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 <linux/atomic.h>
40 #include <linux/bitops.h>
41 #include <linux/module.h>
42 #include <linux/completion.h>
43 #include <linux/moduleparam.h>
44 #include <linux/percpu.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/mutex.h>
48 #include <linux/time.h>
49 #include <linux/kernel_stat.h>
50 #include <linux/wait.h>
51 #include <linux/kthread.h>
52 #include <linux/prefetch.h>
55 #include <trace/events/rcu.h>
59 /* Data structures. */
61 static struct lock_class_key rcu_node_class
[NUM_RCU_LVLS
];
63 #define RCU_STATE_INITIALIZER(structname) { \
64 .level = { &structname##_state.node[0] }, \
66 NUM_RCU_LVL_0, /* root of hierarchy. */ \
70 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
72 .signaled = RCU_GP_IDLE, \
75 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
76 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
78 .n_force_qs_ngp = 0, \
79 .name = #structname, \
82 struct rcu_state rcu_sched_state
= RCU_STATE_INITIALIZER(rcu_sched
);
83 DEFINE_PER_CPU(struct rcu_data
, rcu_sched_data
);
85 struct rcu_state rcu_bh_state
= RCU_STATE_INITIALIZER(rcu_bh
);
86 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
);
88 static struct rcu_state
*rcu_state
;
91 * The rcu_scheduler_active variable transitions from zero to one just
92 * before the first task is spawned. So when this variable is zero, RCU
93 * can assume that there is but one task, allowing RCU to (for example)
94 * optimized synchronize_sched() to a simple barrier(). When this variable
95 * is one, RCU must actually do all the hard work required to detect real
96 * grace periods. This variable is also used to suppress boot-time false
97 * positives from lockdep-RCU error checking.
99 int rcu_scheduler_active __read_mostly
;
100 EXPORT_SYMBOL_GPL(rcu_scheduler_active
);
103 * The rcu_scheduler_fully_active variable transitions from zero to one
104 * during the early_initcall() processing, which is after the scheduler
105 * is capable of creating new tasks. So RCU processing (for example,
106 * creating tasks for RCU priority boosting) must be delayed until after
107 * rcu_scheduler_fully_active transitions from zero to one. We also
108 * currently delay invocation of any RCU callbacks until after this point.
110 * It might later prove better for people registering RCU callbacks during
111 * early boot to take responsibility for these callbacks, but one step at
114 static int rcu_scheduler_fully_active __read_mostly
;
116 #ifdef CONFIG_RCU_BOOST
119 * Control variables for per-CPU and per-rcu_node kthreads. These
120 * handle all flavors of RCU.
122 static DEFINE_PER_CPU(struct task_struct
*, rcu_cpu_kthread_task
);
123 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status
);
124 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu
);
125 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops
);
126 DEFINE_PER_CPU(char, rcu_cpu_has_work
);
128 #endif /* #ifdef CONFIG_RCU_BOOST */
130 static void rcu_node_kthread_setaffinity(struct rcu_node
*rnp
, int outgoingcpu
);
131 static void invoke_rcu_core(void);
132 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
);
134 #define RCU_KTHREAD_PRIO 1 /* RT priority for per-CPU kthreads. */
137 * Track the rcutorture test sequence number and the update version
138 * number within a given test. The rcutorture_testseq is incremented
139 * on every rcutorture module load and unload, so has an odd value
140 * when a test is running. The rcutorture_vernum is set to zero
141 * when rcutorture starts and is incremented on each rcutorture update.
142 * These variables enable correlating rcutorture output with the
143 * RCU tracing information.
145 unsigned long rcutorture_testseq
;
146 unsigned long rcutorture_vernum
;
149 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
150 * permit this function to be invoked without holding the root rcu_node
151 * structure's ->lock, but of course results can be subject to change.
153 static int rcu_gp_in_progress(struct rcu_state
*rsp
)
155 return ACCESS_ONCE(rsp
->completed
) != ACCESS_ONCE(rsp
->gpnum
);
159 * Note a quiescent state. Because we do not need to know
160 * how many quiescent states passed, just if there was at least
161 * one since the start of the grace period, this just sets a flag.
163 void rcu_sched_qs(int cpu
)
165 struct rcu_data
*rdp
= &per_cpu(rcu_sched_data
, cpu
);
167 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
169 rdp
->passed_quiesc
= 1;
172 void rcu_bh_qs(int cpu
)
174 struct rcu_data
*rdp
= &per_cpu(rcu_bh_data
, cpu
);
176 rdp
->passed_quiesc_completed
= rdp
->gpnum
- 1;
178 rdp
->passed_quiesc
= 1;
182 * Note a context switch. This is a quiescent state for RCU-sched,
183 * and requires special handling for preemptible RCU.
185 void rcu_note_context_switch(int cpu
)
187 trace_rcu_utilization("Start context switch");
189 rcu_preempt_note_context_switch(cpu
);
190 trace_rcu_utilization("End context switch");
192 EXPORT_SYMBOL_GPL(rcu_note_context_switch
);
195 DEFINE_PER_CPU(struct rcu_dynticks
, rcu_dynticks
) = {
196 .dynticks_nesting
= 1,
197 .dynticks
= ATOMIC_INIT(1),
199 #endif /* #ifdef CONFIG_NO_HZ */
201 static int blimit
= 10; /* Maximum callbacks per softirq. */
202 static int qhimark
= 10000; /* If this many pending, ignore blimit. */
203 static int qlowmark
= 100; /* Once only this many pending, use blimit. */
205 module_param(blimit
, int, 0);
206 module_param(qhimark
, int, 0);
207 module_param(qlowmark
, int, 0);
209 int rcu_cpu_stall_suppress __read_mostly
;
210 module_param(rcu_cpu_stall_suppress
, int, 0644);
212 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
);
213 static int rcu_pending(int cpu
);
216 * Return the number of RCU-sched batches processed thus far for debug & stats.
218 long rcu_batches_completed_sched(void)
220 return rcu_sched_state
.completed
;
222 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched
);
225 * Return the number of RCU BH batches processed thus far for debug & stats.
227 long rcu_batches_completed_bh(void)
229 return rcu_bh_state
.completed
;
231 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
234 * Force a quiescent state for RCU BH.
236 void rcu_bh_force_quiescent_state(void)
238 force_quiescent_state(&rcu_bh_state
, 0);
240 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state
);
243 * Record the number of times rcutorture tests have been initiated and
244 * terminated. This information allows the debugfs tracing stats to be
245 * correlated to the rcutorture messages, even when the rcutorture module
246 * is being repeatedly loaded and unloaded. In other words, we cannot
247 * store this state in rcutorture itself.
249 void rcutorture_record_test_transition(void)
251 rcutorture_testseq
++;
252 rcutorture_vernum
= 0;
254 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition
);
257 * Record the number of writer passes through the current rcutorture test.
258 * This is also used to correlate debugfs tracing stats with the rcutorture
261 void rcutorture_record_progress(unsigned long vernum
)
265 EXPORT_SYMBOL_GPL(rcutorture_record_progress
);
268 * Force a quiescent state for RCU-sched.
270 void rcu_sched_force_quiescent_state(void)
272 force_quiescent_state(&rcu_sched_state
, 0);
274 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state
);
277 * Does the CPU have callbacks ready to be invoked?
280 cpu_has_callbacks_ready_to_invoke(struct rcu_data
*rdp
)
282 return &rdp
->nxtlist
!= rdp
->nxttail
[RCU_DONE_TAIL
];
286 * Does the current CPU require a yet-as-unscheduled grace period?
289 cpu_needs_another_gp(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
291 return *rdp
->nxttail
[RCU_DONE_TAIL
] && !rcu_gp_in_progress(rsp
);
295 * Return the root node of the specified rcu_state structure.
297 static struct rcu_node
*rcu_get_root(struct rcu_state
*rsp
)
299 return &rsp
->node
[0];
305 * If the specified CPU is offline, tell the caller that it is in
306 * a quiescent state. Otherwise, whack it with a reschedule IPI.
307 * Grace periods can end up waiting on an offline CPU when that
308 * CPU is in the process of coming online -- it will be added to the
309 * rcu_node bitmasks before it actually makes it online. The same thing
310 * can happen while a CPU is in the process of coming online. Because this
311 * race is quite rare, we check for it after detecting that the grace
312 * period has been delayed rather than checking each and every CPU
313 * each and every time we start a new grace period.
