• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 #define pr_fmt(fmt) "%s: " fmt, __func__
3 
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/wait.h>
7 #include <linux/slab.h>
8 #include <linux/percpu-refcount.h>
9 
10 /*
11  * Initially, a percpu refcount is just a set of percpu counters. Initially, we
12  * don't try to detect the ref hitting 0 - which means that get/put can just
13  * increment or decrement the local counter. Note that the counter on a
14  * particular cpu can (and will) wrap - this is fine, when we go to shutdown the
15  * percpu counters will all sum to the correct value
16  *
17  * (More precisely: because modular arithmetic is commutative the sum of all the
18  * percpu_count vars will be equal to what it would have been if all the gets
19  * and puts were done to a single integer, even if some of the percpu integers
20  * overflow or underflow).
21  *
22  * The real trick to implementing percpu refcounts is shutdown. We can't detect
23  * the ref hitting 0 on every put - this would require global synchronization
24  * and defeat the whole purpose of using percpu refs.
25  *
26  * What we do is require the user to keep track of the initial refcount; we know
27  * the ref can't hit 0 before the user drops the initial ref, so as long as we
28  * convert to non percpu mode before the initial ref is dropped everything
29  * works.
30  *
31  * Converting to non percpu mode is done with some RCUish stuff in
32  * percpu_ref_kill. Additionally, we need a bias value so that the
33  * atomic_long_t can't hit 0 before we've added up all the percpu refs.
34  */
35 
36 #define PERCPU_COUNT_BIAS	(1LU << (BITS_PER_LONG - 1))
37 
38 static DEFINE_SPINLOCK(percpu_ref_switch_lock);
39 static DECLARE_WAIT_QUEUE_HEAD(percpu_ref_switch_waitq);
40 
percpu_count_ptr(struct percpu_ref * ref)41 static unsigned long __percpu *percpu_count_ptr(struct percpu_ref *ref)
42 {
43 	return (unsigned long __percpu *)
44 		(ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC_DEAD);
45 }
46 
47 /**
48  * percpu_ref_init - initialize a percpu refcount
49  * @ref: percpu_ref to initialize
50  * @release: function which will be called when refcount hits 0
51  * @flags: PERCPU_REF_INIT_* flags
52  * @gfp: allocation mask to use
53  *
54  * Initializes @ref.  @ref starts out in percpu mode with a refcount of 1 unless
55  * @flags contains PERCPU_REF_INIT_ATOMIC or PERCPU_REF_INIT_DEAD.  These flags
56  * change the start state to atomic with the latter setting the initial refcount
57  * to 0.  See the definitions of PERCPU_REF_INIT_* flags for flag behaviors.
58  *
59  * Note that @release must not sleep - it may potentially be called from RCU
60  * callback context by percpu_ref_kill().
61  */
percpu_ref_init(struct percpu_ref * ref,percpu_ref_func_t * release,unsigned int flags,gfp_t gfp)62 int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
63 		    unsigned int flags, gfp_t gfp)
64 {
65 	size_t align = max_t(size_t, 1 << __PERCPU_REF_FLAG_BITS,
66 			     __alignof__(unsigned long));
67 	unsigned long start_count = 0;
68 	struct percpu_ref_data *data;
69 
70 	ref->percpu_count_ptr = (unsigned long)
71 		__alloc_percpu_gfp(sizeof(unsigned long), align, gfp);
72 	if (!ref->percpu_count_ptr)
73 		return -ENOMEM;
74 
75 	data = kzalloc(sizeof(*ref->data), gfp);
76 	if (!data) {
77 		free_percpu((void __percpu *)ref->percpu_count_ptr);
78 		return -ENOMEM;
79 	}
80 
81 	data->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
82 	data->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT;
83 
84 	if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD)) {
85 		ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
86 		data->allow_reinit = true;
87 	} else {
88 		start_count += PERCPU_COUNT_BIAS;
89 	}
90 
91 	if (flags & PERCPU_REF_INIT_DEAD)
92 		ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
93 	else
94 		start_count++;
95 
96 	atomic_long_set(&data->count, start_count);
97 
98 	data->release = release;
99 	data->confirm_switch = NULL;
100 	data->ref = ref;
101 	ref->data = data;
102 	return 0;
103 }
104 EXPORT_SYMBOL_GPL(percpu_ref_init);
105 
__percpu_ref_exit(struct percpu_ref * ref)106 static void __percpu_ref_exit(struct percpu_ref *ref)
107 {
108 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
109 
110 	if (percpu_count) {
111 		/* non-NULL confirm_switch indicates switching in progress */
112 		WARN_ON_ONCE(ref->data && ref->data->confirm_switch);
113 		free_percpu(percpu_count);
114 		ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
115 	}
116 }
117 
118 /**
119  * percpu_ref_exit - undo percpu_ref_init()
120  * @ref: percpu_ref to exit
121  *
122  * This function exits @ref.  The caller is responsible for ensuring that
123  * @ref is no longer in active use.  The usual places to invoke this
124  * function from are the @ref->release() callback or in init failure path
125  * where percpu_ref_init() succeeded but other parts of the initialization
126  * of the embedding object failed.
