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 sychronize_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 struct percpu_ref { 67 atomic_long_t count; 68 /* 69 * The low bit of the pointer indicates whether the ref is in percpu 70 * mode; if set, then get/put will manipulate the atomic_t. 71 */ 72 unsigned long percpu_count_ptr; 73 percpu_ref_func_t *release; 74 percpu_ref_func_t *confirm_switch; 75 struct rcu_head rcu; 76 }; 77 78 int __must_check percpu_ref_init(struct percpu_ref *ref, 79 percpu_ref_func_t *release, gfp_t gfp); 80 void percpu_ref_exit(struct percpu_ref *ref); 81 void percpu_ref_switch_to_atomic(struct percpu_ref *ref, 82 percpu_ref_func_t *confirm_switch); 83 void percpu_ref_switch_to_percpu(struct percpu_ref *ref); 84 void percpu_ref_kill_and_confirm(struct percpu_ref *ref, 85 percpu_ref_func_t *confirm_kill); 86 void percpu_ref_reinit(struct percpu_ref *ref); 87 88 /** 89 * percpu_ref_kill - drop the initial ref 90 * @ref: percpu_ref to kill 91 * 92 * Must be used to drop the initial ref on a percpu refcount; must be called 93 * precisely once before shutdown. 94 * 95 * Puts @ref in non percpu mode, then does a call_rcu() before gathering up the 96 * percpu counters and dropping the initial ref. 97 */ 98 static inline void percpu_ref_kill(struct percpu_ref *ref) 99 { 100 return percpu_ref_kill_and_confirm(ref, NULL); 101 } 102 103 /* 104 * Internal helper. Don't use outside percpu-refcount proper. The 105 * function doesn't return the pointer and let the caller test it for NULL 106 * because doing so forces the compiler to generate two conditional 107 * branches as it can't assume that @ref->percpu_count is not NULL. 108 */ 109 static inline bool __ref_is_percpu(struct percpu_ref *ref, 110 unsigned long __percpu **percpu_countp) 111 { 112 unsigned long percpu_ptr = ACCESS_ONCE(ref->percpu_count_ptr); 113 114 /* paired with smp_store_release() in percpu_ref_reinit() */ 115 smp_read_barrier_depends(); 116 117 if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC)) 118 return false; 119 120 *percpu_countp = (unsigned long __percpu *)percpu_ptr; 121 return true; 122 } 123 124 /** 125 * percpu_ref_get - increment a percpu refcount 126 * @ref: percpu_ref to get 127 * 128 * Analagous to atomic_long_inc(). 129 * 130 * This function is safe to call as long as @ref is between init and exit. 131 */ 132 static inline void percpu_ref_get(struct percpu_ref *ref) 133 { 134 unsigned long __percpu *percpu_count; 135 136 rcu_read_lock_sched(); 137 138 if (__ref_is_percpu(ref, &percpu_count)) 139 this_cpu_inc(*percpu_count); 140 else 141 atomic_long_inc(&ref->count); 142 143 rcu_read_unlock_sched(); 144 } 145 146 /** 147 * percpu_ref_tryget - try to increment a percpu refcount 148 * @ref: percpu_ref to try-get 149 * 150 * Increment a percpu refcount unless its count already reached zero. 151 * Returns %true on success; %false on failure. 152 * 153 * This function is safe to call as long as @ref is between init and exit. 154 */ 155 static inline bool percpu_ref_tryget(struct percpu_ref *ref) 156 { 157 unsigned long __percpu *percpu_count; 158 int ret; 159 160 rcu_read_lock_sched(); 161 162 if (__ref_is_percpu(ref, &percpu_count)) { 163 this_cpu_inc(*percpu_count); 164 ret = true; 165 } else { 166 ret = atomic_long_inc_not_zero(&ref->count); 167 } 168 169 rcu_read_unlock_sched(); 170 171 return ret; 172 } 173 174 /** 175 * percpu_ref_tryget_live - try to increment a live percpu refcount 176 * @ref: percpu_ref to try-get 177 * 178 * Increment a percpu refcount unless it has already been killed. Returns 179 * %true on success; %false on failure. 180 * 181 * Completion of percpu_ref_kill() in itself doesn't guarantee that this 182 * function will fail. For such guarantee, percpu_ref_kill_and_confirm() 183 * should be used. After the confirm_kill callback is invoked, it's 184 * guaranteed that no new reference will be given out by 185 * percpu_ref_tryget_live(). 186 * 187 * This function is safe to call as long as @ref is between init and exit. 188 */ 189 static inline bool percpu_ref_tryget_live(struct percpu_ref *ref) 190 { 191 unsigned long __percpu *percpu_count; 192 int ret = false; 193 194 rcu_read_lock_sched(); 195 196 if (__ref_is_percpu(ref, &percpu_count)) { 197 this_cpu_inc(*percpu_count); 198 ret = true; 199 } else if (!(ACCESS_ONCE(ref->percpu_count_ptr) & __PERCPU_REF_DEAD)) { 200 ret = atomic_long_inc_not_zero(&ref->count); 201 } 202 203 rcu_read_unlock_sched(); 204 205 return ret; 206 } 207 208 /** 209 * percpu_ref_put - decrement a percpu refcount 210 * @ref: percpu_ref to put 211 * 212 * Decrement the refcount, and if 0, call the release function (which was passed 213 * to percpu_ref_init()) 214 * 215 * This function is safe to call as long as @ref is between init and exit. 216 */ 217 static inline void percpu_ref_put(struct percpu_ref *ref) 218 { 219 unsigned long __percpu *percpu_count; 220 221 rcu_read_lock_sched(); 222 223 if (__ref_is_percpu(ref, &percpu_count)) 224 this_cpu_dec(*percpu_count); 225 else if (unlikely(atomic_long_dec_and_test(&ref->count))) 226 ref->release(ref); 227 228 rcu_read_unlock_sched(); 229 } 230 231 /** 232 * percpu_ref_is_zero - test whether a percpu refcount reached zero 233 * @ref: percpu_ref to test 234 * 235 * Returns %true if @ref reached zero. 236 * 237 * This function is safe to call as long as @ref is between init and exit. 238 */ 239 static inline bool percpu_ref_is_zero(struct percpu_ref *ref) 240 { 241 unsigned long __percpu *percpu_count; 242 243 if (__ref_is_percpu(ref, &percpu_count)) 244 return false; 245 return !atomic_long_read(&ref->count); 246 } 247 248 #endif 249