projects/linux·blob v5.16 · refs

1660 lines · 44.3 KiB

1// SPDX-License-Identifier: GPL-2.0
2/* kernel/rwsem.c: R/W semaphores, public implementation
3 *
4 * Written by David Howells ([email protected]).
5 * Derived from asm-i386/semaphore.h
6 *
7 * Writer lock-stealing by Alex Shi <[email protected]>
8 * and Michel Lespinasse <[email protected]>
9 *
10 * Optimistic spinning by Tim Chen <[email protected]>
11 * and Davidlohr Bueso <[email protected]>. Based on mutexes.
12 *
13 * Rwsem count bit fields re-definition and rwsem rearchitecture by
14 * Waiman Long <[email protected]> and
15 * Peter Zijlstra <[email protected]>.
16 */
17
18#include <linux/types.h>
19#include <linux/kernel.h>
20#include <linux/sched.h>
21#include <linux/sched/rt.h>
22#include <linux/sched/task.h>
23#include <linux/sched/debug.h>
24#include <linux/sched/wake_q.h>
25#include <linux/sched/signal.h>
26#include <linux/sched/clock.h>
27#include <linux/export.h>
28#include <linux/rwsem.h>
29#include <linux/atomic.h>
30
31#ifndef CONFIG_PREEMPT_RT
32#include "lock_events.h"
33
34/*
35 * The least significant 2 bits of the owner value has the following
36 * meanings when set.
37 * - Bit 0: RWSEM_READER_OWNED - The rwsem is owned by readers
38 * - Bit 1: RWSEM_NONSPINNABLE - Cannot spin on a reader-owned lock
39 *
40 * When the rwsem is reader-owned and a spinning writer has timed out,
41 * the nonspinnable bit will be set to disable optimistic spinning.
42
43 * When a writer acquires a rwsem, it puts its task_struct pointer
44 * into the owner field. It is cleared after an unlock.
45 *
46 * When a reader acquires a rwsem, it will also puts its task_struct
47 * pointer into the owner field with the RWSEM_READER_OWNED bit set.
48 * On unlock, the owner field will largely be left untouched. So
49 * for a free or reader-owned rwsem, the owner value may contain
50 * information about the last reader that acquires the rwsem.
51 *
52 * That information may be helpful in debugging cases where the system
53 * seems to hang on a reader owned rwsem especially if only one reader
54 * is involved. Ideally we would like to track all the readers that own
55 * a rwsem, but the overhead is simply too big.
56 *
57 * A fast path reader optimistic lock stealing is supported when the rwsem
58 * is previously owned by a writer and the following conditions are met:
59 * - rwsem is not currently writer owned
60 * - the handoff isn't set.
61 */
62#define RWSEM_READER_OWNED (1UL << 0)
63#define RWSEM_NONSPINNABLE (1UL << 1)
64#define RWSEM_OWNER_FLAGS_MASK (RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
65
66#ifdef CONFIG_DEBUG_RWSEMS
67# define DEBUG_RWSEMS_WARN_ON(c, sem) do { \
68 if (!debug_locks_silent && \
69 WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
70 #c, atomic_long_read(&(sem)->count), \
71 (unsigned long) sem->magic, \
72 atomic_long_read(&(sem)->owner), (long)current, \
73 list_empty(&(sem)->wait_list) ? "" : "not ")) \
74 debug_locks_off(); \
75 } while (0)
76#else
77# define DEBUG_RWSEMS_WARN_ON(c, sem)
78#endif
79
80/*
81 * On 64-bit architectures, the bit definitions of the count are:
82 *
83 * Bit 0 - writer locked bit
84 * Bit 1 - waiters present bit
85 * Bit 2 - lock handoff bit
86 * Bits 3-7 - reserved
87 * Bits 8-62 - 55-bit reader count
88 * Bit 63 - read fail bit
89 *
90 * On 32-bit architectures, the bit definitions of the count are:
91 *
92 * Bit 0 - writer locked bit
93 * Bit 1 - waiters present bit
94 * Bit 2 - lock handoff bit
95 * Bits 3-7 - reserved
96 * Bits 8-30 - 23-bit reader count
97 * Bit 31 - read fail bit
98 *
99 * It is not likely that the most significant bit (read fail bit) will ever
100 * be set. This guard bit is still checked anyway in the down_read() fastpath
101 * just in case we need to use up more of the reader bits for other purpose
102 * in the future.
103 *
104 * atomic_long_fetch_add() is used to obtain reader lock, whereas
105 * atomic_long_cmpxchg() will be used to obtain writer lock.
106 *
107 * There are three places where the lock handoff bit may be set or cleared.
108 * 1) rwsem_mark_wake() for readers -- set, clear
109 * 2) rwsem_try_write_lock() for writers -- set, clear
110 * 3) rwsem_del_waiter() -- clear
111 *
112 * For all the above cases, wait_lock will be held. A writer must also
113 * be the first one in the wait_list to be eligible for setting the handoff
114 * bit. So concurrent setting/clearing of handoff bit is not possible.
115 */
116#define RWSEM_WRITER_LOCKED (1UL << 0)
117#define RWSEM_FLAG_WAITERS (1UL << 1)
118#define RWSEM_FLAG_HANDOFF (1UL << 2)
119#define RWSEM_FLAG_READFAIL (1UL << (BITS_PER_LONG - 1))
120
121#define RWSEM_READER_SHIFT 8
122#define RWSEM_READER_BIAS (1UL << RWSEM_READER_SHIFT)
123#define RWSEM_READER_MASK (~(RWSEM_READER_BIAS - 1))
124#define RWSEM_WRITER_MASK RWSEM_WRITER_LOCKED
125#define RWSEM_LOCK_MASK (RWSEM_WRITER_MASK|RWSEM_READER_MASK)
126#define RWSEM_READ_FAILED_MASK (RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
127 RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
128
129/*
130 * All writes to owner are protected by WRITE_ONCE() to make sure that
131 * store tearing can't happen as optimistic spinners may read and use
132 * the owner value concurrently without lock. Read from owner, however,
133 * may not need READ_ONCE() as long as the pointer value is only used
134 * for comparison and isn't being dereferenced.
135 */
136static inline void rwsem_set_owner(struct rw_semaphore *sem)
137{
138 atomic_long_set(&sem->owner, (long)current);
139}
140
141static inline void rwsem_clear_owner(struct rw_semaphore *sem)
142{
143 atomic_long_set(&sem->owner, 0);
144}
145
146/*
147 * Test the flags in the owner field.
