v5-0002-Recycle-pages-deleted-during-same-VACUUM.patch

From 435f9b4d82e70665603a7c0eb19c9065010eb5d0 Mon Sep 17 00:00:00 2001
From: Peter Geoghegan <pg@bowt.ie>
Date: Sun, 7 Feb 2021 19:24:03 -0800
Subject: [PATCH v5 2/3] Recycle pages deleted during same VACUUM.

Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-Wzk76_P=67iUscb1UN44-gyZL-KgpsXbSxq_bdcMa7Q+wQ@mail.gmail.com
---
 src/include/access/nbtree.h         |  38 +++++++++-
 src/backend/access/nbtree/README    |  31 ++++++++
 src/backend/access/nbtree/nbtpage.c |  90 ++++++++++++++++------
 src/backend/access/nbtree/nbtree.c  | 111 ++++++++++++++++++++++++----
 4 files changed, 229 insertions(+), 41 deletions(-)

diff --git a/src/include/access/nbtree.h b/src/include/access/nbtree.h
index 901b6f4dc8..5c197fc5c1 100644
--- a/src/include/access/nbtree.h
+++ b/src/include/access/nbtree.h
@@ -280,7 +280,8 @@ BTPageGetDeleteXid(Page page)
  * Is an existing page recyclable?
  *
  * This exists to centralize the policy on which deleted pages are now safe to
- * re-use.
+ * re-use.  The _bt_newly_deleted_pages_recycle() optimization behaves more
+ * aggressively, though that has certain known limitations.
  *
  * Note: PageIsNew() pages are always safe to recycle, but we can't deal with
  * them here (caller is responsible for that case themselves).  Caller might
@@ -313,6 +314,39 @@ BTPageIsRecyclable(Page page)
 	return false;
 }
 
+/*
+ * BTVacState is nbtree.c state used during VACUUM.  It is exported for use by
+ * page deletion related code in nbtpage.c.
+ */
+typedef struct BTPendingRecycle
+{
+	BlockNumber blkno;
+	FullTransactionId safexid;
+} BTPendingRecycle;
+
+typedef struct BTVacState
+{
+	/*
+	 * VACUUM operation state
+	 */
+	IndexVacuumInfo *info;
+	IndexBulkDeleteResult *stats;
+	IndexBulkDeleteCallback callback;
+	void	   *callback_state;
+	BTCycleId	cycleid;
+
+	/*
+	 * Page deletion state for VACUUM
+	 */
+	MemoryContext pagedelcontext;
+	BTPendingRecycle *deleted;
+	bool		grow;
+	bool		full;
+	uint32		ndeletedspace;
+	uint64		maxndeletedspace;
+	uint32		ndeleted;
+} BTVacState;
+
 /*
  *	Lehman and Yao's algorithm requires a ``high key'' on every non-rightmost
  *	page.  The high key is not a tuple that is used to visit the heap.  It is
@@ -1182,7 +1216,7 @@ extern void _bt_delitems_vacuum(Relation rel, Buffer buf,
 extern void _bt_delitems_delete_check(Relation rel, Buffer buf,
 									  Relation heapRel,
 									  TM_IndexDeleteOp *delstate);
-extern uint32 _bt_pagedel(Relation rel, Buffer leafbuf);
+extern void _bt_pagedel(Relation rel, Buffer leafbuf, BTVacState *vstate);
 
 /*
  * prototypes for functions in nbtsearch.c
diff --git a/src/backend/access/nbtree/README b/src/backend/access/nbtree/README
index 46d49bf025..265814ea46 100644
--- a/src/backend/access/nbtree/README
+++ b/src/backend/access/nbtree/README
@@ -430,6 +430,37 @@ whenever it is subsequently taken from the FSM for reuse.  The deleted
 page's contents will be overwritten by the split operation (it will become
 the new right sibling page).
 
