1/* SPDX-License-Identifier: GPL-2.0 */
  2/*
  3 * Percpu refcounts:
  4 * (C) 2012 Google, Inc.
  5 * Author: Kent Overstreet <koverstreet@google.com>
  6 *
  7 * This implements a refcount with similar semantics to atomic_t - atomic_inc(),
  8 * atomic_dec_and_test() - but percpu.
  9 *
 10 * There's one important difference between percpu refs and normal atomic_t
 11 * refcounts; you have to keep track of your initial refcount, and then when you
 12 * start shutting down you call percpu_ref_kill() _before_ dropping the initial
 13 * refcount.
 14 *
 15 * The refcount will have a range of 0 to ((1U << 31) - 1), i.e. one bit less
 16 * than an atomic_t - this is because of the way shutdown works, see
 17 * percpu_ref_kill()/PERCPU_COUNT_BIAS.
 18 *
 19 * Before you call percpu_ref_kill(), percpu_ref_put() does not check for the
 20 * refcount hitting 0 - it can't, if it was in percpu mode. percpu_ref_kill()
 21 * puts the ref back in single atomic_t mode, collecting the per cpu refs and
 22 * issuing the appropriate barriers, and then marks the ref as shutting down so
 23 * that percpu_ref_put() will check for the ref hitting 0.  After it returns,
 24 * it's safe to drop the initial ref.
 25 *
 26 * USAGE:
 27 *
 28 * See fs/aio.c for some example usage; it's used there for struct kioctx, which
 29 * is created when userspaces calls io_setup(), and destroyed when userspace
 30 * calls io_destroy() or the process exits.
 31 *
 32 * In the aio code, kill_ioctx() is called when we wish to destroy a kioctx; it
 33 * removes the kioctx from the proccess's table of kioctxs and kills percpu_ref.
 34 * After that, there can't be any new users of the kioctx (from lookup_ioctx())
 35 * and it's then safe to drop the initial ref with percpu_ref_put().
 36 *
 37 * Note that the free path, free_ioctx(), needs to go through explicit call_rcu()
 38 * to synchronize with RCU protected lookup_ioctx().  percpu_ref operations don't
 39 * imply RCU grace periods of any kind and if a user wants to combine percpu_ref
 40 * with RCU protection, it must be done explicitly.
 41 *
 42 * Code that does a two stage shutdown like this often needs some kind of
 43 * explicit synchronization to ensure the initial refcount can only be dropped
 44 * once - percpu_ref_kill() does this for you, it returns true once and false if
 45 * someone else already called it. The aio code uses it this way, but it's not
 46 * necessary if the code has some other mechanism to synchronize teardown.
 47 * around.
 48 */
 49
 50#ifndef _LINUX_PERCPU_REFCOUNT_H
 51#define _LINUX_PERCPU_REFCOUNT_H
 52
 53#include <linux/atomic.h>
 54#include <linux/kernel.h>
 55#include <linux/percpu.h>
 56#include <linux/rcupdate.h>
 57#include <linux/gfp.h>
 58
 59struct percpu_ref;
 60typedef void (percpu_ref_func_t)(struct percpu_ref *);
 61
 62/* flags set in the lower bits of percpu_ref->percpu_count_ptr */
 63enum {
 64	__PERCPU_REF_ATOMIC	= 1LU << 0,	/* operating in atomic mode */
 65	__PERCPU_REF_DEAD	= 1LU << 1,	/* (being) killed */
 66	__PERCPU_REF_ATOMIC_DEAD = __PERCPU_REF_ATOMIC | __PERCPU_REF_DEAD,
 67
 68	__PERCPU_REF_FLAG_BITS	= 2,
 69};
 70
 71/* @flags for percpu_ref_init() */
 72enum {
 73	/*
 74	 * Start w/ ref == 1 in atomic mode.  Can be switched to percpu
 75	 * operation using percpu_ref_switch_to_percpu().  If initialized
 76	 * with this flag, the ref will stay in atomic mode until
 77	 * percpu_ref_switch_to_percpu() is invoked on it.
 78	 * Implies ALLOW_REINIT.
 79	 */
 80	PERCPU_REF_INIT_ATOMIC	= 1 << 0,
 81
 82	/*
 83	 * Start dead w/ ref == 0 in atomic mode.  Must be revived with
 84	 * percpu_ref_reinit() before used.  Implies INIT_ATOMIC and
 85	 * ALLOW_REINIT.
