postgresql/src/backend/access/spgist/spgxlog.c

1014 lines
26 KiB
C
Raw Normal View History

/*-------------------------------------------------------------------------
*
* spgxlog.c
* WAL replay logic for SP-GiST
*
*
* Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
* IDENTIFICATION
* src/backend/access/spgist/spgxlog.c
*
*-------------------------------------------------------------------------
*/
#include "postgres.h"
#include "access/bufmask.h"
#include "access/spgist_private.h"
#include "access/spgxlog.h"
#include "access/transam.h"
#include "access/xlog.h"
#include "access/xlogutils.h"
#include "storage/standby.h"
#include "utils/memutils.h"
static MemoryContext opCtx; /* working memory for operations */
/*
* Prepare a dummy SpGistState, with just the minimum info needed for replay.
*
* At present, all we need is enough info to support spgFormDeadTuple(),
* plus the isBuild flag.
*/
static void
fillFakeState(SpGistState *state, spgxlogState stateSrc)
{
memset(state, 0, sizeof(*state));
state->myXid = stateSrc.myXid;
state->isBuild = stateSrc.isBuild;
state->deadTupleStorage = palloc0(SGDTSIZE);
}
/*
* Add a leaf tuple, or replace an existing placeholder tuple. This is used
* to replay SpGistPageAddNewItem() operations. If the offset points at an
* existing tuple, it had better be a placeholder tuple.
*/
static void
addOrReplaceTuple(Page page, Item tuple, int size, OffsetNumber offset)
{
if (offset <= PageGetMaxOffsetNumber(page))
{
SpGistDeadTuple dt = (SpGistDeadTuple) PageGetItem(page,
PageGetItemId(page, offset));
if (dt->tupstate != SPGIST_PLACEHOLDER)
elog(ERROR, "SPGiST tuple to be replaced is not a placeholder");
Assert(SpGistPageGetOpaque(page)->nPlaceholder > 0);
SpGistPageGetOpaque(page)->nPlaceholder--;
PageIndexTupleDelete(page, offset);
}
Assert(offset <= PageGetMaxOffsetNumber(page) + 1);
if (PageAddItem(page, tuple, size, offset, false, false) != offset)
elog(ERROR, "failed to add item of size %u to SPGiST index page",
size);
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoAddLeaf(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogAddLeaf *xldata = (spgxlogAddLeaf *) ptr;
char *leafTuple;
SpGistLeafTupleData leafTupleHdr;
Buffer buffer;
Page page;
XLogRedoAction action;
ptr += sizeof(spgxlogAddLeaf);
leafTuple = ptr;
/* the leaf tuple is unaligned, so make a copy to access its header */
memcpy(&leafTupleHdr, leafTuple, sizeof(SpGistLeafTupleData));
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* In normal operation we would have both current and parent pages locked
* simultaneously; but in WAL replay it should be safe to update the leaf
* page before updating the parent.
