bt_split.c revision 1.7 1 1.7 christos /* $NetBSD: bt_split.c,v 1.7 1997/07/13 18:51:59 christos Exp $ */
2 1.5 cgd
3 1.1 cgd /*-
4 1.4 cgd * Copyright (c) 1990, 1993, 1994
5 1.1 cgd * The Regents of the University of California. All rights reserved.
6 1.1 cgd *
7 1.1 cgd * This code is derived from software contributed to Berkeley by
8 1.1 cgd * Mike Olson.
9 1.1 cgd *
10 1.1 cgd * Redistribution and use in source and binary forms, with or without
11 1.1 cgd * modification, are permitted provided that the following conditions
12 1.1 cgd * are met:
13 1.1 cgd * 1. Redistributions of source code must retain the above copyright
14 1.1 cgd * notice, this list of conditions and the following disclaimer.
15 1.1 cgd * 2. Redistributions in binary form must reproduce the above copyright
16 1.1 cgd * notice, this list of conditions and the following disclaimer in the
17 1.1 cgd * documentation and/or other materials provided with the distribution.
18 1.1 cgd * 3. All advertising materials mentioning features or use of this software
19 1.1 cgd * must display the following acknowledgement:
20 1.1 cgd * This product includes software developed by the University of
21 1.1 cgd * California, Berkeley and its contributors.
22 1.1 cgd * 4. Neither the name of the University nor the names of its contributors
23 1.1 cgd * may be used to endorse or promote products derived from this software
24 1.1 cgd * without specific prior written permission.
25 1.1 cgd *
26 1.1 cgd * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 1.1 cgd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 1.1 cgd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 1.1 cgd * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 1.1 cgd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 1.1 cgd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 1.1 cgd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 1.1 cgd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 1.1 cgd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 1.1 cgd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 1.1 cgd * SUCH DAMAGE.
37 1.1 cgd */
38 1.1 cgd
39 1.7 christos #include <sys/cdefs.h>
40 1.1 cgd #if defined(LIBC_SCCS) && !defined(lint)
41 1.5 cgd #if 0
42 1.6 cgd static char sccsid[] = "@(#)bt_split.c 8.9 (Berkeley) 7/26/94";
43 1.5 cgd #else
44 1.7 christos __RCSID("$NetBSD: bt_split.c,v 1.7 1997/07/13 18:51:59 christos Exp $");
45 1.5 cgd #endif
46 1.1 cgd #endif /* LIBC_SCCS and not lint */
47 1.1 cgd
48 1.1 cgd #include <sys/types.h>
49 1.1 cgd
50 1.1 cgd #include <limits.h>
51 1.1 cgd #include <stdio.h>
52 1.1 cgd #include <stdlib.h>
53 1.1 cgd #include <string.h>
54 1.1 cgd
55 1.1 cgd #include <db.h>
56 1.1 cgd #include "btree.h"
57 1.1 cgd
58 1.1 cgd static int bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
59 1.1 cgd static PAGE *bt_page
60 1.4 cgd __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
61 1.1 cgd static int bt_preserve __P((BTREE *, pgno_t));
62 1.1 cgd static PAGE *bt_psplit
63 1.4 cgd __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
64 1.1 cgd static PAGE *bt_root
65 1.4 cgd __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
66 1.1 cgd static int bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
67 1.1 cgd static recno_t rec_total __P((PAGE *));
68 1.1 cgd
69 1.1 cgd #ifdef STATISTICS
70 1.1 cgd u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
71 1.1 cgd #endif
72 1.1 cgd
73 1.1 cgd /*
74 1.1 cgd * __BT_SPLIT -- Split the tree.
75 1.1 cgd *
76 1.1 cgd * Parameters:
77 1.1 cgd * t: tree
78 1.1 cgd * sp: page to split
79 1.1 cgd * key: key to insert
80 1.1 cgd * data: data to insert
81 1.1 cgd * flags: BIGKEY/BIGDATA flags
82 1.1 cgd * ilen: insert length
83 1.1 cgd * skip: index to leave open
84 1.1 cgd *
85 1.1 cgd * Returns:
86 1.1 cgd * RET_ERROR, RET_SUCCESS
87 1.1 cgd */
88 1.1 cgd int
89 1.4 cgd __bt_split(t, sp, key, data, flags, ilen, argskip)
90 1.1 cgd BTREE *t;
91 1.1 cgd PAGE *sp;
92 1.1 cgd const DBT *key, *data;
93 1.4 cgd int flags;
94 1.1 cgd size_t ilen;
95 1.4 cgd u_int32_t argskip;
96 1.1 cgd {
97 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
98 1.7 christos BLEAF *bl = NULL, *tbl; /* pacify gcc */
99 1.1 cgd DBT a, b;
100 1.1 cgd EPGNO *parent;
101 1.1 cgd PAGE *h, *l, *r, *lchild, *rchild;
102 1.1 cgd indx_t nxtindex;
103 1.4 cgd u_int16_t skip;
104 1.7 christos u_int32_t n, nbytes, nksize = 0; /* pacify gcc */
105 1.1 cgd int parentsplit;
106 1.1 cgd char *dest;
107 1.1 cgd
108 1.1 cgd /*
109 1.1 cgd * Split the page into two pages, l and r. The split routines return
110 1.1 cgd * a pointer to the page into which the key should be inserted and with
111 1.1 cgd * skip set to the offset which should be used. Additionally, l and r
112 1.1 cgd * are pinned.
