Hello,

Following up on the torn-page problem Shihao found and Vignesh
reproduced, and on Amit's comments.  I reproduced it, tested the fixes,
and have reworked the series so the torn-page fix comes first.

The case of the torn-page:  The bug is on master today, independent of
the pin-only rework.  I ran Vignesh's reproducer (adapted so it finds
the primary bucket page from a record that registers it, rather than
from the removed bucket_blkno field) against three trees:

  - plain master: recovery fails replaying the overflow DELETE, "invalid
    page in block 7", before the later full-page image of block 7.
  - the pin-only series: same failure, reached through the appended
    block number instead of block 0.
  - master + the fix: recovery succeeds and the index returns the right
    rows.

The hazard was spotted during the original hash WAL work.  In 2017
Robert raised it on the write-ahead-logging thread [1]:

  "the page is registered and therefore an FPI will be taken if this is
   the first reference to this page since the checkpoint, but
   hash_xlog_delete won't restore the FPI.  I think that exposes us to a
   torn-page hazard.  _hash_freeovflpage/hash_xlog_squeeze_page seems to
   have the same disease, as does
   _hash_squeezebucket/hash_xlog_move_page_contents."

Redo was then switched to XLogReadBufferForRedoExtended() so it restores
an image when the record has one, but these records register the primary
bucket page pin-only with REGBUF_NO_IMAGE, so there is nothing to
restore and the RBM_NORMAL read still reaches the torn page.  So the
hazard stayed open; Shihao and Vignesh have now hit it.

On the fix: RBM_ZERO_ON_ERROR (Vignesh) does fix the reproducer, but as
Amit noted it is documented for non-critical forks where the caller
repairs, and this is a main-fork index page.  Guarding the read with
InHotStandby (Amit's suggestion) also fixes it and is the right one:

  - The read exists only to take the cleanup lock, which only matters to
    serialize against a scan on a hot standby.  When hot standby is not
    active no such scan can exist, so neither the read nor the lock is
    needed.
  - That is exactly the window where the torn-page hazard lives.  The
    repairing image always precedes the point where hot standby becomes
    active, so once InHotStandby is true any earlier image has already
    been replayed and the page is no longer torn.
  - It matches existing practice: nbtree and gist already gate their
    redo cleanup-lock work on InHotStandby.

Since the bug is pre-existing, the fix is patch 0001: it stands on its
own against master, applies with or without the rework, and (the
back-branch question aside) is separate from the refactor.  The rest of
the series then builds on it; in 0002 the same guard lives in the one
helper the refactor introduces, hash_xlog_lock_bucket().

  0001  Fix hash index redo reading a torn primary bucket page
  0002  Don't register pin-only primary bucket pages in hash WAL records
        (merges the old 0001 and the four-byte optimization, and
        renumbers the registered blocks per Amit's comment)
  0003  Add a recovery test for the hash WAL pin-only bucket page fix
  0004  Remove REGBUF_NO_CHANGE

For 0002 I folded the append-only-when-needed change into the main patch
so there is no common-case size increase at any point, and renumbered
the registered blocks so each record starts at 0 again now that the
primary bucket page is no longer block 0.

Testing (on current master): the new 068a recovery test fails on master
and on the pin-only series, and passes with 0001; the full recovery,
isolation and regress suites pass with PG_TEST_EXTRA=
wal_consistency_checking across the series.

best.

-greg

[1] 
https://postgr.es/m/caa4ek1lwzk0mak0lnczz6hsqchnamfjb47na8xuwrp4cylf...@mail.gmail.com

From 709c222f2c13e62646ca53947dff6d9a6c3d1693 Mon Sep 17 00:00:00 2001
From: Greg Burd <[email protected]>
Date: Fri, 30 Oct 2026 10:17:46 -0400
Subject: [PATCH v6 4/4] Remove REGBUF_NO_CHANGE

With the previous commits, the hash index AM no longer registers any
pin-only or otherwise-unchanged buffer, so REGBUF_NO_CHANGE has no users
left.  Remove the flag and restore the invariant that every registered
buffer is content-locked and dirty: XLogRegisterBuffer() asserts it
unconditionally again.

This also reverts the defensive measures added in 27caf4ca5ae5, which
let a REGBUF_NO_CHANGE buffer skip the full-page image forced by
wal_consistency_checking: XLogRecordAssemble() no longer special-cases
the flag, and verifyBackupPageConsistency() again asserts that a
registered block carries an image.  27caf4ca5ae5 remains in the back
branches, where this hash rework is not applied.

Discussion: https://postgr.es/m/[email protected]
---
 src/backend/access/transam/README         |  5 ----
 src/backend/access/transam/xloginsert.c   | 28 ++++++-----------------
 src/backend/access/transam/xlogrecovery.c |  7 +-----
 src/include/access/xloginsert.h           |  3 ---
 4 files changed, 8 insertions(+), 35 deletions(-)

diff --git a/src/backend/access/transam/README b/src/backend/access/transam/README
index 595afa01fa7..231106270fd 100644
--- a/src/backend/access/transam/README
+++ b/src/backend/access/transam/README
@@ -586,11 +586,6 @@ void XLogRegisterBuffer(uint8 block_id, Buffer buf, uint8 flags);
     XLogRegisterBufData() is included in the WAL record even if a full-page
     image is taken.
 
-    The REGBUF_NO_CHANGE flag indicates that the buffer is registered without
-    being modified by this record, and possibly without an exclusive content
-    lock.  wal_consistency_checking does not force a full-page image for such
-    a buffer.
-
 void XLogRegisterData(const void *data, int len);
 
     XLogRegisterData is used to include arbitrary data in the WAL record.  If
diff --git a/src/backend/access/transam/xloginsert.c b/src/backend/access/transam/xloginsert.c
index cd3359434e1..2d031e38594 100644
--- a/src/backend/access/transam/xloginsert.c
+++ b/src/backend/access/transam/xloginsert.c
@@ -257,22 +257,13 @@ XLogRegisterBuffer(uint8 block_id, Buffer buffer, uint8 flags)
 	Assert(begininsert_called);
 
 	/*
-	 * Ordinarily, the buffer should be exclusive-locked (or share-exclusive
-	 * in case of hint bits) and marked dirty before we get here, otherwise we
-	 * could end up violating one of the rules in access/transam/README.
-	 *
-	 * Some callers intentionally register a clean page and never update that
-	 * page's LSN; in that case they can pass the flag REGBUF_NO_CHANGE to
-	 * bypass these checks.
+	 * The buffer should be exclusive-locked (or share-exclusive in case of
+	 * hint bits) and marked dirty before we get here, otherwise we could end
+	 * up violating one of the rules in access/transam/README.
 	 */
-#ifdef USE_ASSERT_CHECKING
-	if (!(flags & REGBUF_NO_CHANGE))
-	{
-		Assert(BufferIsDirty(buffer));
-		Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE) ||
-			   BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE));
-	}
-#endif
+	Assert(BufferIsDirty(buffer));
+	Assert(BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_EXCLUSIVE) ||
+		   BufferIsLockedByMeInMode(buffer, BUFFER_LOCK_SHARE_EXCLUSIVE));
 
 	if (block_id >= max_registered_block_id)
 	{
@@ -722,13 +713,8 @@ XLogRecordAssemble(RmgrId rmid, uint8 info,
 		/*
 		 * If needs_backup is true or WAL checking is enabled for current
 		 * resource manager, log a full-page write for the current block.
-		 *
-		 * Do not force an image for a REGBUF_NO_CHANGE buffer.
 		 */
-		if (regbuf->flags & REGBUF_NO_CHANGE)
-			include_image = needs_backup;
-		else
-			include_image = needs_backup || (info & XLR_CHECK_CONSISTENCY) != 0;
+		include_image = needs_backup || (info & XLR_CHECK_CONSISTENCY) != 0;
 
 		if (include_image)
 		{
diff --git a/src/backend/access/transam/xlogrecovery.c b/src/backend/access/transam/xlogrecovery.c
index 52ac6d58423..7a1de0567a8 100644
--- a/src/backend/access/transam/xlogrecovery.c
+++ b/src/backend/access/transam/xlogrecovery.c
@@ -2488,12 +2488,7 @@ verifyBackupPageConsistency(XLogReaderState *record)
 			continue;
 		}
 
