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];
184 * If the specified CPU is offline, tell the caller that it is in
185 * a quiescent state. Otherwise, whack it with a reschedule IPI.
186 * Grace periods can end up waiting on an offline CPU when that
187 * CPU is in the process of coming online -- it will be added to the
188 * rcu_node bitmasks before it actually makes it online. The same thing
189 * can happen while a CPU is in the process of coming online. Because this
190 * race is quite rare, we check for it after detecting that the grace
191 * period has been delayed rather than checking each and every CPU
192 * each and every time we start a new grace period.
194 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
197 * If the CPU is offline, it is in a quiescent state. We can
198 * trust its state not to change because interrupts are disabled.
200 if (cpu_is_offline(rdp->cpu)) {
205 /* If preemptable RCU, no point in sending reschedule IPI. */
206 if (rdp->preemptable)
209 /* The CPU is online, so send it a reschedule IPI. */
210 if (rdp->cpu != smp_processor_id())
211 smp_send_reschedule(rdp->cpu);
218 #endif /* #ifdef CONFIG_SMP */
223 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
225 * Enter nohz mode, in other words, -leave- the mode in which RCU
226 * read-side critical sections can occur. (Though RCU read-side
227 * critical sections can occur in irq handlers in nohz mode, a possibility
228 * handled by rcu_irq_enter() and rcu_irq_exit()).
230 void rcu_enter_nohz(void)
233 struct rcu_dynticks *rdtp;
235 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
236 local_irq_save(flags);
237 rdtp = &__get_cpu_var(rcu_dynticks);
239 rdtp->dynticks_nesting--;
240 WARN_ON_ONCE(rdtp->dynticks & 0x1);
241 local_irq_restore(flags);
245 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
247 * Exit nohz mode, in other words, -enter- the mode in which RCU
248 * read-side critical sections normally occur.
250 void rcu_exit_nohz(void)
253 struct rcu_dynticks *rdtp;
255 local_irq_save(flags);
256 rdtp = &__get_cpu_var(rcu_dynticks);
258 rdtp->dynticks_nesting++;
259 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
260 local_irq_restore(flags);
261 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
265 * rcu_nmi_enter - inform RCU of entry to NMI context
267 * If the CPU was idle with dynamic ticks active, and there is no
268 * irq handler running, this updates rdtp->dynticks_nmi to let the
269 * RCU grace-period handling know that the CPU is active.
271 void rcu_nmi_enter(void)
273 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
275 if (rdtp->dynticks & 0x1)
277 rdtp->dynticks_nmi++;
278 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
279 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
283 * rcu_nmi_exit - inform RCU of exit from NMI context
285 * If the CPU was idle with dynamic ticks active, and there is no
286 * irq handler running, this updates rdtp->dynticks_nmi to let the
287 * RCU grace-period handling know that the CPU is no longer active.
289 void rcu_nmi_exit(void)
291 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
293 if (rdtp->dynticks & 0x1)
295 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
296 rdtp->dynticks_nmi++;
297 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
301 * rcu_irq_enter - inform RCU of entry to hard irq context
303 * If the CPU was idle with dynamic ticks active, this updates the
304 * rdtp->dynticks to let the RCU handling know that the CPU is active.
306 void rcu_irq_enter(void)
308 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
310 if (rdtp->dynticks_nesting++)
313 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
314 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
318 * rcu_irq_exit - inform RCU of exit from hard irq context
320 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
321 * to put let the RCU handling be aware that the CPU is going back to idle
324 void rcu_irq_exit(void)
326 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
328 if (--rdtp->dynticks_nesting)
330 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
332 WARN_ON_ONCE(rdtp->dynticks & 0x1);
334 /* If the interrupt queued a callback, get out of dyntick mode. */
335 if (__get_cpu_var(rcu_sched_data).nxtlist ||
336 __get_cpu_var(rcu_bh_data).nxtlist)
341 * Record the specified "completed" value, which is later used to validate
342 * dynticks counter manipulations. Specify "rsp->completed - 1" to
343 * unconditionally invalidate any future dynticks manipulations (which is
344 * useful at the beginning of a grace period).
346 static void dyntick_record_completed(struct rcu_state *rsp, long comp)
348 rsp->dynticks_completed = comp;
354 * Recall the previously recorded value of the completion for dynticks.
356 static long dyntick_recall_completed(struct rcu_state *rsp)
358 return rsp->dynticks_completed;
362 * Snapshot the specified CPU's dynticks counter so that we can later
363 * credit them with an implicit quiescent state. Return 1 if this CPU
364 * is in dynticks idle mode, which is an extended quiescent state.
