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 *sem,
12 const char *name, struct lock_class_key *rwsem_key)
13{
14 sem->read_count = alloc_percpu(int);
15 if (unlikely(!sem->read_count))
16 return -ENOMEM;
17
18 /* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */
19 rcu_sync_init(&sem->rss, RCU_SCHED_SYNC);
20 __init_rwsem(&sem->rw_sem, name, rwsem_key);
21 init_waitqueue_head(&sem->writer);
22 sem->readers_block = 0;
23 return 0;
24}
25EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
26
27void percpu_free_rwsem(struct percpu_rw_semaphore *sem)
28{
29 /*
30 * XXX: temporary kludge. The error path in alloc_super()
31 * assumes that percpu_free_rwsem() is safe after kzalloc().
32 */
33 if (!sem->read_count)
34 return;
35
36 rcu_sync_dtor(&sem->rss);
37 free_percpu(sem->read_count);
38 sem->read_count = NULL; /* catch use after free bugs */
39}
40EXPORT_SYMBOL_GPL(percpu_free_rwsem);
41
42int __percpu_down_read(struct percpu_rw_semaphore *sem, int try)
43{
44 /*
45 * Due to having preemption disabled the decrement happens on
46 * the same CPU as the increment, avoiding the
47 * increment-on-one-CPU-and-decrement-on-another problem.
48 *
49 * If the reader misses the writer's assignment of readers_block, then
50 * the writer is guaranteed to see the reader's increment.
51 *
52 * Conversely, any readers that increment their sem->read_count after
53 * the writer looks are guaranteed to see the readers_block value,
54 * which in turn means that they are guaranteed to immediately
55 * decrement their sem->read_count, so that it doesn't matter that the
56 * writer missed them.
57 */
58
59 smp_mb(); /* A matches D */
60
61 /*
62 * If !readers_block the critical section starts here, matched by the
63 * release in percpu_up_write().
64 */
65 if (likely(!smp_load_acquire(&sem->readers_block)))
66 return 1;
67
68 /*
69 * Per the above comment; we still have preemption disabled and
70 * will thus decrement on the same CPU as we incremented.
71 */
72 __percpu_up_read(sem);
73
74 if (try)
75 return 0;
76
77 /*
78 * We either call schedule() in the wait, or we'll fall through
79 * and reschedule on the preempt_enable() in percpu_down_read().
80 */
81 preempt_enable_no_resched();
82
83 /*
84 * Avoid lockdep for the down/up_read() we already have them.
85 */
86 __down_read(&sem->rw_sem);
87 this_cpu_inc(*sem->read_count);
88 __up_read(&sem->rw_sem);
89
90 preempt_disable();
91 return 1;
92}
93EXPORT_SYMBOL_GPL(__percpu_down_read);
94
95void __percpu_up_read(struct percpu_rw_semaphore *sem)
96{
97 smp_mb(); /* B matches C */
98 /*
99 * In other words, if they see our decrement (presumably to aggregate
100 * zero, as that is the only time it matters) they will also see our
101 * critical section.
102 */
103 __this_cpu_dec(*sem->read_count);
104
105 /* Prod writer to recheck readers_active */
106 wake_up(&sem->writer);
107}
108EXPORT_SYMBOL_GPL(__percpu_up_read);
109
110#define per_cpu_sum(var) \
111({ \
112 typeof(var) __sum = 0; \
113 int cpu; \
114 compiletime_assert_atomic_type(__sum); \
115 for_each_possible_cpu(cpu) \
116 __sum += per_cpu(var, cpu); \
117 __sum; \
118})
119
120/*
121 * Return true if the modular sum of the sem->read_count per-CPU variable is
122 * zero. If this sum is zero, then it is stable due to the fact that if any
123 * newly arriving readers increment a given counter, they will immediately
124 * decrement that same counter.
125 */
126static bool readers_active_check(struct percpu_rw_semaphore *sem)
127{
128 if (per_cpu_sum(*sem->read_count) != 0)
129 return false;
130
131 /*
132 * If we observed the decrement; ensure we see the entire critical
133 * section.
134 */
135
136 smp_mb(); /* C matches B */
137
138 return true;
139}
140
141void percpu_down_write(struct percpu_rw_semaphore *sem)
142{
143 /* Notify readers to take the slow path. */
144 rcu_sync_enter(&sem->rss);
145
146 down_write(&sem->rw_sem);
147
148 /*
149 * Notify new readers to block; up until now, and thus throughout the
150 * longish rcu_sync_enter() above, new readers could still come in.
151 */
152 WRITE_ONCE(sem->readers_block, 1);
153
154 smp_mb(); /* D matches A */
155
156 /*
157 * If they don't see our writer of readers_block, then we are
158 * guaranteed to see their sem->read_count increment, and therefore
159 * will wait for them.
160 */
161
162 /* Wait for all now active readers to complete. */
163 wait_event(sem->writer, readers_active_check(sem));
164}
165EXPORT_SYMBOL_GPL(percpu_down_write);
166
167void percpu_up_write(struct percpu_rw_semaphore *sem)
168{
169 /*
170 * Signal the writer is done, no fast path yet.
