/*-------------------------------------------------------------------------
*
- * hashpage.c--
- * Hash table page management code for the Postgres hash access method
+ * hashpage.c
+ * Hash table page management code for the Postgres hash access method
*
- * Copyright (c) 1994, Regents of the University of California
+ * Portions Copyright (c) 1996-2006, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
*
*
* IDENTIFICATION
- * $Header: /cvsroot/pgsql/src/backend/access/hash/hashpage.c,v 1.2 1996/10/20 06:34:21 scrappy Exp $
+ * $PostgreSQL: pgsql/src/backend/access/hash/hashpage.c,v 1.58 2006/07/02 02:23:18 momjian Exp $
*
* NOTES
- * Postgres hash pages look like ordinary relation pages. The opaque
- * data at high addresses includes information about the page including
- * whether a page is an overflow page or a true bucket, the block
- * numbers of the preceding and following pages, and the overflow
- * address of the page if it is an overflow page.
+ * Postgres hash pages look like ordinary relation pages. The opaque
+ * data at high addresses includes information about the page including
+ * whether a page is an overflow page or a true bucket, the bucket
+ * number, and the block numbers of the preceding and following pages
+ * in the same bucket.
*
- * The first page in a hash relation, page zero, is special -- it stores
- * information describing the hash table; it is referred to as teh
- * "meta page." Pages one and higher store the actual data.
+ * The first page in a hash relation, page zero, is special -- it stores
+ * information describing the hash table; it is referred to as the
+ * "meta page." Pages one and higher store the actual data.
+ *
+ * There are also bitmap pages, which are not manipulated here;
+ * see hashovfl.c.
*
*-------------------------------------------------------------------------
*/
-
#include "postgres.h"
-#include "utils/rel.h"
+
+#include "access/genam.h"
#include "access/hash.h"
-#include "storage/bufmgr.h"
-
-static void _hash_setpagelock(Relation rel, BlockNumber blkno, int access);
-static void _hash_unsetpagelock(Relation rel, BlockNumber blkno, int access);
-static void _hash_splitpage(Relation rel, Buffer metabuf, Bucket obucket, Bucket nbucket);
-
-/*
- * We use high-concurrency locking on hash indices. There are two cases in
- * which we don't do locking. One is when we're building the index.
- * Since the creating transaction has not committed, no one can see
- * the index, and there's no reason to share locks. The second case
- * is when we're just starting up the database system. We use some
- * special-purpose initialization code in the relation cache manager
- * (see utils/cache/relcache.c) to allow us to do indexed scans on
- * the system catalogs before we'd normally be able to. This happens
- * before the lock table is fully initialized, so we can't use it.
- * Strictly speaking, this violates 2pl, but we don't do 2pl on the
- * system catalogs anyway.
+#include "catalog/index.h"
+#include "miscadmin.h"
+#include "storage/lmgr.h"
+#include "utils/lsyscache.h"
+
+
+static void _hash_splitbucket(Relation rel, Buffer metabuf,
+ Bucket obucket, Bucket nbucket,
+ BlockNumber start_oblkno,
+ BlockNumber start_nblkno,
+ uint32 maxbucket,
+ uint32 highmask, uint32 lowmask);
+
+
+/*
+ * We use high-concurrency locking on hash indexes (see README for an overview
+ * of the locking rules). However, we can skip taking lmgr locks when the
+ * index is local to the current backend (ie, either temp or new in the
+ * current transaction). No one else can see it, so there's no reason to
+ * take locks. We still take buffer-level locks, but not lmgr locks.
*/
+#define USELOCKING(rel) (!RELATION_IS_LOCAL(rel))
-#define USELOCKING (!BuildingHash && !IsInitProcessingMode())
+/*
+ * _hash_getlock() -- Acquire an lmgr lock.
+ *
+ * 'whichlock' should be zero to acquire the split-control lock, or the
+ * block number of a bucket's primary bucket page to acquire the per-bucket
+ * lock. (See README for details of the use of these locks.)
+ *
+ * 'access' must be HASH_SHARE or HASH_EXCLUSIVE.
+ */
+void
+_hash_getlock(Relation rel, BlockNumber whichlock, int access)
+{
+ if (USELOCKING(rel))
+ LockPage(rel, whichlock, access);
+}
+/*
+ * _hash_try_getlock() -- Acquire an lmgr lock, but only if it's free.
+ *
+ * Same as above except we return FALSE without blocking if lock isn't free.
+ */
+bool
+_hash_try_getlock(Relation rel, BlockNumber whichlock, int access)
+{
+ if (USELOCKING(rel))
+ return ConditionalLockPage(rel, whichlock, access);
+ else
+ return true;
+}
/*
- * _hash_metapinit() -- Initialize the metadata page of a hash index,
- * the two buckets that we begin with and the initial
- * bitmap page.
+ * _hash_droplock() -- Release an lmgr lock.
*/
void
-_hash_metapinit(Relation rel)
+_hash_droplock(Relation rel, BlockNumber whichlock, int access)
{
- HashMetaPage metap;
- HashPageOpaque pageopaque;
- Buffer metabuf;
- Buffer buf;
- Page pg;
- int nbuckets;
- uint32 nelem; /* number elements */
- uint32 lg2nelem; /* _hash_log2(nelem) */
- uint32 nblocks;
- uint16 i;
-
- /* can't be sharing this with anyone, now... */
- if (USELOCKING)
- RelationSetLockForWrite(rel);
-
- if ((nblocks = RelationGetNumberOfBlocks(rel)) != 0) {
- elog(WARN, "Cannot initialize non-empty hash table %s",
- RelationGetRelationName(rel));
- }
-
- metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_WRITE);
- pg = BufferGetPage(metabuf);
- metap = (HashMetaPage) pg;
- _hash_pageinit(pg, BufferGetPageSize(metabuf));
-
- metap->hashm_magic = HASH_MAGIC;
- metap->hashm_version = HASH_VERSION;
- metap->hashm_nkeys = 0;
- metap->hashm_nmaps = 0;
- metap->hashm_ffactor = DEFAULT_FFACTOR;
- metap->hashm_bsize = BufferGetPageSize(metabuf);
- metap->hashm_bshift = _hash_log2(metap->hashm_bsize);
- for (i = metap->hashm_bshift; i > 0; --i) {
- if ((1 << i) < (metap->hashm_bsize -
- (DOUBLEALIGN(sizeof(PageHeaderData)) +
- DOUBLEALIGN(sizeof(HashPageOpaqueData))))) {
- break;
- }
- }
- Assert(i);
- metap->hashm_bmsize = 1 << i;
- metap->hashm_procid = index_getprocid(rel, 1, HASHPROC);
-
- /*
- * Make nelem = 2 rather than 0 so that we end up allocating space
- * for the next greater power of two number of buckets.
