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(structname) { \
58 .level = { &structname.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(&structname.onofflock), \
70 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
72 .n_force_qs_ngp = 0, \
73 .name = #structname, \
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);
82 int rcu_scheduler_active __read_mostly;
83 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
86 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
87 * permit this function to be invoked without holding the root rcu_node
88 * structure's ->lock, but of course results can be subject to change.
90 static int rcu_gp_in_progress(struct rcu_state *rsp)
92 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
96 * Note a quiescent state. Because we do not need to know
97 * how many quiescent states passed, just if there was at least
98 * one since the start of the grace period, this just sets a flag.
100 void rcu_sched_qs(int cpu)
102 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
104 rdp->passed_quiesc_completed = rdp->gpnum - 1;
106 rdp->passed_quiesc = 1;
109 void rcu_bh_qs(int cpu)
111 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
113 rdp->passed_quiesc_completed = rdp->gpnum - 1;
115 rdp->passed_quiesc = 1;
119 * Note a context switch. This is a quiescent state for RCU-sched,
120 * and requires special handling for preemptible RCU.
122 void rcu_note_context_switch(int cpu)
125 rcu_preempt_note_context_switch(cpu);
129 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
130 .dynticks_nesting = 1,
131 .dynticks = ATOMIC_INIT(1),
133 #endif /* #ifdef CONFIG_NO_HZ */
135 static int blimit = 10; /* Maximum callbacks per softirq. */
136 static int qhimark = 10000; /* If this many pending, ignore blimit. */
137 static int qlowmark = 100; /* Once only this many pending, use blimit. */
139 module_param(blimit, int, 0);
140 module_param(qhimark, int, 0);
141 module_param(qlowmark, int, 0);
143 int rcu_cpu_stall_suppress __read_mostly;
144 module_param(rcu_cpu_stall_suppress, int, 0644);
146 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
147 static int rcu_pending(int cpu);
150 * Return the number of RCU-sched batches processed thus far for debug & stats.
152 long rcu_batches_completed_sched(void)
154 return rcu_sched_state.completed;
156 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
159 * Return the number of RCU BH batches processed thus far for debug & stats.
161 long rcu_batches_completed_bh(void)
163 return rcu_bh_state.completed;
165 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
168 * Force a quiescent state for RCU BH.
170 void rcu_bh_force_quiescent_state(void)
172 force_quiescent_state(&rcu_bh_state, 0);
174 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
177 * Force a quiescent state for RCU-sched.
179 void rcu_sched_force_quiescent_state(void)
181 force_quiescent_state(&rcu_sched_state, 0);
183 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
186 * Does the CPU have callbacks ready to be invoked?
189 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
191 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
195 * Does the current CPU require a yet-as-unscheduled grace period?
198 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
200 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
204 * Return the root node of the specified rcu_state structure.
206 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
208 return &rsp->node[0];
214 * If the specified CPU is offline, tell the caller that it is in
215 * a quiescent state. Otherwise, whack it with a reschedule IPI.
216 * Grace periods can end up waiting on an offline CPU when that
217 * CPU is in the process of coming online -- it will be added to the
218 * rcu_node bitmasks before it actually makes it online. The same thing
219 * can happen while a CPU is in the process of coming online. Because this
220 * race is quite rare, we check for it after detecting that the grace
221 * period has been delayed rather than checking each and every CPU
222 * each and every time we start a new grace period.
224 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
227 * If the CPU is offline, it is in a quiescent state. We can
228 * trust its state not to change because interrupts are disabled.
230 if (cpu_is_offline(rdp->cpu)) {
235 /* If preemptable RCU, no point in sending reschedule IPI. */
236 if (rdp->preemptable)
239 /* The CPU is online, so send it a reschedule IPI. */
240 if (rdp->cpu != smp_processor_id())
241 smp_send_reschedule(rdp->cpu);
248 #endif /* #ifdef CONFIG_SMP */
253 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
255 * Enter nohz mode, in other words, -leave- the mode in which RCU
256 * read-side critical sections can occur. (Though RCU read-side
257 * critical sections can occur in irq handlers in nohz mode, a possibility
258 * handled by rcu_irq_enter() and rcu_irq_exit()).
260 void rcu_enter_nohz(void)
263 struct rcu_dynticks *rdtp;
265 local_irq_save(flags);
266 rdtp = &__get_cpu_var(rcu_dynticks);
267 if (--rdtp->dynticks_nesting) {
268 local_irq_restore(flags);
271 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
272 smp_mb__before_atomic_inc(); /* See above. */
273 atomic_inc(&rdtp->dynticks);
274 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
275 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
276 local_irq_restore(flags);
278 /* If the interrupt queued a callback, get out of dyntick mode. */
280 (__get_cpu_var(rcu_sched_data).nxtlist ||
281 __get_cpu_var(rcu_bh_data).nxtlist ||
282 rcu_preempt_needs_cpu(smp_processor_id())))
287 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
289 * Exit nohz mode, in other words, -enter- the mode in which RCU
290 * read-side critical sections normally occur.
292 void rcu_exit_nohz(void)
295 struct rcu_dynticks *rdtp;
297 local_irq_save(flags);
298 rdtp = &__get_cpu_var(rcu_dynticks);
299 if (rdtp->dynticks_nesting++) {
300 local_irq_restore(flags);
303 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
304 atomic_inc(&rdtp->dynticks);
305 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
306 smp_mb__after_atomic_inc(); /* See above. */
307 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
308 local_irq_restore(flags);
312 * rcu_nmi_enter - inform RCU of entry to 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 active.
