2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
52 /* Data structures. */
54 static struct lock_class_key rcu_root_class;
56 #define RCU_STATE_INITIALIZER(name) { \
57 .level = { &name.node[0] }, \
59 NUM_RCU_LVL_0, /* root of hierarchy. */ \
62 NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \
64 .signaled = RCU_GP_IDLE, \
67 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
68 .orphan_cbs_list = NULL, \
69 .orphan_cbs_tail = &name.orphan_cbs_list, \
71 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
73 .n_force_qs_ngp = 0, \
76 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
77 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
79 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
80 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
84 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
85 * permit this function to be invoked without holding the root rcu_node
86 * structure's ->lock, but of course results can be subject to change.
88 static int rcu_gp_in_progress(struct rcu_state *rsp)
90 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
94 * Note a quiescent state. Because we do not need to know
95 * how many quiescent states passed, just if there was at least
96 * one since the start of the grace period, this just sets a flag.
98 void rcu_sched_qs(int cpu)
100 struct rcu_data *rdp;
102 rdp = &per_cpu(rcu_sched_data, cpu);
103 rdp->passed_quiesc_completed = rdp->gpnum - 1;
105 rdp->passed_quiesc = 1;
106 rcu_preempt_note_context_switch(cpu);
109 void rcu_bh_qs(int cpu)
111 struct rcu_data *rdp;
113 rdp = &per_cpu(rcu_bh_data, cpu);
114 rdp->passed_quiesc_completed = rdp->gpnum - 1;
116 rdp->passed_quiesc = 1;
120 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
121 .dynticks_nesting = 1,
124 #endif /* #ifdef CONFIG_NO_HZ */
126 static int blimit = 10; /* Maximum callbacks per softirq. */
127 static int qhimark = 10000; /* If this many pending, ignore blimit. */
128 static int qlowmark = 100; /* Once only this many pending, use blimit. */
130 module_param(blimit, int, 0);
131 module_param(qhimark, int, 0);
132 module_param(qlowmark, int, 0);
134 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
135 static int rcu_pending(int cpu);
138 * Return the number of RCU-sched batches processed thus far for debug & stats.
140 long rcu_batches_completed_sched(void)
142 return rcu_sched_state.completed;
144 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
147 * Return the number of RCU BH batches processed thus far for debug & stats.
149 long rcu_batches_completed_bh(void)
151 return rcu_bh_state.completed;
153 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
156 * Does the CPU have callbacks ready to be invoked?
159 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
161 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
165 * Does the current CPU require a yet-as-unscheduled grace period?
168 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
170 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
174 * Return the root node of the specified rcu_state structure.
176 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
178 return &rsp->node[0];
184 * If the specified CPU is offline, tell the caller that it is in
185 * a quiescent state. Otherwise, whack it with a reschedule IPI.
186 * Grace periods can end up waiting on an offline CPU when that
187 * CPU is in the process of coming online -- it will be added to the
188 * rcu_node bitmasks before it actually makes it online. The same thing
189 * can happen while a CPU is in the process of coming online. Because this
190 * race is quite rare, we check for it after detecting that the grace
191 * period has been delayed rather than checking each and every CPU
192 * each and every time we start a new grace period.
194 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
197 * If the CPU is offline, it is in a quiescent state. We can
198 * trust its state not to change because interrupts are disabled.
200 if (cpu_is_offline(rdp->cpu)) {
205 /* If preemptable RCU, no point in sending reschedule IPI. */
206 if (rdp->preemptable)
209 /* The CPU is online, so send it a reschedule IPI. */
210 if (rdp->cpu != smp_processor_id())
211 smp_send_reschedule(rdp->cpu);
218 #endif /* #ifdef CONFIG_SMP */
223 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
225 * Enter nohz mode, in other words, -leave- the mode in which RCU
226 * read-side critical sections can occur. (Though RCU read-side
227 * critical sections can occur in irq handlers in nohz mode, a possibility
228 * handled by rcu_irq_enter() and rcu_irq_exit()).
230 void rcu_enter_nohz(void)
233 struct rcu_dynticks *rdtp;
235 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
236 local_irq_save(flags);
237 rdtp = &__get_cpu_var(rcu_dynticks);
239 rdtp->dynticks_nesting--;
240 WARN_ON_ONCE(rdtp->dynticks & 0x1);
241 local_irq_restore(flags);
245 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
247 * Exit nohz mode, in other words, -enter- the mode in which RCU
248 * read-side critical sections normally occur.
250 void rcu_exit_nohz(void)
253 struct rcu_dynticks *rdtp;
255 local_irq_save(flags);
256 rdtp = &__get_cpu_var(rcu_dynticks);
258 rdtp->dynticks_nesting++;
259 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
260 local_irq_restore(flags);
261 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
265 * rcu_nmi_enter - inform RCU of entry to NMI context
267 * If the CPU was idle with dynamic ticks active, and there is no
268 * irq handler running, this updates rdtp->dynticks_nmi to let the
269 * RCU grace-period handling know that the CPU is active.
271 void rcu_nmi_enter(void)
273 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
275 if (rdtp->dynticks & 0x1)
277 rdtp->dynticks_nmi++;
278 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
279 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
283 * rcu_nmi_exit - inform RCU of exit from NMI context
285 * If the CPU was idle with dynamic ticks active, and there is no
286 * irq handler running, this updates rdtp->dynticks_nmi to let the
287 * RCU grace-period handling know that the CPU is no longer active.
