bt_split.c revision 1.12 1 1.12 mycroft /* $NetBSD: bt_split.c,v 1.12 2000/01/23 00:57:50 mycroft 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 1.12 mycroft __RCSID("$NetBSD: bt_split.c,v 1.12 2000/01/23 00:57:50 mycroft Exp $");
45 1.5 cgd #endif
46 1.1 cgd #endif /* LIBC_SCCS and not lint */
47 1.1 cgd
48 1.8 jtc #include "namespace.h"
49 1.1 cgd #include <sys/types.h>
50 1.1 cgd
51 1.1 cgd #include <limits.h>
52 1.1 cgd #include <stdio.h>
53 1.1 cgd #include <stdlib.h>
54 1.1 cgd #include <string.h>
55 1.1 cgd
56 1.1 cgd #include <db.h>
57 1.1 cgd #include "btree.h"
58 1.1 cgd
59 1.1 cgd static int bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
60 1.1 cgd static PAGE *bt_page
61 1.4 cgd __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
62 1.1 cgd static int bt_preserve __P((BTREE *, pgno_t));
63 1.1 cgd static PAGE *bt_psplit
64 1.4 cgd __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
65 1.1 cgd static PAGE *bt_root
66 1.4 cgd __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
67 1.1 cgd static int bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
68 1.1 cgd static recno_t rec_total __P((PAGE *));
69 1.1 cgd
70 1.1 cgd #ifdef STATISTICS
71 1.1 cgd u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
72 1.1 cgd #endif
73 1.1 cgd
74 1.1 cgd /*
75 1.1 cgd * __BT_SPLIT -- Split the tree.
76 1.1 cgd *
77 1.1 cgd * Parameters:
78 1.1 cgd * t: tree
79 1.1 cgd * sp: page to split
80 1.1 cgd * key: key to insert
81 1.1 cgd * data: data to insert
82 1.1 cgd * flags: BIGKEY/BIGDATA flags
83 1.1 cgd * ilen: insert length
84 1.1 cgd * skip: index to leave open
85 1.1 cgd *
86 1.1 cgd * Returns:
87 1.1 cgd * RET_ERROR, RET_SUCCESS
88 1.1 cgd */
89 1.1 cgd int
90 1.4 cgd __bt_split(t, sp, key, data, flags, ilen, argskip)
91 1.1 cgd BTREE *t;
92 1.1 cgd PAGE *sp;
93 1.1 cgd const DBT *key, *data;
94 1.4 cgd int flags;
95 1.1 cgd size_t ilen;
96 1.4 cgd u_int32_t argskip;
97 1.1 cgd {
98 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
99 1.7 christos BLEAF *bl = NULL, *tbl; /* pacify gcc */
100 1.1 cgd DBT a, b;
101 1.1 cgd EPGNO *parent;
102 1.1 cgd PAGE *h, *l, *r, *lchild, *rchild;
103 1.1 cgd indx_t nxtindex;
104 1.4 cgd u_int16_t skip;
105 1.7 christos u_int32_t n, nbytes, nksize = 0; /* pacify gcc */
106 1.1 cgd int parentsplit;
107 1.1 cgd char *dest;
108 1.1 cgd
109 1.1 cgd /*
110 1.1 cgd * Split the page into two pages, l and r. The split routines return
111 1.1 cgd * a pointer to the page into which the key should be inserted and with
112 1.1 cgd * skip set to the offset which should be used. Additionally, l and r
113 1.1 cgd * are pinned.
114 1.1 cgd */
115 1.4 cgd skip = argskip;
116 1.1 cgd h = sp->pgno == P_ROOT ?
117 1.1 cgd bt_root(t, sp, &l, &r, &skip, ilen) :
118 1.1 cgd bt_page(t, sp, &l, &r, &skip, ilen);
119 1.1 cgd if (h == NULL)
120 1.1 cgd return (RET_ERROR);
121 1.1 cgd
122 1.1 cgd /*
123 1.1 cgd * Insert the new key/data pair into the leaf page. (Key inserts
124 1.1 cgd * always cause a leaf page to split first.)
125 1.1 cgd */
126 1.1 cgd h->linp[skip] = h->upper -= ilen;
127 1.10 christos dest = (char *)(void *)h + h->upper;
128 1.6 cgd if (F_ISSET(t, R_RECNO))
129 1.1 cgd WR_RLEAF(dest, data, flags)
130 1.1 cgd else
131 1.1 cgd WR_BLEAF(dest, key, data, flags)
132 1.1 cgd
133 1.1 cgd /* If the root page was split, make it look right. */
134 1.1 cgd if (sp->pgno == P_ROOT &&
135 1.6 cgd (F_ISSET(t, R_RECNO) ?