113 1.1 cgd */
114 1.4 cgd skip = argskip;
115 1.1 cgd h = sp->pgno == P_ROOT ?
116 1.1 cgd bt_root(t, sp, &l, &r, &skip, ilen) :
117 1.1 cgd bt_page(t, sp, &l, &r, &skip, ilen);
118 1.1 cgd if (h == NULL)
119 1.1 cgd return (RET_ERROR);
120 1.1 cgd
121 1.1 cgd /*
122 1.1 cgd * Insert the new key/data pair into the leaf page. (Key inserts
123 1.1 cgd * always cause a leaf page to split first.)
124 1.1 cgd */
125 1.1 cgd h->linp[skip] = h->upper -= ilen;
126 1.1 cgd dest = (char *)h + h->upper;
127 1.6 cgd if (F_ISSET(t, R_RECNO))
128 1.1 cgd WR_RLEAF(dest, data, flags)
129 1.1 cgd else
130 1.1 cgd WR_BLEAF(dest, key, data, flags)
131 1.1 cgd
132 1.1 cgd /* If the root page was split, make it look right. */
133 1.1 cgd if (sp->pgno == P_ROOT &&
134 1.6 cgd (F_ISSET(t, R_RECNO) ?
135 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
136 1.1 cgd goto err2;
137 1.1 cgd
138 1.1 cgd /*
139 1.1 cgd * Now we walk the parent page stack -- a LIFO stack of the pages that
140 1.1 cgd * were traversed when we searched for the page that split. Each stack
141 1.1 cgd * entry is a page number and a page index offset. The offset is for
142 1.1 cgd * the page traversed on the search. We've just split a page, so we
143 1.1 cgd * have to insert a new key into the parent page.
144 1.1 cgd *
145 1.1 cgd * If the insert into the parent page causes it to split, may have to
146 1.1 cgd * continue splitting all the way up the tree. We stop if the root
147 1.1 cgd * splits or the page inserted into didn't have to split to hold the
148 1.1 cgd * new key. Some algorithms replace the key for the old page as well
149 1.1 cgd * as the new page. We don't, as there's no reason to believe that the
150 1.1 cgd * first key on the old page is any better than the key we have, and,
151 1.1 cgd * in the case of a key being placed at index 0 causing the split, the
152 1.1 cgd * key is unavailable.
153 1.1 cgd *
154 1.1 cgd * There are a maximum of 5 pages pinned at any time. We keep the left
155 1.1 cgd * and right pages pinned while working on the parent. The 5 are the
156 1.1 cgd * two children, left parent and right parent (when the parent splits)
157 1.1 cgd * and the root page or the overflow key page when calling bt_preserve.
158 1.1 cgd * This code must make sure that all pins are released other than the
159 1.1 cgd * root page or overflow page which is unlocked elsewhere.
160 1.1 cgd */
161 1.1 cgd while ((parent = BT_POP(t)) != NULL) {
162 1.1 cgd lchild = l;
163 1.1 cgd rchild = r;
164 1.1 cgd
165 1.1 cgd /* Get the parent page. */
166 1.1 cgd if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
167 1.1 cgd goto err2;
168 1.1 cgd
169 1.1 cgd /*
170 1.1 cgd * The new key goes ONE AFTER the index, because the split
171 1.1 cgd * was to the right.
172 1.1 cgd */
173 1.1 cgd skip = parent->index + 1;
174 1.1 cgd
175 1.1 cgd /*
176 1.1 cgd * Calculate the space needed on the parent page.
177 1.1 cgd *
178 1.1 cgd * Prefix trees: space hack when inserting into BINTERNAL
179 1.1 cgd * pages. Retain only what's needed to distinguish between
180 1.1 cgd * the new entry and the LAST entry on the page to its left.
181 1.1 cgd * If the keys compare equal, retain the entire key. Note,
182 1.1 cgd * we don't touch overflow keys, and the entire key must be
183 1.1 cgd * retained for the next-to-left most key on the leftmost
184 1.1 cgd * page of each level, or the search will fail. Applicable
185 1.1 cgd * ONLY to internal pages that have leaf pages as children.
186 1.1 cgd * Further reduction of the key between pairs of internal
187 1.1 cgd * pages loses too much information.
