1// SPDX-License-Identifier: GPL-2.0-only
  2
  3/*
  4 * RT-specific reader/writer semaphores and reader/writer locks
  5 *
  6 * down_write/write_lock()
  7 *  1) Lock rtmutex
  8 *  2) Remove the reader BIAS to force readers into the slow path
  9 *  3) Wait until all readers have left the critical section
 10 *  4) Mark it write locked
 11 *
 12 * up_write/write_unlock()
 13 *  1) Remove the write locked marker
 14 *  2) Set the reader BIAS, so readers can use the fast path again
 15 *  3) Unlock rtmutex, to release blocked readers
 16 *
 17 * down_read/read_lock()
 18 *  1) Try fast path acquisition (reader BIAS is set)
 19 *  2) Take tmutex::wait_lock, which protects the writelocked flag
 20 *  3) If !writelocked, acquire it for read
 21 *  4) If writelocked, block on tmutex
 22 *  5) unlock rtmutex, goto 1)
 23 *
 24 * up_read/read_unlock()
 25 *  1) Try fast path release (reader count != 1)
 26 *  2) Wake the writer waiting in down_write()/write_lock() #3
 27 *
 28 * down_read/read_lock()#3 has the consequence, that rw semaphores and rw
 29 * locks on RT are not writer fair, but writers, which should be avoided in
 30 * RT tasks (think mmap_sem), are subject to the rtmutex priority/DL
 31 * inheritance mechanism.
 32 *
 33 * It's possible to make the rw primitives writer fair by keeping a list of
 34 * active readers. A blocked writer would force all newly incoming readers
 35 * to block on the rtmutex, but the rtmutex would have to be proxy locked
 36 * for one reader after the other. We can't use multi-reader inheritance
 37 * because there is no way to support that with SCHED_DEADLINE.
 38 * Implementing the one by one reader boosting/handover mechanism is a
 39 * major surgery for a very dubious value.
 40 *
 41 * The risk of writer starvation is there, but the pathological use cases
 42 * which trigger it are not necessarily the typical RT workloads.
 43 *
 44 * Fast-path orderings:
 45 * The lock/unlock of readers can run in fast paths: lock and unlock are only
 46 * atomic ops, and there is no inner lock to provide ACQUIRE and RELEASE
 47 * semantics of rwbase_rt. Atomic ops should thus provide _acquire()
 48 * and _release() (or stronger).
 49 *
 50 * Common code shared between RT rw_semaphore and rwlock
 51 */
 52
 53static __always_inline int rwbase_read_trylock(struct rwbase_rt *rwb)
 54{
 55	int r;
 56
 57	/*
 58	 * Increment reader count, if sem->readers < 0, i.e. READER_BIAS is
 59	 * set.
 60	 */
 61	for (r = atomic_read(&rwb->readers); r < 0;) {
 62		if (likely(atomic_try_cmpxchg_acquire(&rwb->readers, &r, r + 1)))
 63			return 1;
 64	}
 65	return 0;
 66}
 67
 68static int __sched __rwbase_read_lock(struct rwbase_rt *rwb,
 69				      unsigned int state)
 70{
 71	struct rt_mutex_base *rtm = &rwb->rtmutex;
 72	int ret;
 73
 74	raw_spin_lock_irq(&rtm->wait_lock);
 75	/*
 76	 * Allow readers, as long as the writer has not completely
 77	 * acquired the semaphore for write.
 78	 */
 79	if (atomic_read(&rwb->readers) != WRITER_BIAS) {
 80		atomic_inc(&rwb->readers);
 81		raw_spin_unlock_irq(&rtm->wait_lock);
 82		return 0;
 83	}
 84
 85	/*
 86	 * Call into the slow lock path with the rtmutex->wait_lock
 87	 * held, so this can't result in the following race:
 88	 *
 89	 * Reader1		Reader2		Writer
 90	 *			down_read()
 91	 *					down_write()
 92	 *					rtmutex_lock(m)
 93	 *					wait()
 94	 * down_read()
 95	 * unlock(m->wait_lock)
 96	 *			up_read()
 97	 *			wake(Writer)
 98	 *					lock(m->wait_lock)
 99	 *					sem->writelocked=true
100	 *					unlock(m->wait_lock)
101	 *
102	 *					up_write()
103	 *					sem->writelocked=false
104	 *					rtmutex_unlock(m)
105	 *			down_read()
106	 *					down_write()
107	 *					rtmutex_lock(m)
108	 *					wait()
109	 * rtmutex_lock(m)
110	 *
111	 * That would put Reader1 behind the writer waiting on
112	 * Reader2 to call up_read(), which might be unbound.
113	 */
114
115	trace_contention_begin(rwb, LCB_F_RT | LCB_F_READ);
116
117	/*
118	 * For rwlocks this returns 0 unconditionally, so the below
119	 * !ret conditionals are optimized out.
120	 */
121	ret = rwbase_rtmutex_slowlock_locked(rtm, state);
122
123	/*
124	 * On success the rtmutex is held, so there can't be a writer
125	 * active. Increment the reader count and immediately drop the
126	 * rtmutex again.
127	 *
128	 * rtmutex->wait_lock has to be unlocked in any case of course.
