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

Commit226a100

Browse files
committed
Code review for btree page split WAL reduction patch. Make it actually work
(original code *always* created a full-page image for the left page, thusleaving the intended savings unrealized), avoid risk of not having enough roomon the page during xlog restore, squeeze out another couple bytes in the xlogrecord, clean up neglected comments.
1 parentf7424b0 commit226a100

File tree

4 files changed

+167
-127
lines changed

4 files changed

+167
-127
lines changed

‎src/backend/access/nbtree/nbtinsert.c‎

Lines changed: 62 additions & 45 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,7 @@
88
*
99
*
1010
* IDENTIFICATION
11-
* $PostgreSQL: pgsql/src/backend/access/nbtree/nbtinsert.c,v 1.155 2007/03/25 19:45:14 tgl Exp $
11+
* $PostgreSQL: pgsql/src/backend/access/nbtree/nbtinsert.c,v 1.156 2007/04/11 20:47:37 tgl Exp $
1212
*
1313
*-------------------------------------------------------------------------
1414
*/
@@ -49,7 +49,7 @@ static TransactionId _bt_check_unique(Relation rel, IndexTuple itup,
4949
RelationheapRel,Bufferbuf,OffsetNumberioffset,
5050
ScanKeyitup_scankey);
5151
staticvoid_bt_findinsertloc(Relationrel,
52-
Buffer*bufptr,
52+
Buffer*bufptr,
5353
OffsetNumber*offsetptr,
5454
intkeysz,
5555
ScanKeyscankey,
@@ -66,7 +66,7 @@ static OffsetNumber _bt_findsplitloc(Relation rel, Page page,
6666
OffsetNumbernewitemoff,
6767
Sizenewitemsz,
6868
bool*newitemonleft);
69-
staticvoid_bt_checksplitloc(FindSplitData*state,
69+
staticvoid_bt_checksplitloc(FindSplitData*state,
7070
OffsetNumberfirstoldonright,boolnewitemonleft,
7171
intdataitemstoleft,Sizefirstoldonrightsz);
7272
staticvoid_bt_pgaddtup(Relationrel,Pagepage,
@@ -459,7 +459,7 @@ _bt_findinsertloc(Relation rel,
459459
* the hint supplied by the caller invalid */
460460
vacuumed= true;
461461

462-
if (PageGetFreeSpace(page) >=itemsz)
462+
if (PageGetFreeSpace(page) >=itemsz)
463463
break;/* OK, now we have enough space */
464464
}
465465

@@ -506,7 +506,7 @@ _bt_findinsertloc(Relation rel,
506506
* moved right at all, we know we should insert at the start of the
507507
* page. If we didn't move right, we can use the firstlegaloff hint
508508
* if the caller supplied one, unless we vacuumed the page which
509-
* might have moved tuples around making the hint invalid. If we
509+
* might have moved tuples around making the hint invalid. If we
510510
* didn't move right or can't use the hint, find the position
511511
* by searching.
512512
*/
@@ -779,8 +779,6 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
779779
Buffersbuf=InvalidBuffer;
780780
Pagespage=NULL;
781781
BTPageOpaquesopaque=NULL;
782-
OffsetNumberitup_off=0;
783-
BlockNumberitup_blkno=0;
784782
Sizeitemsz;
785783
ItemIditemid;
786784
IndexTupleitem;
@@ -798,6 +796,14 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
798796
_bt_pageinit(leftpage,BufferGetPageSize(buf));
799797
/* rightpage was already initialized by _bt_getbuf */
800798

799+
/*
800+
* Copy the original page's LSN and TLI into leftpage, which will become
801+
* the updated version of the page. We need this because XLogInsert will
802+
* examine these fields and possibly dump them in a page image.
803+
*/
804+
PageSetLSN(leftpage,PageGetLSN(origpage));
805+
PageSetTLI(leftpage,PageGetTLI(origpage));
806+
801807
/* init btree private data */
802808
oopaque= (BTPageOpaque)PageGetSpecialPointer(origpage);
803809
lopaque= (BTPageOpaque)PageGetSpecialPointer(leftpage);
@@ -864,7 +870,10 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
864870
leftoff=OffsetNumberNext(leftoff);
865871

866872
/*
867-
* Now transfer all the data items to the appropriate page
873+
* Now transfer all the data items to the appropriate page.
874+
*
875+
* Note: we *must* insert at least the right page's items in item-number
876+
* order, for the benefit of _bt_restore_page().
868877
*/
869878
maxoff=PageGetMaxOffsetNumber(origpage);
870879

