xref: /openbmc/linux/kernel/locking/rwsem.c (revision d179ebed)
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  */
rwsem_set_owner(struct rw_semaphore * sem)140 static 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 
rwsem_clear_owner(struct rw_semaphore * sem)146 static 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  */
rwsem_test_oflags(struct rw_semaphore * sem,long flags)155 static 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  */
__rwsem_set_reader_owned(struct rw_semaphore * sem,struct task_struct * owner)170 static 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 
rwsem_set_reader_owned(struct rw_semaphore * sem)179 static 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  */
is_rwsem_reader_owned(struct rw_semaphore * sem)187 static 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  */
rwsem_clear_reader_owned(struct rw_semaphore * sem)208 static 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
rwsem_clear_reader_owned(struct rw_semaphore * sem)219 static 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  */
rwsem_set_nonspinnable(struct rw_semaphore * sem)228 static 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 
rwsem_read_trylock(struct rw_semaphore * sem,long * cntp)241 static 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 
rwsem_write_trylock(struct rw_semaphore * sem)256 static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
257 {
258 	long tmp = RWSEM_UNLOCKED_VALUE;
259 
260 	if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp, RWSEM_WRITER_LOCKED)) {
261 		rwsem_set_owner(sem);
262 		return true;
263 	}
264 
265 	return false;
266 }
267 
268 /*
269  * Return just the real task structure pointer of the owner
270  */
rwsem_owner(struct rw_semaphore * sem)271 static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
272 {
273 	return (struct task_struct *)
274 		(atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
275 }
276 
277 /*
278  * Return the real task structure pointer of the owner and the embedded
279  * flags in the owner. pflags must be non-NULL.
280  */
281 static inline struct task_struct *
rwsem_owner_flags(struct rw_semaphore * sem,unsigned long * pflags)282 rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
283 {
284 	unsigned long owner = atomic_long_read(&sem->owner);
285 
286 	*pflags = owner & RWSEM_OWNER_FLAGS_MASK;
287 	return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
288 }
289 
290 /*
291  * Guide to the rw_semaphore's count field.
292  *
293  * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
294  * by a writer.
295  *
296  * The lock is owned by readers when
297  * (1) the RWSEM_WRITER_LOCKED isn't set in count,
298  * (2) some of the reader bits are set in count, and
299  * (3) the owner field has RWSEM_READ_OWNED bit set.
300  *
301  * Having some reader bits set is not enough to guarantee a readers owned
302  * lock as the readers may be in the process of backing out from the count
303  * and a writer has just released the lock. So another writer may steal
304  * the lock immediately after that.
305  */
306 
307 /*
308  * Initialize an rwsem:
309  */
__init_rwsem(struct rw_semaphore * sem,const char * name,struct lock_class_key * key)310 void __init_rwsem(struct rw_semaphore *sem, const char *name,
311 		  struct lock_class_key *key)
312 {
313 #ifdef CONFIG_DEBUG_LOCK_ALLOC
314 	/*
315 	 * Make sure we are not reinitializing a held semaphore:
316 	 */
317 	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
318 	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
319 #endif
320 #ifdef CONFIG_DEBUG_RWSEMS
321 	sem->magic = sem;
322 #endif
323 	atomic_long_set(&sem->count, RWSEM_UNLOCKED_VALUE);
324 	raw_spin_lock_init(&sem->wait_lock);
325 	INIT_LIST_HEAD(&sem->wait_list);
326 	atomic_long_set(&sem->owner, 0L);
327 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
328 	osq_lock_init(&sem->osq);
329 #endif
330 }
331 EXPORT_SYMBOL(__init_rwsem);
332 
333 enum rwsem_waiter_type {
334 	RWSEM_WAITING_FOR_WRITE,
335 	RWSEM_WAITING_FOR_READ
336 };
337 
338 struct rwsem_waiter {
339 	struct list_head list;
340 	struct task_struct *task;
341 	enum rwsem_waiter_type type;
342 	unsigned long timeout;
343 	bool handoff_set;
344 };
345 #define rwsem_first_waiter(sem) \
346 	list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
347 
348 enum rwsem_wake_type {
349 	RWSEM_WAKE_ANY,		/* Wake whatever's at head of wait list */
350 	RWSEM_WAKE_READERS,	/* Wake readers only */
351 	RWSEM_WAKE_READ_OWNED	/* Waker thread holds the read lock */
352 };
353 
354 /*
355  * The typical HZ value is either 250 or 1000. So set the minimum waiting
356  * time to at least 4ms or 1 jiffy (if it is higher than 4ms) in the wait
357  * queue before initiating the handoff protocol.
358  */
359 #define RWSEM_WAIT_TIMEOUT	DIV_ROUND_UP(HZ, 250)
360 
361 /*
362  * Magic number to batch-wakeup waiting readers, even when writers are
363  * also present in the queue. This both limits the amount of work the
364  * waking thread must do and also prevents any potential counter overflow,
365  * however unlikely.
366  */
367 #define MAX_READERS_WAKEUP	0x100
368 
369 static inline void
rwsem_add_waiter(struct rw_semaphore * sem,struct rwsem_waiter * waiter)370 rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
371 {
372 	lockdep_assert_held(&sem->wait_lock);
373 	list_add_tail(&waiter->list, &sem->wait_list);
374 	/* caller will set RWSEM_FLAG_WAITERS */
375 }
376 
377 /*
378  * Remove a waiter from the wait_list and clear flags.
379  *
380  * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
381  * this function. Modify with care.
382  *
383  * Return: true if wait_list isn't empty and false otherwise
384  */
385 static inline bool
rwsem_del_waiter(struct rw_semaphore * sem,struct rwsem_waiter * waiter)386 rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
387 {
388 	lockdep_assert_held(&sem->wait_lock);
389 	list_del(&waiter->list);
390 	if (likely(!list_empty(&sem->wait_list)))
391 		return true;
392 
393 	atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, &sem->count);
394 	return false;
395 }
396 
397 /*
398  * handle the lock release when processes blocked on it that can now run
399  * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
400  *   have been set.
401  * - there must be someone on the queue
402  * - the wait_lock must be held by the caller
403  * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
404  *   to actually wakeup the blocked task(s) and drop the reference count,
405  *   preferably when the wait_lock is released
406  * - woken process blocks are discarded from the list after having task zeroed
407  * - writers are only marked woken if downgrading is false
408  *
409  * Implies rwsem_del_waiter() for all woken readers.
410  */
rwsem_mark_wake(struct rw_semaphore * sem,enum rwsem_wake_type wake_type,struct wake_q_head * wake_q)411 static void rwsem_mark_wake(struct rw_semaphore *sem,
412 			    enum rwsem_wake_type wake_type,
413 			    struct wake_q_head *wake_q)
414 {
415 	struct rwsem_waiter *waiter, *tmp;
416 	long oldcount, woken = 0, adjustment = 0;
417 	struct list_head wlist;
418 
419 	lockdep_assert_held(&sem->wait_lock);
420 
421 	/*
422 	 * Take a peek at the queue head waiter such that we can determine
423 	 * the wakeup(s) to perform.
