1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Percpu refcounts:
4  * (C) 2012 Google, Inc.
5  * Author: Kent Overstreet <[email protected]>
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 	struct rcu_head		rcu;
106 };
107 
108 int __must_check percpu_ref_init(struct percpu_ref *ref,
109 				 percpu_ref_func_t *release, unsigned int flags,
110 				 gfp_t gfp);
111 void percpu_ref_exit(struct percpu_ref *ref);
112 void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
113 				 percpu_ref_func_t *confirm_switch);
114 void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref);
115 void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
116 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
117 				 percpu_ref_func_t *confirm_kill);
118 void percpu_ref_resurrect(struct percpu_ref *ref);
119 void percpu_ref_reinit(struct percpu_ref *ref);
120 
121 /**
122  * percpu_ref_kill - drop the initial ref
123  * @ref: percpu_ref to kill
124  *
125  * Must be used to drop the initial ref on a percpu refcount; must be called
126  * precisely once before shutdown.
127  *
128  * Switches @ref into atomic mode before gathering up the percpu counters
129  * and dropping the initial ref.
130  *
131  * There are no implied RCU grace periods between kill and release.
132  */
133 static inline void percpu_ref_kill(struct percpu_ref *ref)
134 {
135 	percpu_ref_kill_and_confirm(ref, NULL);
136 }
137 
138 /*
139  * Internal helper.  Don't use outside percpu-refcount proper.  The
140  * function doesn't return the pointer and let the caller test it for NULL
141  * because doing so forces the compiler to generate two conditional
142  * branches as it can't assume that @ref->percpu_count is not NULL.
143  */
144 static inline bool __ref_is_percpu(struct percpu_ref *ref,
145 					  unsigned long __percpu **percpu_countp)
146 {
147 	unsigned long percpu_ptr;
148 
149 	/*
150 	 * The value of @ref->percpu_count_ptr is tested for
151 	 * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then
152 	 * used as a pointer.  If the compiler generates a separate fetch
153 	 * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in
154 	 * between contaminating the pointer value, meaning that
155 	 * READ_ONCE() is required when fetching it.
156 	 *
157 	 * The smp_read_barrier_depends() implied by READ_ONCE() pairs
158 	 * with smp_store_release() in __percpu_ref_switch_to_percpu().
159 	 */
160 	percpu_ptr = READ_ONCE(ref->percpu_count_ptr);
161 
162 	/*
163 	 * Theoretically, the following could test just ATOMIC; however,
164 	 * then we'd have to mask off DEAD separately as DEAD may be
165 	 * visible without ATOMIC if we race with percpu_ref_kill().  DEAD
166 	 * implies ATOMIC anyway.  Test them together.
167 	 */
168 	if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD))
169 		return false;
170 
171 	*percpu_countp = (unsigned long __percpu *)percpu_ptr;
172 	return true;
173 }
174 
175 /**
176  * percpu_ref_get_many - increment a percpu refcount
177  * @ref: percpu_ref to get
178  * @nr: number of references to get
179  *
180  * Analogous to atomic_long_add().
181  *
182  * This function is safe to call as long as @ref is between init and exit.
183  */
184 static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
185 {
186 	unsigned long __percpu *percpu_count;
187 
188 	rcu_read_lock_sched();
189 
190 	if (__ref_is_percpu(ref, &percpu_count))
191 		this_cpu_add(*percpu_count, nr);
192 	else
193 		atomic_long_add(nr, &ref->count);
194 
195 	rcu_read_unlock_sched();
196 }
197 
198 /**
199  * percpu_ref_get - increment a percpu refcount
200  * @ref: percpu_ref to get
201  *
202  * Analagous to atomic_long_inc().
203  *
204  * This function is safe to call as long as @ref is between init and exit.
205  */
206 static inline void percpu_ref_get(struct percpu_ref *ref)
207 {
208 	percpu_ref_get_many(ref, 1);
209 }
210 
211 /**
212  * percpu_ref_tryget - try to increment a percpu refcount
213  * @ref: percpu_ref to try-get
214  *
215  * Increment a percpu refcount unless its count already reached zero.
