2 * Read-Copy Update mechanism for mutual exclusion
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright IBM Corporation, 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 -
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>
50 #include <linux/time.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);
56 EXPORT_SYMBOL_GPL(rcu_lock_map);
60 /* Definition for rcupdate control block. */
61 static struct rcu_ctrlblk rcu_ctrlblk = {
65 .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock),
66 .cpumask = CPU_MASK_NONE,
68 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
72 .lock = __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk.lock),
73 .cpumask = CPU_MASK_NONE,
76 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
77 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
79 static int blimit = 10;
80 static int qhimark = 10000;
81 static int qlowmark = 100;
84 static void force_quiescent_state(struct rcu_data *rdp,
85 struct rcu_ctrlblk *rcp)
90 if (unlikely(!rcp->signaled)) {
93 * Don't send IPI to itself. With irqs disabled,
94 * rdp->cpu is the current cpu.
96 * cpu_online_map is updated by the _cpu_down()
97 * using __stop_machine(). Since we're in irqs disabled
98 * section, __stop_machine() is not exectuting, hence
99 * the cpu_online_map is stable.
101 * However, a cpu might have been offlined _just_ before
102 * we disabled irqs while entering here.
103 * And rcu subsystem might not yet have handled the CPU_DEAD
104 * notification, leading to the offlined cpu's bit
105 * being set in the rcp->cpumask.
107 * Hence cpumask = (rcp->cpumask & cpu_online_map) to prevent
108 * sending smp_reschedule() to an offlined CPU.
110 cpus_and(cpumask, rcp->cpumask, cpu_online_map);
111 cpu_clear(rdp->cpu, cpumask);
112 for_each_cpu_mask_nr(cpu, cpumask)
113 smp_send_reschedule(cpu);
117 static inline void force_quiescent_state(struct rcu_data *rdp,
118 struct rcu_ctrlblk *rcp)
124 static void __call_rcu(struct rcu_head *head, struct rcu_ctrlblk *rcp,
125 struct rcu_data *rdp)
128 smp_mb(); /* reads the most recently updated value of rcu->cur. */
131 * Determine the batch number of this callback.
133 * Using ACCESS_ONCE to avoid the following error when gcc eliminates
134 * local variable "batch" and emits codes like this:
135 * 1) rdp->batch = rcp->cur + 1 # gets old value
137 * 2)rcu_batch_after(rcp->cur + 1, rdp->batch) # gets new value
138 * then [*nxttail[0], *nxttail[1]) may contain callbacks
139 * that batch# = rdp->batch, see the comment of struct rcu_data.
141 batch = ACCESS_ONCE(rcp->cur) + 1;
143 if (rdp->nxtlist && rcu_batch_after(batch, rdp->batch)) {
144 /* process callbacks */
145 rdp->nxttail[0] = rdp->nxttail[1];
146 rdp->nxttail[1] = rdp->nxttail[2];
147 if (rcu_batch_after(batch - 1, rdp->batch))
148 rdp->nxttail[0] = rdp->nxttail[2];
152 *rdp->nxttail[2] = head;
153 rdp->nxttail[2] = &head->next;
155 if (unlikely(++rdp->qlen > qhimark)) {
156 rdp->blimit = INT_MAX;
157 force_quiescent_state(rdp, &rcu_ctrlblk);
162 * call_rcu - Queue an RCU callback for invocation after a grace period.
163 * @head: structure to be used for queueing the RCU updates.
164 * @func: actual update function to be invoked after the grace period
166 * The update function will be invoked some time after a full grace
167 * period elapses, in other words after all currently executing RCU
168 * read-side critical sections have completed. RCU read-side critical
169 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
172 void call_rcu(struct rcu_head *head,
173 void (*func)(struct rcu_head *rcu))
179 local_irq_save(flags);
180 __call_rcu(head, &rcu_ctrlblk, &__get_cpu_var(rcu_data));
181 local_irq_restore(flags);
183 EXPORT_SYMBOL_GPL(call_rcu);
186 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
187 * @head: structure to be used for queueing the RCU updates.
188 * @func: actual update function to be invoked after the grace period
190 * The update function will be invoked some time after a full grace
191 * period elapses, in other words after all currently executing RCU
192 * read-side critical sections have completed. call_rcu_bh() assumes
193 * that the read-side critical sections end on completion of a softirq
194 * handler. This means that read-side critical sections in process
195 * context must not be interrupted by softirqs. This interface is to be
196 * used when most of the read-side critical sections are in softirq context.