289 void rcu_nmi_exit(void)
291 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
293 if (rdtp->dynticks & 0x1)
295 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
296 rdtp->dynticks_nmi++;
297 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
301 * rcu_irq_enter - inform RCU of entry to hard irq context
303 * If the CPU was idle with dynamic ticks active, this updates the
304 * rdtp->dynticks to let the RCU handling know that the CPU is active.
306 void rcu_irq_enter(void)
308 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
310 if (rdtp->dynticks_nesting++)
313 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
314 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
318 * rcu_irq_exit - inform RCU of exit from hard irq context
320 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
321 * to put let the RCU handling be aware that the CPU is going back to idle
324 void rcu_irq_exit(void)
326 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
328 if (--rdtp->dynticks_nesting)
330 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
332 WARN_ON_ONCE(rdtp->dynticks & 0x1);
334 /* If the interrupt queued a callback, get out of dyntick mode. */
335 if (__get_cpu_var(rcu_sched_data).nxtlist ||
336 __get_cpu_var(rcu_bh_data).nxtlist)
343 * Snapshot the specified CPU's dynticks counter so that we can later
344 * credit them with an implicit quiescent state. Return 1 if this CPU
345 * is in dynticks idle mode, which is an extended quiescent state.
347 static int dyntick_save_progress_counter(struct rcu_data *rdp)
353 snap = rdp->dynticks->dynticks;
354 snap_nmi = rdp->dynticks->dynticks_nmi;
355 smp_mb(); /* Order sampling of snap with end of grace period. */
356 rdp->dynticks_snap = snap;
357 rdp->dynticks_nmi_snap = snap_nmi;
358 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
365 * Return true if the specified CPU has passed through a quiescent
366 * state by virtue of being in or having passed through an dynticks
367 * idle state since the last call to dyntick_save_progress_counter()
370 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
377 curr = rdp->dynticks->dynticks;
378 snap = rdp->dynticks_snap;
379 curr_nmi = rdp->dynticks->dynticks_nmi;
380 snap_nmi = rdp->dynticks_nmi_snap;
381 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
384 * If the CPU passed through or entered a dynticks idle phase with
385 * no active irq/NMI handlers, then we can safely pretend that the CPU
386 * already acknowledged the request to pass through a quiescent
387 * state. Either way, that CPU cannot possibly be in an RCU
388 * read-side critical section that started before the beginning
389 * of the current RCU grace period.
391 if ((curr != snap || (curr & 0x1) == 0) &&
392 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
397 /* Go check for the CPU being offline. */
398 return rcu_implicit_offline_qs(rdp);
401 #endif /* #ifdef CONFIG_SMP */
403 #else /* #ifdef CONFIG_NO_HZ */
407 static int dyntick_save_progress_counter(struct rcu_data *rdp)
412 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
414 return rcu_implicit_offline_qs(rdp);
417 #endif /* #ifdef CONFIG_SMP */
419 #endif /* #else #ifdef CONFIG_NO_HZ */
421 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
423 static void record_gp_stall_check_time(struct rcu_state *rsp)
425 rsp->gp_start = jiffies;
426 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
429 static void print_other_cpu_stall(struct rcu_state *rsp)
434 struct rcu_node *rnp = rcu_get_root(rsp);
436 /* Only let one CPU complain about others per time interval. */
438 spin_lock_irqsave(&rnp->lock, flags);
439 delta = jiffies - rsp->jiffies_stall;
440 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
441 spin_unlock_irqrestore(&rnp->lock, flags);
444 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
447 * Now rat on any tasks that got kicked up to the root rcu_node
448 * due to CPU offlining.
450 rcu_print_task_stall(rnp);
451 spin_unlock_irqrestore(&rnp->lock, flags);
453 /* OK, time to rat on our buddy... */
455 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
456 rcu_for_each_leaf_node(rsp, rnp) {
457 rcu_print_task_stall(rnp);
458 if (rnp->qsmask == 0)
460 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
461 if (rnp->qsmask & (1UL << cpu))
462 printk(" %d", rnp->grplo + cpu);
464 printk(" (detected by %d, t=%ld jiffies)\n",
465 smp_processor_id(), (long)(jiffies - rsp->gp_start));
466 trigger_all_cpu_backtrace();
468 force_quiescent_state(rsp, 0); /* Kick them all. */
471 static void print_cpu_stall(struct rcu_state *rsp)
474 struct rcu_node *rnp = rcu_get_root(rsp);
476 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
477 smp_processor_id(), jiffies - rsp->gp_start);
478 trigger_all_cpu_backtrace();
480 spin_lock_irqsave(&rnp->lock, flags);
481 if ((long)(jiffies - rsp->jiffies_stall) >= 0)
483 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
484 spin_unlock_irqrestore(&rnp->lock, flags);
486 set_need_resched(); /* kick ourselves to get things going. */
489 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
492 struct rcu_node *rnp;
494 delta = jiffies - rsp->jiffies_stall;
496 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
498 /* We haven't checked in, so go dump stack. */
499 print_cpu_stall(rsp);
501 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
503 /* They had two time units to dump stack, so complain. */
504 print_other_cpu_stall(rsp);
508 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
510 static void record_gp_stall_check_time(struct rcu_state *rsp)
514 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
518 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
521 * Update CPU-local rcu_data state to record the newly noticed grace period.
522 * This is used both when we started the grace period and when we notice
523 * that someone else started the grace period. The caller must hold the
524 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
525 * and must have irqs disabled.
