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 <linux/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>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
56 /* Data structures. */
58 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
60 #define RCU_STATE_INITIALIZER(structname) { \
61 .level = { &structname.node[0] }, \
63 NUM_RCU_LVL_0, /* root of hierarchy. */ \
67 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
69 .signaled = RCU_GP_IDLE, \
72 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
73 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
75 .n_force_qs_ngp = 0, \
76 .name = #structname, \
79 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
80 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
82 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
83 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
85 static struct rcu_state *rcu_state;
87 int rcu_scheduler_active __read_mostly;
88 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
91 * Control variables for per-CPU and per-rcu_node kthreads. These
92 * handle all flavors of RCU.
94 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
95 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
96 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
97 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
98 DEFINE_PER_CPU(char, rcu_cpu_has_work);
99 static char rcu_kthreads_spawnable;
101 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
102 static void invoke_rcu_cpu_kthread(void);
103 static void __invoke_rcu_cpu_kthread(void);
105 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
108 * Track the rcutorture test sequence number and the update version
109 * number within a given test. The rcutorture_testseq is incremented
110 * on every rcutorture module load and unload, so has an odd value
111 * when a test is running. The rcutorture_vernum is set to zero
112 * when rcutorture starts and is incremented on each rcutorture update.
113 * These variables enable correlating rcutorture output with the
114 * RCU tracing information.
116 unsigned long rcutorture_testseq;
117 unsigned long rcutorture_vernum;
120 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
121 * permit this function to be invoked without holding the root rcu_node
122 * structure's ->lock, but of course results can be subject to change.
124 static int rcu_gp_in_progress(struct rcu_state *rsp)
126 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
130 * Note a quiescent state. Because we do not need to know
131 * how many quiescent states passed, just if there was at least
132 * one since the start of the grace period, this just sets a flag.
134 void rcu_sched_qs(int cpu)
136 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
138 rdp->passed_quiesc_completed = rdp->gpnum - 1;
140 rdp->passed_quiesc = 1;
143 void rcu_bh_qs(int cpu)
145 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
147 rdp->passed_quiesc_completed = rdp->gpnum - 1;
149 rdp->passed_quiesc = 1;
153 * Note a context switch. This is a quiescent state for RCU-sched,
154 * and requires special handling for preemptible RCU.
156 void rcu_note_context_switch(int cpu)
159 rcu_preempt_note_context_switch(cpu);
161 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
164 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
165 .dynticks_nesting = 1,
166 .dynticks = ATOMIC_INIT(1),
168 #endif /* #ifdef CONFIG_NO_HZ */
170 static int blimit = 10; /* Maximum callbacks per softirq. */
171 static int qhimark = 10000; /* If this many pending, ignore blimit. */
172 static int qlowmark = 100; /* Once only this many pending, use blimit. */
174 module_param(blimit, int, 0);
175 module_param(qhimark, int, 0);
176 module_param(qlowmark, int, 0);
178 int rcu_cpu_stall_suppress __read_mostly;
179 module_param(rcu_cpu_stall_suppress, int, 0644);
181 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
182 static int rcu_pending(int cpu);
185 * Return the number of RCU-sched batches processed thus far for debug & stats.
187 long rcu_batches_completed_sched(void)
189 return rcu_sched_state.completed;
191 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
194 * Return the number of RCU BH batches processed thus far for debug & stats.
196 long rcu_batches_completed_bh(void)
198 return rcu_bh_state.completed;
200 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
203 * Force a quiescent state for RCU BH.
205 void rcu_bh_force_quiescent_state(void)
207 force_quiescent_state(&rcu_bh_state, 0);
209 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
212 * Record the number of times rcutorture tests have been initiated and
213 * terminated. This information allows the debugfs tracing stats to be
214 * correlated to the rcutorture messages, even when the rcutorture module
215 * is being repeatedly loaded and unloaded. In other words, we cannot
216 * store this state in rcutorture itself.
218 void rcutorture_record_test_transition(void)
220 rcutorture_testseq++;
221 rcutorture_vernum = 0;
223 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
226 * Record the number of writer passes through the current rcutorture test.
227 * This is also used to correlate debugfs tracing stats with the rcutorture
230 void rcutorture_record_progress(unsigned long vernum)
234 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
237 * Force a quiescent state for RCU-sched.
239 void rcu_sched_force_quiescent_state(void)
241 force_quiescent_state(&rcu_sched_state, 0);
243 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
246 * Does the CPU have callbacks ready to be invoked?
249 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
251 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
255 * Does the current CPU require a yet-as-unscheduled grace period?
258 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
260 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
264 * Return the root node of the specified rcu_state structure.
266 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
268 return &rsp->node[0];
274 * If the specified CPU is offline, tell the caller that it is in
275 * a quiescent state. Otherwise, whack it with a reschedule IPI.
276 * Grace periods can end up waiting on an offline CPU when that
277 * CPU is in the process of coming online -- it will be added to the
278 * rcu_node bitmasks before it actually makes it online. The same thing
279 * can happen while a CPU is in the process of coming online. Because this
280 * race is quite rare, we check for it after detecting that the grace
281 * period has been delayed rather than checking each and every CPU
282 * each and every time we start a new grace period.
284 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
287 * If the CPU is offline, it is in a quiescent state. We can
288 * trust its state not to change because interrupts are disabled.
290 if (cpu_is_offline(rdp->cpu)) {
295 /* If preemptible RCU, no point in sending reschedule IPI. */
296 if (rdp->preemptible)
299 /* The CPU is online, so send it a reschedule IPI. */
300 if (rdp->cpu != smp_processor_id())
301 smp_send_reschedule(rdp->cpu);
308 #endif /* #ifdef CONFIG_SMP */
313 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
315 * Enter nohz mode, in other words, -leave- the mode in which RCU
316 * read-side critical sections can occur. (Though RCU read-side
317 * critical sections can occur in irq handlers in nohz mode, a possibility
318 * handled by rcu_irq_enter() and rcu_irq_exit()).
320 void rcu_enter_nohz(void)
323 struct rcu_dynticks *rdtp;
325 local_irq_save(flags);
326 rdtp = &__get_cpu_var(rcu_dynticks);
327 if (--rdtp->dynticks_nesting) {
328 local_irq_restore(flags);
331 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
332 smp_mb__before_atomic_inc(); /* See above. */
333 atomic_inc(&rdtp->dynticks);
334 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
335 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
336 local_irq_restore(flags);
338 /* If the interrupt queued a callback, get out of dyntick mode. */
340 (__get_cpu_var(rcu_sched_data).nxtlist ||
341 __get_cpu_var(rcu_bh_data).nxtlist ||
342 rcu_preempt_needs_cpu(smp_processor_id())))
347 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
349 * Exit nohz mode, in other words, -enter- the mode in which RCU
350 * read-side critical sections normally occur.
352 void rcu_exit_nohz(void)
355 struct rcu_dynticks *rdtp;
357 local_irq_save(flags);
358 rdtp = &__get_cpu_var(rcu_dynticks);
359 if (rdtp->dynticks_nesting++) {
360 local_irq_restore(flags);
363 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
364 atomic_inc(&rdtp->dynticks);
365 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
366 smp_mb__after_atomic_inc(); /* See above. */
367 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
368 local_irq_restore(flags);
372 * rcu_nmi_enter - inform RCU of entry to NMI context
374 * If the CPU was idle with dynamic ticks active, and there is no
375 * irq handler running, this updates rdtp->dynticks_nmi to let the
376 * RCU grace-period handling know that the CPU is active.
