Commit | Line | Data |
---|---|---|
b2441318 | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
215e262f KO |
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 | |
eecc16ba | 17 | * percpu_ref_kill()/PERCPU_COUNT_BIAS. |
215e262f KO |
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 | |
b3a5d111 TH |
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. | |
215e262f KO |
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> | |
a34375ef | 57 | #include <linux/gfp.h> |
215e262f KO |
58 | |
59 | struct percpu_ref; | |
ac899061 | 60 | typedef void (percpu_ref_func_t)(struct percpu_ref *); |
215e262f | 61 | |
9e804d1f TH |
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 */ | |
27344a90 TH |
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, | |
9e804d1f TH |
69 | }; |
70 | ||
2aad2a86 TH |
71 | /* @flags for percpu_ref_init() */ |
72 | enum { | |
73 | /* | |
74 | * Start w/ ref == 1 in atomic mode. Can be switched to percpu | |
1cae13e7 TH |
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. | |
2aad2a86 TH |
78 | */ |
79 | PERCPU_REF_INIT_ATOMIC = 1 << 0, | |
80 | ||
81 | /* | |
82 | * Start dead w/ ref == 0 in atomic mode. Must be revived with | |
83 | * percpu_ref_reinit() before used. Implies INIT_ATOMIC. | |
84 | */ | |
85 | PERCPU_REF_INIT_DEAD = 1 << 1, | |
86 | }; | |
87 | ||
215e262f | 88 | struct percpu_ref { |
e625305b | 89 | atomic_long_t count; |
215e262f KO |
90 | /* |
91 | * The low bit of the pointer indicates whether the ref is in percpu | |
9a1049da | 92 | * mode; if set, then get/put will manipulate the atomic_t. |
215e262f | 93 | */ |
eecc16ba | 94 | unsigned long percpu_count_ptr; |
ac899061 | 95 | percpu_ref_func_t *release; |
9e804d1f | 96 | percpu_ref_func_t *confirm_switch; |
1cae13e7 | 97 | bool force_atomic:1; |
215e262f KO |
98 | struct rcu_head rcu; |
99 | }; | |
100 | ||
acac7883 | 101 | int __must_check percpu_ref_init(struct percpu_ref *ref, |
2aad2a86 TH |
102 | percpu_ref_func_t *release, unsigned int flags, |
103 | gfp_t gfp); | |
9a1049da | 104 | void percpu_ref_exit(struct percpu_ref *ref); |
490c79a6 TH |
105 | void percpu_ref_switch_to_atomic(struct percpu_ref *ref, |
106 | percpu_ref_func_t *confirm_switch); | |
210f7cdc | 107 | void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref); |
f47ad457 | 108 | void percpu_ref_switch_to_percpu(struct percpu_ref *ref); |
dbece3a0 TH |
109 | void percpu_ref_kill_and_confirm(struct percpu_ref *ref, |
110 | percpu_ref_func_t *confirm_kill); | |
18c9a6bb | 111 | void percpu_ref_resurrect(struct percpu_ref *ref); |
f47ad457 | 112 | void percpu_ref_reinit(struct percpu_ref *ref); |
dbece3a0 TH |
113 | |
114 | /** | |
115 | * percpu_ref_kill - drop the initial ref | |
116 | * @ref: percpu_ref to kill | |
117 | * | |
118 | * Must be used to drop the initial ref on a percpu refcount; must be called | |
119 | * precisely once before shutdown. | |
120 | * | |
b3a5d111 TH |
121 | * Switches @ref into atomic mode before gathering up the percpu counters |
122 | * and dropping the initial ref. | |
123 | * | |
124 | * There are no implied RCU grace periods between kill and release. | |
dbece3a0 TH |
125 | */ |
126 | static inline void percpu_ref_kill(struct percpu_ref *ref) | |
127 | { | |
4d414269 | 128 | percpu_ref_kill_and_confirm(ref, NULL); |
dbece3a0 | 129 | } |
215e262f | 130 | |
eae7975d TH |
131 | /* |
132 | * Internal helper. Don't use outside percpu-refcount proper. The | |
133 | * function doesn't return the pointer and let the caller test it for NULL | |
134 | * because doing so forces the compiler to generate two conditional | |
eecc16ba | 135 | * branches as it can't assume that @ref->percpu_count is not NULL. |
eae7975d | 136 | */ |
9e804d1f TH |
137 | static inline bool __ref_is_percpu(struct percpu_ref *ref, |
138 | unsigned long __percpu **percpu_countp) | |
eae7975d | 139 | { |
6810e4a3 TH |
140 | unsigned long percpu_ptr; |
141 | ||
142 | /* | |