171 *
172 * One reason that we cannot just immediately flip to readers_fast is
173 * that new readers might fail to see the results of this writer's
174 * critical section.
175 *
176 * Therefore we force it through the slow path which guarantees an
177 * acquire and thereby guarantees the critical section's consistency.
178 */
179 smp_store_release(&sem->readers_block, 0);
180
181 /*
182 * Release the write lock, this will allow readers back in the game.
183 */
184 up_write(&sem->rw_sem);
185
186 /*
187 * Once this completes (at least one RCU-sched grace period hence) the
188 * reader fast path will be available again. Safe to use outside the
189 * exclusive write lock because its counting.
190 */
191 rcu_sync_exit(&sem->rss);
192}
193EXPORT_SYMBOL_GPL(percpu_up_write);
1// SPDX-License-Identifier: GPL-2.0-only
2#include <linux/atomic.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/sched/task.h>
10#include <linux/sched/debug.h>
11#include <linux/errno.h>
12#include <trace/events/lock.h>
13
14int __percpu_init_rwsem(struct percpu_rw_semaphore *sem,
15 const char *name, struct lock_class_key *key)
16{
17 sem->read_count = alloc_percpu(int);
18 if (unlikely(!sem->read_count))
19 return -ENOMEM;
20
21 rcu_sync_init(&sem->rss);
22 rcuwait_init(&sem->writer);
23 init_waitqueue_head(&sem->waiters);
24 atomic_set(&sem->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}
31EXPORT_SYMBOL_GPL(__percpu_init_rwsem);
32
33void 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}
46EXPORT_SYMBOL_GPL(percpu_free_rwsem);
47
48static bool __percpu_down_read_trylock(struct percpu_rw_semaphore *sem)
49{
50 this_cpu_inc(*sem->read_count);
51
52 /*
53 * Due to having preemption disabled the decrement happens on
54 * the same CPU as the increment, avoiding the
55 * increment-on-one-CPU-and-decrement-on-another problem.
56 *
57 * If the reader misses the writer's assignment of sem->block, then the
58 * writer is guaranteed to see the reader's increment.
59 *
60 * Conversely, any readers that increment their sem->read_count after
61 * the writer looks are guaranteed to see the sem->block value, which
62 * in turn means that they are guaranteed to immediately decrement
63 * their sem->read_count, so that it doesn't matter that the writer
64 * missed them.
65 */
66
67 smp_mb(); /* A matches D */
68
69 /*
70 * If !sem->block the critical section starts here, matched by the
71 * release in percpu_up_write().
72 */
73 if (likely(!atomic_read_acquire(&sem->block)))
74 return true;
75
76 this_cpu_dec(*sem->read_count);
77
78 /* Prod writer to re-evaluate readers_active_check() */
79 rcuwait_wake_up(&sem->writer);
80
81 return false;
82}
83
84static inline bool __percpu_down_write_trylock(struct percpu_rw_semaphore *sem)
85{
86 if (atomic_read(&sem->block))
87 return false;
88
89 return atomic_xchg(&sem->block, 1) == 0;
90}
91
92static bool __percpu_rwsem_trylock(struct percpu_rw_semaphore *sem, bool reader)
93{
94 if (reader) {
95 bool ret;
96
97 preempt_disable();
98 ret = __percpu_down_read_trylock(sem);
99 preempt_enable();
100
101 return ret;
102 }
103 return __percpu_down_write_trylock(sem);
104}
105
106/*
107 * The return value of wait_queue_entry::func means:
108 *
109 * <0 - error, wakeup is terminated and the error is returned
110 * 0 - no wakeup, a next waiter is tried
111 * >0 - woken, if EXCLUSIVE, counted towards @nr_exclusive.
112 *
113 * We use EXCLUSIVE for both readers and writers to preserve FIFO order,
114 * and play games with the return value to allow waking multiple readers.
115 *
116 * Specifically, we wake readers until we've woken a single writer, or until a
117 * trylock fails.
118 */
119static int percpu_rwsem_wake_function(struct wait_queue_entry *wq_entry,
120 unsigned int mode, int wake_flags,
121 void *key)
122{
123 bool reader = wq_entry->flags & WQ_FLAG_CUSTOM;
124 struct percpu_rw_semaphore *sem = key;
125 struct task_struct *p;
126
127 /* concurrent against percpu_down_write(), can get stolen */
128 if (!__percpu_rwsem_trylock(sem, reader))
129 return 1;
130
131 p = get_task_struct(wq_entry->private);
132 list_del_init(&wq_entry->entry);
133 smp_store_release(&wq_entry->private, NULL);
134
135 wake_up_process(p);
136 put_task_struct(p);
137
138 return !reader; /* wake (readers until) 1 writer */
139}
140
141static void percpu_rwsem_wait(struct percpu_rw_semaphore *sem, bool reader)
142{
143 DEFINE_WAIT_FUNC(wq_entry, percpu_rwsem_wake_function);
144 bool wait;
145
146 spin_lock_irq(&sem->waiters.lock);
147 /*
148 * Serialize against the wakeup in percpu_up_write(), if we fail
149 * the trylock, the wakeup must see us on the list.