136 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
137 1.1 cgd goto err2;
138 1.1 cgd
139 1.1 cgd /*
140 1.1 cgd * Now we walk the parent page stack -- a LIFO stack of the pages that
141 1.1 cgd * were traversed when we searched for the page that split. Each stack
142 1.1 cgd * entry is a page number and a page index offset. The offset is for
143 1.1 cgd * the page traversed on the search. We've just split a page, so we
144 1.1 cgd * have to insert a new key into the parent page.
145 1.1 cgd *
146 1.1 cgd * If the insert into the parent page causes it to split, may have to
147 1.1 cgd * continue splitting all the way up the tree. We stop if the root
148 1.1 cgd * splits or the page inserted into didn't have to split to hold the
149 1.1 cgd * new key. Some algorithms replace the key for the old page as well
150 1.1 cgd * as the new page. We don't, as there's no reason to believe that the
151 1.1 cgd * first key on the old page is any better than the key we have, and,
152 1.1 cgd * in the case of a key being placed at index 0 causing the split, the
153 1.1 cgd * key is unavailable.
154 1.1 cgd *
155 1.1 cgd * There are a maximum of 5 pages pinned at any time. We keep the left
156 1.1 cgd * and right pages pinned while working on the parent. The 5 are the
157 1.1 cgd * two children, left parent and right parent (when the parent splits)
158 1.1 cgd * and the root page or the overflow key page when calling bt_preserve.
159 1.1 cgd * This code must make sure that all pins are released other than the
160 1.1 cgd * root page or overflow page which is unlocked elsewhere.
161 1.1 cgd */
162 1.1 cgd while ((parent = BT_POP(t)) != NULL) {
163 1.1 cgd lchild = l;
164 1.1 cgd rchild = r;
165 1.1 cgd
166 1.1 cgd /* Get the parent page. */
167 1.1 cgd if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
168 1.1 cgd goto err2;
169 1.1 cgd
170 1.1 cgd /*
171 1.1 cgd * The new key goes ONE AFTER the index, because the split
172 1.1 cgd * was to the right.
173 1.1 cgd */
174 1.1 cgd skip = parent->index + 1;
175 1.1 cgd
176 1.1 cgd /*
177 1.1 cgd * Calculate the space needed on the parent page.
178 1.1 cgd *
179 1.1 cgd * Prefix trees: space hack when inserting into BINTERNAL
180 1.1 cgd * pages. Retain only what's needed to distinguish between
181 1.1 cgd * the new entry and the LAST entry on the page to its left.
182 1.1 cgd * If the keys compare equal, retain the entire key. Note,
183 1.1 cgd * we don't touch overflow keys, and the entire key must be
184 1.1 cgd * retained for the next-to-left most key on the leftmost
185 1.1 cgd * page of each level, or the search will fail. Applicable
186 1.1 cgd * ONLY to internal pages that have leaf pages as children.
187 1.1 cgd * Further reduction of the key between pairs of internal
188 1.1 cgd * pages loses too much information.
189 1.1 cgd */
190 1.1 cgd switch (rchild->flags & P_TYPE) {
191 1.1 cgd case P_BINTERNAL:
192 1.1 cgd bi = GETBINTERNAL(rchild, 0);
193 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
194 1.1 cgd break;
195 1.1 cgd case P_BLEAF:
196 1.1 cgd bl = GETBLEAF(rchild, 0);
197 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
198 1.1 cgd if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
199 1.1 cgd (h->prevpg != P_INVALID || skip > 1)) {
200 1.1 cgd tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
201 1.1 cgd a.size = tbl->ksize;
202 1.1 cgd a.data = tbl->bytes;
203 1.1 cgd b.size = bl->ksize;
204 1.1 cgd b.data = bl->bytes;
205 1.1 cgd nksize = t->bt_pfx(&a, &b);
206 1.1 cgd n = NBINTERNAL(nksize);
207 1.1 cgd if (n < nbytes) {
208 1.1 cgd #ifdef STATISTICS
209 1.1 cgd bt_pfxsaved += nbytes - n;
210 1.1 cgd #endif
211 1.1 cgd nbytes = n;
212 1.1 cgd } else
213 1.1 cgd nksize = 0;
214 1.1 cgd } else
215 1.1 cgd nksize = 0;
216 1.1 cgd break;
217 1.1 cgd case P_RINTERNAL:
218 1.1 cgd case P_RLEAF:
219 1.1 cgd nbytes = NRINTERNAL;
220 1.1 cgd break;
221 1.1 cgd default:
222 1.1 cgd abort();
223 1.1 cgd }
224 1.1 cgd
225 1.1 cgd /* Split the parent page if necessary or shift the indices. */
226 1.1 cgd if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
227 1.1 cgd sp = h;
228 1.1 cgd h = h->pgno == P_ROOT ?