216  * Returns %true on success; %false on failure.
217  *
218  * This function is safe to call as long as @ref is between init and exit.
219  */
220 static inline bool percpu_ref_tryget(struct percpu_ref *ref)
221 {
222 	unsigned long __percpu *percpu_count;
223 	bool ret;
224 
225 	rcu_read_lock_sched();
226 
227 	if (__ref_is_percpu(ref, &percpu_count)) {
228 		this_cpu_inc(*percpu_count);
229 		ret = true;
230 	} else {
231 		ret = atomic_long_inc_not_zero(&ref->count);
232 	}
233 
234 	rcu_read_unlock_sched();
235 
236 	return ret;
237 }
238 
239 /**
240  * percpu_ref_tryget_live - try to increment a live percpu refcount
241  * @ref: percpu_ref to try-get
242  *
243  * Increment a percpu refcount unless it has already been killed.  Returns
244  * %true on success; %false on failure.
245  *
246  * Completion of percpu_ref_kill() in itself doesn't guarantee that this
247  * function will fail.  For such guarantee, percpu_ref_kill_and_confirm()
248  * should be used.  After the confirm_kill callback is invoked, it's
249  * guaranteed that no new reference will be given out by
250  * percpu_ref_tryget_live().
251  *
252  * This function is safe to call as long as @ref is between init and exit.
253  */
254 static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
255 {
256 	unsigned long __percpu *percpu_count;
257 	bool ret = false;
258 
259 	rcu_read_lock_sched();
260 
261 	if (__ref_is_percpu(ref, &percpu_count)) {
262 		this_cpu_inc(*percpu_count);
263 		ret = true;
264 	} else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
265 		ret = atomic_long_inc_not_zero(&ref->count);
266 	}
267 
268 	rcu_read_unlock_sched();
269 
270 	return ret;
271 }
272 
273 /**
274  * percpu_ref_put_many - decrement a percpu refcount
275  * @ref: percpu_ref to put
276  * @nr: number of references to put
277  *
278  * Decrement the refcount, and if 0, call the release function (which was passed
279  * to percpu_ref_init())
280  *
281  * This function is safe to call as long as @ref is between init and exit.
282  */
283 static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
284 {
285 	unsigned long __percpu *percpu_count;
286 
287 	rcu_read_lock_sched();
288 
289 	if (__ref_is_percpu(ref, &percpu_count))
290 		this_cpu_sub(*percpu_count, nr);
291 	else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
292 		ref->release(ref);
293 
294 	rcu_read_unlock_sched();
295 }
296 
297 /**
298  * percpu_ref_put - decrement a percpu refcount
299  * @ref: percpu_ref to put
300  *
301  * Decrement the refcount, and if 0, call the release function (which was passed
302  * to percpu_ref_init())
303  *
304  * This function is safe to call as long as @ref is between init and exit.
305  */
306 static inline void percpu_ref_put(struct percpu_ref *ref)
307 {
308 	percpu_ref_put_many(ref, 1);
309 }
310 
311 /**
312  * percpu_ref_is_dying - test whether a percpu refcount is dying or dead
313  * @ref: percpu_ref to test
314  *
315  * Returns %true if @ref is dying or dead.
316  *
317  * This function is safe to call as long as @ref is between init and exit
318  * and the caller is responsible for synchronizing against state changes.
319  */
320 static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
321 {
322 	return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
323 }
324 
325 /**
326  * percpu_ref_is_zero - test whether a percpu refcount reached zero
327  * @ref: percpu_ref to test
328  *
329  * Returns %true if @ref reached zero.
330  *
331  * This function is safe to call as long as @ref is between init and exit.
332  */
333 static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
334 {
335 	unsigned long __percpu *percpu_count;
336 
337 	if (__ref_is_percpu(ref, &percpu_count))
338 		return false;
339 	return !atomic_long_read(&ref->count);
340 }
341 
342 #endif
343