ACPI: thinkpad-acpi: add development version tag
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / rcutree.c
blob3ec8160fc75ffa76121cb5d04c948b0f16f3e317
1 /*
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 -
28 * Documentation/RCU
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>
50 #include "rcutree.h"
52 /* Data structures. */
54 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
56 #define RCU_STATE_INITIALIZER(name) { \
57 .level = { &name.node[0] }, \
58 .levelcnt = { \
59 NUM_RCU_LVL_0, /* root of hierarchy. */ \
60 NUM_RCU_LVL_1, \
61 NUM_RCU_LVL_2, \
62 NUM_RCU_LVL_3, \
63 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
64 }, \
65 .signaled = RCU_GP_IDLE, \
66 .gpnum = -300, \
67 .completed = -300, \
68 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&name.onofflock), \
69 .orphan_cbs_list = NULL, \
70 .orphan_cbs_tail = &name.orphan_cbs_list, \
71 .orphan_qlen = 0, \
72 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&name.fqslock), \
73 .n_force_qs = 0, \
74 .n_force_qs_ngp = 0, \
77 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
78 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
80 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
81 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;
104 barrier();
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;
115 barrier();
116 rdp->passed_quiesc = 1;
119 #ifdef CONFIG_NO_HZ
120 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
121 .dynticks_nesting = 1,
122 .dynticks = 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 * Force a quiescent state for RCU BH.
158 void rcu_bh_force_quiescent_state(void)
160 force_quiescent_state(&rcu_bh_state, 0);
162 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
165 * Force a quiescent state for RCU-sched.
167 void rcu_sched_force_quiescent_state(void)
169 force_quiescent_state(&rcu_sched_state, 0);
171 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
174 * Does the CPU have callbacks ready to be invoked?
176 static int
177 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
179 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
183 * Does the current CPU require a yet-as-unscheduled grace period?
185 static int
186 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
188 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
192 * Return the root node of the specified rcu_state structure.
194 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
196 return &rsp->node[0];
199 #ifdef CONFIG_SMP
202 * If the specified CPU is offline, tell the caller that it is in
203 * a quiescent state. Otherwise, whack it with a reschedule IPI.
204 * Grace periods can end up waiting on an offline CPU when that
205 * CPU is in the process of coming online -- it will be added to the
206 * rcu_node bitmasks before it actually makes it online. The same thing
207 * can happen while a CPU is in the process of coming online. Because this
208 * race is quite rare, we check for it after detecting that the grace
209 * period has been delayed rather than checking each and every CPU
210 * each and every time we start a new grace period.
212 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
215 * If the CPU is offline, it is in a quiescent state. We can
216 * trust its state not to change because interrupts are disabled.
218 if (cpu_is_offline(rdp->cpu)) {
219 rdp->offline_fqs++;
220 return 1;
223 /* If preemptable RCU, no point in sending reschedule IPI. */
224 if (rdp->preemptable)
225 return 0;
227 /* The CPU is online, so send it a reschedule IPI. */
228 if (rdp->cpu != smp_processor_id())
229 smp_send_reschedule(rdp->cpu);
230 else
231 set_need_resched();
232 rdp->resched_ipi++;
233 return 0;
236 #endif /* #ifdef CONFIG_SMP */
238 #ifdef CONFIG_NO_HZ
241 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
243 * Enter nohz mode, in other words, -leave- the mode in which RCU
244 * read-side critical sections can occur. (Though RCU read-side
245 * critical sections can occur in irq handlers in nohz mode, a possibility
246 * handled by rcu_irq_enter() and rcu_irq_exit()).
248 void rcu_enter_nohz(void)
250 unsigned long flags;
251 struct rcu_dynticks *rdtp;
253 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
254 local_irq_save(flags);
255 rdtp = &__get_cpu_var(rcu_dynticks);
256 rdtp->dynticks++;
257 rdtp->dynticks_nesting--;
258 WARN_ON_ONCE(rdtp->dynticks & 0x1);
259 local_irq_restore(flags);
263 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
265 * Exit nohz mode, in other words, -enter- the mode in which RCU
266 * read-side critical sections normally occur.
268 void rcu_exit_nohz(void)
270 unsigned long flags;
271 struct rcu_dynticks *rdtp;
273 local_irq_save(flags);
274 rdtp = &__get_cpu_var(rcu_dynticks);
275 rdtp->dynticks++;
276 rdtp->dynticks_nesting++;
277 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
278 local_irq_restore(flags);
279 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
283 * rcu_nmi_enter - inform RCU of entry to 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 active.
289 void rcu_nmi_enter(void)
291 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
293 if (rdtp->dynticks & 0x1)
294 return;
295 rdtp->dynticks_nmi++;
296 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
297 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
301 * rcu_nmi_exit - inform RCU of exit from NMI context
303 * If the CPU was idle with dynamic ticks active, and there is no
304 * irq handler running, this updates rdtp->dynticks_nmi to let the
305 * RCU grace-period handling know that the CPU is no longer active.
307 void rcu_nmi_exit(void)
309 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
311 if (rdtp->dynticks & 0x1)
312 return;
313 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
314 rdtp->dynticks_nmi++;
315 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
319 * rcu_irq_enter - inform RCU of entry to hard irq context
321 * If the CPU was idle with dynamic ticks active, this updates the
322 * rdtp->dynticks to let the RCU handling know that the CPU is active.
324 void rcu_irq_enter(void)
326 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
328 if (rdtp->dynticks_nesting++)
329 return;
330 rdtp->dynticks++;
331 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
332 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
336 * rcu_irq_exit - inform RCU of exit from hard irq context
338 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
339 * to put let the RCU handling be aware that the CPU is going back to idle
340 * with no ticks.
342 void rcu_irq_exit(void)
344 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
346 if (--rdtp->dynticks_nesting)
347 return;
348 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
349 rdtp->dynticks++;
350 WARN_ON_ONCE(rdtp->dynticks & 0x1);
352 /* If the interrupt queued a callback, get out of dyntick mode. */
353 if (__get_cpu_var(rcu_sched_data).nxtlist ||
354 __get_cpu_var(rcu_bh_data).nxtlist)
355 set_need_resched();
358 #ifdef CONFIG_SMP
361 * Snapshot the specified CPU's dynticks counter so that we can later
362 * credit them with an implicit quiescent state. Return 1 if this CPU
363 * is in dynticks idle mode, which is an extended quiescent state.