148 */
149static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
150{
151 return atomic_long_read(&sem->owner) & flags;
152}
153
154/*
155 * The task_struct pointer of the last owning reader will be left in
156 * the owner field.
157 *
158 * Note that the owner value just indicates the task has owned the rwsem
159 * previously, it may not be the real owner or one of the real owners
160 * anymore when that field is examined, so take it with a grain of salt.
161 *
162 * The reader non-spinnable bit is preserved.
163 */
164static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
165 struct task_struct *owner)
166{
167 unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
168 (atomic_long_read(&sem->owner) & RWSEM_NONSPINNABLE);
169
170 atomic_long_set(&sem->owner, val);
171}
172
173static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
174{
175 __rwsem_set_reader_owned(sem, current);
176}
177
178/*
179 * Return true if the rwsem is owned by a reader.
180 */
181static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
182{
183#ifdef CONFIG_DEBUG_RWSEMS
184 /*
185 * Check the count to see if it is write-locked.
186 */
187 long count = atomic_long_read(&sem->count);
188
189 if (count & RWSEM_WRITER_MASK)
190 return false;
191#endif
192 return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
193}
194
195#ifdef CONFIG_DEBUG_RWSEMS
196/*
197 * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
198 * is a task pointer in owner of a reader-owned rwsem, it will be the
199 * real owner or one of the real owners. The only exception is when the
200 * unlock is done by up_read_non_owner().
201 */
202static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
203{
204 unsigned long val = atomic_long_read(&sem->owner);
205
206 while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
207 if (atomic_long_try_cmpxchg(&sem->owner, &val,
208 val & RWSEM_OWNER_FLAGS_MASK))
209 return;
210 }
211}
212#else
213static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
214{
215}
216#endif
217
218/*
219 * Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
220 * remains set. Otherwise, the operation will be aborted.
221 */
222static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
223{
224 unsigned long owner = atomic_long_read(&sem->owner);
225
226 do {
227 if (!(owner & RWSEM_READER_OWNED))
228 break;
229 if (owner & RWSEM_NONSPINNABLE)
230 break;
231 } while (!atomic_long_try_cmpxchg(&sem->owner, &owner,
232 owner | RWSEM_NONSPINNABLE));
233}
234
235static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cntp)
236{
237 *cntp = atomic_long_add_return_acquire(RWSEM_READER_BIAS, &sem->count);
238
239 if (WARN_ON_ONCE(*cntp < 0))
240 rwsem_set_nonspinnable(sem);
241
242 if (!(*cntp & RWSEM_READ_FAILED_MASK)) {
243 rwsem_set_reader_owned(sem);
244 return true;
245 }
246
247 return false;
248}
249
250static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
251{
252 long tmp = RWSEM_UNLOCKED_VALUE;
253
254 if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp, RWSEM_WRITER_LOCKED)) {
255 rwsem_set_owner(sem);
256 return true;
257 }
258
259 return false;
260}
261
262/*
263 * Return just the real task structure pointer of the owner
264 */
265static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
266{
267 return (struct task_struct *)
268 (atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
269}
270
271/*
272 * Return the real task structure pointer of the owner and the embedded
273 * flags in the owner. pflags must be non-NULL.
274 */
275static inline struct task_struct *
276rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
277{
278 unsigned long owner = atomic_long_read(&sem->owner);
279
280 *pflags = owner & RWSEM_OWNER_FLAGS_MASK;
281 return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
282}
283
284/*
285 * Guide to the rw_semaphore's count field.
286 *
287 * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
288 * by a writer.
289 *
290 * The lock is owned by readers when
291 * (1) the RWSEM_WRITER_LOCKED isn't set in count,
292 * (2) some of the reader bits are set in count, and
293 * (3) the owner field has RWSEM_READ_OWNED bit set.
294 *
295 * Having some reader bits set is not enough to guarantee a readers owned
296 * lock as the readers may be in the process of backing out from the count
297 * and a writer has just released the lock. So another writer may steal
298 * the lock immediately after that.
299 */
300
301/*
302 * Initialize an rwsem:
303 */
304void __init_rwsem(struct rw_semaphore *sem, const char *name,
305 struct lock_class_key *key)
306{
307#ifdef CONFIG_DEBUG_LOCK_ALLOC
308 /*
309 * Make sure we are not reinitializing a held semaphore:
310 */
311 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
312 lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
313#endif
314#ifdef CONFIG_DEBUG_RWSEMS
315 sem->magic = sem;
316#endif
317 atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
318 raw_spin_lock_init(&sem->wait_lock);
319 INIT_LIST_HEAD(&sem->wait_list);
320 atomic_long_set(&sem->owner, 0L);
321#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
322 osq_lock_init(&sem->osq);
323#endif
324}
325EXPORT_SYMBOL(__init_rwsem);
326
327enum rwsem_waiter_type {
328 RWSEM_WAITING_FOR_WRITE,
329 RWSEM_WAITING_FOR_READ
330};
331
332struct rwsem_waiter {
333 struct list_head list;
334 struct task_struct *task;
335 enum rwsem_waiter_type type;
336 unsigned long timeout;
337
338 /* Writer only, not initialized in reader */
339 bool handoff_set;
340};
341#define rwsem_first_waiter(sem) \
342 list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
343
344enum rwsem_wake_type {
345 RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
346 RWSEM_WAKE_READERS, /* Wake readers only */
347 RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
348};
349
350/*
351 * The typical HZ value is either 250 or 1000. So set the minimum waiting
352 * time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
353 * queue before initiating the handoff protocol.
354 */
355#define RWSEM_WAIT_TIMEOUT DIV_ROUND_UP(HZ, 250)
356
357/*
358 * Magic number to batch-wakeup waiting readers, even when writers are
359 * also present in the queue. This both limits the amount of work the
360 * waking thread must do and also prevents any potential counter overflow,
361 * however unlikely.
362 */
363#define MAX_READERS_WAKEUP 0x100
364
365static inline void
366rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
367{
368 lockdep_assert_held(&sem->wait_lock);
369 list_add_tail(&waiter->list, &sem->wait_list);
370 /* caller will set RWSEM_FLAG_WAITERS */
371}
372
373/*
374 * Remove a waiter from the wait_list and clear flags.
375 *
376 * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
377 * this function. Modify with care.
