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/cpu.h>
49 #include <linux/mutex.h>
51 #ifdef CONFIG_DEBUG_LOCK_ALLOC
52 static struct lock_class_key rcu_lock_key;
53 struct lockdep_map rcu_lock_map =
54 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
56 EXPORT_SYMBOL_GPL(rcu_lock_map);
59 /* Definition for rcupdate control block. */
60 static struct rcu_ctrlblk rcu_ctrlblk = {
63 .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock),
64 .cpumask = CPU_MASK_NONE,
66 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
69 .lock = __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk.lock),
70 .cpumask = CPU_MASK_NONE,
73 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
74 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
76 static int blimit = 10;
77 static int qhimark = 10000;
78 static int qlowmark = 100;
80 static atomic_t rcu_barrier_cpu_count;
81 static DEFINE_MUTEX(rcu_barrier_mutex);
82 static struct completion rcu_barrier_completion;
85 static void force_quiescent_state(struct rcu_data *rdp,
86 struct rcu_ctrlblk *rcp)
91 if (unlikely(!rcp->signaled)) {
94 * Don't send IPI to itself. With irqs disabled,
95 * rdp->cpu is the current cpu.
97 cpumask = rcp->cpumask;
98 cpu_clear(rdp->cpu, cpumask);
99 for_each_cpu_mask(cpu, cpumask)
100 smp_send_reschedule(cpu);
104 static inline void force_quiescent_state(struct rcu_data *rdp,
105 struct rcu_ctrlblk *rcp)
112 * call_rcu - Queue an RCU callback for invocation after a grace period.
113 * @head: structure to be used for queueing the RCU updates.
114 * @func: actual update function to be invoked after the grace period
116 * The update function will be invoked some time after a full grace
117 * period elapses, in other words after all currently executing RCU
118 * read-side critical sections have completed. RCU read-side critical
119 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
122 void fastcall call_rcu(struct rcu_head *head,
123 void (*func)(struct rcu_head *rcu))
126 struct rcu_data *rdp;
130 local_irq_save(flags);
131 rdp = &__get_cpu_var(rcu_data);
132 *rdp->nxttail = head;
133 rdp->nxttail = &head->next;
134 if (unlikely(++rdp->qlen > qhimark)) {
135 rdp->blimit = INT_MAX;
136 force_quiescent_state(rdp, &rcu_ctrlblk);
138 local_irq_restore(flags);
142 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
143 * @head: structure to be used for queueing the RCU updates.
144 * @func: actual update function to be invoked after the grace period
146 * The update function will be invoked some time after a full grace
147 * period elapses, in other words after all currently executing RCU
148 * read-side critical sections have completed. call_rcu_bh() assumes
149 * that the read-side critical sections end on completion of a softirq
150 * handler. This means that read-side critical sections in process
151 * context must not be interrupted by softirqs. This interface is to be
152 * used when most of the read-side critical sections are in softirq context.
153 * RCU read-side critical sections are delimited by rcu_read_lock() and
154 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
155 * and rcu_read_unlock_bh(), if in process context. These may be nested.
157 void fastcall call_rcu_bh(struct rcu_head *head,
158 void (*func)(struct rcu_head *rcu))
161 struct rcu_data *rdp;
165 local_irq_save(flags);
166 rdp = &__get_cpu_var(rcu_bh_data);
167 *rdp->nxttail = head;
168 rdp->nxttail = &head->next;
170 if (unlikely(++rdp->qlen > qhimark)) {
171 rdp->blimit = INT_MAX;
172 force_quiescent_state(rdp, &rcu_bh_ctrlblk);
175 local_irq_restore(flags);
179 * Return the number of RCU batches processed thus far. Useful
180 * for debug and statistics.
182 long rcu_batches_completed(void)
184 return rcu_ctrlblk.completed;
188 * Return the number of RCU batches processed thus far. Useful
189 * for debug and statistics.
191 long rcu_batches_completed_bh(void)
193 return rcu_bh_ctrlblk.completed;
196 static void rcu_barrier_callback(struct rcu_head *notused)
198 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
199 complete(&rcu_barrier_completion);
203 * Called with preemption disabled, and from cross-cpu IRQ context.
205 static void rcu_barrier_func(void *notused)
207 int cpu = smp_processor_id();
208 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
209 struct rcu_head *head;
211 head = &rdp->barrier;
212 atomic_inc(&rcu_barrier_cpu_count);
213 call_rcu(head, rcu_barrier_callback);
217 * rcu_barrier - Wait until all the in-flight RCUs are complete.
