1// SPDX-License-Identifier: GPL-2.0
2/* kernel/rwsem.c: R/W semaphores, public implementation
3 *
4 * Written by David Howells (dhowells@redhat.com).
5 * Derived from asm-i386/semaphore.h
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 * Rwsem count bit fields re-definition and rwsem rearchitecture by
14 * Waiman Long <longman@redhat.com> and
15 * Peter Zijlstra <peterz@infradead.org>.
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 rwsem_can_spin_on_owner(struct rw_semaphore *sem)
662{
663	struct task_struct *owner;
664	unsigned long flags;
665	bool ret = true;
666
667	if (need_resched()) {
668		lockevent_inc(rwsem_opt_fail);
669		return false;
670	}
671
672	preempt_disable();
673	/*
674	 * Disable preemption is equal to the RCU read-side crital section,
675	 * thus the task_strcut structure won't go away.
676	 */
677	owner = rwsem_owner_flags(sem, &flags);
678	/*
679	 * Don't check the read-owner as the entry may be stale.
680	 */
681	if ((flags & RWSEM_NONSPINNABLE) ||
682	    (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
683		ret = false;
684	preempt_enable();
685
686	lockevent_cond_inc(rwsem_opt_fail, !ret);
687	return ret;
688}
689
690#define OWNER_SPINNABLE		(OWNER_NULL | OWNER_WRITER | OWNER_READER)
691
692static inline enum owner_state
693rwsem_owner_state(struct task_struct *owner, unsigned long flags)
694{
695	if (flags & RWSEM_NONSPINNABLE)
696		return OWNER_NONSPINNABLE;
697
698	if (flags & RWSEM_READER_OWNED)
699		return OWNER_READER;
700
701	return owner ? OWNER_WRITER : OWNER_NULL;
702}
703
704static noinline enum owner_state
705rwsem_spin_on_owner(struct rw_semaphore *sem)
706{
707	struct task_struct *new, *owner;
708	unsigned long flags, new_flags;
709	enum owner_state state;
710
711	lockdep_assert_preemption_disabled();
712
713	owner = rwsem_owner_flags(sem, &flags);
714	state = rwsem_owner_state(owner, flags);
715	if (state != OWNER_WRITER)
716		return state;
717
718	for (;;) {
719		/*
720		 * When a waiting writer set the handoff flag, it may spin
721		 * on the owner as well. Once that writer acquires the lock,
722		 * we can spin on it. So we don't need to quit even when the
723		 * handoff bit is set.
724		 */
725		new = rwsem_owner_flags(sem, &new_flags);
726		if ((new != owner) || (new_flags != flags)) {
727			state = rwsem_owner_state(new, new_flags);
728			break;
729		}
730
731		/*
732		 * Ensure we emit the owner->on_cpu, dereference _after_
733		 * checking sem->owner still matches owner, if that fails,
734		 * owner might point to free()d memory, if it still matches,
735		 * our spinning context already disabled preemption which is
736		 * equal to RCU read-side crital section ensures the memory
737		 * stays valid.
738		 */
739		barrier();
740
741		if (need_resched() || !owner_on_cpu(owner)) {
742			state = OWNER_NONSPINNABLE;
743			break;
744		}
745
746		cpu_relax();
747	}
748
749	return state;
750}
751
752/*
753 * Calculate reader-owned rwsem spinning threshold for writer
754 *
755 * The more readers own the rwsem, the longer it will take for them to
756 * wind down and free the rwsem. So the empirical formula used to
757 * determine the actual spinning time limit here is:
758 *
759 *   Spinning threshold = (10 + nr_readers/2)us
760 *
761 * The limit is capped to a maximum of 25us (30 readers). This is just
762 * a heuristic and is subjected to change in the future.
763 */
764static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
765{
766	long count = atomic_long_read(&sem->count);
767	int readers = count >> RWSEM_READER_SHIFT;
768	u64 delta;
769
770	if (readers > 30)
771		readers = 30;
772	delta = (20 + readers) * NSEC_PER_USEC / 2;
773
774	return sched_clock() + delta;
775}
776
777static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
778{
779	bool taken = false;
780	int prev_owner_state = OWNER_NULL;
781	int loop = 0;
782	u64 rspin_threshold = 0;
783
784	preempt_disable();
785
786	/* sem->wait_lock should not be held when doing optimistic spinning */
787	if (!osq_lock(&sem->osq))
788		goto done;
789
790	/*
791	 * Optimistically spin on the owner field and attempt to acquire the
792	 * lock whenever the owner changes. Spinning will be stopped when:
793	 *  1) the owning writer isn't running; or
794	 *  2) readers own the lock and spinning time has exceeded limit.
