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