blob: fe03c6d527611937e196395f83f7348da86f8f06 [file] [log] [blame]
Kent Overstreet215e2622013-05-31 15:26:45 -07001#define pr_fmt(fmt) "%s: " fmt "\n", __func__
2
3#include <linux/kernel.h>
Tejun Heo490c79a2014-09-24 13:31:49 -04004#include <linux/sched.h>
5#include <linux/wait.h>
Kent Overstreet215e2622013-05-31 15:26:45 -07006#include <linux/percpu-refcount.h>
7
8/*
9 * Initially, a percpu refcount is just a set of percpu counters. Initially, we
10 * don't try to detect the ref hitting 0 - which means that get/put can just
11 * increment or decrement the local counter. Note that the counter on a
12 * particular cpu can (and will) wrap - this is fine, when we go to shutdown the
13 * percpu counters will all sum to the correct value
14 *
Bogdan Sikorabdb428c2015-12-27 14:58:23 +010015 * (More precisely: because modular arithmetic is commutative the sum of all the
Tejun Heoeecc16b2014-09-24 13:31:48 -040016 * percpu_count vars will be equal to what it would have been if all the gets
17 * and puts were done to a single integer, even if some of the percpu integers
Kent Overstreet215e2622013-05-31 15:26:45 -070018 * overflow or underflow).
19 *
20 * The real trick to implementing percpu refcounts is shutdown. We can't detect
21 * the ref hitting 0 on every put - this would require global synchronization
22 * and defeat the whole purpose of using percpu refs.
23 *
24 * What we do is require the user to keep track of the initial refcount; we know
25 * the ref can't hit 0 before the user drops the initial ref, so as long as we
26 * convert to non percpu mode before the initial ref is dropped everything
27 * works.
28 *
29 * Converting to non percpu mode is done with some RCUish stuff in
Tejun Heoe6253052014-09-20 01:27:25 -040030 * percpu_ref_kill. Additionally, we need a bias value so that the
31 * atomic_long_t can't hit 0 before we've added up all the percpu refs.
Kent Overstreet215e2622013-05-31 15:26:45 -070032 */
33
Tejun Heoeecc16b2014-09-24 13:31:48 -040034#define PERCPU_COUNT_BIAS (1LU << (BITS_PER_LONG - 1))
Kent Overstreet215e2622013-05-31 15:26:45 -070035
Tejun Heo33e465c2015-09-29 17:47:20 -040036static DEFINE_SPINLOCK(percpu_ref_switch_lock);
Tejun Heo490c79a2014-09-24 13:31:49 -040037static DECLARE_WAIT_QUEUE_HEAD(percpu_ref_switch_waitq);
38
Tejun Heoeecc16b2014-09-24 13:31:48 -040039static unsigned long __percpu *percpu_count_ptr(struct percpu_ref *ref)
Tejun Heoeae79752014-06-28 08:10:13 -040040{
Tejun Heoeecc16b2014-09-24 13:31:48 -040041 return (unsigned long __percpu *)
Tejun Heo27344a92014-09-24 13:31:49 -040042 (ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC_DEAD);
Tejun Heoeae79752014-06-28 08:10:13 -040043}
44
Kent Overstreet215e2622013-05-31 15:26:45 -070045/**
46 * percpu_ref_init - initialize a percpu refcount
Tejun Heoac899062013-06-12 20:43:06 -070047 * @ref: percpu_ref to initialize
48 * @release: function which will be called when refcount hits 0
Tejun Heo2aad2a82014-09-24 13:31:50 -040049 * @flags: PERCPU_REF_INIT_* flags
Tejun Heoa34375e2014-09-08 09:51:30 +090050 * @gfp: allocation mask to use
Kent Overstreet215e2622013-05-31 15:26:45 -070051 *
Tejun Heo2aad2a82014-09-24 13:31:50 -040052 * Initializes @ref. If @flags is zero, @ref starts in percpu mode with a
53 * refcount of 1; analagous to atomic_long_set(ref, 1). See the
54 * definitions of PERCPU_REF_INIT_* flags for flag behaviors.
Kent Overstreet215e2622013-05-31 15:26:45 -070055 *
56 * Note that @release must not sleep - it may potentially be called from RCU
57 * callback context by percpu_ref_kill().
