~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/locking/rwsem-xadd.c

Version: ~ [ linux-5.8-rc4 ] ~ [ linux-5.7.7 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.50 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.131 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.187 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.229 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.229 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0
  2 /* rwsem.c: R/W semaphores: contention handling functions
  3  *
  4  * Written by David Howells (dhowells@redhat.com).
  5  * Derived from arch/i386/kernel/semaphore.c
  6  *
  7  * Writer lock-stealing by Alex Shi <alex.shi@intel.com>
  8  * and Michel Lespinasse <walken@google.com>
  9  *
 10  * Optimistic spinning by Tim Chen <tim.c.chen@intel.com>
 11  * and Davidlohr Bueso <davidlohr@hp.com>. Based on mutexes.
 12  */
 13 #include <linux/rwsem.h>
 14 #include <linux/init.h>
 15 #include <linux/export.h>
 16 #include <linux/sched/signal.h>
 17 #include <linux/sched/rt.h>
 18 #include <linux/sched/wake_q.h>
 19 #include <linux/sched/debug.h>
 20 #include <linux/osq_lock.h>
 21 
 22 #include "rwsem.h"
 23 
 24 /*
 25  * Guide to the rw_semaphore's count field for common values.
 26  * (32-bit case illustrated, similar for 64-bit)
 27  *
 28  * 0x0000000X   (1) X readers active or attempting lock, no writer waiting
 29  *                  X = #active_readers + #readers attempting to lock
 30  *                  (X*ACTIVE_BIAS)
 31  *
 32  * 0x00000000   rwsem is unlocked, and no one is waiting for the lock or
 33  *              attempting to read lock or write lock.
 34  *
 35  * 0xffff000X   (1) X readers active or attempting lock, with waiters for lock
 36  *                  X = #active readers + # readers attempting lock
 37  *                  (X*ACTIVE_BIAS + WAITING_BIAS)
 38  *              (2) 1 writer attempting lock, no waiters for lock
 39  *                  X-1 = #active readers + #readers attempting lock
 40  *                  ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
 41  *              (3) 1 writer active, no waiters for lock
 42  *                  X-1 = #active readers + #readers attempting lock
 43  *                  ((X-1)*ACTIVE_BIAS + ACTIVE_WRITE_BIAS)
 44  *
 45  * 0xffff0001   (1) 1 reader active or attempting lock, waiters for lock
 46  *                  (WAITING_BIAS + ACTIVE_BIAS)
 47  *              (2) 1 writer active or attempting lock, no waiters for lock
 48  *                  (ACTIVE_WRITE_BIAS)
 49  *
 50  * 0xffff0000   (1) There are writers or readers queued but none active
 51  *                  or in the process of attempting lock.
 52  *                  (WAITING_BIAS)
 53  *              Note: writer can attempt to steal lock for this count by adding
 54  *              ACTIVE_WRITE_BIAS in cmpxchg and checking the old count
 55  *
 56  * 0xfffe0001   (1) 1 writer active, or attempting lock. Waiters on queue.
 57  *                  (ACTIVE_WRITE_BIAS + WAITING_BIAS)
 58  *
 59  * Note: Readers attempt to lock by adding ACTIVE_BIAS in down_read and checking
 60  *       the count becomes more than 0 for successful lock acquisition,
 61  *       i.e. the case where there are only readers or nobody has lock.
 62  *       (1st and 2nd case above).
 63  *
 64  *       Writers attempt to lock by adding ACTIVE_WRITE_BIAS in down_write and
 65  *       checking the count becomes ACTIVE_WRITE_BIAS for successful lock
 66  *       acquisition (i.e. nobody else has lock or attempts lock).  If
 67  *       unsuccessful, in rwsem_down_write_failed, we'll check to see if there
 68  *       are only waiters but none active (5th case above), and attempt to
 69  *       steal the lock.
