Exemple #1
0
/*
 *	_hash_squeezebucket(rel, bucket)
 *
 *	Try to squeeze the tuples onto pages occurring earlier in the
 *	bucket chain in an attempt to free overflow pages. When we start
 *	the "squeezing", the page from which we start taking tuples (the
 *	"read" page) is the last bucket in the bucket chain and the page
 *	onto which we start squeezing tuples (the "write" page) is the
 *	first page in the bucket chain.  The read page works backward and
 *	the write page works forward; the procedure terminates when the
 *	read page and write page are the same page.
 *
 *	At completion of this procedure, it is guaranteed that all pages in
 *	the bucket are nonempty, unless the bucket is totally empty (in
 *	which case all overflow pages will be freed).  The original implementation
 *	required that to be true on entry as well, but it's a lot easier for
 *	callers to leave empty overflow pages and let this guy clean it up.
 *
 *	Caller must acquire cleanup lock on the primary page of the target
 *	bucket to exclude any scans that are in progress, which could easily
 *	be confused into returning the same tuple more than once or some tuples
 *	not at all by the rearrangement we are performing here.  To prevent
 *	any concurrent scan to cross the squeeze scan we use lock chaining
 *	similar to hasbucketcleanup.  Refer comments atop hashbucketcleanup.
 *
 *	We need to retain a pin on the primary bucket to ensure that no concurrent
 *	split can start.
 *
 *	Since this function is invoked in VACUUM, we provide an access strategy
 *	parameter that controls fetches of the bucket pages.
 */
void
_hash_squeezebucket(Relation rel,
					Bucket bucket,
					BlockNumber bucket_blkno,
					Buffer bucket_buf,
					BufferAccessStrategy bstrategy)
{
	BlockNumber wblkno;
	BlockNumber rblkno;
	Buffer		wbuf;
	Buffer		rbuf;
	Page		wpage;
	Page		rpage;
	HashPageOpaque wopaque;
	HashPageOpaque ropaque;

	/*
	 * start squeezing into the primary bucket page.
	 */
	wblkno = bucket_blkno;
	wbuf = bucket_buf;
	wpage = BufferGetPage(wbuf);
	wopaque = (HashPageOpaque) PageGetSpecialPointer(wpage);

	/*
	 * if there aren't any overflow pages, there's nothing to squeeze. caller
	 * is responsible for releasing the pin on primary bucket page.
	 */
	if (!BlockNumberIsValid(wopaque->hasho_nextblkno))
	{
		LockBuffer(wbuf, BUFFER_LOCK_UNLOCK);
		return;
	}

	/*
	 * Find the last page in the bucket chain by starting at the base bucket
	 * page and working forward.  Note: we assume that a hash bucket chain is
	 * usually smaller than the buffer ring being used by VACUUM, else using
	 * the access strategy here would be counterproductive.
	 */
	rbuf = InvalidBuffer;
	ropaque = wopaque;
	do
	{
		rblkno = ropaque->hasho_nextblkno;
		if (rbuf != InvalidBuffer)
			_hash_relbuf(rel, rbuf);
		rbuf = _hash_getbuf_with_strategy(rel,
										  rblkno,
										  HASH_WRITE,
										  LH_OVERFLOW_PAGE,
										  bstrategy);
		rpage = BufferGetPage(rbuf);
		ropaque = (HashPageOpaque) PageGetSpecialPointer(rpage);
		Assert(ropaque->hasho_bucket == bucket);
	} while (BlockNumberIsValid(ropaque->hasho_nextblkno));

	/*
	 * squeeze the tuples.
	 */
	for (;;)
	{
		OffsetNumber roffnum;
		OffsetNumber maxroffnum;
		OffsetNumber deletable[MaxOffsetNumber];
		IndexTuple	itups[MaxIndexTuplesPerPage];
		Size		tups_size[MaxIndexTuplesPerPage];
		OffsetNumber itup_offsets[MaxIndexTuplesPerPage];
		uint16		ndeletable = 0;
		uint16		nitups = 0;
		Size		all_tups_size = 0;
		int			i;
		bool		retain_pin = false;

readpage:
		/* Scan each tuple in "read" page */
		maxroffnum = PageGetMaxOffsetNumber(rpage);
		for (roffnum = FirstOffsetNumber;
			 roffnum <= maxroffnum;
			 roffnum = OffsetNumberNext(roffnum))
		{
			IndexTuple	itup;
			Size		itemsz;

