1 /* $OpenBSD: bt_split.c,v 1.8 2002/02/16 21:27:22 millert 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. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 */ 38 39 #if defined(LIBC_SCCS) && !defined(lint) 40 #if 0 41 static char rcsid[] = "$OpenBSD: bt_split.c,v 1.8 2002/02/16 21:27:22 millert Exp $"; 42 #else 43 static char sccsid[] = "@(#)bt_split.c 8.10 (Berkeley) 1/9/95"; 44 #endif 45 #endif /* LIBC_SCCS and not lint */ 46 47 #include <sys/types.h> 48 49 #include <limits.h> 50 #include <stdio.h> 51 #include <stdlib.h> 52 #include <string.h> 53 54 #include <db.h> 55 #include "btree.h" 56 57 static int bt_broot(BTREE *, PAGE *, PAGE *, PAGE *); 58 static PAGE *bt_page 59 (BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t); 60 static int bt_preserve(BTREE *, pgno_t); 61 static PAGE *bt_psplit 62 (BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t); 63 static PAGE *bt_root 64 (BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t); 65 static int bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *); 66 static recno_t rec_total(PAGE *); 67 68 #ifdef STATISTICS 69 u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved; 70 #endif 71 72 /* 73 * __BT_SPLIT -- Split the tree. 74 * 75 * Parameters: 76 * t: tree 77 * sp: page to split 78 * key: key to insert 79 * data: data to insert 80 * flags: BIGKEY/BIGDATA flags 81 * ilen: insert length 82 * skip: index to leave open 83 * 84 * Returns: 85 * RET_ERROR, RET_SUCCESS 86 */ 87 int 88 __bt_split(t, sp, key, data, flags, ilen, argskip) 89 BTREE *t; 90 PAGE *sp; 91 const DBT *key, *data; 92 int flags; 93 size_t ilen; 94 u_int32_t argskip; 95 { 96 BINTERNAL *bi; 97 BLEAF *bl, *tbl; 98 DBT a, b; 99 EPGNO *parent; 100 PAGE *h, *l, *r, *lchild, *rchild; 101 indx_t nxtindex; 102 u_int16_t skip; 103 u_int32_t n, nbytes, nksize; 104 int parentsplit; 105 char *dest; 106 107 /* 108 * Split the page into two pages, l and r. The split routines return 109 * a pointer to the page into which the key should be inserted and with 110 * skip set to the offset which should be used. Additionally, l and r 111 * are pinned. 112 */ 113 skip = argskip; 114 h = sp->pgno == P_ROOT ? 115 bt_root(t, sp, &l, &r, &skip, ilen) : 116 bt_page(t, sp, &l, &r, &skip, ilen); 117 if (h == NULL) 118 return (RET_ERROR); 119 120 /* 121 * Insert the new key/data pair into the leaf page. (Key inserts 122 * always cause a leaf page to split first.) 123 */ 124 h->linp[skip] = h->upper -= ilen; 125 dest = (char *)h + h->upper; 126 if (F_ISSET(t, R_RECNO)) 127 WR_RLEAF(dest, data, flags) 128 else 129 WR_BLEAF(dest, key, data, flags) 130 131 /* If the root page was split, make it look right. */ 132 if (sp->pgno == P_ROOT && 133 (F_ISSET(t, R_RECNO) ? 134 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 135 goto err2; 136 137 /* 138 * Now we walk the parent page stack -- a LIFO stack of the pages that 139 * were traversed when we searched for the page that split. Each stack 140 * entry is a page number and a page index offset. The offset is for 141 * the page traversed on the search. We've just split a page, so we 142 * have to insert a new key into the parent page. 143 * 144 * If the insert into the parent page causes it to split, may have to 145 * continue splitting all the way up the tree. We stop if the root 146 * splits or the page inserted into didn't have to split to hold the 147 * new key. Some algorithms replace the key for the old page as well 148 * as the new page. We don't, as there's no reason to believe that the 149 * first key on the old page is any better than the key we have, and, 150 * in the case of a key being placed at index 0 causing the split, the 151 * key is unavailable. 