bt_split.c revision 1.19 1 1.19 christos /* $NetBSD: bt_split.c,v 1.19 2009/04/22 18:44:06 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 1.13 agc * 3. Neither the name of the University nor the names of its contributors
19 1.1 cgd * may be used to endorse or promote products derived from this software
20 1.1 cgd * without specific prior written permission.
21 1.1 cgd *
22 1.1 cgd * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 1.1 cgd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 1.1 cgd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 1.1 cgd * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 1.1 cgd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 1.1 cgd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 1.1 cgd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 1.1 cgd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 1.1 cgd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 1.1 cgd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 1.1 cgd * SUCH DAMAGE.
33 1.1 cgd */
34 1.1 cgd
35 1.17 joerg #if HAVE_NBTOOL_CONFIG_H
36 1.17 joerg #include "nbtool_config.h"
37 1.17 joerg #endif
38 1.17 joerg
39 1.7 christos #include <sys/cdefs.h>
40 1.19 christos __RCSID("$NetBSD: bt_split.c,v 1.19 2009/04/22 18:44:06 christos Exp $");
41 1.1 cgd
42 1.8 jtc #include "namespace.h"
43 1.1 cgd #include <sys/types.h>
44 1.1 cgd
45 1.14 christos #include <assert.h>
46 1.1 cgd #include <limits.h>
47 1.1 cgd #include <stdio.h>
48 1.1 cgd #include <stdlib.h>
49 1.1 cgd #include <string.h>
50 1.1 cgd
51 1.1 cgd #include <db.h>
52 1.1 cgd #include "btree.h"
53 1.1 cgd
54 1.14 christos static int bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
55 1.14 christos static PAGE *bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
56 1.14 christos static int bt_preserve(BTREE *, pgno_t);
57 1.14 christos static PAGE *bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
58 1.14 christos static PAGE *bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
59 1.14 christos static int bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
60 1.14 christos static recno_t rec_total(PAGE *);
61 1.1 cgd
62 1.1 cgd #ifdef STATISTICS
63 1.15 joerg unsigned long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
64 1.1 cgd #endif
65 1.1 cgd
66 1.1 cgd /*
67 1.1 cgd * __BT_SPLIT -- Split the tree.
68 1.1 cgd *
69 1.1 cgd * Parameters:
70 1.1 cgd * t: tree
71 1.1 cgd * sp: page to split
72 1.1 cgd * key: key to insert
73 1.1 cgd * data: data to insert
74 1.1 cgd * flags: BIGKEY/BIGDATA flags
75 1.1 cgd * ilen: insert length
76 1.1 cgd * skip: index to leave open
77 1.1 cgd *
78 1.1 cgd * Returns:
79 1.1 cgd * RET_ERROR, RET_SUCCESS
80 1.1 cgd */
81 1.1 cgd int
82 1.14 christos __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
83 1.15 joerg size_t ilen, uint32_t argskip)
84 1.1 cgd {
85 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
86 1.7 christos BLEAF *bl = NULL, *tbl; /* pacify gcc */
87 1.1 cgd DBT a, b;
88 1.1 cgd EPGNO *parent;
89 1.1 cgd PAGE *h, *l, *r, *lchild, *rchild;
90 1.1 cgd indx_t nxtindex;
91 1.15 joerg uint16_t skip;
92 1.15 joerg uint32_t n, nbytes, nksize = 0; /* pacify gcc */
93 1.1 cgd int parentsplit;
94 1.1 cgd char *dest;
95 1.1 cgd
96 1.1 cgd /*
97 1.1 cgd * Split the page into two pages, l and r. The split routines return
98 1.1 cgd * a pointer to the page into which the key should be inserted and with
99 1.1 cgd * skip set to the offset which should be used. Additionally, l and r
100 1.1 cgd * are pinned.
101 1.1 cgd */
102 1.4 cgd skip = argskip;
103 1.1 cgd h = sp->pgno == P_ROOT ?
104 1.1 cgd bt_root(t, sp, &l, &r, &skip, ilen) :
105 1.1 cgd bt_page(t, sp, &l, &r, &skip, ilen);
106 1.1 cgd if (h == NULL)
107 1.1 cgd return (RET_ERROR);
108 1.1 cgd
109 1.1 cgd /*
110 1.1 cgd * Insert the new key/data pair into the leaf page. (Key inserts
111 1.1 cgd * always cause a leaf page to split first.)
112 1.1 cgd */
113 1.14 christos _DBFIT(ilen, indx_t);
114 1.14 christos h->upper -= (indx_t)ilen;
115 1.14 christos h->linp[skip] = h->upper;
116 1.10 christos dest = (char *)(void *)h + h->upper;
117 1.6 cgd if (F_ISSET(t, R_RECNO))
118 1.14 christos WR_RLEAF(dest, data, flags);
119 1.1 cgd else
120 1.14 christos WR_BLEAF(dest, key, data, flags);
121 1.1 cgd
122 1.1 cgd /* If the root page was split, make it look right. */
123 1.1 cgd if (sp->pgno == P_ROOT &&
124 1.6 cgd (F_ISSET(t, R_RECNO) ?