129	 */
130	if (!ret)
131		atomic_inc(&rwb->readers);
132	raw_spin_unlock_irq(&rtm->wait_lock);
133	if (!ret)
134		rwbase_rtmutex_unlock(rtm);
135
136	trace_contention_end(rwb, ret);
137	return ret;
138}
139
140static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb,
141					    unsigned int state)
142{
143	if (rwbase_read_trylock(rwb))
144		return 0;
145
146	return __rwbase_read_lock(rwb, state);
147}
148
149static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
150					 unsigned int state)
151{
152	struct rt_mutex_base *rtm = &rwb->rtmutex;
153	struct task_struct *owner;
154	DEFINE_RT_WAKE_Q(wqh);
155
156	raw_spin_lock_irq(&rtm->wait_lock);
157	/*
158	 * Wake the writer, i.e. the rtmutex owner. It might release the
159	 * rtmutex concurrently in the fast path (due to a signal), but to
160	 * clean up rwb->readers it needs to acquire rtm->wait_lock. The
161	 * worst case which can happen is a spurious wakeup.
162	 */
163	owner = rt_mutex_owner(rtm);
164	if (owner)
165		rt_mutex_wake_q_add_task(&wqh, owner, state);
166
167	/* Pairs with the preempt_enable in rt_mutex_wake_up_q() */
168	preempt_disable();
169	raw_spin_unlock_irq(&rtm->wait_lock);
170	rt_mutex_wake_up_q(&wqh);
171}
172
173static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
174					       unsigned int state)
175{
176	/*
177	 * rwb->readers can only hit 0 when a writer is waiting for the
178	 * active readers to leave the critical section.
179	 *
180	 * dec_and_test() is fully ordered, provides RELEASE.
181	 */
182	if (unlikely(atomic_dec_and_test(&rwb->readers)))
183		__rwbase_read_unlock(rwb, state);
184}
185
186static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
187					 unsigned long flags)
188{
189	struct rt_mutex_base *rtm = &rwb->rtmutex;
190
191	/*
192	 * _release() is needed in case that reader is in fast path, pairing
193	 * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
194	 */
195	(void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
196	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
197	rwbase_rtmutex_unlock(rtm);
198}
199
200static inline void rwbase_write_unlock(struct rwbase_rt *rwb)
201{
202	struct rt_mutex_base *rtm = &rwb->rtmutex;
203	unsigned long flags;
204
205	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
206	__rwbase_write_unlock(rwb, WRITER_BIAS, flags);
207}
208
209static inline void rwbase_write_downgrade(struct rwbase_rt *rwb)
210{
211	struct rt_mutex_base *rtm = &rwb->rtmutex;
212	unsigned long flags;
213
214	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
215	/* Release it and account current as reader */
216	__rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags);
217}
218
219static inline bool __rwbase_write_trylock(struct rwbase_rt *rwb)
220{
221	/* Can do without CAS because we're serialized by wait_lock. */
222	lockdep_assert_held(&rwb->rtmutex.wait_lock);
223
224	/*
225	 * _acquire is needed in case the reader is in the fast path, pairing
226	 * with rwbase_read_unlock(), provides ACQUIRE.
227	 */
228	if (!atomic_read_acquire(&rwb->readers)) {
229		atomic_set(&rwb->readers, WRITER_BIAS);
230		return 1;
231	}
232
233	return 0;
234}
235
236static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
237				     unsigned int state)
238{
239	struct rt_mutex_base *rtm = &rwb->rtmutex;
240	unsigned long flags;
241
242	/* Take the rtmutex as a first step */
243	if (rwbase_rtmutex_lock_state(rtm, state))
244		return -EINTR;
245
246	/* Force readers into slow path */
247	atomic_sub(READER_BIAS, &rwb->readers);
248
249	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
250	if (__rwbase_write_trylock(rwb))
251		goto out_unlock;
252
253	rwbase_set_and_save_current_state(state);
254	trace_contention_begin(rwb, LCB_F_RT | LCB_F_WRITE);
255	for (;;) {
256		/* Optimized out for rwlocks */
257		if (rwbase_signal_pending_state(state, current)) {
258			rwbase_restore_current_state();
259			__rwbase_write_unlock(rwb, 0, flags);
260			trace_contention_end(rwb, -EINTR);
261			return -EINTR;
262		}
263
264		if (__rwbase_write_trylock(rwb))
265			break;
266
267		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
268		rwbase_schedule();
269		raw_spin_lock_irqsave(&rtm->wait_lock, flags);
270
271		set_current_state(state);
272	}
273	rwbase_restore_current_state();
274	trace_contention_end(rwb, 0);
275
276out_unlock:
277	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
278	return 0;
279}
280
281static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
282{
283	struct rt_mutex_base *rtm = &rwb->rtmutex;
284	unsigned long flags;
285
286	if (!rwbase_rtmutex_trylock(rtm))
287		return 0;
288
289	atomic_sub(READER_BIAS, &rwb->readers);
290
291	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
292	if (__rwbase_write_trylock(rwb)) {
293		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
294		return 1;
295	}
296	__rwbase_write_unlock(rwb, 0, flags);
297	return 0;
298}