2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
53 /* Data structures. */
55 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
57 #define RCU_STATE_INITIALIZER(name) { \
58 .level = { &name.node[0] }, \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
64 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
66 .signaled = RCU_GP_IDLE, \
69 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&name.onofflock), \
70 .orphan_cbs_list = NULL, \
71 .orphan_cbs_tail = &name.orphan_cbs_list, \
73 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&name.fqslock), \
75 .n_force_qs_ngp = 0, \
78 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
79 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
81 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
82 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84 int rcu_scheduler_active __read_mostly;
85 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
88 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
89 * permit this function to be invoked without holding the root rcu_node
90 * structure's ->lock, but of course results can be subject to change.
92 static int rcu_gp_in_progress(struct rcu_state *rsp)
94 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
98 * Note a quiescent state. Because we do not need to know
99 * how many quiescent states passed, just if there was at least
100 * one since the start of the grace period, this just sets a flag.
102 void rcu_sched_qs(int cpu)
104 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
106 rdp->passed_quiesc_completed = rdp->gpnum - 1;
108 rdp->passed_quiesc = 1;
111 void rcu_bh_qs(int cpu)
113 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
115 rdp->passed_quiesc_completed = rdp->gpnum - 1;
117 rdp->passed_quiesc = 1;
121 * Note a context switch. This is a quiescent state for RCU-sched,
122 * and requires special handling for preemptible RCU.
124 void rcu_note_context_switch(int cpu)
127 rcu_preempt_note_context_switch(cpu);
131 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
132 .dynticks_nesting = 1,
135 #endif /* #ifdef CONFIG_NO_HZ */
137 static int blimit = 10; /* Maximum callbacks per softirq. */
138 static int qhimark = 10000; /* If this many pending, ignore blimit. */
139 static int qlowmark = 100; /* Once only this many pending, use blimit. */
141 module_param(blimit, int, 0);
142 module_param(qhimark, int, 0);
143 module_param(qlowmark, int, 0);
145 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
146 static int rcu_pending(int cpu);
149 * Return the number of RCU-sched batches processed thus far for debug & stats.
151 long rcu_batches_completed_sched(void)
153 return rcu_sched_state.completed;
155 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
158 * Return the number of RCU BH batches processed thus far for debug & stats.
160 long rcu_batches_completed_bh(void)
162 return rcu_bh_state.completed;
164 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
167 * Force a quiescent state for RCU BH.
169 void rcu_bh_force_quiescent_state(void)
171 force_quiescent_state(&rcu_bh_state, 0);
173 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
176 * Force a quiescent state for RCU-sched.
178 void rcu_sched_force_quiescent_state(void)
180 force_quiescent_state(&rcu_sched_state, 0);
182 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
185 * Does the CPU have callbacks ready to be invoked?
188 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
190 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
194 * Does the current CPU require a yet-as-unscheduled grace period?
197 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
199 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
203 * Return the root node of the specified rcu_state structure.
205 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
207 return &rsp->node[0];
213 * If the specified CPU is offline, tell the caller that it is in
214 * a quiescent state. Otherwise, whack it with a reschedule IPI.
215 * Grace periods can end up waiting on an offline CPU when that
216 * CPU is in the process of coming online -- it will be added to the
217 * rcu_node bitmasks before it actually makes it online. The same thing
218 * can happen while a CPU is in the process of coming online. Because this
219 * race is quite rare, we check for it after detecting that the grace
220 * period has been delayed rather than checking each and every CPU
221 * each and every time we start a new grace period.
223 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
226 * If the CPU is offline, it is in a quiescent state. We can
227 * trust its state not to change because interrupts are disabled.
229 if (cpu_is_offline(rdp->cpu)) {
234 /* If preemptable RCU, no point in sending reschedule IPI. */
235 if (rdp->preemptable)
238 /* The CPU is online, so send it a reschedule IPI. */
239 if (rdp->cpu != smp_processor_id())
240 smp_send_reschedule(rdp->cpu);
247 #endif /* #ifdef CONFIG_SMP */
252 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
254 * Enter nohz mode, in other words, -leave- the mode in which RCU
255 * read-side critical sections can occur. (Though RCU read-side
256 * critical sections can occur in irq handlers in nohz mode, a possibility
257 * handled by rcu_irq_enter() and rcu_irq_exit()).
259 void rcu_enter_nohz(void)
262 struct rcu_dynticks *rdtp;
264 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
265 local_irq_save(flags);
266 rdtp = &__get_cpu_var(rcu_dynticks);
268 rdtp->dynticks_nesting--;
269 WARN_ON_ONCE(rdtp->dynticks & 0x1);
270 local_irq_restore(flags);
274 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
276 * Exit nohz mode, in other words, -enter- the mode in which RCU
277 * read-side critical sections normally occur.
279 void rcu_exit_nohz(void)
282 struct rcu_dynticks *rdtp;
284 local_irq_save(flags);
285 rdtp = &__get_cpu_var(rcu_dynticks);
287 rdtp->dynticks_nesting++;
288 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
289 local_irq_restore(flags);
290 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
294 * rcu_nmi_enter - inform RCU of entry to NMI context
296 * If the CPU was idle with dynamic ticks active, and there is no
297 * irq handler running, this updates rdtp->dynticks_nmi to let the
298 * RCU grace-period handling know that the CPU is active.
300 void rcu_nmi_enter(void)
302 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
304 if (rdtp->dynticks & 0x1)
306 rdtp->dynticks_nmi++;
307 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
308 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
312 * rcu_nmi_exit - inform RCU of exit from NMI context
314 * If the CPU was idle with dynamic ticks active, and there is no
315 * irq handler running, this updates rdtp->dynticks_nmi to let the
316 * RCU grace-period handling know that the CPU is no longer active.