125 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
126 1.1 cgd goto err2;
127 1.1 cgd
128 1.1 cgd /*
129 1.1 cgd * Now we walk the parent page stack -- a LIFO stack of the pages that
130 1.1 cgd * were traversed when we searched for the page that split. Each stack
131 1.1 cgd * entry is a page number and a page index offset. The offset is for
132 1.1 cgd * the page traversed on the search. We've just split a page, so we
133 1.1 cgd * have to insert a new key into the parent page.
134 1.1 cgd *
135 1.1 cgd * If the insert into the parent page causes it to split, may have to
136 1.1 cgd * continue splitting all the way up the tree. We stop if the root
137 1.1 cgd * splits or the page inserted into didn't have to split to hold the
138 1.1 cgd * new key. Some algorithms replace the key for the old page as well
139 1.1 cgd * as the new page. We don't, as there's no reason to believe that the
140 1.1 cgd * first key on the old page is any better than the key we have, and,
141 1.1 cgd * in the case of a key being placed at index 0 causing the split, the
142 1.1 cgd * key is unavailable.
143 1.1 cgd *
144 1.1 cgd * There are a maximum of 5 pages pinned at any time. We keep the left
145 1.1 cgd * and right pages pinned while working on the parent. The 5 are the
146 1.1 cgd * two children, left parent and right parent (when the parent splits)
147 1.1 cgd * and the root page or the overflow key page when calling bt_preserve.
148 1.1 cgd * This code must make sure that all pins are released other than the
149 1.1 cgd * root page or overflow page which is unlocked elsewhere.
150 1.1 cgd */
151 1.1 cgd while ((parent = BT_POP(t)) != NULL) {
152 1.1 cgd lchild = l;
153 1.1 cgd rchild = r;
154 1.1 cgd
155 1.1 cgd /* Get the parent page. */
156 1.1 cgd if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
157 1.1 cgd goto err2;
158 1.1 cgd
159 1.1 cgd /*
160 1.1 cgd * The new key goes ONE AFTER the index, because the split
161 1.1 cgd * was to the right.
162 1.1 cgd */
163 1.1 cgd skip = parent->index + 1;
164 1.1 cgd
165 1.1 cgd /*
166 1.1 cgd * Calculate the space needed on the parent page.
167 1.1 cgd *
168 1.1 cgd * Prefix trees: space hack when inserting into BINTERNAL
169 1.1 cgd * pages. Retain only what's needed to distinguish between
170 1.1 cgd * the new entry and the LAST entry on the page to its left.
171 1.1 cgd * If the keys compare equal, retain the entire key. Note,
172 1.1 cgd * we don't touch overflow keys, and the entire key must be
173 1.1 cgd * retained for the next-to-left most key on the leftmost
174 1.1 cgd * page of each level, or the search will fail. Applicable
175 1.1 cgd * ONLY to internal pages that have leaf pages as children.
176 1.1 cgd * Further reduction of the key between pairs of internal
177 1.1 cgd * pages loses too much information.
178 1.1 cgd */
179 1.1 cgd switch (rchild->flags & P_TYPE) {
180 1.1 cgd case P_BINTERNAL:
181 1.1 cgd bi = GETBINTERNAL(rchild, 0);
182 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
183 1.1 cgd break;
184 1.1 cgd case P_BLEAF:
185 1.1 cgd bl = GETBLEAF(rchild, 0);
186 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
187 1.1 cgd if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
188 1.1 cgd (h->prevpg != P_INVALID || skip > 1)) {
189 1.14 christos size_t temp;
190 1.1 cgd tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
191 1.1 cgd a.size = tbl->ksize;
192 1.1 cgd a.data = tbl->bytes;
193 1.1 cgd b.size = bl->ksize;
194 1.1 cgd b.data = bl->bytes;
195 1.14 christos temp = t->bt_pfx(&a, &b);
196 1.15 joerg _DBFIT(temp, uint32_t);
197 1.15 joerg nksize = (uint32_t)temp;
198 1.1 cgd n = NBINTERNAL(nksize);
199 1.1 cgd if (n < nbytes) {
200 1.1 cgd #ifdef STATISTICS
201 1.1 cgd bt_pfxsaved += nbytes - n;
202 1.1 cgd #endif
203 1.1 cgd nbytes = n;
204 1.1 cgd } else
205 1.1 cgd nksize = 0;
206 1.1 cgd } else
207 1.1 cgd nksize = 0;
208 1.1 cgd break;
209 1.1 cgd case P_RINTERNAL:
210 1.1 cgd case P_RLEAF:
211 1.1 cgd nbytes = NRINTERNAL;
212 1.1 cgd break;
213 1.1 cgd default:
214 1.1 cgd abort();
215 1.1 cgd }
216 1.1 cgd
217 1.1 cgd /* Split the parent page if necessary or shift the indices. */
218 1.18 lukem if ((uint32_t)h->upper - (uint32_t)h->lower < nbytes + sizeof(indx_t)) {
219 1.1 cgd sp = h;
220 1.1 cgd h = h->pgno == P_ROOT ?