365 static int dyntick_save_progress_counter(struct rcu_data *rdp)
367 int ret;
368 int snap;
369 int snap_nmi;
371 snap = rdp->dynticks->dynticks;
372 snap_nmi = rdp->dynticks->dynticks_nmi;
373 smp_mb(); /* Order sampling of snap with end of grace period. */
374 rdp->dynticks_snap = snap;
375 rdp->dynticks_nmi_snap = snap_nmi;
376 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
377 if (ret)
378 rdp->dynticks_fqs++;
379 return ret;
383 * Return true if the specified CPU has passed through a quiescent
384 * state by virtue of being in or having passed through an dynticks
385 * idle state since the last call to dyntick_save_progress_counter()
386 * for this same CPU.
388 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
390 long curr;
391 long curr_nmi;
392 long snap;
393 long snap_nmi;
395 curr = rdp->dynticks->dynticks;
396 snap = rdp->dynticks_snap;
397 curr_nmi = rdp->dynticks->dynticks_nmi;
398 snap_nmi = rdp->dynticks_nmi_snap;
399 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
402 * If the CPU passed through or entered a dynticks idle phase with
403 * no active irq/NMI handlers, then we can safely pretend that the CPU
404 * already acknowledged the request to pass through a quiescent
405 * state. Either way, that CPU cannot possibly be in an RCU
406 * read-side critical section that started before the beginning
407 * of the current RCU grace period.
409 if ((curr != snap || (curr & 0x1) == 0) &&
410 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
411 rdp->dynticks_fqs++;
412 return 1;
415 /* Go check for the CPU being offline. */
416 return rcu_implicit_offline_qs(rdp);
419 #endif /* #ifdef CONFIG_SMP */
421 #else /* #ifdef CONFIG_NO_HZ */
423 #ifdef CONFIG_SMP
425 static int dyntick_save_progress_counter(struct rcu_data *rdp)
427 return 0;
430 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
432 return rcu_implicit_offline_qs(rdp);
435 #endif /* #ifdef CONFIG_SMP */
437 #endif /* #else #ifdef CONFIG_NO_HZ */
439 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
441 static void record_gp_stall_check_time(struct rcu_state *rsp)
443 rsp->gp_start = jiffies;
444 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
447 static void print_other_cpu_stall(struct rcu_state *rsp)
449 int cpu;
450 long delta;
451 unsigned long flags;
452 struct rcu_node *rnp = rcu_get_root(rsp);
454 /* Only let one CPU complain about others per time interval. */
456 raw_spin_lock_irqsave(&rnp->lock, flags);
457 delta = jiffies - rsp->jiffies_stall;
458 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
459 raw_spin_unlock_irqrestore(&rnp->lock, flags);
460 return;
462 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
465 * Now rat on any tasks that got kicked up to the root rcu_node
466 * due to CPU offlining.
468 rcu_print_task_stall(rnp);
469 raw_spin_unlock_irqrestore(&rnp->lock, flags);
471 /* OK, time to rat on our buddy... */
473 printk(KERN_ERR "INFO: RCU detected CPU stalls:");
474 rcu_for_each_leaf_node(rsp, rnp) {
475 raw_spin_lock_irqsave(&rnp->lock, flags);
476 rcu_print_task_stall(rnp);
477 raw_spin_unlock_irqrestore(&rnp->lock, flags);
478 if (rnp->qsmask == 0)
479 continue;
480 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
481 if (rnp->qsmask & (1UL << cpu))
482 printk(" %d", rnp->grplo + cpu);
484 printk(" (detected by %d, t=%ld jiffies)\n",
485 smp_processor_id(), (long)(jiffies - rsp->gp_start));
486 trigger_all_cpu_backtrace();
488 /* If so configured, complain about tasks blocking the grace period. */
490 rcu_print_detail_task_stall(rsp);
492 force_quiescent_state(rsp, 0); /* Kick them all. */
495 static void print_cpu_stall(struct rcu_state *rsp)
497 unsigned long flags;
498 struct rcu_node *rnp = rcu_get_root(rsp);
500 printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
501 smp_processor_id(), jiffies - rsp->gp_start);
502 trigger_all_cpu_backtrace();
504 raw_spin_lock_irqsave(&rnp->lock, flags);
505 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
506 rsp->jiffies_stall =
507 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
508 raw_spin_unlock_irqrestore(&rnp->lock, flags);
510 set_need_resched(); /* kick ourselves to get things going. */
513 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
515 long delta;
516 struct rcu_node *rnp;
518 delta = jiffies - rsp->jiffies_stall;
519 rnp = rdp->mynode;
520 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
522 /* We haven't checked in, so go dump stack. */
523 print_cpu_stall(rsp);
525 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
527 /* They had two time units to dump stack, so complain. */
528 print_other_cpu_stall(rsp);
532 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
534 static void record_gp_stall_check_time(struct rcu_state *rsp)
538 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
542 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
545 * Update CPU-local rcu_data state to record the newly noticed grace period.
546 * This is used both when we started the grace period and when we notice
547 * that someone else started the grace period. The caller must hold the
548 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
549 * and must have irqs disabled.
551 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
553 if (rdp->gpnum != rnp->gpnum) {
554 rdp->qs_pending = 1;
555 rdp->passed_quiesc = 0;
556 rdp->gpnum = rnp->gpnum;
560 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
562 unsigned long flags;
563 struct rcu_node *rnp;
565 local_irq_save(flags);
566 rnp = rdp->mynode;
567 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
568 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
569 local_irq_restore(flags);
570 return;
572 __note_new_gpnum(rsp, rnp, rdp);
573 raw_spin_unlock_irqrestore(&rnp->lock, flags);
577 * Did someone else start a new RCU grace period start since we last
578 * checked? Update local state appropriately if so. Must be called
579 * on the CPU corresponding to rdp.