229 1.1 cgd bt_root(t, h, &l, &r, &skip, nbytes) :
230 1.1 cgd bt_page(t, h, &l, &r, &skip, nbytes);
231 1.1 cgd if (h == NULL)
232 1.1 cgd goto err1;
233 1.1 cgd parentsplit = 1;
234 1.1 cgd } else {
235 1.1 cgd if (skip < (nxtindex = NEXTINDEX(h)))
236 1.1 cgd memmove(h->linp + skip + 1, h->linp + skip,
237 1.1 cgd (nxtindex - skip) * sizeof(indx_t));
238 1.1 cgd h->lower += sizeof(indx_t);
239 1.1 cgd parentsplit = 0;
240 1.1 cgd }
241 1.1 cgd
242 1.1 cgd /* Insert the key into the parent page. */
243 1.6 cgd switch (rchild->flags & P_TYPE) {
244 1.1 cgd case P_BINTERNAL:
245 1.1 cgd h->linp[skip] = h->upper -= nbytes;
246 1.10 christos dest = (char *)(void *)h + h->linp[skip];
247 1.1 cgd memmove(dest, bi, nbytes);
248 1.10 christos ((BINTERNAL *)(void *)dest)->pgno = rchild->pgno;
249 1.1 cgd break;
250 1.1 cgd case P_BLEAF:
251 1.1 cgd h->linp[skip] = h->upper -= nbytes;
252 1.10 christos dest = (char *)(void *)h + h->linp[skip];
253 1.1 cgd WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
254 1.1 cgd rchild->pgno, bl->flags & P_BIGKEY);
255 1.1 cgd memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
256 1.1 cgd if (bl->flags & P_BIGKEY &&
257 1.10 christos bt_preserve(t, *(pgno_t *)(void *)bl->bytes) ==
258 1.10 christos RET_ERROR)
259 1.1 cgd goto err1;
260 1.1 cgd break;
261 1.1 cgd case P_RINTERNAL:
262 1.1 cgd /*
263 1.1 cgd * Update the left page count. If split
264 1.1 cgd * added at index 0, fix the correct page.
265 1.1 cgd */
266 1.1 cgd if (skip > 0)
267 1.10 christos dest = (char *)(void *)h + h->linp[skip - 1];
268 1.1 cgd else
269 1.10 christos dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
270 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild);
271 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
272 1.1 cgd
273 1.1 cgd /* Update the right page count. */
274 1.1 cgd h->linp[skip] = h->upper -= nbytes;
275 1.10 christos dest = (char *)(void *)h + h->linp[skip];
276 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild);
277 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
278 1.1 cgd break;
279 1.1 cgd case P_RLEAF:
280 1.1 cgd /*
281 1.1 cgd * Update the left page count. If split
282 1.1 cgd * added at index 0, fix the correct page.
283 1.1 cgd */
284 1.1 cgd if (skip > 0)
285 1.10 christos dest = (char *)(void *)h + h->linp[skip - 1];
286 1.1 cgd else
287 1.10 christos dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
288 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild);
289 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
290 1.1 cgd
291 1.1 cgd /* Update the right page count. */
292 1.1 cgd h->linp[skip] = h->upper -= nbytes;
293 1.10 christos dest = (char *)(void *)h + h->linp[skip];
294 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild);
295 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
296 1.1 cgd break;
297 1.1 cgd default:
298 1.1 cgd abort();
299 1.1 cgd }
300 1.1 cgd
301 1.1 cgd /* Unpin the held pages. */
302 1.1 cgd if (!parentsplit) {
303 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
304 1.1 cgd break;
305 1.1 cgd }
306 1.1 cgd
307 1.1 cgd /* If the root page was split, make it look right. */
308 1.1 cgd if (sp->pgno == P_ROOT &&
309 1.6 cgd (F_ISSET(t, R_RECNO) ?
310 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
311 1.1 cgd goto err1;
312 1.1 cgd
313 1.1 cgd mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
314 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
315 1.1 cgd }
316 1.1 cgd
317 1.1 cgd /* Unpin the held pages. */
318 1.1 cgd mpool_put(t->bt_mp, l, MPOOL_DIRTY);
319 1.1 cgd mpool_put(t->bt_mp, r, MPOOL_DIRTY);
320 1.1 cgd
321 1.1 cgd /* Clear any pages left on the stack. */
322 1.1 cgd return (RET_SUCCESS);
323 1.1 cgd
324 1.1 cgd /*
325 1.1 cgd * If something fails in the above loop we were already walking back
326 1.1 cgd * up the tree and the tree is now inconsistent. Nothing much we can
327 1.1 cgd * do about it but release any memory we're holding.