424 	 */
425 	waiter = rwsem_first_waiter(sem);
426 
427 	if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
428 		if (wake_type == RWSEM_WAKE_ANY) {
429 			/*
430 			 * Mark writer at the front of the queue for wakeup.
431 			 * Until the task is actually later awoken later by
432 			 * the caller, other writers are able to steal it.
433 			 * Readers, on the other hand, will block as they
434 			 * will notice the queued writer.
435 			 */
436 			wake_q_add(wake_q, waiter->task);
437 			lockevent_inc(rwsem_wake_writer);
438 		}
439 
440 		return;
441 	}
442 
443 	/*
444 	 * No reader wakeup if there are too many of them already.
445 	 */
446 	if (unlikely(atomic_long_read(&sem->count) < 0))
447 		return;
448 
449 	/*
450 	 * Writers might steal the lock before we grant it to the next reader.
451 	 * We prefer to do the first reader grant before counting readers
452 	 * so we can bail out early if a writer stole the lock.
453 	 */
454 	if (wake_type != RWSEM_WAKE_READ_OWNED) {
455 		struct task_struct *owner;
456 
457 		adjustment = RWSEM_READER_BIAS;
458 		oldcount = atomic_long_fetch_add(adjustment, &sem->count);
459 		if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
460 			/*
461 			 * When we've been waiting "too" long (for writers
462 			 * to give up the lock), request a HANDOFF to
463 			 * force the issue.
464 			 */
465 			if (time_after(jiffies, waiter->timeout)) {
466 				if (!(oldcount & RWSEM_FLAG_HANDOFF)) {
467 					adjustment -= RWSEM_FLAG_HANDOFF;
468 					lockevent_inc(rwsem_rlock_handoff);
469 				}
470 				waiter->handoff_set = true;
471 			}
472 
473 			atomic_long_add(-adjustment, &sem->count);
474 			return;
475 		}
476 		/*
477 		 * Set it to reader-owned to give spinners an early
478 		 * indication that readers now have the lock.
479 		 * The reader nonspinnable bit seen at slowpath entry of
480 		 * the reader is copied over.
481 		 */
482 		owner = waiter->task;
483 		__rwsem_set_reader_owned(sem, owner);
484 	}
485 
486 	/*
487 	 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
488 	 * queue. We know that the woken will be at least 1 as we accounted
489 	 * for above. Note we increment the 'active part' of the count by the
490 	 * number of readers before waking any processes up.
491 	 *
492 	 * This is an adaptation of the phase-fair R/W locks where at the
493 	 * reader phase (first waiter is a reader), all readers are eligible
494 	 * to acquire the lock at the same time irrespective of their order
495 	 * in the queue. The writers acquire the lock according to their
496 	 * order in the queue.
497 	 *
498 	 * We have to do wakeup in 2 passes to prevent the possibility that
499 	 * the reader count may be decremented before it is incremented. It
500 	 * is because the to-be-woken waiter may not have slept yet. So it
501 	 * may see waiter->task got cleared, finish its critical section and
502 	 * do an unlock before the reader count increment.
503 	 *
504 	 * 1) Collect the read-waiters in a separate list, count them and
505 	 *    fully increment the reader count in rwsem.
506 	 * 2) For each waiters in the new list, clear waiter->task and
507 	 *    put them into wake_q to be woken up later.
508 	 */
509 	INIT_LIST_HEAD(&wlist);
510 	list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
511 		if (waiter->type == RWSEM_WAITING_FOR_WRITE)
512 			continue;
513 
514 		woken++;
515 		list_move_tail(&waiter->list, &wlist);
516 
517 		/*
518 		 * Limit # of readers that can be woken up per wakeup call.
519 		 */
520 		if (unlikely(woken >= MAX_READERS_WAKEUP))
521 			break;
522 	}
523 
524 	adjustment = woken * RWSEM_READER_BIAS - adjustment;
525 	lockevent_cond_inc(rwsem_wake_reader, woken);
526 
527 	oldcount = atomic_long_read(&sem->count);
528 	if (list_empty(&sem->wait_list)) {
529 		/*
530 		 * Combined with list_move_tail() above, this implies
531 		 * rwsem_del_waiter().
532 		 */
533 		adjustment -= RWSEM_FLAG_WAITERS;
534 		if (oldcount & RWSEM_FLAG_HANDOFF)
535 			adjustment -= RWSEM_FLAG_HANDOFF;
536 	} else if (woken) {
537 		/*
538 		 * When we've woken a reader, we no longer need to force
539 		 * writers to give up the lock and we can clear HANDOFF.
540 		 */
541 		if (oldcount & RWSEM_FLAG_HANDOFF)
542 			adjustment -= RWSEM_FLAG_HANDOFF;
543 	}
544 
545 	if (adjustment)
546 		atomic_long_add(adjustment, &sem->count);
547 
548 	/* 2nd pass */
549 	list_for_each_entry_safe(waiter, tmp, &wlist, list) {
550 		struct task_struct *tsk;
551 
552 		tsk = waiter->task;
553 		get_task_struct(tsk);
554 
555 		/*
556 		 * Ensure calling get_task_struct() before setting the reader
557 		 * waiter to nil such that rwsem_down_read_slowpath() cannot
558 		 * race with do_exit() by always holding a reference count
559 		 * to the task to wakeup.
560 		 */
561 		smp_store_release(&waiter->task, NULL);
562 		/*
563 		 * Ensure issuing the wakeup (either by us or someone else)
564 		 * after setting the reader waiter to nil.
565 		 */
566 		wake_q_add_safe(wake_q, tsk);
567 	}
568 }
569 
570 /*
571  * Remove a waiter and try to wake up other waiters in the wait queue
572  * This function is called from the out_nolock path of both the reader and
573  * writer slowpaths with wait_lock held. It releases the wait_lock and
574  * optionally wake up waiters before it returns.
575  */
576 static inline void
rwsem_del_wake_waiter(struct rw_semaphore * sem,struct rwsem_waiter * waiter,struct wake_q_head * wake_q)577 rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,
578 		      struct wake_q_head *wake_q)
579 		      __releases(&sem->wait_lock)
580 {
581 	bool first = rwsem_first_waiter(sem) == waiter;
582 
583 	wake_q_init(wake_q);
584 
585 	/*
586 	 * If the wait_list isn't empty and the waiter to be deleted is
587 	 * the first waiter, we wake up the remaining waiters as they may
588 	 * be eligible to acquire or spin on the lock.