127  */
percpu_ref_exit(struct percpu_ref * ref)128 void percpu_ref_exit(struct percpu_ref *ref)
129 {
130 	struct percpu_ref_data *data = ref->data;
131 	unsigned long flags;
132 
133 	__percpu_ref_exit(ref);
134 
135 	if (!data)
136 		return;
137 
138 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
139 	ref->percpu_count_ptr |= atomic_long_read(&ref->data->count) <<
140 		__PERCPU_REF_FLAG_BITS;
141 	ref->data = NULL;
142 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
143 
144 	kfree(data);
145 }
146 EXPORT_SYMBOL_GPL(percpu_ref_exit);
147 
percpu_ref_call_confirm_rcu(struct rcu_head * rcu)148 static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
149 {
150 	struct percpu_ref_data *data = container_of(rcu,
151 			struct percpu_ref_data, rcu);
152 	struct percpu_ref *ref = data->ref;
153 
154 	data->confirm_switch(ref);
155 	data->confirm_switch = NULL;
156 	wake_up_all(&percpu_ref_switch_waitq);
157 
158 	if (!data->allow_reinit)
159 		__percpu_ref_exit(ref);
160 
161 	/* drop ref from percpu_ref_switch_to_atomic() */
162 	percpu_ref_put(ref);
163 }
164 
percpu_ref_switch_to_atomic_rcu(struct rcu_head * rcu)165 static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
166 {
167 	struct percpu_ref_data *data = container_of(rcu,
168 			struct percpu_ref_data, rcu);
169 	struct percpu_ref *ref = data->ref;
170 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
171 	unsigned long count = 0;
172 	int cpu;
173 
174 	for_each_possible_cpu(cpu)
175 		count += *per_cpu_ptr(percpu_count, cpu);
176 
177 	pr_debug("global %lu percpu %lu\n",
178 		 atomic_long_read(&data->count), count);
179 
180 	/*
181 	 * It's crucial that we sum the percpu counters _before_ adding the sum
182 	 * to &ref->count; since gets could be happening on one cpu while puts
183 	 * happen on another, adding a single cpu's count could cause
184 	 * @ref->count to hit 0 before we've got a consistent value - but the
185 	 * sum of all the counts will be consistent and correct.
186 	 *
187 	 * Subtracting the bias value then has to happen _after_ adding count to
188 	 * &ref->count; we need the bias value to prevent &ref->count from
189 	 * reaching 0 before we add the percpu counts. But doing it at the same
190 	 * time is equivalent and saves us atomic operations:
191 	 */
192 	atomic_long_add((long)count - PERCPU_COUNT_BIAS, &data->count);
193 
194 	WARN_ONCE(atomic_long_read(&data->count) <= 0,
195 		  "percpu ref (%ps) <= 0 (%ld) after switching to atomic",
196 		  data->release, atomic_long_read(&data->count));
197 
198 	/* @ref is viewed as dead on all CPUs, send out switch confirmation */
199 	percpu_ref_call_confirm_rcu(rcu);
200 }
201 
percpu_ref_noop_confirm_switch(struct percpu_ref * ref)202 static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref)
203 {
204 }
205 
__percpu_ref_switch_to_atomic(struct percpu_ref * ref,percpu_ref_func_t * confirm_switch)206 static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
207 					  percpu_ref_func_t *confirm_switch)
208 {
209 	if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) {
210 		if (confirm_switch)
211 			confirm_switch(ref);
212 		return;
213 	}
214 
215 	/* switching from percpu to atomic */
216 	ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
217 
218 	/*
219 	 * Non-NULL ->confirm_switch is used to indicate that switching is
220 	 * in progress.  Use noop one if unspecified.
