]>
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 TH |
71 | ref->force_atomic = flags & PERCPU_REF_INIT_ATOMIC; |
72 | ||
2aad2a86 TH |
73 | 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 | ||
215e262f KO |
85 | ref->release = release; |
86 | return 0; | |
87 | } | |
5e9dd373 | 88 | EXPORT_SYMBOL_GPL(percpu_ref_init); |
215e262f | 89 | |
bc497bd3 | 90 | /** |
9a1049da TH |
91 | * percpu_ref_exit - undo percpu_ref_init() |
92 | * @ref: percpu_ref to exit | |
bc497bd3 | 93 | * |
9a1049da TH |
94 | * This function exits @ref. The caller is responsible for ensuring that |
95 | * @ref is no longer in active use. The usual places to invoke this | |
96 | * function from are the @ref->release() callback or in init failure path | |
97 | * where percpu_ref_init() succeeded but other parts of the initialization | |
98 | * of the embedding object failed. | |
bc497bd3 | 99 | */ |
9a1049da | 100 | void percpu_ref_exit(struct percpu_ref *ref) |
bc497bd3 | 101 | { |
eecc16ba | 102 | unsigned long __percpu *percpu_count = percpu_count_ptr(ref); |
bc497bd3 | 103 | |
eecc16ba TH |
104 | if (percpu_count) { |
105 | free_percpu(percpu_count); | |
27344a90 | 106 | ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD; |
bc497bd3 TH |
107 | } |
108 | } | |
9a1049da | 109 | EXPORT_SYMBOL_GPL(percpu_ref_exit); |
bc497bd3 | 110 | |
490c79a6 TH |
111 | static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu) |
112 | { | |
113 | struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu); | |
114 | ||
115 | ref->confirm_switch(ref); | |
116 | ref->confirm_switch = NULL; | |
117 | wake_up_all(&percpu_ref_switch_waitq); | |
118 | ||
119 | /* drop ref from percpu_ref_switch_to_atomic() */ | |
120 | percpu_ref_put(ref); | |
121 | } | |
122 | ||
123 | static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu) | |
215e262f KO |
124 | { |
125 | struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu); | |
eecc16ba | 126 | unsigned long __percpu *percpu_count = percpu_count_ptr(ref); |
e625305b | 127 | unsigned long count = 0; |
215e262f KO |
128 | int cpu; |
129 | ||
215e262f | 130 | for_each_possible_cpu(cpu) |
eecc16ba | 131 | count += *per_cpu_ptr(percpu_count, cpu); |
215e262f | 132 | |
eecc16ba | 133 | pr_debug("global %ld percpu %ld", |
e625305b | 134 | atomic_long_read(&ref->count), (long)count); |
215e262f KO |
135 | |
136 | /* | |
137 | * It's crucial that we sum the percpu counters _before_ adding the sum | |
138 | * to &ref->count; since gets could be happening on one cpu while puts | |
139 | * happen on another, adding a single cpu's count could cause | |
140 | * @ref->count to hit 0 before we've got a consistent value - but the | |
141 | * sum of all the counts will be consistent and correct. | |
142 | * | |
143 | * Subtracting the bias value then has to happen _after_ adding count to | |
144 | * &ref->count; we need the bias value to prevent &ref->count from | |
145 | * reaching 0 before we add the percpu counts. But doing it at the same | |
146 | * time is equivalent and saves us atomic operations: | |
147 | */ | |
eecc16ba | 148 | atomic_long_add((long)count - PERCPU_COUNT_BIAS, &ref->count); |
215e262f | 149 | |
e625305b | 150 | WARN_ONCE(atomic_long_read(&ref->count) <= 0, |
490c79a6 | 151 | "percpu ref (%pf) <= 0 (%ld) after switching to atomic", |
e625305b | 152 | ref->release, atomic_long_read(&ref->count)); |
687b0ad2 | 153 | |
490c79a6 TH |
154 | /* @ref is viewed as dead on all CPUs, send out switch confirmation */ |
155 | percpu_ref_call_confirm_rcu(rcu); | |
156 | } | |
dbece3a0 | 157 | |
490c79a6 TH |
158 | static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref) |
159 | { | |
160 | } | |
161 | ||
162 | static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref, | |
163 | percpu_ref_func_t *confirm_switch) | |
164 | { | |
b2302c7f | 165 | if (ref->percpu_count_ptr & __PERCPU_REF_ATOMIC) { |
18808354 | 166 | if (confirm_switch) |
b2302c7f | 167 | confirm_switch(ref); |
b2302c7f | 168 | return; |
490c79a6 | 169 | } |
215e262f | 170 | |
b2302c7f TH |
171 | /* switching from percpu to atomic */ |
172 | ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC; | |
