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 <linux/hung_task.h>
31#include <trace/events/lock.h>
32
33#ifndef CONFIG_PREEMPT_RT
34#include "lock_events.h"
35
36/*
37 * The least significant 2 bits of the owner value has the following
38 * meanings when set.
39 * - Bit 0: RWSEM_READER_OWNED - rwsem may be owned by readers (just a hint)
40 * - Bit 1: RWSEM_NONSPINNABLE - Cannot spin on a reader-owned lock
41 *
42 * When the rwsem is reader-owned and a spinning writer has timed out,
43 * the nonspinnable bit will be set to disable optimistic spinning.
44
45 * When a writer acquires a rwsem, it puts its task_struct pointer
46 * into the owner field. It is cleared after an unlock.
47 *
48 * When a reader acquires a rwsem, it will also puts its task_struct
49 * pointer into the owner field with the RWSEM_READER_OWNED bit set.
50 * On unlock, the owner field will largely be left untouched. So
51 * for a free or reader-owned rwsem, the owner value may contain
52 * information about the last reader that acquires the rwsem.
53 *
54 * That information may be helpful in debugging cases where the system
55 * seems to hang on a reader owned rwsem especially if only one reader
56 * is involved. Ideally we would like to track all the readers that own
57 * a rwsem, but the overhead is simply too big.
58 *
59 * A fast path reader optimistic lock stealing is supported when the rwsem
60 * is previously owned by a writer and the following conditions are met:
61 * - rwsem is not currently writer owned
62 * - the handoff isn't set.
63 */
64#define RWSEM_READER_OWNED (1UL << 0)
65#define RWSEM_NONSPINNABLE (1UL << 1)
66#define RWSEM_OWNER_FLAGS_MASK (RWSEM_READER_OWNED | RWSEM_NONSPINNABLE)
67
68#ifdef CONFIG_DEBUG_RWSEMS
69# define DEBUG_RWSEMS_WARN_ON(c, sem) do { \
70 if (!debug_locks_silent && \
71 WARN_ONCE(c, "DEBUG_RWSEMS_WARN_ON(%s): count = 0x%lx, magic = 0x%lx, owner = 0x%lx, curr 0x%lx, list %sempty\n",\
72 #c, atomic_long_read(&(sem)->count), \
73 (unsigned long) sem->magic, \
74 atomic_long_read(&(sem)->owner), (long)current, \
75 list_empty(&(sem)->wait_list) ? "" : "not ")) \
76 debug_locks_off(); \
77 } while (0)
78#else
79# define DEBUG_RWSEMS_WARN_ON(c, sem)
80#endif
81
82/*
83 * On 64-bit architectures, the bit definitions of the count are:
84 *
85 * Bit 0 - writer locked bit
86 * Bit 1 - waiters present bit
87 * Bit 2 - lock handoff bit
88 * Bits 3-7 - reserved
89 * Bits 8-62 - 55-bit reader count
90 * Bit 63 - read fail bit
91 *
92 * On 32-bit architectures, the bit definitions of the count are:
93 *
94 * Bit 0 - writer locked bit
95 * Bit 1 - waiters present bit
96 * Bit 2 - lock handoff bit
97 * Bits 3-7 - reserved
98 * Bits 8-30 - 23-bit reader count
99 * Bit 31 - read fail bit
100 *
101 * It is not likely that the most significant bit (read fail bit) will ever
102 * be set. This guard bit is still checked anyway in the down_read() fastpath
103 * just in case we need to use up more of the reader bits for other purpose
104 * in the future.
105 *
106 * atomic_long_fetch_add() is used to obtain reader lock, whereas
107 * atomic_long_cmpxchg() will be used to obtain writer lock.
108 *
109 * There are three places where the lock handoff bit may be set or cleared.
110 * 1) rwsem_mark_wake() for readers -- set, clear
111 * 2) rwsem_try_write_lock() for writers -- set, clear
112 * 3) rwsem_del_waiter() -- clear
113 *
114 * For all the above cases, wait_lock will be held. A writer must also
115 * be the first one in the wait_list to be eligible for setting the handoff
116 * bit. So concurrent setting/clearing of handoff bit is not possible.
117 */
118#define RWSEM_WRITER_LOCKED (1UL << 0)
119#define RWSEM_FLAG_WAITERS (1UL << 1)
120#define RWSEM_FLAG_HANDOFF (1UL << 2)
121#define RWSEM_FLAG_READFAIL (1UL << (BITS_PER_LONG - 1))
122
123#define RWSEM_READER_SHIFT 8
124#define RWSEM_READER_BIAS (1UL << RWSEM_READER_SHIFT)
125#define RWSEM_READER_MASK (~(RWSEM_READER_BIAS - 1))
126#define RWSEM_WRITER_MASK RWSEM_WRITER_LOCKED
127#define RWSEM_LOCK_MASK (RWSEM_WRITER_MASK|RWSEM_READER_MASK)
128#define RWSEM_READ_FAILED_MASK (RWSEM_WRITER_MASK|RWSEM_FLAG_WAITERS|\
129 RWSEM_FLAG_HANDOFF|RWSEM_FLAG_READFAIL)
130
131/*
132 * All writes to owner are protected by WRITE_ONCE() to make sure that
133 * store tearing can't happen as optimistic spinners may read and use
134 * the owner value concurrently without lock. Read from owner, however,
135 * may not need READ_ONCE() as long as the pointer value is only used
136 * for comparison and isn't being dereferenced.
137 *
138 * Both rwsem_{set,clear}_owner() functions should be in the same
139 * preempt disable section as the atomic op that changes sem->count.
140 */
141static inline void rwsem_set_owner(struct rw_semaphore *sem)
142{
143 lockdep_assert_preemption_disabled();
144 atomic_long_set(v: &sem->owner, i: (long)current);
145}
146
147static inline void rwsem_clear_owner(struct rw_semaphore *sem)
148{
149 lockdep_assert_preemption_disabled();
150 atomic_long_set(v: &sem->owner, i: 0);
151}
152
153/*
154 * Test the flags in the owner field.
155 */
156static inline bool rwsem_test_oflags(struct rw_semaphore *sem, long flags)
157{
158 return atomic_long_read(v: &sem->owner) & flags;
159}
160
161/*
162 * The task_struct pointer of the last owning reader will be left in
163 * the owner field.
164 *
165 * Note that the owner value just indicates the task has owned the rwsem
166 * previously, it may not be the real owner or one of the real owners
167 * anymore when that field is examined, so take it with a grain of salt.
168 *
169 * The reader non-spinnable bit is preserved.
170 */
171static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
172 struct task_struct *owner)
173{
174 unsigned long val = (unsigned long)owner | RWSEM_READER_OWNED |
175 (atomic_long_read(v: &sem->owner) & RWSEM_NONSPINNABLE);
176
177 atomic_long_set(v: &sem->owner, i: val);
178}
179
180static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
181{
182 __rwsem_set_reader_owned(sem, current);
183}
184
185#if defined(CONFIG_DEBUG_RWSEMS) || defined(CONFIG_DETECT_HUNG_TASK_BLOCKER)
186/*
187 * Return just the real task structure pointer of the owner
188 */
189struct task_struct *rwsem_owner(struct rw_semaphore *sem)
190{
191 return (struct task_struct *)
192 (atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
193}
194
195/*
196 * Return true if the rwsem is owned by a reader.
197 */
198bool is_rwsem_reader_owned(struct rw_semaphore *sem)
199{
200 /*
201 * Check the count to see if it is write-locked.