 70  *
 71  */
 72 
 73 /*
 74  * Initialize an rwsem:
 75  */
 76 void __init_rwsem(struct rw_semaphore *sem, const char *name,
 77                   struct lock_class_key *key)
 78 {
 79 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 80         /*
 81          * Make sure we are not reinitializing a held semaphore:
 82          */
 83         debug_check_no_locks_freed((void *)sem, sizeof(*sem));
 84         lockdep_init_map(&sem->dep_map, name, key, 0);
 85 #endif
 86         atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
 87         raw_spin_lock_init(&sem->wait_lock);
 88         INIT_LIST_HEAD(&sem->wait_list);
 89 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
 90         sem->owner = NULL;
 91         osq_lock_init(&sem->osq);
 92 #endif
 93 }
 94 
 95 EXPORT_SYMBOL(__init_rwsem);
 96 
 97 enum rwsem_waiter_type {
 98         RWSEM_WAITING_FOR_WRITE,
 99         RWSEM_WAITING_FOR_READ
100 };
101 
102 struct rwsem_waiter {
103         struct list_head list;
104         struct task_struct *task;
105         enum rwsem_waiter_type type;
106 };
107 
108 enum rwsem_wake_type {
109         RWSEM_WAKE_ANY,         /* Wake whatever's at head of wait list */
110         RWSEM_WAKE_READERS,     /* Wake readers only */
111         RWSEM_WAKE_READ_OWNED   /* Waker thread holds the read lock */
112 };
113 
114 /*
115  * handle the lock release when processes blocked on it that can now run
116  * - if we come here from up_xxxx(), then:
117  *   - the 'active part' of count (&0x0000ffff) reached 0 (but may have changed)
118  *   - the 'waiting part' of count (&0xffff0000) is -ve (and will still be so)
119  * - there must be someone on the queue
120  * - the wait_lock must be held by the caller
121  * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
122  *   to actually wakeup the blocked task(s) and drop the reference count,
123  *   preferably when the wait_lock is released
124  * - woken process blocks are discarded from the list after having task zeroed
125  * - writers are only marked woken if downgrading is false
126  */
127 static void __rwsem_mark_wake(struct rw_semaphore *sem,
128                               enum rwsem_wake_type wake_type,
129                               struct wake_q_head *wake_q)
130 {
131         struct rwsem_waiter *waiter, *tmp;
132         long oldcount, woken = 0, adjustment = 0;
133 
134         /*
135          * Take a peek at the queue head waiter such that we can determine
136          * the wakeup(s) to perform.
137          */
138         waiter = list_first_entry(&sem->wait_list, struct rwsem_waiter, list);
139 
140         if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
141                 if (wake_type == RWSEM_WAKE_ANY) {
142                         /*
143                          * Mark writer at the front of the queue for wakeup.
144                          * Until the task is actually later awoken later by
145                          * the caller, other writers are able to steal it.
146                          * Readers, on the other hand, will block as they
147                          * will notice the queued writer.
148                          */
149                         wake_q_add(wake_q, waiter->task);
150                 }
151 
152                 return;
153         }
154 
155         /*
156          * Writers might steal the lock before we grant it to the next reader.
157          * We prefer to do the first reader grant before counting readers
158          * so we can bail out early if a writer stole the lock.
159          */
160         if (wake_type != RWSEM_WAKE_READ_OWNED) {
161                 adjustment = RWSEM_ACTIVE_READ_BIAS;
162  try_reader_grant:
163                 oldcount = atomic_long_fetch_add(adjustment, &sem->count);
164                 if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
165                         /*
166                          * If the count is still less than RWSEM_WAITING_BIAS
167                          * after removing the adjustment, it is assumed that
168                          * a writer has stolen the lock. We have to undo our
169                          * reader grant.
170                          */
171                         if (atomic_long_add_return(-adjustment, &sem->count) <
172                             RWSEM_WAITING_BIAS)
173                                 return;
174 
175                         /* Last active locker left. Retry waking readers. */
176                         goto try_reader_grant;
177                 }
178                 /*
179                  * It is not really necessary to set it to reader-owned here,
180                  * but it gives the spinners an early indication that the
181                  * readers now have the lock.
182                  */
183                 rwsem_set_reader_owned(sem);
184         }
185 
186         /*
187          * Grant an infinite number of read locks to the readers at the front
188          * of the queue. We know that woken will be at least 1 as we accounted
189          * for above. Note we increment the 'active part' of the count by the
190          * number of readers before waking any processes up.
