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->completed;
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->completed;
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->dynticks_completed = 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->dynticks_completed;
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.
545 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
548 rdp->passed_quiesc = 0;
549 rdp->gpnum = rsp->gpnum;
553 * Did someone else start a new RCU grace period start since we last
554 * checked? Update local state appropriately if so. Must be called
555 * on the CPU corresponding to rdp.
558 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
563 local_irq_save(flags);
564 if (rdp->gpnum != rsp->gpnum) {
565 note_new_gpnum(rsp, rdp);
568 local_irq_restore(flags);
573 * Advance this CPU's callbacks, but only if the current grace period
574 * has ended. This may be called only from the CPU to whom the rdp
575 * belongs. In addition, the corresponding leaf rcu_node structure's
576 * ->lock must be held by the caller, with irqs disabled.
579 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
581 /* Did another grace period end? */
582 if (rdp->completed != rnp->completed) {
584 /* Advance callbacks. No harm if list empty. */
585 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
586 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
587 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
589 /* Remember that we saw this grace-period completion. */
590 rdp->completed = rnp->completed;
595 * Advance this CPU's callbacks, but only if the current grace period
596 * has ended. This may be called only from the CPU to whom the rdp
600 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
603 struct rcu_node *rnp;
605 local_irq_save(flags);
607 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
608 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
609 local_irq_restore(flags);
612 __rcu_process_gp_end(rsp, rnp, rdp);
613 spin_unlock_irqrestore(&rnp->lock, flags);
617 * Do per-CPU grace-period initialization for running CPU. The caller
618 * must hold the lock of the leaf rcu_node structure corresponding to
622 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
624 /* Prior grace period ended, so advance callbacks for current CPU. */
625 __rcu_process_gp_end(rsp, rnp, rdp);
628 * Because this CPU just now started the new grace period, we know
629 * that all of its callbacks will be covered by this upcoming grace
630 * period, even the ones that were registered arbitrarily recently.
631 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
633 * Other CPUs cannot be sure exactly when the grace period started.
634 * Therefore, their recently registered callbacks must pass through
635 * an additional RCU_NEXT_READY stage, so that they will be handled
636 * by the next RCU grace period.
638 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
639 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
643 * Start a new RCU grace period if warranted, re-initializing the hierarchy
644 * in preparation for detecting the next grace period. The caller must hold
645 * the root node's ->lock, which is released before return. Hard irqs must
649 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
650 __releases(rcu_get_root(rsp)->lock)
652 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
653 struct rcu_node *rnp = rcu_get_root(rsp);
655 if (!cpu_needs_another_gp(rsp, rdp)) {
656 spin_unlock_irqrestore(&rnp->lock, flags);
660 /* Advance to a new grace period and initialize state. */
662 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
663 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
664 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
665 record_gp_stall_check_time(rsp);
666 dyntick_record_completed(rsp, rsp->completed - 1);
667 note_new_gpnum(rsp, rdp);
669 /* Special-case the common single-level case. */
670 if (NUM_RCU_NODES == 1) {
671 rcu_preempt_check_blocked_tasks(rnp);
672 rnp->qsmask = rnp->qsmaskinit;
673 rnp->gpnum = rsp->gpnum;
674 rnp->completed = rsp->completed;
675 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
676 rcu_start_gp_per_cpu(rsp, rnp, rdp);
677 spin_unlock_irqrestore(&rnp->lock, flags);
681 spin_unlock(&rnp->lock); /* leave irqs disabled. */
684 /* Exclude any concurrent CPU-hotplug operations. */
685 spin_lock(&rsp->onofflock); /* irqs already disabled. */
688 * Set the quiescent-state-needed bits in all the rcu_node
689 * structures for all currently online CPUs in breadth-first
690 * order, starting from the root rcu_node structure. This
691 * operation relies on the layout of the hierarchy within the
692 * rsp->node[] array. Note that other CPUs will access only
693 * the leaves of the hierarchy, which still indicate that no
694 * grace period is in progress, at least until the corresponding
695 * leaf node has been initialized. In addition, we have excluded
696 * CPU-hotplug operations.