318 void rcu_nmi_exit(void)
320 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
322 if (rdtp->dynticks & 0x1)
324 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
325 rdtp->dynticks_nmi++;
326 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
330 * rcu_irq_enter - inform RCU of entry to hard irq context
332 * If the CPU was idle with dynamic ticks active, this updates the
333 * rdtp->dynticks to let the RCU handling know that the CPU is active.
335 void rcu_irq_enter(void)
337 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
339 if (rdtp->dynticks_nesting++)
342 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
343 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
347 * rcu_irq_exit - inform RCU of exit from hard irq context
349 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
350 * to put let the RCU handling be aware that the CPU is going back to idle
353 void rcu_irq_exit(void)
355 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
357 if (--rdtp->dynticks_nesting)
359 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
361 WARN_ON_ONCE(rdtp->dynticks & 0x1);
363 /* If the interrupt queued a callback, get out of dyntick mode. */
364 if (__get_cpu_var(rcu_sched_data).nxtlist ||
365 __get_cpu_var(rcu_bh_data).nxtlist)
372 * Snapshot the specified CPU's dynticks counter so that we can later
373 * credit them with an implicit quiescent state. Return 1 if this CPU
374 * is in dynticks idle mode, which is an extended quiescent state.
376 static int dyntick_save_progress_counter(struct rcu_data *rdp)
382 snap = rdp->dynticks->dynticks;
383 snap_nmi = rdp->dynticks->dynticks_nmi;
384 smp_mb(); /* Order sampling of snap with end of grace period. */
385 rdp->dynticks_snap = snap;
386 rdp->dynticks_nmi_snap = snap_nmi;
387 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
394 * Return true if the specified CPU has passed through a quiescent
395 * state by virtue of being in or having passed through an dynticks
396 * idle state since the last call to dyntick_save_progress_counter()
399 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
406 curr = rdp->dynticks->dynticks;
407 snap = rdp->dynticks_snap;
408 curr_nmi = rdp->dynticks->dynticks_nmi;
409 snap_nmi = rdp->dynticks_nmi_snap;
410 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
413 * If the CPU passed through or entered a dynticks idle phase with
414 * no active irq/NMI handlers, then we can safely pretend that the CPU
415 * already acknowledged the request to pass through a quiescent
416 * state. Either way, that CPU cannot possibly be in an RCU
417 * read-side critical section that started before the beginning
418 * of the current RCU grace period.
420 if ((curr != snap || (curr & 0x1) == 0) &&
421 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
426 /* Go check for the CPU being offline. */
427 return rcu_implicit_offline_qs(rdp);
430 #endif /* #ifdef CONFIG_SMP */
432 #else /* #ifdef CONFIG_NO_HZ */
436 static int dyntick_save_progress_counter(struct rcu_data *rdp)
441 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
443 return rcu_implicit_offline_qs(rdp);
446 #endif /* #ifdef CONFIG_SMP */
448 #endif /* #else #ifdef CONFIG_NO_HZ */
450 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
452 static void record_gp_stall_check_time(struct rcu_state *rsp)
454 rsp->gp_start = jiffies;
455 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
458 static void print_other_cpu_stall(struct rcu_state *rsp)
463 struct rcu_node *rnp = rcu_get_root(rsp);
465 /* Only let one CPU complain about others per time interval. */
467 raw_spin_lock_irqsave(&rnp->lock, flags);
468 delta = jiffies - rsp->jiffies_stall;
469 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
470 raw_spin_unlock_irqrestore(&rnp->lock, flags);
473 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
476 * Now rat on any tasks that got kicked up to the root rcu_node
477 * due to CPU offlining.
479 rcu_print_task_stall(rnp);
480 raw_spin_unlock_irqrestore(&rnp->lock, flags);
482 /* OK, time to rat on our buddy... */
484 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
485 rcu_for_each_leaf_node(rsp, rnp) {
486 raw_spin_lock_irqsave(&rnp->lock, flags);
487 rcu_print_task_stall(rnp);
488 raw_spin_unlock_irqrestore(&rnp->lock, flags);
489 if (rnp->qsmask == 0)
491 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
492 if (rnp->qsmask & (1UL << cpu))
493 printk(" %d", rnp->grplo + cpu);
495 printk(" (detected by %d, t=%ld jiffies)\n",
496 smp_processor_id(), (long)(jiffies - rsp->gp_start));
497 trigger_all_cpu_backtrace();
499 /* If so configured, complain about tasks blocking the grace period. */
501 rcu_print_detail_task_stall(rsp);
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 raw_spin_lock_irqsave(&rnp->lock, flags);
516 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
518 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
519 raw_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. The caller must hold the
559 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
560 * and must have irqs disabled.
562 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
564 if (rdp->gpnum != rnp->gpnum) {
566 rdp->passed_quiesc = 0;
567 rdp->gpnum = rnp->gpnum;
571 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
574 struct rcu_node *rnp;
576 local_irq_save(flags);
578 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
579 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
580 local_irq_restore(flags);
583 __note_new_gpnum(rsp, rnp, rdp);
584 raw_spin_unlock_irqrestore(&rnp->lock, flags);
588 * Did someone else start a new RCU grace period start since we last
589 * checked? Update local state appropriately if so. Must be called
590 * on the CPU corresponding to rdp.