366 static int dyntick_save_progress_counter(struct rcu_data *rdp)
372 snap = rdp->dynticks->dynticks;
373 snap_nmi = rdp->dynticks->dynticks_nmi;
374 smp_mb(); /* Order sampling of snap with end of grace period. */
375 rdp->dynticks_snap = snap;
376 rdp->dynticks_nmi_snap = snap_nmi;
377 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
384 * Return true if the specified CPU has passed through a quiescent
385 * state by virtue of being in or having passed through an dynticks
386 * idle state since the last call to dyntick_save_progress_counter()
389 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
396 curr = rdp->dynticks->dynticks;
397 snap = rdp->dynticks_snap;
398 curr_nmi = rdp->dynticks->dynticks_nmi;
399 snap_nmi = rdp->dynticks_nmi_snap;
400 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
403 * If the CPU passed through or entered a dynticks idle phase with
404 * no active irq/NMI handlers, then we can safely pretend that the CPU
405 * already acknowledged the request to pass through a quiescent
406 * state. Either way, that CPU cannot possibly be in an RCU
407 * read-side critical section that started before the beginning
408 * of the current RCU grace period.
410 if ((curr != snap || (curr & 0x1) == 0) &&
411 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
416 /* Go check for the CPU being offline. */
417 return rcu_implicit_offline_qs(rdp);
420 #endif /* #ifdef CONFIG_SMP */
422 #else /* #ifdef CONFIG_NO_HZ */
424 static void dyntick_record_completed(struct rcu_state *rsp, long comp)
431 * If there are no dynticks, then the only way that a CPU can passively
432 * be in a quiescent state is to be offline. Unlike dynticks idle, which
433 * is a point in time during the prior (already finished) grace period,
434 * an offline CPU is always in a quiescent state, and thus can be
435 * unconditionally applied. So just return the current value of completed.
437 static long dyntick_recall_completed(struct rcu_state *rsp)
439 return rsp->completed;
442 static int dyntick_save_progress_counter(struct rcu_data *rdp)
447 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
449 return rcu_implicit_offline_qs(rdp);
452 #endif /* #ifdef CONFIG_SMP */
454 #endif /* #else #ifdef CONFIG_NO_HZ */
456 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
458 static void record_gp_stall_check_time(struct rcu_state *rsp)
460 rsp->gp_start = jiffies;
461 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
464 static void print_other_cpu_stall(struct rcu_state *rsp)
469 struct rcu_node *rnp = rcu_get_root(rsp);
471 /* Only let one CPU complain about others per time interval. */
473 spin_lock_irqsave(&rnp->lock, flags);
474 delta = jiffies - rsp->jiffies_stall;
475 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
476 spin_unlock_irqrestore(&rnp->lock, flags);
479 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
482 * Now rat on any tasks that got kicked up to the root rcu_node
483 * due to CPU offlining.
485 rcu_print_task_stall(rnp);
486 spin_unlock_irqrestore(&rnp->lock, flags);
488 /* OK, time to rat on our buddy... */
490 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
491 rcu_for_each_leaf_node(rsp, rnp) {
492 rcu_print_task_stall(rnp);
493 if (rnp->qsmask == 0)
495 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
496 if (rnp->qsmask & (1UL << cpu))
497 printk(" %d", rnp->grplo + cpu);
499 printk(" (detected by %d, t=%ld jiffies)\n",
500 smp_processor_id(), (long)(jiffies - rsp->gp_start));
501 trigger_all_cpu_backtrace();
503 force_quiescent_state(rsp, 0); /* Kick them all. */
506 static void print_cpu_stall(struct rcu_state *rsp)
509 struct rcu_node *rnp = rcu_get_root(rsp);
511 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
512 smp_processor_id(), jiffies - rsp->gp_start);
513 trigger_all_cpu_backtrace();
515 spin_lock_irqsave(&rnp->lock, flags);
516 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
518 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
519 spin_unlock_irqrestore(&rnp->lock, flags);
521 set_need_resched(); /* kick ourselves to get things going. */
524 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
527 struct rcu_node *rnp;
529 delta = jiffies - rsp->jiffies_stall;
531 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
533 /* We haven't checked in, so go dump stack. */
534 print_cpu_stall(rsp);
536 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
538 /* They had two time units to dump stack, so complain. */
539 print_other_cpu_stall(rsp);
543 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
545 static void record_gp_stall_check_time(struct rcu_state *rsp)
549 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
553 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
556 * Update CPU-local rcu_data state to record the newly noticed grace period.
557 * This is used both when we started the grace period and when we notice
558 * that someone else started the grace period.
560 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
563 rdp->passed_quiesc = 0;
564 rdp->gpnum = rsp->gpnum;
568 * Did someone else start a new RCU grace period start since we last
569 * checked? Update local state appropriately if so. Must be called
570 * on the CPU corresponding to rdp.
