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 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
144 int rcu_cpu_stall_suppress __read_mostly = RCU_CPU_STALL_SUPPRESS_INIT;
145 module_param(rcu_cpu_stall_suppress, int, 0644);
146 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
148 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
149 static int rcu_pending(int cpu);
152 * Return the number of RCU-sched batches processed thus far for debug & stats.
154 long rcu_batches_completed_sched(void)
156 return rcu_sched_state.completed;
158 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
161 * Return the number of RCU BH batches processed thus far for debug & stats.
163 long rcu_batches_completed_bh(void)
165 return rcu_bh_state.completed;
167 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
170 * Force a quiescent state for RCU BH.
172 void rcu_bh_force_quiescent_state(void)
174 force_quiescent_state(&rcu_bh_state, 0);
176 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
179 * Force a quiescent state for RCU-sched.
181 void rcu_sched_force_quiescent_state(void)
183 force_quiescent_state(&rcu_sched_state, 0);
185 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
188 * Does the CPU have callbacks ready to be invoked?
191 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
193 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
197 * Does the current CPU require a yet-as-unscheduled grace period?
200 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
202 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
206 * Return the root node of the specified rcu_state structure.
208 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
210 return &rsp->node[0];
216 * If the specified CPU is offline, tell the caller that it is in
217 * a quiescent state. Otherwise, whack it with a reschedule IPI.
218 * Grace periods can end up waiting on an offline CPU when that
219 * CPU is in the process of coming online -- it will be added to the
220 * rcu_node bitmasks before it actually makes it online. The same thing
221 * can happen while a CPU is in the process of coming online. Because this
222 * race is quite rare, we check for it after detecting that the grace
223 * period has been delayed rather than checking each and every CPU
224 * each and every time we start a new grace period.
226 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
229 * If the CPU is offline, it is in a quiescent state. We can
230 * trust its state not to change because interrupts are disabled.
232 if (cpu_is_offline(rdp->cpu)) {
237 /* If preemptable RCU, no point in sending reschedule IPI. */
238 if (rdp->preemptable)
241 /* The CPU is online, so send it a reschedule IPI. */
242 if (rdp->cpu != smp_processor_id())
243 smp_send_reschedule(rdp->cpu);
250 #endif /* #ifdef CONFIG_SMP */
255 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
257 * Enter nohz mode, in other words, -leave- the mode in which RCU
258 * read-side critical sections can occur. (Though RCU read-side
259 * critical sections can occur in irq handlers in nohz mode, a possibility
260 * handled by rcu_irq_enter() and rcu_irq_exit()).
262 void rcu_enter_nohz(void)
265 struct rcu_dynticks *rdtp;
267 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
268 local_irq_save(flags);
269 rdtp = &__get_cpu_var(rcu_dynticks);
271 rdtp->dynticks_nesting--;
272 WARN_ON_ONCE(rdtp->dynticks & 0x1);
273 local_irq_restore(flags);
277 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
279 * Exit nohz mode, in other words, -enter- the mode in which RCU
280 * read-side critical sections normally occur.
282 void rcu_exit_nohz(void)
285 struct rcu_dynticks *rdtp;
287 local_irq_save(flags);
288 rdtp = &__get_cpu_var(rcu_dynticks);
290 rdtp->dynticks_nesting++;
291 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
292 local_irq_restore(flags);
293 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
297 * rcu_nmi_enter - inform RCU of entry to NMI context
299 * If the CPU was idle with dynamic ticks active, and there is no
300 * irq handler running, this updates rdtp->dynticks_nmi to let the
301 * RCU grace-period handling know that the CPU is active.
303 void rcu_nmi_enter(void)
305 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
307 if (rdtp->dynticks & 0x1)
309 rdtp->dynticks_nmi++;
310 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
311 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
315 * rcu_nmi_exit - inform RCU of exit from NMI context
317 * If the CPU was idle with dynamic ticks active, and there is no
318 * irq handler running, this updates rdtp->dynticks_nmi to let the
319 * RCU grace-period handling know that the CPU is no longer active.
