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 (C) IBM Corporation, 2001
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
23 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
24 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
27 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * http://lse.sourceforge.net/locking/rcupdate.html
33 #include <linux/types.h>
34 #include <linux/kernel.h>
35 #include <linux/init.h>
36 #include <linux/spinlock.h>
37 #include <linux/smp.h>
38 #include <linux/rcupdate.h>
39 #include <linux/interrupt.h>
40 #include <linux/sched.h>
41 #include <asm/atomic.h>
42 #include <linux/bitops.h>
43 #include <linux/module.h>
44 #include <linux/completion.h>
45 #include <linux/moduleparam.h>
46 #include <linux/percpu.h>
47 #include <linux/notifier.h>
48 #include <linux/rcupdate.h>
49 #include <linux/cpu.h>
50 #include <linux/mutex.h>
52 #ifdef CONFIG_DEBUG_LOCK_ALLOC
53 static struct lock_class_key rcu_lock_key
;
54 struct lockdep_map rcu_lock_map
=
55 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key
);
57 EXPORT_SYMBOL_GPL(rcu_lock_map
);
60 /* Definition for rcupdate control block. */
61 static struct rcu_ctrlblk rcu_ctrlblk
= {
64 .lock
= __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk
.lock
),
65 .cpumask
= CPU_MASK_NONE
,
67 static struct rcu_ctrlblk rcu_bh_ctrlblk
= {
70 .lock
= __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk
.lock
),
71 .cpumask
= CPU_MASK_NONE
,
74 DEFINE_PER_CPU(struct rcu_data
, rcu_data
) = { 0L };
75 DEFINE_PER_CPU(struct rcu_data
, rcu_bh_data
) = { 0L };
77 /* Fake initialization required by compiler */
78 static DEFINE_PER_CPU(struct tasklet_struct
, rcu_tasklet
) = {NULL
};
79 static int blimit
= 10;
80 static int qhimark
= 10000;
81 static int qlowmark
= 100;
83 static atomic_t rcu_barrier_cpu_count
;
84 static DEFINE_MUTEX(rcu_barrier_mutex
);
85 static struct completion rcu_barrier_completion
;
88 static void force_quiescent_state(struct rcu_data
*rdp
,
89 struct rcu_ctrlblk
*rcp
)
94 if (unlikely(!rcp
->signaled
)) {
97 * Don't send IPI to itself. With irqs disabled,
98 * rdp->cpu is the current cpu.
100 cpumask
= rcp
->cpumask
;
101 cpu_clear(rdp
->cpu
, cpumask
);
102 for_each_cpu_mask(cpu
, cpumask
)
103 smp_send_reschedule(cpu
);
107 static inline void force_quiescent_state(struct rcu_data
*rdp
,
108 struct rcu_ctrlblk
*rcp
)
115 * call_rcu - Queue an RCU callback for invocation after a grace period.
116 * @head: structure to be used for queueing the RCU updates.
117 * @func: actual update function to be invoked after the grace period
119 * The update function will be invoked some time after a full grace
120 * period elapses, in other words after all currently executing RCU
121 * read-side critical sections have completed. RCU read-side critical
122 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
125 void fastcall
call_rcu(struct rcu_head
*head
,
126 void (*func
)(struct rcu_head
*rcu
))
129 struct rcu_data
*rdp
;
133 local_irq_save(flags
);
134 rdp
= &__get_cpu_var(rcu_data
);
135 *rdp
->nxttail
= head
;
136 rdp
->nxttail
= &head
->next
;
137 if (unlikely(++rdp
->qlen
> qhimark
)) {
138 rdp
->blimit
= INT_MAX
;
139 force_quiescent_state(rdp
, &rcu_ctrlblk
);
141 local_irq_restore(flags
);
145 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
146 * @head: structure to be used for queueing the RCU updates.