795	 */
796	for (;;) {
797		enum owner_state owner_state;
798
799		owner_state = rwsem_spin_on_owner(sem);
800		if (!(owner_state & OWNER_SPINNABLE))
801			break;
802
803		/*
804		 * Try to acquire the lock
805		 */
806		taken = rwsem_try_write_lock_unqueued(sem);
807
808		if (taken)
809			break;
810
811		/*
812		 * Time-based reader-owned rwsem optimistic spinning
813		 */
814		if (owner_state == OWNER_READER) {
815			/*
816			 * Re-initialize rspin_threshold every time when
817			 * the owner state changes from non-reader to reader.
818			 * This allows a writer to steal the lock in between
819			 * 2 reader phases and have the threshold reset at
820			 * the beginning of the 2nd reader phase.
821			 */
822			if (prev_owner_state != OWNER_READER) {
823				if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
824					break;
825				rspin_threshold = rwsem_rspin_threshold(sem);
826				loop = 0;
827			}
828
829			/*
830			 * Check time threshold once every 16 iterations to
831			 * avoid calling sched_clock() too frequently so
832			 * as to reduce the average latency between the times
833			 * when the lock becomes free and when the spinner
834			 * is ready to do a trylock.
835			 */
836			else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
837				rwsem_set_nonspinnable(sem);
838				lockevent_inc(rwsem_opt_nospin);
839				break;
840			}
841		}
842
843		/*
844		 * An RT task cannot do optimistic spinning if it cannot
845		 * be sure the lock holder is running or live-lock may
846		 * happen if the current task and the lock holder happen
847		 * to run in the same CPU. However, aborting optimistic
848		 * spinning while a NULL owner is detected may miss some
849		 * opportunity where spinning can continue without causing
850		 * problem.
851		 *
852		 * There are 2 possible cases where an RT task may be able
853		 * to continue spinning.
854		 *
855		 * 1) The lock owner is in the process of releasing the
856		 *    lock, sem->owner is cleared but the lock has not
857		 *    been released yet.
858		 * 2) The lock was free and owner cleared, but another
859		 *    task just comes in and acquire the lock before
860		 *    we try to get it. The new owner may be a spinnable
861		 *    writer.
862		 *
863		 * To take advantage of two scenarios listed above, the RT
864		 * task is made to retry one more time to see if it can
865		 * acquire the lock or continue spinning on the new owning
866		 * writer. Of course, if the time lag is long enough or the
867		 * new owner is not a writer or spinnable, the RT task will
868		 * quit spinning.
869		 *
870		 * If the owner is a writer, the need_resched() check is
871		 * done inside rwsem_spin_on_owner(). If the owner is not
872		 * a writer, need_resched() check needs to be done here.
873		 */
874		if (owner_state != OWNER_WRITER) {
875			if (need_resched())
876				break;
877			if (rt_task(current) &&
878			   (prev_owner_state != OWNER_WRITER))
879				break;
880		}
881		prev_owner_state = owner_state;
882
883		/*
884		 * The cpu_relax() call is a compiler barrier which forces
885		 * everything in this loop to be re-loaded. We don't need
886		 * memory barriers as we'll eventually observe the right
887		 * values at the cost of a few extra spins.
888		 */
889		cpu_relax();
890	}
891	osq_unlock(&sem->osq);
892done:
893	preempt_enable();
894	lockevent_cond_inc(rwsem_opt_fail, !taken);
895	return taken;
896}
897
898/*
899 * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
900 * only be called when the reader count reaches 0.
