blob: 18f9651ff23d411e96737ec070d4fc6bc29c50fe [file] [log] [blame]
Eric Dumazet536533e2008-11-16 19:41:14 -08001Using hlist_nulls to protect read-mostly linked lists and
2objects using SLAB_DESTROY_BY_RCU allocations.
3
4Please read the basics in Documentation/RCU/listRCU.txt
5
6Using special makers (called 'nulls') is a convenient way
7to solve following problem :
8
9A typical RCU linked list managing objects which are
10allocated with SLAB_DESTROY_BY_RCU kmem_cache can
11use following algos :
12
131) Lookup algo
14--------------
15rcu_read_lock()
16begin:
17obj = lockless_lookup(key);
18if (obj) {
19 if (!try_get_ref(obj)) // might fail for free objects
20 goto begin;
21 /*
22 * Because a writer could delete object, and a writer could
23 * reuse these object before the RCU grace period, we
Jesper Dangaard Broueredd40702009-03-31 09:36:48 +000024 * must check key after getting the reference on object
Eric Dumazet536533e2008-11-16 19:41:14 -080025 */
26 if (obj->key != key) { // not the object we expected
27 put_ref(obj);
28 goto begin;
29 }
30}
31rcu_read_unlock();
32
33Beware that lockless_lookup(key) cannot use traditional hlist_for_each_entry_rcu()
34but a version with an additional memory barrier (smp_rmb())
35
36lockless_lookup(key)
37{
38 struct hlist_node *node, *next;
39 for (pos = rcu_dereference((head)->first);
40 pos && ({ next = pos->next; smp_rmb(); prefetch(next); 1; }) &&
41 ({ tpos = hlist_entry(pos, typeof(*tpos), member); 1; });
42 pos = rcu_dereference(next))
43 if (obj->key == key)
44 return obj;
45 return NULL;
46
47And note the traditional hlist_for_each_entry_rcu() misses this smp_rmb() :
48
49 struct hlist_node *node;
50 for (pos = rcu_dereference((head)->first);
51 pos && ({ prefetch(pos->next); 1; }) &&
52 ({ tpos = hlist_entry(pos, typeof(*tpos), member); 1; });
53 pos = rcu_dereference(pos->next))
54 if (obj->key == key)
55 return obj;
56 return NULL;
57}
58
59Quoting Corey Minyard :
60
61"If the object is moved from one list to another list in-between the
62 time the hash is calculated and the next field is accessed, and the
63 object has moved to the end of a new list, the traversal will not
64 complete properly on the list it should have, since the object will
65 be on the end of the new list and there's not a way to tell it's on a
66 new list and restart the list traversal. I think that this can be
67 solved by pre-fetching the "next" field (with proper barriers) before
68 checking the key."
69
702) Insert algo :
71----------------
72
73We need to make sure a reader cannot read the new 'obj->obj_next' value
74and previous value of 'obj->key'. Or else, an item could be deleted
75from a chain, and inserted into another chain. If new chain was empty
76before the move, 'next' pointer is NULL, and lockless reader can
77not detect it missed following items in original chain.
78
79/*
80 * Please note that new inserts are done at the head of list,
81 * not in the middle or end.
82 */
83obj = kmem_cache_alloc(...);
84lock_chain(); // typically a spin_lock()
85obj->key = key;
Eric Dumazet536533e2008-11-16 19:41:14 -080086/*
87 * we need to make sure obj->key is updated before obj->next
Eric Dumazet941297f2009-07-16 14:03:40 +020088 * or obj->refcnt
Eric Dumazet536533e2008-11-16 19:41:14 -080089 */
90smp_wmb();
Eric Dumazet941297f2009-07-16 14:03:40 +020091atomic_set(&obj->refcnt, 1);
Eric Dumazet536533e2008-11-16 19:41:14 -080092hlist_add_head_rcu(&obj->obj_node, list);
93unlock_chain(); // typically a spin_unlock()
94
95
963) Remove algo
97--------------
98Nothing special here, we can use a standard RCU hlist deletion.
99But thanks to SLAB_DESTROY_BY_RCU, beware a deleted object can be reused
100very very fast (before the end of RCU grace period)
101
102if (put_last_reference_on(obj) {
103 lock_chain(); // typically a spin_lock()
104 hlist_del_init_rcu(&obj->obj_node);
105 unlock_chain(); // typically a spin_unlock()
106 kmem_cache_free(cachep, obj);
107}
108
109
110
111--------------------------------------------------------------------------
112With hlist_nulls we can avoid extra smp_rmb() in lockless_lookup()
113and extra smp_wmb() in insert function.
114
115For example, if we choose to store the slot number as the 'nulls'
116end-of-list marker for each slot of the hash table, we can detect
117a race (some writer did a delete and/or a move of an object
118to another chain) checking the final 'nulls' value if
119the lookup met the end of chain. If final 'nulls' value
120is not the slot number, then we must restart the lookup at
Jesper Dangaard Broueredd40702009-03-31 09:36:48 +0000121the beginning. If the object was moved to the same chain,
Matt LaPlante19f59462009-04-27 15:06:31 +0200122then the reader doesn't care : It might eventually
Eric Dumazet536533e2008-11-16 19:41:14 -0800123scan the list again without harm.
124
125
1261) lookup algo
127
128 head = &table[slot];
129 rcu_read_lock();
130begin:
131 hlist_nulls_for_each_entry_rcu(obj, node, head, member) {
132 if (obj->key == key) {
133 if (!try_get_ref(obj)) // might fail for free objects
134 goto begin;
135 if (obj->key != key) { // not the object we expected
136 put_ref(obj);
137 goto begin;
138 }
139 goto out;
140 }
141/*
142 * if the nulls value we got at the end of this lookup is
143 * not the expected one, we must restart lookup.
144 * We probably met an item that was moved to another chain.
145 */
146 if (get_nulls_value(node) != slot)
147 goto begin;
148 obj = NULL;
149
150out:
151 rcu_read_unlock();
152
1532) Insert function :
154--------------------
155
156/*
157 * Please note that new inserts are done at the head of list,
158 * not in the middle or end.
159 */
160obj = kmem_cache_alloc(cachep);
161lock_chain(); // typically a spin_lock()
162obj->key = key;
Eric Dumazet941297f2009-07-16 14:03:40 +0200163/*
164 * changes to obj->key must be visible before refcnt one
165 */
166smp_wmb();
Eric Dumazet536533e2008-11-16 19:41:14 -0800167atomic_set(&obj->refcnt, 1);
168/*
169 * insert obj in RCU way (readers might be traversing chain)
170 */
171hlist_nulls_add_head_rcu(&obj->obj_node, list);
172unlock_chain(); // typically a spin_unlock()