2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
53 /* Data structures. */
55 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
57 #define RCU_STATE_INITIALIZER(name) { \
58 .level = { &name.node[0] }, \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
64 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
66 .signaled = RCU_GP_IDLE, \
69 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
70 .orphan_cbs_list = NULL, \
71 .orphan_cbs_tail = &name.orphan_cbs_list, \
73 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
75 .n_force_qs_ngp = 0, \
78 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
79 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
81 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
82 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84 static int rcu_scheduler_active __read_mostly;
88 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
89 * permit this function to be invoked without holding the root rcu_node
90 * structure's ->lock, but of course results can be subject to change.
92 static int rcu_gp_in_progress(struct rcu_state *rsp)
94 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
98 * Note a quiescent state. Because we do not need to know
99 * how many quiescent states passed, just if there was at least
100 * one since the start of the grace period, this just sets a flag.
102 void rcu_sched_qs(int cpu)
104 struct rcu_data *rdp;
106 rdp = &per_cpu(rcu_sched_data, cpu);
107 rdp->passed_quiesc_completed = rdp->gpnum - 1;
109 rdp->passed_quiesc = 1;
110 rcu_preempt_note_context_switch(cpu);
113 void rcu_bh_qs(int cpu)
115 struct rcu_data *rdp;
117 rdp = &per_cpu(rcu_bh_data, cpu);
118 rdp->passed_quiesc_completed = rdp->gpnum - 1;
120 rdp->passed_quiesc = 1;
124 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
125 .dynticks_nesting = 1,
128 #endif /* #ifdef CONFIG_NO_HZ */
130 static int blimit = 10; /* Maximum callbacks per softirq. */
131 static int qhimark = 10000; /* If this many pending, ignore blimit. */
132 static int qlowmark = 100; /* Once only this many pending, use blimit. */
134 module_param(blimit, int, 0);
135 module_param(qhimark, int, 0);
136 module_param(qlowmark, int, 0);
138 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
139 static int rcu_pending(int cpu);
142 * Return the number of RCU-sched batches processed thus far for debug & stats.
144 long rcu_batches_completed_sched(void)
146 return rcu_sched_state.completed;
148 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
151 * Return the number of RCU BH batches processed thus far for debug & stats.
153 long rcu_batches_completed_bh(void)
155 return rcu_bh_state.completed;
157 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
160 * Does the CPU have callbacks ready to be invoked?
163 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
165 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
169 * Does the current CPU require a yet-as-unscheduled grace period?
172 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
174 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
178 * Return the root node of the specified rcu_state structure.
180 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
182 return &rsp->node[0];
188 * If the specified CPU is offline, tell the caller that it is in
189 * a quiescent state. Otherwise, whack it with a reschedule IPI.
190 * Grace periods can end up waiting on an offline CPU when that
191 * CPU is in the process of coming online -- it will be added to the
192 * rcu_node bitmasks before it actually makes it online. The same thing
193 * can happen while a CPU is in the process of coming online. Because this
194 * race is quite rare, we check for it after detecting that the grace
195 * period has been delayed rather than checking each and every CPU
196 * each and every time we start a new grace period.
198 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
201 * If the CPU is offline, it is in a quiescent state. We can
202 * trust its state not to change because interrupts are disabled.
204 if (cpu_is_offline(rdp->cpu)) {
209 /* If preemptable RCU, no point in sending reschedule IPI. */
210 if (rdp->preemptable)
213 /* The CPU is online, so send it a reschedule IPI. */
214 if (rdp->cpu != smp_processor_id())
215 smp_send_reschedule(rdp->cpu);
222 #endif /* #ifdef CONFIG_SMP */
227 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
229 * Enter nohz mode, in other words, -leave- the mode in which RCU
230 * read-side critical sections can occur. (Though RCU read-side
231 * critical sections can occur in irq handlers in nohz mode, a possibility
232 * handled by rcu_irq_enter() and rcu_irq_exit()).
234 void rcu_enter_nohz(void)
237 struct rcu_dynticks *rdtp;
239 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
240 local_irq_save(flags);
241 rdtp = &__get_cpu_var(rcu_dynticks);
243 rdtp->dynticks_nesting--;
244 WARN_ON_ONCE(rdtp->dynticks & 0x1);
245 local_irq_restore(flags);
249 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
251 * Exit nohz mode, in other words, -enter- the mode in which RCU
252 * read-side critical sections normally occur.
254 void rcu_exit_nohz(void)
257 struct rcu_dynticks *rdtp;
259 local_irq_save(flags);
260 rdtp = &__get_cpu_var(rcu_dynticks);
262 rdtp->dynticks_nesting++;
263 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
264 local_irq_restore(flags);
265 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
269 * rcu_nmi_enter - inform RCU of entry to NMI context
271 * If the CPU was idle with dynamic ticks active, and there is no
272 * irq handler running, this updates rdtp->dynticks_nmi to let the
273 * RCU grace-period handling know that the CPU is active.
275 void rcu_nmi_enter(void)
277 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
279 if (rdtp->dynticks & 0x1)
281 rdtp->dynticks_nmi++;
282 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
283 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
287 * rcu_nmi_exit - inform RCU of exit from NMI context
289 * If the CPU was idle with dynamic ticks active, and there is no
290 * irq handler running, this updates rdtp->dynticks_nmi to let the
291 * RCU grace-period handling know that the CPU is no longer active.
293 void rcu_nmi_exit(void)
295 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
297 if (rdtp->dynticks & 0x1)
299 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
300 rdtp->dynticks_nmi++;
301 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
305 * rcu_irq_enter - inform RCU of entry to hard irq context
307 * If the CPU was idle with dynamic ticks active, this updates the
308 * rdtp->dynticks to let the RCU handling know that the CPU is active.