- */
- nelem = 2;
- lg2nelem = 1; /*_hash_log2(MAX(nelem, 2)) */
- nbuckets = 2; /*1 << lg2nelem */
-
- memset((char *) metap->hashm_spares, 0, sizeof(metap->hashm_spares));
- memset((char *) metap->hashm_mapp, 0, sizeof(metap->hashm_mapp));
-
- metap->hashm_spares[lg2nelem] = 2; /* lg2nelem + 1 */
- metap->hashm_spares[lg2nelem + 1] = 2; /* lg2nelem + 1 */
- metap->hashm_ovflpoint = 1; /* lg2nelem */
- metap->hashm_lastfreed = 2;
-
- metap->hashm_maxbucket = metap->hashm_lowmask = 1; /* nbuckets - 1 */
- metap->hashm_highmask = 3; /* (nbuckets << 1) - 1 */
-
- pageopaque = (HashPageOpaque) PageGetSpecialPointer(pg);
- pageopaque->hasho_oaddr = InvalidOvflAddress;
- pageopaque->hasho_prevblkno = InvalidBlockNumber;
- pageopaque->hasho_nextblkno = InvalidBlockNumber;
- pageopaque->hasho_flag = LH_META_PAGE;
- pageopaque->hasho_bucket = -1;
-
- /*
- * First bitmap page is at: splitpoint lg2nelem page offset 1 which
- * turns out to be page 3. Couldn't initialize page 3 until we created
- * the first two buckets above.
- */
- if (_hash_initbitmap(rel, metap, OADDR_OF(lg2nelem, 1), lg2nelem + 1, 0))
- elog(WARN, "Problem with _hash_initbitmap.");
-
- /* all done */
- _hash_wrtnorelbuf(rel, metabuf);
-
- /*
- * initialize the first two buckets
- */
- for (i = 0; i <= 1; i++) {
- buf = _hash_getbuf(rel, BUCKET_TO_BLKNO(i), HASH_WRITE);
- pg = BufferGetPage(buf);
- _hash_pageinit(pg, BufferGetPageSize(buf));
- pageopaque = (HashPageOpaque) PageGetSpecialPointer(pg);
- pageopaque->hasho_oaddr = InvalidOvflAddress;
- pageopaque->hasho_prevblkno = InvalidBlockNumber;
- pageopaque->hasho_nextblkno = InvalidBlockNumber;
- pageopaque->hasho_flag = LH_BUCKET_PAGE;
- pageopaque->hasho_bucket = i;
- _hash_wrtbuf(rel, buf);
- }
-
- _hash_relbuf(rel, metabuf, HASH_WRITE);
-
- if (USELOCKING)
- RelationUnsetLockForWrite(rel);
+ if (USELOCKING(rel))
+ UnlockPage(rel, whichlock, access);
}
/*
- * _hash_getbuf() -- Get a buffer by block number for read or write.
+ * _hash_getbuf() -- Get a buffer by block number for read or write.
+ *
+ * 'access' must be HASH_READ, HASH_WRITE, or HASH_NOLOCK.
*
- * When this routine returns, the appropriate lock is set on the
- * requested buffer its reference count is correct.
+ * When this routine returns, the appropriate lock is set on the
+ * requested buffer and its reference count has been incremented
+ * (ie, the buffer is "locked and pinned").
*
- * XXX P_NEW is not used because, unlike the tree structures, we
- * need the bucket blocks to be at certain block numbers. we must
- * depend on the caller to call _hash_pageinit on the block if it
- * knows that this is a new block.
+ * XXX P_NEW is not used because, unlike the tree structures, we
+ * need the bucket blocks to be at certain block numbers.
+ *
+ * All call sites should call either _hash_pageinit or _hash_checkpage
+ * on the returned page, depending on whether the block is expected
+ * to be new or not.
*/
Buffer
_hash_getbuf(Relation rel, BlockNumber blkno, int access)
{
- Buffer buf;
-
- if (blkno == P_NEW) {
- elog(WARN, "_hash_getbuf: internal error: hash AM does not use P_NEW");
- }
- switch (access) {
- case HASH_WRITE:
- case HASH_READ:
- _hash_setpagelock(rel, blkno, access);
- break;
- default:
- elog(WARN, "_hash_getbuf: invalid access (%d) on new blk: %.*s",
- access, NAMEDATALEN, RelationGetRelationName(rel));
- break;
- }
- buf = ReadBuffer(rel, blkno);
-
- /* ref count and lock type are correct */
- return (buf);
+ Buffer buf;
+
+ if (blkno == P_NEW)
+ elog(ERROR, "hash AM does not use P_NEW");
+
+ buf = ReadBuffer(rel, blkno);
+
+ if (access != HASH_NOLOCK)
+ LockBuffer(buf, access);
+
+ /* ref count and lock type are correct */
+ return buf;
}
/*
- * _hash_relbuf() -- release a locked buffer.
+ * _hash_relbuf() -- release a locked buffer.
+ *
+ * Lock and pin (refcount) are both dropped.
*/
void
-_hash_relbuf(Relation rel, Buffer buf, int access)
+_hash_relbuf(Relation rel, Buffer buf)
{
- BlockNumber blkno;
-
- blkno = BufferGetBlockNumber(buf);
-
- switch (access) {
- case HASH_WRITE:
- case HASH_READ:
- _hash_unsetpagelock(rel, blkno, access);
- break;
- default:
- elog(WARN, "_hash_relbuf: invalid access (%d) on blk %x: %.*s",
- access, blkno, NAMEDATALEN, RelationGetRelationName(rel));
- }
-
- ReleaseBuffer(buf);
+ UnlockReleaseBuffer(buf);
}
/*
- * _hash_wrtbuf() -- write a hash page to disk.
