<!-- doc/src/sgml/runtime.sgml -->

<chapter id="runtime">
 <title>Server Setup and Operation</title>

 <para>
  This chapter discusses how to set up and run the database server,
  and its interactions with the operating system.
 </para>

 <para>
  The directions in this chapter assume that you are working with
  plain <productname>PostgreSQL</productname> without any additional
  infrastructure, for example a copy that you built from source
  according to the directions in the preceding chapters.
  If you are working with a pre-packaged or vendor-supplied
  version of <productname>PostgreSQL</productname>, it is likely that
  the packager has made special provisions for installing and starting
  the database server according to your system's conventions.
  Consult the package-level documentation for details.
 </para>

 <sect1 id="postgres-user">
  <title>The <productname>PostgreSQL</productname> User Account</title>

  <indexterm>
   <primary>postgres user</primary>
  </indexterm>

  <para>
   As with any server daemon that is accessible to the outside world,
   it is advisable to run <productname>PostgreSQL</productname> under a
   separate user account. This user account should only own the data
   that is managed by the server, and should not be shared with other
   daemons. (For example, using the user <literal>nobody</literal> is a bad
   idea.) In particular, it is advisable that this user account not own
   the <productname>PostgreSQL</productname> executable files, to ensure
   that a compromised server process could not modify those executables.
  </para>

  <para>
   Pre-packaged versions of <productname>PostgreSQL</productname> will
   typically create a suitable user account automatically during
   package installation.
  </para>

  <para>
   To add a Unix user account to your system, look for a command
   <command>useradd</command> or <command>adduser</command>. The user
   name <systemitem>postgres</systemitem> is often used, and is assumed
   throughout this book, but you can use another name if you like.
  </para>
 </sect1>

 <sect1 id="creating-cluster">
  <title>Creating a Database Cluster</title>

  <indexterm>
   <primary>database cluster</primary>
  </indexterm>

  <indexterm>
   <primary>data area</primary>
   <see>database cluster</see>
  </indexterm>

  <para>
   Before you can do anything, you must initialize a database storage
   area on disk. We call this a <firstterm>database cluster</firstterm>.
   (The <acronym>SQL</acronym> standard uses the term catalog cluster.) A
   database cluster is a collection of databases that is managed by a
   single instance of a running database server. After initialization, a
   database cluster will contain a database named <literal>postgres</literal>,
   which is meant as a default database for use by utilities, users and third
   party applications.  The database server itself does not require the
   <literal>postgres</literal> database to exist, but many external utility
   programs assume it exists.  Another database created within each cluster
   during initialization is called
   <literal>template1</literal>.  As the name suggests, this will be used
   as a template for subsequently created databases; it should not be
   used for actual work.  (See <xref linkend="managing-databases"/> for
   information about creating new databases within a cluster.)
  </para>

  <para>
   In file system terms, a database cluster is a single directory
   under which all data will be stored. We call this the <firstterm>data
   directory</firstterm> or <firstterm>data area</firstterm>. It is
   completely up to you where you choose to store your data.  There is no
   default, although locations such as
   <filename>/usr/local/pgsql/data</filename> or
   <filename>/var/lib/pgsql/data</filename> are popular.
   The data directory must be initialized before being used, using the program
   <xref linkend="app-initdb"/><indexterm><primary>initdb</primary></indexterm>
   which is installed with <productname>PostgreSQL</productname>.
  </para>

  <para>
   If you are using a pre-packaged version
   of <productname>PostgreSQL</productname>, it may well have a specific
   convention for where to place the data directory, and it may also
   provide a script for creating the data directory.  In that case you
   should use that script in preference to
   running <command>initdb</command> directly.
   Consult the package-level documentation for details.
  </para>

  <para>
   To initialize a database cluster manually,
   run <command>initdb</command> and specify the desired
   file system location of the database cluster with the
   <option>-D</option> option, for example:
<screen>
<prompt>$</prompt> <userinput>initdb -D /usr/local/pgsql/data</userinput>
</screen>
   Note that you must execute this command while logged into the
   <productname>PostgreSQL</productname> user account, which is
   described in the previous section.
  </para>

  <tip>
   <para>
    As an alternative to the <option>-D</option> option, you can set
    the environment variable <envar>PGDATA</envar>.
    <indexterm><primary><envar>PGDATA</envar></primary></indexterm>
   </para>
  </tip>

