Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 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. |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame] | 78 | * Implies ALLOW_REINIT. |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 79 | */ |
| 80 | PERCPU_REF_INIT_ATOMIC = 1 << 0, |
| 81 | |
| 82 | /* |
| 83 | * Start dead w/ ref == 0 in atomic mode. Must be revived with |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame] | 84 | * percpu_ref_reinit() before used. Implies INIT_ATOMIC and |
| 85 | * ALLOW_REINIT. |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 86 | */ |
| 87 | PERCPU_REF_INIT_DEAD = 1 << 1, |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame] | 88 | |
| 89 | /* |
| 90 | * Allow switching from atomic mode to percpu mode. |
| 91 | */ |
| 92 | PERCPU_REF_ALLOW_REINIT = 1 << 2, |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 93 | }; |
| 94 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 95 | struct percpu_ref_data { |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 96 | atomic_long_t count; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 97 | percpu_ref_func_t *release; |
| 98 | percpu_ref_func_t *confirm_switch; |
| 99 | bool force_atomic:1; |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame] | 100 | bool allow_reinit:1; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 101 | struct rcu_head rcu; |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 102 | struct percpu_ref *ref; |
| 103 | }; |
| 104 | |
| 105 | struct percpu_ref { |
| 106 | /* |
| 107 | * The low bit of the pointer indicates whether the ref is in percpu |
| 108 | * mode; if set, then get/put will manipulate the atomic_t. |
| 109 | */ |
| 110 | unsigned long percpu_count_ptr; |
| 111 | |
| 112 | /* |
| 113 | * 'percpu_ref' is often embedded into user structure, and only |
| 114 | * 'percpu_count_ptr' is required in fast path, move other fields |
| 115 | * into 'percpu_ref_data', so we can reduce memory footprint in |
| 116 | * fast path. |
| 117 | */ |
| 118 | struct percpu_ref_data *data; |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 119 | }; |
| 120 | |
| 121 | int __must_check percpu_ref_init(struct percpu_ref *ref, |
| 122 | percpu_ref_func_t *release, unsigned int flags, |
| 123 | gfp_t gfp); |
| 124 | void percpu_ref_exit(struct percpu_ref *ref); |
| 125 | void percpu_ref_switch_to_atomic(struct percpu_ref *ref, |
| 126 | percpu_ref_func_t *confirm_switch); |
| 127 | void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref); |
| 128 | void percpu_ref_switch_to_percpu(struct percpu_ref *ref); |
| 129 | void percpu_ref_kill_and_confirm(struct percpu_ref *ref, |
| 130 | percpu_ref_func_t *confirm_kill); |
David Brazdil | 0f672f6 | 2019-12-10 10:32:29 +0000 | [diff] [blame] | 131 | void percpu_ref_resurrect(struct percpu_ref *ref); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 132 | void percpu_ref_reinit(struct percpu_ref *ref); |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 133 | bool percpu_ref_is_zero(struct percpu_ref *ref); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 134 | |
| 135 | /** |
| 136 | * percpu_ref_kill - drop the initial ref |
| 137 | * @ref: percpu_ref to kill |
| 138 | * |
| 139 | * Must be used to drop the initial ref on a percpu refcount; must be called |
| 140 | * precisely once before shutdown. |
| 141 | * |
| 142 | * Switches @ref into atomic mode before gathering up the percpu counters |
| 143 | * and dropping the initial ref. |
| 144 | * |
| 145 | * There are no implied RCU grace periods between kill and release. |
| 146 | */ |
| 147 | static inline void percpu_ref_kill(struct percpu_ref *ref) |
| 148 | { |
| 149 | percpu_ref_kill_and_confirm(ref, NULL); |
| 150 | } |
| 151 | |
| 152 | /* |
| 153 | * Internal helper. Don't use outside percpu-refcount proper. The |
| 154 | * function doesn't return the pointer and let the caller test it for NULL |
| 155 | * because doing so forces the compiler to generate two conditional |
| 156 | * branches as it can't assume that @ref->percpu_count is not NULL. |
| 157 | */ |
| 158 | static inline bool __ref_is_percpu(struct percpu_ref *ref, |