527 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
529 if (rdp->gpnum != rnp->gpnum) {
531 rdp->passed_quiesc = 0;
532 rdp->gpnum = rnp->gpnum;
536 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
539 struct rcu_node *rnp;
541 local_irq_save(flags);
543 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
544 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
545 local_irq_restore(flags);
548 __note_new_gpnum(rsp, rnp, rdp);
549 spin_unlock_irqrestore(&rnp->lock, flags);
553 * Did someone else start a new RCU grace period start since we last
554 * checked? Update local state appropriately if so. Must be called
555 * on the CPU corresponding to rdp.
558 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
563 local_irq_save(flags);
564 if (rdp->gpnum != rsp->gpnum) {
565 note_new_gpnum(rsp, rdp);
568 local_irq_restore(flags);
573 * Advance this CPU's callbacks, but only if the current grace period
574 * has ended. This may be called only from the CPU to whom the rdp
575 * belongs. In addition, the corresponding leaf rcu_node structure's
576 * ->lock must be held by the caller, with irqs disabled.
579 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
581 /* Did another grace period end? */
582 if (rdp->completed != rnp->completed) {
584 /* Advance callbacks. No harm if list empty. */
585 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
586 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
587 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
589 /* Remember that we saw this grace-period completion. */
590 rdp->completed = rnp->completed;
595 * Advance this CPU's callbacks, but only if the current grace period
596 * has ended. This may be called only from the CPU to whom the rdp
600 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
603 struct rcu_node *rnp;
605 local_irq_save(flags);
607 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
608 !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
609 local_irq_restore(flags);
612 __rcu_process_gp_end(rsp, rnp, rdp);
613 spin_unlock_irqrestore(&rnp->lock, flags);
617 * Do per-CPU grace-period initialization for running CPU. The caller
618 * must hold the lock of the leaf rcu_node structure corresponding to
622 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
624 /* Prior grace period ended, so advance callbacks for current CPU. */
625 __rcu_process_gp_end(rsp, rnp, rdp);
628 * Because this CPU just now started the new grace period, we know
629 * that all of its callbacks will be covered by this upcoming grace
630 * period, even the ones that were registered arbitrarily recently.
631 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
633 * Other CPUs cannot be sure exactly when the grace period started.
634 * Therefore, their recently registered callbacks must pass through
635 * an additional RCU_NEXT_READY stage, so that they will be handled
636 * by the next RCU grace period.
638 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
639 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
641 /* Set state so that this CPU will detect the next quiescent state. */
642 __note_new_gpnum(rsp, rnp, rdp);
646 * Start a new RCU grace period if warranted, re-initializing the hierarchy
647 * in preparation for detecting the next grace period. The caller must hold
648 * the root node's ->lock, which is released before return. Hard irqs must
652 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
653 __releases(rcu_get_root(rsp)->lock)
655 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
656 struct rcu_node *rnp = rcu_get_root(rsp);
658 if (!cpu_needs_another_gp(rsp, rdp)) {
659 if (rnp->completed == rsp->completed) {
660 spin_unlock_irqrestore(&rnp->lock, flags);
663 spin_unlock(&rnp->lock); /* irqs remain disabled. */
666 * Propagate new ->completed value to rcu_node structures
667 * so that other CPUs don't have to wait until the start
668 * of the next grace period to process their callbacks.
670 rcu_for_each_node_breadth_first(rsp, rnp) {
671 spin_lock(&rnp->lock); /* irqs already disabled. */
672 rnp->completed = rsp->completed;
673 spin_unlock(&rnp->lock); /* irqs remain disabled. */
675 local_irq_restore(flags);
679 /* Advance to a new grace period and initialize state. */
681 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
682 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
683 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
684 record_gp_stall_check_time(rsp);
686 /* Special-case the common single-level case. */
687 if (NUM_RCU_NODES == 1) {
688 rcu_preempt_check_blocked_tasks(rnp);
689 rnp->qsmask = rnp->qsmaskinit;
690 rnp->gpnum = rsp->gpnum;
691 rnp->completed = rsp->completed;
692 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
693 rcu_start_gp_per_cpu(rsp, rnp, rdp);
694 spin_unlock_irqrestore(&rnp->lock, flags);
698 spin_unlock(&rnp->lock); /* leave irqs disabled. */
701 /* Exclude any concurrent CPU-hotplug operations. */
702 spin_lock(&rsp->onofflock); /* irqs already disabled. */
705 * Set the quiescent-state-needed bits in all the rcu_node
706 * structures for all currently online CPUs in breadth-first
707 * order, starting from the root rcu_node structure. This
708 * operation relies on the layout of the hierarchy within the
709 * rsp->node[] array. Note that other CPUs will access only
710 * the leaves of the hierarchy, which still indicate that no
711 * grace period is in progress, at least until the corresponding
712 * leaf node has been initialized. In addition, we have excluded
713 * CPU-hotplug operations.