  <para>
   Alternatively, you can run <command>initdb</command> via
   the <xref linkend="app-pg-ctl"/>
   program<indexterm><primary>pg_ctl</primary></indexterm> like so:
<screen>
<prompt>$</prompt> <userinput>pg_ctl -D /usr/local/pgsql/data initdb</userinput>
</screen>
   This may be more intuitive if you are
   using <command>pg_ctl</command> for starting and stopping the
   server (see <xref linkend="server-start"/>), so
   that <command>pg_ctl</command> would be the sole command you use
   for managing the database server instance.
  </para>

  <para>
   <command>initdb</command> will attempt to create the directory you
   specify if it does not already exist.  Of course, this will fail if
   <command>initdb</command> does not have permissions to write in the
   parent directory.  It's generally recommendable that the
   <productname>PostgreSQL</productname> user own not just the data
   directory but its parent directory as well, so that this should not
   be a problem.  If the desired parent directory doesn't exist either,
   you will need to create it first, using root privileges if the
   grandparent directory isn't writable.  So the process might look
   like this:
<screen>
root# <userinput>mkdir /usr/local/pgsql</userinput>
root# <userinput>chown postgres /usr/local/pgsql</userinput>
root# <userinput>su postgres</userinput>
postgres$ <userinput>initdb -D /usr/local/pgsql/data</userinput>
</screen>
  </para>

  <para>
   <command>initdb</command> will refuse to run if the data directory
   exists and already contains files; this is to prevent accidentally
   overwriting an existing installation.
  </para>

  <para>
   Because the data directory contains all the data stored in the
   database, it is essential that it be secured from unauthorized
   access. <command>initdb</command> therefore revokes access
   permissions from everyone but the
   <productname>PostgreSQL</productname> user, and optionally, group.
   Group access, when enabled, is read-only.  This allows an unprivileged
   user in the same group as the cluster owner to take a backup of the
   cluster data or perform other operations that only require read access.
  </para>

  <para>
   Note that enabling or disabling group access on an existing cluster requires
   the cluster to be shut down and the appropriate mode to be set on all
   directories and files before restarting
   <productname>PostgreSQL</productname>.  Otherwise, a mix of modes might
   exist in the data directory.  For clusters that allow access only by the
   owner, the appropriate modes are <literal>0700</literal> for directories
   and <literal>0600</literal> for files.  For clusters that also allow
   reads by the group, the appropriate modes are <literal>0750</literal>
   for directories and <literal>0640</literal> for files.
  </para>

  <para>
   However, while the directory contents are secure, the default
   client authentication setup allows any local user to connect to the
   database and even become the database superuser. If you do not
   trust other local users, we recommend you use one of
   <command>initdb</command>'s <option>-W</option>, <option>--pwprompt</option>
   or <option>--pwfile</option> options to assign a password to the
   database superuser.<indexterm>
     <primary>password</primary>
     <secondary>of the superuser</secondary>
   </indexterm>
   Also, specify <option>-A scram-sha-256</option>
   so that the default <literal>trust</literal> authentication
   mode is not used; or modify the generated <filename>pg_hba.conf</filename>
   file after running <command>initdb</command>, but
   <emphasis>before</emphasis> you start the server for the first time. (Other
   reasonable approaches include using <literal>peer</literal> authentication
   or file system permissions to restrict connections. See <xref
   linkend="client-authentication"/> for more information.)
  </para>

  <para>
   <command>initdb</command> also initializes the default
   locale<indexterm><primary>locale</primary></indexterm> for the database cluster.
   Normally, it will just take the locale settings in the environment
   and apply them to the initialized database.  It is possible to
   specify a different locale for the database; more information about
   that can be found in <xref linkend="locale"/>.  The default sort order used
   within the particular database cluster is set by
   <command>initdb</command>, and while you can create new databases using
   different sort order, the order used in the template databases that initdb
   creates cannot be changed without dropping and recreating them.
   There is also a performance impact for using locales
   other than <literal>C</literal> or <literal>POSIX</literal>. Therefore, it is
   important to make this choice correctly the first time.
  </para>

  <para>
   <command>initdb</command> also sets the default character set encoding
   for the database cluster.  Normally this should be chosen to match the
   locale setting.  For details see <xref linkend="multibyte"/>.
  </para>

  <para>
   Non-<literal>C</literal> and non-<literal>POSIX</literal> locales rely on the
   operating system's collation library for character set ordering.
   This controls the ordering of keys stored in indexes.  For this reason,
   a cluster cannot switch to an incompatible collation library version,
   either through snapshot restore, binary streaming replication, a
   different operating system, or an operating system upgrade.
  </para>

  <sect2 id="creating-cluster-mount-points">
   <title>Use of Secondary File Systems</title>

   <indexterm zone="creating-cluster-mount-points">
    <primary>file system mount points</primary>
   </indexterm>

