bt_split.c revision 1.16
1 /* $NetBSD: bt_split.c,v 1.16 2008/09/10 17:52:35 joerg Exp $ */ 2 3 /*- 4 * Copyright (c) 1990, 1993, 1994 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * Mike Olson. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __RCSID("$NetBSD: bt_split.c,v 1.16 2008/09/10 17:52:35 joerg Exp $"); 37 38 #include "namespace.h" 39 #include <sys/types.h> 40 41 #include <assert.h> 42 #include <limits.h> 43 #include <stdio.h> 44 #include <stdlib.h> 45 #include <string.h> 46 47 #include <db.h> 48 #include "btree.h" 49 50 static int bt_broot(BTREE *, PAGE *, PAGE *, PAGE *); 51 static PAGE *bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t); 52 static int bt_preserve(BTREE *, pgno_t); 53 static PAGE *bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t); 54 static PAGE *bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t); 55 static int bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *); 56 static recno_t rec_total(PAGE *); 57 58 #ifdef STATISTICS 59 unsigned long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved; 60 #endif 61 62 /* 63 * __BT_SPLIT -- Split the tree. 64 * 65 * Parameters: 66 * t: tree 67 * sp: page to split 68 * key: key to insert 69 * data: data to insert 70 * flags: BIGKEY/BIGDATA flags 71 * ilen: insert length 72 * skip: index to leave open 73 * 74 * Returns: 75 * RET_ERROR, RET_SUCCESS 76 */ 77 int 78 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags, 79 size_t ilen, uint32_t argskip) 80 { 81 BINTERNAL *bi = NULL; /* pacify gcc */ 82 BLEAF *bl = NULL, *tbl; /* pacify gcc */ 83 DBT a, b; 84 EPGNO *parent; 85 PAGE *h, *l, *r, *lchild, *rchild; 86 indx_t nxtindex; 87 uint16_t skip; 88 uint32_t n, nbytes, nksize = 0; /* pacify gcc */ 89 int parentsplit; 90 char *dest; 91 92 /* 93 * Split the page into two pages, l and r. The split routines return 94 * a pointer to the page into which the key should be inserted and with 95 * skip set to the offset which should be used. Additionally, l and r 96 * are pinned. 97 */ 98 skip = argskip; 99 h = sp->pgno == P_ROOT ? 100 bt_root(t, sp, &l, &r, &skip, ilen) : 101 bt_page(t, sp, &l, &r, &skip, ilen); 102 if (h == NULL) 103 return (RET_ERROR); 104 105 /* 106 * Insert the new key/data pair into the leaf page. (Key inserts 107 * always cause a leaf page to split first.) 108 */ 109 _DBFIT(ilen, indx_t); 110 h->upper -= (indx_t)ilen; 111 h->linp[skip] = h->upper; 112 dest = (char *)(void *)h + h->upper; 113 if (F_ISSET(t, R_RECNO)) 114 WR_RLEAF(dest, data, flags); 115 else 116 WR_BLEAF(dest, key, data, flags); 117 118 /* If the root page was split, make it look right. */ 119 if (sp->pgno == P_ROOT && 120 (F_ISSET(t, R_RECNO) ? 121 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 122 goto err2; 123 124 /* 125 * Now we walk the parent page stack -- a LIFO stack of the pages that 126 * were traversed when we searched for the page that split. Each stack 127 * entry is a page number and a page index offset. The offset is for 128 * the page traversed on the search. We've just split a page, so we 129 * have to insert a new key into the parent page. 130 * 131 * If the insert into the parent page causes it to split, may have to 132 * continue splitting all the way up the tree. We stop if the root 133 * splits or the page inserted into didn't have to split to hold the 134 * new key. Some algorithms replace the key for the old page as well 135 * as the new page. We don't, as there's no reason to believe that the 136 * first key on the old page is any better than the key we have, and, 137 * in the case of a key being placed at index 0 causing the split, the 138 * key is unavailable. 139 * 140 * There are a maximum of 5 pages pinned at any time. We keep the left 141 * and right pages pinned while working on the parent. The 5 are the 142 * two children, left parent and right parent (when the parent splits) 143 * and the root page or the overflow key page when calling bt_preserve. 