+Prior to PostgreSQL 14, VACUUM was only able to recycle pages that were
+deleted by a previous VACUUM operation (VACUUM typically placed all pages
+deleted by the last VACUUM into the FSM, though there were and are no
+certainties here).  This had the obvious disadvantage of creating
+uncertainty about when and how pages get recycled, especially with bursty
+workloads.  It was naive, even within the constraints of the design, since
+there is no reason to think that it will take long for a deleted page to
+become recyclable.  It's convenient to use XIDs to implement the drain
+technique, but that is totally unrelated to any of the other things that
+VACUUM needs to do with XIDs.
+
+VACUUM operations now consider if it's possible to recycle any pages that
+the same operation deleted after the physical scan of the index, the last
+point it's convenient to do one last check.  This changes nothing about
+the basic design, and so it might still not be possible to recycle any
+pages at that time (e.g., there might not even be one single new
+transactions after an index page deletion, but before VACUUM ends).  But
+we have little to lose and plenty to gain by trying.  We only need to keep
+around a little information about recently deleted pages in local memory.
+We don't even have to access the deleted pages a second time.
+
+Currently VACUUM delays considering the possibility of recycling its own
+recently deleted page until the end of its btbulkdelete scan (or until the
+end of btvacuumcleanup in cases where there were no tuples to delete in
+the index).  It would be slightly more effective if btbulkdelete page
+deletions were deferred until btvacuumcleanup, simply because more time
+will have passed.  Our current approach works well enough in practice,
+especially in cases where it really matters: cases where we're vacuuming a
+large index, where recycling pages sooner rather than later is
+particularly likely to matter.
+
 Fastpath For Index Insertion
 ----------------------------
 
diff --git a/src/backend/access/nbtree/nbtpage.c b/src/backend/access/nbtree/nbtpage.c
index 8ae16428d7..55395c87c1 100644
--- a/src/backend/access/nbtree/nbtpage.c
+++ b/src/backend/access/nbtree/nbtpage.c
@@ -50,7 +50,7 @@ static bool _bt_mark_page_halfdead(Relation rel, Buffer leafbuf,
 static bool _bt_unlink_halfdead_page(Relation rel, Buffer leafbuf,
 									 BlockNumber scanblkno,
 									 bool *rightsib_empty,
-									 uint32 *ndeleted);
+									 BTVacState *vstate);
 static bool _bt_lock_subtree_parent(Relation rel, BlockNumber child,
 									BTStack stack,
 									Buffer *subtreeparent,
@@ -1761,20 +1761,22 @@ _bt_rightsib_halfdeadflag(Relation rel, BlockNumber leafrightsib)
  * should never pass a buffer containing an existing deleted page here.  The
  * lock and pin on caller's buffer will be dropped before we return.
  *
- * Returns the number of pages successfully deleted (zero if page cannot
- * be deleted now; could be more than one if parent or right sibling pages
- * were deleted too).  Note that this does not include pages that we delete
- * that the btvacuumscan scan has yet to reach; they'll get counted later
- * instead.
+ * Maintains bulk delete stats for caller, which are taken from vstate.  We
+ * need to cooperate closely with caller here so that whole VACUUM operation
+ * reliably avoids any double counting of subsidiary-to-leafbuf pages that we
+ * delete in passing.  If such pages happen to be from a block number that is
+ * ahead of the current scanblkno position, then caller is expected to count
+ * them directly later on.  It's simpler for us to understand caller's
+ * requirements than it would be for caller to understand when or how a
+ * deleted page became deleted after the fact.
  *
  * NOTE: this leaks memory.  Rather than trying to clean up everything
  * carefully, it's better to run it in a temp context that can be reset
  * frequently.
  */
-uint32
-_bt_pagedel(Relation rel, Buffer leafbuf)
+void
+_bt_pagedel(Relation rel, Buffer leafbuf, BTVacState *vstate)
 {
-	uint32		ndeleted = 0;
 	BlockNumber rightsib;
 	bool		rightsib_empty;
 	Page		page;
@@ -1782,7 +1784,8 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 
 	/*
 	 * Save original leafbuf block number from caller.  Only deleted blocks
-	 * that are <= scanblkno get counted in ndeleted return value.
+	 * that are <= scanblkno are added to bulk delete stat's pages_deleted
+	 * count.
 	 */
 	BlockNumber scanblkno = BufferGetBlockNumber(leafbuf);
 
@@ -1844,7 +1847,7 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 										 RelationGetRelationName(rel))));
 