-		/*
-		 * Blocks registered with REGBUF_NO_CHANGE may have no image, see
-		 * XLogRecordAssemble().  Nothing to compare for these.
-		 */
-		if (!XLogRecHasBlockImage(record, block_id))
-			continue;
+		Assert(XLogRecHasBlockImage(record, block_id));
 
 		if (XLogRecBlockImageApply(record, block_id))
 		{
diff --git a/src/include/access/xloginsert.h b/src/include/access/xloginsert.h
index 86264564ebb..ddb1a85db59 100644
--- a/src/include/access/xloginsert.h
+++ b/src/include/access/xloginsert.h
@@ -38,9 +38,6 @@
 									 * will be skipped) */
 #define REGBUF_KEEP_DATA	0x10	/* include data even if a full-page image
 									 * is taken */
-#define REGBUF_NO_CHANGE	0x20	/* intentionally register clean buffer; no
-									 * image is forced for WAL consistency
-									 * checking */
 
 /* prototypes for public functions in xloginsert.c: */
 extern void XLogBeginInsert(void);
-- 
2.47.3

From 76d2bb84fffaa04564ceeb3ac92c38d2228010ac Mon Sep 17 00:00:00 2001
From: Greg Burd <[email protected]>
Date: Thu, 29 Oct 2026 10:30:14 -0400
Subject: [PATCH v6 2/4] Don't register pin-only primary bucket pages in hash
 WAL records

Replay of XLOG_HASH_DELETE, XLOG_HASH_MOVE_PAGE_CONTENTS and
XLOG_HASH_SQUEEZE_PAGE must take a cleanup lock on the primary bucket
page so that no scan on a hot standby is in progress in the bucket, and
none can start, while tuples are removed or moved.  That lock was
obtained by registering the page (pin-only) with REGBUF_NO_CHANGE; redo
never reads data from the block, the registration existed only for the
lock.  That broke the rule that a registered buffer is one the caller
holds exclusively locked and dirty, and under wal_consistency_checking
it forced a full-page image of a page a concurrent inserter could still
modify, so the record CRC was computed over shifting data.

Stop registering the primary bucket page when the record does not change
it.  When the page is not one of the pages the record changes, the
operation holds only a pin on it, so the record carries the page's block
number in its main data and replay reads that block and takes a cleanup
lock on it directly, at the same point and in the same order as before.
The block number is appended only in that case, so the records are no
larger than before in the common case where the primary bucket page is
registered.

After this change hash registers only buffers it holds exclusively
locked and dirty, and there are no REGBUF_NO_CHANGE users left.

The three record formats change, so bump XLOG_PAGE_MAGIC; master only.

Co-authored-by: Vignesh C <[email protected]>
Suggested-by: Amit Kapila <[email protected]>
Discussion: https://postgr.es/m/auNs8b_vS2OPdTL8@floki
---
 src/backend/access/hash/README      |   9 ++
 src/backend/access/hash/hash.c      |  17 ++--
 src/backend/access/hash/hash_xlog.c | 139 ++++++++++++++++------------
 src/backend/access/hash/hashovfl.c  | 100 +++++++++-----------
 src/include/access/hash_xlog.h      |  37 +++++---
 src/include/access/xlog_internal.h  |   2 +-
 6 files changed, 165 insertions(+), 139 deletions(-)

diff --git a/src/backend/access/hash/README b/src/backend/access/hash/README
index fc9031117c9..7774e1d2fe0 100644
--- a/src/backend/access/hash/README
+++ b/src/backend/access/hash/README
@@ -641,6 +641,15 @@ entire bucket.  After recovery, the operations will work correctly, but
 the index will remain bloated and this can impact performance of read and
 insert operations until the next vacuum squeeze the bucket completely.
 
+Replay of the delete, move-page-contents and squeeze-page records takes a
+cleanup lock on the primary bucket page before touching any other page, so
+that no scan on a hot standby is in progress in the bucket or can start in
+it while tuples are removed or moved; such a scan could otherwise miss
+tuples or return them twice.  When the primary bucket page is not one of
+the pages the record changes, the operation holds only a pin on it, so it
+is not registered in the record.  The record then carries its block number
+instead, and replay locks that block directly.
+
 
 Other Notes
 -----------
diff --git a/src/backend/access/hash/hash.c b/src/backend/access/hash/hash.c
index 63846c10840..7a79030303e 100644
--- a/src/backend/access/hash/hash.c
+++ b/src/backend/access/hash/hash.c
@@ -909,19 +909,16 @@ hashbucketcleanup(Relation rel, Bucket cur_bucket, Buffer bucket_buf,
 				XLogRegisterData(&xlrec, SizeOfHashDelete);
 
 				/*
-				 * bucket buffer was not changed, but still needs to be
-				 * registered to ensure that we can acquire a cleanup lock on
-				 * it during replay.
+				 * If buf is an overflow page, the primary bucket page is not
+				 * changed and we hold only a pin on it, so it is not
+				 * registered.  Append its block number instead, and replay
+				 * takes a cleanup lock on it.  It is not needed otherwise.
 				 */
 				if (!xlrec.is_primary_bucket_page)
-				{
-					uint8		flags = REGBUF_STANDARD | REGBUF_NO_IMAGE | REGBUF_NO_CHANGE;
-
-					XLogRegisterBuffer(0, bucket_buf, flags);
-				}
+					XLogRegisterData(&bucket_blkno, sizeof(BlockNumber));
 
-				XLogRegisterBuffer(1, buf, REGBUF_STANDARD);
-				XLogRegisterBufData(1, deletable,
+				XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
+				XLogRegisterBufData(0, deletable,
 									ndeletable * sizeof(OffsetNumber));
 
 				recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_DELETE);
diff --git a/src/backend/access/hash/hash_xlog.c b/src/backend/access/hash/hash_xlog.c
index 35791b087be..187de1f54ea 100644
--- a/src/backend/access/hash/hash_xlog.c
+++ b/src/backend/access/hash/hash_xlog.c
@@ -469,6 +469,54 @@ hash_xlog_split_complete(XLogReaderState *record)
 		UnlockReleaseBuffer(newbuf);
 }
 