143 | * The value of @ref->percpu_count_ptr is tested for | |
144 | * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then | |
145 | * used as a pointer. If the compiler generates a separate fetch | |
146 | * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in | |
147 | * between contaminating the pointer value, meaning that | |
ed8ebd1d | 148 | * READ_ONCE() is required when fetching it. |
b393e8b3 PM |
149 | * |
150 | * The smp_read_barrier_depends() implied by READ_ONCE() pairs | |
151 | * with smp_store_release() in __percpu_ref_switch_to_percpu(). | |
6810e4a3 | 152 | */ |
ed8ebd1d TH |
153 | percpu_ptr = READ_ONCE(ref->percpu_count_ptr); |
154 | ||
4aab3b5b TH |
155 | /* |
156 | * Theoretically, the following could test just ATOMIC; however, | |
157 | * then we'd have to mask off DEAD separately as DEAD may be | |
158 | * visible without ATOMIC if we race with percpu_ref_kill(). DEAD | |
159 | * implies ATOMIC anyway. Test them together. | |
160 | */ | |
161 | if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD)) | |
eae7975d TH |
162 | return false; |
163 | ||
eecc16ba | 164 | *percpu_countp = (unsigned long __percpu *)percpu_ptr; |
eae7975d TH |
165 | return true; |
166 | } | |
215e262f KO |
167 | |
168 | /** | |
e8ea14cc | 169 | * percpu_ref_get_many - increment a percpu refcount |
ac899061 | 170 | * @ref: percpu_ref to get |
e8ea14cc | 171 | * @nr: number of references to get |
215e262f | 172 | * |
e8ea14cc | 173 | * Analogous to atomic_long_add(). |
6251f997 TH |
174 | * |
175 | * This function is safe to call as long as @ref is between init and exit. | |
176 | */ | |
e8ea14cc | 177 | static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr) |
215e262f | 178 | { |
eecc16ba | 179 | unsigned long __percpu *percpu_count; |
215e262f | 180 | |
a4244454 | 181 | rcu_read_lock_sched(); |
215e262f | 182 | |
9e804d1f | 183 | if (__ref_is_percpu(ref, &percpu_count)) |
e8ea14cc | 184 | this_cpu_add(*percpu_count, nr); |
215e262f | 185 | else |
e8ea14cc | 186 | atomic_long_add(nr, &ref->count); |
215e262f | 187 | |
a4244454 | 188 | rcu_read_unlock_sched(); |
215e262f KO |
189 | } |
190 | ||
e8ea14cc JW |
191 | /** |
192 | * percpu_ref_get - increment a percpu refcount | |
193 | * @ref: percpu_ref to get | |
194 | * | |
195 | * Analagous to atomic_long_inc(). | |
196 | * | |
197 | * This function is safe to call as long as @ref is between init and exit. | |
198 | */ | |
199 | static inline void percpu_ref_get(struct percpu_ref *ref) | |
200 | { | |
201 | percpu_ref_get_many(ref, 1); | |
202 | } | |
203 | ||
4fb6e250 TH |
204 | /** |
205 | * percpu_ref_tryget - try to increment a percpu refcount | |
206 | * @ref: percpu_ref to try-get | |
207 | * | |
208 | * Increment a percpu refcount unless its count already reached zero. | |
209 | * Returns %true on success; %false on failure. | |
210 | * | |
6251f997 | 211 | * This function is safe to call as long as @ref is between init and exit. |
4fb6e250 TH |
212 | */ |
213 | static inline bool percpu_ref_tryget(struct percpu_ref *ref) | |
214 | { | |
eecc16ba | 215 | unsigned long __percpu *percpu_count; |
966d2b04 | 216 | bool ret; |
4fb6e250 TH |
217 | |
218 | rcu_read_lock_sched(); | |
219 | ||
9e804d1f | 220 | if (__ref_is_percpu(ref, &percpu_count)) { |
eecc16ba | 221 | this_cpu_inc(*percpu_count); |
4fb6e250 TH |
222 | ret = true; |
223 | } else { | |
e625305b | 224 | ret = atomic_long_inc_not_zero(&ref->count); |
4fb6e250 TH |
225 | } |
226 | ||
227 | rcu_read_unlock_sched(); | |
228 | ||
229 | return ret; | |
230 | } | |
231 | ||
dbece3a0 | 232 | /** |
2070d50e | 233 | * percpu_ref_tryget_live - try to increment a live percpu refcount |
dbece3a0 TH |
234 | * @ref: percpu_ref to try-get |
235 | * | |
236 | * Increment a percpu refcount unless it has already been killed. Returns | |
237 | * %true on success; %false on failure. | |
238 | * | |
6251f997 TH |
239 | * Completion of percpu_ref_kill() in itself doesn't guarantee that this |
