• 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 		ref->percpu_count_ptr = 0;
79 		return -ENOMEM;
80 	}
81 
82 	data->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
83 	data->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT;
84 
85 	if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD)) {
86 		ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
87 		data->allow_reinit = true;
88 	} else {
89 		start_count += PERCPU_COUNT_BIAS;
90 	}
91 
92 	if (flags & PERCPU_REF_INIT_DEAD)
93 		ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
94 	else
95 		start_count++;
96 
97 	atomic_long_set(&data->count, start_count);
98 
99 	data->release = release;
100 	data->confirm_switch = NULL;
101 	data->ref = ref;
102 	ref->data = data;
103 	return 0;
104 }
105 EXPORT_SYMBOL_GPL(percpu_ref_init);
106 
__percpu_ref_exit(struct percpu_ref * ref)107 static void __percpu_ref_exit(struct percpu_ref *ref)
108 {
109 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
110 
111 	if (percpu_count) {
112 		/* non-NULL confirm_switch indicates switching in progress */
113 		WARN_ON_ONCE(ref->data && ref->data->confirm_switch);
114 		free_percpu(percpu_count);
115 		ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
116 	}
117 }
118 
119 /**
120  * percpu_ref_exit - undo percpu_ref_init()
121  * @ref: percpu_ref to exit
122  *
123  * This function exits @ref.  The caller is responsible for ensuring that
124  * @ref is no longer in active use.  The usual places to invoke this
125  * function from are the @ref->release() callback or in init failure path
126  * where percpu_ref_init() succeeded but other parts of the initialization
127  * of the embedding object failed.
128  */
percpu_ref_exit(struct percpu_ref * ref)129 void percpu_ref_exit(struct percpu_ref *ref)
130 {
131 	struct percpu_ref_data *data = ref->data;
132 	unsigned long flags;
133 
134 	__percpu_ref_exit(ref);
135 
136 	if (!data)
137 		return;
138 
139 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
140 	ref->percpu_count_ptr |= atomic_long_read(&ref->data->count) <<
141 		__PERCPU_REF_FLAG_BITS;
142 	ref->data = NULL;
143 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
144 
145 	kfree(data);
146 }
147 EXPORT_SYMBOL_GPL(percpu_ref_exit);
148 
percpu_ref_call_confirm_rcu(struct rcu_head * rcu)149 static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
150 {
151 	struct percpu_ref_data *data = container_of(rcu,
152 			struct percpu_ref_data, rcu);
153 	struct percpu_ref *ref = data->ref;
154 
155 	data->confirm_switch(ref);
156 	data->confirm_switch = NULL;
157 	wake_up_all(&percpu_ref_switch_waitq);
158 
159 	if (!data->allow_reinit)
160 		__percpu_ref_exit(ref);
161 
162 	/* drop ref from percpu_ref_switch_to_atomic() */
163 	percpu_ref_put(ref);
164 }
165 
percpu_ref_switch_to_atomic_rcu(struct rcu_head * rcu)166 static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
167 {
168 	struct percpu_ref_data *data = container_of(rcu,
169 			struct percpu_ref_data, rcu);
170 	struct percpu_ref *ref = data->ref;
171 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
172 	unsigned long count = 0;
173 	int cpu;
174 
175 	for_each_possible_cpu(cpu)
176 		count += *per_cpu_ptr(percpu_count, cpu);
177 
178 	pr_debug("global %lu percpu %lu\n",
179 		 atomic_long_read(&data->count), count);
180 
181 	/*
182 	 * It's crucial that we sum the percpu counters _before_ adding the sum
183 	 * to &ref->count; since gets could be happening on one cpu while puts
184 	 * happen on another, adding a single cpu's count could cause
185 	 * @ref->count to hit 0 before we've got a consistent value - but the
186 	 * sum of all the counts will be consistent and correct.
187 	 *
188 	 * Subtracting the bias value then has to happen _after_ adding count to
189 	 * &ref->count; we need the bias value to prevent &ref->count from
190 	 * reaching 0 before we add the percpu counts. But doing it at the same
191 	 * time is equivalent and saves us atomic operations:
192 	 */
193 	atomic_long_add((long)count - PERCPU_COUNT_BIAS, &data->count);
194 
195 	WARN_ONCE(atomic_long_read(&data->count) <= 0,
196 		  "percpu ref (%ps) <= 0 (%ld) after switching to atomic",
197 		  data->release, atomic_long_read(&data->count));
198 
199 	/* @ref is viewed as dead on all CPUs, send out switch confirmation */
200 	percpu_ref_call_confirm_rcu(rcu);
201 }
202 
percpu_ref_noop_confirm_switch(struct percpu_ref * ref)203 static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref)
204 {
205 }
206 
__percpu_ref_switch_to_atomic(struct percpu_ref * ref,percpu_ref_func_t * confirm_switch)207 static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
208 					  percpu_ref_func_t *confirm_switch)
209 {
210 	if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) {
211 		if (confirm_switch)
212 			confirm_switch(ref);
213 		return;
214 	}
215 
216 	/* switching from percpu to atomic */
217 	ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
218 
219 	/*
220 	 * Non-NULL ->confirm_switch is used to indicate that switching is
221 	 * in progress.  Use noop one if unspecified.