378 void rcu_nmi_enter(void)
380 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
382 if (rdtp->dynticks_nmi_nesting == 0 &&
383 (atomic_read(&rdtp->dynticks) & 0x1))
385 rdtp->dynticks_nmi_nesting++;
386 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
387 atomic_inc(&rdtp->dynticks);
388 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
389 smp_mb__after_atomic_inc(); /* See above. */
390 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
394 * rcu_nmi_exit - inform RCU of exit from NMI context
396 * If the CPU was idle with dynamic ticks active, and there is no
397 * irq handler running, this updates rdtp->dynticks_nmi to let the
398 * RCU grace-period handling know that the CPU is no longer active.
400 void rcu_nmi_exit(void)
402 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
404 if (rdtp->dynticks_nmi_nesting == 0 ||
405 --rdtp->dynticks_nmi_nesting != 0)
407 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
408 smp_mb__before_atomic_inc(); /* See above. */
409 atomic_inc(&rdtp->dynticks);
410 smp_mb__after_atomic_inc(); /* Force delay to next write. */
411 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
415 * rcu_irq_enter - inform RCU of entry to hard irq context
417 * If the CPU was idle with dynamic ticks active, this updates the
418 * rdtp->dynticks to let the RCU handling know that the CPU is active.
420 void rcu_irq_enter(void)
426 * rcu_irq_exit - inform RCU of exit from hard irq context
428 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
429 * to put let the RCU handling be aware that the CPU is going back to idle
432 void rcu_irq_exit(void)
440 * Snapshot the specified CPU's dynticks counter so that we can later
441 * credit them with an implicit quiescent state. Return 1 if this CPU
442 * is in dynticks idle mode, which is an extended quiescent state.
444 static int dyntick_save_progress_counter(struct rcu_data *rdp)
446 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
451 * Return true if the specified CPU has passed through a quiescent
452 * state by virtue of being in or having passed through an dynticks
453 * idle state since the last call to dyntick_save_progress_counter()
456 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
461 curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
462 snap = (unsigned long)rdp->dynticks_snap;
465 * If the CPU passed through or entered a dynticks idle phase with
466 * no active irq/NMI handlers, then we can safely pretend that the CPU
467 * already acknowledged the request to pass through a quiescent
468 * state. Either way, that CPU cannot possibly be in an RCU
469 * read-side critical section that started before the beginning
470 * of the current RCU grace period.
472 if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
477 /* Go check for the CPU being offline. */
478 return rcu_implicit_offline_qs(rdp);
481 #endif /* #ifdef CONFIG_SMP */
483 #else /* #ifdef CONFIG_NO_HZ */
487 static int dyntick_save_progress_counter(struct rcu_data *rdp)
492 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
494 return rcu_implicit_offline_qs(rdp);
497 #endif /* #ifdef CONFIG_SMP */
499 #endif /* #else #ifdef CONFIG_NO_HZ */
501 int rcu_cpu_stall_suppress __read_mostly;
503 static void record_gp_stall_check_time(struct rcu_state *rsp)
505 rsp->gp_start = jiffies;
506 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
509 static void print_other_cpu_stall(struct rcu_state *rsp)
514 struct rcu_node *rnp = rcu_get_root(rsp);
516 /* Only let one CPU complain about others per time interval. */
518 raw_spin_lock_irqsave(&rnp->lock, flags);
519 delta = jiffies - rsp->jiffies_stall;
520 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
521 raw_spin_unlock_irqrestore(&rnp->lock, flags);
524 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
527 * Now rat on any tasks that got kicked up to the root rcu_node
528 * due to CPU offlining.
530 rcu_print_task_stall(rnp);
531 raw_spin_unlock_irqrestore(&rnp->lock, flags);
534 * OK, time to rat on our buddy...
535 * See Documentation/RCU/stallwarn.txt for info on how to debug
536 * RCU CPU stall warnings.
538 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
540 rcu_for_each_leaf_node(rsp, rnp) {
541 raw_spin_lock_irqsave(&rnp->lock, flags);
542 rcu_print_task_stall(rnp);
543 raw_spin_unlock_irqrestore(&rnp->lock, flags);
544 if (rnp->qsmask == 0)
546 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
547 if (rnp->qsmask & (1UL << cpu))
548 printk(" %d", rnp->grplo + cpu);
550 printk("} (detected by %d, t=%ld jiffies)\n",
551 smp_processor_id(), (long)(jiffies - rsp->gp_start));
552 trigger_all_cpu_backtrace();
554 /* If so configured, complain about tasks blocking the grace period. */
556 rcu_print_detail_task_stall(rsp);
558 force_quiescent_state(rsp, 0); /* Kick them all. */
561 static void print_cpu_stall(struct rcu_state *rsp)
564 struct rcu_node *rnp = rcu_get_root(rsp);
567 * OK, time to rat on ourselves...
568 * See Documentation/RCU/stallwarn.txt for info on how to debug
569 * RCU CPU stall warnings.
571 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
572 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
573 trigger_all_cpu_backtrace();
575 raw_spin_lock_irqsave(&rnp->lock, flags);
576 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
578 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
579 raw_spin_unlock_irqrestore(&rnp->lock, flags);
581 set_need_resched(); /* kick ourselves to get things going. */
584 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
588 struct rcu_node *rnp;
590 if (rcu_cpu_stall_suppress)
592 j = ACCESS_ONCE(jiffies);
593 js = ACCESS_ONCE(rsp->jiffies_stall);
595 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
597 /* We haven't checked in, so go dump stack. */
598 print_cpu_stall(rsp);
600 } else if (rcu_gp_in_progress(rsp) &&
601 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
603 /* They had a few time units to dump stack, so complain. */
604 print_other_cpu_stall(rsp);
608 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
610 rcu_cpu_stall_suppress = 1;
615 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
617 * Set the stall-warning timeout way off into the future, thus preventing
618 * any RCU CPU stall-warning messages from appearing in the current set of
621 * The caller must disable hard irqs.
623 void rcu_cpu_stall_reset(void)
625 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
626 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
627 rcu_preempt_stall_reset();
630 static struct notifier_block rcu_panic_block = {
631 .notifier_call = rcu_panic,
634 static void __init check_cpu_stall_init(void)
636 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
640 * Update CPU-local rcu_data state to record the newly noticed grace period.
641 * This is used both when we started the grace period and when we notice
642 * that someone else started the grace period. The caller must hold the
643 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
644 * and must have irqs disabled.
646 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
648 if (rdp->gpnum != rnp->gpnum) {
650 * If the current grace period is waiting for this CPU,
651 * set up to detect a quiescent state, otherwise don't
652 * go looking for one.
654 rdp->gpnum = rnp->gpnum;
655 if (rnp->qsmask & rdp->grpmask) {
657 rdp->passed_quiesc = 0;
663 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
666 struct rcu_node *rnp;
668 local_irq_save(flags);
670 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
671 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
672 local_irq_restore(flags);
675 __note_new_gpnum(rsp, rnp, rdp);
676 raw_spin_unlock_irqrestore(&rnp->lock, flags);
680 * Did someone else start a new RCU grace period start since we last
681 * checked? Update local state appropriately if so. Must be called
682 * on the CPU corresponding to rdp.
685 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
690 local_irq_save(flags);
691 if (rdp->gpnum != rsp->gpnum) {
692 note_new_gpnum(rsp, rdp);
695 local_irq_restore(flags);
700 * Advance this CPU's callbacks, but only if the current grace period
701 * has ended. This may be called only from the CPU to whom the rdp
702 * belongs. In addition, the corresponding leaf rcu_node structure's
703 * ->lock must be held by the caller, with irqs disabled.