147 * @func: actual update function to be invoked after the grace period
149 * The update function will be invoked some time after a full grace
150 * period elapses, in other words after all currently executing RCU
151 * read-side critical sections have completed. call_rcu_bh() assumes
152 * that the read-side critical sections end on completion of a softirq
153 * handler. This means that read-side critical sections in process
154 * context must not be interrupted by softirqs. This interface is to be
155 * used when most of the read-side critical sections are in softirq context.
156 * RCU read-side critical sections are delimited by rcu_read_lock() and
157 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
158 * and rcu_read_unlock_bh(), if in process context. These may be nested.
160 void fastcall
call_rcu_bh(struct rcu_head
*head
,
161 void (*func
)(struct rcu_head
*rcu
))
164 struct rcu_data
*rdp
;
168 local_irq_save(flags
);
169 rdp
= &__get_cpu_var(rcu_bh_data
);
170 *rdp
->nxttail
= head
;
171 rdp
->nxttail
= &head
->next
;
173 if (unlikely(++rdp
->qlen
> qhimark
)) {
174 rdp
->blimit
= INT_MAX
;
175 force_quiescent_state(rdp
, &rcu_bh_ctrlblk
);
178 local_irq_restore(flags
);
182 * Return the number of RCU batches processed thus far. Useful
183 * for debug and statistics.
185 long rcu_batches_completed(void)
187 return rcu_ctrlblk
.completed
;
191 * Return the number of RCU batches processed thus far. Useful
192 * for debug and statistics.
194 long rcu_batches_completed_bh(void)
196 return rcu_bh_ctrlblk
.completed
;
199 static void rcu_barrier_callback(struct rcu_head
*notused
)
201 if (atomic_dec_and_test(&rcu_barrier_cpu_count
))
202 complete(&rcu_barrier_completion
);
206 * Called with preemption disabled, and from cross-cpu IRQ context.
208 static void rcu_barrier_func(void *notused
)
210 int cpu
= smp_processor_id();
211 struct rcu_data
*rdp
= &per_cpu(rcu_data
, cpu
);
212 struct rcu_head
*head
;
214 head
= &rdp
->barrier
;
215 atomic_inc(&rcu_barrier_cpu_count
);
216 call_rcu(head
, rcu_barrier_callback
);
220 * rcu_barrier - Wait until all the in-flight RCUs are complete.
222 void rcu_barrier(void)
224 BUG_ON(in_interrupt());
225 /* Take cpucontrol mutex to protect against CPU hotplug */
226 mutex_lock(&rcu_barrier_mutex
);
227 init_completion(&rcu_barrier_completion
);
228 atomic_set(&rcu_barrier_cpu_count
, 0);
229 on_each_cpu(rcu_barrier_func
, NULL
, 0, 1);
230 wait_for_completion(&rcu_barrier_completion
);
231 mutex_unlock(&rcu_barrier_mutex
);
233 EXPORT_SYMBOL_GPL(rcu_barrier
);
236 * Invoke the completed RCU callbacks. They are expected to be in
239 static void rcu_do_batch(struct rcu_data
*rdp
)
241 struct rcu_head
*next
, *list
;
244 list
= rdp
->donelist
;
250 if (++count
>= rdp
->blimit
)
253 rdp
->donelist
= list
;
258 if (rdp
->blimit
== INT_MAX
&& rdp
->qlen
<= qlowmark
)
259 rdp
->blimit
= blimit
;
262 rdp
->donetail
= &rdp
->donelist
;
264 tasklet_schedule(&per_cpu(rcu_tasklet
, rdp
->cpu
));
268 * Grace period handling:
269 * The grace period handling consists out of two steps:
270 * - A new grace period is started.
271 * This is done by rcu_start_batch. The start is not broadcasted to
272 * all cpus, they must pick this up by comparing rcp->cur with
273 * rdp->quiescbatch. All cpus are recorded in the
274 * rcu_ctrlblk.cpumask bitmap.
275 * - All cpus must go through a quiescent state.
276 * Since the start of the grace period is not broadcasted, at least two
277 * calls to rcu_check_quiescent_state are required:
278 * The first call just notices that a new grace period is running. The
279 * following calls check if there was a quiescent state since the beginning
280 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
281 * the bitmap is empty, then the grace period is completed.