@@ -881,16 +890,12 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
881890
{
882891
_bt_pgaddtup(rel,leftpage,newitemsz,newitem,leftoff,
883892
"left sibling");
884-
itup_off=leftoff;
885-
itup_blkno=BufferGetBlockNumber(buf);
886893
leftoff=OffsetNumberNext(leftoff);
887894
}
888895
else
889896
{
890897
_bt_pgaddtup(rel,rightpage,newitemsz,newitem,rightoff,
891898
"right sibling");
892-
itup_off=rightoff;
893-
itup_blkno=BufferGetBlockNumber(rbuf);
894899
rightoff=OffsetNumberNext(rightoff);
895900
}
896901
}
@@ -921,8 +926,6 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
921926
Assert(!newitemonleft);
922927
_bt_pgaddtup(rel,rightpage,newitemsz,newitem,rightoff,
923928
"right sibling");
924-
itup_off=rightoff;
925-
itup_blkno=BufferGetBlockNumber(rbuf);
926929
rightoff=OffsetNumberNext(rightoff);
927930
}
928931

@@ -961,7 +964,7 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
961964

962965
/*
963966
* Right sibling is locked, new siblings are prepared, but original page
964-
* is not updated yet. Log changes before continuing.
967+
* is not updated yet.
965968
*
966969
* NO EREPORT(ERROR) till right sibling is updated. We can get away with
967970
* not starting the critical section till here because we haven't been
@@ -970,15 +973,6 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
970973
*/
971974
START_CRIT_SECTION();
972975

973-
MarkBufferDirty(buf);
974-
MarkBufferDirty(rbuf);
975-
976-
if (!P_RIGHTMOST(ropaque))
977-
{
978-
sopaque->btpo_prev=BufferGetBlockNumber(rbuf);
979-
MarkBufferDirty(sbuf);
980-
}
981-
982976
/*
983977
* By here, the original data page has been split into two new halves, and
984978
* these are correct. The algorithm requires that the left page never
@@ -994,6 +988,15 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
994988
*/
995989
PageRestoreTempPage(leftpage,origpage);
996990

991+
MarkBufferDirty(buf);
992+
MarkBufferDirty(rbuf);
993+
994+
if (!P_RIGHTMOST(ropaque))
995+
{
996+
sopaque->btpo_prev=BufferGetBlockNumber(rbuf);
997+
MarkBufferDirty(sbuf);
998+
}
999+
9971000
/* XLOG stuff */
9981001
if (!rel->rd_istemp)
9991002
{
@@ -1006,9 +1009,9 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
10061009
xlrec.node=rel->rd_node;
10071010
xlrec.leftsib=BufferGetBlockNumber(buf);
10081011
xlrec.rightsib=BufferGetBlockNumber(rbuf);
1009-
xlrec.firstright=firstright;
10101012
xlrec.rnext=ropaque->btpo_next;
10111013
xlrec.level=ropaque->btpo.level;
1014+
xlrec.firstright=firstright;
10121015

10131016
rdata[0].data= (char*)&xlrec;
10141017
rdata[0].len=SizeOfBtreeSplit;
@@ -1027,54 +1030,68 @@ _bt_split(Relation rel, Buffer buf, OffsetNumber firstright,
10271030
lastrdata->buffer=InvalidBuffer;
10281031
}
10291032

1030-
/* Log the new item, if it was inserted on the left page. If it was
1031-
* put on the right page, we don't need to explicitly WAL log it
1032-
* because it's included with all the other items on the right page.
1033+
/*
1034+
* Log the new item and its offset, if it was inserted on the left
1035+
* page. (If it was put on the right page, we don't need to explicitly
1036+
* WAL log it because it's included with all the other items on the
1037+
* right page.) Show these as belonging to the left page buffer,
1038+
* so that they are not stored if XLogInsert decides it needs a
1039+
* full-page image of the left page.
10331040
*/
1034-
lastrdata->next=lastrdata+1;
1035-
lastrdata++;
10361041
if (newitemonleft)
10371042
{
1043+
lastrdata->next=lastrdata+1;
1044+
lastrdata++;
10381045
lastrdata->data= (char*)&newitemoff;
10391046
lastrdata->len=sizeof(OffsetNumber);
10401047
lastrdata->buffer=buf;/* backup block 1 */
10411048
lastrdata->buffer_std= true;
10421049