221 1.1 cgd bt_root(t, h, &l, &r, &skip, nbytes) :
222 1.1 cgd bt_page(t, h, &l, &r, &skip, nbytes);
223 1.1 cgd if (h == NULL)
224 1.1 cgd goto err1;
225 1.1 cgd parentsplit = 1;
226 1.1 cgd } else {
227 1.1 cgd if (skip < (nxtindex = NEXTINDEX(h)))
228 1.1 cgd memmove(h->linp + skip + 1, h->linp + skip,
229 1.1 cgd (nxtindex - skip) * sizeof(indx_t));
230 1.1 cgd h->lower += sizeof(indx_t);
231 1.1 cgd parentsplit = 0;
232 1.1 cgd }
233 1.1 cgd
234 1.1 cgd /* Insert the key into the parent page. */
235 1.6 cgd switch (rchild->flags & P_TYPE) {
236 1.1 cgd case P_BINTERNAL:
237 1.1 cgd h->linp[skip] = h->upper -= nbytes;
238 1.10 christos dest = (char *)(void *)h + h->linp[skip];
239 1.1 cgd memmove(dest, bi, nbytes);
240 1.10 christos ((BINTERNAL *)(void *)dest)->pgno = rchild->pgno;
241 1.1 cgd break;
242 1.1 cgd case P_BLEAF:
243 1.1 cgd h->linp[skip] = h->upper -= nbytes;
244 1.10 christos dest = (char *)(void *)h + h->linp[skip];
245 1.1 cgd WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
246 1.1 cgd rchild->pgno, bl->flags & P_BIGKEY);
247 1.1 cgd memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
248 1.1 cgd if (bl->flags & P_BIGKEY &&
249 1.10 christos bt_preserve(t, *(pgno_t *)(void *)bl->bytes) ==
250 1.10 christos RET_ERROR)
251 1.1 cgd goto err1;
252 1.1 cgd break;
253 1.1 cgd case P_RINTERNAL:
254 1.1 cgd /*
255 1.1 cgd * Update the left page count. If split
256 1.1 cgd * added at index 0, fix the correct page.
257 1.1 cgd */
258 1.1 cgd if (skip > 0)
259 1.10 christos dest = (char *)(void *)h + h->linp[skip - 1];
260 1.1 cgd else
261 1.10 christos dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
262 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild);
263 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
264 1.1 cgd
265 1.1 cgd /* Update the right page count. */
266 1.1 cgd h->linp[skip] = h->upper -= nbytes;
267 1.10 christos dest = (char *)(void *)h + h->linp[skip];
268 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild);
269 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
270 1.1 cgd break;
271 1.1 cgd case P_RLEAF:
272 1.1 cgd /*
273 1.1 cgd * Update the left page count. If split
274 1.1 cgd * added at index 0, fix the correct page.
275 1.1 cgd */
276 1.1 cgd if (skip > 0)
277 1.10 christos dest = (char *)(void *)h + h->linp[skip - 1];
278 1.1 cgd else
279 1.10 christos dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
280 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild);
281 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
282 1.1 cgd
283 1.1 cgd /* Update the right page count. */
284 1.1 cgd h->linp[skip] = h->upper -= nbytes;
285 1.10 christos dest = (char *)(void *)h + h->linp[skip];
286 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild);
287 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
288 1.1 cgd break;
289 1.1 cgd default:
290 1.1 cgd abort();
291 1.1 cgd }
292 1.1 cgd
293 1.1 cgd /* Unpin the held pages. */
294 1.1 cgd if (!parentsplit) {
295 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
296 1.1 cgd break;
297 1.1 cgd }
298 1.1 cgd
299 1.1 cgd /* If the root page was split, make it look right. */
300 1.1 cgd if (sp->pgno == P_ROOT &&
301 1.6 cgd (F_ISSET(t, R_RECNO) ?
302 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
303 1.1 cgd goto err1;
304 1.1 cgd
305 1.1 cgd mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
306 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
307 1.1 cgd }
308 1.1 cgd
309 1.1 cgd /* Unpin the held pages. */
310 1.1 cgd mpool_put(t->bt_mp, l, MPOOL_DIRTY);
311 1.1 cgd mpool_put(t->bt_mp, r, MPOOL_DIRTY);
312 1.1 cgd
313 1.1 cgd /* Clear any pages left on the stack. */
314 1.1 cgd return (RET_SUCCESS);
315 1.1 cgd
316 1.1 cgd /*
317 1.1 cgd * If something fails in the above loop we were already walking back
318 1.1 cgd * up the tree and the tree is now inconsistent. Nothing much we can
319 1.1 cgd * do about it but release any memory we're holding.
