public inbox for [email protected]  
help / color / mirror / Atom feed
From: Heikki Linnakangas <[email protected]>
Subject: [PATCH v34 1/1] Support checksum enable/disable in a running cluster v33
Date: Wed, 27 Jan 2021 17:20:44 +0200

This allows data checksums to be enabled, or disabled, in a running
cluster without restricting access to the cluster during processing.

A dynamic background worker is responsible for launching a per-database
worker which will mark all buffers dirty for all relation with storage
in order for them to have data checksums on write. A new in-progress
state is introduced which during processing ensures that data checksums
are written but not verified to avoid false negatives. State changes
across backends are synchronized using a procsignalbarrier.

Authors: Daniel Gustafsson, Magnus Hagander
Reviewed-by: Heikki Linnakangas, Robert Haas, Andres Freund, Tomas Vondra, Michael Banck, Andrey Borodin
Discussion: https://postgr.es/m/CABUevExz9hUUOLnJVr2kpw9Cx=o4MCr1SVKwbupzuxP7ckNutA@mail.gmail.com
Discussion: https://postgr.es/m/[email protected]
Discussion: https://postgr.es/m/CABUevEwE3urLtwxxqdgd5O2oQz9J717ZzMbh+ziCSa5YLLU_BA@mail.gmail.com
---
 doc/src/sgml/catalogs.sgml                   |   11 +
 doc/src/sgml/func.sgml                       |   68 +
 doc/src/sgml/monitoring.sgml                 |    6 +-
 doc/src/sgml/ref/pg_checksums.sgml           |    6 +
 doc/src/sgml/wal.sgml                        |   57 +-
 src/backend/access/heap/heapam.c             |    9 +-
 src/backend/access/rmgrdesc/xlogdesc.c       |   18 +
 src/backend/access/transam/xlog.c            |  452 +++++-
 src/backend/access/transam/xlogfuncs.c       |   47 +
 src/backend/catalog/heap.c                   |    7 +
 src/backend/catalog/system_views.sql         |    5 +
 src/backend/postmaster/Makefile              |    1 +
 src/backend/postmaster/bgworker.c            |   10 +
 src/backend/postmaster/datachecksumsworker.c | 1530 ++++++++++++++++++
 src/backend/postmaster/pgstat.c              |    6 +
 src/backend/replication/basebackup.c         |    9 +-
 src/backend/replication/logical/decode.c     |    1 +
 src/backend/storage/buffer/bufmgr.c          |    5 +
 src/backend/storage/ipc/ipci.c               |    3 +
 src/backend/storage/ipc/procsignal.c         |   33 +-
 src/backend/storage/lmgr/lwlocknames.txt     |    1 +
 src/backend/storage/page/README              |    4 +-
 src/backend/storage/page/bufpage.c           |   29 +-
 src/backend/utils/adt/pgstatfuncs.c          |    6 -
 src/backend/utils/cache/relcache.c           |   60 +-
 src/backend/utils/init/miscinit.c            |    6 +
 src/backend/utils/init/postinit.c            |    5 +
 src/backend/utils/misc/guc.c                 |   37 +-
 src/bin/pg_checksums/pg_checksums.c          |    2 +-
 src/bin/pg_upgrade/controldata.c             |    9 +
 src/bin/pg_upgrade/pg_upgrade.h              |    2 +-
 src/include/access/xlog.h                    |   19 +-
 src/include/access/xlog_internal.h           |    7 +
 src/include/catalog/pg_class.h               |    3 +
 src/include/catalog/pg_control.h             |    1 +
 src/include/catalog/pg_proc.dat              |   16 +
 src/include/miscadmin.h                      |    2 +
 src/include/pgstat.h                         |    2 +
 src/include/postmaster/datachecksumsworker.h |   30 +
 src/include/storage/bufpage.h                |    3 +
 src/include/storage/checksum.h               |    8 +
 src/include/storage/procsignal.h             |   10 +-
 src/test/Makefile                            |    2 +-
 src/test/checksum/.gitignore                 |    2 +
 src/test/checksum/Makefile                   |   23 +
 src/test/checksum/README                     |   22 +
 src/test/checksum/t/001_basic.pl             |   92 ++
 src/test/checksum/t/002_restarts.pl          |  117 ++
 src/test/checksum/t/003_standby_checksum.pl  |  127 ++
 src/test/checksum/t/004_offline.pl           |  105 ++
 src/test/perl/PostgresNode.pm                |   36 +
 51 files changed, 2985 insertions(+), 87 deletions(-)
 create mode 100644 src/backend/postmaster/datachecksumsworker.c
 create mode 100644 src/include/postmaster/datachecksumsworker.h
 create mode 100644 src/test/checksum/.gitignore
 create mode 100644 src/test/checksum/Makefile
 create mode 100644 src/test/checksum/README
 create mode 100644 src/test/checksum/t/001_basic.pl
 create mode 100644 src/test/checksum/t/002_restarts.pl
 create mode 100644 src/test/checksum/t/003_standby_checksum.pl
 create mode 100644 src/test/checksum/t/004_offline.pl

diff --git a/doc/src/sgml/catalogs.sgml b/doc/src/sgml/catalogs.sgml
index 865e826fb0b..75cc1588a5c 100644
--- a/doc/src/sgml/catalogs.sgml
+++ b/doc/src/sgml/catalogs.sgml
@@ -2166,6 +2166,17 @@ SCRAM-SHA-256$<replaceable>&lt;iteration count&gt;</replaceable>:<replaceable>&l
       </para></entry>
      </row>
 
+     <row>
+      <entry role="catalog_table_entry"><para role="column_definition">
+       <structfield>relhaschecksums</structfield> <type>bool</type>
+      </para>
+      <para>
+        True if relation has data checksums on all pages. This state is only
+        used during checksum processing; this field should never be consulted
+        for cluster checksum status.
+      </para></entry>
+     </row>
+
      <row>
       <entry role="catalog_table_entry"><para role="column_definition">
        <structfield>relrewrite</structfield> <type>oid</type>
diff --git a/doc/src/sgml/func.sgml b/doc/src/sgml/func.sgml
index aa99665e2eb..94182fb7b1e 100644
--- a/doc/src/sgml/func.sgml
+++ b/doc/src/sgml/func.sgml
@@ -25839,6 +25839,74 @@ postgres=# SELECT * FROM pg_walfile_name_offset(pg_stop_backup());
 
   </sect2>
 
+  <sect2 id="functions-admin-checksum">
+   <title>Data Checksum Functions</title>
+
+   <para>
+    The functions shown in <xref linkend="functions-checksums-table" /> can
+    be used to enable or disable data checksums in a running cluster.
+    See <xref linkend="checksums" /> for details.
+   </para>
+
+   <table id="functions-checksums-table">
+    <title>Data Checksum Functions</title>
+    <tgroup cols="1">
+     <thead>
+      <row>
+       <entry role="func_table_entry"><para role="func_signature">
+        Function
+       </para>
+       <para>
+        Description
+       </para></entry>
+      </row>
+     </thead>
+
+     <tbody>
+      <row>
+       <entry role="func_table_entry"><para role="func_signature">
+        <indexterm>
+         <primary>pg_enable_data_checksums</primary>
+        </indexterm>
+        <function>pg_enable_data_checksums</function> ( <optional><parameter>cost_delay</parameter> <type>int</type>, <parameter>cost_limit</parameter> <type>int</type></optional> )
+        <returnvalue>void</returnvalue>
+       </para>
+       <para>
+        Initiates data checksums for the cluster. This will switch the data
+        checksums mode to <literal>inprogress-on</literal> as well as start a
+        background worker that will process all data in the database and enable
+        checksums for it. When all data pages have had checksums enabled, the
+        cluster will automatically switch data checksums mode to
+        <literal>on</literal>.
+       </para>
+       <para>
+        If <parameter>cost_delay</parameter> and <parameter>cost_limit</parameter> are
+        specified, the speed of the process is throttled using the same principles as
+        <link linkend="runtime-config-resource-vacuum-cost">Cost-based Vacuum Delay</link>.
+       </para></entry>
+      </row>
+
+      <row>
+       <entry role="func_table_entry"><para role="func_signature">
+        <indexterm>
+         <primary>pg_disable_data_checksums</primary>
+        </indexterm>
+        <function>pg_disable_data_checksums</function> ()
+        <returnvalue>void</returnvalue>
+       </para>
+       <para>
+        Disables data checksums for the cluster. This will switch the data
+        checksum mode to <literal>inprogress-off</literal> while data checksums
+        are being disabled. When all active backends have ceased to validate
+        data checksums, the data checksum mode will be changed to <literal>off</literal>.
+       </para></entry>
+      </row>
+     </tbody>
+    </tgroup>
+   </table>
+
+  </sect2>
+
   <sect2 id="functions-admin-dbobject">
    <title>Database Object Management Functions</title>
 
diff --git a/doc/src/sgml/monitoring.sgml b/doc/src/sgml/monitoring.sgml
index 9496f76b1fb..7e170ec4299 100644
--- a/doc/src/sgml/monitoring.sgml
+++ b/doc/src/sgml/monitoring.sgml
@@ -3695,8 +3695,7 @@ SELECT pid, wait_event_type, wait_event FROM pg_stat_activity WHERE wait_event i
       </para>
       <para>
        Number of data page checksum failures detected in this
-       database (or on a shared object), or NULL if data checksums are not
-       enabled.
+       database.
       </para></entry>
      </row>
 
@@ -3706,8 +3705,7 @@ SELECT pid, wait_event_type, wait_event FROM pg_stat_activity WHERE wait_event i
       </para>
       <para>
        Time at which the last data page checksum failure was detected in
-       this database (or on a shared object), or NULL if data checksums are not
-       enabled.
+       this database (or on a shared object).
       </para></entry>
      </row>
 
diff --git a/doc/src/sgml/ref/pg_checksums.sgml b/doc/src/sgml/ref/pg_checksums.sgml
index c84bc5c5b23..d879550e81c 100644
--- a/doc/src/sgml/ref/pg_checksums.sgml
+++ b/doc/src/sgml/ref/pg_checksums.sgml
@@ -45,6 +45,12 @@ PostgreSQL documentation
    exit status is nonzero if the operation failed.
   </para>
 
+  <para>
+   When enabling checksums, if checksums were in the process of being enabled
+   when the cluster was shut down, <application>pg_checksums</application>
+   will still process all relations regardless of the online processing.
+  </para>
+
   <para>
    When verifying checksums, every file in the cluster is scanned. When
    enabling checksums, every file in the cluster is rewritten in-place.
diff --git a/doc/src/sgml/wal.sgml b/doc/src/sgml/wal.sgml
index 66de1ee2f81..48890ccc9d3 100644
--- a/doc/src/sgml/wal.sgml
+++ b/doc/src/sgml/wal.sgml
@@ -247,9 +247,10 @@
   <para>
    Checksums are normally enabled when the cluster is initialized using <link
    linkend="app-initdb-data-checksums"><application>initdb</application></link>.
-   They can also be enabled or disabled at a later time as an offline
-   operation. Data checksums are enabled or disabled at the full cluster
-   level, and cannot be specified individually for databases or tables.
+   They can also be enabled or disabled at a later time either as an offline
+   operation or online in a running cluster allowing concurrent access. Data
+   checksums are enabled or disabled at the full cluster level, and cannot be
+   specified individually for databases or tables.
   </para>
 
   <para>
@@ -266,7 +267,7 @@
   </para>
 
   <sect2 id="checksums-offline-enable-disable">
-   <title>Off-line Enabling of Checksums</title>
+   <title>Offline Enabling of Checksums</title>
 
    <para>
     The <link linkend="app-pgchecksums"><application>pg_checksums</application></link>
@@ -275,6 +276,54 @@
    </para>
 
   </sect2>
+
+  <sect2 id="checksums-online-enable-disable">
+   <title>Online Enabling of Checksums</title>
+
+   <para>
+    Checksums can be enabled or disabled online, by calling the appropriate
+    <link linkend="functions-admin-checksum">functions</link>.
+   </para>
+
+   <para>
+    Enabling checksums will put the cluster checksum mode in
+    <literal>inprogress-on</literal> mode.  During this time, checksums will be
+    written but not verified. In addition to this, a background worker process
+    is started that enables checksums on all existing data in the cluster. Once
+    this worker has completed processing all databases in the cluster, the
+    checksum mode will automatically switch to <literal>on</literal>. The
+    processing will consume a background worker process, make sure that
+    <varname>max_worker_processes</varname> allows for at least one more
+    additional process.
+   </para>
+
+   <para>
+    The process will initially wait for all open transactions to finish before
+    it starts, so that it can be certain that there are no tables that have been
+    created inside a transaction that has not committed yet and thus would not
+    be visible to the process enabling checksums. It will also, for each database,
+    wait for all pre-existing temporary tables to get removed before it finishes.
+    If long-lived temporary tables are used in the application it may be necessary
+    to terminate these application connections to allow the process to complete.
+   </para>
+
+   <para>
+    If the cluster is stopped while in <literal>inprogress-on</literal> mode, for
+    any reason, then this process must be restarted manually. To do this,
+    re-execute the function <function>pg_enable_data_checksums()</function>
+    once the cluster has been restarted. The background worker will attempt
+    to resume the work from where it was interrupted.
+   </para>
+
+   <note>
+    <para>
+     Enabling checksums can cause significant I/O to the system, as most of the
+     database pages will need to be rewritten, and will be written both to the
+     data files and the WAL.
+    </para>
+   </note>
+
+  </sect2>
  </sect1>
 
   <sect1 id="wal-intro">
diff --git a/src/backend/access/heap/heapam.c b/src/backend/access/heap/heapam.c
index 9926e2bd546..ffcd8899082 100644
--- a/src/backend/access/heap/heapam.c
+++ b/src/backend/access/heap/heapam.c
@@ -7927,7 +7927,7 @@ log_heap_freeze(Relation reln, Buffer buffer, TransactionId cutoff_xid,
  * and dirtied.
  *
  * If checksums are enabled, we also generate a full-page image of
- * heap_buffer, if necessary.
+ * heap_buffer.
  */
 XLogRecPtr
 log_heap_visible(RelFileNode rnode, Buffer heap_buffer, Buffer vm_buffer,
@@ -7948,11 +7948,18 @@ log_heap_visible(RelFileNode rnode, Buffer heap_buffer, Buffer vm_buffer,
 	XLogRegisterBuffer(0, vm_buffer, 0);
 
 	flags = REGBUF_STANDARD;
+	/*
+	 * Hold interrupts for the duration of xlogging to avoid the state of data
+	 * checksums changing during the processing which would alter the premise
+	 * for xlogging hint bits.
+	 */
+	HOLD_INTERRUPTS();
 	if (!XLogHintBitIsNeeded())
 		flags |= REGBUF_NO_IMAGE;
 	XLogRegisterBuffer(1, heap_buffer, flags);
 
 	recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE);
+	RESUME_INTERRUPTS();
 
 	return recptr;
 }
diff --git a/src/backend/access/rmgrdesc/xlogdesc.c b/src/backend/access/rmgrdesc/xlogdesc.c
index 92cc7ea0735..fa074c6046f 100644
--- a/src/backend/access/rmgrdesc/xlogdesc.c
+++ b/src/backend/access/rmgrdesc/xlogdesc.c
@@ -18,6 +18,7 @@
 #include "access/xlog.h"
 #include "access/xlog_internal.h"
 #include "catalog/pg_control.h"
+#include "storage/bufpage.h"
 #include "utils/guc.h"
 #include "utils/timestamp.h"
 
@@ -140,6 +141,20 @@ xlog_desc(StringInfo buf, XLogReaderState *record)
 						 xlrec.ThisTimeLineID, xlrec.PrevTimeLineID,
 						 timestamptz_to_str(xlrec.end_time));
 	}
+	else if (info == XLOG_CHECKSUMS)
+	{
+		xl_checksum_state xlrec;
+
+		memcpy(&xlrec, rec, sizeof(xl_checksum_state));
+		if (xlrec.new_checksumtype == PG_DATA_CHECKSUM_VERSION)
+			appendStringInfo(buf, "on");
+		else if (xlrec.new_checksumtype == PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION)
+			appendStringInfo(buf, "inprogress-off");
+		else if (xlrec.new_checksumtype == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION)
+			appendStringInfo(buf, "inprogress-on");
+		else
+			appendStringInfo(buf, "off");
+	}
 }
 
 const char *
@@ -185,6 +200,9 @@ xlog_identify(uint8 info)
 		case XLOG_FPI_FOR_HINT:
 			id = "FPI_FOR_HINT";
 			break;
+		case XLOG_CHECKSUMS:
+			id = "CHECKSUMS";
+			break;
 	}
 
 	return id;
diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c
index cc007b8963e..8531def93c1 100644
--- a/src/backend/access/transam/xlog.c
+++ b/src/backend/access/transam/xlog.c
@@ -38,6 +38,7 @@
 #include "access/xlogreader.h"
 #include "access/xlogutils.h"
 #include "catalog/catversion.h"
+#include "catalog/pg_class.h"
 #include "catalog/pg_control.h"
 #include "catalog/pg_database.h"
 #include "commands/progress.h"
@@ -50,6 +51,7 @@
 #include "port/atomics.h"
 #include "port/pg_iovec.h"
 #include "postmaster/bgwriter.h"
+#include "postmaster/datachecksumsworker.h"
 #include "postmaster/startup.h"
 #include "postmaster/walwriter.h"
 #include "replication/basebackup.h"
@@ -253,6 +255,16 @@ static bool LocalPromoteIsTriggered = false;
  */
 static int	LocalXLogInsertAllowed = -1;
 