593 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
598 local_irq_save(flags);
599 if (rdp->gpnum != rsp->gpnum) {
600 note_new_gpnum(rsp, rdp);
603 local_irq_restore(flags);
608 * Advance this CPU's callbacks, but only if the current grace period
609 * has ended. This may be called only from the CPU to whom the rdp
610 * belongs. In addition, the corresponding leaf rcu_node structure's
611 * ->lock must be held by the caller, with irqs disabled.
614 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
616 /* Did another grace period end? */
617 if (rdp->completed != rnp->completed) {
619 /* Advance callbacks. No harm if list empty. */
620 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
621 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
622 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
624 /* Remember that we saw this grace-period completion. */
625 rdp->completed = rnp->completed;
630 * Advance this CPU's callbacks, but only if the current grace period
631 * has ended. This may be called only from the CPU to whom the rdp
635 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
638 struct rcu_node *rnp;
640 local_irq_save(flags);
642 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
643 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
644 local_irq_restore(flags);
647 __rcu_process_gp_end(rsp, rnp, rdp);
648 raw_spin_unlock_irqrestore(&rnp->lock, flags);
652 * Do per-CPU grace-period initialization for running CPU. The caller
653 * must hold the lock of the leaf rcu_node structure corresponding to
657 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
659 /* Prior grace period ended, so advance callbacks for current CPU. */
660 __rcu_process_gp_end(rsp, rnp, rdp);
663 * Because this CPU just now started the new grace period, we know
664 * that all of its callbacks will be covered by this upcoming grace
665 * period, even the ones that were registered arbitrarily recently.
666 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
668 * Other CPUs cannot be sure exactly when the grace period started.
669 * Therefore, their recently registered callbacks must pass through
670 * an additional RCU_NEXT_READY stage, so that they will be handled
671 * by the next RCU grace period.
673 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
674 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
676 /* Set state so that this CPU will detect the next quiescent state. */
677 __note_new_gpnum(rsp, rnp, rdp);
681 * Start a new RCU grace period if warranted, re-initializing the hierarchy
682 * in preparation for detecting the next grace period. The caller must hold
683 * the root node's ->lock, which is released before return. Hard irqs must
687 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
688 __releases(rcu_get_root(rsp)->lock)
690 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
691 struct rcu_node *rnp = rcu_get_root(rsp);
693 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
694 if (cpu_needs_another_gp(rsp, rdp))
695 rsp->fqs_need_gp = 1;
696 if (rnp->completed == rsp->completed) {
697 raw_spin_unlock_irqrestore(&rnp->lock, flags);
700 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
703 * Propagate new ->completed value to rcu_node structures
704 * so that other CPUs don't have to wait until the start
705 * of the next grace period to process their callbacks.
707 rcu_for_each_node_breadth_first(rsp, rnp) {
708 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
709 rnp->completed = rsp->completed;
710 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
712 local_irq_restore(flags);
716 /* Advance to a new grace period and initialize state. */
718 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
719 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
720 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
721 record_gp_stall_check_time(rsp);
723 /* Special-case the common single-level case. */
724 if (NUM_RCU_NODES == 1) {
725 rcu_preempt_check_blocked_tasks(rnp);
726 rnp->qsmask = rnp->qsmaskinit;
727 rnp->gpnum = rsp->gpnum;
728 rnp->completed = rsp->completed;
729 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
730 rcu_start_gp_per_cpu(rsp, rnp, rdp);
731 raw_spin_unlock_irqrestore(&rnp->lock, flags);
735 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
738 /* Exclude any concurrent CPU-hotplug operations. */
739 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
742 * Set the quiescent-state-needed bits in all the rcu_node
743 * structures for all currently online CPUs in breadth-first
744 * order, starting from the root rcu_node structure. This
745 * operation relies on the layout of the hierarchy within the
746 * rsp->node[] array. Note that other CPUs will access only
747 * the leaves of the hierarchy, which still indicate that no
748 * grace period is in progress, at least until the corresponding
749 * leaf node has been initialized. In addition, we have excluded
750 * CPU-hotplug operations.
752 * Note that the grace period cannot complete until we finish
753 * the initialization process, as there will be at least one
754 * qsmask bit set in the root node until that time, namely the
755 * one corresponding to this CPU, due to the fact that we have
758 rcu_for_each_node_breadth_first(rsp, rnp) {
759 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
760 rcu_preempt_check_blocked_tasks(rnp);
761 rnp->qsmask = rnp->qsmaskinit;
762 rnp->gpnum = rsp->gpnum;
763 rnp->completed = rsp->completed;
764 if (rnp == rdp->mynode)
765 rcu_start_gp_per_cpu(rsp, rnp, rdp);
766 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
769 rnp = rcu_get_root(rsp);
770 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
771 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
772 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
773 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
777 * Report a full set of quiescent states to the specified rcu_state
778 * data structure. This involves cleaning up after the prior grace
779 * period and letting rcu_start_gp() start up the next grace period
780 * if one is needed. Note that the caller must hold rnp->lock, as
781 * required by rcu_start_gp(), which will release it.
783 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
784 __releases(rcu_get_root(rsp)->lock)
786 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
787 rsp->completed = rsp->gpnum;
788 rsp->signaled = RCU_GP_IDLE;
789 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
793 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
794 * Allows quiescent states for a group of CPUs to be reported at one go
795 * to the specified rcu_node structure, though all the CPUs in the group
796 * must be represented by the same rcu_node structure (which need not be
797 * a leaf rcu_node structure, though it often will be). That structure's
798 * lock must be held upon entry, and it is released before return.
