Update. cvs/libc-ud-971111

1997-11-11 21:30  Ulrich Drepper  <drepper@cygnus.com>

	* include/sys/stat.h: Define stat, fstat, lstat and *64 variants
	as macros so the the library compiles correctly even without
	optimization.
	* io/fstat.c: Undef fstat.
	* io/fstat64.c: Undef fstat64
	* io/lstat.c: Undef lstat.
	* io/lstat64.c: Undef lstat64
	* io/stat.c: Undef stat.
	* io/stat64.c: Undef stat64
	* io/fts.c: Include <include/sys/stat.h> to get macro definitions.
	* io/ftw.c: Likewise.
	* io/getdirname.c: Likewise.

	* Makefile (install): Run test-installation.pl if possible.

	* db2/Makefile: Update from db-2.3.12.
	* db2/db.h: Likewise.
	* db2/db_int.h: Likewise.
	* db2/btree/bt_cursor.c: Likewise.
	* db2/btree/bt_delete.c: Likewise.
	* db2/btree/bt_open.c: Likewise.
	* db2/btree/bt_put.c: Likewise.
	* db2/btree/bt_rec.c: Likewise.
	* db2/btree/bt_recno.c: Likewise.
	* db2/btree/bt_search.c: Likewise.
	* db2/btree/bt_split.c: Likewise.
	* db2/btree/bt_stat.c: Likewise.
	* db2/btree/btree.src: Likewise.
	* db2/btree/btree_auto.c: Likewise.
	* db2/btree/bt_cursor.c: Likewise.
	* db2/btree/bt_delete.c: Likewise.
	* db2/btree/bt_open.c: Likewise.
	* db2/btree/bt_put.c: Likewise.
	* db2/btree/bt_rec.c: Likewise.
	* db2/btree/bt_recno.c: Likewise.
	* db2/btree/bt_search.c: Likewise.
	* db2/btree/bt_split.c: Likewise.
	* db2/btree/bt_stat.c: Likewise.
	* db2/btree/btree.src: Likewise.
	* db2/btree/btree_auto.c: Likewise.
	* db2/common/db_appinit.c: Likewise.
	* db2/common/db_apprec.c: Likewise.
	* db2/common/db_byteorder.c: Likewise.
	* db2/common/db_region.c: Likewise.
	* db2/db/db.c: Likewise
	* db2/db/db.src: Likewise
	* db2/db/db_auto.c: Likewise
	* db2/db/db_dispatch.c: Likewise
	* db2/db/db_dup.c: Likewise
	* db2/db/db_overflow.c: Likewise
	* db2/db/db_pr.c: Likewise
	* db2/db/db_rec.c: Likewise
	* db2/db/db_ret.c: Likewise
	* db2/db/db_thread.c: Likewise
	* db2/db185/db185.c: Likewise.
	* db2/hash/hash.c: Likewise.
	* db2/hash/hash.src: Likewise.
	* db2/hash/hash_auto.c: Likewise.
	* db2/hash/hash_dup.c: Likewise.
	* db2/hash/hash_page.c: Likewise.
	* db2/hash/hash_rec.c: Likewise.
	* db2/include/btree_auto.h: Likewise.
	* db2/include/btree_ext.h: Likewise.
	* db2/include/clib_ext.h: Likewise.
	* db2/include/common_ext.h: Likewise.
	* db2/include/db.h.src: Likewise.
	* db2/include/db_am.h: Likewise.