58 */
Tejun Heoa34375e2014-09-08 09:51:30 +090059int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
Tejun Heo2aad2a82014-09-24 13:31:50 -040060 unsigned int flags, gfp_t gfp)
Kent Overstreet215e2622013-05-31 15:26:45 -070061{
Tejun Heo27344a92014-09-24 13:31:49 -040062 size_t align = max_t(size_t, 1 << __PERCPU_REF_FLAG_BITS,
63 __alignof__(unsigned long));
Tejun Heo2aad2a82014-09-24 13:31:50 -040064 unsigned long start_count = 0;
Kent Overstreet215e2622013-05-31 15:26:45 -070065
Tejun Heo27344a92014-09-24 13:31:49 -040066 ref->percpu_count_ptr = (unsigned long)
67 __alloc_percpu_gfp(sizeof(unsigned long), align, gfp);
Tejun Heoeecc16b2014-09-24 13:31:48 -040068 if (!ref->percpu_count_ptr)
Kent Overstreet215e2622013-05-31 15:26:45 -070069 return -ENOMEM;
70
Tejun Heo1cae13e2014-09-24 13:31:50 -040071 ref->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
72
Tejun Heo2aad2a82014-09-24 13:31:50 -040073 if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD))
74 ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
75 else
76 start_count += PERCPU_COUNT_BIAS;
77
78 if (flags & PERCPU_REF_INIT_DEAD)
79 ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
80 else
81 start_count++;
82
83 atomic_long_set(&ref->count, start_count);
84
Kent Overstreet215e2622013-05-31 15:26:45 -070085 ref->release = release;
Roman Pena67823c2016-08-11 19:27:09 +020086 ref->confirm_switch = NULL;
Kent Overstreet215e2622013-05-31 15:26:45 -070087 return 0;
88}
Matias Bjorling5e9dd372013-10-16 13:47:01 -070089EXPORT_SYMBOL_GPL(percpu_ref_init);
Kent Overstreet215e2622013-05-31 15:26:45 -070090
Tejun Heobc497bd2013-06-12 20:52:35 -070091/**
Tejun Heo9a1049d2014-06-28 08:10:14 -040092 * percpu_ref_exit - undo percpu_ref_init()
93 * @ref: percpu_ref to exit
Tejun Heobc497bd2013-06-12 20:52:35 -070094 *
Tejun Heo9a1049d2014-06-28 08:10:14 -040095 * This function exits @ref. The caller is responsible for ensuring that
96 * @ref is no longer in active use. The usual places to invoke this
97 * function from are the @ref->release() callback or in init failure path
98 * where percpu_ref_init() succeeded but other parts of the initialization
99 * of the embedding object failed.
Tejun Heobc497bd2013-06-12 20:52:35 -0700100 */
Tejun Heo9a1049d2014-06-28 08:10:14 -0400101void percpu_ref_exit(struct percpu_ref *ref)
Tejun Heobc497bd2013-06-12 20:52:35 -0700102{
Tejun Heoeecc16b2014-09-24 13:31:48 -0400103 unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
Tejun Heobc497bd2013-06-12 20:52:35 -0700104
Tejun Heoeecc16b2014-09-24 13:31:48 -0400105 if (percpu_count) {
Roman Pena67823c2016-08-11 19:27:09 +0200106 /* non-NULL confirm_switch indicates switching in progress */
107 WARN_ON_ONCE(ref->confirm_switch);
Tejun Heoeecc16b2014-09-24 13:31:48 -0400108 free_percpu(percpu_count);
Tejun Heo27344a92014-09-24 13:31:49 -0400109 ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
Tejun Heobc497bd2013-06-12 20:52:35 -0700110 }
111}
Tejun Heo9a1049d2014-06-28 08:10:14 -0400112EXPORT_SYMBOL_GPL(percpu_ref_exit);
Tejun Heobc497bd2013-06-12 20:52:35 -0700113
Tejun Heo490c79a2014-09-24 13:31:49 -0400114static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
115{
116 struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
117
118 ref->confirm_switch(ref);
119 ref->confirm_switch = NULL;
120 wake_up_all(&percpu_ref_switch_waitq);
121
122 /* drop ref from percpu_ref_switch_to_atomic() */
123 percpu_ref_put(ref);
124}
125
126static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
Kent Overstreet215e2622013-05-31 15:26:45 -0700127{
128 struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
Tejun Heoeecc16b2014-09-24 13:31:48 -0400129 unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
Tejun Heoe6253052014-09-20 01:27:25 -0400130 unsigned long count = 0;
Kent Overstreet215e2622013-05-31 15:26:45 -0700131 int cpu;
132
Kent Overstreet215e2622013-05-31 15:26:45 -0700133 for_each_possible_cpu(cpu)
Tejun Heoeecc16b2014-09-24 13:31:48 -0400134 count += *per_cpu_ptr(percpu_count, cpu);
Kent Overstreet215e2622013-05-31 15:26:45 -0700135
Tejun Heoeecc16b2014-09-24 13:31:48 -0400136 pr_debug("global %ld percpu %ld",
Tejun Heoe6253052014-09-20 01:27:25 -0400137 atomic_long_read(&ref->count), (long)count);
Kent Overstreet215e2622013-05-31 15:26:45 -0700138
139 /*
140 * It's crucial that we sum the percpu counters _before_ adding the sum
141 * to &ref->count; since gets could be happening on one cpu while puts
142 * happen on another, adding a single cpu's count could cause
143 * @ref->count to hit 0 before we've got a consistent value - but the
144 * sum of all the counts will be consistent and correct.