310 void rcu_irq_enter(void)
312 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
314 if (rdtp->dynticks_nesting++)
317 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
318 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
322 * rcu_irq_exit - inform RCU of exit from hard irq context
324 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
325 * to put let the RCU handling be aware that the CPU is going back to idle
328 void rcu_irq_exit(void)
330 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
332 if (--rdtp->dynticks_nesting)
334 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
336 WARN_ON_ONCE(rdtp->dynticks & 0x1);
338 /* If the interrupt queued a callback, get out of dyntick mode. */
339 if (__get_cpu_var(rcu_sched_data).nxtlist ||
340 __get_cpu_var(rcu_bh_data).nxtlist)
347 * Snapshot the specified CPU's dynticks counter so that we can later
348 * credit them with an implicit quiescent state. Return 1 if this CPU
349 * is in dynticks idle mode, which is an extended quiescent state.
351 static int dyntick_save_progress_counter(struct rcu_data *rdp)
357 snap = rdp->dynticks->dynticks;
358 snap_nmi = rdp->dynticks->dynticks_nmi;
359 smp_mb(); /* Order sampling of snap with end of grace period. */
360 rdp->dynticks_snap = snap;
361 rdp->dynticks_nmi_snap = snap_nmi;
362 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
369 * Return true if the specified CPU has passed through a quiescent
370 * state by virtue of being in or having passed through an dynticks
371 * idle state since the last call to dyntick_save_progress_counter()
374 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
381 curr = rdp->dynticks->dynticks;
382 snap = rdp->dynticks_snap;
383 curr_nmi = rdp->dynticks->dynticks_nmi;
384 snap_nmi = rdp->dynticks_nmi_snap;
385 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
388 * If the CPU passed through or entered a dynticks idle phase with
389 * no active irq/NMI handlers, then we can safely pretend that the CPU
390 * already acknowledged the request to pass through a quiescent
391 * state. Either way, that CPU cannot possibly be in an RCU
392 * read-side critical section that started before the beginning
393 * of the current RCU grace period.
395 if ((curr != snap || (curr & 0x1) == 0) &&
396 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
401 /* Go check for the CPU being offline. */
402 return rcu_implicit_offline_qs(rdp);
405 #endif /* #ifdef CONFIG_SMP */
407 #else /* #ifdef CONFIG_NO_HZ */
411 static int dyntick_save_progress_counter(struct rcu_data *rdp)
416 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
418 return rcu_implicit_offline_qs(rdp);
421 #endif /* #ifdef CONFIG_SMP */
423 #endif /* #else #ifdef CONFIG_NO_HZ */
425 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
427 static void record_gp_stall_check_time(struct rcu_state *rsp)
429 rsp->gp_start = jiffies;
430 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
433 static void print_other_cpu_stall(struct rcu_state *rsp)
438 struct rcu_node *rnp = rcu_get_root(rsp);
440 /* Only let one CPU complain about others per time interval. */
442 spin_lock_irqsave(&rnp->lock, flags);
443 delta = jiffies - rsp->jiffies_stall;
444 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
445 spin_unlock_irqrestore(&rnp->lock, flags);
448 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
451 * Now rat on any tasks that got kicked up to the root rcu_node
452 * due to CPU offlining.
454 rcu_print_task_stall(rnp);
455 spin_unlock_irqrestore(&rnp->lock, flags);
457 /* OK, time to rat on our buddy... */
459 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
460 rcu_for_each_leaf_node(rsp, rnp) {
461 rcu_print_task_stall(rnp);
462 if (rnp->qsmask == 0)
464 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
465 if (rnp->qsmask & (1UL << cpu))
466 printk(" %d", rnp->grplo + cpu);
468 printk(" (detected by %d, t=%ld jiffies)\n",
469 smp_processor_id(), (long)(jiffies - rsp->gp_start));
470 trigger_all_cpu_backtrace();
472 force_quiescent_state(rsp, 0); /* Kick them all. */
475 static void print_cpu_stall(struct rcu_state *rsp)
478 struct rcu_node *rnp = rcu_get_root(rsp);
480 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
481 smp_processor_id(), jiffies - rsp->gp_start);
482 trigger_all_cpu_backtrace();
484 spin_lock_irqsave(&rnp->lock, flags);
485 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
487 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
488 spin_unlock_irqrestore(&rnp->lock, flags);
490 set_need_resched(); /* kick ourselves to get things going. */
493 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
496 struct rcu_node *rnp;
498 delta = jiffies - rsp->jiffies_stall;
500 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
502 /* We haven't checked in, so go dump stack. */
503 print_cpu_stall(rsp);
505 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
507 /* They had two time units to dump stack, so complain. */
508 print_other_cpu_stall(rsp);
512 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
514 static void record_gp_stall_check_time(struct rcu_state *rsp)
518 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
522 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
525 * Update CPU-local rcu_data state to record the newly noticed grace period.
526 * This is used both when we started the grace period and when we notice
527 * that someone else started the grace period. The caller must hold the
528 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
529 * and must have irqs disabled.
531 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
533 if (rdp->gpnum != rnp->gpnum) {
535 rdp->passed_quiesc = 0;
536 rdp->gpnum = rnp->gpnum;
540 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
543 struct rcu_node *rnp;
545 local_irq_save(flags);
547 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
548 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
549 local_irq_restore(flags);
552 __note_new_gpnum(rsp, rnp, rdp);
553 spin_unlock_irqrestore(&rnp->lock, flags);
557 * Did someone else start a new RCU grace period start since we last
558 * checked? Update local state appropriately if so. Must be called
559 * on the CPU corresponding to rdp.
562 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
567 local_irq_save(flags);
568 if (rdp->gpnum != rsp->gpnum) {
569 note_new_gpnum(rsp, rdp);
572 local_irq_restore(flags);
577 * Advance this CPU's callbacks, but only if the current grace period
578 * has ended. This may be called only from the CPU to whom the rdp
579 * belongs. In addition, the corresponding leaf rcu_node structure's
580 * ->lock must be held by the caller, with irqs disabled.