318 void rcu_nmi_enter(void)
320 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
322 if (rdtp->dynticks_nmi_nesting == 0 &&
323 (atomic_read(&rdtp->dynticks) & 0x1))
325 rdtp->dynticks_nmi_nesting++;
326 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
327 atomic_inc(&rdtp->dynticks);
328 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
329 smp_mb__after_atomic_inc(); /* See above. */
330 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
334 * rcu_nmi_exit - inform RCU of exit from NMI context
336 * If the CPU was idle with dynamic ticks active, and there is no
337 * irq handler running, this updates rdtp->dynticks_nmi to let the
338 * RCU grace-period handling know that the CPU is no longer active.
340 void rcu_nmi_exit(void)
342 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
344 if (rdtp->dynticks_nmi_nesting == 0 ||
345 --rdtp->dynticks_nmi_nesting != 0)
347 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
348 smp_mb__before_atomic_inc(); /* See above. */
349 atomic_inc(&rdtp->dynticks);
350 smp_mb__after_atomic_inc(); /* Force delay to next write. */
351 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
355 * rcu_irq_enter - inform RCU of entry to hard irq context
357 * If the CPU was idle with dynamic ticks active, this updates the
358 * rdtp->dynticks to let the RCU handling know that the CPU is active.
360 void rcu_irq_enter(void)
366 * rcu_irq_exit - inform RCU of exit from hard irq context
368 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
369 * to put let the RCU handling be aware that the CPU is going back to idle
372 void rcu_irq_exit(void)
380 * Snapshot the specified CPU's dynticks counter so that we can later
381 * credit them with an implicit quiescent state. Return 1 if this CPU
382 * is in dynticks idle mode, which is an extended quiescent state.
384 static int dyntick_save_progress_counter(struct rcu_data *rdp)
386 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
391 * Return true if the specified CPU has passed through a quiescent
392 * state by virtue of being in or having passed through an dynticks
393 * idle state since the last call to dyntick_save_progress_counter()
396 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
401 curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
402 snap = (unsigned long)rdp->dynticks_snap;
405 * If the CPU passed through or entered a dynticks idle phase with
406 * no active irq/NMI handlers, then we can safely pretend that the CPU
407 * already acknowledged the request to pass through a quiescent
408 * state. Either way, that CPU cannot possibly be in an RCU
409 * read-side critical section that started before the beginning
410 * of the current RCU grace period.
412 if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
417 /* Go check for the CPU being offline. */
418 return rcu_implicit_offline_qs(rdp);
421 #endif /* #ifdef CONFIG_SMP */
423 #else /* #ifdef CONFIG_NO_HZ */
427 static int dyntick_save_progress_counter(struct rcu_data *rdp)
432 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
434 return rcu_implicit_offline_qs(rdp);
437 #endif /* #ifdef CONFIG_SMP */
439 #endif /* #else #ifdef CONFIG_NO_HZ */
441 int rcu_cpu_stall_suppress __read_mostly;
443 static void record_gp_stall_check_time(struct rcu_state *rsp)
445 rsp->gp_start = jiffies;
446 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
449 static void print_other_cpu_stall(struct rcu_state *rsp)
454 struct rcu_node *rnp = rcu_get_root(rsp);
456 /* Only let one CPU complain about others per time interval. */
458 raw_spin_lock_irqsave(&rnp->lock, flags);
459 delta = jiffies - rsp->jiffies_stall;
460 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
461 raw_spin_unlock_irqrestore(&rnp->lock, flags);
464 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
467 * Now rat on any tasks that got kicked up to the root rcu_node
468 * due to CPU offlining.
470 rcu_print_task_stall(rnp);
471 raw_spin_unlock_irqrestore(&rnp->lock, flags);
474 * OK, time to rat on our buddy...
475 * See Documentation/RCU/stallwarn.txt for info on how to debug
476 * RCU CPU stall warnings.
478 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
480 rcu_for_each_leaf_node(rsp, rnp) {
481 raw_spin_lock_irqsave(&rnp->lock, flags);
482 rcu_print_task_stall(rnp);
483 raw_spin_unlock_irqrestore(&rnp->lock, flags);
484 if (rnp->qsmask == 0)
486 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
487 if (rnp->qsmask & (1UL << cpu))
488 printk(" %d", rnp->grplo + cpu);
490 printk("} (detected by %d, t=%ld jiffies)\n",
491 smp_processor_id(), (long)(jiffies - rsp->gp_start));
492 trigger_all_cpu_backtrace();
494 /* If so configured, complain about tasks blocking the grace period. */
496 rcu_print_detail_task_stall(rsp);
498 force_quiescent_state(rsp, 0); /* Kick them all. */
501 static void print_cpu_stall(struct rcu_state *rsp)
504 struct rcu_node *rnp = rcu_get_root(rsp);
507 * OK, time to rat on ourselves...
508 * See Documentation/RCU/stallwarn.txt for info on how to debug
509 * RCU CPU stall warnings.
511 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
512 rsp->name, 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 if (rcu_cpu_stall_suppress)
531 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
533 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && delta >= 0) {
535 /* We haven't checked in, so go dump stack. */
536 print_cpu_stall(rsp);
538 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
540 /* They had two time units to dump stack, so complain. */
541 print_other_cpu_stall(rsp);
545 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
547 rcu_cpu_stall_suppress = 1;
552 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
554 * Set the stall-warning timeout way off into the future, thus preventing
555 * any RCU CPU stall-warning messages from appearing in the current set of
558 * The caller must disable hard irqs.
560 void rcu_cpu_stall_reset(void)
562 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
563 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
564 rcu_preempt_stall_reset();
567 static struct notifier_block rcu_panic_block = {
568 .notifier_call = rcu_panic,
571 static void __init check_cpu_stall_init(void)
573 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
577 * Update CPU-local rcu_data state to record the newly noticed grace period.
578 * This is used both when we started the grace period and when we notice
579 * that someone else started the grace period. The caller must hold the
580 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
581 * and must have irqs disabled.
