Delete empty pages during GiST VACUUM.

To do this, we scan GiST two times. In the first pass we make note of
empty leaf pages and internal pages. At second pass we scan through
internal pages, looking for downlinks to the empty pages.

Deleting internal pages is still not supported, like in nbtree, the last
child of an internal page is never deleted. That means that if you have a
workload where new keys are always inserted to different area than where
old keys are removed, the index will still grow without bound. But the rate
of growth will be an order of magnitude slower than before.

Author: Andrey Borodin
Discussion: https://www.postgresql.org/message-id/B1E4DF12-6CD3-4706-BDBD-BF3283328F60@yandex-team.ru

1 files changed, 343 insertions, 35 deletions

diff --git a/src/backend/access/gist/gistvacuum.c b/src/backend/access/gist/gistvacuum.c
index 3c1d75691e8..28d5ec9ce42 100644
--- a/src/backend/access/gist/gistvacuum.c
+++ b/src/backend/access/gist/gistvacuum.c
@@ -16,26 +16,65 @@
 
 #include "access/genam.h"
 #include "access/gist_private.h"
+#include "access/transam.h"
 #include "commands/vacuum.h"
+#include "lib/integerset.h"
 #include "miscadmin.h"
 #include "storage/indexfsm.h"
 #include "storage/lmgr.h"
+#include "utils/memutils.h"
 
-/* Working state needed by gistbulkdelete */
+/*
+ * State kept across vacuum stages.
+ */
 typedef struct
 {
+	IndexBulkDeleteResult stats;	/* must be first */
+
 	IndexVacuumInfo *info;
-	IndexBulkDeleteResult *stats;
+
+	/*
+	 * These are used to memorize all internal and empty leaf pages in the 1st
+	 * vacuum stage.  They are used in the 2nd stage, to delete all the empty
+	 * pages.
+	 */
+	IntegerSet *internal_page_set;
+	IntegerSet *empty_leaf_set;
+	MemoryContext page_set_context;
+} GistBulkDeleteResult;
+
+/* Working state needed by gistbulkdelete */
+typedef struct
+{
+	GistBulkDeleteResult *stats;
 	IndexBulkDeleteCallback callback;
 	void	   *callback_state;
 	GistNSN		startNSN;
-	BlockNumber totFreePages;	/* true total # of free pages */
 } GistVacState;
 
-static void gistvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
+static void gistvacuumscan(IndexVacuumInfo *info, GistBulkDeleteResult *stats,
 			   IndexBulkDeleteCallback callback, void *callback_state);
 static void gistvacuumpage(GistVacState *vstate, BlockNumber blkno,
 			   BlockNumber orig_blkno);
+static void gistvacuum_delete_empty_pages(GistBulkDeleteResult *stats);
+static bool gistdeletepage(GistBulkDeleteResult *stats,
+			   Buffer buffer, OffsetNumber downlink,
+			   Buffer leafBuffer);
+
+/* allocate the 'stats' struct that's kept over vacuum stages */
+static GistBulkDeleteResult *
+create_GistBulkDeleteResult(void)
+{
+	GistBulkDeleteResult *gist_stats;
+
+	gist_stats = (GistBulkDeleteResult *) palloc0(sizeof(GistBulkDeleteResult));
+	gist_stats->page_set_context =
+		GenerationContextCreate(CurrentMemoryContext,
+								"GiST VACUUM page set context",
+								16 * 1024);
+
+	return gist_stats;
+}
 
 /*
  * VACUUM bulkdelete stage: remove index entries.
@@ -44,21 +83,25 @@ IndexBulkDeleteResult *
 gistbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 			   IndexBulkDeleteCallback callback, void *callback_state)
 {
+	GistBulkDeleteResult *gist_stats = (GistBulkDeleteResult *) stats;
+
 	/* allocate stats if first time through, else re-use existing struct */
-	if (stats == NULL)
-		stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
+	if (gist_stats == NULL)
+		gist_stats = create_GistBulkDeleteResult();
 
-	gistvacuumscan(info, stats, callback, callback_state);
+	gistvacuumscan(info, gist_stats, callback, callback_state);
 