152 * 153 * There are a maximum of 5 pages pinned at any time. We keep the left 154 * and right pages pinned while working on the parent. The 5 are the 155 * two children, left parent and right parent (when the parent splits) 156 * and the root page or the overflow key page when calling bt_preserve. 157 * This code must make sure that all pins are released other than the 158 * root page or overflow page which is unlocked elsewhere. 159 */ 160 while ((parent = BT_POP(t)) != NULL) { 161 lchild = l; 162 rchild = r; 163 164 /* Get the parent page. */ 165 if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL) 166 goto err2; 167 168 /* 169 * The new key goes ONE AFTER the index, because the split 170 * was to the right. 171 */ 172 skip = parent->index + 1; 173 174 /* 175 * Calculate the space needed on the parent page. 176 * 177 * Prefix trees: space hack when inserting into BINTERNAL 178 * pages. Retain only what's needed to distinguish between 179 * the new entry and the LAST entry on the page to its left. 180 * If the keys compare equal, retain the entire key. Note, 181 * we don't touch overflow keys, and the entire key must be 182 * retained for the next-to-left most key on the leftmost 183 * page of each level, or the search will fail. Applicable 184 * ONLY to internal pages that have leaf pages as children. 185 * Further reduction of the key between pairs of internal 186 * pages loses too much information. 187 */ 188 switch (rchild->flags & P_TYPE) { 189 case P_BINTERNAL: 190 bi = GETBINTERNAL(rchild, 0); 191 nbytes = NBINTERNAL(bi->ksize); 192 break; 193 case P_BLEAF: 194 bl = GETBLEAF(rchild, 0); 195 nbytes = NBINTERNAL(bl->ksize); 196 if (t->bt_pfx && !(bl->flags & P_BIGKEY) && 197 (h->prevpg != P_INVALID || skip > 1)) { 198 tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1); 199 a.size = tbl->ksize; 200 a.data = tbl->bytes; 201 b.size = bl->ksize; 202 b.data = bl->bytes; 203 nksize = t->bt_pfx(&a, &b); 204 n = NBINTERNAL(nksize); 205 if (n < nbytes) { 206 #ifdef STATISTICS 207 bt_pfxsaved += nbytes - n; 208 #endif 209 nbytes = n; 210 } else 211 nksize = 0; 212 } else 213 nksize = 0; 214 break; 215 case P_RINTERNAL: 216 case P_RLEAF: 217 nbytes = NRINTERNAL; 218 break; 219 default: 220 abort(); 221 } 222 223 /* Split the parent page if necessary or shift the indices. */ 224 if (h->upper - h->lower < nbytes + sizeof(indx_t)) { 225 sp = h; 226 h = h->pgno == P_ROOT ? 227 bt_root(t, h, &l, &r, &skip, nbytes) : 228 bt_page(t, h, &l, &r, &skip, nbytes); 229 if (h == NULL) 230 goto err1; 231 parentsplit = 1; 232 } else { 233 if (skip < (nxtindex = NEXTINDEX(h))) 234 memmove(h->linp + skip + 1, h->linp + skip, 235 (nxtindex - skip) * sizeof(indx_t)); 236 h->lower += sizeof(indx_t); 237 parentsplit = 0; 238 } 239 240 /* Insert the key into the parent page. */ 241 switch (rchild->flags & P_TYPE) { 242 case P_BINTERNAL: 243 h->linp[skip] = h->upper -= nbytes; 244 dest = (char *)h + h->linp[skip]; 245 memmove(dest, bi, nbytes); 246 ((BINTERNAL *)dest)->pgno = rchild->pgno; 247 break; 248 case P_BLEAF: 249 h->linp[skip] = h->upper -= nbytes; 250 dest = (char *)h + h->linp[skip]; 251 WR_BINTERNAL(dest, nksize ? nksize : bl->ksize, 252 rchild->pgno, bl->flags & P_BIGKEY); 253 memmove(dest, bl->bytes, nksize ? nksize : bl->ksize); 254 if (bl->flags & P_BIGKEY && 255 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR) 256 goto err1; 257 break; 258 case P_RINTERNAL: 259 /* 260 * Update the left page count. If split 261 * added at index 0, fix the correct page. 262 */ 263 if (skip > 0) 264 dest = (char *)h + h->linp[skip - 1]; 265 else 266 dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; 267 ((RINTERNAL *)dest)->nrecs = rec_total(lchild); 268 ((RINTERNAL *)dest)->pgno = lchild->pgno; 269 270 /* Update the right page count. */ 271 