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_root_class;
56 #define RCU_STATE_INITIALIZER(name) { \
57 .level = { &name.node[0] }, \
59 NUM_RCU_LVL_0, /* root of hierarchy. */ \
62 NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \
64 .signaled = RCU_GP_IDLE, \
67 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
68 .orphan_cbs_list = NULL, \
69 .orphan_cbs_tail = &name.orphan_cbs_list, \
71 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
73 .n_force_qs_ngp = 0, \
76 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
77 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
79 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
80 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
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;
102 rdp = &per_cpu(rcu_sched_data, cpu);
103 rdp->passed_quiesc_completed = rdp->gpnum - 1;
105 rdp->passed_quiesc = 1;
106 rcu_preempt_note_context_switch(cpu);
109 void rcu_bh_qs(int cpu)
111 struct rcu_data *rdp;
113 rdp = &per_cpu(rcu_bh_data, cpu);
114 rdp->passed_quiesc_completed = rdp->gpnum - 1;
116 rdp->passed_quiesc = 1;
120 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
121 .dynticks_nesting = 1,
124 #endif /* #ifdef CONFIG_NO_HZ */
126 static int blimit = 10; /* Maximum callbacks per softirq. */
127 static int qhimark = 10000; /* If this many pending, ignore blimit. */
128 static int qlowmark = 100; /* Once only this many pending, use blimit. */
130 module_param(blimit, int, 0);
131 module_param(qhimark, int, 0);
132 module_param(qlowmark, int, 0);
134 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
135 static int rcu_pending(int cpu);
138 * Return the number of RCU-sched batches processed thus far for debug & stats.
140 long rcu_batches_completed_sched(void)
142 return rcu_sched_state.completed;
144 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
147 * Return the number of RCU BH batches processed thus far for debug & stats.
149 long rcu_batches_completed_bh(void)
151 return rcu_bh_state.completed;
153 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
156 * Does the CPU have callbacks ready to be invoked?
159 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
161 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
165 * Does the current CPU require a yet-as-unscheduled grace period?
168 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
170 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
174 * Return the root node of the specified rcu_state structure.
176 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
178 return &rsp->node[0];
182 * Record the specified "completed" value, which is later used to validate
183 * dynticks counter manipulations and CPU-offline checks. Specify
184 * "rsp->completed - 1" to unconditionally invalidate any future dynticks
185 * manipulations and CPU-offline checks. Such invalidation is useful at
186 * the beginning of a grace period.
188 static void dyntick_record_completed(struct rcu_state *rsp, long comp)
190 rsp->completed_fqs = comp;
196 * Recall the previously recorded value of the completion for dynticks.
198 static long dyntick_recall_completed(struct rcu_state *rsp)
200 return rsp->completed_fqs;
204 * If the specified CPU is offline, tell the caller that it is in
205 * a quiescent state. Otherwise, whack it with a reschedule IPI.
206 * Grace periods can end up waiting on an offline CPU when that
207 * CPU is in the process of coming online -- it will be added to the
208 * rcu_node bitmasks before it actually makes it online. The same thing
209 * can happen while a CPU is in the process of coming online. Because this
210 * race is quite rare, we check for it after detecting that the grace
211 * period has been delayed rather than checking each and every CPU
212 * each and every time we start a new grace period.
214 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
217 * If the CPU is offline, it is in a quiescent state. We can
218 * trust its state not to change because interrupts are disabled.
220 if (cpu_is_offline(rdp->cpu)) {
225 /* If preemptable RCU, no point in sending reschedule IPI. */
226 if (rdp->preemptable)
229 /* The CPU is online, so send it a reschedule IPI. */
230 if (rdp->cpu != smp_processor_id())
231 smp_send_reschedule(rdp->cpu);
238 #endif /* #ifdef CONFIG_SMP */
243 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
245 * Enter nohz mode, in other words, -leave- the mode in which RCU
246 * read-side critical sections can occur. (Though RCU read-side
247 * critical sections can occur in irq handlers in nohz mode, a possibility
248 * handled by rcu_irq_enter() and rcu_irq_exit()).
250 void rcu_enter_nohz(void)
253 struct rcu_dynticks *rdtp;
255 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
256 local_irq_save(flags);
257 rdtp = &__get_cpu_var(rcu_dynticks);
259 rdtp->dynticks_nesting--;
260 WARN_ON_ONCE(rdtp->dynticks & 0x1);
261 local_irq_restore(flags);
265 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
267 * Exit nohz mode, in other words, -enter- the mode in which RCU
268 * read-side critical sections normally occur.
270 void rcu_exit_nohz(void)
273 struct rcu_dynticks *rdtp;
275 local_irq_save(flags);
276 rdtp = &__get_cpu_var(rcu_dynticks);
278 rdtp->dynticks_nesting++;
279 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
280 local_irq_restore(flags);
281 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
285 * rcu_nmi_enter - inform RCU of entry to NMI context
287 * If the CPU was idle with dynamic ticks active, and there is no
288 * irq handler running, this updates rdtp->dynticks_nmi to let the
289 * RCU grace-period handling know that the CPU is active.
291 void rcu_nmi_enter(void)
293 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
295 if (rdtp->dynticks & 0x1)
297 rdtp->dynticks_nmi++;
298 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
299 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
303 * rcu_nmi_exit - inform RCU of exit from NMI context
305 * If the CPU was idle with dynamic ticks active, and there is no
306 * irq handler running, this updates rdtp->dynticks_nmi to let the
307 * RCU grace-period handling know that the CPU is no longer active.