589 	 */
590 	if (rwsem_del_waiter(sem, waiter) && first)
591 		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, wake_q);
592 	raw_spin_unlock_irq(&sem->wait_lock);
593 	if (!wake_q_empty(wake_q))
594 		wake_up_q(wake_q);
595 }
596 
597 /*
598  * This function must be called with the sem->wait_lock held to prevent
599  * race conditions between checking the rwsem wait list and setting the
600  * sem->count accordingly.
601  *
602  * Implies rwsem_del_waiter() on success.
603  */
rwsem_try_write_lock(struct rw_semaphore * sem,struct rwsem_waiter * waiter)604 static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
605 					struct rwsem_waiter *waiter)
606 {
607 	struct rwsem_waiter *first = rwsem_first_waiter(sem);
608 	long count, new;
609 
610 	lockdep_assert_held(&sem->wait_lock);
611 
612 	count = atomic_long_read(&sem->count);
613 	do {
614 		bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
615 
616 		if (has_handoff) {
617 			/*
618 			 * Honor handoff bit and yield only when the first
619 			 * waiter is the one that set it. Otherwisee, we
620 			 * still try to acquire the rwsem.
621 			 */
622 			if (first->handoff_set && (waiter != first))
623 				return false;
624 		}
625 
626 		new = count;
627 
628 		if (count & RWSEM_LOCK_MASK) {
629 			/*
630 			 * A waiter (first or not) can set the handoff bit
631 			 * if it is an RT task or wait in the wait queue
632 			 * for too long.
633 			 */
634 			if (has_handoff || (!rt_task(waiter->task) &&
635 					    !time_after(jiffies, waiter->timeout)))
636 				return false;
637 
638 			new |= RWSEM_FLAG_HANDOFF;
639 		} else {
640 			new |= RWSEM_WRITER_LOCKED;
641 			new &= ~RWSEM_FLAG_HANDOFF;
642 
643 			if (list_is_singular(&sem->wait_list))
644 				new &= ~RWSEM_FLAG_WAITERS;
645 		}
646 	} while (!atomic_long_try_cmpxchg_acquire(&sem->count, &count, new));
647 
648 	/*
649 	 * We have either acquired the lock with handoff bit cleared or set
650 	 * the handoff bit. Only the first waiter can have its handoff_set
651 	 * set here to enable optimistic spinning in slowpath loop.
652 	 */
653 	if (new & RWSEM_FLAG_HANDOFF) {
654 		first->handoff_set = true;
655 		lockevent_inc(rwsem_wlock_handoff);
656 		return false;
657 	}
658 
659 	/*
660 	 * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
661 	 * success.
662 	 */
663 	list_del(&waiter->list);
664 	rwsem_set_owner(sem);
665 	return true;
666 }
667 
668 /*
669  * The rwsem_spin_on_owner() function returns the following 4 values
670  * depending on the lock owner state.
671  *   OWNER_NULL  : owner is currently NULL
672  *   OWNER_WRITER: when owner changes and is a writer
673  *   OWNER_READER: when owner changes and the new owner may be a reader.
674  *   OWNER_NONSPINNABLE:
675  *		   when optimistic spinning has to stop because either the
676  *		   owner stops running, is unknown, or its timeslice has
677  *		   been used up.
678  */
679 enum owner_state {
680 	OWNER_NULL		= 1 << 0,
681 	OWNER_WRITER		= 1 << 1,
682 	OWNER_READER		= 1 << 2,
683 	OWNER_NONSPINNABLE	= 1 << 3,
684 };
685 
686 #ifdef CONFIG_RWSEM_SPIN_ON_OWNER
687 /*
688  * Try to acquire write lock before the writer has been put on wait queue.
689  */
rwsem_try_write_lock_unqueued(struct rw_semaphore * sem)690 static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
691 {
692 	long count = atomic_long_read(&sem->count);
693 
694 	while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
695 		if (atomic_long_try_cmpxchg_acquire(&sem->count, &count,
696 					count | RWSEM_WRITER_LOCKED)) {
697 			rwsem_set_owner(sem);
698 			lockevent_inc(rwsem_opt_lock);
699 			return true;
700 		}
701 	}
702 	return false;
703 }
704 
rwsem_can_spin_on_owner(struct rw_semaphore * sem)705 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
706 {
707 	struct task_struct *owner;
708 	unsigned long flags;
709 	bool ret = true;
710 
711 	if (need_resched()) {
712 		lockevent_inc(rwsem_opt_fail);
713 		return false;
714 	}
715 
716 	/*
717 	 * Disable preemption is equal to the RCU read-side crital section,
718 	 * thus the task_strcut structure won't go away.
719 	 */
720 	owner = rwsem_owner_flags(sem, &flags);
721 	/*
722 	 * Don't check the read-owner as the entry may be stale.
723 	 */
724 	if ((flags & RWSEM_NONSPINNABLE) ||
725 	    (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
726 		ret = false;
727 
728 	lockevent_cond_inc(rwsem_opt_fail, !ret);
729 	return ret;
730 }
731 
732 #define OWNER_SPINNABLE		(OWNER_NULL | OWNER_WRITER | OWNER_READER)
733 
734 static inline enum owner_state
rwsem_owner_state(struct task_struct * owner,unsigned long flags)735 rwsem_owner_state(struct task_struct *owner, unsigned long flags)
736 {
737 	if (flags & RWSEM_NONSPINNABLE)
738 		return OWNER_NONSPINNABLE;
739 
740 	if (flags & RWSEM_READER_OWNED)
741 		return OWNER_READER;
742 
743 	return owner ? OWNER_WRITER : OWNER_NULL;
744 }
745 
746 static noinline enum owner_state
rwsem_spin_on_owner(struct rw_semaphore * sem)747 rwsem_spin_on_owner(struct rw_semaphore *sem)
748 {
749 	struct task_struct *new, *owner;
750 	unsigned long flags, new_flags;
751 	enum owner_state state;
752 
753 	lockdep_assert_preemption_disabled();
754 
755 	owner = rwsem_owner_flags(sem, &flags);
756 	state = rwsem_owner_state(owner, flags);
757 	if (state != OWNER_WRITER)
758 		return state;
759 
760 	for (;;) {
761 		/*
762 		 * When a waiting writer set the handoff flag, it may spin
763 		 * on the owner as well. Once that writer acquires the lock,
764 		 * we can spin on it. So we don't need to quit even when the
765 		 * handoff bit is set.
766 		 */
767 		new = rwsem_owner_flags(sem, &new_flags);
768 		if ((new != owner) || (new_flags != flags)) {
769 			state = rwsem_owner_state(new, new_flags);
770 			break;
771 		}
772 
773 		/*
774 		 * Ensure we emit the owner->on_cpu, dereference _after_
775 		 * checking sem->owner still matches owner, if that fails,
776 		 * owner might point to free()d memory, if it still matches,
777 		 * our spinning context already disabled preemption which is
778 		 * equal to RCU read-side crital section ensures the memory
779 		 * stays valid.
