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>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
55 /* Data structures. */
57 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
59 #define RCU_STATE_INITIALIZER(structname) { \
60 .level = { &structname.node[0] }, \
62 NUM_RCU_LVL_0, /* root of hierarchy. */ \
66 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
68 .signaled = RCU_GP_IDLE, \
71 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
72 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
74 .n_force_qs_ngp = 0, \
75 .name = #structname, \
78 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
79 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
81 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
82 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84 static struct rcu_state *rcu_state;
86 int rcu_scheduler_active __read_mostly;
87 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
90 * Control variables for per-CPU and per-rcu_node kthreads. These
91 * handle all flavors of RCU.
93 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
94 static DEFINE_PER_CPU(wait_queue_head_t, rcu_cpu_wq);
95 static DEFINE_PER_CPU(char, rcu_cpu_has_work);
96 static char rcu_kthreads_spawnable;
98 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
99 static void invoke_rcu_cpu_kthread(void);
101 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
104 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
105 * permit this function to be invoked without holding the root rcu_node
106 * structure's ->lock, but of course results can be subject to change.
108 static int rcu_gp_in_progress(struct rcu_state *rsp)
110 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
114 * Note a quiescent state. Because we do not need to know
115 * how many quiescent states passed, just if there was at least
116 * one since the start of the grace period, this just sets a flag.
118 void rcu_sched_qs(int cpu)
120 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
122 rdp->passed_quiesc_completed = rdp->gpnum - 1;
124 rdp->passed_quiesc = 1;
127 void rcu_bh_qs(int cpu)
129 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
131 rdp->passed_quiesc_completed = rdp->gpnum - 1;
133 rdp->passed_quiesc = 1;
137 * Note a context switch. This is a quiescent state for RCU-sched,
138 * and requires special handling for preemptible RCU.
140 void rcu_note_context_switch(int cpu)
143 rcu_preempt_note_context_switch(cpu);
147 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
148 .dynticks_nesting = 1,
149 .dynticks = ATOMIC_INIT(1),
151 #endif /* #ifdef CONFIG_NO_HZ */
153 static int blimit = 10; /* Maximum callbacks per softirq. */
154 static int qhimark = 10000; /* If this many pending, ignore blimit. */
155 static int qlowmark = 100; /* Once only this many pending, use blimit. */
157 module_param(blimit, int, 0);
158 module_param(qhimark, int, 0);
159 module_param(qlowmark, int, 0);
161 int rcu_cpu_stall_suppress __read_mostly;
162 module_param(rcu_cpu_stall_suppress, int, 0644);
164 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
165 static int rcu_pending(int cpu);
168 * Return the number of RCU-sched batches processed thus far for debug & stats.
170 long rcu_batches_completed_sched(void)
172 return rcu_sched_state.completed;
174 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
177 * Return the number of RCU BH batches processed thus far for debug & stats.
179 long rcu_batches_completed_bh(void)
181 return rcu_bh_state.completed;
183 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
186 * Force a quiescent state for RCU BH.
188 void rcu_bh_force_quiescent_state(void)
190 force_quiescent_state(&rcu_bh_state, 0);
192 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
195 * Force a quiescent state for RCU-sched.
197 void rcu_sched_force_quiescent_state(void)
199 force_quiescent_state(&rcu_sched_state, 0);
201 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
204 * Does the CPU have callbacks ready to be invoked?
207 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
209 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
213 * Does the current CPU require a yet-as-unscheduled grace period?
216 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
218 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
222 * Return the root node of the specified rcu_state structure.
224 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
226 return &rsp->node[0];
232 * If the specified CPU is offline, tell the caller that it is in
233 * a quiescent state. Otherwise, whack it with a reschedule IPI.
234 * Grace periods can end up waiting on an offline CPU when that
235 * CPU is in the process of coming online -- it will be added to the
236 * rcu_node bitmasks before it actually makes it online. The same thing
237 * can happen while a CPU is in the process of coming online. Because this
238 * race is quite rare, we check for it after detecting that the grace
239 * period has been delayed rather than checking each and every CPU
240 * each and every time we start a new grace period.
242 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
245 * If the CPU is offline, it is in a quiescent state. We can
246 * trust its state not to change because interrupts are disabled.
248 if (cpu_is_offline(rdp->cpu)) {
253 /* If preemptable RCU, no point in sending reschedule IPI. */
254 if (rdp->preemptable)
257 /* The CPU is online, so send it a reschedule IPI. */
258 if (rdp->cpu != smp_processor_id())
259 smp_send_reschedule(rdp->cpu);
266 #endif /* #ifdef CONFIG_SMP */
271 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
273 * Enter nohz mode, in other words, -leave- the mode in which RCU
274 * read-side critical sections can occur. (Though RCU read-side
275 * critical sections can occur in irq handlers in nohz mode, a possibility
276 * handled by rcu_irq_enter() and rcu_irq_exit()).
278 void rcu_enter_nohz(void)
281 struct rcu_dynticks *rdtp;
283 local_irq_save(flags);
284 rdtp = &__get_cpu_var(rcu_dynticks);
285 if (--rdtp->dynticks_nesting) {
286 local_irq_restore(flags);
289 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
290 smp_mb__before_atomic_inc(); /* See above. */
291 atomic_inc(&rdtp->dynticks);
292 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
293 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
294 local_irq_restore(flags);
296 /* If the interrupt queued a callback, get out of dyntick mode. */
298 (__get_cpu_var(rcu_sched_data).nxtlist ||
299 __get_cpu_var(rcu_bh_data).nxtlist ||
300 rcu_preempt_needs_cpu(smp_processor_id())))
305 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
307 * Exit nohz mode, in other words, -enter- the mode in which RCU
308 * read-side critical sections normally occur.
310 void rcu_exit_nohz(void)
313 struct rcu_dynticks *rdtp;
315 local_irq_save(flags);
316 rdtp = &__get_cpu_var(rcu_dynticks);
317 if (rdtp->dynticks_nesting++) {
318 local_irq_restore(flags);
321 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
322 atomic_inc(&rdtp->dynticks);
323 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
324 smp_mb__after_atomic_inc(); /* See above. */
325 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
326 local_irq_restore(flags);
330 * rcu_nmi_enter - inform RCU of entry to NMI context
332 * If the CPU was idle with dynamic ticks active, and there is no
333 * irq handler running, this updates rdtp->dynticks_nmi to let the
334 * RCU grace-period handling know that the CPU is active.
336 void rcu_nmi_enter(void)
338 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
340 if (rdtp->dynticks_nmi_nesting == 0 &&
341 (atomic_read(&rdtp->dynticks) & 0x1))
343 rdtp->dynticks_nmi_nesting++;
344 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
345 atomic_inc(&rdtp->dynticks);
346 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
347 smp_mb__after_atomic_inc(); /* See above. */
348 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
352 * rcu_nmi_exit - inform RCU of exit from NMI context
354 * If the CPU was idle with dynamic ticks active, and there is no
355 * irq handler running, this updates rdtp->dynticks_nmi to let the
356 * RCU grace-period handling know that the CPU is no longer active.
