percpu-refcount.h 9.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328
  1. /*
  2. * Percpu refcounts:
  3. * (C) 2012 Google, Inc.
  4. * Author: Kent Overstreet <koverstreet@google.com>
  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. #ifndef _LINUX_PERCPU_REFCOUNT_H
  45. #define _LINUX_PERCPU_REFCOUNT_H
  46. #include <linux/atomic.h>
  47. #include <linux/kernel.h>
  48. #include <linux/percpu.h>
  49. #include <linux/rcupdate.h>
  50. #include <linux/gfp.h>
  51. struct percpu_ref;
  52. typedef void (percpu_ref_func_t)(struct percpu_ref *);
  53. /* flags set in the lower bits of percpu_ref->percpu_count_ptr */
  54. enum {
  55. __PERCPU_REF_ATOMIC = 1LU << 0, /* operating in atomic mode */
  56. __PERCPU_REF_DEAD = 1LU << 1, /* (being) killed */
  57. __PERCPU_REF_ATOMIC_DEAD = __PERCPU_REF_ATOMIC | __PERCPU_REF_DEAD,
  58. __PERCPU_REF_FLAG_BITS = 2,
  59. };
  60. /* @flags for percpu_ref_init() */
  61. enum {
  62. /*
  63. * Start w/ ref == 1 in atomic mode. Can be switched to percpu
  64. * operation using percpu_ref_switch_to_percpu(). If initialized
  65. * with this flag, the ref will stay in atomic mode until
  66. * percpu_ref_switch_to_percpu() is invoked on it.
  67. */
  68. PERCPU_REF_INIT_ATOMIC = 1 << 0,
  69. /*
  70. * Start dead w/ ref == 0 in atomic mode. Must be revived with
  71. * percpu_ref_reinit() before used. Implies INIT_ATOMIC.
  72. */
  73. PERCPU_REF_INIT_DEAD = 1 << 1,
  74. };
  75. struct percpu_ref {
  76. atomic_long_t count;
  77. /*
  78. * The low bit of the pointer indicates whether the ref is in percpu
  79. * mode; if set, then get/put will manipulate the atomic_t.
  80. */
  81. unsigned long percpu_count_ptr;
  82. percpu_ref_func_t *release;
  83. percpu_ref_func_t *confirm_switch;
  84. bool force_atomic:1;
  85. struct rcu_head rcu;
  86. };
  87. int __must_check percpu_ref_init(struct percpu_ref *ref,
  88. percpu_ref_func_t *release, unsigned int flags,
  89. gfp_t gfp);
  90. void percpu_ref_exit(struct percpu_ref *ref);
  91. void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
  92. percpu_ref_func_t *confirm_switch);
  93. void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
  94. void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
  95. percpu_ref_func_t *confirm_kill);
  96. void percpu_ref_reinit(struct percpu_ref *ref);
  97. /**
  98. * percpu_ref_kill - drop the initial ref
  99. * @ref: percpu_ref to kill
  100. *
  101. * Must be used to drop the initial ref on a percpu refcount; must be called
  102. * precisely once before shutdown.
  103. *
  104. * Puts @ref in non percpu mode, then does a call_rcu() before gathering up the
  105. * percpu counters and dropping the initial ref.
  106. */
  107. static inline void percpu_ref_kill(struct percpu_ref *ref)
  108. {
  109. return percpu_ref_kill_and_confirm(ref, NULL);
  110. }
  111. /*
  112. * Internal helper. Don't use outside percpu-refcount proper. The
  113. * function doesn't return the pointer and let the caller test it for NULL
  114. * because doing so forces the compiler to generate two conditional
  115. * branches as it can't assume that @ref->percpu_count is not NULL.
  116. */
  117. static inline bool __ref_is_percpu(struct percpu_ref *ref,
  118. unsigned long __percpu **percpu_countp)
  119. {
  120. unsigned long percpu_ptr;
  121. /*
  122. * The value of @ref->percpu_count_ptr is tested for
  123. * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then
  124. * used as a pointer. If the compiler generates a separate fetch
  125. * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in
  126. * between contaminating the pointer value, meaning that
  127. * ACCESS_ONCE() is required when fetching it.
  128. *
  129. * Also, we need a data dependency barrier to be paired with
  130. * smp_store_release() in __percpu_ref_switch_to_percpu().
  131. *
  132. * Use lockless deref which contains both.
  133. */
  134. percpu_ptr = lockless_dereference(ref->percpu_count_ptr);
  135. /*
  136. * Theoretically, the following could test just ATOMIC; however,
  137. * then we'd have to mask off DEAD separately as DEAD may be
  138. * visible without ATOMIC if we race with percpu_ref_kill(). DEAD
  139. * implies ATOMIC anyway. Test them together.
  140. */
  141. if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD))
  142. return false;
  143. *percpu_countp = (unsigned long __percpu *)percpu_ptr;
  144. return true;
  145. }
  146. /**
  147. * percpu_ref_get_many - increment a percpu refcount
  148. * @ref: percpu_ref to get
  149. * @nr: number of references to get
  150. *
  151. * Analogous to atomic_long_add().