583 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
585 /* Did another grace period end? */
586 if (rdp->completed != rnp->completed) {
588 /* Advance callbacks. No harm if list empty. */
589 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
590 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
591 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
593 /* Remember that we saw this grace-period completion. */
594 rdp->completed = rnp->completed;
599 * Advance this CPU's callbacks, but only if the current grace period
600 * has ended. This may be called only from the CPU to whom the rdp
604 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
607 struct rcu_node *rnp;
609 local_irq_save(flags);
611 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
612 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
613 local_irq_restore(flags);
616 __rcu_process_gp_end(rsp, rnp, rdp);
617 spin_unlock_irqrestore(&rnp->lock, flags);
621 * Do per-CPU grace-period initialization for running CPU. The caller
622 * must hold the lock of the leaf rcu_node structure corresponding to
626 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
628 /* Prior grace period ended, so advance callbacks for current CPU. */
629 __rcu_process_gp_end(rsp, rnp, rdp);
632 * Because this CPU just now started the new grace period, we know
633 * that all of its callbacks will be covered by this upcoming grace
634 * period, even the ones that were registered arbitrarily recently.
635 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
637 * Other CPUs cannot be sure exactly when the grace period started.
638 * Therefore, their recently registered callbacks must pass through
639 * an additional RCU_NEXT_READY stage, so that they will be handled
640 * by the next RCU grace period.
642 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
643 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
645 /* Set state so that this CPU will detect the next quiescent state. */
646 __note_new_gpnum(rsp, rnp, rdp);
650 * Start a new RCU grace period if warranted, re-initializing the hierarchy
651 * in preparation for detecting the next grace period. The caller must hold
652 * the root node's ->lock, which is released before return. Hard irqs must
656 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
657 __releases(rcu_get_root(rsp)->lock)
659 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
660 struct rcu_node *rnp = rcu_get_root(rsp);
662 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
663 if (rnp->completed == rsp->completed) {
664 spin_unlock_irqrestore(&rnp->lock, flags);
667 spin_unlock(&rnp->lock); /* irqs remain disabled. */
670 * Propagate new ->completed value to rcu_node structures
671 * so that other CPUs don't have to wait until the start
672 * of the next grace period to process their callbacks.
674 rcu_for_each_node_breadth_first(rsp, rnp) {
675 spin_lock(&rnp->lock); /* irqs already disabled. */
676 rnp->completed = rsp->completed;
677 spin_unlock(&rnp->lock); /* irqs remain disabled. */
679 local_irq_restore(flags);
683 /* Advance to a new grace period and initialize state. */
685 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
686 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
687 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
688 record_gp_stall_check_time(rsp);
690 /* Special-case the common single-level case. */
691 if (NUM_RCU_NODES == 1) {
692 rcu_preempt_check_blocked_tasks(rnp);
693 rnp->qsmask = rnp->qsmaskinit;
694 rnp->gpnum = rsp->gpnum;
695 rnp->completed = rsp->completed;
696 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
697 rcu_start_gp_per_cpu(rsp, rnp, rdp);
698 spin_unlock_irqrestore(&rnp->lock, flags);
702 spin_unlock(&rnp->lock); /* leave irqs disabled. */
705 /* Exclude any concurrent CPU-hotplug operations. */
706 spin_lock(&rsp->onofflock); /* irqs already disabled. */
709 * Set the quiescent-state-needed bits in all the rcu_node
710 * structures for all currently online CPUs in breadth-first
711 * order, starting from the root rcu_node structure. This
712 * operation relies on the layout of the hierarchy within the
713 * rsp->node[] array. Note that other CPUs will access only
714 * the leaves of the hierarchy, which still indicate that no
715 * grace period is in progress, at least until the corresponding
716 * leaf node has been initialized. In addition, we have excluded
717 * CPU-hotplug operations.
719 * Note that the grace period cannot complete until we finish
720 * the initialization process, as there will be at least one
721 * qsmask bit set in the root node until that time, namely the
722 * one corresponding to this CPU, due to the fact that we have
725 rcu_for_each_node_breadth_first(rsp, rnp) {
726 spin_lock(&rnp->lock); /* irqs already disabled. */
727 rcu_preempt_check_blocked_tasks(rnp);
728 rnp->qsmask = rnp->qsmaskinit;
729 rnp->gpnum = rsp->gpnum;
730 rnp->completed = rsp->completed;
731 if (rnp == rdp->mynode)
732 rcu_start_gp_per_cpu(rsp, rnp, rdp);
733 spin_unlock(&rnp->lock); /* irqs remain disabled. */
736 rnp = rcu_get_root(rsp);
737 spin_lock(&rnp->lock); /* irqs already disabled. */
738 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
739 spin_unlock(&rnp->lock); /* irqs remain disabled. */
740 spin_unlock_irqrestore(&rsp->onofflock, flags);
744 * Report a full set of quiescent states to the specified rcu_state
745 * data structure. This involves cleaning up after the prior grace
746 * period and letting rcu_start_gp() start up the next grace period
747 * if one is needed. Note that the caller must hold rnp->lock, as
748 * required by rcu_start_gp(), which will release it.
750 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
751 __releases(rcu_get_root(rsp)->lock)
753 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
754 rsp->completed = rsp->gpnum;
755 rsp->signaled = RCU_GP_IDLE;
756 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
760 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
761 * Allows quiescent states for a group of CPUs to be reported at one go
762 * to the specified rcu_node structure, though all the CPUs in the group
763 * must be represented by the same rcu_node structure (which need not be
764 * a leaf rcu_node structure, though it often will be). That structure's
765 * lock must be held upon entry, and it is released before return.
