Commit | Line | Data |
---|---|---|
215e262f KO |
1 | #define pr_fmt(fmt) "%s: " fmt "\n", __func__ |
2 | ||
3 | #include <linux/kernel.h> | |
4 | #include <linux/percpu-refcount.h> | |
5 | ||
6 | /* | |
7 | * Initially, a percpu refcount is just a set of percpu counters. Initially, we | |
8 | * don't try to detect the ref hitting 0 - which means that get/put can just | |
9 | * increment or decrement the local counter. Note that the counter on a | |
10 | * particular cpu can (and will) wrap - this is fine, when we go to shutdown the | |
11 | * percpu counters will all sum to the correct value | |
12 | * | |
13 | * (More precisely: because moduler arithmatic is commutative the sum of all the | |
14 | * pcpu_count vars will be equal to what it would have been if all the gets and | |
15 | * puts were done to a single integer, even if some of the percpu integers | |
16 | * overflow or underflow). | |
17 | * | |
18 | * The real trick to implementing percpu refcounts is shutdown. We can't detect | |
19 | * the ref hitting 0 on every put - this would require global synchronization | |
20 | * and defeat the whole purpose of using percpu refs. | |
21 | * | |
22 | * What we do is require the user to keep track of the initial refcount; we know | |
23 | * the ref can't hit 0 before the user drops the initial ref, so as long as we | |
24 | * convert to non percpu mode before the initial ref is dropped everything | |
25 | * works. | |
26 | * | |
27 | * Converting to non percpu mode is done with some RCUish stuff in | |
28 | * percpu_ref_kill. Additionally, we need a bias value so that the atomic_t | |
29 | * can't hit 0 before we've added up all the percpu refs. | |
30 | */ | |
31 | ||
32 | #define PCPU_COUNT_BIAS (1U << 31) | |
33 | ||
eae7975d TH |
34 | static unsigned __percpu *pcpu_count_ptr(struct percpu_ref *ref) |
35 | { | |
7d742075 | 36 | return (unsigned __percpu *)(ref->pcpu_count_ptr & ~PCPU_REF_DEAD); |
eae7975d TH |
37 | } |
38 | ||
215e262f KO |
39 | /** |
40 | * percpu_ref_init - initialize a percpu refcount | |
ac899061 TH |
41 | * @ref: percpu_ref to initialize |
42 | * @release: function which will be called when refcount hits 0 | |
215e262f KO |
43 | * |
44 | * Initializes the refcount in single atomic counter mode with a refcount of 1; | |
45 | * analagous to atomic_set(ref, 1). | |
46 | * | |
47 | * Note that @release must not sleep - it may potentially be called from RCU | |
48 | * callback context by percpu_ref_kill(). | |
49 | */ | |
ac899061 | 50 | int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release) |
215e262f KO |
51 | { |
52 | atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS); | |
53 | ||
7d742075 TH |
54 | ref->pcpu_count_ptr = (unsigned long)alloc_percpu(unsigned); |
55 | if (!ref->pcpu_count_ptr) | |
215e262f KO |
56 | return -ENOMEM; |
57 | ||
58 | ref->release = release; | |
59 | return 0; | |
60 | } | |
5e9dd373 | 61 | EXPORT_SYMBOL_GPL(percpu_ref_init); |
215e262f | 62 | |
2d722782 TH |
63 | /** |
64 | * percpu_ref_reinit - re-initialize a percpu refcount | |
65 | * @ref: perpcu_ref to re-initialize | |
66 | * | |
67 | * Re-initialize @ref so that it's in the same state as when it finished | |
68 | * percpu_ref_init(). @ref must have been initialized successfully, killed | |
69 | * and reached 0 but not exited. | |
70 | * | |
71 | * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while | |
72 | * this function is in progress. | |
73 | */ | |
74 | void percpu_ref_reinit(struct percpu_ref *ref) | |
75 | { | |
76 | unsigned __percpu *pcpu_count = pcpu_count_ptr(ref); | |
77 | int cpu; | |
78 | ||
79 | BUG_ON(!pcpu_count); | |
80 | WARN_ON(!percpu_ref_is_zero(ref)); | |
81 | ||
82 | atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS); | |
83 | ||
84 | /* | |
85 | * Restore per-cpu operation. smp_store_release() is paired with | |
86 | * smp_read_barrier_depends() in __pcpu_ref_alive() and guarantees | |
87 | * that the zeroing is visible to all percpu accesses which can see | |
88 | * the following PCPU_REF_DEAD clearing. | |
89 | */ | |
90 | for_each_possible_cpu(cpu) | |
91 | *per_cpu_ptr(pcpu_count, cpu) = 0; | |
92 | ||
93 | smp_store_release(&ref->pcpu_count_ptr, | |
94 | ref->pcpu_count_ptr & ~PCPU_REF_DEAD); | |
95 | } | |
96 | EXPORT_SYMBOL_GPL(percpu_ref_reinit); | |
97 | ||
bc497bd3 | 98 | /** |
9a1049da TH |
99 | * percpu_ref_exit - undo percpu_ref_init() |
100 | * @ref: percpu_ref to exit | |
bc497bd3 | 101 | * |
9a1049da TH |
102 | * This function exits @ref. The caller is responsible for ensuring that |
103 | * @ref is no longer in active use. The usual places to invoke this | |
