postgresql/src/include/access/ginblock.h

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

347 lines
13 KiB
C
Raw Normal View History

/*--------------------------------------------------------------------------
* ginblock.h
* details of structures stored in GIN index blocks
*
* Copyright (c) 2006-2022, PostgreSQL Global Development Group
*
* src/include/access/ginblock.h
*--------------------------------------------------------------------------
*/
#ifndef GINBLOCK_H
#define GINBLOCK_H
#include "access/transam.h"
#include "storage/block.h"
snapshot scalability: Don't compute global horizons while building snapshots. To make GetSnapshotData() more scalable, it cannot not look at at each proc's xmin: While snapshot contents do not need to change whenever a read-only transaction commits or a snapshot is released, a proc's xmin is modified in those cases. The frequency of xmin modifications leads to, particularly on higher core count systems, many cache misses inside GetSnapshotData(), despite the data underlying a snapshot not changing. That is the most significant source of GetSnapshotData() scaling poorly on larger systems. Without accessing xmins, GetSnapshotData() cannot calculate accurate horizons / thresholds as it has so far. But we don't really have to: The horizons don't actually change that much between GetSnapshotData() calls. Nor are the horizons actually used every time a snapshot is built. The trick this commit introduces is to delay computation of accurate horizons until there use and using horizon boundaries to determine whether accurate horizons need to be computed. The use of RecentGlobal[Data]Xmin to decide whether a row version could be removed has been replaces with new GlobalVisTest* functions. These use two thresholds to determine whether a row can be pruned: 1) definitely_needed, indicating that rows deleted by XIDs >= definitely_needed are definitely still visible. 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can definitely be removed GetSnapshotData() updates definitely_needed to be the xmin of the computed snapshot. When testing whether a row can be removed (with GlobalVisTestIsRemovableXid()) and the tested XID falls in between the two (i.e. XID >= maybe_needed && XID < definitely_needed) the boundaries can be recomputed to be more accurate. As it is not cheap to compute accurate boundaries, we limit the number of times that happens in short succession. As the boundaries used by GlobalVisTestIsRemovableXid() are never reset (with maybe_needed updated by GetSnapshotData()), it is likely that further test can benefit from an earlier computation of accurate horizons. To avoid regressing performance when old_snapshot_threshold is set (as that requires an accurate horizon to be computed), heap_page_prune_opt() doesn't unconditionally call TransactionIdLimitedForOldSnapshots() anymore. Both the computation of the limited horizon, and the triggering of errors (with SetOldSnapshotThresholdTimestamp()) is now only done when necessary to remove tuples. This commit just removes the accesses to PGXACT->xmin from GetSnapshotData(), but other members of PGXACT residing in the same cache line are accessed. Therefore this in itself does not result in a significant improvement. Subsequent commits will take advantage of the fact that GetSnapshotData() now does not need to access xmins anymore. Note: This contains a workaround in heap_page_prune_opt() to keep the snapshot_too_old tests working. While that workaround is ugly, the tests currently are not meaningful, and it seems best to address them separately. Author: Andres Freund <andres@anarazel.de> Reviewed-By: Robert Haas <robertmhaas@gmail.com> Reviewed-By: Thomas Munro <thomas.munro@gmail.com> Reviewed-By: David Rowley <dgrowleyml@gmail.com> Discussion: https://postgr.es/m/20200301083601.ews6hz5dduc3w2se@alap3.anarazel.de
2020-08-13 01:03:49 +02:00
#include "storage/bufpage.h"
#include "storage/itemptr.h"
#include "storage/off.h"
/*
* Page opaque data in an inverted index page.
*
* Note: GIN does not include a page ID word as do the other index types.
* This is OK because the opaque data is only 8 bytes and so can be reliably
* distinguished by size. Revisit this if the size ever increases.
* Further note: as of 9.2, SP-GiST also uses 8-byte special space, as does
* BRIN as of 9.5. This is still OK, as long as GIN isn't using all of the
* high-order bits in its flags word, because that way the flags word cannot
* match the page IDs used by SP-GiST and BRIN.
