1 #ifndef CEPH_CRUSH_CRUSH_H
2 #define CEPH_CRUSH_CRUSH_H
3
4 #include <linux/types.h>
5
6 /*
7 * CRUSH is a pseudo-random data distribution algorithm that
8 * efficiently distributes input values (typically, data objects)
9 * across a heterogeneous, structured storage cluster.
10 *
11 * The algorithm was originally described in detail in this paper
12 * (although the algorithm has evolved somewhat since then):
13 *
14 * http://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf
15 *
16 * LGPL2
17 */
18
19
20 #define CRUSH_MAGIC 0x00010000ul /* for detecting algorithm revisions */
21
22
23 #define CRUSH_MAX_DEPTH 10 /* max crush hierarchy depth */
24 #define CRUSH_MAX_SET 10 /* max size of a mapping result */
25
26
27 /*
28 * CRUSH uses user-defined "rules" to describe how inputs should be
29 * mapped to devices. A rule consists of sequence of steps to perform
30 * to generate the set of output devices.
31 */
32 struct crush_rule_step {
33 __u32 op;
34 __s32 arg1;
35 __s32 arg2;
36 };
37
38 /* step op codes */
39 enum {
40 CRUSH_RULE_NOOP = 0,
41 CRUSH_RULE_TAKE = 1, /* arg1 = value to start with */
42 CRUSH_RULE_CHOOSE_FIRSTN = 2, /* arg1 = num items to pick */
43 /* arg2 = type */
44 CRUSH_RULE_CHOOSE_INDEP = 3, /* same */
45 CRUSH_RULE_EMIT = 4, /* no args */
46 CRUSH_RULE_CHOOSE_LEAF_FIRSTN = 6,
47 CRUSH_RULE_CHOOSE_LEAF_INDEP = 7,
48 };
49
50 /*
51 * for specifying choose num (arg1) relative to the max parameter
52 * passed to do_rule
53 */
54 #define CRUSH_CHOOSE_N 0
55 #define CRUSH_CHOOSE_N_MINUS(x) (-(x))
56
57 /*
58 * The rule mask is used to describe what the rule is intended for.
59 * Given a ruleset and size of output set, we search through the
60 * rule list for a matching rule_mask.
61 */
62 struct crush_rule_mask {
63 __u8 ruleset;
64 __u8 type;
65 __u8 min_size;
66 __u8 max_size;
67 };
68
69 struct crush_rule {
70 __u32 len;
71 struct crush_rule_mask mask;
72 struct crush_rule_step steps[0];
73 };
74
75 #define crush_rule_size(len) (sizeof(struct crush_rule) + \
76 (len)*sizeof(struct crush_rule_step))
77
78
79
80 /*
81 * A bucket is a named container of other items (either devices or
82 * other buckets). Items within a bucket are chosen using one of a
83 * few different algorithms. The table summarizes how the speed of
84 * each option measures up against mapping stability when items are
85 * added or removed.
86 *
87 * Bucket Alg Speed Additions Removals
88 * ------------------------------------------------
89 * uniform O(1) poor poor
90 * list O(n) optimal poor
91 * tree O(log n) good good
92 * straw O(n) optimal optimal
93 */
94 enum {
95 CRUSH_BUCKET_UNIFORM = 1,
96 CRUSH_BUCKET_LIST = 2,
97 CRUSH_BUCKET_TREE = 3,
98 CRUSH_BUCKET_STRAW = 4
99 };
100 extern const char *crush_bucket_alg_name(int alg);
101
102 struct crush_bucket {
103 __s32 id; /* this'll be negative */
104 __u16 type; /* non-zero; type=0 is reserved for devices */
105 __u8 alg; /* one of CRUSH_BUCKET_* */
106 __u8 hash; /* which hash function to use, CRUSH_HASH_* */
107 __u32 weight; /* 16-bit fixed point */
108 __u32 size; /* num items */
109 __s32 *items;
110
111 /*
112 * cached random permutation: used for uniform bucket and for
113 * the linear search fallback for the other bucket types.
114 */
115 __u32 perm_x; /* @x for which *perm is defined */
116 __u32 perm_n; /* num elements of *perm that are permuted/defined */
117 __u32 *perm;
118 };
119
120 struct crush_bucket_uniform {
121 struct crush_bucket h;
122 __u32 item_weight; /* 16-bit fixed point; all items equally weighted */
123 };
124
125 struct crush_bucket_list {
126 struct crush_bucket h;
127 __u32 *item_weights; /* 16-bit fixed point */
128 __u32 *sum_weights; /* 16-bit fixed point. element i is sum
129 of weights 0..i, inclusive */
130 };
131
132 struct crush_bucket_tree {
133 struct crush_bucket h; /* note: h.size is _tree_ size, not number of
134 actual items */
135 __u8 num_nodes;
136 __u32 *node_weights;
137 };
138
139 struct crush_bucket_straw {
140 struct crush_bucket h;
141 __u32 *item_weights; /* 16-bit fixed point */
142 __u32 *straws; /* 16-bit fixed point */
143 };
144
145
146
147 /*
148 * CRUSH map includes all buckets, rules, etc.
149 */
150 struct crush_map {
151 struct crush_bucket **buckets;
152 struct crush_rule **rules;
153
154 /*
155 * Parent pointers to identify the parent bucket a device or
156 * bucket in the hierarchy. If an item appears more than
157 * once, this is the _last_ time it appeared (where buckets
158 * are processed in bucket id order, from -1 on down to
159 * -max_buckets.
160 */
161 __u32 *bucket_parents;
162 __u32 *device_parents;
163
164 __s32 max_buckets;
165 __u32 max_rules;
166 __s32 max_devices;
167 };
168
169
170 /* crush.c */
171 extern int crush_get_bucket_item_weight(const struct crush_bucket *b, int pos);
172 extern void crush_calc_parents(struct crush_map *map);
173 extern void crush_destroy_bucket_uniform(struct crush_bucket_uniform *b);
174 extern void crush_destroy_bucket_list(struct crush_bucket_list *b);
175 extern void crush_destroy_bucket_tree(struct crush_bucket_tree *b);
176 extern void crush_destroy_bucket_straw(struct crush_bucket_straw *b);
177 extern void crush_destroy_bucket(struct crush_bucket *b);
178 extern void crush_destroy(struct crush_map *map);
179
crush_calc_tree_node(int i)180 static inline int crush_calc_tree_node(int i)
181 {
182 return ((i+1) << 1)-1;
183 }
184
185 #endif
186