715 * Note that the grace period cannot complete until we finish
716 * the initialization process, as there will be at least one
717 * qsmask bit set in the root node until that time, namely the
718 * one corresponding to this CPU, due to the fact that we have
721 rcu_for_each_node_breadth_first(rsp, rnp) {
722 spin_lock(&rnp->lock); /* irqs already disabled. */
723 rcu_preempt_check_blocked_tasks(rnp);
724 rnp->qsmask = rnp->qsmaskinit;
725 rnp->gpnum = rsp->gpnum;
726 rnp->completed = rsp->completed;
727 if (rnp == rdp->mynode)
728 rcu_start_gp_per_cpu(rsp, rnp, rdp);
729 spin_unlock(&rnp->lock); /* irqs remain disabled. */
732 rnp = rcu_get_root(rsp);
733 spin_lock(&rnp->lock); /* irqs already disabled. */
734 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
735 spin_unlock(&rnp->lock); /* irqs remain disabled. */
736 spin_unlock_irqrestore(&rsp->onofflock, flags);
740 * Clean up after the prior grace period and let rcu_start_gp() start up
741 * the next grace period if one is needed. Note that the caller must
742 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
744 static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
745 __releases(rcu_get_root(rsp)->lock)
747 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
748 rsp->completed = rsp->gpnum;
749 rsp->signaled = RCU_GP_IDLE;
750 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
754 * Similar to cpu_quiet(), for which it is a helper function. Allows
755 * a group of CPUs to be quieted at one go, though all the CPUs in the
756 * group must be represented by the same leaf rcu_node structure.
757 * That structure's lock must be held upon entry, and it is released
761 cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp,
763 __releases(rnp->lock)
765 struct rcu_node *rnp_c;
767 /* Walk up the rcu_node hierarchy. */
769 if (!(rnp->qsmask & mask)) {
771 /* Our bit has already been cleared, so done. */
772 spin_unlock_irqrestore(&rnp->lock, flags);
775 rnp->qsmask &= ~mask;
776 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
778 /* Other bits still set at this level, so done. */
779 spin_unlock_irqrestore(&rnp->lock, flags);
783 if (rnp->parent == NULL) {
785 /* No more levels. Exit loop holding root lock. */
789 spin_unlock_irqrestore(&rnp->lock, flags);
792 spin_lock_irqsave(&rnp->lock, flags);
793 WARN_ON_ONCE(rnp_c->qsmask);
797 * Get here if we are the last CPU to pass through a quiescent
798 * state for this grace period. Invoke cpu_quiet_msk_finish()
799 * to clean up and start the next grace period if one is needed.
801 cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */
805 * Record a quiescent state for the specified CPU, which must either be
806 * the current CPU. The lastcomp argument is used to make sure we are
807 * still in the grace period of interest. We don't want to end the current
808 * grace period based on quiescent states detected in an earlier grace
812 cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
816 struct rcu_node *rnp;
819 spin_lock_irqsave(&rnp->lock, flags);
820 if (lastcomp != ACCESS_ONCE(rsp->completed)) {
823 * Someone beat us to it for this grace period, so leave.
824 * The race with GP start is resolved by the fact that we
825 * hold the leaf rcu_node lock, so that the per-CPU bits
826 * cannot yet be initialized -- so we would simply find our
827 * CPU's bit already cleared in cpu_quiet_msk() if this race
830 rdp->passed_quiesc = 0; /* try again later! */
831 spin_unlock_irqrestore(&rnp->lock, flags);
835 if ((rnp->qsmask & mask) == 0) {
836 spin_unlock_irqrestore(&rnp->lock, flags);
841 * This GP can't end until cpu checks in, so all of our
842 * callbacks can be processed during the next GP.
844 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
846 cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */
851 * Check to see if there is a new grace period of which this CPU
852 * is not yet aware, and if so, set up local rcu_data state for it.
853 * Otherwise, see if this CPU has just passed through its first
854 * quiescent state for this grace period, and record that fact if so.
857 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
859 /* If there is now a new grace period, record and return. */
860 if (check_for_new_grace_period(rsp, rdp))
864 * Does this CPU still need to do its part for current grace period?
865 * If no, return and let the other CPUs do their part as well.
867 if (!rdp->qs_pending)
871 * Was there a quiescent state since the beginning of the grace
872 * period? If no, then exit and wait for the next call.
874 if (!rdp->passed_quiesc)
877 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
878 cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
881 #ifdef CONFIG_HOTPLUG_CPU
884 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
885 * specified flavor of RCU. The callbacks will be adopted by the next
886 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
887 * comes first. Because this is invoked from the CPU_DYING notifier,
888 * irqs are already disabled.
890 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
893 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
895 if (rdp->nxtlist == NULL)
896 return; /* irqs disabled, so comparison is stable. */
897 spin_lock(&rsp->onofflock); /* irqs already disabled. */
898 *rsp->orphan_cbs_tail = rdp->nxtlist;
899 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
901 for (i = 0; i < RCU_NEXT_SIZE; i++)
902 rdp->nxttail[i] = &rdp->nxtlist;
903 rsp->orphan_qlen += rdp->qlen;
905 spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
909 * Adopt previously orphaned RCU callbacks.
911 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
914 struct rcu_data *rdp;
916 spin_lock_irqsave(&rsp->onofflock, flags);
917 rdp = rsp->rda[smp_processor_id()];
918 if (rsp->orphan_cbs_list == NULL) {
919 spin_unlock_irqrestore(&rsp->onofflock, flags);
922 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
923 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
924 rdp->qlen += rsp->orphan_qlen;
925 rsp->orphan_cbs_list = NULL;
926 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
927 rsp->orphan_qlen = 0;
928 spin_unlock_irqrestore(&rsp->onofflock, flags);
932 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
933 * and move all callbacks from the outgoing CPU to the current one.