+ * _hash_dropbuf() -- release an unlocked buffer.
*
- * This routine releases the lock held on the buffer and our reference
- * to it. It is an error to call _hash_wrtbuf() without a write lock
- * or a reference to the buffer.
+ * This is used to unpin a buffer on which we hold no lock.
*/
void
-_hash_wrtbuf(Relation rel, Buffer buf)
+_hash_dropbuf(Relation rel, Buffer buf)
{
- BlockNumber blkno;
-
- blkno = BufferGetBlockNumber(buf);
- WriteBuffer(buf);
- _hash_unsetpagelock(rel, blkno, HASH_WRITE);
+ ReleaseBuffer(buf);
}
/*
- * _hash_wrtnorelbuf() -- write a hash page to disk, but do not release
- * our reference or lock.
+ * _hash_wrtbuf() -- write a hash page to disk.
*
- * It is an error to call _hash_wrtnorelbuf() without a write lock
- * or a reference to the buffer.
+ * This routine releases the lock held on the buffer and our refcount
+ * for it. It is an error to call _hash_wrtbuf() without a write lock
+ * and a pin on the buffer.
+ *
+ * NOTE: this routine should go away when/if hash indexes are WAL-ified.
+ * The correct sequence of operations is to mark the buffer dirty, then
+ * write the WAL record, then release the lock and pin; so marking dirty
+ * can't be combined with releasing.
*/
void
-_hash_wrtnorelbuf(Relation rel, Buffer buf)
+_hash_wrtbuf(Relation rel, Buffer buf)
{
- BlockNumber blkno;
-
- blkno = BufferGetBlockNumber(buf);
- WriteNoReleaseBuffer(buf);
+ MarkBufferDirty(buf);
+ UnlockReleaseBuffer(buf);
}
-Page
+/*
+ * _hash_chgbufaccess() -- Change the lock type on a buffer, without
+ * dropping our pin on it.
+ *
+ * from_access and to_access may be HASH_READ, HASH_WRITE, or HASH_NOLOCK,
+ * the last indicating that no buffer-level lock is held or wanted.
+ *
+ * When from_access == HASH_WRITE, we assume the buffer is dirty and tell
+ * bufmgr it must be written out. If the caller wants to release a write
+ * lock on a page that's not been modified, it's okay to pass from_access
+ * as HASH_READ (a bit ugly, but handy in some places).
+ */
+void
_hash_chgbufaccess(Relation rel,
- Buffer *bufp,
- int from_access,
- int to_access)
+ Buffer buf,
+ int from_access,
+ int to_access)
{
- BlockNumber blkno;
-
- blkno = BufferGetBlockNumber(*bufp);
-
- switch (from_access) {
- case HASH_WRITE:
- _hash_wrtbuf(rel, *bufp);
- break;
- case HASH_READ:
- _hash_relbuf(rel, *bufp, from_access);
- break;
- default:
- elog(WARN, "_hash_chgbufaccess: invalid access (%d) on blk %x: %.*s",
- from_access, blkno, NAMEDATALEN, RelationGetRelationName(rel));
- break;
- }
- *bufp = _hash_getbuf(rel, blkno, to_access);
- return (BufferGetPage(*bufp));
+ if (from_access == HASH_WRITE)
+ MarkBufferDirty(buf);
+ if (from_access != HASH_NOLOCK)
+ LockBuffer(buf, BUFFER_LOCK_UNLOCK);
+ if (to_access != HASH_NOLOCK)
+ LockBuffer(buf, to_access);
}
+
/*
- * _hash_pageinit() -- Initialize a new page.
+ * _hash_metapinit() -- Initialize the metadata page of a hash index,
+ * the two buckets that we begin with and the initial
+ * bitmap page.
+ *
+ * We are fairly cavalier about locking here, since we know that no one else
+ * could be accessing this index. In particular the rule about not holding
+ * multiple buffer locks is ignored.
*/
void
-_hash_pageinit(Page page, Size size)
+_hash_metapinit(Relation rel)
{
- Assert(((PageHeader) page)->pd_lower == 0);
- Assert(((PageHeader) page)->pd_upper == 0);
- Assert(((PageHeader) page)->pd_special == 0);
-
- /*
- * Cargo-cult programming -- don't really need this to be zero, but
- * creating new pages is an infrequent occurrence and it makes me feel
- * good when I know they're empty.
- */
- memset(page, 0, size);
-
- PageInit(page, size, sizeof(HashPageOpaqueData));
-}
+ HashMetaPage metap;
+ HashPageOpaque pageopaque;
+ Buffer metabuf;
+ Buffer buf;
+ Page pg;
+ int32 data_width;
+ int32 item_width;
+ int32 ffactor;
+ uint16 i;
-static void
-_hash_setpagelock(Relation rel,
- BlockNumber blkno,
- int access)
-{
- ItemPointerData iptr;
-
- if (USELOCKING) {
- ItemPointerSet(&iptr, blkno, 1);
-
- switch (access) {
- case HASH_WRITE:
- RelationSetSingleWLockPage(rel, &iptr);
- break;
- case HASH_READ:
- RelationSetSingleRLockPage(rel, &iptr);
- break;
- default:
- elog(WARN, "_hash_setpagelock: invalid access (%d) on blk %x: %.*s",
- access, blkno, NAMEDATALEN, RelationGetRelationName(rel));
- break;
+ /* safety check */
+ if (RelationGetNumberOfBlocks(rel) != 0)
+ elog(ERROR, "cannot initialize non-empty hash index \"%s\"",
+ RelationGetRelationName(rel));
+
+ /*
+ * Determine the target fill factor (tuples per bucket) for this index.
+ * The idea is to make the fill factor correspond to pages about 3/4ths
+ * full. We can compute it exactly if the index datatype is fixed-width,
+ * but for var-width there's some guessing involved.
