2 * Sleepable 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, 2006
19 * Copyright (C) Fujitsu, 2012
21 * Author: Paul McKenney <paulmck@us.ibm.com>
22 * Lai Jiangshan <laijs@cn.fujitsu.com>
24 * For detailed explanation of Read-Copy Update mechanism see -
25 * Documentation/RCU/ *.txt
29 #include <linux/export.h>
30 #include <linux/mutex.h>
31 #include <linux/percpu.h>
32 #include <linux/preempt.h>
33 #include <linux/rcupdate.h>
34 #include <linux/sched.h>
35 #include <linux/smp.h>
36 #include <linux/delay.h>
37 #include <linux/srcu.h>
40 * Initialize an rcu_batch structure to empty.
42 static inline void rcu_batch_init(struct rcu_batch *b)
49 * Enqueue a callback onto the tail of the specified rcu_batch structure.
51 static inline void rcu_batch_queue(struct rcu_batch *b, struct rcu_head *head)
54 b->tail = &head->next;
58 * Is the specified rcu_batch structure empty?
60 static inline bool rcu_batch_empty(struct rcu_batch *b)
62 return b->tail == &b->head;
66 * Remove the callback at the head of the specified rcu_batch structure
67 * and return a pointer to it, or return NULL if the structure is empty.
69 static inline struct rcu_head *rcu_batch_dequeue(struct rcu_batch *b)
71 struct rcu_head *head;
73 if (rcu_batch_empty(b))
78 if (b->tail == &head->next)
85 * Move all callbacks from the rcu_batch structure specified by "from" to
86 * the structure specified by "to".
88 static inline void rcu_batch_move(struct rcu_batch *to, struct rcu_batch *from)
90 if (!rcu_batch_empty(from)) {
91 *to->tail = from->head;
92 to->tail = from->tail;
97 static int init_srcu_struct_fields(struct srcu_struct *sp)
100 spin_lock_init(&sp->queue_lock);
102 rcu_batch_init(&sp->batch_queue);
103 rcu_batch_init(&sp->batch_check0);
104 rcu_batch_init(&sp->batch_check1);
105 rcu_batch_init(&sp->batch_done);
106 INIT_DELAYED_WORK(&sp->work, process_srcu);
107 sp->per_cpu_ref = alloc_percpu(struct srcu_struct_array);
108 return sp->per_cpu_ref ? 0 : -ENOMEM;
111 #ifdef CONFIG_DEBUG_LOCK_ALLOC
113 int __init_srcu_struct(struct srcu_struct *sp, const char *name,
114 struct lock_class_key *key)
116 /* Don't re-initialize a lock while it is held. */
117 debug_check_no_locks_freed((void *)sp, sizeof(*sp));
118 lockdep_init_map(&sp->dep_map, name, key, 0);
119 return init_srcu_struct_fields(sp);
121 EXPORT_SYMBOL_GPL(__init_srcu_struct);
123 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
126 * init_srcu_struct - initialize a sleep-RCU structure
127 * @sp: structure to initialize.
129 * Must invoke this on a given srcu_struct before passing that srcu_struct
130 * to any other function. Each srcu_struct represents a separate domain
131 * of SRCU protection.
133 int init_srcu_struct(struct srcu_struct *sp)
135 return init_srcu_struct_fields(sp);
137 EXPORT_SYMBOL_GPL(init_srcu_struct);
139 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
142 * Returns approximate total of the readers' ->seq[] values for the
143 * rank of per-CPU counters specified by idx.
145 static unsigned long srcu_readers_seq_idx(struct srcu_struct *sp, int idx)
148 unsigned long sum = 0;
151 for_each_possible_cpu(cpu) {
152 t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->seq[idx]);
159 * Returns approximate number of readers active on the specified rank
160 * of the per-CPU ->c[] counters.