573 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
578 local_irq_save(flags);
579 if (rdp->gpnum != rsp->gpnum) {
580 note_new_gpnum(rsp, rdp);
583 local_irq_restore(flags);
588 * Start a new RCU grace period if warranted, re-initializing the hierarchy
589 * in preparation for detecting the next grace period. The caller must hold
590 * the root node's ->lock, which is released before return. Hard irqs must
594 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
595 __releases(rcu_get_root(rsp)->lock)
597 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
598 struct rcu_node *rnp = rcu_get_root(rsp);
600 if (!cpu_needs_another_gp(rsp, rdp)) {
601 spin_unlock_irqrestore(&rnp->lock, flags);
605 /* Advance to a new grace period and initialize state. */
607 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
608 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
609 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
610 record_gp_stall_check_time(rsp);
611 dyntick_record_completed(rsp, rsp->completed - 1);
612 note_new_gpnum(rsp, rdp);
615 * Because this CPU just now started the new grace period, we know
616 * that all of its callbacks will be covered by this upcoming grace
617 * period, even the ones that were registered arbitrarily recently.
618 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
620 * Other CPUs cannot be sure exactly when the grace period started.
621 * Therefore, their recently registered callbacks must pass through
622 * an additional RCU_NEXT_READY stage, so that they will be handled
623 * by the next RCU grace period.
625 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
626 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
628 /* Special-case the common single-level case. */
629 if (NUM_RCU_NODES == 1) {
630 rcu_preempt_check_blocked_tasks(rnp);
631 rnp->qsmask = rnp->qsmaskinit;
632 rnp->gpnum = rsp->gpnum;
633 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
634 spin_unlock_irqrestore(&rnp->lock, flags);
638 spin_unlock(&rnp->lock); /* leave irqs disabled. */
641 /* Exclude any concurrent CPU-hotplug operations. */
642 spin_lock(&rsp->onofflock); /* irqs already disabled. */
645 * Set the quiescent-state-needed bits in all the rcu_node
646 * structures for all currently online CPUs in breadth-first
647 * order, starting from the root rcu_node structure. This
648 * operation relies on the layout of the hierarchy within the
649 * rsp->node[] array. Note that other CPUs will access only
650 * the leaves of the hierarchy, which still indicate that no
651 * grace period is in progress, at least until the corresponding
652 * leaf node has been initialized. In addition, we have excluded
653 * CPU-hotplug operations.
655 * Note that the grace period cannot complete until we finish
656 * the initialization process, as there will be at least one
657 * qsmask bit set in the root node until that time, namely the
658 * one corresponding to this CPU, due to the fact that we have
661 rcu_for_each_node_breadth_first(rsp, rnp) {
662 spin_lock(&rnp->lock); /* irqs already disabled. */
663 rcu_preempt_check_blocked_tasks(rnp);
664 rnp->qsmask = rnp->qsmaskinit;
665 rnp->gpnum = rsp->gpnum;
666 spin_unlock(&rnp->lock); /* irqs remain disabled. */
669 rnp = rcu_get_root(rsp);
670 spin_lock(&rnp->lock); /* irqs already disabled. */
671 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
672 spin_unlock(&rnp->lock); /* irqs remain disabled. */
673 spin_unlock_irqrestore(&rsp->onofflock, flags);
677 * Advance this CPU's callbacks, but only if the current grace period
678 * has ended. This may be called only from the CPU to whom the rdp
682 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
687 local_irq_save(flags);
688 completed_snap = ACCESS_ONCE(rsp->completed); /* outside of lock. */
690 /* Did another grace period end? */
691 if (rdp->completed != completed_snap) {
693 /* Advance callbacks. No harm if list empty. */
694 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
695 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
696 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
698 /* Remember that we saw this grace-period completion. */
699 rdp->completed = completed_snap;
701 local_irq_restore(flags);
705 * Clean up after the prior grace period and let rcu_start_gp() start up
706 * the next grace period if one is needed. Note that the caller must
707 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
709 static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
710 __releases(rcu_get_root(rsp)->lock)
712 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
713 rsp->completed = rsp->gpnum;
714 rsp->signaled = RCU_GP_IDLE;
715 rcu_process_gp_end(rsp, rsp->rda[smp_processor_id()]);
716 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
720 * Similar to cpu_quiet(), for which it is a helper function. Allows
721 * a group of CPUs to be quieted at one go, though all the CPUs in the
722 * group must be represented by the same leaf rcu_node structure.
723 * That structure's lock must be held upon entry, and it is released
727 cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp,
729 __releases(rnp->lock)
731 struct rcu_node *rnp_c;
733 /* Walk up the rcu_node hierarchy. */
735 if (!(rnp->qsmask & mask)) {
737 /* Our bit has already been cleared, so done. */
738 spin_unlock_irqrestore(&rnp->lock, flags);
741 rnp->qsmask &= ~mask;
742 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
744 /* Other bits still set at this level, so done. */
745 spin_unlock_irqrestore(&rnp->lock, flags);
749 if (rnp->parent == NULL) {
751 /* No more levels. Exit loop holding root lock. */
755 spin_unlock_irqrestore(&rnp->lock, flags);
758 spin_lock_irqsave(&rnp->lock, flags);
759 WARN_ON_ONCE(rnp_c->qsmask);
763 * Get here if we are the last CPU to pass through a quiescent
764 * state for this grace period. Invoke cpu_quiet_msk_finish()
765 * to clean up and start the next grace period if one is needed.
