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>
52 /* Data structures. */
54 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
56 #define RCU_STATE_INITIALIZER(name) { \
57 .level = { &name.node[0] }, \
59 NUM_RCU_LVL_0, /* root of hierarchy. */ \
63 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
65 .signaled = RCU_GP_IDLE, \
68 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&name.onofflock), \
69 .orphan_cbs_list = NULL, \
70 .orphan_cbs_tail = &name.orphan_cbs_list, \
72 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&name.fqslock), \
74 .n_force_qs_ngp = 0, \
77 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
78 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
80 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
81 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
85 * permit this function to be invoked without holding the root rcu_node
86 * structure's ->lock, but of course results can be subject to change.
88 static int rcu_gp_in_progress(struct rcu_state *rsp)
90 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
94 * Note a quiescent state. Because we do not need to know
95 * how many quiescent states passed, just if there was at least
96 * one since the start of the grace period, this just sets a flag.
98 void rcu_sched_qs(int cpu)
100 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
102 rdp->passed_quiesc_completed = rdp->gpnum - 1;
104 rdp->passed_quiesc = 1;
107 void rcu_bh_qs(int cpu)
109 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
111 rdp->passed_quiesc_completed = rdp->gpnum - 1;
113 rdp->passed_quiesc = 1;
117 * Note a context switch. This is a quiescent state for RCU-sched,
118 * and requires special handling for preemptible RCU.
120 void rcu_note_context_switch(int cpu)
123 rcu_preempt_note_context_switch(cpu);
127 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
128 .dynticks_nesting = 1,
131 #endif /* #ifdef CONFIG_NO_HZ */
133 static int blimit = 10; /* Maximum callbacks per softirq. */
134 static int qhimark = 10000; /* If this many pending, ignore blimit. */
135 static int qlowmark = 100; /* Once only this many pending, use blimit. */
137 module_param(blimit, int, 0);
138 module_param(qhimark, int, 0);
139 module_param(qlowmark, int, 0);
141 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
142 static int rcu_pending(int cpu);
145 * Return the number of RCU-sched batches processed thus far for debug & stats.
147 long rcu_batches_completed_sched(void)
149 return rcu_sched_state.completed;
151 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
154 * Return the number of RCU BH batches processed thus far for debug & stats.
156 long rcu_batches_completed_bh(void)
158 return rcu_bh_state.completed;
160 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
163 * Force a quiescent state for RCU BH.
165 void rcu_bh_force_quiescent_state(void)
167 force_quiescent_state(&rcu_bh_state, 0);
169 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
172 * Force a quiescent state for RCU-sched.
174 void rcu_sched_force_quiescent_state(void)
176 force_quiescent_state(&rcu_sched_state, 0);
178 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
181 * Does the CPU have callbacks ready to be invoked?
184 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
186 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
190 * Does the current CPU require a yet-as-unscheduled grace period?
193 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
195 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
199 * Return the root node of the specified rcu_state structure.
201 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
203 return &rsp->node[0];
209 * If the specified CPU is offline, tell the caller that it is in
210 * a quiescent state. Otherwise, whack it with a reschedule IPI.
211 * Grace periods can end up waiting on an offline CPU when that
212 * CPU is in the process of coming online -- it will be added to the
213 * rcu_node bitmasks before it actually makes it online. The same thing
214 * can happen while a CPU is in the process of coming online. Because this
215 * race is quite rare, we check for it after detecting that the grace
216 * period has been delayed rather than checking each and every CPU
217 * each and every time we start a new grace period.
219 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
222 * If the CPU is offline, it is in a quiescent state. We can
223 * trust its state not to change because interrupts are disabled.
225 if (cpu_is_offline(rdp->cpu)) {
230 /* If preemptable RCU, no point in sending reschedule IPI. */
231 if (rdp->preemptable)
234 /* The CPU is online, so send it a reschedule IPI. */
235 if (rdp->cpu != smp_processor_id())
236 smp_send_reschedule(rdp->cpu);
243 #endif /* #ifdef CONFIG_SMP */
248 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
250 * Enter nohz mode, in other words, -leave- the mode in which RCU
251 * read-side critical sections can occur. (Though RCU read-side
252 * critical sections can occur in irq handlers in nohz mode, a possibility
253 * handled by rcu_irq_enter() and rcu_irq_exit()).
255 void rcu_enter_nohz(void)
258 struct rcu_dynticks *rdtp;
260 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
261 local_irq_save(flags);
262 rdtp = &__get_cpu_var(rcu_dynticks);
264 rdtp->dynticks_nesting--;
265 WARN_ON_ONCE(rdtp->dynticks & 0x1);
266 local_irq_restore(flags);
270 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
272 * Exit nohz mode, in other words, -enter- the mode in which RCU
273 * read-side critical sections normally occur.
275 void rcu_exit_nohz(void)
278 struct rcu_dynticks *rdtp;
280 local_irq_save(flags);
281 rdtp = &__get_cpu_var(rcu_dynticks);
283 rdtp->dynticks_nesting++;
284 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
285 local_irq_restore(flags);
286 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
290 * rcu_nmi_enter - inform RCU of entry to NMI context
292 * If the CPU was idle with dynamic ticks active, and there is no
293 * irq handler running, this updates rdtp->dynticks_nmi to let the
294 * RCU grace-period handling know that the CPU is active.
296 void rcu_nmi_enter(void)
298 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
300 if (rdtp->dynticks & 0x1)
302 rdtp->dynticks_nmi++;
303 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
304 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
308 * rcu_nmi_exit - inform RCU of exit from NMI context
310 * If the CPU was idle with dynamic ticks active, and there is no
311 * irq handler running, this updates rdtp->dynticks_nmi to let the
312 * RCU grace-period handling know that the CPU is no longer active.
