xref: /openbmc/linux/kernel/locking/percpu-rwsem.c (revision 8dd3cdea)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/atomic.h>
3 #include <linux/percpu.h>
4 #include <linux/wait.h>
5 #include <linux/lockdep.h>
6 #include <linux/percpu-rwsem.h>
7 #include <linux/rcupdate.h>
8 #include <linux/sched.h>
9 #include <linux/sched/task.h>
10 #include <linux/sched/debug.h>
11 #include <linux/errno.h>
12 
13 int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
14 			const char *name, struct lock_class_key *key)
15 {
16 	sem->read_count = alloc_percpu(int);
17 	if (unlikely(!sem->read_count))
18 		return -ENOMEM;
19 
20 	rcu_sync_init(&sem->rss);
21 	rcuwait_init(&sem->writer);
22 	init_waitqueue_head(&sem->waiters);
23 	atomic_set(&sem->block, 0);
24 #ifdef CONFIG_DEBUG_LOCK_ALLOC
25 	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
26 	lockdep_init_map(&sem->dep_map, name, key, 0);
27 #endif
28 	return 0;
29 }
30 EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
31 
32 void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
33 {
34 	/*
35 	 * XXX: temporary kludge. The error path in alloc_super()
36 	 * assumes that percpu_free_rwsem() is safe after kzalloc().
37 	 */
38 	if (!sem->read_count)
39 		return;
40 
41 	rcu_sync_dtor(&sem->rss);
42 	free_percpu(sem->read_count);
43 	sem->read_count = NULL; /* catch use after free bugs */
44 }
45 EXPORT_SYMBOL_GPL(percpu_free_rwsem);
46 
47 static bool __percpu_down_read_trylock(struct percpu_rw_semaphore *sem)
48 {
49 	this_cpu_inc(*sem->read_count);
50 
51 	/*
52 	 * Due to having preemption disabled the decrement happens on
53 	 * the same CPU as the increment, avoiding the
54 	 * increment-on-one-CPU-and-decrement-on-another problem.
55 	 *
56 	 * If the reader misses the writer's assignment of sem->block, then the
57 	 * writer is guaranteed to see the reader's increment.
58 	 *
59 	 * Conversely, any readers that increment their sem->read_count after
60 	 * the writer looks are guaranteed to see the sem->block value, which
61 	 * in turn means that they are guaranteed to immediately decrement
62 	 * their sem->read_count, so that it doesn't matter that the writer
63 	 * missed them.
64 	 */
65 
66 	smp_mb(); /* A matches D */
67 
68 	/*
69 	 * If !sem->block the critical section starts here, matched by the
70 	 * release in percpu_up_write().
71 	 */
72 	if (likely(!atomic_read_acquire(&sem->block)))
73 		return true;
74 
75 	this_cpu_dec(*sem->read_count);
76 
77 	/* Prod writer to re-evaluate readers_active_check() */
78 	rcuwait_wake_up(&sem->writer);
79 
80 	return false;
81 }
82 
83 static inline bool __percpu_down_write_trylock(struct percpu_rw_semaphore *sem)
84 {
85 	if (atomic_read(&sem->block))
86 		return false;
87 
88 	return atomic_xchg(&sem->block, 1) == 0;
89 }
90 
91 static bool __percpu_rwsem_trylock(struct percpu_rw_semaphore *sem, bool reader)
92 {
93 	if (reader) {
94 		bool ret;
95 
96 		preempt_disable();
97 		ret = __percpu_down_read_trylock(sem);
98 		preempt_enable();
99 
100 		return ret;
101 	}
102 	return __percpu_down_write_trylock(sem);
103 }
104 
105 /*
106  * The return value of wait_queue_entry::func means:
107  *
108  *  <0 - error, wakeup is terminated and the error is returned
109  *   0 - no wakeup, a next waiter is tried
110  *  >0 - woken, if EXCLUSIVE, counted towards @nr_exclusive.
111  *
112  * We use EXCLUSIVE for both readers and writers to preserve FIFO order,
113  * and play games with the return value to allow waking multiple readers.
114  *
115  * Specifically, we wake readers until we've woken a single writer, or until a
116  * trylock fails.
117  */
118 static int percpu_rwsem_wake_function(struct wait_queue_entry *wq_entry,
119 				      unsigned int mode, int wake_flags,
120 				      void *key)
121 {
122 	bool reader = wq_entry->flags & WQ_FLAG_CUSTOM;
123 	struct percpu_rw_semaphore *sem = key;
124 	struct task_struct *p;
125 
126 	/* concurrent against percpu_down_write(), can get stolen */
127 	if (!__percpu_rwsem_trylock(sem, reader))
128 		return 1;
129 
130 	p = get_task_struct(wq_entry->private);
131 	list_del_init(&wq_entry->entry);
132 	smp_store_release(&wq_entry->private, NULL);
133 
134 	wake_up_process(p);
135 	put_task_struct(p);
136 
137 	return !reader; /* wake (readers until) 1 writer */
138 }
139 
140 static void percpu_rwsem_wait(struct percpu_rw_semaphore *sem, bool reader)
141 {
142 	DEFINE_WAIT_FUNC(wq_entry, percpu_rwsem_wake_function);
143 	bool wait;
144 
145 	spin_lock_irq(&sem->waiters.lock);
146 	/*
147 	 * Serialize against the wakeup in percpu_up_write(), if we fail
148 	 * the trylock, the wakeup must see us on the list.
