Home | History | Annotate | Line # | Download | only in hash
hash_page.c revision 1.12
      1 /*	$NetBSD: hash_page.c,v 1.12 1998/10/14 12:06:49 kleink Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1990, 1993, 1994
      5  *	The Regents of the University of California.  All rights reserved.
      6  *
      7  * This code is derived from software contributed to Berkeley by
      8  * Margo Seltzer.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  * 3. All advertising materials mentioning features or use of this software
     19  *    must display the following acknowledgement:
     20  *	This product includes software developed by the University of
     21  *	California, Berkeley and its contributors.
     22  * 4. Neither the name of the University nor the names of its contributors
     23  *    may be used to endorse or promote products derived from this software
     24  *    without specific prior written permission.
     25  *
     26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     36  * SUCH DAMAGE.
     37  */
     38 
     39 #include <sys/cdefs.h>
     40 #if defined(LIBC_SCCS) && !defined(lint)
     41 #if 0
     42 static char sccsid[] = "@(#)hash_page.c	8.7 (Berkeley) 8/16/94";
     43 #else
     44 __RCSID("$NetBSD: hash_page.c,v 1.12 1998/10/14 12:06:49 kleink Exp $");
     45 #endif
     46 #endif /* LIBC_SCCS and not lint */
     47 
     48 /*
     49  * PACKAGE:  hashing
     50  *
     51  * DESCRIPTION:
     52  *	Page manipulation for hashing package.
     53  *
     54  * ROUTINES:
     55  *
     56  * External
     57  *	__get_page
     58  *	__add_ovflpage
     59  * Internal
     60  *	overflow_page
     61  *	open_temp
     62  */
     63 
     64 #include "namespace.h"
     65 
     66 #include <sys/types.h>
     67 
     68 #include <errno.h>
     69 #include <fcntl.h>
     70 #include <signal.h>
     71 #include <stdio.h>
     72 #include <stdlib.h>
     73 #include <string.h>
     74 #include <unistd.h>
     75 #ifdef DEBUG
     76 #include <assert.h>
     77 #endif
     78 
     79 #include <db.h>
     80 #include "hash.h"
     81 #include "page.h"
     82 #include "extern.h"
     83 
     84 static u_int32_t	*fetch_bitmap __P((HTAB *, int));
     85 static u_int32_t	 first_free __P((u_int32_t));
     86 static int	 open_temp __P((HTAB *));
     87 static u_int16_t	 overflow_page __P((HTAB *));
     88 static void	 putpair __P((char *, const DBT *, const DBT *));
     89 static void	 squeeze_key __P((u_int16_t *, const DBT *, const DBT *));
     90 static int	 ugly_split
     91 		    __P((HTAB *, u_int32_t, BUFHEAD *, BUFHEAD *, int, int));
     92 
     93 #define	PAGE_INIT(P) { \
     94 	((u_int16_t *)(P))[0] = 0; \
     95 	((u_int16_t *)(P))[1] = hashp->BSIZE - 3 * sizeof(u_int16_t); \
     96 	((u_int16_t *)(P))[2] = hashp->BSIZE; \
     97 }
     98 
     99 /*
    100  * This is called AFTER we have verified that there is room on the page for
    101  * the pair (PAIRFITS has returned true) so we go right ahead and start moving
    102  * stuff on.
    103  */
    104 static void
    105 putpair(p, key, val)
    106 	char *p;
    107 	const DBT *key, *val;
    108 {
    109 	register u_int16_t *bp, n, off;
    110 
    111 	bp = (u_int16_t *)p;
    112 
    113 	/* Enter the key first. */
    114 	n = bp[0];
    115 
    116 	off = OFFSET(bp) - key->size;
    117 	memmove(p + off, key->data, key->size);
    118 	bp[++n] = off;
    119 
    120 	/* Now the data. */
    121 	off -= val->size;
    122 	memmove(p + off, val->data, val->size);
    123 	bp[++n] = off;
    124 
    125 	/* Adjust page info. */
    126 	bp[0] = n;
    127 	bp[n + 1] = off - ((n + 3) * sizeof(u_int16_t));
    128 	bp[n + 2] = off;
    129 }
    130 
    131 /*
    132  * Returns:
    133  *	 0 OK
    134  *	-1 error
    135  */
    136 extern int
    137 __delpair(hashp, bufp, ndx)
    138 	HTAB *hashp;
    139 	BUFHEAD *bufp;
    140 	register int ndx;
    141 {
    142 	register u_int16_t *bp, newoff;
    143 	register int n;
    144 	u_int16_t pairlen;
    145 