901 */
902static inline void clear_nonspinnable(struct rw_semaphore *sem)
903{
904	if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
905		atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
906}
907
908#else
909static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
910{
911	return false;
912}
913
914static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
915{
916	return false;
917}
918
919static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
920
921static inline enum owner_state
922rwsem_spin_on_owner(struct rw_semaphore *sem)
923{
924	return OWNER_NONSPINNABLE;
925}
926#endif
927
928/*
929 * Wait for the read lock to be granted
930 */
931static struct rw_semaphore __sched *
932rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
933{
934	long adjustment = -RWSEM_READER_BIAS;
935	long rcnt = (count >> RWSEM_READER_SHIFT);
936	struct rwsem_waiter waiter;
937	DEFINE_WAKE_Q(wake_q);
938	bool wake = false;
939
940	/*
941	 * To prevent a constant stream of readers from starving a sleeping
942	 * waiter, don't attempt optimistic lock stealing if the lock is
943	 * currently owned by readers.
944	 */
945	if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
946	    (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
947		goto queue;
948
949	/*
950	 * Reader optimistic lock stealing.
951	 */
952	if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
953		rwsem_set_reader_owned(sem);
954		lockevent_inc(rwsem_rlock_steal);
955
956		/*
957		 * Wake up other readers in the wait queue if it is
958		 * the first reader.
959		 */
960		if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
961			raw_spin_lock_irq(&sem->wait_lock);
962			if (!list_empty(&sem->wait_list))
963				rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
964						&wake_q);
965			raw_spin_unlock_irq(&sem->wait_lock);
966			wake_up_q(&wake_q);
967		}
968		return sem;
969	}
970
971queue:
972	waiter.task = current;
973	waiter.type = RWSEM_WAITING_FOR_READ;
974	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
975
976	raw_spin_lock_irq(&sem->wait_lock);
977	if (list_empty(&sem->wait_list)) {
978		/*
979		 * In case the wait queue is empty and the lock isn't owned
980		 * by a writer or has the handoff bit set, this reader can
981		 * exit the slowpath and return immediately as its
982		 * RWSEM_READER_BIAS has already been set in the count.
983		 */
984		if (!(atomic_long_read(&sem->count) &
985		     (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
986			/* Provide lock ACQUIRE */
987			smp_acquire__after_ctrl_dep();
988			raw_spin_unlock_irq(&sem->wait_lock);
989			rwsem_set_reader_owned(sem);
990			lockevent_inc(rwsem_rlock_fast);
991			return sem;
992		}
993		adjustment += RWSEM_FLAG_WAITERS;
994	}
995	rwsem_add_waiter(sem, &waiter);
996
997	/* we're now waiting on the lock, but no longer actively locking */
998	count = atomic_long_add_return(adjustment, &sem->count);
999
1000	/*
1001	 * If there are no active locks, wake the front queued process(es).
1002	 *
1003	 * If there are no writers and we are first in the queue,
1004	 * wake our own waiter to join the existing active readers !
1005	 */
1006	if (!(count & RWSEM_LOCK_MASK)) {
1007		clear_nonspinnable(sem);
1008		wake = true;
1009	}
1010	if (wake || (!(count & RWSEM_WRITER_MASK) &&
1011		    (adjustment & RWSEM_FLAG_WAITERS)))
1012		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1013
1014	raw_spin_unlock_irq(&sem->wait_lock);
1015	wake_up_q(&wake_q);
1016
1017	/* wait to be given the lock */
1018	for (;;) {
1019		set_current_state(state);
1020		if (!smp_load_acquire(&waiter.task)) {
1021			/* Matches rwsem_mark_wake()'s smp_store_release(). */
1022			break;
1023		}
1024		if (signal_pending_state(state, current)) {
1025			raw_spin_lock_irq(&sem->wait_lock);
1026			if (waiter.task)
1027				goto out_nolock;
1028			raw_spin_unlock_irq(&sem->wait_lock);
1029			/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1030			break;
1031		}
1032		schedule();
1033		lockevent_inc(rwsem_sleep_reader);
1034	}
1035
1036	__set_current_state(TASK_RUNNING);
1037	lockevent_inc(rwsem_rlock);
1038	return sem;
1039
1040out_nolock:
1041	rwsem_del_waiter(sem, &waiter);
1042	raw_spin_unlock_irq(&sem->wait_lock);
1043	__set_current_state(TASK_RUNNING);
1044	lockevent_inc(rwsem_rlock_fail);
1045	return ERR_PTR(-EINTR);
1046}
1047
1048/*
1049 * Wait until we successfully acquire the write lock
1050 */
1051static struct rw_semaphore *
1052rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1053{
1054	long count;
1055	struct rwsem_waiter waiter;
1056	DEFINE_WAKE_Q(wake_q);
1057
1058	/* do optimistic spinning and steal lock if possible */
1059	if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1060		/* rwsem_optimistic_spin() implies ACQUIRE on success */
1061		return sem;
1062	}
1063
1064	/*
1065	 * Optimistic spinning failed, proceed to the slowpath
1066	 * and block until we can acquire the sem.