583 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
585 if (rdp->gpnum != rnp->gpnum) {
587 * If the current grace period is waiting for this CPU,
588 * set up to detect a quiescent state, otherwise don't
589 * go looking for one.
591 rdp->gpnum = rnp->gpnum;
592 if (rnp->qsmask & rdp->grpmask) {
594 rdp->passed_quiesc = 0;
600 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
603 struct rcu_node *rnp;
605 local_irq_save(flags);
607 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
608 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
609 local_irq_restore(flags);
612 __note_new_gpnum(rsp, rnp, rdp);
613 raw_spin_unlock_irqrestore(&rnp->lock, flags);
617 * Did someone else start a new RCU grace period start since we last
618 * checked? Update local state appropriately if so. Must be called
619 * on the CPU corresponding to rdp.
622 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
627 local_irq_save(flags);
628 if (rdp->gpnum != rsp->gpnum) {
629 note_new_gpnum(rsp, rdp);
632 local_irq_restore(flags);
637 * Advance this CPU's callbacks, but only if the current grace period
638 * has ended. This may be called only from the CPU to whom the rdp
639 * belongs. In addition, the corresponding leaf rcu_node structure's
640 * ->lock must be held by the caller, with irqs disabled.
643 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
645 /* Did another grace period end? */
646 if (rdp->completed != rnp->completed) {
648 /* Advance callbacks. No harm if list empty. */
649 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
650 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
651 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
653 /* Remember that we saw this grace-period completion. */
654 rdp->completed = rnp->completed;
657 * If we were in an extended quiescent state, we may have
658 * missed some grace periods that others CPUs handled on
659 * our behalf. Catch up with this state to avoid noting
660 * spurious new grace periods. If another grace period
661 * has started, then rnp->gpnum will have advanced, so
662 * we will detect this later on.
664 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
665 rdp->gpnum = rdp->completed;
668 * If RCU does not need a quiescent state from this CPU,
669 * then make sure that this CPU doesn't go looking for one.
671 if ((rnp->qsmask & rdp->grpmask) == 0)
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)
685 struct rcu_node *rnp;
687 local_irq_save(flags);
689 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
690 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
691 local_irq_restore(flags);
694 __rcu_process_gp_end(rsp, rnp, rdp);
695 raw_spin_unlock_irqrestore(&rnp->lock, flags);
699 * Do per-CPU grace-period initialization for running CPU. The caller
700 * must hold the lock of the leaf rcu_node structure corresponding to
704 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
706 /* Prior grace period ended, so advance callbacks for current CPU. */
707 __rcu_process_gp_end(rsp, rnp, rdp);
710 * Because this CPU just now started the new grace period, we know
711 * that all of its callbacks will be covered by this upcoming grace
712 * period, even the ones that were registered arbitrarily recently.
713 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
715 * Other CPUs cannot be sure exactly when the grace period started.
716 * Therefore, their recently registered callbacks must pass through
717 * an additional RCU_NEXT_READY stage, so that they will be handled
718 * by the next RCU grace period.
720 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
721 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
723 /* Set state so that this CPU will detect the next quiescent state. */
724 __note_new_gpnum(rsp, rnp, rdp);
728 * Start a new RCU grace period if warranted, re-initializing the hierarchy
729 * in preparation for detecting the next grace period. The caller must hold
730 * the root node's ->lock, which is released before return. Hard irqs must
734 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
735 __releases(rcu_get_root(rsp)->lock)
737 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
738 struct rcu_node *rnp = rcu_get_root(rsp);
740 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
741 if (cpu_needs_another_gp(rsp, rdp))
742 rsp->fqs_need_gp = 1;
743 if (rnp->completed == rsp->completed) {
744 raw_spin_unlock_irqrestore(&rnp->lock, flags);
747 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
750 * Propagate new ->completed value to rcu_node structures
751 * so that other CPUs don't have to wait until the start
752 * of the next grace period to process their callbacks.
754 rcu_for_each_node_breadth_first(rsp, rnp) {
755 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
756 rnp->completed = rsp->completed;
757 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
759 local_irq_restore(flags);
763 /* Advance to a new grace period and initialize state. */
765 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
766 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
767 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
768 record_gp_stall_check_time(rsp);
770 /* Special-case the common single-level case. */
771 if (NUM_RCU_NODES == 1) {
772 rcu_preempt_check_blocked_tasks(rnp);
773 rnp->qsmask = rnp->qsmaskinit;
774 rnp->gpnum = rsp->gpnum;
775 rnp->completed = rsp->completed;
776 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
777 rcu_start_gp_per_cpu(rsp, rnp, rdp);
778 raw_spin_unlock_irqrestore(&rnp->lock, flags);
782 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
785 /* Exclude any concurrent CPU-hotplug operations. */
786 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
789 * Set the quiescent-state-needed bits in all the rcu_node
790 * structures for all currently online CPUs in breadth-first
791 * order, starting from the root rcu_node structure. This
792 * operation relies on the layout of the hierarchy within the
793 * rsp->node[] array. Note that other CPUs will access only
794 * the leaves of the hierarchy, which still indicate that no
795 * grace period is in progress, at least until the corresponding
796 * leaf node has been initialized. In addition, we have excluded
797 * CPU-hotplug operations.