+ */
+ data_width = get_typavgwidth(RelationGetDescr(rel)->attrs[0]->atttypid,
+ RelationGetDescr(rel)->attrs[0]->atttypmod);
+ item_width = MAXALIGN(sizeof(IndexTupleData)) + MAXALIGN(data_width) +
+ sizeof(ItemIdData); /* include the line pointer */
+ ffactor = BLCKSZ * IndexGetFillFactor(rel) / 100 / item_width;
+ /* keep to a sane range */
+ if (ffactor < 10)
+ ffactor = 10;
+
+ metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_WRITE);
+ pg = BufferGetPage(metabuf);
+ _hash_pageinit(pg, BufferGetPageSize(metabuf));
+
+ pageopaque = (HashPageOpaque) PageGetSpecialPointer(pg);
+ pageopaque->hasho_prevblkno = InvalidBlockNumber;
+ pageopaque->hasho_nextblkno = InvalidBlockNumber;
+ pageopaque->hasho_bucket = -1;
+ pageopaque->hasho_flag = LH_META_PAGE;
+ pageopaque->hasho_filler = HASHO_FILL;
+
+ metap = (HashMetaPage) pg;
+
+ metap->hashm_magic = HASH_MAGIC;
+ metap->hashm_version = HASH_VERSION;
+ metap->hashm_ntuples = 0;
+ metap->hashm_nmaps = 0;
+ metap->hashm_ffactor = ffactor;
+ metap->hashm_bsize = BufferGetPageSize(metabuf);
+ /* find largest bitmap array size that will fit in page size */
+ for (i = _hash_log2(metap->hashm_bsize); i > 0; --i)
+ {
+ if ((1 << i) <= (metap->hashm_bsize -
+ (MAXALIGN(sizeof(PageHeaderData)) +
+ MAXALIGN(sizeof(HashPageOpaqueData)))))
+ break;
}
- }
-}
+ Assert(i > 0);
+ metap->hashm_bmsize = 1 << i;
+ metap->hashm_bmshift = i + BYTE_TO_BIT;
+ Assert((1 << BMPG_SHIFT(metap)) == (BMPG_MASK(metap) + 1));
-static void
-_hash_unsetpagelock(Relation rel,
- BlockNumber blkno,
- int access)
-{
- ItemPointerData iptr;
-
- if (USELOCKING) {
- ItemPointerSet(&iptr, blkno, 1);
-
- switch (access) {
- case HASH_WRITE:
- RelationUnsetSingleWLockPage(rel, &iptr);
- break;
- case HASH_READ:
- RelationUnsetSingleRLockPage(rel, &iptr);
- break;
- default:
- elog(WARN, "_hash_unsetpagelock: invalid access (%d) on blk %x: %.*s",
- access, blkno, NAMEDATALEN, RelationGetRelationName(rel));
- break;
+ metap->hashm_procid = index_getprocid(rel, 1, HASHPROC);
+
+ /*
+ * We initialize the index with two buckets, 0 and 1, occupying physical
+ * blocks 1 and 2. The first freespace bitmap page is in block 3.
+ */
+ metap->hashm_maxbucket = metap->hashm_lowmask = 1; /* nbuckets - 1 */
+ metap->hashm_highmask = 3; /* (nbuckets << 1) - 1 */
+
+ MemSet(metap->hashm_spares, 0, sizeof(metap->hashm_spares));
+ MemSet(metap->hashm_mapp, 0, sizeof(metap->hashm_mapp));
+
+ metap->hashm_spares[1] = 1; /* the first bitmap page is only spare */
+ metap->hashm_ovflpoint = 1;
+ metap->hashm_firstfree = 0;
+
+ /*
+ * Initialize the first two buckets
+ */
+ for (i = 0; i <= 1; i++)
+ {
+ buf = _hash_getbuf(rel, BUCKET_TO_BLKNO(metap, i), HASH_WRITE);
+ pg = BufferGetPage(buf);
+ _hash_pageinit(pg, BufferGetPageSize(buf));
+ pageopaque = (HashPageOpaque) PageGetSpecialPointer(pg);
+ pageopaque->hasho_prevblkno = InvalidBlockNumber;
+ pageopaque->hasho_nextblkno = InvalidBlockNumber;
+ pageopaque->hasho_bucket = i;
+ pageopaque->hasho_flag = LH_BUCKET_PAGE;
+ pageopaque->hasho_filler = HASHO_FILL;
+ _hash_wrtbuf(rel, buf);
}
- }
+
+ /*
+ * Initialize first bitmap page. Can't do this until we create the first
+ * two buckets, else smgr will complain.
+ */
+ _hash_initbitmap(rel, metap, 3);
+
+ /* all done */
+ _hash_wrtbuf(rel, metabuf);
}
+/*
+ * _hash_pageinit() -- Initialize a new hash index page.
+ */
void
-_hash_pagedel(Relation rel, ItemPointer tid)
+_hash_pageinit(Page page, Size size)
{
- Buffer buf;
- Buffer metabuf;
- Page page;
- BlockNumber blkno;
- OffsetNumber offno;
- HashMetaPage metap;
- HashPageOpaque opaque;
-
- blkno = ItemPointerGetBlockNumber(tid);
- offno = ItemPointerGetOffsetNumber(tid);
-
- buf = _hash_getbuf(rel, blkno, HASH_WRITE);
- page = BufferGetPage(buf);
- _hash_checkpage(page, LH_BUCKET_PAGE|LH_OVERFLOW_PAGE);
- opaque = (HashPageOpaque) PageGetSpecialPointer(page);
-
- PageIndexTupleDelete(page, offno);
- _hash_wrtnorelbuf(rel, buf);
-
- if (PageIsEmpty(page) && (opaque->hasho_flag & LH_OVERFLOW_PAGE)) {
- buf = _hash_freeovflpage(rel, buf);
- if (BufferIsValid(buf)) {
- _hash_relbuf(rel, buf, HASH_WRITE);
- }
- } else {
- _hash_relbuf(rel, buf, HASH_WRITE);
- }
-
- metabuf = _hash_getbuf(rel, HASH_METAPAGE, HASH_WRITE);
- metap = (HashMetaPage) BufferGetPage(metabuf);
- _hash_checkpage((Page) metap, LH_META_PAGE);
- ++metap->hashm_nkeys;
- _hash_wrtbuf(rel, metabuf);
+ Assert(PageIsNew(page));
+ PageInit(page, size, sizeof(HashPageOpaqueData));
}
+/*
+ * Attempt to expand the hash table by creating one new bucket.