309 void rcu_nmi_exit(void)
311 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
313 if (rdtp->dynticks & 0x1)
315 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
316 rdtp->dynticks_nmi++;
317 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
321 * rcu_irq_enter - inform RCU of entry to hard irq context
323 * If the CPU was idle with dynamic ticks active, this updates the
324 * rdtp->dynticks to let the RCU handling know that the CPU is active.
326 void rcu_irq_enter(void)
328 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
330 if (rdtp->dynticks_nesting++)
333 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
334 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
338 * rcu_irq_exit - inform RCU of exit from hard irq context
340 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
341 * to put let the RCU handling be aware that the CPU is going back to idle
344 void rcu_irq_exit(void)
346 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
348 if (--rdtp->dynticks_nesting)
350 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
352 WARN_ON_ONCE(rdtp->dynticks & 0x1);
354 /* If the interrupt queued a callback, get out of dyntick mode. */
355 if (__get_cpu_var(rcu_sched_data).nxtlist ||
356 __get_cpu_var(rcu_bh_data).nxtlist)
363 * Snapshot the specified CPU's dynticks counter so that we can later
364 * credit them with an implicit quiescent state. Return 1 if this CPU
365 * is in dynticks idle mode, which is an extended quiescent state.
367 static int dyntick_save_progress_counter(struct rcu_data *rdp)
373 snap = rdp->dynticks->dynticks;
374 snap_nmi = rdp->dynticks->dynticks_nmi;
375 smp_mb(); /* Order sampling of snap with end of grace period. */
376 rdp->dynticks_snap = snap;
377 rdp->dynticks_nmi_snap = snap_nmi;
378 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
385 * Return true if the specified CPU has passed through a quiescent
386 * state by virtue of being in or having passed through an dynticks
387 * idle state since the last call to dyntick_save_progress_counter()
390 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
397 curr = rdp->dynticks->dynticks;
398 snap = rdp->dynticks_snap;
399 curr_nmi = rdp->dynticks->dynticks_nmi;
400 snap_nmi = rdp->dynticks_nmi_snap;
401 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
404 * If the CPU passed through or entered a dynticks idle phase with
405 * no active irq/NMI handlers, then we can safely pretend that the CPU
406 * already acknowledged the request to pass through a quiescent
407 * state. Either way, that CPU cannot possibly be in an RCU
408 * read-side critical section that started before the beginning
409 * of the current RCU grace period.
411 if ((curr != snap || (curr & 0x1) == 0) &&
412 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
417 /* Go check for the CPU being offline. */
418 return rcu_implicit_offline_qs(rdp);
421 #endif /* #ifdef CONFIG_SMP */
423 #else /* #ifdef CONFIG_NO_HZ */
427 static int dyntick_save_progress_counter(struct rcu_data *rdp)
432 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
434 return rcu_implicit_offline_qs(rdp);
437 #endif /* #ifdef CONFIG_SMP */
439 #endif /* #else #ifdef CONFIG_NO_HZ */
441 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
443 static void record_gp_stall_check_time(struct rcu_state *rsp)
445 rsp->gp_start = jiffies;
446 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
449 static void print_other_cpu_stall(struct rcu_state *rsp)
454 struct rcu_node *rnp = rcu_get_root(rsp);
456 /* Only let one CPU complain about others per time interval. */
458 spin_lock_irqsave(&rnp->lock, flags);
459 delta = jiffies - rsp->jiffies_stall;
460 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
461 spin_unlock_irqrestore(&rnp->lock, flags);
464 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
467 * Now rat on any tasks that got kicked up to the root rcu_node
468 * due to CPU offlining.
470 rcu_print_task_stall(rnp);
471 spin_unlock_irqrestore(&rnp->lock, flags);
473 /* OK, time to rat on our buddy... */
475 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
476 rcu_for_each_leaf_node(rsp, rnp) {
477 rcu_print_task_stall(rnp);
478 if (rnp->qsmask == 0)
480 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
481 if (rnp->qsmask & (1UL << cpu))
482 printk(" %d", rnp->grplo + cpu);
484 printk(" (detected by %d, t=%ld jiffies)\n",
485 smp_processor_id(), (long)(jiffies - rsp->gp_start));
486 trigger_all_cpu_backtrace();
488 force_quiescent_state(rsp, 0); /* Kick them all. */
491 static void print_cpu_stall(struct rcu_state *rsp)
494 struct rcu_node *rnp = rcu_get_root(rsp);
496 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
497 smp_processor_id(), jiffies - rsp->gp_start);
498 trigger_all_cpu_backtrace();
500 spin_lock_irqsave(&rnp->lock, flags);
501 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
503 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
504 spin_unlock_irqrestore(&rnp->lock, flags);
506 set_need_resched(); /* kick ourselves to get things going. */
509 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
512 struct rcu_node *rnp;
514 delta = jiffies - rsp->jiffies_stall;
516 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
518 /* We haven't checked in, so go dump stack. */
519 print_cpu_stall(rsp);
521 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
523 /* They had two time units to dump stack, so complain. */
524 print_other_cpu_stall(rsp);
528 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
530 static void record_gp_stall_check_time(struct rcu_state *rsp)
534 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
538 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
541 * Update CPU-local rcu_data state to record the newly noticed grace period.
542 * This is used both when we started the grace period and when we notice
543 * that someone else started the grace period. The caller must hold the
544 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
545 * and must have irqs disabled.