1067	 */
1068	waiter.task = current;
1069	waiter.type = RWSEM_WAITING_FOR_WRITE;
1070	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1071	waiter.handoff_set = false;
1072
1073	raw_spin_lock_irq(&sem->wait_lock);
1074	rwsem_add_waiter(sem, &waiter);
1075
1076	/* we're now waiting on the lock */
1077	if (rwsem_first_waiter(sem) != &waiter) {
1078		count = atomic_long_read(&sem->count);
1079
1080		/*
1081		 * If there were already threads queued before us and:
1082		 *  1) there are no active locks, wake the front
1083		 *     queued process(es) as the handoff bit might be set.
1084		 *  2) there are no active writers and some readers, the lock
1085		 *     must be read owned; so we try to wake any read lock
1086		 *     waiters that were queued ahead of us.
1087		 */
1088		if (count & RWSEM_WRITER_MASK)
1089			goto wait;
1090
1091		rwsem_mark_wake(sem, (count & RWSEM_READER_MASK)
1092					? RWSEM_WAKE_READERS
1093					: RWSEM_WAKE_ANY, &wake_q);
1094
1095		if (!wake_q_empty(&wake_q)) {
1096			/*
1097			 * We want to minimize wait_lock hold time especially
1098			 * when a large number of readers are to be woken up.
1099			 */
1100			raw_spin_unlock_irq(&sem->wait_lock);
1101			wake_up_q(&wake_q);
1102			wake_q_init(&wake_q);	/* Used again, reinit */
1103			raw_spin_lock_irq(&sem->wait_lock);
1104		}
1105	} else {
1106		atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1107	}
1108
1109wait:
1110	/* wait until we successfully acquire the lock */
1111	set_current_state(state);
1112	for (;;) {
1113		if (rwsem_try_write_lock(sem, &waiter)) {
1114			/* rwsem_try_write_lock() implies ACQUIRE on success */
1115			break;
1116		}
1117
1118		raw_spin_unlock_irq(&sem->wait_lock);
1119
1120		if (signal_pending_state(state, current))
1121			goto out_nolock;
1122
1123		/*
1124		 * After setting the handoff bit and failing to acquire
1125		 * the lock, attempt to spin on owner to accelerate lock
1126		 * transfer. If the previous owner is a on-cpu writer and it
1127		 * has just released the lock, OWNER_NULL will be returned.
1128		 * In this case, we attempt to acquire the lock again
1129		 * without sleeping.
1130		 */
1131		if (waiter.handoff_set) {
1132			enum owner_state owner_state;
1133
1134			preempt_disable();
1135			owner_state = rwsem_spin_on_owner(sem);
1136			preempt_enable();
1137
1138			if (owner_state == OWNER_NULL)
1139				goto trylock_again;
1140		}
1141
1142		schedule();
1143		lockevent_inc(rwsem_sleep_writer);
1144		set_current_state(state);
1145trylock_again:
1146		raw_spin_lock_irq(&sem->wait_lock);
1147	}
1148	__set_current_state(TASK_RUNNING);
1149	raw_spin_unlock_irq(&sem->wait_lock);
1150	lockevent_inc(rwsem_wlock);
1151	return sem;
1152
1153out_nolock:
1154	__set_current_state(TASK_RUNNING);
1155	raw_spin_lock_irq(&sem->wait_lock);
1156	rwsem_del_waiter(sem, &waiter);
1157	if (!list_empty(&sem->wait_list))
1158		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1159	raw_spin_unlock_irq(&sem->wait_lock);
1160	wake_up_q(&wake_q);
1161	lockevent_inc(rwsem_wlock_fail);
1162	return ERR_PTR(-EINTR);
1163}
1164
1165/*
1166 * handle waking up a waiter on the semaphore
1167 * - up_read/up_write has decremented the active part of count if we come here
1168 */
1169static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1170{
1171	unsigned long flags;
1172	DEFINE_WAKE_Q(wake_q);
1173
1174	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1175
1176	if (!list_empty(&sem->wait_list))
1177		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1178
1179	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1180	wake_up_q(&wake_q);
1181
1182	return sem;
1183}
1184
1185/*
1186 * downgrade a write lock into a read lock
1187 * - caller incremented waiting part of count and discovered it still negative
1188 * - just wake up any readers at the front of the queue
1189 */
1190static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1191{
1192	unsigned long flags;
1193	DEFINE_WAKE_Q(wake_q);
