2 * Copyright (c) 1990, 1993, 1994
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
37 #if defined(LIBC_SCCS) && !defined(lint)
38 static char sccsid[] = "@(#)bt_split.c 8.9 (Berkeley) 7/26/94";
39 #endif /* LIBC_SCCS and not lint */
41 #include <sys/types.h>
51 static int bt_broot __P((BTREE *, PAGE *, PAGE *, 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));
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 *));
63 u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
67 * __BT_SPLIT -- Split the tree.
73 * data: data to insert
74 * flags: BIGKEY/BIGDATA flags
76 * skip: index to leave open
79 * RET_ERROR, RET_SUCCESS
82 __bt_split(t, sp, key, data, flags, ilen, argskip)
85 const DBT *key, *data;
94 PAGE *h, *l, *r, *lchild, *rchild;
97 u_int32_t n, nbytes, nksize;
102 * Split the page into two pages, l and r. The split routines return
103 * a pointer to the page into which the key should be inserted and with
104 * skip set to the offset which should be used. Additionally, l and r
108 h = sp->pgno == P_ROOT ?
109 bt_root(t, sp, &l, &r, &skip, ilen) :
110 bt_page(t, sp, &l, &r, &skip, ilen);
115 * Insert the new key/data pair into the leaf page. (Key inserts
116 * always cause a leaf page to split first.)
118 h->linp[skip] = h->upper -= ilen;
119 dest = (char *)h + h->upper;
120 if (F_ISSET(t, R_RECNO))
121 WR_RLEAF(dest, data, flags)
123 WR_BLEAF(dest, key, data, flags)
125 /* If the root page was split, make it look right. */
126 if (sp->pgno == P_ROOT &&
127 (F_ISSET(t, R_RECNO) ?
128 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
132 * Now we walk the parent page stack -- a LIFO stack of the pages that
133 * were traversed when we searched for the page that split. Each stack
134 * entry is a page number and a page index offset. The offset is for
135 * the page traversed on the search. We've just split a page, so we
136 * have to insert a new key into the parent page.
138 * If the insert into the parent page causes it to split, may have to
139 * continue splitting all the way up the tree. We stop if the root
140 * splits or the page inserted into didn't have to split to hold the
141 * new key. Some algorithms replace the key for the old page as well
142 * as the new page. We don't, as there's no reason to believe that the
143 * first key on the old page is any better than the key we have, and,
144 * in the case of a key being placed at index 0 causing the split, the
145 * key is unavailable.
147 * There are a maximum of 5 pages pinned at any time. We keep the left
148 * and right pages pinned while working on the parent. The 5 are the
149 * two children, left parent and right parent (when the parent splits)
150 * and the root page or the overflow key page when calling bt_preserve.
151 * This code must make sure that all pins are released other than the
152 * root page or overflow page which is unlocked elsewhere.
154 while ((parent = BT_POP(t)) != NULL) {
158 /* Get the parent page. */
159 if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
163 * The new key goes ONE AFTER the index, because the split
166 skip = parent->index + 1;
169 * Calculate the space needed on the parent page.
171 * Prefix trees: space hack when inserting into BINTERNAL
172 * pages. Retain only what's needed to distinguish between
173 * the new entry and the LAST entry on the page to its left.
174 * If the keys compare equal, retain the entire key. Note,
175 * we don't touch overflow keys, and the entire key must be
176 * retained for the next-to-left most key on the leftmost
177 * page of each level, or the search will fail. Applicable
178 * ONLY to internal pages that have leaf pages as children.
179 * Further reduction of the key between pairs of internal
180 * pages loses too much information.
182 switch (rchild->flags & P_TYPE) {
184 bi = GETBINTERNAL(rchild, 0);
185 nbytes = NBINTERNAL(bi->ksize);
188 bl = GETBLEAF(rchild, 0);
189 nbytes = NBINTERNAL(bl->ksize);
190 if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
191 (h->prevpg != P_INVALID || skip > 1)) {
192 tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
197 nksize = t->bt_pfx(&a, &b);
198 n = NBINTERNAL(nksize);
201 bt_pfxsaved += nbytes - n;
217 /* Split the parent page if necessary or shift the indices. */
218 if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
220 h = h->pgno == P_ROOT ?