698 * Note that the grace period cannot complete until we finish
699 * the initialization process, as there will be at least one
700 * qsmask bit set in the root node until that time, namely the
701 * one corresponding to this CPU, due to the fact that we have
704 rcu_for_each_node_breadth_first(rsp, rnp) {
705 spin_lock(&rnp->lock); /* irqs already disabled. */
706 rcu_preempt_check_blocked_tasks(rnp);
707 rnp->qsmask = rnp->qsmaskinit;
708 rnp->gpnum = rsp->gpnum;
709 rnp->completed = rsp->completed;
710 if (rnp == rdp->mynode)
711 rcu_start_gp_per_cpu(rsp, rnp, rdp);
712 spin_unlock(&rnp->lock); /* irqs remain disabled. */
715 rnp = rcu_get_root(rsp);
716 spin_lock(&rnp->lock); /* irqs already disabled. */
717 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
718 spin_unlock(&rnp->lock); /* irqs remain disabled. */
719 spin_unlock_irqrestore(&rsp->onofflock, flags);
723 * Clean up after the prior grace period and let rcu_start_gp() start up
724 * the next grace period if one is needed. Note that the caller must
725 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
727 static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
728 __releases(rcu_get_root(rsp)->lock)
730 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
731 rsp->completed = rsp->gpnum;
732 rsp->signaled = RCU_GP_IDLE;
733 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
737 * Similar to cpu_quiet(), for which it is a helper function. Allows
738 * a group of CPUs to be quieted at one go, though all the CPUs in the
739 * group must be represented by the same leaf rcu_node structure.
740 * That structure's lock must be held upon entry, and it is released
744 cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp,
746 __releases(rnp->lock)
748 struct rcu_node *rnp_c;
750 /* Walk up the rcu_node hierarchy. */
752 if (!(rnp->qsmask & mask)) {
754 /* Our bit has already been cleared, so done. */
755 spin_unlock_irqrestore(&rnp->lock, flags);
758 rnp->qsmask &= ~mask;
759 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
761 /* Other bits still set at this level, so done. */
762 spin_unlock_irqrestore(&rnp->lock, flags);
766 if (rnp->parent == NULL) {
768 /* No more levels. Exit loop holding root lock. */
772 spin_unlock_irqrestore(&rnp->lock, flags);
775 spin_lock_irqsave(&rnp->lock, flags);
776 WARN_ON_ONCE(rnp_c->qsmask);
780 * Get here if we are the last CPU to pass through a quiescent
781 * state for this grace period. Invoke cpu_quiet_msk_finish()
782 * to clean up and start the next grace period if one is needed.
784 cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */
788 * Record a quiescent state for the specified CPU, which must either be
789 * the current CPU. The lastcomp argument is used to make sure we are
790 * still in the grace period of interest. We don't want to end the current
791 * grace period based on quiescent states detected in an earlier grace
795 cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
799 struct rcu_node *rnp;
802 spin_lock_irqsave(&rnp->lock, flags);
803 if (lastcomp != ACCESS_ONCE(rsp->completed)) {
806 * Someone beat us to it for this grace period, so leave.
807 * The race with GP start is resolved by the fact that we
808 * hold the leaf rcu_node lock, so that the per-CPU bits
809 * cannot yet be initialized -- so we would simply find our
810 * CPU's bit already cleared in cpu_quiet_msk() if this race
813 rdp->passed_quiesc = 0; /* try again later! */
814 spin_unlock_irqrestore(&rnp->lock, flags);
818 if ((rnp->qsmask & mask) == 0) {
819 spin_unlock_irqrestore(&rnp->lock, flags);
824 * This GP can't end until cpu checks in, so all of our
825 * callbacks can be processed during the next GP.
827 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
829 cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */
834 * Check to see if there is a new grace period of which this CPU
835 * is not yet aware, and if so, set up local rcu_data state for it.