799 * Note that the grace period cannot complete until we finish
800 * the initialization process, as there will be at least one
801 * qsmask bit set in the root node until that time, namely the
802 * one corresponding to this CPU, due to the fact that we have
805 rcu_for_each_node_breadth_first(rsp, rnp) {
806 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
807 rcu_preempt_check_blocked_tasks(rnp);
808 rnp->qsmask = rnp->qsmaskinit;
809 rnp->gpnum = rsp->gpnum;
810 rnp->completed = rsp->completed;
811 if (rnp == rdp->mynode)
812 rcu_start_gp_per_cpu(rsp, rnp, rdp);
813 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
816 rnp = rcu_get_root(rsp);
817 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
818 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
819 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
820 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
824 * Report a full set of quiescent states to the specified rcu_state
825 * data structure. This involves cleaning up after the prior grace
826 * period and letting rcu_start_gp() start up the next grace period
827 * if one is needed. Note that the caller must hold rnp->lock, as
828 * required by rcu_start_gp(), which will release it.
830 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
831 __releases(rcu_get_root(rsp)->lock)
833 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
836 * Ensure that all grace-period and pre-grace-period activity
837 * is seen before the assignment to rsp->completed.
839 smp_mb(); /* See above block comment. */
840 rsp->completed = rsp->gpnum;
841 rsp->signaled = RCU_GP_IDLE;
842 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
846 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
847 * Allows quiescent states for a group of CPUs to be reported at one go
848 * to the specified rcu_node structure, though all the CPUs in the group
849 * must be represented by the same rcu_node structure (which need not be
850 * a leaf rcu_node structure, though it often will be). That structure's
851 * lock must be held upon entry, and it is released before return.
854 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
855 struct rcu_node *rnp, unsigned long flags)
856 __releases(rnp->lock)
858 struct rcu_node *rnp_c;
860 /* Walk up the rcu_node hierarchy. */
862 if (!(rnp->qsmask & mask)) {
864 /* Our bit has already been cleared, so done. */
865 raw_spin_unlock_irqrestore(&rnp->lock, flags);
868 rnp->qsmask &= ~mask;
869 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
871 /* Other bits still set at this level, so done. */
872 raw_spin_unlock_irqrestore(&rnp->lock, flags);
876 if (rnp->parent == NULL) {
878 /* No more levels. Exit loop holding root lock. */
882 raw_spin_unlock_irqrestore(&rnp->lock, flags);
885 raw_spin_lock_irqsave(&rnp->lock, flags);
886 WARN_ON_ONCE(rnp_c->qsmask);
890 * Get here if we are the last CPU to pass through a quiescent
891 * state for this grace period. Invoke rcu_report_qs_rsp()
892 * to clean up and start the next grace period if one is needed.
894 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
898 * Record a quiescent state for the specified CPU to that CPU's rcu_data
899 * structure. This must be either called from the specified CPU, or
900 * called when the specified CPU is known to be offline (and when it is
901 * also known that no other CPU is concurrently trying to help the offline
902 * CPU). The lastcomp argument is used to make sure we are still in the
903 * grace period of interest. We don't want to end the current grace period
904 * based on quiescent states detected in an earlier grace period!
907 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
911 struct rcu_node *rnp;
914 raw_spin_lock_irqsave(&rnp->lock, flags);
915 if (lastcomp != rnp->completed) {
918 * Someone beat us to it for this grace period, so leave.
919 * The race with GP start is resolved by the fact that we
920 * hold the leaf rcu_node lock, so that the per-CPU bits
921 * cannot yet be initialized -- so we would simply find our
922 * CPU's bit already cleared in rcu_report_qs_rnp() if this
925 rdp->passed_quiesc = 0; /* try again later! */
926 raw_spin_unlock_irqrestore(&rnp->lock, flags);
930 if ((rnp->qsmask & mask) == 0) {
931 raw_spin_unlock_irqrestore(&rnp->lock, flags);
936 * This GP can't end until cpu checks in, so all of our
937 * callbacks can be processed during the next GP.
939 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
941 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
946 * Check to see if there is a new grace period of which this CPU
947 * is not yet aware, and if so, set up local rcu_data state for it.
948 * Otherwise, see if this CPU has just passed through its first
949 * quiescent state for this grace period, and record that fact if so.
952 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
954 /* If there is now a new grace period, record and return. */
955 if (check_for_new_grace_period(rsp, rdp))
959 * Does this CPU still need to do its part for current grace period?
960 * If no, return and let the other CPUs do their part as well.
962 if (!rdp->qs_pending)
966 * Was there a quiescent state since the beginning of the grace
967 * period? If no, then exit and wait for the next call.
969 if (!rdp->passed_quiesc)
973 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
976 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
979 #ifdef CONFIG_HOTPLUG_CPU
982 * Move a dying CPU's RCU callbacks to online CPU's callback list.
983 * Synchronization is not required because this function executes
984 * in stop_machine() context.
986 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
989 /* current DYING CPU is cleared in the cpu_online_mask */
990 int receive_cpu = cpumask_any(cpu_online_mask);
991 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
992 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
994 if (rdp->nxtlist == NULL)
995 return; /* irqs disabled, so comparison is stable. */
997 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
998 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
999 receive_rdp->qlen += rdp->qlen;
1000 receive_rdp->n_cbs_adopted += rdp->qlen;
1001 rdp->n_cbs_orphaned += rdp->qlen;
1003 rdp->nxtlist = NULL;
1004 for (i = 0; i < RCU_NEXT_SIZE; i++)
1005 rdp->nxttail[i] = &rdp->nxtlist;
1010 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1011 * and move all callbacks from the outgoing CPU to the current one.
1013 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1015 unsigned long flags;
1017 int need_report = 0;
1018 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1019 struct rcu_node *rnp;
1021 /* Exclude any attempts to start a new grace period. */
1022 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1024 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1025 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1026 mask = rdp->grpmask; /* rnp->grplo is constant. */
1028 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1029 rnp->qsmaskinit &= ~mask;
1030 if (rnp->qsmaskinit != 0) {
1031 if (rnp != rdp->mynode)
1032 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1035 if (rnp == rdp->mynode)
1036 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1038 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1039 mask = rnp->grpmask;
1041 } while (rnp != NULL);
1044 * We still hold the leaf rcu_node structure lock here, and
1045 * irqs are still disabled. The reason for this subterfuge is
1046 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1047 * held leads to deadlock.