581 static int
582 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
584 unsigned long flags;
585 int ret = 0;
587 local_irq_save(flags);
588 if (rdp->gpnum != rsp->gpnum) {
589 note_new_gpnum(rsp, rdp);
590 ret = 1;
592 local_irq_restore(flags);
593 return ret;
597 * Advance this CPU's callbacks, but only if the current grace period
598 * has ended. This may be called only from the CPU to whom the rdp
599 * belongs. In addition, the corresponding leaf rcu_node structure's
600 * ->lock must be held by the caller, with irqs disabled.
602 static void
603 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
605 /* Did another grace period end? */
606 if (rdp->completed != rnp->completed) {
608 /* Advance callbacks. No harm if list empty. */
609 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
610 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
611 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
613 /* Remember that we saw this grace-period completion. */
614 rdp->completed = rnp->completed;
619 * Advance this CPU's callbacks, but only if the current grace period
620 * has ended. This may be called only from the CPU to whom the rdp
621 * belongs.
623 static void
624 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
626 unsigned long flags;
627 struct rcu_node *rnp;
629 local_irq_save(flags);
630 rnp = rdp->mynode;
631 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
632 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
633 local_irq_restore(flags);
634 return;
636 __rcu_process_gp_end(rsp, rnp, rdp);
637 raw_spin_unlock_irqrestore(&rnp->lock, flags);
641 * Do per-CPU grace-period initialization for running CPU. The caller
642 * must hold the lock of the leaf rcu_node structure corresponding to
643 * this CPU.
645 static void
646 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
648 /* Prior grace period ended, so advance callbacks for current CPU. */
649 __rcu_process_gp_end(rsp, rnp, rdp);
652 * Because this CPU just now started the new grace period, we know
653 * that all of its callbacks will be covered by this upcoming grace
654 * period, even the ones that were registered arbitrarily recently.
655 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
657 * Other CPUs cannot be sure exactly when the grace period started.
658 * Therefore, their recently registered callbacks must pass through
659 * an additional RCU_NEXT_READY stage, so that they will be handled
660 * by the next RCU grace period.
662 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
663 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
665 /* Set state so that this CPU will detect the next quiescent state. */
666 __note_new_gpnum(rsp, rnp, rdp);
670 * Start a new RCU grace period if warranted, re-initializing the hierarchy
671 * in preparation for detecting the next grace period. The caller must hold
672 * the root node's ->lock, which is released before return. Hard irqs must
673 * be disabled.
675 static void
676 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
677 __releases(rcu_get_root(rsp)->lock)
679 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
680 struct rcu_node *rnp = rcu_get_root(rsp);
682 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
683 if (cpu_needs_another_gp(rsp, rdp))
684 rsp->fqs_need_gp = 1;
685 if (rnp->completed == rsp->completed) {
686 raw_spin_unlock_irqrestore(&rnp->lock, flags);
687 return;
689 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
692 * Propagate new ->completed value to rcu_node structures
693 * so that other CPUs don't have to wait until the start
694 * of the next grace period to process their callbacks.
696 rcu_for_each_node_breadth_first(rsp, rnp) {
697 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
698 rnp->completed = rsp->completed;
699 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
701 local_irq_restore(flags);
702 return;
705 /* Advance to a new grace period and initialize state. */
706 rsp->gpnum++;
707 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
708 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
709 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
710 record_gp_stall_check_time(rsp);
712 /* Special-case the common single-level case. */
713 if (NUM_RCU_NODES == 1) {
714 rcu_preempt_check_blocked_tasks(rnp);
715 rnp->qsmask = rnp->qsmaskinit;
716 rnp->gpnum = rsp->gpnum;
717 rnp->completed = rsp->completed;
718 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
719 rcu_start_gp_per_cpu(rsp, rnp, rdp);
720 raw_spin_unlock_irqrestore(&rnp->lock, flags);
721 return;
724 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
727 /* Exclude any concurrent CPU-hotplug operations. */
728 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
731 * Set the quiescent-state-needed bits in all the rcu_node
732 * structures for all currently online CPUs in breadth-first
733 * order, starting from the root rcu_node structure. This
734 * operation relies on the layout of the hierarchy within the
735 * rsp->node[] array. Note that other CPUs will access only
736 * the leaves of the hierarchy, which still indicate that no
737 * grace period is in progress, at least until the corresponding
738 * leaf node has been initialized. In addition, we have excluded
739 * CPU-hotplug operations.
741 * Note that the grace period cannot complete until we finish
742 * the initialization process, as there will be at least one
743 * qsmask bit set in the root node until that time, namely the
744 * one corresponding to this CPU, due to the fact that we have
745 * irqs disabled.
747 rcu_for_each_node_breadth_first(rsp, rnp) {
748 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
749 rcu_preempt_check_blocked_tasks(rnp);
750 rnp->qsmask = rnp->qsmaskinit;
751 rnp->gpnum = rsp->gpnum;
752 rnp->completed = rsp->completed;
753 if (rnp == rdp->mynode)
754 rcu_start_gp_per_cpu(rsp, rnp, rdp);
755 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
758 rnp = rcu_get_root(rsp);
759 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
760 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
761 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
762 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
766 * Report a full set of quiescent states to the specified rcu_state
767 * data structure. This involves cleaning up after the prior grace
768 * period and letting rcu_start_gp() start up the next grace period
769 * if one is needed. Note that the caller must hold rnp->lock, as
770 * required by rcu_start_gp(), which will release it.
772 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
773 __releases(rcu_get_root(rsp)->lock)
775 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
776 rsp->completed = rsp->gpnum;
777 rsp->signaled = RCU_GP_IDLE;
778 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
782 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
783 * Allows quiescent states for a group of CPUs to be reported at one go
784 * to the specified rcu_node structure, though all the CPUs in the group
785 * must be represented by the same rcu_node structure (which need not be
786 * a leaf rcu_node structure, though it often will be). That structure's
787 * lock must be held upon entry, and it is released before return.