314 void rcu_nmi_exit(void)
316 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
318 if (rdtp->dynticks & 0x1)
320 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
321 rdtp->dynticks_nmi++;
322 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
326 * rcu_irq_enter - inform RCU of entry to hard irq context
328 * If the CPU was idle with dynamic ticks active, this updates the
329 * rdtp->dynticks to let the RCU handling know that the CPU is active.
331 void rcu_irq_enter(void)
333 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
335 if (rdtp->dynticks_nesting++)
338 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
339 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
343 * rcu_irq_exit - inform RCU of exit from hard irq context
345 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
346 * to put let the RCU handling be aware that the CPU is going back to idle
349 void rcu_irq_exit(void)
351 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
353 if (--rdtp->dynticks_nesting)
355 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
357 WARN_ON_ONCE(rdtp->dynticks & 0x1);
359 /* If the interrupt queued a callback, get out of dyntick mode. */
360 if (__get_cpu_var(rcu_sched_data).nxtlist ||
361 __get_cpu_var(rcu_bh_data).nxtlist)
368 * Snapshot the specified CPU's dynticks counter so that we can later
369 * credit them with an implicit quiescent state. Return 1 if this CPU
370 * is in dynticks idle mode, which is an extended quiescent state.
372 static int dyntick_save_progress_counter(struct rcu_data *rdp)
378 snap = rdp->dynticks->dynticks;
379 snap_nmi = rdp->dynticks->dynticks_nmi;
380 smp_mb(); /* Order sampling of snap with end of grace period. */
381 rdp->dynticks_snap = snap;
382 rdp->dynticks_nmi_snap = snap_nmi;
383 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
390 * Return true if the specified CPU has passed through a quiescent
391 * state by virtue of being in or having passed through an dynticks
392 * idle state since the last call to dyntick_save_progress_counter()
395 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
402 curr = rdp->dynticks->dynticks;
403 snap = rdp->dynticks_snap;
404 curr_nmi = rdp->dynticks->dynticks_nmi;
405 snap_nmi = rdp->dynticks_nmi_snap;
406 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
409 * If the CPU passed through or entered a dynticks idle phase with
410 * no active irq/NMI handlers, then we can safely pretend that the CPU
411 * already acknowledged the request to pass through a quiescent
412 * state. Either way, that CPU cannot possibly be in an RCU
413 * read-side critical section that started before the beginning
414 * of the current RCU grace period.
416 if ((curr != snap || (curr & 0x1) == 0) &&
417 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
422 /* Go check for the CPU being offline. */
423 return rcu_implicit_offline_qs(rdp);
426 #endif /* #ifdef CONFIG_SMP */
428 #else /* #ifdef CONFIG_NO_HZ */
432 static int dyntick_save_progress_counter(struct rcu_data *rdp)
437 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
439 return rcu_implicit_offline_qs(rdp);
442 #endif /* #ifdef CONFIG_SMP */
444 #endif /* #else #ifdef CONFIG_NO_HZ */
446 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
448 static void record_gp_stall_check_time(struct rcu_state *rsp)
450 rsp->gp_start = jiffies;
451 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
454 static void print_other_cpu_stall(struct rcu_state *rsp)
459 struct rcu_node *rnp = rcu_get_root(rsp);
461 /* Only let one CPU complain about others per time interval. */
463 raw_spin_lock_irqsave(&rnp->lock, flags);
464 delta = jiffies - rsp->jiffies_stall;
465 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
466 raw_spin_unlock_irqrestore(&rnp->lock, flags);
469 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
472 * Now rat on any tasks that got kicked up to the root rcu_node
473 * due to CPU offlining.
475 rcu_print_task_stall(rnp);
476 raw_spin_unlock_irqrestore(&rnp->lock, flags);
478 /* OK, time to rat on our buddy... */
480 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
481 rcu_for_each_leaf_node(rsp, rnp) {
482 raw_spin_lock_irqsave(&rnp->lock, flags);
483 rcu_print_task_stall(rnp);
484 raw_spin_unlock_irqrestore(&rnp->lock, flags);
485 if (rnp->qsmask == 0)
487 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
488 if (rnp->qsmask & (1UL << cpu))
489 printk(" %d", rnp->grplo + cpu);
491 printk(" (detected by %d, t=%ld jiffies)\n",
492 smp_processor_id(), (long)(jiffies - rsp->gp_start));
493 trigger_all_cpu_backtrace();
495 /* If so configured, complain about tasks blocking the grace period. */
497 rcu_print_detail_task_stall(rsp);
499 force_quiescent_state(rsp, 0); /* Kick them all. */
502 static void print_cpu_stall(struct rcu_state *rsp)
505 struct rcu_node *rnp = rcu_get_root(rsp);
507 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
508 smp_processor_id(), jiffies - rsp->gp_start);
509 trigger_all_cpu_backtrace();
511 raw_spin_lock_irqsave(&rnp->lock, flags);
512 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
514 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
515 raw_spin_unlock_irqrestore(&rnp->lock, flags);
517 set_need_resched(); /* kick ourselves to get things going. */
520 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
523 struct rcu_node *rnp;
525 delta = jiffies - rsp->jiffies_stall;
527 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
529 /* We haven't checked in, so go dump stack. */
530 print_cpu_stall(rsp);
532 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
534 /* They had two time units to dump stack, so complain. */
535 print_other_cpu_stall(rsp);
539 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
541 static void record_gp_stall_check_time(struct rcu_state *rsp)
545 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
549 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
552 * Update CPU-local rcu_data state to record the newly noticed grace period.