547 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
549 if (rdp->gpnum != rnp->gpnum) {
551 rdp->passed_quiesc = 0;
552 rdp->gpnum = rnp->gpnum;
556 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
559 struct rcu_node *rnp;
561 local_irq_save(flags);
563 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
564 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
565 local_irq_restore(flags);
568 __note_new_gpnum(rsp, rnp, rdp);
569 spin_unlock_irqrestore(&rnp->lock, flags);
573 * Did someone else start a new RCU grace period start since we last
574 * checked? Update local state appropriately if so. Must be called
575 * on the CPU corresponding to rdp.
578 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
583 local_irq_save(flags);
584 if (rdp->gpnum != rsp->gpnum) {
585 note_new_gpnum(rsp, rdp);
588 local_irq_restore(flags);
593 * Advance this CPU's callbacks, but only if the current grace period
594 * has ended. This may be called only from the CPU to whom the rdp
595 * belongs. In addition, the corresponding leaf rcu_node structure's
596 * ->lock must be held by the caller, with irqs disabled.
599 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
601 /* Did another grace period end? */
602 if (rdp->completed != rnp->completed) {
604 /* Advance callbacks. No harm if list empty. */
605 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
606 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
607 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
609 /* Remember that we saw this grace-period completion. */
610 rdp->completed = rnp->completed;
615 * Advance this CPU's callbacks, but only if the current grace period
616 * has ended. This may be called only from the CPU to whom the rdp
620 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
623 struct rcu_node *rnp;
625 local_irq_save(flags);
627 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
628 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
629 local_irq_restore(flags);
632 __rcu_process_gp_end(rsp, rnp, rdp);
633 spin_unlock_irqrestore(&rnp->lock, flags);
637 * Do per-CPU grace-period initialization for running CPU. The caller
638 * must hold the lock of the leaf rcu_node structure corresponding to
642 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
644 /* Prior grace period ended, so advance callbacks for current CPU. */
645 __rcu_process_gp_end(rsp, rnp, rdp);
648 * Because this CPU just now started the new grace period, we know
649 * that all of its callbacks will be covered by this upcoming grace
650 * period, even the ones that were registered arbitrarily recently.
651 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
653 * Other CPUs cannot be sure exactly when the grace period started.
654 * Therefore, their recently registered callbacks must pass through
655 * an additional RCU_NEXT_READY stage, so that they will be handled
656 * by the next RCU grace period.
658 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
659 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
661 /* Set state so that this CPU will detect the next quiescent state. */
662 __note_new_gpnum(rsp, rnp, rdp);
666 * Start a new RCU grace period if warranted, re-initializing the hierarchy
667 * in preparation for detecting the next grace period. The caller must hold
668 * the root node's ->lock, which is released before return. Hard irqs must
672 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
673 __releases(rcu_get_root(rsp)->lock)
675 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
676 struct rcu_node *rnp = rcu_get_root(rsp);
678 if (!cpu_needs_another_gp(rsp, rdp)) {
679 if (rnp->completed == rsp->completed) {
680 spin_unlock_irqrestore(&rnp->lock, flags);
683 spin_unlock(&rnp->lock); /* irqs remain disabled. */
686 * Propagate new ->completed value to rcu_node structures
687 * so that other CPUs don't have to wait until the start
688 * of the next grace period to process their callbacks.
690 rcu_for_each_node_breadth_first(rsp, rnp) {
691 spin_lock(&rnp->lock); /* irqs already disabled. */
692 rnp->completed = rsp->completed;
693 spin_unlock(&rnp->lock); /* irqs remain disabled. */
695 local_irq_restore(flags);
699 /* Advance to a new grace period and initialize state. */
701 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
702 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
703 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
704 record_gp_stall_check_time(rsp);
705 dyntick_record_completed(rsp, rsp->completed - 1);
707 /* Special-case the common single-level case. */
708 if (NUM_RCU_NODES == 1) {
709 rcu_preempt_check_blocked_tasks(rnp);
710 rnp->qsmask = rnp->qsmaskinit;
711 rnp->gpnum = rsp->gpnum;
712 rnp->completed = rsp->completed;
713 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
714 rcu_start_gp_per_cpu(rsp, rnp, rdp);
715 spin_unlock_irqrestore(&rnp->lock, flags);
719 spin_unlock(&rnp->lock); /* leave irqs disabled. */
722 /* Exclude any concurrent CPU-hotplug operations. */
723 spin_lock(&rsp->onofflock); /* irqs already disabled. */
726 * Set the quiescent-state-needed bits in all the rcu_node
727 * structures for all currently online CPUs in breadth-first
728 * order, starting from the root rcu_node structure. This
729 * operation relies on the layout of the hierarchy within the
730 * rsp->node[] array. Note that other CPUs will access only
731 * the leaves of the hierarchy, which still indicate that no
732 * grace period is in progress, at least until the corresponding
733 * leaf node has been initialized. In addition, we have excluded
734 * CPU-hotplug operations.
