2 * Read-Copy Update mechanism for mutual exclusion, realtime implementation
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, 2006
20 * Authors: Paul E. McKenney <paulmck@us.ibm.com>
21 * With thanks to Esben Nielsen, Bill Huey, and Ingo Molnar
22 * for pushing me away from locks and towards counters, and
23 * to Suparna Bhattacharya for pushing me completely away
24 * from atomic instructions on the read side.
26 * - Added handling of Dynamic Ticks
27 * Copyright 2007 - Paul E. Mckenney <paulmck@us.ibm.com>
28 * - Steven Rostedt <srostedt@redhat.com>
30 * Papers: http://www.rdrop.com/users/paulmck/RCU
32 * Design Document: http://lwn.net/Articles/253651/
34 * For detailed explanation of Read-Copy Update mechanism see -
35 * Documentation/RCU/ *.txt
38 #include <linux/types.h>
39 #include <linux/kernel.h>
40 #include <linux/init.h>
41 #include <linux/spinlock.h>
42 #include <linux/smp.h>
43 #include <linux/rcupdate.h>
44 #include <linux/interrupt.h>
45 #include <linux/sched.h>
46 #include <asm/atomic.h>
47 #include <linux/bitops.h>
48 #include <linux/module.h>
49 #include <linux/kthread.h>
50 #include <linux/completion.h>
51 #include <linux/moduleparam.h>
52 #include <linux/percpu.h>
53 #include <linux/notifier.h>
54 #include <linux/cpu.h>
55 #include <linux/random.h>
56 #include <linux/delay.h>
57 #include <linux/cpumask.h>
58 #include <linux/rcupreempt_trace.h>
59 #include <asm/byteorder.h>
62 * PREEMPT_RCU data structures.
66 * GP_STAGES specifies the number of times the state machine has
67 * to go through the all the rcu_try_flip_states (see below)
68 * in a single Grace Period.
70 * GP in GP_STAGES stands for Grace Period ;)
74 raw_spinlock_t lock
; /* Protect rcu_data fields. */
75 long completed
; /* Number of last completed batch. */
77 struct rcu_head
*nextlist
;
78 struct rcu_head
**nexttail
;
79 struct rcu_head
*waitlist
[GP_STAGES
];
80 struct rcu_head
**waittail
[GP_STAGES
];
81 struct rcu_head
*donelist
; /* from waitlist & waitschedlist */
82 struct rcu_head
**donetail
;
84 struct rcu_head
*nextschedlist
;
85 struct rcu_head
**nextschedtail
;
86 struct rcu_head
*waitschedlist
;
87 struct rcu_head
**waitschedtail
;
88 int rcu_sched_sleeping
;
89 #ifdef CONFIG_RCU_TRACE
90 struct rcupreempt_trace trace
;
91 #endif /* #ifdef CONFIG_RCU_TRACE */
95 * States for rcu_try_flip() and friends.
98 enum rcu_try_flip_states
{
101 * Stay here if nothing is happening. Flip the counter if somthing
102 * starts happening. Denoted by "I"
104 rcu_try_flip_idle_state
,
107 * Wait here for all CPUs to notice that the counter has flipped. This
108 * prevents the old set of counters from ever being incremented once
109 * we leave this state, which in turn is necessary because we cannot
110 * test any individual counter for zero -- we can only check the sum.
113 rcu_try_flip_waitack_state
,
116 * Wait here for the sum of the old per-CPU counters to reach zero.
119 rcu_try_flip_waitzero_state
,
122 * Wait here for each of the other CPUs to execute a memory barrier.
123 * This is necessary to ensure that these other CPUs really have
124 * completed executing their RCU read-side critical sections, despite
125 * their CPUs wildly reordering memory. Denoted by "M".
127 rcu_try_flip_waitmb_state
,
131 * States for rcu_ctrlblk.rcu_sched_sleep.
134 enum rcu_sched_sleep_states
{
135 rcu_sched_not_sleeping
, /* Not sleeping, callbacks need GP. */
136 rcu_sched_sleep_prep
, /* Thinking of sleeping, rechecking. */
137 rcu_sched_sleeping
, /* Sleeping, awaken if GP needed. */
141 raw_spinlock_t fliplock
; /* Protect state-machine transitions. */
142 long completed
; /* Number of last completed batch. */
143 enum rcu_try_flip_states rcu_try_flip_state
; /* The current state of
144 the rcu state machine */
145 spinlock_t schedlock
; /* Protect rcu_sched sleep state. */
146 enum rcu_sched_sleep_states sched_sleep
; /* rcu_sched state. */
147 wait_queue_head_t sched_wq
; /* Place for rcu_sched to sleep. */
150 struct rcu_dyntick_sched
{
155 int sched_dynticks_snap
;
158 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_dyntick_sched
, rcu_dyntick_sched
) = {
162 void rcu_qsctr_inc(int cpu
)
164 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
171 void rcu_enter_nohz(void)
173 static DEFINE_RATELIMIT_STATE(rs
, 10 * HZ
, 1);
175 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
176 __get_cpu_var(rcu_dyntick_sched
).dynticks
++;
177 WARN_ON_RATELIMIT(__get_cpu_var(rcu_dyntick_sched
).dynticks
& 0x1, &rs
);
180 void rcu_exit_nohz(void)
182 static DEFINE_RATELIMIT_STATE(rs
, 10 * HZ
, 1);
184 __get_cpu_var(rcu_dyntick_sched
).dynticks
++;
185 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
186 WARN_ON_RATELIMIT(!(__get_cpu_var(rcu_dyntick_sched
).dynticks
& 0x1),
190 #endif /* CONFIG_NO_HZ */
193 static DEFINE_PER_CPU(struct rcu_data
, rcu_data
);
195 static struct rcu_ctrlblk rcu_ctrlblk
= {
196 .fliplock
= RAW_SPIN_LOCK_UNLOCKED(rcu_ctrlblk
.fliplock
),
198 .rcu_try_flip_state
= rcu_try_flip_idle_state
,
199 .schedlock
= __SPIN_LOCK_UNLOCKED(rcu_ctrlblk
.schedlock
),
200 .sched_sleep
= rcu_sched_not_sleeping
,
201 .sched_wq
= __WAIT_QUEUE_HEAD_INITIALIZER(rcu_ctrlblk
.sched_wq
),
204 static struct task_struct
*rcu_sched_grace_period_task
;
206 #ifdef CONFIG_RCU_TRACE
207 static char *rcu_try_flip_state_names
[] =
208 { "idle", "waitack", "waitzero", "waitmb" };
209 #endif /* #ifdef CONFIG_RCU_TRACE */
211 static DECLARE_BITMAP(rcu_cpu_online_map
, NR_CPUS
) __read_mostly
215 * Enum and per-CPU flag to determine when each CPU has seen
216 * the most recent counter flip.