780 		 */
781 		barrier();
782 
783 		if (need_resched() || !owner_on_cpu(owner)) {
784 			state = OWNER_NONSPINNABLE;
785 			break;
786 		}
787 
788 		cpu_relax();
789 	}
790 
791 	return state;
792 }
793 
794 /*
795  * Calculate reader-owned rwsem spinning threshold for writer
796  *
797  * The more readers own the rwsem, the longer it will take for them to
798  * wind down and free the rwsem. So the empirical formula used to
799  * determine the actual spinning time limit here is:
800  *
801  *   Spinning threshold = (10 + nr_readers/2)us
802  *
803  * The limit is capped to a maximum of 25us (30 readers). This is just
804  * a heuristic and is subjected to change in the future.
805  */
rwsem_rspin_threshold(struct rw_semaphore * sem)806 static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
807 {
808 	long count = atomic_long_read(&sem->count);
809 	int readers = count >> RWSEM_READER_SHIFT;
810 	u64 delta;
811 
812 	if (readers > 30)
813 		readers = 30;
814 	delta = (20 + readers) * NSEC_PER_USEC / 2;
815 
816 	return sched_clock() + delta;
817 }
818 
rwsem_optimistic_spin(struct rw_semaphore * sem)819 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
820 {
821 	bool taken = false;
822 	int prev_owner_state = OWNER_NULL;
823 	int loop = 0;
824 	u64 rspin_threshold = 0;
825 
826 	/* sem->wait_lock should not be held when doing optimistic spinning */
827 	if (!osq_lock(&sem->osq))
828 		goto done;
829 
830 	/*
831 	 * Optimistically spin on the owner field and attempt to acquire the
832 	 * lock whenever the owner changes. Spinning will be stopped when:
833 	 *  1) the owning writer isn't running; or
834 	 *  2) readers own the lock and spinning time has exceeded limit.
835 	 */
836 	for (;;) {
837 		enum owner_state owner_state;
838 
839 		owner_state = rwsem_spin_on_owner(sem);
840 		if (!(owner_state & OWNER_SPINNABLE))
841 			break;
842 
843 		/*
844 		 * Try to acquire the lock
845 		 */
846 		taken = rwsem_try_write_lock_unqueued(sem);
847 
848 		if (taken)
849 			break;
850 
851 		/*
852 		 * Time-based reader-owned rwsem optimistic spinning
853 		 */
854 		if (owner_state == OWNER_READER) {
855 			/*
856 			 * Re-initialize rspin_threshold every time when
857 			 * the owner state changes from non-reader to reader.
858 			 * This allows a writer to steal the lock in between
859 			 * 2 reader phases and have the threshold reset at
860 			 * the beginning of the 2nd reader phase.
861 			 */
862 			if (prev_owner_state != OWNER_READER) {
863 				if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
864 					break;
865 				rspin_threshold = rwsem_rspin_threshold(sem);
866 				loop = 0;
867 			}
868 
869 			/*
870 			 * Check time threshold once every 16 iterations to
871 			 * avoid calling sched_clock() too frequently so
872 			 * as to reduce the average latency between the times
873 			 * when the lock becomes free and when the spinner
874 			 * is ready to do a trylock.
875 			 */
876 			else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
877 				rwsem_set_nonspinnable(sem);
878 				lockevent_inc(rwsem_opt_nospin);
879 				break;
880 			}
881 		}
882 
883 		/*
884 		 * An RT task cannot do optimistic spinning if it cannot
885 		 * be sure the lock holder is running or live-lock may
886 		 * happen if the current task and the lock holder happen
887 		 * to run in the same CPU. However, aborting optimistic
888 		 * spinning while a NULL owner is detected may miss some
889 		 * opportunity where spinning can continue without causing
890 		 * problem.
891 		 *
892 		 * There are 2 possible cases where an RT task may be able
893 		 * to continue spinning.
894 		 *
895 		 * 1) The lock owner is in the process of releasing the
896 		 *    lock, sem->owner is cleared but the lock has not
897 		 *    been released yet.
898 		 * 2) The lock was free and owner cleared, but another
899 		 *    task just comes in and acquire the lock before
900 		 *    we try to get it. The new owner may be a spinnable
901 		 *    writer.
902 		 *
903 		 * To take advantage of two scenarios listed above, the RT
904 		 * task is made to retry one more time to see if it can
905 		 * acquire the lock or continue spinning on the new owning
906 		 * writer. Of course, if the time lag is long enough or the
907 		 * new owner is not a writer or spinnable, the RT task will
908 		 * quit spinning.
909 		 *
910 		 * If the owner is a writer, the need_resched() check is
911 		 * done inside rwsem_spin_on_owner(). If the owner is not
912 		 * a writer, need_resched() check needs to be done here.
913 		 */
914 		if (owner_state != OWNER_WRITER) {
915 			if (need_resched())
916 				break;
917 			if (rt_task(current) &&
918 			   (prev_owner_state != OWNER_WRITER))
919 				break;
920 		}
921 		prev_owner_state = owner_state;
922 
923 		/*
924 		 * The cpu_relax() call is a compiler barrier which forces
925 		 * everything in this loop to be re-loaded. We don't need
926 		 * memory barriers as we'll eventually observe the right
927 		 * values at the cost of a few extra spins.
928 		 */
929 		cpu_relax();
930 	}
931 	osq_unlock(&sem->osq);
932 done:
933 	lockevent_cond_inc(rwsem_opt_fail, !taken);
934 	return taken;
935 }
936 
937 /*
938  * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
939  * only be called when the reader count reaches 0.
940  */
clear_nonspinnable(struct rw_semaphore * sem)941 static inline void clear_nonspinnable(struct rw_semaphore *sem)
942 {
943 	if (unlikely(rwsem_test_oflags(sem, RWSEM_NONSPINNABLE)))
944 		atomic_long_andnot(RWSEM_NONSPINNABLE, &sem->owner);
945 }
946 
947 #else
rwsem_can_spin_on_owner(struct rw_semaphore * sem)948 static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
949 {
950 	return false;
951 }
952 
rwsem_optimistic_spin(struct rw_semaphore * sem)953 static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
954 {
955 	return false;
956 }
957 
clear_nonspinnable(struct rw_semaphore * sem)958 static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
959 
960 static inline enum owner_state
rwsem_spin_on_owner(struct rw_semaphore * sem)961 rwsem_spin_on_owner(struct rw_semaphore *sem)
962 {
963 	return OWNER_NONSPINNABLE;
964 }
965 #endif
966 
967 /*
968  * Prepare to wake up waiter(s) in the wait queue by putting them into the
969  * given wake_q if the rwsem lock owner isn't a writer. If rwsem is likely
970  * reader-owned, wake up read lock waiters in queue front or wake up any
971  * front waiter otherwise.
