1*7c478bd9Sstevel@tonic-gate /*-
2*7c478bd9Sstevel@tonic-gate  * See the file LICENSE for redistribution information.
3*7c478bd9Sstevel@tonic-gate  *
4*7c478bd9Sstevel@tonic-gate  * Copyright (c) 1997, 1998
5*7c478bd9Sstevel@tonic-gate  *	Sleepycat Software.  All rights reserved.
6*7c478bd9Sstevel@tonic-gate  */
7*7c478bd9Sstevel@tonic-gate 
8*7c478bd9Sstevel@tonic-gate #include "config.h"
9*7c478bd9Sstevel@tonic-gate 
10*7c478bd9Sstevel@tonic-gate #ifndef lint
11*7c478bd9Sstevel@tonic-gate static const char sccsid[] = "@(#)bt_recno.c	10.53 (Sleepycat) 12/11/98";
12*7c478bd9Sstevel@tonic-gate #endif /* not lint */
13*7c478bd9Sstevel@tonic-gate 
14*7c478bd9Sstevel@tonic-gate #ifndef NO_SYSTEM_INCLUDES
15*7c478bd9Sstevel@tonic-gate #include <sys/types.h>
16*7c478bd9Sstevel@tonic-gate 
17*7c478bd9Sstevel@tonic-gate #include <errno.h>
18*7c478bd9Sstevel@tonic-gate #include <limits.h>
19*7c478bd9Sstevel@tonic-gate #include <string.h>
20*7c478bd9Sstevel@tonic-gate #endif
21*7c478bd9Sstevel@tonic-gate 
22*7c478bd9Sstevel@tonic-gate #include "db_int.h"
23*7c478bd9Sstevel@tonic-gate #include "db_page.h"
24*7c478bd9Sstevel@tonic-gate #include "btree.h"
25*7c478bd9Sstevel@tonic-gate #include "db_ext.h"
26*7c478bd9Sstevel@tonic-gate #include "shqueue.h"
27*7c478bd9Sstevel@tonic-gate #include "db_shash.h"
28*7c478bd9Sstevel@tonic-gate #include "lock.h"
29*7c478bd9Sstevel@tonic-gate #include "lock_ext.h"
30*7c478bd9Sstevel@tonic-gate 
31*7c478bd9Sstevel@tonic-gate static int __ram_add __P((DBC *, db_recno_t *, DBT *, u_int32_t, u_int32_t));
32*7c478bd9Sstevel@tonic-gate static int __ram_delete __P((DB *, DB_TXN *, DBT *, u_int32_t));
33*7c478bd9Sstevel@tonic-gate static int __ram_fmap __P((DBC *, db_recno_t));
34*7c478bd9Sstevel@tonic-gate static int __ram_i_delete __P((DBC *));
35*7c478bd9Sstevel@tonic-gate static int __ram_put __P((DB *, DB_TXN *, DBT *, DBT *, u_int32_t));
36*7c478bd9Sstevel@tonic-gate static int __ram_source __P((DB *, RECNO *, const char *));
37*7c478bd9Sstevel@tonic-gate static int __ram_sync __P((DB *, u_int32_t));
38*7c478bd9Sstevel@tonic-gate static int __ram_update __P((DBC *, db_recno_t, int));
39*7c478bd9Sstevel@tonic-gate static int __ram_vmap __P((DBC *, db_recno_t));
40*7c478bd9Sstevel@tonic-gate static int __ram_writeback __P((DBC *));
41*7c478bd9Sstevel@tonic-gate 
42*7c478bd9Sstevel@tonic-gate /*
43*7c478bd9Sstevel@tonic-gate  * In recno, there are two meanings to the on-page "deleted" flag.  If we're
44*7c478bd9Sstevel@tonic-gate  * re-numbering records, it means the record was implicitly created.  We skip
45*7c478bd9Sstevel@tonic-gate  * over implicitly created records if doing a cursor "next" or "prev", and
46*7c478bd9Sstevel@tonic-gate  * return DB_KEYEMPTY if they're explicitly requested..  If not re-numbering
47*7c478bd9Sstevel@tonic-gate  * records, it means that the record was implicitly created, or was deleted.
48*7c478bd9Sstevel@tonic-gate  * We skip over implicitly created or deleted records if doing a cursor "next"
49*7c478bd9Sstevel@tonic-gate  * or "prev", and return DB_KEYEMPTY if they're explicitly requested.
50*7c478bd9Sstevel@tonic-gate  *
51*7c478bd9Sstevel@tonic-gate  * If we're re-numbering records, then we have to detect in the cursor that
52*7c478bd9Sstevel@tonic-gate  * a record was deleted, and adjust the cursor as necessary on the next get.
53*7c478bd9Sstevel@tonic-gate  * If we're not re-numbering records, then we can detect that a record has
54*7c478bd9Sstevel@tonic-gate  * been deleted by looking at the actual on-page record, so we completely
55*7c478bd9Sstevel@tonic-gate  * ignore the cursor's delete flag.  This is different from the B+tree code.
56*7c478bd9Sstevel@tonic-gate  * It also maintains whether the cursor references a deleted record in the
57*7c478bd9Sstevel@tonic-gate  * cursor, and it doesn't always check the on-page value.
58*7c478bd9Sstevel@tonic-gate  */
59*7c478bd9Sstevel@tonic-gate #define	CD_SET(dbp, cp) {						\
60*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_RE_RENUMBER))				\
61*7c478bd9Sstevel@tonic-gate 		F_SET(cp, C_DELETED);					\
62*7c478bd9Sstevel@tonic-gate }
63*7c478bd9Sstevel@tonic-gate #define	CD_CLR(dbp, cp) {						\
64*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_RE_RENUMBER))				\
65*7c478bd9Sstevel@tonic-gate 		F_CLR(cp, C_DELETED);					\
66*7c478bd9Sstevel@tonic-gate }
67*7c478bd9Sstevel@tonic-gate #define	CD_ISSET(dbp, cp)						\
68*7c478bd9Sstevel@tonic-gate 	(F_ISSET(dbp, DB_RE_RENUMBER) && F_ISSET(cp, C_DELETED))
69*7c478bd9Sstevel@tonic-gate 
70*7c478bd9Sstevel@tonic-gate /*
71*7c478bd9Sstevel@tonic-gate  * __ram_open --
72*7c478bd9Sstevel@tonic-gate  *	Recno open function.
73*7c478bd9Sstevel@tonic-gate  *
74*7c478bd9Sstevel@tonic-gate  * PUBLIC: int __ram_open __P((DB *, DB_INFO *));
75*7c478bd9Sstevel@tonic-gate  */
76*7c478bd9Sstevel@tonic-gate int
__ram_open(dbp,dbinfo)77*7c478bd9Sstevel@tonic-gate __ram_open(dbp, dbinfo)
78*7c478bd9Sstevel@tonic-gate 	DB *dbp;
79*7c478bd9Sstevel@tonic-gate 	DB_INFO *dbinfo;
80*7c478bd9Sstevel@tonic-gate {
81*7c478bd9Sstevel@tonic-gate 	BTREE *t;
82*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
83*7c478bd9Sstevel@tonic-gate 	RECNO *rp;
84*7c478bd9Sstevel@tonic-gate 	int ret, t_ret;
85*7c478bd9Sstevel@tonic-gate 
86*7c478bd9Sstevel@tonic-gate 	/* Allocate and initialize the private btree structure. */
87*7c478bd9Sstevel@tonic-gate 	if ((ret = __os_calloc(1, sizeof(BTREE), &t)) != 0)
88*7c478bd9Sstevel@tonic-gate 		return (ret);
89*7c478bd9Sstevel@tonic-gate 	dbp->internal = t;
90*7c478bd9Sstevel@tonic-gate 	__bam_setovflsize(dbp);
91*7c478bd9Sstevel@tonic-gate 
92*7c478bd9Sstevel@tonic-gate 	/* Allocate and initialize the private recno structure. */
93*7c478bd9Sstevel@tonic-gate 	if ((ret = __os_calloc(1, sizeof(*rp), &rp)) != 0)
94*7c478bd9Sstevel@tonic-gate 		return (ret);
95*7c478bd9Sstevel@tonic-gate 	/* Link in the private recno structure. */
96*7c478bd9Sstevel@tonic-gate 	t->recno = rp;
97*7c478bd9Sstevel@tonic-gate 
98*7c478bd9Sstevel@tonic-gate 	/*
99*7c478bd9Sstevel@tonic-gate 	 * Intention is to make sure all of the user's selections are okay
100*7c478bd9Sstevel@tonic-gate 	 * here and then use them without checking.
101*7c478bd9Sstevel@tonic-gate 	 */
102*7c478bd9Sstevel@tonic-gate 	if (dbinfo == NULL) {
103*7c478bd9Sstevel@tonic-gate 		rp->re_delim = '\n';
104*7c478bd9Sstevel@tonic-gate 		rp->re_pad = ' ';
105*7c478bd9Sstevel@tonic-gate 		rp->re_fd = -1;
106*7c478bd9Sstevel@tonic-gate 		F_SET(rp, RECNO_EOF);
107*7c478bd9Sstevel@tonic-gate 	} else {
108*7c478bd9Sstevel@tonic-gate 		/*
109*7c478bd9Sstevel@tonic-gate 		 * If the user specified a source tree, open it and map it in.
110*7c478bd9Sstevel@tonic-gate 		 *
111*7c478bd9Sstevel@tonic-gate 		 * !!!
112*7c478bd9Sstevel@tonic-gate 		 * We don't complain if the user specified transactions or
113*7c478bd9Sstevel@tonic-gate 		 * threads.  It's possible to make it work, but you'd better
114*7c478bd9Sstevel@tonic-gate 		 * know what you're doing!