			/* skip dead tuples */
			if (ItemIdIsDead(PageGetItemId(rpage, roffnum)))
				continue;

			itup = (IndexTuple) PageGetItem(rpage,
											PageGetItemId(rpage, roffnum));
			itemsz = IndexTupleDSize(*itup);
			itemsz = MAXALIGN(itemsz);

			/*
			 * Walk up the bucket chain, looking for a page big enough for
			 * this item and all other accumulated items.  Exit if we reach
			 * the read page.
			 */
			while (PageGetFreeSpaceForMultipleTuples(wpage, nitups + 1) < (all_tups_size + itemsz))
			{
				Buffer		next_wbuf = InvalidBuffer;
				bool		tups_moved = false;

				Assert(!PageIsEmpty(wpage));

				if (wblkno == bucket_blkno)
					retain_pin = true;

				wblkno = wopaque->hasho_nextblkno;
				Assert(BlockNumberIsValid(wblkno));

				/* don't need to move to next page if we reached the read page */
				if (wblkno != rblkno)
					next_wbuf = _hash_getbuf_with_strategy(rel,
														   wblkno,
														   HASH_WRITE,
														   LH_OVERFLOW_PAGE,
														   bstrategy);

				if (nitups > 0)
				{
					Assert(nitups == ndeletable);

					/*
					 * This operation needs to log multiple tuples, prepare
					 * WAL for that.
					 */
					if (RelationNeedsWAL(rel))
						XLogEnsureRecordSpace(0, 3 + nitups);

					START_CRIT_SECTION();

					/*
					 * we have to insert tuples on the "write" page, being
					 * careful to preserve hashkey ordering.  (If we insert
					 * many tuples into the same "write" page it would be
					 * worth qsort'ing them).
					 */
					_hash_pgaddmultitup(rel, wbuf, itups, itup_offsets, nitups);
					MarkBufferDirty(wbuf);

					/* Delete tuples we already moved off read page */
					PageIndexMultiDelete(rpage, deletable, ndeletable);
					MarkBufferDirty(rbuf);

					/* XLOG stuff */
					if (RelationNeedsWAL(rel))
					{
						XLogRecPtr	recptr;
						xl_hash_move_page_contents xlrec;

						xlrec.ntups = nitups;
						xlrec.is_prim_bucket_same_wrt = (wbuf == bucket_buf) ? true : false;

						XLogBeginInsert();
						XLogRegisterData((char *) &xlrec, SizeOfHashMovePageContents);

						/*
						 * bucket buffer needs to be registered to ensure that
						 * we can acquire a cleanup lock on it during replay.
						 */
						if (!xlrec.is_prim_bucket_same_wrt)
							XLogRegisterBuffer(0, bucket_buf, REGBUF_STANDARD | REGBUF_NO_IMAGE);

						XLogRegisterBuffer(1, wbuf, REGBUF_STANDARD);
						XLogRegisterBufData(1, (char *) itup_offsets,
											nitups * sizeof(OffsetNumber));
						for (i = 0; i < nitups; i++)
							XLogRegisterBufData(1, (char *) itups[i], tups_size[i]);

						XLogRegisterBuffer(2, rbuf, REGBUF_STANDARD);
						XLogRegisterBufData(2, (char *) deletable,
											ndeletable * sizeof(OffsetNumber));

						recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_MOVE_PAGE_CONTENTS);

						PageSetLSN(BufferGetPage(wbuf), recptr);
						PageSetLSN(BufferGetPage(rbuf), recptr);
					}

					END_CRIT_SECTION();

					tups_moved = true;
				}

				/*
				 * release the lock on previous page after acquiring the lock
				 * on next page
				 */
				if (retain_pin)
					LockBuffer(wbuf, BUFFER_LOCK_UNLOCK);
				else
					_hash_relbuf(rel, wbuf);

