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#include <trace/events/lock.h>
31
32#ifndef CONFIG_PREEMPT_RT
33#include "lock_events.h"
34
35/*
36 * The least significant 2 bits of the owner value has the following
37 * meanings when set.
38 *  - Bit 0: RWSEM_READER_OWNED - The rwsem is owned by readers
39 *  - Bit 1: RWSEM_NONSPINNABLE - Cannot spin on a reader-owned lock
40 *
41 * When the rwsem is reader-owned and a spinning writer has timed out,
42 * the nonspinnable bit will be set to disable optimistic spinning.
43
44 * When a writer acquires a rwsem, it puts its task_struct pointer
45 * into the owner field. It is cleared after an unlock.
46 *
47 * When a reader acquires a rwsem, it will also puts its task_struct
48 * pointer into the owner field with the RWSEM_READER_OWNED bit set.
49 * On unlock, the owner field will largely be left untouched. So
50 * for a free or reader-owned rwsem, the owner value may contain
51 * information about the last reader that acquires the rwsem.
52 *
53 * That information may be helpful in debugging cases where the system
54 * seems to hang on a reader owned rwsem especially if only one reader
55 * is involved. Ideally we would like to track all the readers that own
56 * a rwsem, but the overhead is simply too big.
57 *
58 * A fast path reader optimistic lock stealing is supported when the rwsem
59 * is previously owned by a writer and the following conditions are met:
60 *  - rwsem is not currently writer owned
61 *  - the handoff isn't set.
62 */
63#define RWSEM_READER_OWNED	(1UL << 0)
64#define RWSEM_NONSPINNABLE	(1UL << 1)
65#define RWSEM_OWNER_FLAGS_MASK	(RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
66
67#ifdef CONFIG_DEBUG_RWSEMS
68# define DEBUG_RWSEMS_WARN_ON(c, sem)	do {			\
69	if (!debug_locks_silent &&				\
70	    WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
71		#c, atomic_long_read(&(sem)->count),		\
72		(unsigned long) sem->magic,			\
73		atomic_long_read(&(sem)->owner), (long)current,	\
74		list_empty(&(sem)->wait_list) ? "" : "not "))	\
75			debug_locks_off();			\
76	} while (0)
77#else
78# define DEBUG_RWSEMS_WARN_ON(c, sem)
79#endif
80
81/*
82 * On 64-bit architectures, the bit definitions of the count are:
83 *
84 * Bit  0    - writer locked bit
85 * Bit  1    - waiters present bit
86 * Bit  2    - lock handoff bit
87 * Bits 3-7  - reserved
88 * Bits 8-62 - 55-bit reader count
89 * Bit  63   - read fail bit
90 *
91 * On 32-bit architectures, the bit definitions of the count are:
92 *
93 * Bit  0    - writer locked bit
94 * Bit  1    - waiters present bit
95 * Bit  2    - lock handoff bit
96 * Bits 3-7  - reserved
97 * Bits 8-30 - 23-bit reader count
98 * Bit  31   - read fail bit
99 *
100 * It is not likely that the most significant bit (read fail bit) will ever
101 * be set. This guard bit is still checked anyway in the down_read() fastpath
102 * just in case we need to use up more of the reader bits for other purpose
103 * in the future.
104 *
105 * atomic_long_fetch_add() is used to obtain reader lock, whereas
106 * atomic_long_cmpxchg() will be used to obtain writer lock.
107 *
108 * There are three places where the lock handoff bit may be set or cleared.
109 * 1) rwsem_mark_wake() for readers		-- set, clear
110 * 2) rwsem_try_write_lock() for writers	-- set, clear
111 * 3) rwsem_del_waiter()			-- clear
112 *
113 * For all the above cases, wait_lock will be held. A writer must also
114 * be the first one in the wait_list to be eligible for setting the handoff
115 * bit. So concurrent setting/clearing of handoff bit is not possible.
116 */
117#define RWSEM_WRITER_LOCKED	(1UL << 0)
118#define RWSEM_FLAG_WAITERS	(1UL << 1)
119#define RWSEM_FLAG_HANDOFF	(1UL << 2)
120#define RWSEM_FLAG_READFAIL	(1UL << (BITS_PER_LONG - 1))
121
122#define RWSEM_READER_SHIFT	8
123#define RWSEM_READER_BIAS	(1UL << RWSEM_READER_SHIFT)
124#define RWSEM_READER_MASK	(~(RWSEM_READER_BIAS - 1))
125#define RWSEM_WRITER_MASK	RWSEM_WRITER_LOCKED
126#define RWSEM_LOCK_MASK		(RWSEM_WRITER_MASK|RWSEM_READER_MASK)
127#define RWSEM_READ_FAILED_MASK	(RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
128				 RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
129
130/*
131 * All writes to owner are protected by WRITE_ONCE() to make sure that
132 * store tearing can't happen as optimistic spinners may read and use
133 * the owner value concurrently without lock. Read from owner, however,
134 * may not need READ_ONCE() as long as the pointer value is only used
135 * for comparison and isn't being dereferenced.
136 *
137 * Both rwsem_{set,clear}_owner() functions should be in the same
138 * preempt disable section as the atomic op that changes sem->count.
139 */
140static inline void rwsem_set_owner(struct rw_semaphore *sem)
141{
142	lockdep_assert_preemption_disabled();
143	atomic_long_set(&sem->owner, (long)current);
144}
145
146static inline void rwsem_clear_owner(struct rw_semaphore *sem)
147{
148	lockdep_assert_preemption_disabled();
149	atomic_long_set(&sem->owner, 0);
150}
151
152/*
153 * Test the flags in the owner field.
154 */
155static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
156{
157	return atomic_long_read(&sem->owner) & flags;
158}
159
160/*
161 * The task_struct pointer of the last owning reader will be left in
162 * the owner field.
163 *
164 * Note that the owner value just indicates the task has owned the rwsem
165 * previously, it may not be the real owner or one of the real owners
166 * anymore when that field is examined, so take it with a grain of salt.
167 *
168 * The reader non-spinnable bit is preserved.
169 */
170static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
171					    struct task_struct *owner)
172{
173	unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
174		(atomic_long_read(&sem->owner) & RWSEM_NONSPINNABLE);
175
176	atomic_long_set(&sem->owner, val);
177}
178
179static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
180{
181	__rwsem_set_reader_owned(sem, current);
182}
183
184/*
185 * Return true if the rwsem is owned by a reader.
186 */
187static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
188{
189#ifdef CONFIG_DEBUG_RWSEMS
190	/*
191	 * Check the count to see if it is write-locked.