789 static void
790 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
791 struct rcu_node *rnp, unsigned long flags)
792 __releases(rnp->lock)
794 struct rcu_node *rnp_c;
796 /* Walk up the rcu_node hierarchy. */
797 for (;;) {
798 if (!(rnp->qsmask & mask)) {
800 /* Our bit has already been cleared, so done. */
801 raw_spin_unlock_irqrestore(&rnp->lock, flags);
802 return;
804 rnp->qsmask &= ~mask;
805 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
807 /* Other bits still set at this level, so done. */
808 raw_spin_unlock_irqrestore(&rnp->lock, flags);
809 return;
811 mask = rnp->grpmask;
812 if (rnp->parent == NULL) {
814 /* No more levels. Exit loop holding root lock. */
816 break;
818 raw_spin_unlock_irqrestore(&rnp->lock, flags);
819 rnp_c = rnp;
820 rnp = rnp->parent;
821 raw_spin_lock_irqsave(&rnp->lock, flags);
822 WARN_ON_ONCE(rnp_c->qsmask);
826 * Get here if we are the last CPU to pass through a quiescent
827 * state for this grace period. Invoke rcu_report_qs_rsp()
828 * to clean up and start the next grace period if one is needed.
830 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
834 * Record a quiescent state for the specified CPU to that CPU's rcu_data
835 * structure. This must be either called from the specified CPU, or
836 * called when the specified CPU is known to be offline (and when it is
837 * also known that no other CPU is concurrently trying to help the offline
838 * CPU). The lastcomp argument is used to make sure we are still in the
839 * grace period of interest. We don't want to end the current grace period
840 * based on quiescent states detected in an earlier grace period!
842 static void
843 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
845 unsigned long flags;
846 unsigned long mask;
847 struct rcu_node *rnp;
849 rnp = rdp->mynode;
850 raw_spin_lock_irqsave(&rnp->lock, flags);
851 if (lastcomp != rnp->completed) {
854 * Someone beat us to it for this grace period, so leave.
855 * The race with GP start is resolved by the fact that we
856 * hold the leaf rcu_node lock, so that the per-CPU bits
857 * cannot yet be initialized -- so we would simply find our
858 * CPU's bit already cleared in rcu_report_qs_rnp() if this
859 * race occurred.
861 rdp->passed_quiesc = 0; /* try again later! */
862 raw_spin_unlock_irqrestore(&rnp->lock, flags);
863 return;
865 mask = rdp->grpmask;
866 if ((rnp->qsmask & mask) == 0) {
867 raw_spin_unlock_irqrestore(&rnp->lock, flags);
868 } else {
869 rdp->qs_pending = 0;
872 * This GP can't end until cpu checks in, so all of our
873 * callbacks can be processed during the next GP.
875 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
877 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
882 * Check to see if there is a new grace period of which this CPU
883 * is not yet aware, and if so, set up local rcu_data state for it.
884 * Otherwise, see if this CPU has just passed through its first
885 * quiescent state for this grace period, and record that fact if so.
887 static void
888 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
890 /* If there is now a new grace period, record and return. */
891 if (check_for_new_grace_period(rsp, rdp))
892 return;
895 * Does this CPU still need to do its part for current grace period?
896 * If no, return and let the other CPUs do their part as well.
898 if (!rdp->qs_pending)
899 return;
902 * Was there a quiescent state since the beginning of the grace
903 * period? If no, then exit and wait for the next call.
905 if (!rdp->passed_quiesc)
906 return;
909 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
910 * judge of that).
912 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
915 #ifdef CONFIG_HOTPLUG_CPU
918 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
919 * specified flavor of RCU. The callbacks will be adopted by the next
920 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
921 * comes first. Because this is invoked from the CPU_DYING notifier,
922 * irqs are already disabled.
924 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
926 int i;
927 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
929 if (rdp->nxtlist == NULL)
930 return; /* irqs disabled, so comparison is stable. */
931 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
932 *rsp->orphan_cbs_tail = rdp->nxtlist;
933 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
934 rdp->nxtlist = NULL;
935 for (i = 0; i < RCU_NEXT_SIZE; i++)
936 rdp->nxttail[i] = &rdp->nxtlist;
937 rsp->orphan_qlen += rdp->qlen;
938 rdp->qlen = 0;
939 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
943 * Adopt previously orphaned RCU callbacks.
945 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
947 unsigned long flags;
948 struct rcu_data *rdp;
950 raw_spin_lock_irqsave(&rsp->onofflock, flags);
951 rdp = rsp->rda[smp_processor_id()];
952 if (rsp->orphan_cbs_list == NULL) {
953 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
954 return;
956 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
957 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
958 rdp->qlen += rsp->orphan_qlen;
959 rsp->orphan_cbs_list = NULL;
960 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
961 rsp->orphan_qlen = 0;
962 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
966 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
967 * and move all callbacks from the outgoing CPU to the current one.
969 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
971 unsigned long flags;
972 unsigned long mask;
973 int need_report = 0;
974 struct rcu_data *rdp = rsp->rda[cpu];
975 struct rcu_node *rnp;
977 /* Exclude any attempts to start a new grace period. */
978 raw_spin_lock_irqsave(&rsp->onofflock, flags);
980 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
981 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
982 mask = rdp->grpmask; /* rnp->grplo is constant. */
983 do {
984 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
985 rnp->qsmaskinit &= ~mask;
986 if (rnp->qsmaskinit != 0) {
987 if (rnp != rdp->mynode)
988 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
989 break;
991 if (rnp == rdp->mynode)
992 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
993 else
994 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
995 mask = rnp->grpmask;
996 rnp = rnp->parent;
997 } while (rnp != NULL);
1000 * We still hold the leaf rcu_node structure lock here, and
1001 * irqs are still disabled. The reason for this subterfuge is
1002 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1003 * held leads to deadlock.
1005 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1006 rnp = rdp->mynode;
1007 if (need_report & RCU_OFL_TASKS_NORM_GP)
1008 rcu_report_unblock_qs_rnp(rnp, flags);
1009 else
1010 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1011 if (need_report & RCU_OFL_TASKS_EXP_GP)
1012 rcu_report_exp_rnp(rsp, rnp);
1014 rcu_adopt_orphan_cbs(rsp);
1018 * Remove the specified CPU from the RCU hierarchy and move any pending
1019 * callbacks that it might have to the current CPU. This code assumes
1020 * that at least one CPU in the system will remain running at all times.