				/* nothing more to do if we reached the read page */
				if (rblkno == wblkno)
				{
					_hash_relbuf(rel, rbuf);
					return;
				}

				wbuf = next_wbuf;
				wpage = BufferGetPage(wbuf);
				wopaque = (HashPageOpaque) PageGetSpecialPointer(wpage);
				Assert(wopaque->hasho_bucket == bucket);
				retain_pin = false;

				/* be tidy */
				for (i = 0; i < nitups; i++)
					pfree(itups[i]);
				nitups = 0;
				all_tups_size = 0;
				ndeletable = 0;

				/*
				 * after moving the tuples, rpage would have been compacted,
				 * so we need to rescan it.
				 */
				if (tups_moved)
					goto readpage;
			}

			/* remember tuple for deletion from "read" page */
			deletable[ndeletable++] = roffnum;

			/*
			 * we need a copy of index tuples as they can be freed as part of
			 * overflow page, however we need them to write a WAL record in
			 * _hash_freeovflpage.
			 */
			itups[nitups] = CopyIndexTuple(itup);
			tups_size[nitups++] = itemsz;
			all_tups_size += itemsz;
		}

		/*
		 * If we reach here, there are no live tuples on the "read" page ---
		 * it was empty when we got to it, or we moved them all.  So we can
		 * just free the page without bothering with deleting tuples
		 * individually.  Then advance to the previous "read" page.
		 *
		 * Tricky point here: if our read and write pages are adjacent in the
		 * bucket chain, our write lock on wbuf will conflict with
		 * _hash_freeovflpage's attempt to update the sibling links of the
		 * removed page.  In that case, we don't need to lock it again.
		 */
		rblkno = ropaque->hasho_prevblkno;
		Assert(BlockNumberIsValid(rblkno));

		/* free this overflow page (releases rbuf) */
		_hash_freeovflpage(rel, bucket_buf, rbuf, wbuf, itups, itup_offsets,
						   tups_size, nitups, bstrategy);

		/* be tidy */
		for (i = 0; i < nitups; i++)
			pfree(itups[i]);

		/* are we freeing the page adjacent to wbuf? */
		if (rblkno == wblkno)
		{
			/* retain the pin on primary bucket page till end of bucket scan */
			if (wblkno == bucket_blkno)
				LockBuffer(wbuf, BUFFER_LOCK_UNLOCK);
			else
				_hash_relbuf(rel, wbuf);
			return;
		}

		rbuf = _hash_getbuf_with_strategy(rel,
										  rblkno,
										  HASH_WRITE,
										  LH_OVERFLOW_PAGE,
										  bstrategy);
		rpage = BufferGetPage(rbuf);
		ropaque = (HashPageOpaque) PageGetSpecialPointer(rpage);
		Assert(ropaque->hasho_bucket == bucket);
	}

	/* NOTREACHED */
}
/*
 *	_hash_squeezebucket(rel, bucket)
 *
 *	Try to squeeze the tuples onto pages occurring earlier in the
 *	bucket chain in an attempt to free overflow pages. When we start
 *	the "squeezing", the page from which we start taking tuples (the
 *	"read" page) is the last bucket in the bucket chain and the page
 *	onto which we start squeezing tuples (the "write" page) is the
 *	first page in the bucket chain.  The read page works backward and
 *	the write page works forward; the procedure terminates when the
 *	read page and write page are the same page.
 *
 *	At completion of this procedure, it is guaranteed that all pages in
 *	the bucket are nonempty, unless the bucket is totally empty (in
 *	which case all overflow pages will be freed).  The original implementation
 *	required that to be true on entry as well, but it's a lot easier for
 *	callers to leave empty overflow pages and let this guy clean it up.
 *
 *	Caller must hold exclusive lock on the target bucket.  This allows
 *	us to safely lock multiple pages in the bucket.
 *
 *	Since this function is invoked in VACUUM, we provide an access strategy
 *	parameter that controls fetches of the bucket pages.
 */
void
_hash_squeezebucket(Relation rel,
                    Bucket bucket,
                    BlockNumber bucket_blkno,
                    BufferAccessStrategy bstrategy)
{
    BlockNumber wblkno;
    BlockNumber rblkno;
    Buffer		wbuf;
    Buffer		rbuf;
    Page		wpage;
    Page		rpage;
    HashPageOpaque wopaque;
    HashPageOpaque ropaque;
    bool		wbuf_dirty;

