Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Commit2692d32

Browse files
committed
Tweak smgrblindwrt per advice from Vadim: add parameter indicating
whether to do fsync or not, and if so (which should be seldom) justdo the fsync immediately. This way we need not build data structuresin md.c/fd.c for blind writes.
1 parenta447ae2 commit2692d32

File tree

5 files changed

+95
-135
lines changed

5 files changed

+95
-135
lines changed

‎src/backend/storage/buffer/bufmgr.c‎

Lines changed: 9 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,7 @@
88
*
99
*
1010
* IDENTIFICATION
11-
* $Header: /cvsroot/pgsql/src/backend/storage/buffer/bufmgr.c,v 1.78 2000/04/09 04:43:18 tgl Exp $
11+
* $Header: /cvsroot/pgsql/src/backend/storage/buffer/bufmgr.c,v 1.79 2000/04/10 23:41:49 tgl Exp $
1212
*
1313
*-------------------------------------------------------------------------
1414
*/
@@ -1127,7 +1127,8 @@ BufferSync()
11271127
bufHdr->blind.relname,
11281128
bufdb,bufrel,
11291129
bufHdr->tag.blockNum,
1130-
(char*)MAKE_PTR(bufHdr->data));
1130+
(char*)MAKE_PTR(bufHdr->data),
1131+
true);/* must fsync */
11311132
}
11321133
else
11331134
{
@@ -1529,7 +1530,8 @@ BufferReplace(BufferDesc *bufHdr)
15291530
status=smgrblindwrt(DEFAULT_SMGR,bufHdr->blind.dbname,
15301531
bufHdr->blind.relname,bufdb,bufrel,
15311532
bufHdr->tag.blockNum,
1532-
(char*)MAKE_PTR(bufHdr->data));
1533+
(char*)MAKE_PTR(bufHdr->data),
1534+
false);/* no fsync */
15331535
}
15341536

15351537
#ifndefOPTIMIZE_SINGLE
@@ -1544,9 +1546,11 @@ BufferReplace(BufferDesc *bufHdr)
15441546
return FALSE;
15451547

15461548
/* If we had marked this buffer as needing to be fsync'd, we can forget
1547-
* about that, because it's now the storage manager's responsibility.
1549+
* about that, because it's now the storage manager's responsibility
1550+
* (but only if we called smgrwrite, not smgrblindwrt).
15481551
*/
1549-
ClearBufferDirtiedByMe(BufferDescriptorGetBuffer(bufHdr),bufHdr);
1552+
if (reln!= (Relation)NULL)
1553+
ClearBufferDirtiedByMe(BufferDescriptorGetBuffer(bufHdr),bufHdr);
15501554

15511555
BufferFlushCount++;
15521556

‎src/backend/storage/smgr/md.c‎

Lines changed: 68 additions & 119 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,7 @@
88
*
99
*
1010
* IDENTIFICATION
11-
* $Header: /cvsroot/pgsql/src/backend/storage/smgr/md.c,v 1.65 2000/04/09 04:43:20 tgl Exp $
11+
* $Header: /cvsroot/pgsql/src/backend/storage/smgr/md.c,v 1.66 2000/04/10 23:41:51 tgl Exp $
1212
*
1313
*-------------------------------------------------------------------------
1414
*/
@@ -48,11 +48,10 @@
4848
typedefstruct_MdfdVec
4949
{
5050
intmdfd_vfd;/* fd number in vfd pool */
51-
intmdfd_flags;/*free, temporary */
51+
intmdfd_flags;/*fd status flags */
5252

5353
/* these are the assigned bits in mdfd_flags: */
5454
#defineMDFD_FREE(1 << 0)/* unused entry */
55-
#defineMDFD_TEMP(1 << 1)/* close this entry at transaction end */
5655

