projects/linux·blob v5.15 · refs

477 lines · 15.3 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2#define pr_fmt(fmt) "%s: " fmt, __func__
3
4#include <linux/kernel.h>
5#include <linux/sched.h>
6#include <linux/wait.h>
7#include <linux/slab.h>
8#include <linux/mm.h>
9#include <linux/percpu-refcount.h>
10
11/*
12 * Initially, a percpu refcount is just a set of percpu counters. Initially, we
13 * don't try to detect the ref hitting 0 - which means that get/put can just
14 * increment or decrement the local counter. Note that the counter on a
15 * particular cpu can (and will) wrap - this is fine, when we go to shutdown the
16 * percpu counters will all sum to the correct value
17 *
18 * (More precisely: because modular arithmetic is commutative the sum of all the
19 * percpu_count vars will be equal to what it would have been if all the gets
20 * and puts were done to a single integer, even if some of the percpu integers
21 * overflow or underflow).
22 *
23 * The real trick to implementing percpu refcounts is shutdown. We can't detect
24 * the ref hitting 0 on every put - this would require global synchronization
25 * and defeat the whole purpose of using percpu refs.
26 *
27 * What we do is require the user to keep track of the initial refcount; we know
28 * the ref can't hit 0 before the user drops the initial ref, so as long as we
29 * convert to non percpu mode before the initial ref is dropped everything
30 * works.
31 *
32 * Converting to non percpu mode is done with some RCUish stuff in
33 * percpu_ref_kill. Additionally, we need a bias value so that the
34 * atomic_long_t can't hit 0 before we've added up all the percpu refs.
35 */
36
37#define PERCPU_COUNT_BIAS (1LU << (BITS_PER_LONG - 1))
38
39static DEFINE_SPINLOCK(percpu_ref_switch_lock);
40static DECLARE_WAIT_QUEUE_HEAD(percpu_ref_switch_waitq);
41
42static unsigned long __percpu *percpu_count_ptr(struct percpu_ref *ref)
43{
44 return (unsigned long __percpu *)
45 (ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC_DEAD);
46}
47
48/**
49 * percpu_ref_init - initialize a percpu refcount
50 * @ref: percpu_ref to initialize
51 * @release: function which will be called when refcount hits 0
52 * @flags: PERCPU_REF_INIT_* flags
53 * @gfp: allocation mask to use
54 *
55 * Initializes @ref. @ref starts out in percpu mode with a refcount of 1 unless
56 * @flags contains PERCPU_REF_INIT_ATOMIC or PERCPU_REF_INIT_DEAD. These flags
57 * change the start state to atomic with the latter setting the initial refcount
58 * to 0. See the definitions of PERCPU_REF_INIT_* flags for flag behaviors.
59 *
60 * Note that @release must not sleep - it may potentially be called from RCU
61 * callback context by percpu_ref_kill().
62 */
63int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
64 unsigned int flags, gfp_t gfp)
65{
66 size_t align = max_t(size_t, 1 << __PERCPU_REF_FLAG_BITS,
67 __alignof__(unsigned long));
68 unsigned long start_count = 0;
69 struct percpu_ref_data *data;
70
71 ref->percpu_count_ptr = (unsigned long)
72 __alloc_percpu_gfp(sizeof(unsigned long), align, gfp);
73 if (!ref->percpu_count_ptr)
74 return -ENOMEM;
75
76 data = kzalloc(sizeof(*ref->data), gfp);
77 if (!data) {
78 free_percpu((void __percpu *)ref->percpu_count_ptr);
79 return -ENOMEM;
80 }
81
82 data->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
83 data->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT;
84
85 if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD)) {
86 ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
87 data->allow_reinit = true;
88 } else {
89 start_count += PERCPU_COUNT_BIAS;
90 }
91
92 if (flags & PERCPU_REF_INIT_DEAD)
93 ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
94 else
95 start_count++;
96
97 atomic_long_set(&data->count, start_count);
98
99 data->release = release;
100 data->confirm_switch = NULL;
101 data->ref = ref;
102 ref->data = data;
103 return 0;
104}
105EXPORT_SYMBOL_GPL(percpu_ref_init);
106
107static void __percpu_ref_exit(struct percpu_ref *ref)
108{
109 unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
110
111 if (percpu_count) {
112 /* non-NULL confirm_switch indicates switching in progress */
113 WARN_ON_ONCE(ref->data && ref->data->confirm_switch);
114 free_percpu(percpu_count);
115 ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
116 }
117}
118
119/**
120 * percpu_ref_exit - undo percpu_ref_init()
121 * @ref: percpu_ref to exit
122 *
123 * This function exits @ref. The caller is responsible for ensuring that
124 * @ref is no longer in active use. The usual places to invoke this
125 * function from are the @ref->release() callback or in init failure path
126 * where percpu_ref_init() succeeded but other parts of the initialization
127 * of the embedding object failed.