192	 */
193	long count = atomic_long_read(&sem->count);
194
195	if (count & RWSEM_WRITER_MASK)
196		return false;
197#endif
198	return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
199}
200
201#ifdef CONFIG_DEBUG_RWSEMS
202/*
203 * With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
204 * is a task pointer in owner of a reader-owned rwsem, it will be the
205 * real owner or one of the real owners. The only exception is when the
206 * unlock is done by up_read_non_owner().
207 */
208static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
209{
210	unsigned long val = atomic_long_read(&sem->owner);
211
212	while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
213		if (atomic_long_try_cmpxchg(&sem->owner, &val,
214					    val & RWSEM_OWNER_FLAGS_MASK))
215			return;
216	}
217}
218#else
219static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
220{
221}
222#endif
223
224/*
225 * Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
226 * remains set. Otherwise, the operation will be aborted.
227 */
228static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
229{
230	unsigned long owner = atomic_long_read(&sem->owner);
231
232	do {
233		if (!(owner & RWSEM_READER_OWNED))
234			break;
235		if (owner & RWSEM_NONSPINNABLE)
236			break;
237	} while (!atomic_long_try_cmpxchg(&sem->owner, &owner,
238					  owner | RWSEM_NONSPINNABLE));
239}
240
241static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cntp)
242{
243	*cntp = atomic_long_add_return_acquire(RWSEM_READER_BIAS, &sem->count);
244
245	if (WARN_ON_ONCE(*cntp < 0))
246		rwsem_set_nonspinnable(sem);
247
248	if (!(*cntp & RWSEM_READ_FAILED_MASK)) {
249		rwsem_set_reader_owned(sem);
250		return true;
251	}
252
253	return false;
254}
255
256static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
257{
258	long tmp = RWSEM_UNLOCKED_VALUE;
259	bool ret = false;
260
261	preempt_disable();
262	if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp, RWSEM_WRITER_LOCKED)) {
263		rwsem_set_owner(sem);
264		ret = true;
265	}
266
267	preempt_enable();
268	return ret;
269}
270
271/*
272 * Return just the real task structure pointer of the owner
273 */
274static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
275{
276	return (struct task_struct *)
277		(atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
278}
279
280/*
281 * Return the real task structure pointer of the owner and the embedded
282 * flags in the owner. pflags must be non-NULL.
283 */
284static inline struct task_struct *
285rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
286{
287	unsigned long owner = atomic_long_read(&sem->owner);
288
289	*pflags = owner & RWSEM_OWNER_FLAGS_MASK;
290	return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
291}
292
293/*
294 * Guide to the rw_semaphore's count field.
295 *
296 * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
297 * by a writer.
298 *
299 * The lock is owned by readers when
300 * (1) the RWSEM_WRITER_LOCKED isn't set in count,
301 * (2) some of the reader bits are set in count, and
302 * (3) the owner field has RWSEM_READ_OWNED bit set.
303 *
304 * Having some reader bits set is not enough to guarantee a readers owned
305 * lock as the readers may be in the process of backing out from the count
306 * and a writer has just released the lock. So another writer may steal
307 * the lock immediately after that.
308 */
309
310/*
311 * Initialize an rwsem:
312 */
313void __init_rwsem(struct rw_semaphore *sem, const char *name,
314		  struct lock_class_key *key)
315{
316#ifdef CONFIG_DEBUG_LOCK_ALLOC
317	/*
318	 * Make sure we are not reinitializing a held semaphore:
319	 */
320	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
321	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
322#endif
323#ifdef CONFIG_DEBUG_RWSEMS
324	sem->magic = sem;
325#endif
326	atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
327	raw_spin_lock_init(&sem->wait_lock);
328	INIT_LIST_HEAD(&sem->wait_list);
329	atomic_long_set(&sem->owner, 0L);
330#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
331	osq_lock_init(&sem->osq);
332#endif
333}
334EXPORT_SYMBOL(__init_rwsem);
335
336enum rwsem_waiter_type {
337	RWSEM_WAITING_FOR_WRITE,
338	RWSEM_WAITING_FOR_READ
339};
340
341struct rwsem_waiter {
342	struct list_head list;
343	struct task_struct *task;
344	enum rwsem_waiter_type type;
345	unsigned long timeout;
346	bool handoff_set;
347};
348#define rwsem_first_waiter(sem) \
349	list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
350
351enum rwsem_wake_type {
352	RWSEM_WAKE_ANY,		/* Wake whatever's at head of wait list */
353	RWSEM_WAKE_READERS,	/* Wake readers only */
354	RWSEM_WAKE_READ_OWNED	/* Waker thread holds the read lock */
355};
356
357/*
358 * The typical HZ value is either 250 or 1000. So set the minimum waiting
359 * time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
360 * queue before initiating the handoff protocol.
361 */
362#define RWSEM_WAIT_TIMEOUT	DIV_ROUND_UP(HZ, 250)
363
364/*
365 * Magic number to batch-wakeup waiting readers, even when writers are
366 * also present in the queue. This both limits the amount of work the
367 * waking thread must do and also prevents any potential counter overflow,
368 * however unlikely.
369 */
370#define MAX_READERS_WAKEUP	0x100
371
372static inline void
373rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
374{
375	lockdep_assert_held(&sem->wait_lock);
376	list_add_tail(&waiter->list, &sem->wait_list);
377	/* caller will set RWSEM_FLAG_WAITERS */
378}
379
380/*
381 * Remove a waiter from the wait_list and clear flags.
382 *
383 * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
384 * this function. Modify with care.
385 *
386 * Return: true if wait_list isn't empty and false otherwise
387 */
388static inline bool
389rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
390{
391	lockdep_assert_held(&sem->wait_lock);
392	list_del(&waiter->list);
393	if (likely(!list_empty(&sem->wait_list)))
394		return true;
395
396	atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, &sem->count);
397	return false;
398}
399
400/*
401 * handle the lock release when processes blocked on it that can now run
402 * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
403 *   have been set.
404 * - there must be someone on the queue
405 * - the wait_lock must be held by the caller
406 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
407 *   to actually wakeup the blocked task(s) and drop the reference count,
408 *   preferably when the wait_lock is released
409 * - woken process blocks are discarded from the list after having task zeroed
410 * - writers are only marked woken if downgrading is false
411 *
412 * Implies rwsem_del_waiter() for all woken readers.