321 void rcu_nmi_exit(void)
323 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
325 if (rdtp->dynticks & 0x1)
327 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
328 rdtp->dynticks_nmi++;
329 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
333 * rcu_irq_enter - inform RCU of entry to hard irq context
335 * If the CPU was idle with dynamic ticks active, this updates the
336 * rdtp->dynticks to let the RCU handling know that the CPU is active.
338 void rcu_irq_enter(void)
340 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
342 if (rdtp->dynticks_nesting++)
345 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
346 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
350 * rcu_irq_exit - inform RCU of exit from hard irq context
352 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
353 * to put let the RCU handling be aware that the CPU is going back to idle
356 void rcu_irq_exit(void)
358 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
360 if (--rdtp->dynticks_nesting)
362 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
364 WARN_ON_ONCE(rdtp->dynticks & 0x1);
366 /* If the interrupt queued a callback, get out of dyntick mode. */
367 if (__get_cpu_var(rcu_sched_data).nxtlist ||
368 __get_cpu_var(rcu_bh_data).nxtlist)
375 * Snapshot the specified CPU's dynticks counter so that we can later
376 * credit them with an implicit quiescent state. Return 1 if this CPU
377 * is in dynticks idle mode, which is an extended quiescent state.
379 static int dyntick_save_progress_counter(struct rcu_data *rdp)
385 snap = rdp->dynticks->dynticks;
386 snap_nmi = rdp->dynticks->dynticks_nmi;
387 smp_mb(); /* Order sampling of snap with end of grace period. */
388 rdp->dynticks_snap = snap;
389 rdp->dynticks_nmi_snap = snap_nmi;
390 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
397 * Return true if the specified CPU has passed through a quiescent
398 * state by virtue of being in or having passed through an dynticks
399 * idle state since the last call to dyntick_save_progress_counter()
402 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
409 curr = rdp->dynticks->dynticks;
410 snap = rdp->dynticks_snap;
411 curr_nmi = rdp->dynticks->dynticks_nmi;
412 snap_nmi = rdp->dynticks_nmi_snap;
413 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
416 * If the CPU passed through or entered a dynticks idle phase with
417 * no active irq/NMI handlers, then we can safely pretend that the CPU
418 * already acknowledged the request to pass through a quiescent
419 * state. Either way, that CPU cannot possibly be in an RCU
420 * read-side critical section that started before the beginning
421 * of the current RCU grace period.
423 if ((curr != snap || (curr & 0x1) == 0) &&
424 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
429 /* Go check for the CPU being offline. */
430 return rcu_implicit_offline_qs(rdp);
433 #endif /* #ifdef CONFIG_SMP */
435 #else /* #ifdef CONFIG_NO_HZ */
439 static int dyntick_save_progress_counter(struct rcu_data *rdp)
444 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
446 return rcu_implicit_offline_qs(rdp);
449 #endif /* #ifdef CONFIG_SMP */
451 #endif /* #else #ifdef CONFIG_NO_HZ */
453 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
455 int rcu_cpu_stall_suppress __read_mostly;
457 static void record_gp_stall_check_time(struct rcu_state *rsp)
459 rsp->gp_start = jiffies;
460 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
463 static void print_other_cpu_stall(struct rcu_state *rsp)
468 struct rcu_node *rnp = rcu_get_root(rsp);
470 /* Only let one CPU complain about others per time interval. */
472 raw_spin_lock_irqsave(&rnp->lock, flags);
473 delta = jiffies - rsp->jiffies_stall;
474 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
475 raw_spin_unlock_irqrestore(&rnp->lock, flags);
478 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
481 * Now rat on any tasks that got kicked up to the root rcu_node
482 * due to CPU offlining.
484 rcu_print_task_stall(rnp);
485 raw_spin_unlock_irqrestore(&rnp->lock, flags);
488 * OK, time to rat on our buddy...
489 * See Documentation/RCU/stallwarn.txt for info on how to debug
490 * RCU CPU stall warnings.
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);
521 * OK, time to rat on ourselves...
522 * See Documentation/RCU/stallwarn.txt for info on how to debug
523 * RCU CPU stall warnings.