	* db2/include/db_auto.h: Likewise.
	* db2/include/db_cxx.h: Likewise.
	* db2/include/db_ext.h: Likewise.
	* db2/include/db_int.h.src: Likewise.
	* db2/include/hash.h: Likewise.
	* db2/include/hash_auto.h: Likewise.
	* db2/include/hash_ext.h: Likewise.
	* db2/include/lock.h: Likewise.
	* db2/include/lock_ext.h: Likewise.
	* db2/include/log.h: Likewise.
	* db2/include/log_ext.h: Likewise.
	* db2/include/mp.h: Likewise.
	* db2/include/mp_ext.h: Likewise.
	* db2/include/mutex_ext.h: Likewise.
	* db2/include/os_ext.h: Likewise.
	* db2/include/os_func.h: Likewise.
	* db2/include/txn.h: Likewise.
	* db2/include/txn_ext.h: Likewise.
	* db2/lock/lock.c: Likewise.
	* db2/lock/lock_deadlock.c: Likewise.
	* db2/log/log.c: Likewise.
	* db2/log/log_archive.c: Likewise.
	* db2/log/log_auto.c: Likewise.
	* db2/log/log_findckp.c: Likewise.
	* db2/log/log_get.c: Likewise.
	* db2/log/log_put.c: Likewise.
	* db2/log/log_rec.c: Likewise.
	* db2/log/log_register.c: Likewise.
	* db2/mp/mp_bh.c: Likewise.
	* db2/mp/mp_fget.c: Likewise.
	* db2/mp/mp_fopen.c: Likewise.
	* db2/mp/mp_fput.c: Likewise.
	* db2/mp/mp_fset.c: Likewise.
	* db2/mp/mp_open.c: Likewise.
	* db2/mp/mp_pr.c: Likewise.
	* db2/mp/mp_region.c: Likewise.
	* db2/mp/mp_sync.c: Likewise.
	* db2/mutex/mutex.c: Likewise.
	* db2/os/os_abs.c: Likewise.
	* db2/os/os_dir.c: Likewise.
	* db2/os/os_fid.c: Likewise.
	* db2/os/os_fsync.c: Likewise.
	* db2/os/os_func.c: Likewise.
	* db2/os/os_map.c: Likewise.
	* db2/os/os_oflags.c: Likewise.
	* db2/os/os_open.c: Likewise.
	* db2/os/os_rpath.c: Likewise.
	* db2/os/os_rw.c: Likewise.
	* db2/os/os_seek.c: Likewise.
	* db2/os/os_sleep.c: Likewise.
	* db2/os/os_stat.c: Likewise.
	* db2/os/os_unlink.c: Likewise.
	* db2/progs/db_deadlock/db_deadlock.c: Likewise.
	* db2/progs/db_dump/db_dump.c: Likewise.
	* db2/progs/db_load/db_load.c: Likewise.
	* db2/progs/db_recover/db_recover.c: Likewise.
	* db2/progs/db_stat/db_stat.c: Likewise.
	* db2/txn/txn.c: Likewise.
	* db2/txn/txn_auto.c: Likewise.
	* db2/txn/txn_rec.c: Likewise.
	* db2/os/db_os_abs.c: Removed.
	* db2/os/db_os_dir.c: Removed.
	* db2/os/db_os_fid.c: Removed.
	* db2/os/db_os_lseek.c: Removed.
	* db2/os/db_os_mmap.c: Removed.
	* db2/os/db_os_open.c: Removed.
	* db2/os/db_os_rw.c: Removed.
	* db2/os/db_os_sleep.c: Removed.
	* db2/os/db_os_stat.c: Removed.
	* db2/os/db_os_unlink.c: Removed.