1049 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1051 if (need_report & RCU_OFL_TASKS_NORM_GP)
1052 rcu_report_unblock_qs_rnp(rnp, flags);
1054 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1055 if (need_report & RCU_OFL_TASKS_EXP_GP)
1056 rcu_report_exp_rnp(rsp, rnp);
1060 * Remove the specified CPU from the RCU hierarchy and move any pending
1061 * callbacks that it might have to the current CPU. This code assumes
1062 * that at least one CPU in the system will remain running at all times.
1063 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1065 static void rcu_offline_cpu(int cpu)
1067 __rcu_offline_cpu(cpu, &rcu_sched_state);
1068 __rcu_offline_cpu(cpu, &rcu_bh_state);
1069 rcu_preempt_offline_cpu(cpu);
1072 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1074 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1078 static void rcu_offline_cpu(int cpu)
1082 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1085 * Invoke any RCU callbacks that have made it to the end of their grace
1086 * period. Thottle as specified by rdp->blimit.
1088 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1090 unsigned long flags;
1091 struct rcu_head *next, *list, **tail;
1094 /* If no callbacks are ready, just return.*/
1095 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1099 * Extract the list of ready callbacks, disabling to prevent
1100 * races with call_rcu() from interrupt handlers.
1102 local_irq_save(flags);
1103 list = rdp->nxtlist;
1104 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1105 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1106 tail = rdp->nxttail[RCU_DONE_TAIL];
1107 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1108 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1109 rdp->nxttail[count] = &rdp->nxtlist;
1110 local_irq_restore(flags);
1112 /* Invoke callbacks. */
1117 debug_rcu_head_unqueue(list);
1120 if (++count >= rdp->blimit)
1124 local_irq_save(flags);
1126 /* Update count, and requeue any remaining callbacks. */
1128 rdp->n_cbs_invoked += count;
1130 *tail = rdp->nxtlist;
1131 rdp->nxtlist = list;
1132 for (count = 0; count < RCU_NEXT_SIZE; count++)
1133 if (&rdp->nxtlist == rdp->nxttail[count])
1134 rdp->nxttail[count] = tail;
1139 /* Reinstate batch limit if we have worked down the excess. */
1140 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1141 rdp->blimit = blimit;
1143 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1144 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1145 rdp->qlen_last_fqs_check = 0;
1146 rdp->n_force_qs_snap = rsp->n_force_qs;
1147 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1148 rdp->qlen_last_fqs_check = rdp->qlen;
1150 local_irq_restore(flags);
1152 /* Re-raise the RCU softirq if there are callbacks remaining. */
1153 if (cpu_has_callbacks_ready_to_invoke(rdp))
1154 raise_softirq(RCU_SOFTIRQ);
1158 * Check to see if this CPU is in a non-context-switch quiescent state
1159 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1160 * Also schedule the RCU softirq handler.
1162 * This function must be called with hardirqs disabled. It is normally
1163 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1164 * false, there is no point in invoking rcu_check_callbacks().
1166 void rcu_check_callbacks(int cpu, int user)
1169 (idle_cpu(cpu) && rcu_scheduler_active &&
1170 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1173 * Get here if this CPU took its interrupt from user
1174 * mode or from the idle loop, and if this is not a
1175 * nested interrupt. In this case, the CPU is in
1176 * a quiescent state, so note it.
1178 * No memory barrier is required here because both
1179 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1180 * variables that other CPUs neither access nor modify,
1181 * at least not while the corresponding CPU is online.
1187 } else if (!in_softirq()) {
1190 * Get here if this CPU did not take its interrupt from
1191 * softirq, in other words, if it is not interrupting
1192 * a rcu_bh read-side critical section. This is an _bh
1193 * critical section, so note it.
1198 rcu_preempt_check_callbacks(cpu);
1199 if (rcu_pending(cpu))
1200 raise_softirq(RCU_SOFTIRQ);
1206 * Scan the leaf rcu_node structures, processing dyntick state for any that
1207 * have not yet encountered a quiescent state, using the function specified.
1208 * The caller must have suppressed start of new grace periods.
1210 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1214 unsigned long flags;
1216 struct rcu_node *rnp;
1218 rcu_for_each_leaf_node(rsp, rnp) {
1220 raw_spin_lock_irqsave(&rnp->lock, flags);
1221 if (!rcu_gp_in_progress(rsp)) {
1222 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1225 if (rnp->qsmask == 0) {
1226 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1231 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1232 if ((rnp->qsmask & bit) != 0 &&
1233 f(per_cpu_ptr(rsp->rda, cpu)))
1238 /* rcu_report_qs_rnp() releases rnp->lock. */
1239 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1242 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1247 * Force quiescent states on reluctant CPUs, and also detect which
1248 * CPUs are in dyntick-idle mode.