+/*
+ * Take a cleanup lock on the primary bucket page of the bucket that a
+ * squeeze or delete record operates on, when that page is not one of the
+ * pages the record changes and so is not registered in it.  Its block number
+ * is stored in the main data of the record, right after the fixed-size struct
+ * of rec_size bytes.
+ *
+ * This serializes replay against scans on a hot standby, which pin the
+ * primary bucket page for their whole duration; see the comments in the
+ * callers.  The page is not modified, so nothing is done with it besides
+ * locking it.  block_id must be a block of the same index registered in
+ * the record.
+ */
+static Buffer
+hash_xlog_lock_bucket(XLogReaderState *record, uint8 block_id, Size rec_size)
+{
+	RelFileLocator rlocator;
+	ForkNumber	forknum;
+	BlockNumber bucket_blkno;
+	Buffer		buf;
+
+	Assert(XLogRecGetDataLen(record) == rec_size + sizeof(BlockNumber));
+	memcpy(&bucket_blkno, XLogRecGetData(record) + rec_size, sizeof(BlockNumber));
+
+	/*
+	 * The cleanup lock on the primary bucket page serializes replay against
+	 * scans on a hot standby, which pin the page for their whole duration.
+	 * When hot standby is not active no such scan can exist, so there is
+	 * nothing to serialize against and we need not read or lock the page.
+	 *
+	 * Skipping the read also sidesteps a torn copy of the page whose
+	 * repairing full-page image comes later in the WAL: that hazard exists
+	 * only before consistency is reached, which is before hot standby becomes
+	 * active, so once InHotStandby is true any earlier image has already been
+	 * replayed and the page is no longer torn.  Redo never uses the page's
+	 * contents here in any case.
+	 */
+	if (!InHotStandby)
+		return InvalidBuffer;
+
+	XLogRecGetBlockTag(record, block_id, &rlocator, &forknum, NULL);
+	buf = XLogReadBufferExtended(rlocator, forknum, bucket_blkno,
+								 RBM_NORMAL, InvalidBuffer);
+	if (BufferIsValid(buf))
+		LockBufferForCleanup(buf);
+	return buf;
+}
+
 /*
  * replay move of page contents for squeeze operation of hash index
  */
@@ -490,24 +538,12 @@ hash_xlog_move_page_contents(XLogReaderState *record)
 	 * can miss some records or show the same record multiple times.
 	 */
 	if (xldata->is_prim_bucket_same_wrt)
-		action = XLogReadBufferForRedoExtended(record, 1, RBM_NORMAL, true, &writebuf);
+		action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &writebuf);
 	else
 	{
-		/*
-		 * The cleanup lock on the primary bucket page serializes replay
-		 * against scans on a hot standby, which pin the page for their whole
-		 * duration.  When hot standby is not active no such scan can exist,
-		 * so we need not read or lock the page.  Skipping the read also
-		 * sidesteps a torn copy of the page whose repairing full-page image
-		 * comes later in the WAL: that hazard exists only before consistency
-		 * is reached, which is before hot standby becomes active, so once
-		 * InHotStandby is true any earlier image has already been replayed.
-		 * We do not use the page contents here in any case.
-		 */
-		if (InHotStandby)
-			(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
+		bucketbuf = hash_xlog_lock_bucket(record, 0, SizeOfHashMovePageContents);
 
-		action = XLogReadBufferForRedo(record, 1, &writebuf);
+		action = XLogReadBufferForRedo(record, 0, &writebuf);
 	}
 
 	/* replay the record for adding entries in overflow buffer */
@@ -519,7 +555,7 @@ hash_xlog_move_page_contents(XLogReaderState *record)
 		Size		datalen;
 		uint16		ninserted = 0;
 
-		data = begin = XLogRecGetBlockData(record, 1, &datalen);
+		data = begin = XLogRecGetBlockData(record, 0, &datalen);
 
 		writepage = BufferGetPage(writebuf);
 
@@ -558,13 +594,13 @@ hash_xlog_move_page_contents(XLogReaderState *record)
 	}
 
 	/* replay the record for deleting entries from overflow buffer */
-	if (XLogReadBufferForRedo(record, 2, &deletebuf) == BLK_NEEDS_REDO)
+	if (XLogReadBufferForRedo(record, 1, &deletebuf) == BLK_NEEDS_REDO)
 	{
 		Page		page;
 		char	   *ptr;
 		Size		len;
 
-		ptr = XLogRecGetBlockData(record, 2, &len);
+		ptr = XLogRecGetBlockData(record, 1, &len);
 
 		page = BufferGetPage(deletebuf);
 
@@ -625,25 +661,22 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 	 * can miss some records or show the same record multiple times.
 	 */
 	if (xldata->is_prim_bucket_same_wrt)
-		action = XLogReadBufferForRedoExtended(record, 1, RBM_NORMAL, true, &writebuf);
+	{
+		/* block 0 is absent if the primary bucket page is not modified */
+		if (XLogRecHasBlockRef(record, 0))
+			action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &writebuf);
+		else
+		{
+			bucketbuf = hash_xlog_lock_bucket(record, 1, SizeOfHashSqueezePage);
+			action = BLK_NOTFOUND;
+		}
+	}
 	else
 	{
-		/*
-		 * The cleanup lock on the primary bucket page serializes replay
-		 * against scans on a hot standby, which pin the page for their whole
-		 * duration.  When hot standby is not active no such scan can exist,
-		 * so we need not read or lock the page.  Skipping the read also
-		 * sidesteps a torn copy of the page whose repairing full-page image
-		 * comes later in the WAL: that hazard exists only before consistency
-		 * is reached, which is before hot standby becomes active, so once
-		 * InHotStandby is true any earlier image has already been replayed.
-		 * We do not use the page contents here in any case.
-		 */
-		if (InHotStandby)
-			(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
+		bucketbuf = hash_xlog_lock_bucket(record, 1, SizeOfHashSqueezePage);
 
 		if (xldata->ntups > 0 || xldata->is_prev_bucket_same_wrt)
-			action = XLogReadBufferForRedo(record, 1, &writebuf);
+			action = XLogReadBufferForRedo(record, 0, &writebuf);
 		else
 			action = BLK_NOTFOUND;
 	}
@@ -658,7 +691,7 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 		uint16		ninserted = 0;
 		bool		mod_wbuf = false;
 
-		data = begin = XLogRecGetBlockData(record, 1, &datalen);
+		data = begin = XLogRecGetBlockData(record, 0, &datalen);
 
 		writepage = BufferGetPage(writebuf);
 
@@ -724,7 +757,7 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 	}
 
 	/* replay the record for initializing overflow buffer */
-	if (XLogReadBufferForRedo(record, 2, &ovflbuf) == BLK_NEEDS_REDO)
+	if (XLogReadBufferForRedo(record, 1, &ovflbuf) == BLK_NEEDS_REDO)
 	{
 		Page		ovflpage;
 		HashPageOpaque ovflopaque;
@@ -749,7 +782,7 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 
 	/* replay the record for page previous to the freed overflow page */
 	if (!xldata->is_prev_bucket_same_wrt &&
-		XLogReadBufferForRedo(record, 3, &prevbuf) == BLK_NEEDS_REDO)
+		XLogReadBufferForRedo(record, 2, &prevbuf) == BLK_NEEDS_REDO)
 	{
 		Page		prevpage = BufferGetPage(prevbuf);
 		HashPageOpaque prevopaque = HashPageGetOpaque(prevpage);
@@ -763,11 +796,11 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 		UnlockReleaseBuffer(prevbuf);
 
 	/* replay the record for page next to the freed overflow page */
-	if (XLogRecHasBlockRef(record, 4))
+	if (XLogRecHasBlockRef(record, 3))
 	{
 		Buffer		nextbuf;
 
-		if (XLogReadBufferForRedo(record, 4, &nextbuf) == BLK_NEEDS_REDO)
+		if (XLogReadBufferForRedo(record, 3, &nextbuf) == BLK_NEEDS_REDO)
 		{
 			Page		nextpage = BufferGetPage(nextbuf);
 			HashPageOpaque nextopaque = HashPageGetOpaque(nextpage);
@@ -794,7 +827,7 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 	 * index updates can be happening concurrently.
 	 */
 	/* replay the record for bitmap page */
-	if (XLogReadBufferForRedo(record, 5, &mapbuf) == BLK_NEEDS_REDO)
+	if (XLogReadBufferForRedo(record, 4, &mapbuf) == BLK_NEEDS_REDO)
 	{
 		Page		mappage = BufferGetPage(mapbuf);
 		uint32	   *freep = NULL;
@@ -804,7 +837,7 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 
 		freep = HashPageGetBitmap(mappage);
 
-		data = XLogRecGetBlockData(record, 5, &datalen);
+		data = XLogRecGetBlockData(record, 4, &datalen);
 		bitmap_page_bit = (uint32 *) data;
 
 		CLRBIT(freep, *bitmap_page_bit);
@@ -816,11 +849,11 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 		UnlockReleaseBuffer(mapbuf);
 
 	/* replay the record for meta page */
-	if (XLogRecHasBlockRef(record, 6))
+	if (XLogRecHasBlockRef(record, 5))
 	{
 		Buffer		metabuf;
 