358 void rcu_nmi_exit(void)
360 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
362 if (rdtp->dynticks_nmi_nesting == 0 ||
363 --rdtp->dynticks_nmi_nesting != 0)
365 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
366 smp_mb__before_atomic_inc(); /* See above. */
367 atomic_inc(&rdtp->dynticks);
368 smp_mb__after_atomic_inc(); /* Force delay to next write. */
369 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
373 * rcu_irq_enter - inform RCU of entry to hard irq context
375 * If the CPU was idle with dynamic ticks active, this updates the
376 * rdtp->dynticks to let the RCU handling know that the CPU is active.
378 void rcu_irq_enter(void)
384 * rcu_irq_exit - inform RCU of exit from hard irq context
386 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
387 * to put let the RCU handling be aware that the CPU is going back to idle
390 void rcu_irq_exit(void)
398 * Snapshot the specified CPU's dynticks counter so that we can later
399 * credit them with an implicit quiescent state. Return 1 if this CPU
400 * is in dynticks idle mode, which is an extended quiescent state.
402 static int dyntick_save_progress_counter(struct rcu_data *rdp)
404 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
409 * Return true if the specified CPU has passed through a quiescent
410 * state by virtue of being in or having passed through an dynticks
411 * idle state since the last call to dyntick_save_progress_counter()
414 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
419 curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
420 snap = (unsigned long)rdp->dynticks_snap;
423 * If the CPU passed through or entered a dynticks idle phase with
424 * no active irq/NMI handlers, then we can safely pretend that the CPU
425 * already acknowledged the request to pass through a quiescent
426 * state. Either way, that CPU cannot possibly be in an RCU
427 * read-side critical section that started before the beginning
428 * of the current RCU grace period.
430 if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
435 /* Go check for the CPU being offline. */
436 return rcu_implicit_offline_qs(rdp);
439 #endif /* #ifdef CONFIG_SMP */
441 #else /* #ifdef CONFIG_NO_HZ */
445 static int dyntick_save_progress_counter(struct rcu_data *rdp)
450 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
452 return rcu_implicit_offline_qs(rdp);
455 #endif /* #ifdef CONFIG_SMP */
457 #endif /* #else #ifdef CONFIG_NO_HZ */
459 int rcu_cpu_stall_suppress __read_mostly;
461 static void record_gp_stall_check_time(struct rcu_state *rsp)
463 rsp->gp_start = jiffies;
464 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
467 static void print_other_cpu_stall(struct rcu_state *rsp)
472 struct rcu_node *rnp = rcu_get_root(rsp);
474 /* Only let one CPU complain about others per time interval. */
476 raw_spin_lock_irqsave(&rnp->lock, flags);
477 delta = jiffies - rsp->jiffies_stall;
478 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
479 raw_spin_unlock_irqrestore(&rnp->lock, flags);
482 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
485 * Now rat on any tasks that got kicked up to the root rcu_node
486 * due to CPU offlining.
488 rcu_print_task_stall(rnp);
489 raw_spin_unlock_irqrestore(&rnp->lock, flags);
492 * OK, time to rat on our buddy...
493 * See Documentation/RCU/stallwarn.txt for info on how to debug
494 * RCU CPU stall warnings.
496 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
498 rcu_for_each_leaf_node(rsp, rnp) {
499 raw_spin_lock_irqsave(&rnp->lock, flags);
500 rcu_print_task_stall(rnp);
501 raw_spin_unlock_irqrestore(&rnp->lock, flags);
502 if (rnp->qsmask == 0)
504 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
505 if (rnp->qsmask & (1UL << cpu))
506 printk(" %d", rnp->grplo + cpu);
508 printk("} (detected by %d, t=%ld jiffies)\n",
509 smp_processor_id(), (long)(jiffies - rsp->gp_start));
510 trigger_all_cpu_backtrace();
512 /* If so configured, complain about tasks blocking the grace period. */
514 rcu_print_detail_task_stall(rsp);
516 force_quiescent_state(rsp, 0); /* Kick them all. */
519 static void print_cpu_stall(struct rcu_state *rsp)
522 struct rcu_node *rnp = rcu_get_root(rsp);
525 * OK, time to rat on ourselves...
526 * See Documentation/RCU/stallwarn.txt for info on how to debug
527 * RCU CPU stall warnings.
529 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
530 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
531 trigger_all_cpu_backtrace();
533 raw_spin_lock_irqsave(&rnp->lock, flags);
534 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
536 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
537 raw_spin_unlock_irqrestore(&rnp->lock, flags);
539 set_need_resched(); /* kick ourselves to get things going. */
542 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
545 struct rcu_node *rnp;
547 if (rcu_cpu_stall_suppress)
549 delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall);
551 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && delta >= 0) {
553 /* We haven't checked in, so go dump stack. */
554 print_cpu_stall(rsp);
556 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
558 /* They had two time units to dump stack, so complain. */
559 print_other_cpu_stall(rsp);
563 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
565 rcu_cpu_stall_suppress = 1;
570 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
572 * Set the stall-warning timeout way off into the future, thus preventing
573 * any RCU CPU stall-warning messages from appearing in the current set of
576 * The caller must disable hard irqs.
578 void rcu_cpu_stall_reset(void)
580 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
581 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
582 rcu_preempt_stall_reset();
585 static struct notifier_block rcu_panic_block = {
586 .notifier_call = rcu_panic,
589 static void __init check_cpu_stall_init(void)
591 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
595 * Update CPU-local rcu_data state to record the newly noticed grace period.
596 * This is used both when we started the grace period and when we notice
597 * that someone else started the grace period. The caller must hold the
598 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
599 * and must have irqs disabled.
601 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
603 if (rdp->gpnum != rnp->gpnum) {
605 * If the current grace period is waiting for this CPU,
606 * set up to detect a quiescent state, otherwise don't
607 * go looking for one.
609 rdp->gpnum = rnp->gpnum;
610 if (rnp->qsmask & rdp->grpmask) {
612 rdp->passed_quiesc = 0;
618 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
621 struct rcu_node *rnp;
623 local_irq_save(flags);
625 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
626 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
627 local_irq_restore(flags);
630 __note_new_gpnum(rsp, rnp, rdp);
631 raw_spin_unlock_irqrestore(&rnp->lock, flags);
635 * Did someone else start a new RCU grace period start since we last
636 * checked? Update local state appropriately if so. Must be called
637 * on the CPU corresponding to rdp.
640 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
645 local_irq_save(flags);
646 if (rdp->gpnum != rsp->gpnum) {
647 note_new_gpnum(rsp, rdp);
650 local_irq_restore(flags);
655 * Advance this CPU's callbacks, but only if the current grace period
656 * has ended. This may be called only from the CPU to whom the rdp
657 * belongs. In addition, the corresponding leaf rcu_node structure's
658 * ->lock must be held by the caller, with irqs disabled.
661 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
663 /* Did another grace period end? */
664 if (rdp->completed != rnp->completed) {
666 /* Advance callbacks. No harm if list empty. */
667 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
668 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
669 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
671 /* Remember that we saw this grace-period completion. */
672 rdp->completed = rnp->completed;
675 * If we were in an extended quiescent state, we may have
676 * missed some grace periods that others CPUs handled on
677 * our behalf. Catch up with this state to avoid noting
678 * spurious new grace periods. If another grace period
679 * has started, then rnp->gpnum will have advanced, so
680 * we will detect this later on.
