xref: /csrg-svn/lib/libc/db/btree/bt_open.c (revision 51358)
1 /*-
2  * Copyright (c) 1990 The Regents of the University of California.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Mike Olson.
7  *
8  * %sccs.include.redist.c%
9  */
10 
11 #if defined(LIBC_SCCS) && !defined(lint)
12 static char sccsid[] = "@(#)bt_open.c	5.12 (Berkeley) 10/10/91";
13 #endif /* LIBC_SCCS and not lint */
14 
15 /*
16  * Implementation of btree access method for 4.4BSD.
17  *
18  * The design here was originally based on that of the btree access method
19  * used in the Postgres database system at UC Berkeley.  This implementation
20  * is wholly independent of the Postgres code.
21  */
22 
23 #include <sys/param.h>
24 #include <sys/stat.h>
25 #include <fcntl.h>
26 #include <errno.h>
27 #include <limits.h>
28 #define	__DBINTERFACE_PRIVATE
29 #include <db.h>
30 #include <stdio.h>
31 #include <unistd.h>
32 #include <stdlib.h>
33 #include <string.h>
34 #include "btree.h"
35 
36 static int nroot __P((BTREE *));
37 static int tmp __P((void));
38 
39 /*
40  * __BT_OPEN -- Open a btree.
41  *
42  * Creates and fills a DB struct, and calls the routine that actually
43  * opens the btree.
44  *
45  * Parameters:
46  *	fname:	filename (NULL for in-memory trees)
47  *	flags:	open flag bits
48  *	mode:	open permission bits
49  *	b:	BTREEINFO pointer
50  *
51  * Returns:
52  *	NULL on failure, pointer to DB on success.
53  *
54  */
55 DB *
56 __bt_open(fname, flags, mode, openinfo)
57 	const char *fname;
58 	int flags, mode;
59 	const BTREEINFO *openinfo;
60 {
61 	BTMETA m;
62 	BTREE *t;
63 	BTREEINFO b;
64 	DB *dbp;
65 	pgno_t ncache;
66 	struct stat sb;
67 	int nr;
68 
69 	/*
70 	 * Intention is to make sure all of the user's selections are okay
71 	 * here and then use them without checking.  Can't be complete, since
72 	 * we don't know the right page size, lorder or flags until the backing
73 	 * file is opened.  Also, the file's page size can cause the cachesize
74 	 * to change.
75 	 */
76 	if (openinfo) {
77 		b = *openinfo;
78 
79 		/* Flags: R_DUP. */
80 		if (b.flags & ~(R_DUP))
81 			goto einval;
82 
83 		/*
84 		 * Page size must be index_t aligned and >= MINPSIZE.  Default
85 		 * page size is set farther on, based on the underlying file
86 		 * transfer size.
87 		 */
88 		if (b.psize &&
89 		    (b.psize < MINPSIZE || b.psize > MAX_PAGE_OFFSET ||
90 		    b.psize & sizeof(index_t) - 1))
91 				goto einval;
92 
93 		/* Minimum number of keys per page; absolute minimum is 2. */
94 		if (b.minkeypage) {
95 			if (b.minkeypage < 2)
96 				goto einval;
97 		} else
98 			b.minkeypage = DEFMINKEYPAGE;
99 
100 		/* If no comparison, use default comparison and prefix. */
101 		if (b.compare == NULL) {
102 			b.compare = __bt_defcmp;
103 			if (b.prefix == NULL)
104 				b.prefix = __bt_defpfx;
105 		}
106 
107 		if (b.lorder == 0)
108 			b.lorder = BYTE_ORDER;
109 		else if (b.lorder != BIG_ENDIAN && b.lorder != LITTLE_ENDIAN)
110 			goto einval;
111 	} else {
112 		b.flags = 0;
113 		b.minkeypage = DEFMINKEYPAGE;
114 		b.compare = __bt_defcmp;
115 		b.prefix = __bt_defpfx;
116 		b.lorder = BYTE_ORDER;
117 	}
118 
119 	/* Allocate and initialize DB and BTREE structures. */
120 	if ((t = malloc(sizeof(BTREE))) == NULL)
121 		goto err;
122 	t->bt_fd = -1;			/* Don't close unopened fd on error. */
123 	if ((t->bt_dbp = dbp = malloc(sizeof(DB))) == NULL)
124 		goto err;
125 	t->bt_bcursor.pgno = P_INVALID;
126 	t->bt_bcursor.index = 0;
127 	t->bt_stack = NULL;
128 	t->bt_sp = t->bt_maxstack = 0;
129 	t->bt_kbuf = t->bt_dbuf = NULL;
130 	t->bt_kbufsz = t->bt_dbufsz = 0;
131 	t->bt_order = NOT;
132 	t->bt_cmp = b.compare;
133 	t->bt_pfx = b.prefix;
134 
135 	dbp->type = DB_BTREE;
136 	dbp->internal = t;
137 	dbp->close = __bt_close;
138 	dbp->del = __bt_delete;
139 	dbp->get = __bt_get;
140 	dbp->put = __bt_put;
141 	dbp->seq = __bt_seq;
142 	dbp->sync = __bt_sync;
143 
144 	/*
145 	 * If no file name was supplied, this is an in-memory btree and we
146 	 * open a backing temporary file.  Otherwise, it's a disk-based tree.