315 static int rcu_implicit_offline_qs(struct rcu_data
*rdp
)
318 * If the CPU is offline, it is in a quiescent state. We can
319 * trust its state not to change because interrupts are disabled.
321 if (cpu_is_offline(rdp
->cpu
)) {
326 /* If preemptible RCU, no point in sending reschedule IPI. */
327 if (rdp
->preemptible
)
330 /* The CPU is online, so send it a reschedule IPI. */
331 if (rdp
->cpu
!= smp_processor_id())
332 smp_send_reschedule(rdp
->cpu
);
339 #endif /* #ifdef CONFIG_SMP */
344 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
346 * Enter nohz mode, in other words, -leave- the mode in which RCU
347 * read-side critical sections can occur. (Though RCU read-side
348 * critical sections can occur in irq handlers in nohz mode, a possibility
349 * handled by rcu_irq_enter() and rcu_irq_exit()).
351 void rcu_enter_nohz(void)
354 struct rcu_dynticks
*rdtp
;
356 local_irq_save(flags
);
357 rdtp
= &__get_cpu_var(rcu_dynticks
);
358 if (--rdtp
->dynticks_nesting
) {
359 local_irq_restore(flags
);
362 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
363 smp_mb__before_atomic_inc(); /* See above. */
364 atomic_inc(&rdtp
->dynticks
);
365 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
366 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
367 local_irq_restore(flags
);
369 /* If the interrupt queued a callback, get out of dyntick mode. */
371 (__get_cpu_var(rcu_sched_data
).nxtlist
||
372 __get_cpu_var(rcu_bh_data
).nxtlist
||
373 rcu_preempt_needs_cpu(smp_processor_id())))
378 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
380 * Exit nohz mode, in other words, -enter- the mode in which RCU
381 * read-side critical sections normally occur.
383 void rcu_exit_nohz(void)
386 struct rcu_dynticks
*rdtp
;
388 local_irq_save(flags
);
389 rdtp
= &__get_cpu_var(rcu_dynticks
);
390 if (rdtp
->dynticks_nesting
++) {
391 local_irq_restore(flags
);
394 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
395 atomic_inc(&rdtp
->dynticks
);
396 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
397 smp_mb__after_atomic_inc(); /* See above. */
398 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
399 local_irq_restore(flags
);
403 * rcu_nmi_enter - inform RCU of entry to NMI context
405 * If the CPU was idle with dynamic ticks active, and there is no
406 * irq handler running, this updates rdtp->dynticks_nmi to let the
407 * RCU grace-period handling know that the CPU is active.
409 void rcu_nmi_enter(void)
411 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
413 if (rdtp
->dynticks_nmi_nesting
== 0 &&
414 (atomic_read(&rdtp
->dynticks
) & 0x1))
416 rdtp
->dynticks_nmi_nesting
++;
417 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
418 atomic_inc(&rdtp
->dynticks
);
419 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
420 smp_mb__after_atomic_inc(); /* See above. */
421 WARN_ON_ONCE(!(atomic_read(&rdtp
->dynticks
) & 0x1));
425 * rcu_nmi_exit - inform RCU of exit from NMI context
427 * If the CPU was idle with dynamic ticks active, and there is no
428 * irq handler running, this updates rdtp->dynticks_nmi to let the
429 * RCU grace-period handling know that the CPU is no longer active.
431 void rcu_nmi_exit(void)
433 struct rcu_dynticks
*rdtp
= &__get_cpu_var(rcu_dynticks
);
435 if (rdtp
->dynticks_nmi_nesting
== 0 ||
436 --rdtp
->dynticks_nmi_nesting
!= 0)
438 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
439 smp_mb__before_atomic_inc(); /* See above. */
440 atomic_inc(&rdtp
->dynticks
);
441 smp_mb__after_atomic_inc(); /* Force delay to next write. */
442 WARN_ON_ONCE(atomic_read(&rdtp
->dynticks
) & 0x1);
446 * rcu_irq_enter - inform RCU of entry to hard irq context
448 * If the CPU was idle with dynamic ticks active, this updates the
449 * rdtp->dynticks to let the RCU handling know that the CPU is active.
451 void rcu_irq_enter(void)
457 * rcu_irq_exit - inform RCU of exit from hard irq context
459 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
460 * to put let the RCU handling be aware that the CPU is going back to idle
463 void rcu_irq_exit(void)
471 * Snapshot the specified CPU's dynticks counter so that we can later
472 * credit them with an implicit quiescent state. Return 1 if this CPU
473 * is in dynticks idle mode, which is an extended quiescent state.
475 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
477 rdp
->dynticks_snap
= atomic_add_return(0, &rdp
->dynticks
->dynticks
);
482 * Return true if the specified CPU has passed through a quiescent
483 * state by virtue of being in or having passed through an dynticks
484 * idle state since the last call to dyntick_save_progress_counter()
487 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
492 curr
= (unsigned long)atomic_add_return(0, &rdp
->dynticks
->dynticks
);
493 snap
= (unsigned long)rdp
->dynticks_snap
;
496 * If the CPU passed through or entered a dynticks idle phase with
497 * no active irq/NMI handlers, then we can safely pretend that the CPU
498 * already acknowledged the request to pass through a quiescent
499 * state. Either way, that CPU cannot possibly be in an RCU
500 * read-side critical section that started before the beginning
501 * of the current RCU grace period.
503 if ((curr
& 0x1) == 0 || ULONG_CMP_GE(curr
, snap
+ 2)) {
508 /* Go check for the CPU being offline. */
509 return rcu_implicit_offline_qs(rdp
);
512 #endif /* #ifdef CONFIG_SMP */
514 #else /* #ifdef CONFIG_NO_HZ */
518 static int dyntick_save_progress_counter(struct rcu_data
*rdp
)
523 static int rcu_implicit_dynticks_qs(struct rcu_data
*rdp
)
525 return rcu_implicit_offline_qs(rdp
);
528 #endif /* #ifdef CONFIG_SMP */
530 #endif /* #else #ifdef CONFIG_NO_HZ */
532 int rcu_cpu_stall_suppress __read_mostly
;
534 static void record_gp_stall_check_time(struct rcu_state
*rsp
)
536 rsp
->gp_start
= jiffies
;
537 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_CHECK
;
540 static void print_other_cpu_stall(struct rcu_state
*rsp
)
545 struct rcu_node
*rnp
= rcu_get_root(rsp
);
547 /* Only let one CPU complain about others per time interval. */
549 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
550 delta
= jiffies
- rsp
->jiffies_stall
;
551 if (delta
< RCU_STALL_RAT_DELAY
|| !rcu_gp_in_progress(rsp
)) {
552 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
555 rsp
->jiffies_stall
= jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
558 * Now rat on any tasks that got kicked up to the root rcu_node
559 * due to CPU offlining.
561 rcu_print_task_stall(rnp
);
562 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
565 * OK, time to rat on our buddy...
566 * See Documentation/RCU/stallwarn.txt for info on how to debug
567 * RCU CPU stall warnings.
569 printk(KERN_ERR
"INFO: %s detected stalls on CPUs/tasks: {",
571 rcu_for_each_leaf_node(rsp
, rnp
) {
572 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
573 rcu_print_task_stall(rnp
);
574 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
575 if (rnp
->qsmask
== 0)
577 for (cpu
= 0; cpu
<= rnp
->grphi
- rnp
->grplo
; cpu
++)
578 if (rnp
->qsmask
& (1UL << cpu
))
579 printk(" %d", rnp
->grplo
+ cpu
);
581 printk("} (detected by %d, t=%ld jiffies)\n",
582 smp_processor_id(), (long)(jiffies
- rsp
->gp_start
));
583 trigger_all_cpu_backtrace();
585 /* If so configured, complain about tasks blocking the grace period. */
587 rcu_print_detail_task_stall(rsp
);
589 force_quiescent_state(rsp
, 0); /* Kick them all. */
592 static void print_cpu_stall(struct rcu_state
*rsp
)
595 struct rcu_node
*rnp
= rcu_get_root(rsp
);
598 * OK, time to rat on ourselves...