682 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
683 rdp->gpnum = rdp->completed;
686 * If RCU does not need a quiescent state from this CPU,
687 * then make sure that this CPU doesn't go looking for one.
689 if ((rnp->qsmask & rdp->grpmask) == 0)
695 * Advance this CPU's callbacks, but only if the current grace period
696 * has ended. This may be called only from the CPU to whom the rdp
700 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
703 struct rcu_node *rnp;
705 local_irq_save(flags);
707 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
708 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
709 local_irq_restore(flags);
712 __rcu_process_gp_end(rsp, rnp, rdp);
713 raw_spin_unlock_irqrestore(&rnp->lock, flags);
717 * Do per-CPU grace-period initialization for running CPU. The caller
718 * must hold the lock of the leaf rcu_node structure corresponding to
722 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
724 /* Prior grace period ended, so advance callbacks for current CPU. */
725 __rcu_process_gp_end(rsp, rnp, rdp);
728 * Because this CPU just now started the new grace period, we know
729 * that all of its callbacks will be covered by this upcoming grace
730 * period, even the ones that were registered arbitrarily recently.
731 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
733 * Other CPUs cannot be sure exactly when the grace period started.
734 * Therefore, their recently registered callbacks must pass through
735 * an additional RCU_NEXT_READY stage, so that they will be handled
736 * by the next RCU grace period.
738 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
739 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
741 /* Set state so that this CPU will detect the next quiescent state. */
742 __note_new_gpnum(rsp, rnp, rdp);
746 * Start a new RCU grace period if warranted, re-initializing the hierarchy
747 * in preparation for detecting the next grace period. The caller must hold
748 * the root node's ->lock, which is released before return. Hard irqs must
752 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
753 __releases(rcu_get_root(rsp)->lock)
755 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
756 struct rcu_node *rnp = rcu_get_root(rsp);
758 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
759 if (cpu_needs_another_gp(rsp, rdp))
760 rsp->fqs_need_gp = 1;
761 if (rnp->completed == rsp->completed) {
762 raw_spin_unlock_irqrestore(&rnp->lock, flags);
765 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
768 * Propagate new ->completed value to rcu_node structures
769 * so that other CPUs don't have to wait until the start
770 * of the next grace period to process their callbacks.
772 rcu_for_each_node_breadth_first(rsp, rnp) {
773 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
774 rnp->completed = rsp->completed;
775 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
777 local_irq_restore(flags);
781 /* Advance to a new grace period and initialize state. */
783 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
784 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
785 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
786 record_gp_stall_check_time(rsp);
788 /* Special-case the common single-level case. */
789 if (NUM_RCU_NODES == 1) {
790 rcu_preempt_check_blocked_tasks(rnp);
791 rnp->qsmask = rnp->qsmaskinit;
792 rnp->gpnum = rsp->gpnum;
793 rnp->completed = rsp->completed;
794 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
795 rcu_start_gp_per_cpu(rsp, rnp, rdp);
796 rcu_preempt_boost_start_gp(rnp);
797 raw_spin_unlock_irqrestore(&rnp->lock, flags);
801 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
804 /* Exclude any concurrent CPU-hotplug operations. */
805 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
808 * Set the quiescent-state-needed bits in all the rcu_node
809 * structures for all currently online CPUs in breadth-first
810 * order, starting from the root rcu_node structure. This
811 * operation relies on the layout of the hierarchy within the
812 * rsp->node[] array. Note that other CPUs will access only
813 * the leaves of the hierarchy, which still indicate that no
814 * grace period is in progress, at least until the corresponding
815 * leaf node has been initialized. In addition, we have excluded
816 * CPU-hotplug operations.
818 * Note that the grace period cannot complete until we finish
819 * the initialization process, as there will be at least one
820 * qsmask bit set in the root node until that time, namely the
821 * one corresponding to this CPU, due to the fact that we have
824 rcu_for_each_node_breadth_first(rsp, rnp) {
825 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
826 rcu_preempt_check_blocked_tasks(rnp);
827 rnp->qsmask = rnp->qsmaskinit;
828 rnp->gpnum = rsp->gpnum;
829 rnp->completed = rsp->completed;
830 if (rnp == rdp->mynode)
831 rcu_start_gp_per_cpu(rsp, rnp, rdp);
832 rcu_preempt_boost_start_gp(rnp);
833 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
836 rnp = rcu_get_root(rsp);
837 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
838 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
839 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
840 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
844 * Report a full set of quiescent states to the specified rcu_state
845 * data structure. This involves cleaning up after the prior grace
846 * period and letting rcu_start_gp() start up the next grace period
847 * if one is needed. Note that the caller must hold rnp->lock, as
848 * required by rcu_start_gp(), which will release it.
850 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
851 __releases(rcu_get_root(rsp)->lock)
853 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
856 * Ensure that all grace-period and pre-grace-period activity
857 * is seen before the assignment to rsp->completed.
859 smp_mb(); /* See above block comment. */
860 rsp->completed = rsp->gpnum;
861 rsp->signaled = RCU_GP_IDLE;
862 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
866 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
867 * Allows quiescent states for a group of CPUs to be reported at one go
868 * to the specified rcu_node structure, though all the CPUs in the group
869 * must be represented by the same rcu_node structure (which need not be
870 * a leaf rcu_node structure, though it often will be). That structure's
871 * lock must be held upon entry, and it is released before return.
874 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
875 struct rcu_node *rnp, unsigned long flags)
876 __releases(rnp->lock)
878 struct rcu_node *rnp_c;
880 /* Walk up the rcu_node hierarchy. */
882 if (!(rnp->qsmask & mask)) {
884 /* Our bit has already been cleared, so done. */
885 raw_spin_unlock_irqrestore(&rnp->lock, flags);
888 rnp->qsmask &= ~mask;
889 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
891 /* Other bits still set at this level, so done. */
892 raw_spin_unlock_irqrestore(&rnp->lock, flags);
896 if (rnp->parent == NULL) {
898 /* No more levels. Exit loop holding root lock. */
902 raw_spin_unlock_irqrestore(&rnp->lock, flags);
905 raw_spin_lock_irqsave(&rnp->lock, flags);
906 WARN_ON_ONCE(rnp_c->qsmask);
910 * Get here if we are the last CPU to pass through a quiescent
911 * state for this grace period. Invoke rcu_report_qs_rsp()
912 * to clean up and start the next grace period if one is needed.
914 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
918 * Record a quiescent state for the specified CPU to that CPU's rcu_data
919 * structure. This must be either called from the specified CPU, or
920 * called when the specified CPU is known to be offline (and when it is
921 * also known that no other CPU is concurrently trying to help the offline
922 * CPU). The lastcomp argument is used to make sure we are still in the
923 * grace period of interest. We don't want to end the current grace period
924 * based on quiescent states detected in an earlier grace period!
927 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
931 struct rcu_node *rnp;
934 raw_spin_lock_irqsave(&rnp->lock, flags);
935 if (lastcomp != rnp->completed) {
938 * Someone beat us to it for this grace period, so leave.