413 */
414static void rwsem_mark_wake(struct rw_semaphore *sem,
415			    enum rwsem_wake_type wake_type,
416			    struct wake_q_head *wake_q)
417{
418	struct rwsem_waiter *waiter, *tmp;
419	long oldcount, woken = 0, adjustment = 0;
420	struct list_head wlist;
421
422	lockdep_assert_held(&sem->wait_lock);
423
424	/*
425	 * Take a peek at the queue head waiter such that we can determine
426	 * the wakeup(s) to perform.
427	 */
428	waiter = rwsem_first_waiter(sem);
429
430	if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
431		if (wake_type == RWSEM_WAKE_ANY) {
432			/*
433			 * Mark writer at the front of the queue for wakeup.
434			 * Until the task is actually later awoken later by
435			 * the caller, other writers are able to steal it.
436			 * Readers, on the other hand, will block as they
437			 * will notice the queued writer.
438			 */
439			wake_q_add(wake_q, waiter->task);
440			lockevent_inc(rwsem_wake_writer);
441		}
442
443		return;
444	}
445
446	/*
447	 * No reader wakeup if there are too many of them already.
448	 */
449	if (unlikely(atomic_long_read(&sem->count) < 0))
450		return;
451
452	/*
453	 * Writers might steal the lock before we grant it to the next reader.
454	 * We prefer to do the first reader grant before counting readers
455	 * so we can bail out early if a writer stole the lock.
456	 */
457	if (wake_type != RWSEM_WAKE_READ_OWNED) {
458		struct task_struct *owner;
459
460		adjustment = RWSEM_READER_BIAS;
461		oldcount = atomic_long_fetch_add(adjustment, &sem->count);
462		if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
463			/*
464			 * When we've been waiting "too" long (for writers
465			 * to give up the lock), request a HANDOFF to
466			 * force the issue.
467			 */
468			if (time_after(jiffies, waiter->timeout)) {
469				if (!(oldcount & RWSEM_FLAG_HANDOFF)) {
470					adjustment -= RWSEM_FLAG_HANDOFF;
471					lockevent_inc(rwsem_rlock_handoff);
472				}
473				waiter->handoff_set = true;
474			}
475
476			atomic_long_add(-adjustment, &sem->count);
477			return;
478		}
479		/*
480		 * Set it to reader-owned to give spinners an early
481		 * indication that readers now have the lock.
482		 * The reader nonspinnable bit seen at slowpath entry of
483		 * the reader is copied over.
484		 */
485		owner = waiter->task;
486		__rwsem_set_reader_owned(sem, owner);
487	}
488
489	/*
490	 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
491	 * queue. We know that the woken will be at least 1 as we accounted
492	 * for above. Note we increment the 'active part' of the count by the
493	 * number of readers before waking any processes up.
494	 *
495	 * This is an adaptation of the phase-fair R/W locks where at the
496	 * reader phase (first waiter is a reader), all readers are eligible
497	 * to acquire the lock at the same time irrespective of their order
498	 * in the queue. The writers acquire the lock according to their
499	 * order in the queue.
500	 *
501	 * We have to do wakeup in 2 passes to prevent the possibility that
502	 * the reader count may be decremented before it is incremented. It
503	 * is because the to-be-woken waiter may not have slept yet. So it
504	 * may see waiter->task got cleared, finish its critical section and
505	 * do an unlock before the reader count increment.
506	 *
507	 * 1) Collect the read-waiters in a separate list, count them and
508	 *    fully increment the reader count in rwsem.
509	 * 2) For each waiters in the new list, clear waiter->task and
510	 *    put them into wake_q to be woken up later.
511	 */
512	INIT_LIST_HEAD(&wlist);
513	list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
514		if (waiter->type == RWSEM_WAITING_FOR_WRITE)
515			continue;
516
517		woken++;
518		list_move_tail(&waiter->list, &wlist);
519
520		/*
521		 * Limit # of readers that can be woken up per wakeup call.
522		 */
523		if (unlikely(woken >= MAX_READERS_WAKEUP))
524			break;
525	}
526
527	adjustment = woken * RWSEM_READER_BIAS - adjustment;
528	lockevent_cond_inc(rwsem_wake_reader, woken);
529
530	oldcount = atomic_long_read(&sem->count);
531	if (list_empty(&sem->wait_list)) {
532		/*
533		 * Combined with list_move_tail() above, this implies
534		 * rwsem_del_waiter().
535		 */
536		adjustment -= RWSEM_FLAG_WAITERS;
537		if (oldcount & RWSEM_FLAG_HANDOFF)
538			adjustment -= RWSEM_FLAG_HANDOFF;
539	} else if (woken) {
540		/*
541		 * When we've woken a reader, we no longer need to force
542		 * writers to give up the lock and we can clear HANDOFF.
543		 */
544		if (oldcount & RWSEM_FLAG_HANDOFF)
545			adjustment -= RWSEM_FLAG_HANDOFF;
546	}
547
548	if (adjustment)
549		atomic_long_add(adjustment, &sem->count);
550
551	/* 2nd pass */
552	list_for_each_entry_safe(waiter, tmp, &wlist, list) {
553		struct task_struct *tsk;
554
555		tsk = waiter->task;
556		get_task_struct(tsk);
557
558		/*
559		 * Ensure calling get_task_struct() before setting the reader
560		 * waiter to nil such that rwsem_down_read_slowpath() cannot
561		 * race with do_exit() by always holding a reference count
562		 * to the task to wakeup.
563		 */
564		smp_store_release(&waiter->task, NULL);
565		/*
566		 * Ensure issuing the wakeup (either by us or someone else)
567		 * after setting the reader waiter to nil.
568		 */
569		wake_q_add_safe(wake_q, tsk);
570	}
571}
572
573/*
574 * Remove a waiter and try to wake up other waiters in the wait queue
575 * This function is called from the out_nolock path of both the reader and
576 * writer slowpaths with wait_lock held. It releases the wait_lock and
577 * optionally wake up waiters before it returns.
578 */
579static inline void
580rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,
581		      struct wake_q_head *wake_q)
582		      __releases(&sem->wait_lock)
583{
584	bool first = rwsem_first_waiter(sem) == waiter;
585
586	wake_q_init(wake_q);
587
588	/*
589	 * If the wait_list isn't empty and the waiter to be deleted is
590	 * the first waiter, we wake up the remaining waiters as they may
591	 * be eligible to acquire or spin on the lock.