525 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
526 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
527 trigger_all_cpu_backtrace();
529 raw_spin_lock_irqsave(&rnp->lock, flags);
530 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
532 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
533 raw_spin_unlock_irqrestore(&rnp->lock, flags);
535 set_need_resched(); /* kick ourselves to get things going. */
538 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
541 struct rcu_node *rnp;
543 if (rcu_cpu_stall_suppress)
545 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
547 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && delta >= 0) {
549 /* We haven't checked in, so go dump stack. */
550 print_cpu_stall(rsp);
552 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
554 /* They had two time units to dump stack, so complain. */
555 print_other_cpu_stall(rsp);
559 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
561 rcu_cpu_stall_suppress = 1;
566 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
568 * Set the stall-warning timeout way off into the future, thus preventing
569 * any RCU CPU stall-warning messages from appearing in the current set of
572 * The caller must disable hard irqs.
574 void rcu_cpu_stall_reset(void)
576 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
577 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
578 rcu_preempt_stall_reset();
581 static struct notifier_block rcu_panic_block = {
582 .notifier_call = rcu_panic,
585 static void __init check_cpu_stall_init(void)
587 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
590 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
592 static void record_gp_stall_check_time(struct rcu_state *rsp)
596 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
600 void rcu_cpu_stall_reset(void)
604 static void __init check_cpu_stall_init(void)
608 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
611 * Update CPU-local rcu_data state to record the newly noticed grace period.
612 * This is used both when we started the grace period and when we notice
613 * that someone else started the grace period. The caller must hold the
614 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
615 * and must have irqs disabled.
617 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
619 if (rdp->gpnum != rnp->gpnum) {
621 rdp->passed_quiesc = 0;
622 rdp->gpnum = rnp->gpnum;
626 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
629 struct rcu_node *rnp;
631 local_irq_save(flags);
633 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
634 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
635 local_irq_restore(flags);
638 __note_new_gpnum(rsp, rnp, rdp);
639 raw_spin_unlock_irqrestore(&rnp->lock, flags);
643 * Did someone else start a new RCU grace period start since we last
644 * checked? Update local state appropriately if so. Must be called
645 * on the CPU corresponding to rdp.
648 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
653 local_irq_save(flags);
654 if (rdp->gpnum != rsp->gpnum) {
655 note_new_gpnum(rsp, rdp);
658 local_irq_restore(flags);
663 * Advance this CPU's callbacks, but only if the current grace period
664 * has ended. This may be called only from the CPU to whom the rdp
665 * belongs. In addition, the corresponding leaf rcu_node structure's
666 * ->lock must be held by the caller, with irqs disabled.
669 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
671 /* Did another grace period end? */
672 if (rdp->completed != rnp->completed) {
674 /* Advance callbacks. No harm if list empty. */
675 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
676 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
677 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
679 /* Remember that we saw this grace-period completion. */
680 rdp->completed = rnp->completed;
685 * Advance this CPU's callbacks, but only if the current grace period
686 * has ended. This may be called only from the CPU to whom the rdp
690 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
693 struct rcu_node *rnp;
695 local_irq_save(flags);
697 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
698 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
699 local_irq_restore(flags);
702 __rcu_process_gp_end(rsp, rnp, rdp);
703 raw_spin_unlock_irqrestore(&rnp->lock, flags);
707 * Do per-CPU grace-period initialization for running CPU. The caller
708 * must hold the lock of the leaf rcu_node structure corresponding to
712 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
714 /* Prior grace period ended, so advance callbacks for current CPU. */
715 __rcu_process_gp_end(rsp, rnp, rdp);
718 * Because this CPU just now started the new grace period, we know
719 * that all of its callbacks will be covered by this upcoming grace
720 * period, even the ones that were registered arbitrarily recently.
721 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
723 * Other CPUs cannot be sure exactly when the grace period started.
724 * Therefore, their recently registered callbacks must pass through
725 * an additional RCU_NEXT_READY stage, so that they will be handled
726 * by the next RCU grace period.