+ *
+ * This will silently do nothing if it cannot get the needed locks.
+ *
+ * The caller should hold no locks on the hash index.
+ *
+ * The caller must hold a pin, but no lock, on the metapage buffer.
+ * The buffer is returned in the same state.
+ */
void
_hash_expandtable(Relation rel, Buffer metabuf)
{
- HashMetaPage metap;
- Bucket old_bucket;
- Bucket new_bucket;
- uint32 spare_ndx;
-
-/* elog(DEBUG, "_hash_expandtable: expanding..."); */
-
- metap = (HashMetaPage) BufferGetPage(metabuf);
- _hash_checkpage((Page) metap, LH_META_PAGE);
-
- metap = (HashMetaPage) _hash_chgbufaccess(rel, &metabuf, HASH_READ, HASH_WRITE);
- new_bucket = ++metap->MAX_BUCKET;
- metap = (HashMetaPage) _hash_chgbufaccess(rel, &metabuf, HASH_WRITE, HASH_READ);
- old_bucket = (metap->MAX_BUCKET & metap->LOW_MASK);
-
- /*
- * If the split point is increasing (MAX_BUCKET's log base 2
- * * increases), we need to copy the current contents of the spare
- * split bucket to the next bucket.
- */
- spare_ndx = _hash_log2(metap->MAX_BUCKET + 1);
- if (spare_ndx > metap->OVFL_POINT) {
-
- metap = (HashMetaPage) _hash_chgbufaccess(rel, &metabuf, HASH_READ, HASH_WRITE);
- metap->SPARES[spare_ndx] = metap->SPARES[metap->OVFL_POINT];
- metap->OVFL_POINT = spare_ndx;
- metap = (HashMetaPage) _hash_chgbufaccess(rel, &metabuf, HASH_WRITE, HASH_READ);
- }
-
- if (new_bucket > metap->HIGH_MASK) {
-
- /* Starting a new doubling */
- metap = (HashMetaPage) _hash_chgbufaccess(rel, &metabuf, HASH_READ, HASH_WRITE);
- metap->LOW_MASK = metap->HIGH_MASK;
- metap->HIGH_MASK = new_bucket | metap->LOW_MASK;
- metap = (HashMetaPage) _hash_chgbufaccess(rel, &metabuf, HASH_WRITE, HASH_READ);
-
- }
- /* Relocate records to the new bucket */
- _hash_splitpage(rel, metabuf, old_bucket, new_bucket);
+ HashMetaPage metap;
+ Bucket old_bucket;
+ Bucket new_bucket;
+ uint32 spare_ndx;
+ BlockNumber start_oblkno;
+ BlockNumber start_nblkno;
+ uint32 maxbucket;
+ uint32 highmask;
+ uint32 lowmask;
+
+ /*
+ * Obtain the page-zero lock to assert the right to begin a split (see
+ * README).
+ *
+ * Note: deadlock should be impossible here. Our own backend could only be
+ * holding bucket sharelocks due to stopped indexscans; those will not
+ * block other holders of the page-zero lock, who are only interested in
+ * acquiring bucket sharelocks themselves. Exclusive bucket locks are
+ * only taken here and in hashbulkdelete, and neither of these operations
+ * needs any additional locks to complete. (If, due to some flaw in this
+ * reasoning, we manage to deadlock anyway, it's okay to error out; the
+ * index will be left in a consistent state.)
+ */
+ _hash_getlock(rel, 0, HASH_EXCLUSIVE);
+
+ /* Write-lock the meta page */
+ _hash_chgbufaccess(rel, metabuf, HASH_NOLOCK, HASH_WRITE);
+
+ _hash_checkpage(rel, metabuf, LH_META_PAGE);
+ metap = (HashMetaPage) BufferGetPage(metabuf);
+
+ /*
+ * Check to see if split is still needed; someone else might have already
+ * done one while we waited for the lock.
+ *
+ * Make sure this stays in sync with _hash_doinsert()
+ */
+ if (metap->hashm_ntuples <=
+ (double) metap->hashm_ffactor * (metap->hashm_maxbucket + 1))
+ goto fail;
+
+ /*
+ * Determine which bucket is to be split, and attempt to lock the old
+ * bucket. If we can't get the lock, give up.
+ *
+ * The lock protects us against other backends, but not against our own
+ * backend. Must check for active scans separately.
+ *
+ * Ideally we would lock the new bucket too before proceeding, but if we
+ * are about to cross a splitpoint then the BUCKET_TO_BLKNO mapping isn't
+ * correct yet. For simplicity we update the metapage first and then
+ * lock. This should be okay because no one else should be trying to lock
+ * the new bucket yet...
+ */
+ new_bucket = metap->hashm_maxbucket + 1;
+ old_bucket = (new_bucket & metap->hashm_lowmask);
+
+ start_oblkno = BUCKET_TO_BLKNO(metap, old_bucket);
+
+ if (_hash_has_active_scan(rel, old_bucket))
+ goto fail;
+
+ if (!_hash_try_getlock(rel, start_oblkno, HASH_EXCLUSIVE))
+ goto fail;
+
+ /*
+ * Okay to proceed with split. Update the metapage bucket mapping info.
+ *
+ * Since we are scribbling on the metapage data right in the shared
+ * buffer, any failure in this next little bit leaves us with a big
+ * problem: the metapage is effectively corrupt but could get written back
+ * to disk. We don't really expect any failure, but just to be sure,
+ * establish a critical section.
+ */
+ START_CRIT_SECTION();
+
+ metap->hashm_maxbucket = new_bucket;
+
+ if (new_bucket > metap->hashm_highmask)
+ {
+ /* Starting a new doubling */
+ metap->hashm_lowmask = metap->hashm_highmask;
+ metap->hashm_highmask = new_bucket | metap->hashm_lowmask;
+ }
+
+ /*
+ * If the split point is increasing (hashm_maxbucket's log base 2
+ * increases), we need to adjust the hashm_spares[] array and
+ * hashm_ovflpoint so that future overflow pages will be created beyond
+ * this new batch of bucket pages.
