]>
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 | |
e625305b TH |
28 | * percpu_ref_kill. Additionally, we need a bias value so that the |
29 | * atomic_long_t can't hit 0 before we've added up all the percpu refs. | |
215e262f KO |
30 | */ |
31 | ||
e625305b | 32 | #define PCPU_COUNT_BIAS (1LU << (BITS_PER_LONG - 1)) |
215e262f | 33 | |
e625305b | 34 | static unsigned long __percpu *pcpu_count_ptr(struct percpu_ref *ref) |
eae7975d | 35 | { |
e625305b | 36 | return (unsigned long __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 | |
a34375ef | 43 | * @gfp: allocation mask to use |
215e262f KO |
44 | * |
45 | * Initializes the refcount in single atomic counter mode with a refcount of 1; | |
e625305b | 46 | * analagous to atomic_long_set(ref, 1). |
215e262f KO |
47 | * |
48 | * Note that @release must not sleep - it may potentially be called from RCU | |
49 | * callback context by percpu_ref_kill(). | |
50 | */ | |
a34375ef TH |
51 | int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release, |
52 | gfp_t gfp) | |
215e262f | 53 | { |
e625305b | 54 | atomic_long_set(&ref->count, 1 + PCPU_COUNT_BIAS); |
215e262f | 55 | |
e625305b | 56 | ref->pcpu_count_ptr = (unsigned long)alloc_percpu_gfp(unsigned long, gfp); |
7d742075 | 57 | if (!ref->pcpu_count_ptr) |
215e262f KO |
58 | return -ENOMEM; |
59 | ||
60 | ref->release = release; | |
61 | return 0; | |
62 | } | |
5e9dd373 | 63 | EXPORT_SYMBOL_GPL(percpu_ref_init); |
215e262f | 64 | |
2d722782 TH |
65 | /** |
66 | * percpu_ref_reinit - re-initialize a percpu refcount | |
67 | * @ref: perpcu_ref to re-initialize | |
68 | * | |
69 | * Re-initialize @ref so that it's in the same state as when it finished | |
70 | * percpu_ref_init(). @ref must have been initialized successfully, killed | |
71 | * and reached 0 but not exited. | |
72 | * | |
73 | * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while | |
74 | * this function is in progress. | |
75 | */ | |
76 | void percpu_ref_reinit(struct percpu_ref *ref) | |
77 | { | |
e625305b | 78 | unsigned long __percpu *pcpu_count = pcpu_count_ptr(ref); |
2d722782 TH |
79 | int cpu; |
80 | ||
81 | BUG_ON(!pcpu_count); | |
82 | WARN_ON(!percpu_ref_is_zero(ref)); | |
83 | ||
e625305b | 84 | atomic_long_set(&ref->count, 1 + PCPU_COUNT_BIAS); |
2d722782 TH |
85 | |
86 | /* | |
87 | * Restore per-cpu operation. smp_store_release() is paired with | |
88 | * smp_read_barrier_depends() in __pcpu_ref_alive() and guarantees | |
89 | * that the zeroing is visible to all percpu accesses which can see | |
90 | * the following PCPU_REF_DEAD clearing. | |
91 | */ | |
92 | for_each_possible_cpu(cpu) | |
93 | *per_cpu_ptr(pcpu_count, cpu) = 0; | |
94 | ||
95 | smp_store_release(&ref->pcpu_count_ptr, | |
96 | ref->pcpu_count_ptr & ~PCPU_REF_DEAD); | |
97 | } | |
98 | EXPORT_SYMBOL_GPL(percpu_ref_reinit); | |
99 | ||
bc497bd3 | 100 | /** |
9a1049da TH |
101 | * percpu_ref_exit - undo percpu_ref_init() |
102 | * @ref: percpu_ref to exit | |
bc497bd3 | 103 | * |
9a1049da TH |
104 | * This function exits @ref. The caller is responsible for ensuring that |
105 | * @ref is no longer in active use. The usual places to invoke this | |
106 | * function from are the @ref->release() callback or in init failure path | |
107 | * where percpu_ref_init() succeeded but other parts of the initialization | |
108 | * of the embedding object failed. | |
bc497bd3 | 109 | */ |
9a1049da | 110 | void percpu_ref_exit(struct percpu_ref *ref) |
bc497bd3 | 111 | { |
e625305b | 112 | unsigned long __percpu *pcpu_count = pcpu_count_ptr(ref); |
bc497bd3 TH |
113 | |
114 | if (pcpu_count) { | |
eae7975d | 115 | free_percpu(pcpu_count); |
