Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Elastic Binary Trees - macros and structures for Multi-Byte data nodes. |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 3 | * Version 6.0.5 |
| 4 | * (C) 2002-2011 - Willy Tarreau <w@1wt.eu> |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License as published by |
| 8 | * the Free Software Foundation; either version 2 of the License, or |
| 9 | * (at your option) any later version. |
| 10 | * |
| 11 | * This program is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | * GNU General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU General Public License |
| 17 | * along with this program; if not, write to the Free Software |
| 18 | * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
| 19 | */ |
| 20 | |
Willy Tarreau | ead63a0 | 2009-11-02 14:41:23 +0100 | [diff] [blame] | 21 | #ifndef _EBMBTREE_H |
| 22 | #define _EBMBTREE_H |
| 23 | |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 24 | #include <string.h> |
| 25 | #include "ebtree.h" |
| 26 | |
| 27 | /* Return the structure of type <type> whose member <member> points to <ptr> */ |
| 28 | #define ebmb_entry(ptr, type, member) container_of(ptr, type, member) |
| 29 | |
| 30 | #define EBMB_ROOT EB_ROOT |
| 31 | #define EBMB_TREE_HEAD EB_TREE_HEAD |
| 32 | |
| 33 | /* This structure carries a node, a leaf, and a key. It must start with the |
| 34 | * eb_node so that it can be cast into an eb_node. We could also have put some |
| 35 | * sort of transparent union here to reduce the indirection level, but the fact |
| 36 | * is, the end user is not meant to manipulate internals, so this is pointless. |
| 37 | * The 'node.bit' value here works differently from scalar types, as it contains |
| 38 | * the number of identical bits between the two branches. |
| 39 | */ |
| 40 | struct ebmb_node { |
| 41 | struct eb_node node; /* the tree node, must be at the beginning */ |
| 42 | unsigned char key[0]; /* the key, its size depends on the application */ |
| 43 | }; |
| 44 | |
| 45 | /* |
| 46 | * Exported functions and macros. |
| 47 | * Many of them are always inlined because they are extremely small, and |
| 48 | * are generally called at most once or twice in a program. |
| 49 | */ |
| 50 | |
| 51 | /* Return leftmost node in the tree, or NULL if none */ |
| 52 | static forceinline struct ebmb_node *ebmb_first(struct eb_root *root) |
| 53 | { |
| 54 | return ebmb_entry(eb_first(root), struct ebmb_node, node); |
| 55 | } |
| 56 | |
| 57 | /* Return rightmost node in the tree, or NULL if none */ |
| 58 | static forceinline struct ebmb_node *ebmb_last(struct eb_root *root) |
| 59 | { |
| 60 | return ebmb_entry(eb_last(root), struct ebmb_node, node); |
| 61 | } |
| 62 | |
| 63 | /* Return next node in the tree, or NULL if none */ |
| 64 | static forceinline struct ebmb_node *ebmb_next(struct ebmb_node *ebmb) |
| 65 | { |
| 66 | return ebmb_entry(eb_next(&ebmb->node), struct ebmb_node, node); |
| 67 | } |
| 68 | |
| 69 | /* Return previous node in the tree, or NULL if none */ |
| 70 | static forceinline struct ebmb_node *ebmb_prev(struct ebmb_node *ebmb) |
| 71 | { |
| 72 | return ebmb_entry(eb_prev(&ebmb->node), struct ebmb_node, node); |
| 73 | } |
| 74 | |
| 75 | /* Return next node in the tree, skipping duplicates, or NULL if none */ |
| 76 | static forceinline struct ebmb_node *ebmb_next_unique(struct ebmb_node *ebmb) |
| 77 | { |
| 78 | return ebmb_entry(eb_next_unique(&ebmb->node), struct ebmb_node, node); |
| 79 | } |
| 80 | |
| 81 | /* Return previous node in the tree, skipping duplicates, or NULL if none */ |
| 82 | static forceinline struct ebmb_node *ebmb_prev_unique(struct ebmb_node *ebmb) |
| 83 | { |
| 84 | return ebmb_entry(eb_prev_unique(&ebmb->node), struct ebmb_node, node); |
| 85 | } |
| 86 | |
| 87 | /* Delete node from the tree if it was linked in. Mark the node unused. Note |
| 88 | * that this function relies on a non-inlined generic function: eb_delete. |
| 89 | */ |
| 90 | static forceinline void ebmb_delete(struct ebmb_node *ebmb) |
| 91 | { |
| 92 | eb_delete(&ebmb->node); |
| 93 | } |
| 94 | |
| 95 | /* The following functions are not inlined by default. They are declared |
| 96 | * in ebmbtree.c, which simply relies on their inline version. |
| 97 | */ |
| 98 | REGPRM3 struct ebmb_node *ebmb_lookup(struct eb_root *root, const void *x, unsigned int len); |
| 99 | REGPRM3 struct ebmb_node *ebmb_insert(struct eb_root *root, struct ebmb_node *new, unsigned int len); |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 100 | REGPRM2 struct ebmb_node *ebmb_lookup_longest(struct eb_root *root, const void *x); |