553 * This is used both when we started the grace period and when we notice
554 * that someone else started the grace period. The caller must hold the
555 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
556 * and must have irqs disabled.
558 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
560 if (rdp->gpnum != rnp->gpnum) {
562 rdp->passed_quiesc = 0;
563 rdp->gpnum = rnp->gpnum;
567 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
570 struct rcu_node *rnp;
572 local_irq_save(flags);
574 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
575 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
576 local_irq_restore(flags);
579 __note_new_gpnum(rsp, rnp, rdp);
580 raw_spin_unlock_irqrestore(&rnp->lock, flags);
584 * Did someone else start a new RCU grace period start since we last
585 * checked? Update local state appropriately if so. Must be called
586 * on the CPU corresponding to rdp.
589 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
594 local_irq_save(flags);
595 if (rdp->gpnum != rsp->gpnum) {
596 note_new_gpnum(rsp, rdp);
599 local_irq_restore(flags);
604 * Advance this CPU's callbacks, but only if the current grace period
605 * has ended. This may be called only from the CPU to whom the rdp
606 * belongs. In addition, the corresponding leaf rcu_node structure's
607 * ->lock must be held by the caller, with irqs disabled.
610 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
612 /* Did another grace period end? */
613 if (rdp->completed != rnp->completed) {
615 /* Advance callbacks. No harm if list empty. */
616 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
617 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
618 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
620 /* Remember that we saw this grace-period completion. */
621 rdp->completed = rnp->completed;
626 * Advance this CPU's callbacks, but only if the current grace period
627 * has ended. This may be called only from the CPU to whom the rdp
631 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
634 struct rcu_node *rnp;
636 local_irq_save(flags);
638 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
639 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
640 local_irq_restore(flags);
643 __rcu_process_gp_end(rsp, rnp, rdp);
644 raw_spin_unlock_irqrestore(&rnp->lock, flags);
648 * Do per-CPU grace-period initialization for running CPU. The caller
649 * must hold the lock of the leaf rcu_node structure corresponding to
653 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
655 /* Prior grace period ended, so advance callbacks for current CPU. */
656 __rcu_process_gp_end(rsp, rnp, rdp);
659 * Because this CPU just now started the new grace period, we know
660 * that all of its callbacks will be covered by this upcoming grace
661 * period, even the ones that were registered arbitrarily recently.
662 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
664 * Other CPUs cannot be sure exactly when the grace period started.
665 * Therefore, their recently registered callbacks must pass through
666 * an additional RCU_NEXT_READY stage, so that they will be handled
667 * by the next RCU grace period.
669 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
670 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
672 /* Set state so that this CPU will detect the next quiescent state. */
673 __note_new_gpnum(rsp, rnp, rdp);
677 * Start a new RCU grace period if warranted, re-initializing the hierarchy
678 * in preparation for detecting the next grace period. The caller must hold
679 * the root node's ->lock, which is released before return. Hard irqs must
683 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
684 __releases(rcu_get_root(rsp)->lock)
686 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
687 struct rcu_node *rnp = rcu_get_root(rsp);
689 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
690 if (cpu_needs_another_gp(rsp, rdp))
691 rsp->fqs_need_gp = 1;
692 if (rnp->completed == rsp->completed) {
693 raw_spin_unlock_irqrestore(&rnp->lock, flags);
696 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
699 * Propagate new ->completed value to rcu_node structures
700 * so that other CPUs don't have to wait until the start
701 * of the next grace period to process their callbacks.
703 rcu_for_each_node_breadth_first(rsp, rnp) {
704 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
705 rnp->completed = rsp->completed;
706 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
708 local_irq_restore(flags);
712 /* Advance to a new grace period and initialize state. */
714 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
715 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
716 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
717 record_gp_stall_check_time(rsp);
719 /* Special-case the common single-level case. */
720 if (NUM_RCU_NODES == 1) {
721 rcu_preempt_check_blocked_tasks(rnp);
722 rnp->qsmask = rnp->qsmaskinit;
723 rnp->gpnum = rsp->gpnum;
724 rnp->completed = rsp->completed;
725 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
726 rcu_start_gp_per_cpu(rsp, rnp, rdp);
727 raw_spin_unlock_irqrestore(&rnp->lock, flags);
731 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
734 /* Exclude any concurrent CPU-hotplug operations. */
735 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
738 * Set the quiescent-state-needed bits in all the rcu_node
739 * structures for all currently online CPUs in breadth-first
740 * order, starting from the root rcu_node structure. This
741 * operation relies on the layout of the hierarchy within the
742 * rsp->node[] array. Note that other CPUs will access only
743 * the leaves of the hierarchy, which still indicate that no
744 * grace period is in progress, at least until the corresponding
745 * leaf node has been initialized. In addition, we have excluded
746 * CPU-hotplug operations.
748 * Note that the grace period cannot complete until we finish
749 * the initialization process, as there will be at least one
750 * qsmask bit set in the root node until that time, namely the
751 * one corresponding to this CPU, due to the fact that we have
754 rcu_for_each_node_breadth_first(rsp, rnp) {
755 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
756 rcu_preempt_check_blocked_tasks(rnp);
757 rnp->qsmask = rnp->qsmaskinit;
758 rnp->gpnum = rsp->gpnum;
759 rnp->completed = rsp->completed;
760 if (rnp == rdp->mynode)
761 rcu_start_gp_per_cpu(rsp, rnp, rdp);
762 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
765 rnp = rcu_get_root(rsp);
766 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
767 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
768 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
769 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
773 * Report a full set of quiescent states to the specified rcu_state
774 * data structure. This involves cleaning up after the prior grace
775 * period and letting rcu_start_gp() start up the next grace period
776 * if one is needed. Note that the caller must hold rnp->lock, as
777 * required by rcu_start_gp(), which will release it.
