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	rwbase_pre_schedule();
 75	raw_spin_lock_irq(&rtm->wait_lock);
 76
 77	/*
 78	 * Call into the slow lock path with the rtmutex->wait_lock
 79	 * held, so this can't result in the following race:
 80	 *
 81	 * Reader1		Reader2		Writer
 82	 *			down_read()
 83	 *					down_write()
 84	 *					rtmutex_lock(m)
 85	 *					wait()
 86	 * down_read()
 87	 * unlock(m->wait_lock)
 88	 *			up_read()
 89	 *			wake(Writer)
 90	 *					lock(m->wait_lock)
 91	 *					sem->writelocked=true
 92	 *					unlock(m->wait_lock)
 93	 *
 94	 *					up_write()
 95	 *					sem->writelocked=false
 96	 *					rtmutex_unlock(m)
 97	 *			down_read()
 98	 *					down_write()
 99	 *					rtmutex_lock(m)
100	 *					wait()
101	 * rtmutex_lock(m)
102	 *
103	 * That would put Reader1 behind the writer waiting on
104	 * Reader2 to call up_read(), which might be unbound.
105	 */
106
107	trace_contention_begin(rwb, LCB_F_RT | LCB_F_READ);
108
109	/*
110	 * For rwlocks this returns 0 unconditionally, so the below
111	 * !ret conditionals are optimized out.
112	 */
113	ret = rwbase_rtmutex_slowlock_locked(rtm, state);
114
115	/*
116	 * On success the rtmutex is held, so there can't be a writer
117	 * active. Increment the reader count and immediately drop the
118	 * rtmutex again.
119	 *
120	 * rtmutex->wait_lock has to be unlocked in any case of course.
121	 */
122	if (!ret)
123		atomic_inc(&rwb->readers);
124	raw_spin_unlock_irq(&rtm->wait_lock);
125	if (!ret)
126		rwbase_rtmutex_unlock(rtm);
127
128	trace_contention_end(rwb, ret);
129	rwbase_post_schedule();
130	return ret;
131}
132
133static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb,
134					    unsigned int state)
135{
136	lockdep_assert(!current->pi_blocked_on);
137
138	if (rwbase_read_trylock(rwb))
139		return 0;
140
141	return __rwbase_read_lock(rwb, state);
142}
143
144static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
145					 unsigned int state)
146{
147	struct rt_mutex_base *rtm = &rwb->rtmutex;
148	struct task_struct *owner;
149	DEFINE_RT_WAKE_Q(wqh);
150
151	raw_spin_lock_irq(&rtm->wait_lock);
152	/*
153	 * Wake the writer, i.e. the rtmutex owner. It might release the
154	 * rtmutex concurrently in the fast path (due to a signal), but to
155	 * clean up rwb->readers it needs to acquire rtm->wait_lock. The
156	 * worst case which can happen is a spurious wakeup.
157	 */
158	owner = rt_mutex_owner(rtm);
159	if (owner)
160		rt_mutex_wake_q_add_task(&wqh, owner, state);
161
162	/* Pairs with the preempt_enable in rt_mutex_wake_up_q() */
163	preempt_disable();
164	raw_spin_unlock_irq(&rtm->wait_lock);
165	rt_mutex_wake_up_q(&wqh);
166}
167
168static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
169					       unsigned int state)
170{
171	/*
172	 * rwb->readers can only hit 0 when a writer is waiting for the
173	 * active readers to leave the critical section.
174	 *
175	 * dec_and_test() is fully ordered, provides RELEASE.
176	 */
177	if (unlikely(atomic_dec_and_test(&rwb->readers)))
178		__rwbase_read_unlock(rwb, state);
179}
180
181static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
182					 unsigned long flags)
183{
184	struct rt_mutex_base *rtm = &rwb->rtmutex;
185
186	/*
187	 * _release() is needed in case that reader is in fast path, pairing
188	 * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
189	 */
190	(void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
191	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
192	rwbase_rtmutex_unlock(rtm);
193}
194
195static inline void rwbase_write_unlock(struct rwbase_rt *rwb)
196{
197	struct rt_mutex_base *rtm = &rwb->rtmutex;
198	unsigned long flags;
199
200	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
201	__rwbase_write_unlock(rwb, WRITER_BIAS, flags);
202}
203
204static inline void rwbase_write_downgrade(struct rwbase_rt *rwb)
205{
206	struct rt_mutex_base *rtm = &rwb->rtmutex;
207	unsigned long flags;
208
209	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
210	/* Release it and account current as reader */
211	__rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags);
212}
213
214static inline bool __rwbase_write_trylock(struct rwbase_rt *rwb)
215{
216	/* Can do without CAS because we're serialized by wait_lock. */
217	lockdep_assert_held(&rwb->rtmutex.wait_lock);
218
219	/*
220	 * _acquire is needed in case the reader is in the fast path, pairing
221	 * with rwbase_read_unlock(), provides ACQUIRE.
222	 */
223	if (!atomic_read_acquire(&rwb->readers)) {
224		atomic_set(&rwb->readers, WRITER_BIAS);
225		return 1;
226	}
227
228	return 0;
229}
230
231static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
232				     unsigned int state)
233{
234	struct rt_mutex_base *rtm = &rwb->rtmutex;
235	unsigned long flags;
236
237	/* Take the rtmutex as a first step */
238	if (rwbase_rtmutex_lock_state(rtm, state))
239		return -EINTR;
240
241	/* Force readers into slow path */
242	atomic_sub(READER_BIAS, &rwb->readers);
243
244	rwbase_pre_schedule();
245
246	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
247	if (__rwbase_write_trylock(rwb))
248		goto out_unlock;
249
250	rwbase_set_and_save_current_state(state);
251	trace_contention_begin(rwb, LCB_F_RT | LCB_F_WRITE);
252	for (;;) {
253		/* Optimized out for rwlocks */
254		if (rwbase_signal_pending_state(state, current)) {
255			rwbase_restore_current_state();
256			__rwbase_write_unlock(rwb, 0, flags);
257			rwbase_post_schedule();
258			trace_contention_end(rwb, -EINTR);
259			return -EINTR;
260		}
261
262		if (__rwbase_write_trylock(rwb))
263			break;
264
265		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
266		rwbase_schedule();
267		raw_spin_lock_irqsave(&rtm->wait_lock, flags);
268
269		set_current_state(state);
270	}
271	rwbase_restore_current_state();
272	trace_contention_end(rwb, 0);
273
274out_unlock:
275	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
276	rwbase_post_schedule();
277	return 0;
278}
279
280static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
281{
282	struct rt_mutex_base *rtm = &rwb->rtmutex;
283	unsigned long flags;
284
285	if (!rwbase_rtmutex_trylock(rtm))
286		return 0;
287
288	atomic_sub(READER_BIAS, &rwb->readers);
289
290	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
291	if (__rwbase_write_trylock(rwb)) {
292		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
293		return 1;
294	}
295	__rwbase_write_unlock(rwb, 0, flags);
296	return 0;
297}