Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
215e262f KO |
2 | #define pr_fmt(fmt) "%s: " fmt "\n", __func__ |
3 | ||
4 | #include <linux/kernel.h> | |
490c79a6 TH |
5 | #include <linux/sched.h> |
6 | #include <linux/wait.h> | |
215e262f KO |
7 | #include <linux/percpu-refcount.h> |
8 | ||
9 | /* | |
10 | * Initially, a percpu refcount is just a set of percpu counters. Initially, we | |
11 | * don't try to detect the ref hitting 0 - which means that get/put can just | |
12 | * increment or decrement the local counter. Note that the counter on a | |
13 | * particular cpu can (and will) wrap - this is fine, when we go to shutdown the | |
14 | * percpu counters will all sum to the correct value | |
15 | * | |
bdb428c8 | 16 | * (More precisely: because modular arithmetic is commutative the sum of all the |
eecc16ba TH |
17 | * percpu_count vars will be equal to what it would have been if all the gets |
18 | * and puts were done to a single integer, even if some of the percpu integers | |
215e262f KO |
19 | * overflow or underflow). |
20 | * | |
21 | * The real trick to implementing percpu refcounts is shutdown. We can't detect | |
22 | * the ref hitting 0 on every put - this would require global synchronization | |
23 | * and defeat the whole purpose of using percpu refs. | |
24 | * | |
25 | * What we do is require the user to keep track of the initial refcount; we know | |
26 | * the ref can't hit 0 before the user drops the initial ref, so as long as we | |
27 | * convert to non percpu mode before the initial ref is dropped everything | |
28 | * works. | |
29 | * | |
30 | * Converting to non percpu mode is done with some RCUish stuff in | |
e625305b TH |
31 | * percpu_ref_kill. Additionally, we need a bias value so that the |
32 | * atomic_long_t can't hit 0 before we've added up all the percpu refs. | |
215e262f KO |
33 | */ |
34 | ||
eecc16ba | 35 | #define PERCPU_COUNT_BIAS (1LU << (BITS_PER_LONG - 1)) |
215e262f | 36 | |
33e465ce | 37 | static DEFINE_SPINLOCK(percpu_ref_switch_lock); |
490c79a6 TH |
38 | static DECLARE_WAIT_QUEUE_HEAD(percpu_ref_switch_waitq); |
39 | ||
eecc16ba | 40 | static unsigned long __percpu *percpu_count_ptr(struct percpu_ref *ref) |
eae7975d | 41 | { |
eecc16ba | 42 | return (unsigned long __percpu *) |
27344a90 | 43 | (ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC_DEAD); |
eae7975d TH |
44 | } |
45 | ||
215e262f KO |
46 | /** |
47 | * percpu_ref_init - initialize a percpu refcount | |
ac899061 TH |
48 | * @ref: percpu_ref to initialize |
49 | * @release: function which will be called when refcount hits 0 | |
2aad2a86 | 50 | * @flags: PERCPU_REF_INIT_* flags |
a34375ef | 51 | * @gfp: allocation mask to use |
215e262f | 52 | * |
2aad2a86 TH |
53 | * Initializes @ref. If @flags is zero, @ref starts in percpu mode with a |
54 | * refcount of 1; analagous to atomic_long_set(ref, 1). See the | |
55 | * definitions of PERCPU_REF_INIT_* flags for flag behaviors. | |
215e262f KO |
56 | * |
57 | * Note that @release must not sleep - it may potentially be called from RCU | |
58 | * callback context by percpu_ref_kill(). | |
59 | */ | |
a34375ef | 60 | int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release, |
2aad2a86 | 61 | unsigned int flags, gfp_t gfp) |
215e262f | 62 | { |
27344a90 TH |
63 | size_t align = max_t(size_t, 1 << __PERCPU_REF_FLAG_BITS, |
64 | __alignof__(unsigned long)); | |
2aad2a86 | 65 | unsigned long start_count = 0; |
215e262f | 66 | |
27344a90 TH |
67 | ref->percpu_count_ptr = (unsigned long) |
68 | __alloc_percpu_gfp(sizeof(unsigned long), align, gfp); | |
eecc16ba | 69 | if (!ref->percpu_count_ptr) |
215e262f KO |
70 | return -ENOMEM; |
71 | ||
1cae13e7 | 72 | ref->force_atomic = flags & PERCPU_REF_INIT_ATOMIC; |