1250 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1252 unsigned long flags;
1253 struct rcu_node *rnp = rcu_get_root(rsp);
1255 if (!rcu_gp_in_progress(rsp))
1256 return; /* No grace period in progress, nothing to force. */
1257 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1258 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1259 return; /* Someone else is already on the job. */
1261 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1262 goto unlock_fqs_ret; /* no emergency and done recently. */
1264 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1265 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1266 if(!rcu_gp_in_progress(rsp)) {
1267 rsp->n_force_qs_ngp++;
1268 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1269 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1271 rsp->fqs_active = 1;
1272 switch (rsp->signaled) {
1276 break; /* grace period idle or initializing, ignore. */
1278 case RCU_SAVE_DYNTICK:
1279 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1280 break; /* So gcc recognizes the dead code. */
1282 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1284 /* Record dyntick-idle state. */
1285 force_qs_rnp(rsp, dyntick_save_progress_counter);
1286 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1287 if (rcu_gp_in_progress(rsp))
1288 rsp->signaled = RCU_FORCE_QS;
1293 /* Check dyntick-idle state, send IPI to laggarts. */
1294 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1295 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1297 /* Leave state in case more forcing is required. */
1299 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1302 rsp->fqs_active = 0;
1303 if (rsp->fqs_need_gp) {
1304 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1305 rsp->fqs_need_gp = 0;
1306 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1309 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1311 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1314 #else /* #ifdef CONFIG_SMP */
1316 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1321 #endif /* #else #ifdef CONFIG_SMP */
1324 * This does the RCU processing work from softirq context for the
1325 * specified rcu_state and rcu_data structures. This may be called
1326 * only from the CPU to whom the rdp belongs.
1329 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1331 unsigned long flags;
1333 WARN_ON_ONCE(rdp->beenonline == 0);
1336 * If an RCU GP has gone long enough, go check for dyntick
1337 * idle CPUs and, if needed, send resched IPIs.
1339 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1340 force_quiescent_state(rsp, 1);
1343 * Advance callbacks in response to end of earlier grace
1344 * period that some other CPU ended.
1346 rcu_process_gp_end(rsp, rdp);
1348 /* Update RCU state based on any recent quiescent states. */
1349 rcu_check_quiescent_state(rsp, rdp);
1351 /* Does this CPU require a not-yet-started grace period? */
1352 if (cpu_needs_another_gp(rsp, rdp)) {
1353 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1354 rcu_start_gp(rsp, flags); /* releases above lock */
1357 /* If there are callbacks ready, invoke them. */
1358 rcu_do_batch(rsp, rdp);
1362 * Do softirq processing for the current CPU.
1364 static void rcu_process_callbacks(struct softirq_action *unused)
1366 __rcu_process_callbacks(&rcu_sched_state,
1367 &__get_cpu_var(rcu_sched_data));
1368 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1369 rcu_preempt_process_callbacks();
1371 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1372 rcu_needs_cpu_flush();
1376 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1377 struct rcu_state *rsp)
1379 unsigned long flags;
1380 struct rcu_data *rdp;
1382 debug_rcu_head_queue(head);
1386 smp_mb(); /* Ensure RCU update seen before callback registry. */
1389 * Opportunistically note grace-period endings and beginnings.
1390 * Note that we might see a beginning right after we see an
1391 * end, but never vice versa, since this CPU has to pass through
1392 * a quiescent state betweentimes.
1394 local_irq_save(flags);
1395 rdp = this_cpu_ptr(rsp->rda);
1397 /* Add the callback to our list. */
1398 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1399 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1402 * Force the grace period if too many callbacks or too long waiting.
1403 * Enforce hysteresis, and don't invoke force_quiescent_state()
1404 * if some other CPU has recently done so. Also, don't bother
1405 * invoking force_quiescent_state() if the newly enqueued callback
1406 * is the only one waiting for a grace period to complete.
1408 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1410 /* Are we ignoring a completed grace period? */
1411 rcu_process_gp_end(rsp, rdp);
1412 check_for_new_grace_period(rsp, rdp);
1414 /* Start a new grace period if one not already started. */
1415 if (!rcu_gp_in_progress(rsp)) {
1416 unsigned long nestflag;
1417 struct rcu_node *rnp_root = rcu_get_root(rsp);
1419 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1420 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1422 /* Give the grace period a kick. */
1423 rdp->blimit = LONG_MAX;
1424 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1425 *rdp->nxttail[RCU_DONE_TAIL] != head)
1426 force_quiescent_state(rsp, 0);
1427 rdp->n_force_qs_snap = rsp->n_force_qs;
1428 rdp->qlen_last_fqs_check = rdp->qlen;
1430 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1431 force_quiescent_state(rsp, 1);
1432 local_irq_restore(flags);
1436 * Queue an RCU-sched callback for invocation after a grace period.
1438 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1440 __call_rcu(head, func, &rcu_sched_state);
1442 EXPORT_SYMBOL_GPL(call_rcu_sched);
1445 * Queue an RCU for invocation after a quicker grace period.
1447 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1449 __call_rcu(head, func, &rcu_bh_state);
1451 EXPORT_SYMBOL_GPL(call_rcu_bh);
1454 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1456 * Control will return to the caller some time after a full rcu-sched
1457 * grace period has elapsed, in other words after all currently executing
1458 * rcu-sched read-side critical sections have completed. These read-side
1459 * critical sections are delimited by rcu_read_lock_sched() and
1460 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1461 * local_irq_disable(), and so on may be used in place of
1462 * rcu_read_lock_sched().
1464 * This means that all preempt_disable code sequences, including NMI and
1465 * hardware-interrupt handlers, in progress on entry will have completed
1466 * before this primitive returns. However, this does not guarantee that
1467 * softirq handlers will have completed, since in some kernels, these
1468 * handlers can run in process context, and can block.
1470 * This primitive provides the guarantees made by the (now removed)
1471 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1472 * guarantees that rcu_read_lock() sections will have completed.
1473 * In "classic RCU", these two guarantees happen to be one and
1474 * the same, but can differ in realtime RCU implementations.
1476 void synchronize_sched(void)
1478 struct rcu_synchronize rcu;
1480 if (rcu_blocking_is_gp())
1483 init_rcu_head_on_stack(&rcu.head);
1484 init_completion(&rcu.completion);
1485 /* Will wake me after RCU finished. */
1486 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1488 wait_for_completion(&rcu.completion);
1489 destroy_rcu_head_on_stack(&rcu.head);
1491 EXPORT_SYMBOL_GPL(synchronize_sched);
1494 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1496 * Control will return to the caller some time after a full rcu_bh grace
1497 * period has elapsed, in other words after all currently executing rcu_bh
1498 * read-side critical sections have completed. RCU read-side critical
1499 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1500 * and may be nested.