	* libio/stdio.h (fopen): Add __restrict to parameters.

	* manual/process.texi (system): Describe behaviour for NULL argument.

	* stdio-common/printf-parse.h: Parse hh modifier.
	* stdio-common/vfprintf.c: Handle hh modifier.
	* stdio-common/vfscanf.c: Likewise.
	* manual/stdio.texi: Describe hh modifier for scanf/printf.

	* math/complex.h: Don't define _Imaginary_I, but instead _Complex_I.
	gcc does no yet know the `imaginary' keyword.

	* math/test-math.c: Add little test for know gcc bug.

	* math/tgmath.h: Make complex versions of log10() only available
	if __USE_GNU.

	* stdlib/test-canon.c: Fix typo.

	* sysdeps/generic/setenv.c: Avoid compilation warnings.
	Reported by Jim Meyering.

	* sysdeps/generic/bits/errno.h: EILSEQ is an ISO C error number.
	* sysdeps/mach/hurd/bits/errno.h: Likewise.
	* sysdeps/standalone/bits/errno.h: Likewise.
	* sysdeps/unix/sysv/linux/bits/errno.h: Likewise.

	* sysdeps/i386/i586/memcpy.S: New file.
	* sysdeps/i386/i586/mempcpy.S: New file.

	* sysdeps/i386/i586/memset.S: Fix typo.

	* sysdeps/posix/getcwd.c: Define HAVE_MEMPCPY for _LIBC.  Add casts.

	* sysdeps/posix/system.c: Add comment to explain code.

	* sysdeps/wordsize-32/inttypes.h: Include <stddef.h> for wchar_t.
	Define PTRDIFF_{MIN,MAX}, SIG_ATOMIC_{MIN,MAX}, SIZE_MAX,
	WCHAR_{MIN,MAX}, WINT_{MIN,MAX}.
	Define wcstoimax, wcstoumax.
	* sysdeps/wordsize-64/inttypes.h: Likewise.

	* wcsmbs/wchar.h: Define WCHAR_{MIN,MAX} if not already defined.
	Declare __wcsto{l,ul,ll,ull}_internal only if not already done.

	* time/Makefile (routines): Add strfxtime.
	* time/strftime.c: Implement %F and %f format.
	* time/strfxtime.c: New file.
	* time/time.h: Define new types and symbols from ISO C 9X.

	* time/mktime.c: Little comment correction.

1997-11-10  Andreas Jaeger  <aj@arthur.rhein-neckar.de>

	* sysdeps/libm-ieee754/s_sincosl.c: Fix typo.

	* sysdeps/libm-ieee754/s_tanl.c: Fix typo.

	* sysdeps/libm-ieee754/s_floorl.c: Correct typos.

	* sysdeps/libm-ieee754/e_remainderl.c: Replace
	EXTRACT_LDOUBLE_WORDS by GET_LDOUBLE_WORDS.

	* sysdeps/libm-ieee754/e_atan2l.c: Replace EXTRACT_LDOUBLE_WORDS
	by GET_LDOUBLE_WORDS.

	* sysdeps/libm-ieee754/s_scalbnl.c: Replace ";" by "," for correct
	variable declaration.
	* sysdeps/libm-ieee754/s_scalblnl.c: Likewise.

	* sysdeps/libm-ieee754/s_lrint.c (__lrint): Correct function.

	* math/libm-test.c (sqrt_test): Add test for sqrt (0.25).
	(asin_test): Add more test.

1997-11-10 23:34  Ulrich Drepper  <drepper@cygnus.com>

	* sysdeps/libm-ieee754/e_asin.c: Add braces to make code clearer
	and to not confuse the poor compiler.
	* sysdeps/libm-ieee754/e_asinf.c: Likewise.
	Reported by vertex@cagent.com.

1997-11-09  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* elf/dl-lookup.c (do_lookup): Don't accept the base version if we
	require a specific one.

	* libio/oldfreopen.c: Bind old symbols to version GLIBC_2.0.
	* libio/oldiofopen.c: Likewise.
	* libio/oldstdfiles.c: Likewise.
	* libc.map: Export them.

1997-11-10 07:40  H.J. Lu  <hjl@gnu.ai.mit.edu>

	* stdlib/exit.c (exit): Handle recursive calls to exit ().

1997-11-09  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* sysdeps/m68k/fpu/s_llrint.c: Fixed to take double argument
	instead of long double.
	* sysdeps/m68k/fpu/s_llrintf.c: New file.
	* sysdeps/m68k/fpu/s_llrintl.c: New file.

	* sysdeps/libm-ieee754/s_llrint.c: Make compilable and fix
	overflow condition.
	* sysdeps/libm-ieee754/s_llrintf.c: Fix overflow condition.
	* sysdeps/libm-ieee754/s_llrintl.c: Likewise.
	* sysdeps/libm-ieee754/s_llround.c: Likewise.
	* sysdeps/libm-ieee754/s_llroundf.c: Likewise.
	* sysdeps/libm-ieee754/s_llroundl.c: Likewise.
	* sysdeps/libm-ieee754/s_lrint.c: Likewise.
	* sysdeps/libm-ieee754/s_lrintf.c: Likewise.
	* sysdeps/libm-ieee754/s_lrintl.c: Likewise.
	* sysdeps/libm-ieee754/s_lround.c: Likewise.
	* sysdeps/libm-ieee754/s_lroundf.c: Likewise.
	* sysdeps/libm-ieee754/s_lroundl.c: Likewise.

