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. */ \
63 NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \
65 .signaled = RCU_GP_IDLE, \
68 .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \
69 .orphan_cbs_list = NULL, \
70 .orphan_cbs_tail = &name.orphan_cbs_list, \
72 .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \
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
);
83 static int rcu_scheduler_active __read_mostly
;
87 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
88 * permit this function to be invoked without holding the root rcu_node
89 * structure's ->lock, but of course results can be subject to change.
91 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
93 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
97 * Note a quiescent state. Because we do not need to know
98 * how many quiescent states passed, just if there was at least
99 * one since the start of the grace period, this just sets a flag.
101 void rcu_sched_qs(int cpu
)
103 struct rcu_data
*rdp
;
105 rdp
= &per_cpu(rcu_sched_data
, cpu
);
106 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
108 rdp
->passed_quiesc
= 1;
109 rcu_preempt_note_context_switch(cpu
);
112 void rcu_bh_qs(int cpu
)
114 struct rcu_data
*rdp
;
116 rdp
= &per_cpu(rcu_bh_data
, cpu
);
117 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
119 rdp
->passed_quiesc
= 1;
123 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
124 .dynticks_nesting
= 1,
127 #endif /* #ifdef CONFIG_NO_HZ */
129 static int blimit
= 10; /* Maximum callbacks per softirq. */
130 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
131 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
133 module_param(blimit
, int, 0);
134 module_param(qhimark
, int, 0);
135 module_param(qlowmark
, int, 0);
137 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
138 static int rcu_pending(int cpu
);
141 * Return the number of RCU-sched batches processed thus far for debug & stats.
143 long rcu_batches_completed_sched(void)
145 return rcu_sched_state
.completed
;
147 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
150 * Return the number of RCU BH batches processed thus far for debug & stats.
152 long rcu_batches_completed_bh(void)
154 return rcu_bh_state
.completed
;
156 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
159 * Does the CPU have callbacks ready to be invoked?
162 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
164 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
168 * Does the current CPU require a yet-as-unscheduled grace period?
171 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
173 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
177 * Return the root node of the specified rcu_state structure.
179 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
181 return &rsp
->node
[0];
187 * If the specified CPU is offline, tell the caller that it is in
188 * a quiescent state. Otherwise, whack it with a reschedule IPI.
189 * Grace periods can end up waiting on an offline CPU when that
190 * CPU is in the process of coming online -- it will be added to the
191 * rcu_node bitmasks before it actually makes it online. The same thing
192 * can happen while a CPU is in the process of coming online. Because this
193 * race is quite rare, we check for it after detecting that the grace
194 * period has been delayed rather than checking each and every CPU
195 * each and every time we start a new grace period.
197 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
200 * If the CPU is offline, it is in a quiescent state. We can
201 * trust its state not to change because interrupts are disabled.
203 if (cpu_is_offline(rdp
->cpu
)) {
208 /* If preemptable RCU, no point in sending reschedule IPI. */
209 if (rdp
->preemptable
)
212 /* The CPU is online, so send it a reschedule IPI. */
213 if (rdp
->cpu
!= smp_processor_id())
214 smp_send_reschedule(rdp
->cpu
);
221 #endif /* #ifdef CONFIG_SMP */
226 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
228 * Enter nohz mode, in other words, -leave- the mode in which RCU
229 * read-side critical sections can occur. (Though RCU read-side
230 * critical sections can occur in irq handlers in nohz mode, a possibility
231 * handled by rcu_irq_enter() and rcu_irq_exit()).
233 void rcu_enter_nohz(void)
236 struct rcu_dynticks
*rdtp
;
238 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
239 local_irq_save(flags
);
240 rdtp
= &__get_cpu_var(rcu_dynticks
);
242 rdtp
->dynticks_nesting
--;
243 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
244 local_irq_restore(flags
);
248 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
250 * Exit nohz mode, in other words, -enter- the mode in which RCU
251 * read-side critical sections normally occur.
253 void rcu_exit_nohz(void)
256 struct rcu_dynticks
*rdtp
;
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
);
264 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
268 * rcu_nmi_enter - inform RCU of entry to NMI context
270 * If the CPU was idle with dynamic ticks active, and there is no
271 * irq handler running, this updates rdtp->dynticks_nmi to let the
272 * RCU grace-period handling know that the CPU is active.
274 void rcu_nmi_enter(void)
276 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
278 if (rdtp
->dynticks
& 0x1)
280 rdtp
->dynticks_nmi
++;
281 WARN_ON_ONCE(!(rdtp
->dynticks_nmi
& 0x1));
282 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
286 * rcu_nmi_exit - inform RCU of exit from NMI context
288 * If the CPU was idle with dynamic ticks active, and there is no
289 * irq handler running, this updates rdtp->dynticks_nmi to let the
290 * RCU grace-period handling know that the CPU is no longer active.
292 void rcu_nmi_exit(void)
294 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
296 if (rdtp
->dynticks
& 0x1)
298 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
299 rdtp
->dynticks_nmi
++;
300 WARN_ON_ONCE(rdtp
->dynticks_nmi
& 0x1);
304 * rcu_irq_enter - inform RCU of entry to hard irq context
306 * If the CPU was idle with dynamic ticks active, this updates the
307 * rdtp->dynticks to let the RCU handling know that the CPU is active.
309 void rcu_irq_enter(void)
311 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
313 if (rdtp
->dynticks_nesting
++)
316 WARN_ON_ONCE(!(rdtp
->dynticks
& 0x1));
317 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
321 * rcu_irq_exit - inform RCU of exit from hard irq context
323 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
324 * to put let the RCU handling be aware that the CPU is going back to idle
327 void rcu_irq_exit(void)
329 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
331 if (--rdtp
->dynticks_nesting
)
333 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
335 WARN_ON_ONCE(rdtp
->dynticks
& 0x1);
337 /* If the interrupt queued a callback, get out of dyntick mode. */
338 if (__get_cpu_var(rcu_sched_data
).nxtlist
||
339 __get_cpu_var(rcu_bh_data
).nxtlist
)
346 * Snapshot the specified CPU's dynticks counter so that we can later
347 * credit them with an implicit quiescent state. Return 1 if this CPU
348 * is in dynticks idle mode, which is an extended quiescent state.