320 1.1 cgd */
321 1.1 cgd err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
322 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
323 1.1 cgd
324 1.1 cgd err2: mpool_put(t->bt_mp, l, 0);
325 1.1 cgd mpool_put(t->bt_mp, r, 0);
326 1.1 cgd __dbpanic(t->bt_dbp);
327 1.1 cgd return (RET_ERROR);
328 1.1 cgd }
329 1.1 cgd
330 1.1 cgd /*
331 1.1 cgd * BT_PAGE -- Split a non-root page of a btree.
332 1.1 cgd *
333 1.1 cgd * Parameters:
334 1.1 cgd * t: tree
335 1.1 cgd * h: root page
336 1.1 cgd * lp: pointer to left page pointer
337 1.1 cgd * rp: pointer to right page pointer
338 1.1 cgd * skip: pointer to index to leave open
339 1.1 cgd * ilen: insert length
340 1.1 cgd *
341 1.1 cgd * Returns:
342 1.1 cgd * Pointer to page in which to insert or NULL on error.
343 1.1 cgd */
344 1.1 cgd static PAGE *
345 1.14 christos bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
346 1.1 cgd {
347 1.1 cgd PAGE *l, *r, *tp;
348 1.1 cgd pgno_t npg;
349 1.1 cgd
350 1.1 cgd #ifdef STATISTICS
351 1.1 cgd ++bt_split;
352 1.1 cgd #endif
353 1.1 cgd /* Put the new right page for the split into place. */
354 1.1 cgd if ((r = __bt_new(t, &npg)) == NULL)
355 1.1 cgd return (NULL);
356 1.1 cgd r->pgno = npg;
357 1.1 cgd r->lower = BTDATAOFF;
358 1.1 cgd r->upper = t->bt_psize;
359 1.1 cgd r->nextpg = h->nextpg;
360 1.1 cgd r->prevpg = h->pgno;
361 1.1 cgd r->flags = h->flags & P_TYPE;
362 1.1 cgd
363 1.1 cgd /*
364 1.1 cgd * If we're splitting the last page on a level because we're appending
365 1.1 cgd * a key to it (skip is NEXTINDEX()), it's likely that the data is
366 1.1 cgd * sorted. Adding an empty page on the side of the level is less work
367 1.1 cgd * and can push the fill factor much higher than normal. If we're
368 1.1 cgd * wrong it's no big deal, we'll just do the split the right way next
369 1.1 cgd * time. It may look like it's equally easy to do a similar hack for
370 1.1 cgd * reverse sorted data, that is, split the tree left, but it's not.
371 1.1 cgd * Don't even try.
372 1.1 cgd */
373 1.1 cgd if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
374 1.1 cgd #ifdef STATISTICS
375 1.1 cgd ++bt_sortsplit;
376 1.1 cgd #endif
377 1.1 cgd h->nextpg = r->pgno;
378 1.1 cgd r->lower = BTDATAOFF + sizeof(indx_t);
379 1.1 cgd *skip = 0;
380 1.1 cgd *lp = h;
381 1.1 cgd *rp = r;
382 1.1 cgd return (r);
383 1.1 cgd }
384 1.1 cgd
385 1.1 cgd /* Put the new left page for the split into place. */
386 1.19 christos if ((l = calloc(1, t->bt_psize)) == NULL) {
387 1.1 cgd mpool_put(t->bt_mp, r, 0);
388 1.1 cgd return (NULL);
389 1.1 cgd }
390 1.6 cgd #ifdef PURIFY
391 1.6 cgd memset(l, 0xff, t->bt_psize);
392 1.6 cgd #endif
393 1.1 cgd l->pgno = h->pgno;
394 1.1 cgd l->nextpg = r->pgno;
395 1.1 cgd l->prevpg = h->prevpg;
396 1.1 cgd l->lower = BTDATAOFF;
397 1.1 cgd l->upper = t->bt_psize;
398 1.1 cgd l->flags = h->flags & P_TYPE;
399 1.1 cgd
400 1.1 cgd /* Fix up the previous pointer of the page after the split page. */
401 1.1 cgd if (h->nextpg != P_INVALID) {
402 1.1 cgd if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
403 1.1 cgd free(l);
404 1.1 cgd /* XXX mpool_free(t->bt_mp, r->pgno); */
405 1.1 cgd return (NULL);
406 1.1 cgd }
407 1.1 cgd tp->prevpg = r->pgno;
408 1.6 cgd mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
409 1.1 cgd }
410 1.1 cgd
411 1.1 cgd /*
412 1.1 cgd * Split right. The key/data pairs aren't sorted in the btree page so
413 1.1 cgd * it's simpler to copy the data from the split page onto two new pages
414 1.1 cgd * instead of copying half the data to the right page and compacting
415 1.1 cgd * the left page in place. Since the left page can't change, we have
416 1.1 cgd * to swap the original and the allocated left page after the split.
