2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 2008
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
27 * For detailed explanation of Read-Copy Update mechanism see -
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/init.h>
33 #include <linux/spinlock.h>
34 #include <linux/smp.h>
35 #include <linux/rcupdate.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/nmi.h>
39 #include <asm/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
53 /* Data structures. */
55 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
57 #define RCU_STATE_INITIALIZER(name) { \
58 .level = { &name.node[0] }, \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
64 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
66 .signaled = RCU_GP_IDLE, \
69 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&name.onofflock), \
70 .orphan_cbs_list = NULL, \
71 .orphan_cbs_tail = &name.orphan_cbs_list, \
73 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&name.fqslock), \
75 .n_force_qs_ngp = 0, \
78 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched_state
);
79 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
81 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh_state
);
82 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
84 static int rcu_scheduler_active __read_mostly
;
88 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
89 * permit this function to be invoked without holding the root rcu_node
90 * structure's ->lock, but of course results can be subject to change.
92 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
94 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
98 * Note a quiescent state. Because we do not need to know
99 * how many quiescent states passed, just if there was at least
100 * one since the start of the grace period, this just sets a flag.
102 void rcu_sched_qs(int cpu
)
104 struct rcu_data
*rdp
;
106 rdp
= &per_cpu(rcu_sched_data
, cpu
);
107 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
109 rdp
->passed_quiesc
= 1;
110 rcu_preempt_note_context_switch(cpu
);
113 void rcu_bh_qs(int cpu
)
115 struct rcu_data
*rdp
;
117 rdp
= &per_cpu(rcu_bh_data
, cpu
);
118 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
120 rdp
->passed_quiesc
= 1;
124 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
125 .dynticks_nesting
= 1,
128 #endif /* #ifdef CONFIG_NO_HZ */
130 static int blimit
= 10; /* Maximum callbacks per softirq. */
131 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
132 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
134 module_param(blimit
, int, 0);
135 module_param(qhimark
, int, 0);
136 module_param(qlowmark
, int, 0);
138 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
139 static int rcu_pending(int cpu
);
142 * Return the number of RCU-sched batches processed thus far for debug & stats.
144 long rcu_batches_completed_sched(void)
146 return rcu_sched_state
.completed
;
148 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
151 * Return the number of RCU BH batches processed thus far for debug & stats.
153 long rcu_batches_completed_bh(void)
155 return rcu_bh_state
.completed
;
157 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
160 * Force a quiescent state for RCU BH.
162 void rcu_bh_force_quiescent_state(void)
164 force_quiescent_state(&rcu_bh_state
, 0);
166 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
169 * Force a quiescent state for RCU-sched.
171 void rcu_sched_force_quiescent_state(void)
173 force_quiescent_state(&rcu_sched_state
, 0);
175 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
178 * Does the CPU have callbacks ready to be invoked?
181 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
183 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
187 * Does the current CPU require a yet-as-unscheduled grace period?
190 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
192 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
196 * Return the root node of the specified rcu_state structure.
198 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
200 return &rsp
->node
[0];
206 * If the specified CPU is offline, tell the caller that it is in
207 * a quiescent state. Otherwise, whack it with a reschedule IPI.
208 * Grace periods can end up waiting on an offline CPU when that
209 * CPU is in the process of coming online -- it will be added to the
210 * rcu_node bitmasks before it actually makes it online. The same thing
211 * can happen while a CPU is in the process of coming online. Because this
212 * race is quite rare, we check for it after detecting that the grace
213 * period has been delayed rather than checking each and every CPU
214 * each and every time we start a new grace period.
216 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
219 * If the CPU is offline, it is in a quiescent state. We can
220 * trust its state not to change because interrupts are disabled.
222 if (cpu_is_offline(rdp
->cpu
)) {
227 /* If preemptable RCU, no point in sending reschedule IPI. */
228 if (rdp
->preemptable
)
231 /* The CPU is online, so send it a reschedule IPI. */
232 if (rdp
->cpu
!= smp_processor_id())
233 smp_send_reschedule(rdp
->cpu
);
240 #endif /* #ifdef CONFIG_SMP */
245 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
247 * Enter nohz mode, in other words, -leave- the mode in which RCU
248 * read-side critical sections can occur. (Though RCU read-side
249 * critical sections can occur in irq handlers in nohz mode, a possibility
250 * handled by rcu_irq_enter() and rcu_irq_exit()).
252 void rcu_enter_nohz(void)
255 struct rcu_dynticks
*rdtp
;
257 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
258 local_irq_save(flags
);
259 rdtp
= &__get_cpu_var(rcu_dynticks
);
261 rdtp
->dynticks_nesting
--;
262 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
263 local_irq_restore(flags
);
267 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
269 * Exit nohz mode, in other words, -enter- the mode in which RCU
270 * read-side critical sections normally occur.
272 void rcu_exit_nohz(void)
275 struct rcu_dynticks
*rdtp
;
277 local_irq_save(flags
);
278 rdtp
= &__get_cpu_var(rcu_dynticks
);
280 rdtp
->dynticks_nesting
++;
281 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
282 local_irq_restore(flags
);
283 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
287 * rcu_nmi_enter - inform RCU of entry to NMI context
289 * If the CPU was idle with dynamic ticks active, and there is no
290 * irq handler running, this updates rdtp->dynticks_nmi to let the
291 * RCU grace-period handling know that the CPU is active.