219 enum rcu_flip_flag_values
{
220 rcu_flip_seen
, /* Steady/initial state, last flip seen. */
221 /* Only GP detector can update. */
222 rcu_flipped
/* Flip just completed, need confirmation. */
223 /* Only corresponding CPU can update. */
225 static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_flip_flag_values
, rcu_flip_flag
)
229 * Enum and per-CPU flag to determine when each CPU has executed the
230 * needed memory barrier to fence in memory references from its last RCU
231 * read-side critical section in the just-completed grace period.
234 enum rcu_mb_flag_values
{
235 rcu_mb_done
, /* Steady/initial state, no mb()s required. */
236 /* Only GP detector can update. */
237 rcu_mb_needed
/* Flip just completed, need an mb(). */
238 /* Only corresponding CPU can update. */
240 static DEFINE_PER_CPU_SHARED_ALIGNED(enum rcu_mb_flag_values
, rcu_mb_flag
)
244 * RCU_DATA_ME: find the current CPU's rcu_data structure.
245 * RCU_DATA_CPU: find the specified CPU's rcu_data structure.
247 #define RCU_DATA_ME() (&__get_cpu_var(rcu_data))
248 #define RCU_DATA_CPU(cpu) (&per_cpu(rcu_data, cpu))
251 * Helper macro for tracing when the appropriate rcu_data is not
252 * cached in a local variable, but where the CPU number is so cached.
254 #define RCU_TRACE_CPU(f, cpu) RCU_TRACE(f, &(RCU_DATA_CPU(cpu)->trace));
257 * Helper macro for tracing when the appropriate rcu_data is not
258 * cached in a local variable.
260 #define RCU_TRACE_ME(f) RCU_TRACE(f, &(RCU_DATA_ME()->trace));
263 * Helper macro for tracing when the appropriate rcu_data is pointed
264 * to by a local variable.
266 #define RCU_TRACE_RDP(f, rdp) RCU_TRACE(f, &((rdp)->trace));
268 #define RCU_SCHED_BATCH_TIME (HZ / 50)
271 * Return the number of RCU batches processed thus far. Useful
272 * for debug and statistics.
274 long rcu_batches_completed(void)
276 return rcu_ctrlblk
.completed
;
278 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
280 void __rcu_read_lock(void)
283 struct task_struct
*t
= current
;
286 nesting
= ACCESS_ONCE(t
->rcu_read_lock_nesting
);
289 /* An earlier rcu_read_lock() covers us, just count it. */
291 t
->rcu_read_lock_nesting
= nesting
+ 1;
297 * We disable interrupts for the following reasons:
298 * - If we get scheduling clock interrupt here, and we
299 * end up acking the counter flip, it's like a promise
300 * that we will never increment the old counter again.
301 * Thus we will break that promise if that
302 * scheduling clock interrupt happens between the time
303 * we pick the .completed field and the time that we
304 * increment our counter.
306 * - We don't want to be preempted out here.
308 * NMIs can still occur, of course, and might themselves
309 * contain rcu_read_lock().
312 local_irq_save(flags
);
315 * Outermost nesting of rcu_read_lock(), so increment
316 * the current counter for the current CPU. Use volatile
317 * casts to prevent the compiler from reordering.
320 idx
= ACCESS_ONCE(rcu_ctrlblk
.completed
) & 0x1;
321 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr
[idx
])++;
324 * Now that the per-CPU counter has been incremented, we
325 * are protected from races with rcu_read_lock() invoked
326 * from NMI handlers on this CPU. We can therefore safely
327 * increment the nesting counter, relieving further NMIs
328 * of the need to increment the per-CPU counter.
331 ACCESS_ONCE(t
->rcu_read_lock_nesting
) = nesting
+ 1;
334 * Now that we have preventing any NMIs from storing
335 * to the ->rcu_flipctr_idx, we can safely use it to
336 * remember which counter to decrement in the matching
340 ACCESS_ONCE(t
->rcu_flipctr_idx
) = idx
;
341 local_irq_restore(flags
);
344 EXPORT_SYMBOL_GPL(__rcu_read_lock
);
346 void __rcu_read_unlock(void)
349 struct task_struct
*t
= current
;
352 nesting
= ACCESS_ONCE(t
->rcu_read_lock_nesting
);
356 * We are still protected by the enclosing rcu_read_lock(),
357 * so simply decrement the counter.
360 t
->rcu_read_lock_nesting
= nesting
- 1;
366 * Disable local interrupts to prevent the grace-period
367 * detection state machine from seeing us half-done.
368 * NMIs can still occur, of course, and might themselves
369 * contain rcu_read_lock() and rcu_read_unlock().
372 local_irq_save(flags
);
375 * Outermost nesting of rcu_read_unlock(), so we must
376 * decrement the current counter for the current CPU.
377 * This must be done carefully, because NMIs can
378 * occur at any point in this code, and any rcu_read_lock()
379 * and rcu_read_unlock() pairs in the NMI handlers
380 * must interact non-destructively with this code.
381 * Lots of volatile casts, and -very- careful ordering.
383 * Changes to this code, including this one, must be
384 * inspected, validated, and tested extremely carefully!!!
388 * First, pick up the index.
391 idx
= ACCESS_ONCE(t
->rcu_flipctr_idx
);
394 * Now that we have fetched the counter index, it is
395 * safe to decrement the per-task RCU nesting counter.
396 * After this, any interrupts or NMIs will increment and
397 * decrement the per-CPU counters.
399 ACCESS_ONCE(t
->rcu_read_lock_nesting
) = nesting
- 1;
402 * It is now safe to decrement this task's nesting count.