115*7c478bd9Sstevel@tonic-gate 		 */
116*7c478bd9Sstevel@tonic-gate 		if (dbinfo->re_source == NULL) {
117*7c478bd9Sstevel@tonic-gate 			rp->re_fd = -1;
118*7c478bd9Sstevel@tonic-gate 			F_SET(rp, RECNO_EOF);
119*7c478bd9Sstevel@tonic-gate 		} else {
120*7c478bd9Sstevel@tonic-gate 			if ((ret =
121*7c478bd9Sstevel@tonic-gate 			    __ram_source(dbp, rp, dbinfo->re_source)) != 0)
122*7c478bd9Sstevel@tonic-gate 			goto err;
123*7c478bd9Sstevel@tonic-gate 		}
124*7c478bd9Sstevel@tonic-gate 
125*7c478bd9Sstevel@tonic-gate 		/* Copy delimiter, length and padding values. */
126*7c478bd9Sstevel@tonic-gate 		rp->re_delim =
127*7c478bd9Sstevel@tonic-gate 		    F_ISSET(dbp, DB_RE_DELIMITER) ? dbinfo->re_delim : '\n';
128*7c478bd9Sstevel@tonic-gate 		rp->re_pad = F_ISSET(dbp, DB_RE_PAD) ? dbinfo->re_pad : ' ';
129*7c478bd9Sstevel@tonic-gate 
130*7c478bd9Sstevel@tonic-gate 		if (F_ISSET(dbp, DB_RE_FIXEDLEN)) {
131*7c478bd9Sstevel@tonic-gate 			if ((rp->re_len = dbinfo->re_len) == 0) {
132*7c478bd9Sstevel@tonic-gate 				__db_err(dbp->dbenv,
133*7c478bd9Sstevel@tonic-gate 				    "record length must be greater than 0");
134*7c478bd9Sstevel@tonic-gate 				ret = EINVAL;
135*7c478bd9Sstevel@tonic-gate 				goto err;
136*7c478bd9Sstevel@tonic-gate 			}
137*7c478bd9Sstevel@tonic-gate 		} else
138*7c478bd9Sstevel@tonic-gate 			rp->re_len = 0;
139*7c478bd9Sstevel@tonic-gate 	}
140*7c478bd9Sstevel@tonic-gate 
141*7c478bd9Sstevel@tonic-gate 	/* Initialize the remaining fields/methods of the DB. */
142*7c478bd9Sstevel@tonic-gate 	dbp->am_close = __ram_close;
143*7c478bd9Sstevel@tonic-gate 	dbp->del = __ram_delete;
144*7c478bd9Sstevel@tonic-gate 	dbp->put = __ram_put;
145*7c478bd9Sstevel@tonic-gate 	dbp->stat = __bam_stat;
146*7c478bd9Sstevel@tonic-gate 	dbp->sync = __ram_sync;
147*7c478bd9Sstevel@tonic-gate 
148*7c478bd9Sstevel@tonic-gate 	/* Start up the tree. */
149*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_read_root(dbp)) != 0)
150*7c478bd9Sstevel@tonic-gate 		goto err;
151*7c478bd9Sstevel@tonic-gate 
152*7c478bd9Sstevel@tonic-gate 	/* Set the overflow page size. */
153*7c478bd9Sstevel@tonic-gate 	__bam_setovflsize(dbp);
154*7c478bd9Sstevel@tonic-gate 
155*7c478bd9Sstevel@tonic-gate 	/* If we're snapshotting an underlying source file, do it now. */
156*7c478bd9Sstevel@tonic-gate 	if (dbinfo != NULL && F_ISSET(dbinfo, DB_SNAPSHOT)) {
157*7c478bd9Sstevel@tonic-gate 		/* Allocate a cursor. */
158*7c478bd9Sstevel@tonic-gate 		if ((ret = dbp->cursor(dbp, NULL, &dbc, 0)) != 0)
159*7c478bd9Sstevel@tonic-gate 			goto err;
160*7c478bd9Sstevel@tonic-gate 
161*7c478bd9Sstevel@tonic-gate 		/* Do the snapshot. */
162*7c478bd9Sstevel@tonic-gate 		if ((ret = __ram_update(dbc,
163*7c478bd9Sstevel@tonic-gate 		    DB_MAX_RECORDS, 0)) != 0 && ret == DB_NOTFOUND)
164*7c478bd9Sstevel@tonic-gate 			ret = 0;
165*7c478bd9Sstevel@tonic-gate 
166*7c478bd9Sstevel@tonic-gate 		/* Discard the cursor. */
167*7c478bd9Sstevel@tonic-gate 		if ((t_ret = dbc->c_close(dbc)) != 0 && ret == 0)
168*7c478bd9Sstevel@tonic-gate 			ret = t_ret;
169*7c478bd9Sstevel@tonic-gate 
170*7c478bd9Sstevel@tonic-gate 		if (ret != 0)
171*7c478bd9Sstevel@tonic-gate 			goto err;
172*7c478bd9Sstevel@tonic-gate 	}
173*7c478bd9Sstevel@tonic-gate 
174*7c478bd9Sstevel@tonic-gate 	return (0);
175*7c478bd9Sstevel@tonic-gate 
176*7c478bd9Sstevel@tonic-gate err:	/* If we mmap'd a source file, discard it. */
177*7c478bd9Sstevel@tonic-gate 	if (rp->re_smap != NULL)
178*7c478bd9Sstevel@tonic-gate 		(void)__db_unmapfile(rp->re_smap, rp->re_msize);
179*7c478bd9Sstevel@tonic-gate 
180*7c478bd9Sstevel@tonic-gate 	/* If we opened a source file, discard it. */
181*7c478bd9Sstevel@tonic-gate 	if (rp->re_fd != -1)
182*7c478bd9Sstevel@tonic-gate 		(void)__os_close(rp->re_fd);
183*7c478bd9Sstevel@tonic-gate 	if (rp->re_source != NULL)
184*7c478bd9Sstevel@tonic-gate 		__os_freestr(rp->re_source);
185*7c478bd9Sstevel@tonic-gate 
186*7c478bd9Sstevel@tonic-gate 	__os_free(rp, sizeof(*rp));
187*7c478bd9Sstevel@tonic-gate 
188*7c478bd9Sstevel@tonic-gate 	return (ret);
189*7c478bd9Sstevel@tonic-gate }
190*7c478bd9Sstevel@tonic-gate 
191*7c478bd9Sstevel@tonic-gate /*
192*7c478bd9Sstevel@tonic-gate  * __ram_delete --
193*7c478bd9Sstevel@tonic-gate  *	Recno db->del function.
194*7c478bd9Sstevel@tonic-gate  */
195*7c478bd9Sstevel@tonic-gate static int
__ram_delete(dbp,txn,key,flags)196*7c478bd9Sstevel@tonic-gate __ram_delete(dbp, txn, key, flags)
197*7c478bd9Sstevel@tonic-gate 	DB *dbp;
198*7c478bd9Sstevel@tonic-gate 	DB_TXN *txn;
199*7c478bd9Sstevel@tonic-gate 	DBT *key;
200*7c478bd9Sstevel@tonic-gate 	u_int32_t flags;
201*7c478bd9Sstevel@tonic-gate {
202*7c478bd9Sstevel@tonic-gate 	CURSOR *cp;
203*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
204*7c478bd9Sstevel@tonic-gate 	db_recno_t recno;
205*7c478bd9Sstevel@tonic-gate 	int ret, t_ret;
206*7c478bd9Sstevel@tonic-gate 
207*7c478bd9Sstevel@tonic-gate 	DB_PANIC_CHECK(dbp);
208*7c478bd9Sstevel@tonic-gate 
209*7c478bd9Sstevel@tonic-gate 	/* Check for invalid flags. */
210*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_delchk(dbp,
211*7c478bd9Sstevel@tonic-gate 	    key, flags, F_ISSET(dbp, DB_AM_RDONLY))) != 0)
212*7c478bd9Sstevel@tonic-gate 		return (ret);
213*7c478bd9Sstevel@tonic-gate 
214*7c478bd9Sstevel@tonic-gate 	/* Acquire a cursor. */
215*7c478bd9Sstevel@tonic-gate 	if ((ret = dbp->cursor(dbp, txn, &dbc, DB_WRITELOCK)) != 0)
216*7c478bd9Sstevel@tonic-gate 		return (ret);
217*7c478bd9Sstevel@tonic-gate 
218*7c478bd9Sstevel@tonic-gate 	DEBUG_LWRITE(dbc, txn, "ram_delete", key, NULL, flags);
219*7c478bd9Sstevel@tonic-gate 
220*7c478bd9Sstevel@tonic-gate 	/* Check the user's record number and fill in as necessary. */
221*7c478bd9Sstevel@tonic-gate 	if ((ret = __ram_getno(dbc, key, &recno, 0)) != 0)
222*7c478bd9Sstevel@tonic-gate 		goto err;
223*7c478bd9Sstevel@tonic-gate 
224*7c478bd9Sstevel@tonic-gate 	/* Do the delete. */
225*7c478bd9Sstevel@tonic-gate 	cp = dbc->internal;
226*7c478bd9Sstevel@tonic-gate 	cp->recno = recno;
227*7c478bd9Sstevel@tonic-gate 	ret = __ram_i_delete(dbc);
228*7c478bd9Sstevel@tonic-gate 
229*7c478bd9Sstevel@tonic-gate 	/* Release the cursor. */
230*7c478bd9Sstevel@tonic-gate err:	if ((t_ret = dbc->c_close(dbc)) != 0 && ret == 0)
231*7c478bd9Sstevel@tonic-gate 		ret = t_ret;
232*7c478bd9Sstevel@tonic-gate 
233*7c478bd9Sstevel@tonic-gate 	return (ret);
234*7c478bd9Sstevel@tonic-gate }
235*7c478bd9Sstevel@tonic-gate 
236*7c478bd9Sstevel@tonic-gate /*
237*7c478bd9Sstevel@tonic-gate  * __ram_i_delete --
238*7c478bd9Sstevel@tonic-gate  *	Internal version of recno delete, called by __ram_delete and
239*7c478bd9Sstevel@tonic-gate  *	__ram_c_del.
240*7c478bd9Sstevel@tonic-gate  */
241*7c478bd9Sstevel@tonic-gate static int
__ram_i_delete(dbc)242*7c478bd9Sstevel@tonic-gate __ram_i_delete(dbc)
243*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
244*7c478bd9Sstevel@tonic-gate {
245*7c478bd9Sstevel@tonic-gate 	BKEYDATA bk;
246*7c478bd9Sstevel@tonic-gate 	BTREE *t;
247*7c478bd9Sstevel@tonic-gate 	CURSOR *cp;
248*7c478bd9Sstevel@tonic-gate 	DB *dbp;
249*7c478bd9Sstevel@tonic-gate 	DBT hdr, data;
250*7c478bd9Sstevel@tonic-gate 	PAGE *h;
251*7c478bd9Sstevel@tonic-gate 	db_indx_t indx;
252*7c478bd9Sstevel@tonic-gate 	int exact, ret, stack;
253*7c478bd9Sstevel@tonic-gate 
254*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
255*7c478bd9Sstevel@tonic-gate 	cp = dbc->internal;
256*7c478bd9Sstevel@tonic-gate 	t = dbp->internal;
257*7c478bd9Sstevel@tonic-gate 	stack = 0;
258*7c478bd9Sstevel@tonic-gate 
259*7c478bd9Sstevel@tonic-gate 	/*
260*7c478bd9Sstevel@tonic-gate 	 * If this is CDB and this isn't a write cursor, then it's an error.
261*7c478bd9Sstevel@tonic-gate 	 * If it is a write cursor, but we don't yet hold the write lock, then
262*7c478bd9Sstevel@tonic-gate 	 * we need to upgrade to the write lock.
263*7c478bd9Sstevel@tonic-gate 	 */
264*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_AM_CDB)) {
265*7c478bd9Sstevel@tonic-gate 		/* Make sure it's a valid update cursor. */
266*7c478bd9Sstevel@tonic-gate 		if (!F_ISSET(dbc, DBC_RMW | DBC_WRITER))
267*7c478bd9Sstevel@tonic-gate 			return (EINVAL);
268*7c478bd9Sstevel@tonic-gate 
269*7c478bd9Sstevel@tonic-gate 		if (F_ISSET(dbc, DBC_RMW) &&
270*7c478bd9Sstevel@tonic-gate 		    (ret = lock_get(dbp->dbenv->lk_info, dbc->locker,
271*7c478bd9Sstevel@tonic-gate 		    DB_LOCK_UPGRADE, &dbc->lock_dbt, DB_LOCK_WRITE,
272*7c478bd9Sstevel@tonic-gate 		    &dbc->mylock)) != 0)
273*7c478bd9Sstevel@tonic-gate 			return (EAGAIN);
274*7c478bd9Sstevel@tonic-gate 	}
275*7c478bd9Sstevel@tonic-gate 
276*7c478bd9Sstevel@tonic-gate 	/* Search the tree for the key; delete only deletes exact matches. */
277*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_rsearch(dbc, &cp->recno, S_DELETE, 1, &exact)) != 0)
278*7c478bd9Sstevel@tonic-gate 		goto err;
279*7c478bd9Sstevel@tonic-gate 	if (!exact) {
280*7c478bd9Sstevel@tonic-gate 		ret = DB_NOTFOUND;
281*7c478bd9Sstevel@tonic-gate 		goto err;
282*7c478bd9Sstevel@tonic-gate 	}
283*7c478bd9Sstevel@tonic-gate 	stack = 1;
284*7c478bd9Sstevel@tonic-gate 
285*7c478bd9Sstevel@tonic-gate 	h = cp->csp->page;
286*7c478bd9Sstevel@tonic-gate 	indx = cp->csp->indx;
287*7c478bd9Sstevel@tonic-gate 
288*7c478bd9Sstevel@tonic-gate 	/*
289*7c478bd9Sstevel@tonic-gate 	 * If re-numbering records, the on-page deleted flag can only mean
290*7c478bd9Sstevel@tonic-gate 	 * that this record was implicitly created.  Applications aren't
291*7c478bd9Sstevel@tonic-gate 	 * permitted to delete records they never created, return an error.