768 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
769 struct rcu_node *rnp, unsigned long flags)
770 __releases(rnp->lock)
772 struct rcu_node *rnp_c;
774 /* Walk up the rcu_node hierarchy. */
776 if (!(rnp->qsmask & mask)) {
778 /* Our bit has already been cleared, so done. */
779 spin_unlock_irqrestore(&rnp->lock, flags);
782 rnp->qsmask &= ~mask;
783 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
785 /* Other bits still set at this level, so done. */
786 spin_unlock_irqrestore(&rnp->lock, flags);
790 if (rnp->parent == NULL) {
792 /* No more levels. Exit loop holding root lock. */
796 spin_unlock_irqrestore(&rnp->lock, flags);
799 spin_lock_irqsave(&rnp->lock, flags);
800 WARN_ON_ONCE(rnp_c->qsmask);
804 * Get here if we are the last CPU to pass through a quiescent
805 * state for this grace period. Invoke rcu_report_qs_rsp()
806 * to clean up and start the next grace period if one is needed.
808 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
812 * Record a quiescent state for the specified CPU to that CPU's rcu_data
813 * structure. This must be either called from the specified CPU, or
814 * called when the specified CPU is known to be offline (and when it is
815 * also known that no other CPU is concurrently trying to help the offline
816 * CPU). The lastcomp argument is used to make sure we are still in the
817 * grace period of interest. We don't want to end the current grace period
818 * based on quiescent states detected in an earlier grace period!
821 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
825 struct rcu_node *rnp;
828 spin_lock_irqsave(&rnp->lock, flags);
829 if (lastcomp != rnp->completed) {
832 * Someone beat us to it for this grace period, so leave.
833 * The race with GP start is resolved by the fact that we
834 * hold the leaf rcu_node lock, so that the per-CPU bits
835 * cannot yet be initialized -- so we would simply find our
836 * CPU's bit already cleared in rcu_report_qs_rnp() if this
839 rdp->passed_quiesc = 0; /* try again later! */
840 spin_unlock_irqrestore(&rnp->lock, flags);
844 if ((rnp->qsmask & mask) == 0) {
845 spin_unlock_irqrestore(&rnp->lock, flags);
850 * This GP can't end until cpu checks in, so all of our
851 * callbacks can be processed during the next GP.
853 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
855 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
860 * Check to see if there is a new grace period of which this CPU
861 * is not yet aware, and if so, set up local rcu_data state for it.
862 * Otherwise, see if this CPU has just passed through its first
863 * quiescent state for this grace period, and record that fact if so.
866 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
868 /* If there is now a new grace period, record and return. */
869 if (check_for_new_grace_period(rsp, rdp))
873 * Does this CPU still need to do its part for current grace period?
874 * If no, return and let the other CPUs do their part as well.
876 if (!rdp->qs_pending)
880 * Was there a quiescent state since the beginning of the grace
881 * period? If no, then exit and wait for the next call.
883 if (!rdp->passed_quiesc)
887 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
890 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
893 #ifdef CONFIG_HOTPLUG_CPU
896 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
897 * specified flavor of RCU. The callbacks will be adopted by the next
898 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
899 * comes first. Because this is invoked from the CPU_DYING notifier,
900 * irqs are already disabled.
902 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
905 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
907 if (rdp->nxtlist == NULL)
908 return; /* irqs disabled, so comparison is stable. */
909 spin_lock(&rsp->onofflock); /* irqs already disabled. */
910 *rsp->orphan_cbs_tail = rdp->nxtlist;
911 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
913 for (i = 0; i < RCU_NEXT_SIZE; i++)
914 rdp->nxttail[i] = &rdp->nxtlist;
915 rsp->orphan_qlen += rdp->qlen;
917 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
921 * Adopt previously orphaned RCU callbacks.
923 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
926 struct rcu_data *rdp;
928 spin_lock_irqsave(&rsp->onofflock, flags);
929 rdp = rsp->rda[smp_processor_id()];
930 if (rsp->orphan_cbs_list == NULL) {
931 spin_unlock_irqrestore(&rsp->onofflock, flags);
934 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
935 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
936 rdp->qlen += rsp->orphan_qlen;
937 rsp->orphan_cbs_list = NULL;
938 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
939 rsp->orphan_qlen = 0;
940 spin_unlock_irqrestore(&rsp->onofflock, flags);
944 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
945 * and move all callbacks from the outgoing CPU to the current one.
947 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
952 struct rcu_data *rdp = rsp->rda[cpu];
953 struct rcu_node *rnp;
955 /* Exclude any attempts to start a new grace period. */
956 spin_lock_irqsave(&rsp->onofflock, flags);
958 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
959 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
960 mask = rdp->grpmask; /* rnp->grplo is constant. */
962 spin_lock(&rnp->lock); /* irqs already disabled. */
963 rnp->qsmaskinit &= ~mask;
964 if (rnp->qsmaskinit != 0) {
965 if (rnp != rdp->mynode)
966 spin_unlock(&rnp->lock); /* irqs remain disabled. */
969 if (rnp == rdp->mynode)
970 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
972 spin_unlock(&rnp->lock); /* irqs remain disabled. */
975 } while (rnp != NULL);
978 * We still hold the leaf rcu_node structure lock here, and
979 * irqs are still disabled. The reason for this subterfuge is
980 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
981 * held leads to deadlock.
983 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
985 if (need_report & RCU_OFL_TASKS_NORM_GP)
986 rcu_report_unblock_qs_rnp(rnp, flags);
988 spin_unlock_irqrestore(&rnp->lock, flags);
989 if (need_report & RCU_OFL_TASKS_EXP_GP)
990 rcu_report_exp_rnp(rsp, rnp);
992 rcu_adopt_orphan_cbs(rsp);
996 * Remove the specified CPU from the RCU hierarchy and move any pending
997 * callbacks that it might have to the current CPU. This code assumes
998 * that at least one CPU in the system will remain running at all times.
