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 /* Definition for rcupdate control block. */
53 static struct rcu_ctrlblk rcu_ctrlblk = {
56 .lock = SPIN_LOCK_UNLOCKED,
57 .cpumask = CPU_MASK_NONE,
59 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
62 .lock = SPIN_LOCK_UNLOCKED,
63 .cpumask = CPU_MASK_NONE,
66 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
67 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
69 /* Fake initialization required by compiler */
70 static DEFINE_PER_CPU(struct tasklet_struct, rcu_tasklet) = {NULL};
71 static int blimit = 10;
72 static int qhimark = 10000;
73 static int qlowmark = 100;
75 static int rsinterval = 1000;
78 static atomic_t rcu_barrier_cpu_count;
79 static DEFINE_MUTEX(rcu_barrier_mutex);
80 static struct completion rcu_barrier_completion;
83 static void force_quiescent_state(struct rcu_data *rdp,
84 struct rcu_ctrlblk *rcp)
89 if (unlikely(rdp->qlen - rdp->last_rs_qlen > rsinterval)) {
90 rdp->last_rs_qlen = rdp->qlen;
92 * Don't send IPI to itself. With irqs disabled,
93 * rdp->cpu is the current cpu.
95 cpumask = rcp->cpumask;
96 cpu_clear(rdp->cpu, cpumask);
97 for_each_cpu_mask(cpu, cpumask)
98 smp_send_reschedule(cpu);
102 static inline void force_quiescent_state(struct rcu_data *rdp,
103 struct rcu_ctrlblk *rcp)
110 * call_rcu - Queue an RCU callback for invocation after a grace period.
111 * @head: structure to be used for queueing the RCU updates.
112 * @func: actual update function to be invoked after the grace period
114 * The update function will be invoked some time after a full grace
115 * period elapses, in other words after all currently executing RCU
116 * read-side critical sections have completed. RCU read-side critical
117 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
120 void fastcall call_rcu(struct rcu_head *head,
121 void (*func)(struct rcu_head *rcu))
124 struct rcu_data *rdp;
128 local_irq_save(flags);
129 rdp = &__get_cpu_var(rcu_data);
130 *rdp->nxttail = head;
131 rdp->nxttail = &head->next;
132 if (unlikely(++rdp->qlen > qhimark)) {
133 rdp->blimit = INT_MAX;
134 force_quiescent_state(rdp, &rcu_ctrlblk);
136 local_irq_restore(flags);
140 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
141 * @head: structure to be used for queueing the RCU updates.
142 * @func: actual update function to be invoked after the grace period
144 * The update function will be invoked some time after a full grace
145 * period elapses, in other words after all currently executing RCU
146 * read-side critical sections have completed. call_rcu_bh() assumes
147 * that the read-side critical sections end on completion of a softirq
148 * handler. This means that read-side critical sections in process
149 * context must not be interrupted by softirqs. This interface is to be
150 * used when most of the read-side critical sections are in softirq context.
151 * RCU read-side critical sections are delimited by rcu_read_lock() and
152 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
153 * and rcu_read_unlock_bh(), if in process context. These may be nested.
155 void fastcall call_rcu_bh(struct rcu_head *head,
156 void (*func)(struct rcu_head *rcu))
159 struct rcu_data *rdp;
163 local_irq_save(flags);
164 rdp = &__get_cpu_var(rcu_bh_data);
165 *rdp->nxttail = head;
166 rdp->nxttail = &head->next;
168 if (unlikely(++rdp->qlen > qhimark)) {
169 rdp->blimit = INT_MAX;
170 force_quiescent_state(rdp, &rcu_bh_ctrlblk);
173 local_irq_restore(flags);
177 * Return the number of RCU batches processed thus far. Useful
178 * for debug and statistics.
180 long rcu_batches_completed(void)
182 return rcu_ctrlblk.completed;
186 * Return the number of RCU batches processed thus far. Useful
187 * for debug and statistics.
189 long rcu_batches_completed_bh(void)
191 return rcu_bh_ctrlblk.completed;
194 static void rcu_barrier_callback(struct rcu_head *notused)
196 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
197 complete(&rcu_barrier_completion);
201 * Called with preemption disabled, and from cross-cpu IRQ context.
203 static void rcu_barrier_func(void *notused)
205 int cpu = smp_processor_id();
206 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
207 struct rcu_head *head;
209 head = &rdp->barrier;
210 atomic_inc(&rcu_barrier_cpu_count);
211 call_rcu(head, rcu_barrier_callback);
215 * rcu_barrier - Wait until all the in-flight RCUs are complete.
