percpu-refcount.h 10 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335
  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. #ifndef _LINUX_PERCPU_REFCOUNT_H
  50. #define _LINUX_PERCPU_REFCOUNT_H
  51. #include <linux/atomic.h>
  52. #include <linux/kernel.h>
  53. #include <linux/percpu.h>
  54. #include <linux/rcupdate.h>
  55. #include <linux/gfp.h>
  56. struct percpu_ref;
  57. typedef void (percpu_ref_func_t)(struct percpu_ref *);
  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. __PERCPU_REF_FLAG_BITS = 2,
  64. };
  65. /* @flags for percpu_ref_init() */
  66. enum {
  67. /*
  68. * Start w/ ref == 1 in atomic mode. Can be switched to percpu
  69. * operation using percpu_ref_switch_to_percpu(). If initialized
  70. * with this flag, the ref will stay in atomic mode until
  71. * percpu_ref_switch_to_percpu() is invoked on it.
  72. */
  73. PERCPU_REF_INIT_ATOMIC = 1 << 0,
  74. /*
  75. * Start dead w/ ref == 0 in atomic mode. Must be revived with
  76. * percpu_ref_reinit() before used. Implies INIT_ATOMIC.
  77. */
  78. PERCPU_REF_INIT_DEAD = 1 << 1,
  79. };
  80. struct percpu_ref {
  81. atomic_long_t count;
  82. /*
  83. * The low bit of the pointer indicates whether the ref is in percpu
  84. * mode; if set, then get/put will manipulate the atomic_t.
  85. */
  86. unsigned long percpu_count_ptr;
  87. percpu_ref_func_t *release;
  88. percpu_ref_func_t *confirm_switch;
  89. bool force_atomic:1;
  90. struct rcu_head rcu;
  91. };
  92. int __must_check percpu_ref_init(struct percpu_ref *ref,
  93. percpu_ref_func_t *release, unsigned int flags,
  94. gfp_t gfp);
  95. void percpu_ref_exit(struct percpu_ref *ref);
  96. void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
  97. percpu_ref_func_t *confirm_switch);
  98. void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref);
  99. void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
  100. void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
  101. percpu_ref_func_t *confirm_kill);
  102. void percpu_ref_reinit(struct percpu_ref *ref);
  103. /**
  104. * percpu_ref_kill - drop the initial ref
  105. * @ref: percpu_ref to kill
  106. *
  107. * Must be used to drop the initial ref on a percpu refcount; must be called
  108. * precisely once before shutdown.
  109. *
  110. * Switches @ref into atomic mode before gathering up the percpu counters
  111. * and dropping the initial ref.
  112. *
  113. * There are no implied RCU grace periods between kill and release.
  114. */
  115. static inline void percpu_ref_kill(struct percpu_ref *ref)
  116. {
  117. percpu_ref_kill_and_confirm(ref, NULL);
  118. }
  119. /*
  120. * Internal helper. Don't use outside percpu-refcount proper. The
  121. * function doesn't return the pointer and let the caller test it for NULL
  122. * because doing so forces the compiler to generate two conditional
  123. * branches as it can't assume that @ref->percpu_count is not NULL.
  124. */
  125. static inline bool __ref_is_percpu(struct percpu_ref *ref,
  126. unsigned long __percpu **percpu_countp)
  127. {
  128. unsigned long percpu_ptr;
  129. /*
  130. * The value of @ref->percpu_count_ptr is tested for
  131. * !__PERCPU_REF_ATOMIC, which may be set asynchronously, and then
  132. * used as a pointer. If the compiler generates a separate fetch
  133. * when using it as a pointer, __PERCPU_REF_ATOMIC may be set in
  134. * between contaminating the pointer value, meaning that
  135. * READ_ONCE() is required when fetching it.
  136. *
  137. * The smp_read_barrier_depends() implied by READ_ONCE() pairs
  138. * with smp_store_release() in __percpu_ref_switch_to_percpu().
  139. */
  140. percpu_ptr = READ_ONCE(ref->percpu_count_ptr);
  141. /*
  142. * Theoretically, the following could test just ATOMIC; however,
  143. * then we'd have to mask off DEAD separately as DEAD may be
  144. * visible without ATOMIC if we race with percpu_ref_kill(). DEAD
  145. * implies ATOMIC anyway. Test them together.
  146. */
  147. if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD))
  148. return false;
  149. *percpu_countp = (unsigned long __percpu *)percpu_ptr;
  150. return true;
  151. }
  152. /**
  153. * percpu_ref_get_many - increment a percpu refcount
  154. * @ref: percpu_ref to get
  155. * @nr: number of references to get
  156. *
  157. * Analogous to atomic_long_add().
  158. *
  159. * This function is safe to call as long as @ref is between init and exit.