592	 */
593	if (rwsem_del_waiter(sem, waiter) && first)
594		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, wake_q);
595	raw_spin_unlock_irq(&sem->wait_lock);
596	if (!wake_q_empty(wake_q))
597		wake_up_q(wake_q);
598}
599
600/*
601 * This function must be called with the sem->wait_lock held to prevent
602 * race conditions between checking the rwsem wait list and setting the
603 * sem->count accordingly.
604 *
605 * Implies rwsem_del_waiter() on success.
606 */
607static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
608					struct rwsem_waiter *waiter)
609{
610	struct rwsem_waiter *first = rwsem_first_waiter(sem);
611	long count, new;
612
613	lockdep_assert_held(&sem->wait_lock);
614
615	count = atomic_long_read(&sem->count);
616	do {
617		bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
618
619		if (has_handoff) {
620			/*
621			 * Honor handoff bit and yield only when the first
622			 * waiter is the one that set it. Otherwisee, we
623			 * still try to acquire the rwsem.
624			 */
625			if (first->handoff_set && (waiter != first))
626				return false;
627
628			/*
629			 * First waiter can inherit a previously set handoff
630			 * bit and spin on rwsem if lock acquisition fails.
631			 */
632			if (waiter == first)
633				waiter->handoff_set = true;
634		}
635
636		new = count;
637
638		if (count & RWSEM_LOCK_MASK) {
639			if (has_handoff || (!rt_task(waiter->task) &&
640					    !time_after(jiffies, waiter->timeout)))
641				return false;
642
643			new |= RWSEM_FLAG_HANDOFF;
644		} else {
645			new |= RWSEM_WRITER_LOCKED;
646			new &= ~RWSEM_FLAG_HANDOFF;
647
648			if (list_is_singular(&sem->wait_list))
649				new &= ~RWSEM_FLAG_WAITERS;
650		}
651	} while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
652
653	/*
654	 * We have either acquired the lock with handoff bit cleared or
655	 * set the handoff bit.
656	 */
657	if (new & RWSEM_FLAG_HANDOFF) {
658		waiter->handoff_set = true;
659		lockevent_inc(rwsem_wlock_handoff);
660		return false;
661	}
662
663	/*
664	 * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
665	 * success.
666	 */
667	list_del(&waiter->list);
668	rwsem_set_owner(sem);
669	return true;
670}
671
672/*
673 * The rwsem_spin_on_owner() function returns the following 4 values
674 * depending on the lock owner state.
675 *   OWNER_NULL  : owner is currently NULL
676 *   OWNER_WRITER: when owner changes and is a writer
677 *   OWNER_READER: when owner changes and the new owner may be a reader.
678 *   OWNER_NONSPINNABLE:
679 *		   when optimistic spinning has to stop because either the
680 *		   owner stops running, is unknown, or its timeslice has
681 *		   been used up.
682 */
683enum owner_state {
684	OWNER_NULL		= 1 << 0,
685	OWNER_WRITER		= 1 << 1,
686	OWNER_READER		= 1 << 2,
687	OWNER_NONSPINNABLE	= 1 << 3,
688};
689
690#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
691/*
692 * Try to acquire write lock before the writer has been put on wait queue.
693 */
694static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
695{
696	long count = atomic_long_read(&sem->count);
697
698	while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
699		if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
700					count | RWSEM_WRITER_LOCKED)) {
701			rwsem_set_owner(sem);
702			lockevent_inc(rwsem_opt_lock);
703			return true;
704		}
705	}
706	return false;
707}
708
709static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
710{
711	struct task_struct *owner;
712	unsigned long flags;
713	bool ret = true;
714
715	if (need_resched()) {
716		lockevent_inc(rwsem_opt_fail);
717		return false;
718	}
719
720	preempt_disable();
721	/*
722	 * Disable preemption is equal to the RCU read-side crital section,
723	 * thus the task_strcut structure won't go away.
724	 */
725	owner = rwsem_owner_flags(sem, &flags);
726	/*
727	 * Don't check the read-owner as the entry may be stale.
728	 */
729	if ((flags & RWSEM_NONSPINNABLE) ||
730	    (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
731		ret = false;
732	preempt_enable();
733
734	lockevent_cond_inc(rwsem_opt_fail, !ret);
735	return ret;
736}
737
738#define OWNER_SPINNABLE		(OWNER_NULL | OWNER_WRITER | OWNER_READER)
739
740static inline enum owner_state
741rwsem_owner_state(struct task_struct *owner, unsigned long flags)
742{
743	if (flags & RWSEM_NONSPINNABLE)
744		return OWNER_NONSPINNABLE;
745
746	if (flags & RWSEM_READER_OWNED)
747		return OWNER_READER;
748
749	return owner ? OWNER_WRITER : OWNER_NULL;
750}
751
752static noinline enum owner_state
753rwsem_spin_on_owner(struct rw_semaphore *sem)
754{
755	struct task_struct *new, *owner;
756	unsigned long flags, new_flags;
757	enum owner_state state;
758
759	lockdep_assert_preemption_disabled();
760
761	owner = rwsem_owner_flags(sem, &flags);
762	state = rwsem_owner_state(owner, flags);
763	if (state != OWNER_WRITER)
764		return state;
765
766	for (;;) {
767		/*
768		 * When a waiting writer set the handoff flag, it may spin
769		 * on the owner as well. Once that writer acquires the lock,
770		 * we can spin on it. So we don't need to quit even when the
771		 * handoff bit is set.
772		 */
773		new = rwsem_owner_flags(sem, &new_flags);
774		if ((new != owner) || (new_flags != flags)) {
775			state = rwsem_owner_state(new, new_flags);
776			break;
777		}
778
779		/*
780		 * Ensure we emit the owner->on_cpu, dereference _after_
781		 * checking sem->owner still matches owner, if that fails,
782		 * owner might point to free()d memory, if it still matches,
783		 * our spinning context already disabled preemption which is
784		 * equal to RCU read-side crital section ensures the memory
785		 * stays valid.
786		 */
787		barrier();
788
789		if (need_resched() || !owner_on_cpu(owner)) {
790			state = OWNER_NONSPINNABLE;
791			break;
792		}
793
794		cpu_relax();
795	}
796
797	return state;
798}
799
800/*
801 * Calculate reader-owned rwsem spinning threshold for writer
802 *
803 * The more readers own the rwsem, the longer it will take for them to
804 * wind down and free the rwsem. So the empirical formula used to
805 * determine the actual spinning time limit here is:
806 *
807 *   Spinning threshold = (10 + nr_readers/2)us
808 *
809 * The limit is capped to a maximum of 25us (30 readers). This is just
810 * a heuristic and is subjected to change in the future.