+/*
+ * Local state for Controlfile data_checksum_version. After initialization,
+ * this is only updated when absorbing a procsignal barrier during interrupt
+ * processing.  The reason for keeping a copy in backend-private memory is to
+ * avoid locking for interrogating checksum state.  Possible values are the
+ * checksum versions defined in storage/bufpage.h and zero for when checksums
+ * are disabled.
+ */
+static uint32 LocalDataChecksumVersion = 0;
+
 /*
  * When ArchiveRecoveryRequested is set, archive recovery was requested,
  * ie. signal files were present. When InArchiveRecovery is set, we are
@@ -900,6 +912,7 @@ static void SetLatestXTime(TimestampTz xtime);
 static void SetCurrentChunkStartTime(TimestampTz xtime);
 static void CheckRequiredParameterValues(void);
 static void XLogReportParameters(void);
+static void XLogChecksums(ChecksumType new_type);
 static void checkTimeLineSwitch(XLogRecPtr lsn, TimeLineID newTLI,
 								TimeLineID prevTLI);
 static void LocalSetXLogInsertAllowed(void);
@@ -1073,8 +1086,8 @@ XLogInsertRecord(XLogRecData *rdata,
 	 * and fast otherwise.
 	 *
 	 * Also check to see if fullPageWrites or forcePageWrites was just turned
-	 * on; if we weren't already doing full-page writes then go back and
-	 * recompute.
+	 * on, or of we are in the process of enabling checksums in the cluster;
+	 * if we weren't already doing full-page writes then go back and recompute.
 	 *
 	 * If we aren't doing full-page writes then RedoRecPtr doesn't actually
 	 * affect the contents of the XLOG record, so we'll update our local copy
@@ -1087,7 +1100,7 @@ XLogInsertRecord(XLogRecData *rdata,
 		Assert(RedoRecPtr < Insert->RedoRecPtr);
 		RedoRecPtr = Insert->RedoRecPtr;
 	}
-	doPageWrites = (Insert->fullPageWrites || Insert->forcePageWrites);
+	doPageWrites = (Insert->fullPageWrites || Insert->forcePageWrites || DataChecksumsOnInProgress());
 
 	if (doPageWrites &&
 		(!prevDoPageWrites ||
@@ -4915,9 +4928,7 @@ ReadControlFile(void)
 
 	CalculateCheckpointSegments();
 
-	/* Make the initdb settings visible as GUC variables, too */
-	SetConfigOption("data_checksums", DataChecksumsEnabled() ? "yes" : "no",
-					PGC_INTERNAL, PGC_S_OVERRIDE);
+	LocalDataChecksumVersion = ControlFile->data_checksum_version;
 }
 
 /*
@@ -4951,13 +4962,370 @@ GetMockAuthenticationNonce(void)
 }
 
 /*
- * Are checksums enabled for data pages?
+ * DataChecksumsNeedWrite
+ *		Returns whether data checksums must be written or not
+ *
+ * Returns true iff data checksums are enabled or are in the process of being
+ * enabled.   During "inprogress-on" and "inprogress-off" states checksums must
+ * be written even though they are not verified (see datachecksumsworker.c for
+ * a longer discussion).
+ *
+ * This function is intedewd for callsites which are about to write a data page
+ * to storage, and need to know whether to re-calculate the checksum for the
+ * page header. Interrupts must be held off during calling this and until the
+ * write operation has finished to avoid the risk of the checksum state
+ * changing. This implies that calling this function must be performed as close
+ * to write operation as possible to keep the critical section short.
+ */
+bool
+DataChecksumsNeedWrite(void)
+{
+	Assert(InterruptHoldoffCount > 0);
+	return (LocalDataChecksumVersion == PG_DATA_CHECKSUM_VERSION ||
+			LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION ||
+			LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION);
+}
+
+/*
+ * DataChecksumsNeedVerify
+ *		Returns whether data checksums must be verified or not
+ *
+ * Data checksums are only verified if they are fully enabled in the cluster.
+ * During the "inprogress-on" and "inprogress-off" states they are only
+ * updated, not verified (see datachecksumsworker.c for a longer discussion).
+ *
+ * This function is intended for callsites which have read data and are about
+ * to perform checksum validation based on the result of this. To avoid the
+ * the risk of the checksum state changing between reading and performing the
+ * validation (or not), interrupts must be held off. This implies that calling
+ * this function must be performed as close to the validation call as possible
+ * to keep the critical section short. This is in order to protect against
+ * time of check/time of use situations around data checksum validation.
+ */
+bool
+DataChecksumsNeedVerify(void)
+{
+	Assert(InterruptHoldoffCount > 0);
+	return (LocalDataChecksumVersion == PG_DATA_CHECKSUM_VERSION);
+}
+
+/*
+ * DataChecksumsOnInProgress
+ *		Returns whether data checksums are being enabled
+ *
+ * Most operations don't need to worry about the "inprogress" states, and
+ * should use DataChecksumsNeedVerify() or DataChecksumsNeedWrite(). The
+ * "inprogress-on" state for enabling checksums is used when the checksum
+ * worker is setting checksums on all pages, it can thus be used to check for
+ * aborted checksum processing which need to be restarted.
+ */
+inline bool
+DataChecksumsOnInProgress(void)
+{
+	return (LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION);
+}
+
+/*
+ * DataChecksumsOffInProgress
+ *		Returns whether data checksums are being disabled
+ *
+ * The "inprogress-off" state for disabling checksums is used for when the
+ * worker resets the catalog state.  DataChecksumsNeedVerify() or
+ * DataChecksumsNeedWrite() should be used for deciding whether to read/write
+ * checksums.
  */
 bool
-DataChecksumsEnabled(void)
+DataChecksumsOffInProgress(void)
+{
+	return (LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION);
+}
+
+/*
+ * SetDataChecksumsOnInProgress
+ *		Sets the data checksum state to "inprogress-on" to enable checksums
+ *
+ * To start the process of enabling data checksums in a running cluster the
+ * data_checksum_version state must be changed to "inprogress-on". See
+ * SetDataChecksumsOn below for a description on how this state change works.
+ * This function blocks until all backends in the cluster have acknowledged the
+ * state transition.
+ */
+void
+SetDataChecksumsOnInProgress(void)
+{
+	uint64		barrier;
+
+	Assert(ControlFile != NULL);
+
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+
+	/*
+	 * Data checksum state can only be transitioned to "inprogress-on" from
+	 * "off", if data checksums are in any other state then exit.
+	 */
+	if (ControlFile->data_checksum_version != 0)
+	{
+		LWLockRelease(ControlFileLock);
+		return;
+	}
+
+	LWLockRelease(ControlFileLock);
+
+	/*
+	 * The state transition is performed in a critical section with
+	 * checkpoints held off to provide crash safety.
+	 */
+	MyProc->delayChkpt = true;
+	START_CRIT_SECTION();
+
+	XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION);
+
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+	ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION;
+	UpdateControlFile();
+	LWLockRelease(ControlFileLock);
+
+	barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON);
+
+	END_CRIT_SECTION();
+	MyProc->delayChkpt = false;
+
+	/*
+	 * Await state change in all backends to ensure that all backends are in
+	 * "inprogress-on". Once done we know that all backends are writing data
+	 * checksums.
+	 */
+	WaitForProcSignalBarrier(barrier);
+}
+
+/*
+ * SetDataChecksumsOn
+ *		Enables data checksums cluster-wide
+ *
+ * Enabling data checksums is performed using two barriers, the first one to
+ * set the checksums state to "inprogress-on" (which is performed by
+ * SetDataChecksumsOnInProgress()) and the second one to set the state to "on"
+ * (performed here).
+ *
+ * To start the process of enabling data checksums in a running cluster the
+ * data_checksum_version state must be changed to "inprogress-on".  This state
+ * requires data checksums to be written but not verified. This ensures that
+ * all data pages can be checksummed without the risk of false negatives in
+ * validation during the process.  When all existing pages are guaranteed to
+ * have checksums, and all new pages will be initiated with checksums, the
+ * state can be changed to "on". Once the state is "on" checksums will be both
+ * written and verified. See datachecksumsworker.c for a longer discussion on
+ * how data checksums can be enabled in a running cluster.
+ *
+ * This function blocks until all backends in the cluster have acknowledged the
+ * state transition.
+ */
+void
+SetDataChecksumsOn(void)
 {
+	uint64		barrier;
+
 	Assert(ControlFile != NULL);
-	return (ControlFile->data_checksum_version > 0);
+
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+
+	/*
+	 * The only allowed state transition to "on" is from "inprogress-on" since
+	 * that state ensures that all pages will have data checksums written.
+	 */
+	if (ControlFile->data_checksum_version != PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION)
+	{
+		LWLockRelease(ControlFileLock);
+		elog(ERROR, "checksums not in \"inprogress-on\" mode");
+	}
+
+	LWLockRelease(ControlFileLock);
+
+	MyProc->delayChkpt = true;
+	START_CRIT_SECTION();
+
+	XLogChecksums(PG_DATA_CHECKSUM_VERSION);
+
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+	ControlFile->data_checksum_version = PG_DATA_CHECKSUM_VERSION;
+	UpdateControlFile();
+	LWLockRelease(ControlFileLock);
+
+	barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_ON);
+
+	END_CRIT_SECTION();
+	MyProc->delayChkpt = false;
+
+	/*
+	 * Await state transition of "on" in all backends. When done we know that
+	 * data checksums are enabled in all backends and data checksums are both
+	 * written and verified.
+	 */
+	WaitForProcSignalBarrier(barrier);
+}
+
+/*
+ * SetDataChecksumsOff
+ *		Disables data checksums cluster-wide
+ *
+ * Disabling data checksums must be performed with two sets of barriers, each
+ * carrying a different state. The state is first set to "inprogress-off"
+ * during which checksums are still written but not verified. This ensures that
+ * backends which have yet to observe the state change from "on" won't get
+ * validation errors on concurrently modified pages. Once all backends have
+ * changed to "inprogress-off", the barrier for moving to "off" can be emitted.
+ * This function blocks until all backends in the cluster have acknowledged the
+ * state transition.
+ */
+void
+SetDataChecksumsOff(void)
+{
+	uint64		barrier;
+
+	Assert(ControlFile);
+
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+
+	/* If data checksums are already disabled there is nothing to do */
+	if (ControlFile->data_checksum_version == 0)
+	{
+		LWLockRelease(ControlFileLock);
+		return;
+	}
+
+	/*
+	 * If data checksums are currently enabled we first transition to the
+	 * "inprogress-off" state during which backends continue to write
+	 * checksums without verifying them. When all backends are in
+	 * "inprogress-off" the next transition to "off" can be performed, after
+	 * which all data checksum processing is disabled.
+	 */
+	if (ControlFile->data_checksum_version == PG_DATA_CHECKSUM_VERSION)
+	{
+		LWLockRelease(ControlFileLock);
+
+		MyProc->delayChkpt = true;
+		START_CRIT_SECTION();
+
+		XLogChecksums(PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION);
+
+		LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+		ControlFile->data_checksum_version = PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION;
+		UpdateControlFile();
+		LWLockRelease(ControlFileLock);
+
+		barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF);
+
+		END_CRIT_SECTION();
+		MyProc->delayChkpt = false;
+
+		/*
+		 * Update local state in all backends to ensure that any backend in
+		 * "on" state is changed to "inprogress-off".
+		 */
+		WaitForProcSignalBarrier(barrier);
+
+		/*
+		 * At this point we know that no backends are verifying data checksums
+		 * during reading. Next, we can safely move to state "off" to also
+		 * stop writing checksums.
+		 */
+	}
+	else
+	{
+		/*
+		 * Ending up here implies that the checksums state is "inprogress-on"
+		 * or "inprogress-off" and we can transition directly to "off" from
+		 * there.
+		 */
+		LWLockRelease(ControlFileLock);
+	}
+
+	/*
+	 * Ensure that we don't incur a checkpoint during disabling checksums.
+	 */
+	MyProc->delayChkpt = true;
+	START_CRIT_SECTION();
+
+	XLogChecksums(0);
+
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+	ControlFile->data_checksum_version = 0;
+	UpdateControlFile();
+	LWLockRelease(ControlFileLock);
+
+	barrier = EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_OFF);
+
+	END_CRIT_SECTION();
+	MyProc->delayChkpt = false;
+
+	WaitForProcSignalBarrier(barrier);
+}
+
+/*
+ * ProcSignalBarrier absorption functions for enabling and disabling data
+ * checksums in a running cluster. The procsignalbarriers are emitted in the
+ * SetDataChecksums* functions.
+ */
+bool
+AbsorbChecksumsOnInProgressBarrier(void)
+{
+	Assert(LocalDataChecksumVersion == 0 ||
+		   LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION);
+	LocalDataChecksumVersion = PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION;
+	return true;
+}
+
+bool
+AbsorbChecksumsOnBarrier(void)
+{
+	Assert(LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION);
+	LocalDataChecksumVersion = PG_DATA_CHECKSUM_VERSION;
+	return true;
+}
+
+bool
+AbsorbChecksumsOffInProgressBarrier(void)
+{
+	LocalDataChecksumVersion = PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION;
+	return true;
+}
+
+bool
+AbsorbChecksumsOffBarrier(void)
+{
+	LocalDataChecksumVersion = 0;
+	return true;
+}
+
+/*
+ * InitLocalControlData
+ *
+ * Set up backend local caches of controldata variables which may change at
+ * any point during runtime and thus require special cased locking. So far
+ * this only applies to data_checksum_version, but it's intended to be general
+ * purpose enough to handle future cases.
+ */
+void
+InitLocalControldata(void)
+{
+	LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+	LocalDataChecksumVersion = ControlFile->data_checksum_version;
+	LWLockRelease(ControlFileLock);
+}
+
+/* guc hook */
+const char *
+show_data_checksums(void)
+{
+	if (LocalDataChecksumVersion == PG_DATA_CHECKSUM_VERSION)
+		return "on";
+	else if (LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION)
+		return "inprogress-on";
+	else if (LocalDataChecksumVersion == PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION)
+		return "inprogress-off";
+	else
+		return "off";
 }
 
 /*
@@ -7991,6 +8359,32 @@ StartupXLOG(void)
 	 */
 	CompleteCommitTsInitialization();
 
+	/*
+	 * If we reach this point with checksums being enabled ("inprogress-on"
+	 * state), we notify the user that they need to manually restart the
+	 * process to enable checksums. This is because we cannot launch a dynamic
+	 * background worker directly from here, it has to be launched from a
+	 * regular backend.
+	 */
+	if (ControlFile->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION)
+		ereport(WARNING,
+				(errmsg("data checksums are being enabled, but no worker is running"),
+				 errhint("Either disable or enable data checksums by calling the pg_disable_data_checksums() or pg_enable_data_checksums() functions.")));
+
+	/*
+	 * If data checksums were being disabled when the cluster was shutdown, we
+	 * know that we have a state where all backends have stopped validating
+	 * checksums and we can move to off instead.
+	 */
+	if (ControlFile->data_checksum_version == PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION)
+	{
+		XLogChecksums(0);
+
+		LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+		ControlFile->data_checksum_version = 0;
+		LWLockRelease(ControlFileLock);
+	}
+
 	/*
 	 * All done with end-of-recovery actions.
 	 *
@@ -9900,6 +10294,24 @@ XLogReportParameters(void)
 	}
 }
 
+/*
+ * Log the new state of checksums
+ */
+static void
+XLogChecksums(ChecksumType new_type)
+{
+	xl_checksum_state xlrec;
+	XLogRecPtr	recptr;
+
+	xlrec.new_checksumtype = new_type;
+
+	XLogBeginInsert();
+	XLogRegisterData((char *) &xlrec, sizeof(xl_checksum_state));
+
+	recptr = XLogInsert(RM_XLOG_ID, XLOG_CHECKSUMS);
+	XLogFlush(recptr);
+}
+
 /*
  * Update full_page_writes in shared memory, and write an
  * XLOG_FPW_CHANGE record if necessary.
@@ -10355,6 +10767,28 @@ xlog_redo(XLogReaderState *record)
 		/* Keep track of full_page_writes */
 		lastFullPageWrites = fpw;
 	}
+	else if (info == XLOG_CHECKSUMS)
+	{
+		xl_checksum_state state;
+
+		memcpy(&state, XLogRecGetData(record), sizeof(xl_checksum_state));
+
+		LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
+		ControlFile->data_checksum_version = state.new_checksumtype;
+		UpdateControlFile();
+		LWLockRelease(ControlFileLock);
+		if (state.new_checksumtype == PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION)
+			WaitForProcSignalBarrier(EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON));
+		else if (state.new_checksumtype == PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION)
+			WaitForProcSignalBarrier(EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF));
+		else if (state.new_checksumtype == PG_DATA_CHECKSUM_VERSION)
+			WaitForProcSignalBarrier(EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_ON));
+		else
+		{
+			Assert(state.new_checksumtype == 0);
+			WaitForProcSignalBarrier(EmitProcSignalBarrier(PROCSIGNAL_BARRIER_CHECKSUM_OFF));
+		}
+	}
 }
 
 #ifdef WAL_DEBUG
diff --git a/src/backend/access/transam/xlogfuncs.c b/src/backend/access/transam/xlogfuncs.c
index 5e1aab319dd..5d77be8a2dc 100644
--- a/src/backend/access/transam/xlogfuncs.c
+++ b/src/backend/access/transam/xlogfuncs.c
@@ -25,6 +25,7 @@
 #include "catalog/pg_type.h"
 #include "funcapi.h"
 #include "miscadmin.h"
+#include "postmaster/datachecksumsworker.h"
 #include "pgstat.h"
 #include "replication/walreceiver.h"
 #include "storage/fd.h"
@@ -784,3 +785,49 @@ pg_promote(PG_FUNCTION_ARGS)
 			(errmsg("server did not promote within %d seconds", wait_seconds)));
 	PG_RETURN_BOOL(false);
 }
+
+/*
+ * Disables checksums for the cluster, unless already disabled.
+ *
+ * Starts a background worker that turns off data checksums.
+ */
+Datum
+disable_data_checksums(PG_FUNCTION_ARGS)
+{
+	if (!superuser())
+		ereport(ERROR,
+				(errmsg("must be superuser")));
+
+	StartDatachecksumsWorkerLauncher(false, 0, 0);
+
+	PG_RETURN_VOID();
+}
+
+/*
+ * Enables checksums for the cluster, unless already enabled.
+ *
+ * Supports vacuum-like cost-based throttling, to limit system load.
+ * Starts a background worker that updates checksums on existing data.
+ */
+Datum
+enable_data_checksums(PG_FUNCTION_ARGS)
+{
+	int			cost_delay = PG_GETARG_INT32(0);
+	int			cost_limit = PG_GETARG_INT32(1);
+
+	if (!superuser())
+		ereport(ERROR,
+				(errmsg("must be superuser")));
+
+	if (cost_delay < 0)
+		ereport(ERROR,
+				(errmsg("cost delay cannot be less than zero")));
+
+	if (cost_limit <= 0)
+		ereport(ERROR,
+				(errmsg("cost limit must be a positive value")));
+
+	StartDatachecksumsWorkerLauncher(true, cost_delay, cost_limit);
+
+	PG_RETURN_VOID();
+}
diff --git a/src/backend/catalog/heap.c b/src/backend/catalog/heap.c
index 9abc4a1f556..87052b06930 100644
--- a/src/backend/catalog/heap.c
+++ b/src/backend/catalog/heap.c
@@ -974,10 +974,17 @@ InsertPgClassTuple(Relation pg_class_desc,
 	/* relpartbound is set by updating this tuple, if necessary */
 	nulls[Anum_pg_class_relpartbound - 1] = true;
 
