior RCU read-side critical sections
1370 * executed by the interrupted code must be seen before any RCU
1371 * grace-period manipulations below.
1373 smp_mb(); /* See above block comment. */
1375 __rcu_process_callbacks(&rcu_sched_state,
1376 &__get_cpu_var(rcu_sched_data));
1377 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1378 rcu_preempt_process_callbacks();
1381 * Memory references from any later RCU read-side critical sections
1382 * executed by the interrupted code must be seen after any RCU
1383 * grace-period manipulations above.
1385 smp_mb(); /* See above block comment. */
1387 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1388 rcu_needs_cpu_flush();
1392 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1393 struct rcu_state *rsp)
1395 unsigned long flags;
1396 struct rcu_data *rdp;
1398 debug_rcu_head_queue(head);
1402 smp_mb(); /* Ensure RCU update seen before callback registry. */
1405 * Opportunistically note grace-period endings and beginnings.
1406 * Note that we might see a beginning right after we see an
1407 * end, but never vice versa, since this CPU has to pass through
1408 * a quiescent state betweentimes.
1410 local_irq_save(flags);
1411 rdp = this_cpu_ptr(rsp->rda);
1412 rcu_process_gp_end(rsp, rdp);
1413 check_for_new_grace_period(rsp, rdp);
1415 /* Add the callback to our list. */
1416 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1417 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1419 /* Start a new grace period if one not already started. */
1420 if (!rcu_gp_in_progress(rsp)) {
1421 unsigned long nestflag;
1422 struct rcu_node *rnp_root = rcu_get_root(rsp);
1424 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1425 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1429 * Force the grace period if too many callbacks or too long waiting.
1430 * Enforce hysteresis, and don't invoke force_quiescent_state()
1431 * if some other CPU has recently done so. Also, don't bother
1432 * invoking force_quiescent_state() if the newly enqueued callback
1433 * is the only one waiting for a grace period to complete.
1435 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1436 rdp->blimit = LONG_MAX;
1437 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1438 *rdp->nxttail[RCU_DONE_TAIL] != head)
1439 force_quiescent_state(rsp, 0);
1440 rdp->n_force_qs_snap = rsp->n_force_qs;
1441 rdp->qlen_last_fqs_check = rdp->qlen;
1442 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1443 force_quiescent_state(rsp, 1);
1444 local_irq_restore(flags);
1448 * Queue an RCU-sched callback for invocation after a grace period.
1450 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1452 __call_rcu(head, func, &rcu_sched_state);
1454 EXPORT_SYMBOL_GPL(call_rcu_sched);
1457 * Queue an RCU for invocation after a quicker grace period.
1459 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1461 __call_rcu(head, func, &rcu_bh_state);
1463 EXPORT_SYMBOL_GPL(call_rcu_bh);
1466 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1468 * Control will return to the caller some time after a full rcu-sched
1469 * grace period has elapsed, in other words after all currently executing
1470 * rcu-sched read-side critical sections have completed. These read-side
1471 * critical sections are delimited by rcu_read_lock_sched() and
1472 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1473 * local_irq_disable(), and so on may be used in place of
1474 * rcu_read_lock_sched().
1476 * This means that all preempt_disable code sequences, including NMI and
1477 * hardware-interrupt handlers, in progress on entry will have completed
1478 * before this primitive returns. However, this does not guarantee that
1479 * softirq handlers will have completed, since in some kernels, these
1480 * handlers can run in process context, and can block.
1482 * This primitive provides the guarantees made by the (now removed)
1483 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1484 * guarantees that rcu_read_lock() sections will have completed.
1485 * In "classic RCU", these two guarantees happen to be one and
1486 * the same, but can differ in realtime RCU implementations.
1488 void synchronize_sched(void)
1490 struct rcu_synchronize rcu;
1492 if (rcu_blocking_is_gp())
1495 init_rcu_head_on_stack(&rcu.head);
1496 init_completion(&rcu.completion);
1497 /* Will wake me after RCU finished. */
1498 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1500 wait_for_completion(&rcu.completion);
1501 destroy_rcu_head_on_stack(&rcu.head);
1503 EXPORT_SYMBOL_GPL(synchronize_sched);
1506 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1508 * Control will return to the caller some time after a full rcu_bh grace
1509 * period has elapsed, in other words after all currently executing rcu_bh
1510 * read-side critical sections have completed. RCU read-side critical
1511 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1512 * and may be nested.
1514 void synchronize_rcu_bh(void)
1516 struct rcu_synchronize rcu;
1518 if (rcu_blocking_is_gp())
1521 init_rcu_head_on_stack(&rcu.head);
1522 init_completion(&rcu.completion);
1523 /* Will wake me after RCU finished. */
1524 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1526 wait_for_completion(&rcu.completion);
1527 destroy_rcu_head_on_stack(&rcu.head);
1529 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1532 * Check to see if there is any immediate RCU-related work to be done
1533 * by the current CPU, for the specified type of RCU, returning 1 if so.