7d9ab9b6 | 73 | ref->allow_reinit = flags & PERCPU_REF_ALLOW_REINIT; |
1cae13e7 | 74 | |
7d9ab9b6 | 75 | if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD)) { |
2aad2a86 | 76 | ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC; |
7d9ab9b6 RG |
77 | ref->allow_reinit = true; |
78 | } else { | |
2aad2a86 | 79 | start_count += PERCPU_COUNT_BIAS; |
7d9ab9b6 | 80 | } |
2aad2a86 TH |
81 | |
82 | if (flags & PERCPU_REF_INIT_DEAD) | |
83 | ref->percpu_count_ptr |= __PERCPU_REF_DEAD; | |
84 | else | |
85 | start_count++; | |
86 | ||
87 | atomic_long_set(&ref->count, start_count); | |
88 | ||
215e262f | 89 | ref->release = release; |
a67823c1 | 90 | ref->confirm_switch = NULL; |
215e262f KO |
91 | return 0; |
92 | } | |
5e9dd373 | 93 | EXPORT_SYMBOL_GPL(percpu_ref_init); |
215e262f | 94 | |
bc497bd3 | 95 | /** |
9a1049da TH |
96 | * percpu_ref_exit - undo percpu_ref_init() |
97 | * @ref: percpu_ref to exit | |
bc497bd3 | 98 | * |
9a1049da TH |
99 | * This function exits @ref. The caller is responsible for ensuring that |
100 | * @ref is no longer in active use. The usual places to invoke this | |
101 | * function from are the @ref->release() callback or in init failure path | |
102 | * where percpu_ref_init() succeeded but other parts of the initialization | |
103 | * of the embedding object failed. | |
bc497bd3 | 104 | */ |
9a1049da | 105 | void percpu_ref_exit(struct percpu_ref *ref) |
bc497bd3 | 106 | { |
eecc16ba | 107 | unsigned long __percpu *percpu_count = percpu_count_ptr(ref); |
bc497bd3 | 108 | |
eecc16ba | 109 | if (percpu_count) { |
a67823c1 RP |
110 | /* non-NULL confirm_switch indicates switching in progress */ |
111 | WARN_ON_ONCE(ref->confirm_switch); | |
eecc16ba | 112 | free_percpu(percpu_count); |
27344a90 | 113 | ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD; |
bc497bd3 TH |
114 | } |
115 | } | |
9a1049da | 116 | EXPORT_SYMBOL_GPL(percpu_ref_exit); |
bc497bd3 | 117 | |
490c79a6 TH |
118 | static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu) |
119 | { | |
120 | struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu); | |
121 | ||
122 | ref->confirm_switch(ref); | |
123 | ref->confirm_switch = NULL; | |
124 | wake_up_all(&percpu_ref_switch_waitq); | |
125 | ||
7d9ab9b6 RG |
126 | if (!ref->allow_reinit) |
127 | percpu_ref_exit(ref); | |
128 | ||
490c79a6 TH |
129 | /* drop ref from percpu_ref_switch_to_atomic() */ |
130 | percpu_ref_put(ref); | |
131 | } | |
132 | ||
133 | static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu) | |
215e262f KO |
134 | { |
135 | struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu); | |
eecc16ba | 136 | unsigned long __percpu *percpu_count = percpu_count_ptr(ref); |
e625305b | 137 | unsigned long count = 0; |
215e262f KO |
138 | int cpu; |
139 | ||
215e262f | 140 | for_each_possible_cpu(cpu) |
eecc16ba | 141 | count += *per_cpu_ptr(percpu_count, cpu); |
215e262f | 142 | |
eecc16ba | 143 | pr_debug("global %ld percpu %ld", |
e625305b | 144 | atomic_long_read(&ref->count), (long)count); |
215e262f KO |
145 | |
146 | /* | |
147 | * It's crucial that we sum the percpu counters _before_ adding the sum | |
148 | * to &ref->count; since gets could be happening on one cpu while puts | |
149 | * happen on another, adding a single cpu's count could cause | |
150 | * @ref->count to hit 0 before we've got a consistent value - but the | |
151 | * sum of all the counts will be consistent and correct. | |
152 | * | |
153 | * Subtracting the bias value then has to happen _after_ adding count to | |
154 | * &ref->count; we need the bias value to prevent &ref->count from | |
155 | * reaching 0 before we add the percpu counts. But doing it at the same | |
156 | * time is equivalent and saves us atomic operations: | |
157 | */ | |
eecc16ba | 158 | atomic_long_add((long)count - PERCPU_COUNT_BIAS, &ref->count); |
