1/* 2 * CDDL HEADER START 3 * 4 * This file and its contents are supplied under the terms of the 5 * Common Development and Distribution License ("CDDL"), version 1.0. 6 * You may only use this file in accordance with the terms of version 7 * 1.0 of the CDDL. 8 * 9 * A full copy of the text of the CDDL should have accompanied this 10 * source. A copy of the CDDL is also available via the Internet at 11 * http://www.illumos.org/license/CDDL. 12 * 13 * CDDL HEADER END 14 */ 15/* 16 * Copyright (c) 2019 by Delphix. All rights reserved. 17 */ 18 19#ifndef _BTREE_H 20#define _BTREE_H 21 22#ifdef __cplusplus 23extern "C" { 24#endif 25 26#include <sys/zfs_context.h> 27 28/* 29 * This file defines the interface for a B-Tree implementation for ZFS. The 30 * tree can be used to store arbitrary sortable data types with low overhead 31 * and good operation performance. In addition the tree intelligently 32 * optimizes bulk in-order insertions to improve memory use and performance. 33 * 34 * Note that for all B-Tree functions, the values returned are pointers to the 35 * internal copies of the data in the tree. The internal data can only be 36 * safely mutated if the changes cannot change the ordering of the element 37 * with respect to any other elements in the tree. 38 * 39 * The major drawback of the B-Tree is that any returned elements or indexes 40 * are only valid until a side-effectful operation occurs, since these can 41 * result in reallocation or relocation of data. Side effectful operations are 42 * defined as insertion, removal, and zfs_btree_destroy_nodes. 43 * 44 * The B-Tree has two types of nodes: core nodes, and leaf nodes. Core 45 * nodes have an array of children pointing to other nodes, and an array of 46 * elements that act as separators between the elements of the subtrees rooted 47 * at its children. Leaf nodes only contain data elements, and form the bottom 48 * layer of the tree. Unlike B+ Trees, in this B-Tree implementation the 49 * elements in the core nodes are not copies of or references to leaf node 50 * elements. Each element occurs only once in the tree, no matter what kind 51 * of node it is in. 52 * 53 * The tree's height is the same throughout, unlike many other forms of search 54 * tree. Each node (except for the root) must be between half minus one and 55 * completely full of elements (and children) at all times. Any operation that 56 * would put the node outside of that range results in a rebalancing operation 57 * (taking, merging, or splitting). 58 * 59 * This tree was implemented using descriptions from Wikipedia's articles on 60 * B-Trees and B+ Trees. 61 */ 62 63/* 64 * Decreasing these values results in smaller memmove operations, but more of 65 * them, and increased memory overhead. Increasing these values results in 66 * higher variance in operation time, and reduces memory overhead. 67 */ 68#define BTREE_CORE_ELEMS 126 69#define BTREE_LEAF_SIZE 4096 70 71extern kmem_cache_t *zfs_btree_leaf_cache; 72 73typedef struct zfs_btree_hdr { 74 struct zfs_btree_core *bth_parent; 75 /* 76 * Set to -1 to indicate core nodes. Other values represent first 77 * valid element offset for leaf nodes. 78 */ 79 uint32_t bth_first; 80 /* 81 * For both leaf and core nodes, represents the number of elements in 82 * the node. For core nodes, they will have bth_count + 1 children. 83 */ 84 uint32_t bth_count; 85} zfs_btree_hdr_t; 86 87typedef struct zfs_btree_core { 88 zfs_btree_hdr_t btc_hdr; 89 zfs_btree_hdr_t *btc_children[BTREE_CORE_ELEMS + 1]; 90 uint8_t btc_elems[]; 91} zfs_btree_core_t; 92 93typedef struct zfs_btree_leaf { 94 zfs_btree_hdr_t btl_hdr; 95 uint8_t btl_elems[]; 96} zfs_btree_leaf_t; 97 98typedef struct zfs_btree_index { 99 zfs_btree_hdr_t *bti_node; 100 uint32_t bti_offset; 101 /* 102 * True if the location is before the list offset, false if it's at 103 * the listed offset. 