Fix pg_rewind bugs when rewinding a standby server.

If the target is a standby server, its WAL doesn't end at the last
checkpoint record, but at minRecoveryPoint. We must scan all the
WAL from the last common checkpoint all the way up to minRecoveryPoint
for modified pages, and also consider that portion when determining
whether the server needs rewinding.

Backpatch to all supported versions.

Author: Ian Barwick and me
Discussion: https://www.postgresql.org/message-id/CABvVfJU-LDWvoz4-Yow3Ay5LZYTuPD7eSjjE4kGyNZpXC6FrVQ%40mail.gmail.com

Author: hlinnaka

src/bin/pg_rewind/parsexlog.c

@@ -55,6 +55,9 @@ static int	SimpleXLogPageRead(XLogReaderState *xlogreader,
  * Read WAL from the datadir/pg_wal, starting from 'startpoint' on timeline
  * index 'tliIndex' in target timeline history, until 'endpoint'. Make note of
  * the data blocks touched by the WAL records, and return them in a page map.
+ *
+ * 'endpoint' is the end of the last record to read. The record starting at
+ * 'endpoint' is the first one that is not read.
  */
 void
 extractPageMap(const char *datadir, XLogRecPtr startpoint, int tliIndex,
@@ -93,7 +96,13 @@ extractPageMap(const char *datadir, XLogRecPtr startpoint, int tliIndex,
 
 		extractPageInfo(xlogreader);
 
-	} while (xlogreader->ReadRecPtr != endpoint);
+	} while (xlogreader->EndRecPtr < endpoint);
+
+	/*
+	 * If 'endpoint' didn't point exactly at a record boundary, the caller
+	 * messed up.
+	 */
+	Assert(xlogreader->EndRecPtr == endpoint);
 
 	XLogReaderFree(xlogreader);
 	if (xlogreadfd != -1)

src/bin/pg_rewind/pg_rewind.c

@@ -129,6 +129,7 @@ main(int argc, char **argv)
 	XLogRecPtr	chkptrec;
 	TimeLineID	chkpttli;
 	XLogRecPtr	chkptredo;
+	XLogRecPtr	target_wal_endrec;
 	size_t		size;
 	char	   *buffer;
 	bool		no_ensure_shutdown = false;
@@ -335,43 +336,56 @@ main(int argc, char **argv)
 	{
 		pg_log_info("source and target cluster are on the same timeline");
 		rewind_needed = false;
+		target_wal_endrec = 0;
 	}
 	else
 	{
+		XLogRecPtr	chkptendrec;
+
 		findCommonAncestorTimeline(&divergerec, &lastcommontliIndex);
 		pg_log_info("servers diverged at WAL location %X/%X on timeline %u",
 					(uint32) (divergerec >> 32), (uint32) divergerec,
 					targetHistory[lastcommontliIndex].tli);
 
+		/*
+		 * Determine the end-of-WAL on the target.
+		 *
+		 * The WAL ends at the last shutdown checkpoint, or at
+		 * minRecoveryPoint if it was a standby. (If we supported rewinding a
+		 * server that was not shut down cleanly, we would need to replay
+		 * until we reach the first invalid record, like crash recovery does.)
+		 */
+
+		/* read the checkpoint record on the target to see where it ends. */
+		chkptendrec = readOneRecord(datadir_target,
+									ControlFile_target.checkPoint,
+									targetNentries - 1,
+									restore_command);
+
+		if (ControlFile_target.minRecoveryPoint > chkptendrec)
+		{
+			target_wal_endrec = ControlFile_target.minRecoveryPoint;
+		}
+		else
+		{
+			target_wal_endrec = chkptendrec;
+		}
+
 		/*
 		 * Check for the possibility that the target is in fact a direct
 		 * ancestor of the source. In that case, there is no divergent history
 		 * in the target that needs rewinding.
 		 */
-		if (ControlFile_target.checkPoint >= divergerec)
+		if (target_wal_endrec > divergerec)
 		{
 			rewind_needed = true;
 		}
 		else
 		{
-			XLogRecPtr	chkptendrec;
+			/* the last common checkpoint record must be part of target WAL */
+			Assert(target_wal_endrec == divergerec);
 
-			/* Read the checkpoint record on the target to see where it ends. */
-			chkptendrec = readOneRecord(datadir_target,
-										ControlFile_target.checkPoint,
-										targetNentries - 1,
-										restore_command);
-
-			/*
-			 * If the histories diverged exactly at the end of the shutdown
-			 * checkpoint record on the target, there are no WAL records in
-			 * the target that don't belong in the source's history, and no
-			 * rewind is needed.
-			 */
-			if (chkptendrec == divergerec)
-				rewind_needed = false;
-			else
-				rewind_needed = true;
+			rewind_needed = false;
 		}
 	}
 
@@ -407,14 +421,12 @@ main(int argc, char **argv)
 	/*
 	 * Read the target WAL from last checkpoint before the point of fork, to
 	 * extract all the pages that were modified on the target cluster after
-	 * the fork. We can stop reading after reaching the final shutdown record.
-	 * XXX: If we supported rewinding a server that was not shut down cleanly,
-	 * we would need to replay until the end of WAL here.
+	 * the fork.
 	 */
 	if (showprogress)
 		pg_log_info("reading WAL in target");
 	extractPageMap(datadir_target, chkptrec, lastcommontliIndex,
-				   ControlFile_target.checkPoint, restore_command);
+				   target_wal_endrec, restore_command);
 
