Allow on-access pruning to set pages all-visible

Many queries do not modify the underlying relation. For such queries, if
on-access pruning occurs during the scan, we can check whether the page
has become all-visible and update the visibility map accordingly.
Previously, only vacuum and COPY FREEZE marked pages as all-visible or
all-frozen.

Supporting this requires passing information about whether the relation
is modified from the executor down to the scan descriptor.

This commit implements on-access VM setting for sequential scans as well
as for the underlying heap relation in index scans and bitmap heap
scans.

Author: melanieplageman

src/backend/access/heap/heapam.c

@@ -555,6 +555,7 @@ heap_prepare_pagescan(TableScanDesc sscan)
 	Buffer		buffer = scan->rs_cbuf;
 	BlockNumber block = scan->rs_cblock;
 	Snapshot	snapshot;
+	Buffer	   *vmbuffer = NULL;
 	Page		page;
 	int			lines;
 	bool		all_visible;
@@ -569,7 +570,9 @@ heap_prepare_pagescan(TableScanDesc sscan)
 	/*
 	 * Prune and repair fragmentation for the whole page, if possible.
 	 */
-	heap_page_prune_opt(scan->rs_base.rs_rd, buffer);
+	if (sscan->rs_flags & SO_ALLOW_VM_SET)
+		vmbuffer = &scan->rs_vmbuffer;
+	heap_page_prune_opt(scan->rs_base.rs_rd, buffer, vmbuffer);
 
 	/*
 	 * We must hold share lock on the buffer content while examining tuple
@@ -1246,6 +1249,7 @@ heap_beginscan(Relation relation, Snapshot snapshot,
 														  sizeof(TBMIterateResult));
 	}
 
+	scan->rs_vmbuffer = InvalidBuffer;
 
 	return (TableScanDesc) scan;
 }
@@ -1284,6 +1288,12 @@ heap_rescan(TableScanDesc sscan, ScanKey key, bool set_params,
 		scan->rs_cbuf = InvalidBuffer;
 	}
 
+	if (BufferIsValid(scan->rs_vmbuffer))
+	{
+		ReleaseBuffer(scan->rs_vmbuffer);
+		scan->rs_vmbuffer = InvalidBuffer;
+	}
+
 	/*
 	 * SO_TYPE_BITMAPSCAN would be cleaned up here, but it does not hold any
 	 * additional data vs a normal HeapScan
@@ -1316,6 +1326,9 @@ heap_endscan(TableScanDesc sscan)
 	if (BufferIsValid(scan->rs_cbuf))
 		ReleaseBuffer(scan->rs_cbuf);
 
+	if (BufferIsValid(scan->rs_vmbuffer))
+		ReleaseBuffer(scan->rs_vmbuffer);
+
 	/*
 	 * Must free the read stream before freeing the BufferAccessStrategy.
 	 */

src/backend/access/heap/heapam_handler.c

@@ -85,6 +85,7 @@ heapam_index_fetch_begin(Relation rel)
 
 	hscan->xs_base.rel = rel;
 	hscan->xs_cbuf = InvalidBuffer;
+	hscan->xs_vmbuffer = InvalidBuffer;
 
 	return &hscan->xs_base;
 }
@@ -99,6 +100,12 @@ heapam_index_fetch_reset(IndexFetchTableData *scan)
 		ReleaseBuffer(hscan->xs_cbuf);
 		hscan->xs_cbuf = InvalidBuffer;
 	}
+
+	if (BufferIsValid(hscan->xs_vmbuffer))
+	{
+		ReleaseBuffer(hscan->xs_vmbuffer);
+		hscan->xs_vmbuffer = InvalidBuffer;
+	}
 }
 
 static void
@@ -138,7 +145,8 @@ heapam_index_fetch_tuple(struct IndexFetchTableData *scan,
 		 * Prune page, but only if we weren't already on this page
 		 */
 		if (prev_buf != hscan->xs_cbuf)
-			heap_page_prune_opt(hscan->xs_base.rel, hscan->xs_cbuf);
+			heap_page_prune_opt(hscan->xs_base.rel, hscan->xs_cbuf,
+								scan->modifies_base_rel ? NULL : &hscan->xs_vmbuffer);
 	}
 