197 * RCU read-side critical sections are delimited by rcu_read_lock() and
198 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
199 * and rcu_read_unlock_bh(), if in process context. These may be nested.
201 void call_rcu_bh(struct rcu_head *head,
202 void (*func)(struct rcu_head *rcu))
208 local_irq_save(flags);
209 __call_rcu(head, &rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
210 local_irq_restore(flags);
212 EXPORT_SYMBOL_GPL(call_rcu_bh);
215 * Return the number of RCU batches processed thus far. Useful
216 * for debug and statistics.
218 long rcu_batches_completed(void)
220 return rcu_ctrlblk.completed;
222 EXPORT_SYMBOL_GPL(rcu_batches_completed);
225 * Return the number of RCU batches processed thus far. Useful
226 * for debug and statistics.
228 long rcu_batches_completed_bh(void)
230 return rcu_bh_ctrlblk.completed;
232 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
234 /* Raises the softirq for processing rcu_callbacks. */
235 static inline void raise_rcu_softirq(void)
237 raise_softirq(RCU_SOFTIRQ);
241 * Invoke the completed RCU callbacks. They are expected to be in
244 static void rcu_do_batch(struct rcu_data *rdp)
246 struct rcu_head *next, *list;
249 list = rdp->donelist;
255 if (++count >= rdp->blimit)
258 rdp->donelist = list;
263 if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
264 rdp->blimit = blimit;
267 rdp->donetail = &rdp->donelist;
273 * Grace period handling:
274 * The grace period handling consists out of two steps:
275 * - A new grace period is started.
276 * This is done by rcu_start_batch. The start is not broadcasted to
277 * all cpus, they must pick this up by comparing rcp->cur with
278 * rdp->quiescbatch. All cpus are recorded in the
279 * rcu_ctrlblk.cpumask bitmap.
280 * - All cpus must go through a quiescent state.
281 * Since the start of the grace period is not broadcasted, at least two
282 * calls to rcu_check_quiescent_state are required:
283 * The first call just notices that a new grace period is running. The
284 * following calls check if there was a quiescent state since the beginning
285 * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
286 * the bitmap is empty, then the grace period is completed.
287 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
288 * period (if necessary).
291 #ifdef CONFIG_DEBUG_RCU_STALL
293 static inline void record_gp_check_time(struct rcu_ctrlblk *rcp)
295 rcp->gp_check = get_seconds() + 3;
298 static void print_other_cpu_stall(struct rcu_ctrlblk *rcp)
303 /* Only let one CPU complain about others per time interval. */
305 spin_lock(&rcp->lock);
306 delta = get_seconds() - rcp->gp_check;
307 if (delta < 2L || cpus_empty(rcp->cpumask)) {
308 spin_unlock(&rcp->lock);
311 rcp->gp_check = get_seconds() + 30;
312 spin_unlock(&rcp->lock);
314 /* OK, time to rat on our buddy... */
316 printk(KERN_ERR "RCU detected CPU stalls:");
317 for_each_cpu_mask(cpu, rcp->cpumask)
319 printk(" (detected by %d, t=%lu/%lu)\n",
320 smp_processor_id(), get_seconds(), rcp->gp_check);
323 static void print_cpu_stall(struct rcu_ctrlblk *rcp)
325 printk(KERN_ERR "RCU detected CPU %d stall (t=%lu/%lu)\n",
326 smp_processor_id(), get_seconds(), rcp->gp_check);
328 spin_lock(&rcp->lock);
329 if ((long)(get_seconds() - rcp->gp_check) >= 0L)
330 rcp->gp_check = get_seconds() + 30;
331 spin_unlock(&rcp->lock);
334 static void check_cpu_stall(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
338 delta = get_seconds() - rcp->gp_check;
339 if (cpu_isset(smp_processor_id(), rcp->cpumask) && delta >= 0L) {
341 /* We haven't checked in, so go dump stack. */
343 print_cpu_stall(rcp);
346 if (!cpus_empty(rcp->cpumask) && delta >= 2L) {
347 /* They had two seconds to dump stack, so complain. */
348 print_other_cpu_stall(rcp);
353 #else /* #ifdef CONFIG_DEBUG_RCU_STALL */
355 static inline void record_gp_check_time(struct rcu_ctrlblk *rcp)
360 check_cpu_stall(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
364 #endif /* #else #ifdef CONFIG_DEBUG_RCU_STALL */
367 * Register a new batch of callbacks, and start it up if there is currently no
368 * active batch and the batch to be registered has not already occurred.