378 */
379static inline void
380rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
381{
382 lockdep_assert_held(&sem->wait_lock);
383 list_del(&waiter->list);
384 if (likely(!list_empty(&sem->wait_list)))
385 return;
386
387 atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, &sem->count);
388}
389
390/*
391 * handle the lock release when processes blocked on it that can now run
392 * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
393 * have been set.
394 * - there must be someone on the queue
395 * - the wait_lock must be held by the caller
396 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
397 * to actually wakeup the blocked task(s) and drop the reference count,
398 * preferably when the wait_lock is released
399 * - woken process blocks are discarded from the list after having task zeroed
400 * - writers are only marked woken if downgrading is false
401 *
402 * Implies rwsem_del_waiter() for all woken readers.
403 */
404static void rwsem_mark_wake(struct rw_semaphore *sem,
405 enum rwsem_wake_type wake_type,
406 struct wake_q_head *wake_q)
407{
408 struct rwsem_waiter *waiter, *tmp;
409 long oldcount, woken = 0, adjustment = 0;
410 struct list_head wlist;
411
412 lockdep_assert_held(&sem->wait_lock);
413
414 /*
415 * Take a peek at the queue head waiter such that we can determine
416 * the wakeup(s) to perform.
417 */
418 waiter = rwsem_first_waiter(sem);
419
420 if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
421 if (wake_type == RWSEM_WAKE_ANY) {
422 /*
423 * Mark writer at the front of the queue for wakeup.
424 * Until the task is actually later awoken later by
425 * the caller, other writers are able to steal it.
426 * Readers, on the other hand, will block as they
427 * will notice the queued writer.
428 */
429 wake_q_add(wake_q, waiter->task);
430 lockevent_inc(rwsem_wake_writer);
431 }
432
433 return;
434 }
435
436 /*
437 * No reader wakeup if there are too many of them already.
438 */
439 if (unlikely(atomic_long_read(&sem->count) < 0))
440 return;
441
442 /*
443 * Writers might steal the lock before we grant it to the next reader.
444 * We prefer to do the first reader grant before counting readers
445 * so we can bail out early if a writer stole the lock.
446 */
447 if (wake_type != RWSEM_WAKE_READ_OWNED) {
448 struct task_struct *owner;
449
450 adjustment = RWSEM_READER_BIAS;
451 oldcount = atomic_long_fetch_add(adjustment, &sem->count);
452 if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
453 /*
454 * When we've been waiting "too" long (for writers
455 * to give up the lock), request a HANDOFF to
456 * force the issue.
457 */
458 if (!(oldcount & RWSEM_FLAG_HANDOFF) &&
459 time_after(jiffies, waiter->timeout)) {
460 adjustment -= RWSEM_FLAG_HANDOFF;
461 lockevent_inc(rwsem_rlock_handoff);
462 }
463
464 atomic_long_add(-adjustment, &sem->count);
465 return;
466 }
467 /*
468 * Set it to reader-owned to give spinners an early
469 * indication that readers now have the lock.
470 * The reader nonspinnable bit seen at slowpath entry of
471 * the reader is copied over.
472 */
473 owner = waiter->task;
474 __rwsem_set_reader_owned(sem, owner);
475 }
476
477 /*
478 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
479 * queue. We know that the woken will be at least 1 as we accounted
480 * for above. Note we increment the 'active part' of the count by the
481 * number of readers before waking any processes up.
482 *
483 * This is an adaptation of the phase-fair R/W locks where at the
484 * reader phase (first waiter is a reader), all readers are eligible
485 * to acquire the lock at the same time irrespective of their order
486 * in the queue. The writers acquire the lock according to their
487 * order in the queue.
488 *
489 * We have to do wakeup in 2 passes to prevent the possibility that
490 * the reader count may be decremented before it is incremented. It
491 * is because the to-be-woken waiter may not have slept yet. So it
492 * may see waiter->task got cleared, finish its critical section and
493 * do an unlock before the reader count increment.
494 *
495 * 1) Collect the read-waiters in a separate list, count them and
496 * fully increment the reader count in rwsem.
497 * 2) For each waiters in the new list, clear waiter->task and
498 * put them into wake_q to be woken up later.
499 */
500 INIT_LIST_HEAD(&wlist);
501 list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
502 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
503 continue;
504
505 woken++;
506 list_move_tail(&waiter->list, &wlist);
507
508 /*
509 * Limit # of readers that can be woken up per wakeup call.
510 */
511 if (unlikely(woken >= MAX_READERS_WAKEUP))
512 break;
513 }
514
515 adjustment = woken * RWSEM_READER_BIAS - adjustment;
516 lockevent_cond_inc(rwsem_wake_reader, woken);
517
518 oldcount = atomic_long_read(&sem->count);
519 if (list_empty(&sem->wait_list)) {
520 /*
521 * Combined with list_move_tail() above, this implies
522 * rwsem_del_waiter().
523 */
524 adjustment -= RWSEM_FLAG_WAITERS;
525 if (oldcount & RWSEM_FLAG_HANDOFF)
526 adjustment -= RWSEM_FLAG_HANDOFF;
527 } else if (woken) {
528 /*
529 * When we've woken a reader, we no longer need to force
530 * writers to give up the lock and we can clear HANDOFF.
531 */
532 if (oldcount & RWSEM_FLAG_HANDOFF)
533 adjustment -= RWSEM_FLAG_HANDOFF;
534 }
535
536 if (adjustment)
537 atomic_long_add(adjustment, &sem->count);
538
539 /* 2nd pass */
540 list_for_each_entry_safe(waiter, tmp, &wlist, list) {
541 struct task_struct *tsk;
542
543 tsk = waiter->task;
544 get_task_struct(tsk);
545
546 /*
547 * Ensure calling get_task_struct() before setting the reader
548 * waiter to nil such that rwsem_down_read_slowpath() cannot
549 * race with do_exit() by always holding a reference count
550 * to the task to wakeup.
551 */
552 smp_store_release(&waiter->task, NULL);
553 /*
554 * Ensure issuing the wakeup (either by us or someone else)
555 * after setting the reader waiter to nil.
556 */
557 wake_q_add_safe(wake_q, tsk);
558 }
559}
560
561/*
562 * This function must be called with the sem->wait_lock held to prevent
563 * race conditions between checking the rwsem wait list and setting the
564 * sem->count accordingly.
