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 * calls percpu_ref_kill(), then hlist_del_rcu() and synchronize_rcu() to remove 34 * the kioctx from the proccess's list of kioctxs - after that, there can't be 35 * any new users of the kioctx (from lookup_ioctx()) and it's then safe to drop 36 * the initial ref with percpu_ref_put(). 37 * 38 * Code that does a two stage shutdown like this often needs some kind of 39 * explicit synchronization to ensure the initial refcount can only be dropped 40 * once - percpu_ref_kill() does this for you, it returns true once and false if 41 * someone else already called it. The aio code uses it this way, but it's not 42 * necessary if the code has some other mechanism to synchronize teardown. 43 * around. 44 */ 45 46 #ifndef _LINUX_PERCPU_REFCOUNT_H 47 #define _LINUX_PERCPU_REFCOUNT_H 48 49 #include <linux/atomic.h> 50 #include <linux/kernel.h> 51 #include <linux/percpu.h> 52 #include <linux/rcupdate.h> 53 #include <linux/gfp.h> 54 55 struct percpu_ref; 56 typedef void (percpu_ref_func_t)(struct percpu_ref *); 57 58 /* flags set in the lower bits of percpu_ref->percpu_count_ptr */ 59 enum { 60 __PERCPU_REF_ATOMIC = 1LU << 0, /* operating in atomic mode */ 61 __PERCPU_REF_DEAD = 1LU << 1, /* (being) killed */ 62 __PERCPU_REF_ATOMIC_DEAD = __PERCPU_REF_ATOMIC | __PERCPU_REF_DEAD, 63 64 __PERCPU_REF_FLAG_BITS = 2, 65 }; 66 67 /* @flags for percpu_ref_init() */ 68 enum { 69 /* 70 * Start w/ ref == 1 in atomic mode. Can be switched to percpu 71 * operation using percpu_ref_switch_to_percpu(). If initialized 72 * with this flag, the ref will stay in atomic mode until 73 * percpu_ref_switch_to_percpu() is invoked on it. 74 */ 75 PERCPU_REF_INIT_ATOMIC = 1 << 0, 76 77 /* 78 * Start dead w/ ref == 0 in atomic mode. Must be revived with 79 * percpu_ref_reinit() before used. Implies INIT_ATOMIC. 80 */ 81 PERCPU_REF_INIT_DEAD = 1 << 1, 82 }; 83 84 struct percpu_ref { 85 atomic_long_t count; 86 /* 87 * The low bit of the pointer indicates whether the ref is in percpu 88 * mode; if set, then get/put will manipulate the atomic_t. 89 */ 90 unsigned long percpu_count_ptr; 91 percpu_ref_func_t *release; 92 percpu_ref_func_t *confirm_switch; 93 bool force_atomic:1; 94 struct rcu_head rcu; 95 }; 96 97 int __must_check percpu_ref_init(struct percpu_ref *ref, 98 percpu_ref_func_t *release, unsigned int flags, 99 gfp_t gfp); 100 void percpu_ref_exit(struct percpu_ref *ref); 101 void percpu_ref_switch_to_atomic(struct percpu_ref *ref, 102 percpu_ref_func_t *confirm_switch); 103 void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref); 104 void percpu_ref_switch_to_percpu(struct percpu_ref *ref); 105 void percpu_ref_kill_and_confirm(struct percpu_ref *ref, 106 percpu_ref_func_t *confirm_kill); 107 void percpu_ref_reinit(struct percpu_ref *ref); 108 109 /** 110 * percpu_ref_kill - drop the initial ref 111 * @ref: percpu_ref to kill 112 * 113 * Must be used to drop the initial ref on a percpu refcount; must be called 114 * precisely once before shutdown. 115 * 116 * Puts @ref in non percpu mode, then does a call_rcu() before gathering up the 117 * percpu counters and dropping the initial ref. 118 */ 119 static inline void percpu_ref_kill(struct percpu_ref *ref) 120 { 121 percpu_ref_kill_and_confirm(ref, NULL); 122 } 123 124 /* 125 * Internal helper. Don't use outside percpu-refcount proper. The 126 * function doesn't return the pointer and let the caller test it for NULL 127 * because doing so forces the compiler to generate two conditional 128 * branches as it can't assume that @ref->percpu_count is not NULL. 129 */ 130 static inline bool __ref_is_percpu(struct percpu_ref *ref, 131 unsigned long __percpu **percpu_countp) 132 { 133 unsigned long percpu_ptr; 134 135 /* 136 * The value of @ref->percpu_count_ptr is tested for 137 * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then 138 * used as a pointer. If the compiler generates a separate fetch 139 * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in 140 * between contaminating the pointer value, meaning that 141 * READ_ONCE() is required when fetching it. 142 */ 143 percpu_ptr = READ_ONCE(ref->percpu_count_ptr); 144 145 /* paired with smp_store_release() in __percpu_ref_switch_to_percpu() */ 146 smp_read_barrier_depends(); 147 148 /* 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)) 155 return false; 156 157 *percpu_countp = (unsigned long __percpu *)percpu_ptr; 158 return true; 159 } 160 161 /** 162 * 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 */ 170 static 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 /** 185 * percpu_ref_get - increment a percpu refcount 186 * @ref: percpu_ref to get 187 * 188 * Analagous to atomic_long_inc(). 189 * 190 * This function is safe to call as long as @ref is between init and exit. 191 */ 192 static inline void percpu_ref_get(struct percpu_ref *ref) 193 { 194 percpu_ref_get_many(ref, 1); 195 } 196 197 /** 198 * 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 * 204 * This function is safe to call as long as @ref is between init and exit. 205 */ 206 static inline bool percpu_ref_tryget(struct percpu_ref *ref) 207 { 208 unsigned long __percpu *percpu_count; 209 bool ret; 210 211 rcu_read_lock_sched(); 212 213 if (__ref_is_percpu(ref, &percpu_count)) { 214 this_cpu_inc(*percpu_count); 215 ret = true; 216 } else { 217 ret = atomic_long_inc_not_zero(&ref->count); 218 } 219 220 rcu_read_unlock_sched(); 221 222 return ret; 223 } 224 225 /** 226 * percpu_ref_tryget_live - try to increment a live percpu refcount 227 * @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 * 232 * 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(). 237 * 238 * This function is safe to call as long as @ref is between init and exit. 239 */ 240 static inline bool percpu_ref_tryget_live(struct percpu_ref *ref) 241 { 242 unsigned long __percpu *percpu_count; 243 bool ret = false; 244 245 rcu_read_lock_sched(); 246 247 if (__ref_is_percpu(ref, &percpu_count)) { 248 this_cpu_inc(*percpu_count); 249 ret = true; 250 } else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) { 251 ret = atomic_long_inc_not_zero(&ref->count); 252 } 253 254 rcu_read_unlock_sched(); 255 256 return ret; 257 } 258 259 /** 260 * 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 */ 269 static 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 /** 284 * percpu_ref_put - decrement a percpu refcount 285 * @ref: percpu_ref to put 286 * 287 * Decrement the refcount, and if 0, call the release function (which was passed 288 * to percpu_ref_init()) 289 * 290 * This function is safe to call as long as @ref is between init and exit. 291 */ 292 static inline void percpu_ref_put(struct percpu_ref *ref) 293 { 294 percpu_ref_put_many(ref, 1); 295 } 296 297 /** 298 * 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 */ 306 static inline bool percpu_ref_is_dying(struct percpu_ref *ref) 307 { 308 return ref->percpu_count_ptr & __PERCPU_REF_DEAD; 309 } 310 311 /** 312 * 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. 316 * 317 * This function is safe to call as long as @ref is between init and exit. 318 */ 319 static inline bool percpu_ref_is_zero(struct percpu_ref *ref) 320 { 321 unsigned long __percpu *percpu_count; 322 323 if (__ref_is_percpu(ref, &percpu_count)) 324 return false; 325 return !atomic_long_read(&ref->count); 326 } 327 328 #endif 329