706 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
708 /* Did another grace period end? */
709 if (rdp->completed != rnp->completed) {
711 /* Advance callbacks. No harm if list empty. */
712 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
713 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
714 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
716 /* Remember that we saw this grace-period completion. */
717 rdp->completed = rnp->completed;
720 * If we were in an extended quiescent state, we may have
721 * missed some grace periods that others CPUs handled on
722 * our behalf. Catch up with this state to avoid noting
723 * spurious new grace periods. If another grace period
724 * has started, then rnp->gpnum will have advanced, so
725 * we will detect this later on.
727 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
728 rdp->gpnum = rdp->completed;
731 * If RCU does not need a quiescent state from this CPU,
732 * then make sure that this CPU doesn't go looking for one.
734 if ((rnp->qsmask & rdp->grpmask) == 0)
740 * Advance this CPU's callbacks, but only if the current grace period
741 * has ended. This may be called only from the CPU to whom the rdp
745 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
748 struct rcu_node *rnp;
750 local_irq_save(flags);
752 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
753 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
754 local_irq_restore(flags);
757 __rcu_process_gp_end(rsp, rnp, rdp);
758 raw_spin_unlock_irqrestore(&rnp->lock, flags);
762 * Do per-CPU grace-period initialization for running CPU. The caller
763 * must hold the lock of the leaf rcu_node structure corresponding to
767 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
769 /* Prior grace period ended, so advance callbacks for current CPU. */
770 __rcu_process_gp_end(rsp, rnp, rdp);
773 * Because this CPU just now started the new grace period, we know
774 * that all of its callbacks will be covered by this upcoming grace
775 * period, even the ones that were registered arbitrarily recently.
776 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
778 * Other CPUs cannot be sure exactly when the grace period started.
779 * Therefore, their recently registered callbacks must pass through
780 * an additional RCU_NEXT_READY stage, so that they will be handled
781 * by the next RCU grace period.
783 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
784 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
786 /* Set state so that this CPU will detect the next quiescent state. */
787 __note_new_gpnum(rsp, rnp, rdp);
791 * Start a new RCU grace period if warranted, re-initializing the hierarchy
792 * in preparation for detecting the next grace period. The caller must hold
793 * the root node's ->lock, which is released before return. Hard irqs must
797 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
798 __releases(rcu_get_root(rsp)->lock)
800 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
801 struct rcu_node *rnp = rcu_get_root(rsp);
803 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
804 if (cpu_needs_another_gp(rsp, rdp))
805 rsp->fqs_need_gp = 1;
806 if (rnp->completed == rsp->completed) {
807 raw_spin_unlock_irqrestore(&rnp->lock, flags);
810 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
813 * Propagate new ->completed value to rcu_node structures
814 * so that other CPUs don't have to wait until the start
815 * of the next grace period to process their callbacks.
817 rcu_for_each_node_breadth_first(rsp, rnp) {
818 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
819 rnp->completed = rsp->completed;
820 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
822 local_irq_restore(flags);
826 /* Advance to a new grace period and initialize state. */
828 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
829 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
830 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
831 record_gp_stall_check_time(rsp);
833 /* Special-case the common single-level case. */
834 if (NUM_RCU_NODES == 1) {
835 rcu_preempt_check_blocked_tasks(rnp);
836 rnp->qsmask = rnp->qsmaskinit;
837 rnp->gpnum = rsp->gpnum;
838 rnp->completed = rsp->completed;
839 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
840 rcu_start_gp_per_cpu(rsp, rnp, rdp);
841 rcu_preempt_boost_start_gp(rnp);
842 raw_spin_unlock_irqrestore(&rnp->lock, flags);
846 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
849 /* Exclude any concurrent CPU-hotplug operations. */
850 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
853 * Set the quiescent-state-needed bits in all the rcu_node
854 * structures for all currently online CPUs in breadth-first
855 * order, starting from the root rcu_node structure. This
856 * operation relies on the layout of the hierarchy within the
857 * rsp->node[] array. Note that other CPUs will access only
858 * the leaves of the hierarchy, which still indicate that no
859 * grace period is in progress, at least until the corresponding
860 * leaf node has been initialized. In addition, we have excluded
861 * CPU-hotplug operations.
863 * Note that the grace period cannot complete until we finish
864 * the initialization process, as there will be at least one
865 * qsmask bit set in the root node until that time, namely the
866 * one corresponding to this CPU, due to the fact that we have
869 rcu_for_each_node_breadth_first(rsp, rnp) {
870 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
871 rcu_preempt_check_blocked_tasks(rnp);
872 rnp->qsmask = rnp->qsmaskinit;
873 rnp->gpnum = rsp->gpnum;
874 rnp->completed = rsp->completed;
875 if (rnp == rdp->mynode)
876 rcu_start_gp_per_cpu(rsp, rnp, rdp);
877 rcu_preempt_boost_start_gp(rnp);
878 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
881 rnp = rcu_get_root(rsp);
882 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
883 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
884 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
885 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
889 * Report a full set of quiescent states to the specified rcu_state
890 * data structure. This involves cleaning up after the prior grace
891 * period and letting rcu_start_gp() start up the next grace period
892 * if one is needed. Note that the caller must hold rnp->lock, as
893 * required by rcu_start_gp(), which will release it.
895 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
896 __releases(rcu_get_root(rsp)->lock)
898 unsigned long gp_duration;
900 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
903 * Ensure that all grace-period and pre-grace-period activity
904 * is seen before the assignment to rsp->completed.
906 smp_mb(); /* See above block comment. */
907 gp_duration = jiffies - rsp->gp_start;
908 if (gp_duration > rsp->gp_max)
909 rsp->gp_max = gp_duration;
910 rsp->completed = rsp->gpnum;
911 rsp->signaled = RCU_GP_IDLE;
912 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
916 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
917 * Allows quiescent states for a group of CPUs to be reported at one go
918 * to the specified rcu_node structure, though all the CPUs in the group
919 * must be represented by the same rcu_node structure (which need not be
920 * a leaf rcu_node structure, though it often will be). That structure's
921 * lock must be held upon entry, and it is released before return.
924 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
925 struct rcu_node *rnp, unsigned long flags)
926 __releases(rnp->lock)
928 struct rcu_node *rnp_c;
930 /* Walk up the rcu_node hierarchy. */
932 if (!(rnp->qsmask & mask)) {
934 /* Our bit has already been cleared, so done. */
935 raw_spin_unlock_irqrestore(&rnp->lock, flags);
938 rnp->qsmask &= ~mask;
939 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
941 /* Other bits still set at this level, so done. */
942 raw_spin_unlock_irqrestore(&rnp->lock, flags);
946 if (rnp->parent == NULL) {
948 /* No more levels. Exit loop holding root lock. */
952 raw_spin_unlock_irqrestore(&rnp->lock, flags);
955 raw_spin_lock_irqsave(&rnp->lock, flags);
956 WARN_ON_ONCE(rnp_c->qsmask);
960 * Get here if we are the last CPU to pass through a quiescent
961 * state for this grace period. Invoke rcu_report_qs_rsp()
962 * to clean up and start the next grace period if one is needed.
964 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
968 * Record a quiescent state for the specified CPU to that CPU's rcu_data
969 * structure. This must be either called from the specified CPU, or
970 * called when the specified CPU is known to be offline (and when it is
971 * also known that no other CPU is concurrently trying to help the offline
972 * CPU). The lastcomp argument is used to make sure we are still in the
973 * grace period of interest. We don't want to end the current grace period
974 * based on quiescent states detected in an earlier grace period!