202 */
203 long count = atomic_long_read(&sem->count);
204
205 if (count & RWSEM_WRITER_MASK)
206 return false;
207 return rwsem_test_oflags(sem, RWSEM_READER_OWNED);
208}
209
210/*
211 * With CONFIG_DEBUG_RWSEMS or CONFIG_DETECT_HUNG_TASK_BLOCKER configured,
212 * it will make sure that the owner field of a reader-owned rwsem either
213 * points to a real reader-owner(s) or gets cleared. The only exception is
214 * when the unlock is done by up_read_non_owner().
215 */
216static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
217{
218 unsigned long val = atomic_long_read(&sem->owner);
219
220 while ((val & ~RWSEM_OWNER_FLAGS_MASK) == (unsigned long)current) {
221 if (atomic_long_try_cmpxchg(&sem->owner, &val,
222 val & RWSEM_OWNER_FLAGS_MASK))
223 return;
224 }
225}
226#else
227static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
228{
229}
230#endif
231
232/*
233 * Set the RWSEM_NONSPINNABLE bits if the RWSEM_READER_OWNED flag
234 * remains set. Otherwise, the operation will be aborted.
235 */
236static inline void rwsem_set_nonspinnable(struct rw_semaphore *sem)
237{
238 unsigned long owner = atomic_long_read(v: &sem->owner);
239
240 do {
241 if (!(owner & RWSEM_READER_OWNED))
242 break;
243 if (owner & RWSEM_NONSPINNABLE)
244 break;
245 } while (!atomic_long_try_cmpxchg(v: &sem->owner, old: &owner,
246 new: owner | RWSEM_NONSPINNABLE));
247}
248
249static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cntp)
250{
251 *cntp = atomic_long_add_return_acquire(RWSEM_READER_BIAS, v: &sem->count);
252
253 if (WARN_ON_ONCE(*cntp < 0))
254 rwsem_set_nonspinnable(sem);
255
256 if (!(*cntp & RWSEM_READ_FAILED_MASK)) {
257 rwsem_set_reader_owned(sem);
258 return true;
259 }
260
261 return false;
262}
263
264static inline bool rwsem_write_trylock(struct rw_semaphore *sem)
265{
266 long tmp = RWSEM_UNLOCKED_VALUE;
267
268 if (atomic_long_try_cmpxchg_acquire(v: &sem->count, old: &tmp, RWSEM_WRITER_LOCKED)) {
269 rwsem_set_owner(sem);
270 return true;
271 }
272
273 return false;
274}
275
276/*
277 * Return the real task structure pointer of the owner and the embedded
278 * flags in the owner. pflags must be non-NULL.
279 */
280static inline struct task_struct *
281rwsem_owner_flags(struct rw_semaphore *sem, unsigned long *pflags)
282{
283 unsigned long owner = atomic_long_read(v: &sem->owner);
284
285 *pflags = owner & RWSEM_OWNER_FLAGS_MASK;
286 return (struct task_struct *)(owner & ~RWSEM_OWNER_FLAGS_MASK);
287}
288
289/*
290 * Guide to the rw_semaphore's count field.
291 *
292 * When the RWSEM_WRITER_LOCKED bit in count is set, the lock is owned
293 * by a writer.
294 *
295 * The lock is owned by readers when
296 * (1) the RWSEM_WRITER_LOCKED isn't set in count,
297 * (2) some of the reader bits are set in count, and
298 * (3) the owner field has RWSEM_READ_OWNED bit set.
299 *
300 * Having some reader bits set is not enough to guarantee a readers owned
301 * lock as the readers may be in the process of backing out from the count
302 * and a writer has just released the lock. So another writer may steal
303 * the lock immediately after that.
304 */
305
306/*
307 * Initialize an rwsem:
308 */
309void __init_rwsem(struct rw_semaphore *sem, const char *name,
310 struct lock_class_key *key)
311{
312#ifdef CONFIG_DEBUG_LOCK_ALLOC
313 /*
314 * Make sure we are not reinitializing a held semaphore:
315 */
316 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
317 lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
318#endif
319#ifdef CONFIG_DEBUG_RWSEMS
320 sem->magic = sem;
321#endif
322 atomic_long_set(v: &sem->count, RWSEM_UNLOCKED_VALUE);
323 raw_spin_lock_init(&sem->wait_lock);
324 INIT_LIST_HEAD(list: &sem->wait_list);
325 atomic_long_set(v: &sem->owner, i: 0L);
326#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
327 osq_lock_init(lock: &sem->osq);
328#endif
329}
330EXPORT_SYMBOL(__init_rwsem);
331
332enum rwsem_waiter_type {
333 RWSEM_WAITING_FOR_WRITE,
334 RWSEM_WAITING_FOR_READ
335};
336
337struct rwsem_waiter {
338 struct list_head list;
339 struct task_struct *task;
340 enum rwsem_waiter_type type;
341 unsigned long timeout;
342 bool handoff_set;
343};
344#define rwsem_first_waiter(sem) \
345 list_first_entry(&sem->wait_list, struct rwsem_waiter, list)
346
347enum rwsem_wake_type {
348 RWSEM_WAKE_ANY, /* Wake whatever's at head of wait list */
349 RWSEM_WAKE_READERS, /* Wake readers only */
350 RWSEM_WAKE_READ_OWNED /* Waker thread holds the read lock */
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
368static inline void
369rwsem_add_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
370{
371 lockdep_assert_held(&sem->wait_lock);
372 list_add_tail(new: &waiter->list, head: &sem->wait_list);
373 /* caller will set RWSEM_FLAG_WAITERS */
374}
375
376/*
377 * Remove a waiter from the wait_list and clear flags.
378 *
379 * Both rwsem_mark_wake() and rwsem_try_write_lock() contain a full 'copy' of
380 * this function. Modify with care.
381 *
382 * Return: true if wait_list isn't empty and false otherwise
383 */
384static inline bool
385rwsem_del_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter)
386{
387 lockdep_assert_held(&sem->wait_lock);
388 list_del(entry: &waiter->list);
389 if (likely(!list_empty(&sem->wait_list)))
390 return true;
391
392 atomic_long_andnot(RWSEM_FLAG_HANDOFF | RWSEM_FLAG_WAITERS, v: &sem->count);
393 return false;
394}
395
396/*
397 * handle the lock release when processes blocked on it that can now run
398 * - if we come here from up_xxxx(), then the RWSEM_FLAG_WAITERS bit must
399 * have been set.
400 * - there must be someone on the queue
401 * - the wait_lock must be held by the caller
402 * - tasks are marked for wakeup, the caller must later invoke wake_up_q()
403 * to actually wakeup the blocked task(s) and drop the reference count,
404 * preferably when the wait_lock is released
405 * - woken process blocks are discarded from the list after having task zeroed
406 * - writers are only marked woken if downgrading is false
407 *
408 * Implies rwsem_del_waiter() for all woken readers.
409 */
410static void rwsem_mark_wake(struct rw_semaphore *sem,
411 enum rwsem_wake_type wake_type,
412 struct wake_q_head *wake_q)
413{
414 struct rwsem_waiter *waiter, *tmp;
415 long oldcount, woken = 0, adjustment = 0;
416 struct list_head wlist;
417
418 lockdep_assert_held(&sem->wait_lock);
419
420 /*
421 * Take a peek at the queue head waiter such that we can determine
422 * the wakeup(s) to perform.
423 */
424 waiter = rwsem_first_waiter(sem);
425
426 if (waiter->type == RWSEM_WAITING_FOR_WRITE) {
427 if (wake_type == RWSEM_WAKE_ANY) {
428 /*
429 * Mark writer at the front of the queue for wakeup.