128 */
129void percpu_ref_exit(struct percpu_ref *ref)
130{
131 struct percpu_ref_data *data = ref->data;
132 unsigned long flags;
133
134 __percpu_ref_exit(ref);
135
136 if (!data)
137 return;
138
139 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
140 ref->percpu_count_ptr |= atomic_long_read(&ref->data->count) <<
141 __PERCPU_REF_FLAG_BITS;
142 ref->data = NULL;
143 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
144
145 kfree(data);
146}
147EXPORT_SYMBOL_GPL(percpu_ref_exit);
148
149static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
150{
151 struct percpu_ref_data *data = container_of(rcu,
152 struct percpu_ref_data, rcu);
153 struct percpu_ref *ref = data->ref;
154
155 data->confirm_switch(ref);
156 data->confirm_switch = NULL;
157 wake_up_all(&percpu_ref_switch_waitq);
158
159 if (!data->allow_reinit)
160 __percpu_ref_exit(ref);
161
162 /* drop ref from percpu_ref_switch_to_atomic() */
163 percpu_ref_put(ref);
164}
165
166static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
167{
168 struct percpu_ref_data *data = container_of(rcu,
169 struct percpu_ref_data, rcu);
170 struct percpu_ref *ref = data->ref;
171 unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
172 static atomic_t underflows;
173 unsigned long count = 0;
174 int cpu;
175
176 for_each_possible_cpu(cpu)
177 count += *per_cpu_ptr(percpu_count, cpu);
178
179 pr_debug("global %lu percpu %lu\n",
180 atomic_long_read(&data->count), count);
181
182 /*
183 * It's crucial that we sum the percpu counters _before_ adding the sum
184 * to &ref->count; since gets could be happening on one cpu while puts
185 * happen on another, adding a single cpu's count could cause
186 * @ref->count to hit 0 before we've got a consistent value - but the
187 * sum of all the counts will be consistent and correct.
188 *
189 * Subtracting the bias value then has to happen _after_ adding count to
190 * &ref->count; we need the bias value to prevent &ref->count from
191 * reaching 0 before we add the percpu counts. But doing it at the same
192 * time is equivalent and saves us atomic operations:
193 */
194 atomic_long_add((long)count - PERCPU_COUNT_BIAS, &data->count);
195
196 if (WARN_ONCE(atomic_long_read(&data->count) <= 0,
197 "percpu ref (%ps) <= 0 (%ld) after switching to atomic",
198 data->release, atomic_long_read(&data->count)) &&
199 atomic_inc_return(&underflows) < 4) {
200 pr_err("%s(): percpu_ref underflow", __func__);
201 mem_dump_obj(data);
202 }
203
204 /* @ref is viewed as dead on all CPUs, send out switch confirmation */
205 percpu_ref_call_confirm_rcu(rcu);
206}
207
208static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref)
209{
210}
211
212static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
213 percpu_ref_func_t *confirm_switch)
214{
215 if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) {
216 if (confirm_switch)
217 confirm_switch(ref);
218 return;
219 }
220
221 /* switching from percpu to atomic */
222 ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
223
224 /*
225 * Non-NULL ->confirm_switch is used to indicate that switching is
226 * in progress. Use noop one if unspecified.