h->linp[skip] = h->upper -= nbytes; 272 dest = (char *)h + h->linp[skip]; 273 ((RINTERNAL *)dest)->nrecs = rec_total(rchild); 274 ((RINTERNAL *)dest)->pgno = rchild->pgno; 275 break; 276 case P_RLEAF: 277 /* 278 * Update the left page count. If split 279 * added at index 0, fix the correct page. 280 */ 281 if (skip > 0) 282 dest = (char *)h + h->linp[skip - 1]; 283 else 284 dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; 285 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild); 286 ((RINTERNAL *)dest)->pgno = lchild->pgno; 287 288 /* Update the right page count. */ 289 h->linp[skip] = h->upper -= nbytes; 290 dest = (char *)h + h->linp[skip]; 291 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild); 292 ((RINTERNAL *)dest)->pgno = rchild->pgno; 293 break; 294 default: 295 abort(); 296 } 297 298 /* Unpin the held pages. */ 299 if (!parentsplit) { 300 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 301 break; 302 } 303 304 /* If the root page was split, make it look right. */ 305 if (sp->pgno == P_ROOT && 306 (F_ISSET(t, R_RECNO) ? 307 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 308 goto err1; 309 310 mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 311 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 312 } 313 314 /* Unpin the held pages. */ 315 mpool_put(t->bt_mp, l, MPOOL_DIRTY); 316 mpool_put(t->bt_mp, r, MPOOL_DIRTY); 317 318 /* Clear any pages left on the stack. */ 319 return (RET_SUCCESS); 320 321 /* 322 * If something fails in the above loop we were already walking back 323 * up the tree and the tree is now inconsistent. Nothing much we can 324 * do about it but release any memory we're holding. 325 */ 326 err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 327 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 328 329 err2: mpool_put(t->bt_mp, l, 0); 330 mpool_put(t->bt_mp, r, 0); 331 __dbpanic(t->bt_dbp); 332 return (RET_ERROR); 333 } 334 335 /* 336 * BT_PAGE -- Split a non-root page of a btree. 337 * 338 * Parameters: 339 * t: tree 340 * h: root page 341 * lp: pointer to left page pointer 342 * rp: pointer to right page pointer 343 * skip: pointer to index to leave open 344 * ilen: insert length 345 * 346 * Returns: 347 * Pointer to page in which to insert or NULL on error. 348 */ 349 static PAGE * 350 bt_page(t, h, lp, rp, skip, ilen) 351 BTREE *t; 352 PAGE *h, **lp, **rp; 353 indx_t *skip; 354 size_t ilen; 355 { 356 PAGE *l, *r, *tp; 357 pgno_t npg; 358 359 #ifdef STATISTICS 360 ++bt_split; 361 #endif 362 /* Put the new right page for the split into place. */ 363 if ((r = __bt_new(t, &npg)) == NULL) 364 return (NULL); 365 r->pgno = npg; 366 r->lower = BTDATAOFF; 367 r->upper = t->bt_psize; 368 r->nextpg = h->nextpg; 369 r->prevpg = h->pgno; 370 r->flags = h->flags & P_TYPE; 371 372 /* 373 * If we're splitting the last page on a level because we're appending 374 * a key to it (skip is NEXTINDEX()), it's likely that the data is 375 * sorted. Adding an empty page on the side of the level is less work 376 * and can push the fill factor much higher than normal. If we're 377 * wrong it's no big deal, we'll just do the split the right way next 378 * time. It may look like it's equally easy to do a similar hack for 379 * reverse sorted data, that is, split the tree left, but it's not. 380 * Don't even try. 381 */ 382 if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) { 383 #ifdef STATISTICS 384 ++bt_sortsplit; 385 #endif 386 h->nextpg = r->pgno; 387 r->lower = BTDATAOFF + sizeof(indx_t); 388 *skip = 0; 389 *lp = h; 390 *rp = r; 391 return (r); 392 } 393 394 /* Put the new left page for the split into place. */ 395 if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) { 396 mpool_put(t->bt_mp, r, 0); 397 return (NULL); 