728 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
729 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
731 /* Set state so that this CPU will detect the next quiescent state. */
732 __note_new_gpnum(rsp, rnp, rdp);
736 * Start a new RCU grace period if warranted, re-initializing the hierarchy
737 * in preparation for detecting the next grace period. The caller must hold
738 * the root node's ->lock, which is released before return. Hard irqs must
742 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
743 __releases(rcu_get_root(rsp)->lock)
745 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
746 struct rcu_node *rnp = rcu_get_root(rsp);
748 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
749 if (cpu_needs_another_gp(rsp, rdp))
750 rsp->fqs_need_gp = 1;
751 if (rnp->completed == rsp->completed) {
752 raw_spin_unlock_irqrestore(&rnp->lock, flags);
755 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
758 * Propagate new ->completed value to rcu_node structures
759 * so that other CPUs don't have to wait until the start
760 * of the next grace period to process their callbacks.
762 rcu_for_each_node_breadth_first(rsp, rnp) {
763 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
764 rnp->completed = rsp->completed;
765 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
767 local_irq_restore(flags);
771 /* Advance to a new grace period and initialize state. */
773 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
774 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
775 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
776 record_gp_stall_check_time(rsp);
778 /* Special-case the common single-level case. */
779 if (NUM_RCU_NODES == 1) {
780 rcu_preempt_check_blocked_tasks(rnp);
781 rnp->qsmask = rnp->qsmaskinit;
782 rnp->gpnum = rsp->gpnum;
783 rnp->completed = rsp->completed;
784 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
785 rcu_start_gp_per_cpu(rsp, rnp, rdp);
786 raw_spin_unlock_irqrestore(&rnp->lock, flags);
790 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
793 /* Exclude any concurrent CPU-hotplug operations. */
794 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
797 * Set the quiescent-state-needed bits in all the rcu_node
798 * structures for all currently online CPUs in breadth-first
799 * order, starting from the root rcu_node structure. This
800 * operation relies on the layout of the hierarchy within the
801 * rsp->node[] array. Note that other CPUs will access only
802 * the leaves of the hierarchy, which still indicate that no
803 * grace period is in progress, at least until the corresponding
804 * leaf node has been initialized. In addition, we have excluded
805 * CPU-hotplug operations.
807 * Note that the grace period cannot complete until we finish
808 * the initialization process, as there will be at least one
809 * qsmask bit set in the root node until that time, namely the
810 * one corresponding to this CPU, due to the fact that we have
813 rcu_for_each_node_breadth_first(rsp, rnp) {
814 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
815 rcu_preempt_check_blocked_tasks(rnp);
816 rnp->qsmask = rnp->qsmaskinit;
817 rnp->gpnum = rsp->gpnum;
818 rnp->completed = rsp->completed;
819 if (rnp == rdp->mynode)
820 rcu_start_gp_per_cpu(rsp, rnp, rdp);
821 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
824 rnp = rcu_get_root(rsp);
825 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
826 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
827 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
828 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
832 * Report a full set of quiescent states to the specified rcu_state
833 * data structure. This involves cleaning up after the prior grace
834 * period and letting rcu_start_gp() start up the next grace period
835 * if one is needed. Note that the caller must hold rnp->lock, as
836 * required by rcu_start_gp(), which will release it.
838 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
839 __releases(rcu_get_root(rsp)->lock)
841 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
842 rsp->completed = rsp->gpnum;
843 rsp->signaled = RCU_GP_IDLE;
844 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
848 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
849 * Allows quiescent states for a group of CPUs to be reported at one go
850 * to the specified rcu_node structure, though all the CPUs in the group
851 * must be represented by the same rcu_node structure (which need not be
852 * a leaf rcu_node structure, though it often will be). That structure's
853 * lock must be held upon entry, and it is released before return.
856 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
857 struct rcu_node *rnp, unsigned long flags)
858 __releases(rnp->lock)
860 struct rcu_node *rnp_c;
862 /* Walk up the rcu_node hierarchy. */
864 if (!(rnp->qsmask & mask)) {
866 /* Our bit has already been cleared, so done. */
867 raw_spin_unlock_irqrestore(&rnp->lock, flags);
870 rnp->qsmask &= ~mask;
871 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
873 /* Other bits still set at this level, so done. */
874 raw_spin_unlock_irqrestore(&rnp->lock, flags);
878 if (rnp->parent == NULL) {
880 /* No more levels. Exit loop holding root lock. */
884 raw_spin_unlock_irqrestore(&rnp->lock, flags);
887 raw_spin_lock_irqsave(&rnp->lock, flags);
888 WARN_ON_ONCE(rnp_c->qsmask);
892 * Get here if we are the last CPU to pass through a quiescent
893 * state for this grace period. Invoke rcu_report_qs_rsp()
894 * to clean up and start the next grace period if one is needed.