188 1.1 cgd */
189 1.1 cgd switch (rchild->flags & P_TYPE) {
190 1.1 cgd case P_BINTERNAL:
191 1.1 cgd bi = GETBINTERNAL(rchild, 0);
192 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
193 1.1 cgd break;
194 1.1 cgd case P_BLEAF:
195 1.1 cgd bl = GETBLEAF(rchild, 0);
196 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
197 1.1 cgd if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
198 1.1 cgd (h->prevpg != P_INVALID || skip > 1)) {
199 1.1 cgd tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
200 1.1 cgd a.size = tbl->ksize;
201 1.1 cgd a.data = tbl->bytes;
202 1.1 cgd b.size = bl->ksize;
203 1.1 cgd b.data = bl->bytes;
204 1.1 cgd nksize = t->bt_pfx(&a, &b);
205 1.1 cgd n = NBINTERNAL(nksize);
206 1.1 cgd if (n < nbytes) {
207 1.1 cgd #ifdef STATISTICS
208 1.1 cgd bt_pfxsaved += nbytes - n;
209 1.1 cgd #endif
210 1.1 cgd nbytes = n;
211 1.1 cgd } else
212 1.1 cgd nksize = 0;
213 1.1 cgd } else
214 1.1 cgd nksize = 0;
215 1.1 cgd break;
216 1.1 cgd case P_RINTERNAL:
217 1.1 cgd case P_RLEAF:
218 1.1 cgd nbytes = NRINTERNAL;
219 1.1 cgd break;
220 1.1 cgd default:
221 1.1 cgd abort();
222 1.1 cgd }
223 1.1 cgd
224 1.1 cgd /* Split the parent page if necessary or shift the indices. */
225 1.1 cgd if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
226 1.1 cgd sp = h;
227 1.1 cgd h = h->pgno == P_ROOT ?
228 1.1 cgd bt_root(t, h, &l, &r, &skip, nbytes) :
229 1.1 cgd bt_page(t, h, &l, &r, &skip, nbytes);
230 1.1 cgd if (h == NULL)
231 1.1 cgd goto err1;
232 1.1 cgd parentsplit = 1;
233 1.1 cgd } else {
234 1.1 cgd if (skip < (nxtindex = NEXTINDEX(h)))
235 1.1 cgd memmove(h->linp + skip + 1, h->linp + skip,
236 1.1 cgd (nxtindex - skip) * sizeof(indx_t));
237 1.1 cgd h->lower += sizeof(indx_t);
238 1.1 cgd parentsplit = 0;
239 1.1 cgd }
240 1.1 cgd
241 1.1 cgd /* Insert the key into the parent page. */
242 1.6 cgd switch (rchild->flags & P_TYPE) {
243 1.1 cgd case P_BINTERNAL:
244 1.1 cgd h->linp[skip] = h->upper -= nbytes;
245 1.1 cgd dest = (char *)h + h->linp[skip];
246 1.1 cgd memmove(dest, bi, nbytes);
247 1.1 cgd ((BINTERNAL *)dest)->pgno = rchild->pgno;
248 1.1 cgd break;
249 1.1 cgd case P_BLEAF:
250 1.1 cgd h->linp[skip] = h->upper -= nbytes;
251 1.1 cgd dest = (char *)h + h->linp[skip];
252 1.1 cgd WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
253 1.1 cgd rchild->pgno, bl->flags & P_BIGKEY);
254 1.1 cgd memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
255 1.1 cgd if (bl->flags & P_BIGKEY &&
256 1.1 cgd bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
257 1.1 cgd goto err1;
258 1.1 cgd break;
259 1.1 cgd case P_RINTERNAL:
260 1.1 cgd /*
261 1.1 cgd * Update the left page count. If split
262 1.1 cgd * added at index 0, fix the correct page.
263 1.1 cgd */
264 1.1 cgd if (skip > 0)
265 1.1 cgd dest = (char *)h + h->linp[skip - 1];
266 1.1 cgd else
267 1.1 cgd dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
268 1.1 cgd ((RINTERNAL *)dest)->nrecs = rec_total(lchild);
269 1.1 cgd ((RINTERNAL *)dest)->pgno = lchild->pgno;
270 1.1 cgd
271 1.1 cgd /* Update the right page count. */
272 1.1 cgd h->linp[skip] = h->upper -= nbytes;
273 1.1 cgd dest = (char *)h + h->linp[skip];
274 1.1 cgd ((RINTERNAL *)dest)->nrecs = rec_total(rchild);
275 1.1 cgd ((RINTERNAL *)dest)->pgno = rchild->pgno;
276 1.1 cgd break;
277 1.1 cgd case P_RLEAF:
278 1.1 cgd /*
279 1.1 cgd * Update the left page count. If split
280 1.1 cgd * added at index 0, fix the correct page.
281 1.1 cgd */
282 1.1 cgd if (skip > 0)
283 1.1 cgd dest = (char *)h + h->linp[skip - 1];
284 1.1 cgd else
285 1.1 cgd dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
286 1.1 cgd ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
287 1.1 cgd ((RINTERNAL *)dest)->pgno = lchild->pgno;
288 1.1 cgd
289 1.1 cgd /* Update the right page count. */
290 1.1 cgd h->linp[skip] = h->upper -= nbytes;
291 1.1 cgd dest = (char *)h + h->linp[skip];
292 1.1 cgd ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
293 1.1 cgd ((RINTERNAL *)dest)->pgno = rchild->pgno;
294 1.1 cgd break;
295 1.1 cgd default:
296 1.1 cgd abort();
297 1.1 cgd }
298 1.1 cgd
299 1.1 cgd /* Unpin the held pages. */
300 1.1 cgd if (!parentsplit) {
301 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
302 1.1 cgd break;
303 1.1 cgd }
304 1.1 cgd
305 1.1 cgd /* If the root page was split, make it look right. */
306 1.1 cgd if (sp->pgno == P_ROOT &&
307 1.6 cgd (F_ISSET(t, R_RECNO) ?