282 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
283 * period (if necessary).
286 * Register a new batch of callbacks, and start it up if there is currently no
287 * active batch and the batch to be registered has not already occurred.
288 * Caller must hold rcu_ctrlblk.lock.
290 static void rcu_start_batch(struct rcu_ctrlblk
*rcp
)
292 if (rcp
->next_pending
&&
293 rcp
->completed
== rcp
->cur
) {
294 rcp
->next_pending
= 0;
296 * next_pending == 0 must be visible in
297 * __rcu_process_callbacks() before it can see new value of cur.
303 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
304 * Barrier Otherwise it can cause tickless idle CPUs to be
305 * included in rcp->cpumask, which will extend graceperiods
309 cpus_andnot(rcp
->cpumask
, cpu_online_map
, nohz_cpu_mask
);
316 * cpu went through a quiescent state since the beginning of the grace period.
317 * Clear it from the cpu mask and complete the grace period if it was the last
318 * cpu. Start another grace period if someone has further entries pending
320 static void cpu_quiet(int cpu
, struct rcu_ctrlblk
*rcp
)
322 cpu_clear(cpu
, rcp
->cpumask
);
323 if (cpus_empty(rcp
->cpumask
)) {
324 /* batch completed ! */
325 rcp
->completed
= rcp
->cur
;
326 rcu_start_batch(rcp
);
331 * Check if the cpu has gone through a quiescent state (say context
332 * switch). If so and if it already hasn't done so in this RCU
333 * quiescent cycle, then indicate that it has done so.
335 static void rcu_check_quiescent_state(struct rcu_ctrlblk
*rcp
,
336 struct rcu_data
*rdp
)
338 if (rdp
->quiescbatch
!= rcp
->cur
) {
339 /* start new grace period: */
341 rdp
->passed_quiesc
= 0;
342 rdp
->quiescbatch
= rcp
->cur
;
346 /* Grace period already completed for this cpu?
347 * qs_pending is checked instead of the actual bitmap to avoid
348 * cacheline trashing.
350 if (!rdp
->qs_pending
)
354 * Was there a quiescent state since the beginning of the grace
355 * period? If no, then exit and wait for the next call.
357 if (!rdp
->passed_quiesc
)
361 spin_lock(&rcp
->lock
);
363 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
364 * during cpu startup. Ignore the quiescent state.
366 if (likely(rdp
->quiescbatch
== rcp
->cur
))
367 cpu_quiet(rdp
->cpu
, rcp
);
369 spin_unlock(&rcp
->lock
);
373 #ifdef CONFIG_HOTPLUG_CPU
375 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
376 * locking requirements, the list it's pulling from has to belong to a cpu
377 * which is dead and hence not processing interrupts.
379 static void rcu_move_batch(struct rcu_data
*this_rdp
, struct rcu_head
*list
,
380 struct rcu_head
**tail
)
383 *this_rdp
->nxttail
= list
;
385 this_rdp
->nxttail
= tail
;
389 static void __rcu_offline_cpu(struct rcu_data
*this_rdp
,
390 struct rcu_ctrlblk
*rcp
, struct rcu_data
*rdp
)
392 /* if the cpu going offline owns the grace period
393 * we can block indefinitely waiting for it, so flush
396 spin_lock_bh(&rcp
->lock
);
397 if (rcp
->cur
!= rcp
->completed
)
398 cpu_quiet(rdp
->cpu
, rcp
);
399 spin_unlock_bh(&rcp
->lock
);
400 rcu_move_batch(this_rdp
, rdp
->curlist
, rdp
->curtail
);
401 rcu_move_batch(this_rdp
, rdp
->nxtlist
, rdp
->nxttail
);
402 rcu_move_batch(this_rdp
, rdp
->donelist
, rdp
->donetail
);
405 static void rcu_offline_cpu(int cpu
)
407 struct rcu_data
*this_rdp
= &get_cpu_var(rcu_data
);
408 struct rcu_data
*this_bh_rdp
= &get_cpu_var(rcu_bh_data
);
410 __rcu_offline_cpu(this_rdp
, &rcu_ctrlblk
,
411 &per_cpu(rcu_data
, cpu
));
412 __rcu_offline_cpu(this_bh_rdp
, &rcu_bh_ctrlblk
,
413 &per_cpu(rcu_bh_data
, cpu
));
414 put_cpu_var(rcu_data
);
415 put_cpu_var(rcu_bh_data
);
416 tasklet_kill_immediate(&per_cpu(rcu_tasklet
, cpu
), cpu
);
421 static void rcu_offline_cpu(int cpu
)
428 * This does the RCU processing work from tasklet context.
