1 files changed, 215 insertions, 72 deletions

diff --git a/bdb/db/db_meta.c b/bdb/db/db_meta.c
index 5b57c369454..015ef5c8fc7 100644
--- a/bdb/db/db_meta.c
+++ b/bdb/db/db_meta.c
@@ -1,7 +1,7 @@
 /*-
  * See the file LICENSE for redistribution information.
  *
- * Copyright (c) 1996, 1997, 1998, 1999, 2000
+ * Copyright (c) 1996-2002
  *	Sleepycat Software.  All rights reserved.
  */
 /*
@@ -43,7 +43,7 @@
 #include "db_config.h"
 
 #ifndef lint
-static const char revid[] = "$Id: db_meta.c,v 11.26 2001/01/16 21:57:19 ubell Exp $";
+static const char revid[] = "$Id: db_meta.c,v 11.61 2002/08/08 03:57:48 bostic Exp $";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -53,12 +53,37 @@ static const char revid[] = "$Id: db_meta.c,v 11.26 2001/01/16 21:57:19 ubell Ex
 #endif
 
 #include "db_int.h"
-#include "db_page.h"
-#include "db_shash.h"
-#include "lock.h"
-#include "txn.h"
-#include "db_am.h"
-#include "btree.h"
+#include "dbinc/db_page.h"
+#include "dbinc/db_shash.h"
+#include "dbinc/lock.h"
+#include "dbinc/db_am.h"
+
+static void __db_init_meta __P((void *, u_int32_t, db_pgno_t, u_int32_t));
+
+/*
+ * __db_init_meta --
+ *	Helper function for __db_new that initializes the important fields in
+ * a meta-data page (used instead of P_INIT).  We need to make sure that we
+ * retain the page number and LSN of the existing page.
+ */
+static void
+__db_init_meta(p, pgsize, pgno, pgtype)
+	void *p;
+	u_int32_t pgsize;
+	db_pgno_t pgno;
+	u_int32_t pgtype;
+{
+	DB_LSN save_lsn;
+	DBMETA *meta;
+
+	meta = (DBMETA *)p;
+	save_lsn = meta->lsn;
+	memset(meta, 0, sizeof(DBMETA));
+	meta->lsn = save_lsn;
+	meta->pagesize = pgsize;
+	meta->pgno = pgno;
+	meta->type = (u_int8_t)pgtype;
+}
 
 /*
  * __db_new --
@@ -75,60 +100,110 @@ __db_new(dbc, type, pagepp)
 	DBMETA *meta;
 	DB *dbp;
 	DB_LOCK metalock;
+	DB_LSN lsn;
+	DB_MPOOLFILE *mpf;
 	PAGE *h;
-	db_pgno_t pgno;
-	int ret;
+	db_pgno_t pgno, newnext;
+	int meta_flags, extend, ret;
 
-	dbp = dbc->dbp;
 	meta = NULL;
+	meta_flags = 0;
+	dbp = dbc->dbp;
+	mpf = dbp->mpf;
 	h = NULL;
+	newnext = PGNO_INVALID;
 
 	pgno = PGNO_BASE_MD;
 	if ((ret = __db_lget(dbc,
 	    LCK_ALWAYS, pgno, DB_LOCK_WRITE, 0, &metalock)) != 0)
 		goto err;
-	if ((ret = memp_fget(dbp->mpf, &pgno, 0, (PAGE **)&meta)) != 0)
+	if ((ret = mpf->get(mpf, &pgno, 0, (PAGE **)&meta)) != 0)
 		goto err;
-
 	if (meta->free == PGNO_INVALID) {
-		if ((ret = memp_fget(dbp->mpf, &pgno, DB_MPOOL_NEW, &h)) != 0)
-			goto err;
-		ZERO_LSN(h->lsn);
-		h->pgno = pgno;
+		pgno = meta->last_pgno + 1;
+		ZERO_LSN(lsn);
+		extend = 1;
 	} else {
 		pgno = meta->free;
-		if ((ret = memp_fget(dbp->mpf, &pgno, 0, &h)) != 0)
+		if ((ret = mpf->get(mpf, &pgno, 0, &h)) != 0)
 			goto err;
-		meta->free = h->next_pgno;
-		(void)memp_fset(dbp->mpf, (PAGE *)meta, DB_MPOOL_DIRTY);
+
+		/*
+		 * We want to take the first page off the free list and
+		 * then set meta->free to the that page's next_pgno, but
+		 * we need to log the change first.
+		 */
+		newnext = h->next_pgno;
+		lsn = h->lsn;
+		extend = 0;
 	}
 