292*7c478bd9Sstevel@tonic-gate 	 *
293*7c478bd9Sstevel@tonic-gate 	 * If not re-numbering records, the on-page deleted flag means that
294*7c478bd9Sstevel@tonic-gate 	 * this record was implicitly created, or, was deleted at some time.
295*7c478bd9Sstevel@tonic-gate 	 * The former is an error because applications aren't permitted to
296*7c478bd9Sstevel@tonic-gate 	 * delete records they never created, the latter is an error because
297*7c478bd9Sstevel@tonic-gate 	 * if the record was "deleted", we could never have found it.
298*7c478bd9Sstevel@tonic-gate 	 */
299*7c478bd9Sstevel@tonic-gate 	if (B_DISSET(GET_BKEYDATA(h, indx)->type)) {
300*7c478bd9Sstevel@tonic-gate 		ret = DB_KEYEMPTY;
301*7c478bd9Sstevel@tonic-gate 		goto err;
302*7c478bd9Sstevel@tonic-gate 	}
303*7c478bd9Sstevel@tonic-gate 
304*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_RE_RENUMBER)) {
305*7c478bd9Sstevel@tonic-gate 		/* Delete the item, adjust the counts, adjust the cursors. */
306*7c478bd9Sstevel@tonic-gate 		if ((ret = __bam_ditem(dbc, h, indx)) != 0)
307*7c478bd9Sstevel@tonic-gate 			goto err;
308*7c478bd9Sstevel@tonic-gate 		__bam_adjust(dbc, -1);
309*7c478bd9Sstevel@tonic-gate 		__ram_ca(dbp, cp->recno, CA_DELETE);
310*7c478bd9Sstevel@tonic-gate 
311*7c478bd9Sstevel@tonic-gate 		/*
312*7c478bd9Sstevel@tonic-gate 		 * If the page is empty, delete it.   The whole tree is locked
313*7c478bd9Sstevel@tonic-gate 		 * so there are no preparations to make.
314*7c478bd9Sstevel@tonic-gate 		 */
315*7c478bd9Sstevel@tonic-gate 		if (NUM_ENT(h) == 0 && h->pgno != PGNO_ROOT) {
316*7c478bd9Sstevel@tonic-gate 			stack = 0;
317*7c478bd9Sstevel@tonic-gate 			ret = __bam_dpages(dbc);
318*7c478bd9Sstevel@tonic-gate 		}
319*7c478bd9Sstevel@tonic-gate 	} else {
320*7c478bd9Sstevel@tonic-gate 		/* Use a delete/put pair to replace the record with a marker. */
321*7c478bd9Sstevel@tonic-gate 		if ((ret = __bam_ditem(dbc, h, indx)) != 0)
322*7c478bd9Sstevel@tonic-gate 			goto err;
323*7c478bd9Sstevel@tonic-gate 
324*7c478bd9Sstevel@tonic-gate 		B_TSET(bk.type, B_KEYDATA, 1);
325*7c478bd9Sstevel@tonic-gate 		bk.len = 0;
326*7c478bd9Sstevel@tonic-gate 		memset(&hdr, 0, sizeof(hdr));
327*7c478bd9Sstevel@tonic-gate 		hdr.data = &bk;
328*7c478bd9Sstevel@tonic-gate 		hdr.size = SSZA(BKEYDATA, data);
329*7c478bd9Sstevel@tonic-gate 		memset(&data, 0, sizeof(data));
330*7c478bd9Sstevel@tonic-gate 		data.data = (char *)"";
331*7c478bd9Sstevel@tonic-gate 		data.size = 0;
332*7c478bd9Sstevel@tonic-gate 		if ((ret = __db_pitem(dbc,
333*7c478bd9Sstevel@tonic-gate 		    h, indx, BKEYDATA_SIZE(0), &hdr, &data)) != 0)
334*7c478bd9Sstevel@tonic-gate 			goto err;
335*7c478bd9Sstevel@tonic-gate 	}
336*7c478bd9Sstevel@tonic-gate 	F_SET(t->recno, RECNO_MODIFIED);
337*7c478bd9Sstevel@tonic-gate 
338*7c478bd9Sstevel@tonic-gate err:	if (stack)
339*7c478bd9Sstevel@tonic-gate 		__bam_stkrel(dbc, 0);
340*7c478bd9Sstevel@tonic-gate 
341*7c478bd9Sstevel@tonic-gate 	/* If we upgraded the CDB lock upon entry; downgrade it now. */
342*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_AM_CDB) && F_ISSET(dbc, DBC_RMW))
343*7c478bd9Sstevel@tonic-gate 		(void)__lock_downgrade(dbp->dbenv->lk_info, dbc->mylock,
344*7c478bd9Sstevel@tonic-gate 		    DB_LOCK_IWRITE, 0);
345*7c478bd9Sstevel@tonic-gate 	return (ret);
346*7c478bd9Sstevel@tonic-gate }
347*7c478bd9Sstevel@tonic-gate 
348*7c478bd9Sstevel@tonic-gate /*
349*7c478bd9Sstevel@tonic-gate  * __ram_put --
350*7c478bd9Sstevel@tonic-gate  *	Recno db->put function.
351*7c478bd9Sstevel@tonic-gate  */
352*7c478bd9Sstevel@tonic-gate static int
__ram_put(dbp,txn,key,data,flags)353*7c478bd9Sstevel@tonic-gate __ram_put(dbp, txn, key, data, flags)
354*7c478bd9Sstevel@tonic-gate 	DB *dbp;
355*7c478bd9Sstevel@tonic-gate 	DB_TXN *txn;
356*7c478bd9Sstevel@tonic-gate 	DBT *key, *data;
357*7c478bd9Sstevel@tonic-gate 	u_int32_t flags;
358*7c478bd9Sstevel@tonic-gate {
359*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
360*7c478bd9Sstevel@tonic-gate 	db_recno_t recno;
361*7c478bd9Sstevel@tonic-gate 	int ret, t_ret;
362*7c478bd9Sstevel@tonic-gate 
363*7c478bd9Sstevel@tonic-gate 	DB_PANIC_CHECK(dbp);
364*7c478bd9Sstevel@tonic-gate 
365*7c478bd9Sstevel@tonic-gate 	/* Check for invalid flags. */
366*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_putchk(dbp,
367*7c478bd9Sstevel@tonic-gate 	    key, data, flags, F_ISSET(dbp, DB_AM_RDONLY), 0)) != 0)
368*7c478bd9Sstevel@tonic-gate 		return (ret);
369*7c478bd9Sstevel@tonic-gate 
370*7c478bd9Sstevel@tonic-gate 	/* Allocate a cursor. */
371*7c478bd9Sstevel@tonic-gate 	if ((ret = dbp->cursor(dbp, txn, &dbc, DB_WRITELOCK)) != 0)
372*7c478bd9Sstevel@tonic-gate 		return (ret);
373*7c478bd9Sstevel@tonic-gate 
374*7c478bd9Sstevel@tonic-gate 	DEBUG_LWRITE(dbc, txn, "ram_put", key, data, flags);
375*7c478bd9Sstevel@tonic-gate 
376*7c478bd9Sstevel@tonic-gate 	/*
377*7c478bd9Sstevel@tonic-gate 	 * If we're appending to the tree, make sure we've read in all of
378*7c478bd9Sstevel@tonic-gate 	 * the backing source file.  Otherwise, check the user's record
379*7c478bd9Sstevel@tonic-gate 	 * number and fill in as necessary.
380*7c478bd9Sstevel@tonic-gate 	 */
381*7c478bd9Sstevel@tonic-gate 	ret = flags == DB_APPEND ?
382*7c478bd9Sstevel@tonic-gate 	    __ram_update(dbc, DB_MAX_RECORDS, 0) :
383*7c478bd9Sstevel@tonic-gate 	    __ram_getno(dbc, key, &recno, 1);
384*7c478bd9Sstevel@tonic-gate 
385*7c478bd9Sstevel@tonic-gate 	/* Add the record. */
386*7c478bd9Sstevel@tonic-gate 	if (ret == 0)
387*7c478bd9Sstevel@tonic-gate 		ret = __ram_add(dbc, &recno, data, flags, 0);
388*7c478bd9Sstevel@tonic-gate 
389*7c478bd9Sstevel@tonic-gate 	/* Discard the cursor. */
390*7c478bd9Sstevel@tonic-gate 	if ((t_ret = dbc->c_close(dbc)) != 0 && ret == 0)
391*7c478bd9Sstevel@tonic-gate 		ret = t_ret;
392*7c478bd9Sstevel@tonic-gate 
393*7c478bd9Sstevel@tonic-gate 	/* Return the record number if we're appending to the tree. */
394*7c478bd9Sstevel@tonic-gate 	if (ret == 0 && flags == DB_APPEND)
395*7c478bd9Sstevel@tonic-gate 		*(db_recno_t *)key->data = recno;
396*7c478bd9Sstevel@tonic-gate 
397*7c478bd9Sstevel@tonic-gate 	return (ret);
398*7c478bd9Sstevel@tonic-gate }
399*7c478bd9Sstevel@tonic-gate 
400*7c478bd9Sstevel@tonic-gate /*
401*7c478bd9Sstevel@tonic-gate  * __ram_sync --
402*7c478bd9Sstevel@tonic-gate  *	Recno db->sync function.
403*7c478bd9Sstevel@tonic-gate  */
404*7c478bd9Sstevel@tonic-gate static int
__ram_sync(dbp,flags)405*7c478bd9Sstevel@tonic-gate __ram_sync(dbp, flags)
406*7c478bd9Sstevel@tonic-gate 	DB *dbp;
407*7c478bd9Sstevel@tonic-gate 	u_int32_t flags;
408*7c478bd9Sstevel@tonic-gate {
409*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
410*7c478bd9Sstevel@tonic-gate 	int ret, t_ret;
411*7c478bd9Sstevel@tonic-gate 
412*7c478bd9Sstevel@tonic-gate 	/*
413*7c478bd9Sstevel@tonic-gate 	 * Sync the underlying btree.
414*7c478bd9Sstevel@tonic-gate 	 *
415*7c478bd9Sstevel@tonic-gate 	 * !!!
416*7c478bd9Sstevel@tonic-gate 	 * We don't need to do a panic check or flags check, the "real"
417*7c478bd9Sstevel@tonic-gate 	 * sync function does all that for us.