 			_bt_relbuf(rel, leafbuf);
-			return ndeleted;
+			return;
 		}
 
 		/*
@@ -1874,7 +1877,7 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 			Assert(!P_ISHALFDEAD(opaque));
 
 			_bt_relbuf(rel, leafbuf);
-			return ndeleted;
+			return;
 		}
 
 		/*
@@ -1923,8 +1926,7 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 				if (_bt_leftsib_splitflag(rel, leftsib, leafblkno))
 				{
 					ReleaseBuffer(leafbuf);
-					Assert(ndeleted == 0);
-					return ndeleted;
+					return;
 				}
 
 				/* we need an insertion scan key for the search, so build one */
@@ -1965,7 +1967,7 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 			if (!_bt_mark_page_halfdead(rel, leafbuf, stack))
 			{
 				_bt_relbuf(rel, leafbuf);
-				return ndeleted;
+				return;
 			}
 		}
 
@@ -1980,7 +1982,7 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 		{
 			/* Check for interrupts in _bt_unlink_halfdead_page */
 			if (!_bt_unlink_halfdead_page(rel, leafbuf, scanblkno,
-										  &rightsib_empty, &ndeleted))
+										  &rightsib_empty, vstate))
 			{
 				/*
 				 * _bt_unlink_halfdead_page should never fail, since we
@@ -1991,7 +1993,7 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 				 * lock and pin on leafbuf for us.
 				 */
 				Assert(false);
-				return ndeleted;
+				return;
 			}
 		}
 
@@ -2027,8 +2029,6 @@ _bt_pagedel(Relation rel, Buffer leafbuf)
 
 		leafbuf = _bt_getbuf(rel, rightsib, BT_WRITE);
 	}
-
-	return ndeleted;
 }
 
 /*
@@ -2263,9 +2263,10 @@ _bt_mark_page_halfdead(Relation rel, Buffer leafbuf, BTStack stack)
  */
 static bool
 _bt_unlink_halfdead_page(Relation rel, Buffer leafbuf, BlockNumber scanblkno,
-						 bool *rightsib_empty, uint32 *ndeleted)
+						 bool *rightsib_empty, BTVacState *vstate)
 {
 	BlockNumber leafblkno = BufferGetBlockNumber(leafbuf);
+	IndexBulkDeleteResult *stats = vstate->stats;
 	BlockNumber leafleftsib;
 	BlockNumber leafrightsib;
 	BlockNumber target;
@@ -2673,12 +2674,53 @@ _bt_unlink_halfdead_page(Relation rel, Buffer leafbuf, BlockNumber scanblkno,
 		_bt_relbuf(rel, buf);
 
 	/*
-	 * If btvacuumscan won't revisit this page in a future btvacuumpage call
-	 * and count it as deleted then, we count it as deleted by current
-	 * btvacuumpage call
+	 * Maintain pages_deleted in a way that takes into account how
+	 * btvacuumpage() will count deleted pages that have yet to become
+	 * scanblkno -- only count page when it's not going to get that treatment
+	 * later on.
 	 */
 	if (target <= scanblkno)
-		(*ndeleted)++;
+		stats->pages_deleted++;
+
+	/*
+	 * Maintain array of pages that were deleted during current btvacuumscan()
+	 * call.  We may well be able to recycle them in a separate pass at the
+	 * end of the current btvacuumscan().
+	 *
+	 * Need to respect work_mem/maxndeletedspace limitation on size of deleted
+	 * array.  Our strategy when the array can no longer grow within the
+	 * bounds of work_mem is simple: keep earlier entries (which are likelier
+	 * to be recyclable in the end), but stop saving new entries.
+	 */
+	if (vstate->full)
+		return true;
+
+	if (vstate->ndeleted >= vstate->ndeletedspace)
+	{
+		uint64 newndeletedspace;
+
+		if (!vstate->grow)
+		{
+			vstate->full = true;
+			return true;
+		}
+
+		newndeletedspace = vstate->ndeletedspace * 2;
+		if (newndeletedspace > vstate->maxndeletedspace)
+		{
+			newndeletedspace = vstate->maxndeletedspace;
+			vstate->grow = false;
+		}
+		vstate->ndeletedspace = newndeletedspace;
+
+		vstate->deleted =
+			repalloc(vstate->deleted,
+					 sizeof(BTPendingRecycle) * vstate->ndeletedspace);
+	}
+
+	vstate->deleted[vstate->ndeleted].blkno = target;
+	vstate->deleted[vstate->ndeleted].safexid = safexid;
+	vstate->ndeleted++;
 