-	return stats;
+	return (IndexBulkDeleteResult *) gist_stats;
 }
 
 /*
- * VACUUM cleanup stage: update index statistics.
+ * VACUUM cleanup stage: delete empty pages, and update index statistics.
  */
 IndexBulkDeleteResult *
 gistvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
 {
+	GistBulkDeleteResult *gist_stats = (GistBulkDeleteResult *) stats;
+
 	/* No-op in ANALYZE ONLY mode */
 	if (info->analyze_only)
 		return stats;
@@ -68,13 +111,25 @@ gistvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
 	 * stats from the latest gistbulkdelete call.  If it wasn't called, we
 	 * still need to do a pass over the index, to obtain index statistics.
 	 */
-	if (stats == NULL)
+	if (gist_stats == NULL)
 	{
-		stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
-		gistvacuumscan(info, stats, NULL, NULL);
+		gist_stats = create_GistBulkDeleteResult();
+		gistvacuumscan(info, gist_stats, NULL, NULL);
 	}
 
 	/*
+	 * If we saw any empty pages, try to unlink them from the tree so that
+	 * they can be reused.
+	 */
+	gistvacuum_delete_empty_pages(gist_stats);
+
+	/* we don't need the internal and empty page sets anymore */
+	MemoryContextDelete(gist_stats->page_set_context);
+	gist_stats->page_set_context = NULL;
+	gist_stats->internal_page_set = NULL;
+	gist_stats->empty_leaf_set = NULL;
+
+	/*
 	 * It's quite possible for us to be fooled by concurrent page splits into
 	 * double-counting some index tuples, so disbelieve any total that exceeds
 	 * the underlying heap's count ... if we know that accurately.  Otherwise
@@ -82,11 +137,11 @@ gistvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
 	 */
 	if (!info->estimated_count)
 	{
-		if (stats->num_index_tuples > info->num_heap_tuples)
-			stats->num_index_tuples = info->num_heap_tuples;
+		if (gist_stats->stats.num_index_tuples > info->num_heap_tuples)
+			gist_stats->stats.num_index_tuples = info->num_heap_tuples;
 	}
 
-	return stats;
+	return (IndexBulkDeleteResult *) gist_stats;
 }
 
 /*
@@ -97,15 +152,16 @@ gistvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
  * btvacuumcleanup invoke this (the latter only if no btbulkdelete call
  * occurred).
  *
- * This also adds unused/delete pages to the free space map, although that
- * is currently not very useful.  There is currently no support for deleting
- * empty pages, so recycleable pages can only be found if an error occurs
- * while the index is being expanded, leaving an all-zeros page behind.
+ * This also makes note of any empty leaf pages, as well as all internal
+ * pages.  The second stage, gistvacuum_delete_empty_pages(), needs that
+ * information.  Any deleted pages are added directly to the free space map.
+ * (They should've been added there when they were originally deleted, already,
+ * but it's possible that the FSM was lost at a crash, for example.)
  *
  * The caller is responsible for initially allocating/zeroing a stats struct.
  */
 static void
-gistvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
+gistvacuumscan(IndexVacuumInfo *info, GistBulkDeleteResult *stats,
 			   IndexBulkDeleteCallback callback, void *callback_state)
 {
 	Relation	rel = info->index;
@@ -118,12 +174,16 @@ gistvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	 * Reset counts that will be incremented during the scan; needed in case
 	 * of multiple scans during a single VACUUM command.
 	 */
-	stats->estimated_count = false;
-	stats->num_index_tuples = 0;
-	stats->pages_deleted = 0;
+	stats->stats.estimated_count = false;
+	stats->stats.num_index_tuples = 0;
+	stats->stats.pages_deleted = 0;
+	stats->stats.pages_free = 0;
+	MemoryContextReset(stats->page_set_context);
+	stats->internal_page_set = intset_create();
+	stats->empty_leaf_set = intset_create();
 
 	/* Set up info to pass down to gistvacuumpage */
-	vstate.info = info;
+	stats->info = info;
 	vstate.stats = stats;
 	vstate.callback = callback;
 	vstate.callback_state = callback_state;
@@ -131,7 +191,6 @@ gistvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 		vstate.startNSN = GetInsertRecPtr();
 	else
 		vstate.startNSN = gistGetFakeLSN(rel);
-	vstate.totFreePages = 0;
 