144 * This code must make sure that all pins are released other than the 145 * root page or overflow page which is unlocked elsewhere. 146 */ 147 while ((parent = BT_POP(t)) != NULL) { 148 lchild = l; 149 rchild = r; 150 151 /* Get the parent page. */ 152 if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL) 153 goto err2; 154 155 /* 156 * The new key goes ONE AFTER the index, because the split 157 * was to the right. 158 */ 159 skip = parent->index + 1; 160 161 /* 162 * Calculate the space needed on the parent page. 163 * 164 * Prefix trees: space hack when inserting into BINTERNAL 165 * pages. Retain only what's needed to distinguish between 166 * the new entry and the LAST entry on the page to its left. 167 * If the keys compare equal, retain the entire key. Note, 168 * we don't touch overflow keys, and the entire key must be 169 * retained for the next-to-left most key on the leftmost 170 * page of each level, or the search will fail. Applicable 171 * ONLY to internal pages that have leaf pages as children. 172 * Further reduction of the key between pairs of internal 173 * pages loses too much information. 174 */ 175 switch (rchild->flags & P_TYPE) { 176 case P_BINTERNAL: 177 bi = GETBINTERNAL(rchild, 0); 178 nbytes = NBINTERNAL(bi->ksize); 179 break; 180 case P_BLEAF: 181 bl = GETBLEAF(rchild, 0); 182 nbytes = NBINTERNAL(bl->ksize); 183 if (t->bt_pfx && !(bl->flags & P_BIGKEY) && 184 (h->prevpg != P_INVALID || skip > 1)) { 185 size_t temp; 186 tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1); 187 a.size = tbl->ksize; 188 a.data = tbl->bytes; 189 b.size = bl->ksize; 190 b.data = bl->bytes; 191 temp = t->bt_pfx(&a, &b); 192 _DBFIT(temp, uint32_t); 193 nksize = (uint32_t)temp; 194 n = NBINTERNAL(nksize); 195 if (n < nbytes) { 196 #ifdef STATISTICS 197 bt_pfxsaved += nbytes - n; 198 #endif 199 nbytes = n; 200 } else 201 nksize = 0; 202 } else 203 nksize = 0; 204 break; 205 case P_RINTERNAL: 206 case P_RLEAF: 207 nbytes = NRINTERNAL; 208 break; 209 default: 210 abort(); 211 } 212 213 /* Split the parent page if necessary or shift the indices. */ 214 if (h->upper - h->lower < nbytes + sizeof(indx_t)) { 215 sp = h; 216 h = h->pgno == P_ROOT ? 217 bt_root(t, h, &l, &r, &skip, nbytes) : 218 bt_page(t, h, &l, &r, &skip, nbytes); 219 if (h == NULL) 220 goto err1; 221 parentsplit = 1; 222 } else { 223 if (skip < (nxtindex = NEXTINDEX(h))) 224 memmove(h->linp + skip + 1, h->linp + skip, 225 (nxtindex - skip) * sizeof(indx_t)); 226 h->lower += sizeof(indx_t); 227 parentsplit = 0; 228 } 229 230 /* Insert the key into the parent page. */ 231 switch (rchild->flags & P_TYPE) { 232 case P_BINTERNAL: 233 h->linp[skip] = h->upper -= nbytes; 234 dest = (char *)(void *)h + h->linp[skip]; 235 memmove(dest, bi, nbytes); 236 ((BINTERNAL *)(void *)dest)->pgno = rchild->pgno; 237 break; 238 case P_BLEAF: 239 h->linp[skip] = h->upper -= nbytes; 240 dest = (char *)(void *)h + h->linp[skip]; 241 WR_BINTERNAL(dest, nksize ? nksize : bl->ksize, 242 rchild->pgno, bl->flags & P_BIGKEY); 243 memmove(dest, bl->bytes, nksize ? nksize : bl->ksize); 244 if (bl->flags & P_BIGKEY && 245 bt_preserve(t, *(pgno_t *)(void *)bl->bytes) == 246 RET_ERROR) 247 goto err1; 248 break; 249 case P_RINTERNAL: 250 /* 251 * Update the left page count. If split 252 * added at index 0, fix the correct page. 253 */ 254 if (skip > 0) 255 dest = (char *)(void *)h + h->linp[skip - 1]; 256 else 257 dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1]; 258 ((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild); 259 ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno; 260 261 /* Update the right page count. */ 262 h->linp[skip] = h->upper -= nbytes; 263 dest = (char *)(void *)h + h->linp[skip]; 264 ((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild); 265 ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno; 266 break; 267 case P_RLEAF: 268 /* 269 * Update the left page count. If split 270 * added at index 0, fix the correct page. 271 */ 272 if (skip > 0) 273 dest = (char *)(void *)h + h->linp[skip - 1]; 274 else 275 dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1]; 276 ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild); 277 ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno; 278 279 /* Update the right page count. */ 280 h->linp[skip] = h->upper -= nbytes; 281 dest = (char *)(void *)h + h->linp[skip]; 282 ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild); 283 ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno; 284 break; 285 default: 286 abort(); 287 } 288 289 /* Unpin the held pages. */ 290 if (!parentsplit) { 291 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 292 break; 293 } 294 295 /* If the root page was split, make it look right. */ 296 if (sp->pgno == P_ROOT && 297 (F_ISSET(t, R_RECNO) ? 298 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 299 goto err1; 300 301 mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 302 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 303 } 304 305 /* Unpin the held pages. */ 306 mpool_put(t->bt_mp, l, MPOOL_DIRTY); 307 mpool_put(t->bt_mp, r, MPOOL_DIRTY); 308 309 /* Clear any pages left on the stack. */ 310 return (RET_SUCCESS); 311 312 /* 313 * If something fails in the above loop we were already walking back 314 * up the tree and the tree is now inconsistent. Nothing much we can 315 * do about it but release any memory we're holding. 316 */ 317 err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 318 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 319 320 err2: mpool_put(t->bt_mp, l, 0); 321 mpool_put(t->bt_mp, r, 0); 322 __dbpanic(t->bt_dbp); 323 return (RET_ERROR); 324 } 325 326 /* 327 * BT_PAGE -- Split a non-root page of a btree. 328 * 329 * Parameters: 330 * t: tree 331 * h: root page 332 * lp: pointer to left page pointer 333 * rp: pointer to right page pointer 334 * skip: pointer to index to leave open 335 * ilen: insert length 336 * 337 * Returns: 338 * Pointer to page in which to insert or NULL on error. 339 */ 340 static PAGE * 341 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen) 342 { 343 PAGE *l, *r, *tp; 344 pgno_t npg; 345 346 #ifdef STATISTICS 347 ++bt_split; 348 #endif 349 /* Put the new right page for the split into place. */ 350 if ((r = __bt_new(t, &npg)) == NULL) 351 return (NULL); 352 r->pgno = npg; 353 r->lower = BTDATAOFF; 354 r->upper = t->bt_psize; 355 r->nextpg = h->nextpg; 356 r->prevpg = h->pgno; 357 r->flags = h->flags & P_TYPE; 358 359 /* 360 * If we're splitting the last page on a level because we're appending 361 * a key to it (skip is NEXTINDEX()), it's likely that the data is 362 * sorted. Adding an empty page on the side of the level is less work 363 * and can push the fill factor much higher than normal. If we're 364 * wrong it's no big deal, we'll just do the split the right way next 365 * time. It may look like it's equally easy to do a similar hack for 366 * reverse sorted data, that is, split the tree left, but it's not. 367 * Don't even try. 368 */ 369 if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) { 370 #ifdef STATISTICS 371 ++bt_sortsplit; 372 #endif 373 h->nextpg = r->pgno; 374 r->lower = BTDATAOFF + sizeof(indx_t); 375 *skip = 0; 376 *lp = h; 377 *rp = r; 378 return (r); 379 } 380 381 /* Put the new left page for the split into place. */ 382 if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) { 383 mpool_put(t->bt_mp, r, 0); 384 return (NULL); 385 } 386 #ifdef PURIFY 387 memset(l, 0xff, t->bt_psize); 388 #endif 389 l->pgno = h->pgno; 390 l->nextpg = r->pgno; 391 l->prevpg = h->prevpg; 392 l->lower = BTDATAOFF; 393 l->upper = t->bt_psize; 394 l->flags = h->flags & P_TYPE; 395 396 /* Fix up the previous pointer of the page after the split page. */ 397 if (h->nextpg != P_INVALID) { 398 if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) { 399 free(l); 400 /* XXX mpool_free(t->bt_mp, r->pgno); */ 401 return (NULL); 402 } 403 tp->prevpg = r->pgno; 404 mpool_put(t->bt_mp, tp, MPOOL_DIRTY); 405 } 406 407 /* 408 * Split right. The key/data pairs aren't sorted in the btree page so 409 * it's simpler to copy the data from the split page onto two new pages 410 * instead of copying half the data to the right page and compacting 411 * the left page in place. Since the left page can't change, we have 412 * to swap the original and the allocated left page after the split. 413 */ 414 tp = bt_psplit(t, h, l, r, skip, ilen); 415 416 /* Move the new left page onto the old left page. */ 417 memmove(h, l, t->bt_psize); 418 if (tp == l) 419 tp = h; 420 free(l); 421 422 *lp = h; 423 *rp = r; 424 return (tp); 425 } 426 427 /* 428 * BT_ROOT -- Split the root page of a btree. 429 * 430 * Parameters: 431 * t: tree 432 * h: root page 433 * lp: pointer to left page pointer 434 * rp: pointer to right page pointer 435 * skip: pointer to index to leave open 436 * ilen: insert length 437 * 438 * Returns: 439 * Pointer to page in which to insert or NULL on error. 440 */ 441 static PAGE * 442 bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen) 443 { 444 PAGE *l, *r, *tp; 445 pgno_t lnpg, rnpg; 446 447 #ifdef STATISTICS 448 ++bt_split; 449 ++bt_rootsplit; 450 #endif 451 /* Put the new left and right pages for the split into place. */ 452 if ((l = __bt_new(t, &lnpg)) == NULL || 453 (r = __bt_new(t, &rnpg)) == NULL) 454 return (NULL); 455 l->pgno = lnpg; 456 r->pgno = rnpg; 457 l->nextpg = r->pgno; 458 r->prevpg = l->pgno; 459 l->prevpg = r->nextpg = P_INVALID; 460 l->lower = r->lower = BTDATAOFF; 461 l->upper = r->upper = t->bt_psize; 462 l->flags = r->flags = h->flags & P_TYPE; 463 464 /* Split the root page. */ 465 tp = bt_psplit(t, h, l, r, skip, ilen); 466 467 *lp = l; 468 *rp = r; 469 return (tp); 470 } 471 472 /* 473 * BT_RROOT -- Fix up the recno root page after it has been split. 474 * 475 * Parameters: 476 * t: tree 477 * h: root page 478 * l: left page 479 * r: right page 480 * 481 * Returns: 482 * RET_ERROR, RET_SUCCESS 483 */ 484 static int 485 bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r) 486 { 487 char *dest; 488 uint32_t sz; 489 size_t temp; 490 491 temp = t->bt_psize - NRINTERNAL; 492 _DBFIT(temp, uint32_t); 493 sz = (uint32_t)temp; 494 495 /* Insert the left and right keys, set the header information. */ 496 _DBFIT(sz, indx_t); 497 h->linp[0] = h->upper = (indx_t)sz; 498 dest = (char *)(void *)h + h->upper; 499 WR_RINTERNAL(dest, 500 l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno); 501 502 h->linp[1] = h->upper -= NRINTERNAL; 503 dest = (char *)(void *)h + h->upper; 504 WR_RINTERNAL(dest, 505 r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno); 506 507 h->lower = BTDATAOFF + 2 * sizeof(indx_t); 508 509 /* Unpin the root page, set to recno internal page. */ 510 h->flags &= ~P_TYPE; 511 h->flags |= P_RINTERNAL; 512 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 513 514 return (RET_SUCCESS); 515 } 516 517 /* 518 * BT_BROOT -- Fix up the btree root page after it has been split. 