  160. */
  161. static inline void percpu_ref_get_many(struct percpu_ref *ref, unsigned long nr)
  162. {
  163. unsigned long __percpu *percpu_count;
  164. rcu_read_lock_sched();
  165. if (__ref_is_percpu(ref, &percpu_count))
  166. this_cpu_add(*percpu_count, nr);
  167. else
  168. atomic_long_add(nr, &ref->count);
  169. rcu_read_unlock_sched();
  170. }
  171. /**
  172. * percpu_ref_get - increment a percpu refcount
  173. * @ref: percpu_ref to get
  174. *
  175. * Analagous to atomic_long_inc().
  176. *
  177. * This function is safe to call as long as @ref is between init and exit.
  178. */
  179. static inline void percpu_ref_get(struct percpu_ref *ref)
  180. {
  181. percpu_ref_get_many(ref, 1);
  182. }
  183. /**
  184. * percpu_ref_tryget - try to increment a percpu refcount
  185. * @ref: percpu_ref to try-get
  186. *
  187. * Increment a percpu refcount unless its count already reached zero.
  188. * Returns %true on success; %false on failure.
  189. *
  190. * This function is safe to call as long as @ref is between init and exit.
  191. */
  192. static inline bool percpu_ref_tryget(struct percpu_ref *ref)
  193. {
  194. unsigned long __percpu *percpu_count;
  195. bool ret;
  196. rcu_read_lock_sched();
  197. if (__ref_is_percpu(ref, &percpu_count)) {
  198. this_cpu_inc(*percpu_count);
  199. ret = true;
  200. } else {
  201. ret = atomic_long_inc_not_zero(&ref->count);
  202. }
  203. rcu_read_unlock_sched();
  204. return ret;
  205. }
  206. /**
  207. * percpu_ref_tryget_live - try to increment a live percpu refcount
  208. * @ref: percpu_ref to try-get
  209. *
  210. * Increment a percpu refcount unless it has already been killed. Returns
  211. * %true on success; %false on failure.
  212. *
  213. * Completion of percpu_ref_kill() in itself doesn't guarantee that this
  214. * function will fail. For such guarantee, percpu_ref_kill_and_confirm()
  215. * should be used. After the confirm_kill callback is invoked, it's
  216. * guaranteed that no new reference will be given out by
  217. * percpu_ref_tryget_live().
  218. *
  219. * This function is safe to call as long as @ref is between init and exit.
  220. */
  221. static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
  222. {
  223. unsigned long __percpu *percpu_count;
  224. bool ret = false;
  225. rcu_read_lock_sched();
  226. if (__ref_is_percpu(ref, &percpu_count)) {
  227. this_cpu_inc(*percpu_count);
  228. ret = true;
  229. } else if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)) {
  230. ret = atomic_long_inc_not_zero(&ref->count);
  231. }
  232. rcu_read_unlock_sched();
  233. return ret;
  234. }
  235. /**
  236. * percpu_ref_put_many - decrement a percpu refcount
  237. * @ref: percpu_ref to put
  238. * @nr: number of references to put
  239. *
  240. * Decrement the refcount, and if 0, call the release function (which was passed
  241. * to percpu_ref_init())
  242. *
  243. * This function is safe to call as long as @ref is between init and exit.
  244. */
  245. static inline void percpu_ref_put_many(struct percpu_ref *ref, unsigned long nr)
  246. {
  247. unsigned long __percpu *percpu_count;
  248. rcu_read_lock_sched();
  249. if (__ref_is_percpu(ref, &percpu_count))
  250. this_cpu_sub(*percpu_count, nr);
  251. else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
  252. ref->release(ref);
  253. rcu_read_unlock_sched();
  254. }
  255. /**
  256. * percpu_ref_put - decrement a percpu refcount
  257. * @ref: percpu_ref to put
  258. *
  259. * Decrement the refcount, and if 0, call the release function (which was passed
  260. * to percpu_ref_init())
  261. *
  262. * This function is safe to call as long as @ref is between init and exit.
  263. */
  264. static inline void percpu_ref_put(struct percpu_ref *ref)
  265. {
  266. percpu_ref_put_many(ref, 1);
  267. }
  268. /**
  269. * percpu_ref_is_dying - test whether a percpu refcount is dying or dead
  270. * @ref: percpu_ref to test
  271. *
  272. * Returns %true if @ref is dying or dead.
  273. *
  274. * This function is safe to call as long as @ref is between init and exit
  275. * and the caller is responsible for synchronizing against state changes.
  276. */
  277. static inline bool percpu_ref_is_dying(struct percpu_ref *ref)
  278. {
  279. return ref->percpu_count_ptr & __PERCPU_REF_DEAD;
  280. }
  281. /**
  282. * percpu_ref_is_zero - test whether a percpu refcount reached zero
  283. * @ref: percpu_ref to test
  284. *
  285. * Returns %true if @ref reached zero.
  286. *
  287. * This function is safe to call as long as @ref is between init and exit.
  288. */
  289. static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
  290. {
  291. unsigned long __percpu *percpu_count;
  292. if (__ref_is_percpu(ref, &percpu_count))
  293. return false;
  294. return !atomic_long_read(&ref->count);
  295. }
  296. #endif