-		if (XLogReadBufferForRedo(record, 6, &metabuf) == BLK_NEEDS_REDO)
+		if (XLogReadBufferForRedo(record, 5, &metabuf) == BLK_NEEDS_REDO)
 		{
 			HashMetaPage metap;
 			Page		page;
@@ -828,7 +861,7 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 			uint32	   *firstfree_ovflpage;
 			Size		datalen;
 
-			data = XLogRecGetBlockData(record, 6, &datalen);
+			data = XLogRecGetBlockData(record, 5, &datalen);
 			firstfree_ovflpage = (uint32 *) data;
 
 			page = BufferGetPage(metabuf);
@@ -864,24 +897,12 @@ hash_xlog_delete(XLogReaderState *record)
 	 * can miss some records or show the same record multiple times.
 	 */
 	if (xldata->is_primary_bucket_page)
-		action = XLogReadBufferForRedoExtended(record, 1, RBM_NORMAL, true, &deletebuf);
+		action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &deletebuf);
 	else
 	{
-		/*
-		 * The cleanup lock on the primary bucket page serializes replay
-		 * against scans on a hot standby, which pin the page for their whole
-		 * duration.  When hot standby is not active no such scan can exist,
-		 * so we need not read or lock the page.  Skipping the read also
-		 * sidesteps a torn copy of the page whose repairing full-page image
-		 * comes later in the WAL: that hazard exists only before consistency
-		 * is reached, which is before hot standby becomes active, so once
-		 * InHotStandby is true any earlier image has already been replayed.
-		 * We do not use the page contents here in any case.
-		 */
-		if (InHotStandby)
-			(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
+		bucketbuf = hash_xlog_lock_bucket(record, 0, SizeOfHashDelete);
 
-		action = XLogReadBufferForRedo(record, 1, &deletebuf);
+		action = XLogReadBufferForRedo(record, 0, &deletebuf);
 	}
 
 	/* replay the record for deleting entries in bucket page */
@@ -890,7 +911,7 @@ hash_xlog_delete(XLogReaderState *record)
 		char	   *ptr;
 		Size		len;
 
-		ptr = XLogRecGetBlockData(record, 1, &len);
+		ptr = XLogRecGetBlockData(record, 0, &len);
 
 		page = BufferGetPage(deletebuf);
 
diff --git a/src/backend/access/hash/hashovfl.c b/src/backend/access/hash/hashovfl.c
index dbc57ef958c..7ad934ca38f 100644
--- a/src/backend/access/hash/hashovfl.c
+++ b/src/backend/access/hash/hashovfl.c
@@ -580,9 +580,13 @@ _hash_freeovflpage(Relation rel, Buffer bucketbuf, Buffer ovflbuf,
 	/* Get write-lock on metapage to update firstfree */
 	LockBuffer(metabuf, BUFFER_LOCK_EXCLUSIVE);
 
-	/* This operation needs to log multiple tuples, prepare WAL for that */
+	/*
+	 * This operation needs to log multiple tuples, prepare WAL for that.  One
+	 * extra data segment accounts for the primary bucket page's block number,
+	 * appended to the record when that page is pin-only (not registered).
+	 */
 	if (RelationNeedsWAL(rel))
-		XLogEnsureRecordSpace(HASH_XLOG_FREE_OVFL_BUFS, 4 + nitups);
+		XLogEnsureRecordSpace(HASH_XLOG_FREE_OVFL_BUFS, 5 + nitups);
 
 	START_CRIT_SECTION();
 
@@ -656,6 +660,7 @@ _hash_freeovflpage(Relation rel, Buffer bucketbuf, Buffer ovflbuf,
 	if (RelationNeedsWAL(rel))
 	{
 		xl_hash_squeeze_page xlrec;
+		BlockNumber bucket_blkno = BufferGetBlockNumber(bucketbuf);
 
 		xlrec.prevblkno = prevblkno;
 		xlrec.nextblkno = nextblkno;
@@ -667,45 +672,27 @@ _hash_freeovflpage(Relation rel, Buffer bucketbuf, Buffer ovflbuf,
 		XLogRegisterData(&xlrec, SizeOfHashSqueezePage);
 
 		/*
-		 * bucket buffer was not changed, but still needs to be registered to
-		 * ensure that we can acquire a cleanup lock on it during replay.
+		 * Register only the pages this record changes.  The primary bucket
+		 * page is registered only if it is the write page and is modified;
+		 * otherwise we may hold only a pin on it, so append its block number
+		 * and replay takes a cleanup lock on it.
 		 */
-		if (!is_prim_bucket_same_wrt)
-		{
-			uint8		flags = REGBUF_STANDARD | REGBUF_NO_IMAGE | REGBUF_NO_CHANGE;
-
-			XLogRegisterBuffer(0, bucketbuf, flags);
-		}
+		if (!(is_prim_bucket_same_wrt && mod_wbuf))
+			XLogRegisterData(&bucket_blkno, sizeof(BlockNumber));
 
-		if (nitups > 0)
-		{
-			XLogRegisterBuffer(1, wbuf, REGBUF_STANDARD);
-			XLogRegisterBufData(1, itup_offsets,
-								nitups * sizeof(OffsetNumber));
-			for (int i = 0; i < nitups; i++)
-				XLogRegisterBufData(1, itups[i], tups_size[i]);
-		}
-		else if (is_prim_bucket_same_wrt || is_prev_bucket_same_wrt)
+		if (mod_wbuf)
 		{
-			uint8		wbuf_flags;
-
-			/*
-			 * A write buffer needs to be registered even if no tuples are
-			 * added to it to ensure that we can acquire a cleanup lock on it
-			 * if it is the same as primary bucket buffer or update the
-			 * nextblkno if it is same as the previous bucket buffer.
-			 */
-			Assert(nitups == 0);
-
-			wbuf_flags = REGBUF_STANDARD;
-			if (!is_prev_bucket_same_wrt)
-				wbuf_flags |= REGBUF_NO_CHANGE;
-			else
-				Assert(mod_wbuf);
-			XLogRegisterBuffer(1, wbuf, wbuf_flags);
+			XLogRegisterBuffer(0, wbuf, REGBUF_STANDARD);
+			if (nitups > 0)
+			{
+				XLogRegisterBufData(0, itup_offsets,
+									nitups * sizeof(OffsetNumber));
+				for (int i = 0; i < nitups; i++)
+					XLogRegisterBufData(0, itups[i], tups_size[i]);
+			}
 		}
 
-		XLogRegisterBuffer(2, ovflbuf, REGBUF_STANDARD);
+		XLogRegisterBuffer(1, ovflbuf, REGBUF_STANDARD);
 
 		/*
 		 * If prevpage and the writepage (block in which we are moving tuples
@@ -714,18 +701,18 @@ _hash_freeovflpage(Relation rel, Buffer bucketbuf, Buffer ovflbuf,
 		 * writepage.
 		 */
 		if (BufferIsValid(prevbuf) && !is_prev_bucket_same_wrt)
-			XLogRegisterBuffer(3, prevbuf, REGBUF_STANDARD);
+			XLogRegisterBuffer(2, prevbuf, REGBUF_STANDARD);
 
 		if (BufferIsValid(nextbuf))
-			XLogRegisterBuffer(4, nextbuf, REGBUF_STANDARD);
+			XLogRegisterBuffer(3, nextbuf, REGBUF_STANDARD);
 
-		XLogRegisterBuffer(5, mapbuf, REGBUF_STANDARD);
-		XLogRegisterBufData(5, &bitmapbit, sizeof(uint32));
+		XLogRegisterBuffer(4, mapbuf, REGBUF_STANDARD);
+		XLogRegisterBufData(4, &bitmapbit, sizeof(uint32));
 
 		if (update_metap)
 		{
-			XLogRegisterBuffer(6, metabuf, REGBUF_STANDARD);
-			XLogRegisterBufData(6, &metap->hashm_firstfree, sizeof(uint32));
+			XLogRegisterBuffer(5, metabuf, REGBUF_STANDARD);
+			XLogRegisterBufData(5, &metap->hashm_firstfree, sizeof(uint32));
 		}
 
 		recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_SQUEEZE_PAGE);
@@ -966,10 +953,12 @@ readpage:
 