| 159 | unsigned long __percpu **percpu_countp) |
| 160 | { |
| 161 | unsigned long percpu_ptr; |
| 162 | |
| 163 | /* |
| 164 | * The value of @ref->percpu_count_ptr is tested for |
| 165 | * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then |
| 166 | * used as a pointer. If the compiler generates a separate fetch |
| 167 | * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in |
| 168 | * between contaminating the pointer value, meaning that |
| 169 | * READ_ONCE() is required when fetching it. |
| 170 | * |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 171 | * The dependency ordering from the READ_ONCE() pairs |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 172 | * with smp_store_release() in __percpu_ref_switch_to_percpu(). |
| 173 | */ |
| 174 | percpu_ptr = READ_ONCE(ref->percpu_count_ptr); |
| 175 | |
| 176 | /* |
| 177 | * Theoretically, the following could test just ATOMIC; however, |
| 178 | * then we'd have to mask off DEAD separately as DEAD may be |
| 179 | * visible without ATOMIC if we race with percpu_ref_kill(). DEAD |
| 180 | * implies ATOMIC anyway. Test them together. |
| 181 | */ |
| 182 | if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD)) |
| 183 | return false; |
| 184 | |
| 185 | *percpu_countp = (unsigned long __percpu *)percpu_ptr; |
| 186 | return true; |
| 187 | } |
| 188 | |
| 189 | /** |
| 190 | * percpu_ref_get_many - increment a percpu refcount |
| 191 | * @ref: percpu_ref to get |
| 192 | * @nr: number of references to get |
| 193 | * |
| 194 | * Analogous to atomic_long_add(). |
| 195 | * |
| 196 | * This function is safe to call as long as @ref is between init and exit. |
| 197 | */ |
| 198 | static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr) |
| 199 | { |
| 200 | unsigned long __percpu *percpu_count; |
| 201 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 202 | rcu_read_lock(); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 203 | |
| 204 | if (__ref_is_percpu(ref, &percpu_count)) |
| 205 | this_cpu_add(*percpu_count, nr); |
| 206 | else |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 207 | atomic_long_add(nr, &ref->data->count); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 208 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 209 | rcu_read_unlock(); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 210 | } |
| 211 | |
| 212 | /** |
| 213 | * percpu_ref_get - increment a percpu refcount |
| 214 | * @ref: percpu_ref to get |
| 215 | * |
| 216 | * Analagous to atomic_long_inc(). |
| 217 | * |
| 218 | * This function is safe to call as long as @ref is between init and exit. |
| 219 | */ |
| 220 | static inline void percpu_ref_get(struct percpu_ref *ref) |
| 221 | { |
| 222 | percpu_ref_get_many(ref, 1); |
| 223 | } |
| 224 | |
| 225 | /** |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 226 | * percpu_ref_tryget_many - try to increment a percpu refcount |
| 227 | * @ref: percpu_ref to try-get |
| 228 | * @nr: number of references to get |
| 229 | * |
| 230 | * Increment a percpu refcount by @nr unless its count already reached zero. |
| 231 | * Returns %true on success; %false on failure. |
| 232 | * |
| 233 | * This function is safe to call as long as @ref is between init and exit. |
| 234 | */ |
| 235 | static inline bool percpu_ref_tryget_many(struct percpu_ref *ref, |
| 236 | unsigned long nr) |
| 237 | { |
| 238 | unsigned long __percpu *percpu_count; |
| 239 | bool ret; |
| 240 | |
| 241 | rcu_read_lock(); |
| 242 | |
| 243 | if (__ref_is_percpu(ref, &percpu_count)) { |
| 244 | this_cpu_add(*percpu_count, nr); |
| 245 | ret = true; |
| 246 | } else { |
| 247 | ret = atomic_long_add_unless(&ref->data->count, nr, 0); |
| 248 | } |
| 249 | |
| 250 | rcu_read_unlock(); |
| 251 | |
| 252 | return ret; |
| 253 | } |
| 254 | |
| 255 | /** |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 256 | * percpu_ref_tryget - try to increment a percpu refcount |
| 257 | * @ref: percpu_ref to try-get |
| 258 | * |
| 259 | * Increment a percpu refcount unless its count already reached zero. |
| 260 | * Returns %true on success; %false on failure. |
| 261 | * |
| 262 | * This function is safe to call as long as @ref is between init and exit. |