| 101 | REGPRM3 struct ebmb_node *ebmb_lookup_prefix(struct eb_root *root, const void *x, unsigned int pfx); |
| 102 | REGPRM3 struct ebmb_node *ebmb_insert_prefix(struct eb_root *root, struct ebmb_node *new, unsigned int len); |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 103 | |
| 104 | /* The following functions are less likely to be used directly, because their |
| 105 | * code is larger. The non-inlined version is preferred. |
| 106 | */ |
| 107 | |
| 108 | /* Delete node from the tree if it was linked in. Mark the node unused. */ |
| 109 | static forceinline void __ebmb_delete(struct ebmb_node *ebmb) |
| 110 | { |
| 111 | __eb_delete(&ebmb->node); |
| 112 | } |
| 113 | |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 114 | /* Find the first occurence of a key of a least <len> bytes matching <x> in the |
| 115 | * tree <root>. The caller is responsible for ensuring that <len> will not exceed |
| 116 | * the common parts between the tree's keys and <x>. In case of multiple matches, |
| 117 | * the leftmost node is returned. This means that this function can be used to |
| 118 | * lookup string keys by prefix if all keys in the tree are zero-terminated. If |
| 119 | * no match is found, NULL is returned. Returns first node if <len> is zero. |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 120 | */ |
| 121 | static forceinline struct ebmb_node *__ebmb_lookup(struct eb_root *root, const void *x, unsigned int len) |
| 122 | { |
| 123 | struct ebmb_node *node; |
| 124 | eb_troot_t *troot; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 125 | int pos, side; |
| 126 | int node_bit; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 127 | |
| 128 | troot = root->b[EB_LEFT]; |
| 129 | if (unlikely(troot == NULL)) |
| 130 | return NULL; |
| 131 | |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 132 | if (unlikely(len == 0)) |
| 133 | goto walk_down; |
| 134 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 135 | pos = 0; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 136 | while (1) { |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 137 | if (eb_gettag(troot) == EB_LEAF) { |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 138 | node = container_of(eb_untag(troot, EB_LEAF), |
| 139 | struct ebmb_node, node.branches); |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 140 | if (memcmp(node->key + pos, x, len) != 0) |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 141 | return NULL; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 142 | else |
| 143 | return node; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 144 | } |
| 145 | node = container_of(eb_untag(troot, EB_NODE), |
| 146 | struct ebmb_node, node.branches); |
| 147 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 148 | node_bit = node->node.bit; |
| 149 | if (node_bit < 0) { |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 150 | /* We have a dup tree now. Either it's for the same |
| 151 | * value, and we walk down left, or it's a different |
| 152 | * one and we don't have our key. |
| 153 | */ |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 154 | if (memcmp(node->key + pos, x, len) != 0) |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 155 | return NULL; |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 156 | walk_left: |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 157 | troot = node->node.branches.b[EB_LEFT]; |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 158 | walk_down: |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 159 | while (eb_gettag(troot) != EB_LEAF) |
| 160 | troot = (eb_untag(troot, EB_NODE))->b[EB_LEFT]; |
| 161 | node = container_of(eb_untag(troot, EB_LEAF), |
| 162 | struct ebmb_node, node.branches); |
| 163 | return node; |
| 164 | } |
| 165 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 166 | /* OK, normal data node, let's walk down. We check if all full |
| 167 | * bytes are equal, and we start from the last one we did not |
| 168 | * completely check. We stop as soon as we reach the last byte, |
| 169 | * because we must decide to go left/right or abort. |
| 170 | */ |
| 171 | node_bit = ~node_bit + (pos << 3) + 8; // = (pos<<3) + (7 - node_bit) |
| 172 | if (node_bit < 0) { |
| 173 | /* This surprizing construction gives better performance |
| 174 | * because gcc does not try to reorder the loop. Tested to |