*/
typedef struct GinPageOpaqueData
{
BlockNumber rightlink; /* next page if any */
OffsetNumber maxoff; /* number of PostingItems on GIN_DATA &
* ~GIN_LEAF page. On GIN_LIST page, number of
* heap tuples. */
uint16 flags; /* see bit definitions below */
} GinPageOpaqueData;
typedef GinPageOpaqueData *GinPageOpaque;
#define GIN_DATA (1 << 0)
#define GIN_LEAF (1 << 1)
#define GIN_DELETED (1 << 2)
#define GIN_META (1 << 3)
#define GIN_LIST (1 << 4)
#define GIN_LIST_FULLROW (1 << 5) /* makes sense only on GIN_LIST page */
#define GIN_INCOMPLETE_SPLIT (1 << 6) /* page was split, but parent not
* updated */
#define GIN_COMPRESSED (1 << 7)
/* Page numbers of fixed-location pages */
#define GIN_METAPAGE_BLKNO (0)
#define GIN_ROOT_BLKNO (1)
typedef struct GinMetaPageData
{
/*
* Pointers to head and tail of pending list, which consists of GIN_LIST
* pages. These store fast-inserted entries that haven't yet been moved
* into the regular GIN structure.
*/
BlockNumber head;
BlockNumber tail;
/*
* Free space in bytes in the pending list's tail page.
*/
uint32 tailFreeSize;
/*
* We store both number of pages and number of heap tuples that are in the
* pending list.
*/
BlockNumber nPendingPages;
int64 nPendingHeapTuples;
/*
* Statistics for planner use (accurate as of last VACUUM)
*/
BlockNumber nTotalPages;
BlockNumber nEntryPages;
BlockNumber nDataPages;
int64 nEntries;
/*
* GIN version number (ideally this should have been at the front, but too
* late now. Don't move it!)
*
* Currently 2 (for indexes initialized in 9.4 or later)
*
* Version 1 (indexes initialized in version 9.1, 9.2 or 9.3), is
* compatible, but may contain uncompressed posting tree (leaf) pages and
* posting lists. They will be converted to compressed format when
* modified.
*
* Version 0 (indexes initialized in 9.0 or before) is compatible but may
* be missing null entries, including both null keys and placeholders.
* Reject full-index-scan attempts on such indexes.
*/
int32 ginVersion;
} GinMetaPageData;
#define GIN_CURRENT_VERSION 2
#define GinPageGetMeta(p) \
((GinMetaPageData *) PageGetContents(p))
/*
* Macros for accessing a GIN index page's opaque data
*/
#define GinPageGetOpaque(page) ( (GinPageOpaque) PageGetSpecialPointer(page) )
#define GinPageIsLeaf(page) ( (GinPageGetOpaque(page)->flags & GIN_LEAF) != 0 )
#define GinPageSetLeaf(page) ( GinPageGetOpaque(page)->flags |= GIN_LEAF )
#define GinPageSetNonLeaf(page) ( GinPageGetOpaque(page)->flags &= ~GIN_LEAF )
#define GinPageIsData(page) ( (GinPageGetOpaque(page)->flags & GIN_DATA) != 0 )
#define GinPageSetData(page) ( GinPageGetOpaque(page)->flags |= GIN_DATA )
#define GinPageIsList(page) ( (GinPageGetOpaque(page)->flags & GIN_LIST) != 0 )
#define GinPageSetList(page) ( GinPageGetOpaque(page)->flags |= GIN_LIST )
#define GinPageHasFullRow(page) ( (GinPageGetOpaque(page)->flags & GIN_LIST_FULLROW) != 0 )
#define GinPageSetFullRow(page) ( GinPageGetOpaque(page)->flags |= GIN_LIST_FULLROW )
#define GinPageIsCompressed(page) ( (GinPageGetOpaque(page)->flags & GIN_COMPRESSED) != 0 )
#define GinPageSetCompressed(page) ( GinPageGetOpaque(page)->flags |= GIN_COMPRESSED )
#define GinPageIsDeleted(page) ( (GinPageGetOpaque(page)->flags & GIN_DELETED) != 0 )
#define GinPageSetDeleted(page) ( GinPageGetOpaque(page)->flags |= GIN_DELETED)
#define GinPageSetNonDeleted(page) ( GinPageGetOpaque(page)->flags &= ~GIN_DELETED)
#define GinPageIsIncompleteSplit(page) ( (GinPageGetOpaque(page)->flags & GIN_INCOMPLETE_SPLIT) != 0 )
#define GinPageRightMost(page) ( GinPageGetOpaque(page)->rightlink == InvalidBlockNumber)
/*
* We should reclaim deleted page only once every transaction started before
* its deletion is over.