369 * Caller must hold rcu_ctrlblk.lock.
371 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
373 if (rcp->cur != rcp->pending &&
374 rcp->completed == rcp->cur) {
376 record_gp_check_time(rcp);
379 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
380 * Barrier Otherwise it can cause tickless idle CPUs to be
381 * included in rcp->cpumask, which will extend graceperiods
385 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
392 * cpu went through a quiescent state since the beginning of the grace period.
393 * Clear it from the cpu mask and complete the grace period if it was the last
394 * cpu. Start another grace period if someone has further entries pending
396 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
398 cpu_clear(cpu, rcp->cpumask);
399 if (cpus_empty(rcp->cpumask)) {
400 /* batch completed ! */
401 rcp->completed = rcp->cur;
402 rcu_start_batch(rcp);
407 * Check if the cpu has gone through a quiescent state (say context
408 * switch). If so and if it already hasn't done so in this RCU
409 * quiescent cycle, then indicate that it has done so.
411 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
412 struct rcu_data *rdp)
414 if (rdp->quiescbatch != rcp->cur) {
415 /* start new grace period: */
417 rdp->passed_quiesc = 0;
418 rdp->quiescbatch = rcp->cur;
422 /* Grace period already completed for this cpu?
423 * qs_pending is checked instead of the actual bitmap to avoid
424 * cacheline trashing.
426 if (!rdp->qs_pending)
430 * Was there a quiescent state since the beginning of the grace
431 * period? If no, then exit and wait for the next call.
433 if (!rdp->passed_quiesc)
437 spin_lock(&rcp->lock);
439 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
440 * during cpu startup. Ignore the quiescent state.
442 if (likely(rdp->quiescbatch == rcp->cur))
443 cpu_quiet(rdp->cpu, rcp);
445 spin_unlock(&rcp->lock);
449 #ifdef CONFIG_HOTPLUG_CPU
451 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
452 * locking requirements, the list it's pulling from has to belong to a cpu
453 * which is dead and hence not processing interrupts.
455 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
456 struct rcu_head **tail, long batch)
460 this_rdp->batch = batch;
461 *this_rdp->nxttail[2] = list;
462 this_rdp->nxttail[2] = tail;
467 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
468 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
470 /* if the cpu going offline owns the grace period
471 * we can block indefinitely waiting for it, so flush
474 spin_lock_bh(&rcp->lock);
475 if (rcp->cur != rcp->completed)
476 cpu_quiet(rdp->cpu, rcp);
477 spin_unlock_bh(&rcp->lock);
478 /* spin_lock implies smp_mb() */
479 rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail, rcp->cur + 1);
480 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail[2], rcp->cur + 1);
483 this_rdp->qlen += rdp->qlen;
487 static void rcu_offline_cpu(int cpu)
489 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
490 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
492 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
493 &per_cpu(rcu_data, cpu));
494 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
495 &per_cpu(rcu_bh_data, cpu));
496 put_cpu_var(rcu_data);
497 put_cpu_var(rcu_bh_data);
502 static void rcu_offline_cpu(int cpu)
509 * This does the RCU processing work from softirq context.
511 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
512 struct rcu_data *rdp)
518 * move the other grace-period-completed entries to
519 * [rdp->nxtlist, *rdp->nxttail[0]) temporarily
521 if (!rcu_batch_before(rcp->completed, rdp->batch))
522 rdp->nxttail[0] = rdp->nxttail[1] = rdp->nxttail[2];
523 else if (!rcu_batch_before(rcp->completed, rdp->batch - 1))
524 rdp->nxttail[0] = rdp->nxttail[1];
527 * the grace period for entries in
528 * [rdp->nxtlist, *rdp->nxttail[0]) has completed and
529 * move these entries to donelist
531 if (rdp->nxttail[0] != &rdp->nxtlist) {
532 *rdp->donetail = rdp->nxtlist;
533 rdp->donetail = rdp->nxttail[0];
534 rdp->nxtlist = *rdp->nxttail[0];
535 *rdp->donetail = NULL;
537 if (rdp->nxttail[1] == rdp->nxttail[0])
538 rdp->nxttail[1] = &rdp->nxtlist;
539 if (rdp->nxttail[2] == rdp->nxttail[0])
540 rdp->nxttail[2] = &rdp->nxtlist;
541 rdp->nxttail[0] = &rdp->nxtlist;
546 if (rcu_batch_after(rdp->batch, rcp->pending)) {
547 /* and start it/schedule start if it's a new batch */
548 spin_lock(&rcp->lock);
549 if (rcu_batch_after(rdp->batch, rcp->pending)) {
550 rcp->pending = rdp->batch;
551 rcu_start_batch(rcp);
553 spin_unlock(&rcp->lock);
557 rcu_check_quiescent_state(rcp, rdp);
562 static void rcu_process_callbacks(struct softirq_action *unused)
564 __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
565 __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
568 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
570 /* Check for CPU stalls, if enabled. */
571 check_cpu_stall(rcp, rdp);
575 * This cpu has pending rcu entries and the grace period
576 * for them has completed.
