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