328 1.1 cgd */
329 1.1 cgd err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
330 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
331 1.1 cgd
332 1.1 cgd err2: mpool_put(t->bt_mp, l, 0);
333 1.1 cgd mpool_put(t->bt_mp, r, 0);
334 1.1 cgd __dbpanic(t->bt_dbp);
335 1.1 cgd return (RET_ERROR);
336 1.1 cgd }
337 1.1 cgd
338 1.1 cgd /*
339 1.1 cgd * BT_PAGE -- Split a non-root page of a btree.
340 1.1 cgd *
341 1.1 cgd * Parameters:
342 1.1 cgd * t: tree
343 1.1 cgd * h: root page
344 1.1 cgd * lp: pointer to left page pointer
345 1.1 cgd * rp: pointer to right page pointer
346 1.1 cgd * skip: pointer to index to leave open
347 1.1 cgd * ilen: insert length
348 1.1 cgd *
349 1.1 cgd * Returns:
350 1.1 cgd * Pointer to page in which to insert or NULL on error.
351 1.1 cgd */
352 1.1 cgd static PAGE *
353 1.1 cgd bt_page(t, h, lp, rp, skip, ilen)
354 1.1 cgd BTREE *t;
355 1.1 cgd PAGE *h, **lp, **rp;
356 1.4 cgd indx_t *skip;
357 1.1 cgd size_t ilen;
358 1.1 cgd {
359 1.1 cgd PAGE *l, *r, *tp;
360 1.1 cgd pgno_t npg;
361 1.1 cgd
362 1.1 cgd #ifdef STATISTICS
363 1.1 cgd ++bt_split;
364 1.1 cgd #endif
365 1.1 cgd /* Put the new right page for the split into place. */
366 1.1 cgd if ((r = __bt_new(t, &npg)) == NULL)
367 1.1 cgd return (NULL);
368 1.1 cgd r->pgno = npg;
369 1.1 cgd r->lower = BTDATAOFF;
370 1.1 cgd r->upper = t->bt_psize;
371 1.1 cgd r->nextpg = h->nextpg;
372 1.1 cgd r->prevpg = h->pgno;
373 1.1 cgd r->flags = h->flags & P_TYPE;
374 1.1 cgd
375 1.1 cgd /*
376 1.1 cgd * If we're splitting the last page on a level because we're appending
377 1.1 cgd * a key to it (skip is NEXTINDEX()), it's likely that the data is
378 1.1 cgd * sorted. Adding an empty page on the side of the level is less work
379 1.1 cgd * and can push the fill factor much higher than normal. If we're
380 1.1 cgd * wrong it's no big deal, we'll just do the split the right way next
381 1.1 cgd * time. It may look like it's equally easy to do a similar hack for
382 1.1 cgd * reverse sorted data, that is, split the tree left, but it's not.
383 1.1 cgd * Don't even try.
384 1.1 cgd */
385 1.1 cgd if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
386 1.1 cgd #ifdef STATISTICS
387 1.1 cgd ++bt_sortsplit;
388 1.1 cgd #endif
389 1.1 cgd h->nextpg = r->pgno;
390 1.1 cgd r->lower = BTDATAOFF + sizeof(indx_t);
391 1.1 cgd *skip = 0;
392 1.1 cgd *lp = h;
393 1.1 cgd *rp = r;
394 1.1 cgd return (r);
395 1.1 cgd }
396 1.1 cgd
397 1.1 cgd /* Put the new left page for the split into place. */
398 1.4 cgd if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
399 1.1 cgd mpool_put(t->bt_mp, r, 0);
400 1.1 cgd return (NULL);
401 1.1 cgd }
402 1.6 cgd #ifdef PURIFY
403 1.6 cgd memset(l, 0xff, t->bt_psize);
404 1.6 cgd #endif
405 1.1 cgd l->pgno = h->pgno;
406 1.1 cgd l->nextpg = r->pgno;
407 1.1 cgd l->prevpg = h->prevpg;
408 1.1 cgd l->lower = BTDATAOFF;
409 1.1 cgd l->upper = t->bt_psize;
410 1.1 cgd l->flags = h->flags & P_TYPE;
411 1.1 cgd
412 1.1 cgd /* Fix up the previous pointer of the page after the split page. */
413 1.1 cgd if (h->nextpg != P_INVALID) {
414 1.1 cgd if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
415 1.1 cgd free(l);
416 1.1 cgd /* XXX mpool_free(t->bt_mp, r->pgno); */
417 1.1 cgd return (NULL);
418 1.1 cgd }
419 1.1 cgd tp->prevpg = r->pgno;
420 1.6 cgd mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
421 1.1 cgd }
422 1.1 cgd
423 1.1 cgd /*
424 1.1 cgd * Split right. The key/data pairs aren't sorted in the btree page so
425 1.1 cgd * it's simpler to copy the data from the split page onto two new pages
426 1.1 cgd * instead of copying half the data to the right page and compacting
427 1.1 cgd * the left page in place. Since the left page can't change, we have
428 1.1 cgd * to swap the original and the allocated left page after the split.
