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