	* math/libm-test.c: Test all three variants of lrint and llrint.
	Fix typos in lround and llround tests.  Add tests for boundary
	cases for lrint and llround.

1997-11-08  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* manual/arith.texi: Misc doc fixes.
	* manual/ctype.texi: Likewise.
	* manual/pattern.texi: Likewise.
	* manual/terminal.texi: Likewise.

1997-11-08  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* sysdeps/unix/sysv/linux/xstatconv.c: Use struct assignment
	instead of memcpy to let the compiler use whatever it regards as
	optimal.
	* sysdeps/unix/sysv/linux/alpha/xstatconv.c: Likewise.

1997-11-08  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* sysdeps/unix/sysv/linux/Makefile (sysdep_headers)
	[$(subdir)=misc]: Add sys/prctl.h.
	* sysdeps/unix/sysv/linux/Dist: Distribute it.

1997-11-08  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* io/ftwtest-sh: Don't use the unknown which command, instead try
	pwd as /bin/pwd and /usr/bin/pwd.

1997-11-08  Andreas Schwab  <schwab@issan.informatik.uni-dortmund.de>

	* manual/maint.texi (Tools for Installation): Don't recommend
	broken version 3.76.1 of make.
	(Porting): Fix wording.

1997-11-06 06:13  H.J. Lu  <hjl@gnu.ai.mit.edu>

	* config.make.in (build-pic-default): New, defined with
	pic_default.

	* configure.in (pic_default): New, set to yes if PIC is
	default.

	* Makeconfig (CPPFLAGS-.o, CPPFLAGS-.op, CPPFLAGS-.og,
	CPPFLAGS-.ob): Add -DPIC if $(build-pic-default) is yes.

1997-11-09 18:15  Ulrich Drepper  <drepper@cygnus.com>

	* Makerules (libc.so): Fix typo.

	* csu/Makefile (CFLAGS-initfini.s): Correctly fix moving function
	definition.  Patch by Zack Weinberg <zack@rabi.phys.columbia.edu>.

	* stdlib/strtod.c: Handle numbers like 0.0e10000 correctly which
	produce ±0.0.  Reported by Joe Keane <jgk@jgk.org>.

	* sysdeps/libm-ieee754/s_ceill.c: Fix typos.
	* sysdeps/libm-ieee754/s_llrint.c: Correct code, it never worked.

1997-11-06 07:00  H.J. Lu  <hjl@gnu.ai.mit.edu>

	* sysdeps/unix/sysv/i386/i686/time.S: Removed.

1997-11-08 14:07  Thorsten Kukuk  <kukuk@vt.uni-paderborn.de>

	* nis/libnsl.map: Add __do_niscall2 for nis_cachemgr.

	* nis/nis_call.c: Set UDP resend timeout correct.

	* nis/nss_compat/compat-grp.c: Rewritten to make it faster.
	* nis/nss_compat/compat-pwd.c: Likewise.
	* nis/nss_compat/compat-spwd.c: Likewise.
	* nis/ypclnt.c: Fix UDP resend timeout, fix yp_bind/do_ypcall
	interaction.

	* inet/protocols/routed.h: Include sys/socket.h.
	* inet/protocols/talkd.h: Likewise.
	* inet/protocols/timed.h: Include rpc/types.h.
	* sunrpc/rpc/pmap_clnt.h: Include rpc/clnt.h.

1997-11-06 01:39  Ulrich Drepper  <drepper@cygnus.com>

	* Makerules (libc.so): Add missing closing brace.

1997-11-05  Brendan Kehoe  <brendan@lisa.cygnus.com>

	* libio.h (__P): Name its arg `p' instead of `params'.
	This was added solely to work around problems with
	the definition of __P in the Solaris math.h header.