 					/*
 					 * This operation needs to log multiple tuples, prepare
-					 * WAL for that.
+					 * WAL for that.  One extra data segment accounts for the
+					 * primary bucket page's block number, appended to the
+					 * record when that page is pin-only (not registered).
 					 */
 					if (RelationNeedsWAL(rel))
-						XLogEnsureRecordSpace(0, 3 + nitups);
+						XLogEnsureRecordSpace(0, 4 + nitups);
 
 					START_CRIT_SECTION();
 
@@ -998,25 +987,22 @@ readpage:
 						XLogRegisterData(&xlrec, SizeOfHashMovePageContents);
 
 						/*
-						 * bucket buffer was not changed, but still needs to
-						 * be registered to ensure that we can acquire a
-						 * cleanup lock on it during replay.
+						 * If wbuf is an overflow page, the primary bucket
+						 * page is not changed and we hold only a pin on it,
+						 * so it is not registered.  Append its block number
+						 * instead, and replay takes a cleanup lock on it.
 						 */
 						if (!xlrec.is_prim_bucket_same_wrt)
-						{
-							int			flags = REGBUF_STANDARD | REGBUF_NO_IMAGE | REGBUF_NO_CHANGE;
-
-							XLogRegisterBuffer(0, bucket_buf, flags);
-						}
+							XLogRegisterData(&bucket_blkno, sizeof(BlockNumber));
 
-						XLogRegisterBuffer(1, wbuf, REGBUF_STANDARD);
-						XLogRegisterBufData(1, itup_offsets,
+						XLogRegisterBuffer(0, wbuf, REGBUF_STANDARD);
+						XLogRegisterBufData(0, itup_offsets,
 											nitups * sizeof(OffsetNumber));
 						for (i = 0; i < nitups; i++)
-							XLogRegisterBufData(1, itups[i], tups_size[i]);
+							XLogRegisterBufData(0, itups[i], tups_size[i]);
 
-						XLogRegisterBuffer(2, rbuf, REGBUF_STANDARD);
-						XLogRegisterBufData(2, deletable,
+						XLogRegisterBuffer(1, rbuf, REGBUF_STANDARD);
+						XLogRegisterBufData(1, deletable,
 											ndeletable * sizeof(OffsetNumber));
 
 						recptr = XLogInsert(RM_HASH_ID, XLOG_HASH_MOVE_PAGE_CONTENTS);
diff --git a/src/include/access/hash_xlog.h b/src/include/access/hash_xlog.h
index 149aed69dc5..0bb0d0590f5 100644
--- a/src/include/access/hash_xlog.h
+++ b/src/include/access/hash_xlog.h
@@ -129,9 +129,13 @@ typedef struct xl_hash_split_complete
  *
  * This data record is used for XLOG_HASH_MOVE_PAGE_CONTENTS
  *
- * Backup Blk 0: primary bucket page
- * Backup Blk 1: page containing moved tuples
- * Backup Blk 2: page from which tuples will be removed
+ * Backup Blk 0: page containing moved tuples
+ * Backup Blk 1: page from which tuples will be removed
+ *
+ * The primary bucket page is not registered.  If it is not the page moved to
+ * (is_prim_bucket_same_wrt is false), the block number of the primary bucket
+ * page follows the struct as a BlockNumber, and replay takes a cleanup lock
+ * on that block.
  */
 typedef struct xl_hash_move_page_contents
 {
@@ -149,13 +153,18 @@ typedef struct xl_hash_move_page_contents
  *
  * This data record is used for XLOG_HASH_SQUEEZE_PAGE
  *
- * Backup Blk 0: primary bucket page
- * Backup Blk 1: page containing tuples moved from freed overflow page
- * Backup Blk 2: freed overflow page
- * Backup Blk 3: page previous to the freed overflow page
- * Backup Blk 4: page next to the freed overflow page
- * Backup Blk 5: bitmap page containing info of freed overflow page
- * Backup Blk 6: meta page
+ * Backup Blk 0: page containing tuples moved from freed overflow page, if
+ *				 it is modified
+ * Backup Blk 1: freed overflow page
+ * Backup Blk 2: page previous to the freed overflow page
+ * Backup Blk 3: page next to the freed overflow page
+ * Backup Blk 4: bitmap page containing info of freed overflow page
+ * Backup Blk 5: meta page
+ *
+ * The primary bucket page is registered only as block 0, when it is the
+ * modified write page.  Otherwise the block number of the primary bucket
+ * page follows the struct as a BlockNumber, and replay takes a cleanup lock
+ * on that block.
  */
 typedef struct xl_hash_squeeze_page
 {
@@ -179,8 +188,12 @@ typedef struct xl_hash_squeeze_page
  *
  * This data record is used for XLOG_HASH_DELETE
  *
- * Backup Blk 0: primary bucket page
- * Backup Blk 1: page from which tuples are deleted
+ * Backup Blk 0: page from which tuples are deleted
+ *
+ * If block 0 is not the primary bucket page (is_primary_bucket_page is
+ * false), the block number of the primary bucket page follows the struct as a
+ * BlockNumber, and replay takes a cleanup lock on that block.  The primary
+ * bucket page is not registered.
  */
 typedef struct xl_hash_delete
 {
diff --git a/src/include/access/xlog_internal.h b/src/include/access/xlog_internal.h
index bf609c3c703..ed8f488dda5 100644
--- a/src/include/access/xlog_internal.h
+++ b/src/include/access/xlog_internal.h
@@ -32,7 +32,7 @@
 /*
  * Each page of XLOG file has a header like this:
  */
-#define XLOG_PAGE_MAGIC 0xD122	/* can be used as WAL version indicator */
+#define XLOG_PAGE_MAGIC 0xD123	/* can be used as WAL version indicator */
 
 typedef struct XLogPageHeaderData
 {
-- 
2.47.3

From 33501fe1747a460c2e1ec57edd3ebaeacc07dec7 Mon Sep 17 00:00:00 2001
From: Greg Burd <[email protected]>
Date: Wed, 11 Nov 2026 20:58:41 +0000
Subject: [PATCH v6 1/4] Fix hash index redo reading a torn primary bucket page

Replay of XLOG_HASH_DELETE, XLOG_HASH_MOVE_PAGE_CONTENTS and
XLOG_HASH_SQUEEZE_PAGE reads the primary bucket page only to take a
cleanup lock on it, so that no scan on a hot standby runs in the bucket
while tuples are removed or moved.  VACUUM holds only a pin on that page,
so the record does not modify it and carries no full-page image of it;
redo reads it from disk with RBM_NORMAL.

If the page is torn on disk and the record that repairs it with a
full-page image comes later in the WAL, recovery fails when it reaches
one of these records before the image, because the RBM_NORMAL read of
the torn page raises "invalid page".  This needs data checksums (or
full_page_writes catching the tear) to be detected, and the right
interleaving of a VACUUM that removes tuples from an overflow page while
the primary bucket page is unmodified, followed by a later modification
of the primary page.

Robert Haas noted this torn-page hazard during the original hash WAL
work in 2017; redo was switched to XLogReadBufferForRedoExtended() so it
restores an image when the record has one, but these records register
the primary bucket page pin-only with REGBUF_NO_IMAGE, so there is no
image to restore and the RBM_NORMAL read still reaches the torn page.

