Initial revision

CVS patchset: 5630
CVS date: 2002/08/14 16:31:49

5 files changed, 1908 insertions, 0 deletions

diff --git a/db/dbreg/dbreg.c b/db/dbreg/dbreg.c
new file mode 100644
index 000000000..69b2fba22
--- /dev/null
+++ b/db/dbreg/dbreg.c
@@ -0,0 +1,449 @@
+/*-
+ * See the file LICENSE for redistribution information.
+ *
+ * Copyright (c) 1996-2002
+ *	Sleepycat Software.  All rights reserved.
+ */
+#include "db_config.h"
+
+#ifndef lint
+static const char revid[] = "Id: dbreg.c,v 11.66 2002/08/11 02:14:16 margo Exp ";
+#endif /* not lint */
+
+#ifndef NO_SYSTEM_INCLUDES
+#include <sys/types.h>
+
+#include <string.h>
+#endif
+
+#include "db_int.h"
+#include "dbinc/log.h"
+#include "dbinc/txn.h"
+
+/*
+ * The dbreg subsystem, as its name implies, registers database handles so
+ * that we can associate log messages with them without logging a filename
+ * or a full, unique DB ID.  Instead, we assign each dbp an int32_t which is
+ * easy and cheap to log, and use this subsystem to map back and forth.
+ *
+ * Overview of how dbreg ids are managed:
+ *
+ * OPEN
+ *	dbreg_setup (Creates FNAME struct.)
+ *	dbreg_new_id (Assigns new ID to dbp and logs it.  May be postponed
+ *	until we attempt to log something else using that dbp, if the dbp
+ *	was opened on a replication client.)
+ *
+ * CLOSE
+ *	dbreg_close_id  (Logs closure of dbp/revocation of ID.)
+ *	dbreg_revoke_id (As name implies, revokes ID.)
+ *	dbreg_teardown (Destroys FNAME.)
+ *
+ * RECOVERY
+ *	dbreg_setup
+ *	dbreg_assign_id (Assigns a particular ID we have in the log to a dbp.)
+ *
+ *	sometimes: dbreg_revoke_id; dbreg_teardown
+ *	other times: normal close path
+ *
+ * A note about locking:
+ *
+ *	FNAME structures are referenced only by their corresponding dbp's
+ *	until they have a valid id.
+ *
+ *	Once they have a valid id, they must get linked into the log
+ *	region list so they can get logged on checkpoints.
+ *
+ *	An FNAME that may/does have a valid id must be accessed under
+ *	protection of the fq_mutex, with the following exception:
+ *
+ *	We don't want to have to grab the fq_mutex on every log
+ *	record, and it should be safe not to do so when we're just
+ *	looking at the id, because once allocated, the id should
+ *	not change under a handle until the handle is closed.
+ *
+ *	If a handle is closed during an attempt by another thread to
+ *	log with it, well, the application doing the close deserves to
+ *	go down in flames and a lot else is about to fail anyway.
+ *
+ *	When in the course of logging we encounter an invalid id
+ *	and go to allocate it lazily, we *do* need to check again
+ *	after grabbing the mutex, because it's possible to race with
+ *	another thread that has also decided that it needs to allocate
+ *	a id lazily.
+ *
+ * See SR #5623 for further discussion of the new dbreg design.
+ */
+
+/*
+ * __dbreg_setup --
+ *	Allocate and initialize an FNAME structure.  The FNAME structures
+ * live in the log shared region and map one-to-one with open database handles.
+ * When the handle needs to be logged, the FNAME should have a valid fid
+ * allocated.  If the handle currently isn't logged, it still has an FNAME
+ * entry.  If we later discover that the handle needs to be logged, we can
+ * allocate a id for it later.  (This happens when the handle is on a
+ * replication client that later becomes a master.)
+ *
+ * PUBLIC: int __dbreg_setup __P((DB *, const char *, u_int32_t));
+ */
+int
+__dbreg_setup(dbp, name, create_txnid)
+	DB *dbp;
+	const char *name;
+	u_int32_t create_txnid;
+{
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	FNAME *fnp;
+	LOG *lp;
+	int ret;
+	size_t len;
+	void *namep;
+
+	dbenv = dbp->dbenv;
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+
+	fnp = NULL;
+	namep = NULL;
+
+	/* Allocate an FNAME and, if necessary, a buffer for the name itself. */
+	R_LOCK(dbenv, &dblp->reginfo);
+	if ((ret =
+	    __db_shalloc(dblp->reginfo.addr, sizeof(FNAME), 0, &fnp)) != 0) {
+		R_UNLOCK(dbenv, &dblp->reginfo);
+		return (ret);
+	}
+	memset(fnp, 0, sizeof(FNAME));
+	if (name != NULL) {
+		len = strlen(name) + 1;
+		if ((ret = __db_shalloc(dblp->reginfo.addr,
+		    len, 0, &namep)) != 0) {
+			R_UNLOCK(dbenv, &dblp->reginfo);
+			return (ret);
+		}
+		fnp->name_off = R_OFFSET(&dblp->reginfo, namep);
+		memcpy(namep, name, len);
+	} else
+		fnp->name_off = INVALID_ROFF;
+
+	R_UNLOCK(dbenv, &dblp->reginfo);
+
+	/*
+	 * Fill in all the remaining info that we'll need later to register
+	 * the file, if we use it for logging.
+	 */
+	fnp->id = DB_LOGFILEID_INVALID;
+	fnp->s_type = dbp->type;
+	memcpy(fnp->ufid, dbp->fileid, DB_FILE_ID_LEN);
+	fnp->meta_pgno = dbp->meta_pgno;
+	fnp->create_txnid = create_txnid;
+
+	dbp->log_filename = fnp;
+
+	return (0);
+}
+
+/*
+ * __dbreg_teardown --
+ *	Destroy a DB handle's FNAME struct.
+ *
+ * PUBLIC: int __dbreg_teardown __P((DB *));
+ */
+int
+__dbreg_teardown(dbp)
+	DB *dbp;
+{
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	FNAME *fnp;
+
+	dbenv = dbp->dbenv;
+	dblp = dbenv->lg_handle;
+	fnp = dbp->log_filename;
+
+	/*
+	 * We may not have an FNAME if we were never opened.  This is not an
+	 * error.
+	 */
+	if (fnp == NULL)
+		return (0);
+
+	DB_ASSERT(fnp->id == DB_LOGFILEID_INVALID);
+
+	R_LOCK(dbenv, &dblp->reginfo);
+	if (fnp->name_off != INVALID_ROFF)
+		__db_shalloc_free(dblp->reginfo.addr,
+		    R_ADDR(&dblp->reginfo, fnp->name_off));
+	__db_shalloc_free(dblp->reginfo.addr, fnp);
+	R_UNLOCK(dbenv, &dblp->reginfo);
+
+	dbp->log_filename = NULL;
+
+	return (0);
+}
+
+/*
+ * __dbreg_new_id --
+ *	Assign an unused dbreg id to this database handle.
+ *
+ * PUBLIC: int __dbreg_new_id __P((DB *, DB_TXN *));
+ */
+int
+__dbreg_new_id(dbp, txn)
+	DB *dbp;
+	DB_TXN *txn;
+{
+	DBT fid_dbt, r_name;
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	DB_LSN unused;
+	FNAME *fnp;
+	LOG *lp;
+	int32_t id;
+	int ret;
+
+	dbenv = dbp->dbenv;
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+	fnp = dbp->log_filename;
+
+	/* The fq_mutex protects the FNAME list and id management. */
+	MUTEX_LOCK(dbenv, &lp->fq_mutex);
+
+	/*
+	 * It's possible that after deciding we needed to call this function,
+	 * someone else allocated an ID before we grabbed the lock.  Check
+	 * to make sure there was no race and we have something useful to do.
+	 */
+	if (fnp->id != DB_LOGFILEID_INVALID) {
+		MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+		return (0);
+	}
+
+	/* Get an unused ID from the free list. */
+	if ((ret = __dbreg_pop_id(dbenv, &id)) != 0)
+		goto err;
+
+	/* If no ID was found, allocate a new one. */
+	if (id == DB_LOGFILEID_INVALID)
+		id = lp->fid_max++;
+
+	fnp->id = id;
+
+	/* Hook the FNAME into the list of open files. */
+	SH_TAILQ_INSERT_HEAD(&lp->fq, fnp, q, __fname);
+
+	/*
+	 * Log the registry.  We should only request a new ID in situations
+	 * where logging is reasonable.
