blob: 652a8ee8efe95acaf1db28eeed40bba820787222 [file] [log] [blame]
Oleg Nesterov9390ef02012-12-17 16:01:36 -08001#include <linux/atomic.h>
Oleg Nesterova1fd3e22012-12-17 16:01:32 -08002#include <linux/rwsem.h>
3#include <linux/percpu.h>
4#include <linux/wait.h>
Oleg Nesterov8ebe3472012-12-17 16:01:38 -08005#include <linux/lockdep.h>
Oleg Nesterova1fd3e22012-12-17 16:01:32 -08006#include <linux/percpu-rwsem.h>
7#include <linux/rcupdate.h>
8#include <linux/sched.h>
9#include <linux/errno.h>
10
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080011int __percpu_init_rwsem(struct percpu_rw_semaphore *brw,
12 const char *name, struct lock_class_key *rwsem_key)
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080013{
14 brw->fast_read_ctr = alloc_percpu(int);
15 if (unlikely(!brw->fast_read_ctr))
16 return -ENOMEM;
17
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080018 /* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
19 __init_rwsem(&brw->rw_sem, name, rwsem_key);
Oleg Nesterov9390ef02012-12-17 16:01:36 -080020 atomic_set(&brw->write_ctr, 0);
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080021 atomic_set(&brw->slow_read_ctr, 0);
22 init_waitqueue_head(&brw->write_waitq);
23 return 0;
24}
25
26void 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
Oleg Nesterov9390ef02012-12-17 16:01:36 -080034 * there is no pending writer (atomic_read(write_ctr) == 0) and inc/dec the
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080035 * 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 *
Oleg Nesterov9390ef02012-12-17 16:01:36 -080050 * But if it succeeds we do not have any barriers, atomic_read(write_ctr) or
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080051 * __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 */
55static bool update_fast_ctr(struct percpu_rw_semaphore *brw, unsigned int val)
56{
57 bool success = false;
58
59 preempt_disable();
Oleg Nesterov9390ef02012-12-17 16:01:36 -080060 if (likely(!atomic_read(&brw->write_ctr))) {
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080061 __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.
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080072 *
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().
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080076 */
77void percpu_down_read(struct percpu_rw_semaphore *brw)
78{
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080079 might_sleep();
80 if (likely(update_fast_ctr(brw, +1))) {
81 rwsem_acquire_read(&brw->rw_sem.dep_map, 0, 0, _RET_IP_);
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080082 return;
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080083 }
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080084
85 down_read(&brw->rw_sem);
86 atomic_inc(&brw->slow_read_ctr);
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080087 /* avoid up_read()->rwsem_release() */
88 __up_read(&brw->rw_sem);
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080089}
90
91void percpu_up_read(struct percpu_rw_semaphore *brw)
92{
Oleg Nesterov8ebe3472012-12-17 16:01:38 -080093 rwsem_release(&brw->rw_sem.dep_map, 1, _RET_IP_);
94
Oleg Nesterova1fd3e22012-12-17 16:01:32 -080095 if (likely(update_fast_ctr(brw, -1)))
96 return;
97
98 /* false-positive is possible but harmless */
99 if (atomic_dec_and_test(&brw->slow_read_ctr))
100 wake_up_all(&brw->write_waitq);
101}
102
103static int clear_fast_ctr(struct percpu_rw_semaphore *brw)
104{
105 unsigned int sum = 0;
106 int cpu;
107
108 for_each_possible_cpu(cpu) {
109 sum += per_cpu(*brw->fast_read_ctr, cpu);
110 per_cpu(*brw->fast_read_ctr, cpu) = 0;
111 }
112
113 return sum;
114}
115
116/*
Oleg Nesterov9390ef02012-12-17 16:01:36 -0800117 * A writer increments ->write_ctr to force the readers to switch to the
118 * slow mode, note the atomic_read() check in update_fast_ctr().
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800119 *
120 * After that the readers can only inc/dec the slow ->slow_read_ctr counter,
121 * ->fast_read_ctr is stable. Once the writer moves its sum into the slow
122 * counter it represents the number of active readers.
123 *
124 * Finally the writer takes ->rw_sem for writing and blocks the new readers,
125 * then waits until the slow counter becomes zero.
126 */
127void percpu_down_write(struct percpu_rw_semaphore *brw)
128{
Oleg Nesterov9390ef02012-12-17 16:01:36 -0800129 /* tell update_fast_ctr() there is a pending writer */
130 atomic_inc(&brw->write_ctr);
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800131 /*
Oleg Nesterov9390ef02012-12-17 16:01:36 -0800132 * 1. Ensures that write_ctr != 0 is visible to any down_read/up_read
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800133 * so that update_fast_ctr() can't succeed.
134 *
135 * 2. Ensures we see the result of every previous this_cpu_add() in
136 * update_fast_ctr().
137 *
138 * 3. Ensures that if any reader has exited its critical section via
139 * fast-path, it executes a full memory barrier before we return.
140 * See R_W case in the comment above update_fast_ctr().
141 */
142 synchronize_sched_expedited();
143
Oleg Nesterov9390ef02012-12-17 16:01:36 -0800144 /* exclude other writers, and block the new readers completely */
145 down_write(&brw->rw_sem);
146
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800147 /* nobody can use fast_read_ctr, move its sum into slow_read_ctr */
148 atomic_add(clear_fast_ctr(brw), &brw->slow_read_ctr);
149
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800150 /* wait for all readers to complete their percpu_up_read() */
151 wait_event(brw->write_waitq, !atomic_read(&brw->slow_read_ctr));
152}
153
154void percpu_up_write(struct percpu_rw_semaphore *brw)
155{
Oleg Nesterov9390ef02012-12-17 16:01:36 -0800156 /* release the lock, but the readers can't use the fast-path */
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800157 up_write(&brw->rw_sem);
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800158 /*
159 * Insert the barrier before the next fast-path in down_read,
160 * see W_R case in the comment above update_fast_ctr().
161 */
162 synchronize_sched_expedited();
Oleg Nesterov9390ef02012-12-17 16:01:36 -0800163 /* the last writer unblocks update_fast_ctr() */
164 atomic_dec(&brw->write_ctr);
Oleg Nesterova1fd3e22012-12-17 16:01:32 -0800165}