896 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
900 * Record a quiescent state for the specified CPU to that CPU's rcu_data
901 * structure. This must be either called from the specified CPU, or
902 * called when the specified CPU is known to be offline (and when it is
903 * also known that no other CPU is concurrently trying to help the offline
904 * CPU). The lastcomp argument is used to make sure we are still in the
905 * grace period of interest. We don't want to end the current grace period
906 * based on quiescent states detected in an earlier grace period!
909 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
913 struct rcu_node *rnp;
916 raw_spin_lock_irqsave(&rnp->lock, flags);
917 if (lastcomp != rnp->completed) {
920 * Someone beat us to it for this grace period, so leave.
921 * The race with GP start is resolved by the fact that we
922 * hold the leaf rcu_node lock, so that the per-CPU bits
923 * cannot yet be initialized -- so we would simply find our
924 * CPU's bit already cleared in rcu_report_qs_rnp() if this
927 rdp->passed_quiesc = 0; /* try again later! */
928 raw_spin_unlock_irqrestore(&rnp->lock, flags);
932 if ((rnp->qsmask & mask) == 0) {
933 raw_spin_unlock_irqrestore(&rnp->lock, flags);
938 * This GP can't end until cpu checks in, so all of our
939 * callbacks can be processed during the next GP.
941 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
943 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
948 * Check to see if there is a new grace period of which this CPU
949 * is not yet aware, and if so, set up local rcu_data state for it.
950 * Otherwise, see if this CPU has just passed through its first
951 * quiescent state for this grace period, and record that fact if so.
954 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
956 /* If there is now a new grace period, record and return. */
957 if (check_for_new_grace_period(rsp, rdp))
961 * Does this CPU still need to do its part for current grace period?
962 * If no, return and let the other CPUs do their part as well.
964 if (!rdp->qs_pending)
968 * Was there a quiescent state since the beginning of the grace
969 * period? If no, then exit and wait for the next call.
971 if (!rdp->passed_quiesc)
975 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
978 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
981 #ifdef CONFIG_HOTPLUG_CPU
984 * Move a dying CPU's RCU callbacks to online CPU's callback list.
985 * Synchronization is not required because this function executes
986 * in stop_machine() context.
988 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
991 /* current DYING CPU is cleared in the cpu_online_mask */
992 int receive_cpu = cpumask_any(cpu_online_mask);
993 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
994 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
996 if (rdp->nxtlist == NULL)
997 return; /* irqs disabled, so comparison is stable. */
999 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1000 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1001 receive_rdp->qlen += rdp->qlen;
1002 receive_rdp->n_cbs_adopted += rdp->qlen;
1003 rdp->n_cbs_orphaned += rdp->qlen;
1005 rdp->nxtlist = NULL;
1006 for (i = 0; i < RCU_NEXT_SIZE; i++)
1007 rdp->nxttail[i] = &rdp->nxtlist;
1012 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1013 * and move all callbacks from the outgoing CPU to the current one.
1015 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1017 unsigned long flags;
1019 int need_report = 0;
1020 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1021 struct rcu_node *rnp;
1023 /* Exclude any attempts to start a new grace period. */
1024 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1026 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1027 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1028 mask = rdp->grpmask; /* rnp->grplo is constant. */
1030 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1031 rnp->qsmaskinit &= ~mask;
1032 if (rnp->qsmaskinit != 0) {
1033 if (rnp != rdp->mynode)
1034 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1037 if (rnp == rdp->mynode)
1038 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1040 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1041 mask = rnp->grpmask;
1043 } while (rnp != NULL);
1046 * We still hold the leaf rcu_node structure lock here, and
1047 * irqs are still disabled. The reason for this subterfuge is
1048 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1049 * held leads to deadlock.