+	 */
+	DB_ASSERT(!F_ISSET(dbp, DB_AM_RECOVER));
+
+	memset(&fid_dbt, 0, sizeof(fid_dbt));
+	memset(&r_name, 0, sizeof(r_name));
+	if (fnp->name_off != INVALID_ROFF) {
+		r_name.data = R_ADDR(&dblp->reginfo, fnp->name_off);
+		r_name.size = (u_int32_t)strlen((char *)r_name.data) + 1;
+	}
+	fid_dbt.data = dbp->fileid;
+	fid_dbt.size = DB_FILE_ID_LEN;
+	if ((ret = __dbreg_register_log(dbenv, txn, &unused, 0, LOG_OPEN,
+	    r_name.size == 0 ? NULL : &r_name, &fid_dbt, id, fnp->s_type,
+	    fnp->meta_pgno, fnp->create_txnid)) != 0)
+		goto err;
+
+	DB_ASSERT(dbp->type == fnp->s_type);
+	DB_ASSERT(dbp->meta_pgno == fnp->meta_pgno);
+
+	if ((ret = __dbreg_add_dbentry(dbenv, dblp, dbp, id)) != 0)
+		goto err;
+
+err:	MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+	return (ret);
+}
+
+/*
+ * __dbreg_assign_id --
+ *	Assign a particular dbreg id to this database handle.
+ *
+ * PUBLIC: int __dbreg_assign_id __P((DB *, int32_t));
+ */
+int
+__dbreg_assign_id(dbp, id)
+	DB *dbp;
+	int32_t id;
+{
+	DB *close_dbp;
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	FNAME *close_fnp, *fnp;
+	LOG *lp;
+	int ret;
+
+	dbenv = dbp->dbenv;
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+	fnp = dbp->log_filename;
+
+	close_dbp = NULL;
+	close_fnp = NULL;
+
+	/* The fq_mutex protects the FNAME list and id management. */
+	MUTEX_LOCK(dbenv, &lp->fq_mutex);
+
+	/* We should only call this on DB handles that have no ID. */
+	DB_ASSERT(fnp->id == DB_LOGFILEID_INVALID);
+
+	/*
+	 * Make sure there isn't already a file open with this ID. There can
+	 * be in recovery, if we're recovering across a point where an ID got
+	 * reused.
+	 */
+	if (__dbreg_id_to_fname(dblp, id, 1, &close_fnp) == 0) {
+		/*
+		 * We want to save off any dbp we have open with this id.
+		 * We can't safely close it now, because we hold the fq_mutex,
+		 * but we should be able to rely on it being open in this
+		 * process, and we're running recovery, so no other thread
+		 * should muck with it if we just put off closing it until
+		 * we're ready to return.
+		 *
+		 * Once we have the dbp, revoke its id;  we're about to
+		 * reuse it.
+		 */
+		if ((ret =
+		    __dbreg_id_to_db(dbenv, NULL, &close_dbp, id, 0)) != 0)
+			goto err;
+
+		if ((ret = __dbreg_revoke_id(close_dbp, 1)) != 0)
+			goto err;
+	}
+
+	/*
+	 * Remove this ID from the free list, if it's there, and make sure
+	 * we don't allocate it anew.
+	 */
+	if ((ret = __dbreg_pluck_id(dbenv, id)) != 0)
+		goto err;
+	if (id >= lp->fid_max)
+		lp->fid_max = id + 1;
+
+	/* Now go ahead and assign the id to our dbp. */
+	fnp->id = id;
+	SH_TAILQ_INSERT_HEAD(&lp->fq, fnp, q, __fname);
+
+	if ((ret = __dbreg_add_dbentry(dbenv, dblp, dbp, id)) != 0)
+		goto err;
+
+err:	MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+
+	/* There's nothing useful that our caller can do if this close fails. */
+	if (close_dbp != NULL)
+		(void)close_dbp->close(close_dbp, DB_NOSYNC);
+
+	return (ret);
+}
+
+/*
+ * __dbreg_revoke_id --
+ *	Take a log id away from a dbp, in preparation for closing it,
+ *	but without logging the close.
+ *
+ * PUBLIC: int __dbreg_revoke_id __P((DB *, int));
+ */
+int
+__dbreg_revoke_id(dbp, have_lock)
+	DB *dbp;
+	int have_lock;
+{
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	FNAME *fnp;
+	LOG *lp;
+	int32_t id;
+	int ret;
+
+	dbenv = dbp->dbenv;
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+	fnp = dbp->log_filename;
+
+	/* If we lack an ID, this is a null-op. */
+	if (fnp == NULL || fnp->id == DB_LOGFILEID_INVALID)
+		return (0);
+
+	if (!have_lock)
+		MUTEX_LOCK(dbenv, &lp->fq_mutex);
+
+	id = fnp->id;
+	fnp->id = DB_LOGFILEID_INVALID;
+
+	/* Remove the FNAME from the list of open files. */
+	SH_TAILQ_REMOVE(&lp->fq, fnp, q, __fname);
+
+	/* Remove this id from the dbentry table. */
+	__dbreg_rem_dbentry(dblp, id);
+
+	/* Push this id onto the free list. */
+	ret = __dbreg_push_id(dbenv, id);
+
+	if (!have_lock)
+		MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+	return (ret);
+}
+
+/*
+ * __dbreg_close_id --
+ *	Take a dbreg id away from a dbp that we're closing, and log
+ * the unregistry.
+ *
+ * PUBLIC: int __dbreg_close_id __P((DB *, DB_TXN *));
+ */
+int
+__dbreg_close_id(dbp, txn)
+	DB *dbp;
+	DB_TXN *txn;
+{
+	DBT fid_dbt, r_name, *dbtp;
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	DB_LSN r_unused;
+	FNAME *fnp;
+	LOG *lp;
+	int ret;
+
+	dbenv = dbp->dbenv;
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+	fnp = dbp->log_filename;
+
+	/* If we lack an ID, this is a null-op. */
+	if (fnp == NULL || fnp->id == DB_LOGFILEID_INVALID)
+		return (0);
+
+	MUTEX_LOCK(dbenv, &lp->fq_mutex);
+
+	if (fnp->name_off == INVALID_ROFF)
+		dbtp = NULL;
+	else {
+		memset(&r_name, 0, sizeof(r_name));
+		r_name.data = R_ADDR(&dblp->reginfo, fnp->name_off);
+		r_name.size =
+		    (u_int32_t)strlen((char *)r_name.data) + 1;
+		dbtp = &r_name;
+	}
+	memset(&fid_dbt, 0, sizeof(fid_dbt));
+	fid_dbt.data = fnp->ufid;
+	fid_dbt.size = DB_FILE_ID_LEN;
+	if ((ret = __dbreg_register_log(dbenv, txn,
+	    &r_unused, 0, LOG_CLOSE, dbtp, &fid_dbt, fnp->id,
+	    fnp->s_type, fnp->meta_pgno, TXN_INVALID)) != 0)
+		goto err;
+
+	ret = __dbreg_revoke_id(dbp, 1);
+
+err:	MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+	return (ret);
+}
diff --git a/db/dbreg/dbreg.src b/db/dbreg/dbreg.src
new file mode 100644
index 000000000..337ffcb61
--- /dev/null
+++ b/db/dbreg/dbreg.src
@@ -0,0 +1,49 @@
+/*-
+ * See the file LICENSE for redistribution information.
+ *
+ * Copyright (c) 1996-2002
+ *	Sleepycat Software.  All rights reserved.
+ *
+ * Id: dbreg.src,v 10.22 2002/03/27 04:31:44 bostic Exp 
+ */
+
+PREFIX	__dbreg
+DBPRIVATE
+
+INCLUDE	#include "db_config.h"
+INCLUDE
+INCLUDE #ifndef NO_SYSTEM_INCLUDES
+INCLUDE #include <sys/types.h>
+INCLUDE
+INCLUDE #include <ctype.h>
+INCLUDE #include <string.h>
+INCLUDE #endif
+INCLUDE
+INCLUDE #include "db_int.h"
+INCLUDE #include "dbinc/crypto.h"
+INCLUDE #include "dbinc/db_page.h"
+INCLUDE #include "dbinc/db_dispatch.h"
+INCLUDE #include "dbinc/db_am.h"
+INCLUDE #include "dbinc/log.h"
+INCLUDE #include "dbinc/rep.h"
+INCLUDE #include "dbinc/txn.h"
+INCLUDE
+
+/*
+ * Used for registering name/id translations at open or close.