219 void rcu_barrier(void)
221 BUG_ON(in_interrupt());
222 /* Take cpucontrol mutex to protect against CPU hotplug */
223 mutex_lock(&rcu_barrier_mutex);
224 init_completion(&rcu_barrier_completion);
225 atomic_set(&rcu_barrier_cpu_count, 0);
226 on_each_cpu(rcu_barrier_func, NULL, 0, 1);
227 wait_for_completion(&rcu_barrier_completion);
228 mutex_unlock(&rcu_barrier_mutex);
230 EXPORT_SYMBOL_GPL(rcu_barrier);
232 /* Raises the softirq for processing rcu_callbacks. */
233 static inline void raise_rcu_softirq(void)
235 raise_softirq(RCU_SOFTIRQ);
237 * The smp_mb() here is required to ensure that this cpu's
238 * __rcu_process_callbacks() reads the most recently updated
245 * Invoke the completed RCU callbacks. They are expected to be in
248 static void rcu_do_batch(struct rcu_data *rdp)
250 struct rcu_head *next, *list;
253 list = rdp->donelist;
259 if (++count >= rdp->blimit)
262 rdp->donelist = list;
267 if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
268 rdp->blimit = blimit;
271 rdp->donetail = &rdp->donelist;
277 * Grace period handling:
278 * The grace period handling consists out of two steps:
279 * - A new grace period is started.
280 * This is done by rcu_start_batch. The start is not broadcasted to
281 * all cpus, they must pick this up by comparing rcp->cur with
282 * rdp->quiescbatch. All cpus are recorded in the
283 * rcu_ctrlblk.cpumask bitmap.
284 * - All cpus must go through a quiescent state.
285 * Since the start of the grace period is not broadcasted, at least two
286 * calls to rcu_check_quiescent_state are required:
287 * The first call just notices that a new grace period is running. The
288 * following calls check if there was a quiescent state since the beginning
289 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
290 * the bitmap is empty, then the grace period is completed.
291 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
292 * period (if necessary).
295 * Register a new batch of callbacks, and start it up if there is currently no
296 * active batch and the batch to be registered has not already occurred.
297 * Caller must hold rcu_ctrlblk.lock.
299 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
301 if (rcp->next_pending &&
302 rcp->completed == rcp->cur) {
303 rcp->next_pending = 0;
305 * next_pending == 0 must be visible in
306 * __rcu_process_callbacks() before it can see new value of cur.
312 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
313 * Barrier Otherwise it can cause tickless idle CPUs to be
314 * included in rcp->cpumask, which will extend graceperiods
318 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
325 * cpu went through a quiescent state since the beginning of the grace period.
326 * Clear it from the cpu mask and complete the grace period if it was the last
327 * cpu. Start another grace period if someone has further entries pending
329 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
331 cpu_clear(cpu, rcp->cpumask);
332 if (cpus_empty(rcp->cpumask)) {
333 /* batch completed ! */
334 rcp->completed = rcp->cur;
335 rcu_start_batch(rcp);
340 * Check if the cpu has gone through a quiescent state (say context
341 * switch). If so and if it already hasn't done so in this RCU
342 * quiescent cycle, then indicate that it has done so.
344 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
345 struct rcu_data *rdp)
347 if (rdp->quiescbatch != rcp->cur) {
348 /* start new grace period: */
350 rdp->passed_quiesc = 0;
351 rdp->quiescbatch = rcp->cur;
355 /* Grace period already completed for this cpu?
356 * qs_pending is checked instead of the actual bitmap to avoid
357 * cacheline trashing.
359 if (!rdp->qs_pending)
363 * Was there a quiescent state since the beginning of the grace
364 * period? If no, then exit and wait for the next call.
366 if (!rdp->passed_quiesc)
370 spin_lock(&rcp->lock);
372 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
373 * during cpu startup. Ignore the quiescent state.
375 if (likely(rdp->quiescbatch == rcp->cur))
376 cpu_quiet(rdp->cpu, rcp);
378 spin_unlock(&rcp->lock);
382 #ifdef CONFIG_HOTPLUG_CPU
384 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
385 * locking requirements, the list it's pulling from has to belong to a cpu
386 * which is dead and hence not processing interrupts.
388 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
389 struct rcu_head **tail)
392 *this_rdp->nxttail = list;
394 this_rdp->nxttail = tail;
398 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
399 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
401 /* if the cpu going offline owns the grace period
402 * we can block indefinitely waiting for it, so flush
405 spin_lock_bh(&rcp->lock);
406 if (rcp->cur != rcp->completed)
407 cpu_quiet(rdp->cpu, rcp);
408 spin_unlock_bh(&rcp->lock);
409 rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
410 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
411 rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail);
414 static void rcu_offline_cpu(int cpu)
416 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
417 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
419 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
420 &per_cpu(rcu_data, cpu));
421 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
422 &per_cpu(rcu_bh_data, cpu));
423 put_cpu_var(rcu_data);
424 put_cpu_var(rcu_bh_data);
429 static void rcu_offline_cpu(int cpu)
436 * This does the RCU processing work from softirq context.