836 * Otherwise, see if this CPU has just passed through its first
837 * quiescent state for this grace period, and record that fact if so.
840 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
842 /* If there is now a new grace period, record and return. */
843 if (check_for_new_grace_period(rsp, rdp))
847 * Does this CPU still need to do its part for current grace period?
848 * If no, return and let the other CPUs do their part as well.
850 if (!rdp->qs_pending)
854 * Was there a quiescent state since the beginning of the grace
855 * period? If no, then exit and wait for the next call.
857 if (!rdp->passed_quiesc)
860 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
861 cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
864 #ifdef CONFIG_HOTPLUG_CPU
867 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
868 * specified flavor of RCU. The callbacks will be adopted by the next
869 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
870 * comes first. Because this is invoked from the CPU_DYING notifier,
871 * irqs are already disabled.
873 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
876 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
878 if (rdp->nxtlist == NULL)
879 return; /* irqs disabled, so comparison is stable. */
880 spin_lock(&rsp->onofflock); /* irqs already disabled. */
881 *rsp->orphan_cbs_tail = rdp->nxtlist;
882 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
884 for (i = 0; i < RCU_NEXT_SIZE; i++)
885 rdp->nxttail[i] = &rdp->nxtlist;
886 rsp->orphan_qlen += rdp->qlen;
888 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
892 * Adopt previously orphaned RCU callbacks.
894 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
897 struct rcu_data *rdp;
899 spin_lock_irqsave(&rsp->onofflock, flags);
900 rdp = rsp->rda[smp_processor_id()];
901 if (rsp->orphan_cbs_list == NULL) {
902 spin_unlock_irqrestore(&rsp->onofflock, flags);
905 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
906 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
907 rdp->qlen += rsp->orphan_qlen;
908 rsp->orphan_cbs_list = NULL;
909 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
910 rsp->orphan_qlen = 0;
911 spin_unlock_irqrestore(&rsp->onofflock, flags);
915 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
916 * and move all callbacks from the outgoing CPU to the current one.
918 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
923 struct rcu_data *rdp = rsp->rda[cpu];
924 struct rcu_node *rnp;
926 /* Exclude any attempts to start a new grace period. */
927 spin_lock_irqsave(&rsp->onofflock, flags);
929 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
930 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
931 mask = rdp->grpmask; /* rnp->grplo is constant. */
933 spin_lock(&rnp->lock); /* irqs already disabled. */
934 rnp->qsmaskinit &= ~mask;
935 if (rnp->qsmaskinit != 0) {
936 spin_unlock(&rnp->lock); /* irqs remain disabled. */
941 * If there was a task blocking the current grace period,
942 * and if all CPUs have checked in, we need to propagate
943 * the quiescent state up the rcu_node hierarchy. But that
944 * is inconvenient at the moment due to deadlock issues if
945 * this should end the current grace period. So set the
946 * offlined CPU's bit in ->qsmask in order to force the
947 * next force_quiescent_state() invocation to clean up this
948 * mess in a deadlock-free manner.
950 if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
954 spin_unlock(&rnp->lock); /* irqs remain disabled. */
956 } while (rnp != NULL);
957 lastcomp = rsp->completed;
959 spin_unlock_irqrestore(&rsp->onofflock, flags);
961 rcu_adopt_orphan_cbs(rsp);
965 * Remove the specified CPU from the RCU hierarchy and move any pending
966 * callbacks that it might have to the current CPU. This code assumes
967 * that at least one CPU in the system will remain running at all times.
968 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
970 static void rcu_offline_cpu(int cpu)
972 __rcu_offline_cpu(cpu, &rcu_sched_state);
973 __rcu_offline_cpu(cpu, &rcu_bh_state);
974 rcu_preempt_offline_cpu(cpu);
977 #else /* #ifdef CONFIG_HOTPLUG_CPU */
979 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
983 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
987 static void rcu_offline_cpu(int cpu)
991 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
994 * Invoke any RCU callbacks that have made it to the end of their grace
995 * period. Thottle as specified by rdp->blimit.
