blob: 84f542df7ff5cca0932f02b6629f8229244c2d1a [file] [log] [blame]
Kent Overstreet215e2622013-05-31 15:26:45 -07001/*
2 * Percpu refcounts:
3 * (C) 2012 Google, Inc.
4 * Author: Kent Overstreet <koverstreet@google.com>
5 *
6 * This implements a refcount with similar semantics to atomic_t - atomic_inc(),
7 * atomic_dec_and_test() - but percpu.
8 *
9 * There's one important difference between percpu refs and normal atomic_t
10 * refcounts; you have to keep track of your initial refcount, and then when you
11 * start shutting down you call percpu_ref_kill() _before_ dropping the initial
12 * refcount.
13 *
14 * The refcount will have a range of 0 to ((1U << 31) - 1), i.e. one bit less
15 * than an atomic_t - this is because of the way shutdown works, see
Tejun Heoeecc16b2014-09-24 13:31:48 -040016 * percpu_ref_kill()/PERCPU_COUNT_BIAS.
Kent Overstreet215e2622013-05-31 15:26:45 -070017 *
18 * Before you call percpu_ref_kill(), percpu_ref_put() does not check for the
19 * refcount hitting 0 - it can't, if it was in percpu mode. percpu_ref_kill()
20 * puts the ref back in single atomic_t mode, collecting the per cpu refs and
21 * issuing the appropriate barriers, and then marks the ref as shutting down so
22 * that percpu_ref_put() will check for the ref hitting 0. After it returns,
23 * it's safe to drop the initial ref.
24 *
25 * USAGE:
26 *
27 * See fs/aio.c for some example usage; it's used there for struct kioctx, which
28 * is created when userspaces calls io_setup(), and destroyed when userspace
29 * calls io_destroy() or the process exits.
30 *
31 * In the aio code, kill_ioctx() is called when we wish to destroy a kioctx; it
Jesper Dangaard Brouerd7cdb962014-06-20 17:19:06 +020032 * calls percpu_ref_kill(), then hlist_del_rcu() and synchronize_rcu() to remove
Kent Overstreet215e2622013-05-31 15:26:45 -070033 * the kioctx from the proccess's list of kioctxs - after that, there can't be
34 * any new users of the kioctx (from lookup_ioctx()) and it's then safe to drop
35 * the initial ref with percpu_ref_put().
36 *
37 * Code that does a two stage shutdown like this often needs some kind of
38 * explicit synchronization to ensure the initial refcount can only be dropped
39 * once - percpu_ref_kill() does this for you, it returns true once and false if
40 * someone else already called it. The aio code uses it this way, but it's not
41 * necessary if the code has some other mechanism to synchronize teardown.
42 * around.
43 */
44
45#ifndef _LINUX_PERCPU_REFCOUNT_H
46#define _LINUX_PERCPU_REFCOUNT_H
47
48#include <linux/atomic.h>
49#include <linux/kernel.h>
50#include <linux/percpu.h>
51#include <linux/rcupdate.h>
Tejun Heoa34375e2014-09-08 09:51:30 +090052#include <linux/gfp.h>
Kent Overstreet215e2622013-05-31 15:26:45 -070053
54struct percpu_ref;
Tejun Heoac899062013-06-12 20:43:06 -070055typedef void (percpu_ref_func_t)(struct percpu_ref *);
Kent Overstreet215e2622013-05-31 15:26:45 -070056
Tejun Heo9e804d12014-09-24 13:31:48 -040057/* flags set in the lower bits of percpu_ref->percpu_count_ptr */
58enum {
59 __PERCPU_REF_ATOMIC = 1LU << 0, /* operating in atomic mode */
Tejun Heo27344a92014-09-24 13:31:49 -040060 __PERCPU_REF_DEAD = 1LU << 1, /* (being) killed */
61 __PERCPU_REF_ATOMIC_DEAD = __PERCPU_REF_ATOMIC | __PERCPU_REF_DEAD,
62
63 __PERCPU_REF_FLAG_BITS = 2,
Tejun Heo9e804d12014-09-24 13:31:48 -040064};
65
Tejun Heo2aad2a82014-09-24 13:31:50 -040066/* @flags for percpu_ref_init() */
67enum {
68 /*
69 * Start w/ ref == 1 in atomic mode. Can be switched to percpu
Tejun Heo1cae13e2014-09-24 13:31:50 -040070 * operation using percpu_ref_switch_to_percpu(). If initialized
71 * with this flag, the ref will stay in atomic mode until
72 * percpu_ref_switch_to_percpu() is invoked on it.