 	return true;
 }
diff --git a/src/backend/access/nbtree/nbtree.c b/src/backend/access/nbtree/nbtree.c
index b5a674d9e0..b18022936b 100644
--- a/src/backend/access/nbtree/nbtree.c
+++ b/src/backend/access/nbtree/nbtree.c
@@ -21,7 +21,9 @@
 #include "access/nbtree.h"
 #include "access/nbtxlog.h"
 #include "access/relscan.h"
+#include "access/table.h"
 #include "access/xlog.h"
+#include "catalog/index.h"
 #include "commands/progress.h"
 #include "commands/vacuum.h"
 #include "miscadmin.h"
@@ -32,23 +34,13 @@
 #include "storage/indexfsm.h"
 #include "storage/ipc.h"
 #include "storage/lmgr.h"
+#include "storage/procarray.h"
 #include "storage/smgr.h"
 #include "utils/builtins.h"
 #include "utils/index_selfuncs.h"
 #include "utils/memutils.h"
 
 
-/* Working state needed by btvacuumpage */
-typedef struct
-{
-	IndexVacuumInfo *info;
-	IndexBulkDeleteResult *stats;
-	IndexBulkDeleteCallback callback;
-	void	   *callback_state;
-	BTCycleId	cycleid;
-	MemoryContext pagedelcontext;
-} BTVacState;
-
 /*
  * BTPARALLEL_NOT_INITIALIZED indicates that the scan has not started.
  *
@@ -868,6 +860,68 @@ _bt_vacuum_needs_cleanup(IndexVacuumInfo *info)
 	return false;
 }
 
+/*
+ * _bt_newly_deleted_pages_recycle() -- Are _bt_pagedel pages recyclable now?
+ *
+ * Note that we assume that the array is ordered by safexid.  No further
+ * entries can be safe to recycle once we encounter the first non-recyclable
+ * entry in the deleted array.
+ */
+static inline void
+_bt_newly_deleted_pages_recycle(Relation rel, BTVacState *vstate)
+{
+	IndexBulkDeleteResult *stats = vstate->stats;
+	Relation	heapRel;
+
+	/*
+	 * Recompute VACUUM XID boundaries.
+	 *
+	 * We don't actually care about the oldest non-removable XID.  Computing
+	 * the oldest such XID has a useful side-effect: It updates the procarray
+	 * state that tracks XID horizon.  This is not just an optimization; it's
+	 * essential.  It allows the GlobalVisCheckRemovableFullXid() calls we
+	 * make here to notice if and when safexid values from pages this same
+	 * VACUUM operation deleted are sufficiently old to allow recycling to
+	 * take place safely.
+	 */
+	GetOldestNonRemovableTransactionId(NULL);
+
+	/*
+	 * Use the heap relation for GlobalVisCheckRemovableFullXid() calls (don't
+	 * pass NULL rel argument).
+	 *
+	 * This is an optimization; it allows us to be much more aggressive in
+	 * cases involving logical decoding (unless this happens to be a system
+	 * catalog).  We don't simply use BTPageIsRecyclable().
+	 *
+	 * XXX: The BTPageIsRecyclable() criteria creates problems for this
+	 * optimization.  Its safexid test is applied in a redundant manner within
+	 * _bt_getbuf() (via its BTPageIsRecyclable() call).  Consequently,
+	 * _bt_getbuf() may believe that it is still unsafe to recycle a page that
+	 * we know to be recycle safe -- in which case it is unnecessarily
+	 * discarded.
+	 *
+	 * We should get around to fixing this _bt_getbuf() issue some day.  For
+	 * now we can still proceed in the hopes that BTPageIsRecyclable() will
+	 * catch up with us before _bt_getbuf() ever reaches the page.
+	 */
+	heapRel = table_open(IndexGetRelation(RelationGetRelid(rel), false),
+						 AccessShareLock);
+	for (int i = 0; i < vstate->ndeleted; i++)
+	{
+		BlockNumber blkno = vstate->deleted[i].blkno;
+		FullTransactionId safexid = vstate->deleted[i].safexid;
+
+		if (!GlobalVisCheckRemovableFullXid(heapRel, safexid))
+			break;
+
+		RecordFreeIndexPage(rel, blkno);
+		stats->pages_free++;
+	}
+
+	table_close(heapRel, AccessShareLock);
+}
+
 /*
  * Bulk deletion of all index entries pointing to a set of heap tuples.
  * The set of target tuples is specified via a callback routine that tells
@@ -953,6 +1007,14 @@ btvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
 	 * _bt_vacuum_needs_cleanup() to force the next VACUUM to proceed with a
 	 * btvacuumscan() call.
 	 *
+	 * Note: Prior to PostgreSQL 14, we were completely reliant on the next
+	 * VACUUM operation taking care of recycling whatever pages the current
+	 * VACUUM operation found to be empty and then deleted.  It is now usually
+	 * possible for _bt_newly_deleted_pages_recycle() to recycle all of the
+	 * pages that any given VACUUM operation deletes, as part of the same
+	 * VACUUM operation.  As a result, it is rare for num_delpages to actually
+	 * exceed 0, including with indexes where page deletions are frequent.
+	 *
 	 * Note: We must delay the _bt_set_cleanup_info() call until this late
 	 * stage of VACUUM (the btvacuumcleanup() phase), to keep num_heap_tuples
 	 * accurate.  The btbulkdelete()-time num_heap_tuples value is generally
@@ -1041,6 +1103,16 @@ btvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 												  "_bt_pagedel",
 												  ALLOCSET_DEFAULT_SIZES);
 