   <para>
    Many installations create their database clusters on file systems
    (volumes) other than the machine's <quote>root</quote> volume.  If you
    choose to do this, it is not advisable to try to use the secondary
    volume's topmost directory (mount point) as the data directory.
    Best practice is to create a directory within the mount-point
    directory that is owned by the <productname>PostgreSQL</productname>
    user, and then create the data directory within that.  This avoids
    permissions problems, particularly for operations such
    as <application>pg_upgrade</application>, and it also ensures clean failures if
    the secondary volume is taken offline.
   </para>

  </sect2>

  <sect2 id="creating-cluster-filesystem">
   <title>File Systems</title>

   <para>
    Generally, any file system with POSIX semantics can be used for
    PostgreSQL.  Users prefer different file systems for a variety of reasons,
    including vendor support, performance, and familiarity.  Experience
    suggests that, all other things being equal, one should not expect major
    performance or behavior changes merely from switching file systems or
    making minor file system configuration changes.
   </para>

   <sect3 id="creating-cluster-nfs">
    <title>NFS</title>

    <indexterm zone="creating-cluster-nfs">
     <primary>NFS</primary>
    </indexterm>

    <para>
     It is possible to use an <acronym>NFS</acronym> file system for storing
     the <productname>PostgreSQL</productname> data directory.
     <productname>PostgreSQL</productname> does nothing special for
     <acronym>NFS</acronym> file systems, meaning it assumes
     <acronym>NFS</acronym> behaves exactly like locally-connected drives.
     <productname>PostgreSQL</productname> does not use any functionality that
     is known to have nonstandard behavior on <acronym>NFS</acronym>, such as
     file locking.
    </para>

    <para>
     The only firm requirement for using <acronym>NFS</acronym> with
     <productname>PostgreSQL</productname> is that the file system is mounted
     using the <literal>hard</literal> option.  With the
     <literal>hard</literal> option, processes can <quote>hang</quote>
     indefinitely if there are network problems, so this configuration will
     require a careful monitoring setup.  The <literal>soft</literal> option
     will interrupt system calls in case of network problems, but
     <productname>PostgreSQL</productname> will not repeat system calls
     interrupted in this way, so any such interruption will result in an I/O
     error being reported.
    </para>

    <para>
     It is not necessary to use the <literal>sync</literal> mount option.  The
     behavior of the <literal>async</literal> option is sufficient, since
     <productname>PostgreSQL</productname> issues <literal>fsync</literal>
     calls at appropriate times to flush the write caches.  (This is analogous
     to how it works on a local file system.)  However, it is strongly
     recommended to use the <literal>sync</literal> export option on the NFS
     <emphasis>server</emphasis> on systems where it exists (mainly Linux).
     Otherwise, an <literal>fsync</literal> or equivalent on the NFS client is
     not actually guaranteed to reach permanent storage on the server, which
     could cause corruption similar to running with the parameter <xref
     linkend="guc-fsync"/> off.  The defaults of these mount and export
     options differ between vendors and versions, so it is recommended to
     check and perhaps specify them explicitly in any case to avoid any
     ambiguity.
    </para>

    <para>
     In some cases, an external storage product can be accessed either via NFS
     or a lower-level protocol such as iSCSI.  In the latter case, the storage
     appears as a block device and any available file system can be created on
     it.  That approach might relieve the DBA from having to deal with some of
     the idiosyncrasies of NFS, but of course the complexity of managing
     remote storage then happens at other levels.
    </para>
   </sect3>
  </sect2>

 </sect1>

 <sect1 id="server-start">
  <title>Starting the Database Server</title>

  <para>
   Before anyone can access the database, you must start the database
   server. The database server program is called
   <command>postgres</command>.<indexterm><primary>postgres</primary></indexterm>
  </para>

  <para>
   If you are using a pre-packaged version
   of <productname>PostgreSQL</productname>, it almost certainly includes
   provisions for running the server as a background task according to the
   conventions of your operating system.  Using the package's
   infrastructure to start the server will be much less work than figuring
   out how to do this yourself.  Consult the package-level documentation
   for details.
  </para>

  <para>
   The bare-bones way to start the server manually is just to invoke
   <command>postgres</command> directly, specifying the location of the
   data directory with the <option>-D</option> option, for example:
<screen>
$ <userinput>postgres -D /usr/local/pgsql/data</userinput>
</screen>
   which will leave the server running in the foreground. This must be
   done while logged into the <productname>PostgreSQL</productname> user
   account. Without <option>-D</option>, the server will try to use
   the data directory named by the environment variable <envar>PGDATA</envar>.
   If that variable is not provided either, it will fail.
  </para>