 	/* Obtain share-lock on the buffer so we can examine visibility */
@@ -2471,6 +2479,7 @@ BitmapHeapScanNextBlock(TableScanDesc scan,
 	TBMIterateResult *tbmres;
 	OffsetNumber offsets[TBM_MAX_TUPLES_PER_PAGE];
 	int			noffsets = -1;
+	Buffer	   *vmbuffer = NULL;
 
 	Assert(scan->rs_flags & SO_TYPE_BITMAPSCAN);
 	Assert(hscan->rs_read_stream);
@@ -2517,7 +2526,9 @@ BitmapHeapScanNextBlock(TableScanDesc scan,
 	/*
 	 * Prune and repair fragmentation for the whole page, if possible.
 	 */
-	heap_page_prune_opt(scan->rs_rd, buffer);
+	if (scan->rs_flags & SO_ALLOW_VM_SET)
+		vmbuffer = &hscan->rs_vmbuffer;
+	heap_page_prune_opt(scan->rs_rd, buffer, vmbuffer);
 
 	/*
 	 * We must hold share lock on the buffer content while examining tuple

src/backend/access/heap/pruneheap.c

@@ -188,7 +188,9 @@ static bool heap_page_will_set_vis(Relation relation,
 								   Buffer heap_buf,
 								   Buffer vmbuffer,
 								   bool blk_known_av,
-								   const PruneState *prstate,
+								   PruneReason reason,
+								   bool do_prune, bool do_freeze,
+								   PruneState *prstate,
 								   uint8 *vmflags,
 								   bool *do_set_pd_vis);
 
@@ -203,9 +205,13 @@ static bool heap_page_will_set_vis(Relation relation,
  * if there's not any use in pruning.
  *
  * Caller must have pin on the buffer, and must *not* have a lock on it.
+ *
+ * If vmbuffer is not NULL, it is okay for pruning to set the visibility map if
+ * the page is all-visible. We will take care of pinning and, if needed,
+ * reading in the page of the visibility map.
  */
 void
-heap_page_prune_opt(Relation relation, Buffer buffer)
+heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer)
 {
 	Page		page = BufferGetPage(buffer);
 	TransactionId prune_xid;
@@ -271,12 +277,21 @@ heap_page_prune_opt(Relation relation, Buffer buffer)
 			PruneFreezeParams params;
 			PruneFreezeResult presult;
 
+			params.options = 0;
+			params.vmbuffer = InvalidBuffer;
+
+			if (vmbuffer)
+			{
+				visibilitymap_pin(relation, BufferGetBlockNumber(buffer), vmbuffer);
+				params.options = HEAP_PAGE_PRUNE_UPDATE_VIS;
+				params.vmbuffer = *vmbuffer;
+			}
+
 			params.relation = relation;
 			params.buffer = buffer;
 			params.reason = PRUNE_ON_ACCESS;
 			params.vistest = vistest;
 			params.cutoffs = NULL;
-			params.vmbuffer = InvalidBuffer;
 			params.blk_known_av = false;
 
 			/*
@@ -456,6 +471,9 @@ heap_page_will_freeze(Relation relation, Buffer buffer,
  * have examined this page’s VM bits (e.g., VACUUM in the previous
  * heap_vac_scan_next_block() call) and can pass that along.
  *
+ * This should be called only after do_freeze has been decided (and do_prune
+ * has been set), as these factor into our heuristic-based decision.
+ *
  * Returns true if one or both VM bits should be set, along with the desired
  * flags in *vmflags. Also indicates via do_set_pd_vis whether PD_ALL_VISIBLE
  * should be set on the heap page.
@@ -466,7 +484,9 @@ heap_page_will_set_vis(Relation relation,
 					   Buffer heap_buf,
 					   Buffer vmbuffer,
 					   bool blk_known_av,
-					   const PruneState *prstate,
+					   PruneReason reason,
+					   bool do_prune, bool do_freeze,
+					   PruneState *prstate,
 					   uint8 *vmflags,
 					   bool *do_set_pd_vis)
 {
@@ -482,6 +502,23 @@ heap_page_will_set_vis(Relation relation,
 		return false;
 	}
 
+	/*
+	 * If this is an on-access call and we're not actually pruning, avoid
+	 * setting the visibility map if it would newly dirty the heap page or, if
+	 * the page is already dirty, if doing so would require including a
+	 * full-page image (FPI) of the heap page in the WAL. This situation
+	 * should be rare, as on-access pruning is only attempted when
+	 * pd_prune_xid is valid.
+	 */
+	if (reason == PRUNE_ON_ACCESS &&
+		prstate->all_visible &&
+		!do_prune && !do_freeze &&
+		(!BufferIsDirty(heap_buf) || XLogCheckBufferNeedsBackup(heap_buf)))
+	{
+		prstate->all_visible = prstate->all_frozen = false;
+		return false;
+	}
+
 	if (prstate->all_visible && !PageIsAllVisible(heap_page))
 		*do_set_pd_vis = true;
 
@@ -505,6 +542,11 @@ heap_page_will_set_vis(Relation relation,
 	 * page-level bit is clear.  However, it's possible that in vacuum the bit
 	 * got cleared after heap_vac_scan_next_block() was called, so we must
 	 * recheck with buffer lock before concluding that the VM is corrupt.
+	 *
+	 * This will never trigger for on-access pruning because it couldn't have
+	 * done a previous visibility map lookup and thus will always pass
+	 * blk_known_av as false. A future vacuum will have to take care of fixing
+	 * the corruption.
 	 */
 	else if (blk_known_av && !PageIsAllVisible(heap_page) &&
 			 visibilitymap_get_status(relation, heap_blk, &vmbuffer) != 0)
@@ -913,6 +955,14 @@ heap_page_prune_and_freeze(PruneFreezeParams *params,
 		prstate.ndead > 0 ||
 		prstate.nunused > 0;
 
+	/*
+	 * Even if we don't prune anything, if we found a new value for the
+	 * pd_prune_xid field or the page was marked full, we will update the hint
+	 * bit.
+	 */
+	do_hint_prune = ((PageHeader) page)->pd_prune_xid != prstate.new_prune_xid ||
+		PageIsFull(page);
+
 	/*
 	 * After processing all the live tuples on the page, if the newest xmin
 	 * amongst them is not visible to everyone, the page cannot be
@@ -923,14 +973,6 @@ heap_page_prune_and_freeze(PruneFreezeParams *params,
 		!GlobalVisXidVisibleToAll(prstate.vistest, prstate.visibility_cutoff_xid))
 		prstate.all_visible = prstate.all_frozen = false;
 
-	/*
-	 * Even if we don't prune anything, if we found a new value for the
-	 * pd_prune_xid field or the page was marked full, we will update the hint
-	 * bit.
-	 */
-	do_hint_prune = ((PageHeader) page)->pd_prune_xid != prstate.new_prune_xid ||
-		PageIsFull(page);
-
 	/*
 	 * Decide if we want to go ahead with freezing according to the freeze
 	 * plans we prepared, or not.
@@ -974,6 +1016,7 @@ heap_page_prune_and_freeze(PruneFreezeParams *params,
 	 */
 	do_set_vm = heap_page_will_set_vis(params->relation,
 									   blockno, buffer, vmbuffer, params->blk_known_av,
+									   params->reason, do_prune, do_freeze,
 									   &prstate, &new_vmbits, &do_set_pd_vis);
 
 	/* We should only set the VM if PD_ALL_VISIBLE is set or will be */
@@ -2250,7 +2293,7 @@ heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
 
 /*
  * Calculate the conflict horizon for the whole XLOG_HEAP2_PRUNE_VACUUM_SCAN
- * record.
+ * or XLOG_HEAP2_PRUNE_ON_ACCESS record.
  */
 static TransactionId
 get_conflict_xid(bool do_prune, bool do_freeze, bool do_set_vm,
@@ -2319,8 +2362,8 @@ get_conflict_xid(bool do_prune, bool do_freeze, bool do_set_vm,
  * - Reaping: During vacuum phase III, items that are already LP_DEAD are
  *   marked as unused.
  *
- * - VM updates: After vacuum phases I and III, the heap page may be marked
- *   all-visible and all-frozen.
+ * - VM updates: After vacuum phases I and III and on-access, the heap page
+ *   may be marked all-visible and all-frozen.
  *
  * These changes all happen together, so we use a single WAL record for them
  * all.

src/backend/access/index/indexam.c

@@ -289,6 +289,32 @@ index_beginscan(Relation heapRelation,
 	return scan;
 }
 
+/*
+ * Similar to index_beginscan(), but allows the caller to indicate whether the
+ * query modifies the underlying base relation. This is used when the caller
+ * wants to attempt marking pages in the base relation as all-visible in the
+ * visibility map during on-access pruning.
+ */
+IndexScanDesc
+index_beginscan_vmset(Relation heapRelation,
+					  Relation indexRelation,
+					  Snapshot snapshot,
+					  IndexScanInstrumentation *instrument,
+					  int nkeys, int norderbys, bool modifies_base_rel)
+{
+	IndexScanDesc scan;
+
+	scan = index_beginscan(heapRelation,
+						   indexRelation,
+						   snapshot,
+						   instrument,
+						   nkeys, norderbys);
+
+	scan->xs_heapfetch->modifies_base_rel = modifies_base_rel;
+
+	return scan;
+}
+
 /*
  * index_beginscan_bitmap - start a scan of an index with amgetbitmap
  *
@@ -620,6 +646,26 @@ index_beginscan_parallel(Relation heaprel, Relation indexrel,
 	return scan;
 }
 
+/*
+ * Parallel version of index_beginscan_vmset()
+ */
+IndexScanDesc
+index_beginscan_parallel_vmset(Relation heaprel, Relation indexrel,
+							   IndexScanInstrumentation *instrument,
+							   int nkeys, int norderbys,
+							   ParallelIndexScanDesc pscan,
+							   bool modifies_base_rel)
+{
+	IndexScanDesc scan;
+
+	scan = index_beginscan_parallel(heaprel, indexrel,
+									instrument,
+									nkeys, norderbys,
+									pscan);
+	scan->xs_heapfetch->modifies_base_rel = modifies_base_rel;
+	return scan;
+}
+
 /* ----------------
  * index_getnext_tid - get the next TID from a scan
  *

src/backend/access/table/tableam.c

@@ -49,6 +49,10 @@
 char	   *default_table_access_method = DEFAULT_TABLE_ACCESS_METHOD;
 bool		synchronize_seqscans = true;
 
+/* Helper for table_beginscan_parallel() and table_beginscan_parallel_vmset() */
+static TableScanDesc table_beginscan_parallel_common(Relation relation, ParallelTableScanDesc pscan,
+													 uint32 flags);
+
 
 /* ----------------------------------------------------------------------------
  * Slot functions.
@@ -162,12 +166,14 @@ table_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan,
 	}
 }
 
-TableScanDesc
-table_beginscan_parallel(Relation relation, ParallelTableScanDesc pscan)
+/*
+ * Common helper for table_beginscan_parallel() and table_beginscan_parallel_vmset()
+ */
+static TableScanDesc
+table_beginscan_parallel_common(Relation relation, ParallelTableScanDesc pscan,
+								uint32 flags)
 {
 	Snapshot	snapshot;
-	uint32		flags = SO_TYPE_SEQSCAN |
-		SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE;
 
 	Assert(RelFileLocatorEquals(relation->rd_locator, pscan->phs_locator));
 
@@ -188,6 +194,31 @@ table_beginscan_parallel(Relation relation, ParallelTableScanDesc pscan)
 											pscan, flags);
 }
 
+TableScanDesc
+table_beginscan_parallel(Relation relation, ParallelTableScanDesc pscan)
+{
+	uint32		flags = SO_TYPE_SEQSCAN |
+		SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE;
+
+	return table_beginscan_parallel_common(relation, pscan, flags);
+}
+
+/*
+ * Parallel version of table_beginscan_vmset()
+ */
+TableScanDesc
+table_beginscan_parallel_vmset(Relation relation, ParallelTableScanDesc pscan,
+							   bool modifies_rel)
+{
+	uint32		flags = SO_TYPE_SEQSCAN |
+		SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE;
+
+	if (!modifies_rel)
+		flags |= SO_ALLOW_VM_SET;
+
+	return table_beginscan_parallel_common(relation, pscan, flags);
+}
+
 
 /* ----------------------------------------------------------------------------
  * Index scan related functions.

src/backend/executor/execMain.c

@@ -916,6 +916,10 @@ InitPlan(QueryDesc *queryDesc, int eflags)
 					break;
 			}
 
+			/* If it has a rowmark, the relation is modified */
+			estate->es_modified_relids = bms_add_member(estate->es_modified_relids,
+														rc->rti);
+
 			/* Check that relation is a legal target for marking */
 			if (relation)
 				CheckValidRowMarkRel(relation, rc->markType);