5756
intmdfd_lstbcnt;/* most recent block count */
5857
intmdfd_nextFree;/* next free vector */
@@ -72,8 +71,8 @@ static void mdclose_fd(int fd);
7271
staticint_mdfd_getrelnfd(Relationreln);
7372
staticMdfdVec*_mdfd_openseg(Relationreln,intsegno,intoflags);
7473
staticMdfdVec*_mdfd_getseg(Relationreln,intblkno);
75-
staticMdfdVec*_mdfd_blind_getseg(char*dbname,char*relname,
76-
Oiddbid,Oidrelid,intblkno);
74+
staticint_mdfd_blind_getseg(char*dbname,char*relname,
75+
Oiddbid,Oidrelid,intblkno);
7776
staticint_fdvec_alloc(void);
7877
staticvoid_fdvec_free(int);
7978
staticBlockNumber_mdnblocks(Filefile,Sizeblcksz);
@@ -572,23 +571,25 @@ mdflush(Relation reln, BlockNumber blocknum, char *buffer)
572571
*
573572
*We have to be able to do this using only the name and OID of
574573
*the database and relation in which the block belongs. Otherwise
575-
*this is just like mdwrite().
574+
*this is much like mdwrite(). If dofsync is TRUE, then we fsync
575+
*the file, making it more like mdflush().
576576
*/
577577
int
578578
mdblindwrt(char*dbname,
579579
char*relname,
580580
Oiddbid,
581581
Oidrelid,
582582
BlockNumberblkno,
583-
char*buffer)
583+
char*buffer,
584+
booldofsync)
584585
{
585586
intstatus;
586587
longseekpos;
587-
MdfdVec*v;
588+
intfd;
588589

589-
v=_mdfd_blind_getseg(dbname,relname,dbid,relid,blkno);
590+
fd=_mdfd_blind_getseg(dbname,relname,dbid,relid,blkno);
590591

591-
if (v==NULL)
592+
if (fd<0)
592593
returnSM_FAIL;
593594

594595
#ifndefLET_OS_MANAGE_FILESIZE
@@ -601,11 +602,22 @@ mdblindwrt(char *dbname,
601602
seekpos= (long) (BLCKSZ* (blkno));
602603
#endif
603604

604-
if (FileSeek(v->mdfd_vfd,seekpos,SEEK_SET)!=seekpos)
605+
if (lseek(fd,seekpos,SEEK_SET)!=seekpos)
606+
{
607+
close(fd);
605608
returnSM_FAIL;
609+
}
606610

607611
status=SM_SUCCESS;
608-
if (FileWrite(v->mdfd_vfd,buffer,BLCKSZ)!=BLCKSZ)
612+
613+
/* write and optionally sync the block */
614+
if (write(fd,buffer,BLCKSZ)!=BLCKSZ)
615+
status=SM_FAIL;
616+
elseif (dofsync&&
617+
pg_fsync(fd)<0)
618+
status=SM_FAIL;
619+
620+
if (close(fd)<0)
609621
status=SM_FAIL;
610622

611623
returnstatus;
@@ -633,7 +645,8 @@ mdmarkdirty(Relation reln, BlockNumber blkno)
633645
*
634646
*We have to be able to do this using only the name and OID of
635647
*the database and relation in which the block belongs. Otherwise
636-
*this is just like mdmarkdirty().
648+
*this is much like mdmarkdirty(). However, we do the fsync immediately
649+
*rather than building md/fd datastructures to postpone it till later.
637650
*/
638651
int
639652
mdblindmarkdirty(char*dbname,
@@ -642,16 +655,23 @@ mdblindmarkdirty(char *dbname,
642655
Oidrelid,
643656
BlockNumberblkno)
644657
{
645-
MdfdVec*v;
658+
intstatus;
659+
intfd;
646660

647-
v=_mdfd_blind_getseg(dbname,relname,dbid,relid,blkno);
661+
fd=_mdfd_blind_getseg(dbname,relname,dbid,relid,blkno);
648662

649-
if (v==NULL)
663+
if (fd<0)
650664
returnSM_FAIL;
651665

652-
FileMarkDirty(v->mdfd_vfd);
666+
status=SM_SUCCESS;
653667

654-
returnSM_SUCCESS;
668+
if (pg_fsync(fd)<0)
669+
status=SM_FAIL;
670+
671+
if (close(fd)<0)
672+
status=SM_FAIL;
673+
674+
returnstatus;
655675
}
656676