350 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
356 snap
= rdp
->dynticks
->dynticks
;
357 snap_nmi
= rdp
->dynticks
->dynticks_nmi
;
358 smp_mb(); /* Order sampling of snap with end of grace period. */
359 rdp
->dynticks_snap
= snap
;
360 rdp
->dynticks_nmi_snap
= snap_nmi
;
361 ret
= ((snap
& 0x1) == 0) && ((snap_nmi
& 0x1) == 0);
368 * Return true if the specified CPU has passed through a quiescent
369 * state by virtue of being in or having passed through an dynticks
370 * idle state since the last call to dyntick_save_progress_counter()
373 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
380 curr
= rdp
->dynticks
->dynticks
;
381 snap
= rdp
->dynticks_snap
;
382 curr_nmi
= rdp
->dynticks
->dynticks_nmi
;
383 snap_nmi
= rdp
->dynticks_nmi_snap
;
384 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
387 * If the CPU passed through or entered a dynticks idle phase with
388 * no active irq/NMI handlers, then we can safely pretend that the CPU
389 * already acknowledged the request to pass through a quiescent
390 * state. Either way, that CPU cannot possibly be in an RCU
391 * read-side critical section that started before the beginning
392 * of the current RCU grace period.
394 if ((curr
!= snap
|| (curr
& 0x1) == 0) &&
395 (curr_nmi
!= snap_nmi
|| (curr_nmi
& 0x1) == 0)) {
400 /* Go check for the CPU being offline. */
401 return rcu_implicit_offline_qs(rdp
);
404 #endif /* #ifdef CONFIG_SMP */
406 #else /* #ifdef CONFIG_NO_HZ */
410 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
415 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
417 return rcu_implicit_offline_qs(rdp
);
420 #endif /* #ifdef CONFIG_SMP */
422 #endif /* #else #ifdef CONFIG_NO_HZ */
424 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
426 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
428 rsp
->gp_start
= jiffies
;
429 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
432 static void print_other_cpu_stall(struct rcu_state
*rsp
)
437 struct rcu_node
*rnp
= rcu_get_root(rsp
);
439 /* Only let one CPU complain about others per time interval. */
441 spin_lock_irqsave(&rnp
->lock
, flags
);
442 delta
= jiffies
- rsp
->jiffies_stall
;
443 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
444 spin_unlock_irqrestore(&rnp
->lock
, flags
);
447 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
450 * Now rat on any tasks that got kicked up to the root rcu_node
451 * due to CPU offlining.
453 rcu_print_task_stall(rnp
);
454 spin_unlock_irqrestore(&rnp
->lock
, flags
);
456 /* OK, time to rat on our buddy... */
458 printk(KERN_ERR
"INFO: RCU detected CPU stalls:");
459 rcu_for_each_leaf_node(rsp
, rnp
) {
460 rcu_print_task_stall(rnp
);
461 if (rnp
->qsmask
== 0)
463 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
464 if (rnp
->qsmask
& (1UL << cpu
))
465 printk(" %d", rnp
->grplo
+ cpu
);
467 printk(" (detected by %d, t=%ld jiffies)\n",
468 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
469 trigger_all_cpu_backtrace();
471 force_quiescent_state(rsp
, 0); /* Kick them all. */
474 static void print_cpu_stall(struct rcu_state
*rsp
)
477 struct rcu_node
*rnp
= rcu_get_root(rsp
);
479 printk(KERN_ERR
"INFO: RCU detected CPU %d stall (t=%lu jiffies)\n",
480 smp_processor_id(), jiffies
- rsp
->gp_start
);
481 trigger_all_cpu_backtrace();
483 spin_lock_irqsave(&rnp
->lock
, flags
);
484 if ((long)(jiffies
- rsp
->jiffies_stall
) >= 0)
486 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
487 spin_unlock_irqrestore(&rnp
->lock
, flags
);
489 set_need_resched(); /* kick ourselves to get things going. */
492 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
495 struct rcu_node
*rnp
;
497 delta
= jiffies
- rsp
->jiffies_stall
;
499 if ((rnp
->qsmask
& rdp
->grpmask
) && delta
>= 0) {
501 /* We haven't checked in, so go dump stack. */
502 print_cpu_stall(rsp
);
504 } else if (rcu_gp_in_progress(rsp
) && delta
>= RCU_STALL_RAT_DELAY
) {
506 /* They had two time units to dump stack, so complain. */
507 print_other_cpu_stall(rsp
);
511 #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
513 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
517 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
521 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
524 * Update CPU-local rcu_data state to record the newly noticed grace period.
525 * This is used both when we started the grace period and when we notice
526 * that someone else started the grace period. The caller must hold the
527 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
528 * and must have irqs disabled.
530 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
532 if (rdp
->gpnum
!= rnp
->gpnum
) {
534 rdp
->passed_quiesc
= 0;
535 rdp
->gpnum
= rnp
->gpnum
;
539 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
542 struct rcu_node
*rnp
;
544 local_irq_save(flags
);
546 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
547 !spin_trylock(&rnp
->lock
)) { /* irqs already off, retry later. */
548 local_irq_restore(flags
);
551 __note_new_gpnum(rsp
, rnp
, rdp
);
552 spin_unlock_irqrestore(&rnp
->lock
, flags
);
556 * Did someone else start a new RCU grace period start since we last
557 * checked? Update local state appropriately if so. Must be called
558 * on the CPU corresponding to rdp.
561 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
566 local_irq_save(flags
);
567 if (rdp
->gpnum
!= rsp
->gpnum
) {
568 note_new_gpnum(rsp
, rdp
);
571 local_irq_restore(flags
);
576 * Advance this CPU's callbacks, but only if the current grace period
577 * has ended. This may be called only from the CPU to whom the rdp
578 * belongs. In addition, the corresponding leaf rcu_node structure's
579 * ->lock must be held by the caller, with irqs disabled.