162 static unsigned long srcu_readers_active_idx(struct srcu_struct *sp, int idx)
165 unsigned long sum = 0;
168 for_each_possible_cpu(cpu) {
169 t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[idx]);
176 * Return true if the number of pre-existing readers is determined to
177 * be stably zero. An example unstable zero can occur if the call
178 * to srcu_readers_active_idx() misses an __srcu_read_lock() increment,
179 * but due to task migration, sees the corresponding __srcu_read_unlock()
180 * decrement. This can happen because srcu_readers_active_idx() takes
181 * time to sum the array, and might in fact be interrupted or preempted
182 * partway through the summation.
184 static bool srcu_readers_active_idx_check(struct srcu_struct *sp, int idx)
188 seq = srcu_readers_seq_idx(sp, idx);
191 * The following smp_mb() A pairs with the smp_mb() B located in
192 * __srcu_read_lock(). This pairing ensures that if an
193 * __srcu_read_lock() increments its counter after the summation
194 * in srcu_readers_active_idx(), then the corresponding SRCU read-side
195 * critical section will see any changes made prior to the start
196 * of the current SRCU grace period.
198 * Also, if the above call to srcu_readers_seq_idx() saw the
199 * increment of ->seq[], then the call to srcu_readers_active_idx()
200 * must see the increment of ->c[].
205 * Note that srcu_readers_active_idx() can incorrectly return
206 * zero even though there is a pre-existing reader throughout.
207 * To see this, suppose that task A is in a very long SRCU
208 * read-side critical section that started on CPU 0, and that
209 * no other reader exists, so that the sum of the counters
210 * is equal to one. Then suppose that task B starts executing
211 * srcu_readers_active_idx(), summing up to CPU 1, and then that
212 * task C starts reading on CPU 0, so that its increment is not
213 * summed, but finishes reading on CPU 2, so that its decrement
214 * -is- summed. Then when task B completes its sum, it will
215 * incorrectly get zero, despite the fact that task A has been
216 * in its SRCU read-side critical section the whole time.
218 * We therefore do a validation step should srcu_readers_active_idx()
221 if (srcu_readers_active_idx(sp, idx) != 0)
225 * The remainder of this function is the validation step.
226 * The following smp_mb() D pairs with the smp_mb() C in
227 * __srcu_read_unlock(). If the __srcu_read_unlock() was seen
228 * by srcu_readers_active_idx() above, then any destructive
229 * operation performed after the grace period will happen after
230 * the corresponding SRCU read-side critical section.
232 * Note that there can be at most NR_CPUS worth of readers using
233 * the old index, which is not enough to overflow even a 32-bit
234 * integer. (Yes, this does mean that systems having more than
235 * a billion or so CPUs need to be 64-bit systems.) Therefore,
236 * the sum of the ->seq[] counters cannot possibly overflow.
237 * Therefore, the only way that the return values of the two
238 * calls to srcu_readers_seq_idx() can be equal is if there were
239 * no increments of the corresponding rank of ->seq[] counts
240 * in the interim. But the missed-increment scenario laid out
241 * above includes an increment of the ->seq[] counter by
242 * the corresponding __srcu_read_lock(). Therefore, if this
243 * scenario occurs, the return values from the two calls to
244 * srcu_readers_seq_idx() will differ, and thus the validation
245 * step below suffices.
249 return srcu_readers_seq_idx(sp, idx) == seq;
253 * srcu_readers_active - returns approximate number of readers.
254 * @sp: which srcu_struct to count active readers (holding srcu_read_lock).
256 * Note that this is not an atomic primitive, and can therefore suffer
257 * severe errors when invoked on an active srcu_struct. That said, it
258 * can be useful as an error check at cleanup time.
260 static int srcu_readers_active(struct srcu_struct *sp)
263 unsigned long sum = 0;
265 for_each_possible_cpu(cpu) {
266 sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[0]);
267 sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[1]);
273 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
274 * @sp: structure to clean up.