1534 * The checks are in order of increasing expense: checks that can be
1535 * carried out against CPU-local state are performed first. However,
1536 * we must check for CPU stalls first, else we might not get a chance.
1538 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1540 struct rcu_node *rnp = rdp->mynode;
1542 rdp->n_rcu_pending++;
1544 /* Check for CPU stalls, if enabled. */
1545 check_cpu_stall(rsp, rdp);
1547 /* Is the RCU core waiting for a quiescent state from this CPU? */
1548 if (rdp->qs_pending && !rdp->passed_quiesc) {
1551 * If force_quiescent_state() coming soon and this CPU
1552 * needs a quiescent state, and this is either RCU-sched
1553 * or RCU-bh, force a local reschedule.
1555 rdp->n_rp_qs_pending++;
1556 if (!rdp->preemptable &&
1557 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1560 } else if (rdp->qs_pending && rdp->passed_quiesc) {
1561 rdp->n_rp_report_qs++;
1565 /* Does this CPU have callbacks ready to invoke? */
1566 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1567 rdp->n_rp_cb_ready++;
1571 /* Has RCU gone idle with this CPU needing another grace period? */
1572 if (cpu_needs_another_gp(rsp, rdp)) {
1573 rdp->n_rp_cpu_needs_gp++;
1577 /* Has another RCU grace period completed? */
1578 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1579 rdp->n_rp_gp_completed++;
1583 /* Has a new RCU grace period started? */
1584 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1585 rdp->n_rp_gp_started++;
1589 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1590 if (rcu_gp_in_progress(rsp) &&
1591 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1592 rdp->n_rp_need_fqs++;
1597 rdp->n_rp_need_nothing++;
1602 * Check to see if there is any immediate RCU-related work to be done
1603 * by the current CPU, returning 1 if so. This function is part of the
1604 * RCU implementation; it is -not- an exported member of the RCU API.
1606 static int rcu_pending(int cpu)
1608 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1609 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1610 rcu_preempt_pending(cpu);
1614 * Check to see if any future RCU-related work will need to be done
1615 * by the current CPU, even if none need be done immediately, returning
1618 static int rcu_needs_cpu_quick_check(int cpu)
1620 /* RCU callbacks either ready or pending? */
1621 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1622 per_cpu(rcu_bh_data, cpu).nxtlist ||
1623 rcu_preempt_needs_cpu(cpu);
1626 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1627 static atomic_t rcu_barrier_cpu_count;
1628 static DEFINE_MUTEX(rcu_barrier_mutex);
1629 static struct completion rcu_barrier_completion;
1631 static void rcu_barrier_callback(struct rcu_head *notused)
1633 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1634 complete(&rcu_barrier_completion);
1638 * Called with preemption disabled, and from cross-cpu IRQ context.
1640 static void rcu_barrier_func(void *type)
1642 int cpu = smp_processor_id();
1643 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1644 void (*call_rcu_func)(struct rcu_head *head,
1645 void (*func)(struct rcu_head *head));
1647 atomic_inc(&rcu_barrier_cpu_count);
1648 call_rcu_func = type;
1649 call_rcu_func(head, rcu_barrier_callback);
1653 * Orchestrate the specified type of RCU barrier, waiting for all
1654 * RCU callbacks of the specified type to complete.
1656 static void _rcu_barrier(struct rcu_state *rsp,
1657 void (*call_rcu_func)(struct rcu_head *head,
1658 void (*func)(struct rcu_head *head)))
1660 BUG_ON(in_interrupt());
1661 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1662 mutex_lock(&rcu_barrier_mutex);
1663 init_completion(&rcu_barrier_completion);
1665 * Initialize rcu_barrier_cpu_count to 1, then invoke
1666 * rcu_barrier_func() on each CPU, so that each CPU also has
1667 * incremented rcu_barrier_cpu_count. Only then is it safe to
1668 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1669 * might complete its grace period before all of the other CPUs
1670 * did their increment, causing this function to return too
1673 atomic_set(&rcu_barrier_cpu_count, 1);
1674 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1675 rcu_adopt_orphan_cbs(rsp);
1676 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1677 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1678 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1679 complete(&rcu_barrier_completion);
1680 wait_for_completion(&rcu_barrier_completion);
1681 mutex_unlock(&rcu_barrier_mutex);
1685 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1687 void rcu_barrier_bh(void)
1689 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1691 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1694 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1696 void rcu_barrier_sched(void)
1698 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1700 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1703 * Do boot-time initialization of a CPU's per-CPU RCU data.