1194
1195	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1196
1197	if (!list_empty(&sem->wait_list))
1198		rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1199
1200	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1201	wake_up_q(&wake_q);
1202
1203	return sem;
1204}
1205
1206/*
1207 * lock for reading
1208 */
1209static inline int __down_read_common(struct rw_semaphore *sem, int state)
1210{
1211	long count;
1212
1213	if (!rwsem_read_trylock(sem, &count)) {
1214		if (IS_ERR(rwsem_down_read_slowpath(sem, count, state)))
1215			return -EINTR;
1216		DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1217	}
1218	return 0;
1219}
1220
1221static inline void __down_read(struct rw_semaphore *sem)
1222{
1223	__down_read_common(sem, TASK_UNINTERRUPTIBLE);
1224}
1225
1226static inline int __down_read_interruptible(struct rw_semaphore *sem)
1227{
1228	return __down_read_common(sem, TASK_INTERRUPTIBLE);
1229}
1230
1231static inline int __down_read_killable(struct rw_semaphore *sem)
1232{
1233	return __down_read_common(sem, TASK_KILLABLE);
1234}
1235
1236static inline int __down_read_trylock(struct rw_semaphore *sem)
1237{
1238	long tmp;
1239
1240	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1241
1242	tmp = atomic_long_read(&sem->count);
1243	while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1244		if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1245						    tmp + RWSEM_READER_BIAS)) {
1246			rwsem_set_reader_owned(sem);
1247			return 1;
1248		}
1249	}
1250	return 0;
1251}
1252
1253/*
1254 * lock for writing
1255 */
1256static inline int __down_write_common(struct rw_semaphore *sem, int state)
1257{
1258	if (unlikely(!rwsem_write_trylock(sem))) {
1259		if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
1260			return -EINTR;
1261	}
1262
1263	return 0;
1264}
1265
1266static inline void __down_write(struct rw_semaphore *sem)
1267{
1268	__down_write_common(sem, TASK_UNINTERRUPTIBLE);
1269}
1270
1271static inline int __down_write_killable(struct rw_semaphore *sem)
1272{
1273	return __down_write_common(sem, TASK_KILLABLE);
1274}
1275
1276static inline int __down_write_trylock(struct rw_semaphore *sem)
1277{
1278	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1279	return rwsem_write_trylock(sem);
1280}
1281
1282/*
1283 * unlock after reading
1284 */
1285static inline void __up_read(struct rw_semaphore *sem)
1286{
1287	long tmp;
1288
1289	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1290	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1291
1292	rwsem_clear_reader_owned(sem);
1293	tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1294	DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1295	if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1296		      RWSEM_FLAG_WAITERS)) {
1297		clear_nonspinnable(sem);
1298		rwsem_wake(sem);
1299	}
1300}
1301
1302/*
1303 * unlock after writing
1304 */
1305static inline void __up_write(struct rw_semaphore *sem)
1306{
1307	long tmp;
1308
1309	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1310	/*
1311	 * sem->owner may differ from current if the ownership is transferred
1312	 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1313	 */
1314	DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1315			    !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1316
1317	rwsem_clear_owner(sem);
1318	tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1319	if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1320		rwsem_wake(sem);
1321}
1322
1323/*
1324 * downgrade write lock to read lock
1325 */
1326static inline void __downgrade_write(struct rw_semaphore *sem)
1327{
1328	long tmp;
1329
1330	/*
1331	 * When downgrading from exclusive to shared ownership,
1332	 * anything inside the write-locked region cannot leak
1333	 * into the read side. In contrast, anything in the
1334	 * read-locked region is ok to be re-ordered into the
1335	 * write side. As such, rely on RELEASE semantics.