1051 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1053 if (need_report & RCU_OFL_TASKS_NORM_GP)
1054 rcu_report_unblock_qs_rnp(rnp, flags);
1056 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1057 if (need_report & RCU_OFL_TASKS_EXP_GP)
1058 rcu_report_exp_rnp(rsp, rnp);
1062 * Remove the specified CPU from the RCU hierarchy and move any pending
1063 * callbacks that it might have to the current CPU. This code assumes
1064 * that at least one CPU in the system will remain running at all times.
1065 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1067 static void rcu_offline_cpu(int cpu)
1069 __rcu_offline_cpu(cpu, &rcu_sched_state);
1070 __rcu_offline_cpu(cpu, &rcu_bh_state);
1071 rcu_preempt_offline_cpu(cpu);
1074 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1076 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1080 static void rcu_offline_cpu(int cpu)
1084 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1087 * Invoke any RCU callbacks that have made it to the end of their grace
1088 * period. Thottle as specified by rdp->blimit.
1090 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1092 unsigned long flags;
1093 struct rcu_head *next, *list, **tail;
1096 /* If no callbacks are ready, just return.*/
1097 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1101 * Extract the list of ready callbacks, disabling to prevent
1102 * races with call_rcu() from interrupt handlers.
1104 local_irq_save(flags);
1105 list = rdp->nxtlist;
1106 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1107 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1108 tail = rdp->nxttail[RCU_DONE_TAIL];
1109 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1110 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1111 rdp->nxttail[count] = &rdp->nxtlist;
1112 local_irq_restore(flags);
1114 /* Invoke callbacks. */
1119 debug_rcu_head_unqueue(list);
1122 if (++count >= rdp->blimit)
1126 local_irq_save(flags);
1128 /* Update count, and requeue any remaining callbacks. */
1130 rdp->n_cbs_invoked += count;
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 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1675 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1676 complete(&rcu_barrier_completion);
1677 wait_for_completion(&rcu_barrier_completion);
1678 mutex_unlock(&rcu_barrier_mutex);
1682 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1684 void rcu_barrier_bh(void)
1686 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1688 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1691 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1693 void rcu_barrier_sched(void)
1695 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1697 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1700 * Do boot-time initialization of a CPU's per-CPU RCU data.
1703 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1705 unsigned long flags;
1707 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1708 struct rcu_node *rnp = rcu_get_root(rsp);
1710 /* Set up local state, ensuring consistent view of global state. */
1711 raw_spin_lock_irqsave(&rnp->lock, flags);
1712 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1713 rdp->nxtlist = NULL;
1714 for (i = 0; i < RCU_NEXT_SIZE; i++)
1715 rdp->nxttail[i] = &rdp->nxtlist;
1718 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1719 #endif /* #ifdef CONFIG_NO_HZ */
1721 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1725 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1726 * offline event can be happening at a given time. Note also that we
1727 * can accept some slop in the rsp->completed access due to the fact
1728 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1730 static void __cpuinit
1731 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1733 unsigned long flags;
1735 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1736 struct rcu_node *rnp = rcu_get_root(rsp);
1738 /* Set up local state, ensuring consistent view of global state. */
1739 raw_spin_lock_irqsave(&rnp->lock, flags);
1740 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1741 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1742 rdp->beenonline = 1; /* We have now been online. */
1743 rdp->preemptable = preemptable;
1744 rdp->qlen_last_fqs_check = 0;
1745 rdp->n_force_qs_snap = rsp->n_force_qs;
1746 rdp->blimit = blimit;
1747 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1750 * A new grace period might start here. If so, we won't be part
1751 * of it, but that is OK, as we are currently in a quiescent state.
1754 /* Exclude any attempts to start a new GP on large systems. */
1755 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1757 /* Add CPU to rcu_node bitmasks. */
1759 mask = rdp->grpmask;
1761 /* Exclude any attempts to start a new GP on small systems. */
1762 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1763 rnp->qsmaskinit |= mask;
1764 mask = rnp->grpmask;
1765 if (rnp == rdp->mynode) {
1766 rdp->gpnum = rnp->completed; /* if GP in progress... */
1767 rdp->completed = rnp->completed;
1768 rdp->passed_quiesc_completed = rnp->completed - 1;
1770 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1772 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1774 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1777 static void __cpuinit rcu_online_cpu(int cpu)
1779 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1780 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1781 rcu_preempt_init_percpu_data(cpu);
1785 * Handle CPU online/offline notification events.