417 1.1 cgd */
418 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
419 1.1 cgd
420 1.1 cgd /* Move the new left page onto the old left page. */
421 1.1 cgd memmove(h, l, t->bt_psize);
422 1.1 cgd if (tp == l)
423 1.1 cgd tp = h;
424 1.1 cgd free(l);
425 1.1 cgd
426 1.1 cgd *lp = h;
427 1.1 cgd *rp = r;
428 1.1 cgd return (tp);
429 1.1 cgd }
430 1.1 cgd
431 1.1 cgd /*
432 1.1 cgd * BT_ROOT -- Split the root page of a btree.
433 1.1 cgd *
434 1.1 cgd * Parameters:
435 1.1 cgd * t: tree
436 1.1 cgd * h: root page
437 1.1 cgd * lp: pointer to left page pointer
438 1.1 cgd * rp: pointer to right page pointer
439 1.1 cgd * skip: pointer to index to leave open
440 1.1 cgd * ilen: insert length
441 1.1 cgd *
442 1.1 cgd * Returns:
443 1.1 cgd * Pointer to page in which to insert or NULL on error.
444 1.1 cgd */
445 1.1 cgd static PAGE *
446 1.14 christos bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
447 1.1 cgd {
448 1.1 cgd PAGE *l, *r, *tp;
449 1.1 cgd pgno_t lnpg, rnpg;
450 1.1 cgd
451 1.1 cgd #ifdef STATISTICS
452 1.1 cgd ++bt_split;
453 1.1 cgd ++bt_rootsplit;
454 1.1 cgd #endif
455 1.1 cgd /* Put the new left and right pages for the split into place. */
456 1.1 cgd if ((l = __bt_new(t, &lnpg)) == NULL ||
457 1.1 cgd (r = __bt_new(t, &rnpg)) == NULL)
458 1.1 cgd return (NULL);
459 1.1 cgd l->pgno = lnpg;
460 1.1 cgd r->pgno = rnpg;
461 1.1 cgd l->nextpg = r->pgno;
462 1.1 cgd r->prevpg = l->pgno;
463 1.1 cgd l->prevpg = r->nextpg = P_INVALID;
464 1.1 cgd l->lower = r->lower = BTDATAOFF;
465 1.1 cgd l->upper = r->upper = t->bt_psize;
466 1.1 cgd l->flags = r->flags = h->flags & P_TYPE;
467 1.1 cgd
468 1.1 cgd /* Split the root page. */
469 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
470 1.1 cgd
471 1.1 cgd *lp = l;
472 1.1 cgd *rp = r;
473 1.1 cgd return (tp);
474 1.1 cgd }
475 1.1 cgd
476 1.1 cgd /*
477 1.1 cgd * BT_RROOT -- Fix up the recno root page after it has been split.
478 1.1 cgd *
479 1.1 cgd * Parameters:
480 1.1 cgd * t: tree
481 1.1 cgd * h: root page
482 1.1 cgd * l: left page
483 1.1 cgd * r: right page
484 1.1 cgd *
485 1.1 cgd * Returns:
486 1.1 cgd * RET_ERROR, RET_SUCCESS
487 1.1 cgd */
488 1.1 cgd static int
489 1.14 christos bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
490 1.1 cgd {
491 1.1 cgd char *dest;
492 1.15 joerg uint32_t sz;
493 1.14 christos size_t temp;
494 1.14 christos
495 1.14 christos temp = t->bt_psize - NRINTERNAL;
496 1.15 joerg _DBFIT(temp, uint32_t);
497 1.15 joerg sz = (uint32_t)temp;
498 1.1 cgd
499 1.1 cgd /* Insert the left and right keys, set the header information. */
500 1.14 christos _DBFIT(sz, indx_t);
501 1.14 christos h->linp[0] = h->upper = (indx_t)sz;
502 1.10 christos dest = (char *)(void *)h + h->upper;
503 1.1 cgd WR_RINTERNAL(dest,
504 1.1 cgd l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
505 1.1 cgd
506 1.1 cgd h->linp[1] = h->upper -= NRINTERNAL;
507 1.10 christos dest = (char *)(void *)h + h->upper;
508 1.1 cgd WR_RINTERNAL(dest,
509 1.1 cgd r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
510 1.1 cgd
511 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
512 1.1 cgd
513 1.1 cgd /* Unpin the root page, set to recno internal page. */
514 1.1 cgd h->flags &= ~P_TYPE;
515 1.1 cgd h->flags |= P_RINTERNAL;
516 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
517 1.1 cgd
518 1.1 cgd return (RET_SUCCESS);
519 1.1 cgd }
520 1.1 cgd
521 1.1 cgd /*
522 1.1 cgd * BT_BROOT -- Fix up the btree root page after it has been split.