565 *
566 * Implies rwsem_del_waiter() on success.
567 */
568static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
569 struct rwsem_waiter *waiter)
570{
571 bool first = rwsem_first_waiter(sem) == waiter;
572 long count, new;
573
574 lockdep_assert_held(&sem->wait_lock);
575
576 count = atomic_long_read(&sem->count);
577 do {
578 bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
579
580 if (has_handoff) {
581 if (!first)
582 return false;
583
584 /* First waiter inherits a previously set handoff bit */
585 waiter->handoff_set = true;
586 }
587
588 new = count;
589
590 if (count & RWSEM_LOCK_MASK) {
591 if (has_handoff || (!rt_task(waiter->task) &&
592 !time_after(jiffies, waiter->timeout)))
593 return false;
594
595 new |= RWSEM_FLAG_HANDOFF;
596 } else {
597 new |= RWSEM_WRITER_LOCKED;
598 new &= ~RWSEM_FLAG_HANDOFF;
599
600 if (list_is_singular(&sem->wait_list))
601 new &= ~RWSEM_FLAG_WAITERS;
602 }
603 } while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
604
605 /*
606 * We have either acquired the lock with handoff bit cleared or
607 * set the handoff bit.
608 */
609 if (new & RWSEM_FLAG_HANDOFF) {
610 waiter->handoff_set = true;
611 lockevent_inc(rwsem_wlock_handoff);
612 return false;
613 }
614
615 /*
616 * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
617 * success.
618 */
619 list_del(&waiter->list);
620 rwsem_set_owner(sem);
621 return true;
622}
623
624/*
625 * The rwsem_spin_on_owner() function returns the following 4 values
626 * depending on the lock owner state.
627 * OWNER_NULL : owner is currently NULL
628 * OWNER_WRITER: when owner changes and is a writer
629 * OWNER_READER: when owner changes and the new owner may be a reader.
630 * OWNER_NONSPINNABLE:
631 * when optimistic spinning has to stop because either the
632 * owner stops running, is unknown, or its timeslice has
633 * been used up.
634 */
635enum owner_state {
636 OWNER_NULL = 1 << 0,
637 OWNER_WRITER = 1 << 1,
638 OWNER_READER = 1 << 2,
639 OWNER_NONSPINNABLE = 1 << 3,
640};
641
642#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
643/*
644 * Try to acquire write lock before the writer has been put on wait queue.
645 */
646static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
647{
648 long count = atomic_long_read(&sem->count);
649
650 while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
651 if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
652 count | RWSEM_WRITER_LOCKED)) {
653 rwsem_set_owner(sem);
654 lockevent_inc(rwsem_opt_lock);
655 return true;
656 }
657 }
658 return false;
659}
660
661static inline bool owner_on_cpu(struct task_struct *owner)
662{
663 /*
664 * As lock holder preemption issue, we both skip spinning if
665 * task is not on cpu or its cpu is preempted
666 */
667 return owner->on_cpu && !vcpu_is_preempted(task_cpu(owner));
668}
669
670static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
671{
672 struct task_struct *owner;
673 unsigned long flags;
674 bool ret = true;
675
676 if (need_resched()) {
677 lockevent_inc(rwsem_opt_fail);
678 return false;
679 }
680
681 preempt_disable();
682 /*
683 * Disable preemption is equal to the RCU read-side crital section,
684 * thus the task_strcut structure won't go away.
685 */
686 owner = rwsem_owner_flags(sem, &flags);
687 /*
688 * Don't check the read-owner as the entry may be stale.
689 */
690 if ((flags & RWSEM_NONSPINNABLE) ||
691 (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
692 ret = false;
693 preempt_enable();
694
695 lockevent_cond_inc(rwsem_opt_fail, !ret);
696 return ret;
697}
698
699#define OWNER_SPINNABLE (OWNER_NULL | OWNER_WRITER | OWNER_READER)
700
701static inline enum owner_state
702rwsem_owner_state(struct task_struct *owner, unsigned long flags)
703{
704 if (flags & RWSEM_NONSPINNABLE)
705 return OWNER_NONSPINNABLE;
706
707 if (flags & RWSEM_READER_OWNED)
708 return OWNER_READER;
709
710 return owner ? OWNER_WRITER : OWNER_NULL;
711}
712
713static noinline enum owner_state
714rwsem_spin_on_owner(struct rw_semaphore *sem)
715{
716 struct task_struct *new, *owner;
717 unsigned long flags, new_flags;
718 enum owner_state state;
719
720 lockdep_assert_preemption_disabled();
721
722 owner = rwsem_owner_flags(sem, &flags);
723 state = rwsem_owner_state(owner, flags);
724 if (state != OWNER_WRITER)
725 return state;
726
727 for (;;) {
728 /*
729 * When a waiting writer set the handoff flag, it may spin
730 * on the owner as well. Once that writer acquires the lock,
731 * we can spin on it. So we don't need to quit even when the
732 * handoff bit is set.
733 */
734 new = rwsem_owner_flags(sem, &new_flags);
735 if ((new != owner) || (new_flags != flags)) {
736 state = rwsem_owner_state(new, new_flags);
737 break;
738 }
739
740 /*
741 * Ensure we emit the owner->on_cpu, dereference _after_
742 * checking sem->owner still matches owner, if that fails,
743 * owner might point to free()d memory, if it still matches,
744 * our spinning context already disabled preemption which is
745 * equal to RCU read-side crital section ensures the memory
746 * stays valid.
747 */
748 barrier();
749
750 if (need_resched() || !owner_on_cpu(owner)) {
751 state = OWNER_NONSPINNABLE;
752 break;
753 }
754
755 cpu_relax();
756 }
757
758 return state;
759}
760
761/*
762 * Calculate reader-owned rwsem spinning threshold for writer
763 *
764 * The more readers own the rwsem, the longer it will take for them to
765 * wind down and free the rwsem. So the empirical formula used to
766 * determine the actual spinning time limit here is:
767 *
768 * Spinning threshold = (10 + nr_readers/2)us
769 *
770 * The limit is capped to a maximum of 25us (30 readers). This is just
771 * a heuristic and is subjected to change in the future.