+ * opcode: register or unregister
+ * name: file name
+ * fileid: unique file id
+ * ftype: file type
+ * ftype: database type
+ * id: transaction id of the subtransaction that created the fs object
+ */
+BEGIN register		2
+ARG	opcode		u_int32_t	lu
+DBT	name		DBT		s
+DBT	uid		DBT		s
+ARG	fileid		int32_t		ld
+ARG	ftype		DBTYPE		lx
+ARG	meta_pgno	db_pgno_t	lu
+ARG	id		u_int32_t	lx
+END
diff --git a/db/dbreg/dbreg_auto.c b/db/dbreg/dbreg_auto.c
new file mode 100644
index 000000000..91eace3f4
--- /dev/null
+++ b/db/dbreg/dbreg_auto.c
@@ -0,0 +1,358 @@
+/* Do not edit: automatically built by gen_rec.awk. */
+#include "db_config.h"
+
+#ifndef NO_SYSTEM_INCLUDES
+#include <sys/types.h>
+
+#include <ctype.h>
+#include <string.h>
+#endif
+
+#include "db_int.h"
+#include "dbinc/crypto.h"
+#include "dbinc/db_page.h"
+#include "dbinc/db_dispatch.h"
+#include "dbinc/db_am.h"
+#include "dbinc/log.h"
+#include "dbinc/rep.h"
+#include "dbinc/txn.h"
+
+/*
+ * PUBLIC: int __dbreg_register_log __P((DB_ENV *, DB_TXN *,
+ * PUBLIC:     DB_LSN *, u_int32_t, u_int32_t, const DBT *, const DBT *,
+ * PUBLIC:     int32_t, DBTYPE, db_pgno_t, u_int32_t));
+ */
+int
+__dbreg_register_log(dbenv, txnid, ret_lsnp, flags,
+    opcode, name, uid, fileid, ftype, meta_pgno,
+    id)
+	DB_ENV *dbenv;
+	DB_TXN *txnid;
+	DB_LSN *ret_lsnp;
+	u_int32_t flags;
+	u_int32_t opcode;
+	const DBT *name;
+	const DBT *uid;
+	int32_t fileid;
+	DBTYPE ftype;
+	db_pgno_t meta_pgno;
+	u_int32_t id;
+{
+	DBT logrec;
+	DB_LSN *lsnp, null_lsn;
+	u_int32_t zero;
+	u_int32_t uinttmp;
+	u_int32_t npad, rectype, txn_num;
+	int ret;
+	u_int8_t *bp;
+
+	rectype = DB___dbreg_register;
+	npad = 0;
+
+	if (txnid == NULL) {
+		txn_num = 0;
+		null_lsn.file = 0;
+		null_lsn.offset = 0;
+		lsnp = &null_lsn;
+	} else {
+		if (TAILQ_FIRST(&txnid->kids) != NULL &&
+		    (ret = __txn_activekids(dbenv, rectype, txnid)) != 0)
+			return (ret);
+		txn_num = txnid->txnid;
+		lsnp = &txnid->last_lsn;
+	}
+
+	logrec.size = sizeof(rectype) + sizeof(txn_num) + sizeof(DB_LSN)
+	    + sizeof(u_int32_t)
+	    + sizeof(u_int32_t) + (name == NULL ? 0 : name->size)
+	    + sizeof(u_int32_t) + (uid == NULL ? 0 : uid->size)
+	    + sizeof(u_int32_t)
+	    + sizeof(u_int32_t)
+	    + sizeof(u_int32_t)
+	    + sizeof(u_int32_t);
+	if (CRYPTO_ON(dbenv)) {
+		npad =
+		    ((DB_CIPHER *)dbenv->crypto_handle)->adj_size(logrec.size);
+		logrec.size += npad;
+	}
+
+	if ((ret = __os_malloc(dbenv,
+	    logrec.size, &logrec.data)) != 0)
+		return (ret);
+
+	if (npad > 0)
+		memset((u_int8_t *)logrec.data + logrec.size - npad, 0, npad);
+
+	bp = logrec.data;
+
+	memcpy(bp, &rectype, sizeof(rectype));
+	bp += sizeof(rectype);
+
+	memcpy(bp, &txn_num, sizeof(txn_num));
+	bp += sizeof(txn_num);
+
+	memcpy(bp, lsnp, sizeof(DB_LSN));
+	bp += sizeof(DB_LSN);
+
+	uinttmp = (u_int32_t)opcode;
+	memcpy(bp, &uinttmp, sizeof(uinttmp));
+	bp += sizeof(uinttmp);
+
+	if (name == NULL) {
+		zero = 0;
+		memcpy(bp, &zero, sizeof(u_int32_t));
+		bp += sizeof(u_int32_t);
+	} else {
+		memcpy(bp, &name->size, sizeof(name->size));
+		bp += sizeof(name->size);
+		memcpy(bp, name->data, name->size);
+		bp += name->size;
+	}
+
+	if (uid == NULL) {
+		zero = 0;
+		memcpy(bp, &zero, sizeof(u_int32_t));
+		bp += sizeof(u_int32_t);
+	} else {
+		memcpy(bp, &uid->size, sizeof(uid->size));
+		bp += sizeof(uid->size);
+		memcpy(bp, uid->data, uid->size);
+		bp += uid->size;
+	}
+
+	uinttmp = (u_int32_t)fileid;
+	memcpy(bp, &uinttmp, sizeof(uinttmp));
+	bp += sizeof(uinttmp);
+
+	uinttmp = (u_int32_t)ftype;
+	memcpy(bp, &uinttmp, sizeof(uinttmp));
+	bp += sizeof(uinttmp);
+
+	uinttmp = (u_int32_t)meta_pgno;
+	memcpy(bp, &uinttmp, sizeof(uinttmp));
+	bp += sizeof(uinttmp);
+
+	uinttmp = (u_int32_t)id;
+	memcpy(bp, &uinttmp, sizeof(uinttmp));
+	bp += sizeof(uinttmp);
+
+	DB_ASSERT((u_int32_t)(bp - (u_int8_t *)logrec.data) <= logrec.size);
+	ret = dbenv->log_put(dbenv,
+	   ret_lsnp, (DBT *)&logrec, flags | DB_NOCOPY);
+	if (txnid != NULL && ret == 0)
+		txnid->last_lsn = *ret_lsnp;
+#ifdef LOG_DIAGNOSTIC
+	if (ret != 0)
+		(void)__dbreg_register_print(dbenv,
+		    (DBT *)&logrec, ret_lsnp, NULL, NULL);
+#endif
+	__os_free(dbenv, logrec.data);
+	return (ret);
+}
+
+/*
+ * PUBLIC: int __dbreg_register_getpgnos __P((DB_ENV *, DBT *,
+ * PUBLIC:     DB_LSN *, db_recops, void *));
+ */
+int
+__dbreg_register_getpgnos(dbenv, rec, lsnp, notused1, summary)
+	DB_ENV *dbenv;
+	DBT *rec;
+	DB_LSN *lsnp;
+	db_recops notused1;
+	void *summary;
+{
+	TXN_RECS *t;
+	int ret;
+	COMPQUIET(rec, NULL);
+	COMPQUIET(notused1, DB_TXN_ABORT);
+
+	t = (TXN_RECS *)summary;
+
+	if ((ret = __rep_check_alloc(dbenv, t, 1)) != 0)
+		return (ret);
+
+	t->array[t->npages].flags = LSN_PAGE_NOLOCK;
+	t->array[t->npages].lsn = *lsnp;
+	t->array[t->npages].fid = DB_LOGFILEID_INVALID;
+	memset(&t->array[t->npages].pgdesc, 0,
+	    sizeof(t->array[t->npages].pgdesc));
+
+	t->npages++;
+
+	return (0);
+}
+
+/*
+ * PUBLIC: int __dbreg_register_print __P((DB_ENV *, DBT *, DB_LSN *,
+ * PUBLIC:     db_recops, void *));
+ */
+int
+__dbreg_register_print(dbenv, dbtp, lsnp, notused2, notused3)
+	DB_ENV *dbenv;
+	DBT *dbtp;
+	DB_LSN *lsnp;
+	db_recops notused2;
+	void *notused3;
+{
+	__dbreg_register_args *argp;
+	u_int32_t i;
+	int ch;
+	int ret;
+
+	notused2 = DB_TXN_ABORT;
+	notused3 = NULL;
+
+	if ((ret = __dbreg_register_read(dbenv, dbtp->data, &argp)) != 0)
+		return (ret);
+	(void)printf(
+	    "[%lu][%lu]__dbreg_register: rec: %lu txnid %lx prevlsn [%lu][%lu]\n",
+	    (u_long)lsnp->file,
+	    (u_long)lsnp->offset,
+	    (u_long)argp->type,
+	    (u_long)argp->txnid->txnid,
+	    (u_long)argp->prev_lsn.file,
+	    (u_long)argp->prev_lsn.offset);
+	(void)printf("\topcode: %lu\n", (u_long)argp->opcode);
+	(void)printf("\tname: ");
+	for (i = 0; i < argp->name.size; i++) {
+		ch = ((u_int8_t *)argp->name.data)[i];
+		printf(isprint(ch) || ch == 0x0a ? "%c" : "%#x ", ch);
+	}
+	(void)printf("\n");