1502 void synchronize_rcu_bh(void)
1504 struct rcu_synchronize rcu;
1506 if (rcu_blocking_is_gp())
1509 init_rcu_head_on_stack(&rcu.head);
1510 init_completion(&rcu.completion);
1511 /* Will wake me after RCU finished. */
1512 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1514 wait_for_completion(&rcu.completion);
1515 destroy_rcu_head_on_stack(&rcu.head);
1517 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1520 * Check to see if there is any immediate RCU-related work to be done
1521 * by the current CPU, for the specified type of RCU, returning 1 if so.
1522 * The checks are in order of increasing expense: checks that can be
1523 * carried out against CPU-local state are performed first. However,
1524 * we must check for CPU stalls first, else we might not get a chance.
1526 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1528 struct rcu_node *rnp = rdp->mynode;
1530 rdp->n_rcu_pending++;
1532 /* Check for CPU stalls, if enabled. */
1533 check_cpu_stall(rsp, rdp);
1535 /* Is the RCU core waiting for a quiescent state from this CPU? */
1536 if (rdp->qs_pending && !rdp->passed_quiesc) {
1539 * If force_quiescent_state() coming soon and this CPU
1540 * needs a quiescent state, and this is either RCU-sched
1541 * or RCU-bh, force a local reschedule.
1543 rdp->n_rp_qs_pending++;
1544 if (!rdp->preemptable &&
1545 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1548 } else if (rdp->qs_pending && rdp->passed_quiesc) {
1549 rdp->n_rp_report_qs++;
1553 /* Does this CPU have callbacks ready to invoke? */
1554 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1555 rdp->n_rp_cb_ready++;
1559 /* Has RCU gone idle with this CPU needing another grace period? */
1560 if (cpu_needs_another_gp(rsp, rdp)) {
1561 rdp->n_rp_cpu_needs_gp++;
1565 /* Has another RCU grace period completed? */
1566 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1567 rdp->n_rp_gp_completed++;
1571 /* Has a new RCU grace period started? */
1572 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1573 rdp->n_rp_gp_started++;
1577 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1578 if (rcu_gp_in_progress(rsp) &&
1579 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1580 rdp->n_rp_need_fqs++;
1585 rdp->n_rp_need_nothing++;
1590 * Check to see if there is any immediate RCU-related work to be done
1591 * by the current CPU, returning 1 if so. This function is part of the
1592 * RCU implementation; it is -not- an exported member of the RCU API.
1594 static int rcu_pending(int cpu)
1596 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1597 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1598 rcu_preempt_pending(cpu);
1602 * Check to see if any future RCU-related work will need to be done
1603 * by the current CPU, even if none need be done immediately, returning
1606 static int rcu_needs_cpu_quick_check(int cpu)
1608 /* RCU callbacks either ready or pending? */
1609 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1610 per_cpu(rcu_bh_data, cpu).nxtlist ||
1611 rcu_preempt_needs_cpu(cpu);
1614 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1615 static atomic_t rcu_barrier_cpu_count;
1616 static DEFINE_MUTEX(rcu_barrier_mutex);
1617 static struct completion rcu_barrier_completion;
1619 static void rcu_barrier_callback(struct rcu_head *notused)
1621 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1622 complete(&rcu_barrier_completion);
1626 * Called with preemption disabled, and from cross-cpu IRQ context.
1628 static void rcu_barrier_func(void *type)
1630 int cpu = smp_processor_id();
1631 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1632 void (*call_rcu_func)(struct rcu_head *head,
1633 void (*func)(struct rcu_head *head));
1635 atomic_inc(&rcu_barrier_cpu_count);
1636 call_rcu_func = type;
1637 call_rcu_func(head, rcu_barrier_callback);
1641 * Orchestrate the specified type of RCU barrier, waiting for all
1642 * RCU callbacks of the specified type to complete.
1644 static void _rcu_barrier(struct rcu_state *rsp,
1645 void (*call_rcu_func)(struct rcu_head *head,
1646 void (*func)(struct rcu_head *head)))
1648 BUG_ON(in_interrupt());
1649 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1650 mutex_lock(&rcu_barrier_mutex);
1651 init_completion(&rcu_barrier_completion);
1653 * Initialize rcu_barrier_cpu_count to 1, then invoke
1654 * rcu_barrier_func() on each CPU, so that each CPU also has
1655 * incremented rcu_barrier_cpu_count. Only then is it safe to
1656 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1657 * might complete its grace period before all of the other CPUs
1658 * did their increment, causing this function to return too
1659 * early. Note that on_each_cpu() disables irqs, which prevents
1660 * any CPUs from coming online or going offline until each online
1661 * CPU has queued its RCU-barrier callback.
1663 atomic_set(&rcu_barrier_cpu_count, 1);
1664 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1665 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1666 complete(&rcu_barrier_completion);
1667 wait_for_completion(&rcu_barrier_completion);
1668 mutex_unlock(&rcu_barrier_mutex);
1672 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1674 void rcu_barrier_bh(void)
1676 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1678 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1681 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1683 void rcu_barrier_sched(void)
1685 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1687 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1690 * Do boot-time initialization of a CPU's per-CPU RCU data.
