projects/linux·blob v3.8 · refs

656 lines · 19.6 KiB

1/*
2 * Sleepable Read-Copy Update mechanism for mutual exclusion.
3 *
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.
8 *
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.
13 *
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.
17 *
18 * Copyright (C) IBM Corporation, 2006
19 * Copyright (C) Fujitsu, 2012
20 *
21 * Author: Paul McKenney <[email protected]>
22 * Lai Jiangshan <[email protected]>
23 *
24 * For detailed explanation of Read-Copy Update mechanism see -
25 * Documentation/RCU/ *.txt
26 *
27 */
28
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>
38
39#include <trace/events/rcu.h>
40
41#include "rcu.h"
42
43/*
44 * Initialize an rcu_batch structure to empty.
45 */
46static inline void rcu_batch_init(struct rcu_batch *b)
47{
48 b->head = NULL;
49 b->tail = &b->head;
50}
51
52/*
53 * Enqueue a callback onto the tail of the specified rcu_batch structure.
54 */
55static inline void rcu_batch_queue(struct rcu_batch *b, struct rcu_head *head)
56{
57 *b->tail = head;
58 b->tail = &head->next;
59}
60
61/*
62 * Is the specified rcu_batch structure empty?
63 */
64static inline bool rcu_batch_empty(struct rcu_batch *b)
65{
66 return b->tail == &b->head;
67}
68
69/*
70 * Remove the callback at the head of the specified rcu_batch structure
71 * and return a pointer to it, or return NULL if the structure is empty.
72 */
73static inline struct rcu_head *rcu_batch_dequeue(struct rcu_batch *b)
74{
75 struct rcu_head *head;
76
77 if (rcu_batch_empty(b))
78 return NULL;
79
80 head = b->head;
81 b->head = head->next;
82 if (b->tail == &head->next)
83 rcu_batch_init(b);
84
85 return head;
86}
87
88/*
89 * Move all callbacks from the rcu_batch structure specified by "from" to
90 * the structure specified by "to".
91 */
92static inline void rcu_batch_move(struct rcu_batch *to, struct rcu_batch *from)
93{
94 if (!rcu_batch_empty(from)) {
95 *to->tail = from->head;
96 to->tail = from->tail;
97 rcu_batch_init(from);
98 }
99}
100
101static int init_srcu_struct_fields(struct srcu_struct *sp)
102{
103 sp->completed = 0;
104 spin_lock_init(&sp->queue_lock);
105 sp->running = false;
106 rcu_batch_init(&sp->batch_queue);
107 rcu_batch_init(&sp->batch_check0);
108 rcu_batch_init(&sp->batch_check1);
109 rcu_batch_init(&sp->batch_done);
110 INIT_DELAYED_WORK(&sp->work, process_srcu);
111 sp->per_cpu_ref = alloc_percpu(struct srcu_struct_array);
112 return sp->per_cpu_ref ? 0 : -ENOMEM;
113}
114
115#ifdef CONFIG_DEBUG_LOCK_ALLOC
116
117int __init_srcu_struct(struct srcu_struct *sp, const char *name,
118 struct lock_class_key *key)
119{
120 /* Don't re-initialize a lock while it is held. */
121 debug_check_no_locks_freed((void *)sp, sizeof(*sp));
122 lockdep_init_map(&sp->dep_map, name, key, 0);
123 return init_srcu_struct_fields(sp);
124}
125EXPORT_SYMBOL_GPL(__init_srcu_struct);
126
127#else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
128
129/**
130 * init_srcu_struct - initialize a sleep-RCU structure
131 * @sp: structure to initialize.
132 *
133 * Must invoke this on a given srcu_struct before passing that srcu_struct
134 * to any other function. Each srcu_struct represents a separate domain
135 * of SRCU protection.
136 */
137int init_srcu_struct(struct srcu_struct *sp)
138{
139 return init_srcu_struct_fields(sp);
140}
141EXPORT_SYMBOL_GPL(init_srcu_struct);
142
143#endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
144
145/*
146 * Returns approximate total of the readers' ->seq[] values for the
147 * rank of per-CPU counters specified by idx.
148 */
149static unsigned long srcu_readers_seq_idx(struct srcu_struct *sp, int idx)
150{
151 int cpu;
152 unsigned long sum = 0;
153 unsigned long t;
154
155 for_each_possible_cpu(cpu) {
156 t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->seq[idx]);
157 sum += t;
158 }
159 return sum;
160}
161
162/*
163 * Returns approximate number of readers active on the specified rank
164 * of the per-CPU ->c[] counters.
