1#include <linux/atomic.h>
  2#include <linux/rwsem.h>
  3#include <linux/percpu.h>
  4#include <linux/lockdep.h>
  5#include <linux/percpu-rwsem.h>
  6#include <linux/rcupdate.h>
  7#include <linux/sched.h>
  8#include <linux/errno.h>
  9
 10int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
 11			const char *name, struct lock_class_key *rwsem_key)
 12{
 13	sem->read_count = alloc_percpu(int);
 14	if (unlikely(!sem->read_count))
 15		return -ENOMEM;
 16
 17	/* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
 18	rcu_sync_init(&sem->rss, RCU_SCHED_SYNC);
 19	__init_rwsem(&sem->rw_sem, name, rwsem_key);
 20	rcuwait_init(&sem->writer);
 21	sem->readers_block = 0;
 22	return 0;
 23}
 24EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
 25
 26void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
 27{
 28	/*
 29	 * XXX: temporary kludge. The error path in alloc_super()
 30	 * assumes that percpu_free_rwsem() is safe after kzalloc().
 31	 */
 32	if (!sem->read_count)
 33		return;
 34
 35	rcu_sync_dtor(&sem->rss);
 36	free_percpu(sem->read_count);
 37	sem->read_count = NULL; /* catch use after free bugs */
 38}
 39EXPORT_SYMBOL_GPL(percpu_free_rwsem);
 40
 41int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
 42{
 43	/*
 44	 * Due to having preemption disabled the decrement happens on
 45	 * the same CPU as the increment, avoiding the
 46	 * increment-on-one-CPU-and-decrement-on-another problem.
 47	 *
 48	 * If the reader misses the writer's assignment of readers_block, then
 49	 * the writer is guaranteed to see the reader's increment.
 50	 *
 51	 * Conversely, any readers that increment their sem->read_count after
 52	 * the writer looks are guaranteed to see the readers_block value,
 53	 * which in turn means that they are guaranteed to immediately
 54	 * decrement their sem->read_count, so that it doesn't matter that the
 55	 * writer missed them.
 56	 */
 57
 58	smp_mb(); /* A matches D */
 59
 60	/*
 61	 * If !readers_block the critical section starts here, matched by the
 62	 * release in percpu_up_write().
 63	 */
 64	if (likely(!smp_load_acquire(&sem->readers_block)))
 65		return 1;
 66
 67	/*
 68	 * Per the above comment; we still have preemption disabled and
 69	 * will thus decrement on the same CPU as we incremented.
 70	 */
 71	__percpu_up_read(sem);
 72
 73	if (try)
 74		return 0;
 75
 76	/*
 77	 * We either call schedule() in the wait, or we'll fall through
 78	 * and reschedule on the preempt_enable() in percpu_down_read().
 79	 */
 80	preempt_enable_no_resched();
 81
 82	/*
 83	 * Avoid lockdep for the down/up_read() we already have them.
 84	 */
 85	__down_read(&sem->rw_sem);
 86	this_cpu_inc(*sem->read_count);
 87	__up_read(&sem->rw_sem);
 88
 89	preempt_disable();
 90	return 1;
 91}
 92EXPORT_SYMBOL_GPL(__percpu_down_read);
 93
 94void __percpu_up_read(struct percpu_rw_semaphore *sem)
 95{
 96	smp_mb(); /* B matches C */
 97	/*
 98	 * In other words, if they see our decrement (presumably to aggregate
 99	 * zero, as that is the only time it matters) they will also see our
100	 * critical section.
101	 */
102	__this_cpu_dec(*sem->read_count);
103
104	/* Prod writer to recheck readers_active */
105	rcuwait_wake_up(&sem->writer);
106}
107EXPORT_SYMBOL_GPL(__percpu_up_read);
108
109#define per_cpu_sum(var)						\
110({									\
111	typeof(var) __sum = 0;						\
112	int cpu;							\
113	compiletime_assert_atomic_type(__sum);				\
114	for_each_possible_cpu(cpu)					\
115		__sum += per_cpu(var, cpu);				\
116	__sum;								\
117})
118
119/*
120 * Return true if the modular sum of the sem->read_count per-CPU variable is
121 * zero.  If this sum is zero, then it is stable due to the fact that if any
122 * newly arriving readers increment a given counter, they will immediately
123 * decrement that same counter.
124 */
125static bool readers_active_check(struct percpu_rw_semaphore *sem)
126{
127	if (per_cpu_sum(*sem->read_count) != 0)
128		return false;
129
130	/*
131	 * If we observed the decrement; ensure we see the entire critical
132	 * section.
133	 */
134
135	smp_mb(); /* C matches B */
136
137	return true;
138}
139
140void percpu_down_write(struct percpu_rw_semaphore *sem)
141{
142	/* Notify readers to take the slow path. */
143	rcu_sync_enter(&sem->rss);
144
145	down_write(&sem->rw_sem);
146
147	/*
148	 * Notify new readers to block; up until now, and thus throughout the
149	 * longish rcu_sync_enter() above, new readers could still come in.
150	 */
151	WRITE_ONCE(sem->readers_block, 1);
152
153	smp_mb(); /* D matches A */
154
155	/*
156	 * If they don't see our writer of readers_block, then we are
157	 * guaranteed to see their sem->read_count increment, and therefore
158	 * will wait for them.
159	 */
160
161	/* Wait for all now active readers to complete. */
162	rcuwait_wait_event(&sem->writer, readers_active_check(sem));
163}
164EXPORT_SYMBOL_GPL(percpu_down_write);
165
166void percpu_up_write(struct percpu_rw_semaphore *sem)
167{
168	/*
169	 * Signal the writer is done, no fast path yet.
170	 *
171	 * One reason that we cannot just immediately flip to readers_fast is
172	 * that new readers might fail to see the results of this writer's
173	 * critical section.
174	 *
175	 * Therefore we force it through the slow path which guarantees an
176	 * acquire and thereby guarantees the critical section's consistency.
177	 */
178	smp_store_release(&sem->readers_block, 0);
179
180	/*
181	 * Release the write lock, this will allow readers back in the game.
182	 */
183	up_write(&sem->rw_sem);
184
185	/*
186	 * Once this completes (at least one RCU-sched grace period hence) the
187	 * reader fast path will be available again. Safe to use outside the
188	 * exclusive write lock because its counting.
189	 */
190	rcu_sync_exit(&sem->rss);
191}
192EXPORT_SYMBOL_GPL(percpu_up_write);