657677
/*
@@ -820,24 +840,15 @@ mdcommit()
820840
v=&Md_fdvec[i];
821841
if (v->mdfd_flags&MDFD_FREE)
822842
continue;
823-
if (v->mdfd_flags&MDFD_TEMP)
824-
{
825-
/* Sync and close the file */
826-
mdclose_fd(i);
827-
}
828-
else
829-
{
830-
/* Sync, but keep the file entry */
831-
843+
/* Sync the file entry */
832844
#ifndefLET_OS_MANAGE_FILESIZE
833-
for ( ;v!= (MdfdVec*)NULL;v=v->mdfd_chain)
845+
for ( ;v!= (MdfdVec*)NULL;v=v->mdfd_chain)
834846
#else
835-
if (v!= (MdfdVec*)NULL)
847+
if (v!= (MdfdVec*)NULL)
836848
#endif
837-
{
838-
if (FileSync(v->mdfd_vfd)<0)
839-
returnSM_FAIL;
840-
}
849+
{
850+
if (FileSync(v->mdfd_vfd)<0)
851+
returnSM_FAIL;
841852
}
842853
}
843854

@@ -854,21 +865,9 @@ mdcommit()
854865
int
855866
mdabort()
856867
{
857-
inti;
858-
MdfdVec*v;
859-
860-
for (i=0;i<CurFd;i++)
861-
{
862-
v=&Md_fdvec[i];
863-
if (v->mdfd_flags&MDFD_FREE)
864-
continue;
865-
if (v->mdfd_flags&MDFD_TEMP)
866-
{
867-
/* Close the file */
868-
mdclose_fd(i);
869-
}
870-
}
871-
868+
/* We don't actually have to do anything here. fd.c will discard
869+
* fsync-needed bits in its AtEOXact_Files() routine.
870+
*/
872871
returnSM_SUCCESS;
873872
}
874873

@@ -1057,102 +1056,52 @@ _mdfd_getseg(Relation reln, int blkno)
10571056
returnv;
10581057
}
10591058

1060-
/* Find the segment of the relation holding the specified block.
1061-
* This is the same as _mdfd_getseg() except that we must work
1062-
* "blind" with no Relation struct.
1059+
/*
1060+
* Find the segment of the relation holding the specified block.
10631061
*
1064-
* NOTE: we have no easy way to tell whether a FD already exists for the
1065-
* target relation, so we always make a new one. This should probably
1066-
* be improved somehow, but I doubt it's a significant performance issue
1067-
* under normal circumstances. The FD is marked to be closed at end of xact
1068-
* so that we don't accumulate a lot of dead FDs.
1062+
* This performs the same work as _mdfd_getseg() except that we must work
1063+
* "blind" with no Relation struct. We assume that we are not likely to
1064+
* touch the same relation again soon, so we do not create an FD entry for
1065+
* the relation --- we just open a kernel file descriptor which will be
1066+
* used and promptly closed. The return value is the kernel descriptor,
1067+
* or -1 on failure.
10691068
*/
10701069

1071-
staticMdfdVec*
1070+
staticint
10721071
_mdfd_blind_getseg(char*dbname,char*relname,Oiddbid,Oidrelid,
10731072
intblkno)
10741073
{
1075-
MdfdVec*v;
10761074
char*path;
10771075
intfd;
1078-
intvfd;
10791076
#ifndefLET_OS_MANAGE_FILESIZE
10801077
intsegno;
1081-
inttargsegno;
10821078
#endif
10831079

1084-
/* construct the path to thefile and open it */
1080+
/* construct the path to therelation */
10851081
path=relpath_blind(dbname,relname,dbid,relid);
10861082

