1 /* 2 * Percpu refcounts: 3 * (C) 2012 Google, Inc. 4 * Author: Kent Overstreet <[email protected]> 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 16 * percpu_ref_kill()/PERCPU_COUNT_BIAS. 17 * 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 32 * calls percpu_ref_kill(), then hlist_del_rcu() and synchronize_rcu() to remove 33 * 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> 52 #include <linux/gfp.h> 53 54 struct percpu_ref; 55 typedef void (percpu_ref_func_t)(struct percpu_ref *); 56 57 /* flags set in the lower bits of percpu_ref->percpu_count_ptr */ 58 enum { 59 __PERCPU_REF_ATOMIC = 1LU << 0, /* operating in atomic mode */ 60 __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, 64 }; 65 66 /* @flags for percpu_ref_init() */ 67 enum { 68 /* 69 * Start w/ ref == 1 in atomic mode. Can be switched to percpu 70 * 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. 73 */ 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 83 struct percpu_ref { 84 atomic_long_t count; 85 /* 86 * The low bit of the pointer indicates whether the ref is in percpu 87 * mode; if set, then get/put will manipulate the atomic_t. 88 */ 89 unsigned long percpu_count_ptr; 90 percpu_ref_func_t *release; 91 percpu_ref_func_t *confirm_switch; 92 bool force_atomic:1; 93 struct rcu_head rcu; 94 }; 95 96 int __must_check percpu_ref_init(struct percpu_ref *ref, 97 percpu_ref_func_t *release, unsigned int flags, 98 gfp_t gfp); 99 void percpu_ref_exit(struct percpu_ref *ref); 100 void percpu_ref_switch_to_atomic(struct percpu_ref *ref, 101 percpu_ref_func_t *confirm_switch); 102 void percpu_ref_switch_to_percpu(struct percpu_ref *ref); 103 void percpu_ref_kill_and_confirm(struct percpu_ref *ref, 104 percpu_ref_func_t *confirm_kill); 105 void percpu_ref_reinit(struct percpu_ref *ref); 106 107 /** 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 */ 117 static inline void percpu_ref_kill(struct percpu_ref *ref) 118 { 119 return percpu_ref_kill_and_confirm(ref, NULL); 120 } 121 122 /* 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 126 * branches as it can't assume that @ref->percpu_count is not NULL. 127 */ 128 static inline bool __ref_is_percpu(struct percpu_ref *ref, 129 unsigned long __percpu **percpu_countp) 130 { 131 unsigned long percpu_ptr = ACCESS_ONCE(ref->percpu_count_ptr); 132 133 /* paired with smp_store_release() in percpu_ref_reinit() */ 134 smp_read_barrier_depends(); 135 136 /* 137 * Theoretically, the following could test just ATOMIC; however, 138 * then we'd have to mask off DEAD separately as DEAD may be 139 * visible without ATOMIC if we race with percpu_ref_kill(). DEAD 140 * implies ATOMIC anyway. Test them together. 141 */ 142 if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD)) 143 return false; 144 145 *percpu_countp = (unsigned long __percpu *)percpu_ptr; 146 return true; 147 } 148 149 /** 150 * percpu_ref_get_many - increment a percpu refcount 151 * @ref: percpu_ref to get 152 * @nr: number of references to get 153 * 154 * Analogous to atomic_long_add(). 155 * 156 * This function is safe to call as long as @ref is between init and exit. 157 */ 158 static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr) 159 { 160 unsigned long __percpu *percpu_count; 161 162 rcu_read_lock_sched(); 163 164 if (__ref_is_percpu(ref, &percpu_count)) 165 this_cpu_add(*percpu_count, nr); 166 else 167 atomic_long_add(nr, &ref->count); 168 169 rcu_read_unlock_sched(); 170 } 171 172 /** 173 * percpu_ref_get - increment a percpu refcount 174 * @ref: percpu_ref to get 175 * 176 * Analagous to atomic_long_inc(). 