811 */
812static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
813{
814	long count = atomic_long_read(&sem->count);
815	int readers = count >> RWSEM_READER_SHIFT;
816	u64 delta;
817
818	if (readers > 30)
819		readers = 30;
820	delta = (20 + readers) * NSEC_PER_USEC / 2;
821
822	return sched_clock() + delta;
823}
824
825static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
826{
827	bool taken = false;
828	int prev_owner_state = OWNER_NULL;
829	int loop = 0;
830	u64 rspin_threshold = 0;
831
832	preempt_disable();
833
834	/* sem->wait_lock should not be held when doing optimistic spinning */
835	if (!osq_lock(&sem->osq))
836		goto done;
837
838	/*
839	 * Optimistically spin on the owner field and attempt to acquire the
840	 * lock whenever the owner changes. Spinning will be stopped when:
841	 *  1) the owning writer isn't running; or
842	 *  2) readers own the lock and spinning time has exceeded limit.
843	 */
844	for (;;) {
845		enum owner_state owner_state;
846
847		owner_state = rwsem_spin_on_owner(sem);
848		if (!(owner_state & OWNER_SPINNABLE))
849			break;
850
851		/*
852		 * Try to acquire the lock
853		 */
854		taken = rwsem_try_write_lock_unqueued(sem);
855
856		if (taken)
857			break;
858
859		/*
860		 * Time-based reader-owned rwsem optimistic spinning
861		 */
862		if (owner_state == OWNER_READER) {
863			/*
864			 * Re-initialize rspin_threshold every time when
865			 * the owner state changes from non-reader to reader.
866			 * This allows a writer to steal the lock in between
867			 * 2 reader phases and have the threshold reset at
868			 * the beginning of the 2nd reader phase.
869			 */
870			if (prev_owner_state != OWNER_READER) {
871				if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
872					break;
873				rspin_threshold = rwsem_rspin_threshold(sem);
874				loop = 0;
875			}
876
877			/*
878			 * Check time threshold once every 16 iterations to
879			 * avoid calling sched_clock() too frequently so
880			 * as to reduce the average latency between the times
881			 * when the lock becomes free and when the spinner
882			 * is ready to do a trylock.
883			 */
884			else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
885				rwsem_set_nonspinnable(sem);
886				lockevent_inc(rwsem_opt_nospin);
887				break;
888			}
889		}
890
891		/*
892		 * An RT task cannot do optimistic spinning if it cannot
893		 * be sure the lock holder is running or live-lock may
894		 * happen if the current task and the lock holder happen
895		 * to run in the same CPU. However, aborting optimistic
896		 * spinning while a NULL owner is detected may miss some
897		 * opportunity where spinning can continue without causing
898		 * problem.
899		 *
900		 * There are 2 possible cases where an RT task may be able
901		 * to continue spinning.
902		 *
903		 * 1) The lock owner is in the process of releasing the
904		 *    lock, sem->owner is cleared but the lock has not
905		 *    been released yet.
906		 * 2) The lock was free and owner cleared, but another
907		 *    task just comes in and acquire the lock before
908		 *    we try to get it. The new owner may be a spinnable
909		 *    writer.
910		 *
911		 * To take advantage of two scenarios listed above, the RT
912		 * task is made to retry one more time to see if it can
913		 * acquire the lock or continue spinning on the new owning
914		 * writer. Of course, if the time lag is long enough or the
915		 * new owner is not a writer or spinnable, the RT task will
916		 * quit spinning.
917		 *
918		 * If the owner is a writer, the need_resched() check is
919		 * done inside rwsem_spin_on_owner(). If the owner is not
920		 * a writer, need_resched() check needs to be done here.
921		 */
922		if (owner_state != OWNER_WRITER) {
923			if (need_resched())
924				break;
925			if (rt_task(current) &&
926			   (prev_owner_state != OWNER_WRITER))
927				break;
928		}
929		prev_owner_state = owner_state;
930
931		/*
932		 * The cpu_relax() call is a compiler barrier which forces
933		 * everything in this loop to be re-loaded. We don't need
934		 * memory barriers as we'll eventually observe the right
935		 * values at the cost of a few extra spins.
936		 */
937		cpu_relax();
938	}
939	osq_unlock(&sem->osq);
940done:
941	preempt_enable();
942	lockevent_cond_inc(rwsem_opt_fail, !taken);
943	return taken;
944}
945
946/*
947 * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
948 * only be called when the reader count reaches 0.
949 */
950static inline void clear_nonspinnable(struct rw_semaphore *sem)
951{
952	if (unlikely(rwsem_test_oflags(sem, RWSEM_NONSPINNABLE)))
953		atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
954}
955
956#else
957static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
958{
959	return false;
960}
961
962static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
963{
964	return false;
965}
966
967static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
968
969static inline enum owner_state
970rwsem_spin_on_owner(struct rw_semaphore *sem)
971{
972	return OWNER_NONSPINNABLE;
973}
974#endif
975
976/*
977 * Prepare to wake up waiter(s) in the wait queue by putting them into the
978 * given wake_q if the rwsem lock owner isn't a writer. If rwsem is likely
979 * reader-owned, wake up read lock waiters in queue front or wake up any
980 * front waiter otherwise.
981
982 * This is being called from both reader and writer slow paths.
983 */
984static inline void rwsem_cond_wake_waiter(struct rw_semaphore *sem, long count,
985					  struct wake_q_head *wake_q)
986{
987	enum rwsem_wake_type wake_type;
988
989	if (count & RWSEM_WRITER_MASK)
990		return;
991
992	if (count & RWSEM_READER_MASK) {
993		wake_type = RWSEM_WAKE_READERS;
994	} else {
995		wake_type = RWSEM_WAKE_ANY;
996		clear_nonspinnable(sem);
997	}
998	rwsem_mark_wake(sem, wake_type, wake_q);
999}
1000
1001/*
1002 * Wait for the read lock to be granted
1003 */
1004static struct rw_semaphore __sched *
1005rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
1006{
1007	long adjustment = -RWSEM_READER_BIAS;
1008	long rcnt = (count >> RWSEM_READER_SHIFT);
1009	struct rwsem_waiter waiter;
1010	DEFINE_WAKE_Q(wake_q);
1011
1012	/*
1013	 * To prevent a constant stream of readers from starving a sleeping
1014	 * waiter, don't attempt optimistic lock stealing if the lock is
1015	 * currently owned by readers.
1016	 */
1017	if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
1018	    (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
1019		goto queue;
1020
1021	/*
1022	 * Reader optimistic lock stealing.