147 	 */
148 	if (fname) {
149 #define	USEFLAGS	(O_CREAT|O_EXCL|O_RDONLY|O_RDWR|O_TRUNC|O_WRONLY)
150 		if ((t->bt_fd = open(fname, flags & USEFLAGS, mode)) < 0)
151 			goto err;
152 		if ((flags & O_ACCMODE) == O_RDONLY)
153 			SET(t, BTF_RDONLY);
154 
155 	} else {
156 		if ((t->bt_fd = tmp()) == -1)
157 			goto err;
158 		SET(t, BTF_INMEM);
159 	}
160 
161 	if (fcntl(t->bt_fd, F_SETFL, 1) == -1)
162 		goto err;
163 
164 	if (fstat(t->bt_fd, &sb))
165 		goto err;
166 	if (sb.st_size) {
167 		nr = read(t->bt_fd, &m, sizeof(BTMETA));
168 		if (nr < 0)
169 			goto err;
170 		if (nr != sizeof(BTMETA))
171 			goto eftype;
172 
173 		/*
174 		 * Read in the meta-data.  This can change the notion of what
175 		 * the lorder, page size and flags are, and, when the page size
176 		 * changes the cachesize value can change as well.
177 		 *
178 		 * Lorder is always stored in host-independent format.
179 		 */
180 		NTOHL(m.m_lorder);
181 		if (m.m_lorder != BIG_ENDIAN && m.m_lorder != LITTLE_ENDIAN)
182 			goto eftype;
183 		if (m.m_lorder != BYTE_ORDER) {
184 			BLSWAP(m.m_magic);
185 			BLSWAP(m.m_version);
186 			BLSWAP(m.m_psize);
187 			BLSWAP(m.m_free);
188 			BLSWAP(m.m_nrecs);
189 			BLSWAP(m.m_flags);
190 		}
191 		if (m.m_magic != BTREEMAGIC || m.m_version != BTREEVERSION)
192 			goto eftype;
193 		if (m.m_psize < MINPSIZE || m.m_psize > MAX_PAGE_OFFSET ||
194 		    m.m_psize & sizeof(index_t) - 1)
195 			goto eftype;
196 		if (m.m_flags & ~SAVEMETA)
197 			goto eftype;
198 		b.psize = m.m_psize;
199 		t->bt_flags |= m.m_flags;
200 		t->bt_free = m.m_free;
201 		t->bt_lorder = m.m_lorder;
202 		t->bt_nrecs = m.m_nrecs;
203 	} else {
204 		/*
205 		 * Set the page size to the best value for I/O to this file.
206 		 * Don't overflow the page offset type.
