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 
59 struct percpu_ref;
60 typedef void (percpu_ref_func_t)(struct percpu_ref *);
61 
62 /* flags set in the lower bits of percpu_ref->percpu_count_ptr */
63 enum {
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() */
72 enum {
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 
95 struct 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 
109 int __must_check percpu_ref_init(struct percpu_ref *ref,
110 				 percpu_ref_func_t *release, unsigned int flags,
111 				 gfp_t gfp);
112 void percpu_ref_exit(struct percpu_ref *ref);
113 void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
114 				 percpu_ref_func_t *confirm_switch);
115 void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref);
116 void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
117 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
118 				 percpu_ref_func_t *confirm_kill);
119 void percpu_ref_resurrect(struct percpu_ref *ref);
120 void 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  */
134 static 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  */
145 static 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 smp_read_barrier_depends() implied by 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  */
185 static 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_sched();
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_sched();
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  */
207 static inline void percpu_ref_get(struct percpu_ref *ref)
208 {
209 	percpu_ref_get_many(ref, 1);
210 }
211 
212 /**
213  * percpu_ref_tryget - try to increment a percpu refcount
214  * @ref: percpu_ref to try-get
215  *
216  * Increment a percpu refcount unless its count already reached zero.
217  * Returns %true on success; %false on failure.
218  *
219  * This function is safe to call as long as @ref is between init and exit.
220  */
221 static inline bool percpu_ref_tryget(struct percpu_ref *ref)
222 {
223 	unsigned long __percpu *percpu_count;
224 	bool ret;
225 
226 	rcu_read_lock_sched();
227 
228 	if (__ref_is_percpu(ref, &percpu_count)) {
229 		this_cpu_inc(*percpu_count);
230 		ret = true;
231 	} else {
232 		ret = atomic_long_inc_not_zero(&ref->count);
233 	}
234 
235 	rcu_read_unlock_sched();
236 
237 	return ret;
238 }
239 
240 /**
241  * percpu_ref_tryget_live - try to increment a live percpu refcount
242  * @ref: percpu_ref to try-get
243  *
244  * Increment a percpu refcount unless it has already been killed.  Returns
245  * %true on success; %false on failure.
246  *
247  * Completion of percpu_ref_kill() in itself doesn't guarantee that this
248  * function will fail.  For such guarantee, percpu_ref_kill_and_confirm()
249  * should be used.  After the confirm_kill callback is invoked, it's
250  * guaranteed that no new reference will be given out by
251  * percpu_ref_tryget_live().
252  *
253  * This function is safe to call as long as @ref is between init and exit.
254  */
255 static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
256 {
257 	unsigned long __percpu *percpu_count;
258 	bool ret = false;
259 
260 	rcu_read_lock_sched();
261 
262 	if (__ref_is_percpu(ref, &percpu_count)) {
263 		this_cpu_inc(*percpu_count);
264 		ret = true;
265 	} else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
266 		ret = atomic_long_inc_not_zero(&ref->count);
267 	}
268 
269 	rcu_read_unlock_sched();
270 
271 	return ret;
272 }
273 
274 /**
275  * percpu_ref_put_many - decrement a percpu refcount
276  * @ref: percpu_ref to put
277  * @nr: number of references to put
278  *
279  * Decrement the refcount, and if 0, call the release function (which was passed
280  * to percpu_ref_init())
281  *
282  * This function is safe to call as long as @ref is between init and exit.
283  */
284 static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
285 {
286 	unsigned long __percpu *percpu_count;
287 
288 	rcu_read_lock_sched();
289 
290 	if (__ref_is_percpu(ref, &percpu_count))
291 		this_cpu_sub(*percpu_count, nr);
292 	else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
293 		ref->release(ref);
294 
295 	rcu_read_unlock_sched();
296 }
297 
298 /**
299  * percpu_ref_put - decrement a percpu refcount
300  * @ref: percpu_ref to put
301  *
302  * Decrement the refcount, and if 0, call the release function (which was passed
303  * to percpu_ref_init())
304  *
305  * This function is safe to call as long as @ref is between init and exit.
306  */
307 static inline void percpu_ref_put(struct percpu_ref *ref)
308 {
309 	percpu_ref_put_many(ref, 1);
310 }
311 
312 /**
313  * percpu_ref_is_dying - test whether a percpu refcount is dying or dead
314  * @ref: percpu_ref to test
315  *
316  * Returns %true if @ref is dying or dead.
317  *
318  * This function is safe to call as long as @ref is between init and exit
319  * and the caller is responsible for synchronizing against state changes.
320  */
321 static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
322 {
323 	return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
324 }
325 
326 /**
327  * percpu_ref_is_zero - test whether a percpu refcount reached zero
328  * @ref: percpu_ref to test
329  *
330  * Returns %true if @ref reached zero.
331  *
332  * This function is safe to call as long as @ref is between init and exit.
333  */
334 static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
335 {
336 	unsigned long __percpu *percpu_count;
337 
338 	if (__ref_is_percpu(ref, &percpu_count))
339 		return false;
340 	return !atomic_long_read(&ref->count);
341 }
342 
343 #endif
344