104 */ 105 boolean_t bti_before; 106} zfs_btree_index_t; 107 108typedef struct btree zfs_btree_t; 109typedef void * (*bt_find_in_buf_f) (zfs_btree_t *, uint8_t *, uint32_t, 110 const void *, zfs_btree_index_t *); 111 112struct btree { 113 int (*bt_compar) (const void *, const void *); 114 bt_find_in_buf_f bt_find_in_buf; 115 size_t bt_elem_size; 116 size_t bt_leaf_size; 117 uint32_t bt_leaf_cap; 118 int32_t bt_height; 119 uint64_t bt_num_elems; 120 uint64_t bt_num_nodes; 121 zfs_btree_hdr_t *bt_root; 122 zfs_btree_leaf_t *bt_bulk; // non-null if bulk loading 123}; 124 125/* 126 * Implementation of Shar's algorithm designed to accelerate binary search by 127 * eliminating impossible to predict branches. 128 * 129 * For optimality, this should be used to generate the search function in the 130 * same file as the comparator and the comparator should be marked 131 * `__attribute__((always_inline) inline` so that the compiler will inline it. 132 * 133 * Arguments are: 134 * 135 * NAME - The function name for this instance of the search function. Use it 136 * in a subsequent call to zfs_btree_create(). 137 * T - The element type stored inside the B-Tree. 138 * COMP - A comparator to compare two nodes, it must return exactly: -1, 0, 139 * or +1 -1 for <, 0 for ==, and +1 for >. For trivial comparisons, 140 * TREE_CMP() from avl.h can be used in a boilerplate function. 141 */ 142/* BEGIN CSTYLED */ 143#define ZFS_BTREE_FIND_IN_BUF_FUNC(NAME, T, COMP) \ 144_Pragma("GCC diagnostic push") \ 145_Pragma("GCC diagnostic ignored \"-Wunknown-pragmas\"") \ 146static void * \ 147NAME(zfs_btree_t *tree, uint8_t *buf, uint32_t nelems, \ 148 const void *value, zfs_btree_index_t *where) \ 149{ \ 150 T *i = (T *)buf; \ 151 (void) tree; \ 152 _Pragma("GCC unroll 9") \ 153 while (nelems > 1) { \ 154 uint32_t half = nelems / 2; \ 155 nelems -= half; \ 156 i += (COMP(&i[half - 1], value) < 0) * half; \ 157 } \ 158 \ 159 int comp = COMP(i, value); \ 160 where->bti_offset = (i - (T *)buf) + (comp < 0); \ 161 where->bti_before = (comp != 0); \ 162 \ 163 if (comp == 0) { \ 164 return (i); \ 165 } \ 166 \ 167 return (NULL); \ 168} \ 169_Pragma("GCC diagnostic pop") 170/* END CSTYLED */ 171 172/* 173 * Allocate and deallocate caches for btree nodes. 174 */ 175void zfs_btree_init(void); 176void zfs_btree_fini(void); 177 178/* 179 * Initialize an B-Tree. Arguments are: 180 * 181 * tree - the tree to be initialized 182 * compar - function to compare two nodes, it must return exactly: -1, 0, or +1 183 * -1 for <, 0 for ==, and +1 for > 184 * find - optional function to accelerate searches inside B-Tree nodes 185 * through Shar's algorithm and comparator inlining. Setting this to 186 * NULL will use a generic function. The function should be created 187 * using ZFS_BTREE_FIND_IN_BUF_FUNC() in the same file as compar. 188 * compar should be marked `__attribute__((always_inline)) inline` or 189 * performance is unlikely to improve very much. 190 * size - the value of sizeof(struct my_type) 191 * lsize - custom leaf size 192 */ 193void zfs_btree_create(zfs_btree_t *, int (*) (const void *, const void *), 194 bt_find_in_buf_f, size_t); 195void zfs_btree_create_custom(zfs_btree_t *, int (*)(const void *, const void *), 196 bt_find_in_buf_f, size_t, size_t); 197 198/* 199 * Find a node with a matching value in the tree. Returns the matching node 200 * found. If not found, it returns NULL and then if "where" is not NULL it sets 201 * "where" for use with zfs_btree_add_idx() or zfs_btree_nearest(). 