977 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
981 struct rcu_node *rnp;
984 raw_spin_lock_irqsave(&rnp->lock, flags);
985 if (lastcomp != rnp->completed) {
988 * Someone beat us to it for this grace period, so leave.
989 * The race with GP start is resolved by the fact that we
990 * hold the leaf rcu_node lock, so that the per-CPU bits
991 * cannot yet be initialized -- so we would simply find our
992 * CPU's bit already cleared in rcu_report_qs_rnp() if this
995 rdp->passed_quiesc = 0; /* try again later! */
996 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1000 if ((rnp->qsmask & mask) == 0) {
1001 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1003 rdp->qs_pending = 0;
1006 * This GP can't end until cpu checks in, so all of our
1007 * callbacks can be processed during the next GP.
1009 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1011 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1016 * Check to see if there is a new grace period of which this CPU
1017 * is not yet aware, and if so, set up local rcu_data state for it.
1018 * Otherwise, see if this CPU has just passed through its first
1019 * quiescent state for this grace period, and record that fact if so.
1022 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1024 /* If there is now a new grace period, record and return. */
1025 if (check_for_new_grace_period(rsp, rdp))
1029 * Does this CPU still need to do its part for current grace period?
1030 * If no, return and let the other CPUs do their part as well.
1032 if (!rdp->qs_pending)
1036 * Was there a quiescent state since the beginning of the grace
1037 * period? If no, then exit and wait for the next call.
1039 if (!rdp->passed_quiesc)
1043 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1046 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
1049 #ifdef CONFIG_HOTPLUG_CPU
1052 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1053 * Synchronization is not required because this function executes
1054 * in stop_machine() context.
1056 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1059 /* current DYING CPU is cleared in the cpu_online_mask */
1060 int receive_cpu = cpumask_any(cpu_online_mask);
1061 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1062 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1064 if (rdp->nxtlist == NULL)
1065 return; /* irqs disabled, so comparison is stable. */
1067 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1068 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1069 receive_rdp->qlen += rdp->qlen;
1070 receive_rdp->n_cbs_adopted += rdp->qlen;
1071 rdp->n_cbs_orphaned += rdp->qlen;
1073 rdp->nxtlist = NULL;
1074 for (i = 0; i < RCU_NEXT_SIZE; i++)
1075 rdp->nxttail[i] = &rdp->nxtlist;
1080 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1081 * and move all callbacks from the outgoing CPU to the current one.
1082 * There can only be one CPU hotplug operation at a time, so no other
1083 * CPU can be attempting to update rcu_cpu_kthread_task.
1085 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1087 unsigned long flags;
1089 int need_report = 0;
1090 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1091 struct rcu_node *rnp;
1092 struct task_struct *t;
1094 /* Stop the CPU's kthread. */
1095 t = per_cpu(rcu_cpu_kthread_task, cpu);
1097 per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1101 /* Exclude any attempts to start a new grace period. */
1102 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1104 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1105 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1106 mask = rdp->grpmask; /* rnp->grplo is constant. */
1108 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1109 rnp->qsmaskinit &= ~mask;
1110 if (rnp->qsmaskinit != 0) {
1111 if (rnp != rdp->mynode)
1112 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1115 if (rnp == rdp->mynode)
1116 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1118 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1119 mask = rnp->grpmask;
1121 } while (rnp != NULL);
1124 * We still hold the leaf rcu_node structure lock here, and
1125 * irqs are still disabled. The reason for this subterfuge is
1126 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1127 * held leads to deadlock.
1129 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1131 if (need_report & RCU_OFL_TASKS_NORM_GP)
1132 rcu_report_unblock_qs_rnp(rnp, flags);
1134 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1135 if (need_report & RCU_OFL_TASKS_EXP_GP)
1136 rcu_report_exp_rnp(rsp, rnp);
1137 rcu_node_kthread_setaffinity(rnp, -1);
1141 * Remove the specified CPU from the RCU hierarchy and move any pending
1142 * callbacks that it might have to the current CPU. This code assumes
1143 * that at least one CPU in the system will remain running at all times.
1144 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1146 static void rcu_offline_cpu(int cpu)
1148 __rcu_offline_cpu(cpu, &rcu_sched_state);
1149 __rcu_offline_cpu(cpu, &rcu_bh_state);
1150 rcu_preempt_offline_cpu(cpu);
1153 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1155 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1159 static void rcu_offline_cpu(int cpu)
1163 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1166 * Invoke any RCU callbacks that have made it to the end of their grace
1167 * period. Thottle as specified by rdp->blimit.
1169 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1171 unsigned long flags;
1172 struct rcu_head *next, *list, **tail;
1175 /* If no callbacks are ready, just return.*/
1176 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1180 * Extract the list of ready callbacks, disabling to prevent
1181 * races with call_rcu() from interrupt handlers.
1183 local_irq_save(flags);
1184 list = rdp->nxtlist;
1185 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1186 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1187 tail = rdp->nxttail[RCU_DONE_TAIL];
1188 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1189 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1190 rdp->nxttail[count] = &rdp->nxtlist;
1191 local_irq_restore(flags);
1193 /* Invoke callbacks. */
1198 debug_rcu_head_unqueue(list);
1199 __rcu_reclaim(list);
1201 if (++count >= rdp->blimit)
1205 local_irq_save(flags);
1207 /* Update count, and requeue any remaining callbacks. */
1209 rdp->n_cbs_invoked += count;
1211 *tail = rdp->nxtlist;
1212 rdp->nxtlist = list;
1213 for (count = 0; count < RCU_NEXT_SIZE; count++)
1214 if (&rdp->nxtlist == rdp->nxttail[count])
1215 rdp->nxttail[count] = tail;
1220 /* Reinstate batch limit if we have worked down the excess. */
1221 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1222 rdp->blimit = blimit;
1224 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1225 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1226 rdp->qlen_last_fqs_check = 0;
1227 rdp->n_force_qs_snap = rsp->n_force_qs;
1228 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1229 rdp->qlen_last_fqs_check = rdp->qlen;
1231 local_irq_restore(flags);
1233 /* Re-raise the RCU softirq if there are callbacks remaining. */
1234 if (cpu_has_callbacks_ready_to_invoke(rdp))
1235 invoke_rcu_cpu_kthread();
1239 * Check to see if this CPU is in a non-context-switch quiescent state
1240 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1241 * Also schedule the RCU softirq handler.
1243 * This function must be called with hardirqs disabled. It is normally
1244 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1245 * false, there is no point in invoking rcu_check_callbacks().
1247 void rcu_check_callbacks(int cpu, int user)
1250 (idle_cpu(cpu) && rcu_scheduler_active &&
1251 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1254 * Get here if this CPU took its interrupt from user
1255 * mode or from the idle loop, and if this is not a
1256 * nested interrupt. In this case, the CPU is in
1257 * a quiescent state, so note it.
1259 * No memory barrier is required here because both
1260 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1261 * variables that other CPUs neither access nor modify,
1262 * at least not while the corresponding CPU is online.
1268 } else if (!in_softirq()) {
1271 * Get here if this CPU did not take its interrupt from
1272 * softirq, in other words, if it is not interrupting
1273 * a rcu_bh read-side critical section. This is an _bh
1274 * critical section, so note it.