1 files changed, 320 insertions, 73 deletions

diff --git a/db2/mp/mp_sync.c b/db2/mp/mp_sync.c
index 65b2a18267..2f042df9e1 100644
--- a/db2/mp/mp_sync.c
+++ b/db2/mp/mp_sync.c
@@ -7,13 +7,14 @@
 #include "config.h"
 
 #ifndef lint
-static const char sccsid[] = "@(#)mp_sync.c	10.9 (Sleepycat) 8/29/97";
+static const char sccsid[] = "@(#)mp_sync.c	10.15 (Sleepycat) 11/1/97";
 #endif /* not lint */
 
 #ifndef NO_SYSTEM_INCLUDES
 #include <sys/types.h>
 
 #include <errno.h>
+#include <stdlib.h>
 #include <string.h>
 #endif
 
@@ -23,6 +24,8 @@ static const char sccsid[] = "@(#)mp_sync.c	10.9 (Sleepycat) 8/29/97";
 #include "mp.h"
 #include "common_ext.h"
 
+static int __bhcmp __P((const void *, const void *));
+
 /*
  * memp_sync --
  *	Mpool sync function.
@@ -32,27 +35,40 @@ memp_sync(dbmp, lsnp)
 	DB_MPOOL *dbmp;
 	DB_LSN *lsnp;
 {
-	BH *bhp;
+	BH *bhp, **bharray;
 	DB_ENV *dbenv;
 	MPOOL *mp;
 	MPOOLFILE *mfp;
-	int can_write, wrote, lsn_cnt, restart, ret;
+	int ar_cnt, cnt, nalloc, next, notused, ret, wrote;
 
 	dbenv = dbmp->dbenv;
 
-	if (dbmp->dbenv->lg_info == NULL) {
-		__db_err(dbenv, "memp_sync requires logging");
+	if (dbenv->lg_info == NULL) {
+		__db_err(dbenv, "memp_sync: requires logging");
 		return (EINVAL);
 	}
 
+	/*
+	 * 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.
+	 */
+	if ((bharray =
+	    (BH **)__db_malloc((nalloc = 1024) * sizeof(BH *))) == NULL)
+		return (ENOMEM);
+
 	LOCKREGION(dbmp);
 
 	/*
-	 * If the application is asking about a previous call, and we haven't
-	 * found any buffers that the application holding the pin couldn't
-	 * write, return yes or no based on the current count.  Note, if the
-	 * application is asking about a LSN *smaller* than one we've already
-	 * handled, then we return based on the count for that LSN.
+	 * If the application is asking about a previous call to memp_sync(),
+	 * and we haven't found any buffers that the application holding the
+	 * pin couldn't write, return yes or no based on the current count.
+	 * Note, if the application is asking about a LSN *smaller* than one
+	 * 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) {
@@ -61,9 +77,7 @@ memp_sync(dbmp, lsnp)
 			ret = 0;
 		} else
 			ret = DB_INCOMPLETE;
-
-		UNLOCKREGION(dbmp);
-		return (ret);
+		goto done;
 	}
 
 	/* Else, it's a new checkpoint. */
@@ -74,7 +88,7 @@ memp_sync(dbmp, lsnp)
 	 * for which we were already doing a checkpoint.  (BTW, I don't expect
 	 * to see multiple LSN's from the same or multiple processes, but You
 	 * Just Never Know.  Responding as if they all called with the largest
-	 * of the LSNs specified makes everything work.
+	 * of the LSNs specified makes everything work.)
 	 *
 	 * We don't currently use the LSN we save.  We could potentially save
 	 * the last-written LSN in each buffer header and use it to determine
@@ -93,64 +107,127 @@ memp_sync(dbmp, lsnp)
 
 	/*
 	 * Walk the list of buffers and mark all dirty buffers to be written
-	 * and all pinned buffers to be potentially written.  We do this in
-	 * single fell swoop while holding the region locked so that processes
-	 * can't make new buffers dirty, causing us to never finish.  Since
-	 * the application may have restarted the sync, clear any BH_WRITE
-	 * flags that appear to be left over.
+	 * and all pinned buffers to be potentially written (we can't know if
+	 * we'll need to write them until the holding process returns them to
+	 * the cache).  We do this in one pass while holding the region locked
+	 * so that processes can't make new buffers dirty, causing us to never
+	 * finish.  Since the application may have restarted the sync, clear
+	 * any BH_WRITE flags that appear to be left over from previous calls.
+	 *
+	 * 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.
 	 */
-	can_write = lsn_cnt = 0;
-	for (lsn_cnt = 0, bhp = SH_TAILQ_FIRST(&mp->bhq, __bh);
+	ar_cnt = 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->ref != 0) {
 			F_SET(bhp, BH_WRITE);
 