523 1.1 cgd *
524 1.1 cgd * Parameters:
525 1.1 cgd * t: tree
526 1.1 cgd * h: root page
527 1.1 cgd * l: left page
528 1.1 cgd * r: right page
529 1.1 cgd *
530 1.1 cgd * Returns:
531 1.1 cgd * RET_ERROR, RET_SUCCESS
532 1.1 cgd */
533 1.1 cgd static int
534 1.14 christos bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
535 1.1 cgd {
536 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
537 1.1 cgd BLEAF *bl;
538 1.15 joerg uint32_t nbytes;
539 1.1 cgd char *dest;
540 1.1 cgd
541 1.1 cgd /*
542 1.1 cgd * If the root page was a leaf page, change it into an internal page.
543 1.1 cgd * We copy the key we split on (but not the key's data, in the case of
544 1.1 cgd * a leaf page) to the new root page.
545 1.1 cgd *
546 1.1 cgd * The btree comparison code guarantees that the left-most key on any
547 1.1 cgd * level of the tree is never used, so it doesn't need to be filled in.
548 1.1 cgd */
549 1.1 cgd nbytes = NBINTERNAL(0);
550 1.1 cgd h->linp[0] = h->upper = t->bt_psize - nbytes;
551 1.10 christos dest = (char *)(void *)h + h->upper;
552 1.1 cgd WR_BINTERNAL(dest, 0, l->pgno, 0);
553 1.1 cgd
554 1.6 cgd switch (h->flags & P_TYPE) {
555 1.1 cgd case P_BLEAF:
556 1.1 cgd bl = GETBLEAF(r, 0);
557 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
558 1.1 cgd h->linp[1] = h->upper -= nbytes;
559 1.10 christos dest = (char *)(void *)h + h->upper;
560 1.1 cgd WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
561 1.1 cgd memmove(dest, bl->bytes, bl->ksize);
562 1.1 cgd
563 1.1 cgd /*
564 1.1 cgd * If the key is on an overflow page, mark the overflow chain
565 1.1 cgd * so it isn't deleted when the leaf copy of the key is deleted.
566 1.1 cgd */
567 1.1 cgd if (bl->flags & P_BIGKEY &&
568 1.10 christos bt_preserve(t, *(pgno_t *)(void *)bl->bytes) == RET_ERROR)
569 1.1 cgd return (RET_ERROR);
570 1.1 cgd break;
571 1.1 cgd case P_BINTERNAL:
572 1.1 cgd bi = GETBINTERNAL(r, 0);
573 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
574 1.1 cgd h->linp[1] = h->upper -= nbytes;
575 1.10 christos dest = (char *)(void *)h + h->upper;
576 1.1 cgd memmove(dest, bi, nbytes);
577 1.10 christos ((BINTERNAL *)(void *)dest)->pgno = r->pgno;
578 1.1 cgd break;
579 1.1 cgd default:
580 1.1 cgd abort();
581 1.1 cgd }
582 1.1 cgd
583 1.1 cgd /* There are two keys on the page. */
584 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
585 1.1 cgd
586 1.1 cgd /* Unpin the root page, set to btree internal page. */
587 1.1 cgd h->flags &= ~P_TYPE;
588 1.1 cgd h->flags |= P_BINTERNAL;
589 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
590 1.1 cgd
591 1.1 cgd return (RET_SUCCESS);
592 1.1 cgd }
593 1.1 cgd
594 1.1 cgd /*
595 1.1 cgd * BT_PSPLIT -- Do the real work of splitting the page.
596 1.1 cgd *
597 1.1 cgd * Parameters:
598 1.1 cgd * t: tree
599 1.1 cgd * h: page to be split
600 1.1 cgd * l: page to put lower half of data
601 1.1 cgd * r: page to put upper half of data
602 1.1 cgd * pskip: pointer to index to leave open
603 1.1 cgd * ilen: insert length
604 1.1 cgd *
605 1.1 cgd * Returns:
606 1.1 cgd * Pointer to page in which to insert.
607 1.1 cgd */
608 1.1 cgd static PAGE *
609 1.14 christos bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
610 1.1 cgd {
611 1.1 cgd BINTERNAL *bi;
612 1.1 cgd BLEAF *bl;
613 1.6 cgd CURSOR *c;
614 1.1 cgd RLEAF *rl;
615 1.1 cgd PAGE *rval;
616 1.7 christos void *src = NULL; /* pacify gcc */
617 1.1 cgd indx_t full, half, nxt, off, skip, top, used;
618 1.15 joerg uint32_t nbytes;
619 1.14 christos size_t temp;
620 1.1 cgd int bigkeycnt, isbigkey;
621 1.1 cgd
622 1.1 cgd /*
623 1.1 cgd * Split the data to the left and right pages. Leave the skip index
624 1.1 cgd * open. Additionally, make some effort not to split on an overflow
625 1.1 cgd * key. This makes internal page processing faster and can save
626 1.1 cgd * space as overflow keys used by internal pages are never deleted.