736 * Note that the grace period cannot complete until we finish
737 * the initialization process, as there will be at least one
738 * qsmask bit set in the root node until that time, namely the
739 * one corresponding to this CPU, due to the fact that we have
742 rcu_for_each_node_breadth_first(rsp, rnp) {
743 spin_lock(&rnp->lock); /* irqs already disabled. */
744 rcu_preempt_check_blocked_tasks(rnp);
745 rnp->qsmask = rnp->qsmaskinit;
746 rnp->gpnum = rsp->gpnum;
747 rnp->completed = rsp->completed;
748 if (rnp == rdp->mynode)
749 rcu_start_gp_per_cpu(rsp, rnp, rdp);
750 spin_unlock(&rnp->lock); /* irqs remain disabled. */
753 rnp = rcu_get_root(rsp);
754 spin_lock(&rnp->lock); /* irqs already disabled. */
755 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
756 spin_unlock(&rnp->lock); /* irqs remain disabled. */
757 spin_unlock_irqrestore(&rsp->onofflock, flags);
761 * Clean up after the prior grace period and let rcu_start_gp() start up
762 * the next grace period if one is needed. Note that the caller must
763 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
765 static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
766 __releases(rcu_get_root(rsp)->lock)
768 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
769 rsp->completed = rsp->gpnum;
770 rsp->signaled = RCU_GP_IDLE;
771 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
775 * Similar to cpu_quiet(), for which it is a helper function. Allows
776 * a group of CPUs to be quieted at one go, though all the CPUs in the
777 * group must be represented by the same leaf rcu_node structure.
778 * That structure's lock must be held upon entry, and it is released
782 cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp,
784 __releases(rnp->lock)
786 struct rcu_node *rnp_c;
788 /* Walk up the rcu_node hierarchy. */
790 if (!(rnp->qsmask & mask)) {
792 /* Our bit has already been cleared, so done. */
793 spin_unlock_irqrestore(&rnp->lock, flags);
796 rnp->qsmask &= ~mask;
797 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
799 /* Other bits still set at this level, so done. */
800 spin_unlock_irqrestore(&rnp->lock, flags);
804 if (rnp->parent == NULL) {
806 /* No more levels. Exit loop holding root lock. */
810 spin_unlock_irqrestore(&rnp->lock, flags);
813 spin_lock_irqsave(&rnp->lock, flags);
814 WARN_ON_ONCE(rnp_c->qsmask);
818 * Get here if we are the last CPU to pass through a quiescent
819 * state for this grace period. Invoke cpu_quiet_msk_finish()
820 * to clean up and start the next grace period if one is needed.
822 cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */
826 * Record a quiescent state for the specified CPU, which must either be
827 * the current CPU. The lastcomp argument is used to make sure we are
828 * still in the grace period of interest. We don't want to end the current
829 * grace period based on quiescent states detected in an earlier grace
833 cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
837 struct rcu_node *rnp;
840 spin_lock_irqsave(&rnp->lock, flags);
841 if (lastcomp != ACCESS_ONCE(rsp->completed)) {
844 * Someone beat us to it for this grace period, so leave.
845 * The race with GP start is resolved by the fact that we
846 * hold the leaf rcu_node lock, so that the per-CPU bits
847 * cannot yet be initialized -- so we would simply find our
848 * CPU's bit already cleared in cpu_quiet_msk() if this race
851 rdp->passed_quiesc = 0; /* try again later! */
852 spin_unlock_irqrestore(&rnp->lock, flags);
856 if ((rnp->qsmask & mask) == 0) {
857 spin_unlock_irqrestore(&rnp->lock, flags);
862 * This GP can't end until cpu checks in, so all of our
863 * callbacks can be processed during the next GP.
865 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
867 cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */
872 * Check to see if there is a new grace period of which this CPU
873 * is not yet aware, and if so, set up local rcu_data state for it.
874 * Otherwise, see if this CPU has just passed through its first
875 * quiescent state for this grace period, and record that fact if so.
878 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
880 /* If there is now a new grace period, record and return. */
881 if (check_for_new_grace_period(rsp, rdp))
885 * Does this CPU still need to do its part for current grace period?
886 * If no, return and let the other CPUs do their part as well.
888 if (!rdp->qs_pending)
892 * Was there a quiescent state since the beginning of the grace
893 * period? If no, then exit and wait for the next call.
895 if (!rdp->passed_quiesc)
898 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
899 cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
902 #ifdef CONFIG_HOTPLUG_CPU
905 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
906 * specified flavor of RCU. The callbacks will be adopted by the next
907 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
908 * comes first. Because this is invoked from the CPU_DYING notifier,
909 * irqs are already disabled.
911 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
914 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
916 if (rdp->nxtlist == NULL)
917 return; /* irqs disabled, so comparison is stable. */
918 spin_lock(&rsp->onofflock); /* irqs already disabled. */
919 *rsp->orphan_cbs_tail = rdp->nxtlist;
920 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
922 for (i = 0; i < RCU_NEXT_SIZE; i++)
923 rdp->nxttail[i] = &rdp->nxtlist;
924 rsp->orphan_qlen += rdp->qlen;
926 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
930 * Adopt previously orphaned RCU callbacks.
932 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
935 struct rcu_data *rdp;
937 spin_lock_irqsave(&rsp->onofflock, flags);
938 rdp = rsp->rda[smp_processor_id()];
939 if (rsp->orphan_cbs_list == NULL) {
940 spin_unlock_irqrestore(&rsp->onofflock, flags);
943 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
944 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
945 rdp->qlen += rsp->orphan_qlen;
946 rsp->orphan_cbs_list = NULL;
947 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
948 rsp->orphan_qlen = 0;
949 spin_unlock_irqrestore(&rsp->onofflock, flags);
953 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
954 * and move all callbacks from the outgoing CPU to the current one.