 	/*
 	 * The outer loop iterates over all index pages, in physical order (we
@@ -188,12 +247,11 @@ gistvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	 * Note that if no recyclable pages exist, we don't bother vacuuming the
 	 * FSM at all.
 	 */
-	if (vstate.totFreePages > 0)
+	if (stats->stats.pages_free > 0)
 		IndexFreeSpaceMapVacuum(rel);
 
 	/* update statistics */
-	stats->num_pages = num_pages;
-	stats->pages_free = vstate.totFreePages;
+	stats->stats.num_pages = num_pages;
 }
 
 /*
@@ -210,8 +268,8 @@ gistvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 static void
 gistvacuumpage(GistVacState *vstate, BlockNumber blkno, BlockNumber orig_blkno)
 {
-	IndexVacuumInfo *info = vstate->info;
-	IndexBulkDeleteResult *stats = vstate->stats;
+	GistBulkDeleteResult *stats = vstate->stats;
+	IndexVacuumInfo *info = stats->info;
 	IndexBulkDeleteCallback callback = vstate->callback;
 	void	   *callback_state = vstate->callback_state;
 	Relation	rel = info->index;
@@ -235,17 +293,23 @@ restart:
 	LockBuffer(buffer, GIST_EXCLUSIVE);
 	page = (Page) BufferGetPage(buffer);
 
-	if (PageIsNew(page) || GistPageIsDeleted(page))
+	if (gistPageRecyclable(page))
 	{
 		/* Okay to recycle this page */
 		RecordFreeIndexPage(rel, blkno);
-		vstate->totFreePages++;
-		stats->pages_deleted++;
+		stats->stats.pages_free++;
+		stats->stats.pages_deleted++;
+	}
+	else if (GistPageIsDeleted(page))
+	{
+		/* Already deleted, but can't recycle yet */
+		stats->stats.pages_deleted++;
 	}
 	else if (GistPageIsLeaf(page))
 	{
 		OffsetNumber todelete[MaxOffsetNumber];
 		int			ntodelete = 0;
+		int			nremain;
 		GISTPageOpaque opaque = GistPageGetOpaque(page);
 		OffsetNumber maxoff = PageGetMaxOffsetNumber(page);
 
@@ -314,12 +378,27 @@ restart:
 
 			END_CRIT_SECTION();
 
-			stats->tuples_removed += ntodelete;
+			stats->stats.tuples_removed += ntodelete;
 			/* must recompute maxoff */
 			maxoff = PageGetMaxOffsetNumber(page);
 		}
 
-		stats->num_index_tuples += maxoff - FirstOffsetNumber + 1;
+		nremain = maxoff - FirstOffsetNumber + 1;
+		if (nremain == 0)
+		{
+			/*
+			 * The page is now completely empty.  Remember its block number,
+			 * so that we will try to delete the page in the second stage.
+			 *
+			 * Skip this when recursing, because IntegerSet requires that the
+			 * values are added in ascending order.  The next VACUUM will pick
+			 * it up.
+			 */
+			if (blkno == orig_blkno)
+				intset_add_member(stats->empty_leaf_set, blkno);
+		}
+		else
+			stats->stats.num_index_tuples += nremain;
 	}
 	else
 	{
@@ -347,6 +426,14 @@ restart:
 						 errdetail("This is caused by an incomplete page split at crash recovery before upgrading to PostgreSQL 9.1."),
 						 errhint("Please REINDEX it.")));
 		}
+
+		/*
+		 * Remember the block number of this page, so that we can revisit it
+		 * later in gistvacuum_delete_empty_pages(), when we search for
+		 * parents of empty leaf pages.
+		 */
+		if (blkno == orig_blkno)
+			intset_add_member(stats->internal_page_set, blkno);
 	}
 