    146 	bp = (u_int16_t *)bufp->page;
    147 	n = bp[0];
    148 
    149 	if (bp[ndx + 1] < REAL_KEY)
    150 		return (__big_delete(hashp, bufp));
    151 	if (ndx != 1)
    152 		newoff = bp[ndx - 1];
    153 	else
    154 		newoff = hashp->BSIZE;
    155 	pairlen = newoff - bp[ndx + 1];
    156 
    157 	if (ndx != (n - 1)) {
    158 		/* Hard Case -- need to shuffle keys */
    159 		register int i;
    160 		register char *src = bufp->page + (int)OFFSET(bp);
    161 		register char *dst = src + (int)pairlen;
    162 		memmove(dst, src, bp[ndx + 1] - OFFSET(bp));
    163 
    164 		/* Now adjust the pointers */
    165 		for (i = ndx + 2; i <= n; i += 2) {
    166 			if (bp[i + 1] == OVFLPAGE) {
    167 				bp[i - 2] = bp[i];
    168 				bp[i - 1] = bp[i + 1];
    169 			} else {
    170 				bp[i - 2] = bp[i] + pairlen;
    171 				bp[i - 1] = bp[i + 1] + pairlen;
    172 			}
    173 		}
    174 	}
    175 	/* Finally adjust the page data */
    176 	bp[n] = OFFSET(bp) + pairlen;
    177 	bp[n - 1] = bp[n + 1] + pairlen + 2 * sizeof(u_int16_t);
    178 	bp[0] = n - 2;
    179 	hashp->NKEYS--;
    180 
    181 	bufp->flags |= BUF_MOD;
    182 	return (0);
    183 }
    184 /*
    185  * Returns:
    186  *	 0 ==> OK
    187  *	-1 ==> Error
    188  */
    189 extern int
    190 __split_page(hashp, obucket, nbucket)
    191 	HTAB *hashp;
    192 	u_int32_t obucket, nbucket;
    193 {
    194 	register BUFHEAD *new_bufp, *old_bufp;
    195 	register u_int16_t *ino;
    196 	register char *np;
    197 	DBT key, val;
    198 	int n, ndx, retval;
    199 	u_int16_t copyto, diff, off, moved;
    200 	char *op;
    201 
    202 	copyto = (u_int16_t)hashp->BSIZE;
    203 	off = (u_int16_t)hashp->BSIZE;
    204 	old_bufp = __get_buf(hashp, obucket, NULL, 0);
    205 	if (old_bufp == NULL)
    206 		return (-1);
    207 	new_bufp = __get_buf(hashp, nbucket, NULL, 0);
    208 	if (new_bufp == NULL)
    209 		return (-1);
    210 
    211 	old_bufp->flags |= (BUF_MOD | BUF_PIN);
    212 	new_bufp->flags |= (BUF_MOD | BUF_PIN);
    213 
    214 	ino = (u_int16_t *)(op = old_bufp->page);
    215 	np = new_bufp->page;
    216 
    217 	moved = 0;
    218 
    219 	for (n = 1, ndx = 1; n < ino[0]; n += 2) {
    220 		if (ino[n + 1] < REAL_KEY) {
    221 			retval = ugly_split(hashp, obucket, old_bufp, new_bufp,
    222 			    (int)copyto, (int)moved);
    223 			old_bufp->flags &= ~BUF_PIN;
    224 			new_bufp->flags &= ~BUF_PIN;
    225 			return (retval);
    226 
    227 		}
    228 		key.data = (u_char *)op + ino[n];
    229 		key.size = off - ino[n];
    230 
    231 		if (__call_hash(hashp, key.data, key.size) == obucket) {
    232 			/* Don't switch page */
    233 			diff = copyto - off;
    234 			if (diff) {
    235 				copyto = ino[n + 1] + diff;
    236 				memmove(op + copyto, op + ino[n + 1],
    237 				    off - ino[n + 1]);
    238 				ino[ndx] = copyto + ino[n] - ino[n + 1];
    239 				ino[ndx + 1] = copyto;
    240 			} else
    241 				copyto = ino[n + 1];
    242 			ndx += 2;
    243 		} else {
    244 			/* Switch page */
    245 			val.data = (u_char *)op + ino[n + 1];
    246 			val.size = ino[n] - ino[n + 1];
    247 			putpair(np, &key, &val);
    248 			moved += 2;
    249 		}
    250 
    251 		off = ino[n + 1];
    252 	}
    253 
    254 	/* Now clean up the page */
    255 	ino[0] -= moved;
    256 	FREESPACE(ino) = copyto - sizeof(u_int16_t) * (ino[0] + 3);
    257 	OFFSET(ino) = copyto;
    258 
    259 #ifdef DEBUG3
    260 	(void)fprintf(stderr, "split %d/%d\n",
    261 	    ((u_int16_t *)np)[0] / 2,
    262 	    ((u_int16_t *)op)[0] / 2);
    263 #endif
    264 	/* unpin both pages */
    265 	old_bufp->flags &= ~BUF_PIN;
    266 	new_bufp->flags &= ~BUF_PIN;
    267 	return (0);
    268 }
    269 
    270 /*
    271  * Called when we encounter an overflow or big key/data page during split
    272  * handling.  This is special cased since we have to begin checking whether
    273  * the key/data pairs fit on their respective pages and because we may need
    274  * overflow pages for both the old and new pages.