801 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
802 struct rcu_node *rnp, unsigned long flags)
803 __releases(rnp->lock)
805 struct rcu_node *rnp_c;
807 /* Walk up the rcu_node hierarchy. */
809 if (!(rnp->qsmask & mask)) {
811 /* Our bit has already been cleared, so done. */
812 raw_spin_unlock_irqrestore(&rnp->lock, flags);
815 rnp->qsmask &= ~mask;
816 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
818 /* Other bits still set at this level, so done. */
819 raw_spin_unlock_irqrestore(&rnp->lock, flags);
823 if (rnp->parent == NULL) {
825 /* No more levels. Exit loop holding root lock. */
829 raw_spin_unlock_irqrestore(&rnp->lock, flags);
832 raw_spin_lock_irqsave(&rnp->lock, flags);
833 WARN_ON_ONCE(rnp_c->qsmask);
837 * Get here if we are the last CPU to pass through a quiescent
838 * state for this grace period. Invoke rcu_report_qs_rsp()
839 * to clean up and start the next grace period if one is needed.
841 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
845 * Record a quiescent state for the specified CPU to that CPU's rcu_data
846 * structure. This must be either called from the specified CPU, or
847 * called when the specified CPU is known to be offline (and when it is
848 * also known that no other CPU is concurrently trying to help the offline
849 * CPU). The lastcomp argument is used to make sure we are still in the
850 * grace period of interest. We don't want to end the current grace period
851 * based on quiescent states detected in an earlier grace period!
854 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
858 struct rcu_node *rnp;
861 raw_spin_lock_irqsave(&rnp->lock, flags);
862 if (lastcomp != rnp->completed) {
865 * Someone beat us to it for this grace period, so leave.
866 * The race with GP start is resolved by the fact that we
867 * hold the leaf rcu_node lock, so that the per-CPU bits
868 * cannot yet be initialized -- so we would simply find our
869 * CPU's bit already cleared in rcu_report_qs_rnp() if this
872 rdp->passed_quiesc = 0; /* try again later! */
873 raw_spin_unlock_irqrestore(&rnp->lock, flags);
877 if ((rnp->qsmask & mask) == 0) {
878 raw_spin_unlock_irqrestore(&rnp->lock, flags);
883 * This GP can't end until cpu checks in, so all of our
884 * callbacks can be processed during the next GP.
886 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
888 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
893 * Check to see if there is a new grace period of which this CPU
894 * is not yet aware, and if so, set up local rcu_data state for it.
895 * Otherwise, see if this CPU has just passed through its first
896 * quiescent state for this grace period, and record that fact if so.
899 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
901 /* If there is now a new grace period, record and return. */
902 if (check_for_new_grace_period(rsp, rdp))
906 * Does this CPU still need to do its part for current grace period?
907 * If no, return and let the other CPUs do their part as well.
909 if (!rdp->qs_pending)
913 * Was there a quiescent state since the beginning of the grace
914 * period? If no, then exit and wait for the next call.
916 if (!rdp->passed_quiesc)
920 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
923 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
926 #ifdef CONFIG_HOTPLUG_CPU
929 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
930 * specified flavor of RCU. The callbacks will be adopted by the next
931 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
932 * comes first. Because this is invoked from the CPU_DYING notifier,
933 * irqs are already disabled.
935 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
938 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
940 if (rdp->nxtlist == NULL)
941 return; /* irqs disabled, so comparison is stable. */
942 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
943 *rsp->orphan_cbs_tail = rdp->nxtlist;
944 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
946 for (i = 0; i < RCU_NEXT_SIZE; i++)
947 rdp->nxttail[i] = &rdp->nxtlist;
948 rsp->orphan_qlen += rdp->qlen;
950 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
954 * Adopt previously orphaned RCU callbacks.
956 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
959 struct rcu_data *rdp;
961 raw_spin_lock_irqsave(&rsp->onofflock, flags);
962 rdp = rsp->rda[smp_processor_id()];
963 if (rsp->orphan_cbs_list == NULL) {
964 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
967 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
968 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
969 rdp->qlen += rsp->orphan_qlen;
970 rsp->orphan_cbs_list = NULL;
971 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
972 rsp->orphan_qlen = 0;
973 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
977 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
978 * and move all callbacks from the outgoing CPU to the current one.
980 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
985 struct rcu_data *rdp = rsp->rda[cpu];
986 struct rcu_node *rnp;
988 /* Exclude any attempts to start a new grace period. */
989 raw_spin_lock_irqsave(&rsp->onofflock, flags);
991 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
992 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
993 mask = rdp->grpmask; /* rnp->grplo is constant. */
995 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
996 rnp->qsmaskinit &= ~mask;
997 if (rnp->qsmaskinit != 0) {
998 if (rnp != rdp->mynode)
999 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1002 if (rnp == rdp->mynode)
1003 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1005 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1006 mask = rnp->grpmask;
1008 } while (rnp != NULL);
1011 * We still hold the leaf rcu_node structure lock here, and
1012 * irqs are still disabled. The reason for this subterfuge is
1013 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1014 * held leads to deadlock.
1016 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1018 if (need_report & RCU_OFL_TASKS_NORM_GP)
1019 rcu_report_unblock_qs_rnp(rnp, flags);
1021 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1022 if (need_report & RCU_OFL_TASKS_EXP_GP)
1023 rcu_report_exp_rnp(rsp, rnp);
1025 rcu_adopt_orphan_cbs(rsp);
1029 * Remove the specified CPU from the RCU hierarchy and move any pending
1030 * callbacks that it might have to the current CPU. This code assumes
1031 * that at least one CPU in the system will remain running at all times.