956 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
961 struct rcu_data *rdp = rsp->rda[cpu];
962 struct rcu_node *rnp;
964 /* Exclude any attempts to start a new grace period. */
965 spin_lock_irqsave(&rsp->onofflock, flags);
967 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
968 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
969 mask = rdp->grpmask; /* rnp->grplo is constant. */
971 spin_lock(&rnp->lock); /* irqs already disabled. */
972 rnp->qsmaskinit &= ~mask;
973 if (rnp->qsmaskinit != 0) {
974 spin_unlock(&rnp->lock); /* irqs remain disabled. */
979 * If there was a task blocking the current grace period,
980 * and if all CPUs have checked in, we need to propagate
981 * the quiescent state up the rcu_node hierarchy. But that
982 * is inconvenient at the moment due to deadlock issues if
983 * this should end the current grace period. So set the
984 * offlined CPU's bit in ->qsmask in order to force the
985 * next force_quiescent_state() invocation to clean up this
986 * mess in a deadlock-free manner.
988 if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
992 spin_unlock(&rnp->lock); /* irqs remain disabled. */
994 } while (rnp != NULL);
995 lastcomp = rsp->completed;
997 spin_unlock_irqrestore(&rsp->onofflock, flags);
999 rcu_adopt_orphan_cbs(rsp);
1003 * Remove the specified CPU from the RCU hierarchy and move any pending
1004 * callbacks that it might have to the current CPU. This code assumes
1005 * that at least one CPU in the system will remain running at all times.
1006 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1008 static void rcu_offline_cpu(int cpu)
1010 __rcu_offline_cpu(cpu, &rcu_sched_state);
1011 __rcu_offline_cpu(cpu, &rcu_bh_state);
1012 rcu_preempt_offline_cpu(cpu);
1015 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1017 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1021 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1025 static void rcu_offline_cpu(int cpu)
1029 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1032 * Invoke any RCU callbacks that have made it to the end of their grace
1033 * period. Thottle as specified by rdp->blimit.
1035 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1037 unsigned long flags;
1038 struct rcu_head *next, *list, **tail;
1041 /* If no callbacks are ready, just return.*/
1042 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1046 * Extract the list of ready callbacks, disabling to prevent
1047 * races with call_rcu() from interrupt handlers.
1049 local_irq_save(flags);
1050 list = rdp->nxtlist;
1051 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1052 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1053 tail = rdp->nxttail[RCU_DONE_TAIL];
1054 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1055 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1056 rdp->nxttail[count] = &rdp->nxtlist;
1057 local_irq_restore(flags);
1059 /* Invoke callbacks. */
1066 if (++count >= rdp->blimit)
1070 local_irq_save(flags);
1072 /* Update count, and requeue any remaining callbacks. */
1075 *tail = rdp->nxtlist;
1076 rdp->nxtlist = list;
1077 for (count = 0; count < RCU_NEXT_SIZE; count++)
1078 if (&rdp->nxtlist == rdp->nxttail[count])
1079 rdp->nxttail[count] = tail;
1084 /* Reinstate batch limit if we have worked down the excess. */
1085 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1086 rdp->blimit = blimit;
1088 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1089 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1090 rdp->qlen_last_fqs_check = 0;
1091 rdp->n_force_qs_snap = rsp->n_force_qs;
1092 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1093 rdp->qlen_last_fqs_check = rdp->qlen;
1095 local_irq_restore(flags);
1097 /* Re-raise the RCU softirq if there are callbacks remaining. */
1098 if (cpu_has_callbacks_ready_to_invoke(rdp))
1099 raise_softirq(RCU_SOFTIRQ);
1103 * Check to see if this CPU is in a non-context-switch quiescent state
1104 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1105 * Also schedule the RCU softirq handler.
1107 * This function must be called with hardirqs disabled. It is normally
1108 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1109 * false, there is no point in invoking rcu_check_callbacks().
1111 void rcu_check_callbacks(int cpu, int user)
1113 if (!rcu_pending(cpu))
1114 return; /* if nothing for RCU to do. */
1116 (idle_cpu(cpu) && rcu_scheduler_active &&
1117 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1120 * Get here if this CPU took its interrupt from user
1121 * mode or from the idle loop, and if this is not a
1122 * nested interrupt. In this case, the CPU is in
1123 * a quiescent state, so note it.
1125 * No memory barrier is required here because both
1126 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1127 * variables that other CPUs neither access nor modify,
1128 * at least not while the corresponding CPU is online.
1134 } else if (!in_softirq()) {
1137 * Get here if this CPU did not take its interrupt from
1138 * softirq, in other words, if it is not interrupting
1139 * a rcu_bh read-side critical section. This is an _bh
1140 * critical section, so note it.
1145 rcu_preempt_check_callbacks(cpu);
1146 raise_softirq(RCU_SOFTIRQ);
1152 * Scan the leaf rcu_node structures, processing dyntick state for any that
1153 * have not yet encountered a quiescent state, using the function specified.
1154 * Returns 1 if the current grace period ends while scanning (possibly
1155 * because we made it end).
1157 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1158 int (*f)(struct rcu_data *))
1162 unsigned long flags;
1164 struct rcu_node *rnp;
1166 rcu_for_each_leaf_node(rsp, rnp) {
1168 spin_lock_irqsave(&rnp->lock, flags);
1169 if (rsp->completed != lastcomp) {
1170 spin_unlock_irqrestore(&rnp->lock, flags);
1173 if (rnp->qsmask == 0) {
1174 spin_unlock_irqrestore(&rnp->lock, flags);
1179 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1180 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1183 if (mask != 0 && rsp->completed == lastcomp) {
1185 /* cpu_quiet_msk() releases rnp->lock. */
1186 cpu_quiet_msk(mask, rsp, rnp, flags);
1189 spin_unlock_irqrestore(&rnp->lock, flags);
1195 * Force quiescent states on reluctant CPUs, and also detect which
1196 * CPUs are in dyntick-idle mode.