10431050
lastrdata->next=lastrdata+1;
10441051
lastrdata++;
1045-
lastrdata->data= (char*)newitem;
1046-
lastrdata->len=newitemsz;
1052+
lastrdata->data= (char*)newitem;
1053+
lastrdata->len=MAXALIGN(newitemsz);
10471054
lastrdata->buffer=buf;/* backup block 1 */
10481055
lastrdata->buffer_std= true;
10491056
}
10501057
else
10511058
{
1059+
/*
1060+
* Although we don't need to WAL-log the new item, we still
1061+
* need XLogInsert to consider storing a full-page image of the
1062+
* left page, so make an empty entry referencing that buffer.
1063+
* This also ensures that the left page is always backup block 1.
1064+
*/
1065+
lastrdata->next=lastrdata+1;
1066+
lastrdata++;
10521067
lastrdata->data=NULL;
10531068
lastrdata->len=0;
10541069
lastrdata->buffer=buf;/* backup block 1 */
10551070
lastrdata->buffer_std= true;
10561071
}
10571072

1058-
/* Log the contents of the right page in the format understood by
1073+
/*
1074+
* Log the contents of the right page in the format understood by
10591075
* _bt_restore_page(). We set lastrdata->buffer to InvalidBuffer,
1060-
* because we're going to recreate the whole page anyway.
1076+
* because we're going to recreate the whole page anyway, so it
1077+
* should never be stored by XLogInsert.
10611078
*
10621079
* Direct access to page is not good but faster - we should implement
10631080
* some new func in page API. Note we only store the tuples
1064-
* themselves, knowing thatthe item pointers areinthe same order
1065-
* andcan be reconstructed by scanning the tuples. See comments for
1081+
* themselves, knowing thatthey were insertedinitem-number order
1082+
* andso the item pointers can be reconstructed. See comments for
10661083
* _bt_restore_page().
10671084
*/
10681085
lastrdata->next=lastrdata+1;
10691086
lastrdata++;
10701087

1071-
lastrdata->data= (char*)rightpage+
1088+
lastrdata->data= (char*)rightpage+
10721089
((PageHeader)rightpage)->pd_upper;
10731090
lastrdata->len= ((PageHeader)rightpage)->pd_special-
10741091
((PageHeader)rightpage)->pd_upper;
10751092
lastrdata->buffer=InvalidBuffer;
10761093

1077-
/* Log the right sibling, because we've changedit's prev-pointer. */
1094+
/* Log the right sibling, because we've changedits' prev-pointer. */
10781095
if (!P_RIGHTMOST(ropaque))
10791096
{
10801097
lastrdata->next=lastrdata+1;
@@ -1216,7 +1233,7 @@ _bt_findsplitloc(Relation rel,
12161233
olddataitemstoleft=0;
12171234
goodenoughfound= false;
12181235
maxoff=PageGetMaxOffsetNumber(page);
1219-
1236+
12201237
for (offnum=P_FIRSTDATAKEY(opaque);
12211238
offnum <=maxoff;
12221239
offnum=OffsetNumberNext(offnum))
@@ -1234,7 +1251,7 @@ _bt_findsplitloc(Relation rel,
12341251
olddataitemstoleft,itemsz);
12351252

12361253
elseif (offnum<newitemoff)
1237-
_bt_checksplitloc(&state,offnum, false,
1254+
_bt_checksplitloc(&state,offnum, false,
12381255
olddataitemstoleft,itemsz);
12391256
else
12401257
{
@@ -1285,11 +1302,11 @@ _bt_findsplitloc(Relation rel,
12851302
* items go to the left page and only the new item goes to the right page.
12861303
* In that case, firstoldonrightsz is not used.
12871304
*
1288-
* olddataitemstoleft is the total size of all old items to the left of
1289-
* firstoldonright.
1305+
* olddataitemstoleft is the total size of all old items to the left of
1306+
* firstoldonright.
12901307
*/
12911308
staticvoid
1292-
_bt_checksplitloc(FindSplitData*state,
1309+
_bt_checksplitloc(FindSplitData*state,
12931310
OffsetNumberfirstoldonright,
12941311
boolnewitemonleft,
12951312
intolddataitemstoleft,
@@ -1311,7 +1328,7 @@ _bt_checksplitloc(FindSplitData *state,
13111328

13121329
/* Account for all the old tuples */
13131330
leftfree=state->leftspace-olddataitemstoleft;
1314-
rightfree=state->rightspace-
1331+
rightfree=state->rightspace-
13151332
(state->olddataitemstotal-olddataitemstoleft);
13161333

13171334
/*
@@ -1854,7 +1871,7 @@ _bt_vacuum_one_page(Relation rel, Buffer buffer)
18541871
BTPageOpaqueopaque= (BTPageOpaque)PageGetSpecialPointer(page);
18551872

18561873
/*
1857-
* Scan over all items to see which ones need to be deleted
1874+
* Scan over all items to see which ones need to be deleted
18581875
* according to LP_DELETE flags.
18591876
*/
18601877
minoff=P_FIRSTDATAKEY(opaque);

0 commit comments

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