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		ref->percpu_count_ptr = 0;
80		return -ENOMEM;
81	}
82
83	data->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
84	data->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT;
85
86	if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD)) {
87		ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
88		data->allow_reinit = true;
89	} else {
90		start_count += PERCPU_COUNT_BIAS;
91	}
92
93	if (flags & PERCPU_REF_INIT_DEAD)
94		ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
95	else
96		start_count++;
97
98	atomic_long_set(&data->count, start_count);
99
100	data->release = release;
101	data->confirm_switch = NULL;
102	data->ref = ref;
103	ref->data = data;
104	return 0;
105}
106EXPORT_SYMBOL_GPL(percpu_ref_init);
107
108static void __percpu_ref_exit(struct percpu_ref *ref)
109{
110	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
111
112	if (percpu_count) {
113		/* non-NULL confirm_switch indicates switching in progress */
114		WARN_ON_ONCE(ref->data && ref->data->confirm_switch);
115		free_percpu(percpu_count);
116		ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
117	}
118}
119
120/**
121 * percpu_ref_exit - undo percpu_ref_init()
122 * @ref: percpu_ref to exit
123 *
124 * This function exits @ref.  The caller is responsible for ensuring that
125 * @ref is no longer in active use.  The usual places to invoke this
126 * function from are the @ref->release() callback or in init failure path
127 * where percpu_ref_init() succeeded but other parts of the initialization
128 * of the embedding object failed.
129 */
130void percpu_ref_exit(struct percpu_ref *ref)
131{
132	struct percpu_ref_data *data = ref->data;
133	unsigned long flags;
134
135	__percpu_ref_exit(ref);
136
137	if (!data)
138		return;
139
140	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
141	ref->percpu_count_ptr |= atomic_long_read(&ref->data->count) <<
142		__PERCPU_REF_FLAG_BITS;
143	ref->data = NULL;
144	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
145
146	kfree(data);
147}
148EXPORT_SYMBOL_GPL(percpu_ref_exit);
149
150static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
151{
152	struct percpu_ref_data *data = container_of(rcu,
153			struct percpu_ref_data, rcu);
154	struct percpu_ref *ref = data->ref;
155
156	data->confirm_switch(ref);
157	data->confirm_switch = NULL;
158	wake_up_all(&percpu_ref_switch_waitq);
159
160	if (!data->allow_reinit)
161		__percpu_ref_exit(ref);
162
163	/* drop ref from percpu_ref_switch_to_atomic() */
164	percpu_ref_put(ref);
165}
166
167static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
168{
169	struct percpu_ref_data *data = container_of(rcu,
170			struct percpu_ref_data, rcu);
171	struct percpu_ref *ref = data->ref;
172	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
173	static atomic_t underflows;
174	unsigned long count = 0;
175	int cpu;
176
177	for_each_possible_cpu(cpu)
178		count += *per_cpu_ptr(percpu_count, cpu);
179
180	pr_debug("global %lu percpu %lu\n",
181		 atomic_long_read(&data->count), count);
182
183	/*
184	 * It's crucial that we sum the percpu counters _before_ adding the sum
185	 * to &ref->count; since gets could be happening on one cpu while puts
186	 * happen on another, adding a single cpu's count could cause
187	 * @ref->count to hit 0 before we've got a consistent value - but the
188	 * sum of all the counts will be consistent and correct.
189	 *
190	 * Subtracting the bias value then has to happen _after_ adding count to
191	 * &ref->count; we need the bias value to prevent &ref->count from
192	 * reaching 0 before we add the percpu counts. But doing it at the same
193	 * time is equivalent and saves us atomic operations:
194	 */
195	atomic_long_add((long)count - PERCPU_COUNT_BIAS, &data->count);
196
197	if (WARN_ONCE(atomic_long_read(&data->count) <= 0,
198		      "percpu ref (%ps) <= 0 (%ld) after switching to atomic",
199		      data->release, atomic_long_read(&data->count)) &&
200	    atomic_inc_return(&underflows) < 4) {
201		pr_err("%s(): percpu_ref underflow", __func__);
202		mem_dump_obj(data);
203	}
204
205	/* @ref is viewed as dead on all CPUs, send out switch confirmation */
206	percpu_ref_call_confirm_rcu(rcu);
207}
208
209static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref)
210{
211}
212
213static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
214					  percpu_ref_func_t *confirm_switch)
215{
216	if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) {
217		if (confirm_switch)
218			confirm_switch(ref);
219		return;
220	}
221
222	/* switching from percpu to atomic */
223	ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
224
225	/*
226	 * Non-NULL ->confirm_switch is used to indicate that switching is
227	 * in progress.  Use noop one if unspecified.