767 cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */
771 * Record a quiescent state for the specified CPU, which must either be
772 * the current CPU. The lastcomp argument is used to make sure we are
773 * still in the grace period of interest. We don't want to end the current
774 * grace period based on quiescent states detected in an earlier grace
778 cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
782 struct rcu_node *rnp;
785 spin_lock_irqsave(&rnp->lock, flags);
786 if (lastcomp != ACCESS_ONCE(rsp->completed)) {
789 * Someone beat us to it for this grace period, so leave.
790 * The race with GP start is resolved by the fact that we
791 * hold the leaf rcu_node lock, so that the per-CPU bits
792 * cannot yet be initialized -- so we would simply find our
793 * CPU's bit already cleared in cpu_quiet_msk() if this race
796 rdp->passed_quiesc = 0; /* try again later! */
797 spin_unlock_irqrestore(&rnp->lock, flags);
801 if ((rnp->qsmask & mask) == 0) {
802 spin_unlock_irqrestore(&rnp->lock, flags);
807 * This GP can't end until cpu checks in, so all of our
808 * callbacks can be processed during the next GP.
810 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
812 cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */
817 * Check to see if there is a new grace period of which this CPU
818 * is not yet aware, and if so, set up local rcu_data state for it.
819 * Otherwise, see if this CPU has just passed through its first
820 * quiescent state for this grace period, and record that fact if so.
823 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
825 /* If there is now a new grace period, record and return. */
826 if (check_for_new_grace_period(rsp, rdp))
830 * Does this CPU still need to do its part for current grace period?
831 * If no, return and let the other CPUs do their part as well.
833 if (!rdp->qs_pending)
837 * Was there a quiescent state since the beginning of the grace
838 * period? If no, then exit and wait for the next call.
840 if (!rdp->passed_quiesc)
843 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
844 cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
847 #ifdef CONFIG_HOTPLUG_CPU
850 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
851 * specified flavor of RCU. The callbacks will be adopted by the next
852 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
853 * comes first. Because this is invoked from the CPU_DYING notifier,
854 * irqs are already disabled.
856 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
859 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
861 if (rdp->nxtlist == NULL)
862 return; /* irqs disabled, so comparison is stable. */
863 spin_lock(&rsp->onofflock); /* irqs already disabled. */
864 *rsp->orphan_cbs_tail = rdp->nxtlist;
865 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
867 for (i = 0; i < RCU_NEXT_SIZE; i++)
868 rdp->nxttail[i] = &rdp->nxtlist;
869 rsp->orphan_qlen += rdp->qlen;
871 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
875 * Adopt previously orphaned RCU callbacks.
877 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
880 struct rcu_data *rdp;
882 spin_lock_irqsave(&rsp->onofflock, flags);
883 rdp = rsp->rda[smp_processor_id()];
884 if (rsp->orphan_cbs_list == NULL) {
885 spin_unlock_irqrestore(&rsp->onofflock, flags);
888 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
889 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
890 rdp->qlen += rsp->orphan_qlen;
891 rsp->orphan_cbs_list = NULL;
892 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
893 rsp->orphan_qlen = 0;
894 spin_unlock_irqrestore(&rsp->onofflock, flags);
898 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
899 * and move all callbacks from the outgoing CPU to the current one.
901 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
906 struct rcu_data *rdp = rsp->rda[cpu];
907 struct rcu_node *rnp;
909 /* Exclude any attempts to start a new grace period. */
910 spin_lock_irqsave(&rsp->onofflock, flags);
912 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
913 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
914 mask = rdp->grpmask; /* rnp->grplo is constant. */
916 spin_lock(&rnp->lock); /* irqs already disabled. */
917 rnp->qsmaskinit &= ~mask;
918 if (rnp->qsmaskinit != 0) {
919 spin_unlock(&rnp->lock); /* irqs remain disabled. */
924 * If there was a task blocking the current grace period,
925 * and if all CPUs have checked in, we need to propagate
926 * the quiescent state up the rcu_node hierarchy. But that
927 * is inconvenient at the moment due to deadlock issues if
928 * this should end the current grace period. So set the
929 * offlined CPU's bit in ->qsmask in order to force the
930 * next force_quiescent_state() invocation to clean up this
931 * mess in a deadlock-free manner.
933 if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
937 spin_unlock(&rnp->lock); /* irqs remain disabled. */
939 } while (rnp != NULL);
940 lastcomp = rsp->completed;
942 spin_unlock_irqrestore(&rsp->onofflock, flags);
944 rcu_adopt_orphan_cbs(rsp);
948 * Remove the specified CPU from the RCU hierarchy and move any pending
949 * callbacks that it might have to the current CPU. This code assumes
950 * that at least one CPU in the system will remain running at all times.
951 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
953 static void rcu_offline_cpu(int cpu)
955 __rcu_offline_cpu(cpu, &rcu_sched_state);
956 __rcu_offline_cpu(cpu, &rcu_bh_state);
957 rcu_preempt_offline_cpu(cpu);
960 #else /* #ifdef CONFIG_HOTPLUG_CPU */
962 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
966 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
970 static void rcu_offline_cpu(int cpu)
974 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
977 * Invoke any RCU callbacks that have made it to the end of their grace
978 * period. Thottle as specified by rdp->blimit.