997 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1000 struct rcu_head *next, *list, **tail;
1003 /* If no callbacks are ready, just return.*/
1004 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1008 * Extract the list of ready callbacks, disabling to prevent
1009 * races with call_rcu() from interrupt handlers.
1011 local_irq_save(flags);
1012 list = rdp->nxtlist;
1013 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1014 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1015 tail = rdp->nxttail[RCU_DONE_TAIL];
1016 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1017 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1018 rdp->nxttail[count] = &rdp->nxtlist;
1019 local_irq_restore(flags);
1021 /* Invoke callbacks. */
1028 if (++count >= rdp->blimit)
1032 local_irq_save(flags);
1034 /* Update count, and requeue any remaining callbacks. */
1037 *tail = rdp->nxtlist;
1038 rdp->nxtlist = list;
1039 for (count = 0; count < RCU_NEXT_SIZE; count++)
1040 if (&rdp->nxtlist == rdp->nxttail[count])
1041 rdp->nxttail[count] = tail;
1046 /* Reinstate batch limit if we have worked down the excess. */
1047 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1048 rdp->blimit = blimit;
1050 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1051 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1052 rdp->qlen_last_fqs_check = 0;
1053 rdp->n_force_qs_snap = rsp->n_force_qs;
1054 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1055 rdp->qlen_last_fqs_check = rdp->qlen;
1057 local_irq_restore(flags);
1059 /* Re-raise the RCU softirq if there are callbacks remaining. */
1060 if (cpu_has_callbacks_ready_to_invoke(rdp))
1061 raise_softirq(RCU_SOFTIRQ);
1065 * Check to see if this CPU is in a non-context-switch quiescent state
1066 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1067 * Also schedule the RCU softirq handler.
1069 * This function must be called with hardirqs disabled. It is normally
1070 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1071 * false, there is no point in invoking rcu_check_callbacks().
1073 void rcu_check_callbacks(int cpu, int user)
1075 if (!rcu_pending(cpu))
1076 return; /* if nothing for RCU to do. */
1078 (idle_cpu(cpu) && rcu_scheduler_active &&
1079 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1082 * Get here if this CPU took its interrupt from user
1083 * mode or from the idle loop, and if this is not a
1084 * nested interrupt. In this case, the CPU is in
1085 * a quiescent state, so note it.
1087 * No memory barrier is required here because both
1088 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1089 * variables that other CPUs neither access nor modify,
1090 * at least not while the corresponding CPU is online.
1096 } else if (!in_softirq()) {
1099 * Get here if this CPU did not take its interrupt from
1100 * softirq, in other words, if it is not interrupting
1101 * a rcu_bh read-side critical section. This is an _bh
1102 * critical section, so note it.
1107 rcu_preempt_check_callbacks(cpu);
1108 raise_softirq(RCU_SOFTIRQ);
1114 * Scan the leaf rcu_node structures, processing dyntick state for any that
1115 * have not yet encountered a quiescent state, using the function specified.
1116 * Returns 1 if the current grace period ends while scanning (possibly
1117 * because we made it end).
1119 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1120 int (*f)(struct rcu_data *))
1124 unsigned long flags;
1126 struct rcu_node *rnp;
1128 rcu_for_each_leaf_node(rsp, rnp) {
1130 spin_lock_irqsave(&rnp->lock, flags);
1131 if (rsp->completed != lastcomp) {
1132 spin_unlock_irqrestore(&rnp->lock, flags);
1135 if (rnp->qsmask == 0) {
1136 spin_unlock_irqrestore(&rnp->lock, flags);
1141 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1142 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1145 if (mask != 0 && rsp->completed == lastcomp) {
1147 /* cpu_quiet_msk() releases rnp->lock. */
1148 cpu_quiet_msk(mask, rsp, rnp, flags);
1151 spin_unlock_irqrestore(&rnp->lock, flags);
1157 * Force quiescent states on reluctant CPUs, and also detect which
1158 * CPUs are in dyntick-idle mode.
