From ec239360d13518a13f572b635d036c7d10028010 Mon Sep 17 00:00:00 2001
From: Ulrich Drepper <drepper@redhat.com>
Date: Sun, 13 Jun 1999 13:36:34 +0000
Subject: Update.

	* db2/Makefile (distribute): Remove files which do not exist
	anymore.
---
 db2/mp/mp_bh.c     | 176 ++++++++++++++++++++++++++++-------------------------
 db2/mp/mp_fget.c   |  27 ++++----
 db2/mp/mp_fopen.c  | 123 ++++++++++++++++++++++++++-----------
 db2/mp/mp_fput.c   |  14 +++--
 db2/mp/mp_fset.c   |   4 +-
 db2/mp/mp_open.c   |  45 ++++++++++----
 db2/mp/mp_pr.c     |  26 ++++----
 db2/mp/mp_region.c |  42 ++++++++++---
 db2/mp/mp_sync.c   | 166 ++++++++++++++++++++++++++++++--------------------
 9 files changed, 382 insertions(+), 241 deletions(-)

(limited to 'db2/mp')

diff --git a/db2/mp/mp_bh.c b/db2/mp/mp_bh.c
index d89f9c2..12c5341 100644
--- a/db2/mp/mp_bh.c
+++ b/db2/mp/mp_bh.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_bh.c	10.38 (Sleepycat) 5/20/98";
+static const char sccsid[] = "@(#)mp_bh.c	10.45 (Sleepycat) 11/25/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -42,11 +42,13 @@ __memp_bhwrite(dbmp, mfp, bhp, restartp, wrotep)
 {
 	DB_MPOOLFILE *dbmfp;
 	DB_MPREG *mpreg;
+	int incremented, ret;
 
 	if (restartp != NULL)
 		*restartp = 0;
 	if (wrotep != NULL)
 		*wrotep = 0;
+	incremented = 0;
 
 	/*
 	 * Walk the process' DB_MPOOLFILE list and find a file descriptor for
@@ -63,6 +65,13 @@ __memp_bhwrite(dbmp, mfp, bhp, restartp, wrotep)
 				UNLOCKHANDLE(dbmp, dbmp->mutexp);
 				return (0);
 			}
+
+			/*
+			 * Increment the reference count -- see the comment in
+			 * memp_fclose().
+			 */
+			++dbmfp->ref;
+			incremented = 1;
 			break;
 		}
 	UNLOCKHANDLE(dbmp, dbmp->mutexp);
@@ -117,7 +126,15 @@ __memp_bhwrite(dbmp, mfp, bhp, restartp, wrotep)
 	    0, 0, mfp->stat.st_pagesize, 0, NULL, &dbmfp) != 0)
 		return (0);
 
-found:	return (__memp_pgwrite(dbmfp, bhp, restartp, wrotep));
+found:	ret = __memp_pgwrite(dbmfp, bhp, restartp, wrotep);
+
+	if (incremented) {
+		LOCKHANDLE(dbmp, dbmp->mutexp);
+		--dbmfp->ref;
+		UNLOCKHANDLE(dbmp, dbmp->mutexp);
+	}
+
+	return (ret);
 }
 
 /*
@@ -132,11 +149,12 @@ __memp_pgread(dbmfp, bhp, can_create)
 	BH *bhp;
 	int can_create;
 {
+	DB_IO db_io;
 	DB_MPOOL *dbmp;
 	MPOOLFILE *mfp;
-	size_t pagesize;
+	size_t len, pagesize;
 	ssize_t nr;
-	int ret;
+	int created, ret;
 
 	dbmp = dbmfp->dbmp;
 	mfp = dbmfp->mfp;
@@ -147,70 +165,63 @@ __memp_pgread(dbmfp, bhp, can_create)
 	UNLOCKREGION(dbmp);
 
 	/*
-	 * Temporary files may not yet have been created.
-	 *
-	 * Seek to the page location.
+	 * Temporary files may not yet have been created.  We don't create
+	 * them now, we create them when the pages have to be flushed.
 	 */
-	ret = 0;
-	LOCKHANDLE(dbmp, dbmfp->mutexp);
-	if (dbmfp->fd == -1 || (ret =
-	    __db_seek(dbmfp->fd, pagesize, bhp->pgno, 0, 0, SEEK_SET)) != 0) {
-		if (!can_create) {
-			if (dbmfp->fd == -1)
-				ret = EINVAL;
-			UNLOCKHANDLE(dbmp, dbmfp->mutexp);
+	nr = 0;
+	if (dbmfp->fd == -1)
+		ret = 0;
+	else {
+		/*
+		 * Ignore read errors if we have permission to create the page.
+		 * Assume that the page doesn't exist, and that we'll create it
+		 * when we write it out.
+		 */
+		db_io.fd_io = dbmfp->fd;
+		db_io.fd_lock = dbmp->reginfo.fd;
+		db_io.mutexp =
+		    F_ISSET(dbmp, MP_LOCKHANDLE) ? dbmfp->mutexp : NULL;
+		db_io.pagesize = db_io.bytes = pagesize;
+		db_io.pgno = bhp->pgno;
+		db_io.buf = bhp->buf;
+
+		ret = __os_io(&db_io, DB_IO_READ, &nr);
+	}
+
+	created = 0;
+	if (nr < (ssize_t)pagesize) {
+		if (can_create)
+			created = 1;
+		else {
+			/* If we had a short read, ret may be 0. */
+			if (ret == 0)
+				ret = EIO;
 			__db_err(dbmp->dbenv,
 			    "%s: page %lu doesn't exist, create flag not set",
 			    __memp_fn(dbmfp), (u_long)bhp->pgno);
 			goto err;
 		}
-		UNLOCKHANDLE(dbmp, dbmfp->mutexp);
-
-		/* Clear the created page. */
-		if (mfp->clear_len == 0)
-			memset(bhp->buf, 0, pagesize);
-		else {
-			memset(bhp->buf, 0, mfp->clear_len);
-#ifdef DIAGNOSTIC
-			memset(bhp->buf + mfp->clear_len,
-			    0xff, pagesize - mfp->clear_len);
-#endif
-		}
-
-		goto pgin;
 	}
 
 	/*
-	 * Read the page; short reads are treated like creates, although
-	 * any valid data is preserved.
+	 * Clear any bytes we didn't read that need to be cleared.  If we're
+	 * running in diagnostic mode, smash any bytes on the page that are
+	 * unknown quantities for the caller.
 	 */
-	ret = __db_read(dbmfp->fd, bhp->buf, pagesize, &nr);
-	UNLOCKHANDLE(dbmp, dbmfp->mutexp);
-	if (ret != 0)
-		goto err;
-
-	if (nr == (ssize_t)pagesize)
-		can_create = 0;
-	else {
-		if (!can_create) {
-			ret = EINVAL;
-			goto err;
-		}
-
-		/*
-		 * If we didn't fail until we tried the read, don't clear the
-		 * whole page, it wouldn't be insane for a filesystem to just
-		 * always behave that way.  Else, clear any uninitialized data.
-		 */
-		if (nr == 0)
-			memset(bhp->buf, 0,
-			    mfp->clear_len == 0 ? pagesize : mfp->clear_len);
-		else
-			memset(bhp->buf + nr, 0, pagesize - nr);
+	if (nr != (ssize_t)pagesize) {
+		len = mfp->clear_len == 0 ? pagesize : mfp->clear_len;
+		if (nr < (ssize_t)len)
+			memset(bhp->buf + nr, 0, len - nr);
+#ifdef DIAGNOSTIC
+		if (nr > (ssize_t)len)
+			len = nr;
+		if (len < pagesize)
+			memset(bhp->buf + len, 0xdb, pagesize - len);
+#endif
 	}
 