+	/* Allocate _bt_newly_deleted_pages_recycle related information */
+	vstate.ndeletedspace = 512;
+	vstate.grow = true;
+	vstate.full = false;
+	vstate.maxndeletedspace = ((work_mem * 1024L) / sizeof(BTPendingRecycle));
+	vstate.maxndeletedspace = Min(vstate.maxndeletedspace, MaxBlockNumber);
+	vstate.maxndeletedspace = Max(vstate.maxndeletedspace, vstate.ndeletedspace);
+	vstate.ndeleted = 0;
+	vstate.deleted = palloc(sizeof(BTPendingRecycle) * vstate.ndeletedspace);
+
 	/*
 	 * The outer loop iterates over all index pages except the metapage, in
 	 * physical order (we hope the kernel will cooperate in providing
@@ -1109,7 +1181,18 @@ btvacuumscan(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	 *
 	 * Note that if no recyclable pages exist, we don't bother vacuuming the
 	 * FSM at all.
+	 *
+	 * Before vacuuming the FSM, try to make the most of the pages we
+	 * ourselves deleted: see if they can be recycled already (try to avoid
+	 * waiting until the next VACUUM operation to recycle).  Our approach is
+	 * to check the local array of pages that were newly deleted during this
+	 * VACUUM.
 	 */
+	if (vstate.ndeleted > 0)
+		_bt_newly_deleted_pages_recycle(rel, &vstate);
+
+	pfree(vstate.deleted);
+
 	if (stats->pages_free > 0)
 		IndexFreeSpaceMapVacuum(rel);
 }
@@ -1448,12 +1531,10 @@ backtrack:
 		oldcontext = MemoryContextSwitchTo(vstate->pagedelcontext);
 
 		/*
-		 * We trust the _bt_pagedel return value because it does not include
-		 * any page that a future call here from btvacuumscan is expected to
-		 * count.  There will be no double-counting.
+		 * _bt_pagedel maintains the bulk delete stats on our behalf
 		 */
 		Assert(blkno == scanblkno);
-		stats->pages_deleted += _bt_pagedel(rel, buf);
+		_bt_pagedel(rel, buf, vstate);
 
 		MemoryContextSwitchTo(oldcontext);
 		/* pagedel released buffer, so we shouldn't */
-- 
2.27.0