    275  *
    276  * The first page might be a page with regular key/data pairs in which case
    277  * we have a regular overflow condition and just need to go on to the next
    278  * page or it might be a big key/data pair in which case we need to fix the
    279  * big key/data pair.
    280  *
    281  * Returns:
    282  *	 0 ==> success
    283  *	-1 ==> failure
    284  */
    285 static int
    286 ugly_split(hashp, obucket, old_bufp, new_bufp, copyto, moved)
    287 	HTAB *hashp;
    288 	u_int32_t obucket;	/* Same as __split_page. */
    289 	BUFHEAD *old_bufp, *new_bufp;
    290 	int copyto;	/* First byte on page which contains key/data values. */
    291 	int moved;	/* Number of pairs moved to new page. */
    292 {
    293 	register BUFHEAD *bufp;	/* Buffer header for ino */
    294 	register u_int16_t *ino;	/* Page keys come off of */
    295 	register u_int16_t *np;	/* New page */
    296 	register u_int16_t *op;	/* Page keys go on to if they aren't moving */
    297 
    298 	BUFHEAD *last_bfp;	/* Last buf header OVFL needing to be freed */
    299 	DBT key, val;
    300 	SPLIT_RETURN ret;
    301 	u_int16_t n, off, ov_addr, scopyto;
    302 	char *cino;		/* Character value of ino */
    303 
    304 	bufp = old_bufp;
    305 	ino = (u_int16_t *)old_bufp->page;
    306 	np = (u_int16_t *)new_bufp->page;
    307 	op = (u_int16_t *)old_bufp->page;
    308 	last_bfp = NULL;
    309 	scopyto = (u_int16_t)copyto;	/* ANSI */
    310 
    311 	n = ino[0] - 1;
    312 	while (n < ino[0]) {
    313 		if (ino[2] < REAL_KEY && ino[2] != OVFLPAGE) {
    314 			if (__big_split(hashp, old_bufp,
    315 			    new_bufp, bufp, bufp->addr, obucket, &ret))
    316 				return (-1);
    317 			old_bufp = ret.oldp;
    318 			if (!old_bufp)
    319 				return (-1);
    320 			op = (u_int16_t *)old_bufp->page;
    321 			new_bufp = ret.newp;
    322 			if (!new_bufp)
    323 				return (-1);
    324 			np = (u_int16_t *)new_bufp->page;
    325 			bufp = ret.nextp;
    326 			if (!bufp)
    327 				return (0);
    328 			cino = (char *)bufp->page;
    329 			ino = (u_int16_t *)cino;
    330 			last_bfp = ret.nextp;
    331 		} else if (ino[n + 1] == OVFLPAGE) {
    332 			ov_addr = ino[n];
    333 			/*
    334 			 * Fix up the old page -- the extra 2 are the fields
    335 			 * which contained the overflow information.
    336 			 */
    337 			ino[0] -= (moved + 2);
    338 			FREESPACE(ino) =
    339 			    scopyto - sizeof(u_int16_t) * (ino[0] + 3);
    340 			OFFSET(ino) = scopyto;
    341 
    342 			bufp = __get_buf(hashp, ov_addr, bufp, 0);
    343 			if (!bufp)
    344 				return (-1);
    345 
    346 			ino = (u_int16_t *)bufp->page;
    347 			n = 1;
    348 			scopyto = hashp->BSIZE;
    349 			moved = 0;
    350 
    351 			if (last_bfp)
    352 				__free_ovflpage(hashp, last_bfp);
    353 			last_bfp = bufp;
    354 		}
    355 		/* Move regular sized pairs of there are any */
    356 		off = hashp->BSIZE;
    357 		for (n = 1; (n < ino[0]) && (ino[n + 1] >= REAL_KEY); n += 2) {
    358 			cino = (char *)ino;
    359 			key.data = (u_char *)cino + ino[n];
    360 			key.size = off - ino[n];
    361 			val.data = (u_char *)cino + ino[n + 1];
    362 			val.size = ino[n] - ino[n + 1];
    363 			off = ino[n + 1];
    364 
    365 			if (__call_hash(hashp, key.data, key.size) == obucket) {
    366 				/* Keep on old page */
    367 				if (PAIRFITS(op, (&key), (&val)))
    368 					putpair((char *)op, &key, &val);
    369 				else {
    370 					old_bufp =
    371 					    __add_ovflpage(hashp, old_bufp);
    372 					if (!old_bufp)
    373 						return (-1);
    374 					op = (u_int16_t *)old_bufp->page;
    375 					putpair((char *)op, &key, &val);
    376 				}
    377 				old_bufp->flags |= BUF_MOD;
    378 			} else {
    379 				/* Move to new page */
    380 				if (PAIRFITS(np, (&key), (&val)))
    381 					putpair((char *)np, &key, &val);