+	/*
+	 * Hold off interrupts to ensure that the observed data checksum state
+	 * cannot change as we form and insert the tuple.
+	 */
+	HOLD_INTERRUPTS();
+	values[Anum_pg_class_relhaschecksums - 1] = BoolGetDatum(DataChecksumsNeedWrite());
 	tup = heap_form_tuple(RelationGetDescr(pg_class_desc), values, nulls);
 
 	/* finally insert the new tuple, update the indexes, and clean up */
 	CatalogTupleInsert(pg_class_desc, tup);
+	RESUME_INTERRUPTS();
 
 	heap_freetuple(tup);
 }
diff --git a/src/backend/catalog/system_views.sql b/src/backend/catalog/system_views.sql
index fa58afd9d78..516ae666b7a 100644
--- a/src/backend/catalog/system_views.sql
+++ b/src/backend/catalog/system_views.sql
@@ -1264,6 +1264,11 @@ CREATE OR REPLACE FUNCTION
   RETURNS boolean STRICT VOLATILE LANGUAGE INTERNAL AS 'pg_promote'
   PARALLEL SAFE;
 
+CREATE OR REPLACE FUNCTION pg_enable_data_checksums (
+        cost_delay int DEFAULT 0, cost_limit int DEFAULT 100)
+  RETURNS void STRICT VOLATILE LANGUAGE internal AS 'enable_data_checksums'
+  PARALLEL RESTRICTED;
+
 -- legacy definition for compatibility with 9.3
 CREATE OR REPLACE FUNCTION
   json_populate_record(base anyelement, from_json json, use_json_as_text boolean DEFAULT false)
diff --git a/src/backend/postmaster/Makefile b/src/backend/postmaster/Makefile
index bfdf6a833db..59b82ee9ce5 100644
--- a/src/backend/postmaster/Makefile
+++ b/src/backend/postmaster/Makefile
@@ -17,6 +17,7 @@ OBJS = \
 	bgworker.o \
 	bgwriter.o \
 	checkpointer.o \
+	datachecksumsworker.o \
 	fork_process.o \
 	interrupt.o \
 	pgarch.o \
diff --git a/src/backend/postmaster/bgworker.c b/src/backend/postmaster/bgworker.c
index dd3dad3de35..8afbf762afc 100644
--- a/src/backend/postmaster/bgworker.c
+++ b/src/backend/postmaster/bgworker.c
@@ -18,6 +18,7 @@
 #include "pgstat.h"
 #include "port/atomics.h"
 #include "postmaster/bgworker_internals.h"
+#include "postmaster/datachecksumsworker.h"
 #include "postmaster/interrupt.h"
 #include "postmaster/postmaster.h"
 #include "replication/logicallauncher.h"
@@ -128,6 +129,15 @@ static const struct
 	},
 	{
 		"ApplyWorkerMain", ApplyWorkerMain
+	},
+	{
+		"DatachecksumsWorkerLauncherMain", DatachecksumsWorkerLauncherMain
+	},
+	{
+		"DatachecksumsWorkerMain", DatachecksumsWorkerMain
+	},
+	{
+		"ResetDataChecksumsStateInDatabase", ResetDataChecksumsStateInDatabase
 	}
 };
 