+ *
+ * XXX should initialize new bucket pages to prevent out-of-order page
+ * creation? Don't wanna do it right here though.
+ */
+ spare_ndx = _hash_log2(metap->hashm_maxbucket + 1);
+ if (spare_ndx > metap->hashm_ovflpoint)
+ {
+ Assert(spare_ndx == metap->hashm_ovflpoint + 1);
+ metap->hashm_spares[spare_ndx] = metap->hashm_spares[metap->hashm_ovflpoint];
+ metap->hashm_ovflpoint = spare_ndx;
+ }
+
+ /* now we can compute the new bucket's primary block number */
+ start_nblkno = BUCKET_TO_BLKNO(metap, new_bucket);
+
+ Assert(!_hash_has_active_scan(rel, new_bucket));
+
+ if (!_hash_try_getlock(rel, start_nblkno, HASH_EXCLUSIVE))
+ elog(PANIC, "could not get lock on supposedly new bucket");
+
+ /* Done mucking with metapage */
+ END_CRIT_SECTION();
+
+ /*
+ * Copy bucket mapping info now; this saves re-accessing the meta page
+ * inside _hash_splitbucket's inner loop. Note that once we drop the
+ * split lock, other splits could begin, so these values might be out of
+ * date before _hash_splitbucket finishes. That's okay, since all it
+ * needs is to tell which of these two buckets to map hashkeys into.
+ */
+ maxbucket = metap->hashm_maxbucket;
+ highmask = metap->hashm_highmask;
+ lowmask = metap->hashm_lowmask;
+
+ /* Write out the metapage and drop lock, but keep pin */
+ _hash_chgbufaccess(rel, metabuf, HASH_WRITE, HASH_NOLOCK);
+
+ /* Release split lock; okay for other splits to occur now */
+ _hash_droplock(rel, 0, HASH_EXCLUSIVE);
+
+ /* Relocate records to the new bucket */
+ _hash_splitbucket(rel, metabuf, old_bucket, new_bucket,
+ start_oblkno, start_nblkno,
+ maxbucket, highmask, lowmask);
+
+ /* Release bucket locks, allowing others to access them */
+ _hash_droplock(rel, start_oblkno, HASH_EXCLUSIVE);
+ _hash_droplock(rel, start_nblkno, HASH_EXCLUSIVE);
+
+ return;
+
+ /* Here if decide not to split or fail to acquire old bucket lock */
+fail:
+
+ /* We didn't write the metapage, so just drop lock */
+ _hash_chgbufaccess(rel, metabuf, HASH_READ, HASH_NOLOCK);
+
+ /* Release split lock */
+ _hash_droplock(rel, 0, HASH_EXCLUSIVE);
}
/*
- * _hash_splitpage -- split 'obucket' into 'obucket' and 'nbucket'
+ * _hash_splitbucket -- split 'obucket' into 'obucket' and 'nbucket'
+ *
+ * We are splitting a bucket that consists of a base bucket page and zero
+ * or more overflow (bucket chain) pages. We must relocate tuples that
+ * belong in the new bucket, and compress out any free space in the old
+ * bucket.
+ *
+ * The caller must hold exclusive locks on both buckets to ensure that
+ * no one else is trying to access them (see README).
*
- * this routine is actually misnamed -- we are splitting a bucket that
- * consists of a base bucket page and zero or more overflow (bucket
- * chain) pages.
+ * The caller must hold a pin, but no lock, on the metapage buffer.
+ * The buffer is returned in the same state. (The metapage is only
+ * touched if it becomes necessary to add or remove overflow pages.)
*/
static void
-_hash_splitpage(Relation rel,
- Buffer metabuf,
- Bucket obucket,
- Bucket nbucket)
+_hash_splitbucket(Relation rel,
+ Buffer metabuf,
+ Bucket obucket,
+ Bucket nbucket,
+ BlockNumber start_oblkno,
+ BlockNumber start_nblkno,
+ uint32 maxbucket,
+ uint32 highmask,
+ uint32 lowmask)
{
- Bucket bucket;
- Buffer obuf;
- Buffer nbuf;
- Buffer ovflbuf;
- BlockNumber oblkno;
- BlockNumber nblkno;
- bool null;
- Datum datum;
- HashItem hitem;
- HashPageOpaque oopaque;
- HashPageOpaque nopaque;
- HashMetaPage metap;
- IndexTuple itup;
- int itemsz;
- OffsetNumber ooffnum;
- OffsetNumber noffnum;
- OffsetNumber omaxoffnum;
- Page opage;
- Page npage;
- TupleDesc itupdesc;
-
-/* elog(DEBUG, "_hash_splitpage: splitting %d into %d,%d",
- obucket, obucket, nbucket);
-*/
- metap = (HashMetaPage) BufferGetPage(metabuf);
- _hash_checkpage((Page) metap, LH_META_PAGE);
-
- /* get the buffers & pages */
- oblkno = BUCKET_TO_BLKNO(obucket);
- nblkno = BUCKET_TO_BLKNO(nbucket);
- obuf = _hash_getbuf(rel, oblkno, HASH_WRITE);
- nbuf = _hash_getbuf(rel, nblkno, HASH_WRITE);
- opage = BufferGetPage(obuf);
- npage = BufferGetPage(nbuf);
-
- /* initialize the new bucket */
- _hash_pageinit(npage, BufferGetPageSize(nbuf));
- nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
- nopaque->hasho_prevblkno = InvalidBlockNumber;
- nopaque->hasho_nextblkno = InvalidBlockNumber;
- nopaque->hasho_flag = LH_BUCKET_PAGE;
- nopaque->hasho_oaddr = InvalidOvflAddress;
- nopaque->hasho_bucket = nbucket;
- _hash_wrtnorelbuf(rel, nbuf);
-
- /*
- * make sure the old bucket isn't empty. advance 'opage' and
- * friends through the overflow bucket chain until we find a
- * non-empty page.
- *
- * XXX we should only need this once, if we are careful to
- * preserve the invariant that overflow pages are never empty.
- */
- _hash_checkpage(opage, LH_BUCKET_PAGE);
- oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
- if (PageIsEmpty(opage)) {
- oblkno = oopaque->hasho_nextblkno;
- _hash_relbuf(rel, obuf, HASH_WRITE);
- if (!BlockNumberIsValid(oblkno)) {
- /*
- * the old bucket is completely empty; of course, the new
- * bucket will be as well, but since it's a base bucket
- * page we don't care.
- */
- _hash_relbuf(rel, nbuf, HASH_WRITE);
- return;
- }
+ Bucket bucket;
+ Buffer obuf;
+ Buffer nbuf;
+ BlockNumber oblkno;
+ BlockNumber nblkno;
+ bool null;
+ Datum datum;
+ HashPageOpaque oopaque;
+ HashPageOpaque nopaque;
+ IndexTuple itup;
+ Size itemsz;
+ OffsetNumber ooffnum;
+ OffsetNumber noffnum;
+ OffsetNumber omaxoffnum;
+ Page opage;
+ Page npage;
+ TupleDesc itupdesc = RelationGetDescr(rel);
+
+ /*
+ * It should be okay to simultaneously write-lock pages from each bucket,
+ * since no one else can be trying to acquire buffer lock on pages of
+ * either bucket.
+ */
+ oblkno = start_oblkno;
obuf = _hash_getbuf(rel, oblkno, HASH_WRITE);
+ _hash_checkpage(rel, obuf, LH_BUCKET_PAGE);
opage = BufferGetPage(obuf);
- _hash_checkpage(opage, LH_OVERFLOW_PAGE);
- if (PageIsEmpty(opage)) {
- elog(WARN, "_hash_splitpage: empty overflow page %d", oblkno);
- }
oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
- }
-
- /*
- * we are now guaranteed that 'opage' is not empty. partition the
- * tuples in the old bucket between the old bucket and the new
- * bucket, advancing along their respective overflow bucket chains
- * and adding overflow pages as needed.
- */
- ooffnum = FirstOffsetNumber;
- omaxoffnum = PageGetMaxOffsetNumber(opage);
- for (;;) {
+
+ nblkno = start_nblkno;
+ nbuf = _hash_getbuf(rel, nblkno, HASH_WRITE);
+ npage = BufferGetPage(nbuf);
+
+ /* initialize the new bucket's primary page */
+ _hash_pageinit(npage, BufferGetPageSize(nbuf));
+ nopaque = (HashPageOpaque) PageGetSpecialPointer(npage);
+ nopaque->hasho_prevblkno = InvalidBlockNumber;
+ nopaque->hasho_nextblkno = InvalidBlockNumber;
+ nopaque->hasho_bucket = nbucket;
+ nopaque->hasho_flag = LH_BUCKET_PAGE;
+ nopaque->hasho_filler = HASHO_FILL;
+
/*
- * at each iteration through this loop, each of these variables
- * should be up-to-date: obuf opage oopaque ooffnum omaxoffnum
+ * Partition the tuples in the old bucket between the old bucket and the
+ * new bucket, advancing along the old bucket's overflow bucket chain and
+ * adding overflow pages to the new bucket as needed.
*/
-
- /* check if we're at the end of the page */
- if (ooffnum > omaxoffnum) {
- /* at end of page, but check for overflow page */
- oblkno = oopaque->hasho_nextblkno;
- if (BlockNumberIsValid(oblkno)) {
+ ooffnum = FirstOffsetNumber;
+ omaxoffnum = PageGetMaxOffsetNumber(opage);
+ for (;;)
+ {
/*
- * we ran out of tuples on this particular page, but
- * we have more overflow pages; re-init values.
+ * at each iteration through this loop, each of these variables should
+ * be up-to-date: obuf opage oopaque ooffnum omaxoffnum
*/
- _hash_wrtbuf(rel, obuf);
- obuf = _hash_getbuf(rel, oblkno, HASH_WRITE);
- opage = BufferGetPage(obuf);
- _hash_checkpage(opage, LH_OVERFLOW_PAGE);
- oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
-
- /* we're guaranteed that an ovfl page has at least 1 tuple */
- if (PageIsEmpty(opage)) {
- elog(WARN, "_hash_splitpage: empty ovfl page %d!",
- oblkno);
+
+ /* check if we're at the end of the page */
+ if (ooffnum > omaxoffnum)
+ {
+ /* at end of page, but check for an(other) overflow page */
+ oblkno = oopaque->hasho_nextblkno;
+ if (!BlockNumberIsValid(oblkno))
+ break;
+
+ /*
+ * we ran out of tuples on this particular page, but we have more
+ * overflow pages; advance to next page.
+ */
+ _hash_wrtbuf(rel, obuf);
+
+ obuf = _hash_getbuf(rel, oblkno, HASH_WRITE);
+ _hash_checkpage(rel, obuf, LH_OVERFLOW_PAGE);
+ opage = BufferGetPage(obuf);
+ oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
+ ooffnum = FirstOffsetNumber;
+ omaxoffnum = PageGetMaxOffsetNumber(opage);
+ continue;
}
- ooffnum = FirstOffsetNumber;
- omaxoffnum = PageGetMaxOffsetNumber(opage);
- } else {
+
/*
- * we're at the end of the bucket chain, so now we're
- * really done with everything. before quitting, call
- * _hash_squeezebucket to ensure the tuples in the
- * bucket (including the overflow pages) are packed as
- * tightly as possible.
+ * Re-hash the tuple to determine which bucket it now belongs in.
+ *
+ * It is annoying to call the hash function while holding locks, but
+ * releasing and relocking the page for each tuple is unappealing too.
*/
- _hash_wrtbuf(rel, obuf);
- _hash_wrtbuf(rel, nbuf);
- _hash_squeezebucket(rel, metap, obucket);
- return;
- }
- }
-
- /* hash on the tuple */
- hitem = (HashItem) PageGetItem(opage, PageGetItemId(opage, ooffnum));
- itup = &(hitem->hash_itup);
- itupdesc = RelationGetTupleDescriptor(rel);
- datum = index_getattr(itup, 1, itupdesc, &null);
- bucket = _hash_call(rel, metap, datum);
-
- if (bucket == nbucket) {
- /*
- * insert the tuple into the new bucket. if it doesn't
- * fit on the current page in the new bucket, we must
- * allocate a new overflow page and place the tuple on
- * that page instead.