| 175 | * be fine with 2.95 to 4.2. |
| 176 | */ |
| 177 | while (1) { |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 178 | if (node->key[pos++] ^ *(unsigned char*)(x++)) |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 179 | return NULL; /* more than one full byte is different */ |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 180 | if (--len == 0) |
| 181 | goto walk_left; /* return first node if all bytes matched */ |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 182 | node_bit += 8; |
| 183 | if (node_bit >= 0) |
| 184 | break; |
| 185 | } |
| 186 | } |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 187 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 188 | /* here we know that only the last byte differs, so node_bit < 8. |
| 189 | * We have 2 possibilities : |
| 190 | * - more than the last bit differs => return NULL |
| 191 | * - walk down on side = (x[pos] >> node_bit) & 1 |
| 192 | */ |
| 193 | side = *(unsigned char *)x >> node_bit; |
| 194 | if (((node->key[pos] >> node_bit) ^ side) > 1) |
| 195 | return NULL; |
| 196 | side &= 1; |
| 197 | troot = node->node.branches.b[side]; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 198 | } |
| 199 | } |
| 200 | |
| 201 | /* Insert ebmb_node <new> into subtree starting at node root <root>. |
| 202 | * Only new->key needs be set with the key. The ebmb_node is returned. |
| 203 | * If root->b[EB_RGHT]==1, the tree may only contain unique keys. The |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 204 | * len is specified in bytes. It is absolutely mandatory that this length |
| 205 | * is the same for all keys in the tree. This function cannot be used to |
| 206 | * insert strings. |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 207 | */ |
| 208 | static forceinline struct ebmb_node * |
| 209 | __ebmb_insert(struct eb_root *root, struct ebmb_node *new, unsigned int len) |
| 210 | { |
| 211 | struct ebmb_node *old; |
| 212 | unsigned int side; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 213 | eb_troot_t *troot, **up_ptr; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 214 | eb_troot_t *root_right = root; |
| 215 | int diff; |
| 216 | int bit; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 217 | eb_troot_t *new_left, *new_rght; |
| 218 | eb_troot_t *new_leaf; |
| 219 | int old_node_bit; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 220 | |
| 221 | side = EB_LEFT; |
| 222 | troot = root->b[EB_LEFT]; |
| 223 | root_right = root->b[EB_RGHT]; |
| 224 | if (unlikely(troot == NULL)) { |
| 225 | /* Tree is empty, insert the leaf part below the left branch */ |
| 226 | root->b[EB_LEFT] = eb_dotag(&new->node.branches, EB_LEAF); |
| 227 | new->node.leaf_p = eb_dotag(root, EB_LEFT); |
| 228 | new->node.node_p = NULL; /* node part unused */ |
| 229 | return new; |
| 230 | } |
| 231 | |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 232 | /* The tree descent is fairly easy : |
| 233 | * - first, check if we have reached a leaf node |
| 234 | * - second, check if we have gone too far |
| 235 | * - third, reiterate |
| 236 | * Everywhere, we use <new> for the node node we are inserting, <root> |
| 237 | * for the node we attach it to, and <old> for the node we are |
| 238 | * displacing below <new>. <troot> will always point to the future node |
| 239 | * (tagged with its type). <side> carries the side the node <new> is |
| 240 | * attached to below its parent, which is also where previous node |
| 241 | * was attached. |
| 242 | */ |
| 243 | |
| 244 | bit = 0; |
| 245 | while (1) { |
| 246 | if (unlikely(eb_gettag(troot) == EB_LEAF)) { |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 247 | /* insert above a leaf */ |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 248 | old = container_of(eb_untag(troot, EB_LEAF), |
| 249 | struct ebmb_node, node.branches); |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 250 | new->node.node_p = old->node.leaf_p; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 251 | up_ptr = &old->node.leaf_p; |
| 252 | goto check_bit_and_break; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 253 | } |
| 254 | |
| 255 | /* OK we're walking down this link */ |
| 256 | old = container_of(eb_untag(troot, EB_NODE), |
| 257 | struct ebmb_node, node.branches); |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 258 | old_node_bit = old->node.bit; |
| 259 | |
| 260 | if (unlikely(old->node.bit < 0)) { |
| 261 | /* We're above a duplicate tree, so we must compare the whole value */ |