 	/* Call any pgin function. */
-pgin:	ret = mfp->ftype == 0 ? 0 : __memp_pg(dbmfp, bhp, 1);
+	ret = mfp->ftype == 0 ? 0 : __memp_pg(dbmfp, bhp, 1);
 
 	/* Unlock the buffer and reacquire the region lock. */
 err:	UNLOCKBUFFER(dbmp, bhp);
@@ -225,7 +236,7 @@ err:	UNLOCKBUFFER(dbmp, bhp);
 		F_CLR(bhp, BH_TRASH);
 
 		/* Update the statistics. */
-		if (can_create) {
+		if (created) {
 			++dbmp->mp->stat.st_page_create;
 			++mfp->stat.st_page_create;
 		} else {
@@ -250,12 +261,12 @@ __memp_pgwrite(dbmfp, bhp, restartp, wrotep)
 	int *restartp, *wrotep;
 {
 	DB_ENV *dbenv;
+	DB_IO db_io;
 	DB_LOG *lg_info;
 	DB_LSN lsn;
 	DB_MPOOL *dbmp;
 	MPOOL *mp;
 	MPOOLFILE *mfp;
-	size_t pagesize;
 	ssize_t nw;
 	int callpgin, ret, syncfail;
 	const char *fail;
@@ -270,7 +281,6 @@ __memp_pgwrite(dbmfp, bhp, restartp, wrotep)
 	if (wrotep != NULL)
 		*wrotep = 0;
 	callpgin = 0;
-	pagesize = mfp->stat.st_pagesize;
 
 	/*
 	 * Check the dirty bit -- this buffer may have been written since we
@@ -326,34 +336,32 @@ __memp_pgwrite(dbmfp, bhp, restartp, wrotep)
 	}
 
 	/* Temporary files may not yet have been created. */
-	LOCKHANDLE(dbmp, dbmfp->mutexp);
-	if (dbmfp->fd == -1 &&
-	    ((ret = __db_appname(dbenv, DB_APP_TMP, NULL, NULL,
-	    DB_CREATE | DB_EXCL | DB_TEMPORARY, &dbmfp->fd, NULL)) != 0 ||
-	    dbmfp->fd == -1)) {
+	if (dbmfp->fd == -1) {
+		LOCKHANDLE(dbmp, dbmfp->mutexp);
+		if (dbmfp->fd == -1 && ((ret = __db_appname(dbenv,
+		    DB_APP_TMP, NULL, NULL, DB_CREATE | DB_EXCL | DB_TEMPORARY,
+		    &dbmfp->fd, NULL)) != 0 || dbmfp->fd == -1)) {
+			UNLOCKHANDLE(dbmp, dbmfp->mutexp);
+			__db_err(dbenv,
+			    "unable to create temporary backing file");
+			goto err;
+		}
 		UNLOCKHANDLE(dbmp, dbmfp->mutexp);
-		__db_err(dbenv, "unable to create temporary backing file");
-		goto err;
 	}
 
-	/*
-	 * Write the page out.
-	 *
-	 * XXX
-	 * Shut the compiler up; it doesn't understand the correlation between
-	 * the failing clauses to __db_lseek and __db_write and this ret != 0.
-	 */
-	COMPQUIET(fail, NULL);
-	if ((ret =
-	    __db_seek(dbmfp->fd, pagesize, bhp->pgno, 0, 0, SEEK_SET)) != 0)
-		fail = "seek";
-	else if ((ret = __db_write(dbmfp->fd, bhp->buf, pagesize, &nw)) != 0)
+	/* Write the page. */
+	db_io.fd_io = dbmfp->fd;
+	db_io.fd_lock = dbmp->reginfo.fd;
+	db_io.mutexp = F_ISSET(dbmp, MP_LOCKHANDLE) ? dbmfp->mutexp : NULL;
+	db_io.pagesize = db_io.bytes = mfp->stat.st_pagesize;
+	db_io.pgno = bhp->pgno;
+	db_io.buf = bhp->buf;
+	if ((ret = __os_io(&db_io, DB_IO_WRITE, &nw)) != 0) {
+		__db_panic(dbenv, ret);
 		fail = "write";
-	UNLOCKHANDLE(dbmp, dbmfp->mutexp);
-	if (ret != 0)
 		goto syserr;
-
-	if (nw != (ssize_t)pagesize) {
+	}
+	if (nw != (ssize_t)mfp->stat.st_pagesize) {
 		ret = EIO;
 		fail = "write";
 		goto syserr;
@@ -394,7 +402,7 @@ __memp_pgwrite(dbmfp, bhp, restartp, wrotep)
 	if (F_ISSET(bhp, BH_WRITE)) {
 		if (mfp->lsn_cnt == 1) {
 			UNLOCKREGION(dbmp);
-			syncfail = __db_fsync(dbmfp->fd) != 0;
+			syncfail = __os_fsync(dbmfp->fd) != 0;
 			LOCKREGION(dbmp);
 			if (syncfail)
 				F_SET(mp, MP_LSN_RETRY);
@@ -574,11 +582,11 @@ __memp_upgrade(dbmp, dbmfp, mfp)
 		ret = 1;
 	} else {
 		/* Swap the descriptors and set the upgrade flag. */
-		(void)__db_close(dbmfp->fd);
+		(void)__os_close(dbmfp->fd);
 		dbmfp->fd = fd;
 		F_SET(dbmfp, MP_UPGRADE);
 		ret = 0;
 	}
-	FREES(rpath);
+	__os_freestr(rpath);
 	return (ret);
 }
diff --git a/db2/mp/mp_fget.c b/db2/mp/mp_fget.c
index 0777aa7..f159dc2 100644
--- a/db2/mp/mp_fget.c
+++ b/db2/mp/mp_fget.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_fget.c	10.48 (Sleepycat) 6/2/98";
+static const char sccsid[] = "@(#)mp_fget.c	10.53 (Sleepycat) 11/16/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -46,6 +46,8 @@ memp_fget(dbmfp, pgnoaddr, flags, addrp)
 	mp = dbmp->mp;
 	mfp = dbmfp->mfp;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/*
 	 * Validate arguments.
 	 *
@@ -79,12 +81,11 @@ memp_fget(dbmfp, pgnoaddr, flags, addrp)
 #ifdef DIAGNOSTIC
 	/*
 	 * XXX
-	 * We want to switch threads as often as possible.  Sleep every time
-	 * we get a new page to make it more likely.
+	 * We want to switch threads as often as possible.  Yield every time
+	 * we get a new page to ensure contention.
 	 */
-	if (DB_GLOBAL(db_pageyield) &&
-	    (__db_yield == NULL || __db_yield() != 0))
-		__db_sleep(0, 1);
+	if (DB_GLOBAL(db_pageyield))
+		__os_yield(1);
 #endif
 