999 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1001 static void rcu_offline_cpu(int cpu)
1003 __rcu_offline_cpu(cpu, &rcu_sched_state);
1004 __rcu_offline_cpu(cpu, &rcu_bh_state);
1005 rcu_preempt_offline_cpu(cpu);
1008 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1010 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1014 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1018 static void rcu_offline_cpu(int cpu)
1022 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1025 * Invoke any RCU callbacks that have made it to the end of their grace
1026 * period. Thottle as specified by rdp->blimit.
1028 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1030 unsigned long flags;
1031 struct rcu_head *next, *list, **tail;
1034 /* If no callbacks are ready, just return.*/
1035 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1039 * Extract the list of ready callbacks, disabling to prevent
1040 * races with call_rcu() from interrupt handlers.
1042 local_irq_save(flags);
1043 list = rdp->nxtlist;
1044 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1045 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1046 tail = rdp->nxttail[RCU_DONE_TAIL];
1047 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1048 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1049 rdp->nxttail[count] = &rdp->nxtlist;
1050 local_irq_restore(flags);
1052 /* Invoke callbacks. */
1059 if (++count >= rdp->blimit)
1063 local_irq_save(flags);
1065 /* Update count, and requeue any remaining callbacks. */
1068 *tail = rdp->nxtlist;
1069 rdp->nxtlist = list;
1070 for (count = 0; count < RCU_NEXT_SIZE; count++)
1071 if (&rdp->nxtlist == rdp->nxttail[count])
1072 rdp->nxttail[count] = tail;
1077 /* Reinstate batch limit if we have worked down the excess. */
1078 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1079 rdp->blimit = blimit;
1081 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1082 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1083 rdp->qlen_last_fqs_check = 0;
1084 rdp->n_force_qs_snap = rsp->n_force_qs;
1085 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1086 rdp->qlen_last_fqs_check = rdp->qlen;
1088 local_irq_restore(flags);
1090 /* Re-raise the RCU softirq if there are callbacks remaining. */
1091 if (cpu_has_callbacks_ready_to_invoke(rdp))
1092 raise_softirq(RCU_SOFTIRQ);
1096 * Check to see if this CPU is in a non-context-switch quiescent state
1097 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1098 * Also schedule the RCU softirq handler.
1100 * This function must be called with hardirqs disabled. It is normally
1101 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1102 * false, there is no point in invoking rcu_check_callbacks().
1104 void rcu_check_callbacks(int cpu, int user)
1106 if (!rcu_pending(cpu))
1107 return; /* if nothing for RCU to do. */
1109 (idle_cpu(cpu) && rcu_scheduler_active &&
1110 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1113 * Get here if this CPU took its interrupt from user
1114 * mode or from the idle loop, and if this is not a
1115 * nested interrupt. In this case, the CPU is in
1116 * a quiescent state, so note it.
1118 * No memory barrier is required here because both
1119 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1120 * variables that other CPUs neither access nor modify,
1121 * at least not while the corresponding CPU is online.
1127 } else if (!in_softirq()) {
1130 * Get here if this CPU did not take its interrupt from
1131 * softirq, in other words, if it is not interrupting
1132 * a rcu_bh read-side critical section. This is an _bh
1133 * critical section, so note it.
1138 rcu_preempt_check_callbacks(cpu);
1139 raise_softirq(RCU_SOFTIRQ);
1145 * Scan the leaf rcu_node structures, processing dyntick state for any that
1146 * have not yet encountered a quiescent state, using the function specified.
1147 * The caller must have suppressed start of new grace periods.
1149 static void rcu_process_dyntick(struct rcu_state *rsp,
1150 int (*f)(struct rcu_data *))
1154 unsigned long flags;
1156 struct rcu_node *rnp;
1158 rcu_for_each_leaf_node(rsp, rnp) {
1160 spin_lock_irqsave(&rnp->lock, flags);
1161 if (!rcu_gp_in_progress(rsp)) {
1162 spin_unlock_irqrestore(&rnp->lock, flags);
1165 if (rnp->qsmask == 0) {
1166 spin_unlock_irqrestore(&rnp->lock, flags);
1171 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1172 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1175 if (mask != 0 && rcu_gp_in_progress(rsp)) {
1177 /* rcu_report_qs_rnp() releases rnp->lock. */
1178 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1181 spin_unlock_irqrestore(&rnp->lock, flags);
1186 * Force quiescent states on reluctant CPUs, and also detect which
1187 * CPUs are in dyntick-idle mode.
1189 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1191 unsigned long flags;
1192 struct rcu_node *rnp = rcu_get_root(rsp);
1194 if (!rcu_gp_in_progress(rsp))
1195 return; /* No grace period in progress, nothing to force. */
1196 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1197 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1198 return; /* Someone else is already on the job. */
1201 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1202 goto unlock_fqs_ret; /* no emergency and done recently. */
1204 spin_lock(&rnp->lock); /* irqs already disabled */
1205 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1206 if(!rcu_gp_in_progress(rsp)) {
1207 rsp->n_force_qs_ngp++;
1208 spin_unlock(&rnp->lock); /* irqs remain disabled */
1209 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1211 rsp->fqs_active = 1;
1212 switch (rsp->signaled) {
1216 break; /* grace period idle or initializing, ignore. */
1218 case RCU_SAVE_DYNTICK:
1220 spin_unlock(&rnp->lock); /* irqs remain disabled */
1221 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1222 break; /* So gcc recognizes the dead code. */
1224 /* Record dyntick-idle state. */
1225 rcu_process_dyntick(rsp, dyntick_save_progress_counter);
1226 spin_lock(&rnp->lock); /* irqs already disabled */
1227 if (rcu_gp_in_progress(rsp))
1228 rsp->signaled = RCU_FORCE_QS;
1233 /* Check dyntick-idle state, send IPI to laggarts. */
1234 spin_unlock(&rnp->lock); /* irqs remain disabled */
1235 rcu_process_dyntick(rsp, rcu_implicit_dynticks_qs);
1237 /* Leave state in case more forcing is required. */
1239 spin_lock(&rnp->lock); /* irqs already disabled */
1242 rsp->fqs_active = 0;
1243 spin_unlock(&rnp->lock); /* irqs remain disabled */
1245 spin_unlock_irqrestore(&rsp->fqslock, flags);
1248 #else /* #ifdef CONFIG_SMP */
1250 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1255 #endif /* #else #ifdef CONFIG_SMP */
1258 * This does the RCU processing work from softirq context for the
1259 * specified rcu_state and rcu_data structures. This may be called
1260 * only from the CPU to whom the rdp belongs.