222 	 */
223 	ref->data->confirm_switch = confirm_switch ?:
224 		percpu_ref_noop_confirm_switch;
225 
226 	percpu_ref_get(ref);	/* put after confirmation */
227 	call_rcu(&ref->data->rcu, percpu_ref_switch_to_atomic_rcu);
228 }
229 
__percpu_ref_switch_to_percpu(struct percpu_ref * ref)230 static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
231 {
232 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
233 	int cpu;
234 
235 	BUG_ON(!percpu_count);
236 
237 	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
238 		return;
239 
240 	if (WARN_ON_ONCE(!ref->data->allow_reinit))
241 		return;
242 
243 	atomic_long_add(PERCPU_COUNT_BIAS, &ref->data->count);
244 
245 	/*
246 	 * Restore per-cpu operation.  smp_store_release() is paired
247 	 * with READ_ONCE() in __ref_is_percpu() and guarantees that the
248 	 * zeroing is visible to all percpu accesses which can see the
249 	 * following __PERCPU_REF_ATOMIC clearing.
250 	 */
251 	for_each_possible_cpu(cpu)
252 		*per_cpu_ptr(percpu_count, cpu) = 0;
253 
254 	smp_store_release(&ref->percpu_count_ptr,
255 			  ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);
256 }
257 
__percpu_ref_switch_mode(struct percpu_ref * ref,percpu_ref_func_t * confirm_switch)258 static void __percpu_ref_switch_mode(struct percpu_ref *ref,
259 				     percpu_ref_func_t *confirm_switch)
260 {
261 	struct percpu_ref_data *data = ref->data;
262 
263 	lockdep_assert_held(&percpu_ref_switch_lock);
264 
265 	/*
266 	 * If the previous ATOMIC switching hasn't finished yet, wait for
267 	 * its completion.  If the caller ensures that ATOMIC switching
268 	 * isn't in progress, this function can be called from any context.
269 	 */
270 	wait_event_lock_irq(percpu_ref_switch_waitq, !data->confirm_switch,
271 			    percpu_ref_switch_lock);
272 
273 	if (data->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD))
274 		__percpu_ref_switch_to_atomic(ref, confirm_switch);
275 	else
276 		__percpu_ref_switch_to_percpu(ref);
277 }
278 
279 /**
280  * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode
281  * @ref: percpu_ref to switch to atomic mode
282  * @confirm_switch: optional confirmation callback
283  *
284  * There's no reason to use this function for the usual reference counting.
285  * Use percpu_ref_kill[_and_confirm]().
286  *
287  * Schedule switching of @ref to atomic mode.  All its percpu counts will
288  * be collected to the main atomic counter.  On completion, when all CPUs
289  * are guaraneed to be in atomic mode, @confirm_switch, which may not
290  * block, is invoked.  This function may be invoked concurrently with all
291  * the get/put operations and can safely be mixed with kill and reinit
292  * operations.  Note that @ref will stay in atomic mode across kill/reinit
293  * cycles until percpu_ref_switch_to_percpu() is called.
294  *
295  * This function may block if @ref is in the process of switching to atomic
296  * mode.  If the caller ensures that @ref is not in the process of
297  * switching to atomic mode, this function can be called from any context.
298  */
percpu_ref_switch_to_atomic(struct percpu_ref * ref,percpu_ref_func_t * confirm_switch)299 void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
300 				 percpu_ref_func_t *confirm_switch)
301 {
302 	unsigned long flags;
303 
304 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
305 
306 	ref->data->force_atomic = true;
307 	__percpu_ref_switch_mode(ref, confirm_switch);
308 
309 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
310 }
311 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic);
312 
313 /**
314  * percpu_ref_switch_to_atomic_sync - switch a percpu_ref to atomic mode
315  * @ref: percpu_ref to switch to atomic mode
316  *
317  * Schedule switching the ref to atomic mode, and wait for the
318  * switch to complete.  Caller must ensure that no other thread
319  * will switch back to percpu mode.
320  */
percpu_ref_switch_to_atomic_sync(struct percpu_ref * ref)321 void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref)
322 {
323 	percpu_ref_switch_to_atomic(ref, NULL);
324 	wait_event(percpu_ref_switch_waitq, !ref->data->confirm_switch);
325 }
326 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync);
327 
328 /**
329  * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode
330  * @ref: percpu_ref to switch to percpu mode
331  *
332  * There's no reason to use this function for the usual reference counting.
333  * To re-use an expired ref, use percpu_ref_reinit().
334  *
335  * Switch @ref to percpu mode.  This function may be invoked concurrently
336  * with all the get/put operations and can safely be mixed with kill and
337  * reinit operations.  This function reverses the sticky atomic state set
338  * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic().  If @ref is
339  * dying or dead, the actual switching takes place on the following
340  * percpu_ref_reinit().