438 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
439 struct rcu_data *rdp)
441 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
442 *rdp->donetail = rdp->curlist;
443 rdp->donetail = rdp->curtail;
445 rdp->curtail = &rdp->curlist;
448 if (rdp->nxtlist && !rdp->curlist) {
450 rdp->curlist = rdp->nxtlist;
451 rdp->curtail = rdp->nxttail;
453 rdp->nxttail = &rdp->nxtlist;
457 * start the next batch of callbacks
460 /* determine batch number */
461 rdp->batch = rcp->cur + 1;
462 /* see the comment and corresponding wmb() in
463 * the rcu_start_batch()
467 if (!rcp->next_pending) {
468 /* and start it/schedule start if it's a new batch */
469 spin_lock(&rcp->lock);
470 rcp->next_pending = 1;
471 rcu_start_batch(rcp);
472 spin_unlock(&rcp->lock);
476 rcu_check_quiescent_state(rcp, rdp);
481 static void rcu_process_callbacks(struct softirq_action *unused)
483 __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
484 __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
487 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
489 /* This cpu has pending rcu entries and the grace period
490 * for them has completed.
492 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch))
495 /* This cpu has no pending entries, but there are new entries */
496 if (!rdp->curlist && rdp->nxtlist)
499 /* This cpu has finished callbacks to invoke */
503 /* The rcu core waits for a quiescent state from the cpu */
504 if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
512 * Check to see if there is any immediate RCU-related work to be done
513 * by the current CPU, returning 1 if so. This function is part of the
514 * RCU implementation; it is -not- an exported member of the RCU API.
516 int rcu_pending(int cpu)
518 return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
519 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
523 * Check to see if any future RCU-related work will need to be done
524 * by the current CPU, even if none need be done immediately, returning
525 * 1 if so. This function is part of the RCU implementation; it is -not-
526 * an exported member of the RCU API.
528 int rcu_needs_cpu(int cpu)
530 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
531 struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
533 return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
536 void rcu_check_callbacks(int cpu, int user)
539 (idle_cpu(cpu) && !in_softirq() &&
540 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
542 rcu_bh_qsctr_inc(cpu);
543 } else if (!in_softirq())
544 rcu_bh_qsctr_inc(cpu);
548 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
549 struct rcu_data *rdp)
551 memset(rdp, 0, sizeof(*rdp));
552 rdp->curtail = &rdp->curlist;
553 rdp->nxttail = &rdp->nxtlist;
554 rdp->donetail = &rdp->donelist;
555 rdp->quiescbatch = rcp->completed;
558 rdp->blimit = blimit;
561 static void __cpuinit rcu_online_cpu(int cpu)
563 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
564 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
566 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
567 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
568 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks, NULL);
571 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
572 unsigned long action, void *hcpu)
574 long cpu = (long)hcpu;
577 case CPU_UP_PREPARE_FROZEN:
581 case CPU_DEAD_FROZEN:
582 rcu_offline_cpu(cpu);
590 static struct notifier_block __cpuinitdata rcu_nb = {
591 .notifier_call = rcu_cpu_notify,
595 * Initializes rcu mechanism. Assumed to be called early.
596 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
597 * Note that rcu_qsctr and friends are implicitly
598 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
600 void __init rcu_init(void)
602 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
603 (void *)(long)smp_processor_id());
604 /* Register notifier for non-boot CPUs */
605 register_cpu_notifier(&rcu_nb);
608 struct rcu_synchronize {
609 struct rcu_head head;
610 struct completion completion;
613 /* Because of FASTCALL declaration of complete, we use this wrapper */
614 static void wakeme_after_rcu(struct rcu_head *head)
616 struct rcu_synchronize *rcu;
618 rcu = container_of(head, struct rcu_synchronize, head);
619 complete(&rcu->completion);
623 * synchronize_rcu - wait until a grace period has elapsed.
625 * Control will return to the caller some time after a full grace
626 * period has elapsed, in other words after all currently executing RCU
627 * read-side critical sections have completed. RCU read-side critical
628 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
631 * If your read-side code is not protected by rcu_read_lock(), do -not-
632 * use synchronize_rcu().
634 void synchronize_rcu(void)
636 struct rcu_synchronize rcu;
638 init_completion(&rcu.completion);
639 /* Will wake me after RCU finished */
640 call_rcu(&rcu.head, wakeme_after_rcu);
643 wait_for_completion(&rcu.completion);
646 module_param(blimit, int, 0);
647 module_param(qhimark, int, 0);
648 module_param(qlowmark, int, 0);
649 EXPORT_SYMBOL_GPL(rcu_batches_completed);
650 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
651 EXPORT_SYMBOL_GPL(call_rcu);
652 EXPORT_SYMBOL_GPL(call_rcu_bh);
653 EXPORT_SYMBOL_GPL(synchronize_rcu);