582 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
584 /* Did another grace period end? */
585 if (rdp
->completed
!= rnp
->completed
) {
587 /* Advance callbacks. No harm if list empty. */
588 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
589 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
590 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
592 /* Remember that we saw this grace-period completion. */
593 rdp
->completed
= rnp
->completed
;
598 * Advance this CPU's callbacks, but only if the current grace period
599 * has ended. This may be called only from the CPU to whom the rdp
603 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
606 struct rcu_node
*rnp
;
608 local_irq_save(flags
);
610 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
611 !spin_trylock(&rnp
->lock
)) { /* irqs already off, retry later. */
612 local_irq_restore(flags
);
615 __rcu_process_gp_end(rsp
, rnp
, rdp
);
616 spin_unlock_irqrestore(&rnp
->lock
, flags
);
620 * Do per-CPU grace-period initialization for running CPU. The caller
621 * must hold the lock of the leaf rcu_node structure corresponding to
625 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
627 /* Prior grace period ended, so advance callbacks for current CPU. */
628 __rcu_process_gp_end(rsp
, rnp
, rdp
);
631 * Because this CPU just now started the new grace period, we know
632 * that all of its callbacks will be covered by this upcoming grace
633 * period, even the ones that were registered arbitrarily recently.
634 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
636 * Other CPUs cannot be sure exactly when the grace period started.
637 * Therefore, their recently registered callbacks must pass through
638 * an additional RCU_NEXT_READY stage, so that they will be handled
639 * by the next RCU grace period.
641 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
642 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
644 /* Set state so that this CPU will detect the next quiescent state. */
645 __note_new_gpnum(rsp
, rnp
, rdp
);
649 * Start a new RCU grace period if warranted, re-initializing the hierarchy
650 * in preparation for detecting the next grace period. The caller must hold
651 * the root node's ->lock, which is released before return. Hard irqs must
655 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
656 __releases(rcu_get_root(rsp
)->lock
)
658 struct rcu_data
*rdp
= rsp
->rda
[smp_processor_id()];
659 struct rcu_node
*rnp
= rcu_get_root(rsp
);
661 if (!cpu_needs_another_gp(rsp
, rdp
)) {
662 if (rnp
->completed
== rsp
->completed
) {
663 spin_unlock_irqrestore(&rnp
->lock
, flags
);
666 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
669 * Propagate new ->completed value to rcu_node structures
670 * so that other CPUs don't have to wait until the start
671 * of the next grace period to process their callbacks.
673 rcu_for_each_node_breadth_first(rsp
, rnp
) {
674 spin_lock(&rnp
->lock
); /* irqs already disabled. */
675 rnp
->completed
= rsp
->completed
;
676 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
678 local_irq_restore(flags
);
682 /* Advance to a new grace period and initialize state. */
684 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
685 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
686 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
687 record_gp_stall_check_time(rsp
);
689 /* Special-case the common single-level case. */
690 if (NUM_RCU_NODES
== 1) {
691 rcu_preempt_check_blocked_tasks(rnp
);
692 rnp
->qsmask
= rnp
->qsmaskinit
;
693 rnp
->gpnum
= rsp
->gpnum
;
694 rnp
->completed
= rsp
->completed
;
695 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
696 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
697 spin_unlock_irqrestore(&rnp
->lock
, flags
);
701 spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
704 /* Exclude any concurrent CPU-hotplug operations. */
705 spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
708 * Set the quiescent-state-needed bits in all the rcu_node
709 * structures for all currently online CPUs in breadth-first
710 * order, starting from the root rcu_node structure. This
711 * operation relies on the layout of the hierarchy within the
712 * rsp->node[] array. Note that other CPUs will access only
713 * the leaves of the hierarchy, which still indicate that no
714 * grace period is in progress, at least until the corresponding
715 * leaf node has been initialized. In addition, we have excluded
716 * CPU-hotplug operations.
718 * Note that the grace period cannot complete until we finish
719 * the initialization process, as there will be at least one
720 * qsmask bit set in the root node until that time, namely the
721 * one corresponding to this CPU, due to the fact that we have
724 rcu_for_each_node_breadth_first(rsp
, rnp
) {
725 spin_lock(&rnp
->lock
); /* irqs already disabled. */
726 rcu_preempt_check_blocked_tasks(rnp
);
727 rnp
->qsmask
= rnp
->qsmaskinit
;
728 rnp
->gpnum
= rsp
->gpnum
;
729 rnp
->completed
= rsp
->completed
;
730 if (rnp
== rdp
->mynode
)
731 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
732 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
735 rnp
= rcu_get_root(rsp
);
736 spin_lock(&rnp
->lock
); /* irqs already disabled. */
737 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
738 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
739 spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
743 * Clean up after the prior grace period and let rcu_start_gp() start up
744 * the next grace period if one is needed. Note that the caller must
745 * hold rnp->lock, as required by rcu_start_gp(), which will release it.
747 static void cpu_quiet_msk_finish(struct rcu_state
*rsp
, unsigned long flags
)
748 __releases(rcu_get_root(rsp
)->lock
)
750 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
751 rsp
->completed
= rsp
->gpnum
;
752 rsp
->signaled
= RCU_GP_IDLE
;
753 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
757 * Similar to cpu_quiet(), for which it is a helper function. Allows
758 * a group of CPUs to be quieted at one go, though all the CPUs in the
759 * group must be represented by the same leaf rcu_node structure.