772 */
773static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
774{
775 long count = atomic_long_read(&sem->count);
776 int readers = count >> RWSEM_READER_SHIFT;
777 u64 delta;
778
779 if (readers > 30)
780 readers = 30;
781 delta = (20 + readers) * NSEC_PER_USEC / 2;
782
783 return sched_clock() + delta;
784}
785
786static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
787{
788 bool taken = false;
789 int prev_owner_state = OWNER_NULL;
790 int loop = 0;
791 u64 rspin_threshold = 0;
792
793 preempt_disable();
794
795 /* sem->wait_lock should not be held when doing optimistic spinning */
796 if (!osq_lock(&sem->osq))
797 goto done;
798
799 /*
800 * Optimistically spin on the owner field and attempt to acquire the
801 * lock whenever the owner changes. Spinning will be stopped when:
802 * 1) the owning writer isn't running; or
803 * 2) readers own the lock and spinning time has exceeded limit.
804 */
805 for (;;) {
806 enum owner_state owner_state;
807
808 owner_state = rwsem_spin_on_owner(sem);
809 if (!(owner_state & OWNER_SPINNABLE))
810 break;
811
812 /*
813 * Try to acquire the lock
814 */
815 taken = rwsem_try_write_lock_unqueued(sem);
816
817 if (taken)
818 break;
819
820 /*
821 * Time-based reader-owned rwsem optimistic spinning
822 */
823 if (owner_state == OWNER_READER) {
824 /*
825 * Re-initialize rspin_threshold every time when
826 * the owner state changes from non-reader to reader.
827 * This allows a writer to steal the lock in between
828 * 2 reader phases and have the threshold reset at
829 * the beginning of the 2nd reader phase.
830 */
831 if (prev_owner_state != OWNER_READER) {
832 if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
833 break;
834 rspin_threshold = rwsem_rspin_threshold(sem);
835 loop = 0;
836 }
837
838 /*
839 * Check time threshold once every 16 iterations to
840 * avoid calling sched_clock() too frequently so
841 * as to reduce the average latency between the times
842 * when the lock becomes free and when the spinner
843 * is ready to do a trylock.
844 */
845 else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
846 rwsem_set_nonspinnable(sem);
847 lockevent_inc(rwsem_opt_nospin);
848 break;
849 }
850 }
851
852 /*
853 * An RT task cannot do optimistic spinning if it cannot
854 * be sure the lock holder is running or live-lock may
855 * happen if the current task and the lock holder happen
856 * to run in the same CPU. However, aborting optimistic
857 * spinning while a NULL owner is detected may miss some
858 * opportunity where spinning can continue without causing
859 * problem.
860 *
861 * There are 2 possible cases where an RT task may be able
862 * to continue spinning.
863 *
864 * 1) The lock owner is in the process of releasing the
865 * lock, sem->owner is cleared but the lock has not
866 * been released yet.
867 * 2) The lock was free and owner cleared, but another
868 * task just comes in and acquire the lock before
869 * we try to get it. The new owner may be a spinnable
870 * writer.
871 *
872 * To take advantage of two scenarios listed above, the RT
873 * task is made to retry one more time to see if it can
874 * acquire the lock or continue spinning on the new owning
875 * writer. Of course, if the time lag is long enough or the
876 * new owner is not a writer or spinnable, the RT task will
877 * quit spinning.
878 *
879 * If the owner is a writer, the need_resched() check is
880 * done inside rwsem_spin_on_owner(). If the owner is not
881 * a writer, need_resched() check needs to be done here.
882 */
883 if (owner_state != OWNER_WRITER) {
884 if (need_resched())
885 break;
886 if (rt_task(current) &&
887 (prev_owner_state != OWNER_WRITER))
888 break;
889 }
890 prev_owner_state = owner_state;
891
892 /*
893 * The cpu_relax() call is a compiler barrier which forces
894 * everything in this loop to be re-loaded. We don't need
895 * memory barriers as we'll eventually observe the right
896 * values at the cost of a few extra spins.
897 */
898 cpu_relax();
899 }
900 osq_unlock(&sem->osq);
901done:
902 preempt_enable();
903 lockevent_cond_inc(rwsem_opt_fail, !taken);
904 return taken;
905}
906
907/*
908 * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
909 * only be called when the reader count reaches 0.
910 */
911static inline void clear_nonspinnable(struct rw_semaphore *sem)
912{
913 if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
914 atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
915}
916
917#else
918static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
919{
920 return false;
921}
922
923static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
924{
925 return false;
926}
927
928static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
929
930static inline enum owner_state
931rwsem_spin_on_owner(struct rw_semaphore *sem)
932{
933 return OWNER_NONSPINNABLE;
934}
935#endif
936
937/*
938 * Wait for the read lock to be granted
939 */
940static struct rw_semaphore __sched *
941rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
942{
943 long adjustment = -RWSEM_READER_BIAS;
944 long rcnt = (count >> RWSEM_READER_SHIFT);
945 struct rwsem_waiter waiter;
946 DEFINE_WAKE_Q(wake_q);
947 bool wake = false;
948
949 /*
950 * To prevent a constant stream of readers from starving a sleeping
951 * waiter, don't attempt optimistic lock stealing if the lock is
952 * currently owned by readers.
953 */
954 if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
955 (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
956 goto queue;
957
958 /*
959 * Reader optimistic lock stealing.
960 */
961 if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
962 rwsem_set_reader_owned(sem);
963 lockevent_inc(rwsem_rlock_steal);
964
965 /*
966 * Wake up other readers in the wait queue if it is
967 * the first reader.
968 */
969 if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
970 raw_spin_lock_irq(&sem->wait_lock);
971 if (!list_empty(&sem->wait_list))
972 rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
973 &wake_q);
974 raw_spin_unlock_irq(&sem->wait_lock);
975 wake_up_q(&wake_q);
976 }
977 return sem;
978 }
979
980queue:
981 waiter.task = current;
982 waiter.type = RWSEM_WAITING_FOR_READ;
983 waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
984
985 raw_spin_lock_irq(&sem->wait_lock);
986 if (list_empty(&sem->wait_list)) {
987 /*
988 * In case the wait queue is empty and the lock isn't owned
989 * by a writer or has the handoff bit set, this reader can
990 * exit the slowpath and return immediately as its
991 * RWSEM_READER_BIAS has already been set in the count.