 	/*
 	 * We have collected all information we need from both systems. Decide

src/bin/pg_rewind/t/008_min_recovery_point.pl

@@ -0,0 +1,156 @@
+#
+# Test situation where a target data directory contains
+# WAL records beyond both the last checkpoint and the divergence
+# point:
+#
+# Target WAL (TLI 2):
+#
+# backup ... Checkpoint A ... INSERT 'rewind this'
+#            (TLI 1 -> 2)
+#
+#            ^ last common                        ^ minRecoveryPoint
+#              checkpoint
+#
+# Source WAL (TLI 3):
+#
+# backup ... Checkpoint A ... Checkpoint B ... INSERT 'keep this'
+#            (TLI 1 -> 2)     (TLI 2 -> 3)
+#
+#
+# The last common checkpoint is Checkpoint A. But there is WAL on TLI 2
+# after the last common checkpoint that needs to be rewound. We used to
+# have a bug where minRecoveryPoint was ignored, and pg_rewind concluded
+# that the target doesn't need rewinding in this scenario, because the
+# last checkpoint on the target TLI was an ancestor of the source TLI.
+#
+#
+# This test does not make use of RewindTest as it requires three
+# nodes.
+
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More tests => 3;
+
+use File::Copy;
+
+my $tmp_folder = TestLib::tempdir;
+
+my $node_1 = get_new_node('node_1');
+$node_1->init(allows_streaming => 1);
+$node_1->append_conf('postgresql.conf', qq(
+wal_keep_size='100 MB'
+));
+
+$node_1->start;
+
+# Create a couple of test tables
+$node_1->safe_psql('postgres', 'CREATE TABLE public.foo (t TEXT)');
+$node_1->safe_psql('postgres', 'CREATE TABLE public.bar (t TEXT)');
+$node_1->safe_psql('postgres', "INSERT INTO public.bar VALUES ('in both')");
+
+
+# Take backup
+my $backup_name = 'my_backup';
+$node_1->backup($backup_name);
+
+# Create streaming standby from backup
+my $node_2 = get_new_node('node_2');
+$node_2->init_from_backup($node_1, $backup_name,
+	has_streaming => 1);
+$node_2->start;
+
+# Create streaming standby from backup
+my $node_3 = get_new_node('node_3');
+$node_3->init_from_backup($node_1, $backup_name,
+	has_streaming => 1);
+$node_3->start;
+
+# Stop node_1
+
+$node_1->stop('fast');
+
+# Promote node_3
+$node_3->promote;
+
+# node_1 rejoins node_3
+
+my $node_3_connstr = $node_3->connstr;
+
+$node_1->append_conf('postgresql.conf', qq(
+primary_conninfo='$node_3_connstr'
+));
+$node_1->set_standby_mode();
+$node_1->start();
+
+# node_2 follows node_3
+
+$node_2->append_conf('postgresql.conf', qq(
+primary_conninfo='$node_3_connstr'
+));
+$node_2->restart();
+
+# Promote node_1
+
+$node_1->promote;
+
+# We now have a split-brain with two primaries. Insert a row on both to
+# demonstratively create a split brain. After the rewind, we should only
+# see the insert on 1, as the insert on node 3 is rewound away.
+$node_1->safe_psql('postgres', "INSERT INTO public.foo (t) VALUES ('keep this')");
+
+# Insert more rows in node 1, to bump up the XID counter. Otherwise, if
+# rewind doesn't correctly rewind the changes made on the other node,
+# we might fail to notice if the inserts are invisible because the XIDs
+# are not marked as committed.
+$node_1->safe_psql('postgres', "INSERT INTO public.foo (t) VALUES ('and this')");
+$node_1->safe_psql('postgres', "INSERT INTO public.foo (t) VALUES ('and this too')");
+
+# Also insert a row in 'bar' on node 3. It is unmodified in node 1, so it won't get
+# overwritten by replaying the WAL from node 1.
+$node_3->safe_psql('postgres', "INSERT INTO public.bar (t) VALUES ('rewind this')");
+
+# Wait for node 2 to catch up
+$node_2->poll_query_until('postgres',
+	q|SELECT COUNT(*) > 1 FROM public.bar|, 't');
+
+# At this point node_2 will shut down without a shutdown checkpoint,
+# but with WAL entries beyond the preceding shutdown checkpoint.
+$node_2->stop('fast');
+$node_3->stop('fast');
+
+my $node_2_pgdata = $node_2->data_dir;
+my $node_1_connstr = $node_1->connstr;
+
+# Keep a temporary postgresql.conf or it would be overwritten during the rewind.
+copy(
+	"$node_2_pgdata/postgresql.conf",
+	"$tmp_folder/node_2-postgresql.conf.tmp");
+
+command_ok(
+    [
+        'pg_rewind',
+        "--source-server=$node_1_connstr",
+        "--target-pgdata=$node_2_pgdata"
+    ],
+	'pg_rewind detects rewind needed');
+
+# Now move back postgresql.conf with old settings
+move(
+	"$tmp_folder/node_2-postgresql.conf.tmp",
+	"$node_2_pgdata/postgresql.conf");
+
+$node_2->start;
+
+# Check contents of the test tables after rewind. The rows inserted in node 3
+# before rewind should've been overwritten with the data from node 1.
+my $result;
+$result = $node_2->safe_psql('postgres', 'checkpoint');
+$result = $node_2->safe_psql('postgres', 'SELECT * FROM public.foo');
+is($result, qq(keep this
+and this
+and this too), 'table foo after rewind');
+
+$result = $node_2->safe_psql('postgres', 'SELECT * FROM public.bar');
+is($result, qq(in both), 'table bar after rewind');