980 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
983 struct rcu_head *next, *list, **tail;
986 /* If no callbacks are ready, just return.*/
987 if (!cpu_has_callbacks_ready_to_invoke(rdp))
991 * Extract the list of ready callbacks, disabling to prevent
992 * races with call_rcu() from interrupt handlers.
994 local_irq_save(flags);
996 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
997 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
998 tail = rdp->nxttail[RCU_DONE_TAIL];
999 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1000 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1001 rdp->nxttail[count] = &rdp->nxtlist;
1002 local_irq_restore(flags);
1004 /* Invoke callbacks. */
1011 if (++count >= rdp->blimit)
1015 local_irq_save(flags);
1017 /* Update count, and requeue any remaining callbacks. */
1020 *tail = rdp->nxtlist;
1021 rdp->nxtlist = list;
1022 for (count = 0; count < RCU_NEXT_SIZE; count++)
1023 if (&rdp->nxtlist == rdp->nxttail[count])
1024 rdp->nxttail[count] = tail;
1029 /* Reinstate batch limit if we have worked down the excess. */
1030 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1031 rdp->blimit = blimit;
1033 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1034 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1035 rdp->qlen_last_fqs_check = 0;
1036 rdp->n_force_qs_snap = rsp->n_force_qs;
1037 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1038 rdp->qlen_last_fqs_check = rdp->qlen;
1040 local_irq_restore(flags);
1042 /* Re-raise the RCU softirq if there are callbacks remaining. */
1043 if (cpu_has_callbacks_ready_to_invoke(rdp))
1044 raise_softirq(RCU_SOFTIRQ);
1048 * Check to see if this CPU is in a non-context-switch quiescent state
1049 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1050 * Also schedule the RCU softirq handler.
1052 * This function must be called with hardirqs disabled. It is normally
1053 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1054 * false, there is no point in invoking rcu_check_callbacks().
1056 void rcu_check_callbacks(int cpu, int user)
1058 if (!rcu_pending(cpu))
1059 return; /* if nothing for RCU to do. */
1061 (idle_cpu(cpu) && rcu_scheduler_active &&
1062 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1065 * Get here if this CPU took its interrupt from user
1066 * mode or from the idle loop, and if this is not a
1067 * nested interrupt. In this case, the CPU is in
1068 * a quiescent state, so note it.
1070 * No memory barrier is required here because both
1071 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1072 * variables that other CPUs neither access nor modify,
1073 * at least not while the corresponding CPU is online.
1079 } else if (!in_softirq()) {
1082 * Get here if this CPU did not take its interrupt from
1083 * softirq, in other words, if it is not interrupting
1084 * a rcu_bh read-side critical section. This is an _bh
1085 * critical section, so note it.
1090 rcu_preempt_check_callbacks(cpu);
1091 raise_softirq(RCU_SOFTIRQ);
1097 * Scan the leaf rcu_node structures, processing dyntick state for any that
1098 * have not yet encountered a quiescent state, using the function specified.
1099 * Returns 1 if the current grace period ends while scanning (possibly
1100 * because we made it end).
1102 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1103 int (*f)(struct rcu_data *))
1107 unsigned long flags;
1109 struct rcu_node *rnp;
1111 rcu_for_each_leaf_node(rsp, rnp) {
1113 spin_lock_irqsave(&rnp->lock, flags);
1114 if (rsp->completed != lastcomp) {
1115 spin_unlock_irqrestore(&rnp->lock, flags);
1118 if (rnp->qsmask == 0) {
1119 spin_unlock_irqrestore(&rnp->lock, flags);
1124 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1125 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1128 if (mask != 0 && rsp->completed == lastcomp) {
1130 /* cpu_quiet_msk() releases rnp->lock. */
1131 cpu_quiet_msk(mask, rsp, rnp, flags);
1134 spin_unlock_irqrestore(&rnp->lock, flags);
1140 * Force quiescent states on reluctant CPUs, and also detect which
1141 * CPUs are in dyntick-idle mode.