779 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
780 __releases(rcu_get_root(rsp)->lock)
782 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
783 rsp->completed = rsp->gpnum;
784 rsp->signaled = RCU_GP_IDLE;
785 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
789 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
790 * Allows quiescent states for a group of CPUs to be reported at one go
791 * to the specified rcu_node structure, though all the CPUs in the group
792 * must be represented by the same rcu_node structure (which need not be
793 * a leaf rcu_node structure, though it often will be). That structure's
794 * lock must be held upon entry, and it is released before return.
797 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
798 struct rcu_node *rnp, unsigned long flags)
799 __releases(rnp->lock)
801 struct rcu_node *rnp_c;
803 /* Walk up the rcu_node hierarchy. */
805 if (!(rnp->qsmask & mask)) {
807 /* Our bit has already been cleared, so done. */
808 raw_spin_unlock_irqrestore(&rnp->lock, flags);
811 rnp->qsmask &= ~mask;
812 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
814 /* Other bits still set at this level, so done. */
815 raw_spin_unlock_irqrestore(&rnp->lock, flags);
819 if (rnp->parent == NULL) {
821 /* No more levels. Exit loop holding root lock. */
825 raw_spin_unlock_irqrestore(&rnp->lock, flags);
828 raw_spin_lock_irqsave(&rnp->lock, flags);
829 WARN_ON_ONCE(rnp_c->qsmask);
833 * Get here if we are the last CPU to pass through a quiescent
834 * state for this grace period. Invoke rcu_report_qs_rsp()
835 * to clean up and start the next grace period if one is needed.
837 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
841 * Record a quiescent state for the specified CPU to that CPU's rcu_data
842 * structure. This must be either called from the specified CPU, or
843 * called when the specified CPU is known to be offline (and when it is
844 * also known that no other CPU is concurrently trying to help the offline
845 * CPU). The lastcomp argument is used to make sure we are still in the
846 * grace period of interest. We don't want to end the current grace period
847 * based on quiescent states detected in an earlier grace period!
850 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
854 struct rcu_node *rnp;
857 raw_spin_lock_irqsave(&rnp->lock, flags);
858 if (lastcomp != rnp->completed) {
861 * Someone beat us to it for this grace period, so leave.
862 * The race with GP start is resolved by the fact that we
863 * hold the leaf rcu_node lock, so that the per-CPU bits
864 * cannot yet be initialized -- so we would simply find our
865 * CPU's bit already cleared in rcu_report_qs_rnp() if this
868 rdp->passed_quiesc = 0; /* try again later! */
869 raw_spin_unlock_irqrestore(&rnp->lock, flags);
873 if ((rnp->qsmask & mask) == 0) {
874 raw_spin_unlock_irqrestore(&rnp->lock, flags);
879 * This GP can't end until cpu checks in, so all of our
880 * callbacks can be processed during the next GP.
882 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
884 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
889 * Check to see if there is a new grace period of which this CPU
890 * is not yet aware, and if so, set up local rcu_data state for it.
891 * Otherwise, see if this CPU has just passed through its first
892 * quiescent state for this grace period, and record that fact if so.
895 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
897 /* If there is now a new grace period, record and return. */
898 if (check_for_new_grace_period(rsp, rdp))
902 * Does this CPU still need to do its part for current grace period?
903 * If no, return and let the other CPUs do their part as well.
905 if (!rdp->qs_pending)
909 * Was there a quiescent state since the beginning of the grace
910 * period? If no, then exit and wait for the next call.
912 if (!rdp->passed_quiesc)
916 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
919 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
922 #ifdef CONFIG_HOTPLUG_CPU
925 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
926 * specified flavor of RCU. The callbacks will be adopted by the next
927 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
928 * comes first. Because this is invoked from the CPU_DYING notifier,
929 * irqs are already disabled.
931 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
934 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
936 if (rdp->nxtlist == NULL)
937 return; /* irqs disabled, so comparison is stable. */
938 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
939 *rsp->orphan_cbs_tail = rdp->nxtlist;
940 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
942 for (i = 0; i < RCU_NEXT_SIZE; i++)
943 rdp->nxttail[i] = &rdp->nxtlist;
944 rsp->orphan_qlen += rdp->qlen;
946 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
950 * Adopt previously orphaned RCU callbacks.
952 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
955 struct rcu_data *rdp;
957 raw_spin_lock_irqsave(&rsp->onofflock, flags);
958 rdp = rsp->rda[smp_processor_id()];
959 if (rsp->orphan_cbs_list == NULL) {
960 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
963 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
964 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
965 rdp->qlen += rsp->orphan_qlen;
966 rsp->orphan_cbs_list = NULL;
967 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
968 rsp->orphan_qlen = 0;
969 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
973 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
974 * and move all callbacks from the outgoing CPU to the current one.
976 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
981 struct rcu_data *rdp = rsp->rda[cpu];
982 struct rcu_node *rnp;
984 /* Exclude any attempts to start a new grace period. */
985 raw_spin_lock_irqsave(&rsp->onofflock, flags);
987 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
988 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
989 mask = rdp->grpmask; /* rnp->grplo is constant. */
991 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
992 rnp->qsmaskinit &= ~mask;
993 if (rnp->qsmaskinit != 0) {
994 if (rnp != rdp->mynode)
995 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
998 if (rnp == rdp->mynode)
999 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1001 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1002 mask = rnp->grpmask;
1004 } while (rnp != NULL);
1007 * We still hold the leaf rcu_node structure lock here, and
1008 * irqs are still disabled. The reason for this subterfuge is
1009 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1010 * held leads to deadlock.