diff --git a/src/backend/postmaster/datachecksumsworker.c b/src/backend/postmaster/datachecksumsworker.c
new file mode 100644
index 00000000000..b26c31e8924
--- /dev/null
+++ b/src/backend/postmaster/datachecksumsworker.c
@@ -0,0 +1,1530 @@
+/*-------------------------------------------------------------------------
+ *
+ * datachecksumsworker.c
+ *	  Background worker for enabling or disabling data checksums online
+ *
+ * When enabling data checksums on a database at initdb time or with
+ * pg_checksums, no extra process is required as each page is checksummed, and
+ * verified, when accessed.  When enabling checksums on an already running
+ * cluster, which does not run with checksums enabled, this worker will ensure
+ * that all pages are checksummed before verification of the checksums is
+ * turned on. In the case of disabling checksums, the state transition is
+ * recorded in the catalog and control file, and no changes are performed
+ * on the data pages or in the catalog.
+ *
+ * Checksums can be either enabled or disabled cluster-wide, with on/off being
+ * the end state for data_checksums.
+ *
+ * Enabling checksums
+ * ------------------
+ * When enabling checksums in an online cluster, data_checksums will be set to
+ * "inprogress-on" which signals that write operations MUST compute and write
+ * the checksum on the data page, but during reading the checksum SHALL NOT be
+ * verified. This ensures that all objects created during checksumming will
+ * have checksums set, but no reads will fail due to incorrect checksum. The
+ * DataChecksumsWorker will compile a list of databases which exist at the
+ * start of checksumming, and all of these which haven't been dropped during
+ * the processing MUST have been processed successfully in order for checksums
+ * to be enabled. Any new relation created during processing will see the
+ * in-progress state and will automatically be checksummed as well as have its
+ * state recorded in the catalog to avoid the datachecksumsworker having to
+ * process it when already checksummed.
+ *
+ * For each database, all relations which have storage are read and every data
+ * page is marked dirty to force a write with the checksum. This will generate
+ * a lot of WAL as the entire database is read and written. Once all data pages
+ * in a relation have been written, pg_class.relhaschecksums is set to true to
+ * indicate that the relation is done.
+ *
+ * If the processing is interrupted by a cluster restart, it will be restarted
+ * from where it left off given that pg_class.relhaschecksums track state of
+ * processed relations and the in-progress state will ensure all new writes
+ * performed with checksums. Each database will be reprocessed, but relations
+ * where pg_class.relhaschecksums is true are skipped.
+ *
+ * If data checksums are enabled, then disabled, and then re-enabled, every
+ * relation's pg_class.relhaschecksums field will be reset to false before
+ * entering the in-progress mode.
+ *
+ *
+ * Disabling checksums
+ * -------------------
+ * When disabling checksums, data_checksums will be set to "inprogress-off"
+ * which signals that checksums are written but no longer verified. This ensure
+ * that backends which have yet to move from the "on" state will still be able
+ * to process data checksum validation. During "inprogress-off", the catalog
+ * state pg_class.relhaschecksums is cleared for all relations.
+ *
+ *
+ * Synchronization and Correctness
+ * -------------------------------
+ * The processes involved in enabling, or disabling, data checksums in an
+ * online cluster must be properly synchronized with the normal backends
+ * serving concurrent queries to ensure correctness. Correctness is defined
+ * as the following:
+ *
+ *    - Backends SHALL NOT violate local datachecksum state
+ *    - Data checksums SHALL NOT be considered enabled cluster-wide until all
+ *      currently connected backends have the local state "enabled"
+ *
+ * There are two levels of synchronization required for enabling data checksums
+ * in an online cluster: (i) changing state in the active backends ("on",
+ * "off", "inprogress-on" and "inprogress-off"), and (ii) ensuring no
+ * incompatible objects and processes are left in a database when workers end.
+ * The former deals with cluster-wide agreement on data checksum state and the
+ * latter with ensuring that any concurrent activity cannot break the data
+ * checksum contract during processing.
+ *
+ * Synchronizing the state change is done with procsignal barriers, where the
+ * WAL logging backend updating the global state in the controlfile will wait
+ * for all other backends to absorb the barrier. Barrier absorption will happen
+ * during interrupt processing, which means that connected backends will change
+ * state at different times. To prevent data checksum state changes when
+ * writing and verifying checksums, interrupts shall be held off before
+ * interrogating state and resumed when the IO operation has been performed.
+ *
+ *   When Enabling Data Checksums
+ *   ----------------------------
+ *   A process which fails to observe data checksums being enabled can induce
+ *   two types of errors: failing to write the checksum when modifying the page
+ *   and failing to validate the data checksum on the page when reading it.
+ *
+ *   When processing starts all backends belong to one of the below sets, with
+ *   one set being empty:
+ *
+ *   Bd: Backends in "off" state
+ *   Bi: Backends in "inprogress-on" state
+ *
+ *   If processing is started in an online cluster then all backends are in Bd.
+ *   If processing was halted by the cluster shutting down, the controlfile
+ *   state "inprogress-on" will be observed on system startup and all backends
+ *   will be in Bd. Backends transition Bd -> Bi via a procsignalbarrier.  When
+ *   the DataChecksumsWorker has finished writing checksums on all pages and
+ *   enables data checksums cluster-wide, there are four sets of backends where
+ *   Bd shall be an empty set:
+ *
+ *   Bg: Backend updating the global state and emitting the procsignalbarrier
+ *   Bd: Backends in "off" state
+ *   Be: Backends in "on" state
+ *   Bi: Backends in "inprogress-on" state
+ *
+ *   Backends in Bi and Be will write checksums when modifying a page, but only
+ *   backends in Be will verify the checksum during reading. The Bg backend is
+ *   blocked waiting for all backends in Bi to process interrupts and move to
+ *   Be. Any backend starting while Bg is waiting on the procsignalbarrier will
+ *   observe the global state being "on" and will thus automatically belong to
+ *   Be.  Checksums are enabled cluster-wide when Bi is an empty set. Bi and Be
+ *   are compatible sets while still operating based on their local state as
+ *   both write data checksums.
+ *
+ *   When Disabling Data Checksums
+ *   -----------------------------
+ *   A process which fails to observe that data checksums have been disabled
+ *   can induce two types of errors: writing the checksum when modifying the
+ *   page and validating a data checksum which is no longer correct due to
+ *   modifications to the page.
+ *
+ *   Bg: Backend updating the global state and emitting the procsignalbarrier
+ *   Bd: Backends in "off" state
+ *   Be: Backends in "on" state
+ *   Bo: Backends in "inprogress-off" state
+ *
+ *   Backends transition from the Be state to Bd like so: Be -> Bo -> Bd
+ *
+ *   The goal is to transition all backends to Bd making the others empty sets.
+ *   Backends in Bo write data checksums, but don't validate them, such that
+ *   backends still in Be can continue to validate pages until the barrier has
+ *   been absorbed such that they are in Bo. Once all backends are in Bo, the
+ *   barrier to transition to "off" can be raised and all backends can safely
+ *   stop writing data checksums as no backend is enforcing data checksum
+ *   validation any longer.
+ *
+ *
+ * Potential optimizations
+ * -----------------------
+ * Below are some potential optimizations and improvements which were brought
+ * up during reviews of this feature, but which weren't implemented in the
+ * initial version. These are ideas listed without any validation on their
+ * feasability or potential payoff. More discussion on these can be found on
+ * the -hackers threads linked to in the commit message of this feature.
+ *
+ *   * Launching datachecksumsworker for resuming operation from the startup
+ *     process: Currently users have to restart processing manually after a
+ *     restart since dynamic background worker cannot be started from the
+ *     postmaster. Changing to the startup process could make resuming the
+ *     processing automatic.
+ *   * Avoid dirtying the page when checksums already match: Iff the checksum
+ *     on the page happens to already match we still dirty the page. It should
+ *     be enough to only do the log_newpage_buffer() call in that case.
+ *   * Invent a lightweight WAL record that doesn't contain the full-page
+ *     image but just the block number: On replay, the redo routine would read
+ *     the page from disk.
+ *   * Teach pg_checksums to avoid checksummed pages when pg_checksums is used
+ *     to enable checksums on a cluster which is in inprogress-on state and
+ *     may have checksummed pages (make pg_checksums be able to resume an
+ *     online operation).
+ *
+ *
+ * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ *
+ * IDENTIFICATION
+ *	  src/backend/postmaster/datachecksumsworker.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include "access/genam.h"
+#include "access/heapam.h"
+#include "access/htup_details.h"
+#include "access/xact.h"
+#include "catalog/indexing.h"
+#include "catalog/pg_class.h"
+#include "catalog/pg_database.h"
+#include "commands/vacuum.h"
+#include "common/relpath.h"
+#include "miscadmin.h"
+#include "pgstat.h"
+#include "postmaster/bgworker.h"
+#include "postmaster/bgwriter.h"
+#include "postmaster/datachecksumsworker.h"
+#include "storage/bufmgr.h"
+#include "storage/checksum.h"
+#include "storage/lmgr.h"
+#include "storage/ipc.h"
+#include "storage/procarray.h"
+#include "storage/smgr.h"
+#include "tcop/tcopprot.h"
+#include "utils/fmgroids.h"
+#include "utils/lsyscache.h"
+#include "utils/ps_status.h"
+#include "utils/syscache.h"
+
+#define DATACHECKSUMSWORKER_MAX_DB_RETRIES 5
+
+#define MAX_OPS 4
+
+typedef enum DataChecksumOperation
+{
+	ENABLE_CHECKSUMS = 1,
+	DISABLE_CHECKSUMS,
+	RESET_STATE,
+	SET_INPROGRESS_ON,
+	SET_CHECKSUMS_ON
+}			DataChecksumOperation;
+
+typedef enum
+{
+	DATACHECKSUMSWORKER_SUCCESSFUL = 0,
+	DATACHECKSUMSWORKER_ABORTED,
+	DATACHECKSUMSWORKER_FAILED,
+	DATACHECKSUMSWORKER_RETRYDB,
+}			DatachecksumsWorkerResult;
+
+/*
+ * Signaling between backends calling pg_enable/disable_checkums, the
+ * checksums launcher process, and the checksums worker process.
+ *
+ * This struct is protected by DatachecksumsWorkerLock
+ */
+typedef struct DatachecksumsWorkerShmemStruct
+{
+	/*
+	 * These are set by pg_enable/disable_checkums, to tell the launcher what
+	 * the target state is.
+	 */
+	bool		launch_enable_checksums;	/* True if checksums are being
+											 * enabled, else false */
+	int			launch_cost_delay;
+	int			launch_cost_limit;
+
+	/*
+	 * Is a launcher process is currently running?
+	 *
+	 * This is set by the launcher process, after it has read the above launch_*
+	 * parameters.
+	 */
+	bool		launcher_running;
+
+	/*
+	 * These fields indicate the target state that the launcher is currently
+	 * working towards. They can be different from the corresponding launch_*
+	 * fields, if a new pg_enable_disable_checksums() call was made while the
+	 * launcher/worker was already running.
+
+	 * The below members are set when the launcher starts, and are only
+	 * accessed read-only by the single worker. Thus, we can access these
+	 * without a lock. If multiple workers, or dynamic cost parameters, are
+	 * supported at some point then this would need to be revisited.
+	 */
+	bool		enabling_checksums;	/* True if checksums are being enabled,
+									 * else false */
+	int			cost_delay;
+	int			cost_limit;
+
+	/*
+	 * Signaling between the launcher and the worker process.
+	 *
+	 * As there is only a single worker, and the launcher
+	 * won't read these until the worker exits, they can be accessed without
+	 * the need for a lock. If multiple workers are supported then this will
+	 * have to be revisited.
+	 */
+	/* result, set by worker before exiting */
+	DatachecksumsWorkerResult success;
+
+	/* tells the worker process whether it should also process the shared catalogs. */
+	bool		process_shared_catalogs;
+} DatachecksumsWorkerShmemStruct;
+
+/* Shared memory segment for datachecksumsworker */
+static DatachecksumsWorkerShmemStruct *DatachecksumsWorkerShmem;
+
+/* Bookkeeping for work to do */
+typedef struct DatachecksumsWorkerDatabase
+{
+	Oid			dboid;
+	char	   *dbname;
+}			DatachecksumsWorkerDatabase;
+
+typedef struct DatachecksumsWorkerResultEntry
+{
+	Oid			dboid;
+	DatachecksumsWorkerResult result;
+	int			retries;
+}			DatachecksumsWorkerResultEntry;
+
+
+/*
+ * Flag set by the interrupt handler
+ */
+static volatile sig_atomic_t abort_requested = false;
+
+/*
+ * Have we set the DatachecksumsWorkerShmemStruct->launcher_running flag?
+ * If we have, we need to clear it before exiting!
+ */
+static volatile sig_atomic_t launcher_running = false;
+
+/*
+ * Are we enabling checkums, or disabling them?
+ */
+static bool enabling_checksums;
+
+/* Prototypes */
+static List *BuildDatabaseList(void);
+static List *BuildRelationList(bool temp_relations, bool include_shared);
+static DatachecksumsWorkerResult ProcessDatabase(DatachecksumsWorkerDatabase *db, const char *bgw_func_name);
+static bool ProcessAllDatabases(bool *already_connected, const char *bgw_func_name);
+static bool ProcessSingleRelationFork(Relation reln, ForkNumber forkNum, BufferAccessStrategy strategy);
+static void launcher_cancel_handler(SIGNAL_ARGS);
+static void SetRelHasChecksums(Oid relOid);
+static void WaitForAllTransactionsToFinish(void);
+
+/*
+ * StartDataChecksumsWorkerLauncher
+ *		Main entry point for datachecksumsworker launcher process
+ *
+ * The main entrypoint for starting data checksums processing for enabling as
+ * well as disabling.
+ */
+void
+StartDatachecksumsWorkerLauncher(bool enable_checksums, int cost_delay, int cost_limit)
+{
+	BackgroundWorker bgw;
+	BackgroundWorkerHandle *bgw_handle;
+	bool		launcher_running;
+
+	/* the cost delay settings have no effect when disabling */
+	Assert(enable_checksums || cost_delay == 0);
+	Assert(enable_checksums || cost_limit == 0);
+
+	/*
+	 * Store the desired state in shared memory.
+	 */
+	LWLockAcquire(DatachecksumsWorkerLock, LW_EXCLUSIVE);
+
+	DatachecksumsWorkerShmem->launch_enable_checksums = enable_checksums;
+	DatachecksumsWorkerShmem->launch_cost_delay = cost_delay;
+	DatachecksumsWorkerShmem->launch_cost_limit = cost_limit;
+
+	/* is the launcher already running? */
+	launcher_running = DatachecksumsWorkerShmem->launcher_running;
+
+	LWLockRelease(DatachecksumsWorkerLock);
+
+	/*
+	 * Launch a new launcher process, if it's not running already.
+	 *
+	 * If the launcher is currently busy enabling the checkums, and we want
+	 * them disabled (or vice versa), the launcher will notice that at latest
+	 * when it's about to exit, and will loop back process the new request.
+	 * So if the launcher is already running, we don't need to do anything
+	 * more here to abort it.
+	 *
+	 * If you call pg_enable/disable_checksums() twice in a row, before the
+	 * launcher has had a chance to start up, we still end up launching it
+	 * twice.  That's OK, the second invocation will see that a launcher is
+	 * already running and exit quickly.
+	 *
+	 * TODO: We could optimize here and skip launching the launcher, if we are
+	 * already in the desired state, i.e. if the checksums are already enabled
+	 * and you call pg_enable_checksums().
+	 */
+	if (!launcher_running)
+	{
+		/*
+		 * Prepare the BackgroundWorker and launch it.
+		 */
+		memset(&bgw, 0, sizeof(bgw));
+		bgw.bgw_flags = BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
+		bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
+		snprintf(bgw.bgw_library_name, BGW_MAXLEN, "postgres");
+		snprintf(bgw.bgw_function_name, BGW_MAXLEN, "DatachecksumsWorkerLauncherMain");
+		snprintf(bgw.bgw_name, BGW_MAXLEN, "datachecksumsworker launcher");
+		snprintf(bgw.bgw_type, BGW_MAXLEN, "datachecksumsworker launcher");
+		bgw.bgw_restart_time = BGW_NEVER_RESTART;
+		bgw.bgw_notify_pid = MyProcPid;
+		bgw.bgw_main_arg = (Datum) 0;
+
+		if (!RegisterDynamicBackgroundWorker(&bgw, &bgw_handle))
+			ereport(ERROR,
+					(errmsg("failed to start background worker to process data checksums")));
+	}
+}
+
+/*
+ * ProcessSingleRelationFork
+ *		Enable data checksums in a single relation/fork.
+ *
+ * Returns true if successful, and false if *aborted*. On error, an actual
+ * error is raised in the lower levels.
+ */
+static bool
+ProcessSingleRelationFork(Relation reln, ForkNumber forkNum, BufferAccessStrategy strategy)
+{
+	BlockNumber numblocks = RelationGetNumberOfBlocksInFork(reln, forkNum);
+	BlockNumber blknum;
+	char		activity[NAMEDATALEN * 2 + 128];
+	char	   *relns;
+
+	relns = get_namespace_name(RelationGetNamespace(reln));
+
+	if (!relns)
+		return false;
+
+	/*
+	 * We are looping over the blocks which existed at the time of process
+	 * start, which is safe since new blocks are created with checksums set
+	 * already due to the state being "inprogress-on".
+	 */
+	for (blknum = 0; blknum < numblocks; blknum++)
+	{
+		Buffer		buf = ReadBufferExtended(reln, forkNum, blknum, RBM_NORMAL, strategy);
+
+		/*
+		 * Report to pgstat every 100 blocks to keep from overwhelming the
+		 * activity reporting with close to identical reports.
+		 */
+		if ((blknum % 100) == 0)
+		{
+			snprintf(activity, sizeof(activity) - 1, "processing: %s.%s (%s block %d/%d)",
+					 relns, RelationGetRelationName(reln),
+					 forkNames[forkNum], blknum, numblocks);
+			pgstat_report_activity(STATE_RUNNING, activity);
+		}
+
+		/* Need to get an exclusive lock before we can flag as dirty */
+		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+
+		/*
+		 * Mark the buffer as dirty and force a full page write.  We have to
+		 * re-write the page to WAL even if the checksum hasn't changed,
+		 * because if there is a replica it might have a slightly different
+		 * version of the page with an invalid checksum, caused by unlogged
+		 * changes (e.g. hintbits) on the master happening while checksums
+		 * were off. This can happen if there was a valid checksum on the page
+		 * at one point in the past, so only when checksums are first on, then
+		 * off, and then turned on again. Iff wal_level is set to "minimal",
+		 * this could be avoided iff the checksum is calculated to be correct.
+		 */
+		START_CRIT_SECTION();
+		MarkBufferDirty(buf);
+		log_newpage_buffer(buf, false);
+		END_CRIT_SECTION();
+
+		UnlockReleaseBuffer(buf);
+
+		/*
+		 * This is the only place where we check if we are asked to abort, the
+		 * abortion will bubble up from here. It's safe to check this without
+		 * a lock, because if we miss it being set, we will try again soon.
+		 */
+		Assert(enabling_checksums);
+		if (!DatachecksumsWorkerShmem->launch_enable_checksums)
+			abort_requested = true;
+		if (abort_requested)
+			return false;
+
+		vacuum_delay_point();
+	}
+
+	pfree(relns);
+	return true;
+}
+
+/*
+ * ProcessSingleRelationByOid
+ *		Process a single relation based on oid.
+ *
+ * Returns true if successful, and false if *aborted*. On error, an actual
+ * error is raised in the lower levels.
+ */
+static bool
+ProcessSingleRelationByOid(Oid relationId, BufferAccessStrategy strategy)
+{
+	Relation	rel;
+	ForkNumber	fnum;
+	bool		aborted = false;
+
+	StartTransactionCommand();
+
+	elog(DEBUG2,
+		 "adding data checksums to relation with OID %u",
+		 relationId);
+
+	rel = try_relation_open(relationId, AccessShareLock);
+	if (rel == NULL)
+	{
+		/*
+		 * Relation no longer exists. We don't consider this an error since
+		 * there are no pages in it that need data checksums, and thus return
+		 * true. The worker operates off a list of relations generated at the
+		 * start of processing, so relations being dropped in the meantime is
+		 * to be expected.
+		 */
+		CommitTransactionCommand();
+		pgstat_report_activity(STATE_IDLE, NULL);
+		return true;
+	}
+	RelationOpenSmgr(rel);
+
+	for (fnum = 0; fnum <= MAX_FORKNUM; fnum++)
+	{
+		if (smgrexists(rel->rd_smgr, fnum))
+		{
+			if (!ProcessSingleRelationFork(rel, fnum, strategy))
+			{
+				aborted = true;
+				break;
+			}
+		}
+	}
+	relation_close(rel, AccessShareLock);
+	elog(DEBUG2,
+		 "data checksum processing done for relation with OID %u: %s",
+		 relationId, (aborted ? "aborted" : "finished"));
+
+	if (!aborted)
+		SetRelHasChecksums(relationId);
+
+	CommitTransactionCommand();
+
+	pgstat_report_activity(STATE_IDLE, NULL);
+
+	return !aborted;
+}
+
+/*
+ * SetRelHasChecksums
+ *
+ * Sets the pg_class.relhaschecksums flag for the relation specified by relOid
+ * to true. The corresponding function for clearing state is
+ * ResetDataChecksumsStateInDatabase which operate on all relations in a
+ * database.
+ */
+static void
+SetRelHasChecksums(Oid relOid)
+{
+	Relation	rel;
+	Relation	heaprel;
+	Form_pg_class pg_class_tuple;
+	HeapTuple	tuple;
+
+	/*
+	 * If the relation has gone away since we checksummed it then that's not
+	 * an errorcase. Exit early and continue on the next relation instead.
+	 */
+	heaprel = try_relation_open(relOid, ShareUpdateExclusiveLock);
+	if (!heaprel)
+		return;
+	rel = table_open(RelationRelationId, RowExclusiveLock);
+
+	tuple = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relOid));
+	if (!HeapTupleIsValid(tuple))
+		elog(ERROR, "cache lookup failed for relation %u", relOid);
+
+	pg_class_tuple = (Form_pg_class) GETSTRUCT(tuple);
+	pg_class_tuple->relhaschecksums = true;
+
+	CatalogTupleUpdate(rel, &tuple->t_self, tuple);
+
+	heap_freetuple(tuple);
+
+	table_close(rel, RowExclusiveLock);
+	relation_close(heaprel, ShareUpdateExclusiveLock);
+}
+
+/*
+ * ProcessDatabase
+ *		Enable data checksums in a single database.
+ *
+ * We do this by launching a dynamic background worker into this database, and
+ * waiting for it to finish.  We have to do this in a separate worker, since
+ * each process can only be connected to one database during its lifetime.
+ */
+static DatachecksumsWorkerResult
+ProcessDatabase(DatachecksumsWorkerDatabase *db, const char *bgw_func_name)
+{
+	BackgroundWorker bgw;
+	BackgroundWorkerHandle *bgw_handle;
+	BgwHandleStatus status;
+	pid_t		pid;
+	char		activity[NAMEDATALEN + 64];
+
+	DatachecksumsWorkerShmem->success = DATACHECKSUMSWORKER_FAILED;
+
+	memset(&bgw, 0, sizeof(bgw));
+	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
+	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
+	snprintf(bgw.bgw_library_name, BGW_MAXLEN, "postgres");
+	snprintf(bgw.bgw_function_name, BGW_MAXLEN, "%s", bgw_func_name);
+	snprintf(bgw.bgw_name, BGW_MAXLEN, "datachecksumsworker worker");
+	snprintf(bgw.bgw_type, BGW_MAXLEN, "datachecksumsworker worker");
+	bgw.bgw_restart_time = BGW_NEVER_RESTART;
+	bgw.bgw_notify_pid = MyProcPid;
+	bgw.bgw_main_arg = ObjectIdGetDatum(db->dboid);
+
+	/*
+	 * If there are no worker slots available, make sure we retry processing
+	 * this database. This will make the datachecksumsworker move on to the
+	 * next database and quite likely fail with the same problem. TODO: Maybe
+	 * we need a backoff to avoid running through all the databases here in
+	 * short order.
+	 */
+	if (!RegisterDynamicBackgroundWorker(&bgw, &bgw_handle))
+	{
+		ereport(WARNING,
+				(errmsg("failed to start worker for enabling data checksums in database \"%s\", retrying",
+						db->dbname),
+				 errhint("The max_worker_processes setting might be too low.")));
+		return DATACHECKSUMSWORKER_RETRYDB;
+	}
+
+	status = WaitForBackgroundWorkerStartup(bgw_handle, &pid);
+	if (status == BGWH_STOPPED)
+	{
+		ereport(WARNING,
+				(errmsg("could not start background worker for enabling data checksums in database \"%s\"",
+						db->dbname),
+				 errhint("More details on the error might be found in the server log.")));
+		return DATACHECKSUMSWORKER_FAILED;
+	}
+
+	/*
+	 * If the postmaster crashed we cannot end up with a processed database so
+	 * we have no alternative other than exiting. When enabling checksums we
+	 * won't at this time have changed the pg_control version to enabled so
+	 * when the cluster comes back up processing will have to be resumed. When
+	 * disabling, the pg_control version will be set to off before this so
+	 * when the cluster comes up checksums will be off as expected. In the
+	 * latter case we might have stale relhaschecksums flags in pg_class which
+	 * it would be nice to handle in some way. Enabling data checksums reset
+	 * the flags so any stale flags won't cause problems at that point, but
+	 * they may cause confusion with users reading pg_class. TODO.
+	 */
+	if (status == BGWH_POSTMASTER_DIED)
+		ereport(FATAL,
+				(errmsg("cannot enable data checksums without the postmaster process"),
+				 errhint("Restart the database and restart data checksum processing by calling pg_enable_data_checksums().")));
+
+	Assert(status == BGWH_STARTED);
+	ereport(DEBUG1,
+			(errmsg("initiating data checksum processing in database \"%s\"",
+					db->dbname)));
+
+	snprintf(activity, sizeof(activity) - 1,
+			 "Waiting for worker in database %s (pid %d)", db->dbname, pid);
+	pgstat_report_activity(STATE_RUNNING, activity);
+
+	status = WaitForBackgroundWorkerShutdown(bgw_handle);
+	if (status == BGWH_POSTMASTER_DIED)
+		ereport(FATAL,
+				(errmsg("postmaster exited during data checksum processing in \"%s\"",
+						db->dbname),
+				 errhint("Restart the database and restart data checksum processing by calling pg_enable_data_checksums().")));