1021 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1023 static void rcu_offline_cpu(int cpu)
1025 __rcu_offline_cpu(cpu, &rcu_sched_state);
1026 __rcu_offline_cpu(cpu, &rcu_bh_state);
1027 rcu_preempt_offline_cpu(cpu);
1030 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1032 static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1036 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1040 static void rcu_offline_cpu(int cpu)
1044 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1047 * Invoke any RCU callbacks that have made it to the end of their grace
1048 * period. Thottle as specified by rdp->blimit.
1050 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1052 unsigned long flags;
1053 struct rcu_head *next, *list, **tail;
1054 int count;
1056 /* If no callbacks are ready, just return.*/
1057 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1058 return;
1061 * Extract the list of ready callbacks, disabling to prevent
1062 * races with call_rcu() from interrupt handlers.
1064 local_irq_save(flags);
1065 list = rdp->nxtlist;
1066 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1067 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1068 tail = rdp->nxttail[RCU_DONE_TAIL];
1069 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1070 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1071 rdp->nxttail[count] = &rdp->nxtlist;
1072 local_irq_restore(flags);
1074 /* Invoke callbacks. */
1075 count = 0;
1076 while (list) {
1077 next = list->next;
1078 prefetch(next);
1079 list->func(list);
1080 list = next;
1081 if (++count >= rdp->blimit)
1082 break;
1085 local_irq_save(flags);
1087 /* Update count, and requeue any remaining callbacks. */
1088 rdp->qlen -= count;
1089 if (list != NULL) {
1090 *tail = rdp->nxtlist;
1091 rdp->nxtlist = list;
1092 for (count = 0; count < RCU_NEXT_SIZE; count++)
1093 if (&rdp->nxtlist == rdp->nxttail[count])
1094 rdp->nxttail[count] = tail;
1095 else
1096 break;
1099 /* Reinstate batch limit if we have worked down the excess. */
1100 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1101 rdp->blimit = blimit;
1103 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1104 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1105 rdp->qlen_last_fqs_check = 0;
1106 rdp->n_force_qs_snap = rsp->n_force_qs;
1107 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1108 rdp->qlen_last_fqs_check = rdp->qlen;
1110 local_irq_restore(flags);
1112 /* Re-raise the RCU softirq if there are callbacks remaining. */
1113 if (cpu_has_callbacks_ready_to_invoke(rdp))
1114 raise_softirq(RCU_SOFTIRQ);
1118 * Check to see if this CPU is in a non-context-switch quiescent state
1119 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1120 * Also schedule the RCU softirq handler.
1122 * This function must be called with hardirqs disabled. It is normally
1123 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1124 * false, there is no point in invoking rcu_check_callbacks().
1126 void rcu_check_callbacks(int cpu, int user)
1128 if (!rcu_pending(cpu))
1129 return; /* if nothing for RCU to do. */
1130 if (user ||
1131 (idle_cpu(cpu) && rcu_scheduler_active &&
1132 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1135 * Get here if this CPU took its interrupt from user
1136 * mode or from the idle loop, and if this is not a
1137 * nested interrupt. In this case, the CPU is in
1138 * a quiescent state, so note it.
1140 * No memory barrier is required here because both
1141 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1142 * variables that other CPUs neither access nor modify,
1143 * at least not while the corresponding CPU is online.
1146 rcu_sched_qs(cpu);
1147 rcu_bh_qs(cpu);
1149 } else if (!in_softirq()) {
1152 * Get here if this CPU did not take its interrupt from
1153 * softirq, in other words, if it is not interrupting
1154 * a rcu_bh read-side critical section. This is an _bh
1155 * critical section, so note it.
1158 rcu_bh_qs(cpu);
1160 rcu_preempt_check_callbacks(cpu);
1161 raise_softirq(RCU_SOFTIRQ);
1164 #ifdef CONFIG_SMP
1167 * Scan the leaf rcu_node structures, processing dyntick state for any that
1168 * have not yet encountered a quiescent state, using the function specified.
1169 * The caller must have suppressed start of new grace periods.
1171 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1173 unsigned long bit;
1174 int cpu;
1175 unsigned long flags;
1176 unsigned long mask;
1177 struct rcu_node *rnp;
1179 rcu_for_each_leaf_node(rsp, rnp) {
1180 mask = 0;
1181 raw_spin_lock_irqsave(&rnp->lock, flags);
1182 if (!rcu_gp_in_progress(rsp)) {
1183 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1184 return;
1186 if (rnp->qsmask == 0) {
1187 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1188 continue;
1190 cpu = rnp->grplo;
1191 bit = 1;
1192 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1193 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
1194 mask |= bit;
1196 if (mask != 0) {
1198 /* rcu_report_qs_rnp() releases rnp->lock. */
1199 rcu_report_qs_rnp(mask, rsp, rnp, flags);
1200 continue;
1202 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1207 * Force quiescent states on reluctant CPUs, and also detect which
1208 * CPUs are in dyntick-idle mode.
1210 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1212 unsigned long flags;
1213 struct rcu_node *rnp = rcu_get_root(rsp);
1215 if (!rcu_gp_in_progress(rsp))
1216 return; /* No grace period in progress, nothing to force. */
1217 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1218 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1219 return; /* Someone else is already on the job. */
1221 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1222 goto unlock_fqs_ret; /* no emergency and done recently. */
1223 rsp->n_force_qs++;
1224 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1225 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1226 if(!rcu_gp_in_progress(rsp)) {
1227 rsp->n_force_qs_ngp++;
1228 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1229 goto unlock_fqs_ret; /* no GP in progress, time updated. */
1231 rsp->fqs_active = 1;
1232 switch (rsp->signaled) {
1233 case RCU_GP_IDLE:
1234 case RCU_GP_INIT:
1236 break; /* grace period idle or initializing, ignore. */
1238 case RCU_SAVE_DYNTICK:
1240 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1241 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1242 break; /* So gcc recognizes the dead code. */
1244 /* Record dyntick-idle state. */
1245 force_qs_rnp(rsp, dyntick_save_progress_counter);
1246 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1247 if (rcu_gp_in_progress(rsp))
1248 rsp->signaled = RCU_FORCE_QS;
1249 break;
1251 case RCU_FORCE_QS:
1253 /* Check dyntick-idle state, send IPI to laggarts. */
1254 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1255 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1257 /* Leave state in case more forcing is required. */
1259 raw_spin_lock(&rnp->lock); /* irqs already disabled */
1260 break;
1262 rsp->fqs_active = 0;
1263 if (rsp->fqs_need_gp) {
1264 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1265 rsp->fqs_need_gp = 0;
1266 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1267 return;
1269 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1270 unlock_fqs_ret:
1271 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1274 #else /* #ifdef CONFIG_SMP */
1276 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1278 set_need_resched();
1281 #endif /* #else #ifdef CONFIG_SMP */
1284 * This does the RCU processing work from softirq context for the
1285 * specified rcu_state and rcu_data structures. This may be called
1286 * only from the CPU to whom the rdp belongs.