1023	 */
1024	if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
1025		rwsem_set_reader_owned(sem);
1026		lockevent_inc(rwsem_rlock_steal);
1027
1028		/*
1029		 * Wake up other readers in the wait queue if it is
1030		 * the first reader.
1031		 */
1032		if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
1033			raw_spin_lock_irq(&sem->wait_lock);
1034			if (!list_empty(&sem->wait_list))
1035				rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
1036						&wake_q);
1037			raw_spin_unlock_irq(&sem->wait_lock);
1038			wake_up_q(&wake_q);
1039		}
1040		return sem;
1041	}
1042
1043queue:
1044	waiter.task = current;
1045	waiter.type = RWSEM_WAITING_FOR_READ;
1046	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1047	waiter.handoff_set = false;
1048
1049	raw_spin_lock_irq(&sem->wait_lock);
1050	if (list_empty(&sem->wait_list)) {
1051		/*
1052		 * In case the wait queue is empty and the lock isn't owned
1053		 * by a writer, this reader can exit the slowpath and return
1054		 * immediately as its RWSEM_READER_BIAS has already been set
1055		 * in the count.
1056		 */
1057		if (!(atomic_long_read(&sem->count) & RWSEM_WRITER_MASK)) {
1058			/* Provide lock ACQUIRE */
1059			smp_acquire__after_ctrl_dep();
1060			raw_spin_unlock_irq(&sem->wait_lock);
1061			rwsem_set_reader_owned(sem);
1062			lockevent_inc(rwsem_rlock_fast);
1063			return sem;
1064		}
1065		adjustment += RWSEM_FLAG_WAITERS;
1066	}
1067	rwsem_add_waiter(sem, &waiter);
1068
1069	/* we're now waiting on the lock, but no longer actively locking */
1070	count = atomic_long_add_return(adjustment, &sem->count);
1071
1072	rwsem_cond_wake_waiter(sem, count, &wake_q);
1073	raw_spin_unlock_irq(&sem->wait_lock);
1074
1075	if (!wake_q_empty(&wake_q))
1076		wake_up_q(&wake_q);
1077
1078	trace_contention_begin(sem, LCB_F_READ);
1079
1080	/* wait to be given the lock */
1081	for (;;) {
1082		set_current_state(state);
1083		if (!smp_load_acquire(&waiter.task)) {
1084			/* Matches rwsem_mark_wake()'s smp_store_release(). */
1085			break;
1086		}
1087		if (signal_pending_state(state, current)) {
1088			raw_spin_lock_irq(&sem->wait_lock);
1089			if (waiter.task)
1090				goto out_nolock;
1091			raw_spin_unlock_irq(&sem->wait_lock);
1092			/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1093			break;
1094		}
1095		schedule();
1096		lockevent_inc(rwsem_sleep_reader);
1097	}
1098
1099	__set_current_state(TASK_RUNNING);
1100	lockevent_inc(rwsem_rlock);
1101	trace_contention_end(sem, 0);
1102	return sem;
1103
1104out_nolock:
1105	rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1106	__set_current_state(TASK_RUNNING);
1107	lockevent_inc(rwsem_rlock_fail);
1108	trace_contention_end(sem, -EINTR);
1109	return ERR_PTR(-EINTR);
1110}
1111
1112/*
1113 * Wait until we successfully acquire the write lock
1114 */
1115static struct rw_semaphore __sched *
1116rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1117{
1118	struct rwsem_waiter waiter;
1119	DEFINE_WAKE_Q(wake_q);
1120
1121	/* do optimistic spinning and steal lock if possible */
1122	if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1123		/* rwsem_optimistic_spin() implies ACQUIRE on success */
1124		return sem;
1125	}
1126
1127	/*
1128	 * Optimistic spinning failed, proceed to the slowpath
1129	 * and block until we can acquire the sem.
1130	 */
1131	waiter.task = current;
1132	waiter.type = RWSEM_WAITING_FOR_WRITE;
1133	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1134	waiter.handoff_set = false;
1135
1136	raw_spin_lock_irq(&sem->wait_lock);
1137	rwsem_add_waiter(sem, &waiter);
1138
1139	/* we're now waiting on the lock */
1140	if (rwsem_first_waiter(sem) != &waiter) {
1141		rwsem_cond_wake_waiter(sem, atomic_long_read(&sem->count),
1142				       &wake_q);
1143		if (!wake_q_empty(&wake_q)) {
1144			/*
1145			 * We want to minimize wait_lock hold time especially
1146			 * when a large number of readers are to be woken up.
1147			 */
1148			raw_spin_unlock_irq(&sem->wait_lock);
1149			wake_up_q(&wake_q);
1150			raw_spin_lock_irq(&sem->wait_lock);
1151		}
1152	} else {
1153		atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1154	}
1155
1156	/* wait until we successfully acquire the lock */
1157	set_current_state(state);
1158	trace_contention_begin(sem, LCB_F_WRITE);
1159
1160	for (;;) {
1161		if (rwsem_try_write_lock(sem, &waiter)) {
1162			/* rwsem_try_write_lock() implies ACQUIRE on success */
1163			break;
1164		}
1165
1166		raw_spin_unlock_irq(&sem->wait_lock);
1167
1168		if (signal_pending_state(state, current))
1169			goto out_nolock;
1170
1171		/*
1172		 * After setting the handoff bit and failing to acquire
1173		 * the lock, attempt to spin on owner to accelerate lock
1174		 * transfer. If the previous owner is a on-cpu writer and it
1175		 * has just released the lock, OWNER_NULL will be returned.
1176		 * In this case, we attempt to acquire the lock again
1177		 * without sleeping.