+
+	if (DatachecksumsWorkerShmem->success == DATACHECKSUMSWORKER_ABORTED)
+		ereport(LOG,
+				(errmsg("data checksums processing was aborted in database \"%s\"",
+						db->dbname)));
+
+	pgstat_report_activity(STATE_IDLE, NULL);
+
+	return DatachecksumsWorkerShmem->success;
+}
+
+/*
+ * launcher_exit
+ *
+ * Internal routine for cleaning up state when the launcher process exits. We
+ * need to clean up the abort flag to ensure that processing can be restarted
+ * again after it was previously aborted.
+ */
+static void
+launcher_exit(int code, Datum arg)
+{
+	if (launcher_running)
+	{
+		LWLockAcquire(DatachecksumsWorkerLock, LW_EXCLUSIVE);
+		launcher_running = false;
+		DatachecksumsWorkerShmem->launcher_running = false;
+		LWLockRelease(DatachecksumsWorkerLock);
+	}
+}
+
+/*
+ * launcher_cancel_handler
+ *
+ * Internal routine for reacting to SIGINT and flagging the worker to abort.
+ * The worker won't be interrupted immediately but will check for abort flag
+ * between each block in a relation.
+ */
+static void
+launcher_cancel_handler(SIGNAL_ARGS)
+{
+	int			save_errno = errno;
+
+	abort_requested = true;
+
+	/*
+	 * There is no sleeping in the main loop, the flag will be checked
+	 * periodically in ProcessSingleRelationFork. The worker does however
+	 * sleep when waiting for concurrent transactions to end so we still need
+	 * to set the latch.
+	 */
+	SetLatch(MyLatch);
+
+	errno = save_errno;
+}
+
+/*
+ * WaitForAllTransactionsToFinish
+ *		Blocks awaiting all current transactions to finish
+ *
+ * Returns when all transactions which are active at the call of the function
+ * have ended, or if the postmaster dies while waiting. If the postmaster dies
+ * the abort flag will be set to indicate that the caller of this shouldn't
+ * proceed.
+ *
+ * NB: this will return early, if aborted by SIGINT or if the target state
+ * is changed while we're running.
+ */
+static void
+WaitForAllTransactionsToFinish(void)
+{
+	TransactionId waitforxid;
+
+	LWLockAcquire(XidGenLock, LW_SHARED);
+	waitforxid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
+	LWLockRelease(XidGenLock);
+
+	while (TransactionIdPrecedes(GetOldestActiveTransactionId(), waitforxid))
+	{
+		char		activity[64];
+		int			rc;
+
+		/* Oldest running xid is older than us, so wait */
+		snprintf(activity,
+				 sizeof(activity),
+				 "Waiting for current transactions to finish (waiting for %u)",
+				 waitforxid);
+		pgstat_report_activity(STATE_RUNNING, activity);
+
+		/* Retry every 5 seconds */
+		ResetLatch(MyLatch);
+		rc = WaitLatch(MyLatch,
+					   WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
+					   5000,
+					   WAIT_EVENT_CHECKSUM_ENABLE_STARTCONDITION);
+
+		/*
+		 * If the postmaster died we won't be able to enable checksums
+		 * cluster-wide so abort and hope to continue when restarted.
+		 */
+		if (rc & WL_POSTMASTER_DEATH)
+			ereport(FATAL,
+					(errmsg("postmaster exited during data checksum processing"),
+					 errhint("Restart the database and restart data checksum processing by calling pg_enable_data_checksums().")));
+
+		LWLockAcquire(DatachecksumsWorkerLock, LW_SHARED);
+		if (DatachecksumsWorkerShmem->launch_enable_checksums != enabling_checksums)
+			abort_requested = true;
+		LWLockRelease(DatachecksumsWorkerLock);
+		if (abort_requested)
+			break;
+	}
+
+	pgstat_report_activity(STATE_IDLE, NULL);
+	return;
+}
+
+/*
+ * DatachecksumsWorkerLauncherMain
+ *
+ * Main function for launching dynamic background workers for processing data
+ * checksums in databases. This function has the bgworker management, with
+ * ProcessAllDatabases being responsible for looping over the databases and
+ * initiating processing.
+ */
+void
+DatachecksumsWorkerLauncherMain(Datum arg)
+{
+	bool		connected = false;
+	bool		status = false;
+	DataChecksumOperation current;
+	int			operations[MAX_OPS];
+
+	on_shmem_exit(launcher_exit, 0);
+
+	ereport(DEBUG1,
+			(errmsg("background worker \"datachecksumsworker\" launcher started")));
+
+	pqsignal(SIGTERM, die);
+	pqsignal(SIGINT, launcher_cancel_handler);
+
+	BackgroundWorkerUnblockSignals();
+
+	InitXLOGAccess();
+
+	MyBackendType = B_DATACHECKSUMSWORKER_LAUNCHER;
+	init_ps_display(NULL);
+
+	LWLockAcquire(DatachecksumsWorkerLock, LW_EXCLUSIVE);
+
+	if (DatachecksumsWorkerShmem->launcher_running)
+	{
+		/* Launcher was already running. Let it finish. */
+		LWLockRelease(DatachecksumsWorkerLock);
+		return;
+	}
+
+	launcher_running = true;
+
+	enabling_checksums = DatachecksumsWorkerShmem->launch_enable_checksums;
+	DatachecksumsWorkerShmem->launcher_running = true;
+	DatachecksumsWorkerShmem->enabling_checksums = enabling_checksums;
+	DatachecksumsWorkerShmem->cost_delay = DatachecksumsWorkerShmem->launch_cost_delay;
+	DatachecksumsWorkerShmem->cost_limit = DatachecksumsWorkerShmem->launch_cost_limit;
+	LWLockRelease(DatachecksumsWorkerLock);
+
+	/*
+	 * The target state can change while we are busy enabling/disabling checksums,
+	 * if the user calls pg_disable/enable_checksums() before we are finished with
+	 * the previous request. In that case, we will loop back here, to process the
+	 * new request.
+	 */
+again:
+
+	memset(operations, 0, sizeof(operations));
+
+	/*
+	 * If we're asked to enable checksums, we need to check if processing was
+	 * previously interrupted such that we should resume rather than start
+	 * from scratch.
+	 */
+	if (enabling_checksums)
+	{
+		/*
+		 * If we are asked to enable checksums in a cluster which already
+		 * has checksums enabled, exit immediately as there is nothing
+		 * more to do.
+		 */
+		if (DataChecksumsNeedVerify())
+			goto done;
+
+		/*
+		 * If the controlfile state is set to "inprogress-on" then we will
+		 * resume from where we left off based on the catalog state. This
+		 * will be safe since new relations created while the checksum-
+		 * worker was disabled will have checksums enabled.
+		 */
+		else if (DataChecksumsOnInProgress())
+		{
+			operations[0] = ENABLE_CHECKSUMS;
+			operations[1] = SET_CHECKSUMS_ON;
+		}
+
+		/*
+		 * If the controlfile state is set to "inprogress-off" then we
+		 * were interrupted while the catalog state was being cleared. In
+		 * this case we need to first reset state and then continue with
+		 * enabling checksums.
+		 */
+		else if (DataChecksumsOffInProgress())
+		{
+			operations[0] = RESET_STATE;
+			operations[1] = SET_INPROGRESS_ON;
+			operations[2] = ENABLE_CHECKSUMS;
+			operations[3] = SET_CHECKSUMS_ON;
+		}
+
+		/*
+		 * Data checksums are off in the cluster, we can proceed with
+		 * enabling them. Just in case we will start by resetting the
+		 * catalog state since we are doing this from scratch and we don't
+		 * want leftover catalog state to cause us to miss a relation.
+		 */
+		else
+		{
+			operations[0] = RESET_STATE;
+			operations[1] = SET_INPROGRESS_ON;
+			operations[2] = ENABLE_CHECKSUMS;
+			operations[3] = SET_CHECKSUMS_ON;
+		}
+	}
+	else
+	{
+		/*
+		 * Regardless of current state in the system, we go through the
+		 * motions when asked to disable checksums. The catalog state is
+		 * only defined to be relevant during the operation of enabling
+		 * checksums, and have no use at any other point in time. That
+		 * being said, a user who sees stale relhaschecksums entries in
+		 * the catalog might run this just in case.
+		 *
+		 * Resetting state must be performed after setting data checksum
+		 * state to off, as there otherwise might (depending on system
+		 * data checksum state) be a window between catalog resetting and
+		 * state transition when new relations are created with the
+		 * catalog state set to true.
+		 */
+		operations[0] = DISABLE_CHECKSUMS;
+		operations[1] = RESET_STATE;
+	}
+
+	for (int i = 0; i < MAX_OPS; i++)
+	{
+		current = operations[i];
+
+		if (!current)
+			break;
+
+		switch (current)
+		{
+			case DISABLE_CHECKSUMS:
+				SetDataChecksumsOff();
+				break;
+
+			case SET_INPROGRESS_ON:
+				SetDataChecksumsOnInProgress();
+				break;
+
+			case SET_CHECKSUMS_ON:
+				SetDataChecksumsOn();
+				break;
+
+			case RESET_STATE:
+				status = ProcessAllDatabases(&connected, "ResetDataChecksumsStateInDatabase");
+				if (!status)
+					ereport(ERROR,
+							(errmsg("unable to reset catalog checksum state")));
+				break;
+
+			case ENABLE_CHECKSUMS:
+				status = ProcessAllDatabases(&connected, "DatachecksumsWorkerMain");
+				if (!status)
+					ereport(ERROR,
+							(errmsg("unable to enable checksums in cluster")));
+				break;
+
+			default:
+				elog(ERROR, "unknown checksum operation requested");
+				break;
+		}
+	}
+
+done:
+	/*
+	 * All done. But before we exit, check if the target state was changed while
+	 * we were running. In that case we will have to start all over again.
+	 */
+	LWLockAcquire(DatachecksumsWorkerLock, LW_EXCLUSIVE);
+	if (DatachecksumsWorkerShmem->launch_enable_checksums != enabling_checksums)
+	{
+		DatachecksumsWorkerShmem->enabling_checksums = DatachecksumsWorkerShmem->launch_enable_checksums;
+		DatachecksumsWorkerShmem->cost_delay = DatachecksumsWorkerShmem->launch_cost_delay;
+		DatachecksumsWorkerShmem->cost_limit = DatachecksumsWorkerShmem->launch_cost_limit;
+		LWLockRelease(DatachecksumsWorkerLock);
+		goto again;
+	}
+
+	launcher_running = false;
+	DatachecksumsWorkerShmem->launcher_running = false;
+	LWLockRelease(DatachecksumsWorkerLock);
+}
+
+/*
+ * ProcessAllDatabases
+ *		Compute the list of all databases and process checksums in each
+ *
+ * This will repeatedly generate a list of databases to process for either
+ * enabling checksums or resetting the checksum catalog tracking. Until no
+ * new databases are found, this will loop around computing a new list and
+ * comparing it to the already seen ones.
+ */
+static bool
+ProcessAllDatabases(bool *already_connected, const char *bgw_func_name)
+{
+	List	   *DatabaseList;
+	HTAB	   *ProcessedDatabases = NULL;
+	ListCell   *lc;
+	HASHCTL		hash_ctl;
+	bool		found_failed = false;
+
+	/* Initialize a hash tracking all processed databases */
+	memset(&hash_ctl, 0, sizeof(hash_ctl));
+	hash_ctl.keysize = sizeof(Oid);
+	hash_ctl.entrysize = sizeof(DatachecksumsWorkerResultEntry);
+	ProcessedDatabases = hash_create("Processed databases",
+									 64,
+									 &hash_ctl,
+									 HASH_ELEM | HASH_BLOBS);
+
+	/*
+	 * Initialize a connection to shared catalogs only.
+	 */
+	if (!*already_connected)
+		BackgroundWorkerInitializeConnection(NULL, NULL, 0);
+
+	*already_connected = true;
+
+	/*
+	 * Set up so first run processes shared catalogs, but not once in every
+	 * db.
+	 */
+	DatachecksumsWorkerShmem->process_shared_catalogs = true;
+
+	/*
+	 * Get a list of all databases to process. This may include databases that
+	 * were created during our runtime.  Since a database can be created as a
+	 * copy of any other database (which may not have existed in our last
+	 * run), we have to repeat this loop until no new databases show up in the
+	 * list.
+	 */
+	DatabaseList = BuildDatabaseList();
+
+	while (true)
+	{
+		int			processed_databases = 0;
+
+		foreach(lc, DatabaseList)
+		{
+			DatachecksumsWorkerDatabase *db = (DatachecksumsWorkerDatabase *) lfirst(lc);
+			DatachecksumsWorkerResult result;
+			DatachecksumsWorkerResultEntry *entry;
+			bool		found;
+
+			elog(DEBUG1,
+				 "starting processing of database %s with oid %u",
+				 db->dbname, db->dboid);
+
+			entry = (DatachecksumsWorkerResultEntry *) hash_search(ProcessedDatabases, &db->dboid,
+																   HASH_FIND, NULL);
+
+			if (entry)
+			{
+				if (entry->result == DATACHECKSUMSWORKER_RETRYDB)
+				{
+					/*
+					 * Limit the number of retries to avoid infinite looping
+					 * in case there simply wont be enough workers in the
+					 * cluster to finish this operation.
+					 */
+					if (entry->retries > DATACHECKSUMSWORKER_MAX_DB_RETRIES)
+						entry->result = DATACHECKSUMSWORKER_FAILED;
+				}
+
+				/* Skip if this database has been processed already */
+				if (entry->result != DATACHECKSUMSWORKER_RETRYDB)
+				{
+					pfree(db->dbname);
+					pfree(db);
+					continue;
+				}
+			}
+
+			result = ProcessDatabase(db, bgw_func_name);
+			processed_databases++;
+
+			if (result == DATACHECKSUMSWORKER_SUCCESSFUL)
+			{
+				/*
+				 * If one database has completed shared catalogs, we don't
+				 * have to process them again.
+				 */
+				if (DatachecksumsWorkerShmem->process_shared_catalogs)
+					DatachecksumsWorkerShmem->process_shared_catalogs = false;
+			}
+			else if (result == DATACHECKSUMSWORKER_ABORTED)
+			{
+				/* Abort flag set, so exit the whole process */
+				return false;
+			}
+
+			entry = hash_search(ProcessedDatabases, &db->dboid, HASH_ENTER, &found);
+			entry->dboid = db->dboid;
+			entry->result = result;
+			if (!found)
+				entry->retries = 0;
+			else
+				entry->retries++;
+
+			pfree(db->dbname);
+			pfree(db);
+		}
+
+		elog(DEBUG1,
+			 "%i databases processed for data checksum enabling, %s",
+			 processed_databases,
+			 (processed_databases ? "process with restart" : "process completed"));
+
+		list_free(DatabaseList);
+
+		/*
+		 * If no databases were processed in this run of the loop, we have now
+		 * finished all databases and no concurrently created ones can exist.
+		 */
+		if (processed_databases == 0)
+			break;
+
+		/*
+		 * Re-generate the list of databases for another pass. Since we wait
+		 * for all pre-existing transactions finish, this way we can be
+		 * certain that there are no databases left without checksums.
+		 */
+		WaitForAllTransactionsToFinish();
+		DatabaseList = BuildDatabaseList();
+	}
+
+	/*
+	 * ProcessedDatabases now has all databases and the results of their
+	 * processing. Failure to enable checksums for a database can be because
+	 * they actually failed for some reason, or because the database was
+	 * dropped between us getting the database list and trying to process it.
+	 * Get a fresh list of databases to detect the second case where the
+	 * database was dropped before we had started processing it. If a database
+	 * still exists, but enabling checksums failed then we fail the entire
+	 * checksumming process and exit with an error.
+	 */
+	WaitForAllTransactionsToFinish();
+	DatabaseList = BuildDatabaseList();
+
+	foreach(lc, DatabaseList)
+	{
+		DatachecksumsWorkerDatabase *db = (DatachecksumsWorkerDatabase *) lfirst(lc);
+		DatachecksumsWorkerResult *entry;
+		bool		found;
+
+		entry = hash_search(ProcessedDatabases, (void *) &db->dboid,
+							HASH_FIND, &found);
+
+		/*
+		 * We are only interested in the databases where the failed database
+		 * still exists.
+		 */
+		if (found && *entry == DATACHECKSUMSWORKER_FAILED)
+		{
+			ereport(WARNING,
+					(errmsg("failed to enable data checksums in \"%s\"",
+							db->dbname)));
+			found_failed = found;
+			continue;
+		}
+	}
+
+	if (found_failed)
+	{
+		/* Disable checksums on cluster, because we failed */
+		SetDataChecksumsOff();
+		ereport(ERROR,
+				(errmsg("checksums failed to get enabled in all databases, aborting"),
+				 errhint("The server log might have more information on the error.")));
+	}
+
+	/*
+	 * Force a checkpoint to get everything out to disk. TODO: we probably
+	 * don't want to use a CHECKPOINT_IMMEDIATE here but it's very convenient
+	 * for testing until the patch is fully baked, as it may otherwise make
+	 * tests take a lot longer.
+	 */
+	RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT | CHECKPOINT_IMMEDIATE);
+
+	return true;
+}
+
+/*
+ * DatachecksumsWorkerShmemSize
+ *		Compute required space for datachecksumsworker-related shared memory
+ */
+Size
+DatachecksumsWorkerShmemSize(void)
+{
+	Size		size;
+
+	size = sizeof(DatachecksumsWorkerShmemStruct);
+	size = MAXALIGN(size);
+
+	return size;
+}
+
+/*
+ * DatachecksumsWorkerShmemInit
+ *		Allocate and initialize datachecksumsworker-related shared memory
+ */
+void
+DatachecksumsWorkerShmemInit(void)
+{
+	bool		found;
+
+	DatachecksumsWorkerShmem = (DatachecksumsWorkerShmemStruct *)
+		ShmemInitStruct("DatachecksumsWorker Data",
+						DatachecksumsWorkerShmemSize(),
+						&found);
+
+	MemSet(DatachecksumsWorkerShmem, 0, DatachecksumsWorkerShmemSize());
+
+	/*
+	 * Even if this is a redundant assignment, we want to be explicit about
+	 * our intent for readability, since we want to be able to query this
+	 * state in case of restartability.
+	 */
+	DatachecksumsWorkerShmem->launch_enable_checksums = false;
+	DatachecksumsWorkerShmem->launcher_running = false;
+}
+
+/*
+ * BuildDatabaseList
+ *		Compile a list of all currently available databases in the cluster
+ *
+ * This creates the list of databases for the datachecksumsworker workers to
+ * add checksums to. If the caller wants to ensure that no concurrently
+ * running CREATE DATABASE calls exist, this needs to be preceeded by a call
+ * to WaitForAllTransactionsToFinish().
+ */
+static List *
+BuildDatabaseList(void)
+{
+	List	   *DatabaseList = NIL;
+	Relation	rel;
+	TableScanDesc scan;
+	HeapTuple	tup;
+	MemoryContext ctx = CurrentMemoryContext;
+	MemoryContext oldctx;
+
+	StartTransactionCommand();
+
+	rel = table_open(DatabaseRelationId, AccessShareLock);
+	scan = table_beginscan_catalog(rel, 0, NULL);
+
+	while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
+	{
+		Form_pg_database pgdb = (Form_pg_database) GETSTRUCT(tup);
+		DatachecksumsWorkerDatabase *db;
+
+		oldctx = MemoryContextSwitchTo(ctx);
+
+		db = (DatachecksumsWorkerDatabase *) palloc(sizeof(DatachecksumsWorkerDatabase));
+
+		db->dboid = pgdb->oid;
+		db->dbname = pstrdup(NameStr(pgdb->datname));
+
+		DatabaseList = lappend(DatabaseList, db);
+
+		MemoryContextSwitchTo(oldctx);
+	}
+
+	table_endscan(scan);
+	table_close(rel, AccessShareLock);
+
+	CommitTransactionCommand();
+
+	return DatabaseList;
+}
+
+/*
+ * BuildRelationList
+ *		Compile a list of relations in the database
+ *
+ * Returns a list of OIDs for the request relation types. If temp_relations
+ * is True then only temporary relations are returned. If temp_relations is
+ * False then non-temporary relations which have data checksums are returned.
+ * If include_shared is True then shared relations are included as well in a
+ * non-temporary list. include_shared has no relevance when building a list of
+ * temporary relations.
+ */
+static List *
+BuildRelationList(bool temp_relations, bool include_shared)
+{
+	List	   *RelationList = NIL;
+	Relation	rel;
+	TableScanDesc scan;
+	HeapTuple	tup;
+	MemoryContext ctx = CurrentMemoryContext;
+	MemoryContext oldctx;
+
+	StartTransactionCommand();
+
+	rel = table_open(RelationRelationId, AccessShareLock);
+	scan = table_beginscan_catalog(rel, 0, NULL);
+
+	while (HeapTupleIsValid(tup = heap_getnext(scan, ForwardScanDirection)))
+	{
+		Form_pg_class pgc = (Form_pg_class) GETSTRUCT(tup);
+
+		/*
+		 * Only include temporary relations when asked for a temp relation
+		 * list.
+		 */
+		if (pgc->relpersistence == RELPERSISTENCE_TEMP)
+		{
+			if (!temp_relations)
+				continue;
+		}
+		else
+		{
+			if (!RELKIND_HAS_STORAGE(pgc->relkind))
+				continue;
+
+			if (pgc->relhaschecksums)
+				continue;
+
+			if (pgc->relisshared && !include_shared)
+				continue;
+		}
+
+		oldctx = MemoryContextSwitchTo(ctx);
+		RelationList = lappend_oid(RelationList, pgc->oid);
+		MemoryContextSwitchTo(oldctx);
+	}
+
+	table_endscan(scan);
+	table_close(rel, AccessShareLock);
+
+	CommitTransactionCommand();
+
+	return RelationList;
+}
+
+/*
+ * ResetDataChecksumsStateInDatabase
+ *		Main worker function for clearing checksums state in the catalog
+ *
+ * Resets the pg_class.relhaschecksums flag to false for all entries in the
+ * current database. This is required to be performed before adding checksums
+ * to a running cluster in order to track the state of the processing.
+ */
+void
+ResetDataChecksumsStateInDatabase(Datum arg)
+{
+	Relation	rel;
+	HeapTuple	tuple;
+	Oid			dboid = DatumGetObjectId(arg);
+	TableScanDesc scan;
+	Form_pg_class pgc;
+
+	pqsignal(SIGTERM, die);
+
+	BackgroundWorkerUnblockSignals();
+
+	MyBackendType = B_DATACHECKSUMSWORKER_WORKER;
+	init_ps_display(NULL);
+
+	ereport(DEBUG1,
+			(errmsg("resetting catalog state for data checksums in database with OID %u",
+					dboid)));
+
+	BackgroundWorkerInitializeConnectionByOid(dboid, InvalidOid, BGWORKER_BYPASS_ALLOWCONN);
+
+	StartTransactionCommand();
+
+	rel = table_open(RelationRelationId, RowExclusiveLock);
+	scan = table_beginscan_catalog(rel, 0, NULL);
+
+	while (HeapTupleIsValid(tuple = heap_getnext(scan, ForwardScanDirection)))
+	{
+		tuple = heap_copytuple(tuple);
+		pgc = (Form_pg_class) GETSTRUCT(tuple);
+
+		if (pgc->relhaschecksums)
+		{
+			pgc->relhaschecksums = false;
+			CatalogTupleUpdate(rel, &tuple->t_self, tuple);
+		}
+
+		heap_freetuple(tuple);
+	}
+
+	table_endscan(scan);
+	table_close(rel, RowExclusiveLock);
+
+	CommitTransactionCommand();
+
+	DatachecksumsWorkerShmem->success = DATACHECKSUMSWORKER_SUCCESSFUL;
+}
+
+/*
+ * DatachecksumsWorkerMain
+ *
+ * Main function for enabling checksums in a single database, This is the
+ * function set as the bgw_function_name in the dynamic background worker
+ * process initiiated for each database by the worker launcher. After enabling
+ * data checksums in each applicable relation in the database, it will wait for
+ * all temporary relations that were present when the function started to
+ * disappear before returning. This is required since we cannot rewrite
+ * existing temporary relations with data checksums.
+ */
+void
+DatachecksumsWorkerMain(Datum arg)
+{
+	Oid			dboid = DatumGetObjectId(arg);
+	List	   *RelationList = NIL;
+	List	   *InitialTempTableList = NIL;
+	ListCell   *lc;
+	BufferAccessStrategy strategy;
+	bool		aborted = false;
+
+	enabling_checksums = true;
+
+	pqsignal(SIGTERM, die);
+
+	BackgroundWorkerUnblockSignals();
+
+	MyBackendType = B_DATACHECKSUMSWORKER_WORKER;
+	init_ps_display(NULL);
+
+	ereport(DEBUG1,
+			(errmsg("starting data checksum processing in database with OID %u",
+					dboid)));
+
+	BackgroundWorkerInitializeConnectionByOid(dboid, InvalidOid,
+											  BGWORKER_BYPASS_ALLOWCONN);
+
+	/*
+	 * Get a list of all temp tables present as we start in this database. We
+	 * need to wait until they are all gone until we are done, since we cannot
+	 * access these relations and modify them.
+	 */
+	InitialTempTableList = BuildRelationList(true, false);
+
+	/*
+	 * Enable vacuum cost delay, if any.
+	 */
+	Assert(DatachecksumsWorkerShmem->enabling_checksums);
+	VacuumCostDelay = DatachecksumsWorkerShmem->cost_delay;
+	VacuumCostLimit = DatachecksumsWorkerShmem->cost_limit;
+	VacuumCostActive = (VacuumCostDelay > 0);
+	VacuumCostBalance = 0;
+	VacuumPageHit = 0;
+	VacuumPageMiss = 0;
+	VacuumPageDirty = 0;
+
+	/*
+	 * Create and set the vacuum strategy as our buffer strategy.
+	 */
+	strategy = GetAccessStrategy(BAS_VACUUM);
+
+	RelationList = BuildRelationList(false,
+									 DatachecksumsWorkerShmem->process_shared_catalogs);
+	foreach(lc, RelationList)
+	{
+		Oid			reloid = lfirst_oid(lc);
+
+		if (!ProcessSingleRelationByOid(reloid, strategy))
+		{
+			aborted = true;
+			break;
+		}
+	}
+	list_free(RelationList);
+
+	if (aborted)
+	{
+		DatachecksumsWorkerShmem->success = DATACHECKSUMSWORKER_ABORTED;
+		ereport(DEBUG1,
+				(errmsg("data checksum processing aborted in database OID %u",
+						dboid)));
+		return;
+	}
+
+	/*
+	 * Wait for all temp tables that existed when we started to go away. This
+	 * is necessary since we cannot "reach" them to enable checksums. Any temp
+	 * tables created after we started will already have checksums in them
+	 * (due to the "inprogress-on" state), so no need to wait for those.
+	 */
+	for (;;)
+	{
+		List	   *CurrentTempTables;
+		ListCell   *lc;
+		int			numleft;
+		char		activity[64];
+
+		CurrentTempTables = BuildRelationList(true, false);
+		numleft = 0;
+		foreach(lc, InitialTempTableList)
+		{
+			if (list_member_oid(CurrentTempTables, lfirst_oid(lc)))
+				numleft++;
+		}
+		list_free(CurrentTempTables);
+
+		if (numleft == 0)
+			break;
+
+		/* At least one temp table is left to wait for */
+		snprintf(activity,
+				 sizeof(activity),
+				 "Waiting for %d temp tables to be removed", numleft);
+		pgstat_report_activity(STATE_RUNNING, activity);
+
+		/* Retry every 5 seconds */
+		ResetLatch(MyLatch);
+		(void) WaitLatch(MyLatch,
+						 WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
+						 5000,
+						 WAIT_EVENT_CHECKSUM_ENABLE_FINISHCONDITION);
+
+		LWLockAcquire(DatachecksumsWorkerLock, LW_EXCLUSIVE);
+		aborted = DatachecksumsWorkerShmem->launch_enable_checksums != enabling_checksums;
+		LWLockRelease(DatachecksumsWorkerLock);
+
+		if (aborted || abort_requested)
+		{
+			DatachecksumsWorkerShmem->success = DATACHECKSUMSWORKER_ABORTED;
+			ereport(DEBUG1,
+					(errmsg("data checksum processing aborted in database OID %u",
+							dboid)));
+			return;
+		}
+	}
+
+	list_free(InitialTempTableList);
+
+	DatachecksumsWorkerShmem->success = DATACHECKSUMSWORKER_SUCCESSFUL;
+	ereport(DEBUG1,
+			(errmsg("data checksum processing completed in database with OID %u",
+					dboid)));
+}
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index f75b52719dd..0fef097eb80 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -4017,6 +4017,12 @@ pgstat_get_wait_ipc(WaitEventIPC w)
 		case WAIT_EVENT_CHECKPOINT_START:
 			event_name = "CheckpointStart";
 			break;
+		case WAIT_EVENT_CHECKSUM_ENABLE_STARTCONDITION:
+			event_name = "ChecksumEnableStartCondition";
+			break;
+		case WAIT_EVENT_CHECKSUM_ENABLE_FINISHCONDITION:
+			event_name = "ChecksumEnableFinishCondition";
+			break;
 		case WAIT_EVENT_EXECUTE_GATHER:
 			event_name = "ExecuteGather";
 			break;
diff --git a/src/backend/replication/basebackup.c b/src/backend/replication/basebackup.c
index 0f54635550b..cc494b6f13d 100644
--- a/src/backend/replication/basebackup.c
+++ b/src/backend/replication/basebackup.c
@@ -1612,7 +1612,7 @@ sendFile(const char *readfilename, const char *tarfilename,
 