| 262 | new->node.node_p = old->node.node_p; |
| 263 | up_ptr = &old->node.node_p; |
| 264 | check_bit_and_break: |
| 265 | bit = equal_bits(new->key, old->key, bit, len << 3); |
| 266 | break; |
| 267 | } |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 268 | |
| 269 | /* Stop going down when we don't have common bits anymore. We |
| 270 | * also stop in front of a duplicates tree because it means we |
| 271 | * have to insert above. Note: we can compare more bits than |
| 272 | * the current node's because as long as they are identical, we |
| 273 | * know we descend along the correct side. |
| 274 | */ |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 275 | |
| 276 | bit = equal_bits(new->key, old->key, bit, old_node_bit); |
| 277 | if (unlikely(bit < old_node_bit)) { |
| 278 | /* The tree did not contain the key, so we insert <new> before the |
| 279 | * node <old>, and set ->bit to designate the lowest bit position in |
| 280 | * <new> which applies to ->branches.b[]. |
| 281 | */ |
| 282 | new->node.node_p = old->node.node_p; |
| 283 | up_ptr = &old->node.node_p; |
| 284 | break; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 285 | } |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 286 | /* we don't want to skip bits for further comparisons, so we must limit <bit>. |
| 287 | * However, since we're going down around <old_node_bit>, we know it will be |
| 288 | * properly matched, so we can skip this bit. |
| 289 | */ |
| 290 | bit = old_node_bit + 1; |
| 291 | |
| 292 | /* walk down */ |
| 293 | root = &old->node.branches; |
| 294 | side = old_node_bit & 7; |
| 295 | side ^= 7; |
| 296 | side = (new->key[old_node_bit >> 3] >> side) & 1; |
| 297 | troot = root->b[side]; |
| 298 | } |
| 299 | |
| 300 | new_left = eb_dotag(&new->node.branches, EB_LEFT); |
| 301 | new_rght = eb_dotag(&new->node.branches, EB_RGHT); |
| 302 | new_leaf = eb_dotag(&new->node.branches, EB_LEAF); |
| 303 | |
| 304 | /* Note: we can compare more bits than |
| 305 | * the current node's because as long as they are identical, we |
| 306 | * know we descend along the correct side. |
| 307 | */ |
| 308 | new->node.bit = bit; |
| 309 | diff = cmp_bits(new->key, old->key, bit); |
| 310 | if (diff == 0) { |
| 311 | new->node.bit = -1; /* mark as new dup tree, just in case */ |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 312 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 313 | if (likely(eb_gettag(root_right))) { |
| 314 | /* we refuse to duplicate this key if the tree is |
| 315 | * tagged as containing only unique keys. |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 316 | */ |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 317 | return old; |
| 318 | } |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 319 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 320 | if (eb_gettag(troot) != EB_LEAF) { |
| 321 | /* there was already a dup tree below */ |
| 322 | struct eb_node *ret; |
| 323 | ret = eb_insert_dup(&old->node, &new->node); |
| 324 | return container_of(ret, struct ebmb_node, node); |
| 325 | } |
| 326 | /* otherwise fall through */ |
| 327 | } |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 328 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 329 | if (diff >= 0) { |
| 330 | new->node.branches.b[EB_LEFT] = troot; |
| 331 | new->node.branches.b[EB_RGHT] = new_leaf; |
| 332 | new->node.leaf_p = new_rght; |
| 333 | *up_ptr = new_left; |
| 334 | } |
| 335 | else if (diff < 0) { |
| 336 | new->node.branches.b[EB_LEFT] = new_leaf; |
| 337 | new->node.branches.b[EB_RGHT] = troot; |
| 338 | new->node.leaf_p = new_left; |
| 339 | *up_ptr = new_rght; |
| 340 | } |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 341 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 342 | /* Ok, now we are inserting <new> between <root> and <old>. <old>'s |
| 343 | * parent is already set to <new>, and the <root>'s branch is still in |
| 344 | * <side>. Update the root's leaf till we have it. Note that we can also |
| 345 | * find the side by checking the side of new->node.node_p. |
| 346 | */ |
| 347 | |
| 348 | root->b[side] = eb_dotag(&new->node.branches, EB_NODE); |
| 349 | return new; |
| 350 | } |
| 351 | |
| 352 | |
| 353 | /* Find the first occurence of the longest prefix matching a key <x> in the |
| 354 | * tree <root>. It's the caller's responsibility to ensure that key <x> is at |
| 355 | * least as long as the keys in the tree. If none can be found, return NULL. |
| 356 | */ |