1143 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1145 unsigned long flags;
1147 struct rcu_node *rnp = rcu_get_root(rsp);
1150 if (!rcu_gp_in_progress(rsp))
1151 return; /* No grace period in progress, nothing to force. */
1152 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1153 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1154 return; /* Someone else is already on the job. */
1157 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1158 goto unlock_ret; /* no emergency and done recently. */
1160 spin_lock(&rnp->lock);
1161 lastcomp = rsp->completed;
1162 signaled = rsp->signaled;
1163 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1164 if (lastcomp == rsp->gpnum) {
1165 rsp->n_force_qs_ngp++;
1166 spin_unlock(&rnp->lock);
1167 goto unlock_ret; /* no GP in progress, time updated. */
1169 spin_unlock(&rnp->lock);
1174 break; /* grace period idle or initializing, ignore. */
1176 case RCU_SAVE_DYNTICK:
1178 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1179 break; /* So gcc recognizes the dead code. */
1181 /* Record dyntick-idle state. */
1182 if (rcu_process_dyntick(rsp, lastcomp,
1183 dyntick_save_progress_counter))
1186 /* Update state, record completion counter. */
1187 spin_lock(&rnp->lock);
1188 if (lastcomp == rsp->completed &&
1189 rsp->signaled == RCU_SAVE_DYNTICK) {
1190 rsp->signaled = RCU_FORCE_QS;
1191 dyntick_record_completed(rsp, lastcomp);
1193 spin_unlock(&rnp->lock);
1198 /* Check dyntick-idle state, send IPI to laggarts. */
1199 if (rcu_process_dyntick(rsp, dyntick_recall_completed(rsp),
1200 rcu_implicit_dynticks_qs))
1203 /* Leave state in case more forcing is required. */
1208 spin_unlock_irqrestore(&rsp->fqslock, flags);
1211 #else /* #ifdef CONFIG_SMP */
1213 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1218 #endif /* #else #ifdef CONFIG_SMP */
1221 * This does the RCU processing work from softirq context for the
1222 * specified rcu_state and rcu_data structures. This may be called
1223 * only from the CPU to whom the rdp belongs.
1226 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1228 unsigned long flags;
1230 WARN_ON_ONCE(rdp->beenonline == 0);
1233 * If an RCU GP has gone long enough, go check for dyntick
1234 * idle CPUs and, if needed, send resched IPIs.
1236 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1237 force_quiescent_state(rsp, 1);
1240 * Advance callbacks in response to end of earlier grace
1241 * period that some other CPU ended.
1243 rcu_process_gp_end(rsp, rdp);
1245 /* Update RCU state based on any recent quiescent states. */
1246 rcu_check_quiescent_state(rsp, rdp);
1248 /* Does this CPU require a not-yet-started grace period? */
1249 if (cpu_needs_another_gp(rsp, rdp)) {
1250 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1251 rcu_start_gp(rsp, flags); /* releases above lock */
1254 /* If there are callbacks ready, invoke them. */
1255 rcu_do_batch(rsp, rdp);
1259 * Do softirq processing for the current CPU.
1261 static void rcu_process_callbacks(struct softirq_action *unused)
1264 * Memory references from any prior RCU read-side critical sections
1265 * executed by the interrupted code must be seen before any RCU
1266 * grace-period manipulations below.
1268 smp_mb(); /* See above block comment. */
1270 __rcu_process_callbacks(&rcu_sched_state,
1271 &__get_cpu_var(rcu_sched_data));
1272 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1273 rcu_preempt_process_callbacks();
1276 * Memory references from any later RCU read-side critical sections
1277 * executed by the interrupted code must be seen after any RCU
1278 * grace-period manipulations above.
1280 smp_mb(); /* See above block comment. */
1284 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1285 struct rcu_state *rsp)
1287 unsigned long flags;
1288 struct rcu_data *rdp;
1293 smp_mb(); /* Ensure RCU update seen before callback registry. */
1296 * Opportunistically note grace-period endings and beginnings.
1297 * Note that we might see a beginning right after we see an
1298 * end, but never vice versa, since this CPU has to pass through
1299 * a quiescent state betweentimes.
1301 local_irq_save(flags);
1302 rdp = rsp->rda[smp_processor_id()];
1303 rcu_process_gp_end(rsp, rdp);
1304 check_for_new_grace_period(rsp, rdp);
1306 /* Add the callback to our list. */
1307 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1308 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1310 /* Start a new grace period if one not already started. */
1311 if (!rcu_gp_in_progress(rsp)) {
1312 unsigned long nestflag;
1313 struct rcu_node *rnp_root = rcu_get_root(rsp);
1315 spin_lock_irqsave(&rnp_root->lock, nestflag);
1316 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1320 * Force the grace period if too many callbacks or too long waiting.
1321 * Enforce hysteresis, and don't invoke force_quiescent_state()
1322 * if some other CPU has recently done so. Also, don't bother
1323 * invoking force_quiescent_state() if the newly enqueued callback
1324 * is the only one waiting for a grace period to complete.
1326 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1327 rdp->blimit = LONG_MAX;
1328 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1329 *rdp->nxttail[RCU_DONE_TAIL] != head)
1330 force_quiescent_state(rsp, 0);
1331 rdp->n_force_qs_snap = rsp->n_force_qs;
1332 rdp->qlen_last_fqs_check = rdp->qlen;
1333 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1334 force_quiescent_state(rsp, 1);
1335 local_irq_restore(flags);
1339 * Queue an RCU-sched callback for invocation after a grace period.
1341 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1343 __call_rcu(head, func, &rcu_sched_state);
1345 EXPORT_SYMBOL_GPL(call_rcu_sched);
1348 * Queue an RCU for invocation after a quicker grace period.