972 
973  * This is being called from both reader and writer slow paths.
974  */
rwsem_cond_wake_waiter(struct rw_semaphore * sem,long count,struct wake_q_head * wake_q)975 static inline void rwsem_cond_wake_waiter(struct rw_semaphore *sem, long count,
976 					  struct wake_q_head *wake_q)
977 {
978 	enum rwsem_wake_type wake_type;
979 
980 	if (count & RWSEM_WRITER_MASK)
981 		return;
982 
983 	if (count & RWSEM_READER_MASK) {
984 		wake_type = RWSEM_WAKE_READERS;
985 	} else {
986 		wake_type = RWSEM_WAKE_ANY;
987 		clear_nonspinnable(sem);
988 	}
989 	rwsem_mark_wake(sem, wake_type, wake_q);
990 }
991 
992 /*
993  * Wait for the read lock to be granted
994  */
995 static struct rw_semaphore __sched *
rwsem_down_read_slowpath(struct rw_semaphore * sem,long count,unsigned int state)996 rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
997 {
998 	long adjustment = -RWSEM_READER_BIAS;
999 	long rcnt = (count >> RWSEM_READER_SHIFT);
1000 	struct rwsem_waiter waiter;
1001 	DEFINE_WAKE_Q(wake_q);
1002 
1003 	/*
1004 	 * To prevent a constant stream of readers from starving a sleeping
1005 	 * waiter, don't attempt optimistic lock stealing if the lock is
1006 	 * currently owned by readers.
1007 	 */
1008 	if ((atomic_long_read(&sem->owner) & RWSEM_READER_OWNED) &&
1009 	    (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
1010 		goto queue;
1011 
1012 	/*
1013 	 * Reader optimistic lock stealing.
1014 	 */
1015 	if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
1016 		rwsem_set_reader_owned(sem);
1017 		lockevent_inc(rwsem_rlock_steal);
1018 
1019 		/*
1020 		 * Wake up other readers in the wait queue if it is
1021 		 * the first reader.
1022 		 */
1023 		if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
1024 			raw_spin_lock_irq(&sem->wait_lock);
1025 			if (!list_empty(&sem->wait_list))
1026 				rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED,
1027 						&wake_q);
1028 			raw_spin_unlock_irq(&sem->wait_lock);
1029 			wake_up_q(&wake_q);
1030 		}
1031 		return sem;
1032 	}
1033 
1034 queue:
1035 	waiter.task = current;
1036 	waiter.type = RWSEM_WAITING_FOR_READ;
1037 	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1038 	waiter.handoff_set = false;
1039 
1040 	raw_spin_lock_irq(&sem->wait_lock);
1041 	if (list_empty(&sem->wait_list)) {
1042 		/*
1043 		 * In case the wait queue is empty and the lock isn't owned
1044 		 * by a writer, this reader can exit the slowpath and return
1045 		 * immediately as its RWSEM_READER_BIAS has already been set
1046 		 * in the count.
1047 		 */
1048 		if (!(atomic_long_read(&sem->count) & RWSEM_WRITER_MASK)) {
1049 			/* Provide lock ACQUIRE */
1050 			smp_acquire__after_ctrl_dep();
1051 			raw_spin_unlock_irq(&sem->wait_lock);
1052 			rwsem_set_reader_owned(sem);
1053 			lockevent_inc(rwsem_rlock_fast);
1054 			return sem;
1055 		}
1056 		adjustment += RWSEM_FLAG_WAITERS;
1057 	}
1058 	rwsem_add_waiter(sem, &waiter);
1059 
1060 	/* we're now waiting on the lock, but no longer actively locking */
1061 	count = atomic_long_add_return(adjustment, &sem->count);
1062 
1063 	rwsem_cond_wake_waiter(sem, count, &wake_q);
1064 	raw_spin_unlock_irq(&sem->wait_lock);
1065 
1066 	if (!wake_q_empty(&wake_q))
1067 		wake_up_q(&wake_q);
1068 
1069 	trace_contention_begin(sem, LCB_F_READ);
1070 
1071 	/* wait to be given the lock */
1072 	for (;;) {
1073 		set_current_state(state);
1074 		if (!smp_load_acquire(&waiter.task)) {
1075 			/* Matches rwsem_mark_wake()'s smp_store_release(). */
1076 			break;
1077 		}
1078 		if (signal_pending_state(state, current)) {
1079 			raw_spin_lock_irq(&sem->wait_lock);
1080 			if (waiter.task)
1081 				goto out_nolock;
1082 			raw_spin_unlock_irq(&sem->wait_lock);
1083 			/* Ordered by sem->wait_lock against rwsem_mark_wake(). */
1084 			break;
1085 		}
1086 		schedule_preempt_disabled();
1087 		lockevent_inc(rwsem_sleep_reader);
1088 	}
1089 
1090 	__set_current_state(TASK_RUNNING);
1091 	lockevent_inc(rwsem_rlock);
1092 	trace_contention_end(sem, 0);
1093 	return sem;
1094 
1095 out_nolock:
1096 	rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1097 	__set_current_state(TASK_RUNNING);
1098 	lockevent_inc(rwsem_rlock_fail);
1099 	trace_contention_end(sem, -EINTR);
1100 	return ERR_PTR(-EINTR);
1101 }
1102 
1103 /*
1104  * Wait until we successfully acquire the write lock
1105  */
1106 static struct rw_semaphore __sched *
rwsem_down_write_slowpath(struct rw_semaphore * sem,int state)1107 rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1108 {
1109 	struct rwsem_waiter waiter;
1110 	DEFINE_WAKE_Q(wake_q);
1111 
1112 	/* do optimistic spinning and steal lock if possible */
1113 	if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1114 		/* rwsem_optimistic_spin() implies ACQUIRE on success */
1115 		return sem;
1116 	}
1117 
1118 	/*
1119 	 * Optimistic spinning failed, proceed to the slowpath
1120 	 * and block until we can acquire the sem.
1121 	 */
1122 	waiter.task = current;
1123 	waiter.type = RWSEM_WAITING_FOR_WRITE;
1124 	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1125 	waiter.handoff_set = false;
1126 
1127 	raw_spin_lock_irq(&sem->wait_lock);
1128 	rwsem_add_waiter(sem, &waiter);
1129 
1130 	/* we're now waiting on the lock */
1131 	if (rwsem_first_waiter(sem) != &waiter) {
1132 		rwsem_cond_wake_waiter(sem, atomic_long_read(&sem->count),
1133 				       &wake_q);
1134 		if (!wake_q_empty(&wake_q)) {
1135 			/*
1136 			 * We want to minimize wait_lock hold time especially
1137 			 * when a large number of readers are to be woken up.