 86	 */
 87	PERCPU_REF_INIT_DEAD	= 1 << 1,
 88
 89	/*
 90	 * Allow switching from atomic mode to percpu mode.
 91	 */
 92	PERCPU_REF_ALLOW_REINIT	= 1 << 2,
 93};
 94
 95struct percpu_ref {
 96	atomic_long_t		count;
 97	/*
 98	 * The low bit of the pointer indicates whether the ref is in percpu
 99	 * mode; if set, then get/put will manipulate the atomic_t.
100	 */
101	unsigned long		percpu_count_ptr;
102	percpu_ref_func_t	*release;
103	percpu_ref_func_t	*confirm_switch;
104	bool			force_atomic:1;
105	bool			allow_reinit:1;
106	struct rcu_head		rcu;
107};
108
109int __must_check percpu_ref_init(struct percpu_ref *ref,
110				 percpu_ref_func_t *release, unsigned int flags,
111				 gfp_t gfp);
112void percpu_ref_exit(struct percpu_ref *ref);
113void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
114				 percpu_ref_func_t *confirm_switch);
115void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref);
116void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
117void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
118				 percpu_ref_func_t *confirm_kill);
119void percpu_ref_resurrect(struct percpu_ref *ref);
120void percpu_ref_reinit(struct percpu_ref *ref);
121
122/**
123 * percpu_ref_kill - drop the initial ref
124 * @ref: percpu_ref to kill
125 *
126 * Must be used to drop the initial ref on a percpu refcount; must be called
127 * precisely once before shutdown.
128 *
129 * Switches @ref into atomic mode before gathering up the percpu counters
130 * and dropping the initial ref.
131 *
132 * There are no implied RCU grace periods between kill and release.
133 */
134static inline void percpu_ref_kill(struct percpu_ref *ref)
135{
136	percpu_ref_kill_and_confirm(ref, NULL);
137}
138
139/*
140 * Internal helper.  Don't use outside percpu-refcount proper.  The
141 * function doesn't return the pointer and let the caller test it for NULL
142 * because doing so forces the compiler to generate two conditional
143 * branches as it can't assume that @ref->percpu_count is not NULL.
144 */
145static inline bool __ref_is_percpu(struct percpu_ref *ref,
146					  unsigned long __percpu **percpu_countp)
147{
148	unsigned long percpu_ptr;
149
150	/*
151	 * The value of @ref->percpu_count_ptr is tested for
152	 * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then
153	 * used as a pointer.  If the compiler generates a separate fetch
154	 * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in
155	 * between contaminating the pointer value, meaning that
156	 * READ_ONCE() is required when fetching it.
157	 *
158	 * The dependency ordering from the READ_ONCE() pairs
159	 * with smp_store_release() in __percpu_ref_switch_to_percpu().
160	 */
161	percpu_ptr = READ_ONCE(ref->percpu_count_ptr);
162
163	/*
164	 * Theoretically, the following could test just ATOMIC; however,
165	 * then we'd have to mask off DEAD separately as DEAD may be
166	 * visible without ATOMIC if we race with percpu_ref_kill().  DEAD
167	 * implies ATOMIC anyway.  Test them together.
168	 */
169	if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD))
170		return false;
171
172	*percpu_countp = (unsigned long __percpu *)percpu_ptr;
173	return true;
174}
175
176/**
177 * percpu_ref_get_many - increment a percpu refcount
178 * @ref: percpu_ref to get
179 * @nr: number of references to get
180 *
181 * Analogous to atomic_long_add().
182 *
183 * This function is safe to call as long as @ref is between init and exit.
184 */
185static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
186{
187	unsigned long __percpu *percpu_count;
188
189	rcu_read_lock();
190
191	if (__ref_is_percpu(ref, &percpu_count))
192		this_cpu_add(*percpu_count, nr);
193	else
194		atomic_long_add(nr, &ref->count);
195
196	rcu_read_unlock();
197}
198
199/**
200 * percpu_ref_get - increment a percpu refcount
201 * @ref: percpu_ref to get
202 *
203 * Analagous to atomic_long_inc().
204 *
205 * This function is safe to call as long as @ref is between init and exit.