*/
if (xldata->newPage)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
buffer = XLogInitBufferForRedo(record, 0);
SpGistInitBuffer(buffer,
SPGIST_LEAF | (xldata->storesNulls ? SPGIST_NULLS : 0));
action = BLK_NEEDS_REDO;
}
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
else
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
action = XLogReadBufferForRedo(record, 0, &buffer);
if (action == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
/* insert new tuple */
if (xldata->offnumLeaf != xldata->offnumHeadLeaf)
{
/* normal cases, tuple was added by SpGistPageAddNewItem */
addOrReplaceTuple(page, (Item) leafTuple, leafTupleHdr.size,
xldata->offnumLeaf);
/* update head tuple's chain link if needed */
if (xldata->offnumHeadLeaf != InvalidOffsetNumber)
{
SpGistLeafTuple head;
head = (SpGistLeafTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumHeadLeaf));
Assert(head->nextOffset == leafTupleHdr.nextOffset);
head->nextOffset = xldata->offnumLeaf;
}
}
else
{
/* replacing a DEAD tuple */
PageIndexTupleDelete(page, xldata->offnumLeaf);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (PageAddItem(page,
(Item) leafTuple, leafTupleHdr.size,
xldata->offnumLeaf, false, false) != xldata->offnumLeaf)
elog(ERROR, "failed to add item of size %u to SPGiST index page",
leafTupleHdr.size);
}
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
/* update parent downlink if necessary */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->offnumParent != InvalidOffsetNumber)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 1, &buffer) == BLK_NEEDS_REDO)
{
SpGistInnerTuple tuple;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
BlockNumber blknoLeaf;
XLogRecGetBlockTag(record, 0, NULL, NULL, &blknoLeaf);
page = BufferGetPage(buffer);
tuple = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
spgUpdateNodeLink(tuple, xldata->nodeI,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoLeaf, xldata->offnumLeaf);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoMoveLeafs(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogMoveLeafs *xldata = (spgxlogMoveLeafs *) ptr;
SpGistState state;
OffsetNumber *toDelete;
OffsetNumber *toInsert;
int nInsert;
Buffer buffer;
Page page;
XLogRedoAction action;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
BlockNumber blknoDst;
XLogRecGetBlockTag(record, 1, NULL, NULL, &blknoDst);
fillFakeState(&state, xldata->stateSrc);
nInsert = xldata->replaceDead ? 1 : xldata->nMoves + 1;
ptr += SizeOfSpgxlogMoveLeafs;
toDelete = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nMoves;
toInsert = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * nInsert;
/* now ptr points to the list of leaf tuples */
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* In normal operation we would have all three pages (source, dest, and
* parent) locked simultaneously; but in WAL replay it should be safe to
* update them one at a time, as long as we do it in the right order.
*/
/* Insert tuples on the dest page (do first, so redirect is valid) */
if (xldata->newPage)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
buffer = XLogInitBufferForRedo(record, 1);
SpGistInitBuffer(buffer,
SPGIST_LEAF | (xldata->storesNulls ? SPGIST_NULLS : 0));
action = BLK_NEEDS_REDO;
}
else
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
action = XLogReadBufferForRedo(record, 1, &buffer);
if (action == BLK_NEEDS_REDO)
{
int i;
page = BufferGetPage(buffer);
for (i = 0; i < nInsert; i++)
{
char *leafTuple;
SpGistLeafTupleData leafTupleHdr;
/*
* the tuples are not aligned, so must copy to access the size
* field.
*/
leafTuple = ptr;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
memcpy(&leafTupleHdr, leafTuple,
sizeof(SpGistLeafTupleData));
addOrReplaceTuple(page, (Item) leafTuple,
leafTupleHdr.size, toInsert[i]);
ptr += leafTupleHdr.size;
}
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
/* Delete tuples from the source page, inserting a redirection pointer */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgPageIndexMultiDelete(&state, page, toDelete, xldata->nMoves,
state.isBuild ? SPGIST_PLACEHOLDER : SPGIST_REDIRECT,
SPGIST_PLACEHOLDER,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoDst,
toInsert[nInsert - 1]);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
/* And update the parent downlink */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 2, &buffer) == BLK_NEEDS_REDO)
{
SpGistInnerTuple tuple;
page = BufferGetPage(buffer);
tuple = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
spgUpdateNodeLink(tuple, xldata->nodeI,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoDst, toInsert[nInsert - 1]);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoAddNode(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogAddNode *xldata = (spgxlogAddNode *) ptr;
char *innerTuple;
SpGistInnerTupleData innerTupleHdr;
SpGistState state;
Buffer buffer;
Page page;
XLogRedoAction action;
ptr += sizeof(spgxlogAddNode);
innerTuple = ptr;
/* the tuple is unaligned, so make a copy to access its header */
memcpy(&innerTupleHdr, innerTuple, sizeof(SpGistInnerTupleData));
fillFakeState(&state, xldata->stateSrc);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (!XLogRecHasBlockRef(record, 1))
{
/* update in place */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
Assert(xldata->parentBlk == -1);
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
PageIndexTupleDelete(page, xldata->offnum);
if (PageAddItem(page, (Item) innerTuple, innerTupleHdr.size,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
xldata->offnum,
false, false) != xldata->offnum)
elog(ERROR, "failed to add item of size %u to SPGiST index page",
innerTupleHdr.size);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
else
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
BlockNumber blkno;
BlockNumber blknoNew;
XLogRecGetBlockTag(record, 0, NULL, NULL, &blkno);
XLogRecGetBlockTag(record, 1, NULL, NULL, &blknoNew);
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* In normal operation we would have all three pages (source, dest,
* and parent) locked simultaneously; but in WAL replay it should be
* safe to update them one at a time, as long as we do it in the right
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
* order. We must insert the new tuple before replacing the old tuple
* with the redirect tuple.