| 357 | static forceinline struct ebmb_node *__ebmb_lookup_longest(struct eb_root *root, const void *x) |
| 358 | { |
| 359 | struct ebmb_node *node; |
| 360 | eb_troot_t *troot, *cover; |
| 361 | int pos, side; |
| 362 | int node_bit; |
| 363 | |
| 364 | troot = root->b[EB_LEFT]; |
| 365 | if (unlikely(troot == NULL)) |
| 366 | return NULL; |
| 367 | |
| 368 | cover = NULL; |
| 369 | pos = 0; |
| 370 | while (1) { |
| 371 | if ((eb_gettag(troot) == EB_LEAF)) { |
| 372 | node = container_of(eb_untag(troot, EB_LEAF), |
| 373 | struct ebmb_node, node.branches); |
| 374 | if (check_bits(x - pos, node->key, pos, node->node.pfx)) |
| 375 | goto not_found; |
| 376 | |
| 377 | return node; |
| 378 | } |
| 379 | node = container_of(eb_untag(troot, EB_NODE), |
| 380 | struct ebmb_node, node.branches); |
| 381 | |
| 382 | node_bit = node->node.bit; |
| 383 | if (node_bit < 0) { |
| 384 | /* We have a dup tree now. Either it's for the same |
| 385 | * value, and we walk down left, or it's a different |
| 386 | * one and we don't have our key. |
| 387 | */ |
| 388 | if (check_bits(x - pos, node->key, pos, node->node.pfx)) |
| 389 | goto not_found; |
| 390 | |
| 391 | troot = node->node.branches.b[EB_LEFT]; |
| 392 | while (eb_gettag(troot) != EB_LEAF) |
| 393 | troot = (eb_untag(troot, EB_NODE))->b[EB_LEFT]; |
| 394 | node = container_of(eb_untag(troot, EB_LEAF), |
| 395 | struct ebmb_node, node.branches); |
| 396 | return node; |
| 397 | } |
| 398 | |
| 399 | node_bit >>= 1; /* strip cover bit */ |
| 400 | node_bit = ~node_bit + (pos << 3) + 8; // = (pos<<3) + (7 - node_bit) |
| 401 | if (node_bit < 0) { |
| 402 | /* This uncommon construction gives better performance |
| 403 | * because gcc does not try to reorder the loop. Tested to |
| 404 | * be fine with 2.95 to 4.2. |
| 405 | */ |
| 406 | while (1) { |
| 407 | x++; pos++; |
| 408 | if (node->key[pos-1] ^ *(unsigned char*)(x-1)) |
| 409 | goto not_found; /* more than one full byte is different */ |
| 410 | node_bit += 8; |
| 411 | if (node_bit >= 0) |
| 412 | break; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 413 | } |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 414 | } |
| 415 | |
| 416 | /* here we know that only the last byte differs, so 0 <= node_bit <= 7. |
| 417 | * We have 2 possibilities : |
| 418 | * - more than the last bit differs => data does not match |
| 419 | * - walk down on side = (x[pos] >> node_bit) & 1 |
| 420 | */ |
| 421 | side = *(unsigned char *)x >> node_bit; |
| 422 | if (((node->key[pos] >> node_bit) ^ side) > 1) |
| 423 | goto not_found; |
| 424 | |
| 425 | if (!(node->node.bit & 1)) { |
| 426 | /* This is a cover node, let's keep a reference to it |
| 427 | * for later. The covering subtree is on the left, and |
| 428 | * the covered subtree is on the right, so we have to |
| 429 | * walk down right. |
| 430 | */ |
| 431 | cover = node->node.branches.b[EB_LEFT]; |
| 432 | troot = node->node.branches.b[EB_RGHT]; |
| 433 | continue; |
| 434 | } |
| 435 | side &= 1; |
| 436 | troot = node->node.branches.b[side]; |
| 437 | } |
| 438 | |
| 439 | not_found: |
| 440 | /* Walk down last cover tre if it exists. It does not matter if cover is NULL */ |
| 441 | return ebmb_entry(eb_walk_down(cover, EB_LEFT), struct ebmb_node, node); |
| 442 | } |
| 443 | |
| 444 | |
| 445 | /* Find the first occurence of a prefix matching a key <x> of <pfx> BITS in the |
Willy Tarreau | 414c4b2 | 2011-01-04 13:21:06 +0100 | [diff] [blame^] | 446 | * tree <root>. It's the caller's responsibility to ensure that key <x> is at |
| 447 | * least as long as the keys in the tree. If none can be found, return NULL. |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 448 | */ |
| 449 | static forceinline struct ebmb_node *__ebmb_lookup_prefix(struct eb_root *root, const void *x, unsigned int pfx) |
| 450 | { |
| 451 | struct ebmb_node *node; |
| 452 | eb_troot_t *troot; |
| 453 | int pos, side; |
| 454 | int node_bit; |
| 455 | |
| 456 | troot = root->b[EB_LEFT]; |
| 457 | if (unlikely(troot == NULL)) |
| 458 | return NULL; |
| 459 | |
| 460 | pos = 0; |
| 461 | while (1) { |
| 462 | if ((eb_gettag(troot) == EB_LEAF)) { |
| 463 | node = container_of(eb_untag(troot, EB_LEAF), |
| 464 | struct ebmb_node, node.branches); |
| 465 | if (node->node.pfx != pfx) |
| 466 | return NULL; |
| 467 | if (check_bits(x - pos, node->key, pos, node->node.pfx)) |
| 468 | return NULL; |
| 469 | return node; |
| 470 | } |
| 471 | node = container_of(eb_untag(troot, EB_NODE), |
| 472 | struct ebmb_node, node.branches); |
| 473 | |