206 */
207static inline void percpu_ref_get(struct percpu_ref *ref)
208{
209	percpu_ref_get_many(ref, 1);
210}
211
212/**
213 * percpu_ref_tryget_many - try to increment a percpu refcount
214 * @ref: percpu_ref to try-get
215 * @nr: number of references to get
216 *
217 * Increment a percpu refcount  by @nr unless its count already reached zero.
218 * Returns %true on success; %false on failure.
219 *
220 * This function is safe to call as long as @ref is between init and exit.
221 */
222static inline bool percpu_ref_tryget_many(struct percpu_ref *ref,
223					  unsigned long nr)
224{
225	unsigned long __percpu *percpu_count;
226	bool ret;
227
228	rcu_read_lock();
229
230	if (__ref_is_percpu(ref, &percpu_count)) {
231		this_cpu_add(*percpu_count, nr);
232		ret = true;
233	} else {
234		ret = atomic_long_add_unless(&ref->count, nr, 0);
235	}
236
237	rcu_read_unlock();
238
239	return ret;
240}
241
242/**
243 * percpu_ref_tryget - try to increment a percpu refcount
244 * @ref: percpu_ref to try-get
245 *
246 * Increment a percpu refcount unless its count already reached zero.
247 * Returns %true on success; %false on failure.
248 *
249 * This function is safe to call as long as @ref is between init and exit.
250 */
251static inline bool percpu_ref_tryget(struct percpu_ref *ref)
252{
253	return percpu_ref_tryget_many(ref, 1);
254}
255
256/**
257 * percpu_ref_tryget_live - try to increment a live percpu refcount
258 * @ref: percpu_ref to try-get
259 *
260 * Increment a percpu refcount unless it has already been killed.  Returns
261 * %true on success; %false on failure.
262 *
263 * Completion of percpu_ref_kill() in itself doesn't guarantee that this
264 * function will fail.  For such guarantee, percpu_ref_kill_and_confirm()
265 * should be used.  After the confirm_kill callback is invoked, it's
266 * guaranteed that no new reference will be given out by
267 * percpu_ref_tryget_live().
268 *
269 * This function is safe to call as long as @ref is between init and exit.
270 */
271static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
272{
273	unsigned long __percpu *percpu_count;
274	bool ret = false;
275
276	rcu_read_lock();
277
278	if (__ref_is_percpu(ref, &percpu_count)) {
279		this_cpu_inc(*percpu_count);
280		ret = true;
281	} else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
282		ret = atomic_long_inc_not_zero(&ref->count);
283	}
284
285	rcu_read_unlock();
286
287	return ret;
288}
289
290/**
291 * percpu_ref_put_many - decrement a percpu refcount
292 * @ref: percpu_ref to put
293 * @nr: number of references to put
294 *
295 * Decrement the refcount, and if 0, call the release function (which was passed
296 * to percpu_ref_init())
297 *
298 * This function is safe to call as long as @ref is between init and exit.
299 */
300static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
301{
302	unsigned long __percpu *percpu_count;
303
304	rcu_read_lock();
305
306	if (__ref_is_percpu(ref, &percpu_count))
307		this_cpu_sub(*percpu_count, nr);
308	else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
309		ref->release(ref);
310
311	rcu_read_unlock();
312}
313
314/**
315 * percpu_ref_put - decrement a percpu refcount
316 * @ref: percpu_ref to put
317 *
318 * Decrement the refcount, and if 0, call the release function (which was passed
319 * to percpu_ref_init())
320 *
321 * This function is safe to call as long as @ref is between init and exit.
322 */
323static inline void percpu_ref_put(struct percpu_ref *ref)
324{
325	percpu_ref_put_many(ref, 1);
326}
327
328/**
329 * percpu_ref_is_dying - test whether a percpu refcount is dying or dead
330 * @ref: percpu_ref to test
331 *
332 * Returns %true if @ref is dying or dead.
333 *
334 * This function is safe to call as long as @ref is between init and exit
335 * and the caller is responsible for synchronizing against state changes.
336 */
337static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
338{
339	return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
340}
341
342/**
343 * percpu_ref_is_zero - test whether a percpu refcount reached zero
344 * @ref: percpu_ref to test
345 *
346 * Returns %true if @ref reached zero.
347 *
348 * This function is safe to call as long as @ref is between init and exit.
349 */
350static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
351{
352	unsigned long __percpu *percpu_count;
353
354	if (__ref_is_percpu(ref, &percpu_count))
355		return false;
356	return !atomic_long_read(&ref->count);
357}
358
359#endif