    382 				else {
    383 					new_bufp =
    384 					    __add_ovflpage(hashp, new_bufp);
    385 					if (!new_bufp)
    386 						return (-1);
    387 					np = (u_int16_t *)new_bufp->page;
    388 					putpair((char *)np, &key, &val);
    389 				}
    390 				new_bufp->flags |= BUF_MOD;
    391 			}
    392 		}
    393 	}
    394 	if (last_bfp)
    395 		__free_ovflpage(hashp, last_bfp);
    396 	return (0);
    397 }
    398 
    399 /*
    400  * Add the given pair to the page
    401  *
    402  * Returns:
    403  *	0 ==> OK
    404  *	1 ==> failure
    405  */
    406 extern int
    407 __addel(hashp, bufp, key, val)
    408 	HTAB *hashp;
    409 	BUFHEAD *bufp;
    410 	const DBT *key, *val;
    411 {
    412 	register u_int16_t *bp, *sop;
    413 	int do_expand;
    414 
    415 	bp = (u_int16_t *)bufp->page;
    416 	do_expand = 0;
    417 	while (bp[0] && (bp[2] < REAL_KEY || bp[bp[0]] < REAL_KEY))
    418 		/* Exception case */
    419 		if (bp[2] == FULL_KEY_DATA && bp[0] == 2)
    420 			/* This is the last page of a big key/data pair
    421 			   and we need to add another page */
    422 			break;
    423 		else if (bp[2] < REAL_KEY && bp[bp[0]] != OVFLPAGE) {
    424 			bufp = __get_buf(hashp, bp[bp[0] - 1], bufp, 0);
    425 			if (!bufp)
    426 				return (-1);
    427 			bp = (u_int16_t *)bufp->page;
    428 		} else
    429 			/* Try to squeeze key on this page */
    430 			if (FREESPACE(bp) > PAIRSIZE(key, val)) {
    431 				squeeze_key(bp, key, val);
    432 				return (0);
    433 			} else {
    434 				bufp = __get_buf(hashp, bp[bp[0] - 1], bufp, 0);
    435 				if (!bufp)
    436 					return (-1);
    437 				bp = (u_int16_t *)bufp->page;
    438 			}
    439 
    440 	if (PAIRFITS(bp, key, val))
    441 		putpair(bufp->page, key, val);
    442 	else {
    443 		do_expand = 1;
    444 		bufp = __add_ovflpage(hashp, bufp);
    445 		if (!bufp)
    446 			return (-1);
    447 		sop = (u_int16_t *)bufp->page;
    448 
    449 		if (PAIRFITS(sop, key, val))
    450 			putpair((char *)sop, key, val);
    451 		else
    452 			if (__big_insert(hashp, bufp, key, val))
    453 				return (-1);
    454 	}
    455 	bufp->flags |= BUF_MOD;
    456 	/*
    457 	 * If the average number of keys per bucket exceeds the fill factor,
    458 	 * expand the table.
    459 	 */
    460 	hashp->NKEYS++;
    461 	if (do_expand ||
    462 	    (hashp->NKEYS / (hashp->MAX_BUCKET + 1) > hashp->FFACTOR))
    463 		return (__expand_table(hashp));
    464 	return (0);
    465 }
    466 
    467 /*
    468  *
    469  * Returns:
    470  *	pointer on success
    471  *	NULL on error
    472  */
    473 extern BUFHEAD *
    474 __add_ovflpage(hashp, bufp)
    475 	HTAB *hashp;
    476 	BUFHEAD *bufp;
    477 {
    478 	register u_int16_t *sp;
    479 	u_int16_t ndx, ovfl_num;
    480 #ifdef DEBUG1
    481 	int tmp1, tmp2;
    482 #endif
    483 	sp = (u_int16_t *)bufp->page;
    484 
    485 	/* Check if we are dynamically determining the fill factor */
    486 	if (hashp->FFACTOR == DEF_FFACTOR) {
    487 		hashp->FFACTOR = sp[0] >> 1;
    488 		if (hashp->FFACTOR < MIN_FFACTOR)
    489 			hashp->FFACTOR = MIN_FFACTOR;
    490 	}
    491 	bufp->flags |= BUF_MOD;
    492 	ovfl_num = overflow_page(hashp);
    493 #ifdef DEBUG1
    494 	tmp1 = bufp->addr;
    495 	tmp2 = bufp->ovfl ? bufp->ovfl->addr : 0;
    496 #endif
    497 	if (!ovfl_num || !(bufp->ovfl = __get_buf(hashp, ovfl_num, bufp, 1)))
    498 		return (NULL);
    499 	bufp->ovfl->flags |= BUF_MOD;
    500 #ifdef DEBUG1
    501 	(void)fprintf(stderr, "ADDOVFLPAGE: %d->ovfl was %d is now %d\n",
    502 	    tmp1, tmp2, bufp->ovfl->addr);
    503 #endif
    504 	ndx = sp[0];
    505 	/*
    506 	 * Since a pair is allocated on a page only if there's room to add
    507 	 * an overflow page, we know that the OVFL information will fit on
    508 	 * the page.
