blob: 7deeb6297a483d7272acb30c4f3b0c0d6b0e76b3 [file] [log] [blame]
Kent Overstreet215e2622013-05-31 15:26:45 -07001#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
34/**
35 * percpu_ref_init - initialize a percpu refcount
Tejun Heoac899062013-06-12 20:43:06 -070036 * @ref: percpu_ref to initialize
37 * @release: function which will be called when refcount hits 0
Kent Overstreet215e2622013-05-31 15:26:45 -070038 *
39 * Initializes the refcount in single atomic counter mode with a refcount of 1;
40 * analagous to atomic_set(ref, 1).
41 *
42 * Note that @release must not sleep - it may potentially be called from RCU
43 * callback context by percpu_ref_kill().
44 */
Tejun Heoac899062013-06-12 20:43:06 -070045int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release)
Kent Overstreet215e2622013-05-31 15:26:45 -070046{
47 atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS);
48
49 ref->pcpu_count = alloc_percpu(unsigned);
50 if (!ref->pcpu_count)
51 return -ENOMEM;
52
53 ref->release = release;
54 return 0;
55}
56
Tejun Heobc497bd2013-06-12 20:52:35 -070057/**
58 * percpu_ref_cancel_init - cancel percpu_ref_init()
59 * @ref: percpu_ref to cancel init for
60 *
61 * Once a percpu_ref is initialized, its destruction is initiated by
62 * percpu_ref_kill() and completes asynchronously, which can be painful to
63 * do when destroying a half-constructed object in init failure path.
64 *
65 * This function destroys @ref without invoking @ref->release and the
66 * memory area containing it can be freed immediately on return. To
67 * prevent accidental misuse, it's required that @ref has finished
68 * percpu_ref_init(), whether successful or not, but never used.
69 *
70 * The weird name and usage restriction are to prevent people from using
71 * this function by mistake for normal shutdown instead of
72 * percpu_ref_kill().
73 */
74void percpu_ref_cancel_init(struct percpu_ref *ref)
75{
76 unsigned __percpu *pcpu_count = ref->pcpu_count;
77 int cpu;
78
79 WARN_ON_ONCE(atomic_read(&ref->count) != 1 + PCPU_COUNT_BIAS);
80
81 if (pcpu_count) {
82 for_each_possible_cpu(cpu)
83 WARN_ON_ONCE(*per_cpu_ptr(pcpu_count, cpu));
84 free_percpu(ref->pcpu_count);
85 }
86}
87
Kent Overstreet215e2622013-05-31 15:26:45 -070088static void percpu_ref_kill_rcu(struct rcu_head *rcu)
89{
90 struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
Tejun Heoacac7882013-06-12 20:52:01 -070091 unsigned __percpu *pcpu_count = ref->pcpu_count;
Kent Overstreet215e2622013-05-31 15:26:45 -070092 unsigned count = 0;
93 int cpu;
94
Kent Overstreet215e2622013-05-31 15:26:45 -070095 /* Mask out PCPU_REF_DEAD */
96 pcpu_count = (unsigned __percpu *)
97 (((unsigned long) pcpu_count) & ~PCPU_STATUS_MASK);
98
99 for_each_possible_cpu(cpu)
100 count += *per_cpu_ptr(pcpu_count, cpu);
101
102 free_percpu(pcpu_count);
103
104 pr_debug("global %i pcpu %i", atomic_read(&ref->count), (int) count);
105
106 /*
107 * It's crucial that we sum the percpu counters _before_ adding the sum
108 * to &ref->count; since gets could be happening on one cpu while puts
109 * happen on another, adding a single cpu's count could cause
110 * @ref->count to hit 0 before we've got a consistent value - but the
111 * sum of all the counts will be consistent and correct.
112 *
113 * Subtracting the bias value then has to happen _after_ adding count to
114 * &ref->count; we need the bias value to prevent &ref->count from
115 * reaching 0 before we add the percpu counts. But doing it at the same
116 * time is equivalent and saves us atomic operations:
117 */
118
119 atomic_add((int) count - PCPU_COUNT_BIAS, &ref->count);
120
Tejun Heodbece3a2013-06-13 19:23:53 -0700121 /* @ref is viewed as dead on all CPUs, send out kill confirmation */
122 if (ref->confirm_kill)
123 ref->confirm_kill(ref);
124
Kent Overstreet215e2622013-05-31 15:26:45 -0700125 /*
126 * Now we're in single atomic_t mode with a consistent refcount, so it's
127 * safe to drop our initial ref:
128 */
129 percpu_ref_put(ref);
130}
131
132/**
Tejun Heodbece3a2013-06-13 19:23:53 -0700133 * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
Tejun Heoac899062013-06-12 20:43:06 -0700134 * @ref: percpu_ref to kill
Tejun Heodbece3a2013-06-13 19:23:53 -0700135 * @confirm_kill: optional confirmation callback
Kent Overstreet215e2622013-05-31 15:26:45 -0700136 *
Tejun Heodbece3a2013-06-13 19:23:53 -0700137 * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
138 * @confirm_kill is not NULL. @confirm_kill, which may not block, will be
139 * called after @ref is seen as dead from all CPUs - all further
140 * invocations of percpu_ref_tryget() will fail. See percpu_ref_tryget()
141 * for more details.
Kent Overstreet215e2622013-05-31 15:26:45 -0700142 *
Tejun Heodbece3a2013-06-13 19:23:53 -0700143 * Due to the way percpu_ref is implemented, @confirm_kill will be called
144 * after at least one full RCU grace period has passed but this is an
145 * implementation detail and callers must not depend on it.
Kent Overstreet215e2622013-05-31 15:26:45 -0700146 */
Tejun Heodbece3a2013-06-13 19:23:53 -0700147void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
148 percpu_ref_func_t *confirm_kill)
Kent Overstreet215e2622013-05-31 15:26:45 -0700149{
Kent Overstreetc1ae6e92013-06-03 16:02:29 -0700150 WARN_ONCE(REF_STATUS(ref->pcpu_count) == PCPU_REF_DEAD,
151 "percpu_ref_kill() called more than once!\n");
Kent Overstreet215e2622013-05-31 15:26:45 -0700152
Kent Overstreetc1ae6e92013-06-03 16:02:29 -0700153 ref->pcpu_count = (unsigned __percpu *)
154 (((unsigned long) ref->pcpu_count)|PCPU_REF_DEAD);
Tejun Heodbece3a2013-06-13 19:23:53 -0700155 ref->confirm_kill = confirm_kill;
Kent Overstreet215e2622013-05-31 15:26:45 -0700156
Tejun Heoa4244452013-06-16 16:12:26 -0700157 call_rcu_sched(&ref->rcu, percpu_ref_kill_rcu);
Kent Overstreet215e2622013-05-31 15:26:45 -0700158}