308 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
309 1.1 cgd goto err1;
310 1.1 cgd
311 1.1 cgd mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
312 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
313 1.1 cgd }
314 1.1 cgd
315 1.1 cgd /* Unpin the held pages. */
316 1.1 cgd mpool_put(t->bt_mp, l, MPOOL_DIRTY);
317 1.1 cgd mpool_put(t->bt_mp, r, MPOOL_DIRTY);
318 1.1 cgd
319 1.1 cgd /* Clear any pages left on the stack. */
320 1.1 cgd return (RET_SUCCESS);
321 1.1 cgd
322 1.1 cgd /*
323 1.1 cgd * If something fails in the above loop we were already walking back
324 1.1 cgd * up the tree and the tree is now inconsistent. Nothing much we can
325 1.1 cgd * do about it but release any memory we're holding.
326 1.1 cgd */
327 1.1 cgd err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
328 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
329 1.1 cgd
330 1.1 cgd err2: mpool_put(t->bt_mp, l, 0);
331 1.1 cgd mpool_put(t->bt_mp, r, 0);
332 1.1 cgd __dbpanic(t->bt_dbp);
333 1.1 cgd return (RET_ERROR);
334 1.1 cgd }
335 1.1 cgd
336 1.1 cgd /*
337 1.1 cgd * BT_PAGE -- Split a non-root page of a btree.
338 1.1 cgd *
339 1.1 cgd * Parameters:
340 1.1 cgd * t: tree
341 1.1 cgd * h: root page
342 1.1 cgd * lp: pointer to left page pointer
343 1.1 cgd * rp: pointer to right page pointer
344 1.1 cgd * skip: pointer to index to leave open
345 1.1 cgd * ilen: insert length
346 1.1 cgd *
347 1.1 cgd * Returns:
348 1.1 cgd * Pointer to page in which to insert or NULL on error.
349 1.1 cgd */
350 1.1 cgd static PAGE *
351 1.1 cgd bt_page(t, h, lp, rp, skip, ilen)
352 1.1 cgd BTREE *t;
353 1.1 cgd PAGE *h, **lp, **rp;
354 1.4 cgd indx_t *skip;
355 1.1 cgd size_t ilen;
356 1.1 cgd {
357 1.1 cgd PAGE *l, *r, *tp;
358 1.1 cgd pgno_t npg;
359 1.1 cgd
360 1.1 cgd #ifdef STATISTICS
361 1.1 cgd ++bt_split;
362 1.1 cgd #endif
363 1.1 cgd /* Put the new right page for the split into place. */
364 1.1 cgd if ((r = __bt_new(t, &npg)) == NULL)
365 1.1 cgd return (NULL);
366 1.1 cgd r->pgno = npg;
367 1.1 cgd r->lower = BTDATAOFF;
368 1.1 cgd r->upper = t->bt_psize;
369 1.1 cgd r->nextpg = h->nextpg;
370 1.1 cgd r->prevpg = h->pgno;
371 1.1 cgd r->flags = h->flags & P_TYPE;
372 1.1 cgd
373 1.1 cgd /*
374 1.1 cgd * If we're splitting the last page on a level because we're appending
375 1.1 cgd * a key to it (skip is NEXTINDEX()), it's likely that the data is
376 1.1 cgd * sorted. Adding an empty page on the side of the level is less work
377 1.1 cgd * and can push the fill factor much higher than normal. If we're
378 1.1 cgd * wrong it's no big deal, we'll just do the split the right way next
379 1.1 cgd * time. It may look like it's equally easy to do a similar hack for
380 1.1 cgd * reverse sorted data, that is, split the tree left, but it's not.
381 1.1 cgd * Don't even try.