The cleanup lock matters only to serialize against hot-standby scans.
When hot standby is not active no such scan can exist, so the read and
lock are unnecessary; skip them.  That also avoids reading the page in
the only window where its torn on-disk copy has not yet been repaired,
since the repairing image always precedes the point where hot standby
becomes active.  nbtree and gist already gate redo cleanup-lock work on
InHotStandby for the same reason.

Reported-by: shihao zhong <[email protected]>
Author: Greg Burd <[email protected]>
Co-authored-by: Vignesh C <[email protected]>
Suggested-by: Amit Kapila <[email protected]>
Discussion: https://postgr.es/m/CAGRkXqT_AiwcsWZg5hnR1L2oWO8UpQpkL0iAjoZR+bey=go...@mail.gmail.com
Discussion: https://postgr.es/m/caa4ek1lwzk0mak0lnczz6hsqchnamfjb47na8xuwrp4cylf...@mail.gmail.com
---
 src/backend/access/hash/hash_xlog.c           |  42 ++++--
 src/test/recovery/meson.build                 |   1 +
 ...68a_hash_torn_primary_page_delete_first.pl | 142 ++++++++++++++++++
 3 files changed, 176 insertions(+), 9 deletions(-)
 create mode 100644 src/test/recovery/t/068a_hash_torn_primary_page_delete_first.pl

diff --git a/src/backend/access/hash/hash_xlog.c b/src/backend/access/hash/hash_xlog.c
index e9a2b9aa9a7..35791b087be 100644
--- a/src/backend/access/hash/hash_xlog.c
+++ b/src/backend/access/hash/hash_xlog.c
@@ -494,10 +494,18 @@ hash_xlog_move_page_contents(XLogReaderState *record)
 	else
 	{
 		/*
-		 * we don't care for return value as the purpose of reading bucketbuf
-		 * is to ensure a cleanup lock on primary bucket page.
+		 * The cleanup lock on the primary bucket page serializes replay
+		 * against scans on a hot standby, which pin the page for their whole
+		 * duration.  When hot standby is not active no such scan can exist,
+		 * so we need not read or lock the page.  Skipping the read also
+		 * sidesteps a torn copy of the page whose repairing full-page image
+		 * comes later in the WAL: that hazard exists only before consistency
+		 * is reached, which is before hot standby becomes active, so once
+		 * InHotStandby is true any earlier image has already been replayed.
+		 * We do not use the page contents here in any case.
 		 */
-		(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
+		if (InHotStandby)
+			(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
 
 		action = XLogReadBufferForRedo(record, 1, &writebuf);
 	}
@@ -621,10 +629,18 @@ hash_xlog_squeeze_page(XLogReaderState *record)
 	else
 	{
 		/*
-		 * we don't care for return value as the purpose of reading bucketbuf
-		 * is to ensure a cleanup lock on primary bucket page.
+		 * The cleanup lock on the primary bucket page serializes replay
+		 * against scans on a hot standby, which pin the page for their whole
+		 * duration.  When hot standby is not active no such scan can exist,
+		 * so we need not read or lock the page.  Skipping the read also
+		 * sidesteps a torn copy of the page whose repairing full-page image
+		 * comes later in the WAL: that hazard exists only before consistency
+		 * is reached, which is before hot standby becomes active, so once
+		 * InHotStandby is true any earlier image has already been replayed.
+		 * We do not use the page contents here in any case.
 		 */
-		(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
+		if (InHotStandby)
+			(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
 
 		if (xldata->ntups > 0 || xldata->is_prev_bucket_same_wrt)
 			action = XLogReadBufferForRedo(record, 1, &writebuf);
@@ -852,10 +868,18 @@ hash_xlog_delete(XLogReaderState *record)
 	else
 	{
 		/*
-		 * we don't care for return value as the purpose of reading bucketbuf
-		 * is to ensure a cleanup lock on primary bucket page.
+		 * The cleanup lock on the primary bucket page serializes replay
+		 * against scans on a hot standby, which pin the page for their whole
+		 * duration.  When hot standby is not active no such scan can exist,
+		 * so we need not read or lock the page.  Skipping the read also
+		 * sidesteps a torn copy of the page whose repairing full-page image
+		 * comes later in the WAL: that hazard exists only before consistency
+		 * is reached, which is before hot standby becomes active, so once
+		 * InHotStandby is true any earlier image has already been replayed.
+		 * We do not use the page contents here in any case.
 		 */
-		(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
+		if (InHotStandby)
+			(void) XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true, &bucketbuf);
 
 		action = XLogReadBufferForRedo(record, 1, &deletebuf);
 	}
diff --git a/src/test/recovery/meson.build b/src/test/recovery/meson.build
index b6ac9dd07a7..06b649be2aa 100644
--- a/src/test/recovery/meson.build
+++ b/src/test/recovery/meson.build
@@ -69,6 +69,7 @@ tests += {
       't/058_shutdown_crash_restart.pl',
       't/059_remote_apply_status_interval.pl',
       't/060_multixact_crashed_updater.pl',
+      't/068a_hash_torn_primary_page_delete_first.pl',
     ],
   },
 }
diff --git a/src/test/recovery/t/068a_hash_torn_primary_page_delete_first.pl b/src/test/recovery/t/068a_hash_torn_primary_page_delete_first.pl
new file mode 100644
index 00000000000..e61a0b98885
--- /dev/null
+++ b/src/test/recovery/t/068a_hash_torn_primary_page_delete_first.pl
@@ -0,0 +1,142 @@
+# Copyright (c) 2026, PostgreSQL Global Development Group
+
+# Test crash recovery when the primary bucket page of a hash bucket is torn,
+# and the full-page image that repairs it comes after the record that reads it.
+#
+
+use strict;
+use warnings FATAL => 'all';
+
+use IPC::Run;
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $node = PostgreSQL::Test::Cluster->new('primary');
+$node->init;
+$node->append_conf(
+	'postgresql.conf', qq{
+autovacuum = off
+fsync = off
+full_page_writes = on
+checkpoint_timeout = 1h
+});
+$node->start;
+
+is($node->safe_psql('postgres', 'SHOW data_checksums'),
+	'on', 'premise: data checksums are on');
+
+# One bucket receives every key: a primary page holding about 407 entries and
+# an overflow page with the rest.
+$node->safe_psql(
+	'postgres', q{
+CREATE TABLE t (id int, k int) WITH (autovacuum_enabled = false);
+CREATE INDEX t_i ON t USING hash (k);
+INSERT INTO t SELECT i, 1 FROM generate_series(1, 600) i;
+});
+$node->safe_psql('postgres', 'CHECKPOINT');
+my $redo = $node->safe_psql('postgres',
+	'SELECT redo_lsn FROM pg_control_checkpoint()');
+
+# The entries of ids 451-600 are on the overflow page, those of ids 1-50 on the
+# primary page.  The DELETE for the overflow page comes first; the VACUUM of
+# the primary page then changes it for the first time since the checkpoint, so
+# its record carries the full-page image of the page.
+$node->safe_psql('postgres', 'DELETE FROM t WHERE id > 450');
+$node->safe_psql('postgres', 'VACUUM t');
+$node->safe_psql('postgres', 'DELETE FROM t WHERE id <= 50');
+$node->safe_psql('postgres', 'VACUUM t');
+
+my $relpath = $node->safe_psql('postgres',
+	q{SELECT pg_relation_filepath('t_i')});
+my $bs = $node->safe_psql('postgres', 'SHOW block_size');
+$node->safe_psql('postgres', 'SELECT pg_switch_wal()');
+
+# Read the Hash records after the checkpoint.
+my ($stdout, $stderr);
+IPC::Run::run(
+	[
+		'pg_waldump',
+		'--path' => $node->data_dir . '/pg_wal',
+		'--start' => $redo,
+		'--rmgr' => 'Hash',
+	],
+	'>' => \$stdout,
+	'2>' => \$stderr);
+my @records = grep { /^rmgr: Hash/ } split(/^(?=rmgr: )/m, $stdout);
+
+# The primary bucket page's block number is not a registered block of the
+# overflow-page DELETE (the record carries it only as appended main data), so
+# find it from a record that does register it: a DELETE or SQUEEZE_PAGE marked
+# is_primary T references the primary bucket page as its first block.
+my $pblk;
+for my $r (@records)
+{
+	if ($r =~ /desc: (?:DELETE|SQUEEZE_PAGE) .*is_primary T/
+		&& $r =~ /blkref #0: rel \S+ blk (\d+)/)
+	{
+		$pblk = $1;
+		last;
+	}
+}
+ok(defined $pblk, 'premise: found the primary bucket page')
+  or diag $stdout;
+
+# The overflow-page DELETE (is_primary F) reads the primary bucket page
+# pin-only to take the cleanup lock; it must come before the primary page's
+# first full-page image.
+my ($idx_delete) =
+  grep { $records[$_] =~ /desc: DELETE .*is_primary F/ } 0 .. $#records;
+ok(defined $idx_delete,
+	'premise: VACUUM wrote a DELETE for an overflow page')
+  or diag $stdout;
+
+# Where the first full-page image of the primary page is.
+my ($idx_fpi) =
+  defined $pblk
+  ? grep { $records[$_] =~ /blkref #\d+: rel \S+ blk $pblk FPW/ }
+  0 .. $#records
+  : ();
+ok(defined $idx_fpi,
+	'premise: the primary page got a full-page image after the checkpoint')
+  or diag $stdout;
+if (defined $idx_delete && defined $idx_fpi)
+{
+	cmp_ok($idx_delete, '<', $idx_fpi,
+		'premise: the DELETE comes before the image of the primary page');
+}
+
+$node->stop('immediate');
+
+# Simulate a torn write of the primary page: keep the first half of its block
+# and overwrite the second half.
+if (defined $pblk)
+{
+	my $path = $node->data_dir . '/' . $relpath;
+	open(my $fh, '+<:raw', $path) or die "could not open $path: $!";
+	sysseek($fh, $pblk * $bs + $bs / 2, 0) or die "seek: $!";
+	my $junk = join('', map { chr(int(rand(256))) } 1 .. $bs / 2);
+	syswrite($fh, $junk) == $bs / 2 or die "write: $!";
+	close($fh);
+}
+
+my $started = $node->start(fail_ok => 1);
+ok($started, 'crash recovery succeeds');
+if ($started)
+{
+	is( $node->safe_psql(
+			'postgres',
+			q{SET enable_seqscan = off; SET enable_bitmapscan = off;
+			  SELECT count(*) FROM t WHERE k = 1}),
+		'400',
+		'the hash index gives the right result after recovery');
+	$node->stop;
+}
+else
+{
+	my $log = slurp_file($node->logfile);
+	diag("the server did not start; the log has:\n"
+		  . join('', grep { /invalid page|FATAL|WAL redo/ } split(/^/m, $log)));
+}
+
+done_testing();
-- 
2.47.3