9a1049da | 116 | ref->pcpu_count_ptr = PCPU_REF_DEAD; |
bc497bd3 TH |
117 | } |
118 | } | |
9a1049da | 119 | EXPORT_SYMBOL_GPL(percpu_ref_exit); |
bc497bd3 | 120 | |
215e262f KO |
121 | static void percpu_ref_kill_rcu(struct rcu_head *rcu) |
122 | { | |
123 | struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu); | |
e625305b TH |
124 | unsigned long __percpu *pcpu_count = pcpu_count_ptr(ref); |
125 | unsigned long count = 0; | |
215e262f KO |
126 | int cpu; |
127 | ||
215e262f KO |
128 | for_each_possible_cpu(cpu) |
129 | count += *per_cpu_ptr(pcpu_count, cpu); | |
130 | ||
e625305b TH |
131 | pr_debug("global %ld pcpu %ld", |
132 | atomic_long_read(&ref->count), (long)count); | |
215e262f KO |
133 | |
134 | /* | |
135 | * It's crucial that we sum the percpu counters _before_ adding the sum | |
136 | * to &ref->count; since gets could be happening on one cpu while puts | |
137 | * happen on another, adding a single cpu's count could cause | |
138 | * @ref->count to hit 0 before we've got a consistent value - but the | |
139 | * sum of all the counts will be consistent and correct. | |
140 | * | |
141 | * Subtracting the bias value then has to happen _after_ adding count to | |
142 | * &ref->count; we need the bias value to prevent &ref->count from | |
143 | * reaching 0 before we add the percpu counts. But doing it at the same | |
144 | * time is equivalent and saves us atomic operations: | |
145 | */ | |
146 | ||
e625305b | 147 | atomic_long_add((long)count - PCPU_COUNT_BIAS, &ref->count); |
215e262f | 148 | |
e625305b TH |
149 | WARN_ONCE(atomic_long_read(&ref->count) <= 0, |
150 | "percpu ref (%pf) <= 0 (%ld) after killed", | |
151 | ref->release, atomic_long_read(&ref->count)); | |
687b0ad2 | 152 | |
dbece3a0 TH |
153 | /* @ref is viewed as dead on all CPUs, send out kill confirmation */ |
154 | if (ref->confirm_kill) | |
155 | ref->confirm_kill(ref); | |
156 | ||
215e262f KO |
157 | /* |
158 | * Now we're in single atomic_t mode with a consistent refcount, so it's | |
159 | * safe to drop our initial ref: | |
160 | */ | |
161 | percpu_ref_put(ref); | |
162 | } | |
163 | ||
164 | /** | |
dbece3a0 | 165 | * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation |
ac899061 | 166 | * @ref: percpu_ref to kill |
dbece3a0 | 167 | * @confirm_kill: optional confirmation callback |
215e262f | 168 | * |
dbece3a0 TH |
169 | * Equivalent to percpu_ref_kill() but also schedules kill confirmation if |
170 | * @confirm_kill is not NULL. @confirm_kill, which may not block, will be | |
171 | * called after @ref is seen as dead from all CPUs - all further | |
172 | * invocations of percpu_ref_tryget() will fail. See percpu_ref_tryget() | |
173 | * for more details. | |
215e262f | 174 | * |
dbece3a0 TH |
175 | * Due to the way percpu_ref is implemented, @confirm_kill will be called |
176 | * after at least one full RCU grace period has passed but this is an | |
177 | * implementation detail and callers must not depend on it. | |
215e262f | 178 | */ |
dbece3a0 TH |
179 | void percpu_ref_kill_and_confirm(struct percpu_ref *ref, |
180 | percpu_ref_func_t *confirm_kill) | |
215e262f | 181 | { |
7d742075 | 182 | WARN_ONCE(ref->pcpu_count_ptr & PCPU_REF_DEAD, |
4843c332 TH |
183 | "percpu_ref_kill() called more than once on %pf!", |
184 | ref->release); | |
215e262f | 185 | |
7d742075 | 186 | ref->pcpu_count_ptr |= PCPU_REF_DEAD; |
dbece3a0 | 187 | ref->confirm_kill = confirm_kill; |
215e262f | 188 | |
a4244454 | 189 | call_rcu_sched(&ref->rcu, percpu_ref_kill_rcu); |
215e262f | 190 | } |
5e9dd373 | 191 | EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm); |