165 */
166static unsigned long srcu_readers_active_idx(struct srcu_struct *sp, int idx)
167{
168 int cpu;
169 unsigned long sum = 0;
170 unsigned long t;
171
172 for_each_possible_cpu(cpu) {
173 t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[idx]);
174 sum += t;
175 }
176 return sum;
177}
178
179/*
180 * Return true if the number of pre-existing readers is determined to
181 * be stably zero. An example unstable zero can occur if the call
182 * to srcu_readers_active_idx() misses an __srcu_read_lock() increment,
183 * but due to task migration, sees the corresponding __srcu_read_unlock()
184 * decrement. This can happen because srcu_readers_active_idx() takes
185 * time to sum the array, and might in fact be interrupted or preempted
186 * partway through the summation.
187 */
188static bool srcu_readers_active_idx_check(struct srcu_struct *sp, int idx)
189{
190 unsigned long seq;
191
192 seq = srcu_readers_seq_idx(sp, idx);
193
194 /*
195 * The following smp_mb() A pairs with the smp_mb() B located in
196 * __srcu_read_lock(). This pairing ensures that if an
197 * __srcu_read_lock() increments its counter after the summation
198 * in srcu_readers_active_idx(), then the corresponding SRCU read-side
199 * critical section will see any changes made prior to the start
200 * of the current SRCU grace period.
201 *
202 * Also, if the above call to srcu_readers_seq_idx() saw the
203 * increment of ->seq[], then the call to srcu_readers_active_idx()
204 * must see the increment of ->c[].
205 */
206 smp_mb(); /* A */
207
208 /*
209 * Note that srcu_readers_active_idx() can incorrectly return
210 * zero even though there is a pre-existing reader throughout.
211 * To see this, suppose that task A is in a very long SRCU
212 * read-side critical section that started on CPU 0, and that
213 * no other reader exists, so that the sum of the counters
214 * is equal to one. Then suppose that task B starts executing
215 * srcu_readers_active_idx(), summing up to CPU 1, and then that
216 * task C starts reading on CPU 0, so that its increment is not
217 * summed, but finishes reading on CPU 2, so that its decrement
218 * -is- summed. Then when task B completes its sum, it will
219 * incorrectly get zero, despite the fact that task A has been
220 * in its SRCU read-side critical section the whole time.
221 *
222 * We therefore do a validation step should srcu_readers_active_idx()
223 * return zero.
224 */
225 if (srcu_readers_active_idx(sp, idx) != 0)
226 return false;
227
228 /*
229 * The remainder of this function is the validation step.
230 * The following smp_mb() D pairs with the smp_mb() C in
231 * __srcu_read_unlock(). If the __srcu_read_unlock() was seen
232 * by srcu_readers_active_idx() above, then any destructive
233 * operation performed after the grace period will happen after
234 * the corresponding SRCU read-side critical section.
235 *
236 * Note that there can be at most NR_CPUS worth of readers using
237 * the old index, which is not enough to overflow even a 32-bit
238 * integer. (Yes, this does mean that systems having more than
239 * a billion or so CPUs need to be 64-bit systems.) Therefore,
240 * the sum of the ->seq[] counters cannot possibly overflow.
241 * Therefore, the only way that the return values of the two
242 * calls to srcu_readers_seq_idx() can be equal is if there were
243 * no increments of the corresponding rank of ->seq[] counts
244 * in the interim. But the missed-increment scenario laid out
245 * above includes an increment of the ->seq[] counter by
246 * the corresponding __srcu_read_lock(). Therefore, if this
247 * scenario occurs, the return values from the two calls to
248 * srcu_readers_seq_idx() will differ, and thus the validation
249 * step below suffices.
250 */
251 smp_mb(); /* D */
252
253 return srcu_readers_seq_idx(sp, idx) == seq;
254}
255
256/**
257 * srcu_readers_active - returns approximate number of readers.
258 * @sp: which srcu_struct to count active readers (holding srcu_read_lock).
259 *
260 * Note that this is not an atomic primitive, and can therefore suffer
261 * severe errors when invoked on an active srcu_struct. That said, it
262 * can be useful as an error check at cleanup time.
263 */
264static int srcu_readers_active(struct srcu_struct *sp)
265{
266 int cpu;
267 unsigned long sum = 0;
268
269 for_each_possible_cpu(cpu) {
270 sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[0]);
271 sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[1]);
272 }
273 return sum;
274}
275
276/**
277 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
278 * @sp: structure to clean up.
279 *
280 * Must invoke this after you are finished using a given srcu_struct that
281 * was initialized via init_srcu_struct(), else you leak memory.