1198 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1200 unsigned long flags;
1202 struct rcu_node *rnp = rcu_get_root(rsp);
1206 if (!rcu_gp_in_progress(rsp))
1207 return; /* No grace period in progress, nothing to force. */
1208 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1209 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1210 return; /* Someone else is already on the job. */
1213 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1214 goto unlock_ret; /* no emergency and done recently. */
1216 spin_lock(&rnp->lock);
1217 lastcomp = rsp->completed;
1218 signaled = rsp->signaled;
1219 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1220 if (lastcomp == rsp->gpnum) {
1221 rsp->n_force_qs_ngp++;
1222 spin_unlock(&rnp->lock);
1223 goto unlock_ret; /* no GP in progress, time updated. */
1225 spin_unlock(&rnp->lock);
1230 break; /* grace period idle or initializing, ignore. */
1232 case RCU_SAVE_DYNTICK:
1234 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1235 break; /* So gcc recognizes the dead code. */
1237 /* Record dyntick-idle state. */
1238 if (rcu_process_dyntick(rsp, lastcomp,
1239 dyntick_save_progress_counter))
1241 /* fall into next case. */
1243 case RCU_SAVE_COMPLETED:
1245 /* Update state, record completion counter. */
1247 spin_lock(&rnp->lock);
1248 if (lastcomp == rsp->completed &&
1249 rsp->signaled == signaled) {
1250 rsp->signaled = RCU_FORCE_QS;
1251 dyntick_record_completed(rsp, lastcomp);
1252 forcenow = signaled == RCU_SAVE_COMPLETED;
1254 spin_unlock(&rnp->lock);
1257 /* fall into next case. */
1261 /* Check dyntick-idle state, send IPI to laggarts. */
1262 if (rcu_process_dyntick(rsp, dyntick_recall_completed(rsp),
1263 rcu_implicit_dynticks_qs))
1266 /* Leave state in case more forcing is required. */
1271 spin_unlock_irqrestore(&rsp->fqslock, flags);
1274 #else /* #ifdef CONFIG_SMP */
1276 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1281 #endif /* #else #ifdef CONFIG_SMP */
1284 * This does the RCU processing work from softirq context for the
1285 * specified rcu_state and rcu_data structures. This may be called
1286 * only from the CPU to whom the rdp belongs.
1289 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1291 unsigned long flags;
1293 WARN_ON_ONCE(rdp->beenonline == 0);
1296 * If an RCU GP has gone long enough, go check for dyntick
1297 * idle CPUs and, if needed, send resched IPIs.
1299 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1300 force_quiescent_state(rsp, 1);
1303 * Advance callbacks in response to end of earlier grace
1304 * period that some other CPU ended.
1306 rcu_process_gp_end(rsp, rdp);
1308 /* Update RCU state based on any recent quiescent states. */
1309 rcu_check_quiescent_state(rsp, rdp);
1311 /* Does this CPU require a not-yet-started grace period? */
1312 if (cpu_needs_another_gp(rsp, rdp)) {
1313 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1314 rcu_start_gp(rsp, flags); /* releases above lock */
1317 /* If there are callbacks ready, invoke them. */
1318 rcu_do_batch(rsp, rdp);
1322 * Do softirq processing for the current CPU.
1324 static void rcu_process_callbacks(struct softirq_action *unused)
1327 * Memory references from any prior RCU read-side critical sections
1328 * executed by the interrupted code must be seen before any RCU
1329 * grace-period manipulations below.
1331 smp_mb(); /* See above block comment. */
1333 __rcu_process_callbacks(&rcu_sched_state,
1334 &__get_cpu_var(rcu_sched_data));
1335 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1336 rcu_preempt_process_callbacks();
1339 * Memory references from any later RCU read-side critical sections
1340 * executed by the interrupted code must be seen after any RCU
1341 * grace-period manipulations above.
1343 smp_mb(); /* See above block comment. */
1347 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1348 struct rcu_state *rsp)
1350 unsigned long flags;
1351 struct rcu_data *rdp;
1356 smp_mb(); /* Ensure RCU update seen before callback registry. */
1359 * Opportunistically note grace-period endings and beginnings.
1360 * Note that we might see a beginning right after we see an
1361 * end, but never vice versa, since this CPU has to pass through
1362 * a quiescent state betweentimes.
1364 local_irq_save(flags);
1365 rdp = rsp->rda[smp_processor_id()];
1366 rcu_process_gp_end(rsp, rdp);
1367 check_for_new_grace_period(rsp, rdp);
1369 /* Add the callback to our list. */
1370 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1371 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1373 /* Start a new grace period if one not already started. */
1374 if (!rcu_gp_in_progress(rsp)) {
1375 unsigned long nestflag;
1376 struct rcu_node *rnp_root = rcu_get_root(rsp);
1378 spin_lock_irqsave(&rnp_root->lock, nestflag);
1379 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1383 * Force the grace period if too many callbacks or too long waiting.