*/
#define GinPageGetDeleteXid(page) ( ((PageHeader) (page))->pd_prune_xid )
#define GinPageSetDeleteXid(page, xid) ( ((PageHeader) (page))->pd_prune_xid = xid)
snapshot scalability: Don't compute global horizons while building snapshots. To make GetSnapshotData() more scalable, it cannot not look at at each proc's xmin: While snapshot contents do not need to change whenever a read-only transaction commits or a snapshot is released, a proc's xmin is modified in those cases. The frequency of xmin modifications leads to, particularly on higher core count systems, many cache misses inside GetSnapshotData(), despite the data underlying a snapshot not changing. That is the most significant source of GetSnapshotData() scaling poorly on larger systems. Without accessing xmins, GetSnapshotData() cannot calculate accurate horizons / thresholds as it has so far. But we don't really have to: The horizons don't actually change that much between GetSnapshotData() calls. Nor are the horizons actually used every time a snapshot is built. The trick this commit introduces is to delay computation of accurate horizons until there use and using horizon boundaries to determine whether accurate horizons need to be computed. The use of RecentGlobal[Data]Xmin to decide whether a row version could be removed has been replaces with new GlobalVisTest* functions. These use two thresholds to determine whether a row can be pruned: 1) definitely_needed, indicating that rows deleted by XIDs >= definitely_needed are definitely still visible. 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can definitely be removed GetSnapshotData() updates definitely_needed to be the xmin of the computed snapshot. When testing whether a row can be removed (with GlobalVisTestIsRemovableXid()) and the tested XID falls in between the two (i.e. XID >= maybe_needed && XID < definitely_needed) the boundaries can be recomputed to be more accurate. As it is not cheap to compute accurate boundaries, we limit the number of times that happens in short succession. As the boundaries used by GlobalVisTestIsRemovableXid() are never reset (with maybe_needed updated by GetSnapshotData()), it is likely that further test can benefit from an earlier computation of accurate horizons. To avoid regressing performance when old_snapshot_threshold is set (as that requires an accurate horizon to be computed), heap_page_prune_opt() doesn't unconditionally call TransactionIdLimitedForOldSnapshots() anymore. Both the computation of the limited horizon, and the triggering of errors (with SetOldSnapshotThresholdTimestamp()) is now only done when necessary to remove tuples. This commit just removes the accesses to PGXACT->xmin from GetSnapshotData(), but other members of PGXACT residing in the same cache line are accessed. Therefore this in itself does not result in a significant improvement. Subsequent commits will take advantage of the fact that GetSnapshotData() now does not need to access xmins anymore. Note: This contains a workaround in heap_page_prune_opt() to keep the snapshot_too_old tests working. While that workaround is ugly, the tests currently are not meaningful, and it seems best to address them separately. Author: Andres Freund <andres@anarazel.de> Reviewed-By: Robert Haas <robertmhaas@gmail.com> Reviewed-By: Thomas Munro <thomas.munro@gmail.com> Reviewed-By: David Rowley <dgrowleyml@gmail.com> Discussion: https://postgr.es/m/20200301083601.ews6hz5dduc3w2se@alap3.anarazel.de
2020-08-13 01:03:49 +02:00
extern bool GinPageIsRecyclable(Page page);
/*
* We use our own ItemPointerGet(BlockNumber|OffsetNumber)
* to avoid Asserts, since sometimes the ip_posid isn't "valid"
*/
#define GinItemPointerGetBlockNumber(pointer) \
(ItemPointerGetBlockNumberNoCheck(pointer))
#define GinItemPointerGetOffsetNumber(pointer) \
(ItemPointerGetOffsetNumberNoCheck(pointer))
#define GinItemPointerSetBlockNumber(pointer, blkno) \
(ItemPointerSetBlockNumber((pointer), (blkno)))
#define GinItemPointerSetOffsetNumber(pointer, offnum) \
(ItemPointerSetOffsetNumber((pointer), (offnum)))
/*
* Special-case item pointer values needed by the GIN search logic.
* MIN: sorts less than any valid item pointer
* MAX: sorts greater than any valid item pointer
* LOSSY PAGE: indicates a whole heap page, sorts after normal item
* pointers for that page
* Note that these are all distinguishable from an "invalid" item pointer
* (which is InvalidBlockNumber/0) as well as from all normal item
* pointers (which have item numbers in the range 1..MaxHeapTuplesPerPage).