939 * The race with GP start is resolved by the fact that we
940 * hold the leaf rcu_node lock, so that the per-CPU bits
941 * cannot yet be initialized -- so we would simply find our
942 * CPU's bit already cleared in rcu_report_qs_rnp() if this
945 rdp->passed_quiesc = 0; /* try again later! */
946 raw_spin_unlock_irqrestore(&rnp->lock, flags);
950 if ((rnp->qsmask & mask) == 0) {
951 raw_spin_unlock_irqrestore(&rnp->lock, flags);
956 * This GP can't end until cpu checks in, so all of our
957 * callbacks can be processed during the next GP.
959 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
961 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
966 * Check to see if there is a new grace period of which this CPU
967 * is not yet aware, and if so, set up local rcu_data state for it.
968 * Otherwise, see if this CPU has just passed through its first
969 * quiescent state for this grace period, and record that fact if so.
972 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
974 /* If there is now a new grace period, record and return. */
975 if (check_for_new_grace_period(rsp, rdp))
979 * Does this CPU still need to do its part for current grace period?
980 * If no, return and let the other CPUs do their part as well.
982 if (!rdp->qs_pending)
986 * Was there a quiescent state since the beginning of the grace
987 * period? If no, then exit and wait for the next call.
989 if (!rdp->passed_quiesc)
993 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
996 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
999 #ifdef CONFIG_HOTPLUG_CPU
1002 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1003 * Synchronization is not required because this function executes
1004 * in stop_machine() context.
1006 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1009 /* current DYING CPU is cleared in the cpu_online_mask */
1010 int receive_cpu = cpumask_any(cpu_online_mask);
1011 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1012 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1014 if (rdp->nxtlist == NULL)
1015 return; /* irqs disabled, so comparison is stable. */
1017 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1018 receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1019 receive_rdp->qlen += rdp->qlen;
1020 receive_rdp->n_cbs_adopted += rdp->qlen;
1021 rdp->n_cbs_orphaned += rdp->qlen;
1023 rdp->nxtlist = NULL;
1024 for (i = 0; i < RCU_NEXT_SIZE; i++)
1025 rdp->nxttail[i] = &rdp->nxtlist;
1030 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1031 * and move all callbacks from the outgoing CPU to the current one.
1032 * There can only be one CPU hotplug operation at a time, so no other
1033 * CPU can be attempting to update rcu_cpu_kthread_task.
1035 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1037 unsigned long flags;
1039 int need_report = 0;
1040 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1041 struct rcu_node *rnp;
1042 struct task_struct *t;
1044 /* Stop the CPU's kthread. */
1045 t = per_cpu(rcu_cpu_kthread_task, cpu);
1047 per_cpu(rcu_cpu_kthread_task, cpu) = NULL;
1051 /* Exclude any attempts to start a new grace period. */
1052 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1054 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1055 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
1056 mask = rdp->grpmask; /* rnp->grplo is constant. */
1058 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1059 rnp->qsmaskinit &= ~mask;
1060 if (rnp->qsmaskinit != 0) {
1061 if (rnp != rdp->mynode)
1062 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1065 if (rnp == rdp->mynode)
1066 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1068 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1069 mask = rnp->grpmask;
1071 } while (rnp != NULL);
1074 * We still hold the leaf rcu_node structure lock here, and
1075 * irqs are still disabled. The reason for this subterfuge is
1076 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1077 * held leads to deadlock.
1079 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1081 if (need_report & RCU_OFL_TASKS_NORM_GP)
1082 rcu_report_unblock_qs_rnp(rnp, flags);
1084 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1085 if (need_report & RCU_OFL_TASKS_EXP_GP)
1086 rcu_report_exp_rnp(rsp, rnp);
1089 * If there are no more online CPUs for this rcu_node structure,
1090 * kill the rcu_node structure's kthread. Otherwise, adjust its
1093 t = rnp->node_kthread_task;
1095 rnp->qsmaskinit == 0) {
1096 raw_spin_lock_irqsave(&rnp->lock, flags);
1097 rnp->node_kthread_task = NULL;
1098 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1100 rcu_stop_boost_kthread(rnp);
1102 rcu_node_kthread_setaffinity(rnp, -1);
1106 * Remove the specified CPU from the RCU hierarchy and move any pending
1107 * callbacks that it might have to the current CPU. This code assumes
1108 * that at least one CPU in the system will remain running at all times.
1109 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1111 static void rcu_offline_cpu(int cpu)
1113 __rcu_offline_cpu(cpu, &rcu_sched_state);
1114 __rcu_offline_cpu(cpu, &rcu_bh_state);
1115 rcu_preempt_offline_cpu(cpu);
1118 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1120 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1124 static void rcu_offline_cpu(int cpu)
1128 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1131 * Invoke any RCU callbacks that have made it to the end of their grace
1132 * period. Thottle as specified by rdp->blimit.
1134 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1136 unsigned long flags;
1137 struct rcu_head *next, *list, **tail;
1140 /* If no callbacks are ready, just return.*/
1141 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1145 * Extract the list of ready callbacks, disabling to prevent
1146 * races with call_rcu() from interrupt handlers.
1148 local_irq_save(flags);
1149 list = rdp->nxtlist;
1150 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1151 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1152 tail = rdp->nxttail[RCU_DONE_TAIL];
1153 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1154 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1155 rdp->nxttail[count] = &rdp->nxtlist;
1156 local_irq_restore(flags);
1158 /* Invoke callbacks. */
1163 debug_rcu_head_unqueue(list);
1166 if (++count >= rdp->blimit)
1170 local_irq_save(flags);
1172 /* Update count, and requeue any remaining callbacks. */
1174 rdp->n_cbs_invoked += count;
1176 *tail = rdp->nxtlist;
1177 rdp->nxtlist = list;
1178 for (count = 0; count < RCU_NEXT_SIZE; count++)
1179 if (&rdp->nxtlist == rdp->nxttail[count])
1180 rdp->nxttail[count] = tail;
1185 /* Reinstate batch limit if we have worked down the excess. */
1186 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1187 rdp->blimit = blimit;
1189 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1190 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1191 rdp->qlen_last_fqs_check = 0;
1192 rdp->n_force_qs_snap = rsp->n_force_qs;
1193 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1194 rdp->qlen_last_fqs_check = rdp->qlen;
1196 local_irq_restore(flags);
1198 /* Re-raise the RCU softirq if there are callbacks remaining. */
1199 if (cpu_has_callbacks_ready_to_invoke(rdp))
1200 invoke_rcu_cpu_kthread();
1204 * Check to see if this CPU is in a non-context-switch quiescent state
1205 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1206 * Also schedule the RCU softirq handler.
1208 * This function must be called with hardirqs disabled. It is normally
1209 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1210 * false, there is no point in invoking rcu_check_callbacks().
1212 void rcu_check_callbacks(int cpu, int user)
1215 (idle_cpu(cpu) && rcu_scheduler_active &&
1216 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1219 * Get here if this CPU took its interrupt from user
1220 * mode or from the idle loop, and if this is not a
1221 * nested interrupt. In this case, the CPU is in
1222 * a quiescent state, so note it.
1224 * No memory barrier is required here because both
1225 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1226 * variables that other CPUs neither access nor modify,
1227 * at least not while the corresponding CPU is online.
1233 } else if (!in_softirq()) {
1236 * Get here if this CPU did not take its interrupt from
1237 * softirq, in other words, if it is not interrupting
1238 * a rcu_bh read-side critical section. This is an _bh
1239 * critical section, so note it.