1384 * Enforce hysteresis, and don't invoke force_quiescent_state()
1385 * if some other CPU has recently done so. Also, don't bother
1386 * invoking force_quiescent_state() if the newly enqueued callback
1387 * is the only one waiting for a grace period to complete.
1389 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1390 rdp->blimit = LONG_MAX;
1391 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1392 *rdp->nxttail[RCU_DONE_TAIL] != head)
1393 force_quiescent_state(rsp, 0);
1394 rdp->n_force_qs_snap = rsp->n_force_qs;
1395 rdp->qlen_last_fqs_check = rdp->qlen;
1396 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1397 force_quiescent_state(rsp, 1);
1398 local_irq_restore(flags);
1402 * Queue an RCU-sched callback for invocation after a grace period.
1404 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1406 __call_rcu(head, func, &rcu_sched_state);
1408 EXPORT_SYMBOL_GPL(call_rcu_sched);
1411 * Queue an RCU for invocation after a quicker grace period.
1413 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1415 __call_rcu(head, func, &rcu_bh_state);
1417 EXPORT_SYMBOL_GPL(call_rcu_bh);
1420 * Check to see if there is any immediate RCU-related work to be done
1421 * by the current CPU, for the specified type of RCU, returning 1 if so.
1422 * The checks are in order of increasing expense: checks that can be
1423 * carried out against CPU-local state are performed first. However,
1424 * we must check for CPU stalls first, else we might not get a chance.
1426 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1428 rdp->n_rcu_pending++;
1430 /* Check for CPU stalls, if enabled. */
1431 check_cpu_stall(rsp, rdp);
1433 /* Is the RCU core waiting for a quiescent state from this CPU? */
1434 if (rdp->qs_pending) {
1435 rdp->n_rp_qs_pending++;
1439 /* Does this CPU have callbacks ready to invoke? */
1440 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1441 rdp->n_rp_cb_ready++;
1445 /* Has RCU gone idle with this CPU needing another grace period? */
1446 if (cpu_needs_another_gp(rsp, rdp)) {
1447 rdp->n_rp_cpu_needs_gp++;
1451 /* Has another RCU grace period completed? */
1452 if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */
1453 rdp->n_rp_gp_completed++;
1457 /* Has a new RCU grace period started? */
1458 if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */
1459 rdp->n_rp_gp_started++;
1463 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1464 if (rcu_gp_in_progress(rsp) &&
1465 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1466 rdp->n_rp_need_fqs++;
1471 rdp->n_rp_need_nothing++;
1476 * Check to see if there is any immediate RCU-related work to be done
1477 * by the current CPU, returning 1 if so. This function is part of the
1478 * RCU implementation; it is -not- an exported member of the RCU API.
1480 static int rcu_pending(int cpu)
1482 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1483 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1484 rcu_preempt_pending(cpu);
1488 * Check to see if any future RCU-related work will need to be done
1489 * by the current CPU, even if none need be done immediately, returning
1490 * 1 if so. This function is part of the RCU implementation; it is -not-
1491 * an exported member of the RCU API.
1493 int rcu_needs_cpu(int cpu)
1495 /* RCU callbacks either ready or pending? */
1496 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1497 per_cpu(rcu_bh_data, cpu).nxtlist ||
1498 rcu_preempt_needs_cpu(cpu);
1501 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1502 static atomic_t rcu_barrier_cpu_count;
1503 static DEFINE_MUTEX(rcu_barrier_mutex);
1504 static struct completion rcu_barrier_completion;
1506 static void rcu_barrier_callback(struct rcu_head *notused)
1508 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1509 complete(&rcu_barrier_completion);
1513 * Called with preemption disabled, and from cross-cpu IRQ context.
1515 static void rcu_barrier_func(void *type)
1517 int cpu = smp_processor_id();
1518 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1519 void (*call_rcu_func)(struct rcu_head *head,
1520 void (*func)(struct rcu_head *head));
1522 atomic_inc(&rcu_barrier_cpu_count);
1523 call_rcu_func = type;
1524 call_rcu_func(head, rcu_barrier_callback);
1528 * Orchestrate the specified type of RCU barrier, waiting for all
1529 * RCU callbacks of the specified type to complete.
1531 static void _rcu_barrier(struct rcu_state *rsp,
1532 void (*call_rcu_func)(struct rcu_head *head,
1533 void (*func)(struct rcu_head *head)))
1535 BUG_ON(in_interrupt());
1536 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1537 mutex_lock(&rcu_barrier_mutex);
1538 init_completion(&rcu_barrier_completion);
1540 * Initialize rcu_barrier_cpu_count to 1, then invoke
1541 * rcu_barrier_func() on each CPU, so that each CPU also has
1542 * incremented rcu_barrier_cpu_count. Only then is it safe to
1543 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1544 * might complete its grace period before all of the other CPUs
1545 * did their increment, causing this function to return too
1548 atomic_set(&rcu_barrier_cpu_count, 1);
1549 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1550 rcu_adopt_orphan_cbs(rsp);
1551 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1552 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1553 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1554 complete(&rcu_barrier_completion);
1555 wait_for_completion(&rcu_barrier_completion);
1556 mutex_unlock(&rcu_barrier_mutex);
1560 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1562 void rcu_barrier_bh(void)
1564 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1566 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1569 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1571 void rcu_barrier_sched(void)
1573 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1575 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1578 * Do boot-time initialization of a CPU's per-CPU RCU data.