- */
- itemsz = IndexTupleDSize(hitem->hash_itup)
- + (sizeof(HashItemData) - sizeof(IndexTupleData));
-
- itemsz = DOUBLEALIGN(itemsz);
-
- if (PageGetFreeSpace(npage) < itemsz) {
- ovflbuf = _hash_addovflpage(rel, &metabuf, nbuf);
- _hash_wrtbuf(rel, nbuf);
- nbuf = ovflbuf;
- npage = BufferGetPage(nbuf);
- _hash_checkpage(npage, LH_BUCKET_PAGE|LH_OVERFLOW_PAGE);
- }
-
- noffnum = OffsetNumberNext(PageGetMaxOffsetNumber(npage));
- (void) PageAddItem(npage, (Item) hitem, itemsz, noffnum, LP_USED);
- _hash_wrtnorelbuf(rel, nbuf);
-
- /*
- * now delete the tuple from the old bucket. after this
- * section of code, 'ooffnum' will actually point to the
- * ItemId to which we would point if we had advanced it
- * before the deletion (PageIndexTupleDelete repacks the
- * ItemId array). this also means that 'omaxoffnum' is
- * exactly one less than it used to be, so we really can
- * just decrement it instead of calling
- * PageGetMaxOffsetNumber.
- */
- PageIndexTupleDelete(opage, ooffnum);
- _hash_wrtnorelbuf(rel, obuf);
- omaxoffnum = OffsetNumberPrev(omaxoffnum);
-
- /*
- * tidy up. if the old page was an overflow page and it
- * is now empty, we must free it (we want to preserve the
- * invariant that overflow pages cannot be empty).
- */
- if (PageIsEmpty(opage) &&
- (oopaque->hasho_flag & LH_OVERFLOW_PAGE)) {
- obuf = _hash_freeovflpage(rel, obuf);
-
- /* check that we're not through the bucket chain */
- if (BufferIsInvalid(obuf)) {
- _hash_wrtbuf(rel, nbuf);
- _hash_squeezebucket(rel, metap, obucket);
- return;
+ itup = (IndexTuple) PageGetItem(opage, PageGetItemId(opage, ooffnum));
+ datum = index_getattr(itup, 1, itupdesc, &null);
+ Assert(!null);
+
+ bucket = _hash_hashkey2bucket(_hash_datum2hashkey(rel, datum),
+ maxbucket, highmask, lowmask);
+
+ if (bucket == nbucket)
+ {
+ /*
+ * insert the tuple into the new bucket. if it doesn't fit on the
+ * current page in the new bucket, we must allocate a new overflow
+ * page and place the tuple on that page instead.
+ */
+ itemsz = IndexTupleDSize(*itup);
+ itemsz = MAXALIGN(itemsz);
+
+ if (PageGetFreeSpace(npage) < itemsz)
+ {
+ /* write out nbuf and drop lock, but keep pin */
+ _hash_chgbufaccess(rel, nbuf, HASH_WRITE, HASH_NOLOCK);
+ /* chain to a new overflow page */
+ nbuf = _hash_addovflpage(rel, metabuf, nbuf);
+ _hash_checkpage(rel, nbuf, LH_OVERFLOW_PAGE);
+ npage = BufferGetPage(nbuf);
+ /* we don't need nopaque within the loop */
+ }
+
+ noffnum = OffsetNumberNext(PageGetMaxOffsetNumber(npage));
+ if (PageAddItem(npage, (Item) itup, itemsz, noffnum, LP_USED)
+ == InvalidOffsetNumber)
+ elog(ERROR, "failed to add index item to \"%s\"",
+ RelationGetRelationName(rel));
+
+ /*
+ * now delete the tuple from the old bucket. after this section
+ * of code, 'ooffnum' will actually point to the ItemId to which
+ * we would point if we had advanced it before the deletion
+ * (PageIndexTupleDelete repacks the ItemId array). this also
+ * means that 'omaxoffnum' is exactly one less than it used to be,
+ * so we really can just decrement it instead of calling
+ * PageGetMaxOffsetNumber.
+ */
+ PageIndexTupleDelete(opage, ooffnum);
+ omaxoffnum = OffsetNumberPrev(omaxoffnum);
}
-
- /*
- * re-init. again, we're guaranteed that an ovfl page
- * has at least one tuple.
- */
- opage = BufferGetPage(obuf);
- _hash_checkpage(opage, LH_OVERFLOW_PAGE);
- oblkno = BufferGetBlockNumber(obuf);
- oopaque = (HashPageOpaque) PageGetSpecialPointer(opage);
- if (PageIsEmpty(opage)) {
- elog(WARN, "_hash_splitpage: empty overflow page %d",
- oblkno);
+ else
+ {
+ /*
+ * the tuple stays on this page. we didn't move anything, so we
+ * didn't delete anything and therefore we don't have to change
+ * 'omaxoffnum'.
+ */
+ Assert(bucket == obucket);
+ ooffnum = OffsetNumberNext(ooffnum);
}
- ooffnum = FirstOffsetNumber;
- omaxoffnum = PageGetMaxOffsetNumber(opage);
- }
- } else {
- /*
- * the tuple stays on this page. we didn't move anything,
- * so we didn't delete anything and therefore we don't
- * have to change 'omaxoffnum'.
- *
- * XXX any hash value from [0, nbucket-1] will map to this
- * bucket, which doesn't make sense to me.
- */
- ooffnum = OffsetNumberNext(ooffnum);
}
- }
- /*NOTREACHED*/
+
+ /*
+ * We're at the end of the old bucket chain, so we're done partitioning
+ * the tuples. Before quitting, call _hash_squeezebucket to ensure the
+ * tuples remaining in the old bucket (including the overflow pages) are
+ * packed as tightly as possible. The new bucket is already tight.
+ */
+ _hash_wrtbuf(rel, obuf);
+ _hash_wrtbuf(rel, nbuf);
+
+ _hash_squeezebucket(rel, obucket, start_oblkno);
}
projects / postgresql / blobdiff