    509 	 */
    510 	sp[ndx + 4] = OFFSET(sp);
    511 	sp[ndx + 3] = FREESPACE(sp) - OVFLSIZE;
    512 	sp[ndx + 1] = ovfl_num;
    513 	sp[ndx + 2] = OVFLPAGE;
    514 	sp[0] = ndx + 2;
    515 #ifdef HASH_STATISTICS
    516 	hash_overflows++;
    517 #endif
    518 	return (bufp->ovfl);
    519 }
    520 
    521 /*
    522  * Returns:
    523  *	 0 indicates SUCCESS
    524  *	-1 indicates FAILURE
    525  */
    526 extern int
    527 __get_page(hashp, p, bucket, is_bucket, is_disk, is_bitmap)
    528 	HTAB *hashp;
    529 	char *p;
    530 	u_int32_t bucket;
    531 	int is_bucket, is_disk, is_bitmap;
    532 {
    533 	register int fd, page, size;
    534 	int rsize;
    535 	u_int16_t *bp;
    536 
    537 	fd = hashp->fp;
    538 	size = hashp->BSIZE;
    539 
    540 	if ((fd == -1) || !is_disk) {
    541 		PAGE_INIT(p);
    542 		return (0);
    543 	}
    544 	if (is_bucket)
    545 		page = BUCKET_TO_PAGE(bucket);
    546 	else
    547 		page = OADDR_TO_PAGE(bucket);
    548 	if ((rsize = pread(fd, p, size, (off_t)page << hashp->BSHIFT)) == -1)
    549 		return (-1);
    550 	bp = (u_int16_t *)p;
    551 	if (!rsize)
    552 		bp[0] = 0;	/* We hit the EOF, so initialize a new page */
    553 	else
    554 		if (rsize != size) {
    555 			errno = EFTYPE;
    556 			return (-1);
    557 		}
    558 	if (!is_bitmap && !bp[0]) {
    559 		PAGE_INIT(p);
    560 	} else
    561 		if (hashp->LORDER != BYTE_ORDER) {
    562 			register int i, max;
    563 
    564 			if (is_bitmap) {
    565 				max = hashp->BSIZE >> 2; /* divide by 4 */
    566 				for (i = 0; i < max; i++)
    567 					M_32_SWAP(((int *)p)[i]);
    568 			} else {
    569 				M_16_SWAP(bp[0]);
    570 				max = bp[0] + 2;
    571 				for (i = 1; i <= max; i++)
    572 					M_16_SWAP(bp[i]);
    573 			}
    574 		}
    575 	return (0);
    576 }
    577 
    578 /*
    579  * Write page p to disk
    580  *
    581  * Returns:
    582  *	 0 ==> OK
    583  *	-1 ==>failure
    584  */
    585 extern int
    586 __put_page(hashp, p, bucket, is_bucket, is_bitmap)
    587 	HTAB *hashp;
    588 	char *p;
    589 	u_int32_t bucket;
    590 	int is_bucket, is_bitmap;
    591 {
    592 	register int fd, page, size;
    593 	int wsize;
    594 
    595 	size = hashp->BSIZE;
    596 	if ((hashp->fp == -1) && open_temp(hashp))
    597 		return (-1);
    598 	fd = hashp->fp;
    599 
    600 	if (hashp->LORDER != BYTE_ORDER) {
    601 		register int i;
    602 		register int max;
    603 
    604 		if (is_bitmap) {
    605 			max = hashp->BSIZE >> 2;	/* divide by 4 */
    606 			for (i = 0; i < max; i++)
    607 				M_32_SWAP(((int *)p)[i]);
    608 		} else {
    609 			max = ((u_int16_t *)p)[0] + 2;
    610 			for (i = 0; i <= max; i++)
    611 				M_16_SWAP(((u_int16_t *)p)[i]);
    612 		}
    613 	}
    614 	if (is_bucket)
    615 		page = BUCKET_TO_PAGE(bucket);
    616 	else
    617 		page = OADDR_TO_PAGE(bucket);
    618 	if ((wsize = pwrite(fd, p, size, (off_t)page << hashp->BSHIFT)) == -1)
    619 		/* Errno is set */
    620 		return (-1);
    621 	if (wsize != size) {
    622 		errno = EFTYPE;
    623 		return (-1);
    624 	}
    625 	return (0);
    626 }
    627 
    628 #define BYTE_MASK	((1 << INT_BYTE_SHIFT) -1)
    629 /*
    630  * Initialize a new bitmap page.  Bitmap pages are left in memory
    631  * once they are read in.
