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/interrupt.h>
39 #include <linux/sched.h>
40 #include <asm/atomic.h>
41 #include <linux/bitops.h>
42 #include <linux/module.h>
43 #include <linux/completion.h>
44 #include <linux/moduleparam.h>
45 #include <linux/percpu.h>
46 #include <linux/notifier.h>
47 #include <linux/rcupdate.h>
48 #include <linux/rcuref.h>
49 #include <linux/cpu.h>
51 /* Definition for rcupdate control block. */
52 struct rcu_ctrlblk rcu_ctrlblk =
53 { .cur = -300, .completed = -300 };
54 struct rcu_ctrlblk rcu_bh_ctrlblk =
55 { .cur = -300, .completed = -300 };
57 /* Bookkeeping of the progress of the grace period */
59 spinlock_t lock; /* Guard this struct and writes to rcu_ctrlblk */
60 cpumask_t cpumask; /* CPUs that need to switch in order */
61 /* for current batch to proceed. */
64 static struct rcu_state rcu_state ____cacheline_maxaligned_in_smp =
65 {.lock = SPIN_LOCK_UNLOCKED, .cpumask = CPU_MASK_NONE };
66 static struct rcu_state rcu_bh_state ____cacheline_maxaligned_in_smp =
67 {.lock = SPIN_LOCK_UNLOCKED, .cpumask = CPU_MASK_NONE };
69 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
70 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
72 /* Fake initialization required by compiler */
73 static DEFINE_PER_CPU(struct tasklet_struct, rcu_tasklet) = {NULL};
74 static int maxbatch = 10000;
76 #ifndef __HAVE_ARCH_CMPXCHG
78 * We use an array of spinlocks for the rcurefs -- similar to ones in sparc
79 * 32 bit atomic_t implementations, and a hash function similar to that
80 * for our refcounting needs.
81 * Can't help multiprocessors which donot have cmpxchg :(
84 spinlock_t __rcuref_hash[RCUREF_HASH_SIZE] = {
85 [0 ... (RCUREF_HASH_SIZE-1)] = SPIN_LOCK_UNLOCKED
90 * call_rcu - Queue an RCU callback for invocation after a grace period.
91 * @head: structure to be used for queueing the RCU updates.
92 * @func: actual update function to be invoked after the grace period
94 * The update function will be invoked some time after a full grace
95 * period elapses, in other words after all currently executing RCU
96 * read-side critical sections have completed. RCU read-side critical
97 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
100 void fastcall call_rcu(struct rcu_head *head,
101 void (*func)(struct rcu_head *rcu))
104 struct rcu_data *rdp;
108 local_irq_save(flags);
109 rdp = &__get_cpu_var(rcu_data);
110 *rdp->nxttail = head;
111 rdp->nxttail = &head->next;
113 if (unlikely(++rdp->count > 10000))
116 local_irq_restore(flags);
119 static atomic_t rcu_barrier_cpu_count;
120 static struct semaphore rcu_barrier_sema;
121 static struct completion rcu_barrier_completion;
124 * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
125 * @head: structure to be used for queueing the RCU updates.
126 * @func: actual update function to be invoked after the grace period
128 * The update function will be invoked some time after a full grace
129 * period elapses, in other words after all currently executing RCU
130 * read-side critical sections have completed. call_rcu_bh() assumes
131 * that the read-side critical sections end on completion of a softirq
132 * handler. This means that read-side critical sections in process
133 * context must not be interrupted by softirqs. This interface is to be
134 * used when most of the read-side critical sections are in softirq context.
135 * RCU read-side critical sections are delimited by rcu_read_lock() and
136 * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
137 * and rcu_read_unlock_bh(), if in process context. These may be nested.
139 void fastcall call_rcu_bh(struct rcu_head *head,
140 void (*func)(struct rcu_head *rcu))
143 struct rcu_data *rdp;
147 local_irq_save(flags);
148 rdp = &__get_cpu_var(rcu_bh_data);
149 *rdp->nxttail = head;
150 rdp->nxttail = &head->next;
153 * Should we directly call rcu_do_batch() here ?
154 * if (unlikely(rdp->count > 10000))
157 local_irq_restore(flags);
161 * Return the number of RCU batches processed thus far. Useful
162 * for debug and statistics.
164 long rcu_batches_completed(void)
166 return rcu_ctrlblk.completed;
169 static void rcu_barrier_callback(struct rcu_head *notused)
171 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
172 complete(&rcu_barrier_completion);
176 * Called with preemption disabled, and from cross-cpu IRQ context.
178 static void rcu_barrier_func(void *notused)
180 int cpu = smp_processor_id();
181 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
182 struct rcu_head *head;
184 head = &rdp->barrier;
185 atomic_inc(&rcu_barrier_cpu_count);
186 call_rcu(head, rcu_barrier_callback);
190 * rcu_barrier - Wait until all the in-flight RCUs are complete.