1032 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1034 static void rcu_offline_cpu(int cpu)
1036 __rcu_offline_cpu(cpu, &rcu_sched_state);
1037 __rcu_offline_cpu(cpu, &rcu_bh_state);
1038 rcu_preempt_offline_cpu(cpu);
1041 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1043 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1047 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1051 static void rcu_offline_cpu(int cpu)
1055 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1058 * Invoke any RCU callbacks that have made it to the end of their grace
1059 * period. Thottle as specified by rdp->blimit.
1061 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1063 unsigned long flags;
1064 struct rcu_head *next, *list, **tail;
1067 /* If no callbacks are ready, just return.*/
1068 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1072 * Extract the list of ready callbacks, disabling to prevent
1073 * races with call_rcu() from interrupt handlers.
1075 local_irq_save(flags);
1076 list = rdp->nxtlist;
1077 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1078 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1079 tail = rdp->nxttail[RCU_DONE_TAIL];
1080 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1081 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1082 rdp->nxttail[count] = &rdp->nxtlist;
1083 local_irq_restore(flags);
1085 /* Invoke callbacks. */
1092 if (++count >= rdp->blimit)
1096 local_irq_save(flags);
1098 /* Update count, and requeue any remaining callbacks. */
1101 *tail = rdp->nxtlist;
1102 rdp->nxtlist = list;
1103 for (count = 0; count < RCU_NEXT_SIZE; count++)
1104 if (&rdp->nxtlist == rdp->nxttail[count])
1105 rdp->nxttail[count] = tail;
1110 /* Reinstate batch limit if we have worked down the excess. */
1111 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1112 rdp->blimit = blimit;
1114 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1115 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1116 rdp->qlen_last_fqs_check = 0;
1117 rdp->n_force_qs_snap = rsp->n_force_qs;
1118 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1119 rdp->qlen_last_fqs_check = rdp->qlen;
1121 local_irq_restore(flags);
1123 /* Re-raise the RCU softirq if there are callbacks remaining. */
1124 if (cpu_has_callbacks_ready_to_invoke(rdp))
1125 raise_softirq(RCU_SOFTIRQ);
1129 * Check to see if this CPU is in a non-context-switch quiescent state
1130 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1131 * Also schedule the RCU softirq handler.
1133 * This function must be called with hardirqs disabled. It is normally
1134 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1135 * false, there is no point in invoking rcu_check_callbacks().
1137 void rcu_check_callbacks(int cpu, int user)
1139 if (!rcu_pending(cpu))
1140 return; /* if nothing for RCU to do. */
1142 (idle_cpu(cpu) && rcu_scheduler_active &&
1143 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1146 * Get here if this CPU took its interrupt from user
1147 * mode or from the idle loop, and if this is not a
1148 * nested interrupt. In this case, the CPU is in
1149 * a quiescent state, so note it.
1151 * No memory barrier is required here because both
1152 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1153 * variables that other CPUs neither access nor modify,
1154 * at least not while the corresponding CPU is online.
1160 } else if (!in_softirq()) {
1163 * Get here if this CPU did not take its interrupt from
1164 * softirq, in other words, if it is not interrupting
1165 * a rcu_bh read-side critical section. This is an _bh
1166 * critical section, so note it.
1171 rcu_preempt_check_callbacks(cpu);
1172 raise_softirq(RCU_SOFTIRQ);
1178 * Scan the leaf rcu_node structures, processing dyntick state for any that
1179 * have not yet encountered a quiescent state, using the function specified.
1180 * The caller must have suppressed start of new grace periods.
1182 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1186 unsigned long flags;
1188 struct rcu_node *rnp;
1190 rcu_for_each_leaf_node(rsp, rnp) {
1192 raw_spin_lock_irqsave(&rnp->lock, flags);
1193 if (!rcu_gp_in_progress(rsp)) {
1194 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1197 if (rnp->qsmask == 0) {
1198 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1203 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1204 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1209 /* rcu_report_qs_rnp() releases rnp->lock. */
1210 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1213 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1218 * Force quiescent states on reluctant CPUs, and also detect which
1219 * CPUs are in dyntick-idle mode.
1221 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1223 unsigned long flags;
1224 struct rcu_node *rnp = rcu_get_root(rsp);
1226 if (!rcu_gp_in_progress(rsp))
1227 return; /* No grace period in progress, nothing to force. */
1228 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1229 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1230 return; /* Someone else is already on the job. */
1232 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1233 goto unlock_fqs_ret; /* no emergency and done recently. */
1235 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1236 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1237 if(!rcu_gp_in_progress(rsp)) {
1238 rsp->n_force_qs_ngp++;
1239 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1240 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1242 rsp->fqs_active = 1;
1243 switch (rsp->signaled) {
1247 break; /* grace period idle or initializing, ignore. */
1249 case RCU_SAVE_DYNTICK:
1250 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1251 break; /* So gcc recognizes the dead code. */
1253 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1255 /* Record dyntick-idle state. */
1256 force_qs_rnp(rsp, dyntick_save_progress_counter);
1257 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1258 if (rcu_gp_in_progress(rsp))
1259 rsp->signaled = RCU_FORCE_QS;
1264 /* Check dyntick-idle state, send IPI to laggarts. */
1265 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1266 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1268 /* Leave state in case more forcing is required. */
1270 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1273 rsp->fqs_active = 0;
1274 if (rsp->fqs_need_gp) {
1275 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1276 rsp->fqs_need_gp = 0;
1277 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1280 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1282 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1285 #else /* #ifdef CONFIG_SMP */
1287 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1292 #endif /* #else #ifdef CONFIG_SMP */
1295 * This does the RCU processing work from softirq context for the
1296 * specified rcu_state and rcu_data structures. This may be called
1297 * only from the CPU to whom the rdp belongs.