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
*/
/* Install new tuple first so redirect is valid */
if (xldata->newPage)
{
/* AddNode is not used for nulls pages */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
buffer = XLogInitBufferForRedo(record, 1);
SpGistInitBuffer(buffer, 0);
action = BLK_NEEDS_REDO;
}
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
else
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
action = XLogReadBufferForRedo(record, 1, &buffer);
if (action == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
addOrReplaceTuple(page, (Item) innerTuple,
innerTupleHdr.size, xldata->offnumNew);
/*
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
* If parent is in this same page, update it now.
*/
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->parentBlk == 1)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
SpGistInnerTuple parentTuple;
parentTuple = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgUpdateNodeLink(parentTuple, xldata->nodeI,
blknoNew, xldata->offnumNew);
}
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
/* Delete old tuple, replacing it with redirect or placeholder tuple */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
SpGistDeadTuple dt;
page = BufferGetPage(buffer);
if (state.isBuild)
dt = spgFormDeadTuple(&state, SPGIST_PLACEHOLDER,
InvalidBlockNumber,
InvalidOffsetNumber);
else
dt = spgFormDeadTuple(&state, SPGIST_REDIRECT,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoNew,
xldata->offnumNew);
PageIndexTupleDelete(page, xldata->offnum);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (PageAddItem(page, (Item) dt, dt->size,
xldata->offnum,
false, false) != xldata->offnum)
elog(ERROR, "failed to add item of size %u to SPGiST index page",
dt->size);
if (state.isBuild)
SpGistPageGetOpaque(page)->nPlaceholder++;
else
SpGistPageGetOpaque(page)->nRedirection++;
/*
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
* If parent is in this same page, update it now.
*/
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->parentBlk == 0)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
SpGistInnerTuple parentTuple;
parentTuple = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgUpdateNodeLink(parentTuple, xldata->nodeI,
blknoNew, xldata->offnumNew);
}
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
/*
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
* Update parent downlink (if we didn't do it as part of the source or
* destination page update already).
*/
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->parentBlk == 2)
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 2, &buffer) == BLK_NEEDS_REDO)
{
SpGistInnerTuple parentTuple;
page = BufferGetPage(buffer);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
parentTuple = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgUpdateNodeLink(parentTuple, xldata->nodeI,
blknoNew, xldata->offnumNew);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
}
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoSplitTuple(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogSplitTuple *xldata = (spgxlogSplitTuple *) ptr;
char *prefixTuple;
SpGistInnerTupleData prefixTupleHdr;
char *postfixTuple;
SpGistInnerTupleData postfixTupleHdr;
Buffer buffer;
Page page;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRedoAction action;
ptr += sizeof(spgxlogSplitTuple);
prefixTuple = ptr;
/* the prefix tuple is unaligned, so make a copy to access its header */
memcpy(&prefixTupleHdr, prefixTuple, sizeof(SpGistInnerTupleData));
ptr += prefixTupleHdr.size;
postfixTuple = ptr;
/* postfix tuple is also unaligned */
memcpy(&postfixTupleHdr, postfixTuple, sizeof(SpGistInnerTupleData));
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* In normal operation we would have both pages locked simultaneously; but
* in WAL replay it should be safe to update them one at a time, as long
* as we do it in the right order.