1160 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1162 unsigned long flags;
1164 struct rcu_node *rnp = rcu_get_root(rsp);
1168 if (!rcu_gp_in_progress(rsp))
1169 return; /* No grace period in progress, nothing to force. */
1170 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1171 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1172 return; /* Someone else is already on the job. */
1175 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1176 goto unlock_ret; /* no emergency and done recently. */
1178 spin_lock(&rnp->lock);
1179 lastcomp = rsp->completed;
1180 signaled = rsp->signaled;
1181 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1182 if (lastcomp == rsp->gpnum) {
1183 rsp->n_force_qs_ngp++;
1184 spin_unlock(&rnp->lock);
1185 goto unlock_ret; /* no GP in progress, time updated. */
1187 spin_unlock(&rnp->lock);
1192 break; /* grace period idle or initializing, ignore. */
1194 case RCU_SAVE_DYNTICK:
1196 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1197 break; /* So gcc recognizes the dead code. */
1199 /* Record dyntick-idle state. */
1200 if (rcu_process_dyntick(rsp, lastcomp,
1201 dyntick_save_progress_counter))
1203 /* fall into next case. */
1205 case RCU_SAVE_COMPLETED:
1207 /* Update state, record completion counter. */
1209 spin_lock(&rnp->lock);
1210 if (lastcomp == rsp->completed &&
1211 rsp->signaled == signaled) {
1212 rsp->signaled = RCU_FORCE_QS;
1213 dyntick_record_completed(rsp, lastcomp);
1214 forcenow = signaled == RCU_SAVE_COMPLETED;
1216 spin_unlock(&rnp->lock);
1219 /* fall into next case. */
1223 /* Check dyntick-idle state, send IPI to laggarts. */
1224 if (rcu_process_dyntick(rsp, dyntick_recall_completed(rsp),
1225 rcu_implicit_dynticks_qs))
1228 /* Leave state in case more forcing is required. */
1233 spin_unlock_irqrestore(&rsp->fqslock, flags);
1236 #else /* #ifdef CONFIG_SMP */
1238 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1243 #endif /* #else #ifdef CONFIG_SMP */
1246 * This does the RCU processing work from softirq context for the
1247 * specified rcu_state and rcu_data structures. This may be called
1248 * only from the CPU to whom the rdp belongs.
1251 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1253 unsigned long flags;
1255 WARN_ON_ONCE(rdp->beenonline == 0);
1258 * If an RCU GP has gone long enough, go check for dyntick
1259 * idle CPUs and, if needed, send resched IPIs.
1261 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1262 force_quiescent_state(rsp, 1);
1265 * Advance callbacks in response to end of earlier grace
1266 * period that some other CPU ended.
1268 rcu_process_gp_end(rsp, rdp);
1270 /* Update RCU state based on any recent quiescent states. */
1271 rcu_check_quiescent_state(rsp, rdp);
1273 /* Does this CPU require a not-yet-started grace period? */
1274 if (cpu_needs_another_gp(rsp, rdp)) {
1275 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1276 rcu_start_gp(rsp, flags); /* releases above lock */
1279 /* If there are callbacks ready, invoke them. */
1280 rcu_do_batch(rsp, rdp);
1284 * Do softirq processing for the current CPU.
1286 static void rcu_process_callbacks(struct softirq_action *unused)
1289 * Memory references from any prior RCU read-side critical sections
1290 * executed by the interrupted code must be seen before any RCU
1291 * grace-period manipulations below.
1293 smp_mb(); /* See above block comment. */
1295 __rcu_process_callbacks(&rcu_sched_state,
1296 &__get_cpu_var(rcu_sched_data));
1297 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1298 rcu_preempt_process_callbacks();
1301 * Memory references from any later RCU read-side critical sections
1302 * executed by the interrupted code must be seen after any RCU
1303 * grace-period manipulations above.
1305 smp_mb(); /* See above block comment. */
1309 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1310 struct rcu_state *rsp)
1312 unsigned long flags;
1313 struct rcu_data *rdp;
1318 smp_mb(); /* Ensure RCU update seen before callback registry. */
1321 * Opportunistically note grace-period endings and beginnings.