282 */
283void cleanup_srcu_struct(struct srcu_struct *sp)
284{
285 int sum;
286
287 sum = srcu_readers_active(sp);
288 WARN_ON(sum); /* Leakage unless caller handles error. */
289 if (sum != 0)
290 return;
291 free_percpu(sp->per_cpu_ref);
292 sp->per_cpu_ref = NULL;
293}
294EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
295
296/*
297 * Counts the new reader in the appropriate per-CPU element of the
298 * srcu_struct. Must be called from process context.
299 * Returns an index that must be passed to the matching srcu_read_unlock().
300 */
301int __srcu_read_lock(struct srcu_struct *sp)
302{
303 int idx;
304
305 preempt_disable();
306 idx = rcu_dereference_index_check(sp->completed,
307 rcu_read_lock_sched_held()) & 0x1;
308 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) += 1;
309 smp_mb(); /* B */ /* Avoid leaking the critical section. */
310 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->seq[idx]) += 1;
311 preempt_enable();
312 return idx;
313}
314EXPORT_SYMBOL_GPL(__srcu_read_lock);
315
316/*
317 * Removes the count for the old reader from the appropriate per-CPU
318 * element of the srcu_struct. Note that this may well be a different
319 * CPU than that which was incremented by the corresponding srcu_read_lock().
320 * Must be called from process context.
321 */
322void __srcu_read_unlock(struct srcu_struct *sp, int idx)
323{
324 preempt_disable();
325 smp_mb(); /* C */ /* Avoid leaking the critical section. */
326 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) -= 1;
327 preempt_enable();
328}
329EXPORT_SYMBOL_GPL(__srcu_read_unlock);
330
331/*
332 * We use an adaptive strategy for synchronize_srcu() and especially for
333 * synchronize_srcu_expedited(). We spin for a fixed time period
334 * (defined below) to allow SRCU readers to exit their read-side critical
335 * sections. If there are still some readers after 10 microseconds,
336 * we repeatedly block for 1-millisecond time periods. This approach
337 * has done well in testing, so there is no need for a config parameter.
338 */
339#define SRCU_RETRY_CHECK_DELAY 5
340#define SYNCHRONIZE_SRCU_TRYCOUNT 2
341#define SYNCHRONIZE_SRCU_EXP_TRYCOUNT 12
342
343/*
344 * @@@ Wait until all pre-existing readers complete. Such readers
345 * will have used the index specified by "idx".
346 * the caller should ensures the ->completed is not changed while checking
347 * and idx = (->completed & 1) ^ 1
348 */
349static bool try_check_zero(struct srcu_struct *sp, int idx, int trycount)
350{
351 for (;;) {
352 if (srcu_readers_active_idx_check(sp, idx))
353 return true;
354 if (--trycount <= 0)
355 return false;
356 udelay(SRCU_RETRY_CHECK_DELAY);
357 }
358}
359
360/*
361 * Increment the ->completed counter so that future SRCU readers will
362 * use the other rank of the ->c[] and ->seq[] arrays. This allows
363 * us to wait for pre-existing readers in a starvation-free manner.
364 */
365static void srcu_flip(struct srcu_struct *sp)
366{
367 sp->completed++;
368}
369
370/*
371 * Enqueue an SRCU callback on the specified srcu_struct structure,
372 * initiating grace-period processing if it is not already running.
373 */
374void call_srcu(struct srcu_struct *sp, struct rcu_head *head,
375 void (*func)(struct rcu_head *head))
376{
377 unsigned long flags;
378
379 head->next = NULL;
380 head->func = func;
381 spin_lock_irqsave(&sp->queue_lock, flags);
382 rcu_batch_queue(&sp->batch_queue, head);
383 if (!sp->running) {
384 sp->running = true;
385 schedule_delayed_work(&sp->work, 0);
386 }
387 spin_unlock_irqrestore(&sp->queue_lock, flags);
388}
389EXPORT_SYMBOL_GPL(call_srcu);
390
391struct rcu_synchronize {
392 struct rcu_head head;
393 struct completion completion;
394};
395
396/*
397 * Awaken the corresponding synchronize_srcu() instance now that a
398 * grace period has elapsed.