228	 */
229	ref->data->confirm_switch = confirm_switch ?:
230		percpu_ref_noop_confirm_switch;
231
232	percpu_ref_get(ref);	/* put after confirmation */
233	call_rcu_hurry(&ref->data->rcu,
234		       percpu_ref_switch_to_atomic_rcu);
235}
236
237static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
238{
239	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
240	int cpu;
241
242	BUG_ON(!percpu_count);
243
244	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
245		return;
246
247	if (WARN_ON_ONCE(!ref->data->allow_reinit))
248		return;
249
250	atomic_long_add(PERCPU_COUNT_BIAS, &ref->data->count);
251
252	/*
253	 * Restore per-cpu operation.  smp_store_release() is paired
254	 * with READ_ONCE() in __ref_is_percpu() and guarantees that the
255	 * zeroing is visible to all percpu accesses which can see the
256	 * following __PERCPU_REF_ATOMIC clearing.
257	 */
258	for_each_possible_cpu(cpu)
259		*per_cpu_ptr(percpu_count, cpu) = 0;
260
261	smp_store_release(&ref->percpu_count_ptr,
262			  ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);
263}
264
265static void __percpu_ref_switch_mode(struct percpu_ref *ref,
266				     percpu_ref_func_t *confirm_switch)
267{
268	struct percpu_ref_data *data = ref->data;
269
270	lockdep_assert_held(&percpu_ref_switch_lock);
271
272	/*
273	 * If the previous ATOMIC switching hasn't finished yet, wait for
274	 * its completion.  If the caller ensures that ATOMIC switching
275	 * isn't in progress, this function can be called from any context.
276	 */
277	wait_event_lock_irq(percpu_ref_switch_waitq, !data->confirm_switch,
278			    percpu_ref_switch_lock);
279
280	if (data->force_atomic || percpu_ref_is_dying(ref))
281		__percpu_ref_switch_to_atomic(ref, confirm_switch);
282	else
283		__percpu_ref_switch_to_percpu(ref);
284}
285
286/**
287 * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode
288 * @ref: percpu_ref to switch to atomic mode
289 * @confirm_switch: optional confirmation callback
290 *
291 * There's no reason to use this function for the usual reference counting.
292 * Use percpu_ref_kill[_and_confirm]().
293 *
294 * Schedule switching of @ref to atomic mode.  All its percpu counts will
295 * be collected to the main atomic counter.  On completion, when all CPUs
296 * are guaraneed to be in atomic mode, @confirm_switch, which may not
297 * block, is invoked.  This function may be invoked concurrently with all
298 * the get/put operations and can safely be mixed with kill and reinit
299 * operations.  Note that @ref will stay in atomic mode across kill/reinit
300 * cycles until percpu_ref_switch_to_percpu() is called.
301 *
302 * This function may block if @ref is in the process of switching to atomic
303 * mode.  If the caller ensures that @ref is not in the process of
304 * switching to atomic mode, this function can be called from any context.
305 */
306void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
307				 percpu_ref_func_t *confirm_switch)
308{
309	unsigned long flags;
310
311	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
312
313	ref->data->force_atomic = true;
314	__percpu_ref_switch_mode(ref, confirm_switch);
315
316	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
317}
318EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
319
320/**
321 * percpu_ref_switch_to_atomic_sync - switch a percpu_ref to atomic mode
322 * @ref: percpu_ref to switch to atomic mode
323 *
324 * Schedule switching the ref to atomic mode, and wait for the
325 * switch to complete.  Caller must ensure that no other thread
326 * will switch back to percpu mode.
327 */
328void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
329{
330	percpu_ref_switch_to_atomic(ref, NULL);
331	wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
332}
333EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);
334
335/**
336 * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode
337 * @ref: percpu_ref to switch to percpu mode
338 *
339 * There's no reason to use this function for the usual reference counting.
340 * To re-use an expired ref, use percpu_ref_reinit().