1178		 */
1179		if (waiter.handoff_set) {
1180			enum owner_state owner_state;
1181
1182			preempt_disable();
1183			owner_state = rwsem_spin_on_owner(sem);
1184			preempt_enable();
1185
1186			if (owner_state == OWNER_NULL)
1187				goto trylock_again;
1188		}
1189
1190		schedule();
1191		lockevent_inc(rwsem_sleep_writer);
1192		set_current_state(state);
1193trylock_again:
1194		raw_spin_lock_irq(&sem->wait_lock);
1195	}
1196	__set_current_state(TASK_RUNNING);
1197	raw_spin_unlock_irq(&sem->wait_lock);
1198	lockevent_inc(rwsem_wlock);
1199	trace_contention_end(sem, 0);
1200	return sem;
1201
1202out_nolock:
1203	__set_current_state(TASK_RUNNING);
1204	raw_spin_lock_irq(&sem->wait_lock);
1205	rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1206	lockevent_inc(rwsem_wlock_fail);
1207	trace_contention_end(sem, -EINTR);
1208	return ERR_PTR(-EINTR);
1209}
1210
1211/*
1212 * handle waking up a waiter on the semaphore
1213 * - up_read/up_write has decremented the active part of count if we come here
1214 */
1215static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1216{
1217	unsigned long flags;
1218	DEFINE_WAKE_Q(wake_q);
1219
1220	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1221
1222	if (!list_empty(&sem->wait_list))
1223		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1224
1225	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1226	wake_up_q(&wake_q);
1227
1228	return sem;
1229}
1230
1231/*
1232 * downgrade a write lock into a read lock
1233 * - caller incremented waiting part of count and discovered it still negative
1234 * - just wake up any readers at the front of the queue
1235 */
1236static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1237{
1238	unsigned long flags;
1239	DEFINE_WAKE_Q(wake_q);
1240
1241	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1242
1243	if (!list_empty(&sem->wait_list))
1244		rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1245
1246	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1247	wake_up_q(&wake_q);
1248
1249	return sem;
1250}
1251
1252/*
1253 * lock for reading
1254 */
1255static inline int __down_read_common(struct rw_semaphore *sem, int state)
1256{
1257	long count;
1258
1259	if (!rwsem_read_trylock(sem, &count)) {
1260		if (IS_ERR(rwsem_down_read_slowpath(sem, count, state)))
1261			return -EINTR;
1262		DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1263	}
1264	return 0;
1265}
1266
1267static inline void __down_read(struct rw_semaphore *sem)
1268{
1269	__down_read_common(sem, TASK_UNINTERRUPTIBLE);
1270}
1271
1272static inline int __down_read_interruptible(struct rw_semaphore *sem)
1273{
1274	return __down_read_common(sem, TASK_INTERRUPTIBLE);
1275}
1276
1277static inline int __down_read_killable(struct rw_semaphore *sem)
1278{
1279	return __down_read_common(sem, TASK_KILLABLE);
1280}
1281
1282static inline int __down_read_trylock(struct rw_semaphore *sem)
1283{
1284	long tmp;
1285
1286	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1287
1288	tmp = atomic_long_read(&sem->count);
1289	while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1290		if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1291						    tmp + RWSEM_READER_BIAS)) {
1292			rwsem_set_reader_owned(sem);
1293			return 1;
1294		}
1295	}
1296	return 0;
1297}
1298
1299/*
1300 * lock for writing
1301 */
1302static inline int __down_write_common(struct rw_semaphore *sem, int state)
1303{
1304	if (unlikely(!rwsem_write_trylock(sem))) {
1305		if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
1306			return -EINTR;
1307	}
1308
1309	return 0;
1310}
1311
1312static inline void __down_write(struct rw_semaphore *sem)
1313{
1314	__down_write_common(sem, TASK_UNINTERRUPTIBLE);
1315}
1316
1317static inline int __down_write_killable(struct rw_semaphore *sem)
1318{
1319	return __down_write_common(sem, TASK_KILLABLE);
1320}
1321
1322static inline int __down_write_trylock(struct rw_semaphore *sem)
1323{
1324	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1325	return rwsem_write_trylock(sem);
1326}
1327
1328/*
1329 * unlock after reading
1330 */
1331static inline void __up_read(struct rw_semaphore *sem)
1332{
1333	long tmp;
1334
1335	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1336	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1337
1338	rwsem_clear_reader_owned(sem);
1339	tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1340	DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1341	if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1342		      RWSEM_FLAG_WAITERS)) {
1343		clear_nonspinnable(sem);
1344		rwsem_wake(sem);
1345	}
1346}
1347
1348/*
1349 * unlock after writing
1350 */
1351static inline void __up_write(struct rw_semaphore *sem)
1352{
1353	long tmp;
1354
1355	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1356	/*
1357	 * sem->owner may differ from current if the ownership is transferred
1358	 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1359	 */
1360	DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1361			    !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1362
1363	preempt_disable();
1364	rwsem_clear_owner(sem);
1365	tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1366	preempt_enable();
1367	if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1368		rwsem_wake(sem);
1369}
1370
1371/*
1372 * downgrade write lock to read lock
1373 */
1374static inline void __downgrade_write(struct rw_semaphore *sem)
1375{
1376	long tmp;
1377
1378	/*
1379	 * When downgrading from exclusive to shared ownership,
1380	 * anything inside the write-locked region cannot leak
1381	 * into the read side. In contrast, anything in the
1382	 * read-locked region is ok to be re-ordered into the
1383	 * write side. As such, rely on RELEASE semantics.