207 		 */
208 		if (b.psize == 0) {
209 			b.psize = sb.st_blksize;
210 			if (b.psize < MINPSIZE)
211 				b.psize = MINPSIZE;
212 			if (b.psize > MAX_PAGE_OFFSET)
213 				b.psize = MAX_PAGE_OFFSET;
214 		}
215 		t->bt_flags |= b.flags & R_DUP ? 0 : BTF_NODUPS;
216 		t->bt_free = P_INVALID;
217 		t->bt_lorder = b.lorder;
218 		t->bt_nrecs = 0;
219 		SET(t, BTF_METADIRTY);
220 	}
221 
222 	t->bt_psize = b.psize;
223 
224 	/* Set the cache size; must be a multiple of the page size. */
225 	if (b.cachesize && b.cachesize & b.psize - 1)
226 		b.cachesize += (~b.cachesize & b.psize - 1) + 1;
227 	if (b.cachesize < b.psize * MINCACHE)
228 		b.cachesize = b.psize * MINCACHE;
229 
230 	/* Calculate number of pages to cache. */
231 	ncache = (b.cachesize + t->bt_psize - 1) / t->bt_psize;
232 
233 	/*
234 	 * The btree data structure requires that at least two keys can fit on
235 	 * a page, but other than that there's no fixed requirement.  The user
236 	 * specified a minimum number per page, and we translated that into the
237 	 * number of bytes a key/data pair can use before being placed on an
238 	 * overflow page.  This calculation includes the page header, the size
239 	 * of the index referencing the leaf item and the size of the leaf item
240 	 * structure.  Also, don't let the user specify a minkeypage such that
241 	 * a key/data pair won't fit even if both key and data are on overflow
242 	 * pages.
243 	 */
244 	t->bt_ovflsize = (t->bt_psize - BTDATAOFF) / b.minkeypage -
245 	    (sizeof(index_t) + NBLEAFDBT(0, 0));
246 	if (t->bt_ovflsize < NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(index_t))
247 		t->bt_ovflsize =
248 		    NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(index_t);
249 
250 	/* Initialize the buffer pool. */
251 	if ((t->bt_mp =
252 	    mpool_open(NULL, t->bt_fd, t->bt_psize, ncache)) == NULL)
253 		goto err;
254 	if (NOTSET(t, BTF_INMEM))
255 		mpool_filter(t->bt_mp, __bt_pgin, __bt_pgout, t);
256 
257 	/* Create a root page if new tree. */
258 	if (nroot(t) == RET_ERROR)
259 		goto err;
260 
261 	return (dbp);
262 
263 einval:	errno = EINVAL;
264 	goto err;
265 
266 eftype:	errno = EFTYPE;
267 	goto err;
268 
269 err:	if (t) {
270 		if (t->bt_dbp)
271 			free(t->bt_dbp);
272 		if (t->bt_fd != -1)
273 			(void)close(t->bt_fd);
274 		free(t);
275 	}
276 	return (NULL);
277 }
278 
279 /*
280  * NROOT -- Create the root of a new tree.
281  *
282  * Parameters:
283  *	t:	tree
284  *
285  * Returns:
286  *	RET_ERROR, RET_SUCCESS
287  */
288 static int
289 nroot(t)
290 	BTREE *t;
291 {
292 	PAGE *meta, *root;
293 	pgno_t npg;
294 
295 	if ((meta = mpool_get(t->bt_mp, 0, 0)) != NULL) {
296 		mpool_put(t->bt_mp, meta, 0);
297 		return (RET_SUCCESS);
298 	}
299 	if (errno != EINVAL)
300 		return (RET_ERROR);
301 
302 	if ((meta = mpool_new(t->bt_mp, &npg)) == NULL)
303 		return (RET_ERROR);
304 
305 	if ((root = mpool_new(t->bt_mp, &npg)) == NULL)
306 		return (RET_ERROR);
307 
308 	if (npg != P_ROOT)
309 		return (RET_ERROR);
310 	root->pgno = npg;
311 	root->prevpg = root->nextpg = P_INVALID;
312 	root->lower = BTDATAOFF;
313 	root->upper = t->bt_psize;
314 	root->flags = P_BLEAF;
315 	bzero(meta, t->bt_psize);
316 	mpool_put(t->bt_mp, meta, MPOOL_DIRTY);
317 	mpool_put(t->bt_mp, root, MPOOL_DIRTY);
318 	return (RET_SUCCESS);
319 }
320 
321 static int
322 tmp()
323 {
324 	sigset_t set, oset;
325 	int fd;
326 	char *envtmp;
327 	char path[MAXPATHLEN];
328 
329 	envtmp = getenv("TMPDIR");
330 	(void)snprintf(path,
331 	    sizeof(path), "%s/bt.XXXXXX", envtmp ? envtmp : "/tmp");
332 
333 	sigfillset(&set);
334 	(void)sigprocmask(SIG_BLOCK, &set, &oset);
335 	if ((fd = mkstemp(path)) != -1)
336 		(void)unlink(path);
337 	(void)sigprocmask(SIG_SETMASK, &oset, NULL);
338 	return(fd);
339 }
340