202 * 203 * node - node that has the value being looked for 204 * where - position for use with zfs_btree_nearest() or zfs_btree_add_idx(), 205 * may be NULL 206 */ 207void *zfs_btree_find(zfs_btree_t *, const void *, zfs_btree_index_t *); 208 209/* 210 * Insert a node into the tree. 211 * 212 * node - the node to insert 213 * where - position as returned from zfs_btree_find() 214 */ 215void zfs_btree_add_idx(zfs_btree_t *, const void *, const zfs_btree_index_t *); 216 217/* 218 * Return the first or last valued node in the tree. Will return NULL if the 219 * tree is empty. The index can be NULL if the location of the first or last 220 * element isn't required. 221 */ 222void *zfs_btree_first(zfs_btree_t *, zfs_btree_index_t *); 223void *zfs_btree_last(zfs_btree_t *, zfs_btree_index_t *); 224 225/* 226 * Return the next or previous valued node in the tree. The second index can 227 * safely be NULL, if the location of the next or previous value isn't 228 * required. 229 */ 230void *zfs_btree_next(zfs_btree_t *, const zfs_btree_index_t *, 231 zfs_btree_index_t *); 232void *zfs_btree_prev(zfs_btree_t *, const zfs_btree_index_t *, 233 zfs_btree_index_t *); 234 235/* 236 * Get a value from a tree and an index. 237 */ 238void *zfs_btree_get(zfs_btree_t *, zfs_btree_index_t *); 239 240/* 241 * Add a single value to the tree. The value must not compare equal to any 242 * other node already in the tree. Note that the value will be copied out, not 243 * inserted directly. It is safe to free or destroy the value once this 244 * function returns. 245 */ 246void zfs_btree_add(zfs_btree_t *, const void *); 247 248/* 249 * Remove a single value from the tree. The value must be in the tree. The 250 * pointer passed in may be a pointer into a tree-controlled buffer, but it 251 * need not be. 252 */ 253void zfs_btree_remove(zfs_btree_t *, const void *); 254 255/* 256 * Remove the value at the given location from the tree. 257 */ 258void zfs_btree_remove_idx(zfs_btree_t *, zfs_btree_index_t *); 259 260/* 261 * Return the number of nodes in the tree 262 */ 263ulong_t zfs_btree_numnodes(zfs_btree_t *); 264 265/* 266 * Used to destroy any remaining nodes in a tree. The cookie argument should 267 * be initialized to NULL before the first call. Returns a node that has been 268 * removed from the tree and may be free()'d. Returns NULL when the tree is 269 * empty. 270 * 271 * Once you call zfs_btree_destroy_nodes(), you can only continuing calling it 272 * and finally zfs_btree_destroy(). No other B-Tree routines will be valid. 273 * 274 * cookie - an index used to save state between calls to 275 * zfs_btree_destroy_nodes() 276 * 277 * EXAMPLE: 278 * zfs_btree_t *tree; 279 * struct my_data *node; 280 * zfs_btree_index_t *cookie; 281 * 282 * cookie = NULL; 283 * while ((node = zfs_btree_destroy_nodes(tree, &cookie)) != NULL) 284 * data_destroy(node); 285 * zfs_btree_destroy(tree); 286 */ 287void *zfs_btree_destroy_nodes(zfs_btree_t *, zfs_btree_index_t **); 288 289/* 290 * Destroys all nodes in the tree quickly. This doesn't give the caller an 291 * opportunity to iterate over each node and do its own cleanup; for that, use 292 * zfs_btree_destroy_nodes(). 293 */ 294void zfs_btree_clear(zfs_btree_t *); 295 296/* 297 * Final destroy of an B-Tree. Arguments are: 298 * 299 * tree - the empty tree to destroy 300 */ 301void zfs_btree_destroy(zfs_btree_t *tree); 302 303/* Runs a variety of self-checks on the btree to verify integrity. */ 304void zfs_btree_verify(zfs_btree_t *tree); 305 306#ifdef __cplusplus 307} 308#endif 309 310#endif /* _BTREE_H */ 311