1787 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1788 unsigned long action, void *hcpu)
1790 long cpu = (long)hcpu;
1793 case CPU_UP_PREPARE:
1794 case CPU_UP_PREPARE_FROZEN:
1795 rcu_online_cpu(cpu);
1798 case CPU_DYING_FROZEN:
1800 * The whole machine is "stopped" except this cpu, so we can
1801 * touch any data without introducing corruption. And we send
1802 * the callbacks to an attribute chosen online cpu.
1804 rcu_send_cbs_to_online(&rcu_bh_state);
1805 rcu_send_cbs_to_online(&rcu_sched_state);
1806 rcu_preempt_send_cbs_to_online();
1809 case CPU_DEAD_FROZEN:
1810 case CPU_UP_CANCELED:
1811 case CPU_UP_CANCELED_FROZEN:
1812 rcu_offline_cpu(cpu);
1821 * This function is invoked towards the end of the scheduler's initialization
1822 * process. Before this is called, the idle task might contain
1823 * RCU read-side critical sections (during which time, this idle
1824 * task is booting the system). After this function is called, the
1825 * idle tasks are prohibited from containing RCU read-side critical
1826 * sections. This function also enables RCU lockdep checking.
1828 void rcu_scheduler_starting(void)
1830 WARN_ON(num_online_cpus() != 1);
1831 WARN_ON(nr_context_switches() > 0);
1832 rcu_scheduler_active = 1;
1836 * Compute the per-level fanout, either using the exact fanout specified
1837 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1839 #ifdef CONFIG_RCU_FANOUT_EXACT
1840 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1844 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1845 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1847 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1848 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1855 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1856 ccur = rsp->levelcnt[i];
1857 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1861 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1864 * Helper function for rcu_init() that initializes one rcu_state structure.
1866 static void __init rcu_init_one(struct rcu_state *rsp,
1867 struct rcu_data __percpu *rda)
1869 static char *buf[] = { "rcu_node_level_0",
1872 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1876 struct rcu_node *rnp;
1878 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1880 /* Initialize the level-tracking arrays. */
1882 for (i = 1; i < NUM_RCU_LVLS; i++)
1883 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1884 rcu_init_levelspread(rsp);
1886 /* Initialize the elements themselves, starting from the leaves. */
1888 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1889 cpustride *= rsp->levelspread[i];
1890 rnp = rsp->level[i];
1891 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1892 raw_spin_lock_init(&rnp->lock);
1893 lockdep_set_class_and_name(&rnp->lock,
1894 &rcu_node_class[i], buf[i]);
1897 rnp->qsmaskinit = 0;
1898 rnp->grplo = j * cpustride;
1899 rnp->grphi = (j + 1) * cpustride - 1;
1900 if (rnp->grphi >= NR_CPUS)
1901 rnp->grphi = NR_CPUS - 1;
1907 rnp->grpnum = j % rsp->levelspread[i - 1];
1908 rnp->grpmask = 1UL << rnp->grpnum;
1909 rnp->parent = rsp->level[i - 1] +
1910 j / rsp->levelspread[i - 1];
1913 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1914 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1915 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1916 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1921 rnp = rsp->level[NUM_RCU_LVLS - 1];
1922 for_each_possible_cpu(i) {
1923 while (i > rnp->grphi)
1925 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
1926 rcu_boot_init_percpu_data(i, rsp);
1930 void __init rcu_init(void)
1934 rcu_bootup_announce();
1935 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
1936 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
1937 __rcu_init_preempt();
1938 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1941 * We don't need protection against CPU-hotplug here because
1942 * this is called early in boot, before either interrupts
1943 * or the scheduler are operational.
1945 cpu_notifier(rcu_cpu_notify, 0);
1946 for_each_online_cpu(cpu)
1947 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1948 check_cpu_stall_init();
1951 #include "rcutree_plugin.h"