191          */
192         list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
193                 struct task_struct *tsk;
194 
195                 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
196                         break;
197 
198                 woken++;
199                 tsk = waiter->task;
200 
201                 wake_q_add(wake_q, tsk);
202                 list_del(&waiter->list);
203                 /*
204                  * Ensure that the last operation is setting the reader
205                  * waiter to nil such that rwsem_down_read_failed() cannot
206                  * race with do_exit() by always holding a reference count
207                  * to the task to wakeup.
208                  */
209                 smp_store_release(&waiter->task, NULL);
210         }
211 
212         adjustment = woken * RWSEM_ACTIVE_READ_BIAS - adjustment;
213         if (list_empty(&sem->wait_list)) {
214                 /* hit end of list above */
215                 adjustment -= RWSEM_WAITING_BIAS;
216         }
217 
218         if (adjustment)
219                 atomic_long_add(adjustment, &sem->count);
220 }
221 
222 /*
223  * Wait for the read lock to be granted
224  */
225 static inline struct rw_semaphore __sched *
226 __rwsem_down_read_failed_common(struct rw_semaphore *sem, int state)
227 {
228         long count, adjustment = -RWSEM_ACTIVE_READ_BIAS;
229         struct rwsem_waiter waiter;
230         DEFINE_WAKE_Q(wake_q);
231 
232         waiter.task = current;
233         waiter.type = RWSEM_WAITING_FOR_READ;
234 
235         raw_spin_lock_irq(&sem->wait_lock);
236         if (list_empty(&sem->wait_list))
237                 adjustment += RWSEM_WAITING_BIAS;
238         list_add_tail(&waiter.list, &sem->wait_list);
239 
240         /* we're now waiting on the lock, but no longer actively locking */
241         count = atomic_long_add_return(adjustment, &sem->count);
242 
243         /*
244          * If there are no active locks, wake the front queued process(es).
245          *
246          * If there are no writers and we are first in the queue,
247          * wake our own waiter to join the existing active readers !
248          */
249         if (count == RWSEM_WAITING_BIAS ||
250             (count > RWSEM_WAITING_BIAS &&
251              adjustment != -RWSEM_ACTIVE_READ_BIAS))
252                 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
253 
254         raw_spin_unlock_irq(&sem->wait_lock);
255         wake_up_q(&wake_q);
256 
257         /* wait to be given the lock */
258         while (true) {
259                 set_current_state(state);
260                 if (!waiter.task)
261                         break;
262                 if (signal_pending_state(state, current)) {
263                         raw_spin_lock_irq(&sem->wait_lock);
264                         if (waiter.task)
265                                 goto out_nolock;
266                         raw_spin_unlock_irq(&sem->wait_lock);
267                         break;
268                 }
269                 schedule();
270         }
271 
272         __set_current_state(TASK_RUNNING);
273         return sem;
274 out_nolock:
275         list_del(&waiter.list);
276         if (list_empty(&sem->wait_list))
277                 atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count);
278         raw_spin_unlock_irq(&sem->wait_lock);
279         __set_current_state(TASK_RUNNING);
280         return ERR_PTR(-EINTR);
281 }
282 
283 __visible struct rw_semaphore * __sched
284 rwsem_down_read_failed(struct rw_semaphore *sem)
285 {
286         return __rwsem_down_read_failed_common(sem, TASK_UNINTERRUPTIBLE);
287 }
288 EXPORT_SYMBOL(rwsem_down_read_failed);
289 
290 __visible struct rw_semaphore * __sched
291 rwsem_down_read_failed_killable(struct rw_semaphore *sem)
292 {
293         return __rwsem_down_read_failed_common(sem, TASK_KILLABLE);
294 }
295 EXPORT_SYMBOL(rwsem_down_read_failed_killable);
296 
297 /*
298  * This function must be called with the sem->wait_lock held to prevent
299  * race conditions between checking the rwsem wait list and setting the
300  * sem->count accordingly.
301  */
302 static inline bool rwsem_try_write_lock(long count, struct rw_semaphore *sem)
303 {
304         /*
305          * Avoid trying to acquire write lock if count isn't RWSEM_WAITING_BIAS.
306          */
307         if (count != RWSEM_WAITING_BIAS)
308                 return false;
309 
310         /*
311          * Acquire the lock by trying to set it to ACTIVE_WRITE_BIAS. If there
312          * are other tasks on the wait list, we need to add on WAITING_BIAS.