1263 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1265 unsigned long flags;
1267 WARN_ON_ONCE(rdp->beenonline == 0);
1270 * If an RCU GP has gone long enough, go check for dyntick
1271 * idle CPUs and, if needed, send resched IPIs.
1273 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1274 force_quiescent_state(rsp, 1);
1277 * Advance callbacks in response to end of earlier grace
1278 * period that some other CPU ended.
1280 rcu_process_gp_end(rsp, rdp);
1282 /* Update RCU state based on any recent quiescent states. */
1283 rcu_check_quiescent_state(rsp, rdp);
1285 /* Does this CPU require a not-yet-started grace period? */
1286 if (cpu_needs_another_gp(rsp, rdp)) {
1287 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1288 rcu_start_gp(rsp, flags); /* releases above lock */
1291 /* If there are callbacks ready, invoke them. */
1292 rcu_do_batch(rsp, rdp);
1296 * Do softirq processing for the current CPU.
1298 static void rcu_process_callbacks(struct softirq_action *unused)
1301 * Memory references from any prior RCU read-side critical sections
1302 * executed by the interrupted code must be seen before any RCU
1303 * grace-period manipulations below.
1305 smp_mb(); /* See above block comment. */
1307 __rcu_process_callbacks(&rcu_sched_state,
1308 &__get_cpu_var(rcu_sched_data));
1309 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1310 rcu_preempt_process_callbacks();
1313 * Memory references from any later RCU read-side critical sections
1314 * executed by the interrupted code must be seen after any RCU
1315 * grace-period manipulations above.
1317 smp_mb(); /* See above block comment. */
1321 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1322 struct rcu_state *rsp)
1324 unsigned long flags;
1325 struct rcu_data *rdp;
1330 smp_mb(); /* Ensure RCU update seen before callback registry. */
1333 * Opportunistically note grace-period endings and beginnings.
1334 * Note that we might see a beginning right after we see an
1335 * end, but never vice versa, since this CPU has to pass through
1336 * a quiescent state betweentimes.
1338 local_irq_save(flags);
1339 rdp = rsp->rda[smp_processor_id()];
1340 rcu_process_gp_end(rsp, rdp);
1341 check_for_new_grace_period(rsp, rdp);
1343 /* Add the callback to our list. */
1344 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1345 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1347 /* Start a new grace period if one not already started. */
1348 if (!rcu_gp_in_progress(rsp)) {
1349 unsigned long nestflag;
1350 struct rcu_node *rnp_root = rcu_get_root(rsp);
1352 spin_lock_irqsave(&rnp_root->lock, nestflag);
1353 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1357 * Force the grace period if too many callbacks or too long waiting.
1358 * Enforce hysteresis, and don't invoke force_quiescent_state()
1359 * if some other CPU has recently done so. Also, don't bother
1360 * invoking force_quiescent_state() if the newly enqueued callback
1361 * is the only one waiting for a grace period to complete.
1363 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1364 rdp->blimit = LONG_MAX;
1365 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1366 *rdp->nxttail[RCU_DONE_TAIL] != head)
1367 force_quiescent_state(rsp, 0);
1368 rdp->n_force_qs_snap = rsp->n_force_qs;
1369 rdp->qlen_last_fqs_check = rdp->qlen;
1370 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1371 force_quiescent_state(rsp, 1);
1372 local_irq_restore(flags);
1376 * Queue an RCU-sched callback for invocation after a grace period.
1378 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1380 __call_rcu(head, func, &rcu_sched_state);
1382 EXPORT_SYMBOL_GPL(call_rcu_sched);
1385 * Queue an RCU for invocation after a quicker grace period.
1387 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1389 __call_rcu(head, func, &rcu_bh_state);
1391 EXPORT_SYMBOL_GPL(call_rcu_bh);
1394 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1396 * Control will return to the caller some time after a full rcu-sched
1397 * grace period has elapsed, in other words after all currently executing
1398 * rcu-sched read-side critical sections have completed. These read-side
1399 * critical sections are delimited by rcu_read_lock_sched() and
1400 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1401 * local_irq_disable(), and so on may be used in place of
1402 * rcu_read_lock_sched().
1404 * This means that all preempt_disable code sequences, including NMI and
1405 * hardware-interrupt handlers, in progress on entry will have completed
1406 * before this primitive returns. However, this does not guarantee that
1407 * softirq handlers will have completed, since in some kernels, these
1408 * handlers can run in process context, and can block.
1410 * This primitive provides the guarantees made by the (now removed)
1411 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1412 * guarantees that rcu_read_lock() sections will have completed.
1413 * In "classic RCU", these two guarantees happen to be one and
1414 * the same, but can differ in realtime RCU implementations.