221 	 */
222 	ref->data->confirm_switch = confirm_switch ?:
223 		percpu_ref_noop_confirm_switch;
224 
225 	percpu_ref_get(ref);	/* put after confirmation */
226 	call_rcu(&ref->data->rcu, percpu_ref_switch_to_atomic_rcu);
227 }
228 
__percpu_ref_switch_to_percpu(struct percpu_ref * ref)229 static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
230 {
231 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
232 	int cpu;
233 
234 	BUG_ON(!percpu_count);
235 
236 	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
237 		return;
238 
239 	if (WARN_ON_ONCE(!ref->data->allow_reinit))
240 		return;
241 
242 	atomic_long_add(PERCPU_COUNT_BIAS, &ref->data->count);
243 
244 	/*
245 	 * Restore per-cpu operation.  smp_store_release() is paired
246 	 * with READ_ONCE() in __ref_is_percpu() and guarantees that the
247 	 * zeroing is visible to all percpu accesses which can see the
248 	 * following __PERCPU_REF_ATOMIC clearing.
249 	 */
250 	for_each_possible_cpu(cpu)
251 		*per_cpu_ptr(percpu_count, cpu) = 0;
252 
253 	smp_store_release(&ref->percpu_count_ptr,
254 			  ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);
255 }
256 
__percpu_ref_switch_mode(struct percpu_ref * ref,percpu_ref_func_t * confirm_switch)257 static void __percpu_ref_switch_mode(struct percpu_ref *ref,
258 				     percpu_ref_func_t *confirm_switch)
259 {
260 	struct percpu_ref_data *data = ref->data;
261 
262 	lockdep_assert_held(&percpu_ref_switch_lock);
263 
264 	/*
265 	 * If the previous ATOMIC switching hasn't finished yet, wait for
266 	 * its completion.  If the caller ensures that ATOMIC switching
267 	 * isn't in progress, this function can be called from any context.
268 	 */
269 	wait_event_lock_irq(percpu_ref_switch_waitq, !data->confirm_switch,
270 			    percpu_ref_switch_lock);
271 
272 	if (data->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD))
273 		__percpu_ref_switch_to_atomic(ref, confirm_switch);
274 	else
275 		__percpu_ref_switch_to_percpu(ref);
276 }
277 
278 /**
279  * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode
280  * @ref: percpu_ref to switch to atomic mode
281  * @confirm_switch: optional confirmation callback
282  *
283  * There's no reason to use this function for the usual reference counting.
284  * Use percpu_ref_kill[_and_confirm]().
285  *
286  * Schedule switching of @ref to atomic mode.  All its percpu counts will
287  * be collected to the main atomic counter.  On completion, when all CPUs
288  * are guaraneed to be in atomic mode, @confirm_switch, which may not
289  * block, is invoked.  This function may be invoked concurrently with all
290  * the get/put operations and can safely be mixed with kill and reinit
291  * operations.  Note that @ref will stay in atomic mode across kill/reinit
292  * cycles until percpu_ref_switch_to_percpu() is called.
293  *
294  * 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.
297  */
percpu_ref_switch_to_atomic(struct percpu_ref * ref,percpu_ref_func_t * confirm_switch)298 void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
299 				 percpu_ref_func_t *confirm_switch)
300 {
301 	unsigned long flags;
302 
303 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
304 
305 	ref->data->force_atomic = true;
306 	__percpu_ref_switch_mode(ref, confirm_switch);
307 
308 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
309 }
310 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
311 
312 /**
313  * percpu_ref_switch_to_atomic_sync - switch a percpu_ref to atomic mode
314  * @ref: percpu_ref to switch to atomic mode
315  *
316  * Schedule switching the ref to atomic mode, and wait for the
317  * switch to complete.  Caller must ensure that no other thread
318  * will switch back to percpu mode.
319  */
percpu_ref_switch_to_atomic_sync(struct percpu_ref * ref)320 void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
321 {
322 	percpu_ref_switch_to_atomic(ref, NULL);
323 	wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
324 }
325 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);
326 
327 /**
328  * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode
329  * @ref: percpu_ref to switch to percpu mode
330  *
331  * There's no reason to use this function for the usual reference counting.
332  * To re-use an expired ref, use percpu_ref_reinit().
333  *
334  * Switch @ref to percpu mode.  This function may be invoked concurrently
335  * with all the get/put operations and can safely be mixed with kill and
336  * reinit operations.  This function reverses the sticky atomic state set
337  * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic().  If @ref is
338  * dying or dead, the actual switching takes place on the following
339  * percpu_ref_reinit().
340  *
341  * This function may block if @ref is in the process of switching to atomic
342  * mode.  If the caller ensures that @ref is not in the process of
343  * switching to atomic mode, this function can be called from any context.
