1 /*- 2 * Copyright (c) 1990, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * Mike Olson. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 */ 36 37 #if defined(LIBC_SCCS) && !defined(lint) 38 static char sccsid[] = "@(#)bt_split.c 8.3 (Berkeley) 2/21/94"; 39 #endif /* LIBC_SCCS and not lint */ 40 41 #include <sys/types.h> 42 43 #include <limits.h> 44 #include <stdio.h> 45 #include <stdlib.h> 46 #include <string.h> 47 48 #include <db.h> 49 #include "btree.h" 50 51 static int bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *)); 52 static PAGE *bt_page 53 __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t)); 54 static int bt_preserve __P((BTREE *, pgno_t)); 55 static PAGE *bt_psplit 56 __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t)); 57 static PAGE *bt_root 58 __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t)); 59 static int bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *)); 60 static recno_t rec_total __P((PAGE *)); 61 62 #ifdef STATISTICS 63 u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved; 64 #endif 65 66 /* 67 * __BT_SPLIT -- Split the tree. 68 * 69 * Parameters: 70 * t: tree 71 * sp: page to split 72 * key: key to insert 73 * data: data to insert 74 * flags: BIGKEY/BIGDATA flags 75 * ilen: insert length 76 * skip: index to leave open 77 * 78 * Returns: 79 * RET_ERROR, RET_SUCCESS 80 */ 81 int 82 __bt_split(t, sp, key, data, flags, ilen, skip) 83 BTREE *t; 84 PAGE *sp; 85 const DBT *key, *data; 86 int flags; 87 size_t ilen; 88 indx_t skip; 89 { 90 BINTERNAL *bi; 91 BLEAF *bl, *tbl; 92 DBT a, b; 93 EPGNO *parent; 94 PAGE *h, *l, *r, *lchild, *rchild; 95 indx_t nxtindex; 96 size_t n, nbytes, nksize; 97 int parentsplit; 98 char *dest; 99 100 /* 101 * Split the page into two pages, l and r. The split routines return 102 * a pointer to the page into which the key should be inserted and with 103 * skip set to the offset which should be used. Additionally, l and r 104 * are pinned. 105 */ 106 h = sp->pgno == P_ROOT ? 107 bt_root(t, sp, &l, &r, &skip, ilen) : 108 bt_page(t, sp, &l, &r, &skip, ilen); 109 if (h == NULL) 110 return (RET_ERROR); 111 112 /* 113 * Insert the new key/data pair into the leaf page. (Key inserts 114 * always cause a leaf page to split first.) 115 */ 116 h->linp[skip] = h->upper -= ilen; 117 dest = (char *)h + h->upper; 118 if (ISSET(t, R_RECNO)) 119 WR_RLEAF(dest, data, flags) 120 else 121 WR_BLEAF(dest, key, data, flags) 122 123 /* If the root page was split, make it look right. */ 124 if (sp->pgno == P_ROOT && 125 (ISSET(t, R_RECNO) ? 126 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 127 goto err2; 128 129 /* 130 * Now we walk the parent page stack -- a LIFO stack of the pages that 131 * were traversed when we searched for the page that split. Each stack 132 * entry is a page number and a page index offset. The offset is for 133 * the page traversed on the search. We've just split a page, so we 134 * have to insert a new key into the parent page. 135 * 136 * If the insert into the parent page causes it to split, may have to 137 * continue splitting all the way up the tree. We stop if the root 138 * splits or the page inserted into didn't have to split to hold the 139 * new key. Some algorithms replace the key for the old page as well 140 * as the new page. We don't, as there's no reason to believe that the 141 * first key on the old page is any better than the key we have, and, 142 * in the case of a key being placed at index 0 causing the split, the 143 * key is unavailable. 144 * 145 * There are a maximum of 5 pages pinned at any time. We keep the left 146 * and right pages pinned while working on the parent. The 5 are the 147 * two children, left parent and right parent (when the parent splits) 148 * and the root page or the overflow key page when calling bt_preserve. 