Loading...
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 *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
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
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 */
55static 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 */
77void 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
91void percpu_up_read(struct percpu_rw_semaphore *brw)
92{
93 rwsem_release(&brw->rw_sem.dep_map, 1, _RET_IP_);
94
95 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/*
117 * 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().
119 *
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{
129 /* tell update_fast_ctr() there is a pending writer */
130 atomic_inc(&brw->write_ctr);
131 /*
132 * 1. Ensures that write_ctr != 0 is visible to any down_read/up_read
133 * 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
144 /* exclude other writers, and block the new readers completely */
145 down_write(&brw->rw_sem);
146
147 /* 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
150 /* 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{
156 /* release the lock, but the readers can't use the fast-path */
157 up_write(&brw->rw_sem);
158 /*
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();
163 /* the last writer unblocks update_fast_ctr() */
164 atomic_dec(&brw->write_ctr);
165}
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
13int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
14 const char *name, struct lock_class_key *rwsem_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, name, rwsem_key);
23 rcuwait_init(&sem->writer);
24 sem->readers_block = 0;
25 return 0;
26}
27EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
28
29void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
30{
31 /*
32 * XXX: temporary kludge. The error path in alloc_super()
33 * assumes that percpu_free_rwsem() is safe after kzalloc().
34 */
35 if (!sem->read_count)
36 return;
37
38 rcu_sync_dtor(&sem->rss);
39 free_percpu(sem->read_count);
40 sem->read_count = NULL; /* catch use after free bugs */
41}
42EXPORT_SYMBOL_GPL(percpu_free_rwsem);
43
44int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
45{
46 /*
47 * Due to having preemption disabled the decrement happens on
48 * the same CPU as the increment, avoiding the
49 * increment-on-one-CPU-and-decrement-on-another problem.
50 *
51 * If the reader misses the writer's assignment of readers_block, then
52 * the writer is guaranteed to see the reader's increment.
53 *
54 * Conversely, any readers that increment their sem->read_count after
55 * the writer looks are guaranteed to see the readers_block value,
56 * which in turn means that they are guaranteed to immediately
57 * decrement their sem->read_count, so that it doesn't matter that the
58 * writer missed them.
59 */
60
61 smp_mb(); /* A matches D */
62
63 /*
64 * If !readers_block the critical section starts here, matched by the
65 * release in percpu_up_write().
66 */
67 if (likely(!smp_load_acquire(&sem->readers_block)))
68 return 1;
69
70 /*
71 * Per the above comment; we still have preemption disabled and
72 * will thus decrement on the same CPU as we incremented.
73 */
74 __percpu_up_read(sem);
75
76 if (try)
77 return 0;
78
79 /*
80 * We either call schedule() in the wait, or we'll fall through
81 * and reschedule on the preempt_enable() in percpu_down_read().
82 */
83 preempt_enable_no_resched();
84
85 /*
86 * Avoid lockdep for the down/up_read() we already have them.
87 */
88 __down_read(&sem->rw_sem);
89 this_cpu_inc(*sem->read_count);
90 __up_read(&sem->rw_sem);
91
92 preempt_disable();
93 return 1;
94}
95EXPORT_SYMBOL_GPL(__percpu_down_read);
96
97void __percpu_up_read(struct percpu_rw_semaphore *sem)
98{
99 smp_mb(); /* B matches C */
100 /*
101 * In other words, if they see our decrement (presumably to aggregate
102 * zero, as that is the only time it matters) they will also see our
103 * critical section.
104 */
105 __this_cpu_dec(*sem->read_count);
106
107 /* Prod writer to recheck readers_active */
108 rcuwait_wake_up(&sem->writer);
109}
110EXPORT_SYMBOL_GPL(__percpu_up_read);
111
112#define per_cpu_sum(var) \
113({ \
114 typeof(var) __sum = 0; \
115 int cpu; \
116 compiletime_assert_atomic_type(__sum); \
117 for_each_possible_cpu(cpu) \
118 __sum += per_cpu(var, cpu); \
119 __sum; \
120})
121
122/*
123 * Return true if the modular sum of the sem->read_count per-CPU variable is
124 * zero. If this sum is zero, then it is stable due to the fact that if any
125 * newly arriving readers increment a given counter, they will immediately
126 * decrement that same counter.
127 */
128static bool readers_active_check(struct percpu_rw_semaphore *sem)
129{
130 if (per_cpu_sum(*sem->read_count) != 0)
131 return false;
132
133 /*
134 * If we observed the decrement; ensure we see the entire critical
135 * section.
136 */
137
138 smp_mb(); /* C matches B */
139
140 return true;
141}
142
143void percpu_down_write(struct percpu_rw_semaphore *sem)
144{
145 /* Notify readers to take the slow path. */
146 rcu_sync_enter(&sem->rss);
147
148 down_write(&sem->rw_sem);
149
150 /*
151 * Notify new readers to block; up until now, and thus throughout the
152 * longish rcu_sync_enter() above, new readers could still come in.
153 */
154 WRITE_ONCE(sem->readers_block, 1);
155
156 smp_mb(); /* D matches A */
157
158 /*
159 * If they don't see our writer of readers_block, then we are
160 * guaranteed to see their sem->read_count increment, and therefore
161 * will wait for them.
162 */
163
164 /* Wait for all now active readers to complete. */
165 rcuwait_wait_event(&sem->writer, readers_active_check(sem));
166}
167EXPORT_SYMBOL_GPL(percpu_down_write);
168
169void percpu_up_write(struct percpu_rw_semaphore *sem)
170{
171 /*
172 * Signal the writer is done, no fast path yet.
173 *
174 * One reason that we cannot just immediately flip to readers_fast is
175 * that new readers might fail to see the results of this writer's
176 * critical section.
177 *
178 * Therefore we force it through the slow path which guarantees an
179 * acquire and thereby guarantees the critical section's consistency.
180 */
181 smp_store_release(&sem->readers_block, 0);
182
183 /*
184 * Release the write lock, this will allow readers back in the game.
185 */
186 up_write(&sem->rw_sem);
187
188 /*
189 * Once this completes (at least one RCU-sched grace period hence) the
190 * reader fast path will be available again. Safe to use outside the
191 * exclusive write lock because its counting.
192 */
193 rcu_sync_exit(&sem->rss);
194}
195EXPORT_SYMBOL_GPL(percpu_up_write);