| 474 | node_bit = node->node.bit; |
| 475 | if (node_bit < 0) { |
| 476 | /* We have a dup tree now. Either it's for the same |
| 477 | * value, and we walk down left, or it's a different |
| 478 | * one and we don't have our key. |
| 479 | */ |
| 480 | if (node->node.pfx != pfx) |
| 481 | return NULL; |
| 482 | if (check_bits(x - pos, node->key, pos, node->node.pfx)) |
| 483 | return NULL; |
| 484 | |
| 485 | troot = node->node.branches.b[EB_LEFT]; |
| 486 | while (eb_gettag(troot) != EB_LEAF) |
| 487 | troot = (eb_untag(troot, EB_NODE))->b[EB_LEFT]; |
| 488 | node = container_of(eb_untag(troot, EB_LEAF), |
| 489 | struct ebmb_node, node.branches); |
| 490 | return node; |
| 491 | } |
| 492 | |
| 493 | node_bit >>= 1; /* strip cover bit */ |
| 494 | node_bit = ~node_bit + (pos << 3) + 8; // = (pos<<3) + (7 - node_bit) |
| 495 | if (node_bit < 0) { |
| 496 | /* This uncommon construction gives better performance |
| 497 | * because gcc does not try to reorder the loop. Tested to |
| 498 | * be fine with 2.95 to 4.2. |
| 499 | */ |
| 500 | while (1) { |
| 501 | x++; pos++; |
| 502 | if (node->key[pos-1] ^ *(unsigned char*)(x-1)) |
| 503 | return NULL; /* more than one full byte is different */ |
| 504 | node_bit += 8; |
| 505 | if (node_bit >= 0) |
| 506 | break; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 507 | } |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 508 | } |
| 509 | |
| 510 | /* here we know that only the last byte differs, so 0 <= node_bit <= 7. |
| 511 | * We have 2 possibilities : |
| 512 | * - more than the last bit differs => data does not match |
| 513 | * - walk down on side = (x[pos] >> node_bit) & 1 |
| 514 | */ |
| 515 | side = *(unsigned char *)x >> node_bit; |
| 516 | if (((node->key[pos] >> node_bit) ^ side) > 1) |
| 517 | return NULL; |
| 518 | |
| 519 | if (!(node->node.bit & 1)) { |
| 520 | /* This is a cover node, it may be the entry we're |
| 521 | * looking for. We already know that it matches all the |
| 522 | * bits, let's compare prefixes and descend the cover |
| 523 | * subtree if they match. |
| 524 | */ |
| 525 | if (node->node.bit >> 1 == pfx) |
| 526 | troot = node->node.branches.b[EB_LEFT]; |
| 527 | else |
| 528 | troot = node->node.branches.b[EB_RGHT]; |
| 529 | continue; |
| 530 | } |
| 531 | side &= 1; |
| 532 | troot = node->node.branches.b[side]; |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | |
| 537 | /* Insert ebmb_node <new> into a prefix subtree starting at node root <root>. |
| 538 | * Only new->key and new->pfx need be set with the key and its prefix length. |
| 539 | * Note that bits between <pfx> and <len> are theorically ignored and should be |
| 540 | * zero, as it is not certain yet that they will always be ignored everywhere |
| 541 | * (eg in bit compare functions). |
| 542 | * The ebmb_node is returned. |
| 543 | * If root->b[EB_RGHT]==1, the tree may only contain unique keys. The |
| 544 | * len is specified in bytes. |
| 545 | */ |
| 546 | static forceinline struct ebmb_node * |
| 547 | __ebmb_insert_prefix(struct eb_root *root, struct ebmb_node *new, unsigned int len) |
| 548 | { |
| 549 | struct ebmb_node *old; |
| 550 | unsigned int side; |
| 551 | eb_troot_t *troot, **up_ptr; |
| 552 | eb_troot_t *root_right = root; |
| 553 | int diff; |
| 554 | int bit; |
| 555 | eb_troot_t *new_left, *new_rght; |
| 556 | eb_troot_t *new_leaf; |
| 557 | int old_node_bit; |
| 558 | |
| 559 | side = EB_LEFT; |
| 560 | troot = root->b[EB_LEFT]; |
| 561 | root_right = root->b[EB_RGHT]; |
| 562 | if (unlikely(troot == NULL)) { |
| 563 | /* Tree is empty, insert the leaf part below the left branch */ |
| 564 | root->b[EB_LEFT] = eb_dotag(&new->node.branches, EB_LEAF); |
| 565 | new->node.leaf_p = eb_dotag(root, EB_LEFT); |
| 566 | new->node.node_p = NULL; /* node part unused */ |
| 567 | return new; |
| 568 | } |
| 569 | |
| 570 | len <<= 3; |
| 571 | if (len > new->node.pfx) |
| 572 | len = new->node.pfx; |
| 573 | |
| 574 | /* The tree descent is fairly easy : |
| 575 | * - first, check if we have reached a leaf node |
| 576 | * - second, check if we have gone too far |
| 577 | * - third, reiterate |
| 578 | * Everywhere, we use <new> for the node node we are inserting, <root> |
| 579 | * for the node we attach it to, and <old> for the node we are |
| 580 | * displacing below <new>. <troot> will always point to the future node |
| 581 | * (tagged with its type). <side> carries the side the node <new> is |
| 582 | * attached to below its parent, which is also where previous node |