430 * Until the task is actually later awoken later by
431 * the caller, other writers are able to steal it.
432 * Readers, on the other hand, will block as they
433 * will notice the queued writer.
434 */
435 wake_q_add(head: wake_q, task: waiter->task);
436 lockevent_inc(rwsem_wake_writer);
437 }
438
439 return;
440 }
441
442 /*
443 * No reader wakeup if there are too many of them already.
444 */
445 if (unlikely(atomic_long_read(&sem->count) < 0))
446 return;
447
448 /*
449 * Writers might steal the lock before we grant it to the next reader.
450 * We prefer to do the first reader grant before counting readers
451 * so we can bail out early if a writer stole the lock.
452 */
453 if (wake_type != RWSEM_WAKE_READ_OWNED) {
454 struct task_struct *owner;
455
456 adjustment = RWSEM_READER_BIAS;
457 oldcount = atomic_long_fetch_add(i: adjustment, v: &sem->count);
458 if (unlikely(oldcount & RWSEM_WRITER_MASK)) {
459 /*
460 * When we've been waiting "too" long (for writers
461 * to give up the lock), request a HANDOFF to
462 * force the issue.
463 */
464 if (time_after(jiffies, waiter->timeout)) {
465 if (!(oldcount & RWSEM_FLAG_HANDOFF)) {
466 adjustment -= RWSEM_FLAG_HANDOFF;
467 lockevent_inc(rwsem_rlock_handoff);
468 }
469 waiter->handoff_set = true;
470 }
471
472 atomic_long_add(i: -adjustment, v: &sem->count);
473 return;
474 }
475 /*
476 * Set it to reader-owned to give spinners an early
477 * indication that readers now have the lock.
478 * The reader nonspinnable bit seen at slowpath entry of
479 * the reader is copied over.
480 */
481 owner = waiter->task;
482 __rwsem_set_reader_owned(sem, owner);
483 }
484
485 /*
486 * Grant up to MAX_READERS_WAKEUP read locks to all the readers in the
487 * queue. We know that the woken will be at least 1 as we accounted
488 * for above. Note we increment the 'active part' of the count by the
489 * number of readers before waking any processes up.
490 *
491 * This is an adaptation of the phase-fair R/W locks where at the
492 * reader phase (first waiter is a reader), all readers are eligible
493 * to acquire the lock at the same time irrespective of their order
494 * in the queue. The writers acquire the lock according to their
495 * order in the queue.
496 *
497 * We have to do wakeup in 2 passes to prevent the possibility that
498 * the reader count may be decremented before it is incremented. It
499 * is because the to-be-woken waiter may not have slept yet. So it
500 * may see waiter->task got cleared, finish its critical section and
501 * do an unlock before the reader count increment.
502 *
503 * 1) Collect the read-waiters in a separate list, count them and
504 * fully increment the reader count in rwsem.
505 * 2) For each waiters in the new list, clear waiter->task and
506 * put them into wake_q to be woken up later.
507 */
508 INIT_LIST_HEAD(list: &wlist);
509 list_for_each_entry_safe(waiter, tmp, &sem->wait_list, list) {
510 if (waiter->type == RWSEM_WAITING_FOR_WRITE)
511 continue;
512
513 woken++;
514 list_move_tail(list: &waiter->list, head: &wlist);
515
516 /*
517 * Limit # of readers that can be woken up per wakeup call.
518 */
519 if (unlikely(woken >= MAX_READERS_WAKEUP))
520 break;
521 }
522
523 adjustment = woken * RWSEM_READER_BIAS - adjustment;
524 lockevent_cond_inc(rwsem_wake_reader, woken);
525
526 oldcount = atomic_long_read(v: &sem->count);
527 if (list_empty(head: &sem->wait_list)) {
528 /*
529 * Combined with list_move_tail() above, this implies
530 * rwsem_del_waiter().
531 */
532 adjustment -= RWSEM_FLAG_WAITERS;
533 if (oldcount & RWSEM_FLAG_HANDOFF)
534 adjustment -= RWSEM_FLAG_HANDOFF;
535 } else if (woken) {
536 /*
537 * When we've woken a reader, we no longer need to force
538 * writers to give up the lock and we can clear HANDOFF.
539 */
540 if (oldcount & RWSEM_FLAG_HANDOFF)
541 adjustment -= RWSEM_FLAG_HANDOFF;
542 }
543
544 if (adjustment)
545 atomic_long_add(i: adjustment, v: &sem->count);
546
547 /* 2nd pass */
548 list_for_each_entry_safe(waiter, tmp, &wlist, list) {
549 struct task_struct *tsk;
550
551 tsk = waiter->task;
552 get_task_struct(t: tsk);
553
554 /*
555 * Ensure calling get_task_struct() before setting the reader
556 * waiter to nil such that rwsem_down_read_slowpath() cannot
557 * race with do_exit() by always holding a reference count
558 * to the task to wakeup.
559 */
560 smp_store_release(&waiter->task, NULL);
561 /*
562 * Ensure issuing the wakeup (either by us or someone else)
563 * after setting the reader waiter to nil.
564 */
565 wake_q_add_safe(head: wake_q, task: tsk);
566 }
567}
568
569/*
570 * Remove a waiter and try to wake up other waiters in the wait queue
571 * This function is called from the out_nolock path of both the reader and
572 * writer slowpaths with wait_lock held. It releases the wait_lock and
573 * optionally wake up waiters before it returns.
574 */
575static inline void
576rwsem_del_wake_waiter(struct rw_semaphore *sem, struct rwsem_waiter *waiter,
577 struct wake_q_head *wake_q)
578 __releases(&sem->wait_lock)
579{
580 bool first = rwsem_first_waiter(sem) == waiter;
581
582 wake_q_init(head: wake_q);
583
584 /*
585 * If the wait_list isn't empty and the waiter to be deleted is
586 * the first waiter, we wake up the remaining waiters as they may
587 * be eligible to acquire or spin on the lock.
588 */
589 if (rwsem_del_waiter(sem, waiter) && first)
590 rwsem_mark_wake(sem, wake_type: RWSEM_WAKE_ANY, wake_q);
591 raw_spin_unlock_irq(&sem->wait_lock);
592 if (!wake_q_empty(head: wake_q))
593 wake_up_q(head: wake_q);
594}
595
596/*
597 * This function must be called with the sem->wait_lock held to prevent
598 * race conditions between checking the rwsem wait list and setting the
599 * sem->count accordingly.
600 *
601 * Implies rwsem_del_waiter() on success.
602 */
603static inline bool rwsem_try_write_lock(struct rw_semaphore *sem,
604 struct rwsem_waiter *waiter)
605{
606 struct rwsem_waiter *first = rwsem_first_waiter(sem);
607 long count, new;
608
609 lockdep_assert_held(&sem->wait_lock);
610
611 count = atomic_long_read(v: &sem->count);
612 do {
613 bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
614
615 if (has_handoff) {
616 /*
617 * Honor handoff bit and yield only when the first
618 * waiter is the one that set it. Otherwisee, we
619 * still try to acquire the rwsem.
620 */
621 if (first->handoff_set && (waiter != first))
622 return false;
623 }
624
625 new = count;
626
627 if (count & RWSEM_LOCK_MASK) {
628 /*
629 * A waiter (first or not) can set the handoff bit
630 * if it is an RT task or wait in the wait queue
631 * for too long.