240 | * function will fail. For such guarantee, percpu_ref_kill_and_confirm() | |
241 | * should be used. After the confirm_kill callback is invoked, it's | |
242 | * guaranteed that no new reference will be given out by | |
243 | * percpu_ref_tryget_live(). | |
4fb6e250 | 244 | * |
6251f997 | 245 | * This function is safe to call as long as @ref is between init and exit. |
dbece3a0 | 246 | */ |
2070d50e | 247 | static inline bool percpu_ref_tryget_live(struct percpu_ref *ref) |
dbece3a0 | 248 | { |
eecc16ba | 249 | unsigned long __percpu *percpu_count; |
966d2b04 | 250 | bool ret = false; |
dbece3a0 | 251 | |
a4244454 | 252 | rcu_read_lock_sched(); |
dbece3a0 | 253 | |
9e804d1f | 254 | if (__ref_is_percpu(ref, &percpu_count)) { |
eecc16ba | 255 | this_cpu_inc(*percpu_count); |
dbece3a0 | 256 | ret = true; |
6810e4a3 | 257 | } else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) { |
490c79a6 | 258 | ret = atomic_long_inc_not_zero(&ref->count); |
dbece3a0 TH |
259 | } |
260 | ||
a4244454 | 261 | rcu_read_unlock_sched(); |
dbece3a0 TH |
262 | |
263 | return ret; | |
264 | } | |
265 | ||
215e262f | 266 | /** |
e8ea14cc | 267 | * percpu_ref_put_many - decrement a percpu refcount |
ac899061 | 268 | * @ref: percpu_ref to put |
e8ea14cc | 269 | * @nr: number of references to put |
215e262f KO |
270 | * |
271 | * Decrement the refcount, and if 0, call the release function (which was passed | |
272 | * to percpu_ref_init()) | |
6251f997 TH |
273 | * |
274 | * This function is safe to call as long as @ref is between init and exit. | |
215e262f | 275 | */ |
e8ea14cc | 276 | static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr) |
215e262f | 277 | { |
eecc16ba | 278 | unsigned long __percpu *percpu_count; |
215e262f | 279 | |
a4244454 | 280 | rcu_read_lock_sched(); |
215e262f | 281 | |
9e804d1f | 282 | if (__ref_is_percpu(ref, &percpu_count)) |
e8ea14cc JW |
283 | this_cpu_sub(*percpu_count, nr); |
284 | else if (unlikely(atomic_long_sub_and_test(nr, &ref->count))) | |
215e262f KO |
285 | ref->release(ref); |
286 | ||
a4244454 | 287 | rcu_read_unlock_sched(); |
215e262f KO |
288 | } |
289 | ||
e8ea14cc JW |
290 | /** |
291 | * percpu_ref_put - decrement a percpu refcount | |
292 | * @ref: percpu_ref to put | |
293 | * | |
294 | * Decrement the refcount, and if 0, call the release function (which was passed | |
295 | * to percpu_ref_init()) | |
296 | * | |
297 | * This function is safe to call as long as @ref is between init and exit. | |
298 | */ | |
299 | static inline void percpu_ref_put(struct percpu_ref *ref) | |
300 | { | |
301 | percpu_ref_put_many(ref, 1); | |
302 | } | |
303 | ||
4c907baf TH |
304 | /** |
305 | * percpu_ref_is_dying - test whether a percpu refcount is dying or dead | |
306 | * @ref: percpu_ref to test | |
307 | * | |
308 | * Returns %true if @ref is dying or dead. | |
309 | * | |
310 | * This function is safe to call as long as @ref is between init and exit | |
311 | * and the caller is responsible for synchronizing against state changes. | |
312 | */ | |
313 | static inline bool percpu_ref_is_dying(struct percpu_ref *ref) | |
314 | { | |
315 | return ref->percpu_count_ptr & __PERCPU_REF_DEAD; | |
316 | } | |
317 | ||
2d722782 TH |
318 | /** |
319 | * percpu_ref_is_zero - test whether a percpu refcount reached zero | |
320 | * @ref: percpu_ref to test | |
321 | * | |
322 | * Returns %true if @ref reached zero. | |
6251f997 TH |
323 | * |
324 | * This function is safe to call as long as @ref is between init and exit. | |
2d722782 TH |
325 | */ |
326 | static inline bool percpu_ref_is_zero(struct percpu_ref *ref) | |
327 | { | |
eecc16ba | 328 | unsigned long __percpu *percpu_count; |
2d722782 | 329 | |
9e804d1f | 330 | if (__ref_is_percpu(ref, &percpu_count)) |
2d722782 | 331 | return false; |
e625305b | 332 | return !atomic_long_read(&ref->count); |
2d722782 TH |
333 | } |
334 | ||
215e262f | 335 | #endif |