418*7c478bd9Sstevel@tonic-gate 	 */
419*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_sync(dbp, flags)) != 0)
420*7c478bd9Sstevel@tonic-gate 		return (ret);
421*7c478bd9Sstevel@tonic-gate 
422*7c478bd9Sstevel@tonic-gate 	/* Allocate a cursor. */
423*7c478bd9Sstevel@tonic-gate 	if ((ret = dbp->cursor(dbp, NULL, &dbc, 0)) != 0)
424*7c478bd9Sstevel@tonic-gate 		return (ret);
425*7c478bd9Sstevel@tonic-gate 
426*7c478bd9Sstevel@tonic-gate 	DEBUG_LWRITE(dbc, NULL, "ram_sync", NULL, NULL, flags);
427*7c478bd9Sstevel@tonic-gate 
428*7c478bd9Sstevel@tonic-gate 	/* Copy back the backing source file. */
429*7c478bd9Sstevel@tonic-gate 	ret = __ram_writeback(dbc);
430*7c478bd9Sstevel@tonic-gate 
431*7c478bd9Sstevel@tonic-gate 	/* Discard the cursor. */
432*7c478bd9Sstevel@tonic-gate 	if ((t_ret = dbc->c_close(dbc)) != 0 && ret == 0)
433*7c478bd9Sstevel@tonic-gate 		ret = t_ret;
434*7c478bd9Sstevel@tonic-gate 
435*7c478bd9Sstevel@tonic-gate 	return (ret);
436*7c478bd9Sstevel@tonic-gate }
437*7c478bd9Sstevel@tonic-gate 
438*7c478bd9Sstevel@tonic-gate /*
439*7c478bd9Sstevel@tonic-gate  * __ram_close --
440*7c478bd9Sstevel@tonic-gate  *	Recno db->close function.
441*7c478bd9Sstevel@tonic-gate  *
442*7c478bd9Sstevel@tonic-gate  * PUBLIC: int __ram_close __P((DB *));
443*7c478bd9Sstevel@tonic-gate  */
444*7c478bd9Sstevel@tonic-gate int
__ram_close(dbp)445*7c478bd9Sstevel@tonic-gate __ram_close(dbp)
446*7c478bd9Sstevel@tonic-gate 	DB *dbp;
447*7c478bd9Sstevel@tonic-gate {
448*7c478bd9Sstevel@tonic-gate 	RECNO *rp;
449*7c478bd9Sstevel@tonic-gate 
450*7c478bd9Sstevel@tonic-gate 	rp = ((BTREE *)dbp->internal)->recno;
451*7c478bd9Sstevel@tonic-gate 
452*7c478bd9Sstevel@tonic-gate 	/* Close any underlying mmap region. */
453*7c478bd9Sstevel@tonic-gate 	if (rp->re_smap != NULL)
454*7c478bd9Sstevel@tonic-gate 		(void)__db_unmapfile(rp->re_smap, rp->re_msize);
455*7c478bd9Sstevel@tonic-gate 
456*7c478bd9Sstevel@tonic-gate 	/* Close any backing source file descriptor. */
457*7c478bd9Sstevel@tonic-gate 	if (rp->re_fd != -1)
458*7c478bd9Sstevel@tonic-gate 		(void)__os_close(rp->re_fd);
459*7c478bd9Sstevel@tonic-gate 
460*7c478bd9Sstevel@tonic-gate 	/* Free any backing source file name. */
461*7c478bd9Sstevel@tonic-gate 	if (rp->re_source != NULL)
462*7c478bd9Sstevel@tonic-gate 		__os_freestr(rp->re_source);
463*7c478bd9Sstevel@tonic-gate 
464*7c478bd9Sstevel@tonic-gate 	/* Free allocated memory. */
465*7c478bd9Sstevel@tonic-gate 	__os_free(rp, sizeof(RECNO));
466*7c478bd9Sstevel@tonic-gate 	((BTREE *)dbp->internal)->recno = NULL;
467*7c478bd9Sstevel@tonic-gate 
468*7c478bd9Sstevel@tonic-gate 	/* Close the underlying btree. */
469*7c478bd9Sstevel@tonic-gate 	return (__bam_close(dbp));
470*7c478bd9Sstevel@tonic-gate }
471*7c478bd9Sstevel@tonic-gate 
472*7c478bd9Sstevel@tonic-gate /*
473*7c478bd9Sstevel@tonic-gate  * __ram_c_del --
474*7c478bd9Sstevel@tonic-gate  *	Recno cursor->c_del function.
475*7c478bd9Sstevel@tonic-gate  *
476*7c478bd9Sstevel@tonic-gate  * PUBLIC: int __ram_c_del __P((DBC *, u_int32_t));
477*7c478bd9Sstevel@tonic-gate  */
478*7c478bd9Sstevel@tonic-gate int
__ram_c_del(dbc,flags)479*7c478bd9Sstevel@tonic-gate __ram_c_del(dbc, flags)
480*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
481*7c478bd9Sstevel@tonic-gate 	u_int32_t flags;
482*7c478bd9Sstevel@tonic-gate {
483*7c478bd9Sstevel@tonic-gate 	CURSOR *cp;
484*7c478bd9Sstevel@tonic-gate 	DB *dbp;
485*7c478bd9Sstevel@tonic-gate 	int ret;
486*7c478bd9Sstevel@tonic-gate 
487*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
488*7c478bd9Sstevel@tonic-gate 	cp = dbc->internal;
489*7c478bd9Sstevel@tonic-gate 
490*7c478bd9Sstevel@tonic-gate 	DB_PANIC_CHECK(dbp);
491*7c478bd9Sstevel@tonic-gate 
492*7c478bd9Sstevel@tonic-gate 	/* Check for invalid flags. */
493*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_cdelchk(dbp, flags,
494*7c478bd9Sstevel@tonic-gate 	    F_ISSET(dbp, DB_AM_RDONLY), cp->recno != RECNO_OOB)) != 0)
495*7c478bd9Sstevel@tonic-gate 		return (ret);
496*7c478bd9Sstevel@tonic-gate 
497*7c478bd9Sstevel@tonic-gate 	DEBUG_LWRITE(dbc, dbc->txn, "ram_c_del", NULL, NULL, flags);
498*7c478bd9Sstevel@tonic-gate 
499*7c478bd9Sstevel@tonic-gate 	/*
500*7c478bd9Sstevel@tonic-gate 	 * If we are running CDB, this had better be either a write
501*7c478bd9Sstevel@tonic-gate 	 * cursor or an immediate writer.
502*7c478bd9Sstevel@tonic-gate 	 */
503*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_AM_CDB))
504*7c478bd9Sstevel@tonic-gate 		if (!F_ISSET(dbc, DBC_RMW | DBC_WRITER))
505*7c478bd9Sstevel@tonic-gate 			return (EINVAL);
506*7c478bd9Sstevel@tonic-gate 
507*7c478bd9Sstevel@tonic-gate 	/*
508*7c478bd9Sstevel@tonic-gate 	 * The semantics of cursors during delete are as follows: if record
509*7c478bd9Sstevel@tonic-gate 	 * numbers are mutable (DB_RE_RENUMBER is set), deleting a record
510*7c478bd9Sstevel@tonic-gate 	 * causes the cursor to automatically point to the record immediately
511*7c478bd9Sstevel@tonic-gate 	 * following.  In this case it is possible to use a single cursor for
512*7c478bd9Sstevel@tonic-gate 	 * repeated delete operations, without intervening operations.
513*7c478bd9Sstevel@tonic-gate 	 *
514*7c478bd9Sstevel@tonic-gate 	 * If record numbers are not mutable, then records are replaced with
515*7c478bd9Sstevel@tonic-gate 	 * a marker containing a delete flag.  If the record referenced by
516*7c478bd9Sstevel@tonic-gate 	 * this cursor has already been deleted, we will detect that as part
517*7c478bd9Sstevel@tonic-gate 	 * of the delete operation, and fail.
518*7c478bd9Sstevel@tonic-gate 	 */
519*7c478bd9Sstevel@tonic-gate 	return (__ram_i_delete(dbc));
520*7c478bd9Sstevel@tonic-gate }
521*7c478bd9Sstevel@tonic-gate 
522*7c478bd9Sstevel@tonic-gate /*
523*7c478bd9Sstevel@tonic-gate  * __ram_c_get --
524*7c478bd9Sstevel@tonic-gate  *	Recno cursor->c_get function.
525*7c478bd9Sstevel@tonic-gate  *
526*7c478bd9Sstevel@tonic-gate  * PUBLIC: int __ram_c_get __P((DBC *, DBT *, DBT *, u_int32_t));
527*7c478bd9Sstevel@tonic-gate  */
528*7c478bd9Sstevel@tonic-gate int
__ram_c_get(dbc,key,data,flags)529*7c478bd9Sstevel@tonic-gate __ram_c_get(dbc, key, data, flags)
530*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
531*7c478bd9Sstevel@tonic-gate 	DBT *key, *data;
532*7c478bd9Sstevel@tonic-gate 	u_int32_t flags;
533*7c478bd9Sstevel@tonic-gate {
534*7c478bd9Sstevel@tonic-gate 	CURSOR *cp, copy;
535*7c478bd9Sstevel@tonic-gate 	DB *dbp;
536*7c478bd9Sstevel@tonic-gate 	PAGE *h;
537*7c478bd9Sstevel@tonic-gate 	db_indx_t indx;
538*7c478bd9Sstevel@tonic-gate 	int exact, ret, stack, tmp_rmw;
539*7c478bd9Sstevel@tonic-gate 
540*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
541*7c478bd9Sstevel@tonic-gate 	cp = dbc->internal;
542*7c478bd9Sstevel@tonic-gate 
543*7c478bd9Sstevel@tonic-gate 	DB_PANIC_CHECK(dbp);
544*7c478bd9Sstevel@tonic-gate 
545*7c478bd9Sstevel@tonic-gate 	/* Check for invalid flags. */
546*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_cgetchk(dbc->dbp,
547*7c478bd9Sstevel@tonic-gate 	    key, data, flags, cp->recno != RECNO_OOB)) != 0)
548*7c478bd9Sstevel@tonic-gate 		return (ret);
549*7c478bd9Sstevel@tonic-gate 
550*7c478bd9Sstevel@tonic-gate 	/* Clear OR'd in additional bits so we can check for flag equality. */
551*7c478bd9Sstevel@tonic-gate 	tmp_rmw = 0;
552*7c478bd9Sstevel@tonic-gate 	if (LF_ISSET(DB_RMW)) {
553*7c478bd9Sstevel@tonic-gate 		if (!F_ISSET(dbp, DB_AM_CDB)) {
554*7c478bd9Sstevel@tonic-gate 			tmp_rmw = 1;
555*7c478bd9Sstevel@tonic-gate 			F_SET(dbc, DBC_RMW);
556*7c478bd9Sstevel@tonic-gate 		}
557*7c478bd9Sstevel@tonic-gate 		LF_CLR(DB_RMW);
558*7c478bd9Sstevel@tonic-gate 	}
559*7c478bd9Sstevel@tonic-gate 
560*7c478bd9Sstevel@tonic-gate 	DEBUG_LREAD(dbc, dbc->txn, "ram_c_get",
561*7c478bd9Sstevel@tonic-gate 	    flags == DB_SET || flags == DB_SET_RANGE ? key : NULL, NULL, flags);
562*7c478bd9Sstevel@tonic-gate 
563*7c478bd9Sstevel@tonic-gate 	/* Initialize the cursor for a new retrieval. */
564*7c478bd9Sstevel@tonic-gate 	copy = *cp;
565*7c478bd9Sstevel@tonic-gate 
566*7c478bd9Sstevel@tonic-gate retry:	/* Update the record number. */
567*7c478bd9Sstevel@tonic-gate 	stack = 0;
568*7c478bd9Sstevel@tonic-gate 	switch (flags) {
569*7c478bd9Sstevel@tonic-gate 	case DB_CURRENT:
570*7c478bd9Sstevel@tonic-gate 		/*
571*7c478bd9Sstevel@tonic-gate 		 * If record numbers are mutable: if we just deleted a record,
572*7c478bd9Sstevel@tonic-gate 		 * there is no action necessary, we return the record following
573*7c478bd9Sstevel@tonic-gate 		 * the deleted item by virtue of renumbering the tree.