430 static void __rcu_process_callbacks(struct rcu_ctrlblk
*rcp
,
431 struct rcu_data
*rdp
)
433 if (rdp
->curlist
&& !rcu_batch_before(rcp
->completed
, rdp
->batch
)) {
434 *rdp
->donetail
= rdp
->curlist
;
435 rdp
->donetail
= rdp
->curtail
;
437 rdp
->curtail
= &rdp
->curlist
;
440 if (rdp
->nxtlist
&& !rdp
->curlist
) {
442 rdp
->curlist
= rdp
->nxtlist
;
443 rdp
->curtail
= rdp
->nxttail
;
445 rdp
->nxttail
= &rdp
->nxtlist
;
449 * start the next batch of callbacks
452 /* determine batch number */
453 rdp
->batch
= rcp
->cur
+ 1;
454 /* see the comment and corresponding wmb() in
455 * the rcu_start_batch()
459 if (!rcp
->next_pending
) {
460 /* and start it/schedule start if it's a new batch */
461 spin_lock(&rcp
->lock
);
462 rcp
->next_pending
= 1;
463 rcu_start_batch(rcp
);
464 spin_unlock(&rcp
->lock
);
468 rcu_check_quiescent_state(rcp
, rdp
);
473 static void rcu_process_callbacks(unsigned long unused
)
475 __rcu_process_callbacks(&rcu_ctrlblk
, &__get_cpu_var(rcu_data
));
476 __rcu_process_callbacks(&rcu_bh_ctrlblk
, &__get_cpu_var(rcu_bh_data
));
479 static int __rcu_pending(struct rcu_ctrlblk
*rcp
, struct rcu_data
*rdp
)
481 /* This cpu has pending rcu entries and the grace period
482 * for them has completed.
484 if (rdp
->curlist
&& !rcu_batch_before(rcp
->completed
, rdp
->batch
))
487 /* This cpu has no pending entries, but there are new entries */
488 if (!rdp
->curlist
&& rdp
->nxtlist
)
491 /* This cpu has finished callbacks to invoke */
495 /* The rcu core waits for a quiescent state from the cpu */
496 if (rdp
->quiescbatch
!= rcp
->cur
|| rdp
->qs_pending
)
504 * Check to see if there is any immediate RCU-related work to be done
505 * by the current CPU, returning 1 if so. This function is part of the
506 * RCU implementation; it is -not- an exported member of the RCU API.
508 int rcu_pending(int cpu
)
510 return __rcu_pending(&rcu_ctrlblk
, &per_cpu(rcu_data
, cpu
)) ||
511 __rcu_pending(&rcu_bh_ctrlblk
, &per_cpu(rcu_bh_data
, cpu
));
515 * Check to see if any future RCU-related work will need to be done
516 * by the current CPU, even if none need be done immediately, returning
517 * 1 if so. This function is part of the RCU implementation; it is -not-
518 * an exported member of the RCU API.
