1 #ifndef CEPH_CRUSH_CRUSH_H
2 #define CEPH_CRUSH_CRUSH_H
4 #include <linux/types.h>
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.
11 * The algorithm was originally described in detail in this paper
12 * (although the algorithm has evolved somewhat since then):
14 * http://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf
20 #define CRUSH_MAGIC 0x00010000ul /* for detecting algorithm revisions */
22 #define CRUSH_MAX_DEPTH 10 /* max crush hierarchy depth */
25 #define CRUSH_ITEM_UNDEF 0x7ffffffe /* undefined result (internal use only) */
26 #define CRUSH_ITEM_NONE 0x7fffffff /* no result */
29 * CRUSH uses user-defined "rules" to describe how inputs should be
30 * mapped to devices. A rule consists of sequence of steps to perform
31 * to generate the set of output devices.
33 struct crush_rule_step {
42 CRUSH_RULE_TAKE = 1, /* arg1 = value to start with */
43 CRUSH_RULE_CHOOSE_FIRSTN = 2, /* arg1 = num items to pick */
45 CRUSH_RULE_CHOOSE_INDEP = 3, /* same */
46 CRUSH_RULE_EMIT = 4, /* no args */
47 CRUSH_RULE_CHOOSE_LEAF_FIRSTN = 6,
48 CRUSH_RULE_CHOOSE_LEAF_INDEP = 7,
50 CRUSH_RULE_SET_CHOOSE_TRIES = 8, /* override choose_total_tries */
51 CRUSH_RULE_SET_CHOOSE_LEAF_TRIES = 9, /* override chooseleaf_descend_once */
55 * for specifying choose num (arg1) relative to the max parameter
58 #define CRUSH_CHOOSE_N 0
59 #define CRUSH_CHOOSE_N_MINUS(x) (-(x))
62 * The rule mask is used to describe what the rule is intended for.
63 * Given a ruleset and size of output set, we search through the
64 * rule list for a matching rule_mask.
66 struct crush_rule_mask {
75 struct crush_rule_mask mask;
76 struct crush_rule_step steps[0];
79 #define crush_rule_size(len) (sizeof(struct crush_rule) + \
80 (len)*sizeof(struct crush_rule_step))
85 * A bucket is a named container of other items (either devices or
86 * other buckets). Items within a bucket are chosen using one of a
87 * few different algorithms. The table summarizes how the speed of
88 * each option measures up against mapping stability when items are
91 * Bucket Alg Speed Additions Removals
92 * ------------------------------------------------
93 * uniform O(1) poor poor
94 * list O(n) optimal poor
95 * tree O(log n) good good
96 * straw O(n) optimal optimal
99 CRUSH_BUCKET_UNIFORM = 1,
100 CRUSH_BUCKET_LIST = 2,
101 CRUSH_BUCKET_TREE = 3,
102 CRUSH_BUCKET_STRAW = 4
104 extern const char *crush_bucket_alg_name(int alg);
106 struct crush_bucket {
107 __s32 id; /* this'll be negative */
108 __u16 type; /* non-zero; type=0 is reserved for devices */
109 __u8 alg; /* one of CRUSH_BUCKET_* */
110 __u8 hash; /* which hash function to use, CRUSH_HASH_* */
111 __u32 weight; /* 16-bit fixed point */
112 __u32 size; /* num items */
116 * cached random permutation: used for uniform bucket and for
117 * the linear search fallback for the other bucket types.
119 __u32 perm_x; /* @x for which *perm is defined */
120 __u32 perm_n; /* num elements of *perm that are permuted/defined */
124 struct crush_bucket_uniform {
125 struct crush_bucket h;
126 __u32 item_weight; /* 16-bit fixed point; all items equally weighted */
129 struct crush_bucket_list {
130 struct crush_bucket h;
131 __u32 *item_weights; /* 16-bit fixed point */
132 __u32 *sum_weights; /* 16-bit fixed point. element i is sum
133 of weights 0..i, inclusive */
136 struct crush_bucket_tree {
137 struct crush_bucket h; /* note: h.size is _tree_ size, not number of
143 struct crush_bucket_straw {
144 struct crush_bucket h;
145 __u32 *item_weights; /* 16-bit fixed point */
146 __u32 *straws; /* 16-bit fixed point */
152 * CRUSH map includes all buckets, rules, etc.
155 struct crush_bucket **buckets;
156 struct crush_rule **rules;
162 /* choose local retries before re-descent */
163 __u32 choose_local_tries;
164 /* choose local attempts using a fallback permutation before
166 __u32 choose_local_fallback_tries;
167 /* choose attempts before giving up */
168 __u32 choose_total_tries;
169 /* attempt chooseleaf inner descent once for firstn mode; on
170 * reject retry outer descent. Note that this does *not*
171 * apply to a collision: in that case we will retry as we used
173 __u32 chooseleaf_descend_once;
178 extern int crush_get_bucket_item_weight(const struct crush_bucket *b, int pos);
179 extern void crush_destroy_bucket_uniform(struct crush_bucket_uniform *b);
180 extern void crush_destroy_bucket_list(struct crush_bucket_list *b);
181 extern void crush_destroy_bucket_tree(struct crush_bucket_tree *b);
182 extern void crush_destroy_bucket_straw(struct crush_bucket_straw *b);
183 extern void crush_destroy_bucket(struct crush_bucket *b);
184 extern void crush_destroy_rule(struct crush_rule *r);
185 extern void crush_destroy(struct crush_map *map);
187 static inline int crush_calc_tree_node(int i)
189 return ((i+1) << 1)-1;