429 1.1 cgd */
430 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
431 1.1 cgd
432 1.1 cgd /* Move the new left page onto the old left page. */
433 1.1 cgd memmove(h, l, t->bt_psize);
434 1.1 cgd if (tp == l)
435 1.1 cgd tp = h;
436 1.1 cgd free(l);
437 1.1 cgd
438 1.1 cgd *lp = h;
439 1.1 cgd *rp = r;
440 1.1 cgd return (tp);
441 1.1 cgd }
442 1.1 cgd
443 1.1 cgd /*
444 1.1 cgd * BT_ROOT -- Split the root page of a btree.
445 1.1 cgd *
446 1.1 cgd * Parameters:
447 1.1 cgd * t: tree
448 1.1 cgd * h: root page
449 1.1 cgd * lp: pointer to left page pointer
450 1.1 cgd * rp: pointer to right page pointer
451 1.1 cgd * skip: pointer to index to leave open
452 1.1 cgd * ilen: insert length
453 1.1 cgd *
454 1.1 cgd * Returns:
455 1.1 cgd * Pointer to page in which to insert or NULL on error.
456 1.1 cgd */
457 1.1 cgd static PAGE *
458 1.1 cgd bt_root(t, h, lp, rp, skip, ilen)
459 1.1 cgd BTREE *t;
460 1.1 cgd PAGE *h, **lp, **rp;
461 1.4 cgd indx_t *skip;
462 1.1 cgd size_t ilen;
463 1.1 cgd {
464 1.1 cgd PAGE *l, *r, *tp;
465 1.1 cgd pgno_t lnpg, rnpg;
466 1.1 cgd
467 1.1 cgd #ifdef STATISTICS
468 1.1 cgd ++bt_split;
469 1.1 cgd ++bt_rootsplit;
470 1.1 cgd #endif
471 1.1 cgd /* Put the new left and right pages for the split into place. */
472 1.1 cgd if ((l = __bt_new(t, &lnpg)) == NULL ||
473 1.1 cgd (r = __bt_new(t, &rnpg)) == NULL)
474 1.1 cgd return (NULL);
475 1.1 cgd l->pgno = lnpg;
476 1.1 cgd r->pgno = rnpg;
477 1.1 cgd l->nextpg = r->pgno;
478 1.1 cgd r->prevpg = l->pgno;
479 1.1 cgd l->prevpg = r->nextpg = P_INVALID;
480 1.1 cgd l->lower = r->lower = BTDATAOFF;
481 1.1 cgd l->upper = r->upper = t->bt_psize;
482 1.1 cgd l->flags = r->flags = h->flags & P_TYPE;
483 1.1 cgd
484 1.1 cgd /* Split the root page. */
485 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
486 1.1 cgd
487 1.1 cgd *lp = l;
488 1.1 cgd *rp = r;
489 1.1 cgd return (tp);
490 1.1 cgd }
491 1.1 cgd
492 1.1 cgd /*
493 1.1 cgd * BT_RROOT -- Fix up the recno root page after it has been split.
494 1.1 cgd *
495 1.1 cgd * Parameters:
496 1.1 cgd * t: tree
497 1.1 cgd * h: root page
498 1.1 cgd * l: left page
499 1.1 cgd * r: right page
500 1.1 cgd *
501 1.1 cgd * Returns:
502 1.1 cgd * RET_ERROR, RET_SUCCESS
503 1.1 cgd */
504 1.1 cgd static int
505 1.1 cgd bt_rroot(t, h, l, r)
506 1.1 cgd BTREE *t;
507 1.1 cgd PAGE *h, *l, *r;
508 1.1 cgd {
509 1.1 cgd char *dest;
510 1.1 cgd
511 1.1 cgd /* Insert the left and right keys, set the header information. */
512 1.1 cgd h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
513 1.10 christos dest = (char *)(void *)h + h->upper;
514 1.1 cgd WR_RINTERNAL(dest,
515 1.1 cgd l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
516 1.1 cgd
517 1.1 cgd h->linp[1] = h->upper -= NRINTERNAL;
518 1.10 christos dest = (char *)(void *)h + h->upper;
519 1.1 cgd WR_RINTERNAL(dest,
520 1.1 cgd r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
521 1.1 cgd
522 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
523 1.1 cgd
524 1.1 cgd /* Unpin the root page, set to recno internal page. */
525 1.1 cgd h->flags &= ~P_TYPE;
526 1.1 cgd h->flags |= P_RINTERNAL;
527 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
528 1.1 cgd
529 1.1 cgd return (RET_SUCCESS);
530 1.1 cgd }
531 1.1 cgd
532 1.1 cgd /*
533 1.1 cgd * BT_BROOT -- Fix up the btree root page after it has been split.