  152. *
  153. * This function is safe to call as long as @ref is between init and exit.
  154. */
  155. static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
  156. {
  157. unsigned long __percpu *percpu_count;
  158. rcu_read_lock_sched();
  159. if (__ref_is_percpu(ref, &percpu_count))
  160. this_cpu_add(*percpu_count, nr);
  161. else
  162. atomic_long_add(nr, &ref->count);
  163. rcu_read_unlock_sched();
  164. }
  165. /**
  166. * percpu_ref_get - increment a percpu refcount
  167. * @ref: percpu_ref to get
  168. *
  169. * Analagous to atomic_long_inc().
  170. *
  171. * This function is safe to call as long as @ref is between init and exit.
  172. */
  173. static inline void percpu_ref_get(struct percpu_ref *ref)
  174. {
  175. percpu_ref_get_many(ref, 1);
  176. }
  177. /**
  178. * percpu_ref_tryget - try to increment a percpu refcount
  179. * @ref: percpu_ref to try-get
  180. *
  181. * Increment a percpu refcount unless its count already reached zero.
  182. * Returns %true on success; %false on failure.
  183. *
  184. * This function is safe to call as long as @ref is between init and exit.
  185. */
  186. static inline bool percpu_ref_tryget(struct percpu_ref *ref)
  187. {
  188. unsigned long __percpu *percpu_count;
  189. bool ret;
  190. rcu_read_lock_sched();
  191. if (__ref_is_percpu(ref, &percpu_count)) {
  192. this_cpu_inc(*percpu_count);
  193. ret = true;
  194. } else {
  195. ret = atomic_long_inc_not_zero(&ref->count);
  196. }
  197. rcu_read_unlock_sched();
  198. return ret;
  199. }
  200. /**
  201. * percpu_ref_tryget_live - try to increment a live percpu refcount
  202. * @ref: percpu_ref to try-get
  203. *
  204. * Increment a percpu refcount unless it has already been killed. Returns
  205. * %true on success; %false on failure.
  206. *
  207. * Completion of percpu_ref_kill() in itself doesn't guarantee that this
  208. * function will fail. For such guarantee, percpu_ref_kill_and_confirm()
  209. * should be used. After the confirm_kill callback is invoked, it's
  210. * guaranteed that no new reference will be given out by
  211. * percpu_ref_tryget_live().
  212. *
  213. * This function is safe to call as long as @ref is between init and exit.
  214. */
  215. static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
  216. {
  217. unsigned long __percpu *percpu_count;
  218. bool ret = false;
  219. rcu_read_lock_sched();
  220. if (__ref_is_percpu(ref, &percpu_count)) {
  221. this_cpu_inc(*percpu_count);
  222. ret = true;
  223. } else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
  224. ret = atomic_long_inc_not_zero(&ref->count);
  225. }
  226. rcu_read_unlock_sched();
  227. return ret;
  228. }
  229. /**
  230. * percpu_ref_put_many - decrement a percpu refcount
  231. * @ref: percpu_ref to put
  232. * @nr: number of references to put
  233. *
  234. * Decrement the refcount, and if 0, call the release function (which was passed
  235. * to percpu_ref_init())
  236. *
  237. * This function is safe to call as long as @ref is between init and exit.
  238. */
  239. static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
  240. {
  241. unsigned long __percpu *percpu_count;
  242. rcu_read_lock_sched();
  243. if (__ref_is_percpu(ref, &percpu_count))
  244. this_cpu_sub(*percpu_count, nr);
  245. else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
  246. ref->release(ref);
  247. rcu_read_unlock_sched();
  248. }
  249. /**
  250. * percpu_ref_put - decrement a percpu refcount
  251. * @ref: percpu_ref to put
  252. *
  253. * Decrement the refcount, and if 0, call the release function (which was passed
  254. * to percpu_ref_init())
  255. *
  256. * This function is safe to call as long as @ref is between init and exit.
  257. */
  258. static inline void percpu_ref_put(struct percpu_ref *ref)
  259. {
  260. percpu_ref_put_many(ref, 1);
  261. }
  262. /**
  263. * percpu_ref_is_dying - test whether a percpu refcount is dying or dead
  264. * @ref: percpu_ref to test
  265. *
  266. * Returns %true if @ref is dying or dead.
  267. *
  268. * This function is safe to call as long as @ref is between init and exit
  269. * and the caller is responsible for synchronizing against state changes.
  270. */
  271. static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
  272. {
  273. return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
  274. }
  275. /**
  276. * percpu_ref_is_zero - test whether a percpu refcount reached zero
  277. * @ref: percpu_ref to test
  278. *
  279. * Returns %true if @ref reached zero.
  280. *
  281. * This function is safe to call as long as @ref is between init and exit.
  282. */
  283. static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
  284. {
  285. unsigned long __percpu *percpu_count;
  286. if (__ref_is_percpu(ref, &percpu_count))
  287. return false;
  288. return !atomic_long_read(&ref->count);
  289. }
  290. #endif