1244 rcu_preempt_check_callbacks(cpu);
1245 if (rcu_pending(cpu))
1246 invoke_rcu_cpu_kthread();
1252 * Scan the leaf rcu_node structures, processing dyntick state for any that
1253 * have not yet encountered a quiescent state, using the function specified.
1254 * Also initiate boosting for any threads blocked on the root rcu_node.
1256 * The caller must have suppressed start of new grace periods.
1258 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1262 unsigned long flags;
1264 struct rcu_node *rnp;
1266 rcu_for_each_leaf_node(rsp, rnp) {
1268 raw_spin_lock_irqsave(&rnp->lock, flags);
1269 if (!rcu_gp_in_progress(rsp)) {
1270 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1273 if (rnp->qsmask == 0) {
1274 rcu_initiate_boost(rnp);
1275 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1280 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1281 if ((rnp->qsmask & bit) != 0 &&
1282 f(per_cpu_ptr(rsp->rda, cpu)))
1287 /* rcu_report_qs_rnp() releases rnp->lock. */
1288 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1291 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1293 rnp = rcu_get_root(rsp);
1294 raw_spin_lock_irqsave(&rnp->lock, flags);
1295 if (rnp->qsmask == 0)
1296 rcu_initiate_boost(rnp);
1297 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1301 * Force quiescent states on reluctant CPUs, and also detect which
1302 * CPUs are in dyntick-idle mode.
1304 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1306 unsigned long flags;
1307 struct rcu_node *rnp = rcu_get_root(rsp);
1309 if (!rcu_gp_in_progress(rsp))
1310 return; /* No grace period in progress, nothing to force. */
1311 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1312 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1313 return; /* Someone else is already on the job. */
1315 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1316 goto unlock_fqs_ret; /* no emergency and done recently. */
1318 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1319 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1320 if(!rcu_gp_in_progress(rsp)) {
1321 rsp->n_force_qs_ngp++;
1322 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1323 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1325 rsp->fqs_active = 1;
1326 switch (rsp->signaled) {
1330 break; /* grace period idle or initializing, ignore. */
1332 case RCU_SAVE_DYNTICK:
1333 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1334 break; /* So gcc recognizes the dead code. */
1336 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1338 /* Record dyntick-idle state. */
1339 force_qs_rnp(rsp, dyntick_save_progress_counter);
1340 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1341 if (rcu_gp_in_progress(rsp))
1342 rsp->signaled = RCU_FORCE_QS;
1347 /* Check dyntick-idle state, send IPI to laggarts. */
1348 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1349 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1351 /* Leave state in case more forcing is required. */
1353 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1356 rsp->fqs_active = 0;
1357 if (rsp->fqs_need_gp) {
1358 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1359 rsp->fqs_need_gp = 0;
1360 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1363 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1365 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1368 #else /* #ifdef CONFIG_SMP */
1370 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1375 #endif /* #else #ifdef CONFIG_SMP */
1378 * This does the RCU processing work from softirq context for the
1379 * specified rcu_state and rcu_data structures. This may be called
1380 * only from the CPU to whom the rdp belongs.
1383 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1385 unsigned long flags;
1387 WARN_ON_ONCE(rdp->beenonline == 0);
1390 * If an RCU GP has gone long enough, go check for dyntick
1391 * idle CPUs and, if needed, send resched IPIs.
1393 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1394 force_quiescent_state(rsp, 1);
1397 * Advance callbacks in response to end of earlier grace
1398 * period that some other CPU ended.
1400 rcu_process_gp_end(rsp, rdp);
1402 /* Update RCU state based on any recent quiescent states. */
1403 rcu_check_quiescent_state(rsp, rdp);
1405 /* Does this CPU require a not-yet-started grace period? */
1406 if (cpu_needs_another_gp(rsp, rdp)) {
1407 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1408 rcu_start_gp(rsp, flags); /* releases above lock */
1411 /* If there are callbacks ready, invoke them. */
1412 rcu_do_batch(rsp, rdp);
1416 * Do softirq processing for the current CPU.
1418 static void rcu_process_callbacks(void)
1420 __rcu_process_callbacks(&rcu_sched_state,
1421 &__get_cpu_var(rcu_sched_data));
1422 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1423 rcu_preempt_process_callbacks();
1425 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1426 rcu_needs_cpu_flush();
1430 * Wake up the current CPU's kthread. This replaces raise_softirq()
1431 * in earlier versions of RCU. Note that because we are running on
1432 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1433 * cannot disappear out from under us.
1435 static void invoke_rcu_cpu_kthread(void)
1437 unsigned long flags;
1438 wait_queue_head_t *q;
1441 local_irq_save(flags);
1442 cpu = smp_processor_id();
1443 per_cpu(rcu_cpu_has_work, cpu) = 1;
1444 if (per_cpu(rcu_cpu_kthread_task, cpu) == NULL) {
1445 local_irq_restore(flags);
1448 q = &per_cpu(rcu_cpu_wq, cpu);
1450 local_irq_restore(flags);
1454 * Wake up the specified per-rcu_node-structure kthread.
1455 * The caller must hold ->lock.
1457 static void invoke_rcu_node_kthread(struct rcu_node *rnp)
1459 struct task_struct *t;
1461 t = rnp->node_kthread_task;
1467 * Timer handler to initiate the waking up of per-CPU kthreads that
1468 * have yielded the CPU due to excess numbers of RCU callbacks.
1469 * We wake up the per-rcu_node kthread, which in turn will wake up
1470 * the booster kthread.
1472 static void rcu_cpu_kthread_timer(unsigned long arg)
1474 unsigned long flags;
1475 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg);
1476 struct rcu_node *rnp = rdp->mynode;
1478 raw_spin_lock_irqsave(&rnp->lock, flags);
1479 rnp->wakemask |= rdp->grpmask;
1480 invoke_rcu_node_kthread(rnp);
1481 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1485 * Drop to non-real-time priority and yield, but only after posting a
1486 * timer that will cause us to regain our real-time priority if we
1487 * remain preempted. Either way, we restore our real-time priority
1490 static void rcu_yield(void (*f)(unsigned long), unsigned long arg)
1492 struct sched_param sp;
1493 struct timer_list yield_timer;
1495 setup_timer_on_stack(&yield_timer, f, arg);
1496 mod_timer(&yield_timer, jiffies + 2);
1497 sp.sched_priority = 0;
1498 sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp);
1500 sp.sched_priority = RCU_KTHREAD_PRIO;
1501 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1502 del_timer(&yield_timer);
1506 * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU.
1507 * This can happen while the corresponding CPU is either coming online
1508 * or going offline. We cannot wait until the CPU is fully online
1509 * before starting the kthread, because the various notifier functions
1510 * can wait for RCU grace periods. So we park rcu_cpu_kthread() until
1511 * the corresponding CPU is online.
1513 * Return 1 if the kthread needs to stop, 0 otherwise.
1515 * Caller must disable bh. This function can momentarily enable it.
1517 static int rcu_cpu_kthread_should_stop(int cpu)
1519 while (cpu_is_offline(cpu) ||
1520 !cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)) ||
1521 smp_processor_id() != cpu) {
1522 if (kthread_should_stop())
1525 schedule_timeout_uninterruptible(1);
1526 if (!cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)))
1527 set_cpus_allowed_ptr(current, cpumask_of(cpu));
1534 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1535 * earlier RCU softirq.
