1#include <linux/atomic.h>
  2#include <linux/rwsem.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/errno.h>
 10
 11int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
 12			const char *name, struct lock_class_key *rwsem_key)
 13{
 14	sem->read_count = alloc_percpu(int);
 15	if (unlikely(!sem->read_count))
 16		return -ENOMEM;
 17
 18	/* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
 19	rcu_sync_init(&sem->rss, RCU_SCHED_SYNC);
 20	__init_rwsem(&sem->rw_sem, name, rwsem_key);
 21	init_waitqueue_head(&sem->writer);
 22	sem->readers_block = 0;
 23	return 0;
 24}
 25EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
 26
 27void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
 28{
 29	/*
 30	 * XXX: temporary kludge. The error path in alloc_super()
 31	 * assumes that percpu_free_rwsem() is safe after kzalloc().
 32	 */
 33	if (!sem->read_count)
 34		return;
 35
 36	rcu_sync_dtor(&sem->rss);
 37	free_percpu(sem->read_count);
 38	sem->read_count = NULL; /* catch use after free bugs */
 39}
 40EXPORT_SYMBOL_GPL(percpu_free_rwsem);
 41
 42int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
 43{
 44	/*
 45	 * Due to having preemption disabled the decrement happens on
 46	 * the same CPU as the increment, avoiding the
 47	 * increment-on-one-CPU-and-decrement-on-another problem.
 48	 *
 49	 * If the reader misses the writer's assignment of readers_block, then
 50	 * the writer is guaranteed to see the reader's increment.
 51	 *
 52	 * Conversely, any readers that increment their sem->read_count after
 53	 * the writer looks are guaranteed to see the readers_block value,
 54	 * which in turn means that they are guaranteed to immediately
 55	 * decrement their sem->read_count, so that it doesn't matter that the
 56	 * writer missed them.
 57	 */
 58
 59	smp_mb(); /* A matches D */
 60
 61	/*
 62	 * If !readers_block the critical section starts here, matched by the
 63	 * release in percpu_up_write().
 64	 */
 65	if (likely(!smp_load_acquire(&sem->readers_block)))
 66		return 1;
 67
 68	/*
 69	 * Per the above comment; we still have preemption disabled and
 70	 * will thus decrement on the same CPU as we incremented.
 71	 */
 72	__percpu_up_read(sem);
 73
 74	if (try)
 75		return 0;
 76
 77	/*
 78	 * We either call schedule() in the wait, or we'll fall through
 79	 * and reschedule on the preempt_enable() in percpu_down_read().
 80	 */
 81	preempt_enable_no_resched();
 82
 83	/*
 84	 * Avoid lockdep for the down/up_read() we already have them.
 85	 */
 86	__down_read(&sem->rw_sem);
 87	this_cpu_inc(*sem->read_count);
 88	__up_read(&sem->rw_sem);
 89
 90	preempt_disable();
 91	return 1;
 92}
 93EXPORT_SYMBOL_GPL(__percpu_down_read);
 94
 95void __percpu_up_read(struct percpu_rw_semaphore *sem)
 96{
 97	smp_mb(); /* B matches C */
 98	/*
 99	 * In other words, if they see our decrement (presumably to aggregate
100	 * zero, as that is the only time it matters) they will also see our
101	 * critical section.
102	 */
103	__this_cpu_dec(*sem->read_count);
104
105	/* Prod writer to recheck readers_active */
106	wake_up(&sem->writer);
107}
108EXPORT_SYMBOL_GPL(__percpu_up_read);
109
110#define per_cpu_sum(var)						\
111({									\
112	typeof(var) __sum = 0;						\
113	int cpu;							\
114	compiletime_assert_atomic_type(__sum);				\
115	for_each_possible_cpu(cpu)					\
116		__sum += per_cpu(var, cpu);				\
117	__sum;								\
118})
119
120/*
121 * Return true if the modular sum of the sem->read_count per-CPU variable is
122 * zero.  If this sum is zero, then it is stable due to the fact that if any
123 * newly arriving readers increment a given counter, they will immediately
124 * decrement that same counter.
125 */
126static bool readers_active_check(struct percpu_rw_semaphore *sem)
127{
128	if (per_cpu_sum(*sem->read_count) != 0)
129		return false;
130
131	/*
132	 * If we observed the decrement; ensure we see the entire critical
133	 * section.
134	 */
135
136	smp_mb(); /* C matches B */
137
138	return true;
139}
140
141void percpu_down_write(struct percpu_rw_semaphore *sem)
142{
143	/* Notify readers to take the slow path. */
144	rcu_sync_enter(&sem->rss);
145
146	down_write(&sem->rw_sem);
147
148	/*
149	 * Notify new readers to block; up until now, and thus throughout the
150	 * longish rcu_sync_enter() above, new readers could still come in.
151	 */
152	WRITE_ONCE(sem->readers_block, 1);
153
154	smp_mb(); /* D matches A */
155
156	/*
157	 * If they don't see our writer of readers_block, then we are
158	 * guaranteed to see their sem->read_count increment, and therefore
159	 * will wait for them.
160	 */
161
162	/* Wait for all now active readers to complete. */
163	wait_event(sem->writer, readers_active_check(sem));
164}
165EXPORT_SYMBOL_GPL(percpu_down_write);
166
167void percpu_up_write(struct percpu_rw_semaphore *sem)
168{
169	/*
170	 * Signal the writer is done, no fast path yet.
171	 *
172	 * One reason that we cannot just immediately flip to readers_fast is
173	 * that new readers might fail to see the results of this writer's
174	 * critical section.
175	 *
176	 * Therefore we force it through the slow path which guarantees an
177	 * acquire and thereby guarantees the critical section's consistency.
178	 */
179	smp_store_release(&sem->readers_block, 0);
180
181	/*
182	 * Release the write lock, this will allow readers back in the game.
183	 */
184	up_write(&sem->rw_sem);
185
186	/*
187	 * Once this completes (at least one RCU-sched grace period hence) the
188	 * reader fast path will be available again. Safe to use outside the
189	 * exclusive write lock because its counting.
190	 */
191	rcu_sync_exit(&sem->rss);
192}
193EXPORT_SYMBOL_GPL(percpu_up_write);