632 */
633 if (has_handoff || (!rt_or_dl_task(p: waiter->task) &&
634 !time_after(jiffies, waiter->timeout)))
635 return false;
636
637 new |= RWSEM_FLAG_HANDOFF;
638 } else {
639 new |= RWSEM_WRITER_LOCKED;
640 new &= ~RWSEM_FLAG_HANDOFF;
641
642 if (list_is_singular(head: &sem->wait_list))
643 new &= ~RWSEM_FLAG_WAITERS;
644 }
645 } while (!atomic_long_try_cmpxchg_acquire(v: &sem->count, old: &count, new));
646
647 /*
648 * We have either acquired the lock with handoff bit cleared or set
649 * the handoff bit. Only the first waiter can have its handoff_set
650 * set here to enable optimistic spinning in slowpath loop.
651 */
652 if (new & RWSEM_FLAG_HANDOFF) {
653 first->handoff_set = true;
654 lockevent_inc(rwsem_wlock_handoff);
655 return false;
656 }
657
658 /*
659 * Have rwsem_try_write_lock() fully imply rwsem_del_waiter() on
660 * success.
661 */
662 list_del(entry: &waiter->list);
663 rwsem_set_owner(sem);
664 return true;
665}
666
667/*
668 * The rwsem_spin_on_owner() function returns the following 4 values
669 * depending on the lock owner state.
670 * OWNER_NULL : owner is currently NULL
671 * OWNER_WRITER: when owner changes and is a writer
672 * OWNER_READER: when owner changes and the new owner may be a reader.
673 * OWNER_NONSPINNABLE:
674 * when optimistic spinning has to stop because either the
675 * owner stops running, is unknown, or its timeslice has
676 * been used up.
677 */
678enum owner_state {
679 OWNER_NULL = 1 << 0,
680 OWNER_WRITER = 1 << 1,
681 OWNER_READER = 1 << 2,
682 OWNER_NONSPINNABLE = 1 << 3,
683};
684
685#ifdef CONFIG_RWSEM_SPIN_ON_OWNER
686/*
687 * Try to acquire write lock before the writer has been put on wait queue.
688 */
689static inline bool rwsem_try_write_lock_unqueued(struct rw_semaphore *sem)
690{
691 long count = atomic_long_read(v: &sem->count);
692
693 while (!(count & (RWSEM_LOCK_MASK|RWSEM_FLAG_HANDOFF))) {
694 if (atomic_long_try_cmpxchg_acquire(v: &sem->count, old: &count,
695 new: count | RWSEM_WRITER_LOCKED)) {
696 rwsem_set_owner(sem);
697 lockevent_inc(rwsem_opt_lock);
698 return true;
699 }
700 }
701 return false;
702}
703
704static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
705{
706 struct task_struct *owner;
707 unsigned long flags;
708 bool ret = true;
709
710 if (need_resched()) {
711 lockevent_inc(rwsem_opt_fail);
712 return false;
713 }
714
715 /*
716 * Disable preemption is equal to the RCU read-side crital section,
717 * thus the task_strcut structure won't go away.
718 */
719 owner = rwsem_owner_flags(sem, pflags: &flags);
720 /*
721 * Don't check the read-owner as the entry may be stale.
722 */
723 if ((flags & RWSEM_NONSPINNABLE) ||
724 (owner && !(flags & RWSEM_READER_OWNED) && !owner_on_cpu(owner)))
725 ret = false;
726
727 lockevent_cond_inc(rwsem_opt_fail, !ret);
728 return ret;
729}
730
731static inline enum owner_state
732rwsem_owner_state(struct task_struct *owner, unsigned long flags)
733{
734 if (flags & RWSEM_NONSPINNABLE)
735 return OWNER_NONSPINNABLE;
736
737 if (flags & RWSEM_READER_OWNED)
738 return OWNER_READER;
739
740 return owner ? OWNER_WRITER : OWNER_NULL;
741}
742
743static noinline enum owner_state
744rwsem_spin_on_owner(struct rw_semaphore *sem)
745{
746 struct task_struct *new, *owner;
747 unsigned long flags, new_flags;
748 enum owner_state state;
749
750 lockdep_assert_preemption_disabled();
751
752 owner = rwsem_owner_flags(sem, pflags: &flags);
753 state = rwsem_owner_state(owner, flags);
754 if (state != OWNER_WRITER)
755 return state;
756
757 for (;;) {
758 /*
759 * When a waiting writer set the handoff flag, it may spin
760 * on the owner as well. Once that writer acquires the lock,
761 * we can spin on it. So we don't need to quit even when the
762 * handoff bit is set.
763 */
764 new = rwsem_owner_flags(sem, pflags: &new_flags);
765 if ((new != owner) || (new_flags != flags)) {
766 state = rwsem_owner_state(owner: new, flags: new_flags);
767 break;
768 }
769
770 /*
771 * Ensure we emit the owner->on_cpu, dereference _after_
772 * checking sem->owner still matches owner, if that fails,
773 * owner might point to free()d memory, if it still matches,
774 * our spinning context already disabled preemption which is
775 * equal to RCU read-side crital section ensures the memory
776 * stays valid.
777 */
778 barrier();
779
780 if (need_resched() || !owner_on_cpu(owner)) {
781 state = OWNER_NONSPINNABLE;
782 break;
783 }
784
785 cpu_relax();
786 }
787
788 return state;
789}
790
791/*
792 * Calculate reader-owned rwsem spinning threshold for writer
793 *
794 * The more readers own the rwsem, the longer it will take for them to
795 * wind down and free the rwsem. So the empirical formula used to
796 * determine the actual spinning time limit here is:
797 *
798 * Spinning threshold = (10 + nr_readers/2)us
799 *
800 * The limit is capped to a maximum of 25us (30 readers). This is just
801 * a heuristic and is subjected to change in the future.
802 */
803static inline u64 rwsem_rspin_threshold(struct rw_semaphore *sem)
804{
805 long count = atomic_long_read(v: &sem->count);
806 int readers = count >> RWSEM_READER_SHIFT;
807 u64 delta;
808
809 if (readers > 30)
810 readers = 30;
811 delta = (20 + readers) * NSEC_PER_USEC / 2;
812
813 return sched_clock() + delta;
814}
815
816static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
817{
818 bool taken = false;
819 int prev_owner_state = OWNER_NULL;
820 int loop = 0;
821 u64 rspin_threshold = 0;
822
823 /* sem->wait_lock should not be held when doing optimistic spinning */
824 if (!osq_lock(lock: &sem->osq))
825 goto done;
826
827 /*
828 * Optimistically spin on the owner field and attempt to acquire the
829 * lock whenever the owner changes. Spinning will be stopped when:
830 * 1) the owning writer isn't running; or
831 * 2) readers own the lock and spinning time has exceeded limit.
832 */
833 for (;;) {
834 enum owner_state owner_state;
835
836 owner_state = rwsem_spin_on_owner(sem);
837 if (owner_state == OWNER_NONSPINNABLE)
838 break;
839
840 /*
841 * Try to acquire the lock
842 */
843 taken = rwsem_try_write_lock_unqueued(sem);
844
845 if (taken)
846 break;
847
848 /*
849 * Time-based reader-owned rwsem optimistic spinning
850 */
851 if (owner_state == OWNER_READER) {
852 /*
853 * Re-initialize rspin_threshold every time when
854 * the owner state changes from non-reader to reader.
855 * This allows a writer to steal the lock in between
856 * 2 reader phases and have the threshold reset at
857 * the beginning of the 2nd reader phase.