1537 static int rcu_cpu_kthread(void *arg)
1539 int cpu = (int)(long)arg;
1540 unsigned long flags;
1542 wait_queue_head_t *wqp = &per_cpu(rcu_cpu_wq, cpu);
1544 char *workp = &per_cpu(rcu_cpu_has_work, cpu);
1547 wait_event_interruptible(*wqp,
1548 *workp != 0 || kthread_should_stop());
1550 if (rcu_cpu_kthread_should_stop(cpu)) {
1554 local_irq_save(flags);
1557 local_irq_restore(flags);
1559 rcu_process_callbacks();
1566 rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu);
1574 * Spawn a per-CPU kthread, setting up affinity and priority.
1575 * Because the CPU hotplug lock is held, no other CPU will be attempting
1576 * to manipulate rcu_cpu_kthread_task. There might be another CPU
1577 * attempting to access it during boot, but the locking in kthread_bind()
1578 * will enforce sufficient ordering.
1580 static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu)
1582 struct sched_param sp;
1583 struct task_struct *t;
1585 if (!rcu_kthreads_spawnable ||
1586 per_cpu(rcu_cpu_kthread_task, cpu) != NULL)
1588 t = kthread_create(rcu_cpu_kthread, (void *)(long)cpu, "rcuc%d", cpu);
1591 kthread_bind(t, cpu);
1592 WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL);
1593 per_cpu(rcu_cpu_kthread_task, cpu) = t;
1595 sp.sched_priority = RCU_KTHREAD_PRIO;
1596 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1601 * Per-rcu_node kthread, which is in charge of waking up the per-CPU
1602 * kthreads when needed. We ignore requests to wake up kthreads
1603 * for offline CPUs, which is OK because force_quiescent_state()
1604 * takes care of this case.
1606 static int rcu_node_kthread(void *arg)
1609 unsigned long flags;
1611 struct rcu_node *rnp = (struct rcu_node *)arg;
1612 struct sched_param sp;
1613 struct task_struct *t;
1616 wait_event_interruptible(rnp->node_wq, rnp->wakemask != 0 ||
1617 kthread_should_stop());
1618 if (kthread_should_stop())
1620 raw_spin_lock_irqsave(&rnp->lock, flags);
1621 mask = rnp->wakemask;
1623 rcu_initiate_boost(rnp);
1624 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1625 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) {
1626 if ((mask & 0x1) == 0)
1629 t = per_cpu(rcu_cpu_kthread_task, cpu);
1630 if (!cpu_online(cpu) || t == NULL) {
1634 per_cpu(rcu_cpu_has_work, cpu) = 1;
1635 sp.sched_priority = RCU_KTHREAD_PRIO;
1636 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1644 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1645 * served by the rcu_node in question. The CPU hotplug lock is still
1646 * held, so the value of rnp->qsmaskinit will be stable.
1648 * We don't include outgoingcpu in the affinity set, use -1 if there is
1649 * no outgoing CPU. If there are no CPUs left in the affinity set,
1650 * this function allows the kthread to execute on any CPU.
1652 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1656 unsigned long mask = rnp->qsmaskinit;
1658 if (rnp->node_kthread_task == NULL || mask == 0)
1660 if (!alloc_cpumask_var(&cm, GFP_KERNEL))
1663 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1664 if ((mask & 0x1) && cpu != outgoingcpu)
1665 cpumask_set_cpu(cpu, cm);
1666 if (cpumask_weight(cm) == 0) {
1668 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1669 cpumask_clear_cpu(cpu, cm);
1670 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1672 set_cpus_allowed_ptr(rnp->node_kthread_task, cm);
1673 rcu_boost_kthread_setaffinity(rnp, cm);
1674 free_cpumask_var(cm);
1678 * Spawn a per-rcu_node kthread, setting priority and affinity.
1679 * Called during boot before online/offline can happen, or, if
1680 * during runtime, with the main CPU-hotplug locks held. So only
1681 * one of these can be executing at a time.
1683 static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp,
1684 struct rcu_node *rnp)
1686 unsigned long flags;
1687 int rnp_index = rnp - &rsp->node[0];
1688 struct sched_param sp;
1689 struct task_struct *t;
1691 if (!rcu_kthreads_spawnable ||
1692 rnp->qsmaskinit == 0)
1694 if (rnp->node_kthread_task == NULL) {
1695 t = kthread_create(rcu_node_kthread, (void *)rnp,
1696 "rcun%d", rnp_index);
1699 raw_spin_lock_irqsave(&rnp->lock, flags);
1700 rnp->node_kthread_task = t;
1701 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1703 sp.sched_priority = 99;
1704 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1706 return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index);
1710 * Spawn all kthreads -- called as soon as the scheduler is running.
1712 static int __init rcu_spawn_kthreads(void)
1715 struct rcu_node *rnp;
1717 rcu_kthreads_spawnable = 1;
1718 for_each_possible_cpu(cpu) {
1719 init_waitqueue_head(&per_cpu(rcu_cpu_wq, cpu));
1720 per_cpu(rcu_cpu_has_work, cpu) = 0;
1721 if (cpu_online(cpu))
1722 (void)rcu_spawn_one_cpu_kthread(cpu);
1724 rnp = rcu_get_root(rcu_state);
1725 init_waitqueue_head(&rnp->node_wq);
1726 rcu_init_boost_waitqueue(rnp);
1727 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1728 if (NUM_RCU_NODES > 1)
1729 rcu_for_each_leaf_node(rcu_state, rnp) {
1730 init_waitqueue_head(&rnp->node_wq);
1731 rcu_init_boost_waitqueue(rnp);
1732 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
1736 early_initcall(rcu_spawn_kthreads);
1739 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1740 struct rcu_state *rsp)
1742 unsigned long flags;
1743 struct rcu_data *rdp;
1745 debug_rcu_head_queue(head);
1749 smp_mb(); /* Ensure RCU update seen before callback registry. */
1752 * Opportunistically note grace-period endings and beginnings.
1753 * Note that we might see a beginning right after we see an
1754 * end, but never vice versa, since this CPU has to pass through
1755 * a quiescent state betweentimes.
1757 local_irq_save(flags);
1758 rdp = this_cpu_ptr(rsp->rda);
1760 /* Add the callback to our list. */
1761 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1762 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1765 * Force the grace period if too many callbacks or too long waiting.
1766 * Enforce hysteresis, and don't invoke force_quiescent_state()
1767 * if some other CPU has recently done so. Also, don't bother
1768 * invoking force_quiescent_state() if the newly enqueued callback
1769 * is the only one waiting for a grace period to complete.
1771 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1773 /* Are we ignoring a completed grace period? */
1774 rcu_process_gp_end(rsp, rdp);
1775 check_for_new_grace_period(rsp, rdp);
1777 /* Start a new grace period if one not already started. */
1778 if (!rcu_gp_in_progress(rsp)) {
1779 unsigned long nestflag;
1780 struct rcu_node *rnp_root = rcu_get_root(rsp);
1782 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1783 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1785 /* Give the grace period a kick. */
1786 rdp->blimit = LONG_MAX;
1787 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1788 *rdp->nxttail[RCU_DONE_TAIL] != head)
1789 force_quiescent_state(rsp, 0);
1790 rdp->n_force_qs_snap = rsp->n_force_qs;
1791 rdp->qlen_last_fqs_check = rdp->qlen;
1793 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1794 force_quiescent_state(rsp, 1);
1795 local_irq_restore(flags);
1799 * Queue an RCU-sched callback for invocation after a grace period.