+	(void)printf("\tuid: ");
+	for (i = 0; i < argp->uid.size; i++) {
+		ch = ((u_int8_t *)argp->uid.data)[i];
+		printf(isprint(ch) || ch == 0x0a ? "%c" : "%#x ", ch);
+	}
+	(void)printf("\n");
+	(void)printf("\tfileid: %ld\n", (long)argp->fileid);
+	(void)printf("\tftype: 0x%lx\n", (u_long)argp->ftype);
+	(void)printf("\tmeta_pgno: %lu\n", (u_long)argp->meta_pgno);
+	(void)printf("\tid: 0x%lx\n", (u_long)argp->id);
+	(void)printf("\n");
+	__os_free(dbenv, argp);
+	return (0);
+}
+
+/*
+ * PUBLIC: int __dbreg_register_read __P((DB_ENV *, void *,
+ * PUBLIC:     __dbreg_register_args **));
+ */
+int
+__dbreg_register_read(dbenv, recbuf, argpp)
+	DB_ENV *dbenv;
+	void *recbuf;
+	__dbreg_register_args **argpp;
+{
+	__dbreg_register_args *argp;
+	u_int32_t uinttmp;
+	u_int8_t *bp;
+	int ret;
+
+	if ((ret = __os_malloc(dbenv,
+	    sizeof(__dbreg_register_args) + sizeof(DB_TXN), &argp)) != 0)
+		return (ret);
+
+	argp->txnid = (DB_TXN *)&argp[1];
+
+	bp = recbuf;
+	memcpy(&argp->type, bp, sizeof(argp->type));
+	bp += sizeof(argp->type);
+
+	memcpy(&argp->txnid->txnid,  bp, sizeof(argp->txnid->txnid));
+	bp += sizeof(argp->txnid->txnid);
+
+	memcpy(&argp->prev_lsn, bp, sizeof(DB_LSN));
+	bp += sizeof(DB_LSN);
+
+	memcpy(&uinttmp, bp, sizeof(uinttmp));
+	argp->opcode = (u_int32_t)uinttmp;
+	bp += sizeof(uinttmp);
+
+	memset(&argp->name, 0, sizeof(argp->name));
+	memcpy(&argp->name.size, bp, sizeof(u_int32_t));
+	bp += sizeof(u_int32_t);
+	argp->name.data = bp;
+	bp += argp->name.size;
+
+	memset(&argp->uid, 0, sizeof(argp->uid));
+	memcpy(&argp->uid.size, bp, sizeof(u_int32_t));
+	bp += sizeof(u_int32_t);
+	argp->uid.data = bp;
+	bp += argp->uid.size;
+
+	memcpy(&uinttmp, bp, sizeof(uinttmp));
+	argp->fileid = (int32_t)uinttmp;
+	bp += sizeof(uinttmp);
+
+	memcpy(&uinttmp, bp, sizeof(uinttmp));
+	argp->ftype = (DBTYPE)uinttmp;
+	bp += sizeof(uinttmp);
+
+	memcpy(&uinttmp, bp, sizeof(uinttmp));
+	argp->meta_pgno = (db_pgno_t)uinttmp;
+	bp += sizeof(uinttmp);
+
+	memcpy(&uinttmp, bp, sizeof(uinttmp));
+	argp->id = (u_int32_t)uinttmp;
+	bp += sizeof(uinttmp);
+
+	*argpp = argp;
+	return (0);
+}
+
+/*
+ * PUBLIC: int __dbreg_init_print __P((DB_ENV *, int (***)(DB_ENV *,
+ * PUBLIC:     DBT *, DB_LSN *, db_recops, void *), size_t *));
+ */
+int
+__dbreg_init_print(dbenv, dtabp, dtabsizep)
+	DB_ENV *dbenv;
+	int (***dtabp)__P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
+	size_t *dtabsizep;
+{
+	int ret;
+
+	if ((ret = __db_add_recovery(dbenv, dtabp, dtabsizep,
+	    __dbreg_register_print, DB___dbreg_register)) != 0)
+		return (ret);
+	return (0);
+}
+
+/*
+ * PUBLIC: int __dbreg_init_getpgnos __P((DB_ENV *,
+ * PUBLIC:     int (***)(DB_ENV *, DBT *, DB_LSN *, db_recops, void *),
+ * PUBLIC:     size_t *));
+ */
+int
+__dbreg_init_getpgnos(dbenv, dtabp, dtabsizep)
+	DB_ENV *dbenv;
+	int (***dtabp)__P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
+	size_t *dtabsizep;
+{
+	int ret;
+
+	if ((ret = __db_add_recovery(dbenv, dtabp, dtabsizep,
+	    __dbreg_register_getpgnos, DB___dbreg_register)) != 0)
+		return (ret);
+	return (0);
+}
+
+/*
+ * PUBLIC: int __dbreg_init_recover __P((DB_ENV *, int (***)(DB_ENV *,
+ * PUBLIC:     DBT *, DB_LSN *, db_recops, void *), size_t *));
+ */
+int
+__dbreg_init_recover(dbenv, dtabp, dtabsizep)
+	DB_ENV *dbenv;
+	int (***dtabp)__P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
+	size_t *dtabsizep;
+{
+	int ret;
+
+	if ((ret = __db_add_recovery(dbenv, dtabp, dtabsizep,
+	    __dbreg_register_recover, DB___dbreg_register)) != 0)
+		return (ret);
+	return (0);
+}
diff --git a/db/dbreg/dbreg_rec.c b/db/dbreg/dbreg_rec.c
new file mode 100644
index 000000000..1deefa88c
--- /dev/null
+++ b/db/dbreg/dbreg_rec.c
@@ -0,0 +1,363 @@
+/*-
+ * See the file LICENSE for redistribution information.
+ *
+ * Copyright (c) 1996-2002
+ *	Sleepycat Software.  All rights reserved.
+ */
+/*
+ * Copyright (c) 1995, 1996
+ *	The President and Fellows of Harvard University.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ * 3. Neither the name of the University nor the names of its contributors
+ *    may be used to endorse or promote products derived from this software
+ *    without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#include "db_config.h"
+
+#ifndef lint
+static const char revid[] = "Id: dbreg_rec.c,v 11.107 2002/08/08 15:44:46 bostic Exp ";
+#endif /* not lint */
+
+#ifndef NO_SYSTEM_INCLUDES
+#include <sys/types.h>
+
+#include <string.h>
+#endif
+
+#include "db_int.h"
+#include "dbinc/db_page.h"
+#include "dbinc/db_am.h"
+#include "dbinc/log.h"
+#include "dbinc/txn.h"
+
+static int __dbreg_open_file __P((DB_ENV *,
+    DB_TXN *, __dbreg_register_args *, void *));
+
+/*
+ * PUBLIC: int __dbreg_register_recover
+ * PUBLIC:     __P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
+ */
+int
+__dbreg_register_recover(dbenv, dbtp, lsnp, op, info)
+	DB_ENV *dbenv;
+	DBT *dbtp;
+	DB_LSN *lsnp;
+	db_recops op;
+	void *info;
+{
+	DB_ENTRY *dbe;
+	DB_LOG *dblp;
+	DB *dbp;
+	__dbreg_register_args *argp;
+	int do_close, do_open, do_rem, ret, t_ret;
+
+	dblp = dbenv->lg_handle;
+	dbp = NULL;
+
+#ifdef DEBUG_RECOVER
+	REC_PRINT(__dbreg_register_print);
+#endif
+	do_open = do_close = 0;
+	if ((ret = __dbreg_register_read(dbenv, dbtp->data, &argp)) != 0)
+		goto out;
+
+	switch (argp->opcode) {
+	case LOG_OPEN:
+		if ((DB_REDO(op) ||
+		    op == DB_TXN_OPENFILES || op == DB_TXN_POPENFILES))
+			do_open = 1;
+		else
+			do_close = 1;
+		break;
+
+	case LOG_CLOSE:
+		if (DB_UNDO(op))
+			do_open = 1;
+		else
+			do_close = 1;
+		break;
+	case LOG_RCLOSE:
+		/*
+		 * LOG_RCLOSE was generated by recover because a file
+		 * was left open.  The POPENFILES pass, which is run
+		 * to open files to abort prepared transactions,
+		 * may not include the open for this file so we
+		 * open it here.  Note that a normal CLOSE is
+		 * not legal before the prepared transaction is
+		 * committed or aborted.