1336	 */
1337	DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1338	tmp = atomic_long_fetch_add_release(
1339		-RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1340	rwsem_set_reader_owned(sem);
1341	if (tmp & RWSEM_FLAG_WAITERS)
1342		rwsem_downgrade_wake(sem);
1343}
1344
1345#else /* !CONFIG_PREEMPT_RT */
1346
1347#define RT_MUTEX_BUILD_MUTEX
1348#include "rtmutex.c"
1349
1350#define rwbase_set_and_save_current_state(state)	\
1351	set_current_state(state)
1352
1353#define rwbase_restore_current_state()			\
1354	__set_current_state(TASK_RUNNING)
1355
1356#define rwbase_rtmutex_lock_state(rtm, state)		\
1357	__rt_mutex_lock(rtm, state)
1358
1359#define rwbase_rtmutex_slowlock_locked(rtm, state)	\
1360	__rt_mutex_slowlock_locked(rtm, NULL, state)
1361
1362#define rwbase_rtmutex_unlock(rtm)			\
1363	__rt_mutex_unlock(rtm)
1364
1365#define rwbase_rtmutex_trylock(rtm)			\
1366	__rt_mutex_trylock(rtm)
1367
1368#define rwbase_signal_pending_state(state, current)	\
1369	signal_pending_state(state, current)
1370
1371#define rwbase_schedule()				\
1372	schedule()
1373
1374#include "rwbase_rt.c"
1375
1376void __init_rwsem(struct rw_semaphore *sem, const char *name,
1377		  struct lock_class_key *key)
1378{
1379	init_rwbase_rt(&(sem)->rwbase);
1380
1381#ifdef CONFIG_DEBUG_LOCK_ALLOC
1382	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1383	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1384#endif
1385}
1386EXPORT_SYMBOL(__init_rwsem);
1387
1388static inline void __down_read(struct rw_semaphore *sem)
1389{
1390	rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1391}
1392
1393static inline int __down_read_interruptible(struct rw_semaphore *sem)
1394{
1395	return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1396}
1397
1398static inline int __down_read_killable(struct rw_semaphore *sem)
1399{
1400	return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1401}
1402
1403static inline int __down_read_trylock(struct rw_semaphore *sem)
1404{
1405	return rwbase_read_trylock(&sem->rwbase);
1406}
1407
1408static inline void __up_read(struct rw_semaphore *sem)
1409{
1410	rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1411}
1412
1413static inline void __sched __down_write(struct rw_semaphore *sem)
1414{
1415	rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1416}
1417
1418static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1419{
1420	return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1421}
1422
1423static inline int __down_write_trylock(struct rw_semaphore *sem)
1424{
1425	return rwbase_write_trylock(&sem->rwbase);
1426}
1427
1428static inline void __up_write(struct rw_semaphore *sem)
1429{
1430	rwbase_write_unlock(&sem->rwbase);
1431}
1432
1433static inline void __downgrade_write(struct rw_semaphore *sem)
1434{
1435	rwbase_write_downgrade(&sem->rwbase);
1436}
1437
1438/* Debug stubs for the common API */
1439#define DEBUG_RWSEMS_WARN_ON(c, sem)
1440
1441static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1442					    struct task_struct *owner)
1443{
1444}
1445
1446static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1447{
1448	int count = atomic_read(&sem->rwbase.readers);
1449
1450	return count < 0 && count != READER_BIAS;
1451}
1452
1453#endif /* CONFIG_PREEMPT_RT */
1454
1455/*
1456 * lock for reading
1457 */
1458void __sched down_read(struct rw_semaphore *sem)
1459{
1460	might_sleep();
1461	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1462
1463	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1464}
1465EXPORT_SYMBOL(down_read);
1466
1467int __sched down_read_interruptible(struct rw_semaphore *sem)
1468{
1469	might_sleep();
1470	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1471
1472	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1473		rwsem_release(&sem->dep_map, _RET_IP_);
1474		return -EINTR;
1475	}
1476
1477	return 0;
1478}
1479EXPORT_SYMBOL(down_read_interruptible);
1480
1481int __sched down_read_killable(struct rw_semaphore *sem)
1482{
1483	might_sleep();
1484	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1485
1486	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1487		rwsem_release(&sem->dep_map, _RET_IP_);