1012 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1014 if (need_report & RCU_OFL_TASKS_NORM_GP)
1015 rcu_report_unblock_qs_rnp(rnp, flags);
1017 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1018 if (need_report & RCU_OFL_TASKS_EXP_GP)
1019 rcu_report_exp_rnp(rsp, rnp);
1021 rcu_adopt_orphan_cbs(rsp);
1025 * Remove the specified CPU from the RCU hierarchy and move any pending
1026 * callbacks that it might have to the current CPU. This code assumes
1027 * that at least one CPU in the system will remain running at all times.
1028 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1030 static void rcu_offline_cpu(int cpu)
1032 __rcu_offline_cpu(cpu, &rcu_sched_state);
1033 __rcu_offline_cpu(cpu, &rcu_bh_state);
1034 rcu_preempt_offline_cpu(cpu);
1037 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1039 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1043 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1047 static void rcu_offline_cpu(int cpu)
1051 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1054 * Invoke any RCU callbacks that have made it to the end of their grace
1055 * period. Thottle as specified by rdp->blimit.
1057 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1059 unsigned long flags;
1060 struct rcu_head *next, *list, **tail;
1063 /* If no callbacks are ready, just return.*/
1064 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1068 * Extract the list of ready callbacks, disabling to prevent
1069 * races with call_rcu() from interrupt handlers.
1071 local_irq_save(flags);
1072 list = rdp->nxtlist;
1073 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1074 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1075 tail = rdp->nxttail[RCU_DONE_TAIL];
1076 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1077 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1078 rdp->nxttail[count] = &rdp->nxtlist;
1079 local_irq_restore(flags);
1081 /* Invoke callbacks. */
1088 if (++count >= rdp->blimit)
1092 local_irq_save(flags);
1094 /* Update count, and requeue any remaining callbacks. */
1097 *tail = rdp->nxtlist;
1098 rdp->nxtlist = list;
1099 for (count = 0; count < RCU_NEXT_SIZE; count++)
1100 if (&rdp->nxtlist == rdp->nxttail[count])
1101 rdp->nxttail[count] = tail;
1106 /* Reinstate batch limit if we have worked down the excess. */
1107 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1108 rdp->blimit = blimit;
1110 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1111 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1112 rdp->qlen_last_fqs_check = 0;
1113 rdp->n_force_qs_snap = rsp->n_force_qs;
1114 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1115 rdp->qlen_last_fqs_check = rdp->qlen;
1117 local_irq_restore(flags);
1119 /* Re-raise the RCU softirq if there are callbacks remaining. */
1120 if (cpu_has_callbacks_ready_to_invoke(rdp))
1121 raise_softirq(RCU_SOFTIRQ);
1125 * Check to see if this CPU is in a non-context-switch quiescent state
1126 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1127 * Also schedule the RCU softirq handler.
1129 * This function must be called with hardirqs disabled. It is normally
1130 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1131 * false, there is no point in invoking rcu_check_callbacks().
1133 void rcu_check_callbacks(int cpu, int user)
1135 if (!rcu_pending(cpu))
1136 return; /* if nothing for RCU to do. */
1138 (idle_cpu(cpu) && rcu_scheduler_active &&
1139 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1142 * Get here if this CPU took its interrupt from user
1143 * mode or from the idle loop, and if this is not a
1144 * nested interrupt. In this case, the CPU is in
1145 * a quiescent state, so note it.
1147 * No memory barrier is required here because both
1148 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1149 * variables that other CPUs neither access nor modify,
1150 * at least not while the corresponding CPU is online.
1156 } else if (!in_softirq()) {
1159 * Get here if this CPU did not take its interrupt from
1160 * softirq, in other words, if it is not interrupting
1161 * a rcu_bh read-side critical section. This is an _bh
1162 * critical section, so note it.
1167 rcu_preempt_check_callbacks(cpu);
1168 raise_softirq(RCU_SOFTIRQ);
1174 * Scan the leaf rcu_node structures, processing dyntick state for any that
1175 * have not yet encountered a quiescent state, using the function specified.
1176 * The caller must have suppressed start of new grace periods.
1178 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1182 unsigned long flags;
1184 struct rcu_node *rnp;
1186 rcu_for_each_leaf_node(rsp, rnp) {
1188 raw_spin_lock_irqsave(&rnp->lock, flags);
1189 if (!rcu_gp_in_progress(rsp)) {
1190 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1193 if (rnp->qsmask == 0) {
1194 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1199 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1200 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1205 /* rcu_report_qs_rnp() releases rnp->lock. */
1206 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1209 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1214 * Force quiescent states on reluctant CPUs, and also detect which
1215 * CPUs are in dyntick-idle mode.