-			if (bhp->ref == 0)
-				can_write = 1;
+			++mp->lsn_cnt;
 
-			mfp = ADDR(dbmp, bhp->mf_offset);
+			mfp = R_ADDR(dbmp, bhp->mf_offset);
 			++mfp->lsn_cnt;
 
-			++lsn_cnt;
+			/*
+			 * If the buffer isn't in use, we should be able to
+			 * write it immediately, so save a reference to it.
+			 */
+			if (bhp->ref == 0) {
+				if (ar_cnt == nalloc) {
+					nalloc *= 2;
+					if ((bharray =
+					    (BH **)__db_realloc(bharray,
+					    nalloc * sizeof(BH *))) == NULL) {
+						ret = ENOMEM;
+						goto err;
+					}
+				}
+				bharray[ar_cnt++] = bhp;
+			}
 		} else
 			F_CLR(bhp, BH_WRITE);
 
-	mp->lsn_cnt = lsn_cnt;
-
-	/* If there no buffers we can write, we're done. */
-	if (!can_write) {
-		UNLOCKREGION(dbmp);
-		return (mp->lsn_cnt ? DB_INCOMPLETE : 0);
+	/* If there no buffers we can write immediately, we're done. */
+	if (ar_cnt == 0) {
+		ret = mp->lsn_cnt ? DB_INCOMPLETE : 0;
+		goto done;
 	}
 
-	/*
-	 * Write any buffers that we can.  Restart the walk after each write,
-	 * __memp_pgwrite() discards and reacquires the region lock during I/O.
-	 */
-retry:	for (bhp = SH_TAILQ_FIRST(&mp->bhq, __bh);
-	    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, q, __bh)) {
-		/* Ignore pinned or locked buffers. */
-		if (!F_ISSET(bhp, BH_WRITE) ||
-		    bhp->ref != 0 || F_ISSET(bhp, BH_LOCKED))
-			continue;
+	/* Lock down the buffers and their contents. */
+	for (cnt = 0; cnt < ar_cnt; ++cnt)
+		++bharray[cnt]->ref;
 
-		mfp = ADDR(dbmp, bhp->mf_offset);
-		if ((ret =
-		    __memp_bhwrite(dbmp, mfp, bhp, &restart, &wrote)) != 0)
-			goto err;
-		if (wrote) {
-			if (restart)
-				goto retry;
+	UNLOCKREGION(dbmp);
+
+	/* Sort the buffers we're going to write. */
+	qsort(bharray, ar_cnt, sizeof(BH *), __bhcmp);
+
+	LOCKREGION(dbmp);
+
+	/* Walk the array, writing buffers. */
+	for (next = 0; next < ar_cnt; ++next) {
+		/*
+		 * 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 (bharray[next]->ref > 1) {
+			--bharray[next]->ref;
 			continue;
 		}
-		__db_err(dbenv, "%s: unable to flush page: %lu",
-		    ADDR(dbmp, mfp->path_off), (u_long)bhp->pgno);
-		ret = EPERM;
-		goto err;
+
+		/* Write the buffer. */
+		mfp = R_ADDR(dbmp, bharray[next]->mf_offset);
+		ret =
+		    __memp_bhwrite(dbmp, mfp, bharray[next], &notused, &wrote);
+
+		/* Release the buffer. */
+		--bharray[next]->ref;
+
+		/* If there's an error, release the rest of the buffers. */
+		if (ret != 0 || !wrote) {
+			while (++next < ar_cnt)
+				--bharray[next]->ref;
+
+			if (ret != 0)
+				goto err;
+
+			/*
+			 * Any process syncing the shared memory buffer pool
+			 * had better be able to write to any underlying file.
+			 * Be understanding, but firm, on this point.
+			 */
+			if (!wrote) {
+				__db_err(dbenv, "%s: unable to flush page: %lu",
+				    R_ADDR(dbmp, mfp->path_off),
+				    (u_long)bharray[next]->pgno);
+				ret = EPERM;
+				goto err;
+			}
+		}
 	}
 	ret = mp->lsn_cnt ? DB_INCOMPLETE : 0;
 
