145876Sbostic /*- 245876Sbostic * Copyright (c) 1990 The Regents of the University of California. 345876Sbostic * All rights reserved. 445876Sbostic * 545876Sbostic * This code is derived from software contributed to Berkeley by 645876Sbostic * Mike Olson. 745876Sbostic * 845876Sbostic * %sccs.include.redist.c% 945876Sbostic */ 1045876Sbostic 1145876Sbostic #if defined(LIBC_SCCS) && !defined(lint) 12*51358Sbostic static char sccsid[] = "@(#)bt_open.c 5.12 (Berkeley) 10/10/91"; 1345876Sbostic #endif /* LIBC_SCCS and not lint */ 1445876Sbostic 1545876Sbostic /* 1650989Sbostic * Implementation of btree access method for 4.4BSD. 1745876Sbostic * 1850989Sbostic * The design here was originally based on that of the btree access method 1950989Sbostic * used in the Postgres database system at UC Berkeley. This implementation 2050989Sbostic * is wholly independent of the Postgres code. 2145876Sbostic */ 2245876Sbostic 2345876Sbostic #include <sys/param.h> 2445876Sbostic #include <sys/stat.h> 2550989Sbostic #include <fcntl.h> 2646561Sbostic #include <errno.h> 2750989Sbostic #include <limits.h> 2850989Sbostic #define __DBINTERFACE_PRIVATE 2945876Sbostic #include <db.h> 3050989Sbostic #include <stdio.h> 3150989Sbostic #include <unistd.h> 3246561Sbostic #include <stdlib.h> 3351098Sbostic #include <string.h> 3446127Smao #include "btree.h" 3545876Sbostic 3650989Sbostic static int nroot __P((BTREE *)); 3750989Sbostic static int tmp __P((void)); 3845876Sbostic 3945876Sbostic /* 4050989Sbostic * __BT_OPEN -- Open a btree. 4145876Sbostic * 4250989Sbostic * Creates and fills a DB struct, and calls the routine that actually 4350989Sbostic * opens the btree. 4445876Sbostic * 4550989Sbostic * Parameters: 4650989Sbostic * fname: filename (NULL for in-memory trees) 4750989Sbostic * flags: open flag bits 4850989Sbostic * mode: open permission bits 4950989Sbostic * b: BTREEINFO pointer 5045876Sbostic * 5150989Sbostic * Returns: 5250989Sbostic * NULL on failure, pointer to DB on success. 5345876Sbostic * 5445876Sbostic */ 5545876Sbostic DB * 5650989Sbostic __bt_open(fname, flags, mode, openinfo) 5750989Sbostic const char *fname; 5850989Sbostic int flags, mode; 5950989Sbostic const BTREEINFO *openinfo; 6045876Sbostic { 6151098Sbostic BTMETA m; 6250989Sbostic BTREE *t; 6351098Sbostic BTREEINFO b; 6450989Sbostic DB *dbp; 6550989Sbostic pgno_t ncache; 6650989Sbostic struct stat sb; 6750989Sbostic int nr; 6845876Sbostic 6945876Sbostic /* 7050989Sbostic * Intention is to make sure all of the user's selections are okay 7150989Sbostic * here and then use them without checking. Can't be complete, since 7250989Sbostic * we don't know the right page size, lorder or flags until the backing 7350989Sbostic * file is opened. Also, the file's page size can cause the cachesize 7450989Sbostic * to change. 