760 * That structure's lock must be held upon entry, and it is released
764 cpu_quiet_msk(unsigned long mask
, struct rcu_state
*rsp
, struct rcu_node
*rnp
,
766 __releases(rnp
->lock
)
768 struct rcu_node
*rnp_c
;
770 /* Walk up the rcu_node hierarchy. */
772 if (!(rnp
->qsmask
& mask
)) {
774 /* Our bit has already been cleared, so done. */
775 spin_unlock_irqrestore(&rnp
->lock
, flags
);
778 rnp
->qsmask
&= ~mask
;
779 if (rnp
->qsmask
!= 0 || rcu_preempted_readers(rnp
)) {
781 /* Other bits still set at this level, so done. */
782 spin_unlock_irqrestore(&rnp
->lock
, flags
);
786 if (rnp
->parent
== NULL
) {
788 /* No more levels. Exit loop holding root lock. */
792 spin_unlock_irqrestore(&rnp
->lock
, flags
);
795 spin_lock_irqsave(&rnp
->lock
, flags
);
796 WARN_ON_ONCE(rnp_c
->qsmask
);
800 * Get here if we are the last CPU to pass through a quiescent
801 * state for this grace period. Invoke cpu_quiet_msk_finish()
802 * to clean up and start the next grace period if one is needed.
804 cpu_quiet_msk_finish(rsp
, flags
); /* releases rnp->lock. */
808 * Record a quiescent state for the specified CPU, which must either be
809 * the current CPU. The lastcomp argument is used to make sure we are
810 * still in the grace period of interest. We don't want to end the current
811 * grace period based on quiescent states detected in an earlier grace
815 cpu_quiet(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
819 struct rcu_node
*rnp
;
822 spin_lock_irqsave(&rnp
->lock
, flags
);
823 if (lastcomp
!= rnp
->completed
) {
826 * Someone beat us to it for this grace period, so leave.
827 * The race with GP start is resolved by the fact that we
828 * hold the leaf rcu_node lock, so that the per-CPU bits
829 * cannot yet be initialized -- so we would simply find our
830 * CPU's bit already cleared in cpu_quiet_msk() if this race
833 rdp
->passed_quiesc
= 0; /* try again later! */
834 spin_unlock_irqrestore(&rnp
->lock
, flags
);
838 if ((rnp
->qsmask
& mask
) == 0) {
839 spin_unlock_irqrestore(&rnp
->lock
, flags
);
844 * This GP can't end until cpu checks in, so all of our
845 * callbacks can be processed during the next GP.
847 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
849 cpu_quiet_msk(mask
, rsp
, rnp
, flags
); /* releases rnp->lock */
854 * Check to see if there is a new grace period of which this CPU
855 * is not yet aware, and if so, set up local rcu_data state for it.
856 * Otherwise, see if this CPU has just passed through its first
857 * quiescent state for this grace period, and record that fact if so.
860 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
862 /* If there is now a new grace period, record and return. */
863 if (check_for_new_grace_period(rsp
, rdp
))
867 * Does this CPU still need to do its part for current grace period?
868 * If no, return and let the other CPUs do their part as well.
870 if (!rdp
->qs_pending
)
874 * Was there a quiescent state since the beginning of the grace
875 * period? If no, then exit and wait for the next call.
877 if (!rdp
->passed_quiesc
)
880 /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */
881 cpu_quiet(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
884 #ifdef CONFIG_HOTPLUG_CPU
887 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
888 * specified flavor of RCU. The callbacks will be adopted by the next
889 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
890 * comes first. Because this is invoked from the CPU_DYING notifier,
891 * irqs are already disabled.
893 static void rcu_send_cbs_to_orphanage(struct rcu_state
*rsp
)
896 struct rcu_data
*rdp
= rsp
->rda
[smp_processor_id()];
898 if (rdp
->nxtlist
== NULL
)
899 return; /* irqs disabled, so comparison is stable. */
900 spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
901 *rsp
->orphan_cbs_tail
= rdp
->nxtlist
;
902 rsp
->orphan_cbs_tail
= rdp
->nxttail
[RCU_NEXT_TAIL
];
904 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
905 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
906 rsp
->orphan_qlen
+= rdp
->qlen
;
908 spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
912 * Adopt previously orphaned RCU callbacks.
914 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
917 struct rcu_data
*rdp
;
919 spin_lock_irqsave(&rsp
->onofflock
, flags
);
920 rdp
= rsp
->rda
[smp_processor_id()];
921 if (rsp
->orphan_cbs_list
== NULL
) {
922 spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
925 *rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_cbs_list
;
926 rdp
->nxttail
[RCU_NEXT_TAIL
] = rsp
->orphan_cbs_tail
;
927 rdp
->qlen
+= rsp
->orphan_qlen
;
928 rsp
->orphan_cbs_list
= NULL
;
929 rsp
->orphan_cbs_tail
= &rsp
->orphan_cbs_list
;
930 rsp
->orphan_qlen
= 0;
931 spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
935 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
936 * and move all callbacks from the outgoing CPU to the current one.
938 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
943 struct rcu_data
*rdp
= rsp
->rda
[cpu
];
944 struct rcu_node
*rnp
;
946 /* Exclude any attempts to start a new grace period. */
947 spin_lock_irqsave(&rsp
->onofflock
, flags
);
949 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
950 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
951 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
953 spin_lock(&rnp
->lock
); /* irqs already disabled. */
954 rnp
->qsmaskinit
&= ~mask
;
955 if (rnp
->qsmaskinit
!= 0) {
956 if (rnp
!= rdp
->mynode
)
957 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
960 if (rnp
== rdp
->mynode
)
961 need_quiet
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
963 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
966 } while (rnp
!= NULL
);
969 * We still hold the leaf rcu_node structure lock here, and
970 * irqs are still disabled. The reason for this subterfuge is
971 * because invoking task_quiet() with ->onofflock held leads
974 spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
977 task_quiet(rnp
, flags
);
979 spin_unlock_irqrestore(&rnp
->lock
, flags
);
981 rcu_adopt_orphan_cbs(rsp
);
985 * Remove the specified CPU from the RCU hierarchy and move any pending
986 * callbacks that it might have to the current CPU. This code assumes
987 * that at least one CPU in the system will remain running at all times.