403 * NMIs that occur after this statement will route their
404 * rcu_read_lock() calls through this "else" clause, and
405 * will thus start incrementing the per-CPU counter on
406 * their own. They will also clobber ->rcu_flipctr_idx,
407 * but that is OK, since we have already fetched it.
410 ACCESS_ONCE(RCU_DATA_ME()->rcu_flipctr
[idx
])--;
411 local_irq_restore(flags
);
414 EXPORT_SYMBOL_GPL(__rcu_read_unlock
);
417 * If a global counter flip has occurred since the last time that we
418 * advanced callbacks, advance them. Hardware interrupts must be
419 * disabled when calling this function.
421 static void __rcu_advance_callbacks(struct rcu_data
*rdp
)
427 if (rdp
->completed
!= rcu_ctrlblk
.completed
) {
428 if (rdp
->waitlist
[GP_STAGES
- 1] != NULL
) {
429 *rdp
->donetail
= rdp
->waitlist
[GP_STAGES
- 1];
430 rdp
->donetail
= rdp
->waittail
[GP_STAGES
- 1];
431 RCU_TRACE_RDP(rcupreempt_trace_move2done
, rdp
);
433 for (i
= GP_STAGES
- 2; i
>= 0; i
--) {
434 if (rdp
->waitlist
[i
] != NULL
) {
435 rdp
->waitlist
[i
+ 1] = rdp
->waitlist
[i
];
436 rdp
->waittail
[i
+ 1] = rdp
->waittail
[i
];
439 rdp
->waitlist
[i
+ 1] = NULL
;
440 rdp
->waittail
[i
+ 1] =
441 &rdp
->waitlist
[i
+ 1];
444 if (rdp
->nextlist
!= NULL
) {
445 rdp
->waitlist
[0] = rdp
->nextlist
;
446 rdp
->waittail
[0] = rdp
->nexttail
;
448 rdp
->nextlist
= NULL
;
449 rdp
->nexttail
= &rdp
->nextlist
;
450 RCU_TRACE_RDP(rcupreempt_trace_move2wait
, rdp
);
452 rdp
->waitlist
[0] = NULL
;
453 rdp
->waittail
[0] = &rdp
->waitlist
[0];
455 rdp
->waitlistcount
= wlc
;
456 rdp
->completed
= rcu_ctrlblk
.completed
;
460 * Check to see if this CPU needs to report that it has seen
461 * the most recent counter flip, thereby declaring that all
462 * subsequent rcu_read_lock() invocations will respect this flip.
465 cpu
= raw_smp_processor_id();
466 if (per_cpu(rcu_flip_flag
, cpu
) == rcu_flipped
) {
467 smp_mb(); /* Subsequent counter accesses must see new value */
468 per_cpu(rcu_flip_flag
, cpu
) = rcu_flip_seen
;
469 smp_mb(); /* Subsequent RCU read-side critical sections */
470 /* seen -after- acknowledgement. */
475 static DEFINE_PER_CPU(int, rcu_update_flag
);
478 * rcu_irq_enter - Called from Hard irq handlers and NMI/SMI.
480 * If the CPU was idle with dynamic ticks active, this updates the
481 * rcu_dyntick_sched.dynticks to let the RCU handling know that the
484 void rcu_irq_enter(void)
486 int cpu
= smp_processor_id();
487 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
489 if (per_cpu(rcu_update_flag
, cpu
))
490 per_cpu(rcu_update_flag
, cpu
)++;
493 * Only update if we are coming from a stopped ticks mode
494 * (rcu_dyntick_sched.dynticks is even).
496 if (!in_interrupt() &&
497 (rdssp
->dynticks
& 0x1) == 0) {
499 * The following might seem like we could have a race
500 * with NMI/SMIs. But this really isn't a problem.
501 * Here we do a read/modify/write, and the race happens
502 * when an NMI/SMI comes in after the read and before
503 * the write. But NMI/SMIs will increment this counter
504 * twice before returning, so the zero bit will not
505 * be corrupted by the NMI/SMI which is the most important
508 * The only thing is that we would bring back the counter
509 * to a postion that it was in during the NMI/SMI.
510 * But the zero bit would be set, so the rest of the
511 * counter would again be ignored.
513 * On return from the IRQ, the counter may have the zero
514 * bit be 0 and the counter the same as the return from
515 * the NMI/SMI. If the state machine was so unlucky to
516 * see that, it still doesn't matter, since all
517 * RCU read-side critical sections on this CPU would
518 * have already completed.
522 * The following memory barrier ensures that any
523 * rcu_read_lock() primitives in the irq handler
524 * are seen by other CPUs to follow the above
525 * increment to rcu_dyntick_sched.dynticks. This is
526 * required in order for other CPUs to correctly
527 * determine when it is safe to advance the RCU
528 * grace-period state machine.
530 smp_mb(); /* see above block comment. */
532 * Since we can't determine the dynamic tick mode from
533 * the rcu_dyntick_sched.dynticks after this routine,
534 * we use a second flag to acknowledge that we came
535 * from an idle state with ticks stopped.
537 per_cpu(rcu_update_flag
, cpu
)++;
539 * If we take an NMI/SMI now, they will also increment
540 * the rcu_update_flag, and will not update the
541 * rcu_dyntick_sched.dynticks on exit. That is for
548 * rcu_irq_exit - Called from exiting Hard irq context.
550 * If the CPU was idle with dynamic ticks active, update the
551 * rcu_dyntick_sched.dynticks to put let the RCU handling be
552 * aware that the CPU is going back to idle with no ticks.
554 void rcu_irq_exit(void)
556 int cpu
= smp_processor_id();
557 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
560 * rcu_update_flag is set if we interrupted the CPU
561 * when it was idle with ticks stopped.
562 * Once this occurs, we keep track of interrupt nesting
563 * because a NMI/SMI could also come in, and we still
564 * only want the IRQ that started the increment of the
565 * rcu_dyntick_sched.dynticks to be the one that modifies
568 if (per_cpu(rcu_update_flag
, cpu
)) {
569 if (--per_cpu(rcu_update_flag
, cpu
))
572 /* This must match the interrupt nesting */
573 WARN_ON(in_interrupt());
576 * If an NMI/SMI happens now we are still
577 * protected by the rcu_dyntick_sched.dynticks being odd.