992 */
993 if (!(atomic_long_read(&sem->count) &
994 (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
995 /* Provide lock ACQUIRE */
996 smp_acquire__after_ctrl_dep();
997 raw_spin_unlock_irq(&sem->wait_lock);
998 rwsem_set_reader_owned(sem);
999 lockevent_inc(rwsem_rlock_fast);
1000 return sem;
1001 }
1002 adjustment += RWSEM_FLAG_WAITERS;
1003 }
1004 rwsem_add_waiter(sem, &waiter);
1005
1006 /* we're now waiting on the lock, but no longer actively locking */
1007 count = atomic_long_add_return(adjustment, &sem->count);
1008
1009 /*
1010 * If there are no active locks, wake the front queued process(es).
1011 *
1012 * If there are no writers and we are first in the queue,
1013 * wake our own waiter to join the existing active readers !
1014 */
1015 if (!(count & RWSEM_LOCK_MASK)) {
1016 clear_nonspinnable(sem);
1017 wake = true;
1018 }
1019 if (wake || (!(count & RWSEM_WRITER_MASK) &&
1020 (adjustment & RWSEM_FLAG_WAITERS)))
1021 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1022
1023 raw_spin_unlock_irq(&sem->wait_lock);
1024 wake_up_q(&wake_q);
1025
1026 /* wait to be given the lock */
1027 for (;;) {
1028 set_current_state(state);
1029 if (!smp_load_acquire(&waiter.task)) {
1030 /* Matches rwsem_mark_wake()'s smp_store_release(). */
1031 break;
1032 }
1033 if (signal_pending_state(state, current)) {
1034 raw_spin_lock_irq(&sem->wait_lock);
1035 if (waiter.task)
1036 goto out_nolock;
1037 raw_spin_unlock_irq(&sem->wait_lock);
1038 /* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1039 break;
1040 }
1041 schedule();
1042 lockevent_inc(rwsem_sleep_reader);
1043 }
1044
1045 __set_current_state(TASK_RUNNING);
1046 lockevent_inc(rwsem_rlock);
1047 return sem;
1048
1049out_nolock:
1050 rwsem_del_waiter(sem, &waiter);
1051 raw_spin_unlock_irq(&sem->wait_lock);
1052 __set_current_state(TASK_RUNNING);
1053 lockevent_inc(rwsem_rlock_fail);
1054 return ERR_PTR(-EINTR);
1055}
1056
1057/*
1058 * Wait until we successfully acquire the write lock
1059 */
1060static struct rw_semaphore *
1061rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1062{
1063 long count;
1064 struct rwsem_waiter waiter;
1065 DEFINE_WAKE_Q(wake_q);
1066
1067 /* do optimistic spinning and steal lock if possible */
1068 if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1069 /* rwsem_optimistic_spin() implies ACQUIRE on success */
1070 return sem;
1071 }
1072
1073 /*
1074 * Optimistic spinning failed, proceed to the slowpath
1075 * and block until we can acquire the sem.
1076 */
1077 waiter.task = current;
1078 waiter.type = RWSEM_WAITING_FOR_WRITE;
1079 waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1080 waiter.handoff_set = false;
1081
1082 raw_spin_lock_irq(&sem->wait_lock);
1083 rwsem_add_waiter(sem, &waiter);
1084
1085 /* we're now waiting on the lock */
1086 if (rwsem_first_waiter(sem) != &waiter) {
1087 count = atomic_long_read(&sem->count);
1088
1089 /*
1090 * If there were already threads queued before us and:
1091 * 1) there are no active locks, wake the front
1092 * queued process(es) as the handoff bit might be set.
1093 * 2) there are no active writers and some readers, the lock
1094 * must be read owned; so we try to wake any read lock
1095 * waiters that were queued ahead of us.
1096 */
1097 if (count & RWSEM_WRITER_MASK)
1098 goto wait;
1099
1100 rwsem_mark_wake(sem, (count & RWSEM_READER_MASK)
1101 ? RWSEM_WAKE_READERS
1102 : RWSEM_WAKE_ANY, &wake_q);
1103
1104 if (!wake_q_empty(&wake_q)) {
1105 /*
1106 * We want to minimize wait_lock hold time especially
1107 * when a large number of readers are to be woken up.
1108 */
1109 raw_spin_unlock_irq(&sem->wait_lock);
1110 wake_up_q(&wake_q);
1111 wake_q_init(&wake_q); /* Used again, reinit */
1112 raw_spin_lock_irq(&sem->wait_lock);
1113 }
1114 } else {
1115 atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1116 }
1117
1118wait:
1119 /* wait until we successfully acquire the lock */
1120 set_current_state(state);
1121 for (;;) {
1122 if (rwsem_try_write_lock(sem, &waiter)) {
1123 /* rwsem_try_write_lock() implies ACQUIRE on success */
1124 break;
1125 }
1126
1127 raw_spin_unlock_irq(&sem->wait_lock);
1128
1129 if (signal_pending_state(state, current))
1130 goto out_nolock;
1131
1132 /*
1133 * After setting the handoff bit and failing to acquire
1134 * the lock, attempt to spin on owner to accelerate lock
1135 * transfer. If the previous owner is a on-cpu writer and it
1136 * has just released the lock, OWNER_NULL will be returned.
1137 * In this case, we attempt to acquire the lock again
1138 * without sleeping.