341 *
342 * Switch @ref to percpu mode.  This function may be invoked concurrently
343 * with all the get/put operations and can safely be mixed with kill and
344 * reinit operations.  This function reverses the sticky atomic state set
345 * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic().  If @ref is
346 * dying or dead, the actual switching takes place on the following
347 * percpu_ref_reinit().
348 *
349 * This function may block if @ref is in the process of switching to atomic
350 * mode.  If the caller ensures that @ref is not in the process of
351 * switching to atomic mode, this function can be called from any context.
352 */
353void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
354{
355	unsigned long flags;
356
357	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
358
359	ref->data->force_atomic = false;
360	__percpu_ref_switch_mode(ref, NULL);
361
362	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
363}
364EXPORT_SYMBOL_GPL(percpu_ref_switch_to_percpu);
365
366/**
367 * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
368 * @ref: percpu_ref to kill
369 * @confirm_kill: optional confirmation callback
370 *
371 * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
372 * @confirm_kill is not NULL.  @confirm_kill, which may not block, will be
373 * called after @ref is seen as dead from all CPUs at which point all
374 * further invocations of percpu_ref_tryget_live() will fail.  See
375 * percpu_ref_tryget_live() for details.
376 *
377 * This function normally doesn't block and can be called from any context
378 * but it may block if @confirm_kill is specified and @ref is in the
379 * process of switching to atomic mode by percpu_ref_switch_to_atomic().
380 *
381 * There are no implied RCU grace periods between kill and release.
382 */
383void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
384				 percpu_ref_func_t *confirm_kill)
385{
386	unsigned long flags;
387
388	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
389
390	WARN_ONCE(percpu_ref_is_dying(ref),
391		  "%s called more than once on %ps!", __func__,
392		  ref->data->release);
393
394	ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
395	__percpu_ref_switch_mode(ref, confirm_kill);
396	percpu_ref_put(ref);
397
398	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
399}
400EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
401
402/**
403 * percpu_ref_is_zero - test whether a percpu refcount reached zero
404 * @ref: percpu_ref to test
405 *
406 * Returns %true if @ref reached zero.
407 *
408 * This function is safe to call as long as @ref is between init and exit.
409 */
410bool percpu_ref_is_zero(struct percpu_ref *ref)
411{
412	unsigned long __percpu *percpu_count;
413	unsigned long count, flags;
414
415	if (__ref_is_percpu(ref, &percpu_count))
416		return false;
417
418	/* protect us from being destroyed */
419	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
420	if (ref->data)
421		count = atomic_long_read(&ref->data->count);
422	else
423		count = ref->percpu_count_ptr >> __PERCPU_REF_FLAG_BITS;
424	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
425
426	return count == 0;
427}
428EXPORT_SYMBOL_GPL(percpu_ref_is_zero);
429
430/**
431 * percpu_ref_reinit - re-initialize a percpu refcount
432 * @ref: perpcu_ref to re-initialize
433 *
434 * Re-initialize @ref so that it's in the same state as when it finished
435 * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD.  @ref must have been
436 * initialized successfully and reached 0 but not exited.
437 *
438 * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
439 * this function is in progress.
440 */
441void percpu_ref_reinit(struct percpu_ref *ref)
442{
443	WARN_ON_ONCE(!percpu_ref_is_zero(ref));
444
445	percpu_ref_resurrect(ref);
446}
447EXPORT_SYMBOL_GPL(percpu_ref_reinit);
448
449/**
450 * percpu_ref_resurrect - modify a percpu refcount from dead to live
451 * @ref: perpcu_ref to resurrect
452 *
453 * Modify @ref so that it's in the same state as before percpu_ref_kill() was
454 * called. @ref must be dead but must not yet have exited.
455 *
456 * If @ref->release() frees @ref then the caller is responsible for
457 * guaranteeing that @ref->release() does not get called while this
458 * function is in progress.
459 *
460 * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
461 * this function is in progress.
462 */
463void percpu_ref_resurrect(struct percpu_ref *ref)
464{
465	unsigned long __percpu *percpu_count;
466	unsigned long flags;
467
468	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
469
470	WARN_ON_ONCE(!percpu_ref_is_dying(ref));
471	WARN_ON_ONCE(__ref_is_percpu(ref, &percpu_count));
472
473	ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD;
474	percpu_ref_get(ref);
475	__percpu_ref_switch_mode(ref, NULL);
476
477	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
478}
479EXPORT_SYMBOL_GPL(percpu_ref_resurrect);
480