 	UnlockReleaseBuffer(buffer);
@@ -364,3 +451,224 @@ restart:
 		goto restart;
 	}
 }
+
+/*
+ * Scan all internal pages, and try to delete their empty child pages.
+ */
+static void
+gistvacuum_delete_empty_pages(GistBulkDeleteResult *stats)
+{
+	IndexVacuumInfo *info = stats->info;
+	Relation	rel = info->index;
+	BlockNumber empty_pages_remaining;
+	uint64		blkno;
+
+	/*
+	 * Rescan all inner pages to find those that have empty child pages.
+	 */
+	empty_pages_remaining = intset_num_entries(stats->empty_leaf_set);
+	intset_begin_iterate(stats->internal_page_set);
+	while (empty_pages_remaining > 0 &&
+		   intset_iterate_next(stats->internal_page_set, &blkno))
+	{
+		Buffer		buffer;
+		Page		page;
+		OffsetNumber off,
+					maxoff;
+		OffsetNumber todelete[MaxOffsetNumber];
+		BlockNumber leafs_to_delete[MaxOffsetNumber];
+		int			ntodelete;
+		int			deleted;
+
+		buffer = ReadBufferExtended(rel, MAIN_FORKNUM, (BlockNumber) blkno,
+									RBM_NORMAL, info->strategy);
+
+		LockBuffer(buffer, GIST_SHARE);
+		page = (Page) BufferGetPage(buffer);
+
+		if (PageIsNew(page) || GistPageIsDeleted(page) || GistPageIsLeaf(page))
+		{
+			/*
+			 * This page was an internal page earlier, but now it's something
+			 * else. Shouldn't happen...
+			 */
+			Assert(false);
+			UnlockReleaseBuffer(buffer);
+			continue;
+		}
+
+		/*
+		 * Scan all the downlinks, and see if any of them point to empty leaf
+		 * pages.
+		 */
+		maxoff = PageGetMaxOffsetNumber(page);
+		ntodelete = 0;
+		for (off = FirstOffsetNumber;
+			 off <= maxoff && ntodelete < maxoff - 1;
+			 off = OffsetNumberNext(off))
+		{
+			ItemId		iid = PageGetItemId(page, off);
+			IndexTuple	idxtuple = (IndexTuple) PageGetItem(page, iid);
+			BlockNumber leafblk;
+
+			leafblk = ItemPointerGetBlockNumber(&(idxtuple->t_tid));
+			if (intset_is_member(stats->empty_leaf_set, leafblk))
+			{
+				leafs_to_delete[ntodelete] = leafblk;
+				todelete[ntodelete++] = off;
+			}
+		}
+
+		/*
+		 * In order to avoid deadlock, child page must be locked before
+		 * parent, so we must release the lock on the parent, lock the child,
+		 * and then re-acquire the lock the parent.  (And we wouldn't want to
+		 * do I/O, while holding a lock, anyway.)
+		 *
+		 * At the instant that we're not holding a lock on the parent, the
+		 * downlink might get moved by a concurrent insert, so we must
+		 * re-check that it still points to the same child page after we have
+		 * acquired both locks.  Also, another backend might have inserted a
+		 * tuple to the page, so that it is no longer empty.  gistdeletepage()
+		 * re-checks all these conditions.
+		 */
+		LockBuffer(buffer, GIST_UNLOCK);
+
+		deleted = 0;
+		for (int i = 0; i < ntodelete; i++)
+		{
+			Buffer		leafbuf;
+
+			/*
+			 * Don't remove the last downlink from the parent.  That would
+			 * confuse the insertion code.
+			 */
+			if (PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
+				break;
+
+			leafbuf = ReadBufferExtended(rel, MAIN_FORKNUM, leafs_to_delete[i],
+										 RBM_NORMAL, info->strategy);
+			LockBuffer(leafbuf, GIST_EXCLUSIVE);
+			gistcheckpage(rel, leafbuf);
+
+			LockBuffer(buffer, GIST_EXCLUSIVE);
+			if (gistdeletepage(stats, buffer, todelete[i] - deleted, leafbuf))
+				deleted++;
+			LockBuffer(buffer, GIST_UNLOCK);
+
+			UnlockReleaseBuffer(leafbuf);
+		}
+
+		ReleaseBuffer(buffer);
+
+		/* update stats */
+		stats->stats.pages_removed += deleted;
+
+		/*
+		 * We can stop the scan as soon as we have seen the downlinks, even if
+		 * we were not able to remove them all.
+		 */
+		empty_pages_remaining -= ntodelete;
+	}
+}
+
+/*
+ * gistdeletepage takes a leaf page, and its parent, and tries to delete the
+ * leaf.  Both pages must be locked.
+ *
+ * Even if the page was empty when we first saw it, a concurrent inserter might
+ * have added a tuple to it since.  Similarly, the downlink might have moved.
+ * We re-check all the conditions, to make sure the page is still deletable,
+ * before modifying anything.
+ *
+ * Returns true, if the page was deleted, and false if a concurrent update
+ * prevented it.
+ */
+static bool
+gistdeletepage(GistBulkDeleteResult *stats,
+			   Buffer parentBuffer, OffsetNumber downlink,
+			   Buffer leafBuffer)
+{
+	Page		parentPage = BufferGetPage(parentBuffer);
+	Page		leafPage = BufferGetPage(leafBuffer);
+	ItemId		iid;
+	IndexTuple	idxtuple;
+	XLogRecPtr	recptr;
+	TransactionId txid;
+
+	/*
+	 * Check that the leaf is still empty and deletable.
+	 */
+	if (!GistPageIsLeaf(leafPage))
+	{
+		/* a leaf page should never become a non-leaf page */
+		Assert(false);
+		return false;
+	}
+
+	if (GistFollowRight(leafPage))
+		return false;			/* don't mess with a concurrent page split */
+
+	if (PageGetMaxOffsetNumber(leafPage) != InvalidOffsetNumber)
+		return false;			/* not empty anymore */
+
+	/*
+	 * Ok, the leaf is deletable.  Is the downlink in the parent page still
+	 * valid?  It might have been moved by a concurrent insert.  We could try
+	 * to re-find it by scanning the page again, possibly moving right if the
+	 * was split.  But for now, let's keep it simple and just give up.  The
+	 * next VACUUM will pick it up.
+	 */
+	if (PageIsNew(parentPage) || GistPageIsDeleted(parentPage) ||
+		GistPageIsLeaf(parentPage))
+	{
+		/* shouldn't happen, internal pages are never deleted */
+		Assert(false);
+		return false;
+	}
+
+	if (PageGetMaxOffsetNumber(parentPage) < downlink
+		|| PageGetMaxOffsetNumber(parentPage) <= FirstOffsetNumber)
+		return false;
+
+	iid = PageGetItemId(parentPage, downlink);
+	idxtuple = (IndexTuple) PageGetItem(parentPage, iid);
+	if (BufferGetBlockNumber(leafBuffer) !=
+		ItemPointerGetBlockNumber(&(idxtuple->t_tid)))
+		return false;
+
+	/*
+	 * All good, proceed with the deletion.
+	 *
+	 * The page cannot be immediately recycled, because in-progress scans that
+	 * saw the downlink might still visit it.  Mark the page with the current
+	 * next-XID counter, so that we know when it can be recycled.  Once that
+	 * XID becomes older than GlobalXmin, we know that all scans that are
+	 * currently in progress must have ended.  (That's much more conservative
+	 * than needed, but let's keep it safe and simple.)
+	 */
+	txid = ReadNewTransactionId();
+
+	START_CRIT_SECTION();
+
+	/* mark the page as deleted */
+	MarkBufferDirty(leafBuffer);
+	GistPageSetDeleteXid(leafPage, txid);
+	GistPageSetDeleted(leafPage);
+	stats->stats.pages_deleted++;
+
+	/* remove the downlink from the parent */
+	MarkBufferDirty(parentBuffer);
+	PageIndexTupleDelete(parentPage, downlink);
+
+	if (RelationNeedsWAL(stats->info->index))
+		recptr = gistXLogPageDelete(leafBuffer, txid, parentBuffer, downlink);
+	else
+		recptr = gistGetFakeLSN(stats->info->index);
+	PageSetLSN(parentPage, recptr);
+	PageSetLSN(leafPage, recptr);
+
+	END_CRIT_SECTION();
+
+	return true;
+}