*/
#define ItemPointerSetMin(p) \
ItemPointerSet((p), (BlockNumber)0, (OffsetNumber)0)
#define ItemPointerIsMin(p) \
(GinItemPointerGetOffsetNumber(p) == (OffsetNumber)0 && \
GinItemPointerGetBlockNumber(p) == (BlockNumber)0)
#define ItemPointerSetMax(p) \
ItemPointerSet((p), InvalidBlockNumber, (OffsetNumber)0xffff)
#define ItemPointerSetLossyPage(p, b) \
ItemPointerSet((p), (b), (OffsetNumber)0xffff)
#define ItemPointerIsLossyPage(p) \
(GinItemPointerGetOffsetNumber(p) == (OffsetNumber)0xffff && \
GinItemPointerGetBlockNumber(p) != InvalidBlockNumber)
/*
* Posting item in a non-leaf posting-tree page
*/
typedef struct
{
/* We use BlockIdData not BlockNumber to avoid padding space wastage */
BlockIdData child_blkno;
ItemPointerData key;
} PostingItem;
#define PostingItemGetBlockNumber(pointer) \
BlockIdGetBlockNumber(&(pointer)->child_blkno)
#define PostingItemSetBlockNumber(pointer, blockNumber) \
BlockIdSet(&((pointer)->child_blkno), (blockNumber))
/*
* Category codes to distinguish placeholder nulls from ordinary NULL keys.
*
* The first two code values were chosen to be compatible with the usual usage
* of bool isNull flags. However, casting between bool and GinNullCategory is
* risky because of the possibility of different bit patterns and type sizes,
* so it is no longer done.
*
* GIN_CAT_EMPTY_QUERY is never stored in the index; and notice that it is
* chosen to sort before not after regular key values.
*/
typedef signed char GinNullCategory;
#define GIN_CAT_NORM_KEY 0 /* normal, non-null key value */
#define GIN_CAT_NULL_KEY 1 /* null key value */
#define GIN_CAT_EMPTY_ITEM 2 /* placeholder for zero-key item */
#define GIN_CAT_NULL_ITEM 3 /* placeholder for null item */
#define GIN_CAT_EMPTY_QUERY (-1) /* placeholder for full-scan query */
/*
* Access macros for null category byte in entry tuples
*/
#define GinCategoryOffset(itup,ginstate) \
(IndexInfoFindDataOffset((itup)->t_info) + \
((ginstate)->oneCol ? 0 : sizeof(int16)))
#define GinGetNullCategory(itup,ginstate) \
(*((GinNullCategory *) ((char*)(itup) + GinCategoryOffset(itup,ginstate))))
#define GinSetNullCategory(itup,ginstate,c) \
(*((GinNullCategory *) ((char*)(itup) + GinCategoryOffset(itup,ginstate))) = (c))
/*
* Access macros for leaf-page entry tuples (see discussion in README)
*/
#define GinGetNPosting(itup) GinItemPointerGetOffsetNumber(&(itup)->t_tid)
#define GinSetNPosting(itup,n) ItemPointerSetOffsetNumber(&(itup)->t_tid,n)
#define GIN_TREE_POSTING ((OffsetNumber)0xffff)
#define GinIsPostingTree(itup) (GinGetNPosting(itup) == GIN_TREE_POSTING)
#define GinSetPostingTree(itup, blkno) ( GinSetNPosting((itup),GIN_TREE_POSTING), ItemPointerSetBlockNumber(&(itup)->t_tid, blkno) )
#define GinGetPostingTree(itup) GinItemPointerGetBlockNumber(&(itup)->t_tid)
#define GIN_ITUP_COMPRESSED (1U << 31)
#define GinGetPostingOffset(itup) (GinItemPointerGetBlockNumber(&(itup)->t_tid) & (~GIN_ITUP_COMPRESSED))
#define GinSetPostingOffset(itup,n) ItemPointerSetBlockNumber(&(itup)->t_tid,(n)|GIN_ITUP_COMPRESSED)
#define GinGetPosting(itup) ((Pointer) ((char*)(itup) + GinGetPostingOffset(itup)))
#define GinItupIsCompressed(itup) ((GinItemPointerGetBlockNumber(&(itup)->t_tid) & GIN_ITUP_COMPRESSED) != 0)
/*
* Maximum size of an item on entry tree page. Make sure that we fit at least
* three items on each page. (On regular B-tree indexes, we must fit at least
* three items: two data items and the "high key". In GIN entry tree, we don't
* currently store the high key explicitly, we just use the rightmost item on
* the page, so it would actually be enough to fit two items.)
*/
#define GinMaxItemSize \
Min(INDEX_SIZE_MASK, \
MAXALIGN_DOWN(((BLCKSZ - \
MAXALIGN(SizeOfPageHeaderData + 3 * sizeof(ItemIdData)) - \
MAXALIGN(sizeof(GinPageOpaqueData))) / 3)))
/*
* Access macros for non-leaf entry tuples
*/
#define GinGetDownlink(itup) GinItemPointerGetBlockNumber(&(itup)->t_tid)
#define GinSetDownlink(itup,blkno) ItemPointerSet(&(itup)->t_tid, blkno, InvalidOffsetNumber)
/*
* Data (posting tree) pages
*
* Posting tree pages don't store regular tuples. Non-leaf pages contain
* PostingItems, which are pairs of ItemPointers and child block numbers.
* Leaf pages contain GinPostingLists and an uncompressed array of item
* pointers.
*
* In a leaf page, the compressed posting lists are stored after the regular
* page header, one after each other. Although we don't store regular tuples,
* pd_lower is used to indicate the end of the posting lists. After that, free
* space follows. This layout is compatible with the "standard" heap and
* index page layout described in bufpage.h, so that we can e.g set buffer_std
* when writing WAL records.
*
* In the special space is the GinPageOpaque struct.
*/
#define GinDataLeafPageGetPostingList(page) \
(GinPostingList *) ((PageGetContents(page) + MAXALIGN(sizeof(ItemPointerData))))
#define GinDataLeafPageGetPostingListSize(page) \
(((PageHeader) page)->pd_lower - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(ItemPointerData)))
#define GinDataLeafPageIsEmpty(page) \
(GinPageIsCompressed(page) ? (GinDataLeafPageGetPostingListSize(page) == 0) : (GinPageGetOpaque(page)->maxoff < FirstOffsetNumber))
#define GinDataLeafPageGetFreeSpace(page) PageGetExactFreeSpace(page)
#define GinDataPageGetRightBound(page) ((ItemPointer) PageGetContents(page))
/*
* Pointer to the data portion of a posting tree page. For internal pages,
* that's the beginning of the array of PostingItems. For compressed leaf
* pages, the first compressed posting list. For uncompressed (pre-9.4) leaf
* pages, it's the beginning of the ItemPointer array.
*/
#define GinDataPageGetData(page) \
(PageGetContents(page) + MAXALIGN(sizeof(ItemPointerData)))
/* non-leaf pages contain PostingItems */
#define GinDataPageGetPostingItem(page, i) \
((PostingItem *) (GinDataPageGetData(page) + ((i)-1) * sizeof(PostingItem)))
/*
* Note: there is no GinDataPageGetDataSize macro, because before version
* 9.4, we didn't set pd_lower on data pages. There can be pages in the index
* that were binary-upgraded from earlier versions and still have an invalid
* pd_lower, so we cannot trust it in general. Compressed posting tree leaf
* pages are new in 9.4, however, so we can trust them; see
* GinDataLeafPageGetPostingListSize.
*/
#define GinDataPageSetDataSize(page, size) \
{ \
Assert(size <= GinDataPageMaxDataSize); \
((PageHeader) page)->pd_lower = (size) + MAXALIGN(SizeOfPageHeaderData) + MAXALIGN(sizeof(ItemPointerData)); \
}
#define GinNonLeafDataPageGetFreeSpace(page) \
(GinDataPageMaxDataSize - \
GinPageGetOpaque(page)->maxoff * sizeof(PostingItem))
#define GinDataPageMaxDataSize \
(BLCKSZ - MAXALIGN(SizeOfPageHeaderData) \
- MAXALIGN(sizeof(ItemPointerData)) \
- MAXALIGN(sizeof(GinPageOpaqueData)))
/*
* List pages
*/
#define GinListPageSize \
( BLCKSZ - SizeOfPageHeaderData - MAXALIGN(sizeof(GinPageOpaqueData)) )
/*
* A compressed posting list.
*
* Note: This requires 2-byte alignment.
*/
typedef struct
{
ItemPointerData first; /* first item in this posting list (unpacked) */
uint16 nbytes; /* number of bytes that follow */
unsigned char bytes[FLEXIBLE_ARRAY_MEMBER]; /* varbyte encoded items */
} GinPostingList;
#define SizeOfGinPostingList(plist) (offsetof(GinPostingList, bytes) + SHORTALIGN((plist)->nbytes) )
#define GinNextPostingListSegment(cur) ((GinPostingList *) (((char *) (cur)) + SizeOfGinPostingList((cur))))
#endif /* GINBLOCK_H */