1279 rcu_preempt_check_callbacks(cpu);
1280 if (rcu_pending(cpu))
1281 invoke_rcu_cpu_kthread();
1287 * Scan the leaf rcu_node structures, processing dyntick state for any that
1288 * have not yet encountered a quiescent state, using the function specified.
1289 * Also initiate boosting for any threads blocked on the root rcu_node.
1291 * The caller must have suppressed start of new grace periods.
1293 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1297 unsigned long flags;
1299 struct rcu_node *rnp;
1301 rcu_for_each_leaf_node(rsp, rnp) {
1303 raw_spin_lock_irqsave(&rnp->lock, flags);
1304 if (!rcu_gp_in_progress(rsp)) {
1305 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1308 if (rnp->qsmask == 0) {
1309 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1314 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1315 if ((rnp->qsmask & bit) != 0 &&
1316 f(per_cpu_ptr(rsp->rda, cpu)))
1321 /* rcu_report_qs_rnp() releases rnp->lock. */
1322 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1325 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1327 rnp = rcu_get_root(rsp);
1328 if (rnp->qsmask == 0) {
1329 raw_spin_lock_irqsave(&rnp->lock, flags);
1330 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1335 * Force quiescent states on reluctant CPUs, and also detect which
1336 * CPUs are in dyntick-idle mode.
1338 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1340 unsigned long flags;
1341 struct rcu_node *rnp = rcu_get_root(rsp);
1343 if (!rcu_gp_in_progress(rsp))
1344 return; /* No grace period in progress, nothing to force. */
1345 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1346 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1347 return; /* Someone else is already on the job. */
1349 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1350 goto unlock_fqs_ret; /* no emergency and done recently. */
1352 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1353 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1354 if(!rcu_gp_in_progress(rsp)) {
1355 rsp->n_force_qs_ngp++;
1356 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1357 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1359 rsp->fqs_active = 1;
1360 switch (rsp->signaled) {
1364 break; /* grace period idle or initializing, ignore. */
1366 case RCU_SAVE_DYNTICK:
1367 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1368 break; /* So gcc recognizes the dead code. */
1370 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1372 /* Record dyntick-idle state. */
1373 force_qs_rnp(rsp, dyntick_save_progress_counter);
1374 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1375 if (rcu_gp_in_progress(rsp))
1376 rsp->signaled = RCU_FORCE_QS;
1381 /* Check dyntick-idle state, send IPI to laggarts. */
1382 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1383 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1385 /* Leave state in case more forcing is required. */
1387 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1390 rsp->fqs_active = 0;
1391 if (rsp->fqs_need_gp) {
1392 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1393 rsp->fqs_need_gp = 0;
1394 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1397 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1399 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1402 #else /* #ifdef CONFIG_SMP */
1404 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1409 #endif /* #else #ifdef CONFIG_SMP */
1412 * This does the RCU processing work from softirq context for the
1413 * specified rcu_state and rcu_data structures. This may be called
1414 * only from the CPU to whom the rdp belongs.
1417 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1419 unsigned long flags;
1421 WARN_ON_ONCE(rdp->beenonline == 0);
1424 * If an RCU GP has gone long enough, go check for dyntick
1425 * idle CPUs and, if needed, send resched IPIs.
1427 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1428 force_quiescent_state(rsp, 1);
1431 * Advance callbacks in response to end of earlier grace
1432 * period that some other CPU ended.
1434 rcu_process_gp_end(rsp, rdp);
1436 /* Update RCU state based on any recent quiescent states. */
1437 rcu_check_quiescent_state(rsp, rdp);
1439 /* Does this CPU require a not-yet-started grace period? */
1440 if (cpu_needs_another_gp(rsp, rdp)) {
1441 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1442 rcu_start_gp(rsp, flags); /* releases above lock */
1445 /* If there are callbacks ready, invoke them. */
1446 if (cpu_has_callbacks_ready_to_invoke(rdp))
1447 __invoke_rcu_cpu_kthread();
1450 static void rcu_kthread_do_work(void)
1452 rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data));
1453 rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1454 rcu_preempt_do_callbacks();
1458 * Do softirq processing for the current CPU.
1460 static void rcu_process_callbacks(struct softirq_action *unused)
1462 __rcu_process_callbacks(&rcu_sched_state,
1463 &__get_cpu_var(rcu_sched_data));
1464 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1465 rcu_preempt_process_callbacks();
1467 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1468 rcu_needs_cpu_flush();
1472 * Wake up the current CPU's kthread. This replaces raise_softirq()
1473 * in earlier versions of RCU. Note that because we are running on
1474 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1475 * cannot disappear out from under us.
1477 static void __invoke_rcu_cpu_kthread(void)
1479 unsigned long flags;
1481 local_irq_save(flags);
1482 __this_cpu_write(rcu_cpu_has_work, 1);
1483 if (__this_cpu_read(rcu_cpu_kthread_task) == NULL) {
1484 local_irq_restore(flags);
1487 wake_up_process(__this_cpu_read(rcu_cpu_kthread_task));
1488 local_irq_restore(flags);
1491 static void invoke_rcu_cpu_kthread(void)
1493 raise_softirq(RCU_SOFTIRQ);
1497 * Wake up the specified per-rcu_node-structure kthread.
1498 * Because the per-rcu_node kthreads are immortal, we don't need
1499 * to do anything to keep them alive.
1501 static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1503 struct task_struct *t;
1505 t = rnp->node_kthread_task;
1511 * Set the specified CPU's kthread to run RT or not, as specified by
1512 * the to_rt argument. The CPU-hotplug locks are held, so the task
1513 * is not going away.
1515 static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
1518 struct sched_param sp;
1519 struct task_struct *t;
1521 t = per_cpu(rcu_cpu_kthread_task, cpu);
1525 policy = SCHED_FIFO;
1526 sp.sched_priority = RCU_KTHREAD_PRIO;
1528 policy = SCHED_NORMAL;
1529 sp.sched_priority = 0;
1531 sched_setscheduler_nocheck(t, policy, &sp);
1535 * Timer handler to initiate the waking up of per-CPU kthreads that
1536 * have yielded the CPU due to excess numbers of RCU callbacks.
1537 * We wake up the per-rcu_node kthread, which in turn will wake up
1538 * the booster kthread.
1540 static void rcu_cpu_kthread_timer(unsigned long arg)
1542 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1543 struct rcu_node *rnp = rdp->mynode;
1545 atomic_or(rdp->grpmask, &rnp->wakemask);
1546 invoke_rcu_node_kthread(rnp);
1550 * Drop to non-real-time priority and yield, but only after posting a
1551 * timer that will cause us to regain our real-time priority if we
1552 * remain preempted. Either way, we restore our real-time priority
1555 static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1557 struct sched_param sp;
1558 struct timer_list yield_timer;
1560 setup_timer_on_stack(&yield_timer, f, arg);
1561 mod_timer(&yield_timer, jiffies + 2);
1562 sp.sched_priority = 0;
1563 sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1564 set_user_nice(current, 19);
1566 sp.sched_priority = RCU_KTHREAD_PRIO;
1567 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1568 del_timer(&yield_timer);
1572 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1573 * This can happen while the corresponding CPU is either coming online
1574 * or going offline. We cannot wait until the CPU is fully online
1575 * before starting the kthread, because the various notifier functions
1576 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1577 * the corresponding CPU is online.
1579 * Return 1 if the kthread needs to stop, 0 otherwise.
1581 * Caller must disable bh. This function can momentarily enable it.