-	DB_ASSERT(TYPE(h) == P_INVALID);
+	/*
+	 * Log the allocation before fetching the new page.  If we
+	 * don't have room in the log then we don't want to tell
+	 * mpool to extend the file.
+	 */
+	if (DBC_LOGGING(dbc)) {
+		if ((ret = __db_pg_alloc_log(dbp, dbc->txn, &LSN(meta), 0,
+		    &LSN(meta), PGNO_BASE_MD, &lsn, pgno,
+		    (u_int32_t)type, newnext)) != 0)
+			goto err;
+	} else
+		LSN_NOT_LOGGED(LSN(meta));
 
-	if (TYPE(h) != P_INVALID)
-		return (__db_panic(dbp->dbenv, EINVAL));
+	meta_flags = DB_MPOOL_DIRTY;
+	meta->free = newnext;
 
-	/* Log the change. */
-	if (DB_LOGGING(dbc)) {
-		if ((ret = __db_pg_alloc_log(dbp->dbenv,
-		    dbc->txn, &LSN(meta), 0, dbp->log_fileid,
-		    &LSN(meta), &h->lsn, h->pgno,
-		    (u_int32_t)type, meta->free)) != 0)
+	if (extend == 1) {
+		meta->last_pgno++;
+		if ((ret = mpf->get(mpf, &pgno, DB_MPOOL_NEW, &h)) != 0)
 			goto err;
-		LSN(h) = LSN(meta);
+		ZERO_LSN(h->lsn);
+		h->pgno = pgno;
+		DB_ASSERT(pgno == meta->last_pgno);
 	}
+	LSN(h) = LSN(meta);
+
+	DB_ASSERT(TYPE(h) == P_INVALID);
+
+	if (TYPE(h) != P_INVALID)
+		return (__db_panic(dbp->dbenv, EINVAL));
 
-	(void)memp_fput(dbp->mpf, (PAGE *)meta, DB_MPOOL_DIRTY);
+	(void)mpf->put(mpf, (PAGE *)meta, DB_MPOOL_DIRTY);
 	(void)__TLPUT(dbc, metalock);
 
-	P_INIT(h, dbp->pgsize, h->pgno, PGNO_INVALID, PGNO_INVALID, 0, type);
+	switch (type) {
+		case P_BTREEMETA:
+		case P_HASHMETA:
+		case P_QAMMETA:
+			__db_init_meta(h, dbp->pgsize, h->pgno, type);
+			break;
+		default:
+			P_INIT(h, dbp->pgsize,
+			    h->pgno, PGNO_INVALID, PGNO_INVALID, 0, type);
+			break;
+	}
+
+	/*
+	 * If dirty reads are enabled and we are in a transaction, we could
+	 * abort this allocation after the page(s) pointing to this
+	 * one have their locks downgraded.  This would permit dirty readers
+	 * to access this page which is ok, but they must be off the
+	 * page when we abort.  This will also prevent updates happening
+	 * to this page until we commit.
+	 */
+	if (F_ISSET(dbc->dbp, DB_AM_DIRTY) && dbc->txn != NULL) {
+		if ((ret = __db_lget(dbc, 0,
+		    h->pgno, DB_LOCK_WWRITE, 0, &metalock)) != 0)
+			goto err;
+	}
 	*pagepp = h;
 	return (0);
 
 err:	if (h != NULL)
-		(void)memp_fput(dbp->mpf, h, 0);
+		(void)mpf->put(mpf, h, 0);
 	if (meta != NULL)
-		(void)memp_fput(dbp->mpf, meta, 0);
+		(void)mpf->put(mpf, meta, meta_flags);
 	(void)__TLPUT(dbc, metalock);
 	return (ret);
 }
@@ -148,11 +223,13 @@ __db_free(dbc, h)
 	DB *dbp;
 	DBT ldbt;
 	DB_LOCK metalock;
+	DB_MPOOLFILE *mpf;
 	db_pgno_t pgno;
 	u_int32_t dirty_flag;
 	int ret, t_ret;
 
 	dbp = dbc->dbp;
+	mpf = dbp->mpf;
 
 	/*
 	 * Retrieve the metadata page and insert the page at the head of
@@ -163,43 +240,44 @@ __db_free(dbc, h)
 	dirty_flag = 0;
 	pgno = PGNO_BASE_MD;
 	if ((ret = __db_lget(dbc,
-	     LCK_ALWAYS, pgno, DB_LOCK_WRITE, 0, &metalock)) != 0)
+	    LCK_ALWAYS, pgno, DB_LOCK_WRITE, 0, &metalock)) != 0)
 		goto err;
-	if ((ret = memp_fget(dbp->mpf, &pgno, 0, (PAGE **)&meta)) != 0) {
+	if ((ret = mpf->get(mpf, &pgno, 0, (PAGE **)&meta)) != 0) {
 		(void)__TLPUT(dbc, metalock);
 		goto err;
 	}
 