1322 * Note that we might see a beginning right after we see an
1323 * end, but never vice versa, since this CPU has to pass through
1324 * a quiescent state betweentimes.
1326 local_irq_save(flags);
1327 rdp = rsp->rda[smp_processor_id()];
1328 rcu_process_gp_end(rsp, rdp);
1329 check_for_new_grace_period(rsp, rdp);
1331 /* Add the callback to our list. */
1332 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1333 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1335 /* Start a new grace period if one not already started. */
1336 if (!rcu_gp_in_progress(rsp)) {
1337 unsigned long nestflag;
1338 struct rcu_node *rnp_root = rcu_get_root(rsp);
1340 spin_lock_irqsave(&rnp_root->lock, nestflag);
1341 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1345 * Force the grace period if too many callbacks or too long waiting.
1346 * Enforce hysteresis, and don't invoke force_quiescent_state()
1347 * if some other CPU has recently done so. Also, don't bother
1348 * invoking force_quiescent_state() if the newly enqueued callback
1349 * is the only one waiting for a grace period to complete.
1351 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1352 rdp->blimit = LONG_MAX;
1353 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1354 *rdp->nxttail[RCU_DONE_TAIL] != head)
1355 force_quiescent_state(rsp, 0);
1356 rdp->n_force_qs_snap = rsp->n_force_qs;
1357 rdp->qlen_last_fqs_check = rdp->qlen;
1358 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1359 force_quiescent_state(rsp, 1);
1360 local_irq_restore(flags);
1364 * Queue an RCU-sched callback for invocation after a grace period.
1366 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1368 __call_rcu(head, func, &rcu_sched_state);
1370 EXPORT_SYMBOL_GPL(call_rcu_sched);
1373 * Queue an RCU for invocation after a quicker grace period.
1375 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1377 __call_rcu(head, func, &rcu_bh_state);
1379 EXPORT_SYMBOL_GPL(call_rcu_bh);
1382 * Check to see if there is any immediate RCU-related work to be done
1383 * by the current CPU, for the specified type of RCU, returning 1 if so.
1384 * The checks are in order of increasing expense: checks that can be
1385 * carried out against CPU-local state are performed first. However,
1386 * we must check for CPU stalls first, else we might not get a chance.
1388 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1390 rdp->n_rcu_pending++;
1392 /* Check for CPU stalls, if enabled. */
1393 check_cpu_stall(rsp, rdp);
1395 /* Is the RCU core waiting for a quiescent state from this CPU? */
1396 if (rdp->qs_pending) {
1397 rdp->n_rp_qs_pending++;
1401 /* Does this CPU have callbacks ready to invoke? */
1402 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1403 rdp->n_rp_cb_ready++;
1407 /* Has RCU gone idle with this CPU needing another grace period? */
1408 if (cpu_needs_another_gp(rsp, rdp)) {
1409 rdp->n_rp_cpu_needs_gp++;
1413 /* Has another RCU grace period completed? */
1414 if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */
1415 rdp->n_rp_gp_completed++;
1419 /* Has a new RCU grace period started? */
1420 if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */
1421 rdp->n_rp_gp_started++;
1425 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1426 if (rcu_gp_in_progress(rsp) &&
1427 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1428 rdp->n_rp_need_fqs++;
1433 rdp->n_rp_need_nothing++;
1438 * Check to see if there is any immediate RCU-related work to be done
1439 * by the current CPU, returning 1 if so. This function is part of the
1440 * RCU implementation; it is -not- an exported member of the RCU API.
1442 static int rcu_pending(int cpu)
1444 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1445 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1446 rcu_preempt_pending(cpu);
1450 * Check to see if any future RCU-related work will need to be done
1451 * by the current CPU, even if none need be done immediately, returning
1452 * 1 if so. This function is part of the RCU implementation; it is -not-
1453 * an exported member of the RCU API.