+		 */
+		if (DB_UNDO(op) || op == DB_TXN_POPENFILES)
+			do_open = 1;
+		else
+			do_close = 1;
+		break;
+
+	case LOG_CHECKPOINT:
+		if (DB_UNDO(op) ||
+		    op == DB_TXN_OPENFILES || op == DB_TXN_POPENFILES)
+			do_open = 1;
+		break;
+	}
+
+	if (do_open) {
+		/*
+		 * We must open the db even if the meta page is not
+		 * yet written as we may be creating subdatabase.
+		 */
+		if (op == DB_TXN_OPENFILES && argp->opcode != LOG_CHECKPOINT)
+			F_SET(dblp, DBLOG_FORCE_OPEN);
+
+		/*
+		 * During an abort or an open pass to recover prepared txns,
+		 * we need to make sure that we use the same locker id on the
+		 * open.  We pass the txnid along to ensure this.
+		 */
+		ret = __dbreg_open_file(dbenv,
+		    op == DB_TXN_ABORT || op == DB_TXN_POPENFILES ?
+		    argp->txnid : NULL, argp, info);
+		if (ret == ENOENT || ret == EINVAL) {
+			/*
+			 * If this is an OPEN while rolling forward, it's
+			 * possible that the file was recreated since last
+			 * time we got here.  In that case, we've got deleted
+			 * set and probably shouldn't, so we need to check
+			 * for that case and possibly retry.
+			 */
+			if (op == DB_TXN_FORWARD_ROLL &&
+			    argp->txnid != 0 &&
+			    dblp->dbentry[argp->fileid].deleted) {
+				dblp->dbentry[argp->fileid].deleted = 0;
+				ret =
+				    __dbreg_open_file(dbenv, NULL, argp, info);
+			}
+			ret = 0;
+		}
+		F_CLR(dblp, DBLOG_FORCE_OPEN);
+	}
+
+	if (do_close) {
+		/*
+		 * If we are undoing an open, or redoing a close,
+		 * then we need to close the file.
+		 *
+		 * If the file is deleted, then we can just ignore this close.
+		 * Otherwise, we should usually have a valid dbp we should
+		 * close or whose reference count should be decremented.
+		 * However, if we shut down without closing a file, we may, in
+		 * fact, not have the file open, and that's OK.
+		 */
+		do_rem = 0;
+		MUTEX_THREAD_LOCK(dbenv, dblp->mutexp);
+		if (argp->fileid < dblp->dbentry_cnt) {
+			/*
+			 * Typically, closes should match an open which means
+			 * that if this is a close, there should be a valid
+			 * entry in the dbentry table when we get here,
+			 * however there is an exception.  If this is an
+			 * OPENFILES pass, then we may have started from
+			 * a log file other than the first, and the
+			 * corresponding open appears in an earlier file.
+			 * We can ignore that case, but all others are errors.
+			 */
+			dbe = &dblp->dbentry[argp->fileid];
+			if (dbe->dbp == NULL && !dbe->deleted) {
+				/* No valid entry here. */
+				if ((argp->opcode != LOG_CLOSE &&
+				    argp->opcode != LOG_RCLOSE) ||
+				    (op != DB_TXN_OPENFILES &&
+				    op !=DB_TXN_POPENFILES)) {
+					__db_err(dbenv,
+					    "Improper file close at %lu/%lu",
+					    (u_long)lsnp->file,
+					    (u_long)lsnp->offset);
+					ret = EINVAL;
+				}
+				MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+				goto done;
+			}
+
+			/* We have either an open entry or a deleted entry. */
+			if ((dbp = dbe->dbp) != NULL) {
+				MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+				(void)__dbreg_revoke_id(dbp, 0);
+
+				/*
+				 * If we're a replication client, it's
+				 * possible to get here with a dbp that
+				 * the user opened, but which we later
+				 * assigned a fileid to.  Be sure that
+				 * we only close dbps that we opened in
+				 * the recovery code;  they should have
+				 * DB_AM_RECOVER set.
+				 *
+				 * The only exception is if we're aborting
+				 * in a normal environment;  then we might
+				 * get here with a non-AM_RECOVER database.
+				 */
+				if (F_ISSET(dbp, DB_AM_RECOVER) ||
+				    op == DB_TXN_ABORT)
+					do_rem = 1;
+			} else if (dbe->deleted) {
+				MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+				__dbreg_rem_dbentry(dblp, argp->fileid);
+			}
+		} else
+			MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+		if (do_rem) {
+			/*
+			 * If we are undoing a create we'd better discard
+			 * any buffers from the memory pool.
+			 */
+			if (dbp != NULL && dbp->mpf != NULL && argp->id != 0) {
+				if ((ret = dbp->mpf->close(dbp->mpf,
+				    DB_MPOOL_DISCARD)) != 0)
+					goto out;
+				dbp->mpf = NULL;
+			}
+
+			/*
+			 * During recovery, all files are closed.  On an abort,
+			 * we only close the file if we opened it during the
+			 * abort (DB_AM_RECOVER set), otherwise we simply do
+			 * a __db_refresh.  For the close case, if remove or
+			 * rename has closed the file, don't request a sync,
+			 * because the NULL mpf would be a problem.
+			 */
+			if (dbp != NULL) {
+				if (op == DB_TXN_ABORT &&
+				    !F_ISSET(dbp, DB_AM_RECOVER))
+					t_ret =
+					    __db_refresh(dbp, NULL, DB_NOSYNC);
+				else
+					t_ret = dbp->close(dbp, DB_NOSYNC);
+				if (t_ret != 0 && ret == 0)
+					ret = t_ret;
+			}
+		}
+	}
+done:	if (ret == 0)
+		*lsnp = argp->prev_lsn;
+out:	if (argp != NULL)
+		__os_free(dbenv, argp);
+	return (ret);
+}
+
+/*
+ * __dbreg_open_file --
+ *	Called during log_register recovery.  Make sure that we have an
+ *	entry in the dbentry table for this ndx.  Returns 0 on success,
+ *	non-zero on error.
+ */
+static int
+__dbreg_open_file(dbenv, txn, argp, info)
+	DB_ENV *dbenv;
+	DB_TXN *txn;
+	__dbreg_register_args *argp;
+	void *info;
+{
+	DB_ENTRY *dbe;
+	DB_LOG *lp;
+	DB *dbp;
+	int ret;
+	u_int32_t id;
+
+	lp = (DB_LOG *)dbenv->lg_handle;
+	/*
+	 * We never re-open temporary files.  Temp files are only
+	 * useful during aborts in which case the dbp was entered
+	 * when the file was registered.  During recovery, we treat
+	 * temp files as properly deleted files, allowing the open to
+	 * fail and not reporting any errors when recovery fails to
+	 * get a valid dbp from __dbreg_id_to_db.
+	 */
+	if (argp->name.size == 0) {
+		(void)__dbreg_add_dbentry(dbenv, lp, NULL, argp->fileid);
+		return (ENOENT);
+	}
+
+	/*
+	 * When we're opening, we have to check that the name we are opening
+	 * is what we expect.  If it's not, then we close the old file and
+	 * open the new one.
+	 */
+	MUTEX_THREAD_LOCK(dbenv, lp->mutexp);
+	if (argp->fileid < lp->dbentry_cnt)
+		dbe = &lp->dbentry[argp->fileid];
+	else
+		dbe = NULL;
+
+	if (dbe != NULL) {
+		if (dbe->deleted) {
+			MUTEX_THREAD_UNLOCK(dbenv, lp->mutexp);
+			return (ENOENT);
+		}
+		if ((dbp = dbe->dbp) != NULL) {
+			if (dbp->meta_pgno != argp->meta_pgno ||
+			    memcmp(dbp->fileid,
+			    argp->uid.data, DB_FILE_ID_LEN) != 0) {
+				MUTEX_THREAD_UNLOCK(dbenv, lp->mutexp);
+				(void)__dbreg_revoke_id(dbp, 0);
+				if (F_ISSET(dbp, DB_AM_RECOVER))
+					dbp->close(dbp, DB_NOSYNC);
+				goto reopen;
+			}
+
+			/*
+			 * We should only get here if we already have the
+			 * dbp from an openfiles pass, in which case, what's
+			 * here had better be the same dbp.
+			 */
+			DB_ASSERT(dbe->dbp == dbp);
+			MUTEX_THREAD_UNLOCK(dbenv, lp->mutexp);
+
+			/*
+			 * This is a successful open.  We need to record that
+			 * in the txnlist so that we know how to handle the
+			 * subtransaction that created the file system object.