215e262f | 159 | |
e625305b | 160 | WARN_ONCE(atomic_long_read(&ref->count) <= 0, |
d75f773c | 161 | "percpu ref (%ps) <= 0 (%ld) after switching to atomic", |
e625305b | 162 | ref->release, atomic_long_read(&ref->count)); |
687b0ad2 | 163 | |
490c79a6 TH |
164 | /* @ref is viewed as dead on all CPUs, send out switch confirmation */ |
165 | percpu_ref_call_confirm_rcu(rcu); | |
166 | } | |
dbece3a0 | 167 | |
490c79a6 TH |
168 | static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref) |
169 | { | |
170 | } | |
171 | ||
172 | static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref, | |
173 | percpu_ref_func_t *confirm_switch) | |
174 | { | |
b2302c7f | 175 | if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) { |
18808354 | 176 | if (confirm_switch) |
b2302c7f | 177 | confirm_switch(ref); |
b2302c7f | 178 | return; |
490c79a6 | 179 | } |
215e262f | 180 | |
b2302c7f TH |
181 | /* switching from percpu to atomic */ |
182 | ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC; | |
183 | ||
184 | /* | |
185 | * Non-NULL ->confirm_switch is used to indicate that switching is | |
186 | * in progress. Use noop one if unspecified. | |
187 | */ | |
b2302c7f TH |
188 | ref->confirm_switch = confirm_switch ?: percpu_ref_noop_confirm_switch; |
189 | ||
190 | percpu_ref_get(ref); /* put after confirmation */ | |
36bd1a8e | 191 | call_rcu(&ref->rcu, percpu_ref_switch_to_atomic_rcu); |
215e262f | 192 | } |
a2237370 | 193 | |
f47ad457 | 194 | static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref) |
a2237370 | 195 | { |
eecc16ba | 196 | unsigned long __percpu *percpu_count = percpu_count_ptr(ref); |
a2237370 TH |
197 | int cpu; |
198 | ||
eecc16ba | 199 | BUG_ON(!percpu_count); |
a2237370 | 200 | |
f47ad457 TH |
201 | if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC)) |
202 | return; | |
203 | ||
7d9ab9b6 RG |
204 | if (WARN_ON_ONCE(!ref->allow_reinit)) |
205 | return; | |
206 | ||
f47ad457 | 207 | atomic_long_add(PERCPU_COUNT_BIAS, &ref->count); |
a2237370 TH |
208 | |
209 | /* | |
b393e8b3 PM |
210 | * Restore per-cpu operation. smp_store_release() is paired |
211 | * with READ_ONCE() in __ref_is_percpu() and guarantees that the | |
212 | * zeroing is visible to all percpu accesses which can see the | |
213 | * following __PERCPU_REF_ATOMIC clearing. | |
a2237370 TH |
214 | */ |
215 | for_each_possible_cpu(cpu) | |
eecc16ba | 216 | *per_cpu_ptr(percpu_count, cpu) = 0; |
a2237370 | 217 | |
eecc16ba | 218 | smp_store_release(&ref->percpu_count_ptr, |
f47ad457 TH |
219 | ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC); |
220 | } | |
221 | ||
3f49bdd9 TH |
222 | static void __percpu_ref_switch_mode(struct percpu_ref *ref, |
223 | percpu_ref_func_t *confirm_switch) | |
224 | { | |
33e465ce TH |
225 | lockdep_assert_held(&percpu_ref_switch_lock); |
226 | ||
3f49bdd9 TH |
227 | /* |
228 | * If the previous ATOMIC switching hasn't finished yet, wait for | |
229 | * its completion. If the caller ensures that ATOMIC switching | |
230 | * isn't in progress, this function can be called from any context. | |
3f49bdd9 | 231 | */ |
33e465ce TH |
232 | wait_event_lock_irq(percpu_ref_switch_waitq, !ref->confirm_switch, |
233 | percpu_ref_switch_lock); | |
3f49bdd9 TH |
234 | |
235 | if (ref->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD)) | |
236 | __percpu_ref_switch_to_atomic(ref, confirm_switch); | |
237 | else | |
238 | __percpu_ref_switch_to_percpu(ref); | |
239 | } | |
240 | ||
b2302c7f TH |
241 | /** |
242 | * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode | |
243 | * @ref: percpu_ref to switch to atomic mode | |
244 | * @confirm_switch: optional confirmation callback | |
245 | * | |
246 | * There's no reason to use this function for the usual reference counting. | |