293 void rcu_nmi_enter(void)
295 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
297 if (rdtp
->dynticks
& 0x1)
299 rdtp
->dynticks_nmi
++;
300 WARN_ON_ONCE(!(rdtp
->dynticks_nmi
& 0x1));
301 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
305 * rcu_nmi_exit - inform RCU of exit from NMI context
307 * If the CPU was idle with dynamic ticks active, and there is no
308 * irq handler running, this updates rdtp->dynticks_nmi to let the
309 * RCU grace-period handling know that the CPU is no longer active.
311 void rcu_nmi_exit(void)
313 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
315 if (rdtp
->dynticks
& 0x1)
317 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
318 rdtp
->dynticks_nmi
++;
319 WARN_ON_ONCE(rdtp
->dynticks_nmi
& 0x1);
323 * rcu_irq_enter - inform RCU of entry to hard irq context
325 * If the CPU was idle with dynamic ticks active, this updates the
326 * rdtp->dynticks to let the RCU handling know that the CPU is active.
328 void rcu_irq_enter(void)
330 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
332 if (rdtp
->dynticks_nesting
++)
335 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
336 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
340 * rcu_irq_exit - inform RCU of exit from hard irq context
342 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
343 * to put let the RCU handling be aware that the CPU is going back to idle
346 void rcu_irq_exit(void)
348 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
350 if (--rdtp
->dynticks_nesting
)
352 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
354 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
356 /* If the interrupt queued a callback, get out of dyntick mode. */
357 if (__get_cpu_var(rcu_sched_data
).nxtlist
||
358 __get_cpu_var(rcu_bh_data
).nxtlist
)
365 * Snapshot the specified CPU's dynticks counter so that we can later
366 * credit them with an implicit quiescent state. Return 1 if this CPU
367 * is in dynticks idle mode, which is an extended quiescent state.
369 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
375 snap
= rdp
->dynticks
->dynticks
;
376 snap_nmi
= rdp
->dynticks
->dynticks_nmi
;
377 smp_mb(); /* Order sampling of snap with end of grace period. */
378 rdp
->dynticks_snap
= snap
;
379 rdp
->dynticks_nmi_snap
= snap_nmi
;
380 ret
= ((snap
& 0x1) == 0) && ((snap_nmi
& 0x1) == 0);
387 * Return true if the specified CPU has passed through a quiescent
388 * state by virtue of being in or having passed through an dynticks
389 * idle state since the last call to dyntick_save_progress_counter()
392 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
399 curr
= rdp
->dynticks
->dynticks
;
400 snap
= rdp
->dynticks_snap
;
401 curr_nmi
= rdp
->dynticks
->dynticks_nmi
;
402 snap_nmi
= rdp
->dynticks_nmi_snap
;
403 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
406 * If the CPU passed through or entered a dynticks idle phase with
407 * no active irq/NMI handlers, then we can safely pretend that the CPU
408 * already acknowledged the request to pass through a quiescent
409 * state. Either way, that CPU cannot possibly be in an RCU
410 * read-side critical section that started before the beginning
411 * of the current RCU grace period.
413 if ((curr
!= snap
|| (curr
& 0x1) == 0) &&
414 (curr_nmi
!= snap_nmi
|| (curr_nmi
& 0x1) == 0)) {
419 /* Go check for the CPU being offline. */
420 return rcu_implicit_offline_qs(rdp
);
423 #endif /* #ifdef CONFIG_SMP */
425 #else /* #ifdef CONFIG_NO_HZ */
429 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
434 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
436 return rcu_implicit_offline_qs(rdp
);
439 #endif /* #ifdef CONFIG_SMP */
441 #endif /* #else #ifdef CONFIG_NO_HZ */
443 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
445 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
447 rsp
->gp_start
= jiffies
;
448 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
451 static void print_other_cpu_stall(struct rcu_state
*rsp
)
456 struct rcu_node
*rnp
= rcu_get_root(rsp
);
458 /* Only let one CPU complain about others per time interval. */
460 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
461 delta
= jiffies
- rsp
->jiffies_stall
;
462 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
463 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
466 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
469 * Now rat on any tasks that got kicked up to the root rcu_node
470 * due to CPU offlining.
472 rcu_print_task_stall(rnp
);
473 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
475 /* OK, time to rat on our buddy... */
477 printk(KERN_ERR
"INFO: RCU detected CPU stalls:");
478 rcu_for_each_leaf_node(rsp
, rnp
) {
479 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
480 rcu_print_task_stall(rnp
);
481 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
482 if (rnp
->qsmask
== 0)
484 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
485 if (rnp
->qsmask
& (1UL << cpu
))
486 printk(" %d", rnp
->grplo
+ cpu
);
488 printk(" (detected by %d, t=%ld jiffies)\n",
489 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
490 trigger_all_cpu_backtrace();
492 /* If so configured, complain about tasks blocking the grace period. */
494 rcu_print_detail_task_stall(rsp
);
496 force_quiescent_state(rsp
, 0); /* Kick them all. */
499 static void print_cpu_stall(struct rcu_state
*rsp
)
502 struct rcu_node
*rnp
= rcu_get_root(rsp
);
504 printk(KERN_ERR
"INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
505 smp_processor_id(), jiffies
- rsp
->gp_start
);
506 trigger_all_cpu_backtrace();
508 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
509 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
511 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
512 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
514 set_need_resched(); /* kick ourselves to get things going. */
517 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
520 struct rcu_node
*rnp
;
522 delta
= jiffies
- rsp
->jiffies_stall
;
524 if ((rnp
->qsmask
& rdp
->grpmask
) && delta
>= 0) {
526 /* We haven't checked in, so go dump stack. */
527 print_cpu_stall(rsp
);
529 } else if (rcu_gp_in_progress(rsp
) && delta
>= RCU_STALL_RAT_DELAY
) {
531 /* They had two time units to dump stack, so complain. */
532 print_other_cpu_stall(rsp
);
536 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
538 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
542 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
546 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
549 * Update CPU-local rcu_data state to record the newly noticed grace period.