1138 			 */
1139 			raw_spin_unlock_irq(&sem->wait_lock);
1140 			wake_up_q(&wake_q);
1141 			raw_spin_lock_irq(&sem->wait_lock);
1142 		}
1143 	} else {
1144 		atomic_long_or(RWSEM_FLAG_WAITERS, &sem->count);
1145 	}
1146 
1147 	/* wait until we successfully acquire the lock */
1148 	set_current_state(state);
1149 	trace_contention_begin(sem, LCB_F_WRITE);
1150 
1151 	for (;;) {
1152 		if (rwsem_try_write_lock(sem, &waiter)) {
1153 			/* rwsem_try_write_lock() implies ACQUIRE on success */
1154 			break;
1155 		}
1156 
1157 		raw_spin_unlock_irq(&sem->wait_lock);
1158 
1159 		if (signal_pending_state(state, current))
1160 			goto out_nolock;
1161 
1162 		/*
1163 		 * After setting the handoff bit and failing to acquire
1164 		 * the lock, attempt to spin on owner to accelerate lock
1165 		 * transfer. If the previous owner is a on-cpu writer and it
1166 		 * has just released the lock, OWNER_NULL will be returned.
1167 		 * In this case, we attempt to acquire the lock again
1168 		 * without sleeping.
1169 		 */
1170 		if (waiter.handoff_set) {
1171 			enum owner_state owner_state;
1172 
1173 			owner_state = rwsem_spin_on_owner(sem);
1174 			if (owner_state == OWNER_NULL)
1175 				goto trylock_again;
1176 		}
1177 
1178 		schedule_preempt_disabled();
1179 		lockevent_inc(rwsem_sleep_writer);
1180 		set_current_state(state);
1181 trylock_again:
1182 		raw_spin_lock_irq(&sem->wait_lock);
1183 	}
1184 	__set_current_state(TASK_RUNNING);
1185 	raw_spin_unlock_irq(&sem->wait_lock);
1186 	lockevent_inc(rwsem_wlock);
1187 	trace_contention_end(sem, 0);
1188 	return sem;
1189 
1190 out_nolock:
1191 	__set_current_state(TASK_RUNNING);
1192 	raw_spin_lock_irq(&sem->wait_lock);
1193 	rwsem_del_wake_waiter(sem, &waiter, &wake_q);
1194 	lockevent_inc(rwsem_wlock_fail);
1195 	trace_contention_end(sem, -EINTR);
1196 	return ERR_PTR(-EINTR);
1197 }
1198 
1199 /*
1200  * handle waking up a waiter on the semaphore
1201  * - up_read/up_write has decremented the active part of count if we come here
1202  */
rwsem_wake(struct rw_semaphore * sem)1203 static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1204 {
1205 	unsigned long flags;
1206 	DEFINE_WAKE_Q(wake_q);
1207 
1208 	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1209 
1210 	if (!list_empty(&sem->wait_list))
1211 		rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
1212 
1213 	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1214 	wake_up_q(&wake_q);
1215 
1216 	return sem;
1217 }
1218 
1219 /*
1220  * downgrade a write lock into a read lock
1221  * - caller incremented waiting part of count and discovered it still negative
1222  * - just wake up any readers at the front of the queue
1223  */
rwsem_downgrade_wake(struct rw_semaphore * sem)1224 static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1225 {
1226 	unsigned long flags;
1227 	DEFINE_WAKE_Q(wake_q);
1228 
1229 	raw_spin_lock_irqsave(&sem->wait_lock, flags);
1230 
1231 	if (!list_empty(&sem->wait_list))
1232 		rwsem_mark_wake(sem, RWSEM_WAKE_READ_OWNED, &wake_q);
1233 
1234 	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1235 	wake_up_q(&wake_q);
1236 
1237 	return sem;
1238 }
1239 
1240 /*
1241  * lock for reading
1242  */
__down_read_common(struct rw_semaphore * sem,int state)1243 static __always_inline int __down_read_common(struct rw_semaphore *sem, int state)
1244 {
1245 	int ret = 0;
1246 	long count;
1247 
1248 	preempt_disable();
1249 	if (!rwsem_read_trylock(sem, &count)) {
1250 		if (IS_ERR(rwsem_down_read_slowpath(sem, count, state))) {
1251 			ret = -EINTR;
1252 			goto out;
1253 		}
1254 		DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1255 	}
1256 out:
1257 	preempt_enable();
1258 	return ret;
1259 }
1260 
__down_read(struct rw_semaphore * sem)1261 static __always_inline void __down_read(struct rw_semaphore *sem)
1262 {
1263 	__down_read_common(sem, TASK_UNINTERRUPTIBLE);
1264 }
1265 
__down_read_interruptible(struct rw_semaphore * sem)1266 static __always_inline int __down_read_interruptible(struct rw_semaphore *sem)
1267 {
1268 	return __down_read_common(sem, TASK_INTERRUPTIBLE);
1269 }
1270 
__down_read_killable(struct rw_semaphore * sem)1271 static __always_inline int __down_read_killable(struct rw_semaphore *sem)
1272 {
1273 	return __down_read_common(sem, TASK_KILLABLE);
1274 }
1275 
__down_read_trylock(struct rw_semaphore * sem)1276 static inline int __down_read_trylock(struct rw_semaphore *sem)
1277 {
1278 	int ret = 0;
1279 	long tmp;
1280 
1281 	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1282 
1283 	preempt_disable();
1284 	tmp = atomic_long_read(&sem->count);
1285 	while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1286 		if (atomic_long_try_cmpxchg_acquire(&sem->count, &tmp,
1287 						    tmp + RWSEM_READER_BIAS)) {
1288 			rwsem_set_reader_owned(sem);
1289 			ret = 1;
1290 			break;
1291 		}
1292 	}
1293 	preempt_enable();
1294 	return ret;
1295 }
1296 
1297 /*
1298  * lock for writing
1299  */
__down_write_common(struct rw_semaphore * sem,int state)1300 static __always_inline int __down_write_common(struct rw_semaphore *sem, int state)
1301 {
1302 	int ret = 0;
1303 
1304 	preempt_disable();
1305 	if (unlikely(!rwsem_write_trylock(sem))) {
1306 		if (IS_ERR(rwsem_down_write_slowpath(sem, state)))
1307 			ret = -EINTR;
1308 	}
1309 	preempt_enable();
1310 	return ret;
1311 }
1312 
__down_write(struct rw_semaphore * sem)1313 static __always_inline void __down_write(struct rw_semaphore *sem)
1314 {
1315 	__down_write_common(sem, TASK_UNINTERRUPTIBLE);
1316 }
1317 
__down_write_killable(struct rw_semaphore * sem)1318 static __always_inline int __down_write_killable(struct rw_semaphore *sem)
1319 {
1320 	return __down_write_common(sem, TASK_KILLABLE);
1321 }
1322 
__down_write_trylock(struct rw_semaphore * sem)1323 static inline int __down_write_trylock(struct rw_semaphore *sem)
1324 {
1325 	int ret;
1326 
1327 	preempt_disable();
1328 	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1329 	ret = rwsem_write_trylock(sem);
1330 	preempt_enable();
1331 
1332 	return ret;
1333 }
1334 
1335 /*
1336  * unlock after reading
1337  */