1801 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1803 __call_rcu(head, func, &rcu_sched_state);
1805 EXPORT_SYMBOL_GPL(call_rcu_sched);
1808 * Queue an RCU for invocation after a quicker grace period.
1810 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1812 __call_rcu(head, func, &rcu_bh_state);
1814 EXPORT_SYMBOL_GPL(call_rcu_bh);
1817 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1819 * Control will return to the caller some time after a full rcu-sched
1820 * grace period has elapsed, in other words after all currently executing
1821 * rcu-sched read-side critical sections have completed. These read-side
1822 * critical sections are delimited by rcu_read_lock_sched() and
1823 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1824 * local_irq_disable(), and so on may be used in place of
1825 * rcu_read_lock_sched().
1827 * This means that all preempt_disable code sequences, including NMI and
1828 * hardware-interrupt handlers, in progress on entry will have completed
1829 * before this primitive returns. However, this does not guarantee that
1830 * softirq handlers will have completed, since in some kernels, these
1831 * handlers can run in process context, and can block.
1833 * This primitive provides the guarantees made by the (now removed)
1834 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1835 * guarantees that rcu_read_lock() sections will have completed.
1836 * In "classic RCU", these two guarantees happen to be one and
1837 * the same, but can differ in realtime RCU implementations.
1839 void synchronize_sched(void)
1841 struct rcu_synchronize rcu;
1843 if (rcu_blocking_is_gp())
1846 init_rcu_head_on_stack(&rcu.head);
1847 init_completion(&rcu.completion);
1848 /* Will wake me after RCU finished. */
1849 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1851 wait_for_completion(&rcu.completion);
1852 destroy_rcu_head_on_stack(&rcu.head);
1854 EXPORT_SYMBOL_GPL(synchronize_sched);
1857 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1859 * Control will return to the caller some time after a full rcu_bh grace
1860 * period has elapsed, in other words after all currently executing rcu_bh
1861 * read-side critical sections have completed. RCU read-side critical
1862 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1863 * and may be nested.
1865 void synchronize_rcu_bh(void)
1867 struct rcu_synchronize rcu;
1869 if (rcu_blocking_is_gp())
1872 init_rcu_head_on_stack(&rcu.head);
1873 init_completion(&rcu.completion);
1874 /* Will wake me after RCU finished. */
1875 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1877 wait_for_completion(&rcu.completion);
1878 destroy_rcu_head_on_stack(&rcu.head);
1880 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1883 * Check to see if there is any immediate RCU-related work to be done
1884 * by the current CPU, for the specified type of RCU, returning 1 if so.
1885 * The checks are in order of increasing expense: checks that can be
1886 * carried out against CPU-local state are performed first. However,
1887 * we must check for CPU stalls first, else we might not get a chance.
1889 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1891 struct rcu_node *rnp = rdp->mynode;
1893 rdp->n_rcu_pending++;
1895 /* Check for CPU stalls, if enabled. */
1896 check_cpu_stall(rsp, rdp);
1898 /* Is the RCU core waiting for a quiescent state from this CPU? */
1899 if (rdp->qs_pending && !rdp->passed_quiesc) {
1902 * If force_quiescent_state() coming soon and this CPU
1903 * needs a quiescent state, and this is either RCU-sched
1904 * or RCU-bh, force a local reschedule.
1906 rdp->n_rp_qs_pending++;
1907 if (!rdp->preemptable &&
1908 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1911 } else if (rdp->qs_pending && rdp->passed_quiesc) {
1912 rdp->n_rp_report_qs++;
1916 /* Does this CPU have callbacks ready to invoke? */
1917 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1918 rdp->n_rp_cb_ready++;
1922 /* Has RCU gone idle with this CPU needing another grace period? */
1923 if (cpu_needs_another_gp(rsp, rdp)) {
1924 rdp->n_rp_cpu_needs_gp++;
1928 /* Has another RCU grace period completed? */
1929 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1930 rdp->n_rp_gp_completed++;
1934 /* Has a new RCU grace period started? */
1935 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1936 rdp->n_rp_gp_started++;
1940 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1941 if (rcu_gp_in_progress(rsp) &&
1942 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1943 rdp->n_rp_need_fqs++;
1948 rdp->n_rp_need_nothing++;
1953 * Check to see if there is any immediate RCU-related work to be done
1954 * by the current CPU, returning 1 if so. This function is part of the
1955 * RCU implementation; it is -not- an exported member of the RCU API.
1957 static int rcu_pending(int cpu)
1959 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1960 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1961 rcu_preempt_pending(cpu);
1965 * Check to see if any future RCU-related work will need to be done
1966 * by the current CPU, even if none need be done immediately, returning
1969 static int rcu_needs_cpu_quick_check(int cpu)
1971 /* RCU callbacks either ready or pending? */
1972 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1973 per_cpu(rcu_bh_data, cpu).nxtlist ||
1974 rcu_preempt_needs_cpu(cpu);
1977 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1978 static atomic_t rcu_barrier_cpu_count;
1979 static DEFINE_MUTEX(rcu_barrier_mutex);
1980 static struct completion rcu_barrier_completion;
1982 static void rcu_barrier_callback(struct rcu_head *notused)
1984 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1985 complete(&rcu_barrier_completion);
1989 * Called with preemption disabled, and from cross-cpu IRQ context.
1991 static void rcu_barrier_func(void *type)
1993 int cpu = smp_processor_id();
1994 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1995 void (*call_rcu_func)(struct rcu_head *head,
1996 void (*func)(struct rcu_head *head));
1998 atomic_inc(&rcu_barrier_cpu_count);
1999 call_rcu_func = type;
2000 call_rcu_func(head, rcu_barrier_callback);
2004 * Orchestrate the specified type of RCU barrier, waiting for all
2005 * RCU callbacks of the specified type to complete.
2007 static void _rcu_barrier(struct rcu_state *rsp,
2008 void (*call_rcu_func)(struct rcu_head *head,
2009 void (*func)(struct rcu_head *head)))
2011 BUG_ON(in_interrupt());
2012 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2013 mutex_lock(&rcu_barrier_mutex);
2014 init_completion(&rcu_barrier_completion);
2016 * Initialize rcu_barrier_cpu_count to 1, then invoke
2017 * rcu_barrier_func() on each CPU, so that each CPU also has
2018 * incremented rcu_barrier_cpu_count. Only then is it safe to
2019 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2020 * might complete its grace period before all of the other CPUs
2021 * did their increment, causing this function to return too
2022 * early. Note that on_each_cpu() disables irqs, which prevents
2023 * any CPUs from coming online or going offline until each online
2024 * CPU has queued its RCU-barrier callback.
2026 atomic_set(&rcu_barrier_cpu_count, 1);
2027 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
2028 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2029 complete(&rcu_barrier_completion);
2030 wait_for_completion(&rcu_barrier_completion);
2031 mutex_unlock(&rcu_barrier_mutex);
2035 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2037 void rcu_barrier_bh(void)
2039 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
2041 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2044 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2046 void rcu_barrier_sched(void)
2048 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
2050 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2053 * Do boot-time initialization of a CPU's per-CPU RCU data.