550 * This is used both when we started the grace period and when we notice
551 * that someone else started the grace period. The caller must hold the
552 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
553 * and must have irqs disabled.
555 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
557 if (rdp
->gpnum
!= rnp
->gpnum
) {
559 rdp
->passed_quiesc
= 0;
560 rdp
->gpnum
= rnp
->gpnum
;
564 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
567 struct rcu_node
*rnp
;
569 local_irq_save(flags
);
571 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
572 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
573 local_irq_restore(flags
);
576 __note_new_gpnum(rsp
, rnp
, rdp
);
577 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
581 * Did someone else start a new RCU grace period start since we last
582 * checked? Update local state appropriately if so. Must be called
583 * on the CPU corresponding to rdp.
586 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
591 local_irq_save(flags
);
592 if (rdp
->gpnum
!= rsp
->gpnum
) {
593 note_new_gpnum(rsp
, rdp
);
596 local_irq_restore(flags
);
601 * Advance this CPU's callbacks, but only if the current grace period
602 * has ended. This may be called only from the CPU to whom the rdp
603 * belongs. In addition, the corresponding leaf rcu_node structure's
604 * ->lock must be held by the caller, with irqs disabled.
607 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
609 /* Did another grace period end? */
610 if (rdp
->completed
!= rnp
->completed
) {
612 /* Advance callbacks. No harm if list empty. */
613 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
614 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
615 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
617 /* Remember that we saw this grace-period completion. */
618 rdp
->completed
= rnp
->completed
;
623 * Advance this CPU's callbacks, but only if the current grace period
624 * has ended. This may be called only from the CPU to whom the rdp
628 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
631 struct rcu_node
*rnp
;
633 local_irq_save(flags
);
635 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
636 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
637 local_irq_restore(flags
);
640 __rcu_process_gp_end(rsp
, rnp
, rdp
);
641 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
645 * Do per-CPU grace-period initialization for running CPU. The caller
646 * must hold the lock of the leaf rcu_node structure corresponding to
650 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
652 /* Prior grace period ended, so advance callbacks for current CPU. */
653 __rcu_process_gp_end(rsp
, rnp
, rdp
);
656 * Because this CPU just now started the new grace period, we know
657 * that all of its callbacks will be covered by this upcoming grace
658 * period, even the ones that were registered arbitrarily recently.
659 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
661 * Other CPUs cannot be sure exactly when the grace period started.
662 * Therefore, their recently registered callbacks must pass through
663 * an additional RCU_NEXT_READY stage, so that they will be handled
664 * by the next RCU grace period.
666 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
667 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
669 /* Set state so that this CPU will detect the next quiescent state. */
670 __note_new_gpnum(rsp
, rnp
, rdp
);
674 * Start a new RCU grace period if warranted, re-initializing the hierarchy
675 * in preparation for detecting the next grace period. The caller must hold
676 * the root node's ->lock, which is released before return. Hard irqs must
680 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
681 __releases(rcu_get_root(rsp
)->lock
)
683 struct rcu_data
*rdp
= rsp
->rda
[smp_processor_id()];
684 struct rcu_node
*rnp
= rcu_get_root(rsp
);
686 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
687 if (cpu_needs_another_gp(rsp
, rdp
))
688 rsp
->fqs_need_gp
= 1;
689 if (rnp
->completed
== rsp
->completed
) {
690 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
693 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
696 * Propagate new ->completed value to rcu_node structures
697 * so that other CPUs don't have to wait until the start
698 * of the next grace period to process their callbacks.
700 rcu_for_each_node_breadth_first(rsp
, rnp
) {
701 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
702 rnp
->completed
= rsp
->completed
;
703 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
705 local_irq_restore(flags
);
709 /* Advance to a new grace period and initialize state. */
711 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
712 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
713 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
714 record_gp_stall_check_time(rsp
);
716 /* Special-case the common single-level case. */
717 if (NUM_RCU_NODES
== 1) {
718 rcu_preempt_check_blocked_tasks(rnp
);
719 rnp
->qsmask
= rnp
->qsmaskinit
;
720 rnp
->gpnum
= rsp
->gpnum
;
721 rnp
->completed
= rsp
->completed
;
722 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
723 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
724 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
728 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
731 /* Exclude any concurrent CPU-hotplug operations. */
732 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
735 * Set the quiescent-state-needed bits in all the rcu_node
736 * structures for all currently online CPUs in breadth-first
737 * order, starting from the root rcu_node structure. This
738 * operation relies on the layout of the hierarchy within the
739 * rsp->node[] array. Note that other CPUs will access only
740 * the leaves of the hierarchy, which still indicate that no
741 * grace period is in progress, at least until the corresponding
742 * leaf node has been initialized. In addition, we have excluded
743 * CPU-hotplug operations.
745 * Note that the grace period cannot complete until we finish
746 * the initialization process, as there will be at least one
747 * qsmask bit set in the root node until that time, namely the
748 * one corresponding to this CPU, due to the fact that we have
751 rcu_for_each_node_breadth_first(rsp
, rnp
) {
752 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
753 rcu_preempt_check_blocked_tasks(rnp
);
754 rnp
->qsmask
= rnp
->qsmaskinit
;
755 rnp
->gpnum
= rsp
->gpnum
;
756 rnp
->completed
= rsp
->completed
;
757 if (rnp
== rdp
->mynode
)
758 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
759 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
762 rnp
= rcu_get_root(rsp
);
763 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
764 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
765 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
766 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
770 * Report a full set of quiescent states to the specified rcu_state
771 * data structure. This involves cleaning up after the prior grace
772 * period and letting rcu_start_gp() start up the next grace period
773 * if one is needed. Note that the caller must hold rnp->lock, as
774 * required by rcu_start_gp(), which will release it.
