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