1581 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1583 unsigned long flags;
1585 struct rcu_data *rdp = rsp->rda[cpu];
1586 struct rcu_node *rnp = rcu_get_root(rsp);
1588 /* Set up local state, ensuring consistent view of global state. */
1589 spin_lock_irqsave(&rnp->lock, flags);
1590 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1591 rdp->nxtlist = NULL;
1592 for (i = 0; i < RCU_NEXT_SIZE; i++)
1593 rdp->nxttail[i] = &rdp->nxtlist;
1596 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1597 #endif /* #ifdef CONFIG_NO_HZ */
1599 spin_unlock_irqrestore(&rnp->lock, flags);
1603 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1604 * offline event can be happening at a given time. Note also that we
1605 * can accept some slop in the rsp->completed access due to the fact
1606 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1608 static void __cpuinit
1609 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1611 unsigned long flags;
1613 struct rcu_data *rdp = rsp->rda[cpu];
1614 struct rcu_node *rnp = rcu_get_root(rsp);
1616 /* Set up local state, ensuring consistent view of global state. */
1617 spin_lock_irqsave(&rnp->lock, flags);
1618 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1619 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1620 rdp->beenonline = 1; /* We have now been online. */
1621 rdp->preemptable = preemptable;
1622 rdp->qlen_last_fqs_check = 0;
1623 rdp->n_force_qs_snap = rsp->n_force_qs;
1624 rdp->blimit = blimit;
1625 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1628 * A new grace period might start here. If so, we won't be part
1629 * of it, but that is OK, as we are currently in a quiescent state.
1632 /* Exclude any attempts to start a new GP on large systems. */
1633 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1635 /* Add CPU to rcu_node bitmasks. */
1637 mask = rdp->grpmask;
1639 /* Exclude any attempts to start a new GP on small systems. */
1640 spin_lock(&rnp->lock); /* irqs already disabled. */
1641 rnp->qsmaskinit |= mask;
1642 mask = rnp->grpmask;
1643 if (rnp == rdp->mynode) {
1644 rdp->gpnum = rnp->completed; /* if GP in progress... */
1645 rdp->completed = rnp->completed;
1646 rdp->passed_quiesc_completed = rnp->completed - 1;
1648 spin_unlock(&rnp->lock); /* irqs already disabled. */
1650 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1652 spin_unlock_irqrestore(&rsp->onofflock, flags);
1655 static void __cpuinit rcu_online_cpu(int cpu)
1657 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1658 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1659 rcu_preempt_init_percpu_data(cpu);
1663 * Handle CPU online/offline notification events.
1665 int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1666 unsigned long action, void *hcpu)
1668 long cpu = (long)hcpu;
1671 case CPU_UP_PREPARE:
1672 case CPU_UP_PREPARE_FROZEN:
1673 rcu_online_cpu(cpu);
1676 case CPU_DYING_FROZEN:
1678 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1679 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1680 * returns, all online cpus have queued rcu_barrier_func().
1681 * The dying CPU clears its cpu_online_mask bit and
1682 * moves all of its RCU callbacks to ->orphan_cbs_list
1683 * in the context of stop_machine(), so subsequent calls
1684 * to _rcu_barrier() will adopt these callbacks and only
1685 * then queue rcu_barrier_func() on all remaining CPUs.
1687 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1688 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1689 rcu_preempt_send_cbs_to_orphanage();
1692 case CPU_DEAD_FROZEN:
1693 case CPU_UP_CANCELED:
1694 case CPU_UP_CANCELED_FROZEN:
1695 rcu_offline_cpu(cpu);
1704 * Compute the per-level fanout, either using the exact fanout specified
1705 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1707 #ifdef CONFIG_RCU_FANOUT_EXACT
1708 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1712 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1713 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1715 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1716 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1723 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1724 ccur = rsp->levelcnt[i];
1725 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1729 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1732 * Helper function for rcu_init() that initializes one rcu_state structure.
1734 static void __init rcu_init_one(struct rcu_state *rsp)
1739 struct rcu_node *rnp;
1741 /* Initialize the level-tracking arrays. */
1743 for (i = 1; i < NUM_RCU_LVLS; i++)
1744 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1745 rcu_init_levelspread(rsp);
1747 /* Initialize the elements themselves, starting from the leaves. */
1749 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1750 cpustride *= rsp->levelspread[i];
1751 rnp = rsp->level[i];
1752 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1753 spin_lock_init(&rnp->lock);
1756 rnp->qsmaskinit = 0;
1757 rnp->grplo = j * cpustride;
1758 rnp->grphi = (j + 1) * cpustride - 1;
1759 if (rnp->grphi >= NR_CPUS)
1760 rnp->grphi = NR_CPUS - 1;
1766 rnp->grpnum = j % rsp->levelspread[i - 1];
1767 rnp->grpmask = 1UL << rnp->grpnum;
1768 rnp->parent = rsp->level[i - 1] +
1769 j / rsp->levelspread[i - 1];
1772 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1773 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1776 lockdep_set_class(&rcu_get_root(rsp)->lock, &rcu_root_class);
1780 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1781 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1784 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1788 struct rcu_node *rnp; \
1790 rcu_init_one(rsp); \
1791 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1793 for_each_possible_cpu(i) { \
1794 if (i > rnp[j].grphi) \
1796 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1797 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1798 rcu_boot_init_percpu_data(i, rsp); \
1802 void __init __rcu_init(void)
1804 rcu_bootup_announce();
1805 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1806 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1807 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1808 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1809 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1810 __rcu_init_preempt();
1811 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1814 #include "rcutree_plugin.h"