217 void rcu_barrier(void)
219 BUG_ON(in_interrupt());
220 /* Take cpucontrol mutex to protect against CPU hotplug */
221 mutex_lock(&rcu_barrier_mutex);
222 init_completion(&rcu_barrier_completion);
223 atomic_set(&rcu_barrier_cpu_count, 0);
224 on_each_cpu(rcu_barrier_func, NULL, 0, 1);
225 wait_for_completion(&rcu_barrier_completion);
226 mutex_unlock(&rcu_barrier_mutex);
228 EXPORT_SYMBOL_GPL(rcu_barrier);
231 * Invoke the completed RCU callbacks. They are expected to be in
234 static void rcu_do_batch(struct rcu_data *rdp)
236 struct rcu_head *next, *list;
239 list = rdp->donelist;
241 next = rdp->donelist = list->next;
245 if (++count >= rdp->blimit)
248 if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
249 rdp->blimit = blimit;
251 rdp->donetail = &rdp->donelist;
253 tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu));
257 * Grace period handling:
258 * The grace period handling consists out of two steps:
259 * - A new grace period is started.
260 * This is done by rcu_start_batch. The start is not broadcasted to
261 * all cpus, they must pick this up by comparing rcp->cur with
262 * rdp->quiescbatch. All cpus are recorded in the
263 * rcu_ctrlblk.cpumask bitmap.
264 * - All cpus must go through a quiescent state.
265 * Since the start of the grace period is not broadcasted, at least two
266 * calls to rcu_check_quiescent_state are required:
267 * The first call just notices that a new grace period is running. The
268 * following calls check if there was a quiescent state since the beginning
269 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
270 * the bitmap is empty, then the grace period is completed.
271 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
272 * period (if necessary).
275 * Register a new batch of callbacks, and start it up if there is currently no
276 * active batch and the batch to be registered has not already occurred.
277 * Caller must hold rcu_ctrlblk.lock.
279 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
281 if (rcp->next_pending &&
282 rcp->completed == rcp->cur) {
283 rcp->next_pending = 0;
285 * next_pending == 0 must be visible in
286 * __rcu_process_callbacks() before it can see new value of cur.
292 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
293 * Barrier Otherwise it can cause tickless idle CPUs to be
294 * included in rcp->cpumask, which will extend graceperiods
298 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
304 * cpu went through a quiescent state since the beginning of the grace period.
305 * Clear it from the cpu mask and complete the grace period if it was the last
306 * cpu. Start another grace period if someone has further entries pending
308 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
310 cpu_clear(cpu, rcp->cpumask);
311 if (cpus_empty(rcp->cpumask)) {
312 /* batch completed ! */
313 rcp->completed = rcp->cur;
314 rcu_start_batch(rcp);
319 * Check if the cpu has gone through a quiescent state (say context
320 * switch). If so and if it already hasn't done so in this RCU
321 * quiescent cycle, then indicate that it has done so.
323 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
324 struct rcu_data *rdp)
326 if (rdp->quiescbatch != rcp->cur) {
327 /* start new grace period: */
329 rdp->passed_quiesc = 0;
330 rdp->quiescbatch = rcp->cur;
334 /* Grace period already completed for this cpu?
335 * qs_pending is checked instead of the actual bitmap to avoid
336 * cacheline trashing.
338 if (!rdp->qs_pending)
342 * Was there a quiescent state since the beginning of the grace
343 * period? If no, then exit and wait for the next call.
345 if (!rdp->passed_quiesc)
349 spin_lock(&rcp->lock);
351 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
352 * during cpu startup. Ignore the quiescent state.
354 if (likely(rdp->quiescbatch == rcp->cur))
355 cpu_quiet(rdp->cpu, rcp);
357 spin_unlock(&rcp->lock);
361 #ifdef CONFIG_HOTPLUG_CPU
363 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
364 * locking requirements, the list it's pulling from has to belong to a cpu
365 * which is dead and hence not processing interrupts.
367 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
368 struct rcu_head **tail)
371 *this_rdp->nxttail = list;
373 this_rdp->nxttail = tail;
377 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
378 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
380 /* if the cpu going offline owns the grace period
381 * we can block indefinitely waiting for it, so flush
384 spin_lock_bh(&rcp->lock);
385 if (rcp->cur != rcp->completed)
386 cpu_quiet(rdp->cpu, rcp);
387 spin_unlock_bh(&rcp->lock);
388 rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
389 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
390 rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail);
393 static void rcu_offline_cpu(int cpu)
395 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
396 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
398 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
399 &per_cpu(rcu_data, cpu));
400 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
401 &per_cpu(rcu_bh_data, cpu));
402 put_cpu_var(rcu_data);
403 put_cpu_var(rcu_bh_data);
404 tasklet_kill_immediate(&per_cpu(rcu_tasklet, cpu), cpu);
409 static void rcu_offline_cpu(int cpu)
416 * This does the RCU processing work from tasklet context.