858 */
859 if (prev_owner_state != OWNER_READER) {
860 if (rwsem_test_oflags(sem, RWSEM_NONSPINNABLE))
861 break;
862 rspin_threshold = rwsem_rspin_threshold(sem);
863 loop = 0;
864 }
865
866 /*
867 * Check time threshold once every 16 iterations to
868 * avoid calling sched_clock() too frequently so
869 * as to reduce the average latency between the times
870 * when the lock becomes free and when the spinner
871 * is ready to do a trylock.
872 */
873 else if (!(++loop & 0xf) && (sched_clock() > rspin_threshold)) {
874 rwsem_set_nonspinnable(sem);
875 lockevent_inc(rwsem_opt_nospin);
876 break;
877 }
878 }
879
880 /*
881 * An RT task cannot do optimistic spinning if it cannot
882 * be sure the lock holder is running or live-lock may
883 * happen if the current task and the lock holder happen
884 * to run in the same CPU. However, aborting optimistic
885 * spinning while a NULL owner is detected may miss some
886 * opportunity where spinning can continue without causing
887 * problem.
888 *
889 * There are 2 possible cases where an RT task may be able
890 * to continue spinning.
891 *
892 * 1) The lock owner is in the process of releasing the
893 * lock, sem->owner is cleared but the lock has not
894 * been released yet.
895 * 2) The lock was free and owner cleared, but another
896 * task just comes in and acquire the lock before
897 * we try to get it. The new owner may be a spinnable
898 * writer.
899 *
900 * To take advantage of two scenarios listed above, the RT
901 * task is made to retry one more time to see if it can
902 * acquire the lock or continue spinning on the new owning
903 * writer. Of course, if the time lag is long enough or the
904 * new owner is not a writer or spinnable, the RT task will
905 * quit spinning.
906 *
907 * If the owner is a writer, the need_resched() check is
908 * done inside rwsem_spin_on_owner(). If the owner is not
909 * a writer, need_resched() check needs to be done here.
910 */
911 if (owner_state != OWNER_WRITER) {
912 if (need_resched())
913 break;
914 if (rt_or_dl_task(current) &&
915 (prev_owner_state != OWNER_WRITER))
916 break;
917 }
918 prev_owner_state = owner_state;
919
920 /*
921 * The cpu_relax() call is a compiler barrier which forces
922 * everything in this loop to be re-loaded. We don't need
923 * memory barriers as we'll eventually observe the right
924 * values at the cost of a few extra spins.
925 */
926 cpu_relax();
927 }
928 osq_unlock(lock: &sem->osq);
929done:
930 lockevent_cond_inc(rwsem_opt_fail, !taken);
931 return taken;
932}
933
934/*
935 * Clear the owner's RWSEM_NONSPINNABLE bit if it is set. This should
936 * only be called when the reader count reaches 0.
937 */
938static inline void clear_nonspinnable(struct rw_semaphore *sem)
939{
940 if (unlikely(rwsem_test_oflags(sem, RWSEM_NONSPINNABLE)))
941 atomic_long_andnot(RWSEM_NONSPINNABLE, v: &sem->owner);
942}
943
944#else
945static inline bool rwsem_can_spin_on_owner(struct rw_semaphore *sem)
946{
947 return false;
948}
949
950static inline bool rwsem_optimistic_spin(struct rw_semaphore *sem)
951{
952 return false;
953}
954
955static inline void clear_nonspinnable(struct rw_semaphore *sem) { }
956
957static inline enum owner_state
958rwsem_spin_on_owner(struct rw_semaphore *sem)
959{
960 return OWNER_NONSPINNABLE;
961}
962#endif
963
964/*
965 * Prepare to wake up waiter(s) in the wait queue by putting them into the
966 * given wake_q if the rwsem lock owner isn't a writer. If rwsem is likely
967 * reader-owned, wake up read lock waiters in queue front or wake up any
968 * front waiter otherwise.
969
970 * This is being called from both reader and writer slow paths.
971 */
972static inline void rwsem_cond_wake_waiter(struct rw_semaphore *sem, long count,
973 struct wake_q_head *wake_q)
974{
975 enum rwsem_wake_type wake_type;
976
977 if (count & RWSEM_WRITER_MASK)
978 return;
979
980 if (count & RWSEM_READER_MASK) {
981 wake_type = RWSEM_WAKE_READERS;
982 } else {
983 wake_type = RWSEM_WAKE_ANY;
984 clear_nonspinnable(sem);
985 }
986 rwsem_mark_wake(sem, wake_type, wake_q);
987}
988
989/*
990 * Wait for the read lock to be granted
991 */
992static struct rw_semaphore __sched *
993rwsem_down_read_slowpath(struct rw_semaphore *sem, long count, unsigned int state)
994{
995 long adjustment = -RWSEM_READER_BIAS;
996 long rcnt = (count >> RWSEM_READER_SHIFT);
997 struct rwsem_waiter waiter;
998 DEFINE_WAKE_Q(wake_q);
999
1000 /*
1001 * To prevent a constant stream of readers from starving a sleeping
1002 * writer, don't attempt optimistic lock stealing if the lock is
1003 * very likely owned by readers.
1004 */
1005 if ((atomic_long_read(v: &sem->owner) & RWSEM_READER_OWNED) &&
1006 (rcnt > 1) && !(count & RWSEM_WRITER_LOCKED))
1007 goto queue;
1008
1009 /*
1010 * Reader optimistic lock stealing.
1011 */
1012 if (!(count & (RWSEM_WRITER_LOCKED | RWSEM_FLAG_HANDOFF))) {
1013 rwsem_set_reader_owned(sem);
1014 lockevent_inc(rwsem_rlock_steal);
1015
1016 /*
1017 * Wake up other readers in the wait queue if it is
1018 * the first reader.
1019 */
1020 if ((rcnt == 1) && (count & RWSEM_FLAG_WAITERS)) {
1021 raw_spin_lock_irq(&sem->wait_lock);
1022 if (!list_empty(head: &sem->wait_list))
1023 rwsem_mark_wake(sem, wake_type: RWSEM_WAKE_READ_OWNED,
1024 wake_q: &wake_q);
1025 raw_spin_unlock_irq(&sem->wait_lock);
1026 wake_up_q(head: &wake_q);
1027 }
1028 return sem;
1029 }
1030
1031queue:
1032 waiter.task = current;
1033 waiter.type = RWSEM_WAITING_FOR_READ;
1034 waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1035 waiter.handoff_set = false;
1036
1037 raw_spin_lock_irq(&sem->wait_lock);
1038 if (list_empty(head: &sem->wait_list)) {
1039 /*
1040 * In case the wait queue is empty and the lock isn't owned
1041 * by a writer, this reader can exit the slowpath and return
1042 * immediately as its RWSEM_READER_BIAS has already been set
1043 * in the count.