1706 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1708 unsigned long flags;
1710 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1711 struct rcu_node *rnp = rcu_get_root(rsp);
1713 /* Set up local state, ensuring consistent view of global state. */
1714 raw_spin_lock_irqsave(&rnp->lock, flags);
1715 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1716 rdp->nxtlist = NULL;
1717 for (i = 0; i < RCU_NEXT_SIZE; i++)
1718 rdp->nxttail[i] = &rdp->nxtlist;
1721 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1722 #endif /* #ifdef CONFIG_NO_HZ */
1724 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1728 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1729 * offline event can be happening at a given time. Note also that we
1730 * can accept some slop in the rsp->completed access due to the fact
1731 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1733 static void __cpuinit
1734 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1736 unsigned long flags;
1738 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1739 struct rcu_node *rnp = rcu_get_root(rsp);
1741 /* Set up local state, ensuring consistent view of global state. */
1742 raw_spin_lock_irqsave(&rnp->lock, flags);
1743 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1744 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1745 rdp->beenonline = 1; /* We have now been online. */
1746 rdp->preemptable = preemptable;
1747 rdp->qlen_last_fqs_check = 0;
1748 rdp->n_force_qs_snap = rsp->n_force_qs;
1749 rdp->blimit = blimit;
1750 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1753 * A new grace period might start here. If so, we won't be part
1754 * of it, but that is OK, as we are currently in a quiescent state.
1757 /* Exclude any attempts to start a new GP on large systems. */
1758 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1760 /* Add CPU to rcu_node bitmasks. */
1762 mask = rdp->grpmask;
1764 /* Exclude any attempts to start a new GP on small systems. */
1765 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1766 rnp->qsmaskinit |= mask;
1767 mask = rnp->grpmask;
1768 if (rnp == rdp->mynode) {
1769 rdp->gpnum = rnp->completed; /* if GP in progress... */
1770 rdp->completed = rnp->completed;
1771 rdp->passed_quiesc_completed = rnp->completed - 1;
1773 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1775 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1777 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1780 static void __cpuinit rcu_online_cpu(int cpu)
1782 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1783 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1784 rcu_preempt_init_percpu_data(cpu);
1788 * Handle CPU online/offline notification events.
1790 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1791 unsigned long action, void *hcpu)
1793 long cpu = (long)hcpu;
1796 case CPU_UP_PREPARE:
1797 case CPU_UP_PREPARE_FROZEN:
1798 rcu_online_cpu(cpu);
1801 case CPU_DYING_FROZEN:
1803 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1804 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1805 * returns, all online cpus have queued rcu_barrier_func().
1806 * The dying CPU clears its cpu_online_mask bit and
1807 * moves all of its RCU callbacks to ->orphan_cbs_list
1808 * in the context of stop_machine(), so subsequent calls
1809 * to _rcu_barrier() will adopt these callbacks and only
1810 * then queue rcu_barrier_func() on all remaining CPUs.
1812 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1813 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1814 rcu_preempt_send_cbs_to_orphanage();
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;
1855 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1856 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1863 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1864 ccur = rsp->levelcnt[i];
1865 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1869 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1872 * Helper function for rcu_init() that initializes one rcu_state structure.
1874 static void __init rcu_init_one(struct rcu_state *rsp,
1875 struct rcu_data __percpu *rda)
1877 static char *buf[] = { "rcu_node_level_0",
1880 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1884 struct rcu_node *rnp;
1886 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1888 /* Initialize the level-tracking arrays. */
1890 for (i = 1; i < NUM_RCU_LVLS; i++)
1891 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1892 rcu_init_levelspread(rsp);
1894 /* Initialize the elements themselves, starting from the leaves. */
1896 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1897 cpustride *= rsp->levelspread[i];
1898 rnp = rsp->level[i];
1899 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1900 raw_spin_lock_init(&rnp->lock);
1901 lockdep_set_class_and_name(&rnp->lock,
1902 &rcu_node_class[i], buf[i]);
1905 rnp->qsmaskinit = 0;
1906 rnp->grplo = j * cpustride;
1907 rnp->grphi = (j + 1) * cpustride - 1;
1908 if (rnp->grphi >= NR_CPUS)
1909 rnp->grphi = NR_CPUS - 1;
1915 rnp->grpnum = j % rsp->levelspread[i - 1];
1916 rnp->grpmask = 1UL << rnp->grpnum;
1917 rnp->parent = rsp->level[i - 1] +
1918 j / rsp->levelspread[i - 1];
1921 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1922 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1923 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1924 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1929 rnp = rsp->level[NUM_RCU_LVLS - 1];
1930 for_each_possible_cpu(i) {
1931 while (i > rnp->grphi)
1933 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
1934 rcu_boot_init_percpu_data(i, rsp);
1938 void __init rcu_init(void)
1942 rcu_bootup_announce();
1943 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
1944 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
1945 __rcu_init_preempt();
1946 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1949 * We don't need protection against CPU-hotplug here because
1950 * this is called early in boot, before either interrupts
1951 * or the scheduler are operational.
1953 cpu_notifier(rcu_cpu_notify, 0);
1954 for_each_online_cpu(cpu)
1955 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1956 check_cpu_stall_init();
1959 #include "rcutree_plugin.h"