    632  */
    633 extern int
    634 __ibitmap(hashp, pnum, nbits, ndx)
    635 	HTAB *hashp;
    636 	int pnum, nbits, ndx;
    637 {
    638 	u_int32_t *ip;
    639 	int clearbytes, clearints;
    640 
    641 	if ((ip = (u_int32_t *)malloc(hashp->BSIZE)) == NULL)
    642 		return (1);
    643 	hashp->nmaps++;
    644 	clearints = ((nbits - 1) >> INT_BYTE_SHIFT) + 1;
    645 	clearbytes = clearints << INT_TO_BYTE;
    646 	(void)memset((char *)ip, 0, clearbytes);
    647 	(void)memset(((char *)ip) + clearbytes, 0xFF,
    648 	    hashp->BSIZE - clearbytes);
    649 	ip[clearints - 1] = ALL_SET << (nbits & BYTE_MASK);
    650 	SETBIT(ip, 0);
    651 	hashp->BITMAPS[ndx] = (u_int16_t)pnum;
    652 	hashp->mapp[ndx] = ip;
    653 	return (0);
    654 }
    655 
    656 static u_int32_t
    657 first_free(map)
    658 	u_int32_t map;
    659 {
    660 	register u_int32_t i, mask;
    661 
    662 	mask = 0x1;
    663 	for (i = 0; i < BITS_PER_MAP; i++) {
    664 		if (!(mask & map))
    665 			return (i);
    666 		mask = mask << 1;
    667 	}
    668 	return (i);
    669 }
    670 
    671 static u_int16_t
    672 overflow_page(hashp)
    673 	HTAB *hashp;
    674 {
    675 	register u_int32_t *freep = NULL;
    676 	register int max_free, offset, splitnum;
    677 	u_int16_t addr;
    678 	int bit, first_page, free_bit, free_page, i, in_use_bits, j;
    679 #ifdef DEBUG2
    680 	int tmp1, tmp2;
    681 #endif
    682 	splitnum = hashp->OVFL_POINT;
    683 	max_free = hashp->SPARES[splitnum];
    684 
    685 	free_page = (max_free - 1) >> (hashp->BSHIFT + BYTE_SHIFT);
    686 	free_bit = (max_free - 1) & ((hashp->BSIZE << BYTE_SHIFT) - 1);
    687 
    688 	/* Look through all the free maps to find the first free block */
    689 	first_page = hashp->LAST_FREED >>(hashp->BSHIFT + BYTE_SHIFT);
    690 	for ( i = first_page; i <= free_page; i++ ) {
    691 		if (!(freep = (u_int32_t *)hashp->mapp[i]) &&
    692 		    !(freep = fetch_bitmap(hashp, i)))
    693 			return (0);
    694 		if (i == free_page)
    695 			in_use_bits = free_bit;
    696 		else
    697 			in_use_bits = (hashp->BSIZE << BYTE_SHIFT) - 1;
    698 
    699 		if (i == first_page) {
    700 			bit = hashp->LAST_FREED &
    701 			    ((hashp->BSIZE << BYTE_SHIFT) - 1);
    702 			j = bit / BITS_PER_MAP;
    703 			bit = bit & ~(BITS_PER_MAP - 1);
    704 		} else {
    705 			bit = 0;
    706 			j = 0;
    707 		}
    708 		for (; bit <= in_use_bits; j++, bit += BITS_PER_MAP)
    709 			if (freep[j] != ALL_SET)
    710 				goto found;
    711 	}
    712 
    713 	/* No Free Page Found */
    714 	hashp->LAST_FREED = hashp->SPARES[splitnum];
    715 	hashp->SPARES[splitnum]++;
    716 	offset = hashp->SPARES[splitnum] -
    717 	    (splitnum ? hashp->SPARES[splitnum - 1] : 0);
    718 
    719 #define	OVMSG	"HASH: Out of overflow pages.  Increase page size\n"
    720 	if (offset > SPLITMASK) {
    721 		if (++splitnum >= NCACHED) {
    722 			(void)write(STDERR_FILENO, OVMSG, sizeof(OVMSG) - 1);
    723 			return (0);
    724 		}
    725 		hashp->OVFL_POINT = splitnum;
    726 		hashp->SPARES[splitnum] = hashp->SPARES[splitnum-1];
    727 		hashp->SPARES[splitnum-1]--;
    728 		offset = 1;
    729 	}
    730 
    731 	/* Check if we need to allocate a new bitmap page */
    732 	if (free_bit == (hashp->BSIZE << BYTE_SHIFT) - 1) {
    733 		free_page++;
    734 		if (free_page >= NCACHED) {
    735 			(void)write(STDERR_FILENO, OVMSG, sizeof(OVMSG) - 1);
    736 			return (0);
    737 		}
    738 		/*
    739 		 * This is tricky.  The 1 indicates that you want the new page
    740 		 * allocated with 1 clear bit.  Actually, you are going to
    741 		 * allocate 2 pages from this map.  The first is going to be
    742 		 * the map page, the second is the overflow page we were
    743 		 * looking for.  The init_bitmap routine automatically, sets
    744 		 * the first bit of itself to indicate that the bitmap itself
    745 		 * is in use.  We would explicitly set the second bit, but
    746 		 * don't have to if we tell init_bitmap not to leave it clear
    747 		 * in the first place.