 	DB_ASSERT(h->pgno != meta->free);
 	/* Log the change. */
-	if (DB_LOGGING(dbc)) {
+	if (DBC_LOGGING(dbc)) {
 		memset(&ldbt, 0, sizeof(ldbt));
 		ldbt.data = h;
-		ldbt.size = P_OVERHEAD;
-		if ((ret = __db_pg_free_log(dbp->dbenv,
-		    dbc->txn, &LSN(meta), 0, dbp->log_fileid, h->pgno,
-		    &LSN(meta), &ldbt, meta->free)) != 0) {
-			(void)memp_fput(dbp->mpf, (PAGE *)meta, 0);
+		ldbt.size = P_OVERHEAD(dbp);
+		if ((ret = __db_pg_free_log(dbp,
+		    dbc->txn, &LSN(meta), 0, h->pgno,
+		    &LSN(meta), PGNO_BASE_MD, &ldbt, meta->free)) != 0) {
+			(void)mpf->put(mpf, (PAGE *)meta, 0);
 			(void)__TLPUT(dbc, metalock);
-			return (ret);
+			goto err;
 		}
-		LSN(h) = LSN(meta);
-	}
+	} else
+		LSN_NOT_LOGGED(LSN(meta));
+	LSN(h) = LSN(meta);
 
 	P_INIT(h, dbp->pgsize, h->pgno, PGNO_INVALID, meta->free, 0, P_INVALID);
 
 	meta->free = h->pgno;
 
 	/* Discard the metadata page. */
-	if ((t_ret = memp_fput(dbp->mpf,
-	    (PAGE *)meta, DB_MPOOL_DIRTY)) != 0 && ret == 0)
+	if ((t_ret =
+	    mpf->put(mpf, (PAGE *)meta, DB_MPOOL_DIRTY)) != 0 && ret == 0)
 		ret = t_ret;
 	if ((t_ret = __TLPUT(dbc, metalock)) != 0 && ret == 0)
 		ret = t_ret;
 
 	/* Discard the caller's page reference. */
 	dirty_flag = DB_MPOOL_DIRTY;
-err:	if ((t_ret = memp_fput(dbp->mpf, h, dirty_flag)) != 0 && ret == 0)
+err:	if ((t_ret = mpf->put(mpf, h, dirty_flag)) != 0 && ret == 0)
 		ret = t_ret;
 
 	/*
@@ -227,44 +305,63 @@ __db_lprint(dbc)
 
 	if (LOCKING_ON(dbp->dbenv)) {
 		req.op = DB_LOCK_DUMP;
-		lock_vec(dbp->dbenv, dbc->locker, 0, &req, 1, NULL);
+		dbp->dbenv->lock_vec(dbp->dbenv, dbc->locker, 0, &req, 1, NULL);
 	}
 	return (0);
 }
 #endif
 
 /*
+ * Implement the rules for transactional locking.  We can release the previous
+ * lock if we are not in a transaction or COUPLE_ALWAYS is specifed (used in
+ * record locking).  If we are doing dirty reads then we can release read locks
+ * and down grade write locks.
+ */
+#define	DB_PUT_ACTION(dbc, action, lockp)				\
+	    (((action == LCK_COUPLE || action == LCK_COUPLE_ALWAYS) &&	\
+	    LOCK_ISSET(*lockp)) ?					\
+	    (dbc->txn == NULL || action == LCK_COUPLE_ALWAYS ||		\
+	    (F_ISSET(dbc, DBC_DIRTY_READ) &&				\
+	    (lockp)->mode == DB_LOCK_DIRTY)) ? LCK_COUPLE :		\
+	    (F_ISSET((dbc)->dbp, DB_AM_DIRTY) &&			\
+	    (lockp)->mode == DB_LOCK_WRITE) ? LCK_DOWNGRADE : 0 : 0)
+
+/*
  * __db_lget --
  *	The standard lock get call.
  *
  * PUBLIC: int __db_lget __P((DBC *,
- * PUBLIC:     int, db_pgno_t, db_lockmode_t, int, DB_LOCK *));
+ * PUBLIC:     int, db_pgno_t, db_lockmode_t, u_int32_t, DB_LOCK *));
  */
 int
-__db_lget(dbc, flags, pgno, mode, lkflags, lockp)
+__db_lget(dbc, action, pgno, mode, lkflags, lockp)
 	DBC *dbc;
-	int flags, lkflags;
+	int action;
 	db_pgno_t pgno;
 	db_lockmode_t mode;
+	u_int32_t lkflags;
 	DB_LOCK *lockp;
 {
 	DB *dbp;
 	DB_ENV *dbenv;
 	DB_LOCKREQ couple[2], *reqp;
-	int ret;
+	DB_TXN *txn;
+	int has_timeout, ret;
 
 	dbp = dbc->dbp;
 	dbenv = dbp->dbenv;
+	txn = dbc->txn;
 
 	/*
 	 * We do not always check if we're configured for locking before
 	 * calling __db_lget to acquire the lock.
 	 */
-	if (CDB_LOCKING(dbenv)
-	    || !LOCKING_ON(dbenv) || F_ISSET(dbc, DBC_COMPENSATE)
-	    || (!LF_ISSET(LCK_ROLLBACK) && F_ISSET(dbc, DBC_RECOVER))
-	    || (!LF_ISSET(LCK_ALWAYS) && F_ISSET(dbc, DBC_OPD))) {
-		lockp->off = LOCK_INVALID;
+	if (CDB_LOCKING(dbenv) ||
+	    !LOCKING_ON(dbenv) || F_ISSET(dbc, DBC_COMPENSATE) ||
+	    (F_ISSET(dbc, DBC_RECOVER) &&
+	    (action != LCK_ROLLBACK || F_ISSET(dbenv, DB_ENV_REP_CLIENT))) ||
+	    (action != LCK_ALWAYS && F_ISSET(dbc, DBC_OPD))) {
+		LOCK_INIT(*lockp);
 		return (0);
 	}
 
@@ -282,27 +379,73 @@ __db_lget(dbc, flags, pgno, mode, lkflags, lockp)
 	if (DB_NONBLOCK(dbc))
 		lkflags |= DB_LOCK_NOWAIT;
 
-	/*
-	 * If the object not currently locked, acquire the lock and return,
-	 * otherwise, lock couple.
-	 */
-	if (LF_ISSET(LCK_COUPLE)) {
-		couple[0].op = DB_LOCK_GET;
+	if (F_ISSET(dbc, DBC_DIRTY_READ) && mode == DB_LOCK_READ)
+		mode = DB_LOCK_DIRTY;
+
+	has_timeout = txn != NULL && F_ISSET(txn, TXN_LOCKTIMEOUT);
+
+	switch (DB_PUT_ACTION(dbc, action, lockp)) {
+	case LCK_COUPLE:
+lck_couple:	couple[0].op = has_timeout? DB_LOCK_GET_TIMEOUT : DB_LOCK_GET;
 		couple[0].obj = &dbc->lock_dbt;
 		couple[0].mode = mode;
-		couple[1].op = DB_LOCK_PUT;
-		couple[1].lock = *lockp;
+		if (action == LCK_COUPLE_ALWAYS)
+			action = LCK_COUPLE;
+		UMRW_SET(couple[0].timeout);
+		if (has_timeout)
+			couple[0].timeout = txn->lock_timeout;
+		if (action == LCK_COUPLE) {
+			couple[1].op = DB_LOCK_PUT;
+			couple[1].lock = *lockp;
+		}
 
-		ret = lock_vec(dbenv,
-		    dbc->locker, lkflags, couple, 2, &reqp);
+		ret = dbenv->lock_vec(dbenv, dbc->locker,
+		    lkflags, couple, action == LCK_COUPLE ? 2 : 1, &reqp);
 		if (ret == 0 || reqp == &couple[1])
 			*lockp = couple[0].lock;
-	} else {
-		ret = lock_get(dbenv,
+		break;
+	case LCK_DOWNGRADE:
+		if ((ret = dbenv->lock_downgrade(
+		    dbenv, lockp, DB_LOCK_WWRITE, 0)) != 0)
+			return (ret);
+		/* FALL THROUGH */
+	default:
+		if (has_timeout)
+			goto lck_couple;
+		ret = dbenv->lock_get(dbenv,
 		    dbc->locker, lkflags, &dbc->lock_dbt, mode, lockp);
+		break;
+	}
+
+	return (ret);
+}
+
+/*
+ * __db_lput --
+ *	The standard lock put call.
+ *
+ * PUBLIC: int __db_lput __P((DBC *, DB_LOCK *));
+ */
+int
+__db_lput(dbc, lockp)
+	DBC *dbc;
+	DB_LOCK *lockp;
+{
+	DB_ENV *dbenv;
+	int ret;
 
-		if (ret != 0)
-			lockp->off = LOCK_INVALID;
+	dbenv = dbc->dbp->dbenv;
+
+	switch (DB_PUT_ACTION(dbc, LCK_COUPLE, lockp)) {
+	case LCK_COUPLE:
+		ret = dbenv->lock_put(dbenv, lockp);
+		break;
+	case LCK_DOWNGRADE:
+		ret = __lock_downgrade(dbenv, lockp, DB_LOCK_WWRITE, 0);
+		break;
+	default:
+		ret = 0;
+		break;
 	}
 
 	return (ret);