+			 */
+			if (argp->id != TXN_INVALID &&
+			    (ret = __db_txnlist_update(dbenv, info,
+			    argp->id, TXN_EXPECTED, NULL)) == TXN_NOTFOUND)
+				ret = __db_txnlist_add(dbenv,
+				    info, argp->id, TXN_EXPECTED, NULL);
+			return (0);
+		}
+	}
+
+	MUTEX_THREAD_UNLOCK(dbenv, lp->mutexp);
+
+	/*
+	 * We are about to pass a recovery txn pointer into the main library.
+	 * We need to make sure that any accessed fields are set appropriately.
+	 */
+reopen:	if (txn != NULL) {
+		id = txn->txnid;
+		memset(txn, 0, sizeof(DB_TXN));
+		txn->txnid = id;
+		txn->mgrp = dbenv->tx_handle;
+	}
+
+	return (__dbreg_do_open(dbenv, txn, lp, argp->uid.data, argp->name.data,
+	    argp->ftype, argp->fileid, argp->meta_pgno, info, argp->id));
+}
diff --git a/db/dbreg/dbreg_util.c b/db/dbreg/dbreg_util.c
new file mode 100644
index 000000000..f625062c7
--- /dev/null
+++ b/db/dbreg/dbreg_util.c
@@ -0,0 +1,689 @@
+/*-
+ * See the file LICENSE for redistribution information.
+ *
+ * Copyright (c) 1997-2002
+ *	Sleepycat Software.  All rights reserved.
+ */
+
+#include "db_config.h"
+
+#ifndef lint
+static const char revid[] = "Id: dbreg_util.c,v 11.13 2002/08/06 06:11:23 bostic Exp ";
+#endif /* not lint */
+
+#ifndef NO_SYSTEM_INCLUDES
+#include <sys/types.h>
+#include <string.h>
+#endif
+
+#include "db_int.h"
+#include "dbinc/db_page.h"
+#include "dbinc/db_am.h"
+#include "dbinc/log.h"
+#include "dbinc/txn.h"
+
+static int __dbreg_check_master __P((DB_ENV *, u_int8_t *, char *));
+
+/*
+ * __dbreg_add_dbentry --
+ *	Adds a DB entry to the dbreg DB entry table.
+ *
+ * PUBLIC: int __dbreg_add_dbentry __P((DB_ENV *, DB_LOG *, DB *, int32_t));
+ */
+int
+__dbreg_add_dbentry(dbenv, dblp, dbp, ndx)
+	DB_ENV *dbenv;
+	DB_LOG *dblp;
+	DB *dbp;
+	int32_t ndx;
+{
+	int32_t i;
+	int ret;
+
+	ret = 0;
+
+	MUTEX_THREAD_LOCK(dbenv, dblp->mutexp);
+
+	/*
+	 * Check if we need to grow the table.  Note, ndx is 0-based (the
+	 * index into the DB entry table) an dbentry_cnt is 1-based, the
+	 * number of available slots.
+	 */
+	if (dblp->dbentry_cnt <= ndx) {
+		if ((ret = __os_realloc(dbenv,
+		    (ndx + DB_GROW_SIZE) * sizeof(DB_ENTRY),
+		    &dblp->dbentry)) != 0)
+			goto err;
+
+		/* Initialize the new entries. */
+		for (i = dblp->dbentry_cnt; i < ndx + DB_GROW_SIZE; i++) {
+			dblp->dbentry[i].dbp = NULL;
+			dblp->dbentry[i].deleted = 0;
+		}
+		dblp->dbentry_cnt = i;
+	}
+
+	DB_ASSERT(dblp->dbentry[ndx].dbp == NULL);
+	dblp->dbentry[ndx].deleted = dbp == NULL;
+	dblp->dbentry[ndx].dbp = dbp;
+
+err:	MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+	return (ret);
+}
+
+/*
+ * __dbreg_rem_dbentry
+ *	Remove an entry from the DB entry table.  Find the appropriate DB and
+ * unlink it from the linked list off the table.  If the DB is NULL, treat
+ * this as a simple refcount decrement.
+ *
+ * PUBLIC: void __dbreg_rem_dbentry __P((DB_LOG *, int32_t));
+ */
+void
+__dbreg_rem_dbentry(dblp, ndx)
+	DB_LOG *dblp;
+	int32_t ndx;
+{
+	MUTEX_THREAD_LOCK(dblp->dbenv, dblp->mutexp);
+	dblp->dbentry[ndx].dbp = NULL;
+	dblp->dbentry[ndx].deleted = 0;
+	MUTEX_THREAD_UNLOCK(dblp->dbenv, dblp->mutexp);
+}
+
+/*
+ * __dbreg_open_files --
+ *	Put a LOG_CHECKPOINT log record for each open database.
+ *
+ * PUBLIC: int __dbreg_open_files __P((DB_ENV *));
+ */
+int
+__dbreg_open_files(dbenv)
+	DB_ENV *dbenv;
+{
+	DB_LOG *dblp;
+	DB_LSN r_unused;
+	DBT *dbtp, fid_dbt, t;
+	FNAME *fnp;
+	LOG *lp;
+	int ret;
+
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+
+	ret = 0;
+
+	MUTEX_LOCK(dbenv, &lp->fq_mutex);
+
+	for (fnp = SH_TAILQ_FIRST(&lp->fq, __fname);
+	    fnp != NULL; fnp = SH_TAILQ_NEXT(fnp, q, __fname)) {
+		if (fnp->name_off == INVALID_ROFF)
+			dbtp = NULL;
+		else {
+			memset(&t, 0, sizeof(t));
+			t.data = R_ADDR(&dblp->reginfo, fnp->name_off);
+			t.size = (u_int32_t)strlen(t.data) + 1;
+			dbtp = &t;
+		}
+		memset(&fid_dbt, 0, sizeof(fid_dbt));
+		fid_dbt.data = fnp->ufid;
+		fid_dbt.size = DB_FILE_ID_LEN;
+		/*
+		 * Output LOG_CHECKPOINT records which will be
+		 * processed during the OPENFILES pass of recovery.
+		 * At the end of recovery we want to output the
+		 * files that were open so that a future recovery
+		 * run will have the correct files open during
+		 * a backward pass.  For this we output LOG_RCLOSE
+		 * records so that the files will be closed on
+		 * the forward pass.
+		 */
+		if ((ret = __dbreg_register_log(dbenv,
+		    NULL, &r_unused, 0,
+		    F_ISSET(dblp, DBLOG_RECOVER) ? LOG_RCLOSE : LOG_CHECKPOINT,
+		    dbtp, &fid_dbt, fnp->id, fnp->s_type, fnp->meta_pgno,
+		    TXN_INVALID)) != 0)
+			break;
+	}
+
+	MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+
+	return (ret);
+}
+
+/*
+ * __dbreg_close_files --
+ *	Close files that were opened by the recovery daemon.  We sync the
+ *	file, unless its mpf pointer has been NULLed by a db_remove or
+ *	db_rename.  We may not have flushed the log_register record that
+ *	closes the file.
+ *
+ * PUBLIC: int __dbreg_close_files __P((DB_ENV *));
+ */
+int
+__dbreg_close_files(dbenv)
+	DB_ENV *dbenv;
+{
+	DB_LOG *dblp;
+	DB *dbp;
+	int ret, t_ret;
+	int32_t i;
+
+	/* If we haven't initialized logging, we have nothing to do. */
+	if (!LOGGING_ON(dbenv))
+		return (0);
+
+	dblp = dbenv->lg_handle;
+	ret = 0;
+	MUTEX_THREAD_LOCK(dbenv, dblp->mutexp);
+	for (i = 0; i < dblp->dbentry_cnt; i++) {
+		/* We only want to close dbps that recovery opened. */
+		if ((dbp = dblp->dbentry[i].dbp) != NULL &&
+		    F_ISSET(dbp, DB_AM_RECOVER)) {
+			/*
+			 * It's unsafe to call DB->close while holding the
+			 * thread lock, because we'll call __dbreg_rem_dbentry
+			 * and grab it again.
+			 *
+			 * Just drop it.  Since dbreg ids go monotonically
+			 * upward, concurrent opens should be safe, and the
+			 * user should have no business closing files while
+			 * we're in this loop anyway--we're in the process of
+			 * making all outstanding dbps invalid.
+			 */
+			MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+			if ((t_ret = dbp->close(dbp,
+			    dbp->mpf == NULL ? DB_NOSYNC : 0)) != 0 && ret == 0)
+				ret = t_ret;
+			MUTEX_THREAD_LOCK(dbenv, dblp->mutexp);
+		}
+		dblp->dbentry[i].deleted = 0;
+		dblp->dbentry[i].dbp = NULL;
+	}
+	MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+	return (ret);
+}
+
+/*
+ * __dbreg_id_to_db --
+ *	Return the DB corresponding to the specified dbreg id.