1350 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1352 __call_rcu(head, func, &rcu_bh_state);
1354 EXPORT_SYMBOL_GPL(call_rcu_bh);
1357 * Check to see if there is any immediate RCU-related work to be done
1358 * by the current CPU, for the specified type of RCU, returning 1 if so.
1359 * The checks are in order of increasing expense: checks that can be
1360 * carried out against CPU-local state are performed first. However,
1361 * we must check for CPU stalls first, else we might not get a chance.
1363 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1365 rdp->n_rcu_pending++;
1367 /* Check for CPU stalls, if enabled. */
1368 check_cpu_stall(rsp, rdp);
1370 /* Is the RCU core waiting for a quiescent state from this CPU? */
1371 if (rdp->qs_pending) {
1372 rdp->n_rp_qs_pending++;
1376 /* Does this CPU have callbacks ready to invoke? */
1377 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1378 rdp->n_rp_cb_ready++;
1382 /* Has RCU gone idle with this CPU needing another grace period? */
1383 if (cpu_needs_another_gp(rsp, rdp)) {
1384 rdp->n_rp_cpu_needs_gp++;
1388 /* Has another RCU grace period completed? */
1389 if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */
1390 rdp->n_rp_gp_completed++;
1394 /* Has a new RCU grace period started? */
1395 if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */
1396 rdp->n_rp_gp_started++;
1400 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1401 if (rcu_gp_in_progress(rsp) &&
1402 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1403 rdp->n_rp_need_fqs++;
1408 rdp->n_rp_need_nothing++;
1413 * Check to see if there is any immediate RCU-related work to be done
1414 * by the current CPU, returning 1 if so. This function is part of the
1415 * RCU implementation; it is -not- an exported member of the RCU API.
1417 static int rcu_pending(int cpu)
1419 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1420 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1421 rcu_preempt_pending(cpu);
1425 * Check to see if any future RCU-related work will need to be done
1426 * by the current CPU, even if none need be done immediately, returning
1427 * 1 if so. This function is part of the RCU implementation; it is -not-
1428 * an exported member of the RCU API.
1430 int rcu_needs_cpu(int cpu)
1432 /* RCU callbacks either ready or pending? */
1433 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1434 per_cpu(rcu_bh_data, cpu).nxtlist ||
1435 rcu_preempt_needs_cpu(cpu);
1438 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1439 static atomic_t rcu_barrier_cpu_count;
1440 static DEFINE_MUTEX(rcu_barrier_mutex);
1441 static struct completion rcu_barrier_completion;
1443 static void rcu_barrier_callback(struct rcu_head *notused)
1445 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1446 complete(&rcu_barrier_completion);
1450 * Called with preemption disabled, and from cross-cpu IRQ context.
1452 static void rcu_barrier_func(void *type)
1454 int cpu = smp_processor_id();
1455 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1456 void (*call_rcu_func)(struct rcu_head *head,
1457 void (*func)(struct rcu_head *head));
1459 atomic_inc(&rcu_barrier_cpu_count);
1460 call_rcu_func = type;
1461 call_rcu_func(head, rcu_barrier_callback);
1465 * Orchestrate the specified type of RCU barrier, waiting for all
1466 * RCU callbacks of the specified type to complete.
1468 static void _rcu_barrier(struct rcu_state *rsp,
1469 void (*call_rcu_func)(struct rcu_head *head,
1470 void (*func)(struct rcu_head *head)))
1472 BUG_ON(in_interrupt());
1473 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1474 mutex_lock(&rcu_barrier_mutex);
1475 init_completion(&rcu_barrier_completion);
1477 * Initialize rcu_barrier_cpu_count to 1, then invoke
1478 * rcu_barrier_func() on each CPU, so that each CPU also has
1479 * incremented rcu_barrier_cpu_count. Only then is it safe to
1480 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1481 * might complete its grace period before all of the other CPUs
1482 * did their increment, causing this function to return too
1485 atomic_set(&rcu_barrier_cpu_count, 1);
1486 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1487 rcu_adopt_orphan_cbs(rsp);
1488 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1489 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1490 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1491 complete(&rcu_barrier_completion);
1492 wait_for_completion(&rcu_barrier_completion);
1493 mutex_unlock(&rcu_barrier_mutex);
1497 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1499 void rcu_barrier_bh(void)
1501 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1503 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1506 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1508 void rcu_barrier_sched(void)
1510 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1512 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1515 * Do boot-time initialization of a CPU's per-CPU RCU data.