-err:	UNLOCKREGION(dbmp);
+done:
+	if (0) {
+err:		/*
+		 * On error, clear:
+		 *	MPOOL->lsn_cnt (the total sync count)
+		 *	MPOOLFILE->lsn_cnt (the per-file sync count)
+		 *	BH_WRITE flag (the scheduled for writing flag)
+		 */
+		mp->lsn_cnt = 0;
+		for (mfp = SH_TAILQ_FIRST(&dbmp->mp->mpfq, __mpoolfile);
+		    mfp != NULL; mfp = SH_TAILQ_NEXT(mfp, q, __mpoolfile))
+			mfp->lsn_cnt = 0;
+		for (bhp = SH_TAILQ_FIRST(&mp->bhq, __bh);
+		    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, q, __bh))
+			F_CLR(bhp, BH_WRITE);
+	}
+	UNLOCKREGION(dbmp);
+	__db_free(bharray);
 	return (ret);
 }
 
@@ -162,10 +239,10 @@ int
 memp_fsync(dbmfp)
 	DB_MPOOLFILE *dbmfp;
 {
-	BH *bhp;
+	BH *bhp, **bharray;
 	DB_MPOOL *dbmp;
 	size_t mf_offset;
-	int pincnt, restart, ret, wrote;
+	int ar_cnt, cnt, nalloc, next, pincnt, notused, ret, wrote;
 
 	/*
 	 * If this handle doesn't have a file descriptor that's open for
@@ -175,35 +252,205 @@ memp_fsync(dbmfp)
 	if (F_ISSET(dbmfp, MP_READONLY | MP_PATH_TEMP))
 		return (0);
 
-	dbmp = dbmfp->dbmp;
 	ret = 0;
+	dbmp = dbmfp->dbmp;
+	mf_offset = R_OFFSET(dbmp, dbmfp->mfp);
 
-	mf_offset = 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.
+	 */
+	nalloc = 1024;
+	if ((bharray =
+	    (BH **)__db_malloc((size_t)nalloc * sizeof(BH *))) == NULL)
+		return (ENOMEM);
 
 	LOCKREGION(dbmp);
 
 	/*
-	 * Walk the list of buffer headers for the MPOOLFILE, and write out any
-	 * dirty buffers that we can.
+	 * Walk the LRU list of buffer headers, and get a list of buffers to
+	 * write for this MPOOLFILE.
 	 */
-retry:	pincnt = 0;
+	ar_cnt = pincnt = 0;
 	for (bhp = SH_TAILQ_FIRST(&dbmp->mp->bhq, __bh);
-	    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, q, __bh))
-		if (F_ISSET(bhp, BH_DIRTY) && bhp->mf_offset == mf_offset) {
-			if (bhp->ref != 0 || F_ISSET(bhp, BH_LOCKED)) {
-				++pincnt;
-				continue;
-			}
-			if ((ret =
-			    __memp_pgwrite(dbmfp, bhp, &restart, &wrote)) != 0)
+	    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;
+			continue;
+		}
+
+		if (ar_cnt == nalloc) {
+			nalloc *= 2;
+			if ((bharray = (BH **)__db_realloc(bharray,
+			    nalloc * sizeof(BH *))) == NULL) {
+				ret = ENOMEM;
 				goto err;
-			if (!wrote)
-				++pincnt;
-			if (restart)
-				goto retry;
+			}
+		}
+
+		bharray[ar_cnt++] = bhp;
+	}
+
+	/* 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);
+
+	LOCKREGION(dbmp);
+
+	/* Walk the array, writing buffers. */
+	for (next = 0; next < ar_cnt; ++next) {
+		/*
+		 * 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 (bharray[next]->ref > 1) {
+			++pincnt;
+
+			--bharray[next]->ref;
+			continue;
 		}
 