*/
/* insert postfix tuple first to avoid dangling link */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (!xldata->postfixBlkSame)
{
if (xldata->newPage)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
buffer = XLogInitBufferForRedo(record, 1);
/* SplitTuple is not used for nulls pages */
SpGistInitBuffer(buffer, 0);
action = BLK_NEEDS_REDO;
}
else
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
action = XLogReadBufferForRedo(record, 1, &buffer);
if (action == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
addOrReplaceTuple(page, (Item) postfixTuple,
postfixTupleHdr.size, xldata->offnumPostfix);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
/* now handle the original page */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
PageIndexTupleDelete(page, xldata->offnumPrefix);
if (PageAddItem(page, (Item) prefixTuple, prefixTupleHdr.size,
xldata->offnumPrefix, false, false) != xldata->offnumPrefix)
elog(ERROR, "failed to add item of size %u to SPGiST index page",
prefixTupleHdr.size);
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->postfixBlkSame)
addOrReplaceTuple(page, (Item) postfixTuple,
postfixTupleHdr.size,
xldata->offnumPostfix);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoPickSplit(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogPickSplit *xldata = (spgxlogPickSplit *) ptr;
char *innerTuple;
SpGistInnerTupleData innerTupleHdr;
SpGistState state;
OffsetNumber *toDelete;
OffsetNumber *toInsert;
uint8 *leafPageSelect;
Buffer srcBuffer;
Buffer destBuffer;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
Buffer innerBuffer;
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
Page srcPage;
Page destPage;
Page page;
int i;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
BlockNumber blknoInner;
XLogRedoAction action;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecGetBlockTag(record, 2, NULL, NULL, &blknoInner);
fillFakeState(&state, xldata->stateSrc);
ptr += SizeOfSpgxlogPickSplit;
toDelete = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nDelete;
toInsert = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nInsert;
leafPageSelect = (uint8 *) ptr;
ptr += sizeof(uint8) * xldata->nInsert;
innerTuple = ptr;
/* the inner tuple is unaligned, so make a copy to access its header */
memcpy(&innerTupleHdr, innerTuple, sizeof(SpGistInnerTupleData));
ptr += innerTupleHdr.size;
/* now ptr points to the list of leaf tuples */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->isRootSplit)
{
/* when splitting root, we touch it only in the guise of new inner */
srcBuffer = InvalidBuffer;
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
srcPage = NULL;
}
else if (xldata->initSrc)
{
/* just re-init the source page */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
srcBuffer = XLogInitBufferForRedo(record, 0);
srcPage = (Page) BufferGetPage(srcBuffer);
SpGistInitBuffer(srcBuffer,
SPGIST_LEAF | (xldata->storesNulls ? SPGIST_NULLS : 0));
/* don't update LSN etc till we're done with it */
}
else
{
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* Delete the specified tuples from source page. (In case we're in
* Hot Standby, we need to hold lock on the page till we're done
* inserting leaf tuples and the new inner tuple, else the added
* redirect tuple will be a dangling link.)
*/
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
srcPage = NULL;
if (XLogReadBufferForRedo(record, 0, &srcBuffer) == BLK_NEEDS_REDO)
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
{
srcPage = BufferGetPage(srcBuffer);
/*
* We have it a bit easier here than in doPickSplit(), because we
* know the inner tuple's location already, so we can inject the
* correct redirection tuple now.
*/
if (!state.isBuild)
spgPageIndexMultiDelete(&state, srcPage,
toDelete, xldata->nDelete,
SPGIST_REDIRECT,
SPGIST_PLACEHOLDER,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoInner,
xldata->offnumInner);
else
spgPageIndexMultiDelete(&state, srcPage,
toDelete, xldata->nDelete,
SPGIST_PLACEHOLDER,
SPGIST_PLACEHOLDER,
InvalidBlockNumber,
InvalidOffsetNumber);
/* don't update LSN etc till we're done with it */
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
}
}
/* try to access dest page if any */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (!XLogRecHasBlockRef(record, 1))
{
destBuffer = InvalidBuffer;
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
destPage = NULL;
}
else if (xldata->initDest)
{
/* just re-init the dest page */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
destBuffer = XLogInitBufferForRedo(record, 1);
destPage = (Page) BufferGetPage(destBuffer);
SpGistInitBuffer(destBuffer,
SPGIST_LEAF | (xldata->storesNulls ? SPGIST_NULLS : 0));
/* don't update LSN etc till we're done with it */
}
else
{
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* We could probably release the page lock immediately in the
* full-page-image case, but for safety let's hold it till later.