 	/* Initialize remaining local variables. */
@@ -205,8 +206,8 @@ memp_fget(dbmfp, pgnoaddr, flags, addrp)
 			 * up running to the end of our CPU quantum as we will
 			 * simply be swapping between the two locks.
 			 */
-			if (!first && (__db_yield == NULL || __db_yield() != 0))
-				__db_sleep(0, 1);
+			if (!first)
+				__os_yield(1);
 
 			LOCKBUFFER(dbmp, bhp);
 			/* Wait for I/O to finish... */
@@ -240,7 +241,7 @@ memp_fget(dbmfp, pgnoaddr, flags, addrp)
 	}
 
 alloc:	/* Allocate new buffer header and data space. */
-	if ((ret = __memp_ralloc(dbmp, sizeof(BH) -
+	if ((ret = __memp_alloc(dbmp, sizeof(BH) -
 	    sizeof(u_int8_t) + mfp->stat.st_pagesize, NULL, &bhp)) != 0)
 		goto err;
 
@@ -285,7 +286,7 @@ alloc:	/* Allocate new buffer header and data space. */
 		else {
 			memset(bhp->buf, 0, mfp->clear_len);
 #ifdef DIAGNOSTIC
-			memset(bhp->buf + mfp->clear_len, 0xff,
+			memset(bhp->buf + mfp->clear_len, 0xdb,
 			    mfp->stat.st_pagesize - mfp->clear_len);
 #endif
 		}
@@ -335,11 +336,9 @@ done:	/* Update the chain search statistics. */
 		mp->stat.st_hash_examined += st_hsearch;
 	}
 
-	UNLOCKREGION(dbmp);
-
-	LOCKHANDLE(dbmp, dbmfp->mutexp);
 	++dbmfp->pinref;
-	UNLOCKHANDLE(dbmp, dbmfp->mutexp);
+
+	UNLOCKREGION(dbmp);
 
 	return (0);
 
diff --git a/db2/mp/mp_fopen.c b/db2/mp/mp_fopen.c
index a4cbac8..dd02662 100644
--- a/db2/mp/mp_fopen.c
+++ b/db2/mp/mp_fopen.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_fopen.c	10.47 (Sleepycat) 5/4/98";
+static const char sccsid[] = "@(#)mp_fopen.c	10.60 (Sleepycat) 1/1/99";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -43,6 +43,8 @@ memp_fopen(dbmp, path, flags, mode, pagesize, finfop, retp)
 {
 	int ret;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/* Validate arguments. */
 	if ((ret = __db_fchk(dbmp->dbenv,
 	    "memp_fopen", flags, DB_CREATE | DB_NOMMAP | DB_RDONLY)) != 0)
@@ -53,6 +55,8 @@ memp_fopen(dbmp, path, flags, mode, pagesize, finfop, retp)
 		__db_err(dbmp->dbenv, "memp_fopen: pagesize not specified");
 		return (EINVAL);
 	}
+	if (finfop != NULL && finfop->clear_len > pagesize)
+		return (EINVAL);
 
 	return (__memp_fopen(dbmp,
 	    NULL, path, flags, mode, pagesize, 1, finfop, retp));
@@ -80,7 +84,7 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 	DB_MPOOLFILE *dbmfp;
 	DB_MPOOL_FINFO finfo;
 	db_pgno_t last_pgno;
-	size_t size;
+	size_t maxmap;
 	u_int32_t mbytes, bytes;
 	int ret;
 	u_int8_t idbuf[DB_FILE_ID_LEN];
@@ -115,13 +119,11 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 	}
 
 	/* Allocate and initialize the per-process structure. */
-	if ((dbmfp =
-	    (DB_MPOOLFILE *)__db_calloc(1, sizeof(DB_MPOOLFILE))) == NULL) {
-		__db_err(dbenv, "memp_fopen: %s", strerror(ENOMEM));
-		return (ENOMEM);
-	}
+	if ((ret = __os_calloc(1, sizeof(DB_MPOOLFILE), &dbmfp)) != 0)
+		return (ret);
 	dbmfp->dbmp = dbmp;
 	dbmfp->fd = -1;
+	dbmfp->ref = 1;
 	if (LF_ISSET(DB_RDONLY))
 		F_SET(dbmfp, MP_READONLY);
 
@@ -132,7 +134,6 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 			ret = EINVAL;
 			goto err;
 		}
-		size = 0;
 		last_pgno = 0;
 	} else {
 		/* Get the real name for this file and open it. */
@@ -146,21 +147,40 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 			goto err;
 		}
 
-		/* Don't permit files that aren't a multiple of the pagesize. */
-		if ((ret = __db_ioinfo(rpath,
+		/*
+		 * Don't permit files that aren't a multiple of the pagesize,
+		 * and find the number of the last page in the file, all the
+		 * time being careful not to overflow 32 bits.
+		 *
+		 * !!!
+		 * We can't use off_t's here, or in any code in the mainline
+		 * library for that matter.  (We have to use them in the os
+		 * stubs, of course, as there are system calls that take them
+		 * as arguments.)  The reason is that some customers build in
+		 * environments where an off_t is 32-bits, but still run where
+		 * offsets are 64-bits, and they pay us a lot of money.
+		 */
+		if ((ret = __os_ioinfo(rpath,
 		    dbmfp->fd, &mbytes, &bytes, NULL)) != 0) {
 			__db_err(dbenv, "%s: %s", rpath, strerror(ret));
 			goto err;
 		}
-		if (bytes % pagesize) {
+
+		/* Page sizes have to be a power-of-two, ignore mbytes. */
+		if (bytes % pagesize != 0) {
 			__db_err(dbenv,
 			    "%s: file size not a multiple of the pagesize",
 			    rpath);
 			ret = EINVAL;
 			goto err;
 		}
-		size = mbytes * MEGABYTE + bytes;
-		last_pgno = size == 0 ? 0 : (size - 1) / pagesize;
+
+		last_pgno = mbytes * (MEGABYTE / pagesize);
+		last_pgno += bytes / pagesize;
+
+		/* Correction: page numbers are zero-based, not 1-based. */
+		if (last_pgno != 0)
+			--last_pgno;
 