__up_read(struct rw_semaphore * sem)1338 static inline void __up_read(struct rw_semaphore *sem)
1339 {
1340 	long tmp;
1341 
1342 	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1343 	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1344 
1345 	preempt_disable();
1346 	rwsem_clear_reader_owned(sem);
1347 	tmp = atomic_long_add_return_release(-RWSEM_READER_BIAS, &sem->count);
1348 	DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1349 	if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1350 		      RWSEM_FLAG_WAITERS)) {
1351 		clear_nonspinnable(sem);
1352 		rwsem_wake(sem);
1353 	}
1354 	preempt_enable();
1355 }
1356 
1357 /*
1358  * unlock after writing
1359  */
__up_write(struct rw_semaphore * sem)1360 static inline void __up_write(struct rw_semaphore *sem)
1361 {
1362 	long tmp;
1363 
1364 	DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1365 	/*
1366 	 * sem->owner may differ from current if the ownership is transferred
1367 	 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1368 	 */
1369 	DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1370 			    !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1371 
1372 	preempt_disable();
1373 	rwsem_clear_owner(sem);
1374 	tmp = atomic_long_fetch_add_release(-RWSEM_WRITER_LOCKED, &sem->count);
1375 	if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1376 		rwsem_wake(sem);
1377 	preempt_enable();
1378 }
1379 
1380 /*
1381  * downgrade write lock to read lock
1382  */
__downgrade_write(struct rw_semaphore * sem)1383 static inline void __downgrade_write(struct rw_semaphore *sem)
1384 {
1385 	long tmp;
1386 
1387 	/*
1388 	 * When downgrading from exclusive to shared ownership,
1389 	 * anything inside the write-locked region cannot leak
1390 	 * into the read side. In contrast, anything in the
1391 	 * read-locked region is ok to be re-ordered into the
1392 	 * write side. As such, rely on RELEASE semantics.
1393 	 */
1394 	DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1395 	preempt_disable();
1396 	tmp = atomic_long_fetch_add_release(
1397 		-RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, &sem->count);
1398 	rwsem_set_reader_owned(sem);
1399 	if (tmp & RWSEM_FLAG_WAITERS)
1400 		rwsem_downgrade_wake(sem);
1401 	preempt_enable();
1402 }
1403 
1404 #else /* !CONFIG_PREEMPT_RT */
1405 
1406 #define RT_MUTEX_BUILD_MUTEX
1407 #include "rtmutex.c"
1408 
1409 #define rwbase_set_and_save_current_state(state)	\
1410 	set_current_state(state)
1411 
1412 #define rwbase_restore_current_state()			\
1413 	__set_current_state(TASK_RUNNING)
1414 
1415 #define rwbase_rtmutex_lock_state(rtm, state)		\
1416 	__rt_mutex_lock(rtm, state)
1417 
1418 #define rwbase_rtmutex_slowlock_locked(rtm, state)	\
1419 	__rt_mutex_slowlock_locked(rtm, NULL, state)
1420 
1421 #define rwbase_rtmutex_unlock(rtm)			\
1422 	__rt_mutex_unlock(rtm)
1423 
1424 #define rwbase_rtmutex_trylock(rtm)			\
1425 	__rt_mutex_trylock(rtm)
1426 
1427 #define rwbase_signal_pending_state(state, current)	\
1428 	signal_pending_state(state, current)
1429 
1430 #define rwbase_schedule()				\
1431 	schedule()
1432 
1433 #include "rwbase_rt.c"
1434 
__init_rwsem(struct rw_semaphore * sem,const char * name,struct lock_class_key * key)1435 void __init_rwsem(struct rw_semaphore *sem, const char *name,
1436 		  struct lock_class_key *key)
1437 {
1438 	init_rwbase_rt(&(sem)->rwbase);
1439 
1440 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1441 	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1442 	lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1443 #endif
1444 }
1445 EXPORT_SYMBOL(__init_rwsem);
1446 
__down_read(struct rw_semaphore * sem)1447 static inline void __down_read(struct rw_semaphore *sem)
1448 {
1449 	rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1450 }
1451 
__down_read_interruptible(struct rw_semaphore * sem)1452 static inline int __down_read_interruptible(struct rw_semaphore *sem)
1453 {
1454 	return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1455 }
1456 
__down_read_killable(struct rw_semaphore * sem)1457 static inline int __down_read_killable(struct rw_semaphore *sem)
1458 {
1459 	return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1460 }
1461 
__down_read_trylock(struct rw_semaphore * sem)1462 static inline int __down_read_trylock(struct rw_semaphore *sem)
1463 {
1464 	return rwbase_read_trylock(&sem->rwbase);
1465 }
1466 
__up_read(struct rw_semaphore * sem)1467 static inline void __up_read(struct rw_semaphore *sem)
1468 {
1469 	rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1470 }
1471 
__down_write(struct rw_semaphore * sem)1472 static inline void __sched __down_write(struct rw_semaphore *sem)
1473 {
1474 	rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1475 }
1476 
__down_write_killable(struct rw_semaphore * sem)1477 static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1478 {
1479 	return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1480 }
1481 
__down_write_trylock(struct rw_semaphore * sem)1482 static inline int __down_write_trylock(struct rw_semaphore *sem)
1483 {
1484 	return rwbase_write_trylock(&sem->rwbase);
1485 }
1486 
__up_write(struct rw_semaphore * sem)1487 static inline void __up_write(struct rw_semaphore *sem)
1488 {
1489 	rwbase_write_unlock(&sem->rwbase);
1490 }
1491 
__downgrade_write(struct rw_semaphore * sem)1492 static inline void __downgrade_write(struct rw_semaphore *sem)
1493 {
1494 	rwbase_write_downgrade(&sem->rwbase);
1495 }
1496 
1497 /* Debug stubs for the common API */
1498 #define DEBUG_RWSEMS_WARN_ON(c, sem)
1499 
__rwsem_set_reader_owned(struct rw_semaphore * sem,struct task_struct * owner)1500 static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1501 					    struct task_struct *owner)
1502 {
1503 }
1504 
is_rwsem_reader_owned(struct rw_semaphore * sem)1505 static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1506 {
1507 	int count = atomic_read(&sem->rwbase.readers);
1508 
1509 	return count < 0 && count != READER_BIAS;
1510 }
1511 
1512 #endif /* CONFIG_PREEMPT_RT */
1513 
1514 /*
1515  * lock for reading
1516  */
down_read(struct rw_semaphore * sem)1517 void __sched down_read(struct rw_semaphore *sem)