149 * This code must make sure that all pins are released other than the 150 * root page or overflow page which is unlocked elsewhere. 151 */ 152 while ((parent = BT_POP(t)) != NULL) { 153 lchild = l; 154 rchild = r; 155 156 /* Get the parent page. */ 157 if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL) 158 goto err2; 159 160 /* 161 * The new key goes ONE AFTER the index, because the split 162 * was to the right. 163 */ 164 skip = parent->index + 1; 165 166 /* 167 * Calculate the space needed on the parent page. 168 * 169 * Prefix trees: space hack when inserting into BINTERNAL 170 * pages. Retain only what's needed to distinguish between 171 * the new entry and the LAST entry on the page to its left. 172 * If the keys compare equal, retain the entire key. Note, 173 * we don't touch overflow keys, and the entire key must be 174 * retained for the next-to-left most key on the leftmost 175 * page of each level, or the search will fail. Applicable 176 * ONLY to internal pages that have leaf pages as children. 177 * Further reduction of the key between pairs of internal 178 * pages loses too much information. 179 */ 180 switch (rchild->flags & P_TYPE) { 181 case P_BINTERNAL: 182 bi = GETBINTERNAL(rchild, 0); 183 nbytes = NBINTERNAL(bi->ksize); 184 break; 185 case P_BLEAF: 186 bl = GETBLEAF(rchild, 0); 187 nbytes = NBINTERNAL(bl->ksize); 188 if (t->bt_pfx && !(bl->flags & P_BIGKEY) && 189 (h->prevpg != P_INVALID || skip > 1)) { 190 tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1); 191 a.size = tbl->ksize; 192 a.data = tbl->bytes; 193 b.size = bl->ksize; 194 b.data = bl->bytes; 195 nksize = t->bt_pfx(&a, &b); 196 n = NBINTERNAL(nksize); 197 if (n < nbytes) { 198 #ifdef STATISTICS 199 bt_pfxsaved += nbytes - n; 200 #endif 201 nbytes = n; 202 } else 203 nksize = 0; 204 } else 205 nksize = 0; 206 break; 207 case P_RINTERNAL: 208 case P_RLEAF: 209 nbytes = NRINTERNAL; 210 break; 211 default: 212 abort(); 213 } 214 215 /* Split the parent page if necessary or shift the indices. */ 216 if (h->upper - h->lower < nbytes + sizeof(indx_t)) { 217 sp = h; 218 h = h->pgno == P_ROOT ? 219 bt_root(t, h, &l, &r, &skip, nbytes) : 220 bt_page(t, h, &l, &r, &skip, nbytes); 221 if (h == NULL) 222 goto err1; 223 parentsplit = 1; 224 } else { 225 if (skip < (nxtindex = NEXTINDEX(h))) 226 memmove(h->linp + skip + 1, h->linp + skip, 227 (nxtindex - skip) * sizeof(indx_t)); 228 h->lower += sizeof(indx_t); 229 parentsplit = 0; 230 } 231 232 /* Insert the key into the parent page. */ 233 switch(rchild->flags & P_TYPE) { 234 case P_BINTERNAL: 235 h->linp[skip] = h->upper -= nbytes; 236 dest = (char *)h + h->linp[skip]; 237 memmove(dest, bi, nbytes); 238 ((BINTERNAL *)dest)->pgno = rchild->pgno; 239 break; 240 case P_BLEAF: 241 h->linp[skip] = h->upper -= nbytes; 242 dest = (char *)h + h->linp[skip]; 243 WR_BINTERNAL(dest, nksize ? nksize : bl->ksize, 244 rchild->pgno, bl->flags & P_BIGKEY); 245 memmove(dest, bl->bytes, nksize ? nksize : bl->ksize); 246 if (bl->flags & P_BIGKEY && 247 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR) 248 goto err1; 249 break; 250 case P_RINTERNAL: 251 /* 252 * Update the left page count. If split 253 * added at index 0, fix the correct page. 254 */ 255 if (skip > 0) 256 dest = (char *)h + h->linp[skip - 1]; 257 else 258 dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; 259 ((RINTERNAL *)dest)->nrecs = rec_total(lchild); 260 ((RINTERNAL *)dest)->pgno = lchild->pgno; 261 262 /* Update the right page count. */ 263 h->linp[skip] = h->upper -= nbytes; 264 dest = (char *)h + h->linp[skip]; 265 ((RINTERNAL *)dest)->nrecs = rec_total(rchild); 266 ((RINTERNAL *)dest)->pgno = rchild->pgno; 267 break; 268 case P_RLEAF: 269 /* 270 * Update the left page count. If split 271 * added at index 0, fix the correct page. 272 */ 273 if (skip > 0) 274 dest = (char *)h + h->linp[skip - 1]; 275 else 276 dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; 277 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild); 278 ((RINTERNAL *)dest)->pgno = lchild->pgno; 279 280 /* Update the right page count. */ 281 h->linp[skip] = h->upper -= nbytes; 282 dest = (char *)h + h->linp[skip]; 283 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild); 284 ((RINTERNAL *)dest)->pgno = rchild->pgno; 285 break; 286 default: 287 abort(); 288 } 289 290 /* Unpin the held pages. */ 291 if (!parentsplit) { 292 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 293 break; 294 } 295 296 /* If the root page was split, make it look right. */ 297 if (sp->pgno == P_ROOT && 298 (ISSET(t, R_RECNO) ? 299 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 300 goto err1; 301 302 mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 303 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 304 } 305 306 /* Unpin the held pages. */ 307 mpool_put(t->bt_mp, l, MPOOL_DIRTY); 308 mpool_put(t->bt_mp, r, MPOOL_DIRTY); 309 310 /* Clear any pages left on the stack. */ 311 return (RET_SUCCESS); 312 313 /* 314 * If something fails in the above loop we were already walking back 315 * up the tree and the tree is now inconsistent. Nothing much we can 316 * do about it but release any memory we're holding. 317 */ 318 err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 319 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 320 321 err2: mpool_put(t->bt_mp, l, 0); 322 mpool_put(t->bt_mp, r, 0); 323 __dbpanic(t->bt_dbp); 324 return (RET_ERROR); 325 } 326 327 /* 328 * BT_PAGE -- Split a non-root page of a btree. 329 * 330 * Parameters: 331 * t: tree 332 * h: root page 333 * lp: pointer to left page pointer 334 * rp: pointer to right page pointer 335 * skip: pointer to index to leave open 336 * ilen: insert length 337 * 338 * Returns: 339 * Pointer to page in which to insert or NULL on error. 340 */ 341 static PAGE * 342 bt_page(t, h, lp, rp, skip, ilen) 343 BTREE *t; 344 PAGE *h, **lp, **rp; 345 indx_t *skip; 346 size_t ilen; 347 { 348 PAGE *l, *r, *tp; 349 pgno_t npg; 350 351 #ifdef STATISTICS 352 ++bt_split; 353 #endif 354 /* Put the new right page for the split into place. */ 355 if ((r = __bt_new(t, &npg)) == NULL) 356 return (NULL); 357 r->pgno = npg; 358 r->lower = BTDATAOFF; 359 r->upper = t->bt_psize; 360 r->nextpg = h->nextpg; 361 r->prevpg = h->pgno; 362 r->flags = h->flags & P_TYPE; 363 364 /* 365 * If we're splitting the last page on a level because we're appending 366 * a key to it (skip is NEXTINDEX()), it's likely that the data is 367 * sorted. Adding an empty page on the side of the level is less work 368 * and can push the fill factor much higher than normal. If we're 369 * wrong it's no big deal, we'll just do the split the right way next 370 * time. It may look like it's equally easy to do a similar hack for 371 * reverse sorted data, that is, split the tree left, but it's not. 372 * Don't even try. 373 */ 374 if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) { 375 #ifdef STATISTICS 376 ++bt_sortsplit; 377 #endif 378 h->nextpg = r->pgno; 379 r->lower = BTDATAOFF + sizeof(indx_t); 380 *skip = 0; 381 *lp = h; 382 *rp = r; 383 return (r); 384 } 385 386 /* Put the new left page for the split into place. */ 387 if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) { 