313          */
314         count = list_is_singular(&sem->wait_list) ?
315                         RWSEM_ACTIVE_WRITE_BIAS :
316                         RWSEM_ACTIVE_WRITE_BIAS + RWSEM_WAITING_BIAS;
317 
318         if (atomic_long_cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS, count)
319                                                         == RWSEM_WAITING_BIAS) {
320                 rwsem_set_owner(sem);
321                 return true;
322         }
323 
324         return false;
325 }
326 
327 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
328 /*
329  * Try to acquire write lock before the writer has been put on wait queue.
330  */
331 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
332 {
333         long old, count = atomic_long_read(&sem->count);
334 
335         while (true) {
336                 if (!(count == 0 || count == RWSEM_WAITING_BIAS))
337                         return false;
338 
339                 old = atomic_long_cmpxchg_acquire(&sem->count, count,
340                                       count + RWSEM_ACTIVE_WRITE_BIAS);
341                 if (old == count) {
342                         rwsem_set_owner(sem);
343                         return true;
344                 }
345 
346                 count = old;
347         }
348 }
349 
350 static inline bool owner_on_cpu(struct task_struct *owner)
351 {
352         /*
353          * As lock holder preemption issue, we both skip spinning if
354          * task is not on cpu or its cpu is preempted
355          */
356         return owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
357 }
358 
359 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
360 {
361         struct task_struct *owner;
362         bool ret = true;
363 
364         BUILD_BUG_ON(!rwsem_has_anonymous_owner(RWSEM_OWNER_UNKNOWN));
365 
366         if (need_resched())
367                 return false;
368 
369         rcu_read_lock();
370         owner = READ_ONCE(sem->owner);
371         if (owner) {
372                 ret = is_rwsem_owner_spinnable(owner) &&
373                       owner_on_cpu(owner);
374         }
375         rcu_read_unlock();
376         return ret;
377 }
378 
379 /*
380  * Return true only if we can still spin on the owner field of the rwsem.
381  */
382 static noinline bool rwsem_spin_on_owner(struct rw_semaphore *sem)
383 {
384         struct task_struct *owner = READ_ONCE(sem->owner);
385 
386         if (!is_rwsem_owner_spinnable(owner))
387                 return false;
388 
389         rcu_read_lock();
390         while (owner && (READ_ONCE(sem->owner) == owner)) {
391                 /*
392                  * Ensure we emit the owner->on_cpu, dereference _after_
393                  * checking sem->owner still matches owner, if that fails,
394                  * owner might point to free()d memory, if it still matches,
395                  * the rcu_read_lock() ensures the memory stays valid.
396                  */
397                 barrier();
398 
399                 /*
400                  * abort spinning when need_resched or owner is not running or
401                  * owner's cpu is preempted.
402                  */
403                 if (need_resched() || !owner_on_cpu(owner)) {
404                         rcu_read_unlock();
405                         return false;
406                 }
407 
408                 cpu_relax();
409         }
410         rcu_read_unlock();
411 
412         /*
413          * If there is a new owner or the owner is not set, we continue
414          * spinning.
415          */
416         return is_rwsem_owner_spinnable(READ_ONCE(sem->owner));
417 }
418 
419 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
420 {
421         bool taken = false;
422 
423         preempt_disable();
424 
425         /* sem->wait_lock should not be held when doing optimistic spinning */
426         if (!rwsem_can_spin_on_owner(sem))
427                 goto done;
428 
429         if (!osq_lock(&sem->osq))
430                 goto done;
431 
432         /*
433          * Optimistically spin on the owner field and attempt to acquire the
434          * lock whenever the owner changes. Spinning will be stopped when:
435          *  1) the owning writer isn't running; or
436          *  2) readers own the lock as we can't determine if they are
437          *     actively running or not.
438          */
439         while (rwsem_spin_on_owner(sem)) {
440                 /*
441                  * Try to acquire the lock
442                  */
443                 if (rwsem_try_write_lock_unqueued(sem)) {
444                         taken = true;
445                         break;
446                 }
447 
448                 /*
449                  * When there's no owner, we might have preempted between the
450                  * owner acquiring the lock and setting the owner field. If
451                  * we're an RT task that will live-lock because we won't let
452                  * the owner complete.