988 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
990 static void rcu_offline_cpu(int cpu
)
992 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
993 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
994 rcu_preempt_offline_cpu(cpu
);
997 #else /* #ifdef CONFIG_HOTPLUG_CPU */
999 static void rcu_send_cbs_to_orphanage(struct rcu_state
*rsp
)
1003 static void rcu_adopt_orphan_cbs(struct rcu_state
*rsp
)
1007 static void rcu_offline_cpu(int cpu
)
1011 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1014 * Invoke any RCU callbacks that have made it to the end of their grace
1015 * period. Thottle as specified by rdp->blimit.
1017 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1019 unsigned long flags
;
1020 struct rcu_head
*next
, *list
, **tail
;
1023 /* If no callbacks are ready, just return.*/
1024 if (!cpu_has_callbacks_ready_to_invoke(rdp
))
1028 * Extract the list of ready callbacks, disabling to prevent
1029 * races with call_rcu() from interrupt handlers.
1031 local_irq_save(flags
);
1032 list
= rdp
->nxtlist
;
1033 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1034 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1035 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1036 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1037 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1038 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1039 local_irq_restore(flags
);
1041 /* Invoke callbacks. */
1048 if (++count
>= rdp
->blimit
)
1052 local_irq_save(flags
);
1054 /* Update count, and requeue any remaining callbacks. */
1057 *tail
= rdp
->nxtlist
;
1058 rdp
->nxtlist
= list
;
1059 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1060 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1061 rdp
->nxttail
[count
] = tail
;
1066 /* Reinstate batch limit if we have worked down the excess. */
1067 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1068 rdp
->blimit
= blimit
;
1070 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1071 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1072 rdp
->qlen_last_fqs_check
= 0;
1073 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1074 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1075 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1077 local_irq_restore(flags
);
1079 /* Re-raise the RCU softirq if there are callbacks remaining. */
1080 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1081 raise_softirq(RCU_SOFTIRQ
);
1085 * Check to see if this CPU is in a non-context-switch quiescent state
1086 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1087 * Also schedule the RCU softirq handler.
1089 * This function must be called with hardirqs disabled. It is normally
1090 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1091 * false, there is no point in invoking rcu_check_callbacks().
1093 void rcu_check_callbacks(int cpu
, int user
)
1095 if (!rcu_pending(cpu
))
1096 return; /* if nothing for RCU to do. */
1098 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1099 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1102 * Get here if this CPU took its interrupt from user
1103 * mode or from the idle loop, and if this is not a
1104 * nested interrupt. In this case, the CPU is in
1105 * a quiescent state, so note it.
1107 * No memory barrier is required here because both
1108 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1109 * variables that other CPUs neither access nor modify,
1110 * at least not while the corresponding CPU is online.
1116 } else if (!in_softirq()) {
1119 * Get here if this CPU did not take its interrupt from
1120 * softirq, in other words, if it is not interrupting
1121 * a rcu_bh read-side critical section. This is an _bh
1122 * critical section, so note it.
1127 rcu_preempt_check_callbacks(cpu
);
1128 raise_softirq(RCU_SOFTIRQ
);
1134 * Scan the leaf rcu_node structures, processing dyntick state for any that
1135 * have not yet encountered a quiescent state, using the function specified.
1136 * Returns 1 if the current grace period ends while scanning (possibly
1137 * because we made it end).
1139 static int rcu_process_dyntick(struct rcu_state
*rsp
, long lastcomp
,
1140 int (*f
)(struct rcu_data
*))
1144 unsigned long flags
;
1146 struct rcu_node
*rnp
;
1148 rcu_for_each_leaf_node(rsp
, rnp
) {
1150 spin_lock_irqsave(&rnp
->lock
, flags
);
1151 if (rnp
->completed
!= lastcomp
) {
1152 spin_unlock_irqrestore(&rnp
->lock
, flags
);
1155 if (rnp
->qsmask
== 0) {
1156 spin_unlock_irqrestore(&rnp
->lock
, flags
);
1161 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1162 if ((rnp
->qsmask
& bit
) != 0 && f(rsp
->rda
[cpu
]))
1165 if (mask
!= 0 && rnp
->completed
== lastcomp
) {
1167 /* cpu_quiet_msk() releases rnp->lock. */
1168 cpu_quiet_msk(mask
, rsp
, rnp
, flags
);
1171 spin_unlock_irqrestore(&rnp
->lock
, flags
);
1177 * Force quiescent states on reluctant CPUs, and also detect which
1178 * CPUs are in dyntick-idle mode.
1180 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1182 unsigned long flags
;
1184 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1188 if (!rcu_gp_in_progress(rsp
))
1189 return; /* No grace period in progress, nothing to force. */
1190 if (!spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1191 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1192 return; /* Someone else is already on the job. */
1195 (long)(rsp
->jiffies_force_qs
- jiffies
) >= 0)
1196 goto unlock_ret
; /* no emergency and done recently. */
1198 spin_lock(&rnp
->lock
);
1199 lastcomp
= rsp
->gpnum
- 1;
1200 signaled
= rsp
->signaled
;
1201 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1202 if(!rcu_gp_in_progress(rsp
)) {
1203 rsp
->n_force_qs_ngp
++;
1204 spin_unlock(&rnp
->lock
);
1205 goto unlock_ret
; /* no GP in progress, time updated. */
1207 spin_unlock(&rnp
->lock
);
1212 break; /* grace period idle or initializing, ignore. */
1214 case RCU_SAVE_DYNTICK
:
1216 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1217 break; /* So gcc recognizes the dead code. */
1219 /* Record dyntick-idle state. */
1220 if (rcu_process_dyntick(rsp
, lastcomp
,
1221 dyntick_save_progress_counter
))
1223 /* fall into next case. */
1225 case RCU_SAVE_COMPLETED
:
1227 /* Update state, record completion counter. */
1229 spin_lock(&rnp
->lock
);
1230 if (lastcomp
+ 1 == rsp
->gpnum
&&
1231 lastcomp
== rsp
->completed
&&
1232 rsp
->signaled
== signaled
) {
1233 rsp
->signaled
= RCU_FORCE_QS
;
1234 rsp
->completed_fqs
= lastcomp
;
1235 forcenow
= signaled
== RCU_SAVE_COMPLETED
;
1237 spin_unlock(&rnp
->lock
);
1240 /* fall into next case. */
1244 /* Check dyntick-idle state, send IPI to laggarts. */
1245 if (rcu_process_dyntick(rsp
, rsp
->completed_fqs
,
1246 rcu_implicit_dynticks_qs
))
1249 /* Leave state in case more forcing is required. */
1254 spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1257 #else /* #ifdef CONFIG_SMP */
1259 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1264 #endif /* #else #ifdef CONFIG_SMP */
1267 * This does the RCU processing work from softirq context for the
1268 * specified rcu_state and rcu_data structures. This may be called
1269 * only from the CPU to whom the rdp belongs.