519 * 520 * Parameters: 521 * t: tree 522 * h: root page 523 * l: left page 524 * r: right page 525 * 526 * Returns: 527 * RET_ERROR, RET_SUCCESS 528 */ 529 static int 530 bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r) 531 { 532 BINTERNAL *bi = NULL; /* pacify gcc */ 533 BLEAF *bl; 534 uint32_t nbytes; 535 char *dest; 536 537 /* 538 * If the root page was a leaf page, change it into an internal page. 539 * We copy the key we split on (but not the key's data, in the case of 540 * a leaf page) to the new root page. 541 * 542 * The btree comparison code guarantees that the left-most key on any 543 * level of the tree is never used, so it doesn't need to be filled in. 544 */ 545 nbytes = NBINTERNAL(0); 546 h->linp[0] = h->upper = t->bt_psize - nbytes; 547 dest = (char *)(void *)h + h->upper; 548 WR_BINTERNAL(dest, 0, l->pgno, 0); 549 550 switch (h->flags & P_TYPE) { 551 case P_BLEAF: 552 bl = GETBLEAF(r, 0); 553 nbytes = NBINTERNAL(bl->ksize); 554 h->linp[1] = h->upper -= nbytes; 555 dest = (char *)(void *)h + h->upper; 556 WR_BINTERNAL(dest, bl->ksize, r->pgno, 0); 557 memmove(dest, bl->bytes, bl->ksize); 558 559 /* 560 * If the key is on an overflow page, mark the overflow chain 561 * so it isn't deleted when the leaf copy of the key is deleted. 562 */ 563 if (bl->flags & P_BIGKEY && 564 bt_preserve(t, *(pgno_t *)(void *)bl->bytes) == RET_ERROR) 565 return (RET_ERROR); 566 break; 567 case P_BINTERNAL: 568 bi = GETBINTERNAL(r, 0); 569 nbytes = NBINTERNAL(bi->ksize); 570 h->linp[1] = h->upper -= nbytes; 571 dest = (char *)(void *)h + h->upper; 572 memmove(dest, bi, nbytes); 573 ((BINTERNAL *)(void *)dest)->pgno = r->pgno; 574 break; 575 default: 576 abort(); 577 } 578 579 /* There are two keys on the page. */ 580 h->lower = BTDATAOFF + 2 * sizeof(indx_t); 581 582 /* Unpin the root page, set to btree internal page. */ 583 h->flags &= ~P_TYPE; 584 h->flags |= P_BINTERNAL; 585 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 586 587 return (RET_SUCCESS); 588 } 589 590 /* 591 * BT_PSPLIT -- Do the real work of splitting the page. 592 * 593 * Parameters: 594 * t: tree 595 * h: page to be split 596 * l: page to put lower half of data 597 * r: page to put upper half of data 598 * pskip: pointer to index to leave open 599 * ilen: insert length 600 * 601 * Returns: 602 * Pointer to page in which to insert. 603 */ 604 static PAGE * 605 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen) 606 { 607 BINTERNAL *bi; 608 BLEAF *bl; 609 CURSOR *c; 610 RLEAF *rl; 611 PAGE *rval; 612 void *src = NULL; /* pacify gcc */ 613 indx_t full, half, nxt, off, skip, top, used; 614 uint32_t nbytes; 615 size_t temp; 616 int bigkeycnt, isbigkey; 617 618 /* 619 * Split the data to the left and right pages. Leave the skip index 620 * open. Additionally, make some effort not to split on an overflow 621 * key. This makes internal page processing faster and can save 622 * space as overflow keys used by internal pages are never deleted. 623 */ 624 bigkeycnt = 0; 625 skip = *pskip; 626 temp = t->bt_psize - BTDATAOFF; 627 _DBFIT(temp, indx_t); 628 full = (indx_t)temp; 629 half = full / 2; 630 used = 0; 631 for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) { 632 if (skip == off) { 633 _DBFIT(ilen, uint32_t); 634 nbytes = (uint32_t)ilen; 635 isbigkey = 0; /* XXX: not really known. */ 636 } else 637 switch (h->flags & P_TYPE) { 638 case P_BINTERNAL: 639 src = bi = GETBINTERNAL(h, nxt); 640 nbytes = NBINTERNAL(bi->ksize); 641 isbigkey = bi->flags & P_BIGKEY; 642 break; 643 case P_BLEAF: 644 src = bl = GETBLEAF(h, nxt); 645 nbytes = NBLEAF(bl); 646 isbigkey = bl->flags & P_BIGKEY; 647 break; 648 case P_RINTERNAL: 649 src = GETRINTERNAL(h, nxt); 650 nbytes = NRINTERNAL; 651 isbigkey = 0; 652 break; 653 case P_RLEAF: 654 src = rl = GETRLEAF(h, nxt); 655 nbytes = NRLEAF(rl); 656 isbigkey = 0; 657 break; 658 default: 659 abort(); 660 } 661 662 /* 663 * If the key/data pairs are substantial fractions of the max 664 * possible size for the page, it's possible to get situations 665 * where we decide to try and copy too much onto the left page. 666 * Make sure that doesn't happen. 