149 	 */
150 	wait = !__percpu_rwsem_trylock(sem, reader);
151 	if (wait) {
152 		wq_entry.flags |= WQ_FLAG_EXCLUSIVE | reader * WQ_FLAG_CUSTOM;
153 		__add_wait_queue_entry_tail(&sem->waiters, &wq_entry);
154 	}
155 	spin_unlock_irq(&sem->waiters.lock);
156 
157 	while (wait) {
158 		set_current_state(TASK_UNINTERRUPTIBLE);
159 		if (!smp_load_acquire(&wq_entry.private))
160 			break;
161 		schedule();
162 	}
163 	__set_current_state(TASK_RUNNING);
164 }
165 
166 bool __sched __percpu_down_read(struct percpu_rw_semaphore *sem, bool try)
167 {
168 	if (__percpu_down_read_trylock(sem))
169 		return true;
170 
171 	if (try)
172 		return false;
173 
174 	preempt_enable();
175 	percpu_rwsem_wait(sem, /* .reader = */ true);
176 	preempt_disable();
177 
178 	return true;
179 }
180 EXPORT_SYMBOL_GPL(__percpu_down_read);
181 
182 #define per_cpu_sum(var)						\
183 ({									\
184 	typeof(var) __sum = 0;						\
185 	int cpu;							\
186 	compiletime_assert_atomic_type(__sum);				\
187 	for_each_possible_cpu(cpu)					\
188 		__sum += per_cpu(var, cpu);				\
189 	__sum;								\
190 })
191 
192 /*
193  * Return true if the modular sum of the sem->read_count per-CPU variable is
194  * zero.  If this sum is zero, then it is stable due to the fact that if any
195  * newly arriving readers increment a given counter, they will immediately
196  * decrement that same counter.
197  *
198  * Assumes sem->block is set.
199  */
200 static bool readers_active_check(struct percpu_rw_semaphore *sem)
201 {
202 	if (per_cpu_sum(*sem->read_count) != 0)
203 		return false;
204 
205 	/*
206 	 * If we observed the decrement; ensure we see the entire critical
207 	 * section.
208 	 */
209 
210 	smp_mb(); /* C matches B */
211 
212 	return true;
213 }
214 
215 void __sched percpu_down_write(struct percpu_rw_semaphore *sem)
216 {
217 	might_sleep();
218 	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
219 
220 	/* Notify readers to take the slow path. */
221 	rcu_sync_enter(&sem->rss);
222 
223 	/*
224 	 * Try set sem->block; this provides writer-writer exclusion.
225 	 * Having sem->block set makes new readers block.
226 	 */
227 	if (!__percpu_down_write_trylock(sem))
228 		percpu_rwsem_wait(sem, /* .reader = */ false);
229 
230 	/* smp_mb() implied by __percpu_down_write_trylock() on success -- D matches A */
231 
232 	/*
233 	 * If they don't see our store of sem->block, then we are guaranteed to
234 	 * see their sem->read_count increment, and therefore will wait for
235 	 * them.
236 	 */
237 
238 	/* Wait for all active readers to complete. */
239 	rcuwait_wait_event(&sem->writer, readers_active_check(sem), TASK_UNINTERRUPTIBLE);
240 }
241 EXPORT_SYMBOL_GPL(percpu_down_write);
242 
243 void percpu_up_write(struct percpu_rw_semaphore *sem)
244 {
245 	rwsem_release(&sem->dep_map, _RET_IP_);
246 
247 	/*
248 	 * Signal the writer is done, no fast path yet.
249 	 *
250 	 * One reason that we cannot just immediately flip to readers_fast is
251 	 * that new readers might fail to see the results of this writer's
252 	 * critical section.
253 	 *
254 	 * Therefore we force it through the slow path which guarantees an
255 	 * acquire and thereby guarantees the critical section's consistency.
256 	 */
257 	atomic_set_release(&sem->block, 0);
258 
259 	/*
260 	 * Prod any pending reader/writer to make progress.
261 	 */
262 	__wake_up(&sem->waiters, TASK_NORMAL, 1, sem);
263 
264 	/*
265 	 * Once this completes (at least one RCU-sched grace period hence) the
266 	 * reader fast path will be available again. Safe to use outside the
267 	 * exclusive write lock because its counting.
268 	 */
269 	rcu_sync_exit(&sem->rss);
270 }
271 EXPORT_SYMBOL_GPL(percpu_up_write);
272