399 */
400static void wakeme_after_rcu(struct rcu_head *head)
401{
402 struct rcu_synchronize *rcu;
403
404 rcu = container_of(head, struct rcu_synchronize, head);
405 complete(&rcu->completion);
406}
407
408static void srcu_advance_batches(struct srcu_struct *sp, int trycount);
409static void srcu_reschedule(struct srcu_struct *sp);
410
411/*
412 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
413 */
414static void __synchronize_srcu(struct srcu_struct *sp, int trycount)
415{
416 struct rcu_synchronize rcu;
417 struct rcu_head *head = &rcu.head;
418 bool done = false;
419
420 rcu_lockdep_assert(!lock_is_held(&sp->dep_map) &&
421 !lock_is_held(&rcu_bh_lock_map) &&
422 !lock_is_held(&rcu_lock_map) &&
423 !lock_is_held(&rcu_sched_lock_map),
424 "Illegal synchronize_srcu() in same-type SRCU (or RCU) read-side critical section");
425
426 init_completion(&rcu.completion);
427
428 head->next = NULL;
429 head->func = wakeme_after_rcu;
430 spin_lock_irq(&sp->queue_lock);
431 if (!sp->running) {
432 /* steal the processing owner */
433 sp->running = true;
434 rcu_batch_queue(&sp->batch_check0, head);
435 spin_unlock_irq(&sp->queue_lock);
436
437 srcu_advance_batches(sp, trycount);
438 if (!rcu_batch_empty(&sp->batch_done)) {
439 BUG_ON(sp->batch_done.head != head);
440 rcu_batch_dequeue(&sp->batch_done);
441 done = true;
442 }
443 /* give the processing owner to work_struct */
444 srcu_reschedule(sp);
445 } else {
446 rcu_batch_queue(&sp->batch_queue, head);
447 spin_unlock_irq(&sp->queue_lock);
448 }
449
450 if (!done)
451 wait_for_completion(&rcu.completion);
452}
453
454/**
455 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
456 * @sp: srcu_struct with which to synchronize.
457 *
458 * Flip the completed counter, and wait for the old count to drain to zero.
459 * As with classic RCU, the updater must use some separate means of
460 * synchronizing concurrent updates. Can block; must be called from
461 * process context.
462 *
463 * Note that it is illegal to call synchronize_srcu() from the corresponding
464 * SRCU read-side critical section; doing so will result in deadlock.
465 * However, it is perfectly legal to call synchronize_srcu() on one
466 * srcu_struct from some other srcu_struct's read-side critical section.
467 */
468void synchronize_srcu(struct srcu_struct *sp)
469{
470 __synchronize_srcu(sp, rcu_expedited
471 ? SYNCHRONIZE_SRCU_EXP_TRYCOUNT
472 : SYNCHRONIZE_SRCU_TRYCOUNT);
473}
474EXPORT_SYMBOL_GPL(synchronize_srcu);
475
476/**
477 * synchronize_srcu_expedited - Brute-force SRCU grace period
478 * @sp: srcu_struct with which to synchronize.
479 *
480 * Wait for an SRCU grace period to elapse, but be more aggressive about
481 * spinning rather than blocking when waiting.
482 *
483 * Note that it is illegal to call this function while holding any lock
484 * that is acquired by a CPU-hotplug notifier. It is also illegal to call
485 * synchronize_srcu_expedited() from the corresponding SRCU read-side
486 * critical section; doing so will result in deadlock. However, it is
487 * perfectly legal to call synchronize_srcu_expedited() on one srcu_struct
488 * from some other srcu_struct's read-side critical section, as long as
489 * the resulting graph of srcu_structs is acyclic.
490 */
491void synchronize_srcu_expedited(struct srcu_struct *sp)
492{
493 __synchronize_srcu(sp, SYNCHRONIZE_SRCU_EXP_TRYCOUNT);
494}
495EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
496
497/**
498 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
499 */
500void srcu_barrier(struct srcu_struct *sp)
501{
502 synchronize_srcu(sp);
503}
504EXPORT_SYMBOL_GPL(srcu_barrier);
505
506/**
507 * srcu_batches_completed - return batches completed.
508 * @sp: srcu_struct on which to report batch completion.
509 *
510 * Report the number of batches, correlated with, but not necessarily
511 * precisely the same as, the number of grace periods that have elapsed.
512 */
513long srcu_batches_completed(struct srcu_struct *sp)
514{
515 return sp->completed;
516}
517EXPORT_SYMBOL_GPL(srcu_batches_completed);
518
519#define SRCU_CALLBACK_BATCH 10
520#define SRCU_INTERVAL 1
521
522/*
523 * Move any new SRCU callbacks to the first stage of the SRCU grace
524 * period pipeline.