1044 */
1045 if (!(atomic_long_read(v: &sem->count) & RWSEM_WRITER_MASK)) {
1046 /* Provide lock ACQUIRE */
1047 smp_acquire__after_ctrl_dep();
1048 raw_spin_unlock_irq(&sem->wait_lock);
1049 rwsem_set_reader_owned(sem);
1050 lockevent_inc(rwsem_rlock_fast);
1051 return sem;
1052 }
1053 adjustment += RWSEM_FLAG_WAITERS;
1054 }
1055 rwsem_add_waiter(sem, waiter: &waiter);
1056
1057 /* we're now waiting on the lock, but no longer actively locking */
1058 count = atomic_long_add_return(i: adjustment, v: &sem->count);
1059
1060 rwsem_cond_wake_waiter(sem, count, wake_q: &wake_q);
1061 raw_spin_unlock_irq(&sem->wait_lock);
1062
1063 if (!wake_q_empty(head: &wake_q))
1064 wake_up_q(head: &wake_q);
1065
1066 trace_contention_begin(lock: sem, LCB_F_READ);
1067 set_current_state(state);
1068
1069 if (state == TASK_UNINTERRUPTIBLE)
1070 hung_task_set_blocker(lock: sem, BLOCKER_TYPE_RWSEM_READER);
1071
1072 /* wait to be given the lock */
1073 for (;;) {
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 set_current_state(state);
1089 }
1090
1091 if (state == TASK_UNINTERRUPTIBLE)
1092 hung_task_clear_blocker();
1093
1094 __set_current_state(TASK_RUNNING);
1095 lockevent_inc(rwsem_rlock);
1096 trace_contention_end(lock: sem, ret: 0);
1097 return sem;
1098
1099out_nolock:
1100 rwsem_del_wake_waiter(sem, waiter: &waiter, wake_q: &wake_q);
1101 __set_current_state(TASK_RUNNING);
1102 lockevent_inc(rwsem_rlock_fail);
1103 trace_contention_end(lock: sem, ret: -EINTR);
1104 return ERR_PTR(error: -EINTR);
1105}
1106
1107/*
1108 * Wait until we successfully acquire the write lock
1109 */
1110static struct rw_semaphore __sched *
1111rwsem_down_write_slowpath(struct rw_semaphore *sem, int state)
1112{
1113 struct rwsem_waiter waiter;
1114 DEFINE_WAKE_Q(wake_q);
1115
1116 /* do optimistic spinning and steal lock if possible */
1117 if (rwsem_can_spin_on_owner(sem) && rwsem_optimistic_spin(sem)) {
1118 /* rwsem_optimistic_spin() implies ACQUIRE on success */
1119 return sem;
1120 }
1121
1122 /*
1123 * Optimistic spinning failed, proceed to the slowpath
1124 * and block until we can acquire the sem.
1125 */
1126 waiter.task = current;
1127 waiter.type = RWSEM_WAITING_FOR_WRITE;
1128 waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
1129 waiter.handoff_set = false;
1130
1131 raw_spin_lock_irq(&sem->wait_lock);
1132 rwsem_add_waiter(sem, waiter: &waiter);
1133
1134 /* we're now waiting on the lock */
1135 if (rwsem_first_waiter(sem) != &waiter) {
1136 rwsem_cond_wake_waiter(sem, count: atomic_long_read(v: &sem->count),
1137 wake_q: &wake_q);
1138 if (!wake_q_empty(head: &wake_q)) {
1139 /*
1140 * We want to minimize wait_lock hold time especially
1141 * when a large number of readers are to be woken up.
1142 */
1143 raw_spin_unlock_irq(&sem->wait_lock);
1144 wake_up_q(head: &wake_q);
1145 raw_spin_lock_irq(&sem->wait_lock);
1146 }
1147 } else {
1148 atomic_long_or(RWSEM_FLAG_WAITERS, v: &sem->count);
1149 }
1150
1151 /* wait until we successfully acquire the lock */
1152 set_current_state(state);
1153 trace_contention_begin(lock: sem, LCB_F_WRITE);
1154
1155 if (state == TASK_UNINTERRUPTIBLE)
1156 hung_task_set_blocker(lock: sem, BLOCKER_TYPE_RWSEM_WRITER);
1157
1158 for (;;) {
1159 if (rwsem_try_write_lock(sem, waiter: &waiter)) {
1160 /* rwsem_try_write_lock() implies ACQUIRE on success */
1161 break;
1162 }
1163
1164 raw_spin_unlock_irq(&sem->wait_lock);
1165
1166 if (signal_pending_state(state, current))
1167 goto out_nolock;
1168
1169 /*
1170 * After setting the handoff bit and failing to acquire
1171 * the lock, attempt to spin on owner to accelerate lock
1172 * transfer. If the previous owner is a on-cpu writer and it
1173 * has just released the lock, OWNER_NULL will be returned.
1174 * In this case, we attempt to acquire the lock again
1175 * without sleeping.
1176 */
1177 if (waiter.handoff_set) {
1178 enum owner_state owner_state;
1179
1180 owner_state = rwsem_spin_on_owner(sem);
1181 if (owner_state == OWNER_NULL)
1182 goto trylock_again;
1183 }
1184
1185 schedule_preempt_disabled();
1186 lockevent_inc(rwsem_sleep_writer);
1187 set_current_state(state);
1188trylock_again:
1189 raw_spin_lock_irq(&sem->wait_lock);
1190 }
1191
1192 if (state == TASK_UNINTERRUPTIBLE)
1193 hung_task_clear_blocker();
1194
1195 __set_current_state(TASK_RUNNING);
1196 raw_spin_unlock_irq(&sem->wait_lock);
1197 lockevent_inc(rwsem_wlock);
1198 trace_contention_end(lock: sem, ret: 0);
1199 return sem;
1200
1201out_nolock:
1202 __set_current_state(TASK_RUNNING);
1203 raw_spin_lock_irq(&sem->wait_lock);
1204 rwsem_del_wake_waiter(sem, waiter: &waiter, wake_q: &wake_q);
1205 lockevent_inc(rwsem_wlock_fail);
1206 trace_contention_end(lock: sem, ret: -EINTR);
1207 return ERR_PTR(error: -EINTR);
1208}
1209
1210/*
1211 * handle waking up a waiter on the semaphore
1212 * - up_read/up_write has decremented the active part of count if we come here
1213 */
1214static struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
1215{
1216 unsigned long flags;
1217 DEFINE_WAKE_Q(wake_q);
1218
1219 raw_spin_lock_irqsave(&sem->wait_lock, flags);
1220
1221 if (!list_empty(head: &sem->wait_list))
1222 rwsem_mark_wake(sem, wake_type: RWSEM_WAKE_ANY, wake_q: &wake_q);
1223
1224 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1225 wake_up_q(head: &wake_q);
1226
1227 return sem;
1228}
1229
1230/*
1231 * downgrade a write lock into a read lock
1232 * - caller incremented waiting part of count and discovered it still negative
1233 * - just wake up any readers at the front of the queue
1234 */
1235static struct rw_semaphore *rwsem_downgrade_wake(struct rw_semaphore *sem)
1236{
1237 unsigned long flags;
1238 DEFINE_WAKE_Q(wake_q);
1239
1240 raw_spin_lock_irqsave(&sem->wait_lock, flags);
1241
1242 if (!list_empty(head: &sem->wait_list))
1243 rwsem_mark_wake(sem, wake_type: RWSEM_WAKE_READ_OWNED, wake_q: &wake_q);
1244
1245 raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
1246 wake_up_q(head: &wake_q);
1247
1248 return sem;
1249}
1250
1251/*
1252 * lock for reading
1253 */
1254static __always_inline int __down_read_common(struct rw_semaphore *sem, int state)
1255{
1256 int ret = 0;
1257 long count;
1258
1259 preempt_disable();
1260 if (!rwsem_read_trylock(sem, cntp: &count)) {
1261 if (IS_ERR(ptr: rwsem_down_read_slowpath(sem, count, state))) {