*/
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 1, &destBuffer) == BLK_NEEDS_REDO)
destPage = (Page) BufferGetPage(destBuffer);
else
destPage = NULL; /* don't do any page updates */
}
/* restore leaf tuples to src and/or dest page */
for (i = 0; i < xldata->nInsert; i++)
{
char *leafTuple;
SpGistLeafTupleData leafTupleHdr;
/* the tuples are not aligned, so must copy to access the size field. */
leafTuple = ptr;
memcpy(&leafTupleHdr, leafTuple, sizeof(SpGistLeafTupleData));
ptr += leafTupleHdr.size;
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
page = leafPageSelect[i] ? destPage : srcPage;
if (page == NULL)
continue; /* no need to touch this page */
addOrReplaceTuple(page, (Item) leafTuple, leafTupleHdr.size,
toInsert[i]);
}
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/* Now update src and dest page LSNs if needed */
if (srcPage != NULL)
{
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
PageSetLSN(srcPage, lsn);
MarkBufferDirty(srcBuffer);
}
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
if (destPage != NULL)
{
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
PageSetLSN(destPage, lsn);
MarkBufferDirty(destBuffer);
}
/* restore new inner tuple */
if (xldata->initInner)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
innerBuffer = XLogInitBufferForRedo(record, 2);
SpGistInitBuffer(innerBuffer, (xldata->storesNulls ? SPGIST_NULLS : 0));
action = BLK_NEEDS_REDO;
}
else
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
action = XLogReadBufferForRedo(record, 2, &innerBuffer);
if (action == BLK_NEEDS_REDO)
{
page = BufferGetPage(innerBuffer);
addOrReplaceTuple(page, (Item) innerTuple, innerTupleHdr.size,
xldata->offnumInner);
/* if inner is also parent, update link while we're here */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (xldata->innerIsParent)
{
SpGistInnerTuple parent;
parent = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
spgUpdateNodeLink(parent, xldata->nodeI,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoInner, xldata->offnumInner);
}
PageSetLSN(page, lsn);
MarkBufferDirty(innerBuffer);
}
if (BufferIsValid(innerBuffer))
UnlockReleaseBuffer(innerBuffer);
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* Now we can release the leaf-page locks. It's okay to do this before
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
* updating the parent downlink.
*/
if (BufferIsValid(srcBuffer))
UnlockReleaseBuffer(srcBuffer);
if (BufferIsValid(destBuffer))
UnlockReleaseBuffer(destBuffer);
/* update parent downlink, unless we did it above */
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogRecHasBlockRef(record, 3))
{
Buffer parentBuffer;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 3, &parentBuffer) == BLK_NEEDS_REDO)
{
SpGistInnerTuple parent;
page = BufferGetPage(parentBuffer);
parent = (SpGistInnerTuple) PageGetItem(page,
PageGetItemId(page, xldata->offnumParent));
spgUpdateNodeLink(parent, xldata->nodeI,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
blknoInner, xldata->offnumInner);
PageSetLSN(page, lsn);
MarkBufferDirty(parentBuffer);
}
if (BufferIsValid(parentBuffer))
UnlockReleaseBuffer(parentBuffer);
}
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
else
Assert(xldata->innerIsParent || xldata->isRootSplit);
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoVacuumLeaf(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogVacuumLeaf *xldata = (spgxlogVacuumLeaf *) ptr;
OffsetNumber *toDead;
OffsetNumber *toPlaceholder;
OffsetNumber *moveSrc;
OffsetNumber *moveDest;
OffsetNumber *chainSrc;
OffsetNumber *chainDest;
SpGistState state;
Buffer buffer;
Page page;
int i;
fillFakeState(&state, xldata->stateSrc);
ptr += SizeOfSpgxlogVacuumLeaf;
toDead = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nDead;
toPlaceholder = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nPlaceholder;
moveSrc = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nMove;
moveDest = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nMove;
chainSrc = (OffsetNumber *) ptr;
ptr += sizeof(OffsetNumber) * xldata->nChain;
chainDest = (OffsetNumber *) ptr;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
spgPageIndexMultiDelete(&state, page,
toDead, xldata->nDead,