1488		return -EINTR;
1489	}
1490
1491	return 0;
1492}
1493EXPORT_SYMBOL(down_read_killable);
1494
1495/*
1496 * trylock for reading -- returns 1 if successful, 0 if contention
1497 */
1498int down_read_trylock(struct rw_semaphore *sem)
1499{
1500	int ret = __down_read_trylock(sem);
1501
1502	if (ret == 1)
1503		rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1504	return ret;
1505}
1506EXPORT_SYMBOL(down_read_trylock);
1507
1508/*
1509 * lock for writing
1510 */
1511void __sched down_write(struct rw_semaphore *sem)
1512{
1513	might_sleep();
1514	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1515	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1516}
1517EXPORT_SYMBOL(down_write);
1518
1519/*
1520 * lock for writing
1521 */
1522int __sched down_write_killable(struct rw_semaphore *sem)
1523{
1524	might_sleep();
1525	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1526
1527	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1528				  __down_write_killable)) {
1529		rwsem_release(&sem->dep_map, _RET_IP_);
1530		return -EINTR;
1531	}
1532
1533	return 0;
1534}
1535EXPORT_SYMBOL(down_write_killable);
1536
1537/*
1538 * trylock for writing -- returns 1 if successful, 0 if contention
1539 */
1540int down_write_trylock(struct rw_semaphore *sem)
1541{
1542	int ret = __down_write_trylock(sem);
1543
1544	if (ret == 1)
1545		rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1546
1547	return ret;
1548}
1549EXPORT_SYMBOL(down_write_trylock);
1550
1551/*
1552 * release a read lock
1553 */
1554void up_read(struct rw_semaphore *sem)
1555{
1556	rwsem_release(&sem->dep_map, _RET_IP_);
1557	__up_read(sem);
1558}
1559EXPORT_SYMBOL(up_read);
1560
1561/*
1562 * release a write lock
1563 */
1564void up_write(struct rw_semaphore *sem)
1565{
1566	rwsem_release(&sem->dep_map, _RET_IP_);
1567	__up_write(sem);
1568}
1569EXPORT_SYMBOL(up_write);
1570
1571/*
1572 * downgrade write lock to read lock
1573 */
1574void downgrade_write(struct rw_semaphore *sem)
1575{
1576	lock_downgrade(&sem->dep_map, _RET_IP_);
1577	__downgrade_write(sem);
1578}
1579EXPORT_SYMBOL(downgrade_write);
1580
1581#ifdef CONFIG_DEBUG_LOCK_ALLOC
1582
1583void down_read_nested(struct rw_semaphore *sem, int subclass)
1584{
1585	might_sleep();
1586	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1587	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1588}
1589EXPORT_SYMBOL(down_read_nested);
1590
1591int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1592{
1593	might_sleep();
1594	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1595
1596	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1597		rwsem_release(&sem->dep_map, _RET_IP_);
1598		return -EINTR;
1599	}
1600
1601	return 0;
1602}
1603EXPORT_SYMBOL(down_read_killable_nested);
1604
1605void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1606{
1607	might_sleep();
1608	rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1609	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1610}
1611EXPORT_SYMBOL(_down_write_nest_lock);
1612
1613void down_read_non_owner(struct rw_semaphore *sem)
1614{
1615	might_sleep();
1616	__down_read(sem);
1617	__rwsem_set_reader_owned(sem, NULL);
1618}
1619EXPORT_SYMBOL(down_read_non_owner);
1620
1621void down_write_nested(struct rw_semaphore *sem, int subclass)
1622{
1623	might_sleep();
1624	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1625	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1626}
1627EXPORT_SYMBOL(down_write_nested);
1628
1629int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1630{
1631	might_sleep();
1632	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1633
1634	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1635				  __down_write_killable)) {
1636		rwsem_release(&sem->dep_map, _RET_IP_);
1637		return -EINTR;
1638	}
1639
1640	return 0;
1641}
1642EXPORT_SYMBOL(down_write_killable_nested);
1643
1644void up_read_non_owner(struct rw_semaphore *sem)
1645{
1646	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1647	__up_read(sem);
1648}
1649EXPORT_SYMBOL(up_read_non_owner);
1650
1651#endif
1652