| 583 | * was attached. |
| 584 | */ |
| 585 | |
| 586 | bit = 0; |
| 587 | while (1) { |
| 588 | if (unlikely(eb_gettag(troot) == EB_LEAF)) { |
| 589 | /* Insert above a leaf. Note that this leaf could very |
| 590 | * well be part of a cover node. |
| 591 | */ |
| 592 | old = container_of(eb_untag(troot, EB_LEAF), |
| 593 | struct ebmb_node, node.branches); |
| 594 | new->node.node_p = old->node.leaf_p; |
| 595 | up_ptr = &old->node.leaf_p; |
| 596 | goto check_bit_and_break; |
| 597 | } |
| 598 | |
| 599 | /* OK we're walking down this link */ |
| 600 | old = container_of(eb_untag(troot, EB_NODE), |
| 601 | struct ebmb_node, node.branches); |
| 602 | old_node_bit = old->node.bit; |
| 603 | /* Note that old_node_bit can be : |
| 604 | * < 0 : dup tree |
| 605 | * = 2N : cover node for N bits |
| 606 | * = 2N+1 : normal node at N bits |
| 607 | */ |
| 608 | |
| 609 | if (unlikely(old_node_bit < 0)) { |
| 610 | /* We're above a duplicate tree, so we must compare the whole value */ |
| 611 | new->node.node_p = old->node.node_p; |
| 612 | up_ptr = &old->node.node_p; |
| 613 | check_bit_and_break: |
| 614 | /* No need to compare everything if the leaves are shorter than the new one. */ |
| 615 | if (len > old->node.pfx) |
| 616 | len = old->node.pfx; |
| 617 | bit = equal_bits(new->key, old->key, bit, len); |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 618 | break; |
| 619 | } |
| 620 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 621 | /* WARNING: for the two blocks below, <bit> is counted in half-bits */ |
| 622 | |
| 623 | bit = equal_bits(new->key, old->key, bit, old_node_bit >> 1); |
| 624 | bit = (bit << 1) + 1; // assume comparisons with normal nodes |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 625 | |
| 626 | /* we must always check that our prefix is larger than the nodes |
| 627 | * we visit, otherwise we have to stop going down. The following |
| 628 | * test is able to stop before both normal and cover nodes. |
| 629 | */ |
| 630 | if (bit >= (new->node.pfx << 1) && (new->node.pfx << 1) < old_node_bit) { |
| 631 | /* insert cover node here on the left */ |
| 632 | new->node.node_p = old->node.node_p; |
| 633 | up_ptr = &old->node.node_p; |
| 634 | new->node.bit = new->node.pfx << 1; |
| 635 | diff = -1; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 636 | goto insert_above; |
| 637 | } |
| 638 | |
| 639 | if (unlikely(bit < old_node_bit)) { |
| 640 | /* The tree did not contain the key, so we insert <new> before the |
| 641 | * node <old>, and set ->bit to designate the lowest bit position in |
| 642 | * <new> which applies to ->branches.b[]. We know that the bit is not |
| 643 | * greater than the prefix length thanks to the test above. |
| 644 | */ |
| 645 | new->node.node_p = old->node.node_p; |
| 646 | up_ptr = &old->node.node_p; |
| 647 | new->node.bit = bit; |
| 648 | diff = cmp_bits(new->key, old->key, bit >> 1); |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 649 | goto insert_above; |
| 650 | } |
| 651 | |
| 652 | if (!(old_node_bit & 1)) { |
| 653 | /* if we encounter a cover node with our exact prefix length, it's |
| 654 | * necessarily the same value, so we insert there as a duplicate on |
| 655 | * the left. For that, we go down on the left and the leaf detection |
| 656 | * code will finish the job. |
| 657 | */ |
| 658 | if ((new->node.pfx << 1) == old_node_bit) { |
| 659 | root = &old->node.branches; |
| 660 | side = EB_LEFT; |
| 661 | troot = root->b[side]; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 662 | continue; |
| 663 | } |
| 664 | |
| 665 | /* cover nodes are always walked through on the right */ |
| 666 | side = EB_RGHT; |
| 667 | bit = old_node_bit >> 1; /* recheck that bit */ |
| 668 | root = &old->node.branches; |
| 669 | troot = root->b[side]; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 670 | continue; |
| 671 | } |
| 672 | |
| 673 | /* we don't want to skip bits for further comparisons, so we must limit <bit>. |
| 674 | * However, since we're going down around <old_node_bit>, we know it will be |
| 675 | * properly matched, so we can skip this bit. |
| 676 | */ |
| 677 | old_node_bit >>= 1; |
| 678 | bit = old_node_bit + 1; |
| 679 | |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 680 | /* walk down */ |
| 681 | root = &old->node.branches; |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 682 | side = old_node_bit & 7; |
| 683 | side ^= 7; |
| 684 | side = (new->key[old_node_bit >> 3] >> side) & 1; |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 685 | troot = root->b[side]; |