382 1.1 cgd */
383 1.1 cgd if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
384 1.1 cgd #ifdef STATISTICS
385 1.1 cgd ++bt_sortsplit;
386 1.1 cgd #endif
387 1.1 cgd h->nextpg = r->pgno;
388 1.1 cgd r->lower = BTDATAOFF + sizeof(indx_t);
389 1.1 cgd *skip = 0;
390 1.1 cgd *lp = h;
391 1.1 cgd *rp = r;
392 1.1 cgd return (r);
393 1.1 cgd }
394 1.1 cgd
395 1.1 cgd /* Put the new left page for the split into place. */
396 1.4 cgd if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
397 1.1 cgd mpool_put(t->bt_mp, r, 0);
398 1.1 cgd return (NULL);
399 1.1 cgd }
400 1.6 cgd #ifdef PURIFY
401 1.6 cgd memset(l, 0xff, t->bt_psize);
402 1.6 cgd #endif
403 1.1 cgd l->pgno = h->pgno;
404 1.1 cgd l->nextpg = r->pgno;
405 1.1 cgd l->prevpg = h->prevpg;
406 1.1 cgd l->lower = BTDATAOFF;
407 1.1 cgd l->upper = t->bt_psize;
408 1.1 cgd l->flags = h->flags & P_TYPE;
409 1.1 cgd
410 1.1 cgd /* Fix up the previous pointer of the page after the split page. */
411 1.1 cgd if (h->nextpg != P_INVALID) {
412 1.1 cgd if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
413 1.1 cgd free(l);
414 1.1 cgd /* XXX mpool_free(t->bt_mp, r->pgno); */
415 1.1 cgd return (NULL);
416 1.1 cgd }
417 1.1 cgd tp->prevpg = r->pgno;
418 1.6 cgd mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
419 1.1 cgd }
420 1.1 cgd
421 1.1 cgd /*
422 1.1 cgd * Split right. The key/data pairs aren't sorted in the btree page so
423 1.1 cgd * it's simpler to copy the data from the split page onto two new pages
424 1.1 cgd * instead of copying half the data to the right page and compacting
425 1.1 cgd * the left page in place. Since the left page can't change, we have
426 1.1 cgd * to swap the original and the allocated left page after the split.
427 1.1 cgd */
428 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
429 1.1 cgd
430 1.1 cgd /* Move the new left page onto the old left page. */
431 1.1 cgd memmove(h, l, t->bt_psize);
432 1.1 cgd if (tp == l)
433 1.1 cgd tp = h;
434 1.1 cgd free(l);
435 1.1 cgd
436 1.1 cgd *lp = h;
437 1.1 cgd *rp = r;
438 1.1 cgd return (tp);
439 1.1 cgd }
440 1.1 cgd
441 1.1 cgd /*
442 1.1 cgd * BT_ROOT -- Split the root page of a btree.
443 1.1 cgd *
444 1.1 cgd * Parameters:
445 1.1 cgd * t: tree
446 1.1 cgd * h: root page
447 1.1 cgd * lp: pointer to left page pointer
448 1.1 cgd * rp: pointer to right page pointer
449 1.1 cgd * skip: pointer to index to leave open
450 1.1 cgd * ilen: insert length
451 1.1 cgd *
452 1.1 cgd * Returns:
453 1.1 cgd * Pointer to page in which to insert or NULL on error.
454 1.1 cgd */
455 1.1 cgd static PAGE *
456 1.1 cgd bt_root(t, h, lp, rp, skip, ilen)
457 1.1 cgd BTREE *t;
458 1.1 cgd PAGE *h, **lp, **rp;
459 1.4 cgd indx_t *skip;
460 1.1 cgd size_t ilen;
461 1.1 cgd {
462 1.1 cgd PAGE *l, *r, *tp;
463 1.1 cgd pgno_t lnpg, rnpg;
464 1.1 cgd
465 1.1 cgd #ifdef STATISTICS
466 1.1 cgd ++bt_split;
467 1.1 cgd ++bt_rootsplit;
468 1.1 cgd #endif
469 1.1 cgd /* Put the new left and right pages for the split into place. */
470 1.1 cgd if ((l = __bt_new(t, &lnpg)) == NULL ||
471 1.1 cgd (r = __bt_new(t, &rnpg)) == NULL)
472 1.1 cgd return (NULL);
473 1.1 cgd l->pgno = lnpg;
474 1.1 cgd r->pgno = rnpg;
475 1.1 cgd l->nextpg = r->pgno;
476 1.1 cgd r->prevpg = l->pgno;
477 1.1 cgd l->prevpg = r->nextpg = P_INVALID;
478 1.1 cgd l->lower = r->lower = BTDATAOFF;
479 1.1 cgd l->upper = r->upper = t->bt_psize;
480 1.1 cgd l->flags = r->flags = h->flags & P_TYPE;
481 1.1 cgd
482 1.1 cgd /* Split the root page. */
483 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
484 1.1 cgd
485 1.1 cgd *lp = l;
486 1.1 cgd *rp = r;
487 1.1 cgd return (tp);
488 1.1 cgd }
489 1.1 cgd
490 1.1 cgd /*
491 1.1 cgd * BT_RROOT -- Fix up the recno root page after it has been split.