1272 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1274 unsigned long flags
;
1276 WARN_ON_ONCE(rdp
->beenonline
== 0);
1279 * If an RCU GP has gone long enough, go check for dyntick
1280 * idle CPUs and, if needed, send resched IPIs.
1282 if ((long)(ACCESS_ONCE(rsp
->jiffies_force_qs
) - jiffies
) < 0)
1283 force_quiescent_state(rsp
, 1);
1286 * Advance callbacks in response to end of earlier grace
1287 * period that some other CPU ended.
1289 rcu_process_gp_end(rsp
, rdp
);
1291 /* Update RCU state based on any recent quiescent states. */
1292 rcu_check_quiescent_state(rsp
, rdp
);
1294 /* Does this CPU require a not-yet-started grace period? */
1295 if (cpu_needs_another_gp(rsp
, rdp
)) {
1296 spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1297 rcu_start_gp(rsp
, flags
); /* releases above lock */
1300 /* If there are callbacks ready, invoke them. */
1301 rcu_do_batch(rsp
, rdp
);
1305 * Do softirq processing for the current CPU.
1307 static void rcu_process_callbacks(struct softirq_action
*unused
)
1310 * Memory references from any prior RCU read-side critical sections
1311 * executed by the interrupted code must be seen before any RCU
1312 * grace-period manipulations below.
1314 smp_mb(); /* See above block comment. */
1316 __rcu_process_callbacks(&rcu_sched_state
,
1317 &__get_cpu_var(rcu_sched_data
));
1318 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1319 rcu_preempt_process_callbacks();
1322 * Memory references from any later RCU read-side critical sections
1323 * executed by the interrupted code must be seen after any RCU
1324 * grace-period manipulations above.
1326 smp_mb(); /* See above block comment. */
1330 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1331 struct rcu_state
*rsp
)
1333 unsigned long flags
;
1334 struct rcu_data
*rdp
;
1339 smp_mb(); /* Ensure RCU update seen before callback registry. */
1342 * Opportunistically note grace-period endings and beginnings.
1343 * Note that we might see a beginning right after we see an
1344 * end, but never vice versa, since this CPU has to pass through
1345 * a quiescent state betweentimes.
1347 local_irq_save(flags
);
1348 rdp
= rsp
->rda
[smp_processor_id()];
1349 rcu_process_gp_end(rsp
, rdp
);
1350 check_for_new_grace_period(rsp
, rdp
);
1352 /* Add the callback to our list. */
1353 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1354 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1356 /* Start a new grace period if one not already started. */
1357 if (!rcu_gp_in_progress(rsp
)) {
1358 unsigned long nestflag
;
1359 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1361 spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1362 rcu_start_gp(rsp
, nestflag
); /* releases rnp_root->lock. */
1366 * Force the grace period if too many callbacks or too long waiting.
1367 * Enforce hysteresis, and don't invoke force_quiescent_state()
1368 * if some other CPU has recently done so. Also, don't bother
1369 * invoking force_quiescent_state() if the newly enqueued callback
1370 * is the only one waiting for a grace period to complete.
1372 if (unlikely(++rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1373 rdp
->blimit
= LONG_MAX
;
1374 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1375 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1376 force_quiescent_state(rsp
, 0);
1377 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1378 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1379 } else if ((long)(ACCESS_ONCE(rsp
->jiffies_force_qs
) - jiffies
) < 0)
1380 force_quiescent_state(rsp
, 1);
1381 local_irq_restore(flags
);
1385 * Queue an RCU-sched callback for invocation after a grace period.
1387 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1389 __call_rcu(head
, func
, &rcu_sched_state
);
1391 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1394 * Queue an RCU for invocation after a quicker grace period.
1396 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1398 __call_rcu(head
, func
, &rcu_bh_state
);
1400 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1403 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1405 * Control will return to the caller some time after a full rcu-sched
1406 * grace period has elapsed, in other words after all currently executing
1407 * rcu-sched read-side critical sections have completed. These read-side
1408 * critical sections are delimited by rcu_read_lock_sched() and
1409 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1410 * local_irq_disable(), and so on may be used in place of
1411 * rcu_read_lock_sched().
1413 * This means that all preempt_disable code sequences, including NMI and
1414 * hardware-interrupt handlers, in progress on entry will have completed
1415 * before this primitive returns. However, this does not guarantee that
1416 * softirq handlers will have completed, since in some kernels, these
1417 * handlers can run in process context, and can block.
1419 * This primitive provides the guarantees made by the (now removed)
1420 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1421 * guarantees that rcu_read_lock() sections will have completed.
1422 * In "classic RCU", these two guarantees happen to be one and
1423 * the same, but can differ in realtime RCU implementations.
1425 void synchronize_sched(void)
1427 struct rcu_synchronize rcu
;
1429 if (rcu_blocking_is_gp())
1432 init_completion(&rcu
.completion
);
1433 /* Will wake me after RCU finished. */
1434 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1436 wait_for_completion(&rcu
.completion
);
1438 EXPORT_SYMBOL_GPL(synchronize_sched
);
1441 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1443 * Control will return to the caller some time after a full rcu_bh grace
1444 * period has elapsed, in other words after all currently executing rcu_bh
1445 * read-side critical sections have completed. RCU read-side critical
1446 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1447 * and may be nested.
1449 void synchronize_rcu_bh(void)
1451 struct rcu_synchronize rcu
;
1453 if (rcu_blocking_is_gp())
1456 init_completion(&rcu
.completion
);
1457 /* Will wake me after RCU finished. */
1458 call_rcu_bh(&rcu
.head
, wakeme_after_rcu
);
1460 wait_for_completion(&rcu
.completion
);
1462 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1465 * Check to see if there is any immediate RCU-related work to be done
1466 * by the current CPU, for the specified type of RCU, returning 1 if so.
1467 * The checks are in order of increasing expense: checks that can be
1468 * carried out against CPU-local state are performed first. However,
1469 * we must check for CPU stalls first, else we might not get a chance.
1471 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1473 struct rcu_node
*rnp
= rdp
->mynode
;
1475 rdp
->n_rcu_pending
++;
1477 /* Check for CPU stalls, if enabled. */
1478 check_cpu_stall(rsp
, rdp
);
1480 /* Is the RCU core waiting for a quiescent state from this CPU? */
1481 if (rdp
->qs_pending
) {
1482 rdp
->n_rp_qs_pending
++;
1486 /* Does this CPU have callbacks ready to invoke? */
1487 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1488 rdp
->n_rp_cb_ready
++;
1492 /* Has RCU gone idle with this CPU needing another grace period? */
1493 if (cpu_needs_another_gp(rsp
, rdp
)) {
1494 rdp
->n_rp_cpu_needs_gp
++;
1498 /* Has another RCU grace period completed? */
1499 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
1500 rdp
->n_rp_gp_completed
++;
1504 /* Has a new RCU grace period started? */
1505 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
1506 rdp
->n_rp_gp_started
++;
1510 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1511 if (rcu_gp_in_progress(rsp
) &&
1512 ((long)(ACCESS_ONCE(rsp
->jiffies_force_qs
) - jiffies
) < 0)) {
1513 rdp
->n_rp_need_fqs
++;
1518 rdp
->n_rp_need_nothing
++;
1523 * Check to see if there is any immediate RCU-related work to be done
1524 * by the current CPU, returning 1 if so. This function is part of the
1525 * RCU implementation; it is -not- an exported member of the RCU API.
1527 static int rcu_pending(int cpu
)
1529 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
1530 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
1531 rcu_preempt_pending(cpu
);
1535 * Check to see if any future RCU-related work will need to be done
1536 * by the current CPU, even if none need be done immediately, returning
1537 * 1 if so. This function is part of the RCU implementation; it is -not-
1538 * an exported member of the RCU API.
1540 int rcu_needs_cpu(int cpu
)
1542 /* RCU callbacks either ready or pending? */
1543 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
1544 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
1545 rcu_preempt_needs_cpu(cpu
);
1549 * This function is invoked towards the end of the scheduler's initialization
1550 * process. Before this is called, the idle task might contain
1551 * RCU read-side critical sections (during which time, this idle
1552 * task is booting the system). After this function is called, the
1553 * idle tasks are prohibited from containing RCU read-side critical
1556 void rcu_scheduler_starting(void)
1558 WARN_ON(num_online_cpus() != 1);
1559 WARN_ON(nr_context_switches() > 0);
1560 rcu_scheduler_active
= 1;
1563 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
1564 static atomic_t rcu_barrier_cpu_count
;
1565 static DEFINE_MUTEX(rcu_barrier_mutex
);
1566 static struct completion rcu_barrier_completion
;
1568 static void rcu_barrier_callback(struct rcu_head
*notused
)
1570 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1571 complete(&rcu_barrier_completion
);
1575 * Called with preemption disabled, and from cross-cpu IRQ context.
1577 static void rcu_barrier_func(void *type
)
1579 int cpu
= smp_processor_id();
1580 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
1581 void (*call_rcu_func
)(struct rcu_head
*head
,
1582 void (*func
)(struct rcu_head
*head
));
1584 atomic_inc(&rcu_barrier_cpu_count
);
1585 call_rcu_func
= type
;
1586 call_rcu_func(head
, rcu_barrier_callback
);
1590 * Orchestrate the specified type of RCU barrier, waiting for all
1591 * RCU callbacks of the specified type to complete.
1593 static void _rcu_barrier(struct rcu_state
*rsp
,
1594 void (*call_rcu_func
)(struct rcu_head
*head
,
1595 void (*func
)(struct rcu_head
*head
)))
1597 BUG_ON(in_interrupt());
1598 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1599 mutex_lock(&rcu_barrier_mutex
);
1600 init_completion(&rcu_barrier_completion
);
1602 * Initialize rcu_barrier_cpu_count to 1, then invoke
1603 * rcu_barrier_func() on each CPU, so that each CPU also has
1604 * incremented rcu_barrier_cpu_count. Only then is it safe to
1605 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1606 * might complete its grace period before all of the other CPUs
1607 * did their increment, causing this function to return too
1610 atomic_set(&rcu_barrier_cpu_count
, 1);
1611 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1612 rcu_adopt_orphan_cbs(rsp
);
1613 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
1614 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
1615 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1616 complete(&rcu_barrier_completion
);
1617 wait_for_completion(&rcu_barrier_completion
);
1618 mutex_unlock(&rcu_barrier_mutex
);
1622 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1624 void rcu_barrier_bh(void)
1626 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
1628 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
1631 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1633 void rcu_barrier_sched(void)
1635 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
1637 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
1640 * Do boot-time initialization of a CPU's per-CPU RCU data.