SPGIST_DEAD, SPGIST_DEAD,
InvalidBlockNumber,
InvalidOffsetNumber);
spgPageIndexMultiDelete(&state, page,
toPlaceholder, xldata->nPlaceholder,
SPGIST_PLACEHOLDER, SPGIST_PLACEHOLDER,
InvalidBlockNumber,
InvalidOffsetNumber);
/* see comments in vacuumLeafPage() */
for (i = 0; i < xldata->nMove; i++)
{
ItemId idSrc = PageGetItemId(page, moveSrc[i]);
ItemId idDest = PageGetItemId(page, moveDest[i]);
ItemIdData tmp;
tmp = *idSrc;
*idSrc = *idDest;
*idDest = tmp;
}
spgPageIndexMultiDelete(&state, page,
moveSrc, xldata->nMove,
SPGIST_PLACEHOLDER, SPGIST_PLACEHOLDER,
InvalidBlockNumber,
InvalidOffsetNumber);
for (i = 0; i < xldata->nChain; i++)
{
SpGistLeafTuple lt;
lt = (SpGistLeafTuple) PageGetItem(page,
PageGetItemId(page, chainSrc[i]));
Assert(lt->tupstate == SPGIST_LIVE);
lt->nextOffset = chainDest[i];
}
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoVacuumRoot(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogVacuumRoot *xldata = (spgxlogVacuumRoot *) ptr;
OffsetNumber *toDelete;
Buffer buffer;
Page page;
toDelete = xldata->offsets;
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
page = BufferGetPage(buffer);
/* The tuple numbers are in order */
PageIndexMultiDelete(page, toDelete, xldata->nDelete);
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
static void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoVacuumRedirect(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
XLogRecPtr lsn = record->EndRecPtr;
char *ptr = XLogRecGetData(record);
spgxlogVacuumRedirect *xldata = (spgxlogVacuumRedirect *) ptr;
OffsetNumber *itemToPlaceholder;
Buffer buffer;
itemToPlaceholder = xldata->offsets;
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
/*
* If any redirection tuples are being removed, make sure there are no
* live Hot Standby transactions that might need to see them.
*/
if (InHotStandby)
{
if (TransactionIdIsValid(xldata->newestRedirectXid))
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
{
RelFileNode node;
XLogRecGetBlockTag(record, 0, &node, NULL, NULL);
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
ResolveRecoveryConflictWithSnapshot(xldata->newestRedirectXid,
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
node);
}
Fix multiple problems in WAL replay. Most of the replay functions for WAL record types that modify more than one page failed to ensure that those pages were locked correctly to ensure that concurrent queries could not see inconsistent page states. This is a hangover from coding decisions made long before Hot Standby was added, when it was hardly necessary to acquire buffer locks during WAL replay at all, let alone hold them for carefully-chosen periods. The key problem was that RestoreBkpBlocks was written to hold lock on each page restored from a full-page image for only as long as it took to update that page. This was guaranteed to break any WAL replay function in which there was any update-ordering constraint between pages, because even if the nominal order of the pages is the right one, any mixture of full-page and non-full-page updates in the same record would result in out-of-order updates. Moreover, it wouldn't work for situations where there's a requirement to maintain lock on one page while updating another. Failure to honor an update ordering constraint in this way is thought to be the cause of bug #7648 from Daniel Farina: what seems to have happened there is that a btree page being split was rewritten from a full-page image before the new right sibling page was written, and because lock on the original page was not maintained it was possible for hot standby queries to try to traverse the page's right-link to the not-yet-existing sibling page. To fix, get rid of RestoreBkpBlocks as such, and instead create a new function RestoreBackupBlock that restores just one full-page image at a time. This function can be invoked by WAL replay functions at the points where they would otherwise perform non-full-page updates; in this way, the physical order of page updates remains the same no matter which pages are replaced by full-page images. We can then further adjust the logic in individual replay functions if it is necessary to hold buffer locks for overlapping periods. A side benefit is that we can simplify the handling of concurrency conflict resolution by moving that code into the record-type-specfic functions; there's no more need to contort the code layout to keep conflict resolution in front of the RestoreBkpBlocks call. In connection with that, standardize on zero-based numbering rather than one-based numbering for referencing the full-page images. In HEAD, I removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are still there in the header files in previous branches, but are no longer used by the code. In addition, fix some other bugs identified in the course of making these changes: spgRedoAddNode could fail to update the parent downlink at all, if the parent tuple is in the same page as either the old or new split tuple and we're not doing a full-page image: it would get fooled by the LSN having been advanced already. This would result in permanent index corruption, not just transient failure of concurrent queries. Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old tail page as a candidate for a full-page image; in the worst case this could result in torn-page corruption. heap_xlog_freeze() was inconsistent about using a cleanup lock or plain exclusive lock: it did the former in the normal path but the latter for a full-page image. A plain exclusive lock seems sufficient, so change to that. Also, remove gistRedoPageDeleteRecord(), which has been dead code since VACUUM FULL was rewritten. Back-patch to 9.0, where hot standby was introduced. Note however that 9.0 had a significantly different WAL-logging scheme for GIST index updates, and it doesn't appear possible to make that scheme safe for concurrent hot standby queries, because it can leave inconsistent states in the index even between WAL records. Given the lack of complaints from the field, we won't work too hard on fixing that branch.
2012-11-13 04:05:08 +01:00
}
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
{
Page page = BufferGetPage(buffer);
SpGistPageOpaque opaque = SpGistPageGetOpaque(page);
int i;
/* Convert redirect pointers to plain placeholders */
for (i = 0; i < xldata->nToPlaceholder; i++)
{
SpGistDeadTuple dt;
dt = (SpGistDeadTuple) PageGetItem(page,
PageGetItemId(page, itemToPlaceholder[i]));
Assert(dt->tupstate == SPGIST_REDIRECT);
dt->tupstate = SPGIST_PLACEHOLDER;
ItemPointerSetInvalid(&dt->pointer);
}
Assert(opaque->nRedirection >= xldata->nToPlaceholder);
opaque->nRedirection -= xldata->nToPlaceholder;
opaque->nPlaceholder += xldata->nToPlaceholder;
/* Remove placeholder tuples at end of page */
if (xldata->firstPlaceholder != InvalidOffsetNumber)
{
int max = PageGetMaxOffsetNumber(page);
OffsetNumber *toDelete;
toDelete = palloc(sizeof(OffsetNumber) * max);
for (i = xldata->firstPlaceholder; i <= max; i++)
toDelete[i - xldata->firstPlaceholder] = i;
i = max - xldata->firstPlaceholder + 1;
Assert(opaque->nPlaceholder >= i);
opaque->nPlaceholder -= i;
/* The array is sorted, so can use PageIndexMultiDelete */
PageIndexMultiDelete(page, toDelete, i);
pfree(toDelete);
}
PageSetLSN(page, lsn);
MarkBufferDirty(buffer);
}
if (BufferIsValid(buffer))
UnlockReleaseBuffer(buffer);
}
void
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spg_redo(XLogReaderState *record)
{
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
uint8 info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
MemoryContext oldCxt;
oldCxt = MemoryContextSwitchTo(opCtx);
switch (info)
{
case XLOG_SPGIST_ADD_LEAF:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoAddLeaf(record);
break;
case XLOG_SPGIST_MOVE_LEAFS:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoMoveLeafs(record);
break;
case XLOG_SPGIST_ADD_NODE:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoAddNode(record);
break;
case XLOG_SPGIST_SPLIT_TUPLE:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoSplitTuple(record);
break;
case XLOG_SPGIST_PICKSPLIT:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoPickSplit(record);
break;
case XLOG_SPGIST_VACUUM_LEAF:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoVacuumLeaf(record);
break;
case XLOG_SPGIST_VACUUM_ROOT:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoVacuumRoot(record);
break;