627 1.1 cgd */
628 1.1 cgd bigkeycnt = 0;
629 1.1 cgd skip = *pskip;
630 1.14 christos temp = t->bt_psize - BTDATAOFF;
631 1.14 christos _DBFIT(temp, indx_t);
632 1.14 christos full = (indx_t)temp;
633 1.1 cgd half = full / 2;
634 1.1 cgd used = 0;
635 1.1 cgd for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
636 1.1 cgd if (skip == off) {
637 1.15 joerg _DBFIT(ilen, uint32_t);
638 1.15 joerg nbytes = (uint32_t)ilen;
639 1.1 cgd isbigkey = 0; /* XXX: not really known. */
640 1.1 cgd } else
641 1.1 cgd switch (h->flags & P_TYPE) {
642 1.1 cgd case P_BINTERNAL:
643 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
644 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
645 1.1 cgd isbigkey = bi->flags & P_BIGKEY;
646 1.1 cgd break;
647 1.1 cgd case P_BLEAF:
648 1.1 cgd src = bl = GETBLEAF(h, nxt);
649 1.1 cgd nbytes = NBLEAF(bl);
650 1.1 cgd isbigkey = bl->flags & P_BIGKEY;
651 1.1 cgd break;
652 1.1 cgd case P_RINTERNAL:
653 1.1 cgd src = GETRINTERNAL(h, nxt);
654 1.1 cgd nbytes = NRINTERNAL;
655 1.1 cgd isbigkey = 0;
656 1.1 cgd break;
657 1.1 cgd case P_RLEAF:
658 1.1 cgd src = rl = GETRLEAF(h, nxt);
659 1.1 cgd nbytes = NRLEAF(rl);
660 1.1 cgd isbigkey = 0;
661 1.1 cgd break;
662 1.1 cgd default:
663 1.1 cgd abort();
664 1.1 cgd }
665 1.1 cgd
666 1.1 cgd /*
667 1.1 cgd * If the key/data pairs are substantial fractions of the max
668 1.1 cgd * possible size for the page, it's possible to get situations
669 1.1 cgd * where we decide to try and copy too much onto the left page.
670 1.1 cgd * Make sure that doesn't happen.
671 1.1 cgd */
672 1.9 is if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
673 1.9 is nxt == top - 1) {
674 1.1 cgd --off;
675 1.1 cgd break;
676 1.1 cgd }
677 1.1 cgd
678 1.1 cgd /* Copy the key/data pair, if not the skipped index. */
679 1.1 cgd if (skip != off) {
680 1.1 cgd ++nxt;
681 1.1 cgd
682 1.1 cgd l->linp[off] = l->upper -= nbytes;
683 1.10 christos memmove((char *)(void *)l + l->upper, src, nbytes);
684 1.1 cgd }
685 1.1 cgd
686 1.14 christos temp = nbytes + sizeof(indx_t);
687 1.14 christos _DBFIT(temp, indx_t);
688 1.14 christos used += (indx_t)temp;
689 1.1 cgd if (used >= half) {
690 1.1 cgd if (!isbigkey || bigkeycnt == 3)
691 1.1 cgd break;
692 1.1 cgd else
693 1.1 cgd ++bigkeycnt;
694 1.1 cgd }
695 1.1 cgd }
696 1.1 cgd
697 1.1 cgd /*
698 1.1 cgd * Off is the last offset that's valid for the left page.
699 1.1 cgd * Nxt is the first offset to be placed on the right page.
700 1.1 cgd */
701 1.14 christos temp = (off + 1) * sizeof(indx_t);
702 1.14 christos _DBFIT(temp, indx_t);
703 1.14 christos l->lower += (indx_t)temp;
704 1.1 cgd
705 1.1 cgd /*
706 1.1 cgd * If splitting the page that the cursor was on, the cursor has to be
707 1.1 cgd * adjusted to point to the same record as before the split. If the
708 1.1 cgd * cursor is at or past the skipped slot, the cursor is incremented by
709 1.1 cgd * one. If the cursor is on the right page, it is decremented by the
710 1.1 cgd * number of records split to the left page.
711 1.1 cgd */
712 1.6 cgd c = &t->bt_cursor;
713 1.6 cgd if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
714 1.6 cgd if (c->pg.index >= skip)
715 1.6 cgd ++c->pg.index;
716 1.6 cgd if (c->pg.index < nxt) /* Left page. */
717 1.6 cgd c->pg.pgno = l->pgno;
718 1.1 cgd else { /* Right page. */
719 1.6 cgd c->pg.pgno = r->pgno;
720 1.6 cgd c->pg.index -= nxt;
721 1.1 cgd }
722 1.1 cgd }
723 1.1 cgd
724 1.1 cgd /*
725 1.1 cgd * If the skipped index was on the left page, just return that page.