    /*
     * start squeezing into the base bucket page.
     */
    wblkno = bucket_blkno;
    wbuf = _hash_getbuf_with_strategy(rel,
                                      wblkno,
                                      HASH_WRITE,
                                      LH_BUCKET_PAGE,
                                      bstrategy);
    wpage = BufferGetPage(wbuf);
    wopaque = (HashPageOpaque) PageGetSpecialPointer(wpage);

    /*
     * if there aren't any overflow pages, there's nothing to squeeze.
     */
    if (!BlockNumberIsValid(wopaque->hasho_nextblkno))
    {
        _hash_relbuf(rel, wbuf);
        return;
    }

    /*
     * Find the last page in the bucket chain by starting at the base bucket
     * page and working forward.  Note: we assume that a hash bucket chain is
     * usually smaller than the buffer ring being used by VACUUM, else using
     * the access strategy here would be counterproductive.
     */
    rbuf = InvalidBuffer;
    ropaque = wopaque;
    do
    {
        rblkno = ropaque->hasho_nextblkno;
        if (rbuf != InvalidBuffer)
            _hash_relbuf(rel, rbuf);
        rbuf = _hash_getbuf_with_strategy(rel,
                                          rblkno,
                                          HASH_WRITE,
                                          LH_OVERFLOW_PAGE,
                                          bstrategy);
        rpage = BufferGetPage(rbuf);
        ropaque = (HashPageOpaque) PageGetSpecialPointer(rpage);
        Assert(ropaque->hasho_bucket == bucket);
    } while (BlockNumberIsValid(ropaque->hasho_nextblkno));

    /*
     * squeeze the tuples.
     */
    wbuf_dirty = false;
    for (;;)
    {
        OffsetNumber roffnum;
        OffsetNumber maxroffnum;
        OffsetNumber deletable[MaxOffsetNumber];
        int			ndeletable = 0;

        /* Scan each tuple in "read" page */
        maxroffnum = PageGetMaxOffsetNumber(rpage);
        for (roffnum = FirstOffsetNumber;
                roffnum <= maxroffnum;
                roffnum = OffsetNumberNext(roffnum))
        {
            IndexTuple	itup;
            Size		itemsz;

            itup = (IndexTuple) PageGetItem(rpage,
                                            PageGetItemId(rpage, roffnum));
            itemsz = IndexTupleDSize(*itup);
            itemsz = MAXALIGN(itemsz);

            /*
             * Walk up the bucket chain, looking for a page big enough for
             * this item.  Exit if we reach the read page.
             */
            while (PageGetFreeSpace(wpage) < itemsz)
            {
                Assert(!PageIsEmpty(wpage));

                wblkno = wopaque->hasho_nextblkno;
                Assert(BlockNumberIsValid(wblkno));

                if (wbuf_dirty)
                    _hash_wrtbuf(rel, wbuf);
                else
                    _hash_relbuf(rel, wbuf);

                /* nothing more to do if we reached the read page */
                if (rblkno == wblkno)
                {
                    if (ndeletable > 0)
                    {
                        /* Delete tuples we already moved off read page */
                        PageIndexMultiDelete(rpage, deletable, ndeletable);
                        _hash_wrtbuf(rel, rbuf);
                    }
                    else
                        _hash_relbuf(rel, rbuf);
                    return;
                }

                wbuf = _hash_getbuf_with_strategy(rel,
                                                  wblkno,
                                                  HASH_WRITE,
                                                  LH_OVERFLOW_PAGE,
                                                  bstrategy);
                wpage = BufferGetPage(wbuf);
                wopaque = (HashPageOpaque) PageGetSpecialPointer(wpage);
                Assert(wopaque->hasho_bucket == bucket);
                wbuf_dirty = false;
            }

            /*
             * we have found room so insert on the "write" page, being careful
             * to preserve hashkey ordering.  (If we insert many tuples into
             * the same "write" page it would be worth qsort'ing instead of
             * doing repeated _hash_pgaddtup.)
             */
            (void) _hash_pgaddtup(rel, wbuf, itemsz, itup);
            wbuf_dirty = true;

            /* remember tuple for deletion from "read" page */
            deletable[ndeletable++] = roffnum;
        }

        /*
         * If we reach here, there are no live tuples on the "read" page ---
         * it was empty when we got to it, or we moved them all.  So we can
         * just free the page without bothering with deleting tuples
         * individually.  Then advance to the previous "read" page.
         *
         * Tricky point here: if our read and write pages are adjacent in the
         * bucket chain, our write lock on wbuf will conflict with
         * _hash_freeovflpage's attempt to update the sibling links of the
         * removed page.  However, in that case we are done anyway, so we can
         * simply drop the write lock before calling _hash_freeovflpage.
         */
        rblkno = ropaque->hasho_prevblkno;
        Assert(BlockNumberIsValid(rblkno));

        /* are we freeing the page adjacent to wbuf? */
        if (rblkno == wblkno)
        {
            /* yes, so release wbuf lock first */
            if (wbuf_dirty)
                _hash_wrtbuf(rel, wbuf);
            else
                _hash_relbuf(rel, wbuf);
            /* free this overflow page (releases rbuf) */
            _hash_freeovflpage(rel, rbuf, bstrategy);
            /* done */
            return;
        }

        /* free this overflow page, then get the previous one */
        _hash_freeovflpage(rel, rbuf, bstrategy);

        rbuf = _hash_getbuf_with_strategy(rel,
                                          rblkno,
                                          HASH_WRITE,
                                          LH_OVERFLOW_PAGE,
                                          bstrategy);
        rpage = BufferGetPage(rbuf);
        ropaque = (HashPageOpaque) PageGetSpecialPointer(rpage);
        Assert(ropaque->hasho_bucket == bucket);
    }

    /* NOTREACHED */
}
/*
 *	_hash_squeezebucket(rel, bucket)
 *
 *	Try to squeeze the tuples onto pages occurring earlier in the
 *	bucket chain in an attempt to free overflow pages. When we start
 *	the "squeezing", the page from which we start taking tuples (the
 *	"read" page) is the last bucket in the bucket chain and the page
 *	onto which we start squeezing tuples (the "write" page) is the
 *	first page in the bucket chain.  The read page works backward and
 *	the write page works forward; the procedure terminates when the
 *	read page and write page are the same page.
 *
 *	At completion of this procedure, it is guaranteed that all pages in
 *	the bucket are nonempty, unless the bucket is totally empty (in
 *	which case all overflow pages will be freed).  The original implementation
 *	required that to be true on entry as well, but it's a lot easier for
 *	callers to leave empty overflow pages and let this guy clean it up.
 *
 *	Caller must hold exclusive lock on the target bucket.  This allows
 *	us to safely lock multiple pages in the bucket.
 */
void
_hash_squeezebucket(Relation rel,
					Bucket bucket,
					BlockNumber bucket_blkno)
{
	Buffer		wbuf;
	Buffer		rbuf = 0;
	BlockNumber wblkno;
	BlockNumber rblkno;
	Page		wpage;
	Page		rpage;
	HashPageOpaque wopaque;
	HashPageOpaque ropaque;
	OffsetNumber woffnum;
	OffsetNumber roffnum;
	IndexTuple	itup;
	Size		itemsz;

	/*
	 * start squeezing into the base bucket page.
	 */
	wblkno = bucket_blkno;
	wbuf = _hash_getbuf(rel, wblkno, HASH_WRITE);
	_hash_checkpage(rel, wbuf, LH_BUCKET_PAGE);
	wpage = BufferGetPage(wbuf);
	wopaque = (HashPageOpaque) PageGetSpecialPointer(wpage);

	/*
	 * if there aren't any overflow pages, there's nothing to squeeze.
	 */
	if (!BlockNumberIsValid(wopaque->hasho_nextblkno))
	{
		_hash_relbuf(rel, wbuf);
		return;
	}

	/*
	 * find the last page in the bucket chain by starting at the base bucket
	 * page and working forward.
	 */
	ropaque = wopaque;
	do
	{
		rblkno = ropaque->hasho_nextblkno;
		if (ropaque != wopaque)
			_hash_relbuf(rel, rbuf);
		rbuf = _hash_getbuf(rel, rblkno, HASH_WRITE);
		_hash_checkpage(rel, rbuf, LH_OVERFLOW_PAGE);
		rpage = BufferGetPage(rbuf);
		ropaque = (HashPageOpaque) PageGetSpecialPointer(rpage);
		Assert(ropaque->hasho_bucket == bucket);
	} while (BlockNumberIsValid(ropaque->hasho_nextblkno));

	/*
	 * squeeze the tuples.
	 */
	roffnum = FirstOffsetNumber;
	for (;;)
	{
		/* this test is needed in case page is empty on entry */
		if (roffnum <= PageGetMaxOffsetNumber(rpage))
		{
			itup = (IndexTuple) PageGetItem(rpage,
											PageGetItemId(rpage, roffnum));
			itemsz = IndexTupleDSize(*itup);
			itemsz = MAXALIGN(itemsz);

			/*
			 * Walk up the bucket chain, looking for a page big enough for
			 * this item.  Exit if we reach the read page.
			 */
			while (PageGetFreeSpace(wpage) < itemsz)
			{
				Assert(!PageIsEmpty(wpage));

				wblkno = wopaque->hasho_nextblkno;
				Assert(BlockNumberIsValid(wblkno));

				_hash_wrtbuf(rel, wbuf);

				if (rblkno == wblkno)
				{
					/* wbuf is already released */
					_hash_wrtbuf(rel, rbuf);
					return;
				}

				wbuf = _hash_getbuf(rel, wblkno, HASH_WRITE);
				_hash_checkpage(rel, wbuf, LH_OVERFLOW_PAGE);
				wpage = BufferGetPage(wbuf);
				wopaque = (HashPageOpaque) PageGetSpecialPointer(wpage);
				Assert(wopaque->hasho_bucket == bucket);
			}

			/*
			 * we have found room so insert on the "write" page.
			 */
			woffnum = OffsetNumberNext(PageGetMaxOffsetNumber(wpage));
			if (PageAddItem(wpage, (Item) itup, itemsz, woffnum, LP_USED)
				== InvalidOffsetNumber)
				elog(ERROR, "failed to add index item to \"%s\"",
					 RelationGetRelationName(rel));

			/*
			 * delete the tuple from the "read" page. PageIndexTupleDelete
			 * repacks the ItemId array, so 'roffnum' will be "advanced" to
			 * the "next" ItemId.
			 */
			PageIndexTupleDelete(rpage, roffnum);
		}

		/*
		 * if the "read" page is now empty because of the deletion (or because
		 * it was empty when we got to it), free it.
		 *
		 * Tricky point here: if our read and write pages are adjacent in the
		 * bucket chain, our write lock on wbuf will conflict with
		 * _hash_freeovflpage's attempt to update the sibling links of the
		 * removed page.  However, in that case we are done anyway, so we can
		 * simply drop the write lock before calling _hash_freeovflpage.
		 */
		if (PageIsEmpty(rpage))
		{
			rblkno = ropaque->hasho_prevblkno;
			Assert(BlockNumberIsValid(rblkno));

			/* are we freeing the page adjacent to wbuf? */
			if (rblkno == wblkno)
			{
				/* yes, so release wbuf lock first */
				_hash_wrtbuf(rel, wbuf);
				/* free this overflow page (releases rbuf) */
				_hash_freeovflpage(rel, rbuf);
				/* done */
				return;
			}

			/* free this overflow page, then get the previous one */
			_hash_freeovflpage(rel, rbuf);

			rbuf = _hash_getbuf(rel, rblkno, HASH_WRITE);
			_hash_checkpage(rel, rbuf, LH_OVERFLOW_PAGE);
			rpage = BufferGetPage(rbuf);
			ropaque = (HashPageOpaque) PageGetSpecialPointer(rpage);
			Assert(ropaque->hasho_bucket == bucket);

			roffnum = FirstOffsetNumber;
		}
	}

	/* NOTREACHED */
}