276 * Must invoke this after you are finished using a given srcu_struct that
277 * was initialized via init_srcu_struct(), else you leak memory.
279 void cleanup_srcu_struct(struct srcu_struct *sp)
283 sum = srcu_readers_active(sp);
284 WARN_ON(sum); /* Leakage unless caller handles error. */
287 free_percpu(sp->per_cpu_ref);
288 sp->per_cpu_ref = NULL;
290 EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
293 * Counts the new reader in the appropriate per-CPU element of the
294 * srcu_struct. Must be called from process context.
295 * Returns an index that must be passed to the matching srcu_read_unlock().
297 int __srcu_read_lock(struct srcu_struct *sp)
302 idx = rcu_dereference_index_check(sp->completed,
303 rcu_read_lock_sched_held()) & 0x1;
304 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) += 1;
305 smp_mb(); /* B */ /* Avoid leaking the critical section. */
306 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->seq[idx]) += 1;
310 EXPORT_SYMBOL_GPL(__srcu_read_lock);
313 * Removes the count for the old reader from the appropriate per-CPU
314 * element of the srcu_struct. Note that this may well be a different
315 * CPU than that which was incremented by the corresponding srcu_read_lock().
316 * Must be called from process context.
318 void __srcu_read_unlock(struct srcu_struct *sp, int idx)
321 smp_mb(); /* C */ /* Avoid leaking the critical section. */
322 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) -= 1;
325 EXPORT_SYMBOL_GPL(__srcu_read_unlock);
328 * We use an adaptive strategy for synchronize_srcu() and especially for
329 * synchronize_srcu_expedited(). We spin for a fixed time period
330 * (defined below) to allow SRCU readers to exit their read-side critical
331 * sections. If there are still some readers after 10 microseconds,
332 * we repeatedly block for 1-millisecond time periods. This approach
333 * has done well in testing, so there is no need for a config parameter.
335 #define SRCU_RETRY_CHECK_DELAY 5
336 #define SYNCHRONIZE_SRCU_TRYCOUNT 2
337 #define SYNCHRONIZE_SRCU_EXP_TRYCOUNT 12
340 * @@@ Wait until all pre-existing readers complete. Such readers
341 * will have used the index specified by "idx".
342 * the caller should ensures the ->completed is not changed while checking
343 * and idx = (->completed & 1) ^ 1
345 static bool try_check_zero(struct srcu_struct *sp, int idx, int trycount)
348 if (srcu_readers_active_idx_check(sp, idx))
352 udelay(SRCU_RETRY_CHECK_DELAY);
357 * Increment the ->completed counter so that future SRCU readers will
358 * use the other rank of the ->c[] and ->seq[] arrays. This allows
359 * us to wait for pre-existing readers in a starvation-free manner.
361 static void srcu_flip(struct srcu_struct *sp)
367 * Enqueue an SRCU callback on the specified srcu_struct structure,
368 * initiating grace-period processing if it is not already running.
370 void call_srcu(struct srcu_struct *sp, struct rcu_head *head,
371 void (*func)(struct rcu_head *head))
377 spin_lock_irqsave(&sp->queue_lock, flags);
378 rcu_batch_queue(&sp->batch_queue, head);
381 schedule_delayed_work(&sp->work, 0);
383 spin_unlock_irqrestore(&sp->queue_lock, flags);
385 EXPORT_SYMBOL_GPL(call_srcu);
387 struct rcu_synchronize {
388 struct rcu_head head;
389 struct completion completion;
393 * Awaken the corresponding synchronize_srcu() instance now that a
394 * grace period has elapsed.