 	_tarWriteHeader(tarfilename, NULL, statbuf, false);
 
-	if (!noverify_checksums && DataChecksumsEnabled())
+	if (!noverify_checksums)
 	{
 		char	   *filename;
 
@@ -1698,7 +1698,14 @@ sendFile(const char *readfilename, const char *tarfilename,
 				 */
 				if (!PageIsNew(page) && PageGetLSN(page) < startptr)
 				{
+					HOLD_INTERRUPTS();
+					if (!DataChecksumsNeedVerify())
+					{
+						RESUME_INTERRUPTS();
+						continue;
+					}
 					checksum = pg_checksum_page((char *) page, blkno + segmentno * RELSEG_SIZE);
+					RESUME_INTERRUPTS();
 					phdr = (PageHeader) page;
 					if (phdr->pd_checksum != checksum)
 					{
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index afa1df00d0e..d9c482454ff 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -223,6 +223,7 @@ DecodeXLogOp(LogicalDecodingContext *ctx, XLogRecordBuffer *buf)
 		case XLOG_FPW_CHANGE:
 		case XLOG_FPI_FOR_HINT:
 		case XLOG_FPI:
+		case XLOG_CHECKSUMS:
 			break;
 		default:
 			elog(ERROR, "unexpected RM_XLOG_ID record type: %u", info);
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index 561c212092f..9362ec00184 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -2944,8 +2944,13 @@ BufferGetLSNAtomic(Buffer buffer)
 	/*
 	 * If we don't need locking for correctness, fastpath out.
 	 */
+	HOLD_INTERRUPTS();
 	if (!XLogHintBitIsNeeded() || BufferIsLocal(buffer))
+	{
+		RESUME_INTERRUPTS();
 		return PageGetLSN(page);
+	}
+	RESUME_INTERRUPTS();
 
 	/* Make sure we've got a real buffer, and that we hold a pin on it. */
 	Assert(BufferIsValid(buffer));
diff --git a/src/backend/storage/ipc/ipci.c b/src/backend/storage/ipc/ipci.c
index f9bbe97b507..c7928f34957 100644
--- a/src/backend/storage/ipc/ipci.c
+++ b/src/backend/storage/ipc/ipci.c
@@ -28,6 +28,7 @@
 #include "postmaster/autovacuum.h"
 #include "postmaster/bgworker_internals.h"
 #include "postmaster/bgwriter.h"
+#include "postmaster/datachecksumsworker.h"
 #include "postmaster/postmaster.h"
 #include "replication/logicallauncher.h"
 #include "replication/origin.h"
@@ -149,6 +150,7 @@ CreateSharedMemoryAndSemaphores(void)
 		size = add_size(size, BTreeShmemSize());
 		size = add_size(size, SyncScanShmemSize());
 		size = add_size(size, AsyncShmemSize());
+		size = add_size(size, DatachecksumsWorkerShmemSize());
 #ifdef EXEC_BACKEND
 		size = add_size(size, ShmemBackendArraySize());
 #endif
@@ -259,6 +261,7 @@ CreateSharedMemoryAndSemaphores(void)
 	WalSndShmemInit();
 	WalRcvShmemInit();
 	ApplyLauncherShmemInit();
+	DatachecksumsWorkerShmemInit();
 
 	/*
 	 * Set up other modules that need some shared memory space
diff --git a/src/backend/storage/ipc/procsignal.c b/src/backend/storage/ipc/procsignal.c
index c43cdd685b4..a3720617f94 100644
--- a/src/backend/storage/ipc/procsignal.c
+++ b/src/backend/storage/ipc/procsignal.c
@@ -18,6 +18,7 @@
 #include <unistd.h>
 
 #include "access/parallel.h"
+#include "access/xlog.h"
 #include "port/pg_bitutils.h"
 #include "commands/async.h"
 #include "miscadmin.h"
@@ -98,7 +99,6 @@ static volatile ProcSignalSlot *MyProcSignalSlot = NULL;
 static bool CheckProcSignal(ProcSignalReason reason);
 static void CleanupProcSignalState(int status, Datum arg);
 static void ResetProcSignalBarrierBits(uint32 flags);
-static bool ProcessBarrierPlaceholder(void);
 
 /*
  * ProcSignalShmemSize
@@ -538,8 +538,17 @@ ProcessProcSignalBarrier(void)
 				type = (ProcSignalBarrierType) pg_rightmost_one_pos32(flags);
 				switch (type)
 				{
-					case PROCSIGNAL_BARRIER_PLACEHOLDER:
-						processed = ProcessBarrierPlaceholder();
+					case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON:
+						processed = AbsorbChecksumsOnInProgressBarrier();
+						break;
+					case PROCSIGNAL_BARRIER_CHECKSUM_ON:
+						processed = AbsorbChecksumsOnBarrier();
+						break;
+					case PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF:
+						processed = AbsorbChecksumsOffInProgressBarrier();
+						break;
+					case PROCSIGNAL_BARRIER_CHECKSUM_OFF:
+						processed = AbsorbChecksumsOffBarrier();
 						break;
 				}
 
@@ -604,24 +613,6 @@ ResetProcSignalBarrierBits(uint32 flags)
 	InterruptPending = true;
 }
 
-static bool
-ProcessBarrierPlaceholder(void)
-{
-	/*
-	 * XXX. This is just a placeholder until the first real user of this
-	 * machinery gets committed. Rename PROCSIGNAL_BARRIER_PLACEHOLDER to
-	 * PROCSIGNAL_BARRIER_SOMETHING_ELSE where SOMETHING_ELSE is something
-	 * appropriately descriptive. Get rid of this function and instead have
-	 * ProcessBarrierSomethingElse. Most likely, that function should live in
-	 * the file pertaining to that subsystem, rather than here.
-	 *
-	 * The return value should be 'true' if the barrier was successfully
-	 * absorbed and 'false' if not. Note that returning 'false' can lead to
-	 * very frequent retries, so try hard to make that an uncommon case.
-	 */
-	return true;
-}
-
 /*
  * CheckProcSignal - check to see if a particular reason has been
  * signaled, and clear the signal flag.  Should be called after receiving
diff --git a/src/backend/storage/lmgr/lwlocknames.txt b/src/backend/storage/lmgr/lwlocknames.txt
index 6c7cf6c2956..5b083749d55 100644
--- a/src/backend/storage/lmgr/lwlocknames.txt
+++ b/src/backend/storage/lmgr/lwlocknames.txt
@@ -53,3 +53,4 @@ XactTruncationLock					44
 # 45 was XactTruncationLock until removal of BackendRandomLock
 WrapLimitsVacuumLock				46
 NotifyQueueTailLock					47
+DatachecksumsWorkerLock				48
diff --git a/src/backend/storage/page/README b/src/backend/storage/page/README
index e30d7ac59ad..78edf57adc8 100644
--- a/src/backend/storage/page/README
+++ b/src/backend/storage/page/README
@@ -10,7 +10,9 @@ http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf, discussed
 2010/12/22 on -hackers list.
 
 Current implementation requires this be enabled system-wide at initdb time, or
-by using the pg_checksums tool on an offline cluster.
+by using the pg_checksums tool on an offline cluster. Checksums can also be
+turned on and off using pg_enable_data_checksums()/pg_disable_data_checksums()
+at runtime.
 
 The checksum is not valid at all times on a data page!!
 The checksum is valid when the page leaves the shared pool and is checked
diff --git a/src/backend/storage/page/bufpage.c b/src/backend/storage/page/bufpage.c
index 9ac556b4ae0..8fbebd98700 100644
--- a/src/backend/storage/page/bufpage.c
+++ b/src/backend/storage/page/bufpage.c
@@ -100,13 +100,20 @@ PageIsVerifiedExtended(Page page, BlockNumber blkno, int flags)
 	 */
 	if (!PageIsNew(page))
 	{
-		if (DataChecksumsEnabled())
+		/*
+		 * Hold interrupts for the duration of the checksum check to ensure
+		 * that the data checksums state cannot change and thus risking a
+		 * false positive or negative.
+		 */
+		HOLD_INTERRUPTS();
+		if (DataChecksumsNeedVerify())
 		{
 			checksum = pg_checksum_page((char *) page, blkno);
 
 			if (checksum != p->pd_checksum)
 				checksum_failure = true;
 		}
+		RESUME_INTERRUPTS();
 
 		/*
 		 * The following checks don't prove the header is correct, only that
@@ -1394,10 +1401,6 @@ PageSetChecksumCopy(Page page, BlockNumber blkno)
 {
 	static char *pageCopy = NULL;
 
-	/* If we don't need a checksum, just return the passed-in data */
-	if (PageIsNew(page) || !DataChecksumsEnabled())
-		return (char *) page;
-
 	/*
 	 * We allocate the copy space once and use it over on each subsequent
 	 * call.  The point of palloc'ing here, rather than having a static char
@@ -1407,8 +1410,17 @@ PageSetChecksumCopy(Page page, BlockNumber blkno)
 	if (pageCopy == NULL)
 		pageCopy = MemoryContextAlloc(TopMemoryContext, BLCKSZ);
 
+	/* If we don't need a checksum, just return the passed-in data */
+	HOLD_INTERRUPTS();
+	if (PageIsNew(page) || !DataChecksumsNeedWrite())
+	{
+		RESUME_INTERRUPTS();
+		return (char *) page;
+	}
+
 	memcpy(pageCopy, (char *) page, BLCKSZ);
 	((PageHeader) pageCopy)->pd_checksum = pg_checksum_page(pageCopy, blkno);
+	RESUME_INTERRUPTS();
 	return pageCopy;
 }
 
@@ -1421,9 +1433,14 @@ PageSetChecksumCopy(Page page, BlockNumber blkno)
 void
 PageSetChecksumInplace(Page page, BlockNumber blkno)
 {
+	HOLD_INTERRUPTS();
 	/* If we don't need a checksum, just return */
-	if (PageIsNew(page) || !DataChecksumsEnabled())
+	if (PageIsNew(page) || !DataChecksumsNeedWrite())
+	{
+		RESUME_INTERRUPTS();
 		return;
+	}
 
 	((PageHeader) page)->pd_checksum = pg_checksum_page((char *) page, blkno);
+	RESUME_INTERRUPTS();
 }
diff --git a/src/backend/utils/adt/pgstatfuncs.c b/src/backend/utils/adt/pgstatfuncs.c
index 62bff52638d..4ac396ccf1e 100644
--- a/src/backend/utils/adt/pgstatfuncs.c
+++ b/src/backend/utils/adt/pgstatfuncs.c
@@ -1567,9 +1567,6 @@ pg_stat_get_db_checksum_failures(PG_FUNCTION_ARGS)
 	int64		result;
 	PgStat_StatDBEntry *dbentry;
 
-	if (!DataChecksumsEnabled())
-		PG_RETURN_NULL();
-
 	if ((dbentry = pgstat_fetch_stat_dbentry(dbid)) == NULL)
 		result = 0;
 	else
@@ -1585,9 +1582,6 @@ pg_stat_get_db_checksum_last_failure(PG_FUNCTION_ARGS)
 	TimestampTz result;
 	PgStat_StatDBEntry *dbentry;
 
-	if (!DataChecksumsEnabled())
-		PG_RETURN_NULL();
-
 	if ((dbentry = pgstat_fetch_stat_dbentry(dbid)) == NULL)
 		result = 0;
 	else
diff --git a/src/backend/utils/cache/relcache.c b/src/backend/utils/cache/relcache.c
index 7ef510cd01b..633821bae59 100644
--- a/src/backend/utils/cache/relcache.c
+++ b/src/backend/utils/cache/relcache.c
@@ -271,7 +271,8 @@ static void write_relcache_init_file(bool shared);
 static void write_item(const void *data, Size len, FILE *fp);
 
 static void formrdesc(const char *relationName, Oid relationReltype,
-					  bool isshared, int natts, const FormData_pg_attribute *attrs);
+					  bool isshared, int natts, const FormData_pg_attribute *attrs,
+					  bool haschecksums);
 
 static HeapTuple ScanPgRelation(Oid targetRelId, bool indexOK, bool force_non_historic);
 static Relation AllocateRelationDesc(Form_pg_class relp);
@@ -1828,7 +1829,8 @@ RelationInitTableAccessMethod(Relation relation)
 static void
 formrdesc(const char *relationName, Oid relationReltype,
 		  bool isshared,
-		  int natts, const FormData_pg_attribute *attrs)
+		  int natts, const FormData_pg_attribute *attrs,
+		  bool haschecksums)
 {
 	Relation	relation;
 	int			i;
@@ -1896,6 +1898,8 @@ formrdesc(const char *relationName, Oid relationReltype,
 	relation->rd_rel->relnatts = (int16) natts;
 	relation->rd_rel->relam = HEAP_TABLE_AM_OID;
 
+	relation->rd_rel->relhaschecksums = haschecksums;
+
 	/*
 	 * initialize attribute tuple form
 	 *
@@ -3548,6 +3552,27 @@ RelationBuildLocalRelation(const char *relname,
 		relkind == RELKIND_MATVIEW)
 		RelationInitTableAccessMethod(rel);
 
+	/*
+	 * Set the data checksum state. Since the data checksum state can change
+	 * at any time, the fetched value might be out of date by the time the
+	 * relation is built.  DataChecksumsNeedWrite returns true when data
+	 * checksums are: enabled; are in the process of being enabled (state:
+	 * "inprogress-on"); are in the process of being disabled (state:
+	 * "inprogress-off"). Since relhaschecksums is only used to track progress
+	 * when data checksums are being enabled, and going from disabled to
+	 * enabled will clear relhaschecksums before starting, it is safe to use
+	 * this value for a concurrent state transition to off.
+	 *
+	 * If DataChecksumsNeedWrite returns false, and is concurrently changed to
+	 * true then that implies that checksums are being enabled. Worst case,
+	 * this will lead to the relation being processed for checksums even
+	 * though each page written will have them already.  Performing this last
+	 * shortens the window, but doesn't avoid it.
+	 */
+	HOLD_INTERRUPTS();
+	rel->rd_rel->relhaschecksums = DataChecksumsNeedWrite();
+	RESUME_INTERRUPTS();
+
 	/*
 	 * Okay to insert into the relcache hash table.
 	 *
@@ -3813,6 +3838,7 @@ void
 RelationCacheInitializePhase2(void)
 {
 	MemoryContext oldcxt;
+	bool		haschecksums;
 