599 * See Documentation/RCU/stallwarn.txt for info on how to debug
600 * RCU CPU stall warnings.
602 printk(KERN_ERR
"INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
603 rsp
->name
, smp_processor_id(), jiffies
- rsp
->gp_start
);
604 trigger_all_cpu_backtrace();
606 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
607 if (ULONG_CMP_GE(jiffies
, rsp
->jiffies_stall
))
609 jiffies
+ RCU_SECONDS_TILL_STALL_RECHECK
;
610 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
612 set_need_resched(); /* kick ourselves to get things going. */
615 static void check_cpu_stall(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
619 struct rcu_node
*rnp
;
621 if (rcu_cpu_stall_suppress
)
623 j
= ACCESS_ONCE(jiffies
);
624 js
= ACCESS_ONCE(rsp
->jiffies_stall
);
626 if ((ACCESS_ONCE(rnp
->qsmask
) & rdp
->grpmask
) && ULONG_CMP_GE(j
, js
)) {
628 /* We haven't checked in, so go dump stack. */
629 print_cpu_stall(rsp
);
631 } else if (rcu_gp_in_progress(rsp
) &&
632 ULONG_CMP_GE(j
, js
+ RCU_STALL_RAT_DELAY
)) {
634 /* They had a few time units to dump stack, so complain. */
635 print_other_cpu_stall(rsp
);
639 static int rcu_panic(struct notifier_block
*this, unsigned long ev
, void *ptr
)
641 rcu_cpu_stall_suppress
= 1;
646 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
648 * Set the stall-warning timeout way off into the future, thus preventing
649 * any RCU CPU stall-warning messages from appearing in the current set of
652 * The caller must disable hard irqs.
654 void rcu_cpu_stall_reset(void)
656 rcu_sched_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
657 rcu_bh_state
.jiffies_stall
= jiffies
+ ULONG_MAX
/ 2;
658 rcu_preempt_stall_reset();
661 static struct notifier_block rcu_panic_block
= {
662 .notifier_call
= rcu_panic
,
665 static void __init
check_cpu_stall_init(void)
667 atomic_notifier_chain_register(&panic_notifier_list
, &rcu_panic_block
);
671 * Update CPU-local rcu_data state to record the newly noticed grace period.
672 * This is used both when we started the grace period and when we notice
673 * that someone else started the grace period. The caller must hold the
674 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
675 * and must have irqs disabled.
677 static void __note_new_gpnum(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
679 if (rdp
->gpnum
!= rnp
->gpnum
) {
681 * If the current grace period is waiting for this CPU,
682 * set up to detect a quiescent state, otherwise don't
683 * go looking for one.
685 rdp
->gpnum
= rnp
->gpnum
;
686 if (rnp
->qsmask
& rdp
->grpmask
) {
688 rdp
->passed_quiesc
= 0;
694 static void note_new_gpnum(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
697 struct rcu_node
*rnp
;
699 local_irq_save(flags
);
701 if (rdp
->gpnum
== ACCESS_ONCE(rnp
->gpnum
) || /* outside lock. */
702 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
703 local_irq_restore(flags
);
706 __note_new_gpnum(rsp
, rnp
, rdp
);
707 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
711 * Did someone else start a new RCU grace period start since we last
712 * checked? Update local state appropriately if so. Must be called
713 * on the CPU corresponding to rdp.
716 check_for_new_grace_period(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
721 local_irq_save(flags
);
722 if (rdp
->gpnum
!= rsp
->gpnum
) {
723 note_new_gpnum(rsp
, rdp
);
726 local_irq_restore(flags
);
731 * Advance this CPU's callbacks, but only if the current grace period
732 * has ended. This may be called only from the CPU to whom the rdp
733 * belongs. In addition, the corresponding leaf rcu_node structure's
734 * ->lock must be held by the caller, with irqs disabled.
737 __rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
739 /* Did another grace period end? */
740 if (rdp
->completed
!= rnp
->completed
) {
742 /* Advance callbacks. No harm if list empty. */
743 rdp
->nxttail
[RCU_DONE_TAIL
] = rdp
->nxttail
[RCU_WAIT_TAIL
];
744 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_READY_TAIL
];
745 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
747 /* Remember that we saw this grace-period completion. */
748 rdp
->completed
= rnp
->completed
;
751 * If we were in an extended quiescent state, we may have
752 * missed some grace periods that others CPUs handled on
753 * our behalf. Catch up with this state to avoid noting
754 * spurious new grace periods. If another grace period
755 * has started, then rnp->gpnum will have advanced, so
756 * we will detect this later on.
758 if (ULONG_CMP_LT(rdp
->gpnum
, rdp
->completed
))
759 rdp
->gpnum
= rdp
->completed
;
762 * If RCU does not need a quiescent state from this CPU,
763 * then make sure that this CPU doesn't go looking for one.
765 if ((rnp
->qsmask
& rdp
->grpmask
) == 0)
771 * Advance this CPU's callbacks, but only if the current grace period
772 * has ended. This may be called only from the CPU to whom the rdp
776 rcu_process_gp_end(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
779 struct rcu_node
*rnp
;
781 local_irq_save(flags
);
783 if (rdp
->completed
== ACCESS_ONCE(rnp
->completed
) || /* outside lock. */
784 !raw_spin_trylock(&rnp
->lock
)) { /* irqs already off, so later. */
785 local_irq_restore(flags
);
788 __rcu_process_gp_end(rsp
, rnp
, rdp
);
789 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
793 * Do per-CPU grace-period initialization for running CPU. The caller
794 * must hold the lock of the leaf rcu_node structure corresponding to
798 rcu_start_gp_per_cpu(struct rcu_state
*rsp
, struct rcu_node
*rnp
, struct rcu_data
*rdp
)
800 /* Prior grace period ended, so advance callbacks for current CPU. */
801 __rcu_process_gp_end(rsp
, rnp
, rdp
);
804 * Because this CPU just now started the new grace period, we know
805 * that all of its callbacks will be covered by this upcoming grace
806 * period, even the ones that were registered arbitrarily recently.
807 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
809 * Other CPUs cannot be sure exactly when the grace period started.
810 * Therefore, their recently registered callbacks must pass through
811 * an additional RCU_NEXT_READY stage, so that they will be handled
812 * by the next RCU grace period.
814 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
815 rdp
->nxttail
[RCU_WAIT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
817 /* Set state so that this CPU will detect the next quiescent state. */
818 __note_new_gpnum(rsp
, rnp
, rdp
);
822 * Start a new RCU grace period if warranted, re-initializing the hierarchy
823 * in preparation for detecting the next grace period. The caller must hold
824 * the root node's ->lock, which is released before return. Hard irqs must
828 rcu_start_gp(struct rcu_state
*rsp
, unsigned long flags
)
829 __releases(rcu_get_root(rsp
)->lock
)
831 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
832 struct rcu_node
*rnp
= rcu_get_root(rsp
);
834 if (!cpu_needs_another_gp(rsp
, rdp
) || rsp
->fqs_active
) {
835 if (cpu_needs_another_gp(rsp
, rdp
))
836 rsp
->fqs_need_gp
= 1;
837 if (rnp
->completed
== rsp
->completed
) {
838 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
841 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
844 * Propagate new ->completed value to rcu_node structures
845 * so that other CPUs don't have to wait until the start
846 * of the next grace period to process their callbacks.
848 rcu_for_each_node_breadth_first(rsp
, rnp
) {
849 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
850 rnp
->completed
= rsp
->completed
;
851 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
853 local_irq_restore(flags
);
857 /* Advance to a new grace period and initialize state. */
859 WARN_ON_ONCE(rsp
->signaled
== RCU_GP_INIT
);
860 rsp
->signaled
= RCU_GP_INIT
; /* Hold off force_quiescent_state. */
861 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
862 record_gp_stall_check_time(rsp
);
864 /* Special-case the common single-level case. */
865 if (NUM_RCU_NODES
== 1) {
866 rcu_preempt_check_blocked_tasks(rnp
);
867 rnp
->qsmask
= rnp
->qsmaskinit
;
868 rnp
->gpnum
= rsp
->gpnum
;
869 rnp
->completed
= rsp
->completed
;
870 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state OK. */
871 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
872 rcu_preempt_boost_start_gp(rnp
);
873 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
877 raw_spin_unlock(&rnp
->lock
); /* leave irqs disabled. */
880 /* Exclude any concurrent CPU-hotplug operations. */
881 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
884 * Set the quiescent-state-needed bits in all the rcu_node
885 * structures for all currently online CPUs in breadth-first
886 * order, starting from the root rcu_node structure. This
887 * operation relies on the layout of the hierarchy within the
888 * rsp->node[] array. Note that other CPUs will access only
889 * the leaves of the hierarchy, which still indicate that no
890 * grace period is in progress, at least until the corresponding
891 * leaf node has been initialized. In addition, we have excluded
892 * CPU-hotplug operations.