145 *
146 * Subtracting the bias value then has to happen _after_ adding count to
147 * &ref->count; we need the bias value to prevent &ref->count from
148 * reaching 0 before we add the percpu counts. But doing it at the same
149 * time is equivalent and saves us atomic operations:
150 */
Tejun Heoeecc16b2014-09-24 13:31:48 -0400151 atomic_long_add((long)count - PERCPU_COUNT_BIAS, &ref->count);
Kent Overstreet215e2622013-05-31 15:26:45 -0700152
Tejun Heoe6253052014-09-20 01:27:25 -0400153 WARN_ONCE(atomic_long_read(&ref->count) <= 0,
Tejun Heo490c79a2014-09-24 13:31:49 -0400154 "percpu ref (%pf) <= 0 (%ld) after switching to atomic",
Tejun Heoe6253052014-09-20 01:27:25 -0400155 ref->release, atomic_long_read(&ref->count));
Kent Overstreet687b0ad2014-01-06 13:13:26 -0800156
Tejun Heo490c79a2014-09-24 13:31:49 -0400157 /* @ref is viewed as dead on all CPUs, send out switch confirmation */
158 percpu_ref_call_confirm_rcu(rcu);
159}
Tejun Heodbece3a2013-06-13 19:23:53 -0700160
Tejun Heo490c79a2014-09-24 13:31:49 -0400161static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref)
162{
163}
164
165static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
166 percpu_ref_func_t *confirm_switch)
167{
Tejun Heob2302c72015-09-29 17:47:17 -0400168 if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) {
Tejun Heo18808352015-09-29 17:47:18 -0400169 if (confirm_switch)
Tejun Heob2302c72015-09-29 17:47:17 -0400170 confirm_switch(ref);
Tejun Heob2302c72015-09-29 17:47:17 -0400171 return;
Tejun Heo490c79a2014-09-24 13:31:49 -0400172 }
Tejun Heob2302c72015-09-29 17:47:17 -0400173
174 /* switching from percpu to atomic */
175 ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
176
177 /*
178 * Non-NULL ->confirm_switch is used to indicate that switching is
179 * in progress. Use noop one if unspecified.
180 */
Tejun Heob2302c72015-09-29 17:47:17 -0400181 ref->confirm_switch = confirm_switch ?: percpu_ref_noop_confirm_switch;
182
183 percpu_ref_get(ref); /* put after confirmation */
184 call_rcu_sched(&ref->rcu, percpu_ref_switch_to_atomic_rcu);
185}
186
187static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
188{
189 unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
190 int cpu;
191
192 BUG_ON(!percpu_count);
193
194 if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
195 return;
196
Tejun Heob2302c72015-09-29 17:47:17 -0400197 atomic_long_add(PERCPU_COUNT_BIAS, &ref->count);
198
199 /*
200 * Restore per-cpu operation. smp_store_release() is paired with
201 * smp_read_barrier_depends() in __ref_is_percpu() and guarantees
202 * that the zeroing is visible to all percpu accesses which can see
203 * the following __PERCPU_REF_ATOMIC clearing.
