xref: /openbmc/linux/kernel/locking/percpu-rwsem.c (revision eb3fcf00)
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 
11 int __percpu_init_rwsem(struct percpu_rw_semaphore *brw,
12 			const char *name, struct lock_class_key *rwsem_key)
13 {
14 	brw->fast_read_ctr = alloc_percpu(int);
15 	if (unlikely(!brw->fast_read_ctr))
16 		return -ENOMEM;
17 
18 	/* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
19 	__init_rwsem(&brw->rw_sem, name, rwsem_key);
20 	atomic_set(&brw->write_ctr, 0);
21 	atomic_set(&brw->slow_read_ctr, 0);
22 	init_waitqueue_head(&brw->write_waitq);
23 	return 0;
24 }
25 
26 void percpu_free_rwsem(struct percpu_rw_semaphore *brw)
27 {
28 	free_percpu(brw->fast_read_ctr);
29 	brw->fast_read_ctr = NULL; /* catch use after free bugs */
30 }
31 
32 /*
33  * This is the fast-path for down_read/up_read, it only needs to ensure
34  * there is no pending writer (atomic_read(write_ctr) == 0) and inc/dec the
35  * fast per-cpu counter. The writer uses synchronize_sched_expedited() to
36  * serialize with the preempt-disabled section below.
37  *
38  * The nontrivial part is that we should guarantee acquire/release semantics
39  * in case when
40  *
41  *	R_W: down_write() comes after up_read(), the writer should see all
42  *	     changes done by the reader
43  * or
44  *	W_R: down_read() comes after up_write(), the reader should see all
45  *	     changes done by the writer
46  *
47  * If this helper fails the callers rely on the normal rw_semaphore and
48  * atomic_dec_and_test(), so in this case we have the necessary barriers.
49  *
50  * But if it succeeds we do not have any barriers, atomic_read(write_ctr) or
51  * __this_cpu_add() below can be reordered with any LOAD/STORE done by the
52  * reader inside the critical section. See the comments in down_write and
53  * up_write below.
54  */
55 static bool update_fast_ctr(struct percpu_rw_semaphore *brw, unsigned int val)
56 {
57 	bool success = false;
58 
59 	preempt_disable();
60 	if (likely(!atomic_read(&brw->write_ctr))) {
61 		__this_cpu_add(*brw->fast_read_ctr, val);
62 		success = true;
63 	}
64 	preempt_enable();
65 
66 	return success;
67 }
68 
69 /*
70  * Like the normal down_read() this is not recursive, the writer can
71  * come after the first percpu_down_read() and create the deadlock.
72  *
73  * Note: returns with lock_is_held(brw->rw_sem) == T for lockdep,
74  * percpu_up_read() does rwsem_release(). This pairs with the usage
75  * of ->rw_sem in percpu_down/up_write().
76  */
77 void percpu_down_read(struct percpu_rw_semaphore *brw)
78 {
79 	might_sleep();
80 	if (likely(update_fast_ctr(brw, +1))) {
81 		rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 0, _RET_IP_);
82 		return;
83 	}
84 
85 	down_read(&brw->rw_sem);
86 	atomic_inc(&brw->slow_read_ctr);
87 	/* avoid up_read()->rwsem_release() */
88 	__up_read(&brw->rw_sem);
89 }
90 
91 int percpu_down_read_trylock(struct percpu_rw_semaphore *brw)
92 {
93 	if (unlikely(!update_fast_ctr(brw, +1))) {
94 		if (!__down_read_trylock(&brw->rw_sem))
95 			return 0;
96 		atomic_inc(&brw->slow_read_ctr);
97 		__up_read(&brw->rw_sem);
98 	}
99 
100 	rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 1, _RET_IP_);
101 	return 1;
102 }
103 
104 void percpu_up_read(struct percpu_rw_semaphore *brw)
105 {
106 	rwsem_release(&brw->rw_sem.dep_map, 1, _RET_IP_);
107 
108 	if (likely(update_fast_ctr(brw, -1)))
109 		return;
110 
111 	/* false-positive is possible but harmless */
112 	if (atomic_dec_and_test(&brw->slow_read_ctr))
113 		wake_up_all(&brw->write_waitq);
114 }
115 
116 static int clear_fast_ctr(struct percpu_rw_semaphore *brw)
117 {
118 	unsigned int sum = 0;
119 	int cpu;
120 
121 	for_each_possible_cpu(cpu) {
122 		sum += per_cpu(*brw->fast_read_ctr, cpu);
123 		per_cpu(*brw->fast_read_ctr, cpu) = 0;
124 	}
125 
126 	return sum;
127 }
128 
129 /*
130  * A writer increments ->write_ctr to force the readers to switch to the
131  * slow mode, note the atomic_read() check in update_fast_ctr().
132  *
133  * After that the readers can only inc/dec the slow ->slow_read_ctr counter,
134  * ->fast_read_ctr is stable. Once the writer moves its sum into the slow
135  * counter it represents the number of active readers.
136  *
137  * Finally the writer takes ->rw_sem for writing and blocks the new readers,
138  * then waits until the slow counter becomes zero.
139  */
140 void percpu_down_write(struct percpu_rw_semaphore *brw)
141 {
142 	/* tell update_fast_ctr() there is a pending writer */
143 	atomic_inc(&brw->write_ctr);
144 	/*
145 	 * 1. Ensures that write_ctr != 0 is visible to any down_read/up_read
146 	 *    so that update_fast_ctr() can't succeed.
147 	 *
148 	 * 2. Ensures we see the result of every previous this_cpu_add() in
149 	 *    update_fast_ctr().
150 	 *
151 	 * 3. Ensures that if any reader has exited its critical section via
152 	 *    fast-path, it executes a full memory barrier before we return.
153 	 *    See R_W case in the comment above update_fast_ctr().
154 	 */
155 	synchronize_sched_expedited();
156 
157 	/* exclude other writers, and block the new readers completely */
158 	down_write(&brw->rw_sem);
159 
160 	/* nobody can use fast_read_ctr, move its sum into slow_read_ctr */
161 	atomic_add(clear_fast_ctr(brw), &brw->slow_read_ctr);
162 
163 	/* wait for all readers to complete their percpu_up_read() */
164 	wait_event(brw->write_waitq, !atomic_read(&brw->slow_read_ctr));
165 }
166 
167 void percpu_up_write(struct percpu_rw_semaphore *brw)
168 {
169 	/* release the lock, but the readers can't use the fast-path */
170 	up_write(&brw->rw_sem);
171 	/*
172 	 * Insert the barrier before the next fast-path in down_read,
173 	 * see W_R case in the comment above update_fast_ctr().
174 	 */
175 	synchronize_sched_expedited();
176 	/* the last writer unblocks update_fast_ctr() */
177 	atomic_dec(&brw->write_ctr);
178 }
179