247 | * Use percpu_ref_kill[_and_confirm](). | |
248 | * | |
249 | * Schedule switching of @ref to atomic mode. All its percpu counts will | |
250 | * be collected to the main atomic counter. On completion, when all CPUs | |
251 | * are guaraneed to be in atomic mode, @confirm_switch, which may not | |
252 | * block, is invoked. This function may be invoked concurrently with all | |
253 | * the get/put operations and can safely be mixed with kill and reinit | |
254 | * operations. Note that @ref will stay in atomic mode across kill/reinit | |
255 | * cycles until percpu_ref_switch_to_percpu() is called. | |
256 | * | |
3f49bdd9 TH |
257 | * This function may block if @ref is in the process of switching to atomic |
258 | * mode. If the caller ensures that @ref is not in the process of | |
259 | * switching to atomic mode, this function can be called from any context. | |
b2302c7f TH |
260 | */ |
261 | void percpu_ref_switch_to_atomic(struct percpu_ref *ref, | |
262 | percpu_ref_func_t *confirm_switch) | |
263 | { | |
33e465ce TH |
264 | unsigned long flags; |
265 | ||
266 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
267 | ||
b2302c7f | 268 | ref->force_atomic = true; |
3f49bdd9 | 269 | __percpu_ref_switch_mode(ref, confirm_switch); |
33e465ce TH |
270 | |
271 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
b2302c7f | 272 | } |
210f7cdc N |
273 | EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic); |
274 | ||
275 | /** | |
276 | * percpu_ref_switch_to_atomic_sync - switch a percpu_ref to atomic mode | |
277 | * @ref: percpu_ref to switch to atomic mode | |
278 | * | |
279 | * Schedule switching the ref to atomic mode, and wait for the | |
280 | * switch to complete. Caller must ensure that no other thread | |
281 | * will switch back to percpu mode. | |
282 | */ | |
283 | void percpu_ref_switch_to_atomic_sync(struct percpu_ref *ref) | |
284 | { | |
285 | percpu_ref_switch_to_atomic(ref, NULL); | |
286 | wait_event(percpu_ref_switch_waitq, !ref->confirm_switch); | |
287 | } | |
288 | EXPORT_SYMBOL_GPL(percpu_ref_switch_to_atomic_sync); | |
b2302c7f | 289 | |
f47ad457 TH |
290 | /** |
291 | * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode | |
292 | * @ref: percpu_ref to switch to percpu mode | |
293 | * | |
294 | * There's no reason to use this function for the usual reference counting. | |
295 | * To re-use an expired ref, use percpu_ref_reinit(). | |
296 | * | |
297 | * Switch @ref to percpu mode. This function may be invoked concurrently | |
298 | * with all the get/put operations and can safely be mixed with kill and | |
1cae13e7 TH |
299 | * reinit operations. This function reverses the sticky atomic state set |
300 | * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic(). If @ref is | |
301 | * dying or dead, the actual switching takes place on the following | |
302 | * percpu_ref_reinit(). | |
f47ad457 | 303 | * |
3f49bdd9 TH |
304 | * This function may block if @ref is in the process of switching to atomic |
305 | * mode. If the caller ensures that @ref is not in the process of | |
306 | * switching to atomic mode, this function can be called from any context. | |
f47ad457 TH |
307 | */ |
308 | void percpu_ref_switch_to_percpu(struct percpu_ref *ref) | |
309 | { | |
33e465ce TH |
310 | unsigned long flags; |
311 | ||
312 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
313 | ||
1cae13e7 | 314 | ref->force_atomic = false; |
3f49bdd9 | 315 | __percpu_ref_switch_mode(ref, NULL); |
33e465ce TH |
316 | |
317 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
a2237370 | 318 | } |
210f7cdc | 319 | EXPORT_SYMBOL_GPL(percpu_ref_switch_to_percpu); |
490c79a6 TH |
320 | |
321 | /** | |
322 | * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation | |
323 | * @ref: percpu_ref to kill | |
324 | * @confirm_kill: optional confirmation callback | |
325 | * | |
326 | * Equivalent to percpu_ref_kill() but also schedules kill confirmation if | |
327 | * @confirm_kill is not NULL. @confirm_kill, which may not block, will be | |
328 | * called after @ref is seen as dead from all CPUs at which point all | |
329 | * further invocations of percpu_ref_tryget_live() will fail. See | |
330 | * percpu_ref_tryget_live() for details. | |
331 | * | |
332 | * This function normally doesn't block and can be called from any context | |
f47ad457 | 333 | * but it may block if @confirm_kill is specified and @ref is in the |
a2f5630c | 334 | * process of switching to atomic mode by percpu_ref_switch_to_atomic(). |
b3a5d111 TH |
335 | * |
336 | * There are no implied RCU grace periods between kill and release. | |
490c79a6 TH |
337 | */ |
338 | void percpu_ref_kill_and_confirm(struct percpu_ref *ref, | |
339 | percpu_ref_func_t *confirm_kill) | |
340 | { | |
33e465ce TH |
341 | unsigned long flags; |
342 | ||
343 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
344 | ||
490c79a6 | 345 | WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD, |
d75f773c | 346 | "%s called more than once on %ps!", __func__, ref->release); |
490c79a6 TH |
347 | |
348 | ref->percpu_count_ptr |= __PERCPU_REF_DEAD; | |
3f49bdd9 | 349 | __percpu_ref_switch_mode(ref, confirm_kill); |
490c79a6 | 350 | percpu_ref_put(ref); |
33e465ce TH |
351 | |
352 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
490c79a6 TH |
353 | } |
354 | EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm); | |
f47ad457 TH |
355 | |
356 | /** | |
357 | * percpu_ref_reinit - re-initialize a percpu refcount | |
358 | * @ref: perpcu_ref to re-initialize | |
359 | * | |
360 | * Re-initialize @ref so that it's in the same state as when it finished | |
1cae13e7 TH |
361 | * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD. @ref must have been |
362 | * initialized successfully and reached 0 but not exited. | |
f47ad457 TH |
363 | * |
364 | * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while | |
365 | * this function is in progress. | |
366 | */ | |
367 | void percpu_ref_reinit(struct percpu_ref *ref) | |
368 | { | |
18c9a6bb BVA |
369 | WARN_ON_ONCE(!percpu_ref_is_zero(ref)); |
370 | ||
371 | percpu_ref_resurrect(ref); | |
372 | } | |
373 | EXPORT_SYMBOL_GPL(percpu_ref_reinit); | |
374 | ||
375 | /** | |
376 | * percpu_ref_resurrect - modify a percpu refcount from dead to live | |
377 | * @ref: perpcu_ref to resurrect | |
378 | * | |
379 | * Modify @ref so that it's in the same state as before percpu_ref_kill() was | |
380 | * called. @ref must be dead but must not yet have exited. | |
381 | * | |
382 | * If @ref->release() frees @ref then the caller is responsible for | |
383 | * guaranteeing that @ref->release() does not get called while this | |
384 | * function is in progress. | |
385 | * | |
386 | * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while | |
387 | * this function is in progress. | |
388 | */ | |
389 | void percpu_ref_resurrect(struct percpu_ref *ref) | |
390 | { | |
391 | unsigned long __percpu *percpu_count; | |
33e465ce TH |
392 | unsigned long flags; |
393 | ||
394 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
395 | ||
18c9a6bb BVA |
396 | WARN_ON_ONCE(!(ref->percpu_count_ptr & __PERCPU_REF_DEAD)); |
397 | WARN_ON_ONCE(__ref_is_percpu(ref, &percpu_count)); | |
f47ad457 TH |
398 | |
399 | ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD; | |
400 | percpu_ref_get(ref); | |
3f49bdd9 | 401 | __percpu_ref_switch_mode(ref, NULL); |
33e465ce TH |
402 | |
403 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
f47ad457 | 404 | } |
18c9a6bb | 405 | EXPORT_SYMBOL_GPL(percpu_ref_resurrect); |