227 */
228 ref->data->confirm_switch = confirm_switch ?:
229 percpu_ref_noop_confirm_switch;
230
231 percpu_ref_get(ref); /* put after confirmation */
232 call_rcu(&ref->data->rcu, percpu_ref_switch_to_atomic_rcu);
233}
234
235static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
236{
237 unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
238 int cpu;
239
240 BUG_ON(!percpu_count);
241
242 if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
243 return;
244
245 if (WARN_ON_ONCE(!ref->data->allow_reinit))
246 return;
247
248 atomic_long_add(PERCPU_COUNT_BIAS, &ref->data->count);
249
250 /*
251 * Restore per-cpu operation. smp_store_release() is paired
252 * with READ_ONCE() in __ref_is_percpu() and guarantees that the
253 * zeroing is visible to all percpu accesses which can see the
254 * following __PERCPU_REF_ATOMIC clearing.
255 */
256 for_each_possible_cpu(cpu)
257 *per_cpu_ptr(percpu_count, cpu) = 0;
258
259 smp_store_release(&ref->percpu_count_ptr,
260 ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);
261}
262
263static void __percpu_ref_switch_mode(struct percpu_ref *ref,
264 percpu_ref_func_t *confirm_switch)
265{
266 struct percpu_ref_data *data = ref->data;
267
268 lockdep_assert_held(&percpu_ref_switch_lock);
269
270 /*
271 * If the previous ATOMIC switching hasn't finished yet, wait for
272 * its completion. If the caller ensures that ATOMIC switching
273 * isn't in progress, this function can be called from any context.
274 */
275 wait_event_lock_irq(percpu_ref_switch_waitq, !data->confirm_switch,
276 percpu_ref_switch_lock);
277
278 if (data->force_atomic || percpu_ref_is_dying(ref))
279 __percpu_ref_switch_to_atomic(ref, confirm_switch);
280 else
281 __percpu_ref_switch_to_percpu(ref);
282}
283
284/**
285 * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode
286 * @ref: percpu_ref to switch to atomic mode
287 * @confirm_switch: optional confirmation callback
288 *
289 * There's no reason to use this function for the usual reference counting.
290 * Use percpu_ref_kill[_and_confirm]().
291 *
292 * Schedule switching of @ref to atomic mode. All its percpu counts will
293 * be collected to the main atomic counter. On completion, when all CPUs
294 * are guaraneed to be in atomic mode, @confirm_switch, which may not
295 * block, is invoked. This function may be invoked concurrently with all
296 * the get/put operations and can safely be mixed with kill and reinit
297 * operations. Note that @ref will stay in atomic mode across kill/reinit
298 * cycles until percpu_ref_switch_to_percpu() is called.
299 *
300 * This function may block if @ref is in the process of switching to atomic
301 * mode. If the caller ensures that @ref is not in the process of
302 * switching to atomic mode, this function can be called from any context.
303 */
304void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
305 percpu_ref_func_t *confirm_switch)
306{
307 unsigned long flags;
308
309 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
310
311 ref->data->force_atomic = true;
312 __percpu_ref_switch_mode(ref, confirm_switch);
313
314 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
315}
316EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
317
318/**
319 * percpu_ref_switch_to_atomic_sync - switch a percpu_ref to atomic mode
320 * @ref: percpu_ref to switch to atomic mode
321 *
322 * Schedule switching the ref to atomic mode, and wait for the
323 * switch to complete. Caller must ensure that no other thread
324 * will switch back to percpu mode.
325 */
326void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
327{
328 percpu_ref_switch_to_atomic(ref, NULL);
329 wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
330}
331EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);
332
333/**
334 * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode
335 * @ref: percpu_ref to switch to percpu mode
336 *
337 * There's no reason to use this function for the usual reference counting.
338 * To re-use an expired ref, use percpu_ref_reinit().
339 *
340 * Switch @ref to percpu mode. This function may be invoked concurrently
341 * with all the get/put operations and can safely be mixed with kill and
342 * reinit operations. This function reverses the sticky atomic state set
343 * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic(). If @ref is
344 * dying or dead, the actual switching takes place on the following
345 * percpu_ref_reinit().