1643 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
1645 unsigned long flags
;
1647 struct rcu_data
*rdp
= rsp
->rda
[cpu
];
1648 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1650 /* Set up local state, ensuring consistent view of global state. */
1651 spin_lock_irqsave(&rnp
->lock
, flags
);
1652 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
1653 rdp
->nxtlist
= NULL
;
1654 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1655 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1658 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
1659 #endif /* #ifdef CONFIG_NO_HZ */
1661 spin_unlock_irqrestore(&rnp
->lock
, flags
);
1665 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1666 * offline event can be happening at a given time. Note also that we
1667 * can accept some slop in the rsp->completed access due to the fact
1668 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1670 static void __cpuinit
1671 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptable
)
1673 unsigned long flags
;
1675 struct rcu_data
*rdp
= rsp
->rda
[cpu
];
1676 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1678 /* Set up local state, ensuring consistent view of global state. */
1679 spin_lock_irqsave(&rnp
->lock
, flags
);
1680 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
1681 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
1682 rdp
->beenonline
= 1; /* We have now been online. */
1683 rdp
->preemptable
= preemptable
;
1684 rdp
->qlen_last_fqs_check
= 0;
1685 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1686 rdp
->blimit
= blimit
;
1687 spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1690 * A new grace period might start here. If so, we won't be part
1691 * of it, but that is OK, as we are currently in a quiescent state.
1694 /* Exclude any attempts to start a new GP on large systems. */
1695 spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
1697 /* Add CPU to rcu_node bitmasks. */
1699 mask
= rdp
->grpmask
;
1701 /* Exclude any attempts to start a new GP on small systems. */
1702 spin_lock(&rnp
->lock
); /* irqs already disabled. */
1703 rnp
->qsmaskinit
|= mask
;
1704 mask
= rnp
->grpmask
;
1705 if (rnp
== rdp
->mynode
) {
1706 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
1707 rdp
->completed
= rnp
->completed
;
1708 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
1710 spin_unlock(&rnp
->lock
); /* irqs already disabled. */
1712 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
1714 spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1717 static void __cpuinit
rcu_online_cpu(int cpu
)
1719 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
1720 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
1721 rcu_preempt_init_percpu_data(cpu
);
1725 * Handle CPU online/offline notification events.
1727 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
1728 unsigned long action
, void *hcpu
)
1730 long cpu
= (long)hcpu
;
1733 case CPU_UP_PREPARE
:
1734 case CPU_UP_PREPARE_FROZEN
:
1735 rcu_online_cpu(cpu
);
1738 case CPU_DYING_FROZEN
:
1740 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
1741 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
1742 * returns, all online cpus have queued rcu_barrier_func().
1743 * The dying CPU clears its cpu_online_mask bit and
1744 * moves all of its RCU callbacks to ->orphan_cbs_list
1745 * in the context of stop_machine(), so subsequent calls
1746 * to _rcu_barrier() will adopt these callbacks and only
1747 * then queue rcu_barrier_func() on all remaining CPUs.
1749 rcu_send_cbs_to_orphanage(&rcu_bh_state
);
1750 rcu_send_cbs_to_orphanage(&rcu_sched_state
);
1751 rcu_preempt_send_cbs_to_orphanage();
1754 case CPU_DEAD_FROZEN
:
1755 case CPU_UP_CANCELED
:
1756 case CPU_UP_CANCELED_FROZEN
:
1757 rcu_offline_cpu(cpu
);
1766 * Compute the per-level fanout, either using the exact fanout specified
1767 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1769 #ifdef CONFIG_RCU_FANOUT_EXACT
1770 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1774 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--)
1775 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
1777 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1778 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1785 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
1786 ccur
= rsp
->levelcnt
[i
];
1787 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
1791 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1794 * Helper function for rcu_init() that initializes one rcu_state structure.
1796 static void __init
rcu_init_one(struct rcu_state
*rsp
)
1801 struct rcu_node
*rnp
;
1803 /* Initialize the level-tracking arrays. */
1805 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
1806 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
1807 rcu_init_levelspread(rsp
);
1809 /* Initialize the elements themselves, starting from the leaves. */
1811 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
1812 cpustride
*= rsp
->levelspread
[i
];
1813 rnp
= rsp
->level
[i
];
1814 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
1815 spin_lock_init(&rnp
->lock
);
1816 lockdep_set_class(&rnp
->lock
, &rcu_node_class
[i
]);
1819 rnp
->qsmaskinit
= 0;
1820 rnp
->grplo
= j
* cpustride
;
1821 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
1822 if (rnp
->grphi
>= NR_CPUS
)
1823 rnp
->grphi
= NR_CPUS
- 1;
1829 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
1830 rnp
->grpmask
= 1UL << rnp
->grpnum
;
1831 rnp
->parent
= rsp
->level
[i
- 1] +
1832 j
/ rsp
->levelspread
[i
- 1];
1835 INIT_LIST_HEAD(&rnp
->blocked_tasks
[0]);
1836 INIT_LIST_HEAD(&rnp
->blocked_tasks
[1]);
1842 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1843 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1846 #define RCU_INIT_FLAVOR(rsp, rcu_data) \
1850 struct rcu_node *rnp; \
1852 rcu_init_one(rsp); \
1853 rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \
1855 for_each_possible_cpu(i) { \
1856 if (i > rnp[j].grphi) \
1858 per_cpu(rcu_data, i).mynode = &rnp[j]; \
1859 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1860 rcu_boot_init_percpu_data(i, rsp); \
1864 void __init
rcu_init(void)
1868 rcu_bootup_announce();
1869 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
1870 printk(KERN_INFO
"RCU-based detection of stalled CPUs is enabled.\n");
1871 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
1872 RCU_INIT_FLAVOR(&rcu_sched_state
, rcu_sched_data
);
1873 RCU_INIT_FLAVOR(&rcu_bh_state
, rcu_bh_data
);
1874 __rcu_init_preempt();
1875 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
1878 * We don't need protection against CPU-hotplug here because
1879 * this is called early in boot, before either interrupts
1880 * or the scheduler are operational.
1882 cpu_notifier(rcu_cpu_notify
, 0);
1883 for_each_online_cpu(i
)
1884 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)i
);
1887 #include "rcutree_plugin.h"