192 void rcu_barrier(void)
194 BUG_ON(in_interrupt());
195 /* Take cpucontrol semaphore to protect against CPU hotplug */
196 down(&rcu_barrier_sema);
197 init_completion(&rcu_barrier_completion);
198 atomic_set(&rcu_barrier_cpu_count, 0);
199 on_each_cpu(rcu_barrier_func, NULL, 0, 1);
200 wait_for_completion(&rcu_barrier_completion);
201 up(&rcu_barrier_sema);
203 EXPORT_SYMBOL_GPL(rcu_barrier);
206 * Invoke the completed RCU callbacks. They are expected to be in
209 static void rcu_do_batch(struct rcu_data *rdp)
211 struct rcu_head *next, *list;
214 list = rdp->donelist;
216 next = rdp->donelist = list->next;
220 if (++count >= maxbatch)
224 rdp->donetail = &rdp->donelist;
226 tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu));
230 * Grace period handling:
231 * The grace period handling consists out of two steps:
232 * - A new grace period is started.
233 * This is done by rcu_start_batch. The start is not broadcasted to
234 * all cpus, they must pick this up by comparing rcp->cur with
235 * rdp->quiescbatch. All cpus are recorded in the
236 * rcu_state.cpumask bitmap.
237 * - All cpus must go through a quiescent state.
238 * Since the start of the grace period is not broadcasted, at least two
239 * calls to rcu_check_quiescent_state are required:
240 * The first call just notices that a new grace period is running. The
241 * following calls check if there was a quiescent state since the beginning
242 * of the grace period. If so, it updates rcu_state.cpumask. If
243 * the bitmap is empty, then the grace period is completed.
244 * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
245 * period (if necessary).
248 * Register a new batch of callbacks, and start it up if there is currently no
249 * active batch and the batch to be registered has not already occurred.
250 * Caller must hold rcu_state.lock.
252 static void rcu_start_batch(struct rcu_ctrlblk *rcp, struct rcu_state *rsp,
256 rcp->next_pending = 1;
258 if (rcp->next_pending &&
259 rcp->completed == rcp->cur) {
260 rcp->next_pending = 0;
262 * next_pending == 0 must be visible in
263 * __rcu_process_callbacks() before it can see new value of cur.
269 * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
270 * Barrier Otherwise it can cause tickless idle CPUs to be
271 * included in rsp->cpumask, which will extend graceperiods
275 cpus_andnot(rsp->cpumask, cpu_online_map, nohz_cpu_mask);
281 * cpu went through a quiescent state since the beginning of the grace period.
282 * Clear it from the cpu mask and complete the grace period if it was the last
283 * cpu. Start another grace period if someone has further entries pending
285 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp, struct rcu_state *rsp)
287 cpu_clear(cpu, rsp->cpumask);
288 if (cpus_empty(rsp->cpumask)) {
289 /* batch completed ! */
290 rcp->completed = rcp->cur;
291 rcu_start_batch(rcp, rsp, 0);
296 * Check if the cpu has gone through a quiescent state (say context
297 * switch). If so and if it already hasn't done so in this RCU
298 * quiescent cycle, then indicate that it has done so.
300 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
301 struct rcu_state *rsp, struct rcu_data *rdp)
303 if (rdp->quiescbatch != rcp->cur) {
304 /* start new grace period: */
306 rdp->passed_quiesc = 0;
307 rdp->quiescbatch = rcp->cur;
311 /* Grace period already completed for this cpu?
312 * qs_pending is checked instead of the actual bitmap to avoid
313 * cacheline trashing.
315 if (!rdp->qs_pending)
319 * Was there a quiescent state since the beginning of the grace
320 * period? If no, then exit and wait for the next call.
322 if (!rdp->passed_quiesc)
326 spin_lock(&rsp->lock);
328 * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
329 * during cpu startup. Ignore the quiescent state.
331 if (likely(rdp->quiescbatch == rcp->cur))
332 cpu_quiet(rdp->cpu, rcp, rsp);
334 spin_unlock(&rsp->lock);
338 #ifdef CONFIG_HOTPLUG_CPU
340 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
341 * locking requirements, the list it's pulling from has to belong to a cpu
342 * which is dead and hence not processing interrupts.
344 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
345 struct rcu_head **tail)
348 *this_rdp->nxttail = list;
350 this_rdp->nxttail = tail;
354 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
355 struct rcu_ctrlblk *rcp, struct rcu_state *rsp, struct rcu_data *rdp)
357 /* if the cpu going offline owns the grace period
358 * we can block indefinitely waiting for it, so flush
361 spin_lock_bh(&rsp->lock);
362 if (rcp->cur != rcp->completed)
363 cpu_quiet(rdp->cpu, rcp, rsp);
364 spin_unlock_bh(&rsp->lock);
365 rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
366 rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
369 static void rcu_offline_cpu(int cpu)
371 struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
372 struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
374 __rcu_offline_cpu(this_rdp, &rcu_ctrlblk, &rcu_state,
375 &per_cpu(rcu_data, cpu));
376 __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk, &rcu_bh_state,
377 &per_cpu(rcu_bh_data, cpu));
378 put_cpu_var(rcu_data);
379 put_cpu_var(rcu_bh_data);
380 tasklet_kill_immediate(&per_cpu(rcu_tasklet, cpu), cpu);
385 static void rcu_offline_cpu(int cpu)
392 * This does the RCU processing work from tasklet context.