150 */
151 wait = !__percpu_rwsem_trylock(sem, reader);
152 if (wait) {
153 wq_entry.flags |= WQ_FLAG_EXCLUSIVE | reader * WQ_FLAG_CUSTOM;
154 __add_wait_queue_entry_tail(&sem->waiters, &wq_entry);
155 }
156 spin_unlock_irq(&sem->waiters.lock);
157
158 while (wait) {
159 set_current_state(TASK_UNINTERRUPTIBLE);
160 if (!smp_load_acquire(&wq_entry.private))
161 break;
162 schedule();
163 }
164 __set_current_state(TASK_RUNNING);
165}
166
167bool __sched __percpu_down_read(struct percpu_rw_semaphore *sem, bool try)
168{
169 if (__percpu_down_read_trylock(sem))
170 return true;
171
172 if (try)
173 return false;
174
175 trace_contention_begin(sem, LCB_F_PERCPU | LCB_F_READ);
176 preempt_enable();
177 percpu_rwsem_wait(sem, /* .reader = */ true);
178 preempt_disable();
179 trace_contention_end(sem, 0);
180
181 return true;
182}
183EXPORT_SYMBOL_GPL(__percpu_down_read);
184
185#define per_cpu_sum(var) \
186({ \
187 typeof(var) __sum = 0; \
188 int cpu; \
189 compiletime_assert_atomic_type(__sum); \
190 for_each_possible_cpu(cpu) \
191 __sum += per_cpu(var, cpu); \
192 __sum; \
193})
194
195bool percpu_is_read_locked(struct percpu_rw_semaphore *sem)
196{
197 return per_cpu_sum(*sem->read_count) != 0 && !atomic_read(&sem->block);
198}
199EXPORT_SYMBOL_GPL(percpu_is_read_locked);
200
201/*
202 * Return true if the modular sum of the sem->read_count per-CPU variable is
203 * zero. If this sum is zero, then it is stable due to the fact that if any
204 * newly arriving readers increment a given counter, they will immediately
205 * decrement that same counter.
206 *
207 * Assumes sem->block is set.
208 */
209static bool readers_active_check(struct percpu_rw_semaphore *sem)
210{
211 if (per_cpu_sum(*sem->read_count) != 0)
212 return false;
213
214 /*
215 * If we observed the decrement; ensure we see the entire critical
216 * section.
217 */
218
219 smp_mb(); /* C matches B */
220
221 return true;
222}
223
224void __sched percpu_down_write(struct percpu_rw_semaphore *sem)
225{
226 might_sleep();
227 rwsem_acquire(&sem->dep_map, 0, 0, _RET_IP_);
228 trace_contention_begin(sem, LCB_F_PERCPU | LCB_F_WRITE);
229
230 /* Notify readers to take the slow path. */
231 rcu_sync_enter(&sem->rss);
232
233 /*
234 * Try set sem->block; this provides writer-writer exclusion.
235 * Having sem->block set makes new readers block.
236 */
237 if (!__percpu_down_write_trylock(sem))
238 percpu_rwsem_wait(sem, /* .reader = */ false);
239
240 /* smp_mb() implied by __percpu_down_write_trylock() on success -- D matches A */
241
242 /*
243 * If they don't see our store of sem->block, then we are guaranteed to
244 * see their sem->read_count increment, and therefore will wait for
245 * them.
246 */
247
248 /* Wait for all active readers to complete. */
249 rcuwait_wait_event(&sem->writer, readers_active_check(sem), TASK_UNINTERRUPTIBLE);
250 trace_contention_end(sem, 0);
251}
252EXPORT_SYMBOL_GPL(percpu_down_write);
253
254void percpu_up_write(struct percpu_rw_semaphore *sem)
255{
256 rwsem_release(&sem->dep_map, _RET_IP_);
257
258 /*
259 * Signal the writer is done, no fast path yet.
260 *
261 * One reason that we cannot just immediately flip to readers_fast is
262 * that new readers might fail to see the results of this writer's
263 * critical section.
264 *
265 * Therefore we force it through the slow path which guarantees an
266 * acquire and thereby guarantees the critical section's consistency.
267 */
268 atomic_set_release(&sem->block, 0);
269
270 /*
271 * Prod any pending reader/writer to make progress.
272 */
273 __wake_up(&sem->waiters, TASK_NORMAL, 1, sem);
274
275 /*
276 * Once this completes (at least one RCU-sched grace period hence) the
277 * reader fast path will be available again. Safe to use outside the
278 * exclusive write lock because its counting.
279 */
280 rcu_sync_exit(&sem->rss);
281}
282EXPORT_SYMBOL_GPL(percpu_up_write);