*** mp/mp_bh.c.orig	Wed Nov 25 10:46:58 1998
--- mp/mp_bh.c	Thu Apr 15 21:19:17 1999
***************
*** 267,273 ****
  	MPOOL *mp;
  	MPOOLFILE *mfp;
  	ssize_t nw;
! 	int callpgin, ret, syncfail;
  	const char *fail;
  
  	dbmp = dbmfp->dbmp;
--- 267,273 ----
  	MPOOL *mp;
  	MPOOLFILE *mfp;
  	ssize_t nw;
! 	int callpgin, dosync, ret, syncfail;
  	const char *fail;
  
  	dbmp = dbmfp->dbmp;
***************
*** 386,425 ****
  	/*
  	 * If we write a buffer for which a checkpoint is waiting, update
  	 * the count of pending buffers (both in the mpool as a whole and
! 	 * for this file).  If the count for this file goes to zero, flush
! 	 * the writes.
! 	 *
! 	 * XXX:
! 	 * Don't lock the region around the sync, fsync(2) has no atomicity
! 	 * issues.
! 	 *
! 	 * XXX:
! 	 * We ignore errors from the sync -- it makes no sense to return an
! 	 * error to the calling process, so set a flag causing the checkpoint
! 	 * to be retried later.
  	 */
  	if (F_ISSET(bhp, BH_WRITE)) {
- 		if (mfp->lsn_cnt == 1) {
- 			UNLOCKREGION(dbmp);
- 			syncfail = __os_fsync(dbmfp->fd) != 0;
- 			LOCKREGION(dbmp);
- 			if (syncfail)
- 				F_SET(mp, MP_LSN_RETRY);
- 
- 		}
- 
  		F_CLR(bhp, BH_WRITE);
  
- 		/*
- 		 * If the buffer just written has a larger LSN than the current
- 		 * max LSN written for this checkpoint, update the saved value.
- 		 */
- 		if (log_compare(&lsn, &mp->lsn) > 0)
- 			mp->lsn = lsn;
- 
  		--mp->lsn_cnt;
! 		--mfp->lsn_cnt;
! 	}
  
  	/* Update the page clean/dirty statistics. */
  	++mp->stat.st_page_clean;
--- 386,401 ----
  	/*
  	 * If we write a buffer for which a checkpoint is waiting, update
  	 * the count of pending buffers (both in the mpool as a whole and
! 	 * for this file).  If the count for this file goes to zero, set a
! 	 * flag so we flush the writes.
  	 */
  	if (F_ISSET(bhp, BH_WRITE)) {
  		F_CLR(bhp, BH_WRITE);
  
  		--mp->lsn_cnt;
! 		dosync = --mfp->lsn_cnt == 0 ? 1 : 0;
! 	} else
! 		dosync = 0;
  
  	/* Update the page clean/dirty statistics. */
  	++mp->stat.st_page_clean;
***************
*** 428,433 ****
--- 404,432 ----
  	/* Update I/O statistics. */
  	++mp->stat.st_page_out;
  	++mfp->stat.st_page_out;
+ 
+ 	/*
+ 	 * Do the sync after everything else has been updated, so any incoming
+ 	 * checkpoint doesn't see inconsistent information.
+ 	 *
+ 	 * XXX:
+ 	 * Don't lock the region around the sync, fsync(2) has no atomicity
+ 	 * issues.
+ 	 *
+ 	 * XXX:
+ 	 * We ignore errors from the sync -- it makes no sense to return an
+ 	 * error to the calling process, so set a flag causing the checkpoint
+ 	 * to be retried later.  There is a possibility, of course, that a
+ 	 * subsequent checkpoint was started and that we're going to force it
+ 	 * to fail.  That should be unlikely, and fixing it would be difficult.
+ 	 */
+ 	if (dosync) {
+ 		UNLOCKREGION(dbmp);
+ 		syncfail = __os_fsync(dbmfp->fd) != 0;
+ 		LOCKREGION(dbmp);
+ 		if (syncfail)
+ 			F_SET(mp, MP_LSN_RETRY);
+ 	}
  
  	return (0);