667 */ 668 if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) || 669 nxt == top - 1) { 670 --off; 671 break; 672 } 673 674 /* Copy the key/data pair, if not the skipped index. */ 675 if (skip != off) { 676 ++nxt; 677 678 l->linp[off] = l->upper -= nbytes; 679 memmove((char *)(void *)l + l->upper, src, nbytes); 680 } 681 682 temp = nbytes + sizeof(indx_t); 683 _DBFIT(temp, indx_t); 684 used += (indx_t)temp; 685 if (used >= half) { 686 if (!isbigkey || bigkeycnt == 3) 687 break; 688 else 689 ++bigkeycnt; 690 } 691 } 692 693 /* 694 * Off is the last offset that's valid for the left page. 695 * Nxt is the first offset to be placed on the right page. 696 */ 697 temp = (off + 1) * sizeof(indx_t); 698 _DBFIT(temp, indx_t); 699 l->lower += (indx_t)temp; 700 701 /* 702 * If splitting the page that the cursor was on, the cursor has to be 703 * adjusted to point to the same record as before the split. If the 704 * cursor is at or past the skipped slot, the cursor is incremented by 705 * one. If the cursor is on the right page, it is decremented by the 706 * number of records split to the left page. 707 */ 708 c = &t->bt_cursor; 709 if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) { 710 if (c->pg.index >= skip) 711 ++c->pg.index; 712 if (c->pg.index < nxt) /* Left page. */ 713 c->pg.pgno = l->pgno; 714 else { /* Right page. */ 715 c->pg.pgno = r->pgno; 716 c->pg.index -= nxt; 717 } 718 } 719 720 /* 721 * If the skipped index was on the left page, just return that page. 722 * Otherwise, adjust the skip index to reflect the new position on 723 * the right page. 724 */ 725 if (skip <= off) { 726 skip = MAX_PAGE_OFFSET; 727 rval = l; 728 } else { 729 rval = r; 730 *pskip -= nxt; 731 } 732 733 for (off = 0; nxt < top; ++off) { 734 if (skip == nxt) { 735 ++off; 736 skip = MAX_PAGE_OFFSET; 737 } 738 switch (h->flags & P_TYPE) { 739 case P_BINTERNAL: 740 src = bi = GETBINTERNAL(h, nxt); 741 nbytes = NBINTERNAL(bi->ksize); 742 break; 743 case P_BLEAF: 744 src = bl = GETBLEAF(h, nxt); 745 nbytes = NBLEAF(bl); 746 break; 747 case P_RINTERNAL: 748 src = GETRINTERNAL(h, nxt); 749 nbytes = NRINTERNAL; 750 break; 751 case P_RLEAF: 752 src = rl = GETRLEAF(h, nxt); 753 nbytes = NRLEAF(rl); 754 break; 755 default: 756 abort(); 757 } 758 ++nxt; 759 r->linp[off] = r->upper -= nbytes; 760 memmove((char *)(void *)r + r->upper, src, nbytes); 761 } 762 temp = off * sizeof(indx_t); 763 _DBFIT(temp, indx_t); 764 r->lower += (indx_t)temp; 765 766 /* If the key is being appended to the page, adjust the index. */ 767 if (skip == top) 768 r->lower += sizeof(indx_t); 769 770 return (rval); 771 } 772 773 /* 774 * BT_PRESERVE -- Mark a chain of pages as used by an internal node. 775 * 776 * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the 777 * record that references them gets deleted. Chains pointed to by internal 778 * pages never get deleted. This routine marks a chain as pointed to by an 779 * internal page. 780 * 781 * Parameters: 782 * t: tree 783 * pg: page number of first page in the chain. 784 * 785 * Returns: 786 * RET_SUCCESS, RET_ERROR. 787 */ 788 static int 789 bt_preserve(BTREE *t, pgno_t pg) 790 { 791 PAGE *h; 792 793 if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL) 794 return (RET_ERROR); 795 h->flags |= P_PRESERVE; 796 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 797 return (RET_SUCCESS); 798 } 799 800 /* 801 * REC_TOTAL -- Return the number of recno entries below a page. 802 * 803 * Parameters: 804 * h: page 805 * 806 * Returns: 807 * The number of recno entries below a page. 808 * 809 * XXX 810 * These values could be set by the bt_psplit routine. The problem is that the 811 * entry has to be popped off of the stack etc. or the values have to be passed 812 * all the way back to bt_split/bt_rroot and it's not very clean. 813 */ 814 static recno_t 815 rec_total(PAGE *h) 816 { 817 recno_t recs; 818 indx_t nxt, top; 819 820 for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt) 821 recs += GETRINTERNAL(h, nxt)->nrecs; 822 return (recs); 823 } 824
Indexes created Tue Sep 30 11:09:46 GMT 2025