525 */
526static void srcu_collect_new(struct srcu_struct *sp)
527{
528 if (!rcu_batch_empty(&sp->batch_queue)) {
529 spin_lock_irq(&sp->queue_lock);
530 rcu_batch_move(&sp->batch_check0, &sp->batch_queue);
531 spin_unlock_irq(&sp->queue_lock);
532 }
533}
534
535/*
536 * Core SRCU state machine. Advance callbacks from ->batch_check0 to
537 * ->batch_check1 and then to ->batch_done as readers drain.
538 */
539static void srcu_advance_batches(struct srcu_struct *sp, int trycount)
540{
541 int idx = 1 ^ (sp->completed & 1);
542
543 /*
544 * Because readers might be delayed for an extended period after
545 * fetching ->completed for their index, at any point in time there
546 * might well be readers using both idx=0 and idx=1. We therefore
547 * need to wait for readers to clear from both index values before
548 * invoking a callback.
549 */
550
551 if (rcu_batch_empty(&sp->batch_check0) &&
552 rcu_batch_empty(&sp->batch_check1))
553 return; /* no callbacks need to be advanced */
554
555 if (!try_check_zero(sp, idx, trycount))
556 return; /* failed to advance, will try after SRCU_INTERVAL */
557
558 /*
559 * The callbacks in ->batch_check1 have already done with their
560 * first zero check and flip back when they were enqueued on
561 * ->batch_check0 in a previous invocation of srcu_advance_batches().
562 * (Presumably try_check_zero() returned false during that
563 * invocation, leaving the callbacks stranded on ->batch_check1.)
564 * They are therefore ready to invoke, so move them to ->batch_done.
565 */
566 rcu_batch_move(&sp->batch_done, &sp->batch_check1);
567
568 if (rcu_batch_empty(&sp->batch_check0))
569 return; /* no callbacks need to be advanced */
570 srcu_flip(sp);
571
572 /*
573 * The callbacks in ->batch_check0 just finished their
574 * first check zero and flip, so move them to ->batch_check1
575 * for future checking on the other idx.
576 */
577 rcu_batch_move(&sp->batch_check1, &sp->batch_check0);
578
579 /*
580 * SRCU read-side critical sections are normally short, so check
581 * at least twice in quick succession after a flip.
582 */
583 trycount = trycount < 2 ? 2 : trycount;
584 if (!try_check_zero(sp, idx^1, trycount))
585 return; /* failed to advance, will try after SRCU_INTERVAL */
586
587 /*
588 * The callbacks in ->batch_check1 have now waited for all
589 * pre-existing readers using both idx values. They are therefore
590 * ready to invoke, so move them to ->batch_done.
591 */
592 rcu_batch_move(&sp->batch_done, &sp->batch_check1);
593}
594
595/*
596 * Invoke a limited number of SRCU callbacks that have passed through
597 * their grace period. If there are more to do, SRCU will reschedule
598 * the workqueue.
599 */
600static void srcu_invoke_callbacks(struct srcu_struct *sp)
601{
602 int i;
603 struct rcu_head *head;
604
605 for (i = 0; i < SRCU_CALLBACK_BATCH; i++) {
606 head = rcu_batch_dequeue(&sp->batch_done);
607 if (!head)
608 break;
609 local_bh_disable();
610 head->func(head);
611 local_bh_enable();
612 }
613}
614
615/*
616 * Finished one round of SRCU grace period. Start another if there are
617 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
618 */
619static void srcu_reschedule(struct srcu_struct *sp)
620{
621 bool pending = true;
622
623 if (rcu_batch_empty(&sp->batch_done) &&
624 rcu_batch_empty(&sp->batch_check1) &&
625 rcu_batch_empty(&sp->batch_check0) &&
626 rcu_batch_empty(&sp->batch_queue)) {
627 spin_lock_irq(&sp->queue_lock);
628 if (rcu_batch_empty(&sp->batch_done) &&
629 rcu_batch_empty(&sp->batch_check1) &&
630 rcu_batch_empty(&sp->batch_check0) &&
631 rcu_batch_empty(&sp->batch_queue)) {
632 sp->running = false;
633 pending = false;
634 }
635 spin_unlock_irq(&sp->queue_lock);
636 }
637
638 if (pending)
639 schedule_delayed_work(&sp->work, SRCU_INTERVAL);
640}
641
642/*
643 * This is the work-queue function that handles SRCU grace periods.
644 */
645void process_srcu(struct work_struct *work)
646{
647 struct srcu_struct *sp;
648
649 sp = container_of(work, struct srcu_struct, work.work);
650
651 srcu_collect_new(sp);
652 srcu_advance_batches(sp, 1);
653 srcu_invoke_callbacks(sp);
654 srcu_reschedule(sp);
655}
656EXPORT_SYMBOL_GPL(process_srcu);