104 | * function from are the @ref->release() callback or in init failure path | |
105 | * where percpu_ref_init() succeeded but other parts of the initialization | |
106 | * of the embedding object failed. | |
bc497bd3 | 107 | */ |
9a1049da | 108 | void percpu_ref_exit(struct percpu_ref *ref) |
bc497bd3 | 109 | { |
eae7975d | 110 | unsigned __percpu *pcpu_count = pcpu_count_ptr(ref); |
bc497bd3 TH |
111 | |
112 | if (pcpu_count) { | |
eae7975d | 113 | free_percpu(pcpu_count); |
9a1049da | 114 | ref->pcpu_count_ptr = PCPU_REF_DEAD; |
bc497bd3 TH |
115 | } |
116 | } | |
9a1049da | 117 | EXPORT_SYMBOL_GPL(percpu_ref_exit); |
bc497bd3 | 118 | |
215e262f KO |
119 | static void percpu_ref_kill_rcu(struct rcu_head *rcu) |
120 | { | |
121 | struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu); | |
eae7975d | 122 | unsigned __percpu *pcpu_count = pcpu_count_ptr(ref); |
215e262f KO |
123 | unsigned count = 0; |
124 | int cpu; | |
125 | ||
215e262f KO |
126 | for_each_possible_cpu(cpu) |
127 | count += *per_cpu_ptr(pcpu_count, cpu); | |
128 | ||
215e262f KO |
129 | pr_debug("global %i pcpu %i", atomic_read(&ref->count), (int) count); |
130 | ||
131 | /* | |
132 | * It's crucial that we sum the percpu counters _before_ adding the sum | |
133 | * to &ref->count; since gets could be happening on one cpu while puts | |
134 | * happen on another, adding a single cpu's count could cause | |
135 | * @ref->count to hit 0 before we've got a consistent value - but the | |
136 | * sum of all the counts will be consistent and correct. | |
137 | * | |
138 | * Subtracting the bias value then has to happen _after_ adding count to | |
139 | * &ref->count; we need the bias value to prevent &ref->count from | |
140 | * reaching 0 before we add the percpu counts. But doing it at the same | |
141 | * time is equivalent and saves us atomic operations: | |
142 | */ | |
143 | ||
144 | atomic_add((int) count - PCPU_COUNT_BIAS, &ref->count); | |
145 | ||
687b0ad2 KO |
146 | WARN_ONCE(atomic_read(&ref->count) <= 0, "percpu ref <= 0 (%i)", |
147 | atomic_read(&ref->count)); | |
148 | ||
dbece3a0 TH |
149 | /* @ref is viewed as dead on all CPUs, send out kill confirmation */ |
150 | if (ref->confirm_kill) | |
151 | ref->confirm_kill(ref); | |
152 | ||
215e262f KO |
153 | /* |
154 | * Now we're in single atomic_t mode with a consistent refcount, so it's | |
155 | * safe to drop our initial ref: | |
156 | */ | |
157 | percpu_ref_put(ref); | |
158 | } | |
159 | ||
160 | /** | |
dbece3a0 | 161 | * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation |
ac899061 | 162 | * @ref: percpu_ref to kill |
dbece3a0 | 163 | * @confirm_kill: optional confirmation callback |
215e262f | 164 | * |
dbece3a0 TH |
165 | * Equivalent to percpu_ref_kill() but also schedules kill confirmation if |
166 | * @confirm_kill is not NULL. @confirm_kill, which may not block, will be | |
167 | * called after @ref is seen as dead from all CPUs - all further | |
168 | * invocations of percpu_ref_tryget() will fail. See percpu_ref_tryget() | |
169 | * for more details. | |
215e262f | 170 | * |
dbece3a0 TH |
171 | * Due to the way percpu_ref is implemented, @confirm_kill will be called |
172 | * after at least one full RCU grace period has passed but this is an | |
173 | * implementation detail and callers must not depend on it. | |
215e262f | 174 | */ |
dbece3a0 TH |
175 | void percpu_ref_kill_and_confirm(struct percpu_ref *ref, |
176 | percpu_ref_func_t *confirm_kill) | |
215e262f | 177 | { |
7d742075 | 178 | WARN_ONCE(ref->pcpu_count_ptr & PCPU_REF_DEAD, |
c1ae6e9b | 179 | "percpu_ref_kill() called more than once!\n"); |
215e262f | 180 | |
7d742075 | 181 | ref->pcpu_count_ptr |= PCPU_REF_DEAD; |
dbece3a0 | 182 | ref->confirm_kill = confirm_kill; |
215e262f | 183 | |
a4244454 | 184 | call_rcu_sched(&ref->rcu, percpu_ref_kill_rcu); |
215e262f | 185 | } |
5e9dd373 | 186 | EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm); |
0a30288d TH |
187 | |
188 | /* | |
189 | * XXX: Temporary kludge to work around SCSI blk-mq stall. Used only by | |
190 | * block/blk-mq.c::blk_mq_freeze_queue(). Will be removed during v3.18 | |
191 | * devel cycle. Do not use anywhere else. | |
192 | */ | |
193 | void __percpu_ref_kill_expedited(struct percpu_ref *ref) | |
194 | { | |
195 | WARN_ONCE(ref->pcpu_count_ptr & PCPU_REF_DEAD, | |
196 | "percpu_ref_kill() called more than once on %pf!", | |
197 | ref->release); | |
198 | ||
199 | ref->pcpu_count_ptr |= PCPU_REF_DEAD; | |
200 | synchronize_sched_expedited(); | |
201 | percpu_ref_kill_rcu(&ref->rcu); | |
202 | } |