204 */
205 for_each_possible_cpu(cpu)
206 *per_cpu_ptr(percpu_count, cpu) = 0;
207
208 smp_store_release(&ref->percpu_count_ptr,
209 ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);
Kent Overstreet215e2622013-05-31 15:26:45 -0700210}
211
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400212static void __percpu_ref_switch_mode(struct percpu_ref *ref,
213 percpu_ref_func_t *confirm_switch)
214{
Tejun Heo33e465c2015-09-29 17:47:20 -0400215 lockdep_assert_held(&percpu_ref_switch_lock);
216
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400217 /*
218 * If the previous ATOMIC switching hasn't finished yet, wait for
219 * its completion. If the caller ensures that ATOMIC switching
220 * isn't in progress, this function can be called from any context.
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400221 */
Tejun Heo33e465c2015-09-29 17:47:20 -0400222 wait_event_lock_irq(percpu_ref_switch_waitq, !ref->confirm_switch,
223 percpu_ref_switch_lock);
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400224
225 if (ref->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD))
226 __percpu_ref_switch_to_atomic(ref, confirm_switch);
227 else
228 __percpu_ref_switch_to_percpu(ref);
229}
230
Kent Overstreet215e2622013-05-31 15:26:45 -0700231/**
Tejun Heo490c79a2014-09-24 13:31:49 -0400232 * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode
233 * @ref: percpu_ref to switch to atomic mode
234 * @confirm_switch: optional confirmation callback
Kent Overstreet215e2622013-05-31 15:26:45 -0700235 *
Tejun Heo490c79a2014-09-24 13:31:49 -0400236 * There's no reason to use this function for the usual reference counting.
237 * Use percpu_ref_kill[_and_confirm]().
Kent Overstreet215e2622013-05-31 15:26:45 -0700238 *
Tejun Heo490c79a2014-09-24 13:31:49 -0400239 * Schedule switching of @ref to atomic mode. All its percpu counts will
240 * be collected to the main atomic counter. On completion, when all CPUs
241 * are guaraneed to be in atomic mode, @confirm_switch, which may not
242 * block, is invoked. This function may be invoked concurrently with all
243 * the get/put operations and can safely be mixed with kill and reinit
Tejun Heo1cae13e2014-09-24 13:31:50 -0400244 * operations. Note that @ref will stay in atomic mode across kill/reinit
245 * cycles until percpu_ref_switch_to_percpu() is called.
Tejun Heo490c79a2014-09-24 13:31:49 -0400246 *
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400247 * This function may block if @ref is in the process of switching to atomic
248 * mode. If the caller ensures that @ref is not in the process of
249 * switching to atomic mode, this function can be called from any context.
Kent Overstreet215e2622013-05-31 15:26:45 -0700250 */
Tejun Heo490c79a2014-09-24 13:31:49 -0400251void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
252 percpu_ref_func_t *confirm_switch)
Kent Overstreet215e2622013-05-31 15:26:45 -0700253{
Tejun Heo33e465c2015-09-29 17:47:20 -0400254 unsigned long flags;
255
256 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
257
Tejun Heo1cae13e2014-09-24 13:31:50 -0400258 ref->force_atomic = true;
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400259 __percpu_ref_switch_mode(ref, confirm_switch);
Tejun Heo33e465c2015-09-29 17:47:20 -0400260
261 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
Kent Overstreet215e2622013-05-31 15:26:45 -0700262}
NeilBrown210f7cd2017-03-15 14:05:14 +1100263EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
264
265/**
266 * percpu_ref_switch_to_atomic_sync - switch a percpu_ref to atomic mode
267 * @ref: percpu_ref to switch to atomic mode
268 *
269 * Schedule switching the ref to atomic mode, and wait for the
270 * switch to complete. Caller must ensure that no other thread
271 * will switch back to percpu mode.
272 */
273void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
274{
275 percpu_ref_switch_to_atomic(ref, NULL);
276 wait_event(percpu_ref_switch_waitq, !ref->confirm_switch);
277}
278EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);
Tejun Heoa2237372014-09-24 13:31:48 -0400279
Tejun Heof47ad452014-09-24 13:31:49 -0400280/**
281 * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode
282 * @ref: percpu_ref to switch to percpu mode
283 *
284 * There's no reason to use this function for the usual reference counting.
285 * To re-use an expired ref, use percpu_ref_reinit().