1217 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1219 unsigned long flags;
1220 struct rcu_node *rnp = rcu_get_root(rsp);
1222 if (!rcu_gp_in_progress(rsp))
1223 return; /* No grace period in progress, nothing to force. */
1224 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1225 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1226 return; /* Someone else is already on the job. */
1228 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1229 goto unlock_fqs_ret; /* no emergency and done recently. */
1231 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1232 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1233 if(!rcu_gp_in_progress(rsp)) {
1234 rsp->n_force_qs_ngp++;
1235 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1236 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1238 rsp->fqs_active = 1;
1239 switch (rsp->signaled) {
1243 break; /* grace period idle or initializing, ignore. */
1245 case RCU_SAVE_DYNTICK:
1246 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1247 break; /* So gcc recognizes the dead code. */
1249 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1251 /* Record dyntick-idle state. */
1252 force_qs_rnp(rsp, dyntick_save_progress_counter);
1253 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1254 if (rcu_gp_in_progress(rsp))
1255 rsp->signaled = RCU_FORCE_QS;
1260 /* Check dyntick-idle state, send IPI to laggarts. */
1261 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1262 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1264 /* Leave state in case more forcing is required. */
1266 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1269 rsp->fqs_active = 0;
1270 if (rsp->fqs_need_gp) {
1271 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1272 rsp->fqs_need_gp = 0;
1273 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1276 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1278 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1281 #else /* #ifdef CONFIG_SMP */
1283 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1288 #endif /* #else #ifdef CONFIG_SMP */
1291 * This does the RCU processing work from softirq context for the
1292 * specified rcu_state and rcu_data structures. This may be called
1293 * only from the CPU to whom the rdp belongs.
1296 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1298 unsigned long flags;
1300 WARN_ON_ONCE(rdp->beenonline == 0);
1303 * If an RCU GP has gone long enough, go check for dyntick
1304 * idle CPUs and, if needed, send resched IPIs.
1306 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1307 force_quiescent_state(rsp, 1);
1310 * Advance callbacks in response to end of earlier grace
1311 * period that some other CPU ended.
1313 rcu_process_gp_end(rsp, rdp);
1315 /* Update RCU state based on any recent quiescent states. */
1316 rcu_check_quiescent_state(rsp, rdp);
1318 /* Does this CPU require a not-yet-started grace period? */
1319 if (cpu_needs_another_gp(rsp, rdp)) {
1320 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1321 rcu_start_gp(rsp, flags); /* releases above lock */
1324 /* If there are callbacks ready, invoke them. */
1325 rcu_do_batch(rsp, rdp);
1329 * Do softirq processing for the current CPU.
1331 static void rcu_process_callbacks(struct softirq_action *unused)
1334 * Memory references from any prior RCU read-side critical sections
1335 * executed by the interrupted code must be seen before any RCU
1336 * grace-period manipulations below.
1338 smp_mb(); /* See above block comment. */
1340 __rcu_process_callbacks(&rcu_sched_state,
1341 &__get_cpu_var(rcu_sched_data));
1342 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1343 rcu_preempt_process_callbacks();
1346 * Memory references from any later RCU read-side critical sections
1347 * executed by the interrupted code must be seen after any RCU
1348 * grace-period manipulations above.
1350 smp_mb(); /* See above block comment. */
1352 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1353 rcu_needs_cpu_flush();
1357 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1358 struct rcu_state *rsp)
1360 unsigned long flags;
1361 struct rcu_data *rdp;
1366 smp_mb(); /* Ensure RCU update seen before callback registry. */
1369 * Opportunistically note grace-period endings and beginnings.
1370 * Note that we might see a beginning right after we see an
1371 * end, but never vice versa, since this CPU has to pass through
1372 * a quiescent state betweentimes.
1374 local_irq_save(flags);
1375 rdp = rsp->rda[smp_processor_id()];
1376 rcu_process_gp_end(rsp, rdp);
1377 check_for_new_grace_period(rsp, rdp);
1379 /* Add the callback to our list. */
1380 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1381 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1383 /* Start a new grace period if one not already started. */
1384 if (!rcu_gp_in_progress(rsp)) {
1385 unsigned long nestflag;
1386 struct rcu_node *rnp_root = rcu_get_root(rsp);
1388 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1389 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1393 * Force the grace period if too many callbacks or too long waiting.
1394 * Enforce hysteresis, and don't invoke force_quiescent_state()
1395 * if some other CPU has recently done so. Also, don't bother
1396 * invoking force_quiescent_state() if the newly enqueued callback
1397 * is the only one waiting for a grace period to complete.
1399 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1400 rdp->blimit = LONG_MAX;
1401 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1402 *rdp->nxttail[RCU_DONE_TAIL] != head)
1403 force_quiescent_state(rsp, 0);
1404 rdp->n_force_qs_snap = rsp->n_force_qs;
1405 rdp->qlen_last_fqs_check = rdp->qlen;
1406 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1407 force_quiescent_state(rsp, 1);
1408 local_irq_restore(flags);
1412 * Queue an RCU-sched callback for invocation after a grace period.
1414 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1416 __call_rcu(head, func, &rcu_sched_state);
1418 EXPORT_SYMBOL_GPL(call_rcu_sched);
1421 * Queue an RCU for invocation after a quicker grace period.
1423 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1425 __call_rcu(head, func, &rcu_bh_state);
1427 EXPORT_SYMBOL_GPL(call_rcu_bh);
1430 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1432 * Control will return to the caller some time after a full rcu-sched
1433 * grace period has elapsed, in other words after all currently executing
1434 * rcu-sched read-side critical sections have completed. These read-side
1435 * critical sections are delimited by rcu_read_lock_sched() and
1436 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1437 * local_irq_disable(), and so on may be used in place of
1438 * rcu_read_lock_sched().
1440 * This means that all preempt_disable code sequences, including NMI and
1441 * hardware-interrupt handlers, in progress on entry will have completed
1442 * before this primitive returns. However, this does not guarantee that
1443 * softirq handlers will have completed, since in some kernels, these
1444 * handlers can run in process context, and can block.
1446 * This primitive provides the guarantees made by the (now removed)
1447 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1448 * guarantees that rcu_read_lock() sections will have completed.
1449 * In "classic RCU", these two guarantees happen to be one and
1450 * the same, but can differ in realtime RCU implementations.