1455 int rcu_needs_cpu(int cpu)
1457 /* RCU callbacks either ready or pending? */
1458 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1459 per_cpu(rcu_bh_data, cpu).nxtlist ||
1460 rcu_preempt_needs_cpu(cpu);
1463 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1464 static atomic_t rcu_barrier_cpu_count;
1465 static DEFINE_MUTEX(rcu_barrier_mutex);
1466 static struct completion rcu_barrier_completion;
1468 static void rcu_barrier_callback(struct rcu_head *notused)
1470 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1471 complete(&rcu_barrier_completion);
1475 * Called with preemption disabled, and from cross-cpu IRQ context.
1477 static void rcu_barrier_func(void *type)
1479 int cpu = smp_processor_id();
1480 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1481 void (*call_rcu_func)(struct rcu_head *head,
1482 void (*func)(struct rcu_head *head));
1484 atomic_inc(&rcu_barrier_cpu_count);
1485 call_rcu_func = type;
1486 call_rcu_func(head, rcu_barrier_callback);
1490 * Orchestrate the specified type of RCU barrier, waiting for all
1491 * RCU callbacks of the specified type to complete.
1493 static void _rcu_barrier(struct rcu_state *rsp,
1494 void (*call_rcu_func)(struct rcu_head *head,
1495 void (*func)(struct rcu_head *head)))
1497 BUG_ON(in_interrupt());
1498 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1499 mutex_lock(&rcu_barrier_mutex);
1500 init_completion(&rcu_barrier_completion);
1502 * Initialize rcu_barrier_cpu_count to 1, then invoke
1503 * rcu_barrier_func() on each CPU, so that each CPU also has
1504 * incremented rcu_barrier_cpu_count. Only then is it safe to
1505 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1506 * might complete its grace period before all of the other CPUs
1507 * did their increment, causing this function to return too
1510 atomic_set(&rcu_barrier_cpu_count, 1);
1511 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1512 rcu_adopt_orphan_cbs(rsp);
1513 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1514 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1515 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1516 complete(&rcu_barrier_completion);
1517 wait_for_completion(&rcu_barrier_completion);
1518 mutex_unlock(&rcu_barrier_mutex);
1522 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1524 void rcu_barrier_bh(void)
1526 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1528 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1531 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1533 void rcu_barrier_sched(void)
1535 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1537 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1540 * Do boot-time initialization of a CPU's per-CPU RCU data.
1543 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1545 unsigned long flags;
1547 struct rcu_data *rdp = rsp->rda[cpu];
1548 struct rcu_node *rnp = rcu_get_root(rsp);
1550 /* Set up local state, ensuring consistent view of global state. */
1551 spin_lock_irqsave(&rnp->lock, flags);
1552 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1553 rdp->nxtlist = NULL;
1554 for (i = 0; i < RCU_NEXT_SIZE; i++)
1555 rdp->nxttail[i] = &rdp->nxtlist;
1558 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1559 #endif /* #ifdef CONFIG_NO_HZ */
1561 spin_unlock_irqrestore(&rnp->lock, flags);
1565 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1566 * offline event can be happening at a given time. Note also that we
1567 * can accept some slop in the rsp->completed access due to the fact
1568 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1570 static void __cpuinit
1571 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1573 unsigned long flags;
1575 struct rcu_data *rdp = rsp->rda[cpu];
1576 struct rcu_node *rnp = rcu_get_root(rsp);
1578 /* Set up local state, ensuring consistent view of global state. */
1579 spin_lock_irqsave(&rnp->lock, flags);
1580 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1581 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1582 rdp->beenonline = 1; /* We have now been online. */
1583 rdp->preemptable = preemptable;
1584 rdp->qlen_last_fqs_check = 0;
1585 rdp->n_force_qs_snap = rsp->n_force_qs;
1586 rdp->blimit = blimit;
1587 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1590 * A new grace period might start here. If so, we won't be part
1591 * of it, but that is OK, as we are currently in a quiescent state.