Tejun Heo2aad2a82014-09-24 13:31:50 -040073 */
74 PERCPU_REF_INIT_ATOMIC = 1 << 0,
75
76 /*
77 * Start dead w/ ref == 0 in atomic mode. Must be revived with
78 * percpu_ref_reinit() before used. Implies INIT_ATOMIC.
79 */
80 PERCPU_REF_INIT_DEAD = 1 << 1,
81};
82
Kent Overstreet215e2622013-05-31 15:26:45 -070083struct percpu_ref {
Tejun Heoe6253052014-09-20 01:27:25 -040084 atomic_long_t count;
Kent Overstreet215e2622013-05-31 15:26:45 -070085 /*
86 * The low bit of the pointer indicates whether the ref is in percpu
Tejun Heo9a1049d2014-06-28 08:10:14 -040087 * mode; if set, then get/put will manipulate the atomic_t.
Kent Overstreet215e2622013-05-31 15:26:45 -070088 */
Tejun Heoeecc16b2014-09-24 13:31:48 -040089 unsigned long percpu_count_ptr;
Tejun Heoac899062013-06-12 20:43:06 -070090 percpu_ref_func_t *release;
Tejun Heo9e804d12014-09-24 13:31:48 -040091 percpu_ref_func_t *confirm_switch;
Tejun Heo1cae13e2014-09-24 13:31:50 -040092 bool force_atomic:1;
Kent Overstreet215e2622013-05-31 15:26:45 -070093 struct rcu_head rcu;
94};
95
Tejun Heoacac7882013-06-12 20:52:01 -070096int __must_check percpu_ref_init(struct percpu_ref *ref,
Tejun Heo2aad2a82014-09-24 13:31:50 -040097 percpu_ref_func_t *release, unsigned int flags,
98 gfp_t gfp);
Tejun Heo9a1049d2014-06-28 08:10:14 -040099void percpu_ref_exit(struct percpu_ref *ref);
Tejun Heo490c79a2014-09-24 13:31:49 -0400100void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
101 percpu_ref_func_t *confirm_switch);
Tejun Heof47ad452014-09-24 13:31:49 -0400102void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
Tejun Heodbece3a2013-06-13 19:23:53 -0700103void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
104 percpu_ref_func_t *confirm_kill);
Tejun Heof47ad452014-09-24 13:31:49 -0400105void percpu_ref_reinit(struct percpu_ref *ref);
Tejun Heodbece3a2013-06-13 19:23:53 -0700106
107/**
108 * percpu_ref_kill - drop the initial ref
109 * @ref: percpu_ref to kill
110 *
111 * Must be used to drop the initial ref on a percpu refcount; must be called
112 * precisely once before shutdown.
113 *
114 * Puts @ref in non percpu mode, then does a call_rcu() before gathering up the
115 * percpu counters and dropping the initial ref.
116 */
117static inline void percpu_ref_kill(struct percpu_ref *ref)
118{
Guillaume Gomez4d414262015-09-23 12:34:30 +0200119 percpu_ref_kill_and_confirm(ref, NULL);
Tejun Heodbece3a2013-06-13 19:23:53 -0700120}
Kent Overstreet215e2622013-05-31 15:26:45 -0700121
Tejun Heoeae79752014-06-28 08:10:13 -0400122/*
123 * Internal helper. Don't use outside percpu-refcount proper. The
124 * function doesn't return the pointer and let the caller test it for NULL
125 * because doing so forces the compiler to generate two conditional
Tejun Heoeecc16b2014-09-24 13:31:48 -0400126 * branches as it can't assume that @ref->percpu_count is not NULL.
Tejun Heoeae79752014-06-28 08:10:13 -0400127 */
Tejun Heo9e804d12014-09-24 13:31:48 -0400128static inline bool __ref_is_percpu(struct percpu_ref *ref,
129 unsigned long __percpu **percpu_countp)
Tejun Heoeae79752014-06-28 08:10:13 -0400130{
Tejun Heo6810e4a2015-01-06 10:26:10 -0500131 unsigned long percpu_ptr;
132
133 /*
134 * The value of @ref->percpu_count_ptr is tested for
135 * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then
136 * used as a pointer. If the compiler generates a separate fetch
137 * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in
138 * between contaminating the pointer value, meaning that
139 * ACCESS_ONCE() is required when fetching it.
140 *
141 * Also, we need a data dependency barrier to be paired with
142 * smp_store_release() in __percpu_ref_switch_to_percpu().
143 *
144 * Use lockless deref which contains both.
145 */
146 percpu_ptr = lockless_dereference(ref->percpu_count_ptr);
Tejun Heo2d722782014-06-28 08:10:14 -0400147
Tejun Heo4aab3b52014-11-22 09:22:42 -0500148 /*
149 * Theoretically, the following could test just ATOMIC; however,
150 * then we'd have to mask off DEAD separately as DEAD may be
151 * visible without ATOMIC if we race with percpu_ref_kill(). DEAD
152 * implies ATOMIC anyway. Test them together.
