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. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 #if defined(LIBC_SCCS) && !defined(lint)
34 static char sccsid[] = "@(#)bt_split.c 8.10 (Berkeley) 1/9/95";
35 #endif /* LIBC_SCCS and not lint */
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
39 #include <sys/param.h>
49 static int bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
50 static PAGE *bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
51 static int bt_preserve(BTREE *, pgno_t);
52 static PAGE *bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
53 static PAGE *bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
54 static int bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
55 static recno_t rec_total(PAGE *);
58 u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
62 * __BT_SPLIT -- Split the tree.
68 * data: data to insert
69 * flags: BIGKEY/BIGDATA flags
71 * skip: index to leave open
74 * RET_ERROR, RET_SUCCESS
77 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
78 size_t ilen, u_int32_t argskip)
84 PAGE *h, *l, *r, *lchild, *rchild;
87 u_int32_t n, nbytes, nksize;
92 * Split the page into two pages, l and r. The split routines return
93 * a pointer to the page into which the key should be inserted and with
94 * skip set to the offset which should be used. Additionally, l and r
98 h = sp->pgno == P_ROOT ?
99 bt_root(t, sp, &l, &r, &skip, ilen) :
100 bt_page(t, sp, &l, &r, &skip, ilen);
105 * Insert the new key/data pair into the leaf page. (Key inserts
106 * always cause a leaf page to split first.)
108 h->linp[skip] = h->upper -= ilen;
109 dest = (char *)h + h->upper;
110 if (F_ISSET(t, R_RECNO))
111 WR_RLEAF(dest, data, flags)
113 WR_BLEAF(dest, key, data, flags)
115 /* If the root page was split, make it look right. */
116 if (sp->pgno == P_ROOT &&
117 (F_ISSET(t, R_RECNO) ?
118 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
122 * Now we walk the parent page stack -- a LIFO stack of the pages that
123 * were traversed when we searched for the page that split. Each stack
124 * entry is a page number and a page index offset. The offset is for
125 * the page traversed on the search. We've just split a page, so we
126 * have to insert a new key into the parent page.
128 * If the insert into the parent page causes it to split, may have to
129 * continue splitting all the way up the tree. We stop if the root
130 * splits or the page inserted into didn't have to split to hold the
131 * new key. Some algorithms replace the key for the old page as well
132 * as the new page. We don't, as there's no reason to believe that the
133 * first key on the old page is any better than the key we have, and,
134 * in the case of a key being placed at index 0 causing the split, the
135 * key is unavailable.
137 * There are a maximum of 5 pages pinned at any time. We keep the left
138 * and right pages pinned while working on the parent. The 5 are the
139 * two children, left parent and right parent (when the parent splits)
140 * and the root page or the overflow key page when calling bt_preserve.
141 * This code must make sure that all pins are released other than the
142 * root page or overflow page which is unlocked elsewhere.
144 while ((parent = BT_POP(t)) != NULL) {
148 /* Get the parent page. */
149 if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
153 * The new key goes ONE AFTER the index, because the split
156 skip = parent->index + 1;
159 * Calculate the space needed on the parent page.
161 * Prefix trees: space hack when inserting into BINTERNAL
162 * pages. Retain only what's needed to distinguish between
163 * the new entry and the LAST entry on the page to its left.
164 * If the keys compare equal, retain the entire key. Note,
165 * we don't touch overflow keys, and the entire key must be
166 * retained for the next-to-left most key on the leftmost
167 * page of each level, or the search will fail. Applicable
168 * ONLY to internal pages that have leaf pages as children.
169 * Further reduction of the key between pairs of internal
170 * pages loses too much information.
172 switch (rchild->flags & P_TYPE) {
174 bi = GETBINTERNAL(rchild, 0);
175 nbytes = NBINTERNAL(bi->ksize);
178 bl = GETBLEAF(rchild, 0);
179 nbytes = NBINTERNAL(bl->ksize);
180 if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
181 (h->prevpg != P_INVALID || skip > 1)) {
182 tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
187 nksize = t->bt_pfx(&a, &b);
188 n = NBINTERNAL(nksize);
191 bt_pfxsaved += nbytes - n;
207 /* Split the parent page if necessary or shift the indices. */
208 if ((u_int32_t)(h->upper - h->lower) < nbytes + sizeof(indx_t)) {
210 h = h->pgno == P_ROOT ?