1452 void synchronize_sched(void)
1454 struct rcu_synchronize rcu;
1456 if (rcu_blocking_is_gp())
1459 init_completion(&rcu.completion);
1460 /* Will wake me after RCU finished. */
1461 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1463 wait_for_completion(&rcu.completion);
1465 EXPORT_SYMBOL_GPL(synchronize_sched);
1468 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1470 * Control will return to the caller some time after a full rcu_bh grace
1471 * period has elapsed, in other words after all currently executing rcu_bh
1472 * read-side critical sections have completed. RCU read-side critical
1473 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1474 * and may be nested.
1476 void synchronize_rcu_bh(void)
1478 struct rcu_synchronize rcu;
1480 if (rcu_blocking_is_gp())
1483 init_completion(&rcu.completion);
1484 /* Will wake me after RCU finished. */
1485 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1487 wait_for_completion(&rcu.completion);
1489 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1492 * Check to see if there is any immediate RCU-related work to be done
1493 * by the current CPU, for the specified type of RCU, returning 1 if so.
1494 * The checks are in order of increasing expense: checks that can be
1495 * carried out against CPU-local state are performed first. However,
1496 * we must check for CPU stalls first, else we might not get a chance.
1498 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1500 struct rcu_node *rnp = rdp->mynode;
1502 rdp->n_rcu_pending++;
1504 /* Check for CPU stalls, if enabled. */
1505 check_cpu_stall(rsp, rdp);
1507 /* Is the RCU core waiting for a quiescent state from this CPU? */
1508 if (rdp->qs_pending) {
1511 * If force_quiescent_state() coming soon and this CPU
1512 * needs a quiescent state, and this is either RCU-sched
1513 * or RCU-bh, force a local reschedule.
1515 if (!rdp->preemptable &&
1516 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1519 rdp->n_rp_qs_pending++;
1523 /* Does this CPU have callbacks ready to invoke? */
1524 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1525 rdp->n_rp_cb_ready++;
1529 /* Has RCU gone idle with this CPU needing another grace period? */
1530 if (cpu_needs_another_gp(rsp, rdp)) {
1531 rdp->n_rp_cpu_needs_gp++;
1535 /* Has another RCU grace period completed? */
1536 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1537 rdp->n_rp_gp_completed++;
1541 /* Has a new RCU grace period started? */
1542 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1543 rdp->n_rp_gp_started++;
1547 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1548 if (rcu_gp_in_progress(rsp) &&
1549 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1550 rdp->n_rp_need_fqs++;
1555 rdp->n_rp_need_nothing++;
1560 * Check to see if there is any immediate RCU-related work to be done
1561 * by the current CPU, returning 1 if so. This function is part of the
1562 * RCU implementation; it is -not- an exported member of the RCU API.
1564 static int rcu_pending(int cpu)
1566 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1567 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1568 rcu_preempt_pending(cpu);
1572 * Check to see if any future RCU-related work will need to be done
1573 * by the current CPU, even if none need be done immediately, returning
1576 static int rcu_needs_cpu_quick_check(int cpu)
1578 /* RCU callbacks either ready or pending? */
1579 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1580 per_cpu(rcu_bh_data, cpu).nxtlist ||
1581 rcu_preempt_needs_cpu(cpu);
1584 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1585 static atomic_t rcu_barrier_cpu_count;
1586 static DEFINE_MUTEX(rcu_barrier_mutex);
1587 static struct completion rcu_barrier_completion;
1589 static void rcu_barrier_callback(struct rcu_head *notused)
1591 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1592 complete(&rcu_barrier_completion);
1596 * Called with preemption disabled, and from cross-cpu IRQ context.
1598 static void rcu_barrier_func(void *type)
1600 int cpu = smp_processor_id();
1601 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1602 void (*call_rcu_func)(struct rcu_head *head,
1603 void (*func)(struct rcu_head *head));
1605 atomic_inc(&rcu_barrier_cpu_count);
1606 call_rcu_func = type;
1607 call_rcu_func(head, rcu_barrier_callback);
1611 * Orchestrate the specified type of RCU barrier, waiting for all
1612 * RCU callbacks of the specified type to complete.
1614 static void _rcu_barrier(struct rcu_state *rsp,
1615 void (*call_rcu_func)(struct rcu_head *head,
1616 void (*func)(struct rcu_head *head)))
1618 BUG_ON(in_interrupt());
1619 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1620 mutex_lock(&rcu_barrier_mutex);
1621 init_completion(&rcu_barrier_completion);
1623 * Initialize rcu_barrier_cpu_count to 1, then invoke
1624 * rcu_barrier_func() on each CPU, so that each CPU also has
1625 * incremented rcu_barrier_cpu_count. Only then is it safe to
1626 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1627 * might complete its grace period before all of the other CPUs
1628 * did their increment, causing this function to return too
1631 atomic_set(&rcu_barrier_cpu_count, 1);
1632 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1633 rcu_adopt_orphan_cbs(rsp);
1634 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1635 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1636 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1637 complete(&rcu_barrier_completion);
1638 wait_for_completion(&rcu_barrier_completion);
1639 mutex_unlock(&rcu_barrier_mutex);
1643 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1645 void rcu_barrier_bh(void)
1647 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1649 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1652 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1654 void rcu_barrier_sched(void)
1656 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1658 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1661 * Do boot-time initialization of a CPU's per-CPU RCU data.