534 1.1 cgd *
535 1.1 cgd * Parameters:
536 1.1 cgd * t: tree
537 1.1 cgd * h: root page
538 1.1 cgd * l: left page
539 1.1 cgd * r: right page
540 1.1 cgd *
541 1.1 cgd * Returns:
542 1.1 cgd * RET_ERROR, RET_SUCCESS
543 1.1 cgd */
544 1.1 cgd static int
545 1.1 cgd bt_broot(t, h, l, r)
546 1.1 cgd BTREE *t;
547 1.1 cgd PAGE *h, *l, *r;
548 1.1 cgd {
549 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
550 1.1 cgd BLEAF *bl;
551 1.4 cgd u_int32_t nbytes;
552 1.1 cgd char *dest;
553 1.1 cgd
554 1.1 cgd /*
555 1.1 cgd * If the root page was a leaf page, change it into an internal page.
556 1.1 cgd * We copy the key we split on (but not the key's data, in the case of
557 1.1 cgd * a leaf page) to the new root page.
558 1.1 cgd *
559 1.1 cgd * The btree comparison code guarantees that the left-most key on any
560 1.1 cgd * level of the tree is never used, so it doesn't need to be filled in.
561 1.1 cgd */
562 1.1 cgd nbytes = NBINTERNAL(0);
563 1.1 cgd h->linp[0] = h->upper = t->bt_psize - nbytes;
564 1.10 christos dest = (char *)(void *)h + h->upper;
565 1.1 cgd WR_BINTERNAL(dest, 0, l->pgno, 0);
566 1.1 cgd
567 1.6 cgd switch (h->flags & P_TYPE) {
568 1.1 cgd case P_BLEAF:
569 1.1 cgd bl = GETBLEAF(r, 0);
570 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
571 1.1 cgd h->linp[1] = h->upper -= nbytes;
572 1.10 christos dest = (char *)(void *)h + h->upper;
573 1.1 cgd WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
574 1.1 cgd memmove(dest, bl->bytes, bl->ksize);
575 1.1 cgd
576 1.1 cgd /*
577 1.1 cgd * If the key is on an overflow page, mark the overflow chain
578 1.1 cgd * so it isn't deleted when the leaf copy of the key is deleted.
579 1.1 cgd */
580 1.1 cgd if (bl->flags & P_BIGKEY &&
581 1.10 christos bt_preserve(t, *(pgno_t *)(void *)bl->bytes) == RET_ERROR)
582 1.1 cgd return (RET_ERROR);
583 1.1 cgd break;
584 1.1 cgd case P_BINTERNAL:
585 1.1 cgd bi = GETBINTERNAL(r, 0);
586 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
587 1.1 cgd h->linp[1] = h->upper -= nbytes;
588 1.10 christos dest = (char *)(void *)h + h->upper;
589 1.1 cgd memmove(dest, bi, nbytes);
590 1.10 christos ((BINTERNAL *)(void *)dest)->pgno = r->pgno;
591 1.1 cgd break;
592 1.1 cgd default:
593 1.1 cgd abort();
594 1.1 cgd }
595 1.1 cgd
596 1.1 cgd /* There are two keys on the page. */
597 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
598 1.1 cgd
599 1.1 cgd /* Unpin the root page, set to btree internal page. */
600 1.1 cgd h->flags &= ~P_TYPE;
601 1.1 cgd h->flags |= P_BINTERNAL;
602 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
603 1.1 cgd
604 1.1 cgd return (RET_SUCCESS);
605 1.1 cgd }
606 1.1 cgd
607 1.1 cgd /*
608 1.1 cgd * BT_PSPLIT -- Do the real work of splitting the page.
609 1.1 cgd *
610 1.1 cgd * Parameters:
611 1.1 cgd * t: tree
612 1.1 cgd * h: page to be split
613 1.1 cgd * l: page to put lower half of data
614 1.1 cgd * r: page to put upper half of data
615 1.1 cgd * pskip: pointer to index to leave open
616 1.1 cgd * ilen: insert length
617 1.1 cgd *
618 1.1 cgd * Returns:
619 1.1 cgd * Pointer to page in which to insert.