286 *
287 * Switch @ref to percpu mode. This function may be invoked concurrently
288 * with all the get/put operations and can safely be mixed with kill and
Tejun Heo1cae13e2014-09-24 13:31:50 -0400289 * reinit operations. This function reverses the sticky atomic state set
290 * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic(). If @ref is
291 * dying or dead, the actual switching takes place on the following
292 * percpu_ref_reinit().
Tejun Heof47ad452014-09-24 13:31:49 -0400293 *
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400294 * This function may block if @ref is in the process of switching to atomic
295 * mode. If the caller ensures that @ref is not in the process of
296 * switching to atomic mode, this function can be called from any context.
Tejun Heof47ad452014-09-24 13:31:49 -0400297 */
298void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
299{
Tejun Heo33e465c2015-09-29 17:47:20 -0400300 unsigned long flags;
301
302 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
303
Tejun Heo1cae13e2014-09-24 13:31:50 -0400304 ref->force_atomic = false;
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400305 __percpu_ref_switch_mode(ref, NULL);
Tejun Heo33e465c2015-09-29 17:47:20 -0400306
307 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
Tejun Heof47ad452014-09-24 13:31:49 -0400308}
NeilBrown210f7cd2017-03-15 14:05:14 +1100309EXPORT_SYMBOL_GPL(percpu_ref_switch_to_percpu);
Tejun Heo490c79a2014-09-24 13:31:49 -0400310
311/**
312 * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
313 * @ref: percpu_ref to kill
314 * @confirm_kill: optional confirmation callback
315 *
316 * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
317 * @confirm_kill is not NULL. @confirm_kill, which may not block, will be
318 * called after @ref is seen as dead from all CPUs at which point all
319 * further invocations of percpu_ref_tryget_live() will fail. See
320 * percpu_ref_tryget_live() for details.
321 *
322 * This function normally doesn't block and can be called from any context
Tejun Heof47ad452014-09-24 13:31:49 -0400323 * but it may block if @confirm_kill is specified and @ref is in the
Tejun Heoa2f56302015-09-29 17:47:16 -0400324 * process of switching to atomic mode by percpu_ref_switch_to_atomic().
Tejun Heo490c79a2014-09-24 13:31:49 -0400325 */
326void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
327 percpu_ref_func_t *confirm_kill)
328{
Tejun Heo33e465c2015-09-29 17:47:20 -0400329 unsigned long flags;
330
331 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
332
Tejun Heo490c79a2014-09-24 13:31:49 -0400333 WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD,
334 "%s called more than once on %pf!", __func__, ref->release);
335
336 ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400337 __percpu_ref_switch_mode(ref, confirm_kill);
Tejun Heo490c79a2014-09-24 13:31:49 -0400338 percpu_ref_put(ref);
Tejun Heo33e465c2015-09-29 17:47:20 -0400339
340 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
Tejun Heo490c79a2014-09-24 13:31:49 -0400341}
342EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
Tejun Heof47ad452014-09-24 13:31:49 -0400343
344/**
345 * percpu_ref_reinit - re-initialize a percpu refcount
346 * @ref: perpcu_ref to re-initialize
347 *
348 * Re-initialize @ref so that it's in the same state as when it finished
Tejun Heo1cae13e2014-09-24 13:31:50 -0400349 * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD. @ref must have been
350 * initialized successfully and reached 0 but not exited.
Tejun Heof47ad452014-09-24 13:31:49 -0400351 *
352 * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
353 * this function is in progress.
354 */
355void percpu_ref_reinit(struct percpu_ref *ref)
356{
Tejun Heo33e465c2015-09-29 17:47:20 -0400357 unsigned long flags;
358
359 spin_lock_irqsave(&percpu_ref_switch_lock, flags);
360
Tejun Heof47ad452014-09-24 13:31:49 -0400361 WARN_ON_ONCE(!percpu_ref_is_zero(ref));
362
363 ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD;
364 percpu_ref_get(ref);
Tejun Heo3f49bdd2015-09-29 17:47:19 -0400365 __percpu_ref_switch_mode(ref, NULL);
Tejun Heo33e465c2015-09-29 17:47:20 -0400366
367 spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
Tejun Heof47ad452014-09-24 13:31:49 -0400368}
369EXPORT_SYMBOL_GPL(percpu_ref_reinit);