| 263 | */ |
| 264 | static inline bool percpu_ref_tryget(struct percpu_ref *ref) |
| 265 | { |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 266 | return percpu_ref_tryget_many(ref, 1); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 267 | } |
| 268 | |
| 269 | /** |
| 270 | * percpu_ref_tryget_live - try to increment a live percpu refcount |
| 271 | * @ref: percpu_ref to try-get |
| 272 | * |
| 273 | * Increment a percpu refcount unless it has already been killed. Returns |
| 274 | * %true on success; %false on failure. |
| 275 | * |
| 276 | * Completion of percpu_ref_kill() in itself doesn't guarantee that this |
| 277 | * function will fail. For such guarantee, percpu_ref_kill_and_confirm() |
| 278 | * should be used. After the confirm_kill callback is invoked, it's |
| 279 | * guaranteed that no new reference will be given out by |
| 280 | * percpu_ref_tryget_live(). |
| 281 | * |
| 282 | * This function is safe to call as long as @ref is between init and exit. |
| 283 | */ |
| 284 | static inline bool percpu_ref_tryget_live(struct percpu_ref *ref) |
| 285 | { |
| 286 | unsigned long __percpu *percpu_count; |
| 287 | bool ret = false; |
| 288 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 289 | rcu_read_lock(); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 290 | |
| 291 | if (__ref_is_percpu(ref, &percpu_count)) { |
| 292 | this_cpu_inc(*percpu_count); |
| 293 | ret = true; |
| 294 | } else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) { |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 295 | ret = atomic_long_inc_not_zero(&ref->data->count); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 296 | } |
| 297 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 298 | rcu_read_unlock(); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 299 | |
| 300 | return ret; |
| 301 | } |
| 302 | |
| 303 | /** |
| 304 | * percpu_ref_put_many - decrement a percpu refcount |
| 305 | * @ref: percpu_ref to put |
| 306 | * @nr: number of references to put |
| 307 | * |
| 308 | * Decrement the refcount, and if 0, call the release function (which was passed |
| 309 | * to percpu_ref_init()) |
| 310 | * |
| 311 | * This function is safe to call as long as @ref is between init and exit. |
| 312 | */ |
| 313 | static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr) |
| 314 | { |
| 315 | unsigned long __percpu *percpu_count; |
| 316 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 317 | rcu_read_lock(); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 318 | |
| 319 | if (__ref_is_percpu(ref, &percpu_count)) |
| 320 | this_cpu_sub(*percpu_count, nr); |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 321 | else if (unlikely(atomic_long_sub_and_test(nr, &ref->data->count))) |
| 322 | ref->data->release(ref); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 323 | |
Olivier Deprez | 157378f | 2022-04-04 15:47:50 +0200 | [diff] [blame] | 324 | rcu_read_unlock(); |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 325 | } |
| 326 | |
| 327 | /** |
| 328 | * percpu_ref_put - decrement a percpu refcount |
| 329 | * @ref: percpu_ref to put |
| 330 | * |
| 331 | * Decrement the refcount, and if 0, call the release function (which was passed |
| 332 | * to percpu_ref_init()) |
| 333 | * |
| 334 | * This function is safe to call as long as @ref is between init and exit. |
| 335 | */ |
| 336 | static inline void percpu_ref_put(struct percpu_ref *ref) |
| 337 | { |
| 338 | percpu_ref_put_many(ref, 1); |
| 339 | } |
| 340 | |
| 341 | /** |
| 342 | * percpu_ref_is_dying - test whether a percpu refcount is dying or dead |
| 343 | * @ref: percpu_ref to test |
| 344 | * |
| 345 | * Returns %true if @ref is dying or dead. |
| 346 | * |
| 347 | * This function is safe to call as long as @ref is between init and exit |
| 348 | * and the caller is responsible for synchronizing against state changes. |
| 349 | */ |
| 350 | static inline bool percpu_ref_is_dying(struct percpu_ref *ref) |
| 351 | { |
| 352 | return ref->percpu_count_ptr & __PERCPU_REF_DEAD; |
| 353 | } |
| 354 | |
Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 355 | #endif |