+ *
+ * PUBLIC: int __dbreg_id_to_db __P((DB_ENV *, DB_TXN *, DB **, int32_t, int));
+ */
+int
+__dbreg_id_to_db(dbenv, txn, dbpp, ndx, inc)
+	DB_ENV *dbenv;
+	DB_TXN *txn;
+	DB **dbpp;
+	int32_t ndx;
+	int inc;
+{
+	DB_LOG *dblp;
+	FNAME *fname;
+	int ret;
+	char *name;
+
+	ret = 0;
+	dblp = dbenv->lg_handle;
+	COMPQUIET(inc, 0);
+
+	MUTEX_THREAD_LOCK(dbenv, dblp->mutexp);
+
+	/*
+	 * Under XA, a process different than the one issuing DB operations
+	 * may abort a transaction.  In this case, the "recovery" routines
+	 * are run by a process that does not necessarily have the file open,
+	 * so we we must open the file explicitly.
+	 */
+	if (ndx >= dblp->dbentry_cnt ||
+	    (!dblp->dbentry[ndx].deleted && dblp->dbentry[ndx].dbp == NULL)) {
+		if (F_ISSET(dblp, DBLOG_RECOVER)) {
+			ret = ENOENT;
+			goto err;
+		}
+
+		/*
+		 * __dbreg_id_to_fname acquires the region's fq_mutex,
+		 * which we can't safely acquire while we hold the thread lock.
+		 * We no longer need it anyway--the dbentry table didn't
+		 * have what we needed.
+		 */
+		MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+
+		if (__dbreg_id_to_fname(dblp, ndx, 0, &fname) != 0)
+			/*
+			 * With transactional opens, we may actually have
+			 * closed this file in the transaction in which
+			 * case this will fail too.  Then it's up to the
+			 * caller to reopen the file.
+			 */
+			return (ENOENT);
+
+		/*
+		 * Note that we're relying on fname not to change, even
+		 * though we released the mutex that protects it (fq_mutex)
+		 * inside __dbreg_id_to_fname.  This should be a safe
+		 * assumption, because the other process that has the file
+		 * open shouldn't be closing it while we're trying to abort.
+		 */
+		name = R_ADDR(&dblp->reginfo, fname->name_off);
+
+		/*
+		 * At this point, we are not holding the thread lock, so exit
+		 * directly instead of going through the exit code at the
+		 * bottom.  If the __dbreg_do_open succeeded, then we don't need
+		 * to do any of the remaining error checking at the end of this
+		 * routine.
+		 * XXX I am sending a NULL txnlist and 0 txnid which may be
+		 * completely broken ;(
+		 */
+		if ((ret = __dbreg_do_open(dbenv, txn, dblp,
+		    fname->ufid, name, fname->s_type,
+		    ndx, fname->meta_pgno, NULL, 0)) != 0)
+			return (ret);
+
+		*dbpp = dblp->dbentry[ndx].dbp;
+		return (0);
+	}
+
+	/*
+	 * Return DB_DELETED if the file has been deleted (it's not an error).
+	 */
+	if (dblp->dbentry[ndx].deleted) {
+		ret = DB_DELETED;
+		goto err;
+	}
+
+	/* It's an error if we don't have a corresponding writeable DB. */
+	if ((*dbpp = dblp->dbentry[ndx].dbp) == NULL)
+		ret = ENOENT;
+
+err:	MUTEX_THREAD_UNLOCK(dbenv, dblp->mutexp);
+	return (ret);
+}
+
+/*
+ * __dbreg_id_to_fname --
+ *	Traverse the shared-memory region looking for the entry that
+ *	matches the passed dbreg id.  Returns 0 on success; -1 on error.
+ *
+ * PUBLIC: int __dbreg_id_to_fname __P((DB_LOG *, int32_t, int, FNAME **));
+ */
+int
+__dbreg_id_to_fname(dblp, lid, have_lock, fnamep)
+	DB_LOG *dblp;
+	int32_t lid;
+	int have_lock;
+	FNAME **fnamep;
+{
+	DB_ENV *dbenv;
+	FNAME *fnp;
+	LOG *lp;
+	int ret;
+
+	dbenv = dblp->dbenv;
+	lp = dblp->reginfo.primary;
+
+	ret = -1;
+
+	if (!have_lock)
+		MUTEX_LOCK(dbenv, &lp->fq_mutex);
+	for (fnp = SH_TAILQ_FIRST(&lp->fq, __fname);
+	    fnp != NULL; fnp = SH_TAILQ_NEXT(fnp, q, __fname)) {
+		if (fnp->id == lid) {
+			*fnamep = fnp;
+			ret = 0;
+			break;
+		}
+	}
+	if (!have_lock)
+		MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+
+	return (ret);
+}
+/*
+ * __dbreg_fid_to_fname --
+ *	Traverse the shared-memory region looking for the entry that
+ *	matches the passed file unique id.  Returns 0 on success; -1 on error.
+ *
+ * PUBLIC: int __dbreg_fid_to_fname __P((DB_LOG *, u_int8_t *, int, FNAME **));
+ */
+int
+__dbreg_fid_to_fname(dblp, fid, have_lock, fnamep)
+	DB_LOG *dblp;
+	u_int8_t *fid;
+	int have_lock;
+	FNAME **fnamep;
+{
+	DB_ENV *dbenv;
+	FNAME *fnp;
+	LOG *lp;
+	int ret;
+
+	dbenv = dblp->dbenv;
+	lp = dblp->reginfo.primary;
+
+	ret = -1;
+
+	if (!have_lock)
+		MUTEX_LOCK(dbenv, &lp->fq_mutex);
+	for (fnp = SH_TAILQ_FIRST(&lp->fq, __fname);
+	    fnp != NULL; fnp = SH_TAILQ_NEXT(fnp, q, __fname)) {
+		if (memcmp(fnp->ufid, fid, DB_FILE_ID_LEN) == 0) {
+			*fnamep = fnp;
+			ret = 0;
+			break;
+		}
+	}
+	if (!have_lock)
+		MUTEX_UNLOCK(dbenv, &lp->fq_mutex);
+
+	return (ret);
+}
+
+/*
+ * __dbreg_get_name
+ *
+ * Interface to get name of registered files.  This is mainly diagnostic
+ * and the name passed could be transient unless there is something
+ * ensuring that the file cannot be closed.
+ *
+ * PUBLIC: int __dbreg_get_name __P((DB_ENV *, u_int8_t *, char **));
+ */
+int
+__dbreg_get_name(dbenv, fid, namep)
+	DB_ENV *dbenv;
+	u_int8_t *fid;
+	char **namep;
+{
+	DB_LOG *dblp;
+	FNAME *fname;
+
+	dblp = dbenv->lg_handle;
+
+	if (dblp != NULL && __dbreg_fid_to_fname(dblp, fid, 0, &fname) == 0) {
+		*namep = R_ADDR(&dblp->reginfo, fname->name_off);
+		return (0);
+	}
+
+	return (-1);
+}
+
+/*
+ * __dbreg_do_open --
+ *	Open files referenced in the log.  This is the part of the open that
+ * is not protected by the thread mutex.
+ * PUBLIC: int __dbreg_do_open __P((DB_ENV *, DB_TXN *, DB_LOG *, u_int8_t *,
+ * PUBLIC:     char *, DBTYPE, int32_t, db_pgno_t, void *, u_int32_t));
+ */
+int
+__dbreg_do_open(dbenv,
+    txn, lp, uid, name, ftype, ndx, meta_pgno, info, id)
+	DB_ENV *dbenv;
+	DB_TXN *txn;
+	DB_LOG *lp;
+	u_int8_t *uid;
+	char *name;
+	DBTYPE ftype;
+	int32_t ndx;
+	db_pgno_t meta_pgno;
+	void *info;
+	u_int32_t id;
+{
+	DB *dbp;
+	int ret;
+	u_int32_t cstat;
+
+	if ((ret = db_create(&dbp, lp->dbenv, 0)) != 0)
+		return (ret);
+
+	/*
+	 * We can open files under a number of different scenarios.
+	 * First, we can open a file during a normal txn_abort, if that file
+	 * was opened and closed during the transaction (as is the master
+	 * database of a sub-database).
+	 * Second, we might be aborting a transaction in XA and not have
+	 * it open in the process that is actually doing the abort.
+	 * Third, we might be in recovery.
+	 * In case 3, there is no locking, so there is no issue.