1416 void synchronize_sched(void)
1418 struct rcu_synchronize rcu;
1420 if (rcu_blocking_is_gp())
1423 init_completion(&rcu.completion);
1424 /* Will wake me after RCU finished. */
1425 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1427 wait_for_completion(&rcu.completion);
1429 EXPORT_SYMBOL_GPL(synchronize_sched);
1432 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1434 * Control will return to the caller some time after a full rcu_bh grace
1435 * period has elapsed, in other words after all currently executing rcu_bh
1436 * read-side critical sections have completed. RCU read-side critical
1437 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1438 * and may be nested.
1440 void synchronize_rcu_bh(void)
1442 struct rcu_synchronize rcu;
1444 if (rcu_blocking_is_gp())
1447 init_completion(&rcu.completion);
1448 /* Will wake me after RCU finished. */
1449 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1451 wait_for_completion(&rcu.completion);
1453 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1456 * Check to see if there is any immediate RCU-related work to be done
1457 * by the current CPU, for the specified type of RCU, returning 1 if so.
1458 * The checks are in order of increasing expense: checks that can be
1459 * carried out against CPU-local state are performed first. However,
1460 * we must check for CPU stalls first, else we might not get a chance.
1462 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1464 struct rcu_node *rnp = rdp->mynode;
1466 rdp->n_rcu_pending++;
1468 /* Check for CPU stalls, if enabled. */
1469 check_cpu_stall(rsp, rdp);
1471 /* Is the RCU core waiting for a quiescent state from this CPU? */
1472 if (rdp->qs_pending) {
1473 rdp->n_rp_qs_pending++;
1477 /* Does this CPU have callbacks ready to invoke? */
1478 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1479 rdp->n_rp_cb_ready++;
1483 /* Has RCU gone idle with this CPU needing another grace period? */
1484 if (cpu_needs_another_gp(rsp, rdp)) {
1485 rdp->n_rp_cpu_needs_gp++;
1489 /* Has another RCU grace period completed? */
1490 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1491 rdp->n_rp_gp_completed++;
1495 /* Has a new RCU grace period started? */
1496 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1497 rdp->n_rp_gp_started++;
1501 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1502 if (rcu_gp_in_progress(rsp) &&
1503 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1504 rdp->n_rp_need_fqs++;
1509 rdp->n_rp_need_nothing++;
1514 * Check to see if there is any immediate RCU-related work to be done
1515 * by the current CPU, returning 1 if so. This function is part of the
1516 * RCU implementation; it is -not- an exported member of the RCU API.
1518 static int rcu_pending(int cpu)
1520 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1521 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1522 rcu_preempt_pending(cpu);
1526 * Check to see if any future RCU-related work will need to be done
1527 * by the current CPU, even if none need be done immediately, returning
1528 * 1 if so. This function is part of the RCU implementation; it is -not-
1529 * an exported member of the RCU API.
1531 int rcu_needs_cpu(int cpu)
1533 /* RCU callbacks either ready or pending? */
1534 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1535 per_cpu(rcu_bh_data, cpu).nxtlist ||
1536 rcu_preempt_needs_cpu(cpu);
1540 * This function is invoked towards the end of the scheduler's initialization
1541 * process. Before this is called, the idle task might contain
1542 * RCU read-side critical sections (during which time, this idle
1543 * task is booting the system). After this function is called, the
1544 * idle tasks are prohibited from containing RCU read-side critical
1547 void rcu_scheduler_starting(void)
1549 WARN_ON(num_online_cpus() != 1);
1550 WARN_ON(nr_context_switches() > 0);
1551 rcu_scheduler_active = 1;
1554 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1555 static atomic_t rcu_barrier_cpu_count;
1556 static DEFINE_MUTEX(rcu_barrier_mutex);
1557 static struct completion rcu_barrier_completion;
1559 static void rcu_barrier_callback(struct rcu_head *notused)
1561 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1562 complete(&rcu_barrier_completion);
1566 * Called with preemption disabled, and from cross-cpu IRQ context.
1568 static void rcu_barrier_func(void *type)
1570 int cpu = smp_processor_id();
1571 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1572 void (*call_rcu_func)(struct rcu_head *head,
1573 void (*func)(struct rcu_head *head));
1575 atomic_inc(&rcu_barrier_cpu_count);
1576 call_rcu_func = type;
1577 call_rcu_func(head, rcu_barrier_callback);
1581 * Orchestrate the specified type of RCU barrier, waiting for all
1582 * RCU callbacks of the specified type to complete.
1584 static void _rcu_barrier(struct rcu_state *rsp,
1585 void (*call_rcu_func)(struct rcu_head *head,
1586 void (*func)(struct rcu_head *head)))
1588 BUG_ON(in_interrupt());
1589 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1590 mutex_lock(&rcu_barrier_mutex);
1591 init_completion(&rcu_barrier_completion);
1593 * Initialize rcu_barrier_cpu_count to 1, then invoke
1594 * rcu_barrier_func() on each CPU, so that each CPU also has
1595 * incremented rcu_barrier_cpu_count. Only then is it safe to
1596 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1597 * might complete its grace period before all of the other CPUs
1598 * did their increment, causing this function to return too
1601 atomic_set(&rcu_barrier_cpu_count, 1);
1602 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1603 rcu_adopt_orphan_cbs(rsp);
1604 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1605 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1606 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1607 complete(&rcu_barrier_completion);
1608 wait_for_completion(&rcu_barrier_completion);
1609 mutex_unlock(&rcu_barrier_mutex);
1613 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1615 void rcu_barrier_bh(void)
1617 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1619 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1622 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1624 void rcu_barrier_sched(void)
1626 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1628 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1631 * Do boot-time initialization of a CPU's per-CPU RCU data.