341  *
342  * This function may block if @ref is in the process of switching to atomic
343  * mode.  If the caller ensures that @ref is not in the process of
344  * switching to atomic mode, this function can be called from any context.
345  */
percpu_ref_switch_to_percpu(struct percpu_ref * ref)346 void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
347 {
348 	unsigned long flags;
349 
350 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
351 
352 	ref->data->force_atomic = false;
353 	__percpu_ref_switch_mode(ref, NULL);
354 
355 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
356 }
357 EXPORT_SYMBOL_GPL(percpu_ref_switch_to_percpu);
358 
359 /**
360  * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
361  * @ref: percpu_ref to kill
362  * @confirm_kill: optional confirmation callback
363  *
364  * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
365  * @confirm_kill is not NULL.  @confirm_kill, which may not block, will be
366  * called after @ref is seen as dead from all CPUs at which point all
367  * further invocations of percpu_ref_tryget_live() will fail.  See
368  * percpu_ref_tryget_live() for details.
369  *
370  * This function normally doesn't block and can be called from any context
371  * but it may block if @confirm_kill is specified and @ref is in the
372  * process of switching to atomic mode by percpu_ref_switch_to_atomic().
373  *
374  * There are no implied RCU grace periods between kill and release.
375  */
percpu_ref_kill_and_confirm(struct percpu_ref * ref,percpu_ref_func_t * confirm_kill)376 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
377 				 percpu_ref_func_t *confirm_kill)
378 {
379 	unsigned long flags;
380 
381 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
382 
383 	WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD,
384 		  "%s called more than once on %ps!", __func__,
385 		  ref->data->release);
386 
387 	ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
388 	__percpu_ref_switch_mode(ref, confirm_kill);
389 	percpu_ref_put(ref);
390 
391 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
392 }
393 EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
394 
395 /**
396  * percpu_ref_is_zero - test whether a percpu refcount reached zero
397  * @ref: percpu_ref to test
398  *
399  * Returns %true if @ref reached zero.
400  *
401  * This function is safe to call as long as @ref is between init and exit.
402  */
percpu_ref_is_zero(struct percpu_ref * ref)403 bool percpu_ref_is_zero(struct percpu_ref *ref)
404 {
405 	unsigned long __percpu *percpu_count;
406 	unsigned long count, flags;
407 
408 	if (__ref_is_percpu(ref, &percpu_count))
409 		return false;
410 
411 	/* protect us from being destroyed */
412 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
413 	if (ref->data)
414 		count = atomic_long_read(&ref->data->count);
415 	else
416 		count = ref->percpu_count_ptr >> __PERCPU_REF_FLAG_BITS;
417 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
418 
419 	return count == 0;
420 }
421 EXPORT_SYMBOL_GPL(percpu_ref_is_zero);
422 
423 /**
424  * percpu_ref_reinit - re-initialize a percpu refcount
425  * @ref: perpcu_ref to re-initialize
426  *
427  * Re-initialize @ref so that it's in the same state as when it finished
428  * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD.  @ref must have been
429  * initialized successfully and reached 0 but not exited.
430  *
431  * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
432  * this function is in progress.
433  */
percpu_ref_reinit(struct percpu_ref * ref)434 void percpu_ref_reinit(struct percpu_ref *ref)
435 {
436 	WARN_ON_ONCE(!percpu_ref_is_zero(ref));
437 
438 	percpu_ref_resurrect(ref);
439 }
440 EXPORT_SYMBOL_GPL(percpu_ref_reinit);
441 
442 /**
443  * percpu_ref_resurrect - modify a percpu refcount from dead to live
444  * @ref: perpcu_ref to resurrect
445  *
446  * Modify @ref so that it's in the same state as before percpu_ref_kill() was
447  * called. @ref must be dead but must not yet have exited.
448  *
449  * If @ref->release() frees @ref then the caller is responsible for
450  * guaranteeing that @ref->release() does not get called while this
451  * function is in progress.
452  *
453  * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
454  * this function is in progress.
455  */
percpu_ref_resurrect(struct percpu_ref * ref)456 void percpu_ref_resurrect(struct percpu_ref *ref)
457 {
458 	unsigned long __percpu *percpu_count;
459 	unsigned long flags;
460 
461 	spin_lock_irqsave(&percpu_ref_switch_lock, flags);
462 
463 	WARN_ON_ONCE(!(ref->percpu_count_ptr & __PERCPU_REF_DEAD));
464 	WARN_ON_ONCE(__ref_is_percpu(ref, &percpu_count));
465 
466 	ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD;
467 	percpu_ref_get(ref);
468 	__percpu_ref_switch_mode(ref, NULL);
469 
470 	spin_unlock_irqrestore(&percpu_ref_switch_lock, flags);
471 }
472 EXPORT_SYMBOL_GPL(percpu_ref_resurrect);
473