1288 static void
1289 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1291 unsigned long flags;
1293 WARN_ON_ONCE(rdp->beenonline == 0);
1296 * If an RCU GP has gone long enough, go check for dyntick
1297 * idle CPUs and, if needed, send resched IPIs.
1299 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1300 force_quiescent_state(rsp, 1);
1303 * Advance callbacks in response to end of earlier grace
1304 * period that some other CPU ended.
1306 rcu_process_gp_end(rsp, rdp);
1308 /* Update RCU state based on any recent quiescent states. */
1309 rcu_check_quiescent_state(rsp, rdp);
1311 /* Does this CPU require a not-yet-started grace period? */
1312 if (cpu_needs_another_gp(rsp, rdp)) {
1313 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1314 rcu_start_gp(rsp, flags); /* releases above lock */
1317 /* If there are callbacks ready, invoke them. */
1318 rcu_do_batch(rsp, rdp);
1322 * Do softirq processing for the current CPU.
1324 static void rcu_process_callbacks(struct softirq_action *unused)
1327 * Memory references from any prior RCU read-side critical sections
1328 * executed by the interrupted code must be seen before any RCU
1329 * grace-period manipulations below.
1331 smp_mb(); /* See above block comment. */
1333 __rcu_process_callbacks(&rcu_sched_state,
1334 &__get_cpu_var(rcu_sched_data));
1335 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1336 rcu_preempt_process_callbacks();
1339 * Memory references from any later RCU read-side critical sections
1340 * executed by the interrupted code must be seen after any RCU
1341 * grace-period manipulations above.
1343 smp_mb(); /* See above block comment. */
1345 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1346 rcu_needs_cpu_flush();
1349 static void
1350 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1351 struct rcu_state *rsp)
1353 unsigned long flags;
1354 struct rcu_data *rdp;
1356 head->func = func;
1357 head->next = NULL;
1359 smp_mb(); /* Ensure RCU update seen before callback registry. */
1362 * Opportunistically note grace-period endings and beginnings.
1363 * Note that we might see a beginning right after we see an
1364 * end, but never vice versa, since this CPU has to pass through
1365 * a quiescent state betweentimes.
1367 local_irq_save(flags);
1368 rdp = rsp->rda[smp_processor_id()];
1369 rcu_process_gp_end(rsp, rdp);
1370 check_for_new_grace_period(rsp, rdp);
1372 /* Add the callback to our list. */
1373 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1374 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1376 /* Start a new grace period if one not already started. */
1377 if (!rcu_gp_in_progress(rsp)) {
1378 unsigned long nestflag;
1379 struct rcu_node *rnp_root = rcu_get_root(rsp);
1381 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1382 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1386 * Force the grace period if too many callbacks or too long waiting.
1387 * Enforce hysteresis, and don't invoke force_quiescent_state()
1388 * if some other CPU has recently done so. Also, don't bother
1389 * invoking force_quiescent_state() if the newly enqueued callback
1390 * is the only one waiting for a grace period to complete.
1392 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1393 rdp->blimit = LONG_MAX;
1394 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1395 *rdp->nxttail[RCU_DONE_TAIL] != head)
1396 force_quiescent_state(rsp, 0);
1397 rdp->n_force_qs_snap = rsp->n_force_qs;
1398 rdp->qlen_last_fqs_check = rdp->qlen;
1399 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1400 force_quiescent_state(rsp, 1);
1401 local_irq_restore(flags);
1405 * Queue an RCU-sched callback for invocation after a grace period.
1407 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1409 __call_rcu(head, func, &rcu_sched_state);
1411 EXPORT_SYMBOL_GPL(call_rcu_sched);
1414 * Queue an RCU for invocation after a quicker grace period.
1416 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1418 __call_rcu(head, func, &rcu_bh_state);
1420 EXPORT_SYMBOL_GPL(call_rcu_bh);
1423 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1425 * Control will return to the caller some time after a full rcu-sched
1426 * grace period has elapsed, in other words after all currently executing
1427 * rcu-sched read-side critical sections have completed. These read-side
1428 * critical sections are delimited by rcu_read_lock_sched() and
1429 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1430 * local_irq_disable(), and so on may be used in place of
1431 * rcu_read_lock_sched().
1433 * This means that all preempt_disable code sequences, including NMI and
1434 * hardware-interrupt handlers, in progress on entry will have completed
1435 * before this primitive returns. However, this does not guarantee that
1436 * softirq handlers will have completed, since in some kernels, these
1437 * handlers can run in process context, and can block.
1439 * This primitive provides the guarantees made by the (now removed)
1440 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1441 * guarantees that rcu_read_lock() sections will have completed.
1442 * In "classic RCU", these two guarantees happen to be one and
1443 * the same, but can differ in realtime RCU implementations.
1445 void synchronize_sched(void)
1447 struct rcu_synchronize rcu;
1449 if (rcu_blocking_is_gp())
1450 return;
1452 init_completion(&rcu.completion);
1453 /* Will wake me after RCU finished. */
1454 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1455 /* Wait for it. */
1456 wait_for_completion(&rcu.completion);
1458 EXPORT_SYMBOL_GPL(synchronize_sched);
1461 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1463 * Control will return to the caller some time after a full rcu_bh grace
1464 * period has elapsed, in other words after all currently executing rcu_bh
1465 * read-side critical sections have completed. RCU read-side critical
1466 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1467 * and may be nested.
1469 void synchronize_rcu_bh(void)
1471 struct rcu_synchronize rcu;
1473 if (rcu_blocking_is_gp())
1474 return;
1476 init_completion(&rcu.completion);
1477 /* Will wake me after RCU finished. */
1478 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1479 /* Wait for it. */
1480 wait_for_completion(&rcu.completion);
1482 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1485 * Check to see if there is any immediate RCU-related work to be done
1486 * by the current CPU, for the specified type of RCU, returning 1 if so.
1487 * The checks are in order of increasing expense: checks that can be
1488 * carried out against CPU-local state are performed first. However,
1489 * we must check for CPU stalls first, else we might not get a chance.
1491 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1493 struct rcu_node *rnp = rdp->mynode;
1495 rdp->n_rcu_pending++;
1497 /* Check for CPU stalls, if enabled. */
1498 check_cpu_stall(rsp, rdp);
1500 /* Is the RCU core waiting for a quiescent state from this CPU? */
1501 if (rdp->qs_pending) {
1502 rdp->n_rp_qs_pending++;
1503 return 1;
1506 /* Does this CPU have callbacks ready to invoke? */
1507 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1508 rdp->n_rp_cb_ready++;
1509 return 1;
1512 /* Has RCU gone idle with this CPU needing another grace period? */
1513 if (cpu_needs_another_gp(rsp, rdp)) {
1514 rdp->n_rp_cpu_needs_gp++;
1515 return 1;
1518 /* Has another RCU grace period completed? */
1519 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1520 rdp->n_rp_gp_completed++;
1521 return 1;
1524 /* Has a new RCU grace period started? */
1525 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1526 rdp->n_rp_gp_started++;
1527 return 1;
1530 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1531 if (rcu_gp_in_progress(rsp) &&
1532 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1533 rdp->n_rp_need_fqs++;
1534 return 1;
1537 /* nothing to do */
1538 rdp->n_rp_need_nothing++;
1539 return 0;
1543 * Check to see if there is any immediate RCU-related work to be done
1544 * by the current CPU, returning 1 if so. This function is part of the
1545 * RCU implementation; it is -not- an exported member of the RCU API.
1547 static int rcu_pending(int cpu)
1549 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1550 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1551 rcu_preempt_pending(cpu);
1555 * Check to see if any future RCU-related work will need to be done
1556 * by the current CPU, even if none need be done immediately, returning
1557 * 1 if so.
1559 static int rcu_needs_cpu_quick_check(int cpu)
1561 /* RCU callbacks either ready or pending? */
1562 return per_cpu(rcu_sched_data, cpu).nxtlist ||
1563 per_cpu(rcu_bh_data, cpu).nxtlist ||
1564 rcu_preempt_needs_cpu(cpu);
1567 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1568 static atomic_t rcu_barrier_cpu_count;
1569 static DEFINE_MUTEX(rcu_barrier_mutex);
1570 static struct completion rcu_barrier_completion;
1572 static void rcu_barrier_callback(struct rcu_head *notused)
1574 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1575 complete(&rcu_barrier_completion);
1579 * Called with preemption disabled, and from cross-cpu IRQ context.
1581 static void rcu_barrier_func(void *type)
1583 int cpu = smp_processor_id();
1584 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1585 void (*call_rcu_func)(struct rcu_head *head,
1586 void (*func)(struct rcu_head *head));
1588 atomic_inc(&rcu_barrier_cpu_count);
1589 call_rcu_func = type;
1590 call_rcu_func(head, rcu_barrier_callback);
1594 * Orchestrate the specified type of RCU barrier, waiting for all
1595 * RCU callbacks of the specified type to complete.
1597 static void _rcu_barrier(struct rcu_state *rsp,
1598 void (*call_rcu_func)(struct rcu_head *head,
1599 void (*func)(struct rcu_head *head)))
1601 BUG_ON(in_interrupt());
1602 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1603 mutex_lock(&rcu_barrier_mutex);
1604 init_completion(&rcu_barrier_completion);
1606 * Initialize rcu_barrier_cpu_count to 1, then invoke
1607 * rcu_barrier_func() on each CPU, so that each CPU also has
1608 * incremented rcu_barrier_cpu_count. Only then is it safe to
1609 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1610 * might complete its grace period before all of the other CPUs
1611 * did their increment, causing this function to return too
1612 * early.
1614 atomic_set(&rcu_barrier_cpu_count, 1);
1615 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1616 rcu_adopt_orphan_cbs(rsp);
1617 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1618 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1619 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1620 complete(&rcu_barrier_completion);
1621 wait_for_completion(&rcu_barrier_completion);
1622 mutex_unlock(&rcu_barrier_mutex);
1626 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1628 void rcu_barrier_bh(void)
1630 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1632 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1635 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1637 void rcu_barrier_sched(void)
1639 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1641 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1644 * Do boot-time initialization of a CPU's per-CPU RCU data.
1646 static void __init
1647 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1649 unsigned long flags;
1650 int i;
1651 struct rcu_data *rdp = rsp->rda[cpu];
1652 struct rcu_node *rnp = rcu_get_root(rsp);
1654 /* Set up local state, ensuring consistent view of global state. */
1655 raw_spin_lock_irqsave(&rnp->lock, flags);
1656 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1657 rdp->nxtlist = NULL;
1658 for (i = 0; i < RCU_NEXT_SIZE; i++)
1659 rdp->nxttail[i] = &rdp->nxtlist;
1660 rdp->qlen = 0;
1661 #ifdef CONFIG_NO_HZ
1662 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1663 #endif /* #ifdef CONFIG_NO_HZ */
1664 rdp->cpu = cpu;
1665 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1669 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1670 * offline event can be happening at a given time. Note also that we
1671 * can accept some slop in the rsp->completed access due to the fact
1672 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1674 static void __cpuinit
1675 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
1677 unsigned long flags;
1678 unsigned long mask;
1679 struct rcu_data *rdp = rsp->rda[cpu];
1680 struct rcu_node *rnp = rcu_get_root(rsp);
1682 /* Set up local state, ensuring consistent view of global state. */
1683 raw_spin_lock_irqsave(&rnp->lock, flags);
1684 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1685 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1686 rdp->beenonline = 1; /* We have now been online. */
1687 rdp->preemptable = preemptable;
1688 rdp->qlen_last_fqs_check = 0;
1689 rdp->n_force_qs_snap = rsp->n_force_qs;
1690 rdp->blimit = blimit;
1691 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1694 * A new grace period might start here. If so, we won't be part
1695 * of it, but that is OK, as we are currently in a quiescent state.
1698 /* Exclude any attempts to start a new GP on large systems. */
1699 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
1701 /* Add CPU to rcu_node bitmasks. */
1702 rnp = rdp->mynode;
1703 mask = rdp->grpmask;
1704 do {
1705 /* Exclude any attempts to start a new GP on small systems. */
1706 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1707 rnp->qsmaskinit |= mask;
1708 mask = rnp->grpmask;
1709 if (rnp == rdp->mynode) {
1710 rdp->gpnum = rnp->completed; /* if GP in progress... */
1711 rdp->completed = rnp->completed;
1712 rdp->passed_quiesc_completed = rnp->completed - 1;
1714 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1715 rnp = rnp->parent;
1716 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1718 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1721 static void __cpuinit rcu_online_cpu(int cpu)
1723 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1724 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1725 rcu_preempt_init_percpu_data(cpu);
1729 * Handle CPU online/offline notification events.
1731 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1732 unsigned long action, void *hcpu)
1734 long cpu = (long)hcpu;
1736 switch (action) {
1737 case CPU_UP_PREPARE:
1738 case CPU_UP_PREPARE_FROZEN:
1739 rcu_online_cpu(cpu);
1740 break;
1741 case CPU_DYING:
1742 case CPU_DYING_FROZEN:
1744 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1745 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1746 * returns, all online cpus have queued rcu_barrier_func().
1747 * The dying CPU clears its cpu_online_mask bit and
1748 * moves all of its RCU callbacks to ->orphan_cbs_list
1749 * in the context of stop_machine(), so subsequent calls
1750 * to _rcu_barrier() will adopt these callbacks and only
1751 * then queue rcu_barrier_func() on all remaining CPUs.
1753 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1754 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1755 rcu_preempt_send_cbs_to_orphanage();
1756 break;
1757 case CPU_DEAD:
1758 case CPU_DEAD_FROZEN:
1759 case CPU_UP_CANCELED:
1760 case CPU_UP_CANCELED_FROZEN:
1761 rcu_offline_cpu(cpu);
1762 break;
1763 default:
1764 break;
1766 return NOTIFY_OK;
1770 * Compute the per-level fanout, either using the exact fanout specified
1771 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1773 #ifdef CONFIG_RCU_FANOUT_EXACT
1774 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1776 int i;
1778 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1779 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1781 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1782 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1784 int ccur;
1785 int cprv;
1786 int i;
1788 cprv = NR_CPUS;
1789 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1790 ccur = rsp->levelcnt[i];
1791 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1792 cprv = ccur;
1795 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1798 * Helper function for rcu_init() that initializes one rcu_state structure.
1800 static void __init rcu_init_one(struct rcu_state *rsp)
1802 static char *buf[] = { "rcu_node_level_0",
1803 "rcu_node_level_1",
1804 "rcu_node_level_2",
1805 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1806 int cpustride = 1;
1807 int i;
1808 int j;
1809 struct rcu_node *rnp;
1811 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1813 /* Initialize the level-tracking arrays. */
1815 for (i = 1; i < NUM_RCU_LVLS; i++)
1816 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1817 rcu_init_levelspread(rsp);
1819 /* Initialize the elements themselves, starting from the leaves. */
1821 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1822 cpustride *= rsp->levelspread[i];
1823 rnp = rsp->level[i];
1824 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
1825 raw_spin_lock_init(&rnp->lock);
1826 lockdep_set_class_and_name(&rnp->lock,
1827 &rcu_node_class[i], buf[i]);
1828 rnp->gpnum = 0;
1829 rnp->qsmask = 0;
1830 rnp->qsmaskinit = 0;
1831 rnp->grplo = j * cpustride;
1832 rnp->grphi = (j + 1) * cpustride - 1;
1833 if (rnp->grphi >= NR_CPUS)
1834 rnp->grphi = NR_CPUS - 1;
1835 if (i == 0) {
1836 rnp->grpnum = 0;
1837 rnp->grpmask = 0;
1838 rnp->parent = NULL;
1839 } else {
1840 rnp->grpnum = j % rsp->levelspread[i - 1];
1841 rnp->grpmask = 1UL << rnp->grpnum;
1842 rnp->parent = rsp->level[i - 1] +
1843 j / rsp->levelspread[i - 1];
1845 rnp->level = i;
1846 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1847 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
1848 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1849 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
1855 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1856 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1857 * structure.
1859 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1860 do { \
1861 int i; \
1862 int j; \
1863 struct rcu_node *rnp; \
1865 rcu_init_one(rsp); \
1866 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1867 j = 0; \
1868 for_each_possible_cpu(i) { \
1869 if (i > rnp[j].grphi) \
1870 j++; \
1871 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1872 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1873 rcu_boot_init_percpu_data(i, rsp); \
1875 } while (0)
1877 void __init rcu_init(void)
1879 int cpu;
1881 rcu_bootup_announce();
1882 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1883 printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n");
1884 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1885 #if NUM_RCU_LVL_4 != 0
1886 printk(KERN_INFO "Experimental four-level hierarchy is enabled.\n");
1887 #endif /* #if NUM_RCU_LVL_4 != 0 */
1888 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1889 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
1890 __rcu_init_preempt();
1891 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
1894 * We don't need protection against CPU-hotplug here because
1895 * this is called early in boot, before either interrupts
1896 * or the scheduler are operational.
1898 cpu_notifier(rcu_cpu_notify, 0);
1899 for_each_online_cpu(cpu)
1900 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1903 #include "rcutree_plugin.h"