396 static void wakeme_after_rcu(struct rcu_head *head)
398 struct rcu_synchronize *rcu;
400 rcu = container_of(head, struct rcu_synchronize, head);
401 complete(&rcu->completion);
404 static void srcu_advance_batches(struct srcu_struct *sp, int trycount);
405 static void srcu_reschedule(struct srcu_struct *sp);
408 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
410 static void __synchronize_srcu(struct srcu_struct *sp, int trycount)
412 struct rcu_synchronize rcu;
413 struct rcu_head *head = &rcu.head;
416 rcu_lockdep_assert(!lock_is_held(&sp->dep_map) &&
417 !lock_is_held(&rcu_bh_lock_map) &&
418 !lock_is_held(&rcu_lock_map) &&
419 !lock_is_held(&rcu_sched_lock_map),
420 "Illegal synchronize_srcu() in same-type SRCU (or RCU) read-side critical section");
422 init_completion(&rcu.completion);
425 head->func = wakeme_after_rcu;
426 spin_lock_irq(&sp->queue_lock);
428 /* steal the processing owner */
430 rcu_batch_queue(&sp->batch_check0, head);
431 spin_unlock_irq(&sp->queue_lock);
433 srcu_advance_batches(sp, trycount);
434 if (!rcu_batch_empty(&sp->batch_done)) {
435 BUG_ON(sp->batch_done.head != head);
436 rcu_batch_dequeue(&sp->batch_done);
439 /* give the processing owner to work_struct */
442 rcu_batch_queue(&sp->batch_queue, head);
443 spin_unlock_irq(&sp->queue_lock);
447 wait_for_completion(&rcu.completion);
451 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
452 * @sp: srcu_struct with which to synchronize.
454 * Flip the completed counter, and wait for the old count to drain to zero.
455 * As with classic RCU, the updater must use some separate means of
456 * synchronizing concurrent updates. Can block; must be called from
459 * Note that it is illegal to call synchronize_srcu() from the corresponding
460 * SRCU read-side critical section; doing so will result in deadlock.
461 * However, it is perfectly legal to call synchronize_srcu() on one
462 * srcu_struct from some other srcu_struct's read-side critical section.
464 void synchronize_srcu(struct srcu_struct *sp)
466 __synchronize_srcu(sp, SYNCHRONIZE_SRCU_TRYCOUNT);
468 EXPORT_SYMBOL_GPL(synchronize_srcu);
471 * synchronize_srcu_expedited - Brute-force SRCU grace period
472 * @sp: srcu_struct with which to synchronize.
474 * Wait for an SRCU grace period to elapse, but be more aggressive about
475 * spinning rather than blocking when waiting.
477 * Note that it is illegal to call this function while holding any lock
478 * that is acquired by a CPU-hotplug notifier. It is also illegal to call
479 * synchronize_srcu_expedited() from the corresponding SRCU read-side
480 * critical section; doing so will result in deadlock. However, it is
481 * perfectly legal to call synchronize_srcu_expedited() on one srcu_struct
482 * from some other srcu_struct's read-side critical section, as long as
483 * the resulting graph of srcu_structs is acyclic.
485 void synchronize_srcu_expedited(struct srcu_struct *sp)
487 __synchronize_srcu(sp, SYNCHRONIZE_SRCU_EXP_TRYCOUNT);
489 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
492 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
494 void srcu_barrier(struct srcu_struct *sp)
496 synchronize_srcu(sp);
498 EXPORT_SYMBOL_GPL(srcu_barrier);
501 * srcu_batches_completed - return batches completed.
502 * @sp: srcu_struct on which to report batch completion.
504 * Report the number of batches, correlated with, but not necessarily
505 * precisely the same as, the number of grace periods that have elapsed.
507 long srcu_batches_completed(struct srcu_struct *sp)
509 return sp->completed;
511 EXPORT_SYMBOL_GPL(srcu_batches_completed);
513 #define SRCU_CALLBACK_BATCH 10
514 #define SRCU_INTERVAL 1
517 * Move any new SRCU callbacks to the first stage of the SRCU grace
520 static void srcu_collect_new(struct srcu_struct *sp)
522 if (!rcu_batch_empty(&sp->batch_queue)) {
523 spin_lock_irq(&sp->queue_lock);
524 rcu_batch_move(&sp->batch_check0, &sp->batch_queue);
525 spin_unlock_irq(&sp->queue_lock);
530 * Core SRCU state machine. Advance callbacks from ->batch_check0 to
531 * ->batch_check1 and then to ->batch_done as readers drain.