221 bt_root(t, h, &l, &r, &skip, nbytes) :
222 bt_page(t, h, &l, &r, &skip, nbytes);
227 if (skip < (nxtindex = NEXTINDEX(h)))
228 memmove(h->linp + skip + 1, h->linp + skip,
229 (nxtindex - skip) * sizeof(indx_t));
230 h->lower += sizeof(indx_t);
234 /* Insert the key into the parent page. */
235 switch (rchild->flags & P_TYPE) {
237 h->linp[skip] = h->upper -= nbytes;
238 dest = (char *)h + h->linp[skip];
239 memmove(dest, bi, nbytes);
240 ((BINTERNAL *)dest)->pgno = rchild->pgno;
243 h->linp[skip] = h->upper -= nbytes;
244 dest = (char *)h + h->linp[skip];
245 WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
246 rchild->pgno, bl->flags & P_BIGKEY);
247 memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
248 if (bl->flags & P_BIGKEY &&
249 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
254 * Update the left page count. If split
255 * added at index 0, fix the correct page.
258 dest = (char *)h + h->linp[skip - 1];
260 dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
261 ((RINTERNAL *)dest)->nrecs = rec_total(lchild);
262 ((RINTERNAL *)dest)->pgno = lchild->pgno;
264 /* Update the right page count. */
265 h->linp[skip] = h->upper -= nbytes;
266 dest = (char *)h + h->linp[skip];
267 ((RINTERNAL *)dest)->nrecs = rec_total(rchild);
268 ((RINTERNAL *)dest)->pgno = rchild->pgno;
272 * Update the left page count. If split
273 * added at index 0, fix the correct page.
276 dest = (char *)h + h->linp[skip - 1];
278 dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
279 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
280 ((RINTERNAL *)dest)->pgno = lchild->pgno;
282 /* Update the right page count. */
283 h->linp[skip] = h->upper -= nbytes;
284 dest = (char *)h + h->linp[skip];
285 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
286 ((RINTERNAL *)dest)->pgno = rchild->pgno;
292 /* Unpin the held pages. */
294 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
298 /* If the root page was split, make it look right. */
299 if (sp->pgno == P_ROOT &&
300 (F_ISSET(t, R_RECNO) ?
301 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
304 mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
305 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
308 /* Unpin the held pages. */
309 mpool_put(t->bt_mp, l, MPOOL_DIRTY);
310 mpool_put(t->bt_mp, r, MPOOL_DIRTY);
312 /* Clear any pages left on the stack. */
313 return (RET_SUCCESS);
316 * If something fails in the above loop we were already walking back
317 * up the tree and the tree is now inconsistent. Nothing much we can
318 * do about it but release any memory we're holding.
320 err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
321 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
323 err2: mpool_put(t->bt_mp, l, 0);
324 mpool_put(t->bt_mp, r, 0);
325 __dbpanic(t->bt_dbp);
330 * BT_PAGE -- Split a non-root page of a btree.
335 * lp: pointer to left page pointer
336 * rp: pointer to right page pointer
337 * skip: pointer to index to leave open
338 * ilen: insert length
341 * Pointer to page in which to insert or NULL on error.
344 bt_page(t, h, lp, rp, skip, ilen)
356 /* Put the new right page for the split into place. */
357 if ((r = __bt_new(t, &npg)) == NULL)
360 r->lower = BTDATAOFF;
361 r->upper = t->bt_psize;
362 r->nextpg = h->nextpg;
364 r->flags = h->flags & P_TYPE;
367 * If we're splitting the last page on a level because we're appending
368 * a key to it (skip is NEXTINDEX()), it's likely that the data is
369 * sorted. Adding an empty page on the side of the level is less work
370 * and can push the fill factor much higher than normal. If we're
371 * wrong it's no big deal, we'll just do the split the right way next
372 * time. It may look like it's equally easy to do a similar hack for
373 * reverse sorted data, that is, split the tree left, but it's not.