1518 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1520 unsigned long flags;
1522 struct rcu_data *rdp = rsp->rda[cpu];
1523 struct rcu_node *rnp = rcu_get_root(rsp);
1525 /* Set up local state, ensuring consistent view of global state. */
1526 spin_lock_irqsave(&rnp->lock, flags);
1527 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1528 rdp->nxtlist = NULL;
1529 for (i = 0; i < RCU_NEXT_SIZE; i++)
1530 rdp->nxttail[i] = &rdp->nxtlist;
1533 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1534 #endif /* #ifdef CONFIG_NO_HZ */
1536 spin_unlock_irqrestore(&rnp->lock, flags);
1540 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1541 * offline event can be happening at a given time. Note also that we
1542 * can accept some slop in the rsp->completed access due to the fact
1543 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1545 static void __cpuinit
1546 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1548 unsigned long flags;
1551 struct rcu_data *rdp = rsp->rda[cpu];
1552 struct rcu_node *rnp = rcu_get_root(rsp);
1554 /* Set up local state, ensuring consistent view of global state. */
1555 spin_lock_irqsave(&rnp->lock, flags);
1556 lastcomp = rsp->completed;
1557 rdp->completed = lastcomp;
1558 rdp->gpnum = lastcomp;
1559 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1560 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1561 rdp->beenonline = 1; /* We have now been online. */
1562 rdp->preemptable = preemptable;
1563 rdp->passed_quiesc_completed = lastcomp - 1;
1564 rdp->qlen_last_fqs_check = 0;
1565 rdp->n_force_qs_snap = rsp->n_force_qs;
1566 rdp->blimit = blimit;
1567 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1570 * A new grace period might start here. If so, we won't be part
1571 * of it, but that is OK, as we are currently in a quiescent state.
1574 /* Exclude any attempts to start a new GP on large systems. */
1575 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1577 /* Add CPU to rcu_node bitmasks. */
1579 mask = rdp->grpmask;
1581 /* Exclude any attempts to start a new GP on small systems. */
1582 spin_lock(&rnp->lock); /* irqs already disabled. */
1583 rnp->qsmaskinit |= mask;
1584 mask = rnp->grpmask;
1585 spin_unlock(&rnp->lock); /* irqs already disabled. */
1587 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1589 spin_unlock_irqrestore(&rsp->onofflock, flags);
1592 static void __cpuinit rcu_online_cpu(int cpu)
1594 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1595 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1596 rcu_preempt_init_percpu_data(cpu);
1600 * Handle CPU online/offline notification events.
1602 int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1603 unsigned long action, void *hcpu)
1605 long cpu = (long)hcpu;
1608 case CPU_UP_PREPARE:
1609 case CPU_UP_PREPARE_FROZEN:
1610 rcu_online_cpu(cpu);
1613 case CPU_DYING_FROZEN:
1615 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1616 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1617 * returns, all online cpus have queued rcu_barrier_func().
1618 * The dying CPU clears its cpu_online_mask bit and
1619 * moves all of its RCU callbacks to ->orphan_cbs_list
1620 * in the context of stop_machine(), so subsequent calls
1621 * to _rcu_barrier() will adopt these callbacks and only
1622 * then queue rcu_barrier_func() on all remaining CPUs.
1624 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1625 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1626 rcu_preempt_send_cbs_to_orphanage();
1629 case CPU_DEAD_FROZEN:
1630 case CPU_UP_CANCELED:
1631 case CPU_UP_CANCELED_FROZEN:
1632 rcu_offline_cpu(cpu);
1641 * Compute the per-level fanout, either using the exact fanout specified
1642 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1644 #ifdef CONFIG_RCU_FANOUT_EXACT
1645 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1649 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1650 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1652 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1653 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1660 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1661 ccur = rsp->levelcnt[i];
1662 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1666 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1669 * Helper function for rcu_init() that initializes one rcu_state structure.
1671 static void __init rcu_init_one(struct rcu_state *rsp)
1676 struct rcu_node *rnp;
1678 /* Initialize the level-tracking arrays. */
1680 for (i = 1; i < NUM_RCU_LVLS; i++)
1681 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1682 rcu_init_levelspread(rsp);
1684 /* Initialize the elements themselves, starting from the leaves. */
1686 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1687 cpustride *= rsp->levelspread[i];
1688 rnp = rsp->level[i];
1689 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1690 spin_lock_init(&rnp->lock);
1693 rnp->qsmaskinit = 0;
1694 rnp->grplo = j * cpustride;
1695 rnp->grphi = (j + 1) * cpustride - 1;
1696 if (rnp->grphi >= NR_CPUS)
1697 rnp->grphi = NR_CPUS - 1;
1703 rnp->grpnum = j % rsp->levelspread[i - 1];
1704 rnp->grpmask = 1UL << rnp->grpnum;
1705 rnp->parent = rsp->level[i - 1] +
1706 j / rsp->levelspread[i - 1];
1709 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1710 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1713 lockdep_set_class(&rcu_get_root(rsp)->lock, &rcu_root_class);
1717 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1718 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1721 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1725 struct rcu_node *rnp; \
1727 rcu_init_one(rsp); \
1728 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1730 for_each_possible_cpu(i) { \
1731 if (i > rnp[j].grphi) \
1733 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1734 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1735 rcu_boot_init_percpu_data(i, rsp); \
1739 void __init __rcu_init(void)
1741 rcu_bootup_announce();
1742 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1743 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1744 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1745 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1746 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1747 __rcu_init_preempt();
1748 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1751 #include "rcutree_plugin.h"