1583 static int rcu_cpu_kthread_should_stop(int cpu)
1585 while (cpu_is_offline(cpu) ||
1586 !cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)) ||
1587 smp_processor_id() != cpu) {
1588 if (kthread_should_stop())
1590 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1591 per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id();
1593 schedule_timeout_uninterruptible(1);
1594 if (!cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)))
1595 set_cpus_allowed_ptr(current, cpumask_of(cpu));
1598 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1603 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1604 * earlier RCU softirq.
1606 static int rcu_cpu_kthread(void *arg)
1608 int cpu = (int)(long)arg;
1609 unsigned long flags;
1611 unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu);
1613 char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1616 *statusp = RCU_KTHREAD_WAITING;
1617 rcu_wait(*workp != 0 || kthread_should_stop());
1619 if (rcu_cpu_kthread_should_stop(cpu)) {
1623 *statusp = RCU_KTHREAD_RUNNING;
1624 per_cpu(rcu_cpu_kthread_loops, cpu)++;
1625 local_irq_save(flags);
1628 local_irq_restore(flags);
1630 rcu_kthread_do_work();
1637 *statusp = RCU_KTHREAD_YIELDING;
1638 rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1642 *statusp = RCU_KTHREAD_STOPPED;
1647 * Spawn a per-CPU kthread, setting up affinity and priority.
1648 * Because the CPU hotplug lock is held, no other CPU will be attempting
1649 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1650 * attempting to access it during boot, but the locking in kthread_bind()
1651 * will enforce sufficient ordering.
1653 * Please note that we cannot simply refuse to wake up the per-CPU
1654 * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state,
1655 * which can result in softlockup complaints if the task ends up being
1656 * idle for more than a couple of minutes.
1658 * However, please note also that we cannot bind the per-CPU kthread to its
1659 * CPU until that CPU is fully online. We also cannot wait until the
1660 * CPU is fully online before we create its per-CPU kthread, as this would
1661 * deadlock the system when CPU notifiers tried waiting for grace
1662 * periods. So we bind the per-CPU kthread to its CPU only if the CPU
1663 * is online. If its CPU is not yet fully online, then the code in
1664 * rcu_cpu_kthread() will wait until it is fully online, and then do
1667 static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1669 struct sched_param sp;
1670 struct task_struct *t;
1672 if (!rcu_kthreads_spawnable ||
1673 per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1675 t = kthread_create(rcu_cpu_kthread, (void *)(long)cpu, "rcuc%d", cpu);
1678 if (cpu_online(cpu))
1679 kthread_bind(t, cpu);
1680 per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu;
1681 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1682 sp.sched_priority = RCU_KTHREAD_PRIO;
1683 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1684 per_cpu(rcu_cpu_kthread_task, cpu) = t;
1685 wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */
1690 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1691 * kthreads when needed. We ignore requests to wake up kthreads
1692 * for offline CPUs, which is OK because force_quiescent_state()
1693 * takes care of this case.
1695 static int rcu_node_kthread(void *arg)
1698 unsigned long flags;
1700 struct rcu_node *rnp = (struct rcu_node *)arg;
1701 struct sched_param sp;
1702 struct task_struct *t;
1705 rnp->node_kthread_status = RCU_KTHREAD_WAITING;
1706 rcu_wait(atomic_read(&rnp->wakemask) != 0);
1707 rnp->node_kthread_status = RCU_KTHREAD_RUNNING;
1708 raw_spin_lock_irqsave(&rnp->lock, flags);
1709 mask = atomic_xchg(&rnp->wakemask, 0);
1710 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1711 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1712 if ((mask & 0x1) == 0)
1715 t = per_cpu(rcu_cpu_kthread_task, cpu);
1716 if (!cpu_online(cpu) || t == NULL) {
1720 per_cpu(rcu_cpu_has_work, cpu) = 1;
1721 sp.sched_priority = RCU_KTHREAD_PRIO;
1722 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1727 rnp->node_kthread_status = RCU_KTHREAD_STOPPED;
1732 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1733 * served by the rcu_node in question. The CPU hotplug lock is still
1734 * held, so the value of rnp->qsmaskinit will be stable.
1736 * We don't include outgoingcpu in the affinity set, use -1 if there is
1737 * no outgoing CPU. If there are no CPUs left in the affinity set,
1738 * this function allows the kthread to execute on any CPU.
1740 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1744 unsigned long mask = rnp->qsmaskinit;
1746 if (rnp->node_kthread_task == NULL)
1748 if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1751 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1752 if ((mask & 0x1) && cpu != outgoingcpu)
1753 cpumask_set_cpu(cpu, cm);
1754 if (cpumask_weight(cm) == 0) {
1756 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1757 cpumask_clear_cpu(cpu, cm);
1758 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1760 set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1761 rcu_boost_kthread_setaffinity(rnp, cm);
1762 free_cpumask_var(cm);
1766 * Spawn a per-rcu_node kthread, setting priority and affinity.
1767 * Called during boot before online/offline can happen, or, if
1768 * during runtime, with the main CPU-hotplug locks held. So only
1769 * one of these can be executing at a time.
1771 static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1772 struct rcu_node *rnp)
1774 unsigned long flags;
1775 int rnp_index = rnp - &rsp->node[0];
1776 struct sched_param sp;
1777 struct task_struct *t;
1779 if (!rcu_kthreads_spawnable ||
1780 rnp->qsmaskinit == 0)
1782 if (rnp->node_kthread_task == NULL) {
1783 t = kthread_create(rcu_node_kthread, (void *)rnp,
1784 "rcun%d", rnp_index);
1787 raw_spin_lock_irqsave(&rnp->lock, flags);
1788 rnp->node_kthread_task = t;
1789 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1790 sp.sched_priority = 99;
1791 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1792 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1794 return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1798 * Spawn all kthreads -- called as soon as the scheduler is running.
1800 static int __init rcu_spawn_kthreads(void)
1803 struct rcu_node *rnp;
1805 rcu_kthreads_spawnable = 1;
1806 for_each_possible_cpu(cpu) {
1807 per_cpu(rcu_cpu_has_work, cpu) = 0;
1808 if (cpu_online(cpu))
1809 (void)rcu_spawn_one_cpu_kthread(cpu);
1811 rnp = rcu_get_root(rcu_state);
1812 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1813 if (NUM_RCU_NODES > 1) {
1814 rcu_for_each_leaf_node(rcu_state, rnp)
1815 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1819 early_initcall(rcu_spawn_kthreads);
1822 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1823 struct rcu_state *rsp)
1825 unsigned long flags;
1826 struct rcu_data *rdp;
1828 debug_rcu_head_queue(head);
1832 smp_mb(); /* Ensure RCU update seen before callback registry. */
1835 * Opportunistically note grace-period endings and beginnings.
1836 * Note that we might see a beginning right after we see an
1837 * end, but never vice versa, since this CPU has to pass through
1838 * a quiescent state betweentimes.
1840 local_irq_save(flags);
1841 rdp = this_cpu_ptr(rsp->rda);
1843 /* Add the callback to our list. */
1844 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1845 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1848 /* If interrupts were disabled, don't dive into RCU core. */
1849 if (irqs_disabled_flags(flags)) {
1850 local_irq_restore(flags);
1855 * Force the grace period if too many callbacks or too long waiting.
1856 * Enforce hysteresis, and don't invoke force_quiescent_state()
1857 * if some other CPU has recently done so. Also, don't bother
1858 * invoking force_quiescent_state() if the newly enqueued callback
1859 * is the only one waiting for a grace period to complete.