case XLOG_SPGIST_VACUUM_REDIRECT:
Revamp the WAL record format. Each WAL record now carries information about the modified relation and block(s) in a standardized format. That makes it easier to write tools that need that information, like pg_rewind, prefetching the blocks to speed up recovery, etc. There's a whole new API for building WAL records, replacing the XLogRecData chains used previously. The new API consists of XLogRegister* functions, which are called for each buffer and chunk of data that is added to the record. The new API also gives more control over when a full-page image is written, by passing flags to the XLogRegisterBuffer function. This also simplifies the XLogReadBufferForRedo() calls. The function can dig the relation and block number from the WAL record, so they no longer need to be passed as arguments. For the convenience of redo routines, XLogReader now disects each WAL record after reading it, copying the main data part and the per-block data into MAXALIGNed buffers. The data chunks are not aligned within the WAL record, but the redo routines can assume that the pointers returned by XLogRecGet* functions are. Redo routines are now passed the XLogReaderState, which contains the record in the already-disected format, instead of the plain XLogRecord. The new record format also makes the fixed size XLogRecord header smaller, by removing the xl_len field. The length of the "main data" portion is now stored at the end of the WAL record, and there's a separate header after XLogRecord for it. The alignment padding at the end of XLogRecord is also removed. This compansates for the fact that the new format would otherwise be more bulky than the old format. Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera, Fujii Masao.
2014-11-20 16:56:26 +01:00
spgRedoVacuumRedirect(record);
break;
default:
elog(PANIC, "spg_redo: unknown op code %u", info);
}
MemoryContextSwitchTo(oldCxt);
MemoryContextReset(opCtx);
}
void
spg_xlog_startup(void)
{
opCtx = AllocSetContextCreate(CurrentMemoryContext,
"SP-GiST temporary context",
Add macros to make AllocSetContextCreate() calls simpler and safer. I found that half a dozen (nearly 5%) of our AllocSetContextCreate calls had typos in the context-sizing parameters. While none of these led to especially significant problems, they did create minor inefficiencies, and it's now clear that expecting people to copy-and-paste those calls accurately is not a great idea. Let's reduce the risk of future errors by introducing single macros that encapsulate the common use-cases. Three such macros are enough to cover all but two special-purpose contexts; those two calls can be left as-is, I think. While this patch doesn't in itself improve matters for third-party extensions, it doesn't break anything for them either, and they can gradually adopt the simplified notation over time. In passing, change TopMemoryContext to use the default allocation parameters. Formerly it could only be extended 8K at a time. That was probably reasonable when this code was written; but nowadays we create many more contexts than we did then, so that it's not unusual to have a couple hundred K in TopMemoryContext, even without considering various dubious code that sticks other things there. There seems no good reason not to let it use growing blocks like most other contexts. Back-patch to 9.6, mostly because that's still close enough to HEAD that it's easy to do so, and keeping the branches in sync can be expected to avoid some future back-patching pain. The bugs fixed by these changes don't seem to be significant enough to justify fixing them further back. Discussion: <21072.1472321324@sss.pgh.pa.us>
2016-08-27 23:50:38 +02:00
ALLOCSET_DEFAULT_SIZES);
}
void
spg_xlog_cleanup(void)
{
MemoryContextDelete(opCtx);
opCtx = NULL;
}
/*
* Mask a SpGist page before performing consistency checks on it.
*/
void
spg_mask(char *pagedata, BlockNumber blkno)
{
Page page = (Page) pagedata;
PageHeader pagehdr = (PageHeader) page;
mask_page_lsn_and_checksum(page);
mask_page_hint_bits(page);
/*
* Mask the unused space, but only if the page's pd_lower appears to have
* been set correctly.
*/
if (pagehdr->pd_lower > SizeOfPageHeaderData)
mask_unused_space(page);
}