574*7c478bd9Sstevel@tonic-gate 		 */
575*7c478bd9Sstevel@tonic-gate 		break;
576*7c478bd9Sstevel@tonic-gate 	case DB_NEXT:
577*7c478bd9Sstevel@tonic-gate 		/*
578*7c478bd9Sstevel@tonic-gate 		 * If record numbers are mutable: if we just deleted a record,
579*7c478bd9Sstevel@tonic-gate 		 * we have to avoid incrementing the record number so that we
580*7c478bd9Sstevel@tonic-gate 		 * return the right record by virtue of renumbering the tree.
581*7c478bd9Sstevel@tonic-gate 		 */
582*7c478bd9Sstevel@tonic-gate 		if (CD_ISSET(dbp, cp))
583*7c478bd9Sstevel@tonic-gate 			break;
584*7c478bd9Sstevel@tonic-gate 
585*7c478bd9Sstevel@tonic-gate 		if (cp->recno != RECNO_OOB) {
586*7c478bd9Sstevel@tonic-gate 			++cp->recno;
587*7c478bd9Sstevel@tonic-gate 			break;
588*7c478bd9Sstevel@tonic-gate 		}
589*7c478bd9Sstevel@tonic-gate 		/* FALLTHROUGH */
590*7c478bd9Sstevel@tonic-gate 	case DB_FIRST:
591*7c478bd9Sstevel@tonic-gate 		flags = DB_NEXT;
592*7c478bd9Sstevel@tonic-gate 		cp->recno = 1;
593*7c478bd9Sstevel@tonic-gate 		break;
594*7c478bd9Sstevel@tonic-gate 	case DB_PREV:
595*7c478bd9Sstevel@tonic-gate 		if (cp->recno != RECNO_OOB) {
596*7c478bd9Sstevel@tonic-gate 			if (cp->recno == 1) {
597*7c478bd9Sstevel@tonic-gate 				ret = DB_NOTFOUND;
598*7c478bd9Sstevel@tonic-gate 				goto err;
599*7c478bd9Sstevel@tonic-gate 			}
600*7c478bd9Sstevel@tonic-gate 			--cp->recno;
601*7c478bd9Sstevel@tonic-gate 			break;
602*7c478bd9Sstevel@tonic-gate 		}
603*7c478bd9Sstevel@tonic-gate 		/* FALLTHROUGH */
604*7c478bd9Sstevel@tonic-gate 	case DB_LAST:
605*7c478bd9Sstevel@tonic-gate 		flags = DB_PREV;
606*7c478bd9Sstevel@tonic-gate 		if (((ret = __ram_update(dbc,
607*7c478bd9Sstevel@tonic-gate 		    DB_MAX_RECORDS, 0)) != 0) && ret != DB_NOTFOUND)
608*7c478bd9Sstevel@tonic-gate 			goto err;
609*7c478bd9Sstevel@tonic-gate 		if ((ret = __bam_nrecs(dbc, &cp->recno)) != 0)
610*7c478bd9Sstevel@tonic-gate 			goto err;
611*7c478bd9Sstevel@tonic-gate 		if (cp->recno == 0) {
612*7c478bd9Sstevel@tonic-gate 			ret = DB_NOTFOUND;
613*7c478bd9Sstevel@tonic-gate 			goto err;
614*7c478bd9Sstevel@tonic-gate 		}
615*7c478bd9Sstevel@tonic-gate 		break;
616*7c478bd9Sstevel@tonic-gate 	case DB_SET:
617*7c478bd9Sstevel@tonic-gate 	case DB_SET_RANGE:
618*7c478bd9Sstevel@tonic-gate 		if ((ret = __ram_getno(dbc, key, &cp->recno, 0)) != 0)
619*7c478bd9Sstevel@tonic-gate 			goto err;
620*7c478bd9Sstevel@tonic-gate 		break;
621*7c478bd9Sstevel@tonic-gate 	}
622*7c478bd9Sstevel@tonic-gate 
623*7c478bd9Sstevel@tonic-gate 	/* Return the key if the user didn't give us one. */
624*7c478bd9Sstevel@tonic-gate 	if (flags != DB_SET && flags != DB_SET_RANGE &&
625*7c478bd9Sstevel@tonic-gate 	    (ret = __db_retcopy(key, &cp->recno, sizeof(cp->recno),
626*7c478bd9Sstevel@tonic-gate 	    &dbc->rkey.data, &dbc->rkey.ulen, dbp->db_malloc)) != 0)
627*7c478bd9Sstevel@tonic-gate 		goto err;
628*7c478bd9Sstevel@tonic-gate 
629*7c478bd9Sstevel@tonic-gate 	/* Search the tree for the record. */
630*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_rsearch(dbc, &cp->recno,
631*7c478bd9Sstevel@tonic-gate 	    F_ISSET(dbc, DBC_RMW) ? S_FIND_WR : S_FIND, 1, &exact)) != 0)
632*7c478bd9Sstevel@tonic-gate 		goto err;
633*7c478bd9Sstevel@tonic-gate 	stack = 1;
634*7c478bd9Sstevel@tonic-gate 	if (!exact) {
635*7c478bd9Sstevel@tonic-gate 		ret = DB_NOTFOUND;
636*7c478bd9Sstevel@tonic-gate 		goto err;
637*7c478bd9Sstevel@tonic-gate 	}
638*7c478bd9Sstevel@tonic-gate 	h = cp->csp->page;
639*7c478bd9Sstevel@tonic-gate 	indx = cp->csp->indx;
640*7c478bd9Sstevel@tonic-gate 
641*7c478bd9Sstevel@tonic-gate 	/*
642*7c478bd9Sstevel@tonic-gate 	 * If re-numbering records, the on-page deleted flag means this record
643*7c478bd9Sstevel@tonic-gate 	 * was implicitly created.  If not re-numbering records, the on-page
644*7c478bd9Sstevel@tonic-gate 	 * deleted flag means this record was implicitly created, or, it was
645*7c478bd9Sstevel@tonic-gate 	 * deleted at some time.  Regardless, we skip such records if doing
646*7c478bd9Sstevel@tonic-gate 	 * cursor next/prev operations, and fail if the application requested
647*7c478bd9Sstevel@tonic-gate 	 * them explicitly.
648*7c478bd9Sstevel@tonic-gate 	 */
649*7c478bd9Sstevel@tonic-gate 	if (B_DISSET(GET_BKEYDATA(h, indx)->type)) {
650*7c478bd9Sstevel@tonic-gate 		if (flags == DB_NEXT || flags == DB_PREV) {
651*7c478bd9Sstevel@tonic-gate 			(void)__bam_stkrel(dbc, 0);
652*7c478bd9Sstevel@tonic-gate 			goto retry;
653*7c478bd9Sstevel@tonic-gate 		}
654*7c478bd9Sstevel@tonic-gate 		ret = DB_KEYEMPTY;
655*7c478bd9Sstevel@tonic-gate 		goto err;
656*7c478bd9Sstevel@tonic-gate 	}
657*7c478bd9Sstevel@tonic-gate 
658*7c478bd9Sstevel@tonic-gate 	/* Return the data item. */
659*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_ret(dbp,
660*7c478bd9Sstevel@tonic-gate 	    h, indx, data, &dbc->rdata.data, &dbc->rdata.ulen)) != 0)
661*7c478bd9Sstevel@tonic-gate 		goto err;
662*7c478bd9Sstevel@tonic-gate 
663*7c478bd9Sstevel@tonic-gate 	/* The cursor was reset, no further delete adjustment is necessary. */
664*7c478bd9Sstevel@tonic-gate 	CD_CLR(dbp, cp);
665*7c478bd9Sstevel@tonic-gate 
666*7c478bd9Sstevel@tonic-gate err:	if (stack)
667*7c478bd9Sstevel@tonic-gate 		(void)__bam_stkrel(dbc, 0);
668*7c478bd9Sstevel@tonic-gate 
669*7c478bd9Sstevel@tonic-gate 	/* Release temporary lock upgrade. */
670*7c478bd9Sstevel@tonic-gate 	if (tmp_rmw)
671*7c478bd9Sstevel@tonic-gate 		F_CLR(dbc, DBC_RMW);
672*7c478bd9Sstevel@tonic-gate 
673*7c478bd9Sstevel@tonic-gate 	if (ret != 0)
674*7c478bd9Sstevel@tonic-gate 		*cp = copy;
675*7c478bd9Sstevel@tonic-gate 
676*7c478bd9Sstevel@tonic-gate 	return (ret);
677*7c478bd9Sstevel@tonic-gate }
678*7c478bd9Sstevel@tonic-gate 
679*7c478bd9Sstevel@tonic-gate /*
680*7c478bd9Sstevel@tonic-gate  * __ram_c_put --
681*7c478bd9Sstevel@tonic-gate  *	Recno cursor->c_put function.
682*7c478bd9Sstevel@tonic-gate  *
683*7c478bd9Sstevel@tonic-gate  * PUBLIC: int __ram_c_put __P((DBC *, DBT *, DBT *, u_int32_t));
684*7c478bd9Sstevel@tonic-gate  */
685*7c478bd9Sstevel@tonic-gate int
__ram_c_put(dbc,key,data,flags)686*7c478bd9Sstevel@tonic-gate __ram_c_put(dbc, key, data, flags)
687*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
688*7c478bd9Sstevel@tonic-gate 	DBT *key, *data;
689*7c478bd9Sstevel@tonic-gate 	u_int32_t flags;
690*7c478bd9Sstevel@tonic-gate {
691*7c478bd9Sstevel@tonic-gate 	CURSOR *cp, copy;
692*7c478bd9Sstevel@tonic-gate 	DB *dbp;
693*7c478bd9Sstevel@tonic-gate 	int exact, ret;
694*7c478bd9Sstevel@tonic-gate 	void *arg;
695*7c478bd9Sstevel@tonic-gate 
696*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
697*7c478bd9Sstevel@tonic-gate 	cp = dbc->internal;
698*7c478bd9Sstevel@tonic-gate 
699*7c478bd9Sstevel@tonic-gate 	DB_PANIC_CHECK(dbp);
700*7c478bd9Sstevel@tonic-gate 
701*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_cputchk(dbc->dbp, key, data, flags,
702*7c478bd9Sstevel@tonic-gate 	    F_ISSET(dbc->dbp, DB_AM_RDONLY), cp->recno != RECNO_OOB)) != 0)
703*7c478bd9Sstevel@tonic-gate 		return (ret);
704*7c478bd9Sstevel@tonic-gate 
705*7c478bd9Sstevel@tonic-gate 	DEBUG_LWRITE(dbc, dbc->txn, "ram_c_put", NULL, data, flags);
706*7c478bd9Sstevel@tonic-gate 
707*7c478bd9Sstevel@tonic-gate 	/*
708*7c478bd9Sstevel@tonic-gate 	 * If we are running CDB, this had better be either a write
709*7c478bd9Sstevel@tonic-gate 	 * cursor or an immediate writer.  If it's a regular writer,
710*7c478bd9Sstevel@tonic-gate 	 * that means we have an IWRITE lock and we need to upgrade
711*7c478bd9Sstevel@tonic-gate 	 * it to a write lock.