 	/*
 	 * relation mapper needs initialized too
@@ -3837,16 +3863,24 @@ RelationCacheInitializePhase2(void)
 	 */
 	if (!load_relcache_init_file(true))
 	{
+		/*
+		 * Our local state can't change at this point, so we can cache the
+		 * checksum state.
+		 */
+		HOLD_INTERRUPTS();
+		haschecksums = DataChecksumsNeedWrite();
+		RESUME_INTERRUPTS();
+
 		formrdesc("pg_database", DatabaseRelation_Rowtype_Id, true,
-				  Natts_pg_database, Desc_pg_database);
+				  Natts_pg_database, Desc_pg_database, haschecksums);
 		formrdesc("pg_authid", AuthIdRelation_Rowtype_Id, true,
-				  Natts_pg_authid, Desc_pg_authid);
+				  Natts_pg_authid, Desc_pg_authid, haschecksums);
 		formrdesc("pg_auth_members", AuthMemRelation_Rowtype_Id, true,
-				  Natts_pg_auth_members, Desc_pg_auth_members);
+				  Natts_pg_auth_members, Desc_pg_auth_members, haschecksums);
 		formrdesc("pg_shseclabel", SharedSecLabelRelation_Rowtype_Id, true,
-				  Natts_pg_shseclabel, Desc_pg_shseclabel);
+				  Natts_pg_shseclabel, Desc_pg_shseclabel, haschecksums);
 		formrdesc("pg_subscription", SubscriptionRelation_Rowtype_Id, true,
-				  Natts_pg_subscription, Desc_pg_subscription);
+				  Natts_pg_subscription, Desc_pg_subscription, haschecksums);
 
 #define NUM_CRITICAL_SHARED_RELS	5	/* fix if you change list above */
 	}
@@ -3875,6 +3909,7 @@ RelationCacheInitializePhase3(void)
 	RelIdCacheEnt *idhentry;
 	MemoryContext oldcxt;
 	bool		needNewCacheFile = !criticalSharedRelcachesBuilt;
+	bool		haschecksums;
 
 	/*
 	 * relation mapper needs initialized too
@@ -3895,15 +3930,18 @@ RelationCacheInitializePhase3(void)
 		!load_relcache_init_file(false))
 	{
 		needNewCacheFile = true;
+		HOLD_INTERRUPTS();
+		haschecksums = DataChecksumsNeedWrite();
+		RESUME_INTERRUPTS();
 
 		formrdesc("pg_class", RelationRelation_Rowtype_Id, false,
-				  Natts_pg_class, Desc_pg_class);
+				  Natts_pg_class, Desc_pg_class, haschecksums);
 		formrdesc("pg_attribute", AttributeRelation_Rowtype_Id, false,
-				  Natts_pg_attribute, Desc_pg_attribute);
+				  Natts_pg_attribute, Desc_pg_attribute, haschecksums);
 		formrdesc("pg_proc", ProcedureRelation_Rowtype_Id, false,
-				  Natts_pg_proc, Desc_pg_proc);
+				  Natts_pg_proc, Desc_pg_proc, haschecksums);
 		formrdesc("pg_type", TypeRelation_Rowtype_Id, false,
-				  Natts_pg_type, Desc_pg_type);
+				  Natts_pg_type, Desc_pg_type, haschecksums);
 
 #define NUM_CRITICAL_LOCAL_RELS 4	/* fix if you change list above */
 	}
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 0f67b99cc55..045da219044 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -275,6 +275,12 @@ GetBackendTypeDesc(BackendType backendType)
 		case B_LOGGER:
 			backendDesc = "logger";
 			break;
+		case B_DATACHECKSUMSWORKER_LAUNCHER:
+			backendDesc = "datachecksumsworker launcher";
+			break;
+		case B_DATACHECKSUMSWORKER_WORKER:
+			backendDesc = "datachecksumsworker worker";
+			break;
 	}
 
 	return backendDesc;
diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c
index e5965bc517d..92367ece4b8 100644
--- a/src/backend/utils/init/postinit.c
+++ b/src/backend/utils/init/postinit.c
@@ -606,6 +606,11 @@ InitPostgres(const char *in_dbname, Oid dboid, const char *username,
 	if (MyBackendId > MaxBackends || MyBackendId <= 0)
 		elog(FATAL, "bad backend ID: %d", MyBackendId);
 
+	/*
+	 * Set up backend local cache of Controldata values.
+	 */
+	InitLocalControldata();
+
 	/* Now that we have a BackendId, we can participate in ProcSignal */
 	ProcSignalInit(MyBackendId);
 
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 17579eeaca9..3b7207afb54 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -36,6 +36,7 @@
 #include "access/transam.h"
 #include "access/twophase.h"
 #include "access/xact.h"
+#include "access/xlog.h"
 #include "access/xlog_internal.h"
 #include "catalog/namespace.h"
 #include "catalog/pg_authid.h"
@@ -76,6 +77,7 @@
 #include "replication/walreceiver.h"
 #include "replication/walsender.h"
 #include "storage/bufmgr.h"
+#include "storage/checksum.h"
 #include "storage/dsm_impl.h"
 #include "storage/fd.h"
 #include "storage/large_object.h"
@@ -500,6 +502,17 @@ static struct config_enum_entry shared_memory_options[] = {
 	{NULL, 0, false}
 };
 
+/*
+ * Options for data_checksums enum.
+ */
+static const struct config_enum_entry data_checksum_options[] = {
+	{"on", DATA_CHECKSUMS_ON, true},
+	{"off", DATA_CHECKSUMS_OFF, true},
+	{"inprogress-on", DATA_CHECKSUMS_INPROGRESS_ON, true},
+	{"inprogress-off", DATA_CHECKSUMS_INPROGRESS_OFF, true},
+	{NULL, 0, false}
+};
+
 /*
  * Options for enum values stored in other modules
  */
@@ -609,7 +622,7 @@ static int	max_identifier_length;
 static int	block_size;
 static int	segment_size;
 static int	wal_block_size;
-static bool data_checksums;
+static int	data_checksums;
 static bool integer_datetimes;
 static bool assert_enabled;
 static bool in_hot_standby;
@@ -1910,17 +1923,6 @@ static struct config_bool ConfigureNamesBool[] =
 		NULL, NULL, NULL
 	},
 
-	{
-		{"data_checksums", PGC_INTERNAL, PRESET_OPTIONS,
-			gettext_noop("Shows whether data checksums are turned on for this cluster."),
-			NULL,
-			GUC_NOT_IN_SAMPLE | GUC_DISALLOW_IN_FILE
-		},
-		&data_checksums,
-		false,
-		NULL, NULL, NULL
-	},
-
 	{
 		{"syslog_sequence_numbers", PGC_SIGHUP, LOGGING_WHERE,
 			gettext_noop("Add sequence number to syslog messages to avoid duplicate suppression."),
@@ -4830,6 +4832,17 @@ static struct config_enum ConfigureNamesEnum[] =
 		NULL, NULL, NULL
 	},
 
+	{
+		{"data_checksums", PGC_INTERNAL, PRESET_OPTIONS,
+			gettext_noop("Shows whether data checksums are turned on for this cluster."),
+			NULL,
+			GUC_NOT_IN_SAMPLE | GUC_DISALLOW_IN_FILE
+		},
+		&data_checksums,
+		DATA_CHECKSUMS_OFF, data_checksum_options,
+		NULL, NULL, show_data_checksums
+	},
+
 	/* End-of-list marker */
 	{
 		{NULL, 0, 0, NULL, NULL}, NULL, 0, NULL, NULL, NULL, NULL
diff --git a/src/bin/pg_checksums/pg_checksums.c b/src/bin/pg_checksums/pg_checksums.c
index 0223ee44082..f3f029f41e5 100644
--- a/src/bin/pg_checksums/pg_checksums.c
+++ b/src/bin/pg_checksums/pg_checksums.c
@@ -600,7 +600,7 @@ main(int argc, char *argv[])
 		exit(1);
 	}
 
-	if (ControlFile->data_checksum_version > 0 &&
+	if (ControlFile->data_checksum_version == DATA_CHECKSUMS_ON &&
 		mode == PG_MODE_ENABLE)
 	{
 		pg_log_error("data checksums are already enabled in cluster");
diff --git a/src/bin/pg_upgrade/controldata.c b/src/bin/pg_upgrade/controldata.c
index 4f647cdf334..12988574583 100644
--- a/src/bin/pg_upgrade/controldata.c
+++ b/src/bin/pg_upgrade/controldata.c
@@ -671,6 +671,15 @@ check_control_data(ControlData *oldctrl,
 	 * check_for_isn_and_int8_passing_mismatch().
 	 */
 
+	/*
+	 * If checksums have been turned on in the old cluster, but the
+	 * datachecksumsworker have yet to finish, then disallow upgrading. The
+	 * user should either let the process finish, or turn off checksums,
+	 * before retrying.
+	 */
+	if (oldctrl->data_checksum_version == 2)
+		pg_fatal("checksum enabling in old cluster is in progress\n");
+
 	/*
 	 * We might eventually allow upgrades from checksum to no-checksum
 	 * clusters.
diff --git a/src/bin/pg_upgrade/pg_upgrade.h b/src/bin/pg_upgrade/pg_upgrade.h
index 919a7849fd0..b35cd4d503a 100644
--- a/src/bin/pg_upgrade/pg_upgrade.h
+++ b/src/bin/pg_upgrade/pg_upgrade.h
@@ -218,7 +218,7 @@ typedef struct
 	uint32		large_object;
 	bool		date_is_int;
 	bool		float8_pass_by_value;
-	bool		data_checksum_version;
+	uint32		data_checksum_version;
 } ControlData;
 
 /*
diff --git a/src/include/access/xlog.h b/src/include/access/xlog.h
index 75ec1073bd0..6947c095914 100644
--- a/src/include/access/xlog.h
+++ b/src/include/access/xlog.h
@@ -198,8 +198,11 @@ extern PGDLLIMPORT int wal_level;
  * individual bits on a page, it's still consistent no matter what combination
  * of the bits make it to disk, but the checksum wouldn't match.  Also WAL-log
  * them if forced by wal_log_hints=on.
+ *
+ * Since XLogHintBitIsNeeded calls DataChecksumsNeedWrite, interrupts must be
+ * held off during this call.
  */
-#define XLogHintBitIsNeeded() (DataChecksumsEnabled() || wal_log_hints)
+#define XLogHintBitIsNeeded() (wal_log_hints || DataChecksumsNeedWrite())
 
 /* Do we need to WAL-log information required only for Hot Standby and logical replication? */
 #define XLogStandbyInfoActive() (wal_level >= WAL_LEVEL_REPLICA)
@@ -318,7 +321,19 @@ extern TimestampTz GetCurrentChunkReplayStartTime(void);
 extern void UpdateControlFile(void);
 extern uint64 GetSystemIdentifier(void);
 extern char *GetMockAuthenticationNonce(void);
-extern bool DataChecksumsEnabled(void);
+extern bool DataChecksumsNeedWrite(void);
+extern bool DataChecksumsNeedVerify(void);
+extern bool DataChecksumsOnInProgress(void);
+extern bool DataChecksumsOffInProgress(void);
+extern void SetDataChecksumsOnInProgress(void);
+extern void SetDataChecksumsOn(void);
+extern void SetDataChecksumsOff(void);
+extern bool AbsorbChecksumsOnInProgressBarrier(void);
+extern bool AbsorbChecksumsOffInProgressBarrier(void);
+extern bool AbsorbChecksumsOnBarrier(void);
+extern bool AbsorbChecksumsOffBarrier(void);
+extern const char *show_data_checksums(void);
+extern void InitLocalControldata(void);
 extern XLogRecPtr GetFakeLSNForUnloggedRel(void);
 extern Size XLOGShmemSize(void);
 extern void XLOGShmemInit(void);
diff --git a/src/include/access/xlog_internal.h b/src/include/access/xlog_internal.h
index 224cae0246f..adbe81e890b 100644
--- a/src/include/access/xlog_internal.h
+++ b/src/include/access/xlog_internal.h
@@ -25,6 +25,7 @@
 #include "lib/stringinfo.h"
 #include "pgtime.h"
 #include "storage/block.h"
+#include "storage/checksum.h"
 #include "storage/relfilenode.h"
 
 
@@ -249,6 +250,12 @@ typedef struct xl_restore_point
 	char		rp_name[MAXFNAMELEN];
 } xl_restore_point;
 
+/* Information logged when checksum level is changed */
+typedef struct xl_checksum_state
+{
+	ChecksumType new_checksumtype;
+}			xl_checksum_state;
+
 /* End of recovery mark, when we don't do an END_OF_RECOVERY checkpoint */
 typedef struct xl_end_of_recovery
 {
diff --git a/src/include/catalog/pg_class.h b/src/include/catalog/pg_class.h
index e8dcd15a55f..bf296625e42 100644
--- a/src/include/catalog/pg_class.h
+++ b/src/include/catalog/pg_class.h
@@ -119,6 +119,9 @@ CATALOG(pg_class,1259,RelationRelationId) BKI_BOOTSTRAP BKI_ROWTYPE_OID(83,Relat
 	/* is relation a partition? */
 	bool		relispartition BKI_DEFAULT(f);
 
+	/* does the relation have checksums enabled */
+	bool		relhaschecksums BKI_DEFAULT(f);
+
 	/* heap for rewrite during DDL, link to original rel */
 	Oid			relrewrite BKI_DEFAULT(0);
 
diff --git a/src/include/catalog/pg_control.h b/src/include/catalog/pg_control.h
index e3f48158ce7..d8229422afc 100644
--- a/src/include/catalog/pg_control.h
+++ b/src/include/catalog/pg_control.h
@@ -76,6 +76,7 @@ typedef struct CheckPoint
 #define XLOG_END_OF_RECOVERY			0x90
 #define XLOG_FPI_FOR_HINT				0xA0
 #define XLOG_FPI						0xB0
+#define XLOG_CHECKSUMS					0xC0
 
 
 /*
diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat
index b5f52d4e4a3..f050f15a58c 100644
--- a/src/include/catalog/pg_proc.dat
+++ b/src/include/catalog/pg_proc.dat
@@ -11301,6 +11301,22 @@
   proname => 'raw_array_subscript_handler', prorettype => 'internal',
   proargtypes => 'internal', prosrc => 'raw_array_subscript_handler' },
 
+{ oid => '9258',
+  descr => 'disable data checksums',
+  proname => 'pg_disable_data_checksums', provolatile => 'v', prorettype => 'void',
+  proparallel => 'r',
+  proargtypes => '',
+  prosrc => 'disable_data_checksums' },
+
+{ oid => '9257',
+  descr => 'enable data checksums',
+  proname => 'pg_enable_data_checksums', provolatile => 'v', prorettype => 'void',
+  proparallel => 'r',
+  proargtypes => 'int4 int4', proallargtypes => '{int4,int4}',
+  proargmodes => '{i,i}',
+  proargnames => '{cost_delay,cost_limit}',
+  prosrc => 'enable_data_checksums' },
+
 # collation management functions
 { oid => '3445', descr => 'import collations from operating system',
   proname => 'pg_import_system_collations', procost => '100',
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 1bdc97e3082..f013acba76a 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -324,6 +324,8 @@ typedef enum BackendType
 	B_ARCHIVER,
 	B_STATS_COLLECTOR,
 	B_LOGGER,
+	B_DATACHECKSUMSWORKER_LAUNCHER,
+	B_DATACHECKSUMSWORKER_WORKER,
 } BackendType;
 
 extern BackendType MyBackendType;
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index 724068cf87e..0974dfadfe4 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -963,6 +963,8 @@ typedef enum
 	WAIT_EVENT_BTREE_PAGE,
 	WAIT_EVENT_CHECKPOINT_DONE,
 	WAIT_EVENT_CHECKPOINT_START,
+	WAIT_EVENT_CHECKSUM_ENABLE_STARTCONDITION,
+	WAIT_EVENT_CHECKSUM_ENABLE_FINISHCONDITION,
 	WAIT_EVENT_EXECUTE_GATHER,
 	WAIT_EVENT_HASH_BATCH_ALLOCATE,
 	WAIT_EVENT_HASH_BATCH_ELECT,
diff --git a/src/include/postmaster/datachecksumsworker.h b/src/include/postmaster/datachecksumsworker.h
new file mode 100644
index 00000000000..845f6bceaae
--- /dev/null
+++ b/src/include/postmaster/datachecksumsworker.h
@@ -0,0 +1,30 @@
+/*-------------------------------------------------------------------------
+ *
+ * datachecksumsworker.h
+ *	  header file for checksum helper background worker
+ *
+ *
+ * Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1994, Regents of the University of California
+ *
+ * src/include/postmaster/datachecksumsworker.h
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef DATACHECKSUMSWORKER_H
+#define DATACHECKSUMSWORKER_H
+
+/* Shared memory */
+extern Size DatachecksumsWorkerShmemSize(void);
+extern void DatachecksumsWorkerShmemInit(void);
+
+/* Start the background processes for enabling or disabling checksums */
+void		StartDatachecksumsWorkerLauncher(bool enable_checksums,
+											 int cost_delay, int cost_limit);
+
+/* Background worker entrypoints */
+void		DatachecksumsWorkerLauncherMain(Datum arg);
+void		DatachecksumsWorkerMain(Datum arg);
+void		ResetDataChecksumsStateInDatabase(Datum arg);
+
+#endif							/* DATACHECKSUMSWORKER_H */
diff --git a/src/include/storage/bufpage.h b/src/include/storage/bufpage.h
index 359b749f7f4..c35b747520a 100644
--- a/src/include/storage/bufpage.h
+++ b/src/include/storage/bufpage.h
@@ -198,6 +198,9 @@ typedef PageHeaderData *PageHeader;
  */
 #define PG_PAGE_LAYOUT_VERSION		4
 #define PG_DATA_CHECKSUM_VERSION	1
+#define PG_DATA_CHECKSUM_INPROGRESS_ON_VERSION		2
+#define PG_DATA_CHECKSUM_INPROGRESS_OFF_VERSION		3
+
 
 /* ----------------------------------------------------------------
  *						page support macros
diff --git a/src/include/storage/checksum.h b/src/include/storage/checksum.h
index 80d23591921..f736b12f986 100644
--- a/src/include/storage/checksum.h
+++ b/src/include/storage/checksum.h
@@ -15,6 +15,14 @@
 
 #include "storage/block.h"
 
+typedef enum ChecksumType
+{
+	DATA_CHECKSUMS_OFF = 0,
+	DATA_CHECKSUMS_ON,
+	DATA_CHECKSUMS_INPROGRESS_ON,
+	DATA_CHECKSUMS_INPROGRESS_OFF
+}			ChecksumType;
+
 /*
  * Compute the checksum for a Postgres page.  The page must be aligned on a
  * 4-byte boundary.
diff --git a/src/include/storage/procsignal.h b/src/include/storage/procsignal.h
index 4ae7dc33b8e..d865796d048 100644
--- a/src/include/storage/procsignal.h
+++ b/src/include/storage/procsignal.h
@@ -48,12 +48,10 @@ typedef enum
 
 typedef enum
 {
-	/*
-	 * XXX. PROCSIGNAL_BARRIER_PLACEHOLDER should be replaced when the first
-	 * real user of the ProcSignalBarrier mechanism is added. It's just here
-	 * for now because we can't have an empty enum.
-	 */
-	PROCSIGNAL_BARRIER_PLACEHOLDER = 0
+	PROCSIGNAL_BARRIER_CHECKSUM_OFF = 0,
+	PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_ON,
+	PROCSIGNAL_BARRIER_CHECKSUM_INPROGRESS_OFF,
+	PROCSIGNAL_BARRIER_CHECKSUM_ON
 } ProcSignalBarrierType;
 