776 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
777 __releases(rcu_get_root(rsp
)->lock
)
779 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
780 rsp
->completed
= rsp
->gpnum
;
781 rsp
->signaled
= RCU_GP_IDLE
;
782 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
786 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
787 * Allows quiescent states for a group of CPUs to be reported at one go
788 * to the specified rcu_node structure, though all the CPUs in the group
789 * must be represented by the same rcu_node structure (which need not be
790 * a leaf rcu_node structure, though it often will be). That structure's
791 * lock must be held upon entry, and it is released before return.
794 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
795 struct rcu_node
*rnp
, unsigned long flags
)
796 __releases(rnp
->lock
)
798 struct rcu_node
*rnp_c
;
800 /* Walk up the rcu_node hierarchy. */
802 if (!(rnp
->qsmask
& mask
)) {
804 /* Our bit has already been cleared, so done. */
805 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
808 rnp
->qsmask
&= ~mask
;
809 if (rnp
->qsmask
!= 0 || rcu_preempted_readers(rnp
)) {
811 /* Other bits still set at this level, so done. */
812 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
816 if (rnp
->parent
== NULL
) {
818 /* No more levels. Exit loop holding root lock. */
822 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
825 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
826 WARN_ON_ONCE(rnp_c
->qsmask
);
830 * Get here if we are the last CPU to pass through a quiescent
831 * state for this grace period. Invoke rcu_report_qs_rsp()
832 * to clean up and start the next grace period if one is needed.
834 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
838 * Record a quiescent state for the specified CPU to that CPU's rcu_data
839 * structure. This must be either called from the specified CPU, or
840 * called when the specified CPU is known to be offline (and when it is
841 * also known that no other CPU is concurrently trying to help the offline
842 * CPU). The lastcomp argument is used to make sure we are still in the
843 * grace period of interest. We don't want to end the current grace period
844 * based on quiescent states detected in an earlier grace period!
847 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
851 struct rcu_node
*rnp
;
854 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
855 if (lastcomp
!= rnp
->completed
) {
858 * Someone beat us to it for this grace period, so leave.
859 * The race with GP start is resolved by the fact that we
860 * hold the leaf rcu_node lock, so that the per-CPU bits
861 * cannot yet be initialized -- so we would simply find our
862 * CPU's bit already cleared in rcu_report_qs_rnp() if this
865 rdp
->passed_quiesc
= 0; /* try again later! */
866 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
870 if ((rnp
->qsmask
& mask
) == 0) {
871 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
876 * This GP can't end until cpu checks in, so all of our
877 * callbacks can be processed during the next GP.
879 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
881 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
886 * Check to see if there is a new grace period of which this CPU
887 * is not yet aware, and if so, set up local rcu_data state for it.
888 * Otherwise, see if this CPU has just passed through its first
889 * quiescent state for this grace period, and record that fact if so.
892 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
894 /* If there is now a new grace period, record and return. */
895 if (check_for_new_grace_period(rsp
, rdp
))
899 * Does this CPU still need to do its part for current grace period?
900 * If no, return and let the other CPUs do their part as well.
902 if (!rdp
->qs_pending
)
906 * Was there a quiescent state since the beginning of the grace
907 * period? If no, then exit and wait for the next call.
909 if (!rdp
->passed_quiesc
)
913 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
916 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
919 #ifdef CONFIG_HOTPLUG_CPU
922 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
923 * specified flavor of RCU. The callbacks will be adopted by the next
924 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
925 * comes first. Because this is invoked from the CPU_DYING notifier,
926 * irqs are already disabled.
928 static void rcu_send_cbs_to_orphanage(struct rcu_state
*rsp
)
931 struct rcu_data
*rdp
= rsp
->rda
[smp_processor_id()];
933 if (rdp
->nxtlist
== NULL
)
934 return; /* irqs disabled, so comparison is stable. */
935 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
936 *rsp
->orphan_cbs_tail
= rdp
->nxtlist
;
937 rsp
->orphan_cbs_tail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
939 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
940 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
941 rsp
->orphan_qlen
+= rdp
->qlen
;
943 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
947 * Adopt previously orphaned RCU callbacks.
949 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
952 struct rcu_data
*rdp
;
954 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
955 rdp
= rsp
->rda
[smp_processor_id()];
956 if (rsp
->orphan_cbs_list
== NULL
) {
957 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
960 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_cbs_list
;
961 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_cbs_tail
;
962 rdp
->qlen
+= rsp
->orphan_qlen
;
963 rsp
->orphan_cbs_list
= NULL
;
964 rsp
->orphan_cbs_tail
= &rsp
->orphan_cbs_list
;
965 rsp
->orphan_qlen
= 0;
966 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
970 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
971 * and move all callbacks from the outgoing CPU to the current one.
973 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
978 struct rcu_data
*rdp
= rsp
->rda
[cpu
];
979 struct rcu_node
*rnp
;
981 /* Exclude any attempts to start a new grace period. */
982 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
984 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
985 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
986 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
988 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
989 rnp
->qsmaskinit
&= ~mask
;
990 if (rnp
->qsmaskinit
!= 0) {
991 if (rnp
!= rdp
->mynode
)
992 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
995 if (rnp
== rdp
->mynode
)
996 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
998 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1001 } while (rnp
!= NULL
);
1004 * We still hold the leaf rcu_node structure lock here, and
1005 * irqs are still disabled. The reason for this subterfuge is
1006 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1007 * held leads to deadlock.