726 1.1 cgd * Otherwise, adjust the skip index to reflect the new position on
727 1.1 cgd * the right page.
728 1.1 cgd */
729 1.1 cgd if (skip <= off) {
730 1.12 mycroft skip = MAX_PAGE_OFFSET;
731 1.1 cgd rval = l;
732 1.1 cgd } else {
733 1.1 cgd rval = r;
734 1.1 cgd *pskip -= nxt;
735 1.1 cgd }
736 1.1 cgd
737 1.1 cgd for (off = 0; nxt < top; ++off) {
738 1.1 cgd if (skip == nxt) {
739 1.1 cgd ++off;
740 1.12 mycroft skip = MAX_PAGE_OFFSET;
741 1.1 cgd }
742 1.1 cgd switch (h->flags & P_TYPE) {
743 1.1 cgd case P_BINTERNAL:
744 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
745 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
746 1.1 cgd break;
747 1.1 cgd case P_BLEAF:
748 1.1 cgd src = bl = GETBLEAF(h, nxt);
749 1.1 cgd nbytes = NBLEAF(bl);
750 1.1 cgd break;
751 1.1 cgd case P_RINTERNAL:
752 1.1 cgd src = GETRINTERNAL(h, nxt);
753 1.1 cgd nbytes = NRINTERNAL;
754 1.1 cgd break;
755 1.1 cgd case P_RLEAF:
756 1.1 cgd src = rl = GETRLEAF(h, nxt);
757 1.1 cgd nbytes = NRLEAF(rl);
758 1.1 cgd break;
759 1.1 cgd default:
760 1.1 cgd abort();
761 1.1 cgd }
762 1.1 cgd ++nxt;
763 1.1 cgd r->linp[off] = r->upper -= nbytes;
764 1.10 christos memmove((char *)(void *)r + r->upper, src, nbytes);
765 1.1 cgd }
766 1.14 christos temp = off * sizeof(indx_t);
767 1.14 christos _DBFIT(temp, indx_t);
768 1.14 christos r->lower += (indx_t)temp;
769 1.1 cgd
770 1.1 cgd /* If the key is being appended to the page, adjust the index. */
771 1.1 cgd if (skip == top)
772 1.1 cgd r->lower += sizeof(indx_t);
773 1.1 cgd
774 1.1 cgd return (rval);
775 1.1 cgd }
776 1.1 cgd
777 1.1 cgd /*
778 1.1 cgd * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
779 1.1 cgd *
780 1.1 cgd * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
781 1.1 cgd * record that references them gets deleted. Chains pointed to by internal
782 1.1 cgd * pages never get deleted. This routine marks a chain as pointed to by an
783 1.1 cgd * internal page.
784 1.1 cgd *
785 1.1 cgd * Parameters:
786 1.1 cgd * t: tree
787 1.1 cgd * pg: page number of first page in the chain.
788 1.1 cgd *
789 1.1 cgd * Returns:
790 1.1 cgd * RET_SUCCESS, RET_ERROR.
791 1.1 cgd */
792 1.1 cgd static int
793 1.14 christos bt_preserve(BTREE *t, pgno_t pg)
794 1.1 cgd {
795 1.1 cgd PAGE *h;
796 1.1 cgd
797 1.1 cgd if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
798 1.1 cgd return (RET_ERROR);
799 1.1 cgd h->flags |= P_PRESERVE;
800 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
801 1.1 cgd return (RET_SUCCESS);
802 1.1 cgd }
803 1.1 cgd
804 1.1 cgd /*
805 1.1 cgd * REC_TOTAL -- Return the number of recno entries below a page.
806 1.1 cgd *
807 1.1 cgd * Parameters:
808 1.1 cgd * h: page
809 1.1 cgd *
810 1.1 cgd * Returns:
811 1.1 cgd * The number of recno entries below a page.
812 1.1 cgd *
813 1.1 cgd * XXX
814 1.1 cgd * These values could be set by the bt_psplit routine. The problem is that the
815 1.1 cgd * entry has to be popped off of the stack etc. or the values have to be passed
816 1.1 cgd * all the way back to bt_split/bt_rroot and it's not very clean.
817 1.1 cgd */
818 1.1 cgd static recno_t
819 1.14 christos rec_total(PAGE *h)
820 1.1 cgd {
821 1.1 cgd recno_t recs;
822 1.1 cgd indx_t nxt, top;
823 1.1 cgd
824 1.1 cgd for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
825 1.1 cgd recs += GETRINTERNAL(h, nxt)->nrecs;
826 1.1 cgd return (recs);
827 1.1 cgd }
828