+	 * In cases 1 and 2, we are guaranteed to already hold any locks
+	 * that we need, since we're still in the same transaction, so by
+	 * setting DB_AM_RECOVER, we guarantee that we don't log and that
+	 * we don't try to acquire locks on behalf of a different locker id.
+	 */
+	F_SET(dbp, DB_AM_RECOVER);
+	if (meta_pgno != PGNO_BASE_MD) {
+		memcpy(dbp->fileid, uid, DB_FILE_ID_LEN);
+		dbp->meta_pgno = meta_pgno;
+	}
+	dbp->type = ftype;
+	if ((ret = __db_dbopen(dbp, txn, name, NULL,
+	    DB_ODDFILESIZE, __db_omode("rw----"), meta_pgno)) == 0) {
+		/*
+		 * Verify that we are opening the same file that we were
+		 * referring to when we wrote this log record.
+		 */
+		if ((meta_pgno != PGNO_BASE_MD &&
+		    __dbreg_check_master(dbenv, uid, name) != 0) ||
+		    memcmp(uid, dbp->fileid, DB_FILE_ID_LEN) != 0)
+			cstat = TXN_IGNORE;
+		else
+			cstat = TXN_EXPECTED;
+
+		/* Assign the specific dbreg id to this dbp. */
+		if ((ret = __dbreg_assign_id(dbp, ndx)) != 0)
+			goto err;
+
+		/*
+		 * If we successfully opened this file, then we need to
+		 * convey that information to the txnlist so that we
+		 * know how to handle the subtransaction that created
+		 * the file system object.
+		 */
+		if (id != TXN_INVALID) {
+			if ((ret = __db_txnlist_update(dbenv,
+			    info, id, cstat, NULL)) == TXN_NOTFOUND)
+				ret = __db_txnlist_add(dbenv,
+				    info, id, cstat, NULL);
+			else if (ret > 0)
+				ret = 0;
+		}
+err:		if (cstat == TXN_IGNORE)
+			goto not_right;
+		return (ret);
+	} else {
+		/* Record that the open failed in the txnlist. */
+		if (id != TXN_INVALID && (ret = __db_txnlist_update(dbenv,
+		    info, id, TXN_UNEXPECTED, NULL)) == TXN_NOTFOUND)
+			ret = __db_txnlist_add(dbenv,
+			    info, id, TXN_UNEXPECTED, NULL);
+	}
+not_right:
+	(void)dbp->close(dbp, 0);
+	/* Add this file as deleted. */
+	(void)__dbreg_add_dbentry(dbenv, lp, NULL, ndx);
+	return (ENOENT);
+}
+
+static int
+__dbreg_check_master(dbenv, uid, name)
+	DB_ENV *dbenv;
+	u_int8_t *uid;
+	char *name;
+{
+	DB *dbp;
+	int ret;
+
+	ret = 0;
+	if ((ret = db_create(&dbp, dbenv, 0)) != 0)
+		return (ret);
+	dbp->type = DB_BTREE;
+	F_SET(dbp, DB_AM_RECOVER);
+	ret = __db_dbopen(dbp,
+	    NULL, name, NULL, 0, __db_omode("rw----"), PGNO_BASE_MD);
+
+	if (ret == 0 && memcmp(uid, dbp->fileid, DB_FILE_ID_LEN) != 0)
+		ret = EINVAL;
+
+	(void)dbp->close(dbp, 0);
+	return (ret);
+}
+
+/*
+ * __dbreg_lazy_id --
+ *	When a replication client gets upgraded to being a replication master,
+ * it may have database handles open that have not been assigned an ID, but
+ * which have become legal to use for logging.
+ *
+ *	This function lazily allocates a new ID for such a function, in a
+ * new transaction created for the purpose.  We need to do this in a new
+ * transaction because we definitely wish to commit the dbreg_register, but
+ * at this point we have no way of knowing whether the log record that incited
+ * us to call this will be part of a committed transaction.
+ *
+ * PUBLIC: int __dbreg_lazy_id __P((DB *));
+ */
+int
+__dbreg_lazy_id(dbp)
+	DB *dbp;
+{
+	DB_ENV *dbenv;
+	DB_TXN *txn;
+	int ret;
+
+	dbenv = dbp->dbenv;
+
+	DB_ASSERT(F_ISSET(dbenv, DB_ENV_REP_MASTER));
+
+	if ((ret = dbenv->txn_begin(dbenv, NULL, &txn, 0)) != 0)
+		return (ret);
+
+	if ((ret = __dbreg_new_id(dbp, txn)) != 0) {
+		(void)txn->abort(txn);
+		return (ret);
+	}
+
+	return (txn->commit(txn, DB_TXN_NOSYNC));
+}
+
+/*
+ * __dbreg_push_id and __dbreg_pop_id --
+ *	Dbreg ids from closed files are kept on a stack in shared memory
+ * for recycling.  (We want to reuse them as much as possible because each
+ * process keeps open files in an array by ID.)  Push them to the stack and
+ * pop them from it, managing memory as appropriate.
+ *
+ * The stack is protected by the fq_mutex, and in both functions we assume
+ * that this is already locked.
+ *
+ * PUBLIC: int __dbreg_push_id __P((DB_ENV *, int32_t));
+ * PUBLIC: int __dbreg_pop_id __P((DB_ENV *, int32_t *));
+ */
+int
+__dbreg_push_id(dbenv, id)
+	DB_ENV *dbenv;
+	int32_t id;
+{
+	DB_LOG *dblp;
+	LOG *lp;
+	int32_t *stack, *newstack;
+	int ret;
+
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+
+	if (lp->free_fid_stack != INVALID_ROFF)
+		stack = R_ADDR(&dblp->reginfo, lp->free_fid_stack);
+	else
+		stack = NULL;
+
+	/* Check if we have room on the stack. */
+	if (lp->free_fids_alloced <= lp->free_fids + 1) {
+		R_LOCK(dbenv, &dblp->reginfo);
+		if ((ret = __db_shalloc(dblp->reginfo.addr,
+		    (lp->free_fids_alloced + 20) * sizeof(u_int32_t), 0,
+		    &newstack)) != 0) {
+			R_UNLOCK(dbenv, &dblp->reginfo);
+			return (ret);
+		}
+
+		memcpy(newstack, stack,
+		    lp->free_fids_alloced * sizeof(u_int32_t));
+		lp->free_fid_stack = R_OFFSET(&dblp->reginfo, newstack);
+		lp->free_fids_alloced += 20;
+
+		if (stack != NULL)
+			__db_shalloc_free(dblp->reginfo.addr, stack);
+
+		stack = newstack;
+		R_UNLOCK(dbenv, &dblp->reginfo);
+	}
+
+	DB_ASSERT(stack != NULL);
+	stack[lp->free_fids++] = id;
+	return (0);
+}
+
+int
+__dbreg_pop_id(dbenv, id)
+	DB_ENV *dbenv;
+	int32_t *id;
+{
+	DB_LOG *dblp;
+	LOG *lp;
+	int32_t *stack;
+
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+
+	/* Do we have anything to pop? */
+	if (lp->free_fid_stack != INVALID_ROFF && lp->free_fids > 0) {
+		stack = R_ADDR(&dblp->reginfo, lp->free_fid_stack);
+		*id = stack[--lp->free_fids];
+	} else
+		*id = DB_LOGFILEID_INVALID;
+
+	return (0);
+}
+
+/*
+ * __dbreg_pluck_id --
+ *	Remove a particular dbreg id from the stack of free ids.  This is
+ * used when we open a file, as in recovery, with a specific ID that might
+ * be on the stack.
+ *
+ * Returns success whether or not the particular id was found, and like
+ * push and pop, assumes that the fq_mutex is locked.
+ *
+ * PUBLIC: int __dbreg_pluck_id __P((DB_ENV *, int32_t));
+ */
+int
+__dbreg_pluck_id(dbenv, id)
+	DB_ENV *dbenv;
+	int32_t id;
+{
+	DB_LOG *dblp;
+	LOG *lp;
+	int32_t *stack;
+	int i;
+
+	dblp = dbenv->lg_handle;
+	lp = dblp->reginfo.primary;
+
+	/* Do we have anything to look at? */
+	if (lp->free_fid_stack != INVALID_ROFF) {
+		stack = R_ADDR(&dblp->reginfo, lp->free_fid_stack);
+		for (i = 0; i < lp->free_fids; i++)
+			if (id == stack[i]) {
+				/*
+				 * Found it.  Overwrite it with the top
+				 * id (which may harmlessly be itself),
+				 * and shorten the stack by one.
+				 */
+				stack[i] = stack[lp->free_fids - 1];
+				lp->free_fids--;
+				return (0);
+			}
+	}
+
+	return (0);
+}