From 0742d3d2d6022703fc154717f1c2270524aff7a1 Mon Sep 17 00:00:00 2001
From: Greg Burd <[email protected]>
Date: Fri, 30 Oct 2026 10:10:03 -0400
Subject: [PATCH v6 3/4] Add a recovery test for the hash WAL pin-only bucket
 page fix

Add recovery/t/061_hash_vacuum_insert_crc.pl, which exercises the
interleaving the previous commits make safe: a hash index VACUUM logging
an overflow-page delete while a concurrent INSERT changes the primary
bucket page it holds pin-only, with wal_consistency_checking='hash'.

The test adds a general, test-only injection point in
XLogRecordAssemble(), right after the record CRC is computed and before
the record is copied into the WAL buffers.  Its argument names the
record as "rmid/info/mainlen" so a test can wait on exactly one kind of
record; here the test waits on the overflow-page XLOG_HASH_DELETE (whose
main data carries the appended primary bucket block number) while VACUUM
holds the primary bucket page pin-only, then lets an INSERT change that
page, then crashes and checks that pg_waldump and crash recovery both
accept the WAL with no "incorrect resource manager data checksum".

Before 27caf4ca5ae5 and this rework, the record's CRC was computed over
the concurrently-modified primary bucket page and replay failed; the
injection point makes that window deterministic so the case is covered
from now on.

Co-authored-by: Vignesh C <[email protected]>
Discussion: https://postgr.es/m/CALDaNm14vRb2TZWyhjSHx=SyT7R=e7+ackrmxuojdwwee40...@mail.gmail.com
---
 src/backend/access/transam/xloginsert.c       |  22 +++
 src/test/recovery/meson.build                 |   1 +
 .../recovery/t/061_hash_vacuum_insert_crc.pl  | 181 ++++++++++++++++++
 3 files changed, 204 insertions(+)
 create mode 100644 src/test/recovery/t/061_hash_vacuum_insert_crc.pl

diff --git a/src/backend/access/transam/xloginsert.c b/src/backend/access/transam/xloginsert.c
index 341deb7b80d..cd3359434e1 100644
--- a/src/backend/access/transam/xloginsert.c
+++ b/src/backend/access/transam/xloginsert.c
@@ -39,6 +39,7 @@
 #include "replication/origin.h"
 #include "storage/bufmgr.h"
 #include "storage/proc.h"
+#include "utils/injection_point.h"
 #include "utils/memutils.h"
 #include "utils/pgstat_internal.h"
 #include "utils/rel.h"
@@ -991,6 +992,27 @@ XLogRecordAssemble(RmgrId rmid, uint8 info,
 	for (rdt = hdr_rdt.next; rdt != NULL; rdt = rdt->next)
 		COMP_CRC32C(rdata_crc, rdt->data, rdt->len);
 