1861 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1863 /* Are we ignoring a completed grace period? */
1864 rcu_process_gp_end(rsp, rdp);
1865 check_for_new_grace_period(rsp, rdp);
1867 /* Start a new grace period if one not already started. */
1868 if (!rcu_gp_in_progress(rsp)) {
1869 unsigned long nestflag;
1870 struct rcu_node *rnp_root = rcu_get_root(rsp);
1872 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1873 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1875 /* Give the grace period a kick. */
1876 rdp->blimit = LONG_MAX;
1877 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1878 *rdp->nxttail[RCU_DONE_TAIL] != head)
1879 force_quiescent_state(rsp, 0);
1880 rdp->n_force_qs_snap = rsp->n_force_qs;
1881 rdp->qlen_last_fqs_check = rdp->qlen;
1883 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1884 force_quiescent_state(rsp, 1);
1885 local_irq_restore(flags);
1889 * Queue an RCU-sched callback for invocation after a grace period.
1891 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1893 __call_rcu(head, func, &rcu_sched_state);
1895 EXPORT_SYMBOL_GPL(call_rcu_sched);
1898 * Queue an RCU for invocation after a quicker grace period.
1900 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1902 __call_rcu(head, func, &rcu_bh_state);
1904 EXPORT_SYMBOL_GPL(call_rcu_bh);
1907 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1909 * Control will return to the caller some time after a full rcu-sched
1910 * grace period has elapsed, in other words after all currently executing
1911 * rcu-sched read-side critical sections have completed. These read-side
1912 * critical sections are delimited by rcu_read_lock_sched() and
1913 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1914 * local_irq_disable(), and so on may be used in place of
1915 * rcu_read_lock_sched().
1917 * This means that all preempt_disable code sequences, including NMI and
1918 * hardware-interrupt handlers, in progress on entry will have completed
1919 * before this primitive returns. However, this does not guarantee that
1920 * softirq handlers will have completed, since in some kernels, these
1921 * handlers can run in process context, and can block.
1923 * This primitive provides the guarantees made by the (now removed)
1924 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1925 * guarantees that rcu_read_lock() sections will have completed.
1926 * In "classic RCU", these two guarantees happen to be one and
1927 * the same, but can differ in realtime RCU implementations.
1929 void synchronize_sched(void)
1931 struct rcu_synchronize rcu;
1933 if (rcu_blocking_is_gp())
1936 init_rcu_head_on_stack(&rcu.head);
1937 init_completion(&rcu.completion);
1938 /* Will wake me after RCU finished. */
1939 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1941 wait_for_completion(&rcu.completion);
1942 destroy_rcu_head_on_stack(&rcu.head);
1944 EXPORT_SYMBOL_GPL(synchronize_sched);
1947 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1949 * Control will return to the caller some time after a full rcu_bh grace
1950 * period has elapsed, in other words after all currently executing rcu_bh
1951 * read-side critical sections have completed. RCU read-side critical
1952 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1953 * and may be nested.
1955 void synchronize_rcu_bh(void)
1957 struct rcu_synchronize rcu;
1959 if (rcu_blocking_is_gp())
1962 init_rcu_head_on_stack(&rcu.head);
1963 init_completion(&rcu.completion);
1964 /* Will wake me after RCU finished. */
1965 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1967 wait_for_completion(&rcu.completion);
1968 destroy_rcu_head_on_stack(&rcu.head);
1970 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1973 * Check to see if there is any immediate RCU-related work to be done
1974 * by the current CPU, for the specified type of RCU, returning 1 if so.
1975 * The checks are in order of increasing expense: checks that can be
1976 * carried out against CPU-local state are performed first. However,
1977 * we must check for CPU stalls first, else we might not get a chance.
1979 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1981 struct rcu_node *rnp = rdp->mynode;
1983 rdp->n_rcu_pending++;
1985 /* Check for CPU stalls, if enabled. */
1986 check_cpu_stall(rsp, rdp);
1988 /* Is the RCU core waiting for a quiescent state from this CPU? */
1989 if (rdp->qs_pending && !rdp->passed_quiesc) {
1992 * If force_quiescent_state() coming soon and this CPU
1993 * needs a quiescent state, and this is either RCU-sched
1994 * or RCU-bh, force a local reschedule.
1996 rdp->n_rp_qs_pending++;
1997 if (!rdp->preemptible &&
1998 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2001 } else if (rdp->qs_pending && rdp->passed_quiesc) {
2002 rdp->n_rp_report_qs++;
2006 /* Does this CPU have callbacks ready to invoke? */
2007 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2008 rdp->n_rp_cb_ready++;
2012 /* Has RCU gone idle with this CPU needing another grace period? */
2013 if (cpu_needs_another_gp(rsp, rdp)) {
2014 rdp->n_rp_cpu_needs_gp++;
2018 /* Has another RCU grace period completed? */
2019 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2020 rdp->n_rp_gp_completed++;
2024 /* Has a new RCU grace period started? */
2025 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2026 rdp->n_rp_gp_started++;
2030 /* Has an RCU GP gone long enough to send resched IPIs &c? */
2031 if (rcu_gp_in_progress(rsp) &&
2032 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
2033 rdp->n_rp_need_fqs++;
2038 rdp->n_rp_need_nothing++;
2043 * Check to see if there is any immediate RCU-related work to be done
2044 * by the current CPU, returning 1 if so. This function is part of the
2045 * RCU implementation; it is -not- an exported member of the RCU API.
2047 static int rcu_pending(int cpu)
2049 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
2050 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2051 rcu_preempt_pending(cpu);
2055 * Check to see if any future RCU-related work will need to be done
2056 * by the current CPU, even if none need be done immediately, returning
2059 static int rcu_needs_cpu_quick_check(int cpu)
2061 /* RCU callbacks either ready or pending? */
2062 return per_cpu(rcu_sched_data, cpu).nxtlist ||
2063 per_cpu(rcu_bh_data, cpu).nxtlist ||
2064 rcu_preempt_needs_cpu(cpu);
2067 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
2068 static atomic_t rcu_barrier_cpu_count;
2069 static DEFINE_MUTEX(rcu_barrier_mutex);
2070 static struct completion rcu_barrier_completion;
2072 static void rcu_barrier_callback(struct rcu_head *notused)
2074 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2075 complete(&rcu_barrier_completion);
2079 * Called with preemption disabled, and from cross-cpu IRQ context.
2081 static void rcu_barrier_func(void *type)
2083 int cpu = smp_processor_id();
2084 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
2085 void (*call_rcu_func)(struct rcu_head *head,
2086 void (*func)(struct rcu_head *head));
2088 atomic_inc(&rcu_barrier_cpu_count);
2089 call_rcu_func = type;
2090 call_rcu_func(head, rcu_barrier_callback);
2094 * Orchestrate the specified type of RCU barrier, waiting for all
2095 * RCU callbacks of the specified type to complete.
2097 static void _rcu_barrier(struct rcu_state *rsp,
2098 void (*call_rcu_func)(struct rcu_head *head,
2099 void (*func)(struct rcu_head *head)))
2101 BUG_ON(in_interrupt());
2102 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2103 mutex_lock(&rcu_barrier_mutex);
2104 init_completion(&rcu_barrier_completion);
2106 * Initialize rcu_barrier_cpu_count to 1, then invoke
2107 * rcu_barrier_func() on each CPU, so that each CPU also has
2108 * incremented rcu_barrier_cpu_count. Only then is it safe to
2109 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2110 * might complete its grace period before all of the other CPUs
2111 * did their increment, causing this function to return too
2112 * early. Note that on_each_cpu() disables irqs, which prevents
2113 * any CPUs from coming online or going offline until each online
2114 * CPU has queued its RCU-barrier callback.