398 } 399 memset(l, 0xff, t->bt_psize); 400 l->pgno = h->pgno; 401 l->nextpg = r->pgno; 402 l->prevpg = h->prevpg; 403 l->lower = BTDATAOFF; 404 l->upper = t->bt_psize; 405 l->flags = h->flags & P_TYPE; 406 407 /* Fix up the previous pointer of the page after the split page. */ 408 if (h->nextpg != P_INVALID) { 409 if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) { 410 free(l); 411 /* XXX mpool_free(t->bt_mp, r->pgno); */ 412 return (NULL); 413 } 414 tp->prevpg = r->pgno; 415 mpool_put(t->bt_mp, tp, MPOOL_DIRTY); 416 } 417 418 /* 419 * Split right. The key/data pairs aren't sorted in the btree page so 420 * it's simpler to copy the data from the split page onto two new pages 421 * instead of copying half the data to the right page and compacting 422 * the left page in place. Since the left page can't change, we have 423 * to swap the original and the allocated left page after the split. 424 */ 425 tp = bt_psplit(t, h, l, r, skip, ilen); 426 427 /* Move the new left page onto the old left page. */ 428 memmove(h, l, t->bt_psize); 429 if (tp == l) 430 tp = h; 431 free(l); 432 433 *lp = h; 434 *rp = r; 435 return (tp); 436 } 437 438 /* 439 * BT_ROOT -- Split the root page of a btree. 440 * 441 * Parameters: 442 * t: tree 443 * h: root page 444 * lp: pointer to left page pointer 445 * rp: pointer to right page pointer 446 * skip: pointer to index to leave open 447 * ilen: insert length 448 * 449 * Returns: 450 * Pointer to page in which to insert or NULL on error. 451 */ 452 static PAGE * 453 bt_root(t, h, lp, rp, skip, ilen) 454 BTREE *t; 455 PAGE *h, **lp, **rp; 456 indx_t *skip; 457 size_t ilen; 458 { 459 PAGE *l, *r, *tp; 460 pgno_t lnpg, rnpg; 461 462 #ifdef STATISTICS 463 ++bt_split; 464 ++bt_rootsplit; 465 #endif 466 /* Put the new left and right pages for the split into place. */ 467 if ((l = __bt_new(t, &lnpg)) == NULL || 468 (r = __bt_new(t, &rnpg)) == NULL) 469 return (NULL); 470 l->pgno = lnpg; 471 r->pgno = rnpg; 472 l->nextpg = r->pgno; 473 r->prevpg = l->pgno; 474 l->prevpg = r->nextpg = P_INVALID; 475 l->lower = r->lower = BTDATAOFF; 476 l->upper = r->upper = t->bt_psize; 477 l->flags = r->flags = h->flags & P_TYPE; 478 479 /* Split the root page. */ 480 tp = bt_psplit(t, h, l, r, skip, ilen); 481 482 *lp = l; 483 *rp = r; 484 return (tp); 485 } 486 487 /* 488 * BT_RROOT -- Fix up the recno root page after it has been split. 489 * 490 * Parameters: 491 * t: tree 492 * h: root page 493 * l: left page 494 * r: right page 495 * 496 * Returns: 497 * RET_ERROR, RET_SUCCESS 498 */ 499 static int 500 bt_rroot(t, h, l, r) 501 BTREE *t; 502 PAGE *h, *l, *r; 503 { 504 char *dest; 505 506 /* Insert the left and right keys, set the header information. */ 507 h->linp[0] = h->upper = t->bt_psize - NRINTERNAL; 508 dest = (char *)h + h->upper; 509 WR_RINTERNAL(dest, 510 l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno); 511 512 h->linp[1] = h->upper -= NRINTERNAL; 513 dest = (char *)h + h->upper; 514 WR_RINTERNAL(dest, 515 r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno); 516 517 h->lower = BTDATAOFF + 2 * sizeof(indx_t); 518 519 /* Unpin the root page, set to recno internal page. */ 520 h->flags &= ~P_TYPE; 521 h->flags |= P_RINTERNAL; 522 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 523 524 return (RET_SUCCESS); 525 } 526 527 /* 528 * BT_BROOT -- Fix up the btree root page after it has been split. 529 * 530 * Parameters: 531 * t: tree 532 * h: root page 533 * l: left page 534 * r: right page 535 * 536 * Returns: 537 * RET_ERROR, RET_SUCCESS 538 */ 539 static int 540 bt_broot(t, h, l, r) 541 BTREE *t; 542 PAGE *h, *l, *r; 543 { 544 BINTERNAL *bi; 545 BLEAF *bl; 546 u_int32_t nbytes; 547 char *dest; 548 549 /* 550 * If the root page was a leaf page, change it into an internal page. 551 * We copy the key we split on (but not the key's data, in the case of 552 * a leaf page) to the new root page. 553 * 554 * The btree comparison code guarantees that the left-most key on any 555 * level of the tree is never used, so it doesn't need to be filled in. 556 */ 557 nbytes = NBINTERNAL(0); 558 h->linp[0] = h->upper = t->bt_psize - nbytes; 559 dest = (char *)h + h->upper; 560 WR_BINTERNAL(dest, 0, l->pgno, 0); 561 562 switch (h->flags & P_TYPE) { 563 case P_BLEAF: 564 bl = GETBLEAF(r, 0); 565 nbytes = NBINTERNAL(bl->ksize); 566 h->linp[1] = h->upper -= nbytes; 567 dest = (char *)h + h->upper; 568 WR_BINTERNAL(dest, bl->ksize, r->pgno, 0); 569 memmove(dest, bl->bytes, bl->ksize); 570 571 /* 572 * If the key is on an overflow page, mark the overflow chain 573 * so it isn't deleted when the leaf copy of the key is deleted. 574 */ 575 if (bl->flags & P_BIGKEY && 576 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR) 577 return (RET_ERROR); 578 break; 579 case P_BINTERNAL: 580 bi = GETBINTERNAL(r, 0); 581 nbytes = NBINTERNAL(bi->ksize); 582 h->linp[1] = h->upper -= nbytes; 583 dest = (char *)h + h->upper; 584 memmove(dest, bi, nbytes); 585 ((BINTERNAL *)dest)->pgno = r->pgno; 586 break; 587 default: 588 abort(); 589 } 590 591 /* There are two keys on the page. */ 592 h->lower = BTDATAOFF + 2 * sizeof(indx_t); 593 594 /* Unpin the root page, set to btree internal page. */ 595 h->flags &= ~P_TYPE; 596 h->flags |= P_BINTERNAL; 597 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 598 599 return (RET_SUCCESS); 600 } 601 602 /* 603 * BT_PSPLIT -- Do the real work of splitting the page. 604 * 605 * Parameters: 606 * t: tree 607 * h: page to be split 608 * l: page to put lower half of data 609 * r: page to put upper half of data 610 * pskip: pointer to index to leave open 611 * ilen: insert length 612 * 613 * Returns: 614 * Pointer to page in which to insert. 615 */ 616 static PAGE * 617 bt_psplit(t, h, l, r, pskip, ilen) 618 BTREE *t; 619 PAGE *h, *l, *r; 620 indx_t *pskip; 621 size_t ilen; 622 { 623 BINTERNAL *bi; 624 BLEAF *bl; 625 CURSOR *c; 626 RLEAF *rl; 627 PAGE *rval; 628 void *src; 629 indx_t full, half, nxt, off, skip, top, used; 630 u_int32_t nbytes; 631 int bigkeycnt, isbigkey; 632 633 /* 634 * Split the data to the left and right pages. Leave the skip index 635 * open. Additionally, make some effort not to split on an overflow 636 * key. This makes internal page processing faster and can save 637 * space as overflow keys used by internal pages are never deleted. 638 */ 639 bigkeycnt = 0; 640 skip = *pskip; 641 full = t->bt_psize - BTDATAOFF; 642 half = full / 2; 643 used = 0; 644 for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) { 645 if (skip == off) { 646 nbytes = ilen; 647 isbigkey = 0; /* XXX: not really known. */ 648 } else 649 switch (h->flags & P_TYPE) { 650 case P_BINTERNAL: 651 src = bi = GETBINTERNAL(h, nxt); 652 nbytes = NBINTERNAL(bi->ksize); 653 isbigkey = bi->flags & P_BIGKEY; 654 break; 655 case P_BLEAF: 656 src = bl = GETBLEAF(h, nxt); 657 nbytes = NBLEAF(bl); 658 isbigkey = bl->flags & P_BIGKEY; 659 break; 660 case P_RINTERNAL: 661 src = GETRINTERNAL(h, nxt); 662 nbytes = NRINTERNAL; 663 isbigkey = 0; 664 break; 665 case P_RLEAF: 666 src = rl = GETRLEAF(h, nxt); 667 nbytes = NRLEAF(rl); 668 isbigkey = 0; 669 break; 670 default: 671 abort(); 672 } 673 674 /* 675 * If the key/data pairs are substantial fractions of the max 676 * possible size for the page, it's possible to get situations 677 * where we decide to try and copy too much onto the left page. 678 * Make sure that doesn't happen. 679 */ 680 if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) || 681 nxt == top - 1) { 682 --off; 683 break; 684 } 685 686 /* Copy the key/data pair, if not the skipped index. */ 687 if (skip != off) { 688 ++nxt; 689 690 l->linp[off] = l->upper -= nbytes; 691 memmove((char *)l + l->upper, src, nbytes); 692 } 693 694 used += nbytes + sizeof(indx_t); 695 if (used >= half) { 696 if (!isbigkey || bigkeycnt == 3) 697 break; 698 else 699 ++bigkeycnt; 700 } 701 } 702 703 /* 704 * Off is the last offset that's valid for the left page. 705 * Nxt is the first offset to be placed on the right page. 706 */ 707 l->lower += (off + 1) * sizeof(indx_t); 708 709 /* 710 * If splitting the page that the cursor was on, the cursor has to be 711 * adjusted to point to the same record as before the split. If the 712 * cursor is at or past the skipped slot, the cursor is incremented by 713 * one. If the cursor is on the right page, it is decremented by the 714 * number of records split to the left page. 715 */ 716 c = &t->bt_cursor; 717 if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) { 718 if (c->pg.index >= skip) 719 ++c->pg.index; 720 if (c->pg.index < nxt) /* Left page. */ 721 c->pg.pgno = l->pgno; 722 else { /* Right page. */ 723 c->pg.pgno = r->pgno; 724 c->pg.index -= nxt; 725 } 726 } 727 728 /* 729 * If the skipped index was on the left page, just return that page. 730 * Otherwise, adjust the skip index to reflect the new position on 731 * the right page. 732 */ 733 if (skip <= off) { 734 skip = MAX_PAGE_OFFSET; 735 rval = l; 736 } else { 737 rval = r; 738 *pskip -= nxt; 739 } 740 741 for (off = 0; nxt < top; ++off) { 742 if (skip == nxt) { 743 ++off; 744 skip = MAX_PAGE_OFFSET; 745 } 746 switch (h->flags & P_TYPE) { 747 case P_BINTERNAL: 748 src = bi = GETBINTERNAL(h, nxt); 749 nbytes = NBINTERNAL(bi->ksize); 750 break; 751 case P_BLEAF: 752 src = bl = GETBLEAF(h, nxt); 753 nbytes = NBLEAF(bl); 754 break; 755 case P_RINTERNAL: 756 src = GETRINTERNAL(h, nxt); 757 nbytes = NRINTERNAL; 758 break; 759 case P_RLEAF: 760 src = rl = GETRLEAF(h, nxt); 761 nbytes = NRLEAF(rl); 762 break; 763 default: 764 abort(); 765 } 766 ++nxt; 767 r->linp[off] = r->upper -= nbytes; 768 memmove((char *)r + r->upper, src, nbytes); 769 } 770 r->lower += off * sizeof(indx_t); 771 772 /* If the key is being appended to the page, adjust the index. */ 773 if (skip == top) 774 r->lower += sizeof(indx_t); 775 776 return (rval); 777 } 778 779 /* 780 * BT_PRESERVE -- Mark a chain of pages as used by an internal node. 781 * 782 * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the 783 * record that references them gets deleted. Chains pointed to by internal 784 * pages never get deleted. This routine marks a chain as pointed to by an 785 * internal page. 786 * 787 * Parameters: 788 * t: tree 789 * pg: page number of first page in the chain. 790 * 791 * Returns: 792 * RET_SUCCESS, RET_ERROR. 793 */ 794 static int 795 bt_preserve(t, pg) 796 BTREE *t; 797 pgno_t pg; 798 { 799 PAGE *h; 800 801 if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL) 802 return (RET_ERROR); 803 h->flags |= P_PRESERVE; 804 mpool_put(t->bt_mp, h, MPOOL_DIRTY); 805 return (RET_SUCCESS); 806 } 807 808 /* 809 * REC_TOTAL -- Return the number of recno entries below a page. 810 * 811 * Parameters: 812 * h: page 813 * 814 * Returns: 815 * The number of recno entries below a page. 816 * 817 * XXX 818 * These values could be set by the bt_psplit routine. The problem is that the 819 * entry has to be popped off of the stack etc. or the values have to be passed 820 * all the way back to bt_split/bt_rroot and it's not very clean. 821 */ 822 static recno_t 823 rec_total(h) 824 PAGE *h; 825 { 826 recno_t recs; 827 indx_t nxt, top; 828 829 for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt) 830 recs += GETRINTERNAL(h, nxt)->nrecs; 831 return (recs); 832 } 833