620 1.1 cgd */
621 1.1 cgd static PAGE *
622 1.1 cgd bt_psplit(t, h, l, r, pskip, ilen)
623 1.1 cgd BTREE *t;
624 1.1 cgd PAGE *h, *l, *r;
625 1.4 cgd indx_t *pskip;
626 1.1 cgd size_t ilen;
627 1.1 cgd {
628 1.1 cgd BINTERNAL *bi;
629 1.1 cgd BLEAF *bl;
630 1.6 cgd CURSOR *c;
631 1.1 cgd RLEAF *rl;
632 1.1 cgd PAGE *rval;
633 1.7 christos void *src = NULL; /* pacify gcc */
634 1.1 cgd indx_t full, half, nxt, off, skip, top, used;
635 1.4 cgd u_int32_t nbytes;
636 1.1 cgd int bigkeycnt, isbigkey;
637 1.1 cgd
638 1.1 cgd /*
639 1.1 cgd * Split the data to the left and right pages. Leave the skip index
640 1.1 cgd * open. Additionally, make some effort not to split on an overflow
641 1.1 cgd * key. This makes internal page processing faster and can save
642 1.1 cgd * space as overflow keys used by internal pages are never deleted.
643 1.1 cgd */
644 1.1 cgd bigkeycnt = 0;
645 1.1 cgd skip = *pskip;
646 1.1 cgd full = t->bt_psize - BTDATAOFF;
647 1.1 cgd half = full / 2;
648 1.1 cgd used = 0;
649 1.1 cgd for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
650 1.1 cgd if (skip == off) {
651 1.1 cgd nbytes = ilen;
652 1.1 cgd isbigkey = 0; /* XXX: not really known. */
653 1.1 cgd } else
654 1.1 cgd switch (h->flags & P_TYPE) {
655 1.1 cgd case P_BINTERNAL:
656 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
657 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
658 1.1 cgd isbigkey = bi->flags & P_BIGKEY;
659 1.1 cgd break;
660 1.1 cgd case P_BLEAF:
661 1.1 cgd src = bl = GETBLEAF(h, nxt);
662 1.1 cgd nbytes = NBLEAF(bl);
663 1.1 cgd isbigkey = bl->flags & P_BIGKEY;
664 1.1 cgd break;
665 1.1 cgd case P_RINTERNAL:
666 1.1 cgd src = GETRINTERNAL(h, nxt);
667 1.1 cgd nbytes = NRINTERNAL;
668 1.1 cgd isbigkey = 0;
669 1.1 cgd break;
670 1.1 cgd case P_RLEAF:
671 1.1 cgd src = rl = GETRLEAF(h, nxt);
672 1.1 cgd nbytes = NRLEAF(rl);
673 1.1 cgd isbigkey = 0;
674 1.1 cgd break;
675 1.1 cgd default:
676 1.1 cgd abort();
677 1.1 cgd }
678 1.1 cgd
679 1.1 cgd /*
680 1.1 cgd * If the key/data pairs are substantial fractions of the max
681 1.1 cgd * possible size for the page, it's possible to get situations
682 1.1 cgd * where we decide to try and copy too much onto the left page.
683 1.1 cgd * Make sure that doesn't happen.
684 1.1 cgd */
685 1.9 is if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
686 1.9 is nxt == top - 1) {
687 1.1 cgd --off;
688 1.1 cgd break;
689 1.1 cgd }
690 1.1 cgd
691 1.1 cgd /* Copy the key/data pair, if not the skipped index. */
692 1.1 cgd if (skip != off) {
693 1.1 cgd ++nxt;
694 1.1 cgd
695 1.1 cgd l->linp[off] = l->upper -= nbytes;
696 1.10 christos memmove((char *)(void *)l + l->upper, src, nbytes);
697 1.1 cgd }
698 1.1 cgd
699 1.9 is used += nbytes + sizeof(indx_t);
700 1.1 cgd if (used >= half) {
701 1.1 cgd if (!isbigkey || bigkeycnt == 3)
702 1.1 cgd break;
703 1.1 cgd else
704 1.1 cgd ++bigkeycnt;
705 1.1 cgd }
706 1.1 cgd }
707 1.1 cgd
708 1.1 cgd /*
709 1.1 cgd * Off is the last offset that's valid for the left page.
710 1.1 cgd * Nxt is the first offset to be placed on the right page.
711 1.1 cgd */
712 1.1 cgd l->lower += (off + 1) * sizeof(indx_t);
713 1.1 cgd
714 1.1 cgd /*
715 1.1 cgd * If splitting the page that the cursor was on, the cursor has to be
716 1.1 cgd * adjusted to point to the same record as before the split. If the
717 1.1 cgd * cursor is at or past the skipped slot, the cursor is incremented by
718 1.1 cgd * one. If the cursor is on the right page, it is decremented by the
719 1.1 cgd * number of records split to the left page.
720 1.1 cgd */
721 1.6 cgd c = &t->bt_cursor;
722 1.6 cgd if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
723 1.6 cgd if (c->pg.index >= skip)
724 1.6 cgd ++c->pg.index;
725 1.6 cgd if (c->pg.index < nxt) /* Left page. */
726 1.6 cgd c->pg.pgno = l->pgno;
727 1.1 cgd else { /* Right page. */
728 1.6 cgd c->pg.pgno = r->pgno;
729 1.6 cgd c->pg.index -= nxt;
730 1.1 cgd }
731 1.1 cgd }
732 1.1 cgd
733 1.1 cgd /*
734 1.1 cgd * If the skipped index was on the left page, just return that page.