935 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
940 struct rcu_data *rdp = rsp->rda[cpu];
941 struct rcu_node *rnp;
943 /* Exclude any attempts to start a new grace period. */
944 spin_lock_irqsave(&rsp->onofflock, flags);
946 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
947 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
948 mask = rdp->grpmask; /* rnp->grplo is constant. */
950 spin_lock(&rnp->lock); /* irqs already disabled. */
951 rnp->qsmaskinit &= ~mask;
952 if (rnp->qsmaskinit != 0) {
953 spin_unlock(&rnp->lock); /* irqs remain disabled. */
958 * If there was a task blocking the current grace period,
959 * and if all CPUs have checked in, we need to propagate
960 * the quiescent state up the rcu_node hierarchy. But that
961 * is inconvenient at the moment due to deadlock issues if
962 * this should end the current grace period. So set the
963 * offlined CPU's bit in ->qsmask in order to force the
964 * next force_quiescent_state() invocation to clean up this
965 * mess in a deadlock-free manner.
967 if (rcu_preempt_offline_tasks(rsp, rnp, rdp) && !rnp->qsmask)
971 spin_unlock(&rnp->lock); /* irqs remain disabled. */
973 } while (rnp != NULL);
974 lastcomp = rsp->completed;
976 spin_unlock_irqrestore(&rsp->onofflock, flags);
978 rcu_adopt_orphan_cbs(rsp);
982 * Remove the specified CPU from the RCU hierarchy and move any pending
983 * callbacks that it might have to the current CPU. This code assumes
984 * that at least one CPU in the system will remain running at all times.
985 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
987 static void rcu_offline_cpu(int cpu)
989 __rcu_offline_cpu(cpu, &rcu_sched_state);
990 __rcu_offline_cpu(cpu, &rcu_bh_state);
991 rcu_preempt_offline_cpu(cpu);
994 #else /* #ifdef CONFIG_HOTPLUG_CPU */
996 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1000 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1004 static void rcu_offline_cpu(int cpu)
1008 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1011 * Invoke any RCU callbacks that have made it to the end of their grace
1012 * period. Thottle as specified by rdp->blimit.
1014 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1016 unsigned long flags;
1017 struct rcu_head *next, *list, **tail;
1020 /* If no callbacks are ready, just return.*/
1021 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1025 * Extract the list of ready callbacks, disabling to prevent
1026 * races with call_rcu() from interrupt handlers.
1028 local_irq_save(flags);
1029 list = rdp->nxtlist;
1030 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1031 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1032 tail = rdp->nxttail[RCU_DONE_TAIL];
1033 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1034 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1035 rdp->nxttail[count] = &rdp->nxtlist;
1036 local_irq_restore(flags);
1038 /* Invoke callbacks. */
1045 if (++count >= rdp->blimit)
1049 local_irq_save(flags);
1051 /* Update count, and requeue any remaining callbacks. */
1054 *tail = rdp->nxtlist;
1055 rdp->nxtlist = list;
1056 for (count = 0; count < RCU_NEXT_SIZE; count++)
1057 if (&rdp->nxtlist == rdp->nxttail[count])
1058 rdp->nxttail[count] = tail;
1063 /* Reinstate batch limit if we have worked down the excess. */
1064 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1065 rdp->blimit = blimit;
1067 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1068 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1069 rdp->qlen_last_fqs_check = 0;
1070 rdp->n_force_qs_snap = rsp->n_force_qs;
1071 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1072 rdp->qlen_last_fqs_check = rdp->qlen;
1074 local_irq_restore(flags);
1076 /* Re-raise the RCU softirq if there are callbacks remaining. */
1077 if (cpu_has_callbacks_ready_to_invoke(rdp))
1078 raise_softirq(RCU_SOFTIRQ);
1082 * Check to see if this CPU is in a non-context-switch quiescent state
1083 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1084 * Also schedule the RCU softirq handler.
1086 * This function must be called with hardirqs disabled. It is normally
1087 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1088 * false, there is no point in invoking rcu_check_callbacks().
1090 void rcu_check_callbacks(int cpu, int user)
1092 if (!rcu_pending(cpu))
1093 return; /* if nothing for RCU to do. */
1095 (idle_cpu(cpu) && rcu_scheduler_active &&
1096 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1099 * Get here if this CPU took its interrupt from user
1100 * mode or from the idle loop, and if this is not a
1101 * nested interrupt. In this case, the CPU is in
1102 * a quiescent state, so note it.
1104 * No memory barrier is required here because both
1105 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1106 * variables that other CPUs neither access nor modify,
1107 * at least not while the corresponding CPU is online.
1113 } else if (!in_softirq()) {
1116 * Get here if this CPU did not take its interrupt from
1117 * softirq, in other words, if it is not interrupting
1118 * a rcu_bh read-side critical section. This is an _bh
1119 * critical section, so note it.
1124 rcu_preempt_check_callbacks(cpu);
1125 raise_softirq(RCU_SOFTIRQ);
1131 * Scan the leaf rcu_node structures, processing dyntick state for any that
1132 * have not yet encountered a quiescent state, using the function specified.
1133 * Returns 1 if the current grace period ends while scanning (possibly
1134 * because we made it end).
1136 static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp,
1137 int (*f)(struct rcu_data *))
1141 unsigned long flags;
1143 struct rcu_node *rnp;
1145 rcu_for_each_leaf_node(rsp, rnp) {
1147 spin_lock_irqsave(&rnp->lock, flags);
1148 if (rsp->completed != lastcomp) {
1149 spin_unlock_irqrestore(&rnp->lock, flags);
1152 if (rnp->qsmask == 0) {
1153 spin_unlock_irqrestore(&rnp->lock, flags);
1158 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1159 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1162 if (mask != 0 && rsp->completed == lastcomp) {
1164 /* cpu_quiet_msk() releases rnp->lock. */
1165 cpu_quiet_msk(mask, rsp, rnp, flags);
1168 spin_unlock_irqrestore(&rnp->lock, flags);
1174 * Force quiescent states on reluctant CPUs, and also detect which
1175 * CPUs are in dyntick-idle mode.