346 *
347 * This function may block if @ref is in the process of switching to atomic
348 * mode. If the caller ensures that @ref is not in the process of
349 * switching to atomic mode, this function can be called from any context.
350 */
351void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
352{
353 unsigned long flags;
354
355 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
356
357 ref->data->force_atomic = false;
358 __percpu_ref_switch_mode(ref, NULL);
359
360 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
361}
362EXPORT_SYMBOL_GPL(percpu_ref_switch_to_percpu);
363
364/**
365 * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
366 * @ref: percpu_ref to kill
367 * @confirm_kill: optional confirmation callback
368 *
369 * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
370 * @confirm_kill is not NULL. @confirm_kill, which may not block, will be
371 * called after @ref is seen as dead from all CPUs at which point all
372 * further invocations of percpu_ref_tryget_live() will fail. See
373 * percpu_ref_tryget_live() for details.
374 *
375 * This function normally doesn't block and can be called from any context
376 * but it may block if @confirm_kill is specified and @ref is in the
377 * process of switching to atomic mode by percpu_ref_switch_to_atomic().
378 *
379 * There are no implied RCU grace periods between kill and release.
380 */
381void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
382 percpu_ref_func_t *confirm_kill)
383{
384 unsigned long flags;
385
386 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
387
388 WARN_ONCE(percpu_ref_is_dying(ref),
389 "%s called more than once on %ps!", __func__,
390 ref->data->release);
391
392 ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
393 __percpu_ref_switch_mode(ref, confirm_kill);
394 percpu_ref_put(ref);
395
396 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
397}
398EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
399
400/**
401 * percpu_ref_is_zero - test whether a percpu refcount reached zero
402 * @ref: percpu_ref to test
403 *
404 * Returns %true if @ref reached zero.
405 *
406 * This function is safe to call as long as @ref is between init and exit.
407 */
408bool percpu_ref_is_zero(struct percpu_ref *ref)
409{
410 unsigned long __percpu *percpu_count;
411 unsigned long count, flags;
412
413 if (__ref_is_percpu(ref, &percpu_count))
414 return false;
415
416 /* protect us from being destroyed */
417 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
418 if (ref->data)
419 count = atomic_long_read(&ref->data->count);
420 else
421 count = ref->percpu_count_ptr >> __PERCPU_REF_FLAG_BITS;
422 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
423
424 return count == 0;
425}
426EXPORT_SYMBOL_GPL(percpu_ref_is_zero);
427
428/**
429 * percpu_ref_reinit - re-initialize a percpu refcount
430 * @ref: perpcu_ref to re-initialize
431 *
432 * Re-initialize @ref so that it's in the same state as when it finished
433 * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD. @ref must have been
434 * initialized successfully and reached 0 but not exited.
435 *
436 * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
437 * this function is in progress.
438 */
439void percpu_ref_reinit(struct percpu_ref *ref)
440{
441 WARN_ON_ONCE(!percpu_ref_is_zero(ref));
442
443 percpu_ref_resurrect(ref);
444}
445EXPORT_SYMBOL_GPL(percpu_ref_reinit);
446
447/**
448 * percpu_ref_resurrect - modify a percpu refcount from dead to live
449 * @ref: perpcu_ref to resurrect
450 *
451 * Modify @ref so that it's in the same state as before percpu_ref_kill() was
452 * called. @ref must be dead but must not yet have exited.
453 *
454 * If @ref->release() frees @ref then the caller is responsible for
455 * guaranteeing that @ref->release() does not get called while this
456 * function is in progress.
457 *
458 * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
459 * this function is in progress.
460 */
461void percpu_ref_resurrect(struct percpu_ref *ref)
462{
463 unsigned long __percpu *percpu_count;
464 unsigned long flags;
465
466 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
467
468 WARN_ON_ONCE(!percpu_ref_is_dying(ref));
469 WARN_ON_ONCE(__ref_is_percpu(ref, &percpu_count));
470
471 ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD;
472 percpu_ref_get(ref);
473 __percpu_ref_switch_mode(ref, NULL);
474
475 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
476}
477EXPORT_SYMBOL_GPL(percpu_ref_resurrect);