894 * Note that the grace period cannot complete until we finish
895 * the initialization process, as there will be at least one
896 * qsmask bit set in the root node until that time, namely the
897 * one corresponding to this CPU, due to the fact that we have
900 rcu_for_each_node_breadth_first(rsp
, rnp
) {
901 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
902 rcu_preempt_check_blocked_tasks(rnp
);
903 rnp
->qsmask
= rnp
->qsmaskinit
;
904 rnp
->gpnum
= rsp
->gpnum
;
905 rnp
->completed
= rsp
->completed
;
906 if (rnp
== rdp
->mynode
)
907 rcu_start_gp_per_cpu(rsp
, rnp
, rdp
);
908 rcu_preempt_boost_start_gp(rnp
);
909 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
912 rnp
= rcu_get_root(rsp
);
913 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
914 rsp
->signaled
= RCU_SIGNAL_INIT
; /* force_quiescent_state now OK. */
915 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
916 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
920 * Report a full set of quiescent states to the specified rcu_state
921 * data structure. This involves cleaning up after the prior grace
922 * period and letting rcu_start_gp() start up the next grace period
923 * if one is needed. Note that the caller must hold rnp->lock, as
924 * required by rcu_start_gp(), which will release it.
926 static void rcu_report_qs_rsp(struct rcu_state
*rsp
, unsigned long flags
)
927 __releases(rcu_get_root(rsp
)->lock
)
929 unsigned long gp_duration
;
931 WARN_ON_ONCE(!rcu_gp_in_progress(rsp
));
934 * Ensure that all grace-period and pre-grace-period activity
935 * is seen before the assignment to rsp->completed.
937 smp_mb(); /* See above block comment. */
938 gp_duration
= jiffies
- rsp
->gp_start
;
939 if (gp_duration
> rsp
->gp_max
)
940 rsp
->gp_max
= gp_duration
;
941 rsp
->completed
= rsp
->gpnum
;
942 rsp
->signaled
= RCU_GP_IDLE
;
943 rcu_start_gp(rsp
, flags
); /* releases root node's rnp->lock. */
947 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
948 * Allows quiescent states for a group of CPUs to be reported at one go
949 * to the specified rcu_node structure, though all the CPUs in the group
950 * must be represented by the same rcu_node structure (which need not be
951 * a leaf rcu_node structure, though it often will be). That structure's
952 * lock must be held upon entry, and it is released before return.
955 rcu_report_qs_rnp(unsigned long mask
, struct rcu_state
*rsp
,
956 struct rcu_node
*rnp
, unsigned long flags
)
957 __releases(rnp
->lock
)
959 struct rcu_node
*rnp_c
;
961 /* Walk up the rcu_node hierarchy. */
963 if (!(rnp
->qsmask
& mask
)) {
965 /* Our bit has already been cleared, so done. */
966 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
969 rnp
->qsmask
&= ~mask
;
970 if (rnp
->qsmask
!= 0 || rcu_preempt_blocked_readers_cgp(rnp
)) {
972 /* Other bits still set at this level, so done. */
973 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
977 if (rnp
->parent
== NULL
) {
979 /* No more levels. Exit loop holding root lock. */
983 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
986 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
987 WARN_ON_ONCE(rnp_c
->qsmask
);
991 * Get here if we are the last CPU to pass through a quiescent
992 * state for this grace period. Invoke rcu_report_qs_rsp()
993 * to clean up and start the next grace period if one is needed.
995 rcu_report_qs_rsp(rsp
, flags
); /* releases rnp->lock. */
999 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1000 * structure. This must be either called from the specified CPU, or
1001 * called when the specified CPU is known to be offline (and when it is
1002 * also known that no other CPU is concurrently trying to help the offline
1003 * CPU). The lastcomp argument is used to make sure we are still in the
1004 * grace period of interest. We don't want to end the current grace period
1005 * based on quiescent states detected in an earlier grace period!
1008 rcu_report_qs_rdp(int cpu
, struct rcu_state
*rsp
, struct rcu_data
*rdp
, long lastcomp
)
1010 unsigned long flags
;
1012 struct rcu_node
*rnp
;
1015 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1016 if (lastcomp
!= rnp
->completed
) {
1019 * Someone beat us to it for this grace period, so leave.
1020 * The race with GP start is resolved by the fact that we
1021 * hold the leaf rcu_node lock, so that the per-CPU bits
1022 * cannot yet be initialized -- so we would simply find our
1023 * CPU's bit already cleared in rcu_report_qs_rnp() if this
1026 rdp
->passed_quiesc
= 0; /* try again later! */
1027 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1030 mask
= rdp
->grpmask
;
1031 if ((rnp
->qsmask
& mask
) == 0) {
1032 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1034 rdp
->qs_pending
= 0;
1037 * This GP can't end until cpu checks in, so all of our
1038 * callbacks can be processed during the next GP.
1040 rdp
->nxttail
[RCU_NEXT_READY_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1042 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
); /* rlses rnp->lock */
1047 * Check to see if there is a new grace period of which this CPU
1048 * is not yet aware, and if so, set up local rcu_data state for it.
1049 * Otherwise, see if this CPU has just passed through its first
1050 * quiescent state for this grace period, and record that fact if so.
1053 rcu_check_quiescent_state(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1055 /* If there is now a new grace period, record and return. */
1056 if (check_for_new_grace_period(rsp
, rdp
))
1060 * Does this CPU still need to do its part for current grace period?
1061 * If no, return and let the other CPUs do their part as well.
1063 if (!rdp
->qs_pending
)
1067 * Was there a quiescent state since the beginning of the grace
1068 * period? If no, then exit and wait for the next call.
1070 if (!rdp
->passed_quiesc
)
1074 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1077 rcu_report_qs_rdp(rdp
->cpu
, rsp
, rdp
, rdp
->passed_quiesc_completed
);
1080 #ifdef CONFIG_HOTPLUG_CPU
1083 * Move a dying CPU's RCU callbacks to online CPU's callback list.
1084 * Synchronization is not required because this function executes
1085 * in stop_machine() context.
1087 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1090 /* current DYING CPU is cleared in the cpu_online_mask */
1091 int receive_cpu
= cpumask_any(cpu_online_mask
);
1092 struct rcu_data
*rdp
= this_cpu_ptr(rsp
->rda
);
1093 struct rcu_data
*receive_rdp
= per_cpu_ptr(rsp
->rda
, receive_cpu
);
1095 if (rdp
->nxtlist
== NULL
)
1096 return; /* irqs disabled, so comparison is stable. */
1098 *receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxtlist
;
1099 receive_rdp
->nxttail
[RCU_NEXT_TAIL
] = rdp
->nxttail
[RCU_NEXT_TAIL
];
1100 receive_rdp
->qlen
+= rdp
->qlen
;
1101 receive_rdp
->n_cbs_adopted
+= rdp
->qlen
;
1102 rdp
->n_cbs_orphaned
+= rdp
->qlen
;
1104 rdp
->nxtlist
= NULL
;
1105 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1106 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1111 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1112 * and move all callbacks from the outgoing CPU to the current one.
1113 * There can only be one CPU hotplug operation at a time, so no other
1114 * CPU can be attempting to update rcu_cpu_kthread_task.