344  */
percpu_ref_switch_to_percpu(struct percpu_ref * ref)345 void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
346 {
347 	unsigned long flags;
348 
349 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
350 
351 	ref->data->force_atomic = false;
352 	__percpu_ref_switch_mode(ref, NULL);
353 
354 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
355 }
356 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_percpu);
357 
358 /**
359  * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
360  * @ref: percpu_ref to kill
361  * @confirm_kill: optional confirmation callback
362  *
363  * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
364  * @confirm_kill is not NULL.  @confirm_kill, which may not block, will be
365  * called after @ref is seen as dead from all CPUs at which point all
366  * further invocations of percpu_ref_tryget_live() will fail.  See
367  * percpu_ref_tryget_live() for details.
368  *
369  * This function normally doesn't block and can be called from any context
370  * but it may block if @confirm_kill is specified and @ref is in the
371  * process of switching to atomic mode by percpu_ref_switch_to_atomic().
372  *
373  * There are no implied RCU grace periods between kill and release.
374  */
percpu_ref_kill_and_confirm(struct percpu_ref * ref,percpu_ref_func_t * confirm_kill)375 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
376 				 percpu_ref_func_t *confirm_kill)
377 {
378 	unsigned long flags;
379 
380 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
381 
382 	WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD,
383 		  "%s called more than once on %ps!", __func__,
384 		  ref->data->release);
385 
386 	ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
387 	__percpu_ref_switch_mode(ref, confirm_kill);
388 	percpu_ref_put(ref);
389 
390 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
391 }
392 EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
393 
394 /**
395  * percpu_ref_is_zero - test whether a percpu refcount reached zero
396  * @ref: percpu_ref to test
397  *
398  * Returns %true if @ref reached zero.
399  *
400  * This function is safe to call as long as @ref is between init and exit.
401  */
percpu_ref_is_zero(struct percpu_ref * ref)402 bool percpu_ref_is_zero(struct percpu_ref *ref)
403 {
404 	unsigned long __percpu *percpu_count;
405 	unsigned long count, flags;
406 
407 	if (__ref_is_percpu(ref, &percpu_count))
408 		return false;
409 
410 	/* protect us from being destroyed */
411 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
412 	if (ref->data)
413 		count = atomic_long_read(&ref->data->count);
414 	else
415 		count = ref->percpu_count_ptr >> __PERCPU_REF_FLAG_BITS;
416 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
417 
418 	return count == 0;
419 }
420 EXPORT_SYMBOL_GPL(percpu_ref_is_zero);
421 
422 /**
423  * percpu_ref_reinit - re-initialize a percpu refcount
424  * @ref: perpcu_ref to re-initialize
425  *
426  * Re-initialize @ref so that it's in the same state as when it finished
427  * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD.  @ref must have been
428  * initialized successfully and reached 0 but not exited.
429  *
430  * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
431  * this function is in progress.
432  */
percpu_ref_reinit(struct percpu_ref * ref)433 void percpu_ref_reinit(struct percpu_ref *ref)
434 {
435 	WARN_ON_ONCE(!percpu_ref_is_zero(ref));
436 
437 	percpu_ref_resurrect(ref);
438 }
439 EXPORT_SYMBOL_GPL(percpu_ref_reinit);
440 
441 /**
442  * percpu_ref_resurrect - modify a percpu refcount from dead to live
443  * @ref: perpcu_ref to resurrect
444  *
445  * Modify @ref so that it's in the same state as before percpu_ref_kill() was
446  * called. @ref must be dead but must not yet have exited.
447  *
448  * If @ref->release() frees @ref then the caller is responsible for
449  * guaranteeing that @ref->release() does not get called while this
450  * function is in progress.
451  *
452  * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
453  * this function is in progress.
454  */
percpu_ref_resurrect(struct percpu_ref * ref)455 void percpu_ref_resurrect(struct percpu_ref *ref)
456 {
457 	unsigned long __percpu *percpu_count;
458 	unsigned long flags;
459 
460 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
461 
462 	WARN_ON_ONCE(!(ref->percpu_count_ptr & __PERCPU_REF_DEAD));
463 	WARN_ON_ONCE(__ref_is_percpu(ref, &percpu_count));
464 
465 	ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD;
466 	percpu_ref_get(ref);
467 	__percpu_ref_switch_mode(ref, NULL);
468 
469 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
470 }
471 EXPORT_SYMBOL_GPL(percpu_ref_resurrect);
472