1300 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1302 unsigned long flags;
1304 WARN_ON_ONCE(rdp->beenonline == 0);
1307 * If an RCU GP has gone long enough, go check for dyntick
1308 * idle CPUs and, if needed, send resched IPIs.
1310 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1311 force_quiescent_state(rsp, 1);
1314 * Advance callbacks in response to end of earlier grace
1315 * period that some other CPU ended.
1317 rcu_process_gp_end(rsp, rdp);
1319 /* Update RCU state based on any recent quiescent states. */
1320 rcu_check_quiescent_state(rsp, rdp);
1322 /* Does this CPU require a not-yet-started grace period? */
1323 if (cpu_needs_another_gp(rsp, rdp)) {
1324 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1325 rcu_start_gp(rsp, flags); /* releases above lock */
1328 /* If there are callbacks ready, invoke them. */
1329 rcu_do_batch(rsp, rdp);
1333 * Do softirq processing for the current CPU.
1335 static void rcu_process_callbacks(struct softirq_action *unused)
1338 * Memory references from any prior RCU read-side critical sections
1339 * executed by the interrupted code must be seen before any RCU
1340 * grace-period manipulations below.
1342 smp_mb(); /* See above block comment. */
1344 __rcu_process_callbacks(&rcu_sched_state,
1345 &__get_cpu_var(rcu_sched_data));
1346 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1347 rcu_preempt_process_callbacks();
1350 * Memory references from any later RCU read-side critical sections
1351 * executed by the interrupted code must be seen after any RCU
1352 * grace-period manipulations above.
1354 smp_mb(); /* See above block comment. */
1356 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1357 rcu_needs_cpu_flush();
1361 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1362 struct rcu_state *rsp)
1364 unsigned long flags;
1365 struct rcu_data *rdp;
1370 smp_mb(); /* Ensure RCU update seen before callback registry. */
1373 * Opportunistically note grace-period endings and beginnings.
1374 * Note that we might see a beginning right after we see an
1375 * end, but never vice versa, since this CPU has to pass through
1376 * a quiescent state betweentimes.
1378 local_irq_save(flags);
1379 rdp = rsp->rda[smp_processor_id()];
1380 rcu_process_gp_end(rsp, rdp);
1381 check_for_new_grace_period(rsp, rdp);
1383 /* Add the callback to our list. */
1384 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1385 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1387 /* Start a new grace period if one not already started. */
1388 if (!rcu_gp_in_progress(rsp)) {
1389 unsigned long nestflag;
1390 struct rcu_node *rnp_root = rcu_get_root(rsp);
1392 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1393 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1397 * Force the grace period if too many callbacks or too long waiting.
1398 * Enforce hysteresis, and don't invoke force_quiescent_state()
1399 * if some other CPU has recently done so. Also, don't bother
1400 * invoking force_quiescent_state() if the newly enqueued callback
1401 * is the only one waiting for a grace period to complete.
1403 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1404 rdp->blimit = LONG_MAX;
1405 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1406 *rdp->nxttail[RCU_DONE_TAIL] != head)
1407 force_quiescent_state(rsp, 0);
1408 rdp->n_force_qs_snap = rsp->n_force_qs;
1409 rdp->qlen_last_fqs_check = rdp->qlen;
1410 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1411 force_quiescent_state(rsp, 1);
1412 local_irq_restore(flags);
1416 * Queue an RCU-sched callback for invocation after a grace period.
1418 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1420 __call_rcu(head, func, &rcu_sched_state);
1422 EXPORT_SYMBOL_GPL(call_rcu_sched);
1425 * Queue an RCU for invocation after a quicker grace period.
1427 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1429 __call_rcu(head, func, &rcu_bh_state);
1431 EXPORT_SYMBOL_GPL(call_rcu_bh);
1434 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1436 * Control will return to the caller some time after a full rcu-sched
1437 * grace period has elapsed, in other words after all currently executing
1438 * rcu-sched read-side critical sections have completed. These read-side
1439 * critical sections are delimited by rcu_read_lock_sched() and
1440 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1441 * local_irq_disable(), and so on may be used in place of
1442 * rcu_read_lock_sched().
1444 * This means that all preempt_disable code sequences, including NMI and
1445 * hardware-interrupt handlers, in progress on entry will have completed
1446 * before this primitive returns. However, this does not guarantee that
1447 * softirq handlers will have completed, since in some kernels, these
1448 * handlers can run in process context, and can block.
1450 * This primitive provides the guarantees made by the (now removed)
1451 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1452 * guarantees that rcu_read_lock() sections will have completed.
1453 * In "classic RCU", these two guarantees happen to be one and
1454 * the same, but can differ in realtime RCU implementations.
1456 void synchronize_sched(void)
1458 struct rcu_synchronize rcu;
1460 if (rcu_blocking_is_gp())
1463 init_completion(&rcu.completion);
1464 /* Will wake me after RCU finished. */
1465 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1467 wait_for_completion(&rcu.completion);
1469 EXPORT_SYMBOL_GPL(synchronize_sched);
1472 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1474 * Control will return to the caller some time after a full rcu_bh grace
1475 * period has elapsed, in other words after all currently executing rcu_bh
1476 * read-side critical sections have completed. RCU read-side critical
1477 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1478 * and may be nested.