1139 */
1140 if (waiter.handoff_set) {
1141 enum owner_state owner_state;
1142
1143 preempt_disable();
1144 owner_state = rwsem_spin_on_owner(sem);
1145 preempt_enable();
1146
1147 if (owner_state == OWNER_NULL)
1148 goto trylock_again;
1149 }
1150
1151 schedule();
1152 lockevent_inc(rwsem_sleep_writer);
1153 set_current_state(state);
1154trylock_again:
1155 raw_spin_lock_irq(&sem->wait_lock);
1156 }
1157 __set_current_state(TASK_RUNNING);
1158 raw_spin_unlock_irq(&sem->wait_lock);
1159 lockevent_inc(rwsem_wlock);
1160 return sem;
1161
1162out_nolock:
1163 __set_current_state(TASK_RUNNING);
1164 raw_spin_lock_irq(&sem->wait_lock);
1165 rwsem_del_waiter(sem, &waiter);
1166 if (!list_empty(&sem->wait_list))
1167 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1168 raw_spin_unlock_irq(&sem->wait_lock);
1169 wake_up_q(&wake_q);
1170 lockevent_inc(rwsem_wlock_fail);
1171 return ERR_PTR(-EINTR);
1172}
1173
1174/*
1175 * handle waking up a waiter on the semaphore
1176 * - up_read/up_write has decremented the active part of count if we come here
1177 */
1178static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1179{
1180 unsigned long flags;
1181 DEFINE_WAKE_Q(wake_q);
1182
1183 raw_spin_lock_irqsave(&sem->wait_lock, flags);
1184
1185 if (!list_empty(&sem->wait_list))
1186 rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1187
1188 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1189 wake_up_q(&wake_q);
1190
1191 return sem;
1192}
1193
1194/*
1195 * downgrade a write lock into a read lock
1196 * - caller incremented waiting part of count and discovered it still negative
1197 * - just wake up any readers at the front of the queue
1198 */
1199static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1200{
1201 unsigned long flags;
1202 DEFINE_WAKE_Q(wake_q);
1203
1204 raw_spin_lock_irqsave(&sem->wait_lock, flags);
1205
1206 if (!list_empty(&sem->wait_list))
1207 rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1208
1209 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1210 wake_up_q(&wake_q);
1211
1212 return sem;
1213}
1214
1215/*
1216 * lock for reading
1217 */
1218static inline int __down_read_common(struct rw_semaphore *sem, int state)
1219{
1220 long count;
1221
1222 if (!rwsem_read_trylock(sem, &count)) {
1223 if (IS_ERR(rwsem_down_read_slowpath(sem, count, state)))
1224 return -EINTR;
1225 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1226 }
1227 return 0;
1228}
1229
1230static inline void __down_read(struct rw_semaphore *sem)
1231{
1232 __down_read_common(sem, TASK_UNINTERRUPTIBLE);
1233}
1234
1235static inline int __down_read_interruptible(struct rw_semaphore *sem)
1236{
1237 return __down_read_common(sem, TASK_INTERRUPTIBLE);
1238}
1239
1240static inline int __down_read_killable(struct rw_semaphore *sem)
1241{
1242 return __down_read_common(sem, TASK_KILLABLE);
1243}
1244
1245static inline int __down_read_trylock(struct rw_semaphore *sem)
1246{
1247 long tmp;
1248
1249 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1250
1251 tmp = atomic_long_read(&sem->count);
1252 while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1253 if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1254 tmp + RWSEM_READER_BIAS)) {
1255 rwsem_set_reader_owned(sem);
1256 return 1;
1257 }
1258 }
1259 return 0;
1260}
1261
1262/*
1263 * lock for writing
1264 */
1265static inline int __down_write_common(struct rw_semaphore *sem, int state)
1266{
1267 if (unlikely(!rwsem_write_trylock(sem))) {
1268 if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
1269 return -EINTR;
1270 }
1271
1272 return 0;
1273}
1274
1275static inline void __down_write(struct rw_semaphore *sem)
1276{
1277 __down_write_common(sem, TASK_UNINTERRUPTIBLE);
1278}
1279
1280static inline int __down_write_killable(struct rw_semaphore *sem)
1281{
1282 return __down_write_common(sem, TASK_KILLABLE);
1283}
1284
1285static inline int __down_write_trylock(struct rw_semaphore *sem)
1286{
1287 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1288 return rwsem_write_trylock(sem);
1289}
1290
1291/*
1292 * unlock after reading
1293 */
1294static inline void __up_read(struct rw_semaphore *sem)
1295{
1296 long tmp;
1297
1298 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1299 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1300
1301 rwsem_clear_reader_owned(sem);
1302 tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1303 DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1304 if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1305 RWSEM_FLAG_WAITERS)) {
1306 clear_nonspinnable(sem);
1307 rwsem_wake(sem);
1308 }
1309}
1310
1311/*
1312 * unlock after writing
1313 */
1314static inline void __up_write(struct rw_semaphore *sem)
1315{
1316 long tmp;
1317
1318 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1319 /*
1320 * sem->owner may differ from current if the ownership is transferred
1321 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1322 */
1323 DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1324 !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1325
1326 rwsem_clear_owner(sem);
1327 tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1328 if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1329 rwsem_wake(sem);
1330}
1331
1332/*
1333 * downgrade write lock to read lock
1334 */
1335static inline void __downgrade_write(struct rw_semaphore *sem)
1336{
1337 long tmp;
1338
1339 /*
1340 * When downgrading from exclusive to shared ownership,
1341 * anything inside the write-locked region cannot leak
1342 * into the read side. In contrast, anything in the
1343 * read-locked region is ok to be re-ordered into the
1344 * write side. As such, rely on RELEASE semantics.
1345 */
1346 DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1347 tmp = atomic_long_fetch_add_release(
1348 -RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1349 rwsem_set_reader_owned(sem);
1350 if (tmp & RWSEM_FLAG_WAITERS)
1351 rwsem_downgrade_wake(sem);
1352}
1353
1354#else /* !CONFIG_PREEMPT_RT */
1355
1356#define RT_MUTEX_BUILD_MUTEX
1357#include "rtmutex.c"
1358
1359#define rwbase_set_and_save_current_state(state) \
1360 set_current_state(state)
1361
1362#define rwbase_restore_current_state() \
1363 __set_current_state(TASK_RUNNING)
1364
1365#define rwbase_rtmutex_lock_state(rtm, state) \
1366 __rt_mutex_lock(rtm, state)
1367
1368#define rwbase_rtmutex_slowlock_locked(rtm, state) \
1369 __rt_mutex_slowlock_locked(rtm, NULL, state)
1370
1371#define rwbase_rtmutex_unlock(rtm) \
1372 __rt_mutex_unlock(rtm)
1373
1374#define rwbase_rtmutex_trylock(rtm) \
1375 __rt_mutex_trylock(rtm)
1376
1377#define rwbase_signal_pending_state(state, current) \
1378 signal_pending_state(state, current)
1379
1380#define rwbase_schedule() \
1381 schedule()
1382
1383#include "rwbase_rt.c"
1384
1385void __init_rwsem(struct rw_semaphore *sem, const char *name,
1386 struct lock_class_key *key)
1387{
1388 init_rwbase_rt(&(sem)->rwbase);
1389