1116 static void __rcu_offline_cpu(int cpu
, struct rcu_state
*rsp
)
1118 unsigned long flags
;
1120 int need_report
= 0;
1121 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1122 struct rcu_node
*rnp
;
1124 rcu_stop_cpu_kthread(cpu
);
1126 /* Exclude any attempts to start a new grace period. */
1127 raw_spin_lock_irqsave(&rsp
->onofflock
, flags
);
1129 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1130 rnp
= rdp
->mynode
; /* this is the outgoing CPU's rnp. */
1131 mask
= rdp
->grpmask
; /* rnp->grplo is constant. */
1133 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1134 rnp
->qsmaskinit
&= ~mask
;
1135 if (rnp
->qsmaskinit
!= 0) {
1136 if (rnp
!= rdp
->mynode
)
1137 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1140 if (rnp
== rdp
->mynode
)
1141 need_report
= rcu_preempt_offline_tasks(rsp
, rnp
, rdp
);
1143 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1144 mask
= rnp
->grpmask
;
1146 } while (rnp
!= NULL
);
1149 * We still hold the leaf rcu_node structure lock here, and
1150 * irqs are still disabled. The reason for this subterfuge is
1151 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1152 * held leads to deadlock.
1154 raw_spin_unlock(&rsp
->onofflock
); /* irqs remain disabled. */
1156 if (need_report
& RCU_OFL_TASKS_NORM_GP
)
1157 rcu_report_unblock_qs_rnp(rnp
, flags
);
1159 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1160 if (need_report
& RCU_OFL_TASKS_EXP_GP
)
1161 rcu_report_exp_rnp(rsp
, rnp
);
1162 rcu_node_kthread_setaffinity(rnp
, -1);
1166 * Remove the specified CPU from the RCU hierarchy and move any pending
1167 * callbacks that it might have to the current CPU. This code assumes
1168 * that at least one CPU in the system will remain running at all times.
1169 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1171 static void rcu_offline_cpu(int cpu
)
1173 __rcu_offline_cpu(cpu
, &rcu_sched_state
);
1174 __rcu_offline_cpu(cpu
, &rcu_bh_state
);
1175 rcu_preempt_offline_cpu(cpu
);
1178 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1180 static void rcu_send_cbs_to_online(struct rcu_state
*rsp
)
1184 static void rcu_offline_cpu(int cpu
)
1188 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1191 * Invoke any RCU callbacks that have made it to the end of their grace
1192 * period. Thottle as specified by rdp->blimit.
1194 static void rcu_do_batch(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1196 unsigned long flags
;
1197 struct rcu_head
*next
, *list
, **tail
;
1200 /* If no callbacks are ready, just return.*/
1201 if (!cpu_has_callbacks_ready_to_invoke(rdp
)) {
1202 trace_rcu_batch_start(rsp
->name
, 0, 0);
1203 trace_rcu_batch_end(rsp
->name
, 0);
1208 * Extract the list of ready callbacks, disabling to prevent
1209 * races with call_rcu() from interrupt handlers.
1211 local_irq_save(flags
);
1213 trace_rcu_batch_start(rsp
->name
, rdp
->qlen
, bl
);
1214 list
= rdp
->nxtlist
;
1215 rdp
->nxtlist
= *rdp
->nxttail
[RCU_DONE_TAIL
];
1216 *rdp
->nxttail
[RCU_DONE_TAIL
] = NULL
;
1217 tail
= rdp
->nxttail
[RCU_DONE_TAIL
];
1218 for (count
= RCU_NEXT_SIZE
- 1; count
>= 0; count
--)
1219 if (rdp
->nxttail
[count
] == rdp
->nxttail
[RCU_DONE_TAIL
])
1220 rdp
->nxttail
[count
] = &rdp
->nxtlist
;
1221 local_irq_restore(flags
);
1223 /* Invoke callbacks. */
1228 debug_rcu_head_unqueue(list
);
1229 __rcu_reclaim(list
);
1235 local_irq_save(flags
);
1236 trace_rcu_batch_end(rsp
->name
, count
);
1238 /* Update count, and requeue any remaining callbacks. */
1240 rdp
->n_cbs_invoked
+= count
;
1242 *tail
= rdp
->nxtlist
;
1243 rdp
->nxtlist
= list
;
1244 for (count
= 0; count
< RCU_NEXT_SIZE
; count
++)
1245 if (&rdp
->nxtlist
== rdp
->nxttail
[count
])
1246 rdp
->nxttail
[count
] = tail
;
1251 /* Reinstate batch limit if we have worked down the excess. */
1252 if (rdp
->blimit
== LONG_MAX
&& rdp
->qlen
<= qlowmark
)
1253 rdp
->blimit
= blimit
;
1255 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1256 if (rdp
->qlen
== 0 && rdp
->qlen_last_fqs_check
!= 0) {
1257 rdp
->qlen_last_fqs_check
= 0;
1258 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1259 } else if (rdp
->qlen
< rdp
->qlen_last_fqs_check
- qhimark
)
1260 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1262 local_irq_restore(flags
);
1264 /* Re-raise the RCU softirq if there are callbacks remaining. */
1265 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1270 * Check to see if this CPU is in a non-context-switch quiescent state
1271 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1272 * Also schedule the RCU softirq handler.
1274 * This function must be called with hardirqs disabled. It is normally
1275 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1276 * false, there is no point in invoking rcu_check_callbacks().
1278 void rcu_check_callbacks(int cpu
, int user
)
1280 trace_rcu_utilization("Start scheduler-tick");
1282 (idle_cpu(cpu
) && rcu_scheduler_active
&&
1283 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
1286 * Get here if this CPU took its interrupt from user
1287 * mode or from the idle loop, and if this is not a
1288 * nested interrupt. In this case, the CPU is in
1289 * a quiescent state, so note it.
1291 * No memory barrier is required here because both
1292 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1293 * variables that other CPUs neither access nor modify,
1294 * at least not while the corresponding CPU is online.
1300 } else if (!in_softirq()) {
1303 * Get here if this CPU did not take its interrupt from
1304 * softirq, in other words, if it is not interrupting
1305 * a rcu_bh read-side critical section. This is an _bh
1306 * critical section, so note it.
1311 rcu_preempt_check_callbacks(cpu
);
1312 if (rcu_pending(cpu
))
1314 trace_rcu_utilization("End scheduler-tick");
1320 * Scan the leaf rcu_node structures, processing dyntick state for any that
1321 * have not yet encountered a quiescent state, using the function specified.
1322 * Also initiate boosting for any threads blocked on the root rcu_node.
1324 * The caller must have suppressed start of new grace periods.
1326 static void force_qs_rnp(struct rcu_state
*rsp
, int (*f
)(struct rcu_data
*))
1330 unsigned long flags
;
1332 struct rcu_node
*rnp
;
1334 rcu_for_each_leaf_node(rsp
, rnp
) {
1336 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1337 if (!rcu_gp_in_progress(rsp
)) {
1338 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1341 if (rnp
->qsmask
== 0) {
1342 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock */
1347 for (; cpu
<= rnp
->grphi
; cpu
++, bit
<<= 1) {
1348 if ((rnp
->qsmask
& bit
) != 0 &&
1349 f(per_cpu_ptr(rsp
->rda
, cpu
)))
1354 /* rcu_report_qs_rnp() releases rnp->lock. */
1355 rcu_report_qs_rnp(mask
, rsp
, rnp
, flags
);
1358 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1360 rnp
= rcu_get_root(rsp
);
1361 if (rnp
->qsmask
== 0) {
1362 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1363 rcu_initiate_boost(rnp
, flags
); /* releases rnp->lock. */
1368 * Force quiescent states on reluctant CPUs, and also detect which
1369 * CPUs are in dyntick-idle mode.