+		/* Write the buffer. */
+		ret = __memp_pgwrite(dbmfp, bharray[next], &notused, &wrote);
+
+		/* Release the buffer. */
+		--bharray[next]->ref;
+
+		/* If there's an error, release the rest of the buffers. */
+		if (ret != 0) {
+			while (++next < ar_cnt)
+				--bharray[next]->ref;
+			goto err;
+		}
+		if (!wrote)
+			++pincnt;
+	}
+
+err:	UNLOCKREGION(dbmp);
+
+	__db_free(bharray);
+
+	/*
+	 * Sync the underlying file as the last thing we do, so that the OS
+	 * has maximal opportunity to flush buffers before we request it.
+	 *
+	 * XXX:
+	 * Don't lock the region around the sync, fsync(2) has no atomicity
+	 * issues.
+	 */
+	if (ret == 0)
+		return (pincnt == 0 ? __db_fsync(dbmfp->fd) : DB_INCOMPLETE);
+	return (ret);
+
+}
+
+/*
+ * memp_trickle --
+ *	Keep a specified percentage of the buffers clean.
+ */
+int
+memp_trickle(dbmp, pct, nwrotep)
+	DB_MPOOL *dbmp;
+	int pct, *nwrotep;
+{
+	BH *bhp;
+	MPOOL *mp;
+	MPOOLFILE *mfp;
+	u_long total;
+	int notused, ret, wrote;
+
+	mp = dbmp->mp;
+	if (nwrotep != NULL)
+		*nwrotep = 0;
+
+	if (pct < 1 || pct > 100)
+		return (EINVAL);
+
+	LOCKREGION(dbmp);
+
+	/*
+	 * If there are sufficient clean buffers, or no buffers or no dirty
+	 * buffers, we're done.
+	 *
+	 * XXX
+	 * Using st_page_clean and st_page_dirty is our only choice at the
+	 * moment, but it's not as correct as we might like in the presence
+	 * of pools with more than one buffer size, as a free 512-byte buffer
+	 * isn't the same as a free 8K buffer.
+	 */
+loop:	total = mp->stat.st_page_clean + mp->stat.st_page_dirty;
+	if (total == 0 || mp->stat.st_page_dirty == 0 ||
+	    (mp->stat.st_page_clean * 100) / total >= (u_long)pct) {
+		UNLOCKREGION(dbmp);
+		return (0);
+	}
+
+	/* Loop until we write a buffer. */
+	for (bhp = SH_TAILQ_FIRST(&mp->bhq, __bh);
+	    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, q, __bh)) {
+		if (bhp->ref != 0 ||
+		    !F_ISSET(bhp, BH_DIRTY) || F_ISSET(bhp, BH_LOCKED))
+			continue;
+
+		mfp = R_ADDR(dbmp, bhp->mf_offset);
+		if ((ret =
+		    __memp_bhwrite(dbmp, mfp, bhp, &notused, &wrote)) != 0)
+			goto err;
+
+		/*
+		 * Any process syncing the shared memory buffer pool
+		 * had better be able to write to any underlying file.
+		 * Be understanding, but firm, on this point.
+		 */
+		if (!wrote) {
+			__db_err(dbmp->dbenv, "%s: unable to flush page: %lu",
+			    R_ADDR(dbmp, mfp->path_off), (u_long)bhp->pgno);
+			ret = EPERM;
+			goto err;
+		}
+
+		++mp->stat.st_page_trickle;
+		if (nwrotep != NULL)
+			++*nwrotep;
+		goto loop;
+	}
+
+	/* No more buffers to write. */
+	return (0);
+
 err:	UNLOCKREGION(dbmp);
+	return (ret);
+}
+
+static int
+__bhcmp(p1, p2)
+	const void *p1, *p2;
+{
+	BH *bhp1, *bhp2;
+
+	bhp1 = *(BH **)p1;
+	bhp2 = *(BH **)p2;
+
+	/* Sort by file (shared memory pool offset). */
+	if (bhp1->mf_offset < bhp2->mf_offset)
+		return (-1);
+	if (bhp1->mf_offset > bhp2->mf_offset)
+		return (1);
 
-	return (ret == 0 ? (pincnt ? DB_INCOMPLETE : 0) : ret);
+	/* Sort by page in file. */
+	return (bhp1->pgno < bhp2->pgno ? -1 : 1);
 }