1262 ret = -EINTR;
1263 goto out;
1264 }
1265 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1266 }
1267out:
1268 preempt_enable();
1269 return ret;
1270}
1271
1272static __always_inline void __down_read(struct rw_semaphore *sem)
1273{
1274 __down_read_common(sem, TASK_UNINTERRUPTIBLE);
1275}
1276
1277static __always_inline int __down_read_interruptible(struct rw_semaphore *sem)
1278{
1279 return __down_read_common(sem, TASK_INTERRUPTIBLE);
1280}
1281
1282static __always_inline int __down_read_killable(struct rw_semaphore *sem)
1283{
1284 return __down_read_common(sem, TASK_KILLABLE);
1285}
1286
1287static inline int __down_read_trylock(struct rw_semaphore *sem)
1288{
1289 int ret = 0;
1290 long tmp;
1291
1292 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1293
1294 preempt_disable();
1295 tmp = atomic_long_read(v: &sem->count);
1296 while (!(tmp & RWSEM_READ_FAILED_MASK)) {
1297 if (atomic_long_try_cmpxchg_acquire(v: &sem->count, old: &tmp,
1298 new: tmp + RWSEM_READER_BIAS)) {
1299 rwsem_set_reader_owned(sem);
1300 ret = 1;
1301 break;
1302 }
1303 }
1304 preempt_enable();
1305 return ret;
1306}
1307
1308/*
1309 * lock for writing
1310 */
1311static __always_inline int __down_write_common(struct rw_semaphore *sem, int state)
1312{
1313 int ret = 0;
1314
1315 preempt_disable();
1316 if (unlikely(!rwsem_write_trylock(sem))) {
1317 if (IS_ERR(ptr: rwsem_down_write_slowpath(sem, state)))
1318 ret = -EINTR;
1319 }
1320 preempt_enable();
1321 return ret;
1322}
1323
1324static __always_inline void __down_write(struct rw_semaphore *sem)
1325{
1326 __down_write_common(sem, TASK_UNINTERRUPTIBLE);
1327}
1328
1329static __always_inline int __down_write_killable(struct rw_semaphore *sem)
1330{
1331 return __down_write_common(sem, TASK_KILLABLE);
1332}
1333
1334static inline int __down_write_trylock(struct rw_semaphore *sem)
1335{
1336 int ret;
1337
1338 preempt_disable();
1339 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1340 ret = rwsem_write_trylock(sem);
1341 preempt_enable();
1342
1343 return ret;
1344}
1345
1346/*
1347 * unlock after reading
1348 */
1349static inline void __up_read(struct rw_semaphore *sem)
1350{
1351 long tmp;
1352
1353 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1354 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1355
1356 preempt_disable();
1357 rwsem_clear_reader_owned(sem);
1358 tmp = atomic_long_add_return_release(i: -RWSEM_READER_BIAS, v: &sem->count);
1359 DEBUG_RWSEMS_WARN_ON(tmp < 0, sem);
1360 if (unlikely((tmp & (RWSEM_LOCK_MASK|RWSEM_FLAG_WAITERS)) ==
1361 RWSEM_FLAG_WAITERS)) {
1362 clear_nonspinnable(sem);
1363 rwsem_wake(sem);
1364 }
1365 preempt_enable();
1366}
1367
1368/*
1369 * unlock after writing
1370 */
1371static inline void __up_write(struct rw_semaphore *sem)
1372{
1373 long tmp;
1374
1375 DEBUG_RWSEMS_WARN_ON(sem->magic != sem, sem);
1376 /*
1377 * sem->owner may differ from current if the ownership is transferred
1378 * to an anonymous writer by setting the RWSEM_NONSPINNABLE bits.
1379 */
1380 DEBUG_RWSEMS_WARN_ON((rwsem_owner(sem) != current) &&
1381 !rwsem_test_oflags(sem, RWSEM_NONSPINNABLE), sem);
1382
1383 preempt_disable();
1384 rwsem_clear_owner(sem);
1385 tmp = atomic_long_fetch_add_release(i: -RWSEM_WRITER_LOCKED, v: &sem->count);
1386 if (unlikely(tmp & RWSEM_FLAG_WAITERS))
1387 rwsem_wake(sem);
1388 preempt_enable();
1389}
1390
1391/*
1392 * downgrade write lock to read lock
1393 */
1394static inline void __downgrade_write(struct rw_semaphore *sem)
1395{
1396 long tmp;
1397
1398 /*
1399 * When downgrading from exclusive to shared ownership,
1400 * anything inside the write-locked region cannot leak
1401 * into the read side. In contrast, anything in the
1402 * read-locked region is ok to be re-ordered into the
1403 * write side. As such, rely on RELEASE semantics.
1404 */
1405 DEBUG_RWSEMS_WARN_ON(rwsem_owner(sem) != current, sem);
1406 preempt_disable();
1407 tmp = atomic_long_fetch_add_release(
1408 i: -RWSEM_WRITER_LOCKED+RWSEM_READER_BIAS, v: &sem->count);
1409 rwsem_set_reader_owned(sem);
1410 if (tmp & RWSEM_FLAG_WAITERS)
1411 rwsem_downgrade_wake(sem);
1412 preempt_enable();
1413}
1414
1415#else /* !CONFIG_PREEMPT_RT */
1416
1417#define RT_MUTEX_BUILD_MUTEX
1418#include "rtmutex.c"
1419
1420#define rwbase_set_and_save_current_state(state) \
1421 set_current_state(state)
1422
1423#define rwbase_restore_current_state() \
1424 __set_current_state(TASK_RUNNING)
1425
1426#define rwbase_rtmutex_lock_state(rtm, state) \
1427 __rt_mutex_lock(rtm, state)
1428
1429#define rwbase_rtmutex_slowlock_locked(rtm, state, wq) \
1430 __rt_mutex_slowlock_locked(rtm, NULL, state, wq)
1431
1432#define rwbase_rtmutex_unlock(rtm) \
1433 __rt_mutex_unlock(rtm)
1434
1435#define rwbase_rtmutex_trylock(rtm) \
1436 __rt_mutex_trylock(rtm)
1437
1438#define rwbase_signal_pending_state(state, current) \
1439 signal_pending_state(state, current)
1440
1441#define rwbase_pre_schedule() \
1442 rt_mutex_pre_schedule()
1443
1444#define rwbase_schedule() \
1445 rt_mutex_schedule()
1446
1447#define rwbase_post_schedule() \
1448 rt_mutex_post_schedule()
1449
1450#include "rwbase_rt.c"
1451
1452void __init_rwsem(struct rw_semaphore *sem, const char *name,
1453 struct lock_class_key *key)
1454{
1455 init_rwbase_rt(&(sem)->rwbase);
1456
1457#ifdef CONFIG_DEBUG_LOCK_ALLOC
1458 debug_check_no_locks_freed((void *)sem, sizeof(*sem));
1459 lockdep_init_map_wait(&sem->dep_map, name, key, 0, LD_WAIT_SLEEP);
1460#endif
1461}
1462EXPORT_SYMBOL(__init_rwsem);
1463
1464static inline void __down_read(struct rw_semaphore *sem)
1465{
1466 rwbase_read_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1467}
1468
1469static inline int __down_read_interruptible(struct rw_semaphore *sem)
1470{
1471 return rwbase_read_lock(&sem->rwbase, TASK_INTERRUPTIBLE);
1472}
1473
1474static inline int __down_read_killable(struct rw_semaphore *sem)
1475{
1476 return rwbase_read_lock(&sem->rwbase, TASK_KILLABLE);
1477}
1478
1479static inline int __down_read_trylock(struct rw_semaphore *sem)
1480{
1481 return rwbase_read_trylock(&sem->rwbase);
1482}
1483
1484static inline void __up_read(struct rw_semaphore *sem)
1485{
1486 rwbase_read_unlock(&sem->rwbase, TASK_NORMAL);
1487}
1488
1489static inline void __sched __down_write(struct rw_semaphore *sem)
1490{