712*7c478bd9Sstevel@tonic-gate 	 */
713*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_AM_CDB)) {
714*7c478bd9Sstevel@tonic-gate 		if (!F_ISSET(dbc, DBC_RMW | DBC_WRITER))
715*7c478bd9Sstevel@tonic-gate 			return (EINVAL);
716*7c478bd9Sstevel@tonic-gate 
717*7c478bd9Sstevel@tonic-gate 		if (F_ISSET(dbc, DBC_RMW) &&
718*7c478bd9Sstevel@tonic-gate 		    (ret = lock_get(dbp->dbenv->lk_info, dbc->locker,
719*7c478bd9Sstevel@tonic-gate 		    DB_LOCK_UPGRADE, &dbc->lock_dbt, DB_LOCK_WRITE,
720*7c478bd9Sstevel@tonic-gate 		    &dbc->mylock)) != 0)
721*7c478bd9Sstevel@tonic-gate 			return (EAGAIN);
722*7c478bd9Sstevel@tonic-gate 	}
723*7c478bd9Sstevel@tonic-gate 
724*7c478bd9Sstevel@tonic-gate 	/* Initialize the cursor for a new retrieval. */
725*7c478bd9Sstevel@tonic-gate 	copy = *cp;
726*7c478bd9Sstevel@tonic-gate 
727*7c478bd9Sstevel@tonic-gate 	/*
728*7c478bd9Sstevel@tonic-gate 	 * To split, we need a valid key for the page.  Since it's a cursor,
729*7c478bd9Sstevel@tonic-gate 	 * we have to build one.
730*7c478bd9Sstevel@tonic-gate 	 *
731*7c478bd9Sstevel@tonic-gate 	 * The split code discards all short-term locks and stack pages.
732*7c478bd9Sstevel@tonic-gate 	 */
733*7c478bd9Sstevel@tonic-gate 	if (0) {
734*7c478bd9Sstevel@tonic-gate split:		arg = &cp->recno;
735*7c478bd9Sstevel@tonic-gate 		if ((ret = __bam_split(dbc, arg)) != 0)
736*7c478bd9Sstevel@tonic-gate 			goto err;
737*7c478bd9Sstevel@tonic-gate 	}
738*7c478bd9Sstevel@tonic-gate 
739*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_rsearch(dbc, &cp->recno, S_INSERT, 1, &exact)) != 0)
740*7c478bd9Sstevel@tonic-gate 		goto err;
741*7c478bd9Sstevel@tonic-gate 	if (!exact) {
742*7c478bd9Sstevel@tonic-gate 		ret = DB_NOTFOUND;
743*7c478bd9Sstevel@tonic-gate 		goto err;
744*7c478bd9Sstevel@tonic-gate 	}
745*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_iitem(dbc, &cp->csp->page,
746*7c478bd9Sstevel@tonic-gate 	    &cp->csp->indx, key, data, flags, 0)) == DB_NEEDSPLIT) {
747*7c478bd9Sstevel@tonic-gate 		if ((ret = __bam_stkrel(dbc, 0)) != 0)
748*7c478bd9Sstevel@tonic-gate 			goto err;
749*7c478bd9Sstevel@tonic-gate 		goto split;
750*7c478bd9Sstevel@tonic-gate 	}
751*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_stkrel(dbc, 0)) != 0)
752*7c478bd9Sstevel@tonic-gate 		goto err;
753*7c478bd9Sstevel@tonic-gate 
754*7c478bd9Sstevel@tonic-gate 	switch (flags) {
755*7c478bd9Sstevel@tonic-gate 	case DB_AFTER:
756*7c478bd9Sstevel@tonic-gate 		/* Adjust the cursors. */
757*7c478bd9Sstevel@tonic-gate 		__ram_ca(dbp, cp->recno, CA_IAFTER);
758*7c478bd9Sstevel@tonic-gate 
759*7c478bd9Sstevel@tonic-gate 		/* Set this cursor to reference the new record. */
760*7c478bd9Sstevel@tonic-gate 		cp->recno = copy.recno + 1;
761*7c478bd9Sstevel@tonic-gate 		break;
762*7c478bd9Sstevel@tonic-gate 	case DB_BEFORE:
763*7c478bd9Sstevel@tonic-gate 		/* Adjust the cursors. */
764*7c478bd9Sstevel@tonic-gate 		__ram_ca(dbp, cp->recno, CA_IBEFORE);
765*7c478bd9Sstevel@tonic-gate 
766*7c478bd9Sstevel@tonic-gate 		/* Set this cursor to reference the new record. */
767*7c478bd9Sstevel@tonic-gate 		cp->recno = copy.recno;
768*7c478bd9Sstevel@tonic-gate 		break;
769*7c478bd9Sstevel@tonic-gate 	}
770*7c478bd9Sstevel@tonic-gate 
771*7c478bd9Sstevel@tonic-gate 	/* The cursor was reset, no further delete adjustment is necessary. */
772*7c478bd9Sstevel@tonic-gate 	CD_CLR(dbp, cp);
773*7c478bd9Sstevel@tonic-gate 
774*7c478bd9Sstevel@tonic-gate err:	if (F_ISSET(dbp, DB_AM_CDB) && F_ISSET(dbc, DBC_RMW))
775*7c478bd9Sstevel@tonic-gate 		(void)__lock_downgrade(dbp->dbenv->lk_info, dbc->mylock,
776*7c478bd9Sstevel@tonic-gate 		    DB_LOCK_IWRITE, 0);
777*7c478bd9Sstevel@tonic-gate 
778*7c478bd9Sstevel@tonic-gate 	if (ret != 0)
779*7c478bd9Sstevel@tonic-gate 		*cp = copy;
780*7c478bd9Sstevel@tonic-gate 
781*7c478bd9Sstevel@tonic-gate 	return (ret);
782*7c478bd9Sstevel@tonic-gate }
783*7c478bd9Sstevel@tonic-gate 
784*7c478bd9Sstevel@tonic-gate /*
785*7c478bd9Sstevel@tonic-gate  * __ram_ca --
786*7c478bd9Sstevel@tonic-gate  *	Adjust cursors.
787*7c478bd9Sstevel@tonic-gate  *
788*7c478bd9Sstevel@tonic-gate  * PUBLIC: void __ram_ca __P((DB *, db_recno_t, ca_recno_arg));
789*7c478bd9Sstevel@tonic-gate  */
790*7c478bd9Sstevel@tonic-gate void
__ram_ca(dbp,recno,op)791*7c478bd9Sstevel@tonic-gate __ram_ca(dbp, recno, op)
792*7c478bd9Sstevel@tonic-gate 	DB *dbp;
793*7c478bd9Sstevel@tonic-gate 	db_recno_t recno;
794*7c478bd9Sstevel@tonic-gate 	ca_recno_arg op;
795*7c478bd9Sstevel@tonic-gate {
796*7c478bd9Sstevel@tonic-gate 	CURSOR *cp;
797*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
798*7c478bd9Sstevel@tonic-gate 
799*7c478bd9Sstevel@tonic-gate 	/*
800*7c478bd9Sstevel@tonic-gate 	 * Adjust the cursors.  See the comment in __bam_ca_delete().
801*7c478bd9Sstevel@tonic-gate 	 */
802*7c478bd9Sstevel@tonic-gate 	DB_THREAD_LOCK(dbp);
803*7c478bd9Sstevel@tonic-gate 	for (dbc = TAILQ_FIRST(&dbp->active_queue);
804*7c478bd9Sstevel@tonic-gate 	    dbc != NULL; dbc = TAILQ_NEXT(dbc, links)) {
805*7c478bd9Sstevel@tonic-gate 		cp = dbc->internal;
806*7c478bd9Sstevel@tonic-gate 		switch (op) {
807*7c478bd9Sstevel@tonic-gate 		case CA_DELETE:
808*7c478bd9Sstevel@tonic-gate 			if (recno > cp->recno)
809*7c478bd9Sstevel@tonic-gate 				--cp->recno;
810*7c478bd9Sstevel@tonic-gate 			if (recno == cp->recno)
811*7c478bd9Sstevel@tonic-gate 				CD_SET(dbp, cp);
812*7c478bd9Sstevel@tonic-gate 			break;
813*7c478bd9Sstevel@tonic-gate 		case CA_IAFTER:
814*7c478bd9Sstevel@tonic-gate 			if (recno > cp->recno)
815*7c478bd9Sstevel@tonic-gate 				++cp->recno;
816*7c478bd9Sstevel@tonic-gate 			break;
817*7c478bd9Sstevel@tonic-gate 		case CA_IBEFORE:
818*7c478bd9Sstevel@tonic-gate 			if (recno >= cp->recno)
819*7c478bd9Sstevel@tonic-gate 				++cp->recno;
820*7c478bd9Sstevel@tonic-gate 			break;
821*7c478bd9Sstevel@tonic-gate 		}
822*7c478bd9Sstevel@tonic-gate 	}
823*7c478bd9Sstevel@tonic-gate 	DB_THREAD_UNLOCK(dbp);
824*7c478bd9Sstevel@tonic-gate }
825*7c478bd9Sstevel@tonic-gate 
826*7c478bd9Sstevel@tonic-gate /*
827*7c478bd9Sstevel@tonic-gate  * __ram_getno --
828*7c478bd9Sstevel@tonic-gate  *	Check the user's record number, and make sure we've seen it.
829*7c478bd9Sstevel@tonic-gate  *
830*7c478bd9Sstevel@tonic-gate  * PUBLIC: int __ram_getno __P((DBC *, const DBT *, db_recno_t *, int));
831*7c478bd9Sstevel@tonic-gate  */
832*7c478bd9Sstevel@tonic-gate int
__ram_getno(dbc,key,rep,can_create)833*7c478bd9Sstevel@tonic-gate __ram_getno(dbc, key, rep, can_create)
834*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
835*7c478bd9Sstevel@tonic-gate 	const DBT *key;
836*7c478bd9Sstevel@tonic-gate 	db_recno_t *rep;
837*7c478bd9Sstevel@tonic-gate 	int can_create;
838*7c478bd9Sstevel@tonic-gate {
839*7c478bd9Sstevel@tonic-gate 	DB *dbp;
840*7c478bd9Sstevel@tonic-gate 	db_recno_t recno;
841*7c478bd9Sstevel@tonic-gate 
842*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
843*7c478bd9Sstevel@tonic-gate 
844*7c478bd9Sstevel@tonic-gate 	/* Check the user's record number. */
845*7c478bd9Sstevel@tonic-gate 	if ((recno = *(db_recno_t *)key->data) == 0) {
846*7c478bd9Sstevel@tonic-gate 		__db_err(dbp->dbenv, "illegal record number of 0");
847*7c478bd9Sstevel@tonic-gate 		return (EINVAL);
848*7c478bd9Sstevel@tonic-gate 	}
849*7c478bd9Sstevel@tonic-gate 	if (rep != NULL)
850*7c478bd9Sstevel@tonic-gate 		*rep = recno;
851*7c478bd9Sstevel@tonic-gate 
852*7c478bd9Sstevel@tonic-gate 	/*
853*7c478bd9Sstevel@tonic-gate 	 * Btree can neither create records nor read them in.  Recno can
854*7c478bd9Sstevel@tonic-gate 	 * do both, see if we can find the record.