1371 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1373 unsigned long flags
;
1374 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1376 trace_rcu_utilization("Start fqs");
1377 if (!rcu_gp_in_progress(rsp
)) {
1378 trace_rcu_utilization("End fqs");
1379 return; /* No grace period in progress, nothing to force. */
1381 if (!raw_spin_trylock_irqsave(&rsp
->fqslock
, flags
)) {
1382 rsp
->n_force_qs_lh
++; /* Inexact, can lose counts. Tough! */
1383 trace_rcu_utilization("End fqs");
1384 return; /* Someone else is already on the job. */
1386 if (relaxed
&& ULONG_CMP_GE(rsp
->jiffies_force_qs
, jiffies
))
1387 goto unlock_fqs_ret
; /* no emergency and done recently. */
1389 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1390 rsp
->jiffies_force_qs
= jiffies
+ RCU_JIFFIES_TILL_FORCE_QS
;
1391 if(!rcu_gp_in_progress(rsp
)) {
1392 rsp
->n_force_qs_ngp
++;
1393 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1394 goto unlock_fqs_ret
; /* no GP in progress, time updated. */
1396 rsp
->fqs_active
= 1;
1397 switch (rsp
->signaled
) {
1401 break; /* grace period idle or initializing, ignore. */
1403 case RCU_SAVE_DYNTICK
:
1404 if (RCU_SIGNAL_INIT
!= RCU_SAVE_DYNTICK
)
1405 break; /* So gcc recognizes the dead code. */
1407 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1409 /* Record dyntick-idle state. */
1410 force_qs_rnp(rsp
, dyntick_save_progress_counter
);
1411 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1412 if (rcu_gp_in_progress(rsp
))
1413 rsp
->signaled
= RCU_FORCE_QS
;
1418 /* Check dyntick-idle state, send IPI to laggarts. */
1419 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1420 force_qs_rnp(rsp
, rcu_implicit_dynticks_qs
);
1422 /* Leave state in case more forcing is required. */
1424 raw_spin_lock(&rnp
->lock
); /* irqs already disabled */
1427 rsp
->fqs_active
= 0;
1428 if (rsp
->fqs_need_gp
) {
1429 raw_spin_unlock(&rsp
->fqslock
); /* irqs remain disabled */
1430 rsp
->fqs_need_gp
= 0;
1431 rcu_start_gp(rsp
, flags
); /* releases rnp->lock */
1432 trace_rcu_utilization("End fqs");
1435 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled */
1437 raw_spin_unlock_irqrestore(&rsp
->fqslock
, flags
);
1438 trace_rcu_utilization("End fqs");
1441 #else /* #ifdef CONFIG_SMP */
1443 static void force_quiescent_state(struct rcu_state
*rsp
, int relaxed
)
1448 #endif /* #else #ifdef CONFIG_SMP */
1451 * This does the RCU processing work from softirq context for the
1452 * specified rcu_state and rcu_data structures. This may be called
1453 * only from the CPU to whom the rdp belongs.
1456 __rcu_process_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1458 unsigned long flags
;
1460 WARN_ON_ONCE(rdp
->beenonline
== 0);
1463 * If an RCU GP has gone long enough, go check for dyntick
1464 * idle CPUs and, if needed, send resched IPIs.
1466 if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1467 force_quiescent_state(rsp
, 1);
1470 * Advance callbacks in response to end of earlier grace
1471 * period that some other CPU ended.
1473 rcu_process_gp_end(rsp
, rdp
);
1475 /* Update RCU state based on any recent quiescent states. */
1476 rcu_check_quiescent_state(rsp
, rdp
);
1478 /* Does this CPU require a not-yet-started grace period? */
1479 if (cpu_needs_another_gp(rsp
, rdp
)) {
1480 raw_spin_lock_irqsave(&rcu_get_root(rsp
)->lock
, flags
);
1481 rcu_start_gp(rsp
, flags
); /* releases above lock */
1484 /* If there are callbacks ready, invoke them. */
1485 if (cpu_has_callbacks_ready_to_invoke(rdp
))
1486 invoke_rcu_callbacks(rsp
, rdp
);
1490 * Do softirq processing for the current CPU.
1492 static void rcu_process_callbacks(struct softirq_action
*unused
)
1494 trace_rcu_utilization("Start RCU core");
1495 __rcu_process_callbacks(&rcu_sched_state
,
1496 &__get_cpu_var(rcu_sched_data
));
1497 __rcu_process_callbacks(&rcu_bh_state
, &__get_cpu_var(rcu_bh_data
));
1498 rcu_preempt_process_callbacks();
1500 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1501 rcu_needs_cpu_flush();
1502 trace_rcu_utilization("End RCU core");
1506 * Wake up the current CPU's kthread. This replaces raise_softirq()
1507 * in earlier versions of RCU. Note that because we are running on
1508 * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1509 * cannot disappear out from under us.
1511 static void invoke_rcu_callbacks(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1513 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active
)))
1515 if (likely(!rsp
->boost
)) {
1516 rcu_do_batch(rsp
, rdp
);
1519 invoke_rcu_callbacks_kthread();
1522 static void invoke_rcu_core(void)
1524 raise_softirq(RCU_SOFTIRQ
);
1528 __call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
),
1529 struct rcu_state
*rsp
)
1531 unsigned long flags
;
1532 struct rcu_data
*rdp
;
1534 debug_rcu_head_queue(head
);
1538 smp_mb(); /* Ensure RCU update seen before callback registry. */
1541 * Opportunistically note grace-period endings and beginnings.
1542 * Note that we might see a beginning right after we see an
1543 * end, but never vice versa, since this CPU has to pass through
1544 * a quiescent state betweentimes.
1546 local_irq_save(flags
);
1547 rdp
= this_cpu_ptr(rsp
->rda
);
1549 /* Add the callback to our list. */
1550 *rdp
->nxttail
[RCU_NEXT_TAIL
] = head
;
1551 rdp
->nxttail
[RCU_NEXT_TAIL
] = &head
->next
;
1554 /* If interrupts were disabled, don't dive into RCU core. */
1555 if (irqs_disabled_flags(flags
)) {
1556 local_irq_restore(flags
);
1561 * Force the grace period if too many callbacks or too long waiting.
1562 * Enforce hysteresis, and don't invoke force_quiescent_state()
1563 * if some other CPU has recently done so. Also, don't bother
1564 * invoking force_quiescent_state() if the newly enqueued callback
1565 * is the only one waiting for a grace period to complete.
1567 if (unlikely(rdp
->qlen
> rdp
->qlen_last_fqs_check
+ qhimark
)) {
1569 /* Are we ignoring a completed grace period? */
1570 rcu_process_gp_end(rsp
, rdp
);
1571 check_for_new_grace_period(rsp
, rdp
);
1573 /* Start a new grace period if one not already started. */
1574 if (!rcu_gp_in_progress(rsp
)) {
1575 unsigned long nestflag
;
1576 struct rcu_node
*rnp_root
= rcu_get_root(rsp
);
1578 raw_spin_lock_irqsave(&rnp_root
->lock
, nestflag
);
1579 rcu_start_gp(rsp
, nestflag
); /* rlses rnp_root->lock */
1581 /* Give the grace period a kick. */
1582 rdp
->blimit
= LONG_MAX
;
1583 if (rsp
->n_force_qs
== rdp
->n_force_qs_snap
&&
1584 *rdp
->nxttail
[RCU_DONE_TAIL
] != head
)
1585 force_quiescent_state(rsp
, 0);
1586 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1587 rdp
->qlen_last_fqs_check
= rdp
->qlen
;
1589 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
))
1590 force_quiescent_state(rsp
, 1);
1591 local_irq_restore(flags
);
1595 * Queue an RCU-sched callback for invocation after a grace period.
1597 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1599 __call_rcu(head
, func
, &rcu_sched_state
);
1601 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1604 * Queue an RCU for invocation after a quicker grace period.
1606 void call_rcu_bh(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1608 __call_rcu(head
, func
, &rcu_bh_state
);
1610 EXPORT_SYMBOL_GPL(call_rcu_bh
);
1613 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1615 * Control will return to the caller some time after a full rcu-sched
1616 * grace period has elapsed, in other words after all currently executing
1617 * rcu-sched read-side critical sections have completed. These read-side
1618 * critical sections are delimited by rcu_read_lock_sched() and
1619 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1620 * local_irq_disable(), and so on may be used in place of
1621 * rcu_read_lock_sched().
1623 * This means that all preempt_disable code sequences, including NMI and
1624 * hardware-interrupt handlers, in progress on entry will have completed
1625 * before this primitive returns. However, this does not guarantee that
1626 * softirq handlers will have completed, since in some kernels, these
1627 * handlers can run in process context, and can block.
1629 * This primitive provides the guarantees made by the (now removed)
1630 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1631 * guarantees that rcu_read_lock() sections will have completed.
1632 * In "classic RCU", these two guarantees happen to be one and
1633 * the same, but can differ in realtime RCU implementations.
1635 void synchronize_sched(void)
1637 if (rcu_blocking_is_gp())
1639 wait_rcu_gp(call_rcu_sched
);
1641 EXPORT_SYMBOL_GPL(synchronize_sched
);
1644 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1646 * Control will return to the caller some time after a full rcu_bh grace
1647 * period has elapsed, in other words after all currently executing rcu_bh
1648 * read-side critical sections have completed. RCU read-side critical
1649 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1650 * and may be nested.
1652 void synchronize_rcu_bh(void)
1654 if (rcu_blocking_is_gp())
1656 wait_rcu_gp(call_rcu_bh
);
1658 EXPORT_SYMBOL_GPL(synchronize_rcu_bh
);
1661 * Check to see if there is any immediate RCU-related work to be done
1662 * by the current CPU, for the specified type of RCU, returning 1 if so.
1663 * The checks are in order of increasing expense: checks that can be
1664 * carried out against CPU-local state are performed first. However,
1665 * we must check for CPU stalls first, else we might not get a chance.
1667 static int __rcu_pending(struct rcu_state
*rsp
, struct rcu_data
*rdp
)
1669 struct rcu_node
*rnp
= rdp
->mynode
;
1671 rdp
->n_rcu_pending
++;
1673 /* Check for CPU stalls, if enabled. */
1674 check_cpu_stall(rsp
, rdp
);
1676 /* Is the RCU core waiting for a quiescent state from this CPU? */
1677 if (rdp
->qs_pending
&& !rdp
->passed_quiesc
) {
1680 * If force_quiescent_state() coming soon and this CPU
1681 * needs a quiescent state, and this is either RCU-sched
1682 * or RCU-bh, force a local reschedule.
1684 rdp
->n_rp_qs_pending
++;
1685 if (!rdp
->preemptible
&&
1686 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
) - 1,
1689 } else if (rdp
->qs_pending
&& rdp
->passed_quiesc
) {
1690 rdp
->n_rp_report_qs
++;
1694 /* Does this CPU have callbacks ready to invoke? */
1695 if (cpu_has_callbacks_ready_to_invoke(rdp
)) {
1696 rdp
->n_rp_cb_ready
++;
1700 /* Has RCU gone idle with this CPU needing another grace period? */
1701 if (cpu_needs_another_gp(rsp
, rdp
)) {
1702 rdp
->n_rp_cpu_needs_gp
++;
1706 /* Has another RCU grace period completed? */
1707 if (ACCESS_ONCE(rnp
->completed
) != rdp
->completed
) { /* outside lock */
1708 rdp
->n_rp_gp_completed
++;
1712 /* Has a new RCU grace period started? */
1713 if (ACCESS_ONCE(rnp
->gpnum
) != rdp
->gpnum
) { /* outside lock */
1714 rdp
->n_rp_gp_started
++;
1718 /* Has an RCU GP gone long enough to send resched IPIs &c? */
1719 if (rcu_gp_in_progress(rsp
) &&
1720 ULONG_CMP_LT(ACCESS_ONCE(rsp
->jiffies_force_qs
), jiffies
)) {
1721 rdp
->n_rp_need_fqs
++;
1726 rdp
->n_rp_need_nothing
++;
1731 * Check to see if there is any immediate RCU-related work to be done
1732 * by the current CPU, returning 1 if so. This function is part of the
1733 * RCU implementation; it is -not- an exported member of the RCU API.
1735 static int rcu_pending(int cpu
)
1737 return __rcu_pending(&rcu_sched_state
, &per_cpu(rcu_sched_data
, cpu
)) ||
1738 __rcu_pending(&rcu_bh_state
, &per_cpu(rcu_bh_data
, cpu
)) ||
1739 rcu_preempt_pending(cpu
);
1743 * Check to see if any future RCU-related work will need to be done
1744 * by the current CPU, even if none need be done immediately, returning
1747 static int rcu_needs_cpu_quick_check(int cpu
)
1749 /* RCU callbacks either ready or pending? */
1750 return per_cpu(rcu_sched_data
, cpu
).nxtlist
||
1751 per_cpu(rcu_bh_data
, cpu
).nxtlist
||
1752 rcu_preempt_needs_cpu(cpu
);
1755 static DEFINE_PER_CPU(struct rcu_head
, rcu_barrier_head
) = {NULL
};
1756 static atomic_t rcu_barrier_cpu_count
;
1757 static DEFINE_MUTEX(rcu_barrier_mutex
);
1758 static struct completion rcu_barrier_completion
;
1760 static void rcu_barrier_callback(struct rcu_head
*notused
)
1762 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1763 complete(&rcu_barrier_completion
);
1767 * Called with preemption disabled, and from cross-cpu IRQ context.
1769 static void rcu_barrier_func(void *type
)
1771 int cpu
= smp_processor_id();
1772 struct rcu_head
*head
= &per_cpu(rcu_barrier_head
, cpu
);
1773 void (*call_rcu_func
)(struct rcu_head
*head
,
1774 void (*func
)(struct rcu_head
*head
));
1776 atomic_inc(&rcu_barrier_cpu_count
);
1777 call_rcu_func
= type
;
1778 call_rcu_func(head
, rcu_barrier_callback
);
1782 * Orchestrate the specified type of RCU barrier, waiting for all
1783 * RCU callbacks of the specified type to complete.
1785 static void _rcu_barrier(struct rcu_state
*rsp
,
1786 void (*call_rcu_func
)(struct rcu_head
*head
,
1787 void (*func
)(struct rcu_head
*head
)))
1789 BUG_ON(in_interrupt());
1790 /* Take mutex to serialize concurrent rcu_barrier() requests. */
1791 mutex_lock(&rcu_barrier_mutex
);
1792 init_completion(&rcu_barrier_completion
);
1794 * Initialize rcu_barrier_cpu_count to 1, then invoke
1795 * rcu_barrier_func() on each CPU, so that each CPU also has
1796 * incremented rcu_barrier_cpu_count. Only then is it safe to
1797 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1798 * might complete its grace period before all of the other CPUs
1799 * did their increment, causing this function to return too
1800 * early. Note that on_each_cpu() disables irqs, which prevents
1801 * any CPUs from coming online or going offline until each online
1802 * CPU has queued its RCU-barrier callback.
1804 atomic_set(&rcu_barrier_cpu_count
, 1);
1805 on_each_cpu(rcu_barrier_func
, (void *)call_rcu_func
, 1);
1806 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
1807 complete(&rcu_barrier_completion
);
1808 wait_for_completion(&rcu_barrier_completion
);
1809 mutex_unlock(&rcu_barrier_mutex
);
1813 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1815 void rcu_barrier_bh(void)
1817 _rcu_barrier(&rcu_bh_state
, call_rcu_bh
);
1819 EXPORT_SYMBOL_GPL(rcu_barrier_bh
);
1822 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1824 void rcu_barrier_sched(void)
1826 _rcu_barrier(&rcu_sched_state
, call_rcu_sched
);
1828 EXPORT_SYMBOL_GPL(rcu_barrier_sched
);
1831 * Do boot-time initialization of a CPU's per-CPU RCU data.
1834 rcu_boot_init_percpu_data(int cpu
, struct rcu_state
*rsp
)
1836 unsigned long flags
;
1838 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1839 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1841 /* Set up local state, ensuring consistent view of global state. */
1842 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1843 rdp
->grpmask
= 1UL << (cpu
- rdp
->mynode
->grplo
);
1844 rdp
->nxtlist
= NULL
;
1845 for (i
= 0; i
< RCU_NEXT_SIZE
; i
++)
1846 rdp
->nxttail
[i
] = &rdp
->nxtlist
;
1849 rdp
->dynticks
= &per_cpu(rcu_dynticks
, cpu
);
1850 #endif /* #ifdef CONFIG_NO_HZ */
1852 raw_spin_unlock_irqrestore(&rnp
->lock
, flags
);
1856 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1857 * offline event can be happening at a given time. Note also that we
1858 * can accept some slop in the rsp->completed access due to the fact
1859 * that this CPU cannot possibly have any RCU callbacks in flight yet.
1861 static void __cpuinit
1862 rcu_init_percpu_data(int cpu
, struct rcu_state
*rsp
, int preemptible
)
1864 unsigned long flags
;
1866 struct rcu_data
*rdp
= per_cpu_ptr(rsp
->rda
, cpu
);
1867 struct rcu_node
*rnp
= rcu_get_root(rsp
);
1869 /* Set up local state, ensuring consistent view of global state. */
1870 raw_spin_lock_irqsave(&rnp
->lock
, flags
);
1871 rdp
->passed_quiesc
= 0; /* We could be racing with new GP, */
1872 rdp
->qs_pending
= 1; /* so set up to respond to current GP. */
1873 rdp
->beenonline
= 1; /* We have now been online. */
1874 rdp
->preemptible
= preemptible
;
1875 rdp
->qlen_last_fqs_check
= 0;
1876 rdp
->n_force_qs_snap
= rsp
->n_force_qs
;
1877 rdp
->blimit
= blimit
;
1878 raw_spin_unlock(&rnp
->lock
); /* irqs remain disabled. */
1881 * A new grace period might start here. If so, we won't be part
1882 * of it, but that is OK, as we are currently in a quiescent state.
1885 /* Exclude any attempts to start a new GP on large systems. */
1886 raw_spin_lock(&rsp
->onofflock
); /* irqs already disabled. */
1888 /* Add CPU to rcu_node bitmasks. */
1890 mask
= rdp
->grpmask
;
1892 /* Exclude any attempts to start a new GP on small systems. */
1893 raw_spin_lock(&rnp
->lock
); /* irqs already disabled. */
1894 rnp
->qsmaskinit
|= mask
;
1895 mask
= rnp
->grpmask
;
1896 if (rnp
== rdp
->mynode
) {
1897 rdp
->gpnum
= rnp
->completed
; /* if GP in progress... */
1898 rdp
->completed
= rnp
->completed
;
1899 rdp
->passed_quiesc_completed
= rnp
->completed
- 1;
1901 raw_spin_unlock(&rnp
->lock
); /* irqs already disabled. */
1903 } while (rnp
!= NULL
&& !(rnp
->qsmaskinit
& mask
));
1905 raw_spin_unlock_irqrestore(&rsp
->onofflock
, flags
);
1908 static void __cpuinit
rcu_prepare_cpu(int cpu
)
1910 rcu_init_percpu_data(cpu
, &rcu_sched_state
, 0);
1911 rcu_init_percpu_data(cpu
, &rcu_bh_state
, 0);
1912 rcu_preempt_init_percpu_data(cpu
);
1916 * Handle CPU online/offline notification events.
1918 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
1919 unsigned long action
, void *hcpu
)
1921 long cpu
= (long)hcpu
;
1922 struct rcu_data
*rdp
= per_cpu_ptr(rcu_state
->rda
, cpu
);
1923 struct rcu_node
*rnp
= rdp
->mynode
;
1925 trace_rcu_utilization("Start CPU hotplug");
1927 case CPU_UP_PREPARE
:
1928 case CPU_UP_PREPARE_FROZEN
:
1929 rcu_prepare_cpu(cpu
);
1930 rcu_prepare_kthreads(cpu
);
1933 case CPU_DOWN_FAILED
:
1934 rcu_node_kthread_setaffinity(rnp
, -1);
1935 rcu_cpu_kthread_setrt(cpu
, 1);
1937 case CPU_DOWN_PREPARE
:
1938 rcu_node_kthread_setaffinity(rnp
, cpu
);
1939 rcu_cpu_kthread_setrt(cpu
, 0);
1942 case CPU_DYING_FROZEN
:
1944 * The whole machine is "stopped" except this CPU, so we can
1945 * touch any data without introducing corruption. We send the
1946 * dying CPU's callbacks to an arbitrarily chosen online CPU.
1948 rcu_send_cbs_to_online(&rcu_bh_state
);
1949 rcu_send_cbs_to_online(&rcu_sched_state
);
1950 rcu_preempt_send_cbs_to_online();
1953 case CPU_DEAD_FROZEN
:
1954 case CPU_UP_CANCELED
:
1955 case CPU_UP_CANCELED_FROZEN
:
1956 rcu_offline_cpu(cpu
);
1961 trace_rcu_utilization("End CPU hotplug");
1966 * This function is invoked towards the end of the scheduler's initialization
1967 * process. Before this is called, the idle task might contain
1968 * RCU read-side critical sections (during which time, this idle
1969 * task is booting the system). After this function is called, the
1970 * idle tasks are prohibited from containing RCU read-side critical
1971 * sections. This function also enables RCU lockdep checking.
1973 void rcu_scheduler_starting(void)
1975 WARN_ON(num_online_cpus() != 1);
1976 WARN_ON(nr_context_switches() > 0);
1977 rcu_scheduler_active
= 1;
1981 * Compute the per-level fanout, either using the exact fanout specified
1982 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1984 #ifdef CONFIG_RCU_FANOUT_EXACT
1985 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
1989 for (i
= NUM_RCU_LVLS
- 1; i
> 0; i
--)
1990 rsp
->levelspread
[i
] = CONFIG_RCU_FANOUT
;
1991 rsp
->levelspread
[0] = RCU_FANOUT_LEAF
;
1993 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1994 static void __init
rcu_init_levelspread(struct rcu_state
*rsp
)
2001 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2002 ccur
= rsp
->levelcnt
[i
];
2003 rsp
->levelspread
[i
] = (cprv
+ ccur
- 1) / ccur
;
2007 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2010 * Helper function for rcu_init() that initializes one rcu_state structure.
2012 static void __init
rcu_init_one(struct rcu_state
*rsp
,
2013 struct rcu_data __percpu
*rda
)
2015 static char *buf
[] = { "rcu_node_level_0",
2018 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
2022 struct rcu_node
*rnp
;
2024 BUILD_BUG_ON(MAX_RCU_LVLS
> ARRAY_SIZE(buf
)); /* Fix buf[] init! */
2026 /* Initialize the level-tracking arrays. */
2028 for (i
= 1; i
< NUM_RCU_LVLS
; i
++)
2029 rsp
->level
[i
] = rsp
->level
[i
- 1] + rsp
->levelcnt
[i
- 1];
2030 rcu_init_levelspread(rsp
);
2032 /* Initialize the elements themselves, starting from the leaves. */
2034 for (i
= NUM_RCU_LVLS
- 1; i
>= 0; i
--) {
2035 cpustride
*= rsp
->levelspread
[i
];
2036 rnp
= rsp
->level
[i
];
2037 for (j
= 0; j
< rsp
->levelcnt
[i
]; j
++, rnp
++) {
2038 raw_spin_lock_init(&rnp
->lock
);
2039 lockdep_set_class_and_name(&rnp
->lock
,
2040 &rcu_node_class
[i
], buf
[i
]);
2043 rnp
->qsmaskinit
= 0;
2044 rnp
->grplo
= j
* cpustride
;
2045 rnp
->grphi
= (j
+ 1) * cpustride
- 1;
2046 if (rnp
->grphi
>= NR_CPUS
)
2047 rnp
->grphi
= NR_CPUS
- 1;
2053 rnp
->grpnum
= j
% rsp
->levelspread
[i
- 1];
2054 rnp
->grpmask
= 1UL << rnp
->grpnum
;
2055 rnp
->parent
= rsp
->level
[i
- 1] +
2056 j
/ rsp
->levelspread
[i
- 1];
2059 INIT_LIST_HEAD(&rnp
->blkd_tasks
);
2064 rnp
= rsp
->level
[NUM_RCU_LVLS
- 1];
2065 for_each_possible_cpu(i
) {
2066 while (i
> rnp
->grphi
)
2068 per_cpu_ptr(rsp
->rda
, i
)->mynode
= rnp
;
2069 rcu_boot_init_percpu_data(i
, rsp
);
2073 void __init
rcu_init(void)
2077 rcu_bootup_announce();
2078 rcu_init_one(&rcu_sched_state
, &rcu_sched_data
);
2079 rcu_init_one(&rcu_bh_state
, &rcu_bh_data
);
2080 __rcu_init_preempt();
2081 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
2084 * We don't need protection against CPU-hotplug here because
2085 * this is called early in boot, before either interrupts
2086 * or the scheduler are operational.
2088 cpu_notifier(rcu_cpu_notify
, 0);
2089 for_each_online_cpu(cpu
)
2090 rcu_cpu_notify(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
2091 check_cpu_stall_init();
2094 #include "rcutree_plugin.h"