1491 rwbase_write_lock(&sem->rwbase, TASK_UNINTERRUPTIBLE);
1492}
1493
1494static inline int __sched __down_write_killable(struct rw_semaphore *sem)
1495{
1496 return rwbase_write_lock(&sem->rwbase, TASK_KILLABLE);
1497}
1498
1499static inline int __down_write_trylock(struct rw_semaphore *sem)
1500{
1501 return rwbase_write_trylock(&sem->rwbase);
1502}
1503
1504static inline void __up_write(struct rw_semaphore *sem)
1505{
1506 rwbase_write_unlock(&sem->rwbase);
1507}
1508
1509static inline void __downgrade_write(struct rw_semaphore *sem)
1510{
1511 rwbase_write_downgrade(&sem->rwbase);
1512}
1513
1514/* Debug stubs for the common API */
1515#define DEBUG_RWSEMS_WARN_ON(c, sem)
1516
1517static inline void __rwsem_set_reader_owned(struct rw_semaphore *sem,
1518 struct task_struct *owner)
1519{
1520}
1521
1522static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
1523{
1524 int count = atomic_read(&sem->rwbase.readers);
1525
1526 return count < 0 && count != READER_BIAS;
1527}
1528
1529#endif /* CONFIG_PREEMPT_RT */
1530
1531/*
1532 * lock for reading
1533 */
1534void __sched down_read(struct rw_semaphore *sem)
1535{
1536 might_sleep();
1537 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1538
1539 LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1540}
1541EXPORT_SYMBOL(down_read);
1542
1543int __sched down_read_interruptible(struct rw_semaphore *sem)
1544{
1545 might_sleep();
1546 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1547
1548 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_interruptible)) {
1549 rwsem_release(&sem->dep_map, _RET_IP_);
1550 return -EINTR;
1551 }
1552
1553 return 0;
1554}
1555EXPORT_SYMBOL(down_read_interruptible);
1556
1557int __sched down_read_killable(struct rw_semaphore *sem)
1558{
1559 might_sleep();
1560 rwsem_acquire_read(&sem->dep_map, 0, 0, _RET_IP_);
1561
1562 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1563 rwsem_release(&sem->dep_map, _RET_IP_);
1564 return -EINTR;
1565 }
1566
1567 return 0;
1568}
1569EXPORT_SYMBOL(down_read_killable);
1570
1571/*
1572 * trylock for reading -- returns 1 if successful, 0 if contention
1573 */
1574int down_read_trylock(struct rw_semaphore *sem)
1575{
1576 int ret = __down_read_trylock(sem);
1577
1578 if (ret == 1)
1579 rwsem_acquire_read(&sem->dep_map, 0, 1, _RET_IP_);
1580 return ret;
1581}
1582EXPORT_SYMBOL(down_read_trylock);
1583
1584/*
1585 * lock for writing
1586 */
1587void __sched down_write(struct rw_semaphore *sem)
1588{
1589 might_sleep();
1590 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1591 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1592}
1593EXPORT_SYMBOL(down_write);
1594
1595/*
1596 * lock for writing
1597 */
1598int __sched down_write_killable(struct rw_semaphore *sem)
1599{
1600 might_sleep();
1601 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
1602
1603 if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1604 __down_write_killable)) {
1605 rwsem_release(&sem->dep_map, _RET_IP_);
1606 return -EINTR;
1607 }
1608
1609 return 0;
1610}
1611EXPORT_SYMBOL(down_write_killable);
1612
1613/*
1614 * trylock for writing -- returns 1 if successful, 0 if contention
1615 */
1616int down_write_trylock(struct rw_semaphore *sem)
1617{
1618 int ret = __down_write_trylock(sem);
1619
1620 if (ret == 1)
1621 rwsem_acquire(&sem->dep_map, 0, 1, _RET_IP_);
1622
1623 return ret;
1624}
1625EXPORT_SYMBOL(down_write_trylock);
1626
1627/*
1628 * release a read lock
1629 */
1630void up_read(struct rw_semaphore *sem)
1631{
1632 rwsem_release(&sem->dep_map, _RET_IP_);
1633 __up_read(sem);
1634}
1635EXPORT_SYMBOL(up_read);
1636
1637/*
1638 * release a write lock
1639 */
1640void up_write(struct rw_semaphore *sem)
1641{
1642 rwsem_release(&sem->dep_map, _RET_IP_);
1643 __up_write(sem);
1644}
1645EXPORT_SYMBOL(up_write);
1646
1647/*
1648 * downgrade write lock to read lock
1649 */
1650void downgrade_write(struct rw_semaphore *sem)
1651{
1652 lock_downgrade(&sem->dep_map, _RET_IP_);
1653 __downgrade_write(sem);
1654}
1655EXPORT_SYMBOL(downgrade_write);
1656
1657#ifdef CONFIG_DEBUG_LOCK_ALLOC
1658
1659void down_read_nested(struct rw_semaphore *sem, int subclass)
1660{
1661 might_sleep();
1662 rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1663 LOCK_CONTENDED(sem, __down_read_trylock, __down_read);
1664}
1665EXPORT_SYMBOL(down_read_nested);
1666
1667int down_read_killable_nested(struct rw_semaphore *sem, int subclass)
1668{
1669 might_sleep();
1670 rwsem_acquire_read(&sem->dep_map, subclass, 0, _RET_IP_);
1671
1672 if (LOCK_CONTENDED_RETURN(sem, __down_read_trylock, __down_read_killable)) {
1673 rwsem_release(&sem->dep_map, _RET_IP_);
1674 return -EINTR;
1675 }
1676
1677 return 0;
1678}
1679EXPORT_SYMBOL(down_read_killable_nested);
1680
1681void _down_write_nest_lock(struct rw_semaphore *sem, struct lockdep_map *nest)
1682{
1683 might_sleep();
1684 rwsem_acquire_nest(&sem->dep_map, 0, 0, nest, _RET_IP_);
1685 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1686}
1687EXPORT_SYMBOL(_down_write_nest_lock);
1688
1689void down_read_non_owner(struct rw_semaphore *sem)
1690{
1691 might_sleep();
1692 __down_read(sem);
1693 /*
1694 * The owner value for a reader-owned lock is mostly for debugging
1695 * purpose only and is not critical to the correct functioning of
1696 * rwsem. So it is perfectly fine to set it in a preempt-enabled
1697 * context here.
1698 */
1699 __rwsem_set_reader_owned(sem, NULL);
1700}
1701EXPORT_SYMBOL(down_read_non_owner);
1702
1703void down_write_nested(struct rw_semaphore *sem, int subclass)
1704{
1705 might_sleep();
1706 rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1707 LOCK_CONTENDED(sem, __down_write_trylock, __down_write);
1708}
1709EXPORT_SYMBOL(down_write_nested);
1710
1711int __sched down_write_killable_nested(struct rw_semaphore *sem, int subclass)
1712{
1713 might_sleep();
1714 rwsem_acquire(&sem->dep_map, subclass, 0, _RET_IP_);
1715
1716 if (LOCK_CONTENDED_RETURN(sem, __down_write_trylock,
1717 __down_write_killable)) {
1718 rwsem_release(&sem->dep_map, _RET_IP_);
1719 return -EINTR;
1720 }
1721
1722 return 0;
1723}
1724EXPORT_SYMBOL(down_write_killable_nested);
1725
1726void up_read_non_owner(struct rw_semaphore *sem)
1727{
1728 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
1729 __up_read(sem);
1730}
1731EXPORT_SYMBOL(up_read_non_owner);
1732
1733#endif
1734