453                  */
454                 if (!sem->owner && (need_resched() || rt_task(current)))
455                         break;
456 
457                 /*
458                  * The cpu_relax() call is a compiler barrier which forces
459                  * everything in this loop to be re-loaded. We don't need
460                  * memory barriers as we'll eventually observe the right
461                  * values at the cost of a few extra spins.
462                  */
463                 cpu_relax();
464         }
465         osq_unlock(&sem->osq);
466 done:
467         preempt_enable();
468         return taken;
469 }
470 
471 /*
472  * Return true if the rwsem has active spinner
473  */
474 static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
475 {
476         return osq_is_locked(&sem->osq);
477 }
478 
479 #else
480 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
481 {
482         return false;
483 }
484 
485 static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
486 {
487         return false;
488 }
489 #endif
490 
491 /*
492  * Wait until we successfully acquire the write lock
493  */
494 static inline struct rw_semaphore *
495 __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state)
496 {
497         long count;
498         bool waiting = true; /* any queued threads before us */
499         struct rwsem_waiter waiter;
500         struct rw_semaphore *ret = sem;
501         DEFINE_WAKE_Q(wake_q);
502 
503         /* undo write bias from down_write operation, stop active locking */
504         count = atomic_long_sub_return(RWSEM_ACTIVE_WRITE_BIAS, &sem->count);
505 
506         /* do optimistic spinning and steal lock if possible */
507         if (rwsem_optimistic_spin(sem))
508                 return sem;
509 
510         /*
511          * Optimistic spinning failed, proceed to the slowpath
512          * and block until we can acquire the sem.
513          */
514         waiter.task = current;
515         waiter.type = RWSEM_WAITING_FOR_WRITE;
516 
517         raw_spin_lock_irq(&sem->wait_lock);
518 
519         /* account for this before adding a new element to the list */
520         if (list_empty(&sem->wait_list))
521                 waiting = false;
522 
523         list_add_tail(&waiter.list, &sem->wait_list);
524 
525         /* we're now waiting on the lock, but no longer actively locking */
526         if (waiting) {
527                 count = atomic_long_read(&sem->count);
528 
529                 /*
530                  * If there were already threads queued before us and there are
531                  * no active writers, the lock must be read owned; so we try to
532                  * wake any read locks that were queued ahead of us.
533                  */
534                 if (count > RWSEM_WAITING_BIAS) {
535                         __rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q);
536                         /*
537                          * The wakeup is normally called _after_ the wait_lock
538                          * is released, but given that we are proactively waking
539                          * readers we can deal with the wake_q overhead as it is
540                          * similar to releasing and taking the wait_lock again
541                          * for attempting rwsem_try_write_lock().
542                          */
543                         wake_up_q(&wake_q);
544 
545                         /*
546                          * Reinitialize wake_q after use.
547                          */
548                         wake_q_init(&wake_q);
549                 }
550 
551         } else
552                 count = atomic_long_add_return(RWSEM_WAITING_BIAS, &sem->count);
553 
554         /* wait until we successfully acquire the lock */
555         set_current_state(state);
556         while (true) {
557                 if (rwsem_try_write_lock(count, sem))
558                         break;
559                 raw_spin_unlock_irq(&sem->wait_lock);
560 
561                 /* Block until there are no active lockers. */
562                 do {
563                         if (signal_pending_state(state, current))
564                                 goto out_nolock;
565 
566                         schedule();
567                         set_current_state(state);
568                 } while ((count = atomic_long_read(&sem->count)) & RWSEM_ACTIVE_MASK);
569 
570                 raw_spin_lock_irq(&sem->wait_lock);
571         }
572         __set_current_state(TASK_RUNNING);
573         list_del(&waiter.list);
574         raw_spin_unlock_irq(&sem->wait_lock);
575 
576         return ret;
577 
578 out_nolock:
579         __set_current_state(TASK_RUNNING);
580         raw_spin_lock_irq(&sem->wait_lock);
581         list_del(&waiter.list);
582         if (list_empty(&sem->wait_list))