1009 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1011 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1012 rcu_report_unblock_qs_rnp(rnp
, flags
);
1014 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1015 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1016 rcu_report_exp_rnp(rsp
, rnp
);
1018 rcu_adopt_orphan_cbs(rsp
);
1022 * Remove the specified CPU from the RCU hierarchy and move any pending
1023 * callbacks that it might have to the current CPU. This code assumes
1024 * that at least one CPU in the system will remain running at all times.
1025 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1027 static void rcu_offline_cpu(int cpu
)
1029 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1030 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1031 rcu_preempt_offline_cpu(cpu
);
1034 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1036 static void rcu_send_cbs_to_orphanage(struct rcu_state
*rsp
)
1040 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1044 static void rcu_offline_cpu(int cpu
)
1048 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1051 * Invoke any RCU callbacks that have made it to the end of their grace
1052 * period. Thottle as specified by rdp->blimit.
1054 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1056 unsigned long flags
;
1057 struct rcu_head
*next
, *list
, **tail
;
1060 /* If no callbacks are ready, just return.*/
1061 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1065 * Extract the list of ready callbacks, disabling to prevent
1066 * races with call_rcu() from interrupt handlers.
1068 local_irq_save(flags
);
1069 list
= rdp
->nxtlist
;
1070 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1071 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1072 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1073 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1074 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1075 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1076 local_irq_restore(flags
);
1078 /* Invoke callbacks. */
1085 if (++count
>= rdp
->blimit
)
1089 local_irq_save(flags
);
1091 /* Update count, and requeue any remaining callbacks. */
1094 *tail
= rdp
->nxtlist
;
1095 rdp
->nxtlist
= list
;
1096 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1097 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1098 rdp
->nxttail
[count
] = tail
;
1103 /* Reinstate batch limit if we have worked down the excess. */
1104 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1105 rdp
->blimit
= blimit
;
1107 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1108 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1109 rdp
->qlen_last_fqs_check
= 0;
1110 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1111 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1112 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1114 local_irq_restore(flags
);
1116 /* Re-raise the RCU softirq if there are callbacks remaining. */
1117 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1118 raise_softirq(RCU_SOFTIRQ
);
1122 * Check to see if this CPU is in a non-context-switch quiescent state
1123 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1124 * Also schedule the RCU softirq handler.
1126 * This function must be called with hardirqs disabled. It is normally
1127 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1128 * false, there is no point in invoking rcu_check_callbacks().
1130 void rcu_check_callbacks(int cpu
, int user
)
1132 if (!rcu_pending(cpu
))
1133 return; /* if nothing for RCU to do. */
1135 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1136 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1139 * Get here if this CPU took its interrupt from user
1140 * mode or from the idle loop, and if this is not a
1141 * nested interrupt. In this case, the CPU is in
1142 * a quiescent state, so note it.
1144 * No memory barrier is required here because both
1145 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1146 * variables that other CPUs neither access nor modify,
1147 * at least not while the corresponding CPU is online.
1153 } else if (!in_softirq()) {
1156 * Get here if this CPU did not take its interrupt from
1157 * softirq, in other words, if it is not interrupting
1158 * a rcu_bh read-side critical section. This is an _bh
1159 * critical section, so note it.
1164 rcu_preempt_check_callbacks(cpu
);
1165 raise_softirq(RCU_SOFTIRQ
);
1171 * Scan the leaf rcu_node structures, processing dyntick state for any that
1172 * have not yet encountered a quiescent state, using the function specified.
1173 * The caller must have suppressed start of new grace periods.
1175 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1179 unsigned long flags
;
1181 struct rcu_node
*rnp
;
1183 rcu_for_each_leaf_node(rsp
, rnp
) {
1185 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1186 if (!rcu_gp_in_progress(rsp
)) {
1187 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1190 if (rnp
->qsmask
== 0) {
1191 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1196 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1197 if ((rnp
->qsmask
& bit
) != 0 && f(rsp
->rda
[cpu
]))
1202 /* rcu_report_qs_rnp() releases rnp->lock. */
1203 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1206 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1211 * Force quiescent states on reluctant CPUs, and also detect which
1212 * CPUs are in dyntick-idle mode.
1214 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1216 unsigned long flags
;
1217 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1219 if (!rcu_gp_in_progress(rsp
))
1220 return; /* No grace period in progress, nothing to force. */
1221 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1222 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1223 return; /* Someone else is already on the job. */
1225 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1226 goto unlock_fqs_ret
; /* no emergency and done recently. */
1228 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1229 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1230 if(!rcu_gp_in_progress(rsp
)) {
1231 rsp
->n_force_qs_ngp
++;
1232 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1233 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1235 rsp
->fqs_active
= 1;
1236 switch (rsp
->signaled
) {
1240 break; /* grace period idle or initializing, ignore. */
1242 case RCU_SAVE_DYNTICK
:
1244 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1245 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1246 break; /* So gcc recognizes the dead code. */
1248 /* Record dyntick-idle state. */
1249 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1250 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1251 if (rcu_gp_in_progress(rsp
))
1252 rsp
->signaled
= RCU_FORCE_QS
;
1257 /* Check dyntick-idle state, send IPI to laggarts. */
1258 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1259 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1261 /* Leave state in case more forcing is required. */
1263 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1266 rsp
->fqs_active
= 0;
1267 if (rsp
->fqs_need_gp
) {
1268 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1269 rsp
->fqs_need_gp
= 0;
1270 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1273 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1275 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1278 #else /* #ifdef CONFIG_SMP */
1280 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1285 #endif /* #else #ifdef CONFIG_SMP */
1288 * This does the RCU processing work from softirq context for the
1289 * specified rcu_state and rcu_data structures. This may be called
1290 * only from the CPU to whom the rdp belongs.
1293 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1295 unsigned long flags
;
1297 WARN_ON_ONCE(rdp
->beenonline
== 0);
1300 * If an RCU GP has gone long enough, go check for dyntick
1301 * idle CPUs and, if needed, send resched IPIs.
1303 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1304 force_quiescent_state(rsp
, 1);
1307 * Advance callbacks in response to end of earlier grace
1308 * period that some other CPU ended.
1310 rcu_process_gp_end(rsp
, rdp
);
1312 /* Update RCU state based on any recent quiescent states. */
1313 rcu_check_quiescent_state(rsp
, rdp
);
1315 /* Does this CPU require a not-yet-started grace period? */
1316 if (cpu_needs_another_gp(rsp
, rdp
)) {
1317 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1318 rcu_start_gp(rsp
, flags
); /* releases above lock */
1321 /* If there are callbacks ready, invoke them. */
1322 rcu_do_batch(rsp
, rdp
);
1326 * Do softirq processing for the current CPU.
1328 static void rcu_process_callbacks(struct softirq_action
*unused
)
1331 * Memory references from any prior RCU read-side critical sections
1332 * executed by the interrupted code must be seen before any RCU
1333 * grace-period manipulations below.
1335 smp_mb(); /* See above block comment. */
1337 __rcu_process_callbacks(&rcu_sched_state
,
1338 &__get_cpu_var(rcu_sched_data
));
1339 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1340 rcu_preempt_process_callbacks();
1343 * Memory references from any later RCU read-side critical sections
1344 * executed by the interrupted code must be seen after any RCU
1345 * grace-period manipulations above.
1347 smp_mb(); /* See above block comment. */
1351 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1352 struct rcu_state
*rsp
)
1354 unsigned long flags
;
1355 struct rcu_data
*rdp
;
1360 smp_mb(); /* Ensure RCU update seen before callback registry. */
1363 * Opportunistically note grace-period endings and beginnings.
1364 * Note that we might see a beginning right after we see an
1365 * end, but never vice versa, since this CPU has to pass through
1366 * a quiescent state betweentimes.
1368 local_irq_save(flags
);
1369 rdp
= rsp
->rda
[smp_processor_id()];
1370 rcu_process_gp_end(rsp
, rdp
);
1371 check_for_new_grace_period(rsp
, rdp
);
1373 /* Add the callback to our list. */
1374 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1375 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1377 /* Start a new grace period if one not already started. */
1378 if (!rcu_gp_in_progress(rsp
)) {
1379 unsigned long nestflag
;
1380 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1382 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1383 rcu_start_gp(rsp
, nestflag
); /* releases rnp_root->lock. */
1387 * Force the grace period if too many callbacks or too long waiting.
1388 * Enforce hysteresis, and don't invoke force_quiescent_state()
1389 * if some other CPU has recently done so. Also, don't bother
1390 * invoking force_quiescent_state() if the newly enqueued callback
1391 * is the only one waiting for a grace period to complete.
1393 if (unlikely(++rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1394 rdp
->blimit
= LONG_MAX
;
1395 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1396 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1397 force_quiescent_state(rsp
, 0);
1398 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1399 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1400 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1401 force_quiescent_state(rsp
, 1);
1402 local_irq_restore(flags
);
1406 * Queue an RCU-sched callback for invocation after a grace period.
1408 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1410 __call_rcu(head
, func
, &rcu_sched_state
);
1412 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1415 * Queue an RCU for invocation after a quicker grace period.
1417 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1419 __call_rcu(head
, func
, &rcu_bh_state
);
1421 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1424 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1426 * Control will return to the caller some time after a full rcu-sched
1427 * grace period has elapsed, in other words after all currently executing
1428 * rcu-sched read-side critical sections have completed. These read-side
1429 * critical sections are delimited by rcu_read_lock_sched() and
1430 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1431 * local_irq_disable(), and so on may be used in place of
1432 * rcu_read_lock_sched().
1434 * This means that all preempt_disable code sequences, including NMI and
1435 * hardware-interrupt handlers, in progress on entry will have completed
1436 * before this primitive returns. However, this does not guarantee that
1437 * softirq handlers will have completed, since in some kernels, these
1438 * handlers can run in process context, and can block.
1440 * This primitive provides the guarantees made by the (now removed)
1441 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1442 * guarantees that rcu_read_lock() sections will have completed.
1443 * In "classic RCU", these two guarantees happen to be one and
1444 * the same, but can differ in realtime RCU implementations.
1446 void synchronize_sched(void)
1448 struct rcu_synchronize rcu
;
1450 if (rcu_blocking_is_gp())
1453 init_completion(&rcu
.completion
);
1454 /* Will wake me after RCU finished. */
1455 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1457 wait_for_completion(&rcu
.completion
);
1459 EXPORT_SYMBOL_GPL(synchronize_sched
);
1462 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1464 * Control will return to the caller some time after a full rcu_bh grace
1465 * period has elapsed, in other words after all currently executing rcu_bh
1466 * read-side critical sections have completed. RCU read-side critical
1467 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1468 * and may be nested.
1470 void synchronize_rcu_bh(void)
1472 struct rcu_synchronize rcu
;
1474 if (rcu_blocking_is_gp())
1477 init_completion(&rcu
.completion
);
1478 /* Will wake me after RCU finished. */
1479 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1481 wait_for_completion(&rcu
.completion
);
1483 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1486 * Check to see if there is any immediate RCU-related work to be done
1487 * by the current CPU, for the specified type of RCU, returning 1 if so.
1488 * The checks are in order of increasing expense: checks that can be
1489 * carried out against CPU-local state are performed first. However,
1490 * we must check for CPU stalls first, else we might not get a chance.
1492 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1494 struct rcu_node
*rnp
= rdp
->mynode
;
1496 rdp
->n_rcu_pending
++;
1498 /* Check for CPU stalls, if enabled. */
1499 check_cpu_stall(rsp
, rdp
);
1501 /* Is the RCU core waiting for a quiescent state from this CPU? */
1502 if (rdp
->qs_pending
) {
1503 rdp
->n_rp_qs_pending
++;
1507 /* Does this CPU have callbacks ready to invoke? */
1508 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1509 rdp
->n_rp_cb_ready
++;
1513 /* Has RCU gone idle with this CPU needing another grace period? */
1514 if (cpu_needs_another_gp(rsp
, rdp
)) {
1515 rdp
->n_rp_cpu_needs_gp
++;
1519 /* Has another RCU grace period completed? */
1520 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
1521 rdp
->n_rp_gp_completed
++;
1525 /* Has a new RCU grace period started? */
1526 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
1527 rdp
->n_rp_gp_started
++;
1531 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1532 if (rcu_gp_in_progress(rsp
) &&
1533 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
1534 rdp
->n_rp_need_fqs
++;
1539 rdp
->n_rp_need_nothing
++;
1544 * Check to see if there is any immediate RCU-related work to be done
1545 * by the current CPU, returning 1 if so. This function is part of the
1546 * RCU implementation; it is -not- an exported member of the RCU API.
1548 static int rcu_pending(int cpu
)
1550 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
1551 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
1552 rcu_preempt_pending(cpu
);
1556 * Check to see if any future RCU-related work will need to be done
1557 * by the current CPU, even if none need be done immediately, returning
1560 static int rcu_needs_cpu_quick_check(int cpu
)
1562 /* RCU callbacks either ready or pending? */
1563 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
1564 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
1565 rcu_preempt_needs_cpu(cpu
);
1569 * This function is invoked towards the end of the scheduler's initialization
1570 * process. Before this is called, the idle task might contain
1571 * RCU read-side critical sections (during which time, this idle
1572 * task is booting the system). After this function is called, the
1573 * idle tasks are prohibited from containing RCU read-side critical
1576 void rcu_scheduler_starting(void)
1578 WARN_ON(num_online_cpus() != 1);
1579 WARN_ON(nr_context_switches() > 0);
1580 rcu_scheduler_active
= 1;
1583 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
1584 static atomic_t rcu_barrier_cpu_count
;
1585 static DEFINE_MUTEX(rcu_barrier_mutex
);
1586 static struct completion rcu_barrier_completion
;
1588 static void rcu_barrier_callback(struct rcu_head
*notused
)
1590 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1591 complete(&rcu_barrier_completion
);
1595 * Called with preemption disabled, and from cross-cpu IRQ context.
1597 static void rcu_barrier_func(void *type
)
1599 int cpu
= smp_processor_id();
1600 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
1601 void (*call_rcu_func
)(struct rcu_head
*head
,
1602 void (*func
)(struct rcu_head
*head
));
1604 atomic_inc(&rcu_barrier_cpu_count
);
1605 call_rcu_func
= type
;
1606 call_rcu_func(head
, rcu_barrier_callback
);
1610 * Orchestrate the specified type of RCU barrier, waiting for all
1611 * RCU callbacks of the specified type to complete.
1613 static void _rcu_barrier(struct rcu_state
*rsp
,
1614 void (*call_rcu_func
)(struct rcu_head
*head
,
1615 void (*func
)(struct rcu_head
*head
)))
1617 BUG_ON(in_interrupt());
1618 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1619 mutex_lock(&rcu_barrier_mutex
);
1620 init_completion(&rcu_barrier_completion
);
1622 * Initialize rcu_barrier_cpu_count to 1, then invoke
1623 * rcu_barrier_func() on each CPU, so that each CPU also has
1624 * incremented rcu_barrier_cpu_count. Only then is it safe to
1625 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1626 * might complete its grace period before all of the other CPUs
1627 * did their increment, causing this function to return too
1630 atomic_set(&rcu_barrier_cpu_count
, 1);
1631 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1632 rcu_adopt_orphan_cbs(rsp
);
1633 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
1634 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1635 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1636 complete(&rcu_barrier_completion
);
1637 wait_for_completion(&rcu_barrier_completion
);
1638 mutex_unlock(&rcu_barrier_mutex
);
1642 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1644 void rcu_barrier_bh(void)
1646 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
1648 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
1651 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1653 void rcu_barrier_sched(void)
1655 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
1657 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
1660 * Do boot-time initialization of a CPU's per-CPU RCU data.
1663 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
1665 unsigned long flags
;
1667 struct rcu_data
*rdp
= rsp
->rda
[cpu
];
1668 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1670 /* Set up local state, ensuring consistent view of global state. */
1671 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1672 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
1673 rdp
->nxtlist
= NULL
;
1674 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1675 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1678 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
1679 #endif /* #ifdef CONFIG_NO_HZ */
1681 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1685 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1686 * offline event can be happening at a given time. Note also that we
1687 * can accept some slop in the rsp->completed access due to the fact
1688 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1690 static void __cpuinit
1691 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptable
)
1693 unsigned long flags
;
1695 struct rcu_data
*rdp
= rsp
->rda
[cpu
];
1696 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1698 /* Set up local state, ensuring consistent view of global state. */
1699 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1700 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
1701 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
1702 rdp
->beenonline
= 1; /* We have now been online. */
1703 rdp
->preemptable
= preemptable
;
1704 rdp
->qlen_last_fqs_check
= 0;
1705 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1706 rdp
->blimit
= blimit
;
1707 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1710 * A new grace period might start here. If so, we won't be part
1711 * of it, but that is OK, as we are currently in a quiescent state.
1714 /* Exclude any attempts to start a new GP on large systems. */
1715 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
1717 /* Add CPU to rcu_node bitmasks. */
1719 mask
= rdp
->grpmask
;
1721 /* Exclude any attempts to start a new GP on small systems. */
1722 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1723 rnp
->qsmaskinit
|= mask
;
1724 mask
= rnp
->grpmask
;
1725 if (rnp
== rdp
->mynode
) {
1726 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
1727 rdp
->completed
= rnp
->completed
;
1728 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
1730 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
1732 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
1734 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1737 static void __cpuinit
rcu_online_cpu(int cpu
)
1739 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
1740 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
1741 rcu_preempt_init_percpu_data(cpu
);
1745 * Handle CPU online/offline notification events.
1747 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
1748 unsigned long action
, void *hcpu
)
1750 long cpu
= (long)hcpu
;
1753 case CPU_UP_PREPARE
:
1754 case CPU_UP_PREPARE_FROZEN
:
1755 rcu_online_cpu(cpu
);
1758 case CPU_DYING_FROZEN
:
1760 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1761 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1762 * returns, all online cpus have queued rcu_barrier_func().
1763 * The dying CPU clears its cpu_online_mask bit and
1764 * moves all of its RCU callbacks to ->orphan_cbs_list
1765 * in the context of stop_machine(), so subsequent calls
1766 * to _rcu_barrier() will adopt these callbacks and only
1767 * then queue rcu_barrier_func() on all remaining CPUs.
1769 rcu_send_cbs_to_orphanage(&rcu_bh_state
);
1770 rcu_send_cbs_to_orphanage(&rcu_sched_state
);
1771 rcu_preempt_send_cbs_to_orphanage();
1774 case CPU_DEAD_FROZEN
:
1775 case CPU_UP_CANCELED
:
1776 case CPU_UP_CANCELED_FROZEN
:
1777 rcu_offline_cpu(cpu
);
1786 * Compute the per-level fanout, either using the exact fanout specified
1787 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1789 #ifdef CONFIG_RCU_FANOUT_EXACT
1790 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1794 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--)
1795 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
1797 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1798 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1805 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
1806 ccur
= rsp
->levelcnt
[i
];
1807 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
1811 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1814 * Helper function for rcu_init() that initializes one rcu_state structure.
1816 static void __init
rcu_init_one(struct rcu_state
*rsp
)
1818 static char *buf
[] = { "rcu_node_level_0",
1821 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
1825 struct rcu_node
*rnp
;
1827 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
1829 /* Initialize the level-tracking arrays. */
1831 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
1832 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
1833 rcu_init_levelspread(rsp
);
1835 /* Initialize the elements themselves, starting from the leaves. */
1837 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
1838 cpustride
*= rsp
->levelspread
[i
];
1839 rnp
= rsp
->level
[i
];
1840 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
1841 raw_spin_lock_init(&rnp
->lock
);
1842 lockdep_set_class_and_name(&rnp
->lock
,
1843 &rcu_node_class
[i
], buf
[i
]);
1846 rnp
->qsmaskinit
= 0;
1847 rnp
->grplo
= j
* cpustride
;
1848 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
1849 if (rnp
->grphi
>= NR_CPUS
)
1850 rnp
->grphi
= NR_CPUS
- 1;
1856 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
1857 rnp
->grpmask
= 1UL << rnp
->grpnum
;
1858 rnp
->parent
= rsp
->level
[i
- 1] +
1859 j
/ rsp
->levelspread
[i
- 1];
1862 INIT_LIST_HEAD(&rnp
->blocked_tasks
[0]);
1863 INIT_LIST_HEAD(&rnp
->blocked_tasks
[1]);
1864 INIT_LIST_HEAD(&rnp
->blocked_tasks
[2]);
1865 INIT_LIST_HEAD(&rnp
->blocked_tasks
[3]);
1871 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1872 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1875 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1879 struct rcu_node *rnp; \
1881 rcu_init_one(rsp); \
1882 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1884 for_each_possible_cpu(i) { \
1885 if (i > rnp[j].grphi) \
1887 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1888 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1889 rcu_boot_init_percpu_data(i, rsp); \
1893 void __init
rcu_init(void)
1897 rcu_bootup_announce();
1898 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1899 printk(KERN_INFO
"RCU-based detection of stalled CPUs is enabled.\n");
1900 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1901 #if NUM_RCU_LVL_4 != 0
1902 printk(KERN_INFO
"Experimental four-level hierarchy is enabled.\n");
1903 #endif /* #if NUM_RCU_LVL_4 != 0 */
1904 RCU_INIT_FLAVOR(&rcu_sched_state
, rcu_sched_data
);
1905 RCU_INIT_FLAVOR(&rcu_bh_state
, rcu_bh_data
);
1906 __rcu_init_preempt();
1907 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
1910 * We don't need protection against CPU-hotplug here because
1911 * this is called early in boot, before either interrupts
1912 * or the scheduler are operational.
1914 cpu_notifier(rcu_cpu_notify
, 0);
1915 for_each_online_cpu(cpu
)
1916 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1919 #include "rcutree_plugin.h"