1480 void synchronize_rcu_bh(void)
1482 struct rcu_synchronize rcu;
1484 if (rcu_blocking_is_gp())
1487 init_completion(&rcu.completion);
1488 /* Will wake me after RCU finished. */
1489 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1491 wait_for_completion(&rcu.completion);
1493 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1496 * Check to see if there is any immediate RCU-related work to be done
1497 * by the current CPU, for the specified type of RCU, returning 1 if so.
1498 * The checks are in order of increasing expense: checks that can be
1499 * carried out against CPU-local state are performed first. However,
1500 * we must check for CPU stalls first, else we might not get a chance.
1502 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1504 struct rcu_node *rnp = rdp->mynode;
1506 rdp->n_rcu_pending++;
1508 /* Check for CPU stalls, if enabled. */
1509 check_cpu_stall(rsp, rdp);
1511 /* Is the RCU core waiting for a quiescent state from this CPU? */
1512 if (rdp->qs_pending) {
1515 * If force_quiescent_state() coming soon and this CPU
1516 * needs a quiescent state, and this is either RCU-sched
1517 * or RCU-bh, force a local reschedule.
1519 if (!rdp->preemptable &&
1520 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1523 rdp->n_rp_qs_pending++;
1527 /* Does this CPU have callbacks ready to invoke? */
1528 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1529 rdp->n_rp_cb_ready++;
1533 /* Has RCU gone idle with this CPU needing another grace period? */
1534 if (cpu_needs_another_gp(rsp, rdp)) {
1535 rdp->n_rp_cpu_needs_gp++;
1539 /* Has another RCU grace period completed? */
1540 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1541 rdp->n_rp_gp_completed++;
1545 /* Has a new RCU grace period started? */
1546 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1547 rdp->n_rp_gp_started++;
1551 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1552 if (rcu_gp_in_progress(rsp) &&
1553 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1554 rdp->n_rp_need_fqs++;
1559 rdp->n_rp_need_nothing++;
1564 * Check to see if there is any immediate RCU-related work to be done
1565 * by the current CPU, returning 1 if so. This function is part of the
1566 * RCU implementation; it is -not- an exported member of the RCU API.
1568 static int rcu_pending(int cpu)
1570 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1571 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1572 rcu_preempt_pending(cpu);
1576 * Check to see if any future RCU-related work will need to be done
1577 * by the current CPU, even if none need be done immediately, returning
1580 static int rcu_needs_cpu_quick_check(int cpu)
1582 /* RCU callbacks either ready or pending? */
1583 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1584 per_cpu(rcu_bh_data, cpu).nxtlist ||
1585 rcu_preempt_needs_cpu(cpu);
1588 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1589 static atomic_t rcu_barrier_cpu_count;
1590 static DEFINE_MUTEX(rcu_barrier_mutex);
1591 static struct completion rcu_barrier_completion;
1593 static void rcu_barrier_callback(struct rcu_head *notused)
1595 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1596 complete(&rcu_barrier_completion);
1600 * Called with preemption disabled, and from cross-cpu IRQ context.
1602 static void rcu_barrier_func(void *type)
1604 int cpu = smp_processor_id();
1605 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1606 void (*call_rcu_func)(struct rcu_head *head,
1607 void (*func)(struct rcu_head *head));
1609 atomic_inc(&rcu_barrier_cpu_count);
1610 call_rcu_func = type;
1611 call_rcu_func(head, rcu_barrier_callback);
1615 * Orchestrate the specified type of RCU barrier, waiting for all
1616 * RCU callbacks of the specified type to complete.
1618 static void _rcu_barrier(struct rcu_state *rsp,
1619 void (*call_rcu_func)(struct rcu_head *head,
1620 void (*func)(struct rcu_head *head)))
1622 BUG_ON(in_interrupt());
1623 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1624 mutex_lock(&rcu_barrier_mutex);
1625 init_completion(&rcu_barrier_completion);
1627 * Initialize rcu_barrier_cpu_count to 1, then invoke
1628 * rcu_barrier_func() on each CPU, so that each CPU also has
1629 * incremented rcu_barrier_cpu_count. Only then is it safe to
1630 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1631 * might complete its grace period before all of the other CPUs
1632 * did their increment, causing this function to return too
1635 atomic_set(&rcu_barrier_cpu_count, 1);
1636 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1637 rcu_adopt_orphan_cbs(rsp);
1638 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1639 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1640 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1641 complete(&rcu_barrier_completion);
1642 wait_for_completion(&rcu_barrier_completion);
1643 mutex_unlock(&rcu_barrier_mutex);
1647 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1649 void rcu_barrier_bh(void)
1651 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1653 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1656 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1658 void rcu_barrier_sched(void)
1660 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1662 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1665 * Do boot-time initialization of a CPU's per-CPU RCU data.
1668 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1670 unsigned long flags;
1672 struct rcu_data *rdp = rsp->rda[cpu];
1673 struct rcu_node *rnp = rcu_get_root(rsp);
1675 /* Set up local state, ensuring consistent view of global state. */
1676 raw_spin_lock_irqsave(&rnp->lock, flags);
1677 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1678 rdp->nxtlist = NULL;
1679 for (i = 0; i < RCU_NEXT_SIZE; i++)
1680 rdp->nxttail[i] = &rdp->nxtlist;
1683 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1684 #endif /* #ifdef CONFIG_NO_HZ */
1686 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1690 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1691 * offline event can be happening at a given time. Note also that we
1692 * can accept some slop in the rsp->completed access due to the fact
1693 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1695 static void __cpuinit
1696 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1698 unsigned long flags;
1700 struct rcu_data *rdp = rsp->rda[cpu];
1701 struct rcu_node *rnp = rcu_get_root(rsp);
1703 /* Set up local state, ensuring consistent view of global state. */
1704 raw_spin_lock_irqsave(&rnp->lock, flags);
1705 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1706 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1707 rdp->beenonline = 1; /* We have now been online. */
1708 rdp->preemptable = preemptable;
1709 rdp->qlen_last_fqs_check = 0;
1710 rdp->n_force_qs_snap = rsp->n_force_qs;
1711 rdp->blimit = blimit;
1712 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1715 * A new grace period might start here. If so, we won't be part
1716 * of it, but that is OK, as we are currently in a quiescent state.