855*7c478bd9Sstevel@tonic-gate 	 */
856*7c478bd9Sstevel@tonic-gate 	return (dbp->type == DB_RECNO ?
857*7c478bd9Sstevel@tonic-gate 	    __ram_update(dbc, recno, can_create) : 0);
858*7c478bd9Sstevel@tonic-gate }
859*7c478bd9Sstevel@tonic-gate 
860*7c478bd9Sstevel@tonic-gate /*
861*7c478bd9Sstevel@tonic-gate  * __ram_update --
862*7c478bd9Sstevel@tonic-gate  *	Ensure the tree has records up to and including the specified one.
863*7c478bd9Sstevel@tonic-gate  */
864*7c478bd9Sstevel@tonic-gate static int
__ram_update(dbc,recno,can_create)865*7c478bd9Sstevel@tonic-gate __ram_update(dbc, recno, can_create)
866*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
867*7c478bd9Sstevel@tonic-gate 	db_recno_t recno;
868*7c478bd9Sstevel@tonic-gate 	int can_create;
869*7c478bd9Sstevel@tonic-gate {
870*7c478bd9Sstevel@tonic-gate 	BTREE *t;
871*7c478bd9Sstevel@tonic-gate 	DB *dbp;
872*7c478bd9Sstevel@tonic-gate 	RECNO *rp;
873*7c478bd9Sstevel@tonic-gate 	db_recno_t nrecs;
874*7c478bd9Sstevel@tonic-gate 	int ret;
875*7c478bd9Sstevel@tonic-gate 
876*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
877*7c478bd9Sstevel@tonic-gate 	t = dbp->internal;
878*7c478bd9Sstevel@tonic-gate 	rp = t->recno;
879*7c478bd9Sstevel@tonic-gate 
880*7c478bd9Sstevel@tonic-gate 	/*
881*7c478bd9Sstevel@tonic-gate 	 * If we can't create records and we've read the entire backing input
882*7c478bd9Sstevel@tonic-gate 	 * file, we're done.
883*7c478bd9Sstevel@tonic-gate 	 */
884*7c478bd9Sstevel@tonic-gate 	if (!can_create && F_ISSET(rp, RECNO_EOF))
885*7c478bd9Sstevel@tonic-gate 		return (0);
886*7c478bd9Sstevel@tonic-gate 
887*7c478bd9Sstevel@tonic-gate 	/*
888*7c478bd9Sstevel@tonic-gate 	 * If we haven't seen this record yet, try to get it from the original
889*7c478bd9Sstevel@tonic-gate 	 * file.
890*7c478bd9Sstevel@tonic-gate 	 */
891*7c478bd9Sstevel@tonic-gate 	if ((ret = __bam_nrecs(dbc, &nrecs)) != 0)
892*7c478bd9Sstevel@tonic-gate 		return (ret);
893*7c478bd9Sstevel@tonic-gate 	if (!F_ISSET(rp, RECNO_EOF) && recno > nrecs) {
894*7c478bd9Sstevel@tonic-gate 		if ((ret = rp->re_irec(dbc, recno)) != 0)
895*7c478bd9Sstevel@tonic-gate 			return (ret);
896*7c478bd9Sstevel@tonic-gate 		if ((ret = __bam_nrecs(dbc, &nrecs)) != 0)
897*7c478bd9Sstevel@tonic-gate 			return (ret);
898*7c478bd9Sstevel@tonic-gate 	}
899*7c478bd9Sstevel@tonic-gate 
900*7c478bd9Sstevel@tonic-gate 	/*
901*7c478bd9Sstevel@tonic-gate 	 * If we can create records, create empty ones up to the requested
902*7c478bd9Sstevel@tonic-gate 	 * record.
903*7c478bd9Sstevel@tonic-gate 	 */
904*7c478bd9Sstevel@tonic-gate 	if (!can_create || recno <= nrecs + 1)
905*7c478bd9Sstevel@tonic-gate 		return (0);
906*7c478bd9Sstevel@tonic-gate 
907*7c478bd9Sstevel@tonic-gate 	dbc->rdata.dlen = 0;
908*7c478bd9Sstevel@tonic-gate 	dbc->rdata.doff = 0;
909*7c478bd9Sstevel@tonic-gate 	dbc->rdata.flags = 0;
910*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_RE_FIXEDLEN)) {
911*7c478bd9Sstevel@tonic-gate 		if (dbc->rdata.ulen < rp->re_len) {
912*7c478bd9Sstevel@tonic-gate 			if ((ret =
913*7c478bd9Sstevel@tonic-gate 			    __os_realloc(&dbc->rdata.data, rp->re_len)) != 0) {
914*7c478bd9Sstevel@tonic-gate 				dbc->rdata.ulen = 0;
915*7c478bd9Sstevel@tonic-gate 				dbc->rdata.data = NULL;
916*7c478bd9Sstevel@tonic-gate 				return (ret);
917*7c478bd9Sstevel@tonic-gate 			}
918*7c478bd9Sstevel@tonic-gate 			dbc->rdata.ulen = rp->re_len;
919*7c478bd9Sstevel@tonic-gate 		}
920*7c478bd9Sstevel@tonic-gate 		dbc->rdata.size = rp->re_len;
921*7c478bd9Sstevel@tonic-gate 		memset(dbc->rdata.data, rp->re_pad, rp->re_len);
922*7c478bd9Sstevel@tonic-gate 	} else
923*7c478bd9Sstevel@tonic-gate 		dbc->rdata.size = 0;
924*7c478bd9Sstevel@tonic-gate 
925*7c478bd9Sstevel@tonic-gate 	while (recno > ++nrecs)
926*7c478bd9Sstevel@tonic-gate 		if ((ret = __ram_add(dbc,
927*7c478bd9Sstevel@tonic-gate 		    &nrecs, &dbc->rdata, 0, BI_DELETED)) != 0)
928*7c478bd9Sstevel@tonic-gate 			return (ret);
929*7c478bd9Sstevel@tonic-gate 	return (0);
930*7c478bd9Sstevel@tonic-gate }
931*7c478bd9Sstevel@tonic-gate 
932*7c478bd9Sstevel@tonic-gate /*
933*7c478bd9Sstevel@tonic-gate  * __ram_source --
934*7c478bd9Sstevel@tonic-gate  *	Load information about the backing file.
935*7c478bd9Sstevel@tonic-gate  */
936*7c478bd9Sstevel@tonic-gate static int
__ram_source(dbp,rp,fname)937*7c478bd9Sstevel@tonic-gate __ram_source(dbp, rp, fname)
938*7c478bd9Sstevel@tonic-gate 	DB *dbp;
939*7c478bd9Sstevel@tonic-gate 	RECNO *rp;
940*7c478bd9Sstevel@tonic-gate 	const char *fname;
941*7c478bd9Sstevel@tonic-gate {
942*7c478bd9Sstevel@tonic-gate 	size_t size;
943*7c478bd9Sstevel@tonic-gate 	u_int32_t bytes, mbytes, oflags;
944*7c478bd9Sstevel@tonic-gate 	int ret;
945*7c478bd9Sstevel@tonic-gate 
946*7c478bd9Sstevel@tonic-gate 	/*
947*7c478bd9Sstevel@tonic-gate 	 * !!!
948*7c478bd9Sstevel@tonic-gate 	 * The caller has full responsibility for cleaning up on error --
949*7c478bd9Sstevel@tonic-gate 	 * (it has to anyway, in case it fails after this routine succeeds).
950*7c478bd9Sstevel@tonic-gate 	 */
951*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_appname(dbp->dbenv,
952*7c478bd9Sstevel@tonic-gate 	    DB_APP_DATA, NULL, fname, 0, NULL, &rp->re_source)) != 0)
953*7c478bd9Sstevel@tonic-gate 		return (ret);
954*7c478bd9Sstevel@tonic-gate 
955*7c478bd9Sstevel@tonic-gate 	oflags = F_ISSET(dbp, DB_AM_RDONLY) ? DB_RDONLY : 0;
956*7c478bd9Sstevel@tonic-gate 	if ((ret =
957*7c478bd9Sstevel@tonic-gate 	    __db_open(rp->re_source, oflags, oflags, 0, &rp->re_fd)) != 0) {
958*7c478bd9Sstevel@tonic-gate 		__db_err(dbp->dbenv, "%s: %s", rp->re_source, strerror(ret));
959*7c478bd9Sstevel@tonic-gate 		return (ret);
960*7c478bd9Sstevel@tonic-gate 	}
961*7c478bd9Sstevel@tonic-gate 
962*7c478bd9Sstevel@tonic-gate 	/*
963*7c478bd9Sstevel@tonic-gate 	 * XXX
964*7c478bd9Sstevel@tonic-gate 	 * We'd like to test to see if the file is too big to mmap.  Since we
965*7c478bd9Sstevel@tonic-gate 	 * don't know what size or type off_t's or size_t's are, or the largest
966*7c478bd9Sstevel@tonic-gate 	 * unsigned integral type is, or what random insanity the local C
967*7c478bd9Sstevel@tonic-gate 	 * compiler will perpetrate, doing the comparison in a portable way is
968*7c478bd9Sstevel@tonic-gate 	 * flatly impossible.  Hope that mmap fails if the file is too large.
969*7c478bd9Sstevel@tonic-gate 	 */
970*7c478bd9Sstevel@tonic-gate 	if ((ret = __os_ioinfo(rp->re_source,
971*7c478bd9Sstevel@tonic-gate 	    rp->re_fd, &mbytes, &bytes, NULL)) != 0) {
972*7c478bd9Sstevel@tonic-gate 		__db_err(dbp->dbenv, "%s: %s", rp->re_source, strerror(ret));
973*7c478bd9Sstevel@tonic-gate 		return (ret);
974*7c478bd9Sstevel@tonic-gate 	}
975*7c478bd9Sstevel@tonic-gate 	if (mbytes == 0 && bytes == 0) {
976*7c478bd9Sstevel@tonic-gate 		F_SET(rp, RECNO_EOF);
977*7c478bd9Sstevel@tonic-gate 		return (0);
978*7c478bd9Sstevel@tonic-gate 	}
979*7c478bd9Sstevel@tonic-gate 
980*7c478bd9Sstevel@tonic-gate 	size = mbytes * MEGABYTE + bytes;
981*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_mapfile(rp->re_source,
982*7c478bd9Sstevel@tonic-gate 	    rp->re_fd, (size_t)size, 1, &rp->re_smap)) != 0)
983*7c478bd9Sstevel@tonic-gate 		return (ret);
984*7c478bd9Sstevel@tonic-gate 	rp->re_cmap = rp->re_smap;
985*7c478bd9Sstevel@tonic-gate 	rp->re_emap = (u_int8_t *)rp->re_smap + (rp->re_msize = size);
986*7c478bd9Sstevel@tonic-gate 	rp->re_irec = F_ISSET(dbp, DB_RE_FIXEDLEN) ?  __ram_fmap : __ram_vmap;
987*7c478bd9Sstevel@tonic-gate 	return (0);
988*7c478bd9Sstevel@tonic-gate }
989*7c478bd9Sstevel@tonic-gate 
990*7c478bd9Sstevel@tonic-gate /*
991*7c478bd9Sstevel@tonic-gate  * __ram_writeback --
992*7c478bd9Sstevel@tonic-gate  *	Rewrite the backing file.