177 * 178 * This function is safe to call as long as @ref is between init and exit. 179 */ 180 static inline void percpu_ref_get(struct percpu_ref *ref) 181 { 182 percpu_ref_get_many(ref, 1); 183 } 184 185 /** 186 * percpu_ref_tryget - try to increment a percpu refcount 187 * @ref: percpu_ref to try-get 188 * 189 * Increment a percpu refcount unless its count already reached zero. 190 * Returns %true on success; %false on failure. 191 * 192 * This function is safe to call as long as @ref is between init and exit. 193 */ 194 static inline bool percpu_ref_tryget(struct percpu_ref *ref) 195 { 196 unsigned long __percpu *percpu_count; 197 int ret; 198 199 rcu_read_lock_sched(); 200 201 if (__ref_is_percpu(ref, &percpu_count)) { 202 this_cpu_inc(*percpu_count); 203 ret = true; 204 } else { 205 ret = atomic_long_inc_not_zero(&ref->count); 206 } 207 208 rcu_read_unlock_sched(); 209 210 return ret; 211 } 212 213 /** 214 * percpu_ref_tryget_live - try to increment a live percpu refcount 215 * @ref: percpu_ref to try-get 216 * 217 * Increment a percpu refcount unless it has already been killed. Returns 218 * %true on success; %false on failure. 219 * 220 * Completion of percpu_ref_kill() in itself doesn't guarantee that this 221 * function will fail. For such guarantee, percpu_ref_kill_and_confirm() 222 * should be used. After the confirm_kill callback is invoked, it's 223 * guaranteed that no new reference will be given out by 224 * percpu_ref_tryget_live(). 225 * 226 * This function is safe to call as long as @ref is between init and exit. 227 */ 228 static inline bool percpu_ref_tryget_live(struct percpu_ref *ref) 229 { 230 unsigned long __percpu *percpu_count; 231 int ret = false; 232 233 rcu_read_lock_sched(); 234 235 if (__ref_is_percpu(ref, &percpu_count)) { 236 this_cpu_inc(*percpu_count); 237 ret = true; 238 } else if (!(ACCESS_ONCE(ref->percpu_count_ptr) & __PERCPU_REF_DEAD)) { 239 ret = atomic_long_inc_not_zero(&ref->count); 240 } 241 242 rcu_read_unlock_sched(); 243 244 return ret; 245 } 246 247 /** 248 * percpu_ref_put_many - decrement a percpu refcount 249 * @ref: percpu_ref to put 250 * @nr: number of references to put 251 * 252 * Decrement the refcount, and if 0, call the release function (which was passed 253 * to percpu_ref_init()) 254 * 255 * This function is safe to call as long as @ref is between init and exit. 256 */ 257 static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr) 258 { 259 unsigned long __percpu *percpu_count; 260 261 rcu_read_lock_sched(); 262 263 if (__ref_is_percpu(ref, &percpu_count)) 264 this_cpu_sub(*percpu_count, nr); 265 else if (unlikely(atomic_long_sub_and_test(nr, &ref->count))) 266 ref->release(ref); 267 268 rcu_read_unlock_sched(); 269 } 270 271 /** 272 * percpu_ref_put - decrement a percpu refcount 273 * @ref: percpu_ref to put 274 * 275 * Decrement the refcount, and if 0, call the release function (which was passed 276 * to percpu_ref_init()) 277 * 278 * This function is safe to call as long as @ref is between init and exit. 279 */ 280 static inline void percpu_ref_put(struct percpu_ref *ref) 281 { 282 percpu_ref_put_many(ref, 1); 283 } 284 285 /** 286 * percpu_ref_is_zero - test whether a percpu refcount reached zero 287 * @ref: percpu_ref to test 288 * 289 * Returns %true if @ref reached zero. 290 * 291 * This function is safe to call as long as @ref is between init and exit. 292 */ 293 static inline bool percpu_ref_is_zero(struct percpu_ref *ref) 294 { 295 unsigned long __percpu *percpu_count; 296 297 if (__ref_is_percpu(ref, &percpu_count)) 298 return false; 299 return !atomic_long_read(&ref->count); 300 } 301 302 #endif 303