7545876Sbostic */ 7650989Sbostic if (openinfo) { 7750989Sbostic b = *openinfo; 7845876Sbostic 7950989Sbostic /* Flags: R_DUP. */ 8051098Sbostic if (b.flags & ~(R_DUP)) 8150989Sbostic goto einval; 8245876Sbostic 8350989Sbostic /* 8450989Sbostic * Page size must be index_t aligned and >= MINPSIZE. Default 8550989Sbostic * page size is set farther on, based on the underlying file 8650989Sbostic * transfer size. 8750989Sbostic */ 8850989Sbostic if (b.psize && 8950989Sbostic (b.psize < MINPSIZE || b.psize > MAX_PAGE_OFFSET || 9050989Sbostic b.psize & sizeof(index_t) - 1)) 9150989Sbostic goto einval; 9245876Sbostic 9351098Sbostic /* Minimum number of keys per page; absolute minimum is 2. */ 9450989Sbostic if (b.minkeypage) { 9550989Sbostic if (b.minkeypage < 2) 9650989Sbostic goto einval; 9750989Sbostic } else 9851098Sbostic b.minkeypage = DEFMINKEYPAGE; 9951098Sbostic 10050989Sbostic /* If no comparison, use default comparison and prefix. */ 10150989Sbostic if (b.compare == NULL) { 10250989Sbostic b.compare = __bt_defcmp; 10350989Sbostic if (b.prefix == NULL) 10450989Sbostic b.prefix = __bt_defpfx; 10545876Sbostic } 10645876Sbostic 10750989Sbostic if (b.lorder == 0) 10850989Sbostic b.lorder = BYTE_ORDER; 10950989Sbostic else if (b.lorder != BIG_ENDIAN && b.lorder != LITTLE_ENDIAN) 11050989Sbostic goto einval; 11150989Sbostic } else { 11250989Sbostic b.flags = 0; 11350989Sbostic b.minkeypage = DEFMINKEYPAGE; 11450989Sbostic b.compare = __bt_defcmp; 11550989Sbostic b.prefix = __bt_defpfx; 11650989Sbostic b.lorder = BYTE_ORDER; 11750989Sbostic } 11845876Sbostic 11950989Sbostic /* Allocate and initialize DB and BTREE structures. */ 12050989Sbostic if ((t = malloc(sizeof(BTREE))) == NULL) 12150989Sbostic goto err; 12250989Sbostic t->bt_fd = -1; /* Don't close unopened fd on error. */ 12350989Sbostic if ((t->bt_dbp = dbp = malloc(sizeof(DB))) == NULL) 12450989Sbostic goto err; 12550989Sbostic t->bt_bcursor.pgno = P_INVALID; 12650989Sbostic t->bt_bcursor.index = 0; 12750989Sbostic t->bt_stack = NULL; 12850989Sbostic t->bt_sp = t->bt_maxstack = 0; 12950989Sbostic t->bt_kbuf = t->bt_dbuf = NULL; 13050989Sbostic t->bt_kbufsz = t->bt_dbufsz = 0; 13150989Sbostic t->bt_order = NOT; 13250989Sbostic t->bt_cmp = b.compare; 13350989Sbostic t->bt_pfx = b.prefix; 13446453Smao 13550989Sbostic dbp->type = DB_BTREE; 13650989Sbostic dbp->internal = t; 13750989Sbostic dbp->close = __bt_close; 13850989Sbostic dbp->del = __bt_delete; 13950989Sbostic dbp->get = __bt_get; 14050989Sbostic dbp->put = __bt_put; 14150989Sbostic dbp->seq = __bt_seq; 14250989Sbostic dbp->sync = __bt_sync; 14345876Sbostic 14450989Sbostic /* 14550989Sbostic * If no file name was supplied, this is an in-memory btree and we 14650989Sbostic * open a backing temporary file. Otherwise, it's a disk-based tree. 14750989Sbostic */ 14850989Sbostic if (fname) { 14950989Sbostic #define USEFLAGS (O_CREAT|O_EXCL|O_RDONLY|O_RDWR|O_TRUNC|O_WRONLY) 15050989Sbostic if ((t->bt_fd = open(fname, flags & USEFLAGS, mode)) < 0) 15150989Sbostic goto err; 15250989Sbostic if ((flags & O_ACCMODE) == O_RDONLY) 15350989Sbostic SET(t, BTF_RDONLY); 15445876Sbostic 15550989Sbostic } else { 15650989Sbostic if ((t->bt_fd = tmp()) == -1) 15750989Sbostic goto err; 15850989Sbostic SET(t, BTF_INMEM); 15950989Sbostic } 16045876Sbostic 16150989Sbostic if (fcntl(t->bt_fd, F_SETFL, 1) == -1) 16250989Sbostic goto err; 16345876Sbostic 16450989Sbostic if (fstat(t->bt_fd, &sb)) 16550989Sbostic goto err; 16650989Sbostic if (sb.st_size) { 16750989Sbostic nr = read(t->bt_fd, &m, sizeof(BTMETA)); 16850989Sbostic if (nr < 0) 16950989Sbostic goto err; 17050989Sbostic if (nr != sizeof(BTMETA)) 17150989Sbostic goto eftype; 17245876Sbostic 17350989Sbostic /* 17450989Sbostic * Read in the meta-data. This can change the notion of what 17550989Sbostic * the lorder, page size and flags are, and, when the page size 17650989Sbostic * changes the cachesize value can change as well. 17750989Sbostic * 17850989Sbostic * Lorder is always stored in host-independent format. 