376 if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
381 r->lower = BTDATAOFF + sizeof(indx_t);
388 /* Put the new left page for the split into place. */
389 if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
390 mpool_put(t->bt_mp, r, 0);
394 memset(l, 0xff, t->bt_psize);
398 l->prevpg = h->prevpg;
399 l->lower = BTDATAOFF;
400 l->upper = t->bt_psize;
401 l->flags = h->flags & P_TYPE;
403 /* Fix up the previous pointer of the page after the split page. */
404 if (h->nextpg != P_INVALID) {
405 if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
407 /* XXX mpool_free(t->bt_mp, r->pgno); */
410 tp->prevpg = r->pgno;
411 mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
415 * Split right. The key/data pairs aren't sorted in the btree page so
416 * it's simpler to copy the data from the split page onto two new pages
417 * instead of copying half the data to the right page and compacting
418 * the left page in place. Since the left page can't change, we have
419 * to swap the original and the allocated left page after the split.
421 tp = bt_psplit(t, h, l, r, skip, ilen);
423 /* Move the new left page onto the old left page. */
424 memmove(h, l, t->bt_psize);
435 * BT_ROOT -- Split the root page of a btree.
440 * lp: pointer to left page pointer
441 * rp: pointer to right page pointer
442 * skip: pointer to index to leave open
443 * ilen: insert length
446 * Pointer to page in which to insert or NULL on error.
449 bt_root(t, h, lp, rp, skip, ilen)
462 /* Put the new left and right pages for the split into place. */
463 if ((l = __bt_new(t, &lnpg)) == NULL ||
464 (r = __bt_new(t, &rnpg)) == NULL)
470 l->prevpg = r->nextpg = P_INVALID;
471 l->lower = r->lower = BTDATAOFF;
472 l->upper = r->upper = t->bt_psize;
473 l->flags = r->flags = h->flags & P_TYPE;
475 /* Split the root page. */
476 tp = bt_psplit(t, h, l, r, skip, ilen);
484 * BT_RROOT -- Fix up the recno root page after it has been split.
493 * RET_ERROR, RET_SUCCESS
502 /* Insert the left and right keys, set the header information. */
503 h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
504 dest = (char *)h + h->upper;
506 l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
508 h->linp[1] = h->upper -= NRINTERNAL;
509 dest = (char *)h + h->upper;
511 r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
513 h->lower = BTDATAOFF + 2 * sizeof(indx_t);
515 /* Unpin the root page, set to recno internal page. */
517 h->flags |= P_RINTERNAL;
518 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
520 return (RET_SUCCESS);
524 * BT_BROOT -- Fix up the btree root page after it has been split.
533 * RET_ERROR, RET_SUCCESS
546 * If the root page was a leaf page, change it into an internal page.
547 * We copy the key we split on (but not the key's data, in the case of
548 * a leaf page) to the new root page.
550 * The btree comparison code guarantees that the left-most key on any
551 * level of the tree is never used, so it doesn't need to be filled in.
553 nbytes = NBINTERNAL(0);
554 h->linp[0] = h->upper = t->bt_psize - nbytes;
555 dest = (char *)h + h->upper;
556 WR_BINTERNAL(dest, 0, l->pgno, 0);
558 switch (h->flags & P_TYPE) {
561 nbytes = NBINTERNAL(bl->ksize);
562 h->linp[1] = h->upper -= nbytes;
563 dest = (char *)h + h->upper;
564 WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
565 memmove(dest, bl->bytes, bl->ksize);
568 * If the key is on an overflow page, mark the overflow chain
569 * so it isn't deleted when the leaf copy of the key is deleted.
571 if (bl->flags & P_BIGKEY &&
572 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
576 bi = GETBINTERNAL(r, 0);
577 nbytes = NBINTERNAL(bi->ksize);
578 h->linp[1] = h->upper -= nbytes;
579 dest = (char *)h + h->upper;
580 memmove(dest, bi, nbytes);
581 ((BINTERNAL *)dest)->pgno = r->pgno;
587 /* There are two keys on the page. */
588 h->lower = BTDATAOFF + 2 * sizeof(indx_t);
590 /* Unpin the root page, set to btree internal page. */
592 h->flags |= P_BINTERNAL;
593 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
595 return (RET_SUCCESS);
599 * BT_PSPLIT -- Do the real work of splitting the page.
603 * h: page to be split
604 * l: page to put lower half of data
605 * r: page to put upper half of data
606 * pskip: pointer to index to leave open
607 * ilen: insert length
610 * Pointer to page in which to insert.
613 bt_psplit(t, h, l, r, pskip, ilen)
625 indx_t full, half, nxt, off, skip, top, used;
627 int bigkeycnt, isbigkey;
630 * Split the data to the left and right pages. Leave the skip index
631 * open. Additionally, make some effort not to split on an overflow
632 * key. This makes internal page processing faster and can save
633 * space as overflow keys used by internal pages are never deleted.
637 full = t->bt_psize - BTDATAOFF;
640 for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
643 isbigkey = 0; /* XXX: not really known. */
645 switch (h->flags & P_TYPE) {
647 src = bi = GETBINTERNAL(h, nxt);
648 nbytes = NBINTERNAL(bi->ksize);
649 isbigkey = bi->flags & P_BIGKEY;
652 src = bl = GETBLEAF(h, nxt);
654 isbigkey = bl->flags & P_BIGKEY;
657 src = GETRINTERNAL(h, nxt);
662 src = rl = GETRLEAF(h, nxt);
671 * If the key/data pairs are substantial fractions of the max
672 * possible size for the page, it's possible to get situations
673 * where we decide to try and copy too much onto the left page.
674 * Make sure that doesn't happen.
677 used + nbytes + sizeof(indx_t) >= full || nxt == top - 1) {
682 /* Copy the key/data pair, if not the skipped index. */
686 l->linp[off] = l->upper -= nbytes;
687 memmove((char *)l + l->upper, src, nbytes);
690 used += nbytes + sizeof(indx_t);
692 if (!isbigkey || bigkeycnt == 3)
700 * Off is the last offset that's valid for the left page.
701 * Nxt is the first offset to be placed on the right page.
703 l->lower += (off + 1) * sizeof(indx_t);
706 * If splitting the page that the cursor was on, the cursor has to be
707 * adjusted to point to the same record as before the split. If the
708 * cursor is at or past the skipped slot, the cursor is incremented by
709 * one. If the cursor is on the right page, it is decremented by the
710 * number of records split to the left page.
713 if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
714 if (c->pg.index >= skip)
716 if (c->pg.index < nxt) /* Left page. */
717 c->pg.pgno = l->pgno;
718 else { /* Right page. */
719 c->pg.pgno = r->pgno;
725 * If the skipped index was on the left page, just return that page.
726 * Otherwise, adjust the skip index to reflect the new position on
737 for (off = 0; nxt < top; ++off) {
742 switch (h->flags & P_TYPE) {
744 src = bi = GETBINTERNAL(h, nxt);
745 nbytes = NBINTERNAL(bi->ksize);
748 src = bl = GETBLEAF(h, nxt);
752 src = GETRINTERNAL(h, nxt);
756 src = rl = GETRLEAF(h, nxt);
763 r->linp[off] = r->upper -= nbytes;
764 memmove((char *)r + r->upper, src, nbytes);
766 r->lower += off * sizeof(indx_t);
768 /* If the key is being appended to the page, adjust the index. */
770 r->lower += sizeof(indx_t);
776 * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
778 * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
779 * record that references them gets deleted. Chains pointed to by internal
780 * pages never get deleted. This routine marks a chain as pointed to by an
785 * pg: page number of first page in the chain.
788 * RET_SUCCESS, RET_ERROR.
797 if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
799 h->flags |= P_PRESERVE;
800 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
801 return (RET_SUCCESS);
805 * REC_TOTAL -- Return the number of recno entries below a page.
811 * The number of recno entries below a page.
814 * These values could be set by the bt_psplit routine. The problem is that the
815 * entry has to be popped off of the stack etc. or the values have to be passed
816 * all the way back to bt_split/bt_rroot and it's not very clean.
825 for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
826 recs += GETRINTERNAL(h, nxt)->nrecs;