1177 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1179 unsigned long flags;
1181 struct rcu_node *rnp = rcu_get_root(rsp);
1185 if (!rcu_gp_in_progress(rsp))
1186 return; /* No grace period in progress, nothing to force. */
1187 if (!spin_trylock_irqsave(&rsp->fqslock, flags)) {
1188 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1189 return; /* Someone else is already on the job. */
1192 (long)(rsp->jiffies_force_qs - jiffies) >= 0)
1193 goto unlock_ret; /* no emergency and done recently. */
1195 spin_lock(&rnp->lock);
1196 lastcomp = rsp->gpnum - 1;
1197 signaled = rsp->signaled;
1198 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1199 if (lastcomp == rsp->gpnum) {
1200 rsp->n_force_qs_ngp++;
1201 spin_unlock(&rnp->lock);
1202 goto unlock_ret; /* no GP in progress, time updated. */
1204 spin_unlock(&rnp->lock);
1209 break; /* grace period idle or initializing, ignore. */
1211 case RCU_SAVE_DYNTICK:
1213 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1214 break; /* So gcc recognizes the dead code. */
1216 /* Record dyntick-idle state. */
1217 if (rcu_process_dyntick(rsp, lastcomp,
1218 dyntick_save_progress_counter))
1220 /* fall into next case. */
1222 case RCU_SAVE_COMPLETED:
1224 /* Update state, record completion counter. */
1226 spin_lock(&rnp->lock);
1227 if (lastcomp == rsp->completed &&
1228 rsp->signaled == signaled) {
1229 rsp->signaled = RCU_FORCE_QS;
1230 rsp->completed_fqs = lastcomp;
1231 forcenow = signaled == RCU_SAVE_COMPLETED;
1233 spin_unlock(&rnp->lock);
1236 /* fall into next case. */
1240 /* Check dyntick-idle state, send IPI to laggarts. */
1241 if (rcu_process_dyntick(rsp, rsp->completed_fqs,
1242 rcu_implicit_dynticks_qs))
1245 /* Leave state in case more forcing is required. */
1250 spin_unlock_irqrestore(&rsp->fqslock, flags);
1253 #else /* #ifdef CONFIG_SMP */
1255 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1260 #endif /* #else #ifdef CONFIG_SMP */
1263 * This does the RCU processing work from softirq context for the
1264 * specified rcu_state and rcu_data structures. This may be called
1265 * only from the CPU to whom the rdp belongs.
1268 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1270 unsigned long flags;
1272 WARN_ON_ONCE(rdp->beenonline == 0);
1275 * If an RCU GP has gone long enough, go check for dyntick
1276 * idle CPUs and, if needed, send resched IPIs.
1278 if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1279 force_quiescent_state(rsp, 1);
1282 * Advance callbacks in response to end of earlier grace
1283 * period that some other CPU ended.
1285 rcu_process_gp_end(rsp, rdp);
1287 /* Update RCU state based on any recent quiescent states. */
1288 rcu_check_quiescent_state(rsp, rdp);
1290 /* Does this CPU require a not-yet-started grace period? */
1291 if (cpu_needs_another_gp(rsp, rdp)) {
1292 spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1293 rcu_start_gp(rsp, flags); /* releases above lock */
1296 /* If there are callbacks ready, invoke them. */
1297 rcu_do_batch(rsp, rdp);
1301 * Do softirq processing for the current CPU.
1303 static void rcu_process_callbacks(struct softirq_action *unused)
1306 * Memory references from any prior RCU read-side critical sections
1307 * executed by the interrupted code must be seen before any RCU
1308 * grace-period manipulations below.
1310 smp_mb(); /* See above block comment. */
1312 __rcu_process_callbacks(&rcu_sched_state,
1313 &__get_cpu_var(rcu_sched_data));
1314 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1315 rcu_preempt_process_callbacks();
1318 * Memory references from any later RCU read-side critical sections
1319 * executed by the interrupted code must be seen after any RCU
1320 * grace-period manipulations above.
1322 smp_mb(); /* See above block comment. */
1326 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1327 struct rcu_state *rsp)
1329 unsigned long flags;
1330 struct rcu_data *rdp;
1335 smp_mb(); /* Ensure RCU update seen before callback registry. */
1338 * Opportunistically note grace-period endings and beginnings.
1339 * Note that we might see a beginning right after we see an
1340 * end, but never vice versa, since this CPU has to pass through
1341 * a quiescent state betweentimes.
1343 local_irq_save(flags);
1344 rdp = rsp->rda[smp_processor_id()];
1345 rcu_process_gp_end(rsp, rdp);
1346 check_for_new_grace_period(rsp, rdp);
1348 /* Add the callback to our list. */
1349 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1350 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1352 /* Start a new grace period if one not already started. */
1353 if (!rcu_gp_in_progress(rsp)) {
1354 unsigned long nestflag;
1355 struct rcu_node *rnp_root = rcu_get_root(rsp);
1357 spin_lock_irqsave(&rnp_root->lock, nestflag);
1358 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1362 * Force the grace period if too many callbacks or too long waiting.
1363 * Enforce hysteresis, and don't invoke force_quiescent_state()
1364 * if some other CPU has recently done so. Also, don't bother
1365 * invoking force_quiescent_state() if the newly enqueued callback
1366 * is the only one waiting for a grace period to complete.