    748 		 */
    749 		if (__ibitmap(hashp,
    750 		    (int)OADDR_OF(splitnum, offset), 1, free_page))
    751 			return (0);
    752 		hashp->SPARES[splitnum]++;
    753 #ifdef DEBUG2
    754 		free_bit = 2;
    755 #endif
    756 		offset++;
    757 		if (offset > SPLITMASK) {
    758 			if (++splitnum >= NCACHED) {
    759 				(void)write(STDERR_FILENO, OVMSG,
    760 				    sizeof(OVMSG) - 1);
    761 				return (0);
    762 			}
    763 			hashp->OVFL_POINT = splitnum;
    764 			hashp->SPARES[splitnum] = hashp->SPARES[splitnum-1];
    765 			hashp->SPARES[splitnum-1]--;
    766 			offset = 0;
    767 		}
    768 	} else {
    769 		/*
    770 		 * Free_bit addresses the last used bit.  Bump it to address
    771 		 * the first available bit.
    772 		 */
    773 		free_bit++;
    774 		SETBIT(freep, free_bit);
    775 	}
    776 
    777 	/* Calculate address of the new overflow page */
    778 	addr = OADDR_OF(splitnum, offset);
    779 #ifdef DEBUG2
    780 	(void)fprintf(stderr, "OVERFLOW_PAGE: ADDR: %d BIT: %d PAGE %d\n",
    781 	    addr, free_bit, free_page);
    782 #endif
    783 	return (addr);
    784 
    785 found:
    786 	bit = bit + first_free(freep[j]);
    787 	SETBIT(freep, bit);
    788 #ifdef DEBUG2
    789 	tmp1 = bit;
    790 	tmp2 = i;
    791 #endif
    792 	/*
    793 	 * Bits are addressed starting with 0, but overflow pages are addressed
    794 	 * beginning at 1. Bit is a bit addressnumber, so we need to increment
    795 	 * it to convert it to a page number.
    796 	 */
    797 	bit = 1 + bit + (i * (hashp->BSIZE << BYTE_SHIFT));
    798 	if (bit >= hashp->LAST_FREED)
    799 		hashp->LAST_FREED = bit - 1;
    800 
    801 	/* Calculate the split number for this page */
    802 	for (i = 0; (i < splitnum) && (bit > hashp->SPARES[i]); i++);
    803 	offset = (i ? bit - hashp->SPARES[i - 1] : bit);
    804 	if (offset >= SPLITMASK)
    805 		return (0);	/* Out of overflow pages */
    806 	addr = OADDR_OF(i, offset);
    807 #ifdef DEBUG2
    808 	(void)fprintf(stderr, "OVERFLOW_PAGE: ADDR: %d BIT: %d PAGE %d\n",
    809 	    addr, tmp1, tmp2);
    810 #endif
    811 
    812 	/* Allocate and return the overflow page */
    813 	return (addr);
    814 }
    815 
    816 /*
    817  * Mark this overflow page as free.
    818  */
    819 extern void
    820 __free_ovflpage(hashp, obufp)
    821 	HTAB *hashp;
    822 	BUFHEAD *obufp;
    823 {
    824 	register u_int16_t addr;
    825 	u_int32_t *freep;
    826 	int bit_address, free_page, free_bit;
    827 	u_int16_t ndx;
    828 
    829 	addr = obufp->addr;
    830 #ifdef DEBUG1
    831 	(void)fprintf(stderr, "Freeing %d\n", addr);
    832 #endif
    833 	ndx = (((u_int16_t)addr) >> SPLITSHIFT);
    834 	bit_address =
    835 	    (ndx ? hashp->SPARES[ndx - 1] : 0) + (addr & SPLITMASK) - 1;
    836 	 if (bit_address < hashp->LAST_FREED)
    837 		hashp->LAST_FREED = bit_address;
    838 	free_page = (bit_address >> (hashp->BSHIFT + BYTE_SHIFT));
    839 	free_bit = bit_address & ((hashp->BSIZE << BYTE_SHIFT) - 1);
    840 
    841 	if (!(freep = hashp->mapp[free_page]))
    842 		freep = fetch_bitmap(hashp, free_page);
    843 #ifdef DEBUG
    844 	/*
    845 	 * This had better never happen.  It means we tried to read a bitmap
    846 	 * that has already had overflow pages allocated off it, and we
    847 	 * failed to read it from the file.