388 mpool_put(t->bt_mp, r, 0); 389 return (NULL); 390 } 391 l->pgno = h->pgno; 392 l->nextpg = r->pgno; 393 l->prevpg = h->prevpg; 394 l->lower = BTDATAOFF; 395 l->upper = t->bt_psize; 396 l->flags = h->flags & P_TYPE; 397 398 /* Fix up the previous pointer of the page after the split page. */ 399 if (h->nextpg != P_INVALID) { 400 if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) { 401 free(l); 402 /* XXX mpool_free(t->bt_mp, r->pgno); */ 403 return (NULL); 404 } 405 tp->prevpg = r->pgno; 406 mpool_put(t->bt_mp, tp, 0); 407 } 408 409 /* 410 * Split right. The key/data pairs aren't sorted in the btree page so 411 * it's simpler to copy the data from the split page onto two new pages 412 * instead of copying half the data to the right page and compacting 413 * the left page in place. Since the left page can't change, we have 414 * to swap the original and the allocated left page after the split. 415 */ 416 tp = bt_psplit(t, h, l, r, skip, ilen); 417 418 /* Move the new left page onto the old left page. */ 419 memmove(h, l, t->bt_psize); 420 if (tp == l) 421 tp = h; 422 free(l); 423 424 *lp = h; 425 *rp = r; 426 return (tp); 427 } 428 429 /* 430 * BT_ROOT -- Split the root page of a btree. 431 * 432 * Parameters: 433 * t: tree 434 * h: root page 435 * lp: pointer to left page pointer 436 * rp: pointer to right page pointer 437 * skip: pointer to index to leave open 438 * ilen: insert length 439 * 440 * Returns: 441 * Pointer to page in which to insert or NULL on error. 442 */ 443 static PAGE * 444 bt_root(t, h, lp, rp, skip, ilen) 445 BTREE *t; 446 PAGE *h, **lp, **rp; 447 indx_t *skip; 448 size_t ilen; 449 { 450 PAGE *l, *r, *tp; 451 pgno_t lnpg, rnpg; 452 453 #ifdef STATISTICS 454 ++bt_split; 455 ++bt_rootsplit; 456 #endif 457 /* Put the new left and right pages for the split into place. */ 458 if ((l = __bt_new(t, &lnpg)) == NULL || 459 (r = __bt_new(t, &rnpg)) == NULL) 460 return (NULL); 461 l->pgno = lnpg; 462 r->pgno = rnpg; 463 l->nextpg = r->pgno; 464 r->prevpg = l->pgno; 465 l->prevpg = r->nextpg = P_INVALID; 466 l->lower = r->lower = BTDATAOFF; 467 l->upper = r->upper = t->bt_psize; 468 l->flags = r->flags = h->flags & P_TYPE; 469 470 /* Split the root page. */ 471 tp = bt_psplit(t, h, l, r, skip, ilen); 472 473 *lp = l; 474 *rp = r; 475 return (tp); 476 } 477 478 /* 479 * BT_RROOT -- Fix up the recno root page after it has been split. 480 * 481 * Parameters: 482 * t: tree 483 * h: root page 484 * l: left page 485 * r: right page 486 * 487 * Returns: 488 * RET_ERROR, RET_SUCCESS 489 */ 490 static int 491 bt_rroot(t, h, l, r) 492 BTREE *t; 493 PAGE *h, *l, *r; 494 { 495 char *dest; 496 497 /* Insert the left and right keys, set the header information. */ 498 h->linp[0] = h->upper = t->bt_psize - NRINTERNAL; 499 dest = (char *)h + h->upper; 500 WR_RINTERNAL(dest, 501 l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno); 502 503 h->linp[1] = h->upper -= NRINTERNAL; 504 dest = (char *)h + h->upper; 505 WR_RINTERNAL(dest, 506 r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno); 507 508 h->lower = BTDATAOFF + 2 * sizeof(indx_t); 509 510 /* Unpin the root page, set to recno internal page. */ 511 h->flags &= ~P_TYPE; 512 h->flags |= P_RINTERNAL; 513 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 514 515 return (RET_SUCCESS); 516 } 517 518 /* 519 * BT_BROOT -- Fix up the btree root page after it has been split. 520 * 521 * Parameters: 522 * t: tree 523 * h: root page 524 * l: left page 525 * r: right page 526 * 527 * Returns: 528 * RET_ERROR, RET_SUCCESS 529 */ 530 static int 531 bt_broot(t, h, l, r) 532 BTREE *t; 533 PAGE *h, *l, *r; 534 { 535 BINTERNAL *bi; 536 BLEAF *bl; 537 size_t nbytes; 538 char *dest; 539 540 /* 541 * If the root page was a leaf page, change it into an internal page. 542 * We copy the key we split on (but not the key's data, in the case of 543 * a leaf page) to the new root page. 544 * 545 * The btree comparison code guarantees that the left-most key on any 546 * level of the tree is never used, so it doesn't need to be filled in. 547 */ 548 nbytes = NBINTERNAL(0); 549 h->linp[0] = h->upper = t->bt_psize - nbytes; 550 dest = (char *)h + h->upper; 551 WR_BINTERNAL(dest, 0, l->pgno, 0); 552 553 switch(h->flags & P_TYPE) { 554 case P_BLEAF: 555 bl = GETBLEAF(r, 0); 556 nbytes = NBINTERNAL(bl->ksize); 557 h->linp[1] = h->upper -= nbytes; 558 dest = (char *)h + h->upper; 559 WR_BINTERNAL(dest, bl->ksize, r->pgno, 0); 560 memmove(dest, bl->bytes, bl->ksize); 561 562 /* 563 * If the key is on an overflow page, mark the overflow chain 564 * so it isn't deleted when the leaf copy of the key is deleted. 565 */ 566 if (bl->flags & P_BIGKEY && 567 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR) 568 return (RET_ERROR); 569 break; 570 case P_BINTERNAL: 571 bi = GETBINTERNAL(r, 0); 572 nbytes = NBINTERNAL(bi->ksize); 573 h->linp[1] = h->upper -= nbytes; 574 dest = (char *)h + h->upper; 575 memmove(dest, bi, nbytes); 576 ((BINTERNAL *)dest)->pgno = r->pgno; 577 break; 578 default: 579 abort(); 580 } 581 582 /* There are two keys on the page. */ 583 h->lower = BTDATAOFF + 2 * sizeof(indx_t); 584 585 /* Unpin the root page, set to btree internal page. */ 586 h->flags &= ~P_TYPE; 587 h->flags |= P_BINTERNAL; 588 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 589 590 return (RET_SUCCESS); 591 } 592 593 /* 594 * BT_PSPLIT -- Do the real work of splitting the page. 595 * 596 * Parameters: 597 * t: tree 598 * h: page to be split 599 * l: page to put lower half of data 600 * r: page to put upper half of data 601 * pskip: pointer to index to leave open 602 * ilen: insert length 603 * 604 * Returns: 605 * Pointer to page in which to insert. 606 */ 607 static PAGE * 608 bt_psplit(t, h, l, r, pskip, ilen) 609 BTREE *t; 610 PAGE *h, *l, *r; 611 indx_t *pskip; 612 size_t ilen; 613 { 614 BINTERNAL *bi; 615 BLEAF *bl; 616 RLEAF *rl; 617 EPGNO *c; 618 PAGE *rval; 619 void *src; 620 indx_t full, half, nxt, off, skip, top, used; 621 size_t nbytes; 622 int bigkeycnt, isbigkey; 623 624 /* 625 * Split the data to the left and right pages. Leave the skip index 626 * open. Additionally, make some effort not to split on an overflow 627 * key. This makes internal page processing faster and can save 628 * space as overflow keys used by internal pages are never deleted. 629 */ 630 bigkeycnt = 0; 631 skip = *pskip; 632 full = t->bt_psize - BTDATAOFF; 633 half = full / 2; 634 used = 0; 635 for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) { 636 if (skip == off) { 637 nbytes = ilen; 638 isbigkey = 0; /* XXX: not really known. */ 639 } else 640 switch (h->flags & P_TYPE) { 641 case P_BINTERNAL: 642 src = bi = GETBINTERNAL(h, nxt); 643 nbytes = NBINTERNAL(bi->ksize); 644 isbigkey = bi->flags & P_BIGKEY; 645 break; 646 case P_BLEAF: 647 src = bl = GETBLEAF(h, nxt); 648 nbytes = NBLEAF(bl); 649 isbigkey = bl->flags & P_BIGKEY; 650 break; 651 case P_RINTERNAL: 652 src = GETRINTERNAL(h, nxt); 653 nbytes = NRINTERNAL; 654 isbigkey = 0; 655 break; 656 case P_RLEAF: 657 src = rl = GETRLEAF(h, nxt); 658 nbytes = NRLEAF(rl); 659 isbigkey = 0; 660 break; 661 default: 662 abort(); 663 } 664 665 /* 666 * If the key/data pairs are substantial fractions of the max 667 * possible size for the page, it's possible to get situations 668 * where we decide to try and copy too much onto the left page. 669 * Make sure that doesn't happen. 670 */ 671 if (skip <= off && used + nbytes >= full) { 672 --off; 673 break; 674 } 675 676 /* Copy the key/data pair, if not the skipped index. */ 677 if (skip != off) { 678 ++nxt; 679 680 l->linp[off] = l->upper -= nbytes; 681 memmove((char *)l + l->upper, src, nbytes); 682 } 683 684 used += nbytes; 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 l->lower += (off + 1) * sizeof(indx_t); 698 699 /* 700 * If splitting the page that the cursor was on, the cursor has to be 701 * adjusted to point to the same record as before the split. If the 702 * cursor is at or past the skipped slot, the cursor is incremented by 703 * one. If the cursor is on the right page, it is decremented by the 704 * number of records split to the left page. 705 * 706 * Don't bother checking for the B_SEQINIT flag, the page number will 707 * be P_INVALID. 708 */ 709 c = &t->bt_bcursor; 710 if (c->pgno == h->pgno) { 711 if (c->index >= skip) 712 ++c->index; 713 if (c->index < nxt) /* Left page. */ 714 c->pgno = l->pgno; 715 else { /* Right page. */ 716 c->pgno = r->pgno; 717 c->index -= nxt; 718 } 719 } 720 721 /* 722 * If the skipped index was on the left page, just return that page. 723 * Otherwise, adjust the skip index to reflect the new position on 724 * the right page. 725 */ 726 if (skip <= off) { 727 skip = 0; 728 rval = l; 729 } else { 730 rval = r; 731 *pskip -= nxt; 732 } 733 734 for (off = 0; nxt < top; ++off) { 735 if (skip == nxt) { 736 ++off; 737 skip = 0; 738 } 739 switch (h->flags & P_TYPE) { 740 case P_BINTERNAL: 741 src = bi = GETBINTERNAL(h, nxt); 742 nbytes = NBINTERNAL(bi->ksize); 743 break; 744 case P_BLEAF: 745 src = bl = GETBLEAF(h, nxt); 746 nbytes = NBLEAF(bl); 747 break; 748 case P_RINTERNAL: 749 src = GETRINTERNAL(h, nxt); 750 nbytes = NRINTERNAL; 751 break; 752 case P_RLEAF: 753 src = rl = GETRLEAF(h, nxt); 754 nbytes = NRLEAF(rl); 755 break; 756 default: 757 abort(); 758 } 759 ++nxt; 760 r->linp[off] = r->upper -= nbytes; 761 memmove((char *)r + r->upper, src, nbytes); 762 } 763 r->lower += off * sizeof(indx_t); 764 765 /* If the key is being appended to the page, adjust the index. */ 766 if (skip == top) 767 r->lower += sizeof(indx_t); 768 769 return (rval); 770 } 771 772 /* 773 * BT_PRESERVE -- Mark a chain of pages as used by an internal node. 774 * 775 * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the 776 * record that references them gets deleted. Chains pointed to by internal 777 * pages never get deleted. This routine marks a chain as pointed to by an 778 * internal page. 779 * 780 * Parameters: 781 * t: tree 782 * pg: page number of first page in the chain. 783 * 784 * Returns: 785 * RET_SUCCESS, RET_ERROR. 786 */ 787 static int 788 bt_preserve(t, pg) 789 BTREE *t; 790 pgno_t pg; 791 { 792 PAGE *h; 793 794 if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL) 795 return (RET_ERROR); 796 h->flags |= P_PRESERVE; 797 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 798 return (RET_SUCCESS); 799 } 800 801 /* 802 * REC_TOTAL -- Return the number of recno entries below a page. 803 * 804 * Parameters: 805 * h: page 806 * 807 * Returns: 808 * The number of recno entries below a page. 809 * 810 * XXX 811 * These values could be set by the bt_psplit routine. The problem is that the 812 * entry has to be popped off of the stack etc. or the values have to be passed 813 * all the way back to bt_split/bt_rroot and it's not very clean. 814 */ 815 static recno_t 816 rec_total(h) 817 PAGE *h; 818 { 819 recno_t recs; 820 indx_t nxt, top; 821 822 for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt) 823 recs += GETRINTERNAL(h, nxt)->nrecs; 824 return (recs); 825 } 826