418 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
419 struct rcu_data *rdp)
421 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
422 *rdp->donetail = rdp->curlist;
423 rdp->donetail = rdp->curtail;
425 rdp->curtail = &rdp->curlist;
428 if (rdp->nxtlist && !rdp->curlist) {
430 rdp->curlist = rdp->nxtlist;
431 rdp->curtail = rdp->nxttail;
433 rdp->nxttail = &rdp->nxtlist;
437 * start the next batch of callbacks
440 /* determine batch number */
441 rdp->batch = rcp->cur + 1;
442 /* see the comment and corresponding wmb() in
443 * the rcu_start_batch()
447 if (!rcp->next_pending) {
448 /* and start it/schedule start if it's a new batch */
449 spin_lock(&rcp->lock);
450 rcp->next_pending = 1;
451 rcu_start_batch(rcp);
452 spin_unlock(&rcp->lock);
456 rcu_check_quiescent_state(rcp, rdp);
461 static void rcu_process_callbacks(unsigned long unused)
463 __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
464 __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
467 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
469 /* This cpu has pending rcu entries and the grace period
470 * for them has completed.
472 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch))
475 /* This cpu has no pending entries, but there are new entries */
476 if (!rdp->curlist && rdp->nxtlist)
479 /* This cpu has finished callbacks to invoke */
483 /* The rcu core waits for a quiescent state from the cpu */
484 if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
492 * Check to see if there is any immediate RCU-related work to be done
493 * by the current CPU, returning 1 if so. This function is part of the
494 * RCU implementation; it is -not- an exported member of the RCU API.
496 int rcu_pending(int cpu)
498 return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
499 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
503 * Check to see if any future RCU-related work will need to be done
504 * by the current CPU, even if none need be done immediately, returning
505 * 1 if so. This function is part of the RCU implementation; it is -not-
506 * an exported member of the RCU API.
508 int rcu_needs_cpu(int cpu)
510 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
511 struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
513 return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
516 void rcu_check_callbacks(int cpu, int user)
519 (idle_cpu(cpu) && !in_softirq() &&
520 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
522 rcu_bh_qsctr_inc(cpu);
523 } else if (!in_softirq())
524 rcu_bh_qsctr_inc(cpu);
525 tasklet_schedule(&per_cpu(rcu_tasklet, cpu));
528 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
529 struct rcu_data *rdp)
531 memset(rdp, 0, sizeof(*rdp));
532 rdp->curtail = &rdp->curlist;
533 rdp->nxttail = &rdp->nxtlist;
534 rdp->donetail = &rdp->donelist;
535 rdp->quiescbatch = rcp->completed;
538 rdp->blimit = blimit;
541 static void __devinit rcu_online_cpu(int cpu)
543 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
544 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
546 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
547 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
548 tasklet_init(&per_cpu(rcu_tasklet, cpu), rcu_process_callbacks, 0UL);
551 static int __devinit rcu_cpu_notify(struct notifier_block *self,
552 unsigned long action, void *hcpu)
554 long cpu = (long)hcpu;
560 rcu_offline_cpu(cpu);
568 static struct notifier_block __devinitdata rcu_nb = {
569 .notifier_call = rcu_cpu_notify,
573 * Initializes rcu mechanism. Assumed to be called early.
574 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
575 * Note that rcu_qsctr and friends are implicitly
576 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
578 void __init rcu_init(void)
580 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
581 (void *)(long)smp_processor_id());
582 /* Register notifier for non-boot CPUs */
583 register_cpu_notifier(&rcu_nb);
586 struct rcu_synchronize {
587 struct rcu_head head;
588 struct completion completion;
591 /* Because of FASTCALL declaration of complete, we use this wrapper */
592 static void wakeme_after_rcu(struct rcu_head *head)
594 struct rcu_synchronize *rcu;
596 rcu = container_of(head, struct rcu_synchronize, head);
597 complete(&rcu->completion);
601 * synchronize_rcu - wait until a grace period has elapsed.
603 * Control will return to the caller some time after a full grace
604 * period has elapsed, in other words after all currently executing RCU
605 * read-side critical sections have completed. RCU read-side critical
606 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
609 * If your read-side code is not protected by rcu_read_lock(), do -not-
610 * use synchronize_rcu().
612 void synchronize_rcu(void)
614 struct rcu_synchronize rcu;
616 init_completion(&rcu.completion);
617 /* Will wake me after RCU finished */
618 call_rcu(&rcu.head, wakeme_after_rcu);
621 wait_for_completion(&rcu.completion);
624 module_param(blimit, int, 0);
625 module_param(qhimark, int, 0);
626 module_param(qlowmark, int, 0);
628 module_param(rsinterval, int, 0);
630 EXPORT_SYMBOL_GPL(rcu_batches_completed);
631 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
632 EXPORT_SYMBOL_GPL(call_rcu);
633 EXPORT_SYMBOL_GPL(call_rcu_bh);
634 EXPORT_SYMBOL_GPL(synchronize_rcu);