394 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
395 struct rcu_state *rsp, struct rcu_data *rdp)
397 if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
398 *rdp->donetail = rdp->curlist;
399 rdp->donetail = rdp->curtail;
401 rdp->curtail = &rdp->curlist;
405 if (rdp->nxtlist && !rdp->curlist) {
406 rdp->curlist = rdp->nxtlist;
407 rdp->curtail = rdp->nxttail;
409 rdp->nxttail = &rdp->nxtlist;
413 * start the next batch of callbacks
416 /* determine batch number */
417 rdp->batch = rcp->cur + 1;
418 /* see the comment and corresponding wmb() in
419 * the rcu_start_batch()
423 if (!rcp->next_pending) {
424 /* and start it/schedule start if it's a new batch */
425 spin_lock(&rsp->lock);
426 rcu_start_batch(rcp, rsp, 1);
427 spin_unlock(&rsp->lock);
432 rcu_check_quiescent_state(rcp, rsp, rdp);
437 static void rcu_process_callbacks(unsigned long unused)
439 __rcu_process_callbacks(&rcu_ctrlblk, &rcu_state,
440 &__get_cpu_var(rcu_data));
441 __rcu_process_callbacks(&rcu_bh_ctrlblk, &rcu_bh_state,
442 &__get_cpu_var(rcu_bh_data));
445 void rcu_check_callbacks(int cpu, int user)
448 (idle_cpu(cpu) && !in_softirq() &&
449 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
451 rcu_bh_qsctr_inc(cpu);
452 } else if (!in_softirq())
453 rcu_bh_qsctr_inc(cpu);
454 tasklet_schedule(&per_cpu(rcu_tasklet, cpu));
457 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
458 struct rcu_data *rdp)
460 memset(rdp, 0, sizeof(*rdp));
461 rdp->curtail = &rdp->curlist;
462 rdp->nxttail = &rdp->nxtlist;
463 rdp->donetail = &rdp->donelist;
464 rdp->quiescbatch = rcp->completed;
469 static void __devinit rcu_online_cpu(int cpu)
471 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
472 struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
474 rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
475 rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
476 tasklet_init(&per_cpu(rcu_tasklet, cpu), rcu_process_callbacks, 0UL);
479 static int __devinit rcu_cpu_notify(struct notifier_block *self,
480 unsigned long action, void *hcpu)
482 long cpu = (long)hcpu;
488 rcu_offline_cpu(cpu);
496 static struct notifier_block __devinitdata rcu_nb = {
497 .notifier_call = rcu_cpu_notify,
501 * Initializes rcu mechanism. Assumed to be called early.
502 * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
503 * Note that rcu_qsctr and friends are implicitly
504 * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
506 void __init rcu_init(void)
508 sema_init(&rcu_barrier_sema, 1);
509 rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
510 (void *)(long)smp_processor_id());
511 /* Register notifier for non-boot CPUs */
512 register_cpu_notifier(&rcu_nb);
515 struct rcu_synchronize {
516 struct rcu_head head;
517 struct completion completion;
520 /* Because of FASTCALL declaration of complete, we use this wrapper */
521 static void wakeme_after_rcu(struct rcu_head *head)
523 struct rcu_synchronize *rcu;
525 rcu = container_of(head, struct rcu_synchronize, head);
526 complete(&rcu->completion);
530 * synchronize_rcu - wait until a grace period has elapsed.
532 * Control will return to the caller some time after a full grace
533 * period has elapsed, in other words after all currently executing RCU
534 * read-side critical sections have completed. RCU read-side critical
535 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
538 * If your read-side code is not protected by rcu_read_lock(), do -not-
539 * use synchronize_rcu().
541 void synchronize_rcu(void)
543 struct rcu_synchronize rcu;
545 init_completion(&rcu.completion);
546 /* Will wake me after RCU finished */
547 call_rcu(&rcu.head, wakeme_after_rcu);
550 wait_for_completion(&rcu.completion);
554 * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
556 void synchronize_kernel(void)
561 module_param(maxbatch, int, 0);
562 EXPORT_SYMBOL_GPL(rcu_batches_completed);
563 EXPORT_SYMBOL(call_rcu); /* WARNING: GPL-only in April 2006. */
564 EXPORT_SYMBOL(call_rcu_bh); /* WARNING: GPL-only in April 2006. */
565 EXPORT_SYMBOL_GPL(synchronize_rcu);
566 EXPORT_SYMBOL(synchronize_kernel); /* WARNING: GPL-only in April 2006. */