2116 atomic_set(&rcu_barrier_cpu_count, 1);
2117 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
2118 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2119 complete(&rcu_barrier_completion);
2120 wait_for_completion(&rcu_barrier_completion);
2121 mutex_unlock(&rcu_barrier_mutex);
2125 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2127 void rcu_barrier_bh(void)
2129 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
2131 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2134 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2136 void rcu_barrier_sched(void)
2138 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
2140 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2143 * Do boot-time initialization of a CPU's per-CPU RCU data.
2146 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2148 unsigned long flags;
2150 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2151 struct rcu_node *rnp = rcu_get_root(rsp);
2153 /* Set up local state, ensuring consistent view of global state. */
2154 raw_spin_lock_irqsave(&rnp->lock, flags);
2155 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2156 rdp->nxtlist = NULL;
2157 for (i = 0; i < RCU_NEXT_SIZE; i++)
2158 rdp->nxttail[i] = &rdp->nxtlist;
2161 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2162 #endif /* #ifdef CONFIG_NO_HZ */
2164 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2168 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2169 * offline event can be happening at a given time. Note also that we
2170 * can accept some slop in the rsp->completed access due to the fact
2171 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2173 static void __cpuinit
2174 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
2176 unsigned long flags;
2178 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2179 struct rcu_node *rnp = rcu_get_root(rsp);
2181 /* Set up local state, ensuring consistent view of global state. */
2182 raw_spin_lock_irqsave(&rnp->lock, flags);
2183 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
2184 rdp->qs_pending = 1; /* so set up to respond to current GP. */
2185 rdp->beenonline = 1; /* We have now been online. */
2186 rdp->preemptible = preemptible;
2187 rdp->qlen_last_fqs_check = 0;
2188 rdp->n_force_qs_snap = rsp->n_force_qs;
2189 rdp->blimit = blimit;
2190 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2193 * A new grace period might start here. If so, we won't be part
2194 * of it, but that is OK, as we are currently in a quiescent state.
2197 /* Exclude any attempts to start a new GP on large systems. */
2198 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2200 /* Add CPU to rcu_node bitmasks. */
2202 mask = rdp->grpmask;
2204 /* Exclude any attempts to start a new GP on small systems. */
2205 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2206 rnp->qsmaskinit |= mask;
2207 mask = rnp->grpmask;
2208 if (rnp == rdp->mynode) {
2209 rdp->gpnum = rnp->completed; /* if GP in progress... */
2210 rdp->completed = rnp->completed;
2211 rdp->passed_quiesc_completed = rnp->completed - 1;
2213 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2215 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2217 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2220 static void __cpuinit rcu_prepare_cpu(int cpu)
2222 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2223 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2224 rcu_preempt_init_percpu_data(cpu);
2227 static void __cpuinit rcu_prepare_kthreads(int cpu)
2229 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2230 struct rcu_node *rnp = rdp->mynode;
2232 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2233 if (rcu_kthreads_spawnable) {
2234 (void)rcu_spawn_one_cpu_kthread(cpu);
2235 if (rnp->node_kthread_task == NULL)
2236 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
2241 * Handle CPU online/offline notification events.
2243 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2244 unsigned long action, void *hcpu)
2246 long cpu = (long)hcpu;
2247 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2248 struct rcu_node *rnp = rdp->mynode;
2251 case CPU_UP_PREPARE:
2252 case CPU_UP_PREPARE_FROZEN:
2253 rcu_prepare_cpu(cpu);
2254 rcu_prepare_kthreads(cpu);
2257 case CPU_DOWN_FAILED:
2258 rcu_node_kthread_setaffinity(rnp, -1);
2259 rcu_cpu_kthread_setrt(cpu, 1);
2261 case CPU_DOWN_PREPARE:
2262 rcu_node_kthread_setaffinity(rnp, cpu);
2263 rcu_cpu_kthread_setrt(cpu, 0);
2266 case CPU_DYING_FROZEN:
2268 * The whole machine is "stopped" except this CPU, so we can
2269 * touch any data without introducing corruption. We send the
2270 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2272 rcu_send_cbs_to_online(&rcu_bh_state);
2273 rcu_send_cbs_to_online(&rcu_sched_state);
2274 rcu_preempt_send_cbs_to_online();
2277 case CPU_DEAD_FROZEN:
2278 case CPU_UP_CANCELED:
2279 case CPU_UP_CANCELED_FROZEN:
2280 rcu_offline_cpu(cpu);
2289 * This function is invoked towards the end of the scheduler's initialization
2290 * process. Before this is called, the idle task might contain
2291 * RCU read-side critical sections (during which time, this idle
2292 * task is booting the system). After this function is called, the
2293 * idle tasks are prohibited from containing RCU read-side critical
2294 * sections. This function also enables RCU lockdep checking.
2296 void rcu_scheduler_starting(void)
2298 WARN_ON(num_online_cpus() != 1);
2299 WARN_ON(nr_context_switches() > 0);
2300 rcu_scheduler_active = 1;
2304 * Compute the per-level fanout, either using the exact fanout specified
2305 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2307 #ifdef CONFIG_RCU_FANOUT_EXACT
2308 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2312 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2313 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2314 rsp->levelspread[0] = RCU_FANOUT_LEAF;
2316 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2317 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2324 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2325 ccur = rsp->levelcnt[i];
2326 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2330 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2333 * Helper function for rcu_init() that initializes one rcu_state structure.
2335 static void __init rcu_init_one(struct rcu_state *rsp,
2336 struct rcu_data __percpu *rda)
2338 static char *buf[] = { "rcu_node_level_0",
2341 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2345 struct rcu_node *rnp;
2347 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2349 /* Initialize the level-tracking arrays. */
2351 for (i = 1; i < NUM_RCU_LVLS; i++)
2352 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2353 rcu_init_levelspread(rsp);
2355 /* Initialize the elements themselves, starting from the leaves. */
2357 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2358 cpustride *= rsp->levelspread[i];
2359 rnp = rsp->level[i];
2360 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2361 raw_spin_lock_init(&rnp->lock);
2362 lockdep_set_class_and_name(&rnp->lock,
2363 &rcu_node_class[i], buf[i]);
2366 rnp->qsmaskinit = 0;
2367 rnp->grplo = j * cpustride;
2368 rnp->grphi = (j + 1) * cpustride - 1;
2369 if (rnp->grphi >= NR_CPUS)
2370 rnp->grphi = NR_CPUS - 1;
2376 rnp->grpnum = j % rsp->levelspread[i - 1];
2377 rnp->grpmask = 1UL << rnp->grpnum;
2378 rnp->parent = rsp->level[i - 1] +
2379 j / rsp->levelspread[i - 1];
2382 INIT_LIST_HEAD(&rnp->blkd_tasks);
2387 rnp = rsp->level[NUM_RCU_LVLS - 1];
2388 for_each_possible_cpu(i) {
2389 while (i > rnp->grphi)
2391 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2392 rcu_boot_init_percpu_data(i, rsp);
2396 void __init rcu_init(void)
2400 rcu_bootup_announce();
2401 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2402 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2403 __rcu_init_preempt();
2404 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2407 * We don't need protection against CPU-hotplug here because
2408 * this is called early in boot, before either interrupts
2409 * or the scheduler are operational.
2411 cpu_notifier(rcu_cpu_notify, 0);
2412 for_each_online_cpu(cpu)
2413 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2414 check_cpu_stall_init();
2417 #include "rcutree_plugin.h"