17950989Sbostic */ 18050989Sbostic NTOHL(m.m_lorder); 18150989Sbostic if (m.m_lorder != BIG_ENDIAN && m.m_lorder != LITTLE_ENDIAN) 18250989Sbostic goto eftype; 18350989Sbostic if (m.m_lorder != BYTE_ORDER) { 18450989Sbostic BLSWAP(m.m_magic); 18550989Sbostic BLSWAP(m.m_version); 18650989Sbostic BLSWAP(m.m_psize); 18750989Sbostic BLSWAP(m.m_free); 18850989Sbostic BLSWAP(m.m_nrecs); 18950989Sbostic BLSWAP(m.m_flags); 19045876Sbostic } 19150989Sbostic if (m.m_magic != BTREEMAGIC || m.m_version != BTREEVERSION) 19250989Sbostic goto eftype; 19350989Sbostic if (m.m_psize < MINPSIZE || m.m_psize > MAX_PAGE_OFFSET || 19450989Sbostic m.m_psize & sizeof(index_t) - 1) 19550989Sbostic goto eftype; 196*51358Sbostic if (m.m_flags & ~SAVEMETA) 19750989Sbostic goto eftype; 19850989Sbostic b.psize = m.m_psize; 19951098Sbostic t->bt_flags |= m.m_flags; 20050989Sbostic t->bt_free = m.m_free; 20150989Sbostic t->bt_lorder = m.m_lorder; 20250989Sbostic t->bt_nrecs = m.m_nrecs; 20350989Sbostic } else { 20450989Sbostic /* 20550989Sbostic * Set the page size to the best value for I/O to this file. 20650989Sbostic * Don't overflow the page offset type. 20750989Sbostic */ 20850989Sbostic if (b.psize == 0) { 20950989Sbostic b.psize = sb.st_blksize; 21050989Sbostic if (b.psize < MINPSIZE) 21150989Sbostic b.psize = MINPSIZE; 21250989Sbostic if (b.psize > MAX_PAGE_OFFSET) 21350989Sbostic b.psize = MAX_PAGE_OFFSET; 21450989Sbostic } 21551098Sbostic t->bt_flags |= b.flags & R_DUP ? 0 : BTF_NODUPS; 21650989Sbostic t->bt_free = P_INVALID; 21750989Sbostic t->bt_lorder = b.lorder; 21850989Sbostic t->bt_nrecs = 0; 21950989Sbostic SET(t, BTF_METADIRTY); 22045876Sbostic } 22145876Sbostic 22250989Sbostic t->bt_psize = b.psize; 22345876Sbostic 22450989Sbostic /* Set the cache size; must be a multiple of the page size. */ 22550989Sbostic if (b.cachesize && b.cachesize & b.psize - 1) 22650989Sbostic b.cachesize += (~b.cachesize & b.psize - 1) + 1; 22750989Sbostic if (b.cachesize < b.psize * MINCACHE) 22850989Sbostic b.cachesize = b.psize * MINCACHE; 22945876Sbostic 23050989Sbostic /* Calculate number of pages to cache. */ 23150989Sbostic ncache = (b.cachesize + t->bt_psize - 1) / t->bt_psize; 23245876Sbostic 23345876Sbostic /* 23451098Sbostic * The btree data structure requires that at least two keys can fit on 23551098Sbostic * a page, but other than that there's no fixed requirement. The user 23651098Sbostic * specified a minimum number per page, and we translated that into the 23751098Sbostic * number of bytes a key/data pair can use before being placed on an 23851098Sbostic * overflow page. This calculation includes the page header, the size 23951098Sbostic * of the index referencing the leaf item and the size of the leaf item 24051098Sbostic * structure. Also, don't let the user specify a minkeypage such that 24151098Sbostic * a key/data pair won't fit even if both key and data are on overflow 24251098Sbostic * pages. 