2056 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
2058 unsigned long flags;
2060 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2061 struct rcu_node *rnp = rcu_get_root(rsp);
2063 /* Set up local state, ensuring consistent view of global state. */
2064 raw_spin_lock_irqsave(&rnp->lock, flags);
2065 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2066 rdp->nxtlist = NULL;
2067 for (i = 0; i < RCU_NEXT_SIZE; i++)
2068 rdp->nxttail[i] = &rdp->nxtlist;
2071 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
2072 #endif /* #ifdef CONFIG_NO_HZ */
2074 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2078 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2079 * offline event can be happening at a given time. Note also that we
2080 * can accept some slop in the rsp->completed access due to the fact
2081 * that this CPU cannot possibly have any RCU callbacks in flight yet.
2083 static void __cpuinit
2084 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
2086 unsigned long flags;
2088 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2089 struct rcu_node *rnp = rcu_get_root(rsp);
2091 /* Set up local state, ensuring consistent view of global state. */
2092 raw_spin_lock_irqsave(&rnp->lock, flags);
2093 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
2094 rdp->qs_pending = 1; /* so set up to respond to current GP. */
2095 rdp->beenonline = 1; /* We have now been online. */
2096 rdp->preemptable = preemptable;
2097 rdp->qlen_last_fqs_check = 0;
2098 rdp->n_force_qs_snap = rsp->n_force_qs;
2099 rdp->blimit = blimit;
2100 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2103 * A new grace period might start here. If so, we won't be part
2104 * of it, but that is OK, as we are currently in a quiescent state.
2107 /* Exclude any attempts to start a new GP on large systems. */
2108 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
2110 /* Add CPU to rcu_node bitmasks. */
2112 mask = rdp->grpmask;
2114 /* Exclude any attempts to start a new GP on small systems. */
2115 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2116 rnp->qsmaskinit |= mask;
2117 mask = rnp->grpmask;
2118 if (rnp == rdp->mynode) {
2119 rdp->gpnum = rnp->completed; /* if GP in progress... */
2120 rdp->completed = rnp->completed;
2121 rdp->passed_quiesc_completed = rnp->completed - 1;
2123 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
2125 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2127 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
2130 static void __cpuinit rcu_online_cpu(int cpu)
2132 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2133 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2134 rcu_preempt_init_percpu_data(cpu);
2137 static void __cpuinit rcu_online_kthreads(int cpu)
2139 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2140 struct rcu_node *rnp = rdp->mynode;
2142 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
2143 if (rcu_kthreads_spawnable) {
2144 (void)rcu_spawn_one_cpu_kthread(cpu);
2145 if (rnp->node_kthread_task == NULL)
2146 (void)rcu_spawn_one_node_kthread(rcu_state, rnp);
2151 * Handle CPU online/offline notification events.
2153 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2154 unsigned long action, void *hcpu)
2156 long cpu = (long)hcpu;
2157 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
2158 struct rcu_node *rnp = rdp->mynode;
2161 case CPU_UP_PREPARE:
2162 case CPU_UP_PREPARE_FROZEN:
2163 rcu_online_cpu(cpu);
2164 rcu_online_kthreads(cpu);
2167 case CPU_DOWN_FAILED:
2168 rcu_node_kthread_setaffinity(rnp, -1);
2170 case CPU_DOWN_PREPARE:
2171 rcu_node_kthread_setaffinity(rnp, cpu);
2174 case CPU_DYING_FROZEN:
2176 * The whole machine is "stopped" except this CPU, so we can
2177 * touch any data without introducing corruption. We send the
2178 * dying CPU's callbacks to an arbitrarily chosen online CPU.
2180 rcu_send_cbs_to_online(&rcu_bh_state);
2181 rcu_send_cbs_to_online(&rcu_sched_state);
2182 rcu_preempt_send_cbs_to_online();
2185 case CPU_DEAD_FROZEN:
2186 case CPU_UP_CANCELED:
2187 case CPU_UP_CANCELED_FROZEN:
2188 rcu_offline_cpu(cpu);
2197 * This function is invoked towards the end of the scheduler's initialization
2198 * process. Before this is called, the idle task might contain
2199 * RCU read-side critical sections (during which time, this idle
2200 * task is booting the system). After this function is called, the
2201 * idle tasks are prohibited from containing RCU read-side critical
2202 * sections. This function also enables RCU lockdep checking.
2204 void rcu_scheduler_starting(void)
2206 WARN_ON(num_online_cpus() != 1);
2207 WARN_ON(nr_context_switches() > 0);
2208 rcu_scheduler_active = 1;
2212 * Compute the per-level fanout, either using the exact fanout specified
2213 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2215 #ifdef CONFIG_RCU_FANOUT_EXACT
2216 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2220 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
2221 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
2222 rsp->levelspread[0] = RCU_FANOUT_LEAF;
2224 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2225 static void __init rcu_init_levelspread(struct rcu_state *rsp)
2232 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2233 ccur = rsp->levelcnt[i];
2234 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2238 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2241 * Helper function for rcu_init() that initializes one rcu_state structure.
2243 static void __init rcu_init_one(struct rcu_state *rsp,
2244 struct rcu_data __percpu *rda)
2246 static char *buf[] = { "rcu_node_level_0",
2249 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2253 struct rcu_node *rnp;
2255 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2257 /* Initialize the level-tracking arrays. */
2259 for (i = 1; i < NUM_RCU_LVLS; i++)
2260 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2261 rcu_init_levelspread(rsp);
2263 /* Initialize the elements themselves, starting from the leaves. */
2265 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2266 cpustride *= rsp->levelspread[i];
2267 rnp = rsp->level[i];
2268 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2269 raw_spin_lock_init(&rnp->lock);
2270 lockdep_set_class_and_name(&rnp->lock,
2271 &rcu_node_class[i], buf[i]);
2274 rnp->qsmaskinit = 0;
2275 rnp->grplo = j * cpustride;
2276 rnp->grphi = (j + 1) * cpustride - 1;
2277 if (rnp->grphi >= NR_CPUS)
2278 rnp->grphi = NR_CPUS - 1;
2284 rnp->grpnum = j % rsp->levelspread[i - 1];
2285 rnp->grpmask = 1UL << rnp->grpnum;
2286 rnp->parent = rsp->level[i - 1] +
2287 j / rsp->levelspread[i - 1];
2290 INIT_LIST_HEAD(&rnp->blkd_tasks);
2295 rnp = rsp->level[NUM_RCU_LVLS - 1];
2296 for_each_possible_cpu(i) {
2297 while (i > rnp->grphi)
2299 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2300 rcu_boot_init_percpu_data(i, rsp);
2304 void __init rcu_init(void)
2308 rcu_bootup_announce();
2309 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2310 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2311 __rcu_init_preempt();
2314 * We don't need protection against CPU-hotplug here because
2315 * this is called early in boot, before either interrupts
2316 * or the scheduler are operational.
2318 cpu_notifier(rcu_cpu_notify, 0);
2319 for_each_online_cpu(cpu)
2320 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2321 check_cpu_stall_init();
2324 #include "rcutree_plugin.h"