1664 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1666 unsigned long flags;
1668 struct rcu_data *rdp = rsp->rda[cpu];
1669 struct rcu_node *rnp = rcu_get_root(rsp);
1671 /* Set up local state, ensuring consistent view of global state. */
1672 raw_spin_lock_irqsave(&rnp->lock, flags);
1673 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1674 rdp->nxtlist = NULL;
1675 for (i = 0; i < RCU_NEXT_SIZE; i++)
1676 rdp->nxttail[i] = &rdp->nxtlist;
1679 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1680 #endif /* #ifdef CONFIG_NO_HZ */
1682 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1686 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1687 * offline event can be happening at a given time. Note also that we
1688 * can accept some slop in the rsp->completed access due to the fact
1689 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1691 static void __cpuinit
1692 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1694 unsigned long flags;
1696 struct rcu_data *rdp = rsp->rda[cpu];
1697 struct rcu_node *rnp = rcu_get_root(rsp);
1699 /* Set up local state, ensuring consistent view of global state. */
1700 raw_spin_lock_irqsave(&rnp->lock, flags);
1701 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1702 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1703 rdp->beenonline = 1; /* We have now been online. */
1704 rdp->preemptable = preemptable;
1705 rdp->qlen_last_fqs_check = 0;
1706 rdp->n_force_qs_snap = rsp->n_force_qs;
1707 rdp->blimit = blimit;
1708 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1711 * A new grace period might start here. If so, we won't be part
1712 * of it, but that is OK, as we are currently in a quiescent state.
1715 /* Exclude any attempts to start a new GP on large systems. */
1716 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1718 /* Add CPU to rcu_node bitmasks. */
1720 mask = rdp->grpmask;
1722 /* Exclude any attempts to start a new GP on small systems. */
1723 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1724 rnp->qsmaskinit |= mask;
1725 mask = rnp->grpmask;
1726 if (rnp == rdp->mynode) {
1727 rdp->gpnum = rnp->completed; /* if GP in progress... */
1728 rdp->completed = rnp->completed;
1729 rdp->passed_quiesc_completed = rnp->completed - 1;
1731 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1733 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1735 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1738 static void __cpuinit rcu_online_cpu(int cpu)
1740 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1741 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1742 rcu_preempt_init_percpu_data(cpu);
1746 * Handle CPU online/offline notification events.
1748 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1749 unsigned long action, void *hcpu)
1751 long cpu = (long)hcpu;
1754 case CPU_UP_PREPARE:
1755 case CPU_UP_PREPARE_FROZEN:
1756 rcu_online_cpu(cpu);
1759 case CPU_DYING_FROZEN:
1761 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1762 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1763 * returns, all online cpus have queued rcu_barrier_func().
1764 * The dying CPU clears its cpu_online_mask bit and
1765 * moves all of its RCU callbacks to ->orphan_cbs_list
1766 * in the context of stop_machine(), so subsequent calls
1767 * to _rcu_barrier() will adopt these callbacks and only
1768 * then queue rcu_barrier_func() on all remaining CPUs.
1770 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1771 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1772 rcu_preempt_send_cbs_to_orphanage();
1775 case CPU_DEAD_FROZEN:
1776 case CPU_UP_CANCELED:
1777 case CPU_UP_CANCELED_FROZEN:
1778 rcu_offline_cpu(cpu);
1787 * Compute the per-level fanout, either using the exact fanout specified
1788 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1790 #ifdef CONFIG_RCU_FANOUT_EXACT
1791 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1795 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1796 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1798 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1799 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1806 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1807 ccur = rsp->levelcnt[i];
1808 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1812 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1815 * Helper function for rcu_init() that initializes one rcu_state structure.
1817 static void __init rcu_init_one(struct rcu_state *rsp)
1819 static char *buf[] = { "rcu_node_level_0",
1822 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1826 struct rcu_node *rnp;
1828 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1830 /* Initialize the level-tracking arrays. */
1832 for (i = 1; i < NUM_RCU_LVLS; i++)
1833 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1834 rcu_init_levelspread(rsp);
1836 /* Initialize the elements themselves, starting from the leaves. */
1838 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1839 cpustride *= rsp->levelspread[i];
1840 rnp = rsp->level[i];
1841 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1842 raw_spin_lock_init(&rnp->lock);
1843 lockdep_set_class_and_name(&rnp->lock,
1844 &rcu_node_class[i], buf[i]);
1847 rnp->qsmaskinit = 0;
1848 rnp->grplo = j * cpustride;
1849 rnp->grphi = (j + 1) * cpustride - 1;
1850 if (rnp->grphi >= NR_CPUS)
1851 rnp->grphi = NR_CPUS - 1;
1857 rnp->grpnum = j % rsp->levelspread[i - 1];
1858 rnp->grpmask = 1UL << rnp->grpnum;
1859 rnp->parent = rsp->level[i - 1] +
1860 j / rsp->levelspread[i - 1];
1863 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1864 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1865 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1866 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1870 rnp = rsp->level[NUM_RCU_LVLS - 1];
1871 for_each_possible_cpu(i) {
1874 rsp->rda[i]->mynode = rnp;
1875 rcu_boot_init_percpu_data(i, rsp);
1880 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1881 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1884 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1888 for_each_possible_cpu(i) { \
1889 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1891 rcu_init_one(rsp); \
1894 void __init rcu_init(void)
1898 rcu_bootup_announce();
1899 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1900 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1901 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1902 #if NUM_RCU_LVL_4 != 0
1903 printk(KERN_INFO "Experimental four-level hierarchy is enabled.\n");
1904 #endif /* #if NUM_RCU_LVL_4 != 0 */
1905 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1906 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1907 __rcu_init_preempt();
1908 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1911 * We don't need protection against CPU-hotplug here because
1912 * this is called early in boot, before either interrupts
1913 * or the scheduler are operational.
1915 cpu_notifier(rcu_cpu_notify, 0);
1916 for_each_online_cpu(cpu)
1917 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1920 #include "rcutree_plugin.h"