1368 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1369 rdp->blimit = LONG_MAX;
1370 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1371 *rdp->nxttail[RCU_DONE_TAIL] != head)
1372 force_quiescent_state(rsp, 0);
1373 rdp->n_force_qs_snap = rsp->n_force_qs;
1374 rdp->qlen_last_fqs_check = rdp->qlen;
1375 } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)
1376 force_quiescent_state(rsp, 1);
1377 local_irq_restore(flags);
1381 * Queue an RCU-sched callback for invocation after a grace period.
1383 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1385 __call_rcu(head, func, &rcu_sched_state);
1387 EXPORT_SYMBOL_GPL(call_rcu_sched);
1390 * Queue an RCU for invocation after a quicker grace period.
1392 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1394 __call_rcu(head, func, &rcu_bh_state);
1396 EXPORT_SYMBOL_GPL(call_rcu_bh);
1399 * Check to see if there is any immediate RCU-related work to be done
1400 * by the current CPU, for the specified type of RCU, returning 1 if so.
1401 * The checks are in order of increasing expense: checks that can be
1402 * carried out against CPU-local state are performed first. However,
1403 * we must check for CPU stalls first, else we might not get a chance.
1405 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1407 rdp->n_rcu_pending++;
1409 /* Check for CPU stalls, if enabled. */
1410 check_cpu_stall(rsp, rdp);
1412 /* Is the RCU core waiting for a quiescent state from this CPU? */
1413 if (rdp->qs_pending) {
1414 rdp->n_rp_qs_pending++;
1418 /* Does this CPU have callbacks ready to invoke? */
1419 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1420 rdp->n_rp_cb_ready++;
1424 /* Has RCU gone idle with this CPU needing another grace period? */
1425 if (cpu_needs_another_gp(rsp, rdp)) {
1426 rdp->n_rp_cpu_needs_gp++;
1430 /* Has another RCU grace period completed? */
1431 if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */
1432 rdp->n_rp_gp_completed++;
1436 /* Has a new RCU grace period started? */
1437 if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */
1438 rdp->n_rp_gp_started++;
1442 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1443 if (rcu_gp_in_progress(rsp) &&
1444 ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) {
1445 rdp->n_rp_need_fqs++;
1450 rdp->n_rp_need_nothing++;
1455 * Check to see if there is any immediate RCU-related work to be done
1456 * by the current CPU, returning 1 if so. This function is part of the
1457 * RCU implementation; it is -not- an exported member of the RCU API.
1459 static int rcu_pending(int cpu)
1461 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1462 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1463 rcu_preempt_pending(cpu);
1467 * Check to see if any future RCU-related work will need to be done
1468 * by the current CPU, even if none need be done immediately, returning
1469 * 1 if so. This function is part of the RCU implementation; it is -not-
1470 * an exported member of the RCU API.
1472 int rcu_needs_cpu(int cpu)
1474 /* RCU callbacks either ready or pending? */
1475 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1476 per_cpu(rcu_bh_data, cpu).nxtlist ||
1477 rcu_preempt_needs_cpu(cpu);
1480 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1481 static atomic_t rcu_barrier_cpu_count;
1482 static DEFINE_MUTEX(rcu_barrier_mutex);
1483 static struct completion rcu_barrier_completion;
1485 static void rcu_barrier_callback(struct rcu_head *notused)
1487 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1488 complete(&rcu_barrier_completion);
1492 * Called with preemption disabled, and from cross-cpu IRQ context.
1494 static void rcu_barrier_func(void *type)
1496 int cpu = smp_processor_id();
1497 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1498 void (*call_rcu_func)(struct rcu_head *head,
1499 void (*func)(struct rcu_head *head));
1501 atomic_inc(&rcu_barrier_cpu_count);
1502 call_rcu_func = type;
1503 call_rcu_func(head, rcu_barrier_callback);
1507 * Orchestrate the specified type of RCU barrier, waiting for all
1508 * RCU callbacks of the specified type to complete.
1510 static void _rcu_barrier(struct rcu_state *rsp,
1511 void (*call_rcu_func)(struct rcu_head *head,
1512 void (*func)(struct rcu_head *head)))
1514 BUG_ON(in_interrupt());
1515 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1516 mutex_lock(&rcu_barrier_mutex);
1517 init_completion(&rcu_barrier_completion);
1519 * Initialize rcu_barrier_cpu_count to 1, then invoke
1520 * rcu_barrier_func() on each CPU, so that each CPU also has
1521 * incremented rcu_barrier_cpu_count. Only then is it safe to
1522 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1523 * might complete its grace period before all of the other CPUs
1524 * did their increment, causing this function to return too
1527 atomic_set(&rcu_barrier_cpu_count, 1);
1528 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1529 rcu_adopt_orphan_cbs(rsp);
1530 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1531 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1532 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1533 complete(&rcu_barrier_completion);
1534 wait_for_completion(&rcu_barrier_completion);
1535 mutex_unlock(&rcu_barrier_mutex);
1539 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1541 void rcu_barrier_bh(void)
1543 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1545 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1548 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1550 void rcu_barrier_sched(void)
1552 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1554 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1557 * Do boot-time initialization of a CPU's per-CPU RCU data.