583                 atomic_long_add(-RWSEM_WAITING_BIAS, &sem->count);
584         else
585                 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
586         raw_spin_unlock_irq(&sem->wait_lock);
587         wake_up_q(&wake_q);
588 
589         return ERR_PTR(-EINTR);
590 }
591 
592 __visible struct rw_semaphore * __sched
593 rwsem_down_write_failed(struct rw_semaphore *sem)
594 {
595         return __rwsem_down_write_failed_common(sem, TASK_UNINTERRUPTIBLE);
596 }
597 EXPORT_SYMBOL(rwsem_down_write_failed);
598 
599 __visible struct rw_semaphore * __sched
600 rwsem_down_write_failed_killable(struct rw_semaphore *sem)
601 {
602         return __rwsem_down_write_failed_common(sem, TASK_KILLABLE);
603 }
604 EXPORT_SYMBOL(rwsem_down_write_failed_killable);
605 
606 /*
607  * handle waking up a waiter on the semaphore
608  * - up_read/up_write has decremented the active part of count if we come here
609  */
610 __visible
611 struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
612 {
613         unsigned long flags;
614         DEFINE_WAKE_Q(wake_q);
615 
616         /*
617         * __rwsem_down_write_failed_common(sem)
618         *   rwsem_optimistic_spin(sem)
619         *     osq_unlock(sem->osq)
620         *   ...
621         *   atomic_long_add_return(&sem->count)
622         *
623         *      - VS -
624         *
625         *              __up_write()
626         *                if (atomic_long_sub_return_release(&sem->count) < 0)
627         *                  rwsem_wake(sem)
628         *                    osq_is_locked(&sem->osq)
629         *
630         * And __up_write() must observe !osq_is_locked() when it observes the
631         * atomic_long_add_return() in order to not miss a wakeup.
632         *
633         * This boils down to:
634         *
635         * [S.rel] X = 1                [RmW] r0 = (Y += 0)
636         *         MB                         RMB
637         * [RmW]   Y += 1               [L]   r1 = X
638         *
639         * exists (r0=1 /\ r1=0)
640         */
641         smp_rmb();
642 
643         /*
644          * If a spinner is present, it is not necessary to do the wakeup.
645          * Try to do wakeup only if the trylock succeeds to minimize
646          * spinlock contention which may introduce too much delay in the
647          * unlock operation.
648          *
649          *    spinning writer           up_write/up_read caller
650          *    ---------------           -----------------------
651          * [S]   osq_unlock()           [L]   osq
652          *       MB                           RMB
653          * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock)
654          *
655          * Here, it is important to make sure that there won't be a missed
656          * wakeup while the rwsem is free and the only spinning writer goes
657          * to sleep without taking the rwsem. Even when the spinning writer
658          * is just going to break out of the waiting loop, it will still do
659          * a trylock in rwsem_down_write_failed() before sleeping. IOW, if
660          * rwsem_has_spinner() is true, it will guarantee at least one
661          * trylock attempt on the rwsem later on.
662          */
663         if (rwsem_has_spinner(sem)) {
664                 /*
665                  * The smp_rmb() here is to make sure that the spinner
666                  * state is consulted before reading the wait_lock.
667                  */
668                 smp_rmb();
669                 if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
670                         return sem;
671                 goto locked;
672         }
673         raw_spin_lock_irqsave(&sem->wait_lock, flags);
674 locked:
675 
676         if (!list_empty(&sem->wait_list))
677                 __rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
678 
679         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
680         wake_up_q(&wake_q);
681 
682         return sem;
683 }
684 EXPORT_SYMBOL(rwsem_wake);
685 
686 /*
687  * downgrade a write lock into a read lock
688  * - caller incremented waiting part of count and discovered it still negative
689  * - just wake up any readers at the front of the queue
690  */
691 __visible
692 struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
693 {
694         unsigned long flags;
695         DEFINE_WAKE_Q(wake_q);
696 
697         raw_spin_lock_irqsave(&sem->wait_lock, flags);
698 
699         if (!list_empty(&sem->wait_list))
700                 __rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
701 
702         raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
703         wake_up_q(&wake_q);
704 
705         return sem;
706 }
707 EXPORT_SYMBOL(rwsem_downgrade_wake);
708 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp