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	DEFINE_WAKE_Q(wake_q);
 73	int ret;
 74
 75	rwbase_pre_schedule();
 76	raw_spin_lock_irq(&rtm->wait_lock);
 77
 78	/*
 79	 * Call into the slow lock path with the rtmutex->wait_lock
 80	 * held, so this can't result in the following race:
 81	 *
 82	 * Reader1		Reader2		Writer
 83	 *			down_read()
 84	 *					down_write()
 85	 *					rtmutex_lock(m)
 86	 *					wait()
 87	 * down_read()
 88	 * unlock(m->wait_lock)
 89	 *			up_read()
 90	 *			wake(Writer)
 91	 *					lock(m->wait_lock)
 92	 *					sem->writelocked=true
 93	 *					unlock(m->wait_lock)
 94	 *
 95	 *					up_write()
 96	 *					sem->writelocked=false
 97	 *					rtmutex_unlock(m)
 98	 *			down_read()
 99	 *					down_write()
100	 *					rtmutex_lock(m)
101	 *					wait()
102	 * rtmutex_lock(m)
103	 *
104	 * That would put Reader1 behind the writer waiting on
105	 * Reader2 to call up_read(), which might be unbound.
106	 */
107
108	trace_contention_begin(rwb, LCB_F_RT | LCB_F_READ);
109
110	/*
111	 * For rwlocks this returns 0 unconditionally, so the below
112	 * !ret conditionals are optimized out.
113	 */
114	ret = rwbase_rtmutex_slowlock_locked(rtm, state, &wake_q);
115
116	/*
117	 * On success the rtmutex is held, so there can't be a writer
118	 * active. Increment the reader count and immediately drop the
119	 * rtmutex again.
120	 *
121	 * rtmutex->wait_lock has to be unlocked in any case of course.
122	 */
123	if (!ret)
124		atomic_inc(&rwb->readers);
125
126	preempt_disable();
127	raw_spin_unlock_irq(&rtm->wait_lock);
128	wake_up_q(&wake_q);
129	preempt_enable();
130
131	if (!ret)
132		rwbase_rtmutex_unlock(rtm);
133
134	trace_contention_end(rwb, ret);
135	rwbase_post_schedule();
136	return ret;
137}
138
139static __always_inline int rwbase_read_lock(struct rwbase_rt *rwb,
140					    unsigned int state)
141{
142	lockdep_assert(!current->pi_blocked_on);
143
144	if (rwbase_read_trylock(rwb))
145		return 0;
146
147	return __rwbase_read_lock(rwb, state);
148}
149
150static void __sched __rwbase_read_unlock(struct rwbase_rt *rwb,
151					 unsigned int state)
152{
153	struct rt_mutex_base *rtm = &rwb->rtmutex;
154	struct task_struct *owner;
155	DEFINE_RT_WAKE_Q(wqh);
156
157	raw_spin_lock_irq(&rtm->wait_lock);
158	/*
159	 * Wake the writer, i.e. the rtmutex owner. It might release the
160	 * rtmutex concurrently in the fast path (due to a signal), but to
161	 * clean up rwb->readers it needs to acquire rtm->wait_lock. The
162	 * worst case which can happen is a spurious wakeup.
163	 */
164	owner = rt_mutex_owner(rtm);
165	if (owner)
166		rt_mutex_wake_q_add_task(&wqh, owner, state);
167
168	/* Pairs with the preempt_enable in rt_mutex_wake_up_q() */
169	preempt_disable();
170	raw_spin_unlock_irq(&rtm->wait_lock);
171	rt_mutex_wake_up_q(&wqh);
172}
173
174static __always_inline void rwbase_read_unlock(struct rwbase_rt *rwb,
175					       unsigned int state)
176{
177	/*
178	 * rwb->readers can only hit 0 when a writer is waiting for the
179	 * active readers to leave the critical section.
180	 *
181	 * dec_and_test() is fully ordered, provides RELEASE.
182	 */
183	if (unlikely(atomic_dec_and_test(&rwb->readers)))
184		__rwbase_read_unlock(rwb, state);
185}
186
187static inline void __rwbase_write_unlock(struct rwbase_rt *rwb, int bias,
188					 unsigned long flags)
189{
190	struct rt_mutex_base *rtm = &rwb->rtmutex;
191
192	/*
193	 * _release() is needed in case that reader is in fast path, pairing
194	 * with atomic_try_cmpxchg_acquire() in rwbase_read_trylock().
195	 */
196	(void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);
197	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
198	rwbase_rtmutex_unlock(rtm);
199}
200
201static inline void rwbase_write_unlock(struct rwbase_rt *rwb)
202{
203	struct rt_mutex_base *rtm = &rwb->rtmutex;
204	unsigned long flags;
205
206	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
207	__rwbase_write_unlock(rwb, WRITER_BIAS, flags);
208}
209
210static inline void rwbase_write_downgrade(struct rwbase_rt *rwb)
211{
212	struct rt_mutex_base *rtm = &rwb->rtmutex;
213	unsigned long flags;
214
215	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
216	/* Release it and account current as reader */
217	__rwbase_write_unlock(rwb, WRITER_BIAS - 1, flags);
218}
219
220static inline bool __rwbase_write_trylock(struct rwbase_rt *rwb)
221{
222	/* Can do without CAS because we're serialized by wait_lock. */
223	lockdep_assert_held(&rwb->rtmutex.wait_lock);
224
225	/*
226	 * _acquire is needed in case the reader is in the fast path, pairing
227	 * with rwbase_read_unlock(), provides ACQUIRE.
228	 */
229	if (!atomic_read_acquire(&rwb->readers)) {
230		atomic_set(&rwb->readers, WRITER_BIAS);
231		return 1;
232	}
233
234	return 0;
235}
236
237static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
238				     unsigned int state)
239{
240	struct rt_mutex_base *rtm = &rwb->rtmutex;
241	unsigned long flags;
242
243	/* Take the rtmutex as a first step */
244	if (rwbase_rtmutex_lock_state(rtm, state))
245		return -EINTR;
246
247	/* Force readers into slow path */
248	atomic_sub(READER_BIAS, &rwb->readers);
249
250	rwbase_pre_schedule();
251
252	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
253	if (__rwbase_write_trylock(rwb))
254		goto out_unlock;
255
256	rwbase_set_and_save_current_state(state);
257	trace_contention_begin(rwb, LCB_F_RT | LCB_F_WRITE);
258	for (;;) {
259		/* Optimized out for rwlocks */
260		if (rwbase_signal_pending_state(state, current)) {
261			rwbase_restore_current_state();
262			__rwbase_write_unlock(rwb, 0, flags);
263			rwbase_post_schedule();
264			trace_contention_end(rwb, -EINTR);
265			return -EINTR;
266		}
267
268		if (__rwbase_write_trylock(rwb))
269			break;
270
271		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
272		rwbase_schedule();
273		raw_spin_lock_irqsave(&rtm->wait_lock, flags);
274
275		set_current_state(state);
276	}
277	rwbase_restore_current_state();
278	trace_contention_end(rwb, 0);
279
280out_unlock:
281	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
282	rwbase_post_schedule();
283	return 0;
284}
285
286static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
287{
288	struct rt_mutex_base *rtm = &rwb->rtmutex;
289	unsigned long flags;
290
291	if (!rwbase_rtmutex_trylock(rtm))
292		return 0;
293
294	atomic_sub(READER_BIAS, &rwb->readers);
295
296	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
297	if (__rwbase_write_trylock(rwb)) {
298		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
299		return 1;
300	}
301	__rwbase_write_unlock(rwb, 0, flags);
302	return 0;
303}