1634 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1636 unsigned long flags;
1638 struct rcu_data *rdp = rsp->rda[cpu];
1639 struct rcu_node *rnp = rcu_get_root(rsp);
1641 /* Set up local state, ensuring consistent view of global state. */
1642 spin_lock_irqsave(&rnp->lock, flags);
1643 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1644 rdp->nxtlist = NULL;
1645 for (i = 0; i < RCU_NEXT_SIZE; i++)
1646 rdp->nxttail[i] = &rdp->nxtlist;
1649 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1650 #endif /* #ifdef CONFIG_NO_HZ */
1652 spin_unlock_irqrestore(&rnp->lock, flags);
1656 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1657 * offline event can be happening at a given time. Note also that we
1658 * can accept some slop in the rsp->completed access due to the fact
1659 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1661 static void __cpuinit
1662 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1664 unsigned long flags;
1666 struct rcu_data *rdp = rsp->rda[cpu];
1667 struct rcu_node *rnp = rcu_get_root(rsp);
1669 /* Set up local state, ensuring consistent view of global state. */
1670 spin_lock_irqsave(&rnp->lock, flags);
1671 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1672 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1673 rdp->beenonline = 1; /* We have now been online. */
1674 rdp->preemptable = preemptable;
1675 rdp->qlen_last_fqs_check = 0;
1676 rdp->n_force_qs_snap = rsp->n_force_qs;
1677 rdp->blimit = blimit;
1678 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1681 * A new grace period might start here. If so, we won't be part
1682 * of it, but that is OK, as we are currently in a quiescent state.
1685 /* Exclude any attempts to start a new GP on large systems. */
1686 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1688 /* Add CPU to rcu_node bitmasks. */
1690 mask = rdp->grpmask;
1692 /* Exclude any attempts to start a new GP on small systems. */
1693 spin_lock(&rnp->lock); /* irqs already disabled. */
1694 rnp->qsmaskinit |= mask;
1695 mask = rnp->grpmask;
1696 if (rnp == rdp->mynode) {
1697 rdp->gpnum = rnp->completed; /* if GP in progress... */
1698 rdp->completed = rnp->completed;
1699 rdp->passed_quiesc_completed = rnp->completed - 1;
1701 spin_unlock(&rnp->lock); /* irqs already disabled. */
1703 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1705 spin_unlock_irqrestore(&rsp->onofflock, flags);
1708 static void __cpuinit rcu_online_cpu(int cpu)
1710 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1711 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1712 rcu_preempt_init_percpu_data(cpu);
1716 * Handle CPU online/offline notification events.
1718 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1719 unsigned long action, void *hcpu)
1721 long cpu = (long)hcpu;
1724 case CPU_UP_PREPARE:
1725 case CPU_UP_PREPARE_FROZEN:
1726 rcu_online_cpu(cpu);
1729 case CPU_DYING_FROZEN:
1731 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1732 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1733 * returns, all online cpus have queued rcu_barrier_func().
1734 * The dying CPU clears its cpu_online_mask bit and
1735 * moves all of its RCU callbacks to ->orphan_cbs_list
1736 * in the context of stop_machine(), so subsequent calls
1737 * to _rcu_barrier() will adopt these callbacks and only
1738 * then queue rcu_barrier_func() on all remaining CPUs.
1740 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1741 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1742 rcu_preempt_send_cbs_to_orphanage();
1745 case CPU_DEAD_FROZEN:
1746 case CPU_UP_CANCELED:
1747 case CPU_UP_CANCELED_FROZEN:
1748 rcu_offline_cpu(cpu);
1757 * Compute the per-level fanout, either using the exact fanout specified
1758 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1760 #ifdef CONFIG_RCU_FANOUT_EXACT
1761 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1765 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1766 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1768 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1769 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1776 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1777 ccur = rsp->levelcnt[i];
1778 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1782 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1785 * Helper function for rcu_init() that initializes one rcu_state structure.
1787 static void __init rcu_init_one(struct rcu_state *rsp)
1792 struct rcu_node *rnp;
1794 /* Initialize the level-tracking arrays. */
1796 for (i = 1; i < NUM_RCU_LVLS; i++)
1797 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1798 rcu_init_levelspread(rsp);
1800 /* Initialize the elements themselves, starting from the leaves. */
1802 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1803 cpustride *= rsp->levelspread[i];
1804 rnp = rsp->level[i];
1805 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1806 spin_lock_init(&rnp->lock);
1807 lockdep_set_class(&rnp->lock, &rcu_node_class[i]);
1810 rnp->qsmaskinit = 0;
1811 rnp->grplo = j * cpustride;
1812 rnp->grphi = (j + 1) * cpustride - 1;
1813 if (rnp->grphi >= NR_CPUS)
1814 rnp->grphi = NR_CPUS - 1;
1820 rnp->grpnum = j % rsp->levelspread[i - 1];
1821 rnp->grpmask = 1UL << rnp->grpnum;
1822 rnp->parent = rsp->level[i - 1] +
1823 j / rsp->levelspread[i - 1];
1826 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1827 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1828 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1829 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1835 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1836 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1839 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1843 struct rcu_node *rnp; \
1845 rcu_init_one(rsp); \
1846 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1848 for_each_possible_cpu(i) { \
1849 if (i > rnp[j].grphi) \
1851 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1852 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1853 rcu_boot_init_percpu_data(i, rsp); \
1857 void __init rcu_init(void)
1861 rcu_bootup_announce();
1862 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1863 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1864 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1865 #if NUM_RCU_LVL_4 != 0
1866 printk(KERN_INFO "Experimental four-level hierarchy is enabled.\n");
1867 #endif /* #if NUM_RCU_LVL_4 != 0 */
1868 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1869 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1870 __rcu_init_preempt();
1871 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1874 * We don't need protection against CPU-hotplug here because
1875 * this is called early in boot, before either interrupts
1876 * or the scheduler are operational.
1878 cpu_notifier(rcu_cpu_notify, 0);
1879 for_each_online_cpu(i)
1880 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)i);
1883 #include "rcutree_plugin.h"