1719 /* Exclude any attempts to start a new GP on large systems. */
1720 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1722 /* Add CPU to rcu_node bitmasks. */
1724 mask = rdp->grpmask;
1726 /* Exclude any attempts to start a new GP on small systems. */
1727 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1728 rnp->qsmaskinit |= mask;
1729 mask = rnp->grpmask;
1730 if (rnp == rdp->mynode) {
1731 rdp->gpnum = rnp->completed; /* if GP in progress... */
1732 rdp->completed = rnp->completed;
1733 rdp->passed_quiesc_completed = rnp->completed - 1;
1735 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1737 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1739 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1742 static void __cpuinit rcu_online_cpu(int cpu)
1744 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1745 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1746 rcu_preempt_init_percpu_data(cpu);
1750 * Handle CPU online/offline notification events.
1752 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1753 unsigned long action, void *hcpu)
1755 long cpu = (long)hcpu;
1758 case CPU_UP_PREPARE:
1759 case CPU_UP_PREPARE_FROZEN:
1760 rcu_online_cpu(cpu);
1763 case CPU_DYING_FROZEN:
1765 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1766 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1767 * returns, all online cpus have queued rcu_barrier_func().
1768 * The dying CPU clears its cpu_online_mask bit and
1769 * moves all of its RCU callbacks to ->orphan_cbs_list
1770 * in the context of stop_machine(), so subsequent calls
1771 * to _rcu_barrier() will adopt these callbacks and only
1772 * then queue rcu_barrier_func() on all remaining CPUs.
1774 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1775 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1776 rcu_preempt_send_cbs_to_orphanage();
1779 case CPU_DEAD_FROZEN:
1780 case CPU_UP_CANCELED:
1781 case CPU_UP_CANCELED_FROZEN:
1782 rcu_offline_cpu(cpu);
1791 * This function is invoked towards the end of the scheduler's initialization
1792 * process. Before this is called, the idle task might contain
1793 * RCU read-side critical sections (during which time, this idle
1794 * task is booting the system). After this function is called, the
1795 * idle tasks are prohibited from containing RCU read-side critical
1796 * sections. This function also enables RCU lockdep checking.
1798 void rcu_scheduler_starting(void)
1800 WARN_ON(num_online_cpus() != 1);
1801 WARN_ON(nr_context_switches() > 0);
1802 rcu_scheduler_active = 1;
1806 * Compute the per-level fanout, either using the exact fanout specified
1807 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1809 #ifdef CONFIG_RCU_FANOUT_EXACT
1810 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1814 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1815 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1817 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1818 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1825 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1826 ccur = rsp->levelcnt[i];
1827 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1831 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1834 * Helper function for rcu_init() that initializes one rcu_state structure.
1836 static void __init rcu_init_one(struct rcu_state *rsp)
1838 static char *buf[] = { "rcu_node_level_0",
1841 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1845 struct rcu_node *rnp;
1847 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1849 /* Initialize the level-tracking arrays. */
1851 for (i = 1; i < NUM_RCU_LVLS; i++)
1852 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1853 rcu_init_levelspread(rsp);
1855 /* Initialize the elements themselves, starting from the leaves. */
1857 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1858 cpustride *= rsp->levelspread[i];
1859 rnp = rsp->level[i];
1860 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1861 raw_spin_lock_init(&rnp->lock);
1862 lockdep_set_class_and_name(&rnp->lock,
1863 &rcu_node_class[i], buf[i]);
1866 rnp->qsmaskinit = 0;
1867 rnp->grplo = j * cpustride;
1868 rnp->grphi = (j + 1) * cpustride - 1;
1869 if (rnp->grphi >= NR_CPUS)
1870 rnp->grphi = NR_CPUS - 1;
1876 rnp->grpnum = j % rsp->levelspread[i - 1];
1877 rnp->grpmask = 1UL << rnp->grpnum;
1878 rnp->parent = rsp->level[i - 1] +
1879 j / rsp->levelspread[i - 1];
1882 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1883 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1884 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1885 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1889 rnp = rsp->level[NUM_RCU_LVLS - 1];
1890 for_each_possible_cpu(i) {
1893 rsp->rda[i]->mynode = rnp;
1894 rcu_boot_init_percpu_data(i, rsp);
1899 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1900 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1903 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1907 for_each_possible_cpu(i) { \
1908 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1910 rcu_init_one(rsp); \
1913 void __init rcu_init(void)
1917 rcu_bootup_announce();
1918 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1919 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1920 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1921 #if NUM_RCU_LVL_4 != 0
1922 printk(KERN_INFO "Experimental four-level hierarchy is enabled.\n");
1923 #endif /* #if NUM_RCU_LVL_4 != 0 */
1924 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1925 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1926 __rcu_init_preempt();
1927 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1930 * We don't need protection against CPU-hotplug here because
1931 * this is called early in boot, before either interrupts
1932 * or the scheduler are operational.
1934 cpu_notifier(rcu_cpu_notify, 0);
1935 for_each_online_cpu(cpu)
1936 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1939 #include "rcutree_plugin.h"