    848 	 */
    849 	if (!freep)
    850 		assert(0);
    851 #endif
    852 	CLRBIT(freep, free_bit);
    853 #ifdef DEBUG2
    854 	(void)fprintf(stderr, "FREE_OVFLPAGE: ADDR: %d BIT: %d PAGE %d\n",
    855 	    obufp->addr, free_bit, free_page);
    856 #endif
    857 	__reclaim_buf(hashp, obufp);
    858 }
    859 
    860 /*
    861  * Returns:
    862  *	 0 success
    863  *	-1 failure
    864  */
    865 static int
    866 open_temp(hashp)
    867 	HTAB *hashp;
    868 {
    869 	sigset_t set, oset;
    870 	char namestr[] = "_hashXXXXXX";
    871 
    872 	/* Block signals; make sure file goes away at process exit. */
    873 	(void)sigfillset(&set);
    874 	(void)sigprocmask(SIG_BLOCK, &set, &oset);
    875 	if ((hashp->fp = mkstemp(namestr)) != -1) {
    876 		(void)unlink(namestr);
    877 		(void)fcntl(hashp->fp, F_SETFD, 1);
    878 	}
    879 	(void)sigprocmask(SIG_SETMASK, &oset, (sigset_t *)NULL);
    880 	return (hashp->fp != -1 ? 0 : -1);
    881 }
    882 
    883 /*
    884  * We have to know that the key will fit, but the last entry on the page is
    885  * an overflow pair, so we need to shift things.
    886  */
    887 static void
    888 squeeze_key(sp, key, val)
    889 	u_int16_t *sp;
    890 	const DBT *key, *val;
    891 {
    892 	register char *p;
    893 	u_int16_t free_space, n, off, pageno;
    894 
    895 	p = (char *)sp;
    896 	n = sp[0];
    897 	free_space = FREESPACE(sp);
    898 	off = OFFSET(sp);
    899 
    900 	pageno = sp[n - 1];
    901 	off -= key->size;
    902 	sp[n - 1] = off;
    903 	memmove(p + off, key->data, key->size);
    904 	off -= val->size;
    905 	sp[n] = off;
    906 	memmove(p + off, val->data, val->size);
    907 	sp[0] = n + 2;
    908 	sp[n + 1] = pageno;
    909 	sp[n + 2] = OVFLPAGE;
    910 	FREESPACE(sp) = free_space - PAIRSIZE(key, val);
    911 	OFFSET(sp) = off;
    912 }
    913 
    914 static u_int32_t *
    915 fetch_bitmap(hashp, ndx)
    916 	HTAB *hashp;
    917 	int ndx;
    918 {
    919 	if (ndx >= hashp->nmaps)
    920 		return (NULL);
    921 	if ((hashp->mapp[ndx] = (u_int32_t *)malloc(hashp->BSIZE)) == NULL)
    922 		return (NULL);
    923 	if (__get_page(hashp,
    924 	    (char *)hashp->mapp[ndx], hashp->BITMAPS[ndx], 0, 1, 1)) {
    925 		free(hashp->mapp[ndx]);
    926 		return (NULL);
    927 	}
    928 	return (hashp->mapp[ndx]);
    929 }
    930 
    931 #ifdef DEBUG4
    932 int
    933 print_chain(addr)
    934 	int addr;
    935 {
    936 	BUFHEAD *bufp;
    937 	short *bp, oaddr;
    938 
    939 	(void)fprintf(stderr, "%d ", addr);
    940 	bufp = __get_buf(hashp, addr, NULL, 0);
    941 	bp = (short *)bufp->page;
    942 	while (bp[0] && ((bp[bp[0]] == OVFLPAGE) ||
    943 		((bp[0] > 2) && bp[2] < REAL_KEY))) {
    944 		oaddr = bp[bp[0] - 1];
    945 		(void)fprintf(stderr, "%d ", (int)oaddr);
    946 		bufp = __get_buf(hashp, (int)oaddr, bufp, 0);
    947 		bp = (short *)bufp->page;
    948 	}
    949 	(void)fprintf(stderr, "\n");
    950 }
    951 #endif
    952