1693 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1695 unsigned long flags;
1697 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1698 struct rcu_node *rnp = rcu_get_root(rsp);
1700 /* Set up local state, ensuring consistent view of global state. */
1701 raw_spin_lock_irqsave(&rnp->lock, flags);
1702 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1703 rdp->nxtlist = NULL;
1704 for (i = 0; i < RCU_NEXT_SIZE; i++)
1705 rdp->nxttail[i] = &rdp->nxtlist;
1708 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1709 #endif /* #ifdef CONFIG_NO_HZ */
1711 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1715 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1716 * offline event can be happening at a given time. Note also that we
1717 * can accept some slop in the rsp->completed access due to the fact
1718 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1720 static void __cpuinit
1721 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1723 unsigned long flags;
1725 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1726 struct rcu_node *rnp = rcu_get_root(rsp);
1728 /* Set up local state, ensuring consistent view of global state. */
1729 raw_spin_lock_irqsave(&rnp->lock, flags);
1730 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1731 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1732 rdp->beenonline = 1; /* We have now been online. */
1733 rdp->preemptable = preemptable;
1734 rdp->qlen_last_fqs_check = 0;
1735 rdp->n_force_qs_snap = rsp->n_force_qs;
1736 rdp->blimit = blimit;
1737 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1740 * A new grace period might start here. If so, we won't be part
1741 * of it, but that is OK, as we are currently in a quiescent state.
1744 /* Exclude any attempts to start a new GP on large systems. */
1745 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1747 /* Add CPU to rcu_node bitmasks. */
1749 mask = rdp->grpmask;
1751 /* Exclude any attempts to start a new GP on small systems. */
1752 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1753 rnp->qsmaskinit |= mask;
1754 mask = rnp->grpmask;
1755 if (rnp == rdp->mynode) {
1756 rdp->gpnum = rnp->completed; /* if GP in progress... */
1757 rdp->completed = rnp->completed;
1758 rdp->passed_quiesc_completed = rnp->completed - 1;
1760 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1762 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1764 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1767 static void __cpuinit rcu_online_cpu(int cpu)
1769 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1770 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1771 rcu_preempt_init_percpu_data(cpu);
1775 * Handle CPU online/offline notification events.
1777 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1778 unsigned long action, void *hcpu)
1780 long cpu = (long)hcpu;
1783 case CPU_UP_PREPARE:
1784 case CPU_UP_PREPARE_FROZEN:
1785 rcu_online_cpu(cpu);
1788 case CPU_DYING_FROZEN:
1790 * The whole machine is "stopped" except this CPU, so we can
1791 * touch any data without introducing corruption. We send the
1792 * dying CPU's callbacks to an arbitrarily chosen online CPU.
1794 rcu_send_cbs_to_online(&rcu_bh_state);
1795 rcu_send_cbs_to_online(&rcu_sched_state);
1796 rcu_preempt_send_cbs_to_online();
1799 case CPU_DEAD_FROZEN:
1800 case CPU_UP_CANCELED:
1801 case CPU_UP_CANCELED_FROZEN:
1802 rcu_offline_cpu(cpu);
1811 * This function is invoked towards the end of the scheduler's initialization
1812 * process. Before this is called, the idle task might contain
1813 * RCU read-side critical sections (during which time, this idle
1814 * task is booting the system). After this function is called, the
1815 * idle tasks are prohibited from containing RCU read-side critical
1816 * sections. This function also enables RCU lockdep checking.
1818 void rcu_scheduler_starting(void)
1820 WARN_ON(num_online_cpus() != 1);
1821 WARN_ON(nr_context_switches() > 0);
1822 rcu_scheduler_active = 1;
1826 * Compute the per-level fanout, either using the exact fanout specified
1827 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1829 #ifdef CONFIG_RCU_FANOUT_EXACT
1830 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1834 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
1835 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1836 rsp->levelspread[0] = RCU_FANOUT_LEAF;
1838 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1839 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1846 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1847 ccur = rsp->levelcnt[i];
1848 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1852 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1855 * Helper function for rcu_init() that initializes one rcu_state structure.
1857 static void __init rcu_init_one(struct rcu_state *rsp,
1858 struct rcu_data __percpu *rda)
1860 static char *buf[] = { "rcu_node_level_0",
1863 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1867 struct rcu_node *rnp;
1869 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1871 /* Initialize the level-tracking arrays. */
1873 for (i = 1; i < NUM_RCU_LVLS; i++)
1874 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1875 rcu_init_levelspread(rsp);
1877 /* Initialize the elements themselves, starting from the leaves. */
1879 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1880 cpustride *= rsp->levelspread[i];
1881 rnp = rsp->level[i];
1882 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1883 raw_spin_lock_init(&rnp->lock);
1884 lockdep_set_class_and_name(&rnp->lock,
1885 &rcu_node_class[i], buf[i]);
1888 rnp->qsmaskinit = 0;
1889 rnp->grplo = j * cpustride;
1890 rnp->grphi = (j + 1) * cpustride - 1;
1891 if (rnp->grphi >= NR_CPUS)
1892 rnp->grphi = NR_CPUS - 1;
1898 rnp->grpnum = j % rsp->levelspread[i - 1];
1899 rnp->grpmask = 1UL << rnp->grpnum;
1900 rnp->parent = rsp->level[i - 1] +
1901 j / rsp->levelspread[i - 1];
1904 INIT_LIST_HEAD(&rnp->blkd_tasks);
1909 rnp = rsp->level[NUM_RCU_LVLS - 1];
1910 for_each_possible_cpu(i) {
1911 while (i > rnp->grphi)
1913 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
1914 rcu_boot_init_percpu_data(i, rsp);
1918 void __init rcu_init(void)
1922 rcu_bootup_announce();
1923 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
1924 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
1925 __rcu_init_preempt();
1926 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1929 * We don't need protection against CPU-hotplug here because
1930 * this is called early in boot, before either interrupts
1931 * or the scheduler are operational.
1933 cpu_notifier(rcu_cpu_notify, 0);
1934 for_each_online_cpu(cpu)
1935 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1936 check_cpu_stall_init();
1939 #include "rcutree_plugin.h"