24345876Sbostic */ 24451098Sbostic t->bt_ovflsize = (t->bt_psize - BTDATAOFF) / b.minkeypage - 24551098Sbostic (sizeof(index_t) + NBLEAFDBT(0, 0)); 24651098Sbostic if (t->bt_ovflsize < NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(index_t)) 24751098Sbostic t->bt_ovflsize = 24851098Sbostic NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(index_t); 24945876Sbostic 25050989Sbostic /* Initialize the buffer pool. */ 25150989Sbostic if ((t->bt_mp = 25250989Sbostic mpool_open(NULL, t->bt_fd, t->bt_psize, ncache)) == NULL) 25350989Sbostic goto err; 25451098Sbostic if (NOTSET(t, BTF_INMEM)) 25551098Sbostic mpool_filter(t->bt_mp, __bt_pgin, __bt_pgout, t); 25645876Sbostic 25750989Sbostic /* Create a root page if new tree. */ 25850989Sbostic if (nroot(t) == RET_ERROR) 25950989Sbostic goto err; 26045876Sbostic 26150989Sbostic return (dbp); 26245876Sbostic 26350989Sbostic einval: errno = EINVAL; 26450989Sbostic goto err; 26545876Sbostic 26650989Sbostic eftype: errno = EFTYPE; 26750989Sbostic goto err; 26849322Sbostic 26950989Sbostic err: if (t) { 27050989Sbostic if (t->bt_dbp) 27150989Sbostic free(t->bt_dbp); 27250989Sbostic if (t->bt_fd != -1) 27350989Sbostic (void)close(t->bt_fd); 27450989Sbostic free(t); 27545876Sbostic } 27650989Sbostic return (NULL); 27745876Sbostic } 27845876Sbostic 27945876Sbostic /* 28050989Sbostic * NROOT -- Create the root of a new tree. 28145876Sbostic * 28250989Sbostic * Parameters: 28350989Sbostic * t: tree 28445876Sbostic * 28550989Sbostic * Returns: 28650989Sbostic * RET_ERROR, RET_SUCCESS 28745876Sbostic */ 28850989Sbostic static int 28950989Sbostic nroot(t) 29050989Sbostic BTREE *t; 29145876Sbostic { 29250989Sbostic PAGE *meta, *root; 29350989Sbostic pgno_t npg; 29445876Sbostic 29550989Sbostic if ((meta = mpool_get(t->bt_mp, 0, 0)) != NULL) { 29650989Sbostic mpool_put(t->bt_mp, meta, 0); 29745876Sbostic return (RET_SUCCESS); 29845876Sbostic } 29950989Sbostic if (errno != EINVAL) 30050989Sbostic return (RET_ERROR); 30145876Sbostic 30250989Sbostic if ((meta = mpool_new(t->bt_mp, &npg)) == NULL) 30350989Sbostic return (RET_ERROR); 30445876Sbostic 30550989Sbostic if ((root = mpool_new(t->bt_mp, &npg)) == NULL) 30650989Sbostic return (RET_ERROR); 30745876Sbostic 30850989Sbostic if (npg != P_ROOT) 30945876Sbostic return (RET_ERROR); 31050989Sbostic root->pgno = npg; 31150989Sbostic root->prevpg = root->nextpg = P_INVALID; 31250989Sbostic root->lower = BTDATAOFF; 31350989Sbostic root->upper = t->bt_psize; 31450989Sbostic root->flags = P_BLEAF; 31551098Sbostic bzero(meta, t->bt_psize); 31650989Sbostic mpool_put(t->bt_mp, meta, MPOOL_DIRTY); 31750989Sbostic mpool_put(t->bt_mp, root, MPOOL_DIRTY); 31850989Sbostic return (RET_SUCCESS); 31945876Sbostic } 32045876Sbostic 32150989Sbostic static int 32250989Sbostic tmp() 32345876Sbostic { 32450989Sbostic sigset_t set, oset; 32550989Sbostic int fd; 32650989Sbostic char *envtmp; 32750989Sbostic char path[MAXPATHLEN]; 32845876Sbostic 32950989Sbostic envtmp = getenv("TMPDIR"); 33050989Sbostic (void)snprintf(path, 33150989Sbostic sizeof(path), "%s/bt.XXXXXX", envtmp ? envtmp : "/tmp"); 33249322Sbostic 33350989Sbostic sigfillset(&set); 33450989Sbostic (void)sigprocmask(SIG_BLOCK, &set, &oset); 33550989Sbostic if ((fd = mkstemp(path)) != -1) 33650989Sbostic (void)unlink(path); 33750989Sbostic (void)sigprocmask(SIG_SETMASK, &oset, NULL); 33850989Sbostic return(fd); 33945876Sbostic } 340