Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git...
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / Documentation / RCU / rcuref.txt
1 Reference-count design for elements of lists/arrays protected by RCU.
2
3 Reference counting on elements of lists which are protected by traditional
4 reader/writer spinlocks or semaphores are straightforward:
5
6 1. 2.
7 add() search_and_reference()
8 { {
9 alloc_object read_lock(&list_lock);
10 ... search_for_element
11 atomic_set(&el->rc, 1); atomic_inc(&el->rc);
12 write_lock(&list_lock); ...
13 add_element read_unlock(&list_lock);
14 ... ...
15 write_unlock(&list_lock); }
16 }
17
18 3. 4.
19 release_referenced() delete()
20 { {
21 ... write_lock(&list_lock);
22 atomic_dec(&el->rc, relfunc) ...
23 ... remove_element
24 } write_unlock(&list_lock);
25 ...
26 if (atomic_dec_and_test(&el->rc))
27 kfree(el);
28 ...
29 }
30
31 If this list/array is made lock free using RCU as in changing the
32 write_lock() in add() and delete() to spin_lock() and changing read_lock()
33 in search_and_reference() to rcu_read_lock(), the atomic_inc() in
34 search_and_reference() could potentially hold reference to an element which
35 has already been deleted from the list/array. Use atomic_inc_not_zero()
36 in this scenario as follows:
37
38 1. 2.
39 add() search_and_reference()
40 { {
41 alloc_object rcu_read_lock();
42 ... search_for_element
43 atomic_set(&el->rc, 1); if (!atomic_inc_not_zero(&el->rc)) {
44 spin_lock(&list_lock); rcu_read_unlock();
45 return FAIL;
46 add_element }
47 ... ...
48 spin_unlock(&list_lock); rcu_read_unlock();
49 } }
50 3. 4.
51 release_referenced() delete()
52 { {
53 ... spin_lock(&list_lock);
54 if (atomic_dec_and_test(&el->rc)) ...
55 call_rcu(&el->head, el_free); remove_element
56 ... spin_unlock(&list_lock);
57 } ...
58 if (atomic_dec_and_test(&el->rc))
59 call_rcu(&el->head, el_free);
60 ...
61 }
62
63 Sometimes, a reference to the element needs to be obtained in the
64 update (write) stream. In such cases, atomic_inc_not_zero() might be
65 overkill, since we hold the update-side spinlock. One might instead
66 use atomic_inc() in such cases.
67
68 It is not always convenient to deal with "FAIL" in the
69 search_and_reference() code path. In such cases, the
70 atomic_dec_and_test() may be moved from delete() to el_free()
71 as follows:
72
73 1. 2.
74 add() search_and_reference()
75 { {
76 alloc_object rcu_read_lock();
77 ... search_for_element
78 atomic_set(&el->rc, 1); atomic_inc(&el->rc);
79 spin_lock(&list_lock); ...
80
81 add_element rcu_read_unlock();
82 ... }
83 spin_unlock(&list_lock); 4.
84 } delete()
85 3. {
86 release_referenced() spin_lock(&list_lock);
87 { ...
88 ... remove_element
89 if (atomic_dec_and_test(&el->rc)) spin_unlock(&list_lock);
90 kfree(el); ...
91 ... call_rcu(&el->head, el_free);
92 } ...
93 5. }
94 void el_free(struct rcu_head *rhp)
95 {
96 release_referenced();
97 }
98
99 The key point is that the initial reference added by add() is not removed
100 until after a grace period has elapsed following removal. This means that
101 search_and_reference() cannot find this element, which means that the value
102 of el->rc cannot increase. Thus, once it reaches zero, there are no
103 readers that can or ever will be able to reference the element. The
104 element can therefore safely be freed. This in turn guarantees that if
105 any reader finds the element, that reader may safely acquire a reference
106 without checking the value of the reference counter.
107
108 In cases where delete() can sleep, synchronize_rcu() can be called from
109 delete(), so that el_free() can be subsumed into delete as follows:
110
111 4.
112 delete()
113 {
114 spin_lock(&list_lock);
115 ...
116 remove_element
117 spin_unlock(&list_lock);
118 ...
119 synchronize_rcu();
120 if (atomic_dec_and_test(&el->rc))
121 kfree(el);
122 ...
123 }