153 */
154 if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD))
Tejun Heoeae79752014-06-28 08:10:13 -0400155 return false;
156
Tejun Heoeecc16b2014-09-24 13:31:48 -0400157 *percpu_countp = (unsigned long __percpu *)percpu_ptr;
Tejun Heoeae79752014-06-28 08:10:13 -0400158 return true;
159}
Kent Overstreet215e2622013-05-31 15:26:45 -0700160
161/**
Johannes Weinere8ea14c2014-12-10 15:42:42 -0800162 * percpu_ref_get_many - increment a percpu refcount
163 * @ref: percpu_ref to get
164 * @nr: number of references to get
165 *
166 * Analogous to atomic_long_add().
167 *
168 * This function is safe to call as long as @ref is between init and exit.
169 */
170static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
171{
172 unsigned long __percpu *percpu_count;
173
174 rcu_read_lock_sched();
175
176 if (__ref_is_percpu(ref, &percpu_count))
177 this_cpu_add(*percpu_count, nr);
178 else
179 atomic_long_add(nr, &ref->count);
180
181 rcu_read_unlock_sched();
182}
183
184/**
Kent Overstreet215e2622013-05-31 15:26:45 -0700185 * percpu_ref_get - increment a percpu refcount
Tejun Heoac899062013-06-12 20:43:06 -0700186 * @ref: percpu_ref to get
Kent Overstreet215e2622013-05-31 15:26:45 -0700187 *
Tejun Heo6251f992014-09-24 13:31:48 -0400188 * Analagous to atomic_long_inc().
189 *
190 * This function is safe to call as long as @ref is between init and exit.
191 */
Kent Overstreet215e2622013-05-31 15:26:45 -0700192static inline void percpu_ref_get(struct percpu_ref *ref)
193{
Johannes Weinere8ea14c2014-12-10 15:42:42 -0800194 percpu_ref_get_many(ref, 1);
Kent Overstreet215e2622013-05-31 15:26:45 -0700195}
196
197/**
Tejun Heo4fb6e252014-05-09 15:11:53 -0400198 * percpu_ref_tryget - try to increment a percpu refcount
199 * @ref: percpu_ref to try-get
200 *
201 * Increment a percpu refcount unless its count already reached zero.
202 * Returns %true on success; %false on failure.
203 *
Tejun Heo6251f992014-09-24 13:31:48 -0400204 * This function is safe to call as long as @ref is between init and exit.
Tejun Heo4fb6e252014-05-09 15:11:53 -0400205 */
206static inline bool percpu_ref_tryget(struct percpu_ref *ref)
207{
Tejun Heoeecc16b2014-09-24 13:31:48 -0400208 unsigned long __percpu *percpu_count;
Tejun Heo6251f992014-09-24 13:31:48 -0400209 int ret;
Tejun Heo4fb6e252014-05-09 15:11:53 -0400210
211 rcu_read_lock_sched();
212
Tejun Heo9e804d12014-09-24 13:31:48 -0400213 if (__ref_is_percpu(ref, &percpu_count)) {
Tejun Heoeecc16b2014-09-24 13:31:48 -0400214 this_cpu_inc(*percpu_count);
Tejun Heo4fb6e252014-05-09 15:11:53 -0400215 ret = true;
216 } else {
Tejun Heoe6253052014-09-20 01:27:25 -0400217 ret = atomic_long_inc_not_zero(&ref->count);
Tejun Heo4fb6e252014-05-09 15:11:53 -0400218 }
219
220 rcu_read_unlock_sched();
221
222 return ret;
223}
224
225/**
Tejun Heo2070d502014-05-09 15:11:53 -0400226 * percpu_ref_tryget_live - try to increment a live percpu refcount
Tejun Heodbece3a2013-06-13 19:23:53 -0700227 * @ref: percpu_ref to try-get
228 *
229 * Increment a percpu refcount unless it has already been killed. Returns
230 * %true on success; %false on failure.
231 *
Tejun Heo6251f992014-09-24 13:31:48 -0400232 * Completion of percpu_ref_kill() in itself doesn't guarantee that this
233 * function will fail. For such guarantee, percpu_ref_kill_and_confirm()
234 * should be used. After the confirm_kill callback is invoked, it's
235 * guaranteed that no new reference will be given out by
236 * percpu_ref_tryget_live().
Tejun Heo4fb6e252014-05-09 15:11:53 -0400237 *
Tejun Heo6251f992014-09-24 13:31:48 -0400238 * This function is safe to call as long as @ref is between init and exit.