492 1.1 cgd *
493 1.1 cgd * Parameters:
494 1.1 cgd * t: tree
495 1.1 cgd * h: root page
496 1.1 cgd * l: left page
497 1.1 cgd * r: right page
498 1.1 cgd *
499 1.1 cgd * Returns:
500 1.1 cgd * RET_ERROR, RET_SUCCESS
501 1.1 cgd */
502 1.1 cgd static int
503 1.1 cgd bt_rroot(t, h, l, r)
504 1.1 cgd BTREE *t;
505 1.1 cgd PAGE *h, *l, *r;
506 1.1 cgd {
507 1.1 cgd char *dest;
508 1.1 cgd
509 1.1 cgd /* Insert the left and right keys, set the header information. */
510 1.1 cgd h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
511 1.1 cgd dest = (char *)h + h->upper;
512 1.1 cgd WR_RINTERNAL(dest,
513 1.1 cgd l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
514 1.1 cgd
515 1.1 cgd h->linp[1] = h->upper -= NRINTERNAL;
516 1.1 cgd dest = (char *)h + h->upper;
517 1.1 cgd WR_RINTERNAL(dest,
518 1.1 cgd r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
519 1.1 cgd
520 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
521 1.1 cgd
522 1.1 cgd /* Unpin the root page, set to recno internal page. */
523 1.1 cgd h->flags &= ~P_TYPE;
524 1.1 cgd h->flags |= P_RINTERNAL;
525 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
526 1.1 cgd
527 1.1 cgd return (RET_SUCCESS);
528 1.1 cgd }
529 1.1 cgd
530 1.1 cgd /*
531 1.1 cgd * BT_BROOT -- Fix up the btree root page after it has been split.
532 1.1 cgd *
533 1.1 cgd * Parameters:
534 1.1 cgd * t: tree
535 1.1 cgd * h: root page
536 1.1 cgd * l: left page
537 1.1 cgd * r: right page
538 1.1 cgd *
539 1.1 cgd * Returns:
540 1.1 cgd * RET_ERROR, RET_SUCCESS
541 1.1 cgd */
542 1.1 cgd static int
543 1.1 cgd bt_broot(t, h, l, r)
544 1.1 cgd BTREE *t;
545 1.1 cgd PAGE *h, *l, *r;
546 1.1 cgd {
547 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
548 1.1 cgd BLEAF *bl;
549 1.4 cgd u_int32_t nbytes;
550 1.1 cgd char *dest;
551 1.1 cgd
552 1.1 cgd /*
553 1.1 cgd * If the root page was a leaf page, change it into an internal page.
554 1.1 cgd * We copy the key we split on (but not the key's data, in the case of
555 1.1 cgd * a leaf page) to the new root page.
556 1.1 cgd *
557 1.1 cgd * The btree comparison code guarantees that the left-most key on any
558 1.1 cgd * level of the tree is never used, so it doesn't need to be filled in.
559 1.1 cgd */
560 1.1 cgd nbytes = NBINTERNAL(0);
561 1.1 cgd h->linp[0] = h->upper = t->bt_psize - nbytes;
562 1.1 cgd dest = (char *)h + h->upper;
563 1.1 cgd WR_BINTERNAL(dest, 0, l->pgno, 0);
564 1.1 cgd
565 1.6 cgd switch (h->flags & P_TYPE) {
566 1.1 cgd case P_BLEAF:
567 1.1 cgd bl = GETBLEAF(r, 0);
568 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
569 1.1 cgd h->linp[1] = h->upper -= nbytes;
570 1.1 cgd dest = (char *)h + h->upper;
571 1.1 cgd WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
572 1.1 cgd memmove(dest, bl->bytes, bl->ksize);
573 1.1 cgd
574 1.1 cgd /*
575 1.1 cgd * If the key is on an overflow page, mark the overflow chain
576 1.1 cgd * so it isn't deleted when the leaf copy of the key is deleted.
577 1.1 cgd */
578 1.1 cgd if (bl->flags & P_BIGKEY &&
579 1.1 cgd bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
580 1.1 cgd return (RET_ERROR);
581 1.1 cgd break;
582 1.1 cgd case P_BINTERNAL:
583 1.1 cgd bi = GETBINTERNAL(r, 0);
584 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
585 1.1 cgd h->linp[1] = h->upper -= nbytes;
586 1.1 cgd dest = (char *)h + h->upper;
587 1.1 cgd memmove(dest, bi, nbytes);
588 1.1 cgd ((BINTERNAL *)dest)->pgno = r->pgno;
589 1.1 cgd break;
590 1.1 cgd default:
591 1.1 cgd abort();
592 1.1 cgd }
593 1.1 cgd
594 1.1 cgd /* There are two keys on the page. */
595 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
596 1.1 cgd
597 1.1 cgd /* Unpin the root page, set to btree internal page. */
598 1.1 cgd h->flags &= ~P_TYPE;
599 1.1 cgd h->flags |= P_BINTERNAL;
600 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
601 1.1 cgd
602 1.1 cgd return (RET_SUCCESS);
603 1.1 cgd }
604 1.1 cgd
605 1.1 cgd /*
606 1.1 cgd * BT_PSPLIT -- Do the real work of splitting the page.
607 1.1 cgd *
608 1.1 cgd * Parameters:
609 1.1 cgd * t: tree
610 1.1 cgd * h: page to be split
611 1.1 cgd * l: page to put lower half of data
612 1.1 cgd * r: page to put upper half of data
613 1.1 cgd * pskip: pointer to index to leave open
614 1.1 cgd * ilen: insert length
615 1.1 cgd *
616 1.1 cgd * Returns:
617 1.1 cgd * Pointer to page in which to insert.