581 * The following memory barrier ensures that any
582 * rcu_read_unlock() primitives in the irq handler
583 * are seen by other CPUs to preceed the following
584 * increment to rcu_dyntick_sched.dynticks. This
585 * is required in order for other CPUs to determine
586 * when it is safe to advance the RCU grace-period
589 smp_mb(); /* see above block comment. */
591 WARN_ON(rdssp
->dynticks
& 0x1);
595 void rcu_nmi_enter(void)
600 void rcu_nmi_exit(void)
605 static void dyntick_save_progress_counter(int cpu
)
607 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
609 rdssp
->dynticks_snap
= rdssp
->dynticks
;
613 rcu_try_flip_waitack_needed(int cpu
)
617 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
619 curr
= rdssp
->dynticks
;
620 snap
= rdssp
->dynticks_snap
;
621 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
624 * If the CPU remained in dynticks mode for the entire time
625 * and didn't take any interrupts, NMIs, SMIs, or whatever,
626 * then it cannot be in the middle of an rcu_read_lock(), so
627 * the next rcu_read_lock() it executes must use the new value
628 * of the counter. So we can safely pretend that this CPU
629 * already acknowledged the counter.
632 if ((curr
== snap
) && ((curr
& 0x1) == 0))
636 * If the CPU passed through or entered a dynticks idle phase with
637 * no active irq handlers, then, as above, we can safely pretend
638 * that this CPU already acknowledged the counter.
641 if ((curr
- snap
) > 2 || (curr
& 0x1) == 0)
644 /* We need this CPU to explicitly acknowledge the counter flip. */
650 rcu_try_flip_waitmb_needed(int cpu
)
654 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
656 curr
= rdssp
->dynticks
;
657 snap
= rdssp
->dynticks_snap
;
658 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
661 * If the CPU remained in dynticks mode for the entire time
662 * and didn't take any interrupts, NMIs, SMIs, or whatever,
663 * then it cannot have executed an RCU read-side critical section
664 * during that time, so there is no need for it to execute a
668 if ((curr
== snap
) && ((curr
& 0x1) == 0))
672 * If the CPU either entered or exited an outermost interrupt,
673 * SMI, NMI, or whatever handler, then we know that it executed
674 * a memory barrier when doing so. So we don't need another one.
679 /* We need the CPU to execute a memory barrier. */
684 static void dyntick_save_progress_counter_sched(int cpu
)
686 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
688 rdssp
->sched_dynticks_snap
= rdssp
->dynticks
;
691 static int rcu_qsctr_inc_needed_dyntick(int cpu
)
695 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
697 curr
= rdssp
->dynticks
;
698 snap
= rdssp
->sched_dynticks_snap
;
699 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
702 * If the CPU remained in dynticks mode for the entire time
703 * and didn't take any interrupts, NMIs, SMIs, or whatever,
704 * then it cannot be in the middle of an rcu_read_lock(), so
705 * the next rcu_read_lock() it executes must use the new value
706 * of the counter. Therefore, this CPU has been in a quiescent
707 * state the entire time, and we don't need to wait for it.
710 if ((curr
== snap
) && ((curr
& 0x1) == 0))
714 * If the CPU passed through or entered a dynticks idle phase with
715 * no active irq handlers, then, as above, this CPU has already
716 * passed through a quiescent state.
719 if ((curr
- snap
) > 2 || (snap
& 0x1) == 0)
722 /* We need this CPU to go through a quiescent state. */
727 #else /* !CONFIG_NO_HZ */
729 # define dyntick_save_progress_counter(cpu) do { } while (0)
730 # define rcu_try_flip_waitack_needed(cpu) (1)
731 # define rcu_try_flip_waitmb_needed(cpu) (1)
733 # define dyntick_save_progress_counter_sched(cpu) do { } while (0)
734 # define rcu_qsctr_inc_needed_dyntick(cpu) (1)
736 #endif /* CONFIG_NO_HZ */
738 static void save_qsctr_sched(int cpu
)
740 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
742 rdssp
->sched_qs_snap
= rdssp
->sched_qs
;
745 static inline int rcu_qsctr_inc_needed(int cpu
)
747 struct rcu_dyntick_sched
*rdssp
= &per_cpu(rcu_dyntick_sched
, cpu
);
750 * If there has been a quiescent state, no more need to wait
754 if (rdssp
->sched_qs
!= rdssp
->sched_qs_snap
) {
755 smp_mb(); /* force ordering with cpu entering schedule(). */
759 /* We need this CPU to go through a quiescent state. */
765 * Get here when RCU is idle. Decide whether we need to
766 * move out of idle state, and return non-zero if so.
767 * "Straightforward" approach for the moment, might later
768 * use callback-list lengths, grace-period duration, or
769 * some such to determine when to exit idle state.
770 * Might also need a pre-idle test that does not acquire
771 * the lock, but let's get the simple case working first...
775 rcu_try_flip_idle(void)
779 RCU_TRACE_ME(rcupreempt_trace_try_flip_i1
);
780 if (!rcu_pending(smp_processor_id())) {
781 RCU_TRACE_ME(rcupreempt_trace_try_flip_ie1
);
789 RCU_TRACE_ME(rcupreempt_trace_try_flip_g1
);
790 rcu_ctrlblk
.completed
++; /* stands in for rcu_try_flip_g2 */
793 * Need a memory barrier so that other CPUs see the new
794 * counter value before they see the subsequent change of all
795 * the rcu_flip_flag instances to rcu_flipped.
798 smp_mb(); /* see above block comment. */
800 /* Now ask each CPU for acknowledgement of the flip. */
802 for_each_cpu(cpu
, to_cpumask(rcu_cpu_online_map
)) {
803 per_cpu(rcu_flip_flag
, cpu
) = rcu_flipped
;
804 dyntick_save_progress_counter(cpu
);
811 * Wait for CPUs to acknowledge the flip.
815 rcu_try_flip_waitack(void)
819 RCU_TRACE_ME(rcupreempt_trace_try_flip_a1
);
820 for_each_cpu(cpu
, to_cpumask(rcu_cpu_online_map
))
821 if (rcu_try_flip_waitack_needed(cpu
) &&
822 per_cpu(rcu_flip_flag
, cpu
) != rcu_flip_seen
) {
823 RCU_TRACE_ME(rcupreempt_trace_try_flip_ae1
);
828 * Make sure our checks above don't bleed into subsequent
829 * waiting for the sum of the counters to reach zero.
832 smp_mb(); /* see above block comment. */
833 RCU_TRACE_ME(rcupreempt_trace_try_flip_a2
);
838 * Wait for collective ``last'' counter to reach zero,
839 * then tell all CPUs to do an end-of-grace-period memory barrier.
843 rcu_try_flip_waitzero(void)
846 int lastidx
= !(rcu_ctrlblk
.completed
& 0x1);
849 /* Check to see if the sum of the "last" counters is zero. */
851 RCU_TRACE_ME(rcupreempt_trace_try_flip_z1
);
852 for_each_cpu(cpu
, to_cpumask(rcu_cpu_online_map
))
853 sum
+= RCU_DATA_CPU(cpu
)->rcu_flipctr
[lastidx
];
855 RCU_TRACE_ME(rcupreempt_trace_try_flip_ze1
);
860 * This ensures that the other CPUs see the call for
861 * memory barriers -after- the sum to zero has been
864 smp_mb(); /* ^^^^^^^^^^^^ */
866 /* Call for a memory barrier from each CPU. */
867 for_each_cpu(cpu
, to_cpumask(rcu_cpu_online_map
)) {
868 per_cpu(rcu_mb_flag
, cpu
) = rcu_mb_needed
;
869 dyntick_save_progress_counter(cpu
);
872 RCU_TRACE_ME(rcupreempt_trace_try_flip_z2
);
877 * Wait for all CPUs to do their end-of-grace-period memory barrier.
878 * Return 0 once all CPUs have done so.
882 rcu_try_flip_waitmb(void)
886 RCU_TRACE_ME(rcupreempt_trace_try_flip_m1
);
887 for_each_cpu(cpu
, to_cpumask(rcu_cpu_online_map
))
888 if (rcu_try_flip_waitmb_needed(cpu
) &&
889 per_cpu(rcu_mb_flag
, cpu
) != rcu_mb_done
) {
890 RCU_TRACE_ME(rcupreempt_trace_try_flip_me1
);
894 smp_mb(); /* Ensure that the above checks precede any following flip. */
895 RCU_TRACE_ME(rcupreempt_trace_try_flip_m2
);
900 * Attempt a single flip of the counters. Remember, a single flip does
901 * -not- constitute a grace period. Instead, the interval between
902 * at least GP_STAGES consecutive flips is a grace period.
904 * If anyone is nuts enough to run this CONFIG_PREEMPT_RCU implementation
905 * on a large SMP, they might want to use a hierarchical organization of
906 * the per-CPU-counter pairs.
908 static void rcu_try_flip(void)
912 RCU_TRACE_ME(rcupreempt_trace_try_flip_1
);
913 if (unlikely(!spin_trylock_irqsave(&rcu_ctrlblk
.fliplock
, flags
))) {
914 RCU_TRACE_ME(rcupreempt_trace_try_flip_e1
);
919 * Take the next transition(s) through the RCU grace-period
920 * flip-counter state machine.
923 switch (rcu_ctrlblk
.rcu_try_flip_state
) {
924 case rcu_try_flip_idle_state
:
925 if (rcu_try_flip_idle())
926 rcu_ctrlblk
.rcu_try_flip_state
=
927 rcu_try_flip_waitack_state
;
929 case rcu_try_flip_waitack_state
:
930 if (rcu_try_flip_waitack())
931 rcu_ctrlblk
.rcu_try_flip_state
=
932 rcu_try_flip_waitzero_state
;
934 case rcu_try_flip_waitzero_state
:
935 if (rcu_try_flip_waitzero())
936 rcu_ctrlblk
.rcu_try_flip_state
=
937 rcu_try_flip_waitmb_state
;
939 case rcu_try_flip_waitmb_state
:
940 if (rcu_try_flip_waitmb())
941 rcu_ctrlblk
.rcu_try_flip_state
=
942 rcu_try_flip_idle_state
;
944 spin_unlock_irqrestore(&rcu_ctrlblk
.fliplock
, flags
);
948 * Check to see if this CPU needs to do a memory barrier in order to
949 * ensure that any prior RCU read-side critical sections have committed
950 * their counter manipulations and critical-section memory references
951 * before declaring the grace period to be completed.
953 static void rcu_check_mb(int cpu
)
955 if (per_cpu(rcu_mb_flag
, cpu
) == rcu_mb_needed
) {
956 smp_mb(); /* Ensure RCU read-side accesses are visible. */
957 per_cpu(rcu_mb_flag
, cpu
) = rcu_mb_done
;
961 void rcu_check_callbacks(int cpu
, int user
)
964 struct rcu_data
*rdp
= RCU_DATA_CPU(cpu
);
967 * If this CPU took its interrupt from user mode or from the
968 * idle loop, and this is not a nested interrupt, then
969 * this CPU has to have exited all prior preept-disable
970 * sections of code. So increment the counter to note this.
972 * The memory barrier is needed to handle the case where
973 * writes from a preempt-disable section of code get reordered
974 * into schedule() by this CPU's write buffer. So the memory
975 * barrier makes sure that the rcu_qsctr_inc() is seen by other
976 * CPUs to happen after any such write.
980 (idle_cpu(cpu
) && !in_softirq() &&
981 hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
982 smp_mb(); /* Guard against aggressive schedule(). */
987 if (rcu_ctrlblk
.completed
== rdp
->completed
)
989 spin_lock_irqsave(&rdp
->lock
, flags
);
990 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks
, rdp
);
991 __rcu_advance_callbacks(rdp
);
992 if (rdp
->donelist
== NULL
) {
993 spin_unlock_irqrestore(&rdp
->lock
, flags
);
995 spin_unlock_irqrestore(&rdp
->lock
, flags
);
996 raise_softirq(RCU_SOFTIRQ
);
1001 * Needed by dynticks, to make sure all RCU processing has finished
1004 void rcu_advance_callbacks(int cpu
, int user
)
1006 unsigned long flags
;
1007 struct rcu_data
*rdp
= RCU_DATA_CPU(cpu
);
1009 if (rcu_ctrlblk
.completed
== rdp
->completed
) {
1011 if (rcu_ctrlblk
.completed
== rdp
->completed
)
1014 spin_lock_irqsave(&rdp
->lock
, flags
);
1015 RCU_TRACE_RDP(rcupreempt_trace_check_callbacks
, rdp
);
1016 __rcu_advance_callbacks(rdp
);
1017 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1020 #ifdef CONFIG_HOTPLUG_CPU
1021 #define rcu_offline_cpu_enqueue(srclist, srctail, dstlist, dsttail) do { \
1022 *dsttail = srclist; \
1023 if (srclist != NULL) { \
1024 dsttail = srctail; \
1026 srctail = &srclist;\
1030 void rcu_offline_cpu(int cpu
)
1033 struct rcu_head
*list
= NULL
;
1034 unsigned long flags
;
1035 struct rcu_data
*rdp
= RCU_DATA_CPU(cpu
);
1036 struct rcu_head
*schedlist
= NULL
;
1037 struct rcu_head
**schedtail
= &schedlist
;
1038 struct rcu_head
**tail
= &list
;
1041 * Remove all callbacks from the newly dead CPU, retaining order.
1042 * Otherwise rcu_barrier() will fail
1045 spin_lock_irqsave(&rdp
->lock
, flags
);
1046 rcu_offline_cpu_enqueue(rdp
->donelist
, rdp
->donetail
, list
, tail
);
1047 for (i
= GP_STAGES
- 1; i
>= 0; i
--)
1048 rcu_offline_cpu_enqueue(rdp
->waitlist
[i
], rdp
->waittail
[i
],
1050 rcu_offline_cpu_enqueue(rdp
->nextlist
, rdp
->nexttail
, list
, tail
);
1051 rcu_offline_cpu_enqueue(rdp
->waitschedlist
, rdp
->waitschedtail
,
1052 schedlist
, schedtail
);
1053 rcu_offline_cpu_enqueue(rdp
->nextschedlist
, rdp
->nextschedtail
,
1054 schedlist
, schedtail
);
1055 rdp
->rcu_sched_sleeping
= 0;
1056 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1057 rdp
->waitlistcount
= 0;
1059 /* Disengage the newly dead CPU from the grace-period computation. */
1061 spin_lock_irqsave(&rcu_ctrlblk
.fliplock
, flags
);
1063 if (per_cpu(rcu_flip_flag
, cpu
) == rcu_flipped
) {
1064 smp_mb(); /* Subsequent counter accesses must see new value */
1065 per_cpu(rcu_flip_flag
, cpu
) = rcu_flip_seen
;
1066 smp_mb(); /* Subsequent RCU read-side critical sections */
1067 /* seen -after- acknowledgement. */
1070 RCU_DATA_ME()->rcu_flipctr
[0] += RCU_DATA_CPU(cpu
)->rcu_flipctr
[0];
1071 RCU_DATA_ME()->rcu_flipctr
[1] += RCU_DATA_CPU(cpu
)->rcu_flipctr
[1];
1073 RCU_DATA_CPU(cpu
)->rcu_flipctr
[0] = 0;
1074 RCU_DATA_CPU(cpu
)->rcu_flipctr
[1] = 0;
1076 cpumask_clear_cpu(cpu
, to_cpumask(rcu_cpu_online_map
));
1078 spin_unlock_irqrestore(&rcu_ctrlblk
.fliplock
, flags
);
1081 * Place the removed callbacks on the current CPU's queue.
1082 * Make them all start a new grace period: simple approach,
1083 * in theory could starve a given set of callbacks, but
1084 * you would need to be doing some serious CPU hotplugging
1085 * to make this happen. If this becomes a problem, adding
1086 * a synchronize_rcu() to the hotplug path would be a simple
1090 local_irq_save(flags
); /* disable preempt till we know what lock. */
1091 rdp
= RCU_DATA_ME();
1092 spin_lock(&rdp
->lock
);
1093 *rdp
->nexttail
= list
;
1095 rdp
->nexttail
= tail
;
1096 *rdp
->nextschedtail
= schedlist
;
1098 rdp
->nextschedtail
= schedtail
;
1099 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1102 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1104 void rcu_offline_cpu(int cpu
)
1108 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1110 void __cpuinit
rcu_online_cpu(int cpu
)
1112 unsigned long flags
;
1113 struct rcu_data
*rdp
;
1115 spin_lock_irqsave(&rcu_ctrlblk
.fliplock
, flags
);
1116 cpumask_set_cpu(cpu
, to_cpumask(rcu_cpu_online_map
));
1117 spin_unlock_irqrestore(&rcu_ctrlblk
.fliplock
, flags
);
1120 * The rcu_sched grace-period processing might have bypassed
1121 * this CPU, given that it was not in the rcu_cpu_online_map
1122 * when the grace-period scan started. This means that the
1123 * grace-period task might sleep. So make sure that if this
1124 * should happen, the first callback posted to this CPU will
1125 * wake up the grace-period task if need be.
1128 rdp
= RCU_DATA_CPU(cpu
);
1129 spin_lock_irqsave(&rdp
->lock
, flags
);
1130 rdp
->rcu_sched_sleeping
= 1;
1131 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1134 static void rcu_process_callbacks(struct softirq_action
*unused
)
1136 unsigned long flags
;
1137 struct rcu_head
*next
, *list
;
1138 struct rcu_data
*rdp
;
1140 local_irq_save(flags
);
1141 rdp
= RCU_DATA_ME();
1142 spin_lock(&rdp
->lock
);
1143 list
= rdp
->donelist
;
1145 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1148 rdp
->donelist
= NULL
;
1149 rdp
->donetail
= &rdp
->donelist
;
1150 RCU_TRACE_RDP(rcupreempt_trace_done_remove
, rdp
);
1151 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1156 RCU_TRACE_ME(rcupreempt_trace_invoke
);
1160 void call_rcu(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1162 unsigned long flags
;
1163 struct rcu_data
*rdp
;
1167 local_irq_save(flags
);
1168 rdp
= RCU_DATA_ME();
1169 spin_lock(&rdp
->lock
);
1170 __rcu_advance_callbacks(rdp
);
1171 *rdp
->nexttail
= head
;
1172 rdp
->nexttail
= &head
->next
;
1173 RCU_TRACE_RDP(rcupreempt_trace_next_add
, rdp
);
1174 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1176 EXPORT_SYMBOL_GPL(call_rcu
);
1178 void call_rcu_sched(struct rcu_head
*head
, void (*func
)(struct rcu_head
*rcu
))
1180 unsigned long flags
;
1181 struct rcu_data
*rdp
;
1186 local_irq_save(flags
);
1187 rdp
= RCU_DATA_ME();
1188 spin_lock(&rdp
->lock
);
1189 *rdp
->nextschedtail
= head
;
1190 rdp
->nextschedtail
= &head
->next
;
1191 if (rdp
->rcu_sched_sleeping
) {
1193 /* Grace-period processing might be sleeping... */
1195 rdp
->rcu_sched_sleeping
= 0;
1198 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1201 /* Wake up grace-period processing, unless someone beat us. */
1203 spin_lock_irqsave(&rcu_ctrlblk
.schedlock
, flags
);
1204 if (rcu_ctrlblk
.sched_sleep
!= rcu_sched_sleeping
)
1206 rcu_ctrlblk
.sched_sleep
= rcu_sched_not_sleeping
;
1207 spin_unlock_irqrestore(&rcu_ctrlblk
.schedlock
, flags
);
1209 wake_up_interruptible(&rcu_ctrlblk
.sched_wq
);
1212 EXPORT_SYMBOL_GPL(call_rcu_sched
);
1215 * Wait until all currently running preempt_disable() code segments
1216 * (including hardware-irq-disable segments) complete. Note that
1217 * in -rt this does -not- necessarily result in all currently executing
1218 * interrupt -handlers- having completed.
1220 void __synchronize_sched(void)
1222 struct rcu_synchronize rcu
;
1224 if (num_online_cpus() == 1)
1225 return; /* blocking is gp if only one CPU! */
1227 init_completion(&rcu
.completion
);
1228 /* Will wake me after RCU finished. */
1229 call_rcu_sched(&rcu
.head
, wakeme_after_rcu
);
1231 wait_for_completion(&rcu
.completion
);
1233 EXPORT_SYMBOL_GPL(__synchronize_sched
);
1236 * kthread function that manages call_rcu_sched grace periods.
1238 static int rcu_sched_grace_period(void *arg
)
1240 int couldsleep
; /* might sleep after current pass. */
1241 int couldsleepnext
= 0; /* might sleep after next pass. */
1243 unsigned long flags
;
1244 struct rcu_data
*rdp
;
1248 * Each pass through the following loop handles one
1249 * rcu_sched grace period cycle.
1252 /* Save each CPU's current state. */
1254 for_each_online_cpu(cpu
) {
1255 dyntick_save_progress_counter_sched(cpu
);
1256 save_qsctr_sched(cpu
);
1260 * Sleep for about an RCU grace-period's worth to
1261 * allow better batching and to consume less CPU.
1263 schedule_timeout_interruptible(RCU_SCHED_BATCH_TIME
);
1266 * If there was nothing to do last time, prepare to
1267 * sleep at the end of the current grace period cycle.
1269 couldsleep
= couldsleepnext
;
1272 spin_lock_irqsave(&rcu_ctrlblk
.schedlock
, flags
);
1273 rcu_ctrlblk
.sched_sleep
= rcu_sched_sleep_prep
;
1274 spin_unlock_irqrestore(&rcu_ctrlblk
.schedlock
, flags
);
1278 * Wait on each CPU in turn to have either visited
1279 * a quiescent state or been in dynticks-idle mode.
1281 for_each_online_cpu(cpu
) {
1282 while (rcu_qsctr_inc_needed(cpu
) &&
1283 rcu_qsctr_inc_needed_dyntick(cpu
)) {
1284 /* resched_cpu(cpu); @@@ */
1285 schedule_timeout_interruptible(1);
1289 /* Advance callbacks for each CPU. */
1291 for_each_online_cpu(cpu
) {
1293 rdp
= RCU_DATA_CPU(cpu
);
1294 spin_lock_irqsave(&rdp
->lock
, flags
);
1297 * We are running on this CPU irq-disabled, so no
1298 * CPU can go offline until we re-enable irqs.
1299 * The current CPU might have already gone
1300 * offline (between the for_each_offline_cpu and
1301 * the spin_lock_irqsave), but in that case all its
1302 * callback lists will be empty, so no harm done.
1304 * Advance the callbacks! We share normal RCU's
1305 * donelist, since callbacks are invoked the
1306 * same way in either case.
1308 if (rdp
->waitschedlist
!= NULL
) {
1309 *rdp
->donetail
= rdp
->waitschedlist
;
1310 rdp
->donetail
= rdp
->waitschedtail
;
1313 * Next rcu_check_callbacks() will
1314 * do the required raise_softirq().
1317 if (rdp
->nextschedlist
!= NULL
) {
1318 rdp
->waitschedlist
= rdp
->nextschedlist
;
1319 rdp
->waitschedtail
= rdp
->nextschedtail
;
1323 rdp
->waitschedlist
= NULL
;
1324 rdp
->waitschedtail
= &rdp
->waitschedlist
;
1326 rdp
->nextschedlist
= NULL
;
1327 rdp
->nextschedtail
= &rdp
->nextschedlist
;
1329 /* Mark sleep intention. */
1331 rdp
->rcu_sched_sleeping
= couldsleep
;
1333 spin_unlock_irqrestore(&rdp
->lock
, flags
);
1336 /* If we saw callbacks on the last scan, go deal with them. */
1341 /* Attempt to block... */
1343 spin_lock_irqsave(&rcu_ctrlblk
.schedlock
, flags
);
1344 if (rcu_ctrlblk
.sched_sleep
!= rcu_sched_sleep_prep
) {
1347 * Someone posted a callback after we scanned.
1348 * Go take care of it.
1350 spin_unlock_irqrestore(&rcu_ctrlblk
.schedlock
, flags
);
1355 /* Block until the next person posts a callback. */
1357 rcu_ctrlblk
.sched_sleep
= rcu_sched_sleeping
;
1358 spin_unlock_irqrestore(&rcu_ctrlblk
.schedlock
, flags
);
1360 __wait_event_interruptible(rcu_ctrlblk
.sched_wq
,
1361 rcu_ctrlblk
.sched_sleep
!= rcu_sched_sleeping
,
1365 * Signals would prevent us from sleeping, and we cannot
1366 * do much with them in any case. So flush them.
1369 flush_signals(current
);
1372 } while (!kthread_should_stop());
1378 * Check to see if any future RCU-related work will need to be done
1379 * by the current CPU, even if none need be done immediately, returning
1380 * 1 if so. Assumes that notifiers would take care of handling any
1381 * outstanding requests from the RCU core.
1383 * This function is part of the RCU implementation; it is -not-
1384 * an exported member of the RCU API.
1386 int rcu_needs_cpu(int cpu
)
1388 struct rcu_data
*rdp
= RCU_DATA_CPU(cpu
);
1390 return (rdp
->donelist
!= NULL
||
1391 !!rdp
->waitlistcount
||
1392 rdp
->nextlist
!= NULL
||
1393 rdp
->nextschedlist
!= NULL
||
1394 rdp
->waitschedlist
!= NULL
);
1397 int rcu_pending(int cpu
)
1399 struct rcu_data
*rdp
= RCU_DATA_CPU(cpu
);
1401 /* The CPU has at least one callback queued somewhere. */
1403 if (rdp
->donelist
!= NULL
||
1404 !!rdp
->waitlistcount
||
1405 rdp
->nextlist
!= NULL
||
1406 rdp
->nextschedlist
!= NULL
||
1407 rdp
->waitschedlist
!= NULL
)
1410 /* The RCU core needs an acknowledgement from this CPU. */
1412 if ((per_cpu(rcu_flip_flag
, cpu
) == rcu_flipped
) ||
1413 (per_cpu(rcu_mb_flag
, cpu
) == rcu_mb_needed
))
1416 /* This CPU has fallen behind the global grace-period number. */
1418 if (rdp
->completed
!= rcu_ctrlblk
.completed
)
1421 /* Nothing needed from this CPU. */
1426 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
1427 unsigned long action
, void *hcpu
)
1429 long cpu
= (long)hcpu
;
1432 case CPU_UP_PREPARE
:
1433 case CPU_UP_PREPARE_FROZEN
:
1434 rcu_online_cpu(cpu
);
1436 case CPU_UP_CANCELED
:
1437 case CPU_UP_CANCELED_FROZEN
:
1439 case CPU_DEAD_FROZEN
:
1440 rcu_offline_cpu(cpu
);
1448 static struct notifier_block __cpuinitdata rcu_nb
= {
1449 .notifier_call
= rcu_cpu_notify
,
1452 void __init
__rcu_init(void)
1456 struct rcu_data
*rdp
;
1458 printk(KERN_NOTICE
"Preemptible RCU implementation.\n");
1459 for_each_possible_cpu(cpu
) {
1460 rdp
= RCU_DATA_CPU(cpu
);
1461 spin_lock_init(&rdp
->lock
);
1463 rdp
->waitlistcount
= 0;
1464 rdp
->nextlist
= NULL
;
1465 rdp
->nexttail
= &rdp
->nextlist
;
1466 for (i
= 0; i
< GP_STAGES
; i
++) {
1467 rdp
->waitlist
[i
] = NULL
;
1468 rdp
->waittail
[i
] = &rdp
->waitlist
[i
];
1470 rdp
->donelist
= NULL
;
1471 rdp
->donetail
= &rdp
->donelist
;
1472 rdp
->rcu_flipctr
[0] = 0;
1473 rdp
->rcu_flipctr
[1] = 0;
1474 rdp
->nextschedlist
= NULL
;
1475 rdp
->nextschedtail
= &rdp
->nextschedlist
;
1476 rdp
->waitschedlist
= NULL
;
1477 rdp
->waitschedtail
= &rdp
->waitschedlist
;
1478 rdp
->rcu_sched_sleeping
= 0;
1480 register_cpu_notifier(&rcu_nb
);
1483 * We don't need protection against CPU-Hotplug here
1485 * a) If a CPU comes online while we are iterating over the
1486 * cpu_online_mask below, we would only end up making a
1487 * duplicate call to rcu_online_cpu() which sets the corresponding
1488 * CPU's mask in the rcu_cpu_online_map.
1490 * b) A CPU cannot go offline at this point in time since the user
1491 * does not have access to the sysfs interface, nor do we
1492 * suspend the system.
1494 for_each_online_cpu(cpu
)
1495 rcu_cpu_notify(&rcu_nb
, CPU_UP_PREPARE
, (void *)(long) cpu
);
1497 open_softirq(RCU_SOFTIRQ
, rcu_process_callbacks
);
1501 * Late-boot-time RCU initialization that must wait until after scheduler
1502 * has been initialized.
1504 void __init
rcu_init_sched(void)
1506 rcu_sched_grace_period_task
= kthread_run(rcu_sched_grace_period
,
1508 "rcu_sched_grace_period");
1509 WARN_ON(IS_ERR(rcu_sched_grace_period_task
));
1512 #ifdef CONFIG_RCU_TRACE
1513 long *rcupreempt_flipctr(int cpu
)
1515 return &RCU_DATA_CPU(cpu
)->rcu_flipctr
[0];
1517 EXPORT_SYMBOL_GPL(rcupreempt_flipctr
);
1519 int rcupreempt_flip_flag(int cpu
)
1521 return per_cpu(rcu_flip_flag
, cpu
);
1523 EXPORT_SYMBOL_GPL(rcupreempt_flip_flag
);
1525 int rcupreempt_mb_flag(int cpu
)
1527 return per_cpu(rcu_mb_flag
, cpu
);
1529 EXPORT_SYMBOL_GPL(rcupreempt_mb_flag
);
1531 char *rcupreempt_try_flip_state_name(void)
1533 return rcu_try_flip_state_names
[rcu_ctrlblk
.rcu_try_flip_state
];
1535 EXPORT_SYMBOL_GPL(rcupreempt_try_flip_state_name
);
1537 struct rcupreempt_trace
*rcupreempt_trace_cpu(int cpu
)
1539 struct rcu_data
*rdp
= RCU_DATA_CPU(cpu
);
1543 EXPORT_SYMBOL_GPL(rcupreempt_trace_cpu
);
1545 #endif /* #ifdef RCU_TRACE */