1560 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1562 unsigned long flags;
1564 struct rcu_data *rdp = rsp->rda[cpu];
1565 struct rcu_node *rnp = rcu_get_root(rsp);
1567 /* Set up local state, ensuring consistent view of global state. */
1568 spin_lock_irqsave(&rnp->lock, flags);
1569 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1570 rdp->nxtlist = NULL;
1571 for (i = 0; i < RCU_NEXT_SIZE; i++)
1572 rdp->nxttail[i] = &rdp->nxtlist;
1575 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1576 #endif /* #ifdef CONFIG_NO_HZ */
1578 spin_unlock_irqrestore(&rnp->lock, flags);
1582 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1583 * offline event can be happening at a given time. Note also that we
1584 * can accept some slop in the rsp->completed access due to the fact
1585 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1587 static void __cpuinit
1588 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1590 unsigned long flags;
1592 struct rcu_data *rdp = rsp->rda[cpu];
1593 struct rcu_node *rnp = rcu_get_root(rsp);
1595 /* Set up local state, ensuring consistent view of global state. */
1596 spin_lock_irqsave(&rnp->lock, flags);
1597 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1598 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1599 rdp->beenonline = 1; /* We have now been online. */
1600 rdp->preemptable = preemptable;
1601 rdp->qlen_last_fqs_check = 0;
1602 rdp->n_force_qs_snap = rsp->n_force_qs;
1603 rdp->blimit = blimit;
1604 spin_unlock(&rnp->lock); /* irqs remain disabled. */
1607 * A new grace period might start here. If so, we won't be part
1608 * of it, but that is OK, as we are currently in a quiescent state.
1611 /* Exclude any attempts to start a new GP on large systems. */
1612 spin_lock(&rsp->onofflock); /* irqs already disabled. */
1614 /* Add CPU to rcu_node bitmasks. */
1616 mask = rdp->grpmask;
1618 /* Exclude any attempts to start a new GP on small systems. */
1619 spin_lock(&rnp->lock); /* irqs already disabled. */
1620 rnp->qsmaskinit |= mask;
1621 mask = rnp->grpmask;
1622 if (rnp == rdp->mynode) {
1623 rdp->gpnum = rnp->completed; /* if GP in progress... */
1624 rdp->completed = rnp->completed;
1625 rdp->passed_quiesc_completed = rnp->completed - 1;
1627 spin_unlock(&rnp->lock); /* irqs already disabled. */
1629 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1631 spin_unlock_irqrestore(&rsp->onofflock, flags);
1634 static void __cpuinit rcu_online_cpu(int cpu)
1636 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1637 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1638 rcu_preempt_init_percpu_data(cpu);
1642 * Handle CPU online/offline notification events.
1644 int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1645 unsigned long action, void *hcpu)
1647 long cpu = (long)hcpu;
1650 case CPU_UP_PREPARE:
1651 case CPU_UP_PREPARE_FROZEN:
1652 rcu_online_cpu(cpu);
1655 case CPU_DYING_FROZEN:
1657 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1658 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1659 * returns, all online cpus have queued rcu_barrier_func().
1660 * The dying CPU clears its cpu_online_mask bit and
1661 * moves all of its RCU callbacks to ->orphan_cbs_list
1662 * in the context of stop_machine(), so subsequent calls
1663 * to _rcu_barrier() will adopt these callbacks and only
1664 * then queue rcu_barrier_func() on all remaining CPUs.
1666 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1667 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1668 rcu_preempt_send_cbs_to_orphanage();
1671 case CPU_DEAD_FROZEN:
1672 case CPU_UP_CANCELED:
1673 case CPU_UP_CANCELED_FROZEN:
1674 rcu_offline_cpu(cpu);
1683 * Compute the per-level fanout, either using the exact fanout specified
1684 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1686 #ifdef CONFIG_RCU_FANOUT_EXACT
1687 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1691 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1692 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1694 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1695 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1702 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1703 ccur = rsp->levelcnt[i];
1704 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1708 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1711 * Helper function for rcu_init() that initializes one rcu_state structure.
1713 static void __init rcu_init_one(struct rcu_state *rsp)
1718 struct rcu_node *rnp;
1720 /* Initialize the level-tracking arrays. */
1722 for (i = 1; i < NUM_RCU_LVLS; i++)
1723 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1724 rcu_init_levelspread(rsp);
1726 /* Initialize the elements themselves, starting from the leaves. */
1728 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1729 cpustride *= rsp->levelspread[i];
1730 rnp = rsp->level[i];
1731 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1732 spin_lock_init(&rnp->lock);
1735 rnp->qsmaskinit = 0;
1736 rnp->grplo = j * cpustride;
1737 rnp->grphi = (j + 1) * cpustride - 1;
1738 if (rnp->grphi >= NR_CPUS)
1739 rnp->grphi = NR_CPUS - 1;
1745 rnp->grpnum = j % rsp->levelspread[i - 1];
1746 rnp->grpmask = 1UL << rnp->grpnum;
1747 rnp->parent = rsp->level[i - 1] +
1748 j / rsp->levelspread[i - 1];
1751 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1752 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1755 lockdep_set_class(&rcu_get_root(rsp)->lock, &rcu_root_class);
1759 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1760 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1763 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1767 struct rcu_node *rnp; \
1769 rcu_init_one(rsp); \
1770 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1772 for_each_possible_cpu(i) { \
1773 if (i > rnp[j].grphi) \
1775 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1776 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1777 rcu_boot_init_percpu_data(i, rsp); \
1781 void __init __rcu_init(void)
1783 rcu_bootup_announce();
1784 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1785 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1786 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1787 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1788 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1789 __rcu_init_preempt();
1790 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1793 #include "rcutree_plugin.h"