618 1.1 cgd */
619 1.1 cgd static PAGE *
620 1.1 cgd bt_psplit(t, h, l, r, pskip, ilen)
621 1.1 cgd BTREE *t;
622 1.1 cgd PAGE *h, *l, *r;
623 1.4 cgd indx_t *pskip;
624 1.1 cgd size_t ilen;
625 1.1 cgd {
626 1.1 cgd BINTERNAL *bi;
627 1.1 cgd BLEAF *bl;
628 1.6 cgd CURSOR *c;
629 1.1 cgd RLEAF *rl;
630 1.1 cgd PAGE *rval;
631 1.7 christos void *src = NULL; /* pacify gcc */
632 1.1 cgd indx_t full, half, nxt, off, skip, top, used;
633 1.4 cgd u_int32_t nbytes;
634 1.1 cgd int bigkeycnt, isbigkey;
635 1.1 cgd
636 1.1 cgd /*
637 1.1 cgd * Split the data to the left and right pages. Leave the skip index
638 1.1 cgd * open. Additionally, make some effort not to split on an overflow
639 1.1 cgd * key. This makes internal page processing faster and can save
640 1.1 cgd * space as overflow keys used by internal pages are never deleted.
641 1.1 cgd */
642 1.1 cgd bigkeycnt = 0;
643 1.1 cgd skip = *pskip;
644 1.1 cgd full = t->bt_psize - BTDATAOFF;
645 1.1 cgd half = full / 2;
646 1.1 cgd used = 0;
647 1.1 cgd for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
648 1.1 cgd if (skip == off) {
649 1.1 cgd nbytes = ilen;
650 1.1 cgd isbigkey = 0; /* XXX: not really known. */
651 1.1 cgd } else
652 1.1 cgd switch (h->flags & P_TYPE) {
653 1.1 cgd case P_BINTERNAL:
654 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
655 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
656 1.1 cgd isbigkey = bi->flags & P_BIGKEY;
657 1.1 cgd break;
658 1.1 cgd case P_BLEAF:
659 1.1 cgd src = bl = GETBLEAF(h, nxt);
660 1.1 cgd nbytes = NBLEAF(bl);
661 1.1 cgd isbigkey = bl->flags & P_BIGKEY;
662 1.1 cgd break;
663 1.1 cgd case P_RINTERNAL:
664 1.1 cgd src = GETRINTERNAL(h, nxt);
665 1.1 cgd nbytes = NRINTERNAL;
666 1.1 cgd isbigkey = 0;
667 1.1 cgd break;
668 1.1 cgd case P_RLEAF:
669 1.1 cgd src = rl = GETRLEAF(h, nxt);
670 1.1 cgd nbytes = NRLEAF(rl);
671 1.1 cgd isbigkey = 0;
672 1.1 cgd break;
673 1.1 cgd default:
674 1.1 cgd abort();
675 1.1 cgd }
676 1.1 cgd
677 1.1 cgd /*
678 1.1 cgd * If the key/data pairs are substantial fractions of the max
679 1.1 cgd * possible size for the page, it's possible to get situations
680 1.1 cgd * where we decide to try and copy too much onto the left page.
681 1.1 cgd * Make sure that doesn't happen.
682 1.1 cgd */
683 1.1 cgd if (skip <= off && used + nbytes >= full) {
684 1.1 cgd --off;
685 1.1 cgd break;
686 1.1 cgd }
687 1.1 cgd
688 1.1 cgd /* Copy the key/data pair, if not the skipped index. */
689 1.1 cgd if (skip != off) {
690 1.1 cgd ++nxt;
691 1.1 cgd
692 1.1 cgd l->linp[off] = l->upper -= nbytes;
693 1.1 cgd memmove((char *)l + l->upper, src, nbytes);
694 1.1 cgd }
695 1.1 cgd
696 1.1 cgd used += nbytes;
697 1.1 cgd if (used >= half) {
698 1.1 cgd if (!isbigkey || bigkeycnt == 3)
699 1.1 cgd break;
700 1.1 cgd else
701 1.1 cgd ++bigkeycnt;
702 1.1 cgd }
703 1.1 cgd }
704 1.1 cgd
705 1.1 cgd /*
706 1.1 cgd * Off is the last offset that's valid for the left page.
707 1.1 cgd * Nxt is the first offset to be placed on the right page.
708 1.1 cgd */
709 1.1 cgd l->lower += (off + 1) * sizeof(indx_t);
710 1.1 cgd
711 1.1 cgd /*
712 1.1 cgd * If splitting the page that the cursor was on, the cursor has to be
713 1.1 cgd * adjusted to point to the same record as before the split. If the
714 1.1 cgd * cursor is at or past the skipped slot, the cursor is incremented by
715 1.1 cgd * one. If the cursor is on the right page, it is decremented by the
716 1.1 cgd * number of records split to the left page.