211 bt_root(t, h, &l, &r, &skip, nbytes) :
212 bt_page(t, h, &l, &r, &skip, nbytes);
217 if (skip < (nxtindex = NEXTINDEX(h)))
218 memmove(h->linp + skip + 1, h->linp + skip,
219 (nxtindex - skip) * sizeof(indx_t));
220 h->lower += sizeof(indx_t);
224 /* Insert the key into the parent page. */
225 switch (rchild->flags & P_TYPE) {
227 h->linp[skip] = h->upper -= nbytes;
228 dest = (char *)h + h->linp[skip];
229 memmove(dest, bi, nbytes);
230 ((BINTERNAL *)dest)->pgno = rchild->pgno;
233 h->linp[skip] = h->upper -= nbytes;
234 dest = (char *)h + h->linp[skip];
235 WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
236 rchild->pgno, bl->flags & P_BIGKEY);
237 memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
238 if (bl->flags & P_BIGKEY) {
240 memcpy(&pgno, bl->bytes, sizeof(pgno));
241 if (bt_preserve(t, pgno) == RET_ERROR)
247 * Update the left page count. If split
248 * added at index 0, fix the correct page.
251 dest = (char *)h + h->linp[skip - 1];
253 dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
254 ((RINTERNAL *)dest)->nrecs = rec_total(lchild);
255 ((RINTERNAL *)dest)->pgno = lchild->pgno;
257 /* Update the right page count. */
258 h->linp[skip] = h->upper -= nbytes;
259 dest = (char *)h + h->linp[skip];
260 ((RINTERNAL *)dest)->nrecs = rec_total(rchild);
261 ((RINTERNAL *)dest)->pgno = rchild->pgno;
265 * Update the left page count. If split
266 * added at index 0, fix the correct page.
269 dest = (char *)h + h->linp[skip - 1];
271 dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
272 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
273 ((RINTERNAL *)dest)->pgno = lchild->pgno;
275 /* Update the right page count. */
276 h->linp[skip] = h->upper -= nbytes;
277 dest = (char *)h + h->linp[skip];
278 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
279 ((RINTERNAL *)dest)->pgno = rchild->pgno;
285 /* Unpin the held pages. */
287 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
291 /* If the root page was split, make it look right. */
292 if (sp->pgno == P_ROOT &&
293 (F_ISSET(t, R_RECNO) ?
294 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
297 mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
298 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
301 /* Unpin the held pages. */
302 mpool_put(t->bt_mp, l, MPOOL_DIRTY);
303 mpool_put(t->bt_mp, r, MPOOL_DIRTY);
305 /* Clear any pages left on the stack. */
306 return (RET_SUCCESS);
309 * If something fails in the above loop we were already walking back
310 * up the tree and the tree is now inconsistent. Nothing much we can
311 * do about it but release any memory we're holding.
313 err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
314 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
316 err2: mpool_put(t->bt_mp, l, 0);
317 mpool_put(t->bt_mp, r, 0);
318 __dbpanic(t->bt_dbp);
323 * BT_PAGE -- Split a non-root page of a btree.
328 * lp: pointer to left page pointer
329 * rp: pointer to right page pointer
330 * skip: pointer to index to leave open
331 * ilen: insert length
334 * Pointer to page in which to insert or NULL on error.
337 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
345 /* Put the new right page for the split into place. */
346 if ((r = __bt_new(t, &npg)) == NULL)
349 r->lower = BTDATAOFF;
350 r->upper = t->bt_psize;
351 r->nextpg = h->nextpg;
353 r->flags = h->flags & P_TYPE;
356 * If we're splitting the last page on a level because we're appending
357 * a key to it (skip is NEXTINDEX()), it's likely that the data is
358 * sorted. Adding an empty page on the side of the level is less work
359 * and can push the fill factor much higher than normal. If we're
360 * wrong it's no big deal, we'll just do the split the right way next
361 * time. It may look like it's equally easy to do a similar hack for
362 * reverse sorted data, that is, split the tree left, but it's not.
365 if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
370 r->lower = BTDATAOFF + sizeof(indx_t);
377 /* Put the new left page for the split into place. */
378 if ((l = (PAGE *)calloc(1, t->bt_psize)) == NULL) {
379 mpool_put(t->bt_mp, r, 0);
384 l->prevpg = h->prevpg;
385 l->lower = BTDATAOFF;
386 l->upper = t->bt_psize;
387 l->flags = h->flags & P_TYPE;
389 /* Fix up the previous pointer of the page after the split page. */
390 if (h->nextpg != P_INVALID) {
391 if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
393 /* XXX mpool_free(t->bt_mp, r->pgno); */
396 tp->prevpg = r->pgno;
397 mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
401 * Split right. The key/data pairs aren't sorted in the btree page so
402 * it's simpler to copy the data from the split page onto two new pages
403 * instead of copying half the data to the right page and compacting
404 * the left page in place. Since the left page can't change, we have
405 * to swap the original and the allocated left page after the split.
407 tp = bt_psplit(t, h, l, r, skip, ilen);
409 /* Move the new left page onto the old left page. */
410 memmove(h, l, t->bt_psize);
421 * BT_ROOT -- Split the root page of a btree.
426 * lp: pointer to left page pointer
427 * rp: pointer to right page pointer
428 * skip: pointer to index to leave open
429 * ilen: insert length
432 * Pointer to page in which to insert or NULL on error.
435 bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
444 /* Put the new left and right pages for the split into place. */
445 if ((l = __bt_new(t, &lnpg)) == NULL ||
446 (r = __bt_new(t, &rnpg)) == NULL)
452 l->prevpg = r->nextpg = P_INVALID;
453 l->lower = r->lower = BTDATAOFF;
454 l->upper = r->upper = t->bt_psize;
455 l->flags = r->flags = h->flags & P_TYPE;
457 /* Split the root page. */
458 tp = bt_psplit(t, h, l, r, skip, ilen);
466 * BT_RROOT -- Fix up the recno root page after it has been split.
475 * RET_ERROR, RET_SUCCESS
478 bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
482 /* Insert the left and right keys, set the header information. */
483 h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
484 dest = (char *)h + h->upper;
486 l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
488 __PAST_END(h->linp, 1) = h->upper -= NRINTERNAL;
489 dest = (char *)h + h->upper;
491 r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
493 h->lower = BTDATAOFF + 2 * sizeof(indx_t);
495 /* Unpin the root page, set to recno internal page. */
497 h->flags |= P_RINTERNAL;
498 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
500 return (RET_SUCCESS);
504 * BT_BROOT -- Fix up the btree root page after it has been split.
513 * RET_ERROR, RET_SUCCESS
516 bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
524 * If the root page was a leaf page, change it into an internal page.
525 * We copy the key we split on (but not the key's data, in the case of
526 * a leaf page) to the new root page.
528 * The btree comparison code guarantees that the left-most key on any
529 * level of the tree is never used, so it doesn't need to be filled in.
531 nbytes = NBINTERNAL(0);
532 h->linp[0] = h->upper = t->bt_psize - nbytes;
533 dest = (char *)h + h->upper;
534 WR_BINTERNAL(dest, 0, l->pgno, 0);
536 switch (h->flags & P_TYPE) {
539 nbytes = NBINTERNAL(bl->ksize);
540 __PAST_END(h->linp, 1) = h->upper -= nbytes;
541 dest = (char *)h + h->upper;
542 WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
543 memmove(dest, bl->bytes, bl->ksize);
546 * If the key is on an overflow page, mark the overflow chain
547 * so it isn't deleted when the leaf copy of the key is deleted.
549 if (bl->flags & P_BIGKEY) {
551 memcpy(&pgno, bl->bytes, sizeof(pgno));
552 if (bt_preserve(t, pgno) == RET_ERROR)
557 bi = GETBINTERNAL(r, 0);
558 nbytes = NBINTERNAL(bi->ksize);
559 __PAST_END(h->linp, 1) = h->upper -= nbytes;
560 dest = (char *)h + h->upper;
561 memmove(dest, bi, nbytes);
562 ((BINTERNAL *)dest)->pgno = r->pgno;
568 /* There are two keys on the page. */
569 h->lower = BTDATAOFF + 2 * sizeof(indx_t);
571 /* Unpin the root page, set to btree internal page. */
573 h->flags |= P_BINTERNAL;
574 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
576 return (RET_SUCCESS);
580 * BT_PSPLIT -- Do the real work of splitting the page.
584 * h: page to be split
585 * l: page to put lower half of data
586 * r: page to put upper half of data
587 * pskip: pointer to index to leave open
588 * ilen: insert length
591 * Pointer to page in which to insert.
594 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
602 indx_t full, half, nxt, off, skip, top, used;
604 int bigkeycnt, isbigkey;
607 * Split the data to the left and right pages. Leave the skip index
608 * open. Additionally, make some effort not to split on an overflow
609 * key. This makes internal page processing faster and can save
610 * space as overflow keys used by internal pages are never deleted.
614 full = t->bt_psize - BTDATAOFF;
617 for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
620 isbigkey = 0; /* XXX: not really known. */
622 switch (h->flags & P_TYPE) {
624 src = bi = GETBINTERNAL(h, nxt);
625 nbytes = NBINTERNAL(bi->ksize);
626 isbigkey = bi->flags & P_BIGKEY;
629 src = bl = GETBLEAF(h, nxt);
631 isbigkey = bl->flags & P_BIGKEY;
634 src = GETRINTERNAL(h, nxt);
639 src = rl = GETRLEAF(h, nxt);
648 * If the key/data pairs are substantial fractions of the max
649 * possible size for the page, it's possible to get situations
650 * where we decide to try and copy too much onto the left page.
651 * Make sure that doesn't happen.
653 if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
659 /* Copy the key/data pair, if not the skipped index. */
663 l->linp[off] = l->upper -= nbytes;
664 memmove((char *)l + l->upper, src, nbytes);
667 used += nbytes + sizeof(indx_t);
669 if (!isbigkey || bigkeycnt == 3)
677 * Off is the last offset that's valid for the left page.
678 * Nxt is the first offset to be placed on the right page.
680 l->lower += (off + 1) * sizeof(indx_t);
683 * If splitting the page that the cursor was on, the cursor has to be
684 * adjusted to point to the same record as before the split. If the
685 * cursor is at or past the skipped slot, the cursor is incremented by
686 * one. If the cursor is on the right page, it is decremented by the
687 * number of records split to the left page.
690 if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
691 if (c->pg.index >= skip)
693 if (c->pg.index < nxt) /* Left page. */
694 c->pg.pgno = l->pgno;
695 else { /* Right page. */
696 c->pg.pgno = r->pgno;
702 * If the skipped index was on the left page, just return that page.
703 * Otherwise, adjust the skip index to reflect the new position on
707 skip = MAX_PAGE_OFFSET;
714 for (off = 0; nxt < top; ++off) {
717 skip = MAX_PAGE_OFFSET;
719 switch (h->flags & P_TYPE) {
721 src = bi = GETBINTERNAL(h, nxt);
722 nbytes = NBINTERNAL(bi->ksize);
725 src = bl = GETBLEAF(h, nxt);
729 src = GETRINTERNAL(h, nxt);
733 src = rl = GETRLEAF(h, nxt);
740 r->linp[off] = r->upper -= nbytes;
741 memmove((char *)r + r->upper, src, nbytes);
743 r->lower += off * sizeof(indx_t);
745 /* If the key is being appended to the page, adjust the index. */
747 r->lower += sizeof(indx_t);
753 * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
755 * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
756 * record that references them gets deleted. Chains pointed to by internal
757 * pages never get deleted. This routine marks a chain as pointed to by an
762 * pg: page number of first page in the chain.
765 * RET_SUCCESS, RET_ERROR.
768 bt_preserve(BTREE *t, pgno_t pg)
772 if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
774 h->flags |= P_PRESERVE;
775 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
776 return (RET_SUCCESS);
780 * REC_TOTAL -- Return the number of recno entries below a page.
786 * The number of recno entries below a page.
789 * These values could be set by the bt_psplit routine. The problem is that the
790 * entry has to be popped off of the stack etc. or the values have to be passed
791 * all the way back to bt_split/bt_rroot and it's not very clean.
799 for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
800 recs += GETRINTERNAL(h, nxt)->nrecs;