 /*
diff --git a/src/test/Makefile b/src/test/Makefile
index ab1ef9a4753..9774816625b 100644
--- a/src/test/Makefile
+++ b/src/test/Makefile
@@ -13,7 +13,7 @@ top_builddir = ../..
 include $(top_builddir)/src/Makefile.global
 
 SUBDIRS = perl regress isolation modules authentication recovery subscription \
-	  locale
+	  locale checksum
 
 # Test suites that are not safe by default but can be run if selected
 # by the user via the whitespace-separated list in variable
diff --git a/src/test/checksum/.gitignore b/src/test/checksum/.gitignore
new file mode 100644
index 00000000000..871e943d50e
--- /dev/null
+++ b/src/test/checksum/.gitignore
@@ -0,0 +1,2 @@
+# Generated by test suite
+/tmp_check/
diff --git a/src/test/checksum/Makefile b/src/test/checksum/Makefile
new file mode 100644
index 00000000000..fd60f7e97f3
--- /dev/null
+++ b/src/test/checksum/Makefile
@@ -0,0 +1,23 @@
+#-------------------------------------------------------------------------
+#
+# Makefile for src/test/checksum
+#
+# Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+# Portions Copyright (c) 1994, Regents of the University of California
+#
+# src/test/checksum/Makefile
+#
+#-------------------------------------------------------------------------
+
+subdir = src/test/checksum
+top_builddir = ../../..
+include $(top_builddir)/src/Makefile.global
+
+check:
+	$(prove_check)
+
+installcheck:
+	$(prove_installcheck)
+
+clean distclean maintainer-clean:
+	rm -rf tmp_check
diff --git a/src/test/checksum/README b/src/test/checksum/README
new file mode 100644
index 00000000000..0f0317060b3
--- /dev/null
+++ b/src/test/checksum/README
@@ -0,0 +1,22 @@
+src/test/checksum/README
+
+Regression tests for data checksums
+===================================
+
+This directory contains a test suite for enabling data checksums
+in a running cluster.
+
+Running the tests
+=================
+
+    make check
+
+or
+
+    make installcheck
+
+NOTE: This creates a temporary installation (in the case of "check"),
+with multiple nodes, be they master or standby(s) for the purpose of
+the tests.
+
+NOTE: This requires the --enable-tap-tests argument to configure.
diff --git a/src/test/checksum/t/001_basic.pl b/src/test/checksum/t/001_basic.pl
new file mode 100644
index 00000000000..3b229de915a
--- /dev/null
+++ b/src/test/checksum/t/001_basic.pl
@@ -0,0 +1,92 @@
+# Test suite for testing enabling data checksums in an online cluster
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+# Initialize node with checksums disabled.
+my $node = get_new_node('main');
+$node->init();
+$node->start();
+
+# Create some content to have un-checksummed data in the cluster
+$node->safe_psql('postgres',
+	"CREATE TABLE t AS SELECT generate_series(1,10000) AS a;");
+
+# Ensure that checksums are turned off
+my $result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'off', 'ensure checksums are disabled');
+
+# No relation in pg_class should have relhaschecksums at this point
+$result = $node->safe_psql('postgres',
+	"SELECT count(*) FROM pg_catalog.pg_class WHERE relhaschecksums;");
+is($result, '0', 'ensure no entries in pg_class has checksums recorded');
+
+# Enable data checksums
+$node->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+
+# Wait for checksums to become enabled
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'on');
+is($result, 1, 'ensure checksums are enabled');
+
+# Check that relations with storage have been marked with relhaschecksums in
+# pg_class
+$result = $node->safe_psql('postgres',
+		"SELECT count(*) FROM pg_catalog.pg_class WHERE NOT relhaschecksums "
+	  . "AND relkind IN ('r', 'i', 'S', 't', 'm');");
+is($result, '0', 'ensure all relations are correctly flagged in the catalog');
+
+# Run a dummy query just to make sure we can read back some data
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure checksummed pages can be read back');
+
+# Enable data checksums again which should be a no-op..
+$node->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+# ..and make sure we can still read/write data
+$node->safe_psql('postgres', "UPDATE t SET a = a + 1;");
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure checksummed pages can be read back');
+
+# Disable checksums again
+$node->safe_psql('postgres', "SELECT pg_disable_data_checksums();");
+
+# Wait for checksums to be disabled. Disabling checksums clear the catalog
+# relhaschecksums state so await that before calling it done.
+$result = $node->poll_query_until('postgres',
+	"SELECT count(*) FROM pg_catalog.pg_class WHERE relhaschecksums;", '0');
+is($result, '1', 'ensure no entries in pg_class has checksums recorded');
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'off');
+is($result, 1, 'ensure checksums are disabled');
+
+# Test reading again
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure previously checksummed pages can be read back');
+
+# Re-enable checksums and make sure that the relhaschecksums flags in the
+# catalog aren't tricking processing into skipping previously checksummed
+# relations
+$node->safe_psql('postgres', "UPDATE t SET a = a + 1;");
+
+$node->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'on');
+is($result, 1, 'ensure checksums are enabled');
+
+# Run a dummy query just to make sure we can read back some data
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure checksummed pages can be read back');
+
+$node->stop;
+
+done_testing();
diff --git a/src/test/checksum/t/002_restarts.pl b/src/test/checksum/t/002_restarts.pl
new file mode 100644
index 00000000000..41a4d640375
--- /dev/null
+++ b/src/test/checksum/t/002_restarts.pl
@@ -0,0 +1,117 @@
+# Test suite for testing enabling data checksums in an online cluster with
+# restarting the processing
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+use IPC::Run qw(pump finish timer);
+
+# If we don't have IO::Pty, forget it, because IPC::Run depends on that
+# to support pty connections
+eval { require IO::Pty; };
+if ($@)
+{
+	plan skip_all => 'IO::Pty is needed to run this test';
+}
+
+# Initialize node with checksums disabled.
+my $node = get_new_node('main');
+$node->init();
+$node->start();
+
+# Create some content to have un-checksummed data in the cluster
+$node->safe_psql('postgres',
+	"CREATE TABLE t AS SELECT generate_series(1,10000) AS a;");
+
+# Ensure that checksums are disabled
+my $result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'off', 'ensure checksums are disabled');
+
+# Create a barrier for checksumming to block on, in this case a pre-existing
+# temporary table which is kept open while processing is started. We can
+# accomplish this by setting up an interactive psql process which keeps the
+# temporary table created as we enable checksums in another psql process.
+my $in    = '';
+my $out   = '';
+my $timer = timer(5);
+
+my $h = $node->interactive_psql('postgres', \$in, \$out, $timer);
+
+$out = '';
+$timer->start(5);
+
+$in .= "CREATE TEMPORARY TABLE tt (a integer);\n";
+pump $h until ($out =~ /CREATE TABLE/ || $timer->is_expired);
+
+# In another session, make sure we can see the blocking temp table but start
+# processing anyways and check that we are blocked with a proper wait event.
+$result = $node->safe_psql('postgres',
+	"SELECT relpersistence FROM pg_catalog.pg_class WHERE relname = 'tt';");
+is($result, 't', 'ensure we can see the temporary table');
+
+$node->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT count(*) FROM pg_catalog.pg_class WHERE NOT relhaschecksums "
+	  . "AND relkind IN ('r', 'i', 'S', 't', 'm');",
+	'1');
+is($result, 1, 'ensure there is a single table left');
+
+$result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'inprogress-on', "ensure checksums aren't enabled yet");
+
+$result = $node->safe_psql('postgres',
+	"SELECT wait_event FROM pg_stat_activity WHERE backend_type = 'datachecksumsworker worker';"
+);
+is($result, 'ChecksumEnableFinishCondition', 'test for correct wait event');
+
+$result = $node->safe_psql('postgres',
+		"SELECT count(*) FROM pg_catalog.pg_class WHERE NOT relhaschecksums "
+	  . "AND relkind IN ('r', 'i', 'S', 't', 'm');");
+is($result, '1',
+	'doublecheck that there is a single table left before restarting');
+
+$node->stop;
+$node->start;
+
+$result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'inprogress-on', "ensure checksums aren't enabled yet");
+
+$result = $node->safe_psql('postgres',
+		"SELECT count(*) FROM pg_catalog.pg_class WHERE NOT relhaschecksums "
+	  . "AND relkind IN ('r', 'i', 'S', 't', 'm');");
+is($result, '0', 'no temporary tables this time around');
+
+$node->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'on');
+is($result, 1, 'ensure checksums are turned on');
+
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure checksummed pages can be read back');
+
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT count(*) FROM pg_stat_activity WHERE backend_type LIKE 'datachecksumsworker%';",
+	'0');
+is($result, 1, 'await datachecksums worker/launcher termination');
+
+$result = $node->safe_psql('postgres', "SELECT pg_disable_data_checksums();");
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'off');
+is($result, 1, 'ensure checksums are turned off');
+
+done_testing();
diff --git a/src/test/checksum/t/003_standby_checksum.pl b/src/test/checksum/t/003_standby_checksum.pl
new file mode 100644
index 00000000000..1555a1694be
--- /dev/null
+++ b/src/test/checksum/t/003_standby_checksum.pl
@@ -0,0 +1,127 @@
+# Test suite for testing enabling data checksums in an online cluster with
+# streaming replication
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+# Initialize primary node
+my $node_primary = get_new_node('primary');
+$node_primary->init(allows_streaming => 1);
+$node_primary->start;
+my $backup_name = 'my_backup';
+
+# Take backup
+$node_primary->backup($backup_name);
+
+# Create streaming standby linking to primary
+my $node_standby_1 = get_new_node('standby_1');
+$node_standby_1->init_from_backup($node_primary, $backup_name,
+	has_streaming => 1);
+$node_standby_1->start;
+
+# Create some content on the primary to have un-checksummed data in the cluster
+$node_primary->safe_psql('postgres',
+	"CREATE TABLE t AS SELECT generate_series(1,10000) AS a;");
+
+# Wait for standbys to catch up
+$node_primary->wait_for_catchup($node_standby_1, 'replay',
+	$node_primary->lsn('insert'));
+
+# Check that checksums are turned off on all nodes
+my $result = $node_primary->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, "off", 'ensure checksums are turned off on primary');
+
+$result = $node_standby_1->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, "off", 'ensure checksums are turned off on standby_1');
+
+# Enable checksums for the cluster
+$node_primary->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+
+# Ensure that the primary switches to "inprogress-on"
+$result = $node_primary->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	"inprogress-on");
+is($result, 1, 'ensure checksums are in progress on primary');
+
+# Wait for checksum enable to be replayed
+$node_primary->wait_for_catchup($node_standby_1, 'replay');
+
+# Ensure that the standby has switched to "inprogress-on" or "on".  Normally it
+# would be "inprogress-on", but it is theoretically possible for the primary to
+# complete the checksum enabling *and* have the standby replay that record
+# before we reach the check below.
+$result = $node_standby_1->poll_query_until(
+	'postgres',
+	"SELECT setting = 'off' FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'f');
+is($result, 1, 'ensure standby has absorbed the inprogress-on barrier');
+$result = $node_standby_1->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+cmp_ok(
+	$result, '~~',
+	[ "inprogress-on", "on" ],
+	'ensure checksums are on, or in progress, on standby_1');
+
+# Insert some more data which should be checksummed on INSERT
+$node_primary->safe_psql('postgres',
+	"INSERT INTO t VALUES (generate_series(1, 10000));");
+
+# Wait for checksums enabled on the primary
+$result = $node_primary->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'on');
+is($result, 1, 'ensure checksums are enabled on the primary');
+
+# Wait for checksums enabled on the standby
+$result = $node_standby_1->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'on');
+is($result, 1, 'ensure checksums are enabled on the standby');
+
+$result = $node_primary->safe_psql('postgres', "SELECT count(a) FROM t");
+is($result, '20000', 'ensure we can safely read all data with checksums');
+
+$result = $node_primary->poll_query_until(
+	'postgres',
+	"SELECT count(*) FROM pg_stat_activity WHERE backend_type LIKE 'datachecksumsworker%';",
+	'0');
+is($result, 1, 'await datachecksums worker/launcher termination');
+
+# Disable checksums and ensure it's propagated to standby and that we can
+# still read all data
+$node_primary->safe_psql('postgres', "SELECT pg_disable_data_checksums();");
+# Wait for checksum disable to be replayed
+$node_primary->wait_for_catchup($node_standby_1, 'replay');
+$result = $node_primary->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'off');
+is($result, 1, 'ensure data checksums are disabled on the primary 2');
+$result = $node_primary->poll_query_until('postgres',
+	"SELECT count(*) FROM pg_catalog.pg_class WHERE relhaschecksums;", '0');
+is($result, '1', 'ensure no entries in pg_class has checksums recorded');
+
+# Ensure that the standby has switched to off
+$result = $node_standby_1->poll_query_until('postgres',
+	"SELECT count(*) FROM pg_catalog.pg_class WHERE relhaschecksums;", '0');
+is($result, '1', 'ensure no entries in pg_class has checksums recorded');
+$result = $node_standby_1->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'off');
+is($result, 1, 'ensure checksums are off on standby_1');
+
+$result = $node_primary->safe_psql('postgres', "SELECT count(a) FROM t");
+is($result, "20000", 'ensure we can safely read all data without checksums');
+
+done_testing();
diff --git a/src/test/checksum/t/004_offline.pl b/src/test/checksum/t/004_offline.pl
new file mode 100644
index 00000000000..2dfca4df235
--- /dev/null
+++ b/src/test/checksum/t/004_offline.pl
@@ -0,0 +1,105 @@
+# Test suite for testing enabling data checksums offline from various states
+# of checksum processing
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+use IPC::Run qw(pump finish timer);
+
+# If we don't have IO::Pty, forget it, because IPC::Run depends on that
+# to support pty connections
+eval { require IO::Pty; };
+if ($@)
+{
+	plan skip_all => 'IO::Pty is needed to run this test';
+}
+
+# Initialize node with checksums disabled.
+my $node = get_new_node('main');
+$node->init();
+$node->start();
+
+# Create some content to have un-checksummed data in the cluster
+$node->safe_psql('postgres',
+	"CREATE TABLE t AS SELECT generate_series(1,10000) AS a;");
+
+# Ensure that checksums are disabled
+my $result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'off', 'ensure checksums are disabled');
+
+# Enable checksums offline using pg_checksums
+$node->stop();
+$node->checksum_enable_offline();
+$node->start();
+
+# Ensure that checksums are enabled
+$result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'on', 'ensure checksums are enabled');
+
+# Run a dummy query just to make sure we can read back some data
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure checksummed pages can be read back');
+
+# Disable checksums offline again using pg_checksums
+$node->stop();
+$node->checksum_disable_offline();
+$node->start();
+
+# Ensure that checksums are disabled
+$result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'off', 'ensure checksums are disabled');
+
+# Create a barrier for checksumming to block on, in this case a pre-existing
+# temporary table which is kept open while processing is started. We can
+# accomplish this by setting up an interactive psql process which keeps the
+# temporary table created as we enable checksums in another psql process.
+my $in    = '';
+my $out   = '';
+my $timer = timer(5);
+
+my $h = $node->interactive_psql('postgres', \$in, \$out, $timer);
+
+$out = '';
+$timer->start(5);
+
+$in .= "CREATE TEMPORARY TABLE tt (a integer);\n";
+pump $h until ($out =~ /CREATE TABLE/ || $timer->is_expired);
+
+# In another session, make sure we can see the blocking temp table but start
+# processing anyways and check that we are blocked with a proper wait event.
+$result = $node->safe_psql('postgres',
+	"SELECT relpersistence FROM pg_catalog.pg_class WHERE relname = 'tt';");
+is($result, 't', 'ensure we can see the temporary table');
+
+$node->safe_psql('postgres', "SELECT pg_enable_data_checksums();");
+
+$result = $node->poll_query_until(
+	'postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';",
+	'inprogress-on');
+is($result, 1, 'ensure checksums are in the process of being enabled');
+
+# Turn the cluster off and enable checksums offline, then start back up
+$node->stop();
+$node->checksum_enable_offline();
+$node->start();
+
+# Ensure that checksums are now enabled even though processing wasn't
+# restarted
+$result = $node->safe_psql('postgres',
+	"SELECT setting FROM pg_catalog.pg_settings WHERE name = 'data_checksums';"
+);
+is($result, 'on', 'ensure checksums are enabled');
+
+# Run a dummy query just to make sure we can read back some data
+$result = $node->safe_psql('postgres', "SELECT count(*) FROM t");
+is($result, '10000', 'ensure checksummed pages can be read back');
+
+done_testing();
diff --git a/src/test/perl/PostgresNode.pm b/src/test/perl/PostgresNode.pm
index 9667f7667ec..b7431a76005 100644
--- a/src/test/perl/PostgresNode.pm
+++ b/src/test/perl/PostgresNode.pm
@@ -2221,6 +2221,42 @@ sub pg_recvlogical_upto
 	}
 }
 
+=item $node->checksum_enable_offline()
+
+Enable data page checksums in an offline cluster with B<pg_checksums>. The
+caller is responsible for ensuring that the cluster is in the right state for
+this operation.
+
+=cut
+
+sub checksum_enable_offline
+{
+	my ($self) = @_;
+
+	print "# Enabling checksums in \"$self->data_dir\"\n";
+	TestLib::system_or_bail('pg_checksums', '-D', $self->data_dir, '-e');
+	print "# Checksums enabled\n";
+	return;
+}
+
+=item checksum_disable_offline
+
+Disable data page checksums in an offline cluster with B<pg_checksums>. The
+caller is responsible for ensuring that the cluster is in the right state for
+this operation.
+
+=cut
+
+sub checksum_disable_offline
+{
+	my ($self) = @_;
+
+	print "# Disabling checksums in \"$self->data_dir\"\n";
+	TestLib::system_or_bail('pg_checksums', '-D', $self->data_dir, '-d');
+	print "# Checksums disabled\n";
+	return;
+}
+
 =pod
 
 =back
-- 
2.29.2


--------------1FF9A53F15934AFFB66C5772--





view thread (27+ messages)  latest in thread

reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients using the --to and --cc options:
  reply via email

  To: [email protected]
  Cc: [email protected]
  Subject: Re: [PATCH v34 1/1] Support checksum enable/disable in a running cluster v33
  In-Reply-To: <no-message-id-1863936@localhost>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox