pg prewarm seq scan

.github/workflows/build.yml

@@ -15,6 +15,7 @@ jobs:
     steps:
       - name: Install Postgres Build Dependencies
         run: |
+          sudo apt-get update
           sudo apt-get install build-essential coreutils libreadline-dev zlib1g-dev flex bison libxml2-dev libxslt-dev libssl-dev libxml2-utils xsltproc git
 
       - name: Checkout postgres repository

contrib/pg_prewarm/pg_prewarm.c

@@ -15,6 +15,7 @@
 #include <sys/stat.h>
 #include <unistd.h>
 
+#include "access/heapam.h"
 #include "access/relation.h"
 #include "fmgr.h"
 #include "miscadmin.h"
@@ -185,29 +186,50 @@ pg_prewarm(PG_FUNCTION_ARGS)
 	}
 	else if (ptype == PREWARM_BUFFER)
 	{
-		BlockNumber prefetch_block = first_block;
-		Oid			nspOid;
-		int			io_concurrency;
+		if (get_relkind_objtype(rel->rd_rel->relkind) == OBJECT_TABLE && forkNumber == MAIN_FORKNUM)
+		{
+			uint32 		scan_flags = SO_TYPE_SEQSCAN | SO_TEMP_SNAPSHOT;
+			HeapTuple 	tuple;
+			Snapshot 	snapshot;
+			TableScanDesc scan;
+			elog(LOG, "pg_prewarm: SeqScan relation \"%s\" starting %ld for %ld blocks", RelationGetRelationName(rel), first_block, last_block - first_block + 1);
+			// Use heap scan to set hint bits on every tuple. SO_ALLOW_PAGEMODE is intentionally NOT SET.
+			// Otherwise, when a page is all visible, tuple hint bits won't be set.
+			snapshot = RegisterSnapshot(GetTransactionSnapshot());
+			scan = heap_beginscan(rel, snapshot, 0, NULL, NULL, scan_flags);
+			heap_setscanlimits(scan, first_block, last_block - first_block + 1);
+			while ((tuple = heap_getnext(scan, ForwardScanDirection)) != NULL)
+			{
+				CHECK_FOR_INTERRUPTS();
+			}
+			heap_endscan(scan);
+		}
+		else
+		{
+			BlockNumber prefetch_block = first_block;
+			Oid			nspOid;
+			int			io_concurrency;
 
-		nspOid = rel->rd_rel->reltablespace;
-		io_concurrency = get_tablespace_maintenance_io_concurrency(nspOid);
+			nspOid = rel->rd_rel->reltablespace;
+			io_concurrency = get_tablespace_maintenance_io_concurrency(nspOid);
 
-		/*
-		 * In buffer mode, we actually pull the data into shared_buffers.
-		 */
-		for (block = first_block; block <= last_block; ++block)
-		{
-			Buffer buf;
-			BlockNumber prefetch_stop = block + Min(last_block - block + 1,
-													io_concurrency);
-			CHECK_FOR_INTERRUPTS();
-			while (prefetch_block < prefetch_stop)
+			/*
+			* In buffer mode, we actually pull the data into shared_buffers.
+			*/
+			for (block = first_block; block <= last_block; ++block)
 			{
-				PrefetchBuffer(rel, forkNumber, prefetch_block++);
+				Buffer buf;
+				BlockNumber prefetch_stop = block + Min(last_block - block + 1,
+														io_concurrency);
+				CHECK_FOR_INTERRUPTS();
+				while (prefetch_block < prefetch_stop)
+				{
+					PrefetchBuffer(rel, forkNumber, prefetch_block++);
+				}
+				buf = ReadBufferExtended(rel, forkNumber, block, RBM_NORMAL, NULL);
+				ReleaseBuffer(buf);
+				++blocks_done;
 			}
-			buf = ReadBufferExtended(rel, forkNumber, block, RBM_NORMAL, NULL);
-			ReleaseBuffer(buf);
-			++blocks_done;
 		}
 	}
 

src/backend/access/nbtree/nbtinsert.c

@@ -1472,6 +1472,8 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 	Page		origpage;
 	Page		leftpage,
 				rightpage;
+	PGAlignedBlock leftpage_buf,
+				rightpage_buf;
 	BlockNumber origpagenumber,
 				rightpagenumber;
 	BTPageOpaque ropaque,
@@ -1545,8 +1547,8 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 	firstrightoff = _bt_findsplitloc(rel, origpage, newitemoff, newitemsz,
 									 newitem, &newitemonleft);
 
-	/* Allocate temp buffer for leftpage */
-	leftpage = PageGetTempPage(origpage);
+	/* Use temporary buffer for leftpage */
+	leftpage = leftpage_buf.data;
 	_bt_pageinit(leftpage, BufferGetPageSize(buf));
 	lopaque = BTPageGetOpaque(leftpage);
 
@@ -1709,19 +1711,23 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 
 	/*
 	 * Acquire a new right page to split into, now that left page has a new
-	 * high key.  From here on, it's not okay to throw an error without
-	 * zeroing rightpage first.  This coding rule ensures that we won't
-	 * confuse future VACUUM operations, which might otherwise try to re-find
-	 * a downlink to a leftover junk page as the page undergoes deletion.
+	 * high key.
 	 *
-	 * It would be reasonable to start the critical section just after the new
-	 * rightpage buffer is acquired instead; that would allow us to avoid
-	 * leftover junk pages without bothering to zero rightpage.  We do it this
-	 * way because it avoids an unnecessary PANIC when either origpage or its
-	 * existing sibling page are corrupt.
+	 * To not confuse future VACUUM operations, we zero the right page and
+	 * work on an in-memory copy of it before writing WAL, then copy its
+	 * contents back to the actual page once we start the critical section
+	 * work.  This simplifies the split work, so as there is no need to zero
+	 * the right page before throwing an error.
 	 */
 	rbuf = _bt_allocbuf(rel, heaprel);
-	rightpage = BufferGetPage(rbuf);
+	rightpage = rightpage_buf.data;
+
+	/*
+	 * Copy the contents of the right page into its temporary location, and
+	 * zero the original space.
+	 */
+	memcpy(rightpage, BufferGetPage(rbuf), BLCKSZ);
+	memset(BufferGetPage(rbuf), 0, BLCKSZ);
 	rightpagenumber = BufferGetBlockNumber(rbuf);
 	/* rightpage was initialized by _bt_getbuf */
 	ropaque = BTPageGetOpaque(rightpage);
@@ -1770,7 +1776,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 		if (PageAddItem(rightpage, (Item) righthighkey, itemsz, afterrightoff,
 						false, false) == InvalidOffsetNumber)
 		{
-			memset(rightpage, 0, BufferGetPageSize(rbuf));
 			elog(ERROR, "failed to add high key to the right sibling"
 				 " while splitting block %u of index \"%s\"",
 				 origpagenumber, RelationGetRelationName(rel));
@@ -1818,7 +1823,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 				if (!_bt_pgaddtup(leftpage, newitemsz, newitem, afterleftoff,
 								  false))
 				{
-					memset(rightpage, 0, BufferGetPageSize(rbuf));
 					elog(ERROR, "failed to add new item to the left sibling"
 						 " while splitting block %u of index \"%s\"",
 						 origpagenumber, RelationGetRelationName(rel));
@@ -1831,7 +1835,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 				if (!_bt_pgaddtup(rightpage, newitemsz, newitem, afterrightoff,
 								  afterrightoff == minusinfoff))
 				{
-					memset(rightpage, 0, BufferGetPageSize(rbuf));
 					elog(ERROR, "failed to add new item to the right sibling"
 						 " while splitting block %u of index \"%s\"",
 						 origpagenumber, RelationGetRelationName(rel));
@@ -1845,7 +1848,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 		{
 			if (!_bt_pgaddtup(leftpage, itemsz, dataitem, afterleftoff, false))
 			{
-				memset(rightpage, 0, BufferGetPageSize(rbuf));
 				elog(ERROR, "failed to add old item to the left sibling"
 					 " while splitting block %u of index \"%s\"",
 					 origpagenumber, RelationGetRelationName(rel));
@@ -1857,7 +1859,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 			if (!_bt_pgaddtup(rightpage, itemsz, dataitem, afterrightoff,
 							  afterrightoff == minusinfoff))
 			{
-				memset(rightpage, 0, BufferGetPageSize(rbuf));
 				elog(ERROR, "failed to add old item to the right sibling"
 					 " while splitting block %u of index \"%s\"",
 					 origpagenumber, RelationGetRelationName(rel));
@@ -1878,7 +1879,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 		if (!_bt_pgaddtup(rightpage, newitemsz, newitem, afterrightoff,
 						  afterrightoff == minusinfoff))
 		{
-			memset(rightpage, 0, BufferGetPageSize(rbuf));
 			elog(ERROR, "failed to add new item to the right sibling"
 				 " while splitting block %u of index \"%s\"",
 				 origpagenumber, RelationGetRelationName(rel));
@@ -1898,7 +1898,6 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 		sopaque = BTPageGetOpaque(spage);
 		if (sopaque->btpo_prev != origpagenumber)
 		{
-			memset(rightpage, 0, BufferGetPageSize(rbuf));
 			ereport(ERROR,
 					(errcode(ERRCODE_INDEX_CORRUPTED),
 					 errmsg_internal("right sibling's left-link doesn't match: "
@@ -1941,9 +1940,19 @@ _bt_split(Relation rel, Relation heaprel, BTScanInsert itup_key, Buffer buf,
 	 * original.  We need to do this before writing the WAL record, so that
 	 * XLogInsert can WAL log an image of the page if necessary.
 	 */
-	PageRestoreTempPage(leftpage, origpage);
+	memcpy(origpage, leftpage, BLCKSZ);
 	/* leftpage, lopaque must not be used below here */
 
+	/*
+	 * Move the contents of the right page from its temporary location to the
+	 * destination buffer, before writing the WAL record.  Unlike the left
+	 * page, the right page and its opaque area are still needed to complete
+	 * the update of the page, so reinitialize them.
+	 */
+	rightpage = BufferGetPage(rbuf);
+	memcpy(rightpage, rightpage_buf.data, BLCKSZ);
+	ropaque = BTPageGetOpaque(rightpage);
+
 	MarkBufferDirty(buf);
 	MarkBufferDirty(rbuf);