1518 {
1519 	might_sleep();
1520 	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1521 
1522 	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1523 }
1524 EXPORT_SYMBOL(down_read);
1525 
down_read_interruptible(struct rw_semaphore * sem)1526 int __sched down_read_interruptible(struct rw_semaphore *sem)
1527 {
1528 	might_sleep();
1529 	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1530 
1531 	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1532 		rwsem_release(&sem->dep_map, _RET_IP_);
1533 		return -EINTR;
1534 	}
1535 
1536 	return 0;
1537 }
1538 EXPORT_SYMBOL(down_read_interruptible);
1539 
down_read_killable(struct rw_semaphore * sem)1540 int __sched down_read_killable(struct rw_semaphore *sem)
1541 {
1542 	might_sleep();
1543 	rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1544 
1545 	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1546 		rwsem_release(&sem->dep_map, _RET_IP_);
1547 		return -EINTR;
1548 	}
1549 
1550 	return 0;
1551 }
1552 EXPORT_SYMBOL(down_read_killable);
1553 
1554 /*
1555  * trylock for reading -- returns 1 if successful, 0 if contention
1556  */
down_read_trylock(struct rw_semaphore * sem)1557 int down_read_trylock(struct rw_semaphore *sem)
1558 {
1559 	int ret = __down_read_trylock(sem);
1560 
1561 	if (ret == 1)
1562 		rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1563 	return ret;
1564 }
1565 EXPORT_SYMBOL(down_read_trylock);
1566 
1567 /*
1568  * lock for writing
1569  */
down_write(struct rw_semaphore * sem)1570 void __sched down_write(struct rw_semaphore *sem)
1571 {
1572 	might_sleep();
1573 	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1574 	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1575 }
1576 EXPORT_SYMBOL(down_write);
1577 
1578 /*
1579  * lock for writing
1580  */
down_write_killable(struct rw_semaphore * sem)1581 int __sched down_write_killable(struct rw_semaphore *sem)
1582 {
1583 	might_sleep();
1584 	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1585 
1586 	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1587 				  __down_write_killable)) {
1588 		rwsem_release(&sem->dep_map, _RET_IP_);
1589 		return -EINTR;
1590 	}
1591 
1592 	return 0;
1593 }
1594 EXPORT_SYMBOL(down_write_killable);
1595 
1596 /*
1597  * trylock for writing -- returns 1 if successful, 0 if contention
1598  */
down_write_trylock(struct rw_semaphore * sem)1599 int down_write_trylock(struct rw_semaphore *sem)
1600 {
1601 	int ret = __down_write_trylock(sem);
1602 
1603 	if (ret == 1)
1604 		rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1605 
1606 	return ret;
1607 }
1608 EXPORT_SYMBOL(down_write_trylock);
1609 
1610 /*
1611  * release a read lock
1612  */
up_read(struct rw_semaphore * sem)1613 void up_read(struct rw_semaphore *sem)
1614 {
1615 	rwsem_release(&sem->dep_map, _RET_IP_);
1616 	__up_read(sem);
1617 }
1618 EXPORT_SYMBOL(up_read);
1619 
1620 /*
1621  * release a write lock
1622  */
up_write(struct rw_semaphore * sem)1623 void up_write(struct rw_semaphore *sem)
1624 {
1625 	rwsem_release(&sem->dep_map, _RET_IP_);
1626 	__up_write(sem);
1627 }
1628 EXPORT_SYMBOL(up_write);
1629 
1630 /*
1631  * downgrade write lock to read lock
1632  */
downgrade_write(struct rw_semaphore * sem)1633 void downgrade_write(struct rw_semaphore *sem)
1634 {
1635 	lock_downgrade(&sem->dep_map, _RET_IP_);
1636 	__downgrade_write(sem);
1637 }
1638 EXPORT_SYMBOL(downgrade_write);
1639 
1640 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1641 
down_read_nested(struct rw_semaphore * sem,int subclass)1642 void down_read_nested(struct rw_semaphore *sem, int subclass)
1643 {
1644 	might_sleep();
1645 	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1646 	LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1647 }
1648 EXPORT_SYMBOL(down_read_nested);
1649 
down_read_killable_nested(struct rw_semaphore * sem,int subclass)1650 int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1651 {
1652 	might_sleep();
1653 	rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1654 
1655 	if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1656 		rwsem_release(&sem->dep_map, _RET_IP_);
1657 		return -EINTR;
1658 	}
1659 
1660 	return 0;
1661 }
1662 EXPORT_SYMBOL(down_read_killable_nested);
1663 
_down_write_nest_lock(struct rw_semaphore * sem,struct lockdep_map * nest)1664 void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1665 {
1666 	might_sleep();
1667 	rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1668 	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1669 }
1670 EXPORT_SYMBOL(_down_write_nest_lock);
1671 
down_read_non_owner(struct rw_semaphore * sem)1672 void down_read_non_owner(struct rw_semaphore *sem)
1673 {
1674 	might_sleep();
1675 	__down_read(sem);
1676 	/*
1677 	 * The owner value for a reader-owned lock is mostly for debugging
1678 	 * purpose only and is not critical to the correct functioning of
1679 	 * rwsem. So it is perfectly fine to set it in a preempt-enabled
1680 	 * context here.
1681 	 */
1682 	__rwsem_set_reader_owned(sem, NULL);
1683 }
1684 EXPORT_SYMBOL(down_read_non_owner);
1685 
down_write_nested(struct rw_semaphore * sem,int subclass)1686 void down_write_nested(struct rw_semaphore *sem, int subclass)
1687 {
1688 	might_sleep();
1689 	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1690 	LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1691 }
1692 EXPORT_SYMBOL(down_write_nested);
1693 
down_write_killable_nested(struct rw_semaphore * sem,int subclass)1694 int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1695 {
1696 	might_sleep();
1697 	rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1698 
1699 	if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1700 				  __down_write_killable)) {
1701 		rwsem_release(&sem->dep_map, _RET_IP_);
1702 		return -EINTR;
1703 	}
1704 
1705 	return 0;
1706 }
1707 EXPORT_SYMBOL(down_write_killable_nested);
1708 
up_read_non_owner(struct rw_semaphore * sem)1709 void up_read_non_owner(struct rw_semaphore *sem)
1710 {
1711 	DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1712 	__up_read(sem);
1713 }
1714 EXPORT_SYMBOL(up_read_non_owner);
1715 
1716 #endif
1717