1087-
#ifndef__CYGWIN32__
1088-
fd=FileNameOpenFile(path,O_RDWR,0600);
1089-
#else
1090-
fd=FileNameOpenFile(path,O_RDWR |O_BINARY,0600);
1091-
#endif
1092-
1093-
if (fd<0)
1094-
returnNULL;
1095-
1096-
vfd=_fdvec_alloc();
1097-
if (vfd<0)
1098-
returnNULL;
1099-
1100-
Md_fdvec[vfd].mdfd_vfd=fd;
1101-
Md_fdvec[vfd].mdfd_flags=MDFD_TEMP;
1102-
Md_fdvec[vfd].mdfd_lstbcnt=_mdnblocks(fd,BLCKSZ);
11031083
#ifndefLET_OS_MANAGE_FILESIZE
1104-
Md_fdvec[vfd].mdfd_chain= (MdfdVec*)NULL;
1105-
1106-
#ifdefDIAGNOSTIC
1107-
if (Md_fdvec[vfd].mdfd_lstbcnt>RELSEG_SIZE)
1108-
elog(FATAL,"segment too big on relopen!");
1109-
#endif
1110-
1111-
targsegno=blkno /RELSEG_SIZE;
1112-
for (v=&Md_fdvec[vfd],segno=1;segno <=targsegno;segno++)
1084+
/* append the '.segno', if needed */
1085+
segno=blkno /RELSEG_SIZE;
1086+
if (segno>0)
11131087
{
1114-
char*segpath;
1115-
MdfdVec*newv;
1116-
MemoryContextoldcxt;
1088+
char*segpath= (char*)palloc(strlen(path)+12);
11171089

1118-
segpath= (char*)palloc(strlen(path)+12);
11191090
sprintf(segpath,"%s.%d",path,segno);
1120-
1121-
#ifndef__CYGWIN32__
1122-
fd=FileNameOpenFile(segpath,O_RDWR |O_CREAT,0600);
1123-
#else
1124-
fd=FileNameOpenFile(segpath,O_RDWR |O_BINARY |O_CREAT,0600);
1091+
pfree(path);
1092+
path=segpath;
1093+
}
11251094
#endif
11261095

1127-
pfree(segpath);
1128-
1129-
if (fd<0)
1130-
return (MdfdVec*)NULL;
1131-
1132-
/* allocate an mdfdvec entry for it */
1133-
oldcxt=MemoryContextSwitchTo(MdCxt);
1134-
newv= (MdfdVec*)palloc(sizeof(MdfdVec));
1135-
MemoryContextSwitchTo(oldcxt);
1136-
1137-
/* fill the entry */
1138-
newv->mdfd_vfd=fd;
1139-
newv->mdfd_flags=MDFD_TEMP;
1140-
newv->mdfd_lstbcnt=_mdnblocks(fd,BLCKSZ);
1141-
newv->mdfd_chain= (MdfdVec*)NULL;
1142-
#ifdefDIAGNOSTIC
1143-
if (newv->mdfd_lstbcnt>RELSEG_SIZE)
1144-
elog(FATAL,"segment too big on open!");
1145-
#endif
1146-
v->mdfd_chain=newv;
1147-
v=newv;
1148-
}
1096+
#ifndef__CYGWIN32__
1097+
fd=open(path,O_RDWR,0600);
11491098
#else
1150-
v=&Md_fdvec[vfd];
1099+
fd=open(path,O_RDWR |O_BINARY,0600);
11511100
#endif
11521101

11531102
pfree(path);
11541103

1155-
returnv;
1104+
returnfd;
11561105
}
11571106

11581107
staticBlockNumber

‎src/backend/storage/smgr/mm.c‎

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -11,7 +11,7 @@
1111
*
1212
*
1313
* IDENTIFICATION
14-
* $Header: /cvsroot/pgsql/src/backend/storage/smgr/Attic/mm.c,v 1.18 2000/01/26 05:57:05 momjian Exp $
14+
* $Header: /cvsroot/pgsql/src/backend/storage/smgr/Attic/mm.c,v 1.19 2000/04/10 23:41:51 tgl Exp $
1515
*
1616
*-------------------------------------------------------------------------
1717
*/
@@ -478,7 +478,8 @@ mmblindwrt(char *dbstr,
478478
Oiddbid,
479479
Oidrelid,
480480
BlockNumberblkno,
481-
char*buffer)
481+
char*buffer,
482+
booldofsync)
482483
{
483484
returnSM_FAIL;
484485
}

0 commit comments

Comments
 (0)
[8]ページ先頭
©2009-2025 Movatter.jp