 		/*
 		 * Get the file id if we weren't given one.  Generated file id's
@@ -168,7 +188,7 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 		 * other process joining the party.
 		 */
 		if (finfop->fileid == NULL) {
-			if ((ret = __db_fileid(dbenv, rpath, 0, idbuf)) != 0)
+			if ((ret = __os_fileid(dbenv, rpath, 0, idbuf)) != 0)
 				goto err;
 			finfop->fileid = idbuf;
 		}
@@ -191,7 +211,7 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 	}
 	if (ret == 0 &&
 	    F_ISSET(dbmp, MP_LOCKHANDLE) && (ret =
-	    __memp_ralloc(dbmp, sizeof(db_mutex_t), NULL, &dbmfp->mutexp)) == 0)
+	    __memp_alloc(dbmp, sizeof(db_mutex_t), NULL, &dbmfp->mutexp)) == 0)
 		LOCKINIT(dbmp, dbmfp->mutexp);
 
 	if (needlock)
@@ -232,13 +252,15 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 			F_CLR(mfp, MP_CAN_MMAP);
 		if (LF_ISSET(DB_NOMMAP))
 			F_CLR(mfp, MP_CAN_MMAP);
-		if (size > (dbenv == NULL || dbenv->mp_mmapsize == 0 ?
-		    DB_MAXMMAPSIZE : dbenv->mp_mmapsize))
+		maxmap = dbenv == NULL || dbenv->mp_mmapsize == 0 ?
+		    DB_MAXMMAPSIZE : dbenv->mp_mmapsize;
+		if (mbytes > maxmap / MEGABYTE ||
+		    (mbytes == maxmap / MEGABYTE && bytes >= maxmap % MEGABYTE))
 			F_CLR(mfp, MP_CAN_MMAP);
 	}
 	dbmfp->addr = NULL;
 	if (F_ISSET(mfp, MP_CAN_MMAP)) {
-		dbmfp->len = size;
+		dbmfp->len = (size_t)mbytes * MEGABYTE + bytes;
 		if (__db_mapfile(rpath,
 		    dbmfp->fd, dbmfp->len, 1, &dbmfp->addr) != 0) {
 			dbmfp->addr = NULL;
@@ -246,7 +268,7 @@ __memp_fopen(dbmp, mfp, path, flags, mode, pagesize, needlock, finfop, retp)
 		}
 	}
 	if (rpath != NULL)
-		FREES(rpath);
+		__os_freestr(rpath);
 
 	LOCKHANDLE(dbmp, dbmp->mutexp);
 	TAILQ_INSERT_TAIL(&dbmp->dbmfq, dbmfp, q);
@@ -260,11 +282,11 @@ err:	/*
 	 * never get to here after we have successfully allocated it.
 	 */
 	if (rpath != NULL)
-		FREES(rpath);
+		__os_freestr(rpath);
 	if (dbmfp->fd != -1)
-		(void)__db_close(dbmfp->fd);
+		(void)__os_close(dbmfp->fd);
 	if (dbmfp != NULL)
-		FREE(dbmfp, sizeof(DB_MPOOLFILE));
+		__os_free(dbmfp, sizeof(DB_MPOOLFILE));
 	return (ret);
 }
 
@@ -315,7 +337,7 @@ __memp_mf_open(dbmp, path, pagesize, last_pgno, finfop, retp)
 		}
 
 	/* Allocate a new MPOOLFILE. */
-	if ((ret = __memp_ralloc(dbmp, sizeof(MPOOLFILE), NULL, &mfp)) != 0)
+	if ((ret = __memp_alloc(dbmp, sizeof(MPOOLFILE), NULL, &mfp)) != 0)
 		return (ret);
 	*retp = mfp;
 
@@ -334,21 +356,22 @@ __memp_mf_open(dbmp, path, pagesize, last_pgno, finfop, retp)
 	mfp->stat.st_pagesize = pagesize;
 	mfp->orig_last_pgno = mfp->last_pgno = last_pgno;
 
-	F_SET(mfp, MP_CAN_MMAP);
 	if (ISTEMPORARY)
 		F_SET(mfp, MP_TEMP);
 	else {
 		/* Copy the file path into shared memory. */
-		if ((ret = __memp_ralloc(dbmp,
+		if ((ret = __memp_alloc(dbmp,
 		    strlen(path) + 1, &mfp->path_off, &p)) != 0)
 			goto err;
 		memcpy(p, path, strlen(path) + 1);
 
 		/* Copy the file identification string into shared memory. */
-		if ((ret = __memp_ralloc(dbmp,
+		if ((ret = __memp_alloc(dbmp,
 		    DB_FILE_ID_LEN, &mfp->fileid_off, &p)) != 0)
 			goto err;
 		memcpy(p, finfop->fileid, DB_FILE_ID_LEN);
+
+		F_SET(mfp, MP_CAN_MMAP);
 	}
 
 	/* Copy the page cookie into shared memory. */
@@ -356,7 +379,7 @@ __memp_mf_open(dbmp, path, pagesize, last_pgno, finfop, retp)
 		mfp->pgcookie_len = 0;
 		mfp->pgcookie_off = 0;
 	} else {
-		if ((ret = __memp_ralloc(dbmp,
+		if ((ret = __memp_alloc(dbmp,
 		    finfop->pgcookie->size, &mfp->pgcookie_off, &p)) != 0)
 			goto err;
 		memcpy(p, finfop->pgcookie->data, finfop->pgcookie->size);
@@ -394,16 +417,48 @@ memp_fclose(dbmfp)
 	dbmp = dbmfp->dbmp;
 	ret = 0;
 
+	MP_PANIC_CHECK(dbmp);
+
+	for (;;) {
+		LOCKHANDLE(dbmp, dbmp->mutexp);
+
+		/*
+		 * We have to reference count DB_MPOOLFILE structures as other
+		 * threads may be using them.  The problem only happens if the
+		 * application makes a bad design choice.  Here's the path:
+		 *
+		 * Thread A opens a database.
+		 * Thread B uses thread A's DB_MPOOLFILE to write a buffer
+		 *    in order to free up memory in the mpool cache.
+		 * Thread A closes the database while thread B is using the
+		 *    DB_MPOOLFILE structure.
+		 *
+		 * By opening all databases before creating the threads, and
+		 * closing them after the threads have exited, applications
+		 * get better performance and avoid the problem path entirely.
+		 *
+		 * Regardless, holding the DB_MPOOLFILE to flush a dirty buffer
+		 * is a short-term lock, even in worst case, since we better be
+		 * the only thread of control using the DB_MPOOLFILE structure
+		 * to read pages *into* the cache.  Wait until we're the only
+		 * reference holder and remove the DB_MPOOLFILE structure from
+		 * the list, so nobody else can even find it.
+		 */
+		if (dbmfp->ref == 1) {
+			TAILQ_REMOVE(&dbmp->dbmfq, dbmfp, q);
+			break;
+		}
+		UNLOCKHANDLE(dbmp, dbmp->mutexp);
+
+		(void)__os_sleep(1, 0);
+	}
+	UNLOCKHANDLE(dbmp, dbmp->mutexp);
+
 	/* Complain if pinned blocks never returned. */
 	if (dbmfp->pinref != 0)
 		__db_err(dbmp->dbenv, "%s: close: %lu blocks left pinned",
 		    __memp_fn(dbmfp), (u_long)dbmfp->pinref);
 
-	/* Remove the DB_MPOOLFILE structure from the list. */
-	LOCKHANDLE(dbmp, dbmp->mutexp);
-	TAILQ_REMOVE(&dbmp->dbmfq, dbmfp, q);
-	UNLOCKHANDLE(dbmp, dbmp->mutexp);
-
 	/* Close the underlying MPOOLFILE. */
 	(void)__memp_mf_close(dbmp, dbmfp);
 
@@ -414,7 +469,7 @@ memp_fclose(dbmfp)
 		    "%s: %s", __memp_fn(dbmfp), strerror(ret));
 
 	/* Close the file; temporary files may not yet have been created. */
-	if (dbmfp->fd != -1 && (t_ret = __db_close(dbmfp->fd)) != 0) {
+	if (dbmfp->fd != -1 && (t_ret = __os_close(dbmfp->fd)) != 0) {
 		__db_err(dbmp->dbenv,
 		    "%s: %s", __memp_fn(dbmfp), strerror(t_ret));
 		if (ret != 0)
@@ -429,7 +484,7 @@ memp_fclose(dbmfp)
 	}
 