+	/*
+	 * Test hook: let a test modify a buffer after the record's CRC has been
+	 * computed but before the record is copied into the WAL buffers, to
+	 * reproduce the window in which a concurrently modified, pin-only
+	 * registered buffer could make the on-disk CRC not match the record.  The
+	 * argument names the record as "rmid/info/mainlen" (resource manager, its
+	 * info bits, and the length of the main data) so a test can wait on
+	 * exactly one kind of record.  This runs inside a critical section, so it
+	 * must not allocate; it fires only in a backend that loaded the point
+	 * beforehand.
+	 */
+#ifdef USE_INJECTION_POINTS
+	{
+		char		pointarg[24];
+
+		snprintf(pointarg, sizeof(pointarg), "%d/%02X/" UINT64_FORMAT,
+				 rmid, info, mainrdata_len);
+		INJECTION_POINT_CACHED("xlog-assemble-after-crc", pointarg);
+	}
+#endif
+
 	/*
 	 * Ensure that the XLogRecord is not too large.
 	 *
diff --git a/src/test/recovery/meson.build b/src/test/recovery/meson.build
index 06b649be2aa..1859ff6f2c2 100644
--- a/src/test/recovery/meson.build
+++ b/src/test/recovery/meson.build
@@ -69,6 +69,7 @@ tests += {
       't/058_shutdown_crash_restart.pl',
       't/059_remote_apply_status_interval.pl',
       't/060_multixact_crashed_updater.pl',
+      't/061_hash_vacuum_insert_crc.pl',
       't/068a_hash_torn_primary_page_delete_first.pl',
     ],
   },
diff --git a/src/test/recovery/t/061_hash_vacuum_insert_crc.pl b/src/test/recovery/t/061_hash_vacuum_insert_crc.pl
new file mode 100644
index 00000000000..aa030abd587
--- /dev/null
+++ b/src/test/recovery/t/061_hash_vacuum_insert_crc.pl
@@ -0,0 +1,181 @@
+# Copyright (c) 2026, PostgreSQL Global Development Group
+
+# Test that the WAL record written by hash index VACUUM for an overflow page is
+# intact when a concurrent INSERT changes the primary bucket page.
+#
+# hashbucketcleanup() keeps only a pin on the primary bucket page while it
+# walks the overflow chain.  Before the hash WAL rework it registered that page
+# (REGBUF_STANDARD | REGBUF_NO_IMAGE | REGBUF_NO_CHANGE) with the
+# XLOG_HASH_DELETE record it writes for each overflow page.  With
+# wal_consistency_checking enabled for the hash resource manager,
+# XLogRecordAssemble() included a full image of every registered block, so it
+# read the unlocked primary bucket page.  If an INSERT added a tuple to that
+# page after the record's CRC was computed but before the page was copied into
+# the WAL buffers, the record was written with a CRC that did not match its
+# contents, and replay stopped at it with "incorrect resource manager data
+# checksum".
+#
+# The window is a few microseconds wide, so the test widens it with an
+# injection point in XLogRecordAssemble(), placed right after the CRC is
+# computed.  It is attached to the VACUUM session only and matches just the
+# overflow-page delete, identified by "rmgr id/xl_info/main-data length".  The
+# point still fires after the fix, so the test exercises the same interleaving
+# on a correct server.
+#
+# The assertions state the correct behavior (all WAL is valid and replays), so
+# the test fails while the bug exists.
+
+use strict;
+use warnings FATAL => 'all';
+
+use IPC::Run;
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+if ($ENV{enable_injection_points} ne 'yes')
+{
+	plan skip_all => 'Injection points not supported by this build';
+}
+
+my $node = PostgreSQL::Test::Cluster->new('primary');
+$node->init;
+$node->append_conf(
+	'postgresql.conf', qq{
+wal_consistency_checking = 'hash'
+autovacuum = off
+fsync = off
+checkpoint_timeout = 1h
+});
+$node->start;
+
+if (!$node->check_extension('injection_points'))
+{
+	plan skip_all => 'Extension injection_points not installed';
+}
+$node->safe_psql('postgres', 'CREATE EXTENSION injection_points');
+
+my $hash_rmid = $node->safe_psql('postgres',
+	q{SELECT rm_id FROM pg_get_wal_resource_managers() WHERE rm_name = 'Hash'}
+);
+
+# Build a single bucket whose primary page stays live and whose overflow page
+# is all dead, so VACUUM logs an XLOG_HASH_DELETE for the overflow page only:
+# the primary page has no dead index entries, so no delete record is written
+# for it, and the only consistency-checked hash DELETE in the window is the
+# overflow one, logged while the primary page is held pin-only.  All rows share
+# the key, so they land in one bucket; the first batch fills most of the
+# primary page and stays live, the second spills onto the overflow page and is
+# deleted.  Free slots are left on the primary page so the concurrent INSERT
+# below lands there.
+#
+# hashbuild() sizes the index from estimate_rel_size() on the heap, which for a
+# table that has never been vacuumed applies a floor of 10 pages.  Vacuuming
+# first sets reltuples to 0, so the estimate is believed and the index starts
+# with the minimum of 2 buckets.
+$node->safe_psql(
+	'postgres', q{
+CREATE TABLE hash_t (k int, tag int) WITH (autovacuum_enabled = false);
+VACUUM hash_t;
+CREATE INDEX hash_i ON hash_t USING hash (k);
+INSERT INTO hash_t SELECT 1, 0 FROM generate_series(1, 390);
+INSERT INTO hash_t SELECT 1, 1 FROM generate_series(1, 210);
+DELETE FROM hash_t WHERE tag = 1;
+});
+
+# Premise: metapage, two bucket pages, a bitmap page and one overflow page.
+is( $node->safe_psql(
+		'postgres',
+		q{SELECT pg_relation_size('hash_i') / current_setting('block_size')::int}
+	),
+	'5',
+	'premise: the index has a primary bucket page and one overflow page');
+
+# Start from a checkpoint, so crash recovery replays everything below and
+# pg_waldump has a record boundary to start from.
+$node->safe_psql('postgres', 'CHECKPOINT');
+my $start_lsn = $node->safe_psql('postgres', 'SELECT pg_current_wal_lsn()');
+
+# The VACUUM session.  The point is attached to this session only and must be
+# loaded before the critical section in which it runs.
+#
+# Wait on the XLOG_HASH_DELETE record for the overflow page, not the one for
+# the primary page.  Both are hash records (rm id $hash_rmid) with info 0x92
+# (XLOG_HASH_DELETE 0x90 | XLR_CHECK_CONSISTENCY 0x02).  They differ in main
+# data length: the primary-page delete carries just xl_hash_delete
+# (SizeOfHashDelete, two bools = 2 bytes), while the overflow-page delete also
+# appends the primary bucket page's block number (a 4-byte BlockNumber), for 6
+# bytes.  Waiting on the 6-byte one parks VACUUM while it holds the primary
+# bucket page pin-only, which is the page the INSERT below changes.
+my $overflow_delete = "$hash_rmid/92/6";
+my $vacuum = $node->background_psql('postgres');
+my $vacuum_pid = $vacuum->query_safe('SELECT pg_backend_pid()');
+$vacuum->query_safe(q{SELECT injection_points_set_local()});
+$vacuum->query_safe(
+	qq{SELECT injection_points_attach('xlog-assemble-after-crc', 'wait', '$overflow_delete')}
+);
+$vacuum->query_safe(
+	q{SELECT injection_points_load('xlog-assemble-after-crc')});
+$vacuum->query_until(
+	qr/starting_vacuum/,
+	q(\echo starting_vacuum
+VACUUM (INDEX_CLEANUP on) hash_t;
+));
+
+# VACUUM has removed the dead entries from the primary page, released its lock
+# (keeping the pin), and is now logging the removal from the overflow page: the
+# CRC of that record is computed and the unlocked primary page is not yet
+# copied.
+$node->poll_query_until('postgres',
+	"SELECT wait_event = 'xlog-assemble-after-crc' FROM pg_stat_activity WHERE pid = $vacuum_pid"
+) or die "VACUUM did not reach the injection point";
+note('VACUUM is between computing the CRC and copying the record');
+
+# Now there is room in the primary page, so this INSERT changes it while VACUUM
+# holds it pin-only during the overflow delete.
+$node->safe_psql('postgres', q{INSERT INTO hash_t VALUES (1, 0)});
+
+$node->safe_psql('postgres',
+	q{SELECT injection_points_wakeup('xlog-assemble-after-crc')});
+$vacuum->quit;
+
+my $end_lsn =
+  $node->safe_psql('postgres', 'SELECT pg_current_wal_insert_lsn()');
+$node->poll_query_until('postgres',
+	"SELECT pg_current_wal_flush_lsn() >= '$end_lsn'")
+  or die "timed out waiting for WAL flush";
+
+# Crash, so that start-up has to replay (and CRC-check) the workload.
+$node->stop('immediate');
+
+# 1. Every record in the range must have a valid CRC.  pg_waldump verifies it
+# for every record it reads, whatever the --rmgr filter.
+my $stderr;
+my $ok = IPC::Run::run(
+	[
+		'pg_waldump',
+		'--path' => $node->data_dir . '/pg_wal',
+		'--start' => $start_lsn,
+		'--end' => $end_lsn,
+		'--rmgr' => 'Hash',
+	],
+	'>' => '/dev/null',
+	'2>' => \$stderr);
+
+ok($ok, 'pg_waldump reads the whole workload without errors');
+unlike(
+	$stderr,
+	qr/incorrect resource manager data checksum/,
+	'no WAL record has a mismatching CRC');
+
+# 2. Crash recovery must replay it without complaint.
+my $log_offset = -s $node->logfile;
+$node->start;
+my $log = slurp_file($node->logfile, $log_offset);
+unlike(
+	$log,
+	qr/incorrect resource manager data checksum|inconsistent page found/,
+	'crash recovery replays the workload without CRC or consistency errors');
+
+$node->stop;
+done_testing();
-- 
2.47.3

Reply via email to