1594 /* Exclude any attempts to start a new GP on large systems. */
1595 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1597 /* Add CPU to rcu_node bitmasks. */
1599 mask = rdp->grpmask;
1601 /* Exclude any attempts to start a new GP on small systems. */
1602 spin_lock(&rnp->lock); /* irqs already disabled. */
1603 rnp->qsmaskinit |= mask;
1604 mask = rnp->grpmask;
1605 if (rnp == rdp->mynode) {
1606 rdp->gpnum = rnp->completed; /* if GP in progress... */
1607 rdp->completed = rnp->completed;
1608 rdp->passed_quiesc_completed = rnp->completed - 1;
1610 spin_unlock(&rnp->lock); /* irqs already disabled. */
1612 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1614 spin_unlock_irqrestore(&rsp->onofflock, flags);
1617 static void __cpuinit rcu_online_cpu(int cpu)
1619 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1620 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1621 rcu_preempt_init_percpu_data(cpu);
1625 * Handle CPU online/offline notification events.
1627 int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1628 unsigned long action, void *hcpu)
1630 long cpu = (long)hcpu;
1633 case CPU_UP_PREPARE:
1634 case CPU_UP_PREPARE_FROZEN:
1635 rcu_online_cpu(cpu);
1638 case CPU_DYING_FROZEN:
1640 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1641 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1642 * returns, all online cpus have queued rcu_barrier_func().
1643 * The dying CPU clears its cpu_online_mask bit and
1644 * moves all of its RCU callbacks to ->orphan_cbs_list
1645 * in the context of stop_machine(), so subsequent calls
1646 * to _rcu_barrier() will adopt these callbacks and only
1647 * then queue rcu_barrier_func() on all remaining CPUs.
1649 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1650 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1651 rcu_preempt_send_cbs_to_orphanage();
1654 case CPU_DEAD_FROZEN:
1655 case CPU_UP_CANCELED:
1656 case CPU_UP_CANCELED_FROZEN:
1657 rcu_offline_cpu(cpu);
1666 * Compute the per-level fanout, either using the exact fanout specified
1667 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1669 #ifdef CONFIG_RCU_FANOUT_EXACT
1670 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1674 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1675 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1677 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1678 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1685 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1686 ccur = rsp->levelcnt[i];
1687 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1691 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1694 * Helper function for rcu_init() that initializes one rcu_state structure.
1696 static void __init rcu_init_one(struct rcu_state *rsp)
1701 struct rcu_node *rnp;
1703 /* Initialize the level-tracking arrays. */
1705 for (i = 1; i < NUM_RCU_LVLS; i++)
1706 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1707 rcu_init_levelspread(rsp);
1709 /* Initialize the elements themselves, starting from the leaves. */
1711 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1712 cpustride *= rsp->levelspread[i];
1713 rnp = rsp->level[i];
1714 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1715 spin_lock_init(&rnp->lock);
1718 rnp->qsmaskinit = 0;
1719 rnp->grplo = j * cpustride;
1720 rnp->grphi = (j + 1) * cpustride - 1;
1721 if (rnp->grphi >= NR_CPUS)
1722 rnp->grphi = NR_CPUS - 1;
1728 rnp->grpnum = j % rsp->levelspread[i - 1];
1729 rnp->grpmask = 1UL << rnp->grpnum;
1730 rnp->parent = rsp->level[i - 1] +
1731 j / rsp->levelspread[i - 1];
1734 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1735 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1738 lockdep_set_class(&rcu_get_root(rsp)->lock, &rcu_root_class);
1742 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1743 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1746 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1750 struct rcu_node *rnp; \
1752 rcu_init_one(rsp); \
1753 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1755 for_each_possible_cpu(i) { \
1756 if (i > rnp[j].grphi) \
1758 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1759 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1760 rcu_boot_init_percpu_data(i, rsp); \
1764 void __init __rcu_init(void)
1766 rcu_bootup_announce();
1767 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1768 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1769 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1770 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1771 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1772 __rcu_init_preempt();
1773 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1776 #include "rcutree_plugin.h"