533 static void srcu_advance_batches(struct srcu_struct *sp, int trycount)
535 int idx = 1 ^ (sp->completed & 1);
538 * Because readers might be delayed for an extended period after
539 * fetching ->completed for their index, at any point in time there
540 * might well be readers using both idx=0 and idx=1. We therefore
541 * need to wait for readers to clear from both index values before
542 * invoking a callback.
545 if (rcu_batch_empty(&sp->batch_check0) &&
546 rcu_batch_empty(&sp->batch_check1))
547 return; /* no callbacks need to be advanced */
549 if (!try_check_zero(sp, idx, trycount))
550 return; /* failed to advance, will try after SRCU_INTERVAL */
553 * The callbacks in ->batch_check1 have already done with their
554 * first zero check and flip back when they were enqueued on
555 * ->batch_check0 in a previous invocation of srcu_advance_batches().
556 * (Presumably try_check_zero() returned false during that
557 * invocation, leaving the callbacks stranded on ->batch_check1.)
558 * They are therefore ready to invoke, so move them to ->batch_done.
560 rcu_batch_move(&sp->batch_done, &sp->batch_check1);
562 if (rcu_batch_empty(&sp->batch_check0))
563 return; /* no callbacks need to be advanced */
567 * The callbacks in ->batch_check0 just finished their
568 * first check zero and flip, so move them to ->batch_check1
569 * for future checking on the other idx.
571 rcu_batch_move(&sp->batch_check1, &sp->batch_check0);
574 * SRCU read-side critical sections are normally short, so check
575 * at least twice in quick succession after a flip.
577 trycount = trycount < 2 ? 2 : trycount;
578 if (!try_check_zero(sp, idx^1, trycount))
579 return; /* failed to advance, will try after SRCU_INTERVAL */
582 * The callbacks in ->batch_check1 have now waited for all
583 * pre-existing readers using both idx values. They are therefore
584 * ready to invoke, so move them to ->batch_done.
586 rcu_batch_move(&sp->batch_done, &sp->batch_check1);
590 * Invoke a limited number of SRCU callbacks that have passed through
591 * their grace period. If there are more to do, SRCU will reschedule
594 static void srcu_invoke_callbacks(struct srcu_struct *sp)
597 struct rcu_head *head;
599 for (i = 0; i < SRCU_CALLBACK_BATCH; i++) {
600 head = rcu_batch_dequeue(&sp->batch_done);
610 * Finished one round of SRCU grace period. Start another if there are
611 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
613 static void srcu_reschedule(struct srcu_struct *sp)
617 if (rcu_batch_empty(&sp->batch_done) &&
618 rcu_batch_empty(&sp->batch_check1) &&
619 rcu_batch_empty(&sp->batch_check0) &&
620 rcu_batch_empty(&sp->batch_queue)) {
621 spin_lock_irq(&sp->queue_lock);
622 if (rcu_batch_empty(&sp->batch_done) &&
623 rcu_batch_empty(&sp->batch_check1) &&
624 rcu_batch_empty(&sp->batch_check0) &&
625 rcu_batch_empty(&sp->batch_queue)) {
629 spin_unlock_irq(&sp->queue_lock);
633 schedule_delayed_work(&sp->work, SRCU_INTERVAL);
637 * This is the work-queue function that handles SRCU grace periods.
639 void process_srcu(struct work_struct *work)
641 struct srcu_struct *sp;
643 sp = container_of(work, struct srcu_struct, work.work);
645 srcu_collect_new(sp);
646 srcu_advance_batches(sp, 1);
647 srcu_invoke_callbacks(sp);
650 EXPORT_SYMBOL_GPL(process_srcu);