| 686 | } |
| 687 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 688 | /* Right here, we have 4 possibilities : |
| 689 | * - the tree does not contain any leaf matching the |
| 690 | * key, and we have new->key < old->key. We insert |
| 691 | * new above old, on the left ; |
| 692 | * |
| 693 | * - the tree does not contain any leaf matching the |
| 694 | * key, and we have new->key > old->key. We insert |
| 695 | * new above old, on the right ; |
| 696 | * |
| 697 | * - the tree does contain the key with the same prefix |
| 698 | * length. We add the new key next to it as a first |
| 699 | * duplicate (since it was alone). |
| 700 | * |
| 701 | * The last two cases can easily be partially merged. |
| 702 | * |
| 703 | * - the tree contains a leaf matching the key, we have |
| 704 | * to insert above it as a cover node. The leaf with |
| 705 | * the shortest prefix becomes the left subtree and |
| 706 | * the leaf with the longest prefix becomes the right |
| 707 | * one. The cover node gets the min of both prefixes |
| 708 | * as its new bit. |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 709 | */ |
| 710 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 711 | /* first we want to ensure that we compare the correct bit, which means |
| 712 | * the largest common to both nodes. |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 713 | */ |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 714 | if (bit > new->node.pfx) |
| 715 | bit = new->node.pfx; |
| 716 | if (bit > old->node.pfx) |
| 717 | bit = old->node.pfx; |
| 718 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 719 | new->node.bit = (bit << 1) + 1; /* assume normal node by default */ |
| 720 | |
| 721 | /* if one prefix is included in the second one, we don't compare bits |
| 722 | * because they won't necessarily match, we just proceed with a cover |
| 723 | * node insertion. |
| 724 | */ |
| 725 | diff = 0; |
| 726 | if (bit < old->node.pfx && bit < new->node.pfx) |
| 727 | diff = cmp_bits(new->key, old->key, bit); |
| 728 | |
| 729 | if (diff == 0) { |
| 730 | /* Both keys match. Either it's a duplicate entry or we have to |
| 731 | * put the shortest prefix left and the largest one right below |
| 732 | * a new cover node. By default, diff==0 means we'll be inserted |
| 733 | * on the right. |
| 734 | */ |
| 735 | new->node.bit--; /* anticipate cover node insertion */ |
| 736 | if (new->node.pfx == old->node.pfx) { |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 737 | new->node.bit = -1; /* mark as new dup tree, just in case */ |
| 738 | |
| 739 | if (unlikely(eb_gettag(root_right))) { |
| 740 | /* we refuse to duplicate this key if the tree is |
| 741 | * tagged as containing only unique keys. |
| 742 | */ |
| 743 | return old; |
| 744 | } |
| 745 | |
| 746 | if (eb_gettag(troot) != EB_LEAF) { |
| 747 | /* there was already a dup tree below */ |
| 748 | struct eb_node *ret; |
| 749 | ret = eb_insert_dup(&old->node, &new->node); |
| 750 | return container_of(ret, struct ebmb_node, node); |
| 751 | } |
| 752 | /* otherwise fall through to insert first duplicate */ |
| 753 | } |
| 754 | /* otherwise we just rely on the tests below to select the right side */ |
| 755 | else if (new->node.pfx < old->node.pfx) |
| 756 | diff = -1; /* force insertion to left side */ |
| 757 | } |
| 758 | |
| 759 | insert_above: |
| 760 | new_left = eb_dotag(&new->node.branches, EB_LEFT); |
| 761 | new_rght = eb_dotag(&new->node.branches, EB_RGHT); |
| 762 | new_leaf = eb_dotag(&new->node.branches, EB_LEAF); |
| 763 | |
| 764 | if (diff >= 0) { |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 765 | new->node.branches.b[EB_LEFT] = troot; |
| 766 | new->node.branches.b[EB_RGHT] = new_leaf; |
| 767 | new->node.leaf_p = new_rght; |
| 768 | *up_ptr = new_left; |
| 769 | } |
| 770 | else { |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 771 | new->node.branches.b[EB_LEFT] = new_leaf; |
| 772 | new->node.branches.b[EB_RGHT] = troot; |
| 773 | new->node.leaf_p = new_left; |
| 774 | *up_ptr = new_rght; |
| 775 | } |
| 776 | |
Willy Tarreau | c218602 | 2009-10-26 19:48:54 +0100 | [diff] [blame] | 777 | root->b[side] = eb_dotag(&new->node.branches, EB_NODE); |
| 778 | return new; |
| 779 | } |
| 780 | |
Willy Tarreau | 3a93244 | 2010-05-09 19:29:23 +0200 | [diff] [blame] | 781 | |
| 782 | |
Willy Tarreau | ead63a0 | 2009-11-02 14:41:23 +0100 | [diff] [blame] | 783 | #endif /* _EBMBTREE_H */ |
| 784 | |