1390#ifdef CONFIG_DEBUG_LOCK_ALLOC
1391 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1392 lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1393#endif
1394}
1395EXPORT_SYMBOL(__init_rwsem);
1396
1397static inline void __down_read(struct rw_semaphore *sem)
1398{
1399 rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1400}
1401
1402static inline int __down_read_interruptible(struct rw_semaphore *sem)
1403{
1404 return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1405}
1406
1407static inline int __down_read_killable(struct rw_semaphore *sem)
1408{
1409 return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1410}
1411
1412static inline int __down_read_trylock(struct rw_semaphore *sem)
1413{
1414 return rwbase_read_trylock(&sem->rwbase);
1415}
1416
1417static inline void __up_read(struct rw_semaphore *sem)
1418{
1419 rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1420}
1421
1422static inline void __sched __down_write(struct rw_semaphore *sem)
1423{
1424 rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1425}
1426
1427static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1428{
1429 return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1430}
1431
1432static inline int __down_write_trylock(struct rw_semaphore *sem)
1433{
1434 return rwbase_write_trylock(&sem->rwbase);
1435}
1436
1437static inline void __up_write(struct rw_semaphore *sem)
1438{
1439 rwbase_write_unlock(&sem->rwbase);
1440}
1441
1442static inline void __downgrade_write(struct rw_semaphore *sem)
1443{
1444 rwbase_write_downgrade(&sem->rwbase);
1445}
1446
1447/* Debug stubs for the common API */
1448#define DEBUG_RWSEMS_WARN_ON(c, sem)
1449
1450static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1451 struct task_struct *owner)
1452{
1453}
1454
1455static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1456{
1457 int count = atomic_read(&sem->rwbase.readers);
1458
1459 return count < 0 && count != READER_BIAS;
1460}
1461
1462#endif /* CONFIG_PREEMPT_RT */
1463
1464/*
1465 * lock for reading
1466 */
1467void __sched down_read(struct rw_semaphore *sem)
1468{
1469 might_sleep();
1470 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1471
1472 LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1473}
1474EXPORT_SYMBOL(down_read);
1475
1476int __sched down_read_interruptible(struct rw_semaphore *sem)
1477{
1478 might_sleep();
1479 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1480
1481 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1482 rwsem_release(&sem->dep_map, _RET_IP_);
1483 return -EINTR;
1484 }
1485
1486 return 0;
1487}
1488EXPORT_SYMBOL(down_read_interruptible);
1489
1490int __sched down_read_killable(struct rw_semaphore *sem)
1491{
1492 might_sleep();
1493 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1494
1495 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1496 rwsem_release(&sem->dep_map, _RET_IP_);
1497 return -EINTR;
1498 }
1499
1500 return 0;
1501}
1502EXPORT_SYMBOL(down_read_killable);
1503
1504/*
1505 * trylock for reading -- returns 1 if successful, 0 if contention
1506 */
1507int down_read_trylock(struct rw_semaphore *sem)
1508{
1509 int ret = __down_read_trylock(sem);
1510
1511 if (ret == 1)
1512 rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1513 return ret;
1514}
1515EXPORT_SYMBOL(down_read_trylock);
1516
1517/*
1518 * lock for writing
1519 */
1520void __sched down_write(struct rw_semaphore *sem)
1521{
1522 might_sleep();
1523 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1524 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1525}
1526EXPORT_SYMBOL(down_write);
1527
1528/*
1529 * lock for writing
1530 */
1531int __sched down_write_killable(struct rw_semaphore *sem)
1532{
1533 might_sleep();
1534 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1535
1536 if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1537 __down_write_killable)) {
1538 rwsem_release(&sem->dep_map, _RET_IP_);
1539 return -EINTR;
1540 }
1541
1542 return 0;
1543}
1544EXPORT_SYMBOL(down_write_killable);
1545
1546/*
1547 * trylock for writing -- returns 1 if successful, 0 if contention
1548 */
1549int down_write_trylock(struct rw_semaphore *sem)
1550{
1551 int ret = __down_write_trylock(sem);
1552
1553 if (ret == 1)
1554 rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1555
1556 return ret;
1557}
1558EXPORT_SYMBOL(down_write_trylock);
1559
1560/*
1561 * release a read lock
1562 */
1563void up_read(struct rw_semaphore *sem)
1564{
1565 rwsem_release(&sem->dep_map, _RET_IP_);
1566 __up_read(sem);
1567}
1568EXPORT_SYMBOL(up_read);
1569
1570/*
1571 * release a write lock
1572 */
1573void up_write(struct rw_semaphore *sem)
1574{
1575 rwsem_release(&sem->dep_map, _RET_IP_);
1576 __up_write(sem);
1577}
1578EXPORT_SYMBOL(up_write);
1579
1580/*
1581 * downgrade write lock to read lock
1582 */
1583void downgrade_write(struct rw_semaphore *sem)
1584{
1585 lock_downgrade(&sem->dep_map, _RET_IP_);
1586 __downgrade_write(sem);
1587}
1588EXPORT_SYMBOL(downgrade_write);
1589
1590#ifdef CONFIG_DEBUG_LOCK_ALLOC
1591
1592void down_read_nested(struct rw_semaphore *sem, int subclass)
1593{
1594 might_sleep();
1595 rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1596 LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1597}
1598EXPORT_SYMBOL(down_read_nested);
1599
1600int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1601{
1602 might_sleep();
1603 rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1604
1605 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1606 rwsem_release(&sem->dep_map, _RET_IP_);
1607 return -EINTR;
1608 }
1609
1610 return 0;
1611}
1612EXPORT_SYMBOL(down_read_killable_nested);
1613
1614void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1615{
1616 might_sleep();
1617 rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1618 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1619}
1620EXPORT_SYMBOL(_down_write_nest_lock);
1621
1622void down_read_non_owner(struct rw_semaphore *sem)
1623{
1624 might_sleep();
1625 __down_read(sem);
1626 __rwsem_set_reader_owned(sem, NULL);
1627}
1628EXPORT_SYMBOL(down_read_non_owner);
1629
1630void down_write_nested(struct rw_semaphore *sem, int subclass)
1631{
1632 might_sleep();
1633 rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1634 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1635}
1636EXPORT_SYMBOL(down_write_nested);
1637
1638int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1639{
1640 might_sleep();
1641 rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1642
1643 if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1644 __down_write_killable)) {
1645 rwsem_release(&sem->dep_map, _RET_IP_);
1646 return -EINTR;
1647 }
1648
1649 return 0;
1650}
1651EXPORT_SYMBOL(down_write_killable_nested);
1652
1653void up_read_non_owner(struct rw_semaphore *sem)
1654{
1655 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1656 __up_read(sem);
1657}
1658EXPORT_SYMBOL(up_read_non_owner);
1659
1660#endif