520 int rcu_needs_cpu(int cpu
)
522 struct rcu_data
*rdp
= &per_cpu(rcu_data
, cpu
);
523 struct rcu_data
*rdp_bh
= &per_cpu(rcu_bh_data
, cpu
);
525 return (!!rdp
->curlist
|| !!rdp_bh
->curlist
|| rcu_pending(cpu
));
528 void rcu_check_callbacks(int cpu
, int user
)
531 (idle_cpu(cpu
) && !in_softirq() &&
532 hardirq_count() <= (1 << HARDIRQ_SHIFT
))) {
534 rcu_bh_qsctr_inc(cpu
);
535 } else if (!in_softirq())
536 rcu_bh_qsctr_inc(cpu
);
537 tasklet_schedule(&per_cpu(rcu_tasklet
, cpu
));
540 static void rcu_init_percpu_data(int cpu
, struct rcu_ctrlblk
*rcp
,
541 struct rcu_data
*rdp
)
543 memset(rdp
, 0, sizeof(*rdp
));
544 rdp
->curtail
= &rdp
->curlist
;
545 rdp
->nxttail
= &rdp
->nxtlist
;
546 rdp
->donetail
= &rdp
->donelist
;
547 rdp
->quiescbatch
= rcp
->completed
;
550 rdp
->blimit
= blimit
;
553 static void __devinit
rcu_online_cpu(int cpu
)
555 struct rcu_data
*rdp
= &per_cpu(rcu_data
, cpu
);
556 struct rcu_data
*bh_rdp
= &per_cpu(rcu_bh_data
, cpu
);
558 rcu_init_percpu_data(cpu
, &rcu_ctrlblk
, rdp
);
559 rcu_init_percpu_data(cpu
, &rcu_bh_ctrlblk
, bh_rdp
);
560 tasklet_init(&per_cpu(rcu_tasklet
, cpu
), rcu_process_callbacks
, 0UL);
563 static int __cpuinit
rcu_cpu_notify(struct notifier_block
*self
,
564 unsigned long action
, void *hcpu
)
566 long cpu
= (long)hcpu
;
569 case CPU_UP_PREPARE_FROZEN
:
573 case CPU_DEAD_FROZEN
:
574 rcu_offline_cpu(cpu
);
582 static struct notifier_block __cpuinitdata rcu_nb
= {
583 .notifier_call
= rcu_cpu_notify
,
587 * Initializes rcu mechanism. Assumed to be called early.
588 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
589 * Note that rcu_qsctr and friends are implicitly
590 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
592 void __init
rcu_init(void)
594 rcu_cpu_notify(&rcu_nb
, CPU_UP_PREPARE
,
595 (void *)(long)smp_processor_id());
596 /* Register notifier for non-boot CPUs */
597 register_cpu_notifier(&rcu_nb
);
600 struct rcu_synchronize
{
601 struct rcu_head head
;
602 struct completion completion
;
605 /* Because of FASTCALL declaration of complete, we use this wrapper */
606 static void wakeme_after_rcu(struct rcu_head
*head
)
608 struct rcu_synchronize
*rcu
;
610 rcu
= container_of(head
, struct rcu_synchronize
, head
);
611 complete(&rcu
->completion
);
615 * synchronize_rcu - wait until a grace period has elapsed.
617 * Control will return to the caller some time after a full grace
618 * period has elapsed, in other words after all currently executing RCU
619 * read-side critical sections have completed. RCU read-side critical
620 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
623 * If your read-side code is not protected by rcu_read_lock(), do -not-
624 * use synchronize_rcu().
626 void synchronize_rcu(void)
628 struct rcu_synchronize rcu
;
630 init_completion(&rcu
.completion
);
631 /* Will wake me after RCU finished */
632 call_rcu(&rcu
.head
, wakeme_after_rcu
);
635 wait_for_completion(&rcu
.completion
);
638 module_param(blimit
, int, 0);
639 module_param(qhimark
, int, 0);
640 module_param(qlowmark
, int, 0);
641 EXPORT_SYMBOL_GPL(rcu_batches_completed
);
642 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh
);
643 EXPORT_SYMBOL_GPL(call_rcu
);
644 EXPORT_SYMBOL_GPL(call_rcu_bh
);
645 EXPORT_SYMBOL_GPL(synchronize_rcu
);