1384	 */
1385	DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1386	tmp = atomic_long_fetch_add_release(
1387		-RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1388	rwsem_set_reader_owned(sem);
1389	if (tmp & RWSEM_FLAG_WAITERS)
1390		rwsem_downgrade_wake(sem);
1391}
1392
1393#else /* !CONFIG_PREEMPT_RT */
1394
1395#define RT_MUTEX_BUILD_MUTEX
1396#include "rtmutex.c"
1397
1398#define rwbase_set_and_save_current_state(state)	\
1399	set_current_state(state)
1400
1401#define rwbase_restore_current_state()			\
1402	__set_current_state(TASK_RUNNING)
1403
1404#define rwbase_rtmutex_lock_state(rtm, state)		\
1405	__rt_mutex_lock(rtm, state)
1406
1407#define rwbase_rtmutex_slowlock_locked(rtm, state)	\
1408	__rt_mutex_slowlock_locked(rtm, NULL, state)
1409
1410#define rwbase_rtmutex_unlock(rtm)			\
1411	__rt_mutex_unlock(rtm)
1412
1413#define rwbase_rtmutex_trylock(rtm)			\
1414	__rt_mutex_trylock(rtm)
1415
1416#define rwbase_signal_pending_state(state, current)	\
1417	signal_pending_state(state, current)
1418
1419#define rwbase_schedule()				\
1420	schedule()
1421
1422#include "rwbase_rt.c"
1423
1424void __init_rwsem(struct rw_semaphore *sem, const char *name,
1425		  struct lock_class_key *key)
1426{
1427	init_rwbase_rt(&(sem)->rwbase);
1428
1429#ifdef CONFIG_DEBUG_LOCK_ALLOC
1430	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1431	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1432#endif
1433}
1434EXPORT_SYMBOL(__init_rwsem);
1435
1436static inline void __down_read(struct rw_semaphore *sem)
1437{
1438	rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1439}
1440
1441static inline int __down_read_interruptible(struct rw_semaphore *sem)
1442{
1443	return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1444}
1445
1446static inline int __down_read_killable(struct rw_semaphore *sem)
1447{
1448	return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1449}
1450
1451static inline int __down_read_trylock(struct rw_semaphore *sem)
1452{
1453	return rwbase_read_trylock(&sem->rwbase);
1454}
1455
1456static inline void __up_read(struct rw_semaphore *sem)
1457{
1458	rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1459}
1460
1461static inline void __sched __down_write(struct rw_semaphore *sem)
1462{
1463	rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1464}
1465
1466static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1467{
1468	return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1469}
1470
1471static inline int __down_write_trylock(struct rw_semaphore *sem)
1472{
1473	return rwbase_write_trylock(&sem->rwbase);
1474}
1475
1476static inline void __up_write(struct rw_semaphore *sem)
1477{
1478	rwbase_write_unlock(&sem->rwbase);
1479}
1480
1481static inline void __downgrade_write(struct rw_semaphore *sem)
1482{
1483	rwbase_write_downgrade(&sem->rwbase);
1484}
1485
1486/* Debug stubs for the common API */
1487#define DEBUG_RWSEMS_WARN_ON(c, sem)
1488
1489static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1490					    struct task_struct *owner)
1491{
1492}
1493
1494static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1495{
1496	int count = atomic_read(&sem->rwbase.readers);
1497
1498	return count < 0 && count != READER_BIAS;
1499}
1500
1501#endif /* CONFIG_PREEMPT_RT */
1502
1503/*
1504 * lock for reading
1505 */
1506void __sched down_read(struct rw_semaphore *sem)
1507{
1508	might_sleep();
1509	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1510
1511	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1512}
1513EXPORT_SYMBOL(down_read);
1514
1515int __sched down_read_interruptible(struct rw_semaphore *sem)
1516{
1517	might_sleep();
1518	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1519
1520	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1521		rwsem_release(&sem->dep_map, _RET_IP_);
1522		return -EINTR;
1523	}
1524
1525	return 0;
1526}
1527EXPORT_SYMBOL(down_read_interruptible);
1528
1529int __sched down_read_killable(struct rw_semaphore *sem)
1530{
1531	might_sleep();
1532	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1533
1534	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1535		rwsem_release(&sem->dep_map, _RET_IP_);
1536		return -EINTR;
1537	}
1538
1539	return 0;
1540}
1541EXPORT_SYMBOL(down_read_killable);
1542
1543/*
1544 * trylock for reading -- returns 1 if successful, 0 if contention
1545 */
1546int down_read_trylock(struct rw_semaphore *sem)
1547{
1548	int ret = __down_read_trylock(sem);
1549
1550	if (ret == 1)
1551		rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1552	return ret;
1553}
1554EXPORT_SYMBOL(down_read_trylock);
1555
1556/*
1557 * lock for writing
1558 */
1559void __sched down_write(struct rw_semaphore *sem)
1560{
1561	might_sleep();
1562	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1563	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1564}
1565EXPORT_SYMBOL(down_write);
1566
1567/*
1568 * lock for writing
1569 */
1570int __sched down_write_killable(struct rw_semaphore *sem)
1571{
1572	might_sleep();
1573	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1574
1575	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1576				  __down_write_killable)) {
1577		rwsem_release(&sem->dep_map, _RET_IP_);
1578		return -EINTR;
1579	}
1580
1581	return 0;
1582}
1583EXPORT_SYMBOL(down_write_killable);
1584
1585/*
1586 * trylock for writing -- returns 1 if successful, 0 if contention
1587 */
1588int down_write_trylock(struct rw_semaphore *sem)
1589{
1590	int ret = __down_write_trylock(sem);
1591
1592	if (ret == 1)
1593		rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1594
1595	return ret;
1596}
1597EXPORT_SYMBOL(down_write_trylock);
1598
1599/*
1600 * release a read lock
1601 */
1602void up_read(struct rw_semaphore *sem)
1603{
1604	rwsem_release(&sem->dep_map, _RET_IP_);
1605	__up_read(sem);
1606}
1607EXPORT_SYMBOL(up_read);
1608
1609/*
1610 * release a write lock
1611 */
1612void up_write(struct rw_semaphore *sem)
1613{
1614	rwsem_release(&sem->dep_map, _RET_IP_);
1615	__up_write(sem);
1616}
1617EXPORT_SYMBOL(up_write);
1618
1619/*
1620 * downgrade write lock to read lock
1621 */
1622void downgrade_write(struct rw_semaphore *sem)
1623{
1624	lock_downgrade(&sem->dep_map, _RET_IP_);
1625	__downgrade_write(sem);
1626}
1627EXPORT_SYMBOL(downgrade_write);
1628
1629#ifdef CONFIG_DEBUG_LOCK_ALLOC
1630
1631void down_read_nested(struct rw_semaphore *sem, int subclass)
1632{
1633	might_sleep();
1634	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1635	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1636}
1637EXPORT_SYMBOL(down_read_nested);
1638
1639int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1640{
1641	might_sleep();
1642	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1643
1644	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1645		rwsem_release(&sem->dep_map, _RET_IP_);
1646		return -EINTR;
1647	}
1648
1649	return 0;
1650}
1651EXPORT_SYMBOL(down_read_killable_nested);
1652
1653void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1654{
1655	might_sleep();
1656	rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1657	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1658}
1659EXPORT_SYMBOL(_down_write_nest_lock);
1660
1661void down_read_non_owner(struct rw_semaphore *sem)
1662{
1663	might_sleep();
1664	__down_read(sem);
1665	__rwsem_set_reader_owned(sem, NULL);
1666}
1667EXPORT_SYMBOL(down_read_non_owner);
1668
1669void down_write_nested(struct rw_semaphore *sem, int subclass)
1670{
1671	might_sleep();
1672	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1673	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1674}
1675EXPORT_SYMBOL(down_write_nested);
1676
1677int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1678{
1679	might_sleep();
1680	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1681
1682	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1683				  __down_write_killable)) {
1684		rwsem_release(&sem->dep_map, _RET_IP_);
1685		return -EINTR;
1686	}
1687
1688	return 0;
1689}
1690EXPORT_SYMBOL(down_write_killable_nested);
1691
1692void up_read_non_owner(struct rw_semaphore *sem)
1693{
1694	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1695	__up_read(sem);
1696}
1697EXPORT_SYMBOL(up_read_non_owner);
1698
1699#endif
1700