993*7c478bd9Sstevel@tonic-gate  */
994*7c478bd9Sstevel@tonic-gate static int
__ram_writeback(dbc)995*7c478bd9Sstevel@tonic-gate __ram_writeback(dbc)
996*7c478bd9Sstevel@tonic-gate 	DBC *dbc;
997*7c478bd9Sstevel@tonic-gate {
998*7c478bd9Sstevel@tonic-gate 	DB *dbp;
999*7c478bd9Sstevel@tonic-gate 	DBT key, data;
1000*7c478bd9Sstevel@tonic-gate 	RECNO *rp;
1001*7c478bd9Sstevel@tonic-gate 	db_recno_t keyno;
1002*7c478bd9Sstevel@tonic-gate 	ssize_t nw;
1003*7c478bd9Sstevel@tonic-gate 	int fd, ret, t_ret;
1004*7c478bd9Sstevel@tonic-gate 	u_int8_t delim, *pad;
1005*7c478bd9Sstevel@tonic-gate 
1006*7c478bd9Sstevel@tonic-gate 	dbp = dbc->dbp;
1007*7c478bd9Sstevel@tonic-gate 	rp = ((BTREE *)dbp->internal)->recno;
1008*7c478bd9Sstevel@tonic-gate 
1009*7c478bd9Sstevel@tonic-gate 	/* If the file wasn't modified, we're done. */
1010*7c478bd9Sstevel@tonic-gate 	if (!F_ISSET(rp, RECNO_MODIFIED))
1011*7c478bd9Sstevel@tonic-gate 		return (0);
1012*7c478bd9Sstevel@tonic-gate 
1013*7c478bd9Sstevel@tonic-gate 	/* If there's no backing source file, we're done. */
1014*7c478bd9Sstevel@tonic-gate 	if (rp->re_source == NULL) {
1015*7c478bd9Sstevel@tonic-gate 		F_CLR(rp, RECNO_MODIFIED);
1016*7c478bd9Sstevel@tonic-gate 		return (0);
1017*7c478bd9Sstevel@tonic-gate 	}
1018*7c478bd9Sstevel@tonic-gate 
1019*7c478bd9Sstevel@tonic-gate 	/*
1020*7c478bd9Sstevel@tonic-gate 	 * Read any remaining records into the tree.
1021*7c478bd9Sstevel@tonic-gate 	 *
1022*7c478bd9Sstevel@tonic-gate 	 * !!!
1023*7c478bd9Sstevel@tonic-gate 	 * This is why we can't support transactions when applications specify
1024*7c478bd9Sstevel@tonic-gate 	 * backing (re_source) files.  At this point we have to read in the
1025*7c478bd9Sstevel@tonic-gate 	 * rest of the records from the file so that we can write all of the
1026*7c478bd9Sstevel@tonic-gate 	 * records back out again, which could modify a page for which we'd
1027*7c478bd9Sstevel@tonic-gate 	 * have to log changes and which we don't have locked.  This could be
1028*7c478bd9Sstevel@tonic-gate 	 * partially fixed by taking a snapshot of the entire file during the
1029*7c478bd9Sstevel@tonic-gate 	 * db_open(), or, since db_open() isn't transaction protected, as part
1030*7c478bd9Sstevel@tonic-gate 	 * of the first DB operation.  But, if a checkpoint occurs then, the
1031*7c478bd9Sstevel@tonic-gate 	 * part of the log holding the copy of the file could be discarded, and
1032*7c478bd9Sstevel@tonic-gate 	 * that would make it impossible to recover in the face of disaster.
1033*7c478bd9Sstevel@tonic-gate 	 * This could all probably be fixed, but it would require transaction
1034*7c478bd9Sstevel@tonic-gate 	 * protecting the backing source file, i.e. mpool would have to know
1035*7c478bd9Sstevel@tonic-gate 	 * about it, and we don't want to go there.
1036*7c478bd9Sstevel@tonic-gate 	 */
1037*7c478bd9Sstevel@tonic-gate 	if ((ret =
1038*7c478bd9Sstevel@tonic-gate 	    __ram_update(dbc, DB_MAX_RECORDS, 0)) != 0 && ret != DB_NOTFOUND)
1039*7c478bd9Sstevel@tonic-gate 		return (ret);
1040*7c478bd9Sstevel@tonic-gate 
1041*7c478bd9Sstevel@tonic-gate 	/*
1042*7c478bd9Sstevel@tonic-gate 	 * !!!
1043*7c478bd9Sstevel@tonic-gate 	 * Close any underlying mmap region.  This is required for Windows NT
1044*7c478bd9Sstevel@tonic-gate 	 * (4.0, Service Pack 2) -- if the file is still mapped, the following
1045*7c478bd9Sstevel@tonic-gate 	 * open will fail.
1046*7c478bd9Sstevel@tonic-gate 	 */
1047*7c478bd9Sstevel@tonic-gate 	if (rp->re_smap != NULL) {
1048*7c478bd9Sstevel@tonic-gate 		(void)__db_unmapfile(rp->re_smap, rp->re_msize);
1049*7c478bd9Sstevel@tonic-gate 		rp->re_smap = NULL;
1050*7c478bd9Sstevel@tonic-gate 	}
1051*7c478bd9Sstevel@tonic-gate 
1052*7c478bd9Sstevel@tonic-gate 	/* Get rid of any backing file descriptor, just on GP's. */
1053*7c478bd9Sstevel@tonic-gate 	if (rp->re_fd != -1) {
1054*7c478bd9Sstevel@tonic-gate 		(void)__os_close(rp->re_fd);
1055*7c478bd9Sstevel@tonic-gate 		rp->re_fd = -1;
1056*7c478bd9Sstevel@tonic-gate 	}
1057*7c478bd9Sstevel@tonic-gate 
1058*7c478bd9Sstevel@tonic-gate 	/* Open the file, truncating it. */
1059*7c478bd9Sstevel@tonic-gate 	if ((ret = __db_open(rp->re_source,
1060*7c478bd9Sstevel@tonic-gate 	    DB_SEQUENTIAL | DB_TRUNCATE,
1061*7c478bd9Sstevel@tonic-gate 	    DB_SEQUENTIAL | DB_TRUNCATE, 0, &fd)) != 0) {
1062*7c478bd9Sstevel@tonic-gate 		__db_err(dbp->dbenv, "%s: %s", rp->re_source, strerror(ret));
1063*7c478bd9Sstevel@tonic-gate 		return (ret);
1064*7c478bd9Sstevel@tonic-gate 	}
1065*7c478bd9Sstevel@tonic-gate 
1066*7c478bd9Sstevel@tonic-gate 	/*
1067*7c478bd9Sstevel@tonic-gate 	 * We step through the records, writing each one out.  Use the record
1068*7c478bd9Sstevel@tonic-gate 	 * number and the dbp->get() function, instead of a cursor, so we find
1069*7c478bd9Sstevel@tonic-gate 	 * and write out "deleted" or non-existent records.
1070*7c478bd9Sstevel@tonic-gate 	 */
1071*7c478bd9Sstevel@tonic-gate 	memset(&key, 0, sizeof(key));
1072*7c478bd9Sstevel@tonic-gate 	memset(&data, 0, sizeof(data));
1073*7c478bd9Sstevel@tonic-gate 	key.size = sizeof(db_recno_t);
1074*7c478bd9Sstevel@tonic-gate 	key.data = &keyno;
1075*7c478bd9Sstevel@tonic-gate 
1076*7c478bd9Sstevel@tonic-gate 	/*
1077*7c478bd9Sstevel@tonic-gate 	 * We'll need the delimiter if we're doing variable-length records,
1078*7c478bd9Sstevel@tonic-gate 	 * and the pad character if we're doing fixed-length records.
1079*7c478bd9Sstevel@tonic-gate 	 */
1080*7c478bd9Sstevel@tonic-gate 	delim = rp->re_delim;
1081*7c478bd9Sstevel@tonic-gate 	if (F_ISSET(dbp, DB_RE_FIXEDLEN)) {
1082*7c478bd9Sstevel@tonic-gate 		if ((ret = __os_malloc(rp->re_len, NULL, &pad)) != 0)
1083*7c478bd9Sstevel@tonic-gate 			goto err;
1084*7c478bd9Sstevel@tonic-gate 		memset(pad, rp->re_pad, rp->re_len);
1085*7c478bd9Sstevel@tonic-gate 	} else
1086*7c478bd9Sstevel@tonic-gate 		COMPQUIET(pad, NULL);
1087*7c478bd9Sstevel@tonic-gate 	for (keyno = 1;; ++keyno) {
1088*7c478bd9Sstevel@tonic-gate 		switch (ret = dbp->get(dbp, NULL, &key, &data, 0)) {
1089*7c478bd9Sstevel@tonic-gate 		case 0:
1090*7c478bd9Sstevel@tonic-gate 			if ((ret =
1091*7c478bd9Sstevel@tonic-gate 			    __os_write(fd, data.data, data.size, &nw)) != 0)
1092*7c478bd9Sstevel@tonic-gate 				goto err;
1093*7c478bd9Sstevel@tonic-gate 			if (nw != (ssize_t)data.size) {
1094*7c478bd9Sstevel@tonic-gate 				ret = EIO;
1095*7c478bd9Sstevel@tonic-gate 				goto err;
1096*7c478bd9Sstevel@tonic-gate 			}
1097*7c478bd9Sstevel@tonic-gate 			break;
1098*7c478bd9Sstevel@tonic-gate 		case DB_KEYEMPTY:
1099*7c478bd9Sstevel@tonic-gate 			if (F_ISSET(dbp, DB_RE_FIXEDLEN)) {
1100*7c478bd9Sstevel@tonic-gate 				if ((ret =
1101*7c478bd9Sstevel@tonic-gate 				    __os_write(fd, pad, rp->re_len, &nw)) != 0)
1102*7c478bd9Sstevel@tonic-gate 					goto err;
1103*7c478bd9Sstevel@tonic-gate 				if (nw != (ssize_t)rp->re_len) {
1104*7c478bd9Sstevel@tonic-gate 					ret = EIO;
1105*7c478bd9Sstevel@tonic-gate 					goto err;
1106*7c478bd9Sstevel@tonic-gate 				}
1107*7c478bd9Sstevel@tonic-gate 			}
1108*7c478bd9Sstevel@tonic-gate 			break;
1109*7c478bd9Sstevel@tonic-gate 		case DB_NOTFOUND:
1110*7c478bd9Sstevel@tonic-gate 			ret = 0;
1111*7c478bd9Sstevel@tonic-gate 			goto done;
1112*7c478bd9Sstevel@tonic-gate 		}
1113*7c478bd9Sstevel@tonic-gate 		if (!F_ISSET(dbp, DB_RE_FIXEDLEN)) {
1114*7c478bd9Sstevel@tonic-gate 			if ((ret = __os_write(fd, &delim, 1, &nw)) != 0)
1115*7c478bd9Sstevel@tonic-gate 				goto err;
1116*7c478bd9Sstevel@tonic-gate 			if (nw != 1) {
1117