173 | ||
174 | /* | |
175 | * Non-NULL ->confirm_switch is used to indicate that switching is | |
176 | * in progress. Use noop one if unspecified. | |
177 | */ | |
b2302c7f TH |
178 | ref->confirm_switch = confirm_switch ?: percpu_ref_noop_confirm_switch; |
179 | ||
180 | percpu_ref_get(ref); /* put after confirmation */ | |
181 | call_rcu_sched(&ref->rcu, percpu_ref_switch_to_atomic_rcu); | |
215e262f | 182 | } |
a2237370 | 183 | |
f47ad457 | 184 | static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref) |
a2237370 | 185 | { |
eecc16ba | 186 | unsigned long __percpu *percpu_count = percpu_count_ptr(ref); |
a2237370 TH |
187 | int cpu; |
188 | ||
eecc16ba | 189 | BUG_ON(!percpu_count); |
a2237370 | 190 | |
f47ad457 TH |
191 | if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC)) |
192 | return; | |
193 | ||
f47ad457 | 194 | atomic_long_add(PERCPU_COUNT_BIAS, &ref->count); |
a2237370 TH |
195 | |
196 | /* | |
197 | * Restore per-cpu operation. smp_store_release() is paired with | |
9e804d1f TH |
198 | * smp_read_barrier_depends() in __ref_is_percpu() and guarantees |
199 | * that the zeroing is visible to all percpu accesses which can see | |
f47ad457 | 200 | * the following __PERCPU_REF_ATOMIC clearing. |
a2237370 TH |
201 | */ |
202 | for_each_possible_cpu(cpu) | |
eecc16ba | 203 | *per_cpu_ptr(percpu_count, cpu) = 0; |
a2237370 | 204 | |
eecc16ba | 205 | smp_store_release(&ref->percpu_count_ptr, |
f47ad457 TH |
206 | ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC); |
207 | } | |
208 | ||
3f49bdd9 TH |
209 | static void __percpu_ref_switch_mode(struct percpu_ref *ref, |
210 | percpu_ref_func_t *confirm_switch) | |
211 | { | |
33e465ce TH |
212 | lockdep_assert_held(&percpu_ref_switch_lock); |
213 | ||
3f49bdd9 TH |
214 | /* |
215 | * If the previous ATOMIC switching hasn't finished yet, wait for | |
216 | * its completion. If the caller ensures that ATOMIC switching | |
217 | * isn't in progress, this function can be called from any context. | |
3f49bdd9 | 218 | */ |
33e465ce TH |
219 | wait_event_lock_irq(percpu_ref_switch_waitq, !ref->confirm_switch, |
220 | percpu_ref_switch_lock); | |
3f49bdd9 TH |
221 | |
222 | if (ref->force_atomic || (ref->percpu_count_ptr & __PERCPU_REF_DEAD)) | |
223 | __percpu_ref_switch_to_atomic(ref, confirm_switch); | |
224 | else | |
225 | __percpu_ref_switch_to_percpu(ref); | |
226 | } | |
227 | ||
b2302c7f TH |
228 | /** |
229 | * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode | |
230 | * @ref: percpu_ref to switch to atomic mode | |
231 | * @confirm_switch: optional confirmation callback | |
232 | * | |
233 | * There's no reason to use this function for the usual reference counting. | |
234 | * Use percpu_ref_kill[_and_confirm](). | |
235 | * | |
236 | * Schedule switching of @ref to atomic mode. All its percpu counts will | |
237 | * be collected to the main atomic counter. On completion, when all CPUs | |
238 | * are guaraneed to be in atomic mode, @confirm_switch, which may not | |
239 | * block, is invoked. This function may be invoked concurrently with all | |
240 | * the get/put operations and can safely be mixed with kill and reinit | |
241 | * operations. Note that @ref will stay in atomic mode across kill/reinit | |
242 | * cycles until percpu_ref_switch_to_percpu() is called. | |
243 | * | |
3f49bdd9 TH |
244 | * This function may block if @ref is in the process of switching to atomic |
245 | * mode. If the caller ensures that @ref is not in the process of | |
246 | * switching to atomic mode, this function can be called from any context. | |
b2302c7f TH |
247 | */ |
248 | void percpu_ref_switch_to_atomic(struct percpu_ref *ref, | |
249 | percpu_ref_func_t *confirm_switch) | |
250 | { | |
33e465ce TH |
251 | unsigned long flags; |
252 | ||
253 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
254 | ||
b2302c7f | 255 | ref->force_atomic = true; |
3f49bdd9 | 256 | __percpu_ref_switch_mode(ref, confirm_switch); |
33e465ce TH |
257 | |
258 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
b2302c7f TH |
259 | } |
260 | ||
f47ad457 TH |
261 | /** |
262 | * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode | |
263 | * @ref: percpu_ref to switch to percpu mode | |
264 | * | |
265 | * There's no reason to use this function for the usual reference counting. | |
266 | * To re-use an expired ref, use percpu_ref_reinit(). | |
267 | * | |
268 | * Switch @ref to percpu mode. This function may be invoked concurrently | |
269 | * with all the get/put operations and can safely be mixed with kill and | |
1cae13e7 TH |
270 | * reinit operations. This function reverses the sticky atomic state set |
271 | * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic(). If @ref is | |
272 | * dying or dead, the actual switching takes place on the following | |
273 | * percpu_ref_reinit(). | |
f47ad457 | 274 | * |
3f49bdd9 TH |
275 | * This function may block if @ref is in the process of switching to atomic |
276 | * mode. If the caller ensures that @ref is not in the process of | |
277 | * switching to atomic mode, this function can be called from any context. | |
f47ad457 TH |
278 | */ |
279 | void percpu_ref_switch_to_percpu(struct percpu_ref *ref) | |
280 | { | |
33e465ce TH |
281 | unsigned long flags; |
282 | ||
283 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
284 | ||
1cae13e7 | 285 | ref->force_atomic = false; |
3f49bdd9 | 286 | __percpu_ref_switch_mode(ref, NULL); |
33e465ce TH |
287 | |
288 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
a2237370 | 289 | } |
490c79a6 TH |
290 | |
291 | /** | |
292 | * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation | |
293 | * @ref: percpu_ref to kill | |
294 | * @confirm_kill: optional confirmation callback | |
295 | * | |
296 | * Equivalent to percpu_ref_kill() but also schedules kill confirmation if | |
297 | * @confirm_kill is not NULL. @confirm_kill, which may not block, will be | |
298 | * called after @ref is seen as dead from all CPUs at which point all | |
299 | * further invocations of percpu_ref_tryget_live() will fail. See | |
300 | * percpu_ref_tryget_live() for details. | |
301 | * | |
302 | * This function normally doesn't block and can be called from any context | |
f47ad457 | 303 | * but it may block if @confirm_kill is specified and @ref is in the |
a2f5630c | 304 | * process of switching to atomic mode by percpu_ref_switch_to_atomic(). |
490c79a6 TH |
305 | */ |
306 | void percpu_ref_kill_and_confirm(struct percpu_ref *ref, | |
307 | percpu_ref_func_t *confirm_kill) | |
308 | { | |
33e465ce TH |
309 | unsigned long flags; |
310 | ||
311 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
312 | ||
490c79a6 TH |
313 | WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD, |
314 | "%s called more than once on %pf!", __func__, ref->release); | |
315 | ||
316 | ref->percpu_count_ptr |= __PERCPU_REF_DEAD; | |
3f49bdd9 | 317 | __percpu_ref_switch_mode(ref, confirm_kill); |
490c79a6 | 318 | percpu_ref_put(ref); |
33e465ce TH |
319 | |
320 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
490c79a6 TH |
321 | } |
322 | EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm); | |
f47ad457 TH |
323 | |
324 | /** | |
325 | * percpu_ref_reinit - re-initialize a percpu refcount | |
326 | * @ref: perpcu_ref to re-initialize | |
327 | * | |
328 | * Re-initialize @ref so that it's in the same state as when it finished | |
1cae13e7 TH |
329 | * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD. @ref must have been |
330 | * initialized successfully and reached 0 but not exited. | |
f47ad457 TH |
331 | * |
332 | * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while | |
333 | * this function is in progress. | |
334 | */ | |
335 | void percpu_ref_reinit(struct percpu_ref *ref) | |
336 | { | |
33e465ce TH |
337 | unsigned long flags; |
338 | ||
339 | spin_lock_irqsave(&percpu_ref_switch_lock, flags); | |
340 | ||
f47ad457 TH |
341 | WARN_ON_ONCE(!percpu_ref_is_zero(ref)); |
342 | ||
343 | ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD; | |
344 | percpu_ref_get(ref); | |
3f49bdd9 | 345 | __percpu_ref_switch_mode(ref, NULL); |
33e465ce TH |
346 | |
347 | spin_unlock_irqrestore(&percpu_ref_switch_lock, flags); | |
f47ad457 TH |
348 | } |
349 | EXPORT_SYMBOL_GPL(percpu_ref_reinit); |