 	/* Discard the DB_MPOOLFILE structure. */
-	FREE(dbmfp, sizeof(DB_MPOOLFILE));
+	__os_free(dbmfp, sizeof(DB_MPOOLFILE));
 
 	return (ret);
 }
diff --git a/db2/mp/mp_fput.c b/db2/mp/mp_fput.c
index 48fdfc3..c551f97 100644
--- a/db2/mp/mp_fput.c
+++ b/db2/mp/mp_fput.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_fput.c	10.22 (Sleepycat) 4/26/98";
+static const char sccsid[] = "@(#)mp_fput.c	10.24 (Sleepycat) 9/27/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -40,6 +40,8 @@ memp_fput(dbmfp, pgaddr, flags)
 	dbmp = dbmfp->dbmp;
 	mp = dbmp->mp;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/* Validate arguments. */
 	if (flags) {
 		if ((ret = __db_fchk(dbmp->dbenv, "memp_fput", flags,
@@ -57,15 +59,15 @@ memp_fput(dbmfp, pgaddr, flags)
 		}
 	}
 
+	LOCKREGION(dbmp);
+
 	/* Decrement the pinned reference count. */
-	LOCKHANDLE(dbmp, dbmfp->mutexp);
 	if (dbmfp->pinref == 0)
 		__db_err(dbmp->dbenv,
 		    "%s: put: more blocks returned than retrieved",
 		    __memp_fn(dbmfp));
 	else
 		--dbmfp->pinref;
-	UNLOCKHANDLE(dbmp, dbmfp->mutexp);
 
 	/*
 	 * If we're mapping the file, there's nothing to do.  Because we can
@@ -74,14 +76,14 @@ memp_fput(dbmfp, pgaddr, flags)
 	 * region.
 	 */
 	if (dbmfp->addr != NULL && pgaddr >= dbmfp->addr &&
-	    (u_int8_t *)pgaddr <= (u_int8_t *)dbmfp->addr + dbmfp->len)
+	    (u_int8_t *)pgaddr <= (u_int8_t *)dbmfp->addr + dbmfp->len) {
+		UNLOCKREGION(dbmp);
 		return (0);
+	}
 
 	/* Convert the page address to a buffer header. */
 	bhp = (BH *)((u_int8_t *)pgaddr - SSZA(BH, buf));
 
-	LOCKREGION(dbmp);
-
 	/* Set/clear the page bits. */
 	if (LF_ISSET(DB_MPOOL_CLEAN) && F_ISSET(bhp, BH_DIRTY)) {
 		++mp->stat.st_page_clean;
diff --git a/db2/mp/mp_fset.c b/db2/mp/mp_fset.c
index 3b352aa..1940d3b 100644
--- a/db2/mp/mp_fset.c
+++ b/db2/mp/mp_fset.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_fset.c	10.15 (Sleepycat) 4/26/98";
+static const char sccsid[] = "@(#)mp_fset.c	10.16 (Sleepycat) 9/27/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -40,6 +40,8 @@ memp_fset(dbmfp, pgaddr, flags)
 	dbmp = dbmfp->dbmp;
 	mp = dbmp->mp;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/* Validate arguments. */
 	if (flags == 0)
 		return (__db_ferr(dbmp->dbenv, "memp_fset", 1));
diff --git a/db2/mp/mp_open.c b/db2/mp/mp_open.c
index fc985bc..4c90fc4 100644
--- a/db2/mp/mp_open.c
+++ b/db2/mp/mp_open.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_open.c	10.23 (Sleepycat) 5/3/98";
+static const char sccsid[] = "@(#)mp_open.c	10.27 (Sleepycat) 10/1/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -52,8 +52,8 @@ memp_open(path, flags, mode, dbenv, retp)
 	cachesize = dbenv == NULL ? 0 : dbenv->mp_size;
 
 	/* Create and initialize the DB_MPOOL structure. */
-	if ((dbmp = (DB_MPOOL *)__db_calloc(1, sizeof(DB_MPOOL))) == NULL)
-		return (ENOMEM);
+	if ((ret = __os_calloc(1, sizeof(DB_MPOOL), &dbmp)) != 0)
+		return (ret);
 	LIST_INIT(&dbmp->dbregq);
 	TAILQ_INIT(&dbmp->dbmfq);
 
@@ -83,7 +83,7 @@ memp_open(path, flags, mode, dbenv, retp)
 	if (LF_ISSET(DB_THREAD)) {
 		F_SET(dbmp, MP_LOCKHANDLE | MP_LOCKREGION);
 		LOCKREGION(dbmp);
-		ret = __memp_ralloc(dbmp,
+		ret = __memp_alloc(dbmp,
 		    sizeof(db_mutex_t), NULL, &dbmp->mutexp);
 		UNLOCKREGION(dbmp);
 		if (ret != 0) {
@@ -97,7 +97,7 @@ memp_open(path, flags, mode, dbenv, retp)
 	return (0);
 
 err:	if (dbmp != NULL)
-		FREE(dbmp, sizeof(DB_MPOOL));
+		__os_free(dbmp, sizeof(DB_MPOOL));
 	return (ret);
 }
 
@@ -115,10 +115,12 @@ memp_close(dbmp)
 
 	ret = 0;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/* Discard DB_MPREGs. */
 	while ((mpreg = LIST_FIRST(&dbmp->dbregq)) != NULL) {
 		LIST_REMOVE(mpreg, q);
-		FREE(mpreg, sizeof(DB_MPREG));
+		__os_free(mpreg, sizeof(DB_MPREG));
 	}
 
 	/* Discard DB_MPOOLFILEs. */
@@ -138,13 +140,27 @@ memp_close(dbmp)
 		ret = t_ret;
 
 	if (dbmp->reginfo.path != NULL)
-		FREES(dbmp->reginfo.path);
-	FREE(dbmp, sizeof(DB_MPOOL));
+		__os_freestr(dbmp->reginfo.path);
+	__os_free(dbmp, sizeof(DB_MPOOL));
 
 	return (ret);
 }
 
 /*
+ * __memp_panic --
+ *	Panic a memory pool.
+ *
+ * PUBLIC: void __memp_panic __P((DB_ENV *));
+ */
+void
+__memp_panic(dbenv)
+	DB_ENV *dbenv;
+{
+	if (dbenv->mp_info != NULL)
+		dbenv->mp_info->mp->rlayout.panic = 1;
+}
+
+/*
  * memp_unlink --
  *	Exit a memory pool.
  */
@@ -160,12 +176,12 @@ memp_unlink(path, force, dbenv)
 	memset(&reginfo, 0, sizeof(reginfo));
 	reginfo.dbenv = dbenv;
 	reginfo.appname = DB_APP_NONE;
-	if (path != NULL && (reginfo.path = __db_strdup(path)) == NULL)
-		return (ENOMEM);
+	if (path != NULL && (ret = __os_strdup(path, &reginfo.path)) != 0)
+		return (ret);
 	reginfo.file = DB_DEFAULT_MPOOL_FILE;
 	ret = __db_runlink(&reginfo, force);
 	if (reginfo.path != NULL)
-		FREES(reginfo.path);
+		__os_freestr(reginfo.path);
 	return (ret);
 }
 
@@ -181,9 +197,12 @@ memp_register(dbmp, ftype, pgin, pgout)
 	int (*pgout) __P((db_pgno_t, void *, DBT *));
 {
 	DB_MPREG *mpr;
+	int ret;
+
+	MP_PANIC_CHECK(dbmp);
 
-	if ((mpr = (DB_MPREG *)__db_malloc(sizeof(DB_MPREG))) == NULL)
-		return (ENOMEM);
+	if ((ret = __os_malloc(sizeof(DB_MPREG), NULL, &mpr)) != 0)
+		return (ret);
 
 	mpr->ftype = ftype;
 	mpr->pgin = pgin;
diff --git a/db2/mp/mp_pr.c b/db2/mp/mp_pr.c
index e83e0f4..84c782e 100644
--- a/db2/mp/mp_pr.c
+++ b/db2/mp/mp_pr.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_pr.c	10.26 (Sleepycat) 5/23/98";
+static const char sccsid[] = "@(#)mp_pr.c	10.30 (Sleepycat) 10/1/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -44,16 +44,17 @@ memp_stat(dbmp, gspp, fspp, db_malloc)
 	DB_MPOOL_FSTAT **tfsp;
 	MPOOLFILE *mfp;
 	size_t len, nlen;
+	int ret;
 	char *name;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/* Allocate space for the global statistics. */
 	if (gspp != NULL) {
 		*gspp = NULL;
 
-		if ((*gspp = db_malloc == NULL ?
-		    (DB_MPOOL_STAT *)__db_malloc(sizeof(**gspp)) :
-		    (DB_MPOOL_STAT *)db_malloc(sizeof(**gspp))) == NULL)
-			return (ENOMEM);
+		if ((ret = __os_malloc(sizeof(**gspp), db_malloc, gspp)) != 0)
+			return (ret);
 
 		LOCKREGION(dbmp);
 
@@ -89,10 +90,8 @@ memp_stat(dbmp, gspp, fspp, db_malloc)
 
 		/* Allocate space for the pointers. */
 		len = (len + 1) * sizeof(DB_MPOOL_FSTAT *);
-		if ((*fspp = db_malloc == NULL ?
-		    (DB_MPOOL_FSTAT **)__db_malloc(len) :
-		    (DB_MPOOL_FSTAT **)db_malloc(len)) == NULL)
-			return (ENOMEM);
+		if ((ret = __os_malloc(len, db_malloc, fspp)) != 0)
+			return (ret);
 
 		LOCKREGION(dbmp);
 
@@ -104,10 +103,8 @@ memp_stat(dbmp, gspp, fspp, db_malloc)
 			name = __memp_fns(dbmp, mfp);
 			nlen = strlen(name);
 			len = sizeof(DB_MPOOL_FSTAT) + nlen + 1;
-			if ((*tfsp = db_malloc == NULL ?
-			    (DB_MPOOL_FSTAT *)__db_malloc(len) :
-			    (DB_MPOOL_FSTAT *)db_malloc(len)) == NULL)
-				return (ENOMEM);
+			if ((ret = __os_malloc(len, db_malloc, tfsp)) != 0)
+				return (ret);
 			**tfsp = mfp->stat;
 			(*tfsp)->file_name = (char *)
 			    (u_int8_t *)*tfsp + sizeof(DB_MPOOL_FSTAT);
@@ -212,8 +209,9 @@ __memp_dump_region(dbmp, area, fp)
 	cnt = 0;
 	for (mfp = SH_TAILQ_FIRST(&dbmp->mp->mpfq, __mpoolfile);
 	    mfp != NULL; mfp = SH_TAILQ_NEXT(mfp, q, __mpoolfile), ++cnt) {
-		(void)fprintf(fp, "file #%d: %s: %lu references: %s\n",
+		(void)fprintf(fp, "file #%d: %s: refs %lu, type %ld, %s\n",
 		    cnt + 1, __memp_fns(dbmp, mfp), (u_long)mfp->ref,
+		    (long)mfp->ftype,
 		    F_ISSET(mfp, MP_CAN_MMAP) ? "mmap" : "read/write");
 		    if (cnt < FMAP_ENTRIES)
 			fmap[cnt] = R_OFFSET(dbmp, mfp);
diff --git a/db2/mp/mp_region.c b/db2/mp/mp_region.c
index b8a7228..b9c92f2 100644
--- a/db2/mp/mp_region.c
+++ b/db2/mp/mp_region.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_region.c	10.30 (Sleepycat) 5/31/98";
+static const char sccsid[] = "@(#)mp_region.c	10.35 (Sleepycat) 12/11/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -24,13 +24,33 @@ static const char sccsid[] = "@(#)mp_region.c	10.30 (Sleepycat) 5/31/98";
 #include "common_ext.h"
 
 /*
- * __memp_ralloc --
+ * __memp_reg_alloc --
+ *	Allocate some space in the mpool region, with locking.
+ *
+ * PUBLIC: int __memp_reg_alloc __P((DB_MPOOL *, size_t, size_t *, void *));
+ */
+int
+__memp_reg_alloc(dbmp, len, offsetp, retp)
+	DB_MPOOL *dbmp;
+	size_t len, *offsetp;
+	void *retp;
+{
+	int ret;
+
+	LOCKREGION(dbmp);
+	ret = __memp_alloc(dbmp, len, offsetp, retp);
+	UNLOCKREGION(dbmp);
+	return (ret);
+}
+
+/*
+ * __memp_alloc --
  *	Allocate some space in the mpool region.
  *
- * PUBLIC: int __memp_ralloc __P((DB_MPOOL *, size_t, size_t *, void *));
+ * PUBLIC: int __memp_alloc __P((DB_MPOOL *, size_t, size_t *, void *));
  */
 int
-__memp_ralloc(dbmp, len, offsetp, retp)
+__memp_alloc(dbmp, len, offsetp, retp)
 	DB_MPOOL *dbmp;
 	size_t len, *offsetp;
 	void *retp;
@@ -52,7 +72,9 @@ alloc:	if ((ret = __db_shalloc(dbmp->addr, len, MUTEX_ALIGNMENT, &p)) == 0) {
 		return (0);
 	}
 	if (nomore) {
-		__db_err(dbmp->dbenv, "%s", strerror(ret));
+		__db_err(dbmp->dbenv,
+	    "Unable to allocate %lu bytes from mpool shared region: %s\n",
+		    (u_long)len, strerror(ret));
 		return (ret);
 	}
 
@@ -91,7 +113,7 @@ alloc:	if ((ret = __db_shalloc(dbmp->addr, len, MUTEX_ALIGNMENT, &p)) == 0) {
 	}
 
 retry:	/* Find a buffer we can flush; pure LRU. */
-	total = 0;
+	restart = total = 0;
 	for (bhp =
 	    SH_TAILQ_FIRST(&mp->bhq, __bh); bhp != NULL; bhp = nbhp) {
 		nbhp = SH_TAILQ_NEXT(bhp, q, __bh);
@@ -222,8 +244,8 @@ __memp_ropen(dbmp, path, cachesize, mode, is_private, flags)
 	if (path == NULL)
 		dbmp->reginfo.path = NULL;
 	else
-		if ((dbmp->reginfo.path = __db_strdup(path)) == NULL)
-			return (ENOMEM);
+		if ((ret = __os_strdup(path, &dbmp->reginfo.path)) != 0)
+			return (ret);
 	dbmp->reginfo.file = DB_DEFAULT_MPOOL_FILE;
 	dbmp->reginfo.mode = mode;
 	dbmp->reginfo.size = rlen;
@@ -244,7 +266,7 @@ __memp_ropen(dbmp, path, cachesize, mode, is_private, flags)
 
 	if ((ret = __db_rattach(&dbmp->reginfo)) != 0) {
 		if (dbmp->reginfo.path != NULL)
-			FREES(dbmp->reginfo.path);
+			__os_freestr(dbmp->reginfo.path);
 		return (ret);
 	}
 
@@ -303,6 +325,6 @@ err:	UNLOCKREGION(dbmp);
 		(void)memp_unlink(path, 1, dbmp->dbenv);
 
 	if (dbmp->reginfo.path != NULL)
-		FREES(dbmp->reginfo.path);
+		__os_freestr(dbmp->reginfo.path);
 	return (ret);
 }
diff --git a/db2/mp/mp_sync.c b/db2/mp/mp_sync.c
index 33218ee..5353485 100644
--- a/db2/mp/mp_sync.c
+++ b/db2/mp/mp_sync.c
@@ -7,7 +7,7 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_sync.c	10.25 (Sleepycat) 4/26/98";
+static const char sccsid[] = "@(#)mp_sync.c	10.31 (Sleepycat) 12/11/98";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
@@ -39,9 +39,12 @@ memp_sync(dbmp, lsnp)
 	DB_ENV *dbenv;
 	MPOOL *mp;
 	MPOOLFILE *mfp;
-	int ar_cnt, cnt, nalloc, next, ret, wrote;
+	int ar_cnt, nalloc, next, maxpin, ret, wrote;
+
+	MP_PANIC_CHECK(dbmp);
 
 	dbenv = dbmp->dbenv;
+	mp = dbmp->mp;
 
 	if (dbenv->lg_info == NULL) {
 		__db_err(dbenv, "memp_sync: requires logging");
@@ -49,16 +52,19 @@ memp_sync(dbmp, lsnp)
 	}
 
 	/*
-	 * We try and write the buffers in page order so that the underlying
-	 * filesystem doesn't have to seek and can write contiguous blocks,
-	 * plus, we don't want to hold the region lock while we write the
-	 * buffers.  Get memory to hold the buffer pointers.  Get a good-size
-	 * block, too, because we realloc while holding the region lock if we
-	 * run out.
+	 * We try and write the buffers in page order: it should reduce seeks
+	 * by the underlying filesystem and possibly reduce the actual number
+	 * of writes.  We don't want to hold the region lock while we write
+	 * the buffers, so only hold it lock while we create a list.  Get a
+	 * good-size block of memory to hold buffer pointers, we don't want
+	 * to run out.
 	 */
-	if ((bharray =
-	    (BH **)__db_malloc((nalloc = 1024) * sizeof(BH *))) == NULL)
-		return (ENOMEM);
+	LOCKREGION(dbmp);
+	nalloc = mp->stat.st_page_dirty + mp->stat.st_page_dirty / 2 + 10;
+	UNLOCKREGION(dbmp);
+
+	if ((ret = __os_malloc(nalloc * sizeof(BH *), NULL, &bharray)) != 0)
+		return (ret);
 
 	LOCKREGION(dbmp);
 
@@ -70,7 +76,6 @@ memp_sync(dbmp, lsnp)
 	 * we've already handled or are currently handling, then we return a
 	 * result based on the count for the larger LSN.
 	 */
-	mp = dbmp->mp;
 	if (!F_ISSET(mp, MP_LSN_RETRY) && log_compare(lsnp, &mp->lsn) <= 0) {
 		if (mp->lsn_cnt == 0) {
 			*lsnp = mp->lsn;
@@ -114,10 +119,15 @@ memp_sync(dbmp, lsnp)
 	 * finish.  Since the application may have restarted the sync, clear
 	 * any BH_WRITE flags that appear to be left over from previous calls.
 	 *
+	 * We don't want to pin down the entire buffer cache, otherwise we'll
+	 * starve threads needing new pages.  Don't pin down more than 80% of
+	 * the cache.
+	 *
 	 * Keep a count of the total number of buffers we need to write in
 	 * MPOOL->lsn_cnt, and for each file, in MPOOLFILE->lsn_count.
 	 */
 	ar_cnt = 0;
+	maxpin = ((mp->stat.st_page_dirty + mp->stat.st_page_clean) * 8) / 10;
 	for (bhp = SH_TAILQ_FIRST(&mp->bhq, __bh);
 	    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, q, __bh))
 		if (F_ISSET(bhp, BH_DIRTY) || bhp->ref != 0) {
@@ -130,19 +140,27 @@ memp_sync(dbmp, lsnp)
 
 			/*
 			 * If the buffer isn't in use, we should be able to
-			 * write it immediately, so save a reference to it.
+			 * write it immediately, so increment the reference
+			 * count to lock it and its contents down, and then
+			 * save a reference to it.
+			 *
+			 * If we've run out space to store buffer references,
+			 * we're screwed.  We don't want to realloc the array
+			 * while holding a region lock, so we set the flag to
+			 * force the checkpoint to be done again, from scratch,
+			 * later.
+			 *
+			 * If we've pinned down too much of the cache stop, and
+			 * set a flag to force the checkpoint to be tried again
+			 * later.
 			 */
 			if (bhp->ref == 0) {
-				if (ar_cnt == nalloc) {
-					nalloc *= 2;
-					if ((bharray =
-					    (BH **)__db_realloc(bharray,
-					    nalloc * sizeof(BH *))) == NULL) {
-						ret = ENOMEM;
-						goto err;
-					}
+				++bhp->ref;
+				bharray[ar_cnt] = bhp;
+				if (++ar_cnt >= nalloc || ar_cnt >= maxpin) {
+					F_SET(mp, MP_LSN_RETRY);
+					break;
 				}
-				bharray[ar_cnt++] = bhp;
 			}
 		} else
 			if (F_ISSET(bhp, BH_WRITE))
@@ -154,10 +172,6 @@ memp_sync(dbmp, lsnp)
 		goto done;
 	}
 
-	/* Lock down the buffers and their contents. */
-	for (cnt = 0; cnt < ar_cnt; ++cnt)
-		++bharray[cnt]->ref;
-
 	UNLOCKREGION(dbmp);
 
 	/* Sort the buffers we're going to write. */
@@ -205,7 +219,8 @@ memp_sync(dbmp, lsnp)
 			goto err;
 		}
 	}
-	ret = mp->lsn_cnt ? DB_INCOMPLETE : 0;
+	ret = mp->lsn_cnt != 0 ||
+	    F_ISSET(mp, MP_LSN_RETRY) ? DB_INCOMPLETE : 0;
 
 done:
 	if (0) {
@@ -224,7 +239,7 @@ err:		/*
 			F_CLR(bhp, BH_WRITE);
 	}
 	UNLOCKREGION(dbmp);
-	__db_free(bharray);
+	__os_free(bharray, nalloc * sizeof(BH *));
 	return (ret);
 }
 
@@ -241,6 +256,8 @@ memp_fsync(dbmfp)
 
 	dbmp = dbmfp->dbmp;
 
+	MP_PANIC_CHECK(dbmp);
+
 	/*
 	 * If this handle doesn't have a file descriptor that's open for
 	 * writing, or if the file is a temporary, there's no reason to
@@ -300,25 +317,29 @@ __memp_fsync(dbmfp)
 {
 	BH *bhp, **bharray;
 	DB_MPOOL *dbmp;
+	MPOOL *mp;
 	size_t mf_offset;
-	int ar_cnt, cnt, nalloc, next, pincnt, ret, wrote;
+	int ar_cnt, incomplete, nalloc, next, ret, wrote;
 
 	ret = 0;
 	dbmp = dbmfp->dbmp;
+	mp = dbmp->mp;
 	mf_offset = R_OFFSET(dbmp, dbmfp->mfp);
 
 	/*
-	 * We try and write the buffers in page order so that the underlying
-	 * filesystem doesn't have to seek and can write contiguous blocks,
-	 * plus, we don't want to hold the region lock while we write the
-	 * buffers.  Get memory to hold the buffer pointers.  Get a good-size
-	 * block, too, because we realloc while holding the region lock if we
-	 * run out.
+	 * We try and write the buffers in page order: it should reduce seeks
+	 * by the underlying filesystem and possibly reduce the actual number
+	 * of writes.  We don't want to hold the region lock while we write
+	 * the buffers, so only hold it lock while we create a list.  Get a
+	 * good-size block of memory to hold buffer pointers, we don't want
+	 * to run out.
 	 */
-	nalloc = 1024;
-	if ((bharray =
-	    (BH **)__db_malloc((size_t)nalloc * sizeof(BH *))) == NULL)
-		return (ENOMEM);
+	LOCKREGION(dbmp);
+	nalloc = mp->stat.st_page_dirty + mp->stat.st_page_dirty / 2 + 10;
+	UNLOCKREGION(dbmp);
+
+	if ((ret = __os_malloc(nalloc * sizeof(BH *), NULL, &bharray)) != 0)
+		return (ret);
 
 	LOCKREGION(dbmp);
 
@@ -326,36 +347,37 @@ __memp_fsync(dbmfp)
 	 * Walk the LRU list of buffer headers, and get a list of buffers to
 	 * write for this MPOOLFILE.
 	 */
-	ar_cnt = pincnt = 0;
-	for (bhp = SH_TAILQ_FIRST(&dbmp->mp->bhq, __bh);
+	ar_cnt = incomplete = 0;
+	for (bhp = SH_TAILQ_FIRST(&mp->bhq, __bh);
 	    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, q, __bh)) {
 		if (!F_ISSET(bhp, BH_DIRTY) || bhp->mf_offset != mf_offset)
 			continue;
 		if (bhp->ref != 0 || F_ISSET(bhp, BH_LOCKED)) {
-			++pincnt;
+			incomplete = 1;
 			continue;
 		}
 
-		if (ar_cnt == nalloc) {
-			nalloc *= 2;
-			if ((bharray = (BH **)__db_realloc(bharray,
-			    nalloc * sizeof(BH *))) == NULL) {
-				ret = ENOMEM;
-				goto err;
-			}
-		}
+		++bhp->ref;
+		bharray[ar_cnt] = bhp;
 
-		bharray[ar_cnt++] = bhp;
+		/*
+		 * If we've run out space to store buffer references, we're
+		 * screwed, as we don't want to realloc the array holding a
+		 * region lock.  Set the incomplete flag -- the only way we
+		 * can get here is if the file is active in the buffer cache,
+		 * which is the same thing as finding pinned buffers.
+		 */
+		if (++ar_cnt >= nalloc) {
+			incomplete = 1;
+			break;
+		}
 	}
 
-	/* Lock down the buffers and their contents. */
-	for (cnt = 0; cnt < ar_cnt; ++cnt)
-		++bharray[cnt]->ref;
-
 	UNLOCKREGION(dbmp);
 
 	/* Sort the buffers we're going to write. */
-	qsort(bharray, ar_cnt, sizeof(BH *), __bhcmp);
+	if (ar_cnt != 0)
+		qsort(bharray, ar_cnt, sizeof(BH *), __bhcmp);
 
 	LOCKREGION(dbmp);
 
@@ -365,11 +387,10 @@ __memp_fsync(dbmfp)
 		 * It's possible for a thread to have gotten the buffer since
 		 * we listed it for writing.  If the reference count is still
 		 * 1, we're the only ones using the buffer, go ahead and write.
-		 * If it's >1, then skip the buffer and assume that it will be
-		 * written when it's returned to the cache.
+		 * If it's >1, then skip the buffer.
 		 */
 		if (bharray[next]->ref > 1) {
-			++pincnt;
+			incomplete = 1;
 
 			--bharray[next]->ref;
 			continue;
@@ -387,13 +408,18 @@ __memp_fsync(dbmfp)
 				--bharray[next]->ref;
 			goto err;
 		}
+
+		/*
+		 * If we didn't write the buffer for some reason, don't return
+		 * success.
+		 */
 		if (!wrote)
-			++pincnt;
+			incomplete = 1;
 	}
 
 err:	UNLOCKREGION(dbmp);
 
-	__db_free(bharray);
+	__os_free(bharray, nalloc * sizeof(BH *));
 
 	/*
 	 * Sync the underlying file as the last thing we do, so that the OS
@@ -404,7 +430,7 @@ err:	UNLOCKREGION(dbmp);
 	 * issues.
 	 */
 	if (ret == 0)
-		return (pincnt == 0 ? __db_fsync(dbmfp->fd) : DB_INCOMPLETE);
+		return (incomplete ? DB_INCOMPLETE : __os_fsync(dbmfp->fd));
 	return (ret);
 }
 
@@ -423,6 +449,8 @@ memp_trickle(dbmp, pct, nwrotep)
 	u_long total;
 	int ret, wrote;
 
+	MP_PANIC_CHECK(dbmp);
+
 	mp = dbmp->mp;
 	if (nwrotep != NULL)
 		*nwrotep = 0;
@@ -487,7 +515,7 @@ loop:	total = mp->stat.st_page_clean + mp->stat.st_page_dirty;
 	}
 
 	/* No more buffers to write. */
-	return (0);
+	ret = 0;
 
 err:	UNLOCKREGION(dbmp);
 	return (ret);
@@ -508,6 +536,14 @@ __bhcmp(p1, p2)
 	if (bhp1->mf_offset > bhp2->mf_offset)
 		return (1);
 
-	/* Sort by page in file. */
-	return (bhp1->pgno < bhp2->pgno ? -1 : 1);
+	/*
+	 * !!!
+	 * Defend against badly written quicksort code calling the comparison
+	 * function with two identical pointers (e.g., WATCOM C++ (Power++)).
+	 */
+	if (bhp1->pgno < bhp2->pgno)
+		return (-1);
+	if (bhp1->pgno > bhp2->pgno)
+		return (1);
+	return (0);
 }
-- 
cgit v1.1