717 1.1 cgd */
718 1.6 cgd c = &t->bt_cursor;
719 1.6 cgd if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
720 1.6 cgd if (c->pg.index >= skip)
721 1.6 cgd ++c->pg.index;
722 1.6 cgd if (c->pg.index < nxt) /* Left page. */
723 1.6 cgd c->pg.pgno = l->pgno;
724 1.1 cgd else { /* Right page. */
725 1.6 cgd c->pg.pgno = r->pgno;
726 1.6 cgd c->pg.index -= nxt;
727 1.1 cgd }
728 1.1 cgd }
729 1.1 cgd
730 1.1 cgd /*
731 1.1 cgd * If the skipped index was on the left page, just return that page.
732 1.1 cgd * Otherwise, adjust the skip index to reflect the new position on
733 1.1 cgd * the right page.
734 1.1 cgd */
735 1.1 cgd if (skip <= off) {
736 1.1 cgd skip = 0;
737 1.1 cgd rval = l;
738 1.1 cgd } else {
739 1.1 cgd rval = r;
740 1.1 cgd *pskip -= nxt;
741 1.1 cgd }
742 1.1 cgd
743 1.1 cgd for (off = 0; nxt < top; ++off) {
744 1.1 cgd if (skip == nxt) {
745 1.1 cgd ++off;
746 1.1 cgd skip = 0;
747 1.1 cgd }
748 1.1 cgd switch (h->flags & P_TYPE) {
749 1.1 cgd case P_BINTERNAL:
750 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
751 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
752 1.1 cgd break;
753 1.1 cgd case P_BLEAF:
754 1.1 cgd src = bl = GETBLEAF(h, nxt);
755 1.1 cgd nbytes = NBLEAF(bl);
756 1.1 cgd break;
757 1.1 cgd case P_RINTERNAL:
758 1.1 cgd src = GETRINTERNAL(h, nxt);
759 1.1 cgd nbytes = NRINTERNAL;
760 1.1 cgd break;
761 1.1 cgd case P_RLEAF:
762 1.1 cgd src = rl = GETRLEAF(h, nxt);
763 1.1 cgd nbytes = NRLEAF(rl);
764 1.1 cgd break;
765 1.1 cgd default:
766 1.1 cgd abort();
767 1.1 cgd }
768 1.1 cgd ++nxt;
769 1.1 cgd r->linp[off] = r->upper -= nbytes;
770 1.1 cgd memmove((char *)r + r->upper, src, nbytes);
771 1.1 cgd }
772 1.1 cgd r->lower += off * sizeof(indx_t);
773 1.1 cgd
774 1.1 cgd /* If the key is being appended to the page, adjust the index. */
775 1.1 cgd if (skip == top)
776 1.1 cgd r->lower += sizeof(indx_t);
777 1.1 cgd
778 1.1 cgd return (rval);
779 1.1 cgd }
780 1.1 cgd
781 1.1 cgd /*
782 1.1 cgd * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
783 1.1 cgd *
784 1.1 cgd * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
785 1.1 cgd * record that references them gets deleted. Chains pointed to by internal
786 1.1 cgd * pages never get deleted. This routine marks a chain as pointed to by an
787 1.1 cgd * internal page.
788 1.1 cgd *
789 1.1 cgd * Parameters:
790 1.1 cgd * t: tree
791 1.1 cgd * pg: page number of first page in the chain.
792 1.1 cgd *
793 1.1 cgd * Returns:
794 1.1 cgd * RET_SUCCESS, RET_ERROR.
795 1.1 cgd */
796 1.1 cgd static int
797 1.1 cgd bt_preserve(t, pg)
798 1.1 cgd BTREE *t;
799 1.1 cgd pgno_t pg;
800 1.1 cgd {
801 1.1 cgd PAGE *h;
802 1.1 cgd
803 1.1 cgd if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
804 1.1 cgd return (RET_ERROR);
805 1.1 cgd h->flags |= P_PRESERVE;
806 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
807 1.1 cgd return (RET_SUCCESS);
808 1.1 cgd }
809 1.1 cgd
810 1.1 cgd /*
811 1.1 cgd * REC_TOTAL -- Return the number of recno entries below a page.
812 1.1 cgd *
813 1.1 cgd * Parameters:
814 1.1 cgd * h: page
815 1.1 cgd *
816 1.1 cgd * Returns:
817 1.1 cgd * The number of recno entries below a page.
818 1.1 cgd *
819 1.1 cgd * XXX
820 1.1 cgd * These values could be set by the bt_psplit routine. The problem is that the
821 1.1 cgd * entry has to be popped off of the stack etc. or the values have to be passed
822 1.1 cgd * all the way back to bt_split/bt_rroot and it's not very clean.
823 1.1 cgd */
824 1.1 cgd static recno_t
825 1.1 cgd rec_total(h)
826 1.1 cgd PAGE *h;
827 1.1 cgd {
828 1.1 cgd recno_t recs;
829 1.1 cgd indx_t nxt, top;
830 1.1 cgd
831 1.1 cgd for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
832 1.1 cgd recs += GETRINTERNAL(h, nxt)->nrecs;
833 1.1 cgd return (recs);
834 1.1 cgd }
835