Tejun Heodbece3a2013-06-13 19:23:53 -0700239 */
Tejun Heo2070d502014-05-09 15:11:53 -0400240static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
Tejun Heodbece3a2013-06-13 19:23:53 -0700241{
Tejun Heoeecc16b2014-09-24 13:31:48 -0400242 unsigned long __percpu *percpu_count;
Tejun Heodbece3a2013-06-13 19:23:53 -0700243 int ret = false;
244
Tejun Heoa4244452013-06-16 16:12:26 -0700245 rcu_read_lock_sched();
Tejun Heodbece3a2013-06-13 19:23:53 -0700246
Tejun Heo9e804d12014-09-24 13:31:48 -0400247 if (__ref_is_percpu(ref, &percpu_count)) {
Tejun Heoeecc16b2014-09-24 13:31:48 -0400248 this_cpu_inc(*percpu_count);
Tejun Heodbece3a2013-06-13 19:23:53 -0700249 ret = true;
Tejun Heo6810e4a2015-01-06 10:26:10 -0500250 } else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
Tejun Heo490c79a2014-09-24 13:31:49 -0400251 ret = atomic_long_inc_not_zero(&ref->count);
Tejun Heodbece3a2013-06-13 19:23:53 -0700252 }
253
Tejun Heoa4244452013-06-16 16:12:26 -0700254 rcu_read_unlock_sched();
Tejun Heodbece3a2013-06-13 19:23:53 -0700255
256 return ret;
257}
258
259/**
Johannes Weinere8ea14c2014-12-10 15:42:42 -0800260 * percpu_ref_put_many - decrement a percpu refcount
261 * @ref: percpu_ref to put
262 * @nr: number of references to put
263 *
264 * Decrement the refcount, and if 0, call the release function (which was passed
265 * to percpu_ref_init())
266 *
267 * This function is safe to call as long as @ref is between init and exit.
268 */
269static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
270{
271 unsigned long __percpu *percpu_count;
272
273 rcu_read_lock_sched();
274
275 if (__ref_is_percpu(ref, &percpu_count))
276 this_cpu_sub(*percpu_count, nr);
277 else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
278 ref->release(ref);
279
280 rcu_read_unlock_sched();
281}
282
283/**
Kent Overstreet215e2622013-05-31 15:26:45 -0700284 * percpu_ref_put - decrement a percpu refcount
Tejun Heoac899062013-06-12 20:43:06 -0700285 * @ref: percpu_ref to put
Kent Overstreet215e2622013-05-31 15:26:45 -0700286 *
287 * Decrement the refcount, and if 0, call the release function (which was passed
288 * to percpu_ref_init())
Tejun Heo6251f992014-09-24 13:31:48 -0400289 *
290 * This function is safe to call as long as @ref is between init and exit.
Kent Overstreet215e2622013-05-31 15:26:45 -0700291 */
292static inline void percpu_ref_put(struct percpu_ref *ref)
293{
Johannes Weinere8ea14c2014-12-10 15:42:42 -0800294 percpu_ref_put_many(ref, 1);
Kent Overstreet215e2622013-05-31 15:26:45 -0700295}
296
Tejun Heo2d722782014-06-28 08:10:14 -0400297/**
Tejun Heo4c907ba2015-01-06 10:26:10 -0500298 * percpu_ref_is_dying - test whether a percpu refcount is dying or dead
299 * @ref: percpu_ref to test
300 *
301 * Returns %true if @ref is dying or dead.
302 *
303 * This function is safe to call as long as @ref is between init and exit
304 * and the caller is responsible for synchronizing against state changes.
305 */
306static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
307{
308 return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
309}
310
311/**
Tejun Heo2d722782014-06-28 08:10:14 -0400312 * percpu_ref_is_zero - test whether a percpu refcount reached zero
313 * @ref: percpu_ref to test
314 *
315 * Returns %true if @ref reached zero.
Tejun Heo6251f992014-09-24 13:31:48 -0400316 *
317 * This function is safe to call as long as @ref is between init and exit.
Tejun Heo2d722782014-06-28 08:10:14 -0400318 */
319static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
320{
Tejun Heoeecc16b2014-09-24 13:31:48 -0400321 unsigned long __percpu *percpu_count;
Tejun Heo2d722782014-06-28 08:10:14 -0400322
Tejun Heo9e804d12014-09-24 13:31:48 -0400323 if (__ref_is_percpu(ref, &percpu_count))
Tejun Heo2d722782014-06-28 08:10:14 -0400324 return false;
Tejun Heoe6253052014-09-20 01:27:25 -0400325 return !atomic_long_read(&ref->count);
Tejun Heo2d722782014-06-28 08:10:14 -0400326}
327
Kent Overstreet215e2622013-05-31 15:26:45 -0700328#endif