578 if (!rcu_batch_before(rcp->completed, rdp->batch))
580 if (!rcu_batch_before(rcp->completed, rdp->batch - 1) &&
581 rdp->nxttail[0] != rdp->nxttail[1])
583 if (rdp->nxttail[0] != &rdp->nxtlist)
587 * This cpu has pending rcu entries and the new batch
588 * for then hasn't been started nor scheduled start
590 if (rcu_batch_after(rdp->batch, rcp->pending))
594 /* This cpu has finished callbacks to invoke */
598 /* The rcu core waits for a quiescent state from the cpu */
599 if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
607 * Check to see if there is any immediate RCU-related work to be done
608 * by the current CPU, returning 1 if so. This function is part of the
609 * RCU implementation; it is -not- an exported member of the RCU API.
611 int rcu_pending(int cpu)
613 return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
614 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
618 * Check to see if any future RCU-related work will need to be done
619 * by the current CPU, even if none need be done immediately, returning
620 * 1 if so. This function is part of the RCU implementation; it is -not-
621 * an exported member of the RCU API.
623 int rcu_needs_cpu(int cpu)
625 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
626 struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
628 return !!rdp->nxtlist || !!rdp_bh->nxtlist || rcu_pending(cpu);
631 void rcu_check_callbacks(int cpu, int user)
634 (idle_cpu(cpu) && !in_softirq() &&
635 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
638 * Get here if this CPU took its interrupt from user
639 * mode or from the idle loop, and if this is not a
640 * nested interrupt. In this case, the CPU is in
641 * a quiescent state, so count it.
643 * Also do a memory barrier. This is needed to handle
644 * the case where writes from a preempt-disable section
645 * of code get reordered into schedule() by this CPU's
646 * write buffer. The memory barrier makes sure that
647 * the rcu_qsctr_inc() and rcu_bh_qsctr_inc() are see
648 * by other CPUs to happen after any such write.
651 smp_mb(); /* See above block comment. */
653 rcu_bh_qsctr_inc(cpu);
655 } else if (!in_softirq()) {
658 * Get here if this CPU did not take its interrupt from
659 * softirq, in other words, if it is not interrupting
660 * a rcu_bh read-side critical section. This is an _bh
661 * critical section, so count it. The memory barrier
662 * is needed for the same reason as is the above one.
665 smp_mb(); /* See above block comment. */
666 rcu_bh_qsctr_inc(cpu);
671 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
672 struct rcu_data *rdp)
674 memset(rdp, 0, sizeof(*rdp));
675 rdp->nxttail[0] = rdp->nxttail[1] = rdp->nxttail[2] = &rdp->nxtlist;
676 rdp->donetail = &rdp->donelist;
677 rdp->quiescbatch = rcp->completed;
680 rdp->blimit = blimit;
683 static void __cpuinit rcu_online_cpu(int cpu)
685 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
686 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
688 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
689 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
690 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
693 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
694 unsigned long action, void *hcpu)
696 long cpu = (long)hcpu;
700 case CPU_UP_PREPARE_FROZEN:
704 case CPU_DEAD_FROZEN:
705 rcu_offline_cpu(cpu);
713 static struct notifier_block __cpuinitdata rcu_nb = {
714 .notifier_call = rcu_cpu_notify,
718 * Initializes rcu mechanism. Assumed to be called early.
719 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
720 * Note that rcu_qsctr and friends are implicitly
721 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
723 void __init __rcu_init(void)
725 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
726 (void *)(long)smp_processor_id());
727 /* Register notifier for non-boot CPUs */
728 register_cpu_notifier(&rcu_nb);
731 module_param(blimit, int, 0);
732 module_param(qhimark, int, 0);
733 module_param(qlowmark, int, 0);