xref: /linux-6.15/kernel/locking/percpu-rwsem.c (revision 206c98ff)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/atomic.h>
3 #include <linux/rwsem.h>
4 #include <linux/percpu.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 
11 #include "rwsem.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 	/* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
21 	rcu_sync_init(&sem->rss);
22 	init_rwsem(&sem->rw_sem);
23 	rcuwait_init(&sem->writer);
24 	sem->readers_block = 0;
25 #ifdef CONFIG_DEBUG_LOCK_ALLOC
26 	debug_check_no_locks_freed((void *)sem, sizeof(*sem));
27 	lockdep_init_map(&sem->dep_map, name, key, 0);
28 #endif
29 	return 0;
30 }
31 EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
32 
33 void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
34 {
35 	/*
36 	 * XXX: temporary kludge. The error path in alloc_super()
37 	 * assumes that percpu_free_rwsem() is safe after kzalloc().
38 	 */
39 	if (!sem->read_count)
40 		return;
41 
42 	rcu_sync_dtor(&sem->rss);
43 	free_percpu(sem->read_count);
44 	sem->read_count = NULL; /* catch use after free bugs */
45 }
46 EXPORT_SYMBOL_GPL(percpu_free_rwsem);
47 
48 bool __percpu_down_read(struct percpu_rw_semaphore *sem, bool try)
49 {
50 	/*
51 	 * Due to having preemption disabled the decrement happens on
52 	 * the same CPU as the increment, avoiding the
53 	 * increment-on-one-CPU-and-decrement-on-another problem.
54 	 *
55 	 * If the reader misses the writer's assignment of readers_block, then
56 	 * the writer is guaranteed to see the reader's increment.
57 	 *
58 	 * Conversely, any readers that increment their sem->read_count after
59 	 * the writer looks are guaranteed to see the readers_block value,
60 	 * which in turn means that they are guaranteed to immediately
61 	 * decrement their sem->read_count, so that it doesn't matter that the
62 	 * writer missed them.
63 	 */
64 
65 	smp_mb(); /* A matches D */
66 
67 	/*
68 	 * If !readers_block the critical section starts here, matched by the
69 	 * release in percpu_up_write().
70 	 */
71 	if (likely(!smp_load_acquire(&sem->readers_block)))
72 		return true;
73 
74 	/*
75 	 * Per the above comment; we still have preemption disabled and
76 	 * will thus decrement on the same CPU as we incremented.
77 	 */
78 	__percpu_up_read(sem);
79 
80 	if (try)
81 		return false;
82 
83 	/*
84 	 * We either call schedule() in the wait, or we'll fall through
85 	 * and reschedule on the preempt_enable() in percpu_down_read().
86 	 */
87 	preempt_enable_no_resched();
88 
89 	/*
90 	 * Avoid lockdep for the down/up_read() we already have them.
91 	 */
92 	__down_read(&sem->rw_sem);
93 	this_cpu_inc(*sem->read_count);
94 	__up_read(&sem->rw_sem);
95 
96 	preempt_disable();
97 	return true;
98 }
99 EXPORT_SYMBOL_GPL(__percpu_down_read);
100 
101 void __percpu_up_read(struct percpu_rw_semaphore *sem)
102 {
103 	smp_mb(); /* B matches C */
104 	/*
105 	 * In other words, if they see our decrement (presumably to aggregate
106 	 * zero, as that is the only time it matters) they will also see our
107 	 * critical section.
108 	 */
109 	__this_cpu_dec(*sem->read_count);
110 
111 	/* Prod writer to recheck readers_active */
112 	rcuwait_wake_up(&sem->writer);
113 }
114 EXPORT_SYMBOL_GPL(__percpu_up_read);
115 
116 #define per_cpu_sum(var)						\
117 ({									\
118 	typeof(var) __sum = 0;						\
119 	int cpu;							\
120 	compiletime_assert_atomic_type(__sum);				\
121 	for_each_possible_cpu(cpu)					\
122 		__sum += per_cpu(var, cpu);				\
123 	__sum;								\
124 })
125 
126 /*
127  * Return true if the modular sum of the sem->read_count per-CPU variable is
128  * zero.  If this sum is zero, then it is stable due to the fact that if any
129  * newly arriving readers increment a given counter, they will immediately
130  * decrement that same counter.
131  */
132 static bool readers_active_check(struct percpu_rw_semaphore *sem)
133 {
134 	if (per_cpu_sum(*sem->read_count) != 0)
135 		return false;
136 
137 	/*
138 	 * If we observed the decrement; ensure we see the entire critical
139 	 * section.
140 	 */
141 
142 	smp_mb(); /* C matches B */
143 
144 	return true;
145 }
146 
147 void percpu_down_write(struct percpu_rw_semaphore *sem)
148 {
149 	rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
150 
151 	/* Notify readers to take the slow path. */
152 	rcu_sync_enter(&sem->rss);
153 
154 	__down_write(&sem->rw_sem);
155 
156 	/*
157 	 * Notify new readers to block; up until now, and thus throughout the
158 	 * longish rcu_sync_enter() above, new readers could still come in.
159 	 */
160 	WRITE_ONCE(sem->readers_block, 1);
161 
162 	smp_mb(); /* D matches A */
163 
164 	/*
165 	 * If they don't see our writer of readers_block, then we are
166 	 * guaranteed to see their sem->read_count increment, and therefore
167 	 * will wait for them.
168 	 */
169 
170 	/* Wait for all now active readers to complete. */
171 	rcuwait_wait_event(&sem->writer, readers_active_check(sem));
172 }
173 EXPORT_SYMBOL_GPL(percpu_down_write);
174 
175 void percpu_up_write(struct percpu_rw_semaphore *sem)
176 {
177 	rwsem_release(&sem->dep_map, _RET_IP_);
178 
179 	/*
180 	 * Signal the writer is done, no fast path yet.
181 	 *
182 	 * One reason that we cannot just immediately flip to readers_fast is
183 	 * that new readers might fail to see the results of this writer's
184 	 * critical section.
185 	 *
186 	 * Therefore we force it through the slow path which guarantees an
187 	 * acquire and thereby guarantees the critical section's consistency.
188 	 */
189 	smp_store_release(&sem->readers_block, 0);
190 
191 	/*
192 	 * Release the write lock, this will allow readers back in the game.
193 	 */
194 	__up_write(&sem->rw_sem);
195 
196 	/*
197 	 * Once this completes (at least one RCU-sched grace period hence) the
198 	 * reader fast path will be available again. Safe to use outside the
199 	 * exclusive write lock because its counting.
200 	 */
201 	rcu_sync_exit(&sem->rss);
202 }
203 EXPORT_SYMBOL_GPL(percpu_up_write);
204