735 1.1 cgd * Otherwise, adjust the skip index to reflect the new position on
736 1.1 cgd * the right page.
737 1.1 cgd */
738 1.1 cgd if (skip <= off) {
739 1.12 mycroft skip = MAX_PAGE_OFFSET;
740 1.1 cgd rval = l;
741 1.1 cgd } else {
742 1.1 cgd rval = r;
743 1.1 cgd *pskip -= nxt;
744 1.1 cgd }
745 1.1 cgd
746 1.1 cgd for (off = 0; nxt < top; ++off) {
747 1.1 cgd if (skip == nxt) {
748 1.1 cgd ++off;
749 1.12 mycroft skip = MAX_PAGE_OFFSET;
750 1.1 cgd }
751 1.1 cgd switch (h->flags & P_TYPE) {
752 1.1 cgd case P_BINTERNAL:
753 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
754 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
755 1.1 cgd break;
756 1.1 cgd case P_BLEAF:
757 1.1 cgd src = bl = GETBLEAF(h, nxt);
758 1.1 cgd nbytes = NBLEAF(bl);
759 1.1 cgd break;
760 1.1 cgd case P_RINTERNAL:
761 1.1 cgd src = GETRINTERNAL(h, nxt);
762 1.1 cgd nbytes = NRINTERNAL;
763 1.1 cgd break;
764 1.1 cgd case P_RLEAF:
765 1.1 cgd src = rl = GETRLEAF(h, nxt);
766 1.1 cgd nbytes = NRLEAF(rl);
767 1.1 cgd break;
768 1.1 cgd default:
769 1.1 cgd abort();
770 1.1 cgd }
771 1.1 cgd ++nxt;
772 1.1 cgd r->linp[off] = r->upper -= nbytes;
773 1.10 christos memmove((char *)(void *)r + r->upper, src, nbytes);
774 1.1 cgd }
775 1.1 cgd r->lower += off * sizeof(indx_t);
776 1.1 cgd
777 1.1 cgd /* If the key is being appended to the page, adjust the index. */
778 1.1 cgd if (skip == top)
779 1.1 cgd r->lower += sizeof(indx_t);
780 1.1 cgd
781 1.1 cgd return (rval);
782 1.1 cgd }
783 1.1 cgd
784 1.1 cgd /*
785 1.1 cgd * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
786 1.1 cgd *
787 1.1 cgd * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
788 1.1 cgd * record that references them gets deleted. Chains pointed to by internal
789 1.1 cgd * pages never get deleted. This routine marks a chain as pointed to by an
790 1.1 cgd * internal page.
791 1.1 cgd *
792 1.1 cgd * Parameters:
793 1.1 cgd * t: tree
794 1.1 cgd * pg: page number of first page in the chain.
795 1.1 cgd *
796 1.1 cgd * Returns:
797 1.1 cgd * RET_SUCCESS, RET_ERROR.
798 1.1 cgd */
799 1.1 cgd static int
800 1.1 cgd bt_preserve(t, pg)
801 1.1 cgd BTREE *t;
802 1.1 cgd pgno_t pg;
803 1.1 cgd {
804 1.1 cgd PAGE *h;
805 1.1 cgd
806 1.1 cgd if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
807 1.1 cgd return (RET_ERROR);
808 1.1 cgd h->flags |= P_PRESERVE;
809 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
810 1.1 cgd return (RET_SUCCESS);
811 1.1 cgd }
812 1.1 cgd
813 1.1 cgd /*
814 1.1 cgd * REC_TOTAL -- Return the number of recno entries below a page.
815 1.1 cgd *
816 1.1 cgd * Parameters:
817 1.1 cgd * h: page
818 1.1 cgd *
819 1.1 cgd * Returns:
820 1.1 cgd * The number of recno entries below a page.
821 1.1 cgd *
822 1.1 cgd * XXX
823 1.1 cgd * These values could be set by the bt_psplit routine. The problem is that the
824 1.1 cgd * entry has to be popped off of the stack etc. or the values have to be passed
825 1.1 cgd * all the way back to bt_split/bt_rroot and it's not very clean.
826 1.1 cgd */
827 1.1 cgd static recno_t
828 1.1 cgd rec_total(h)
829 1.1 cgd PAGE *h;
830 1.1 cgd {
831 1.1 cgd recno_t recs;
832 1.1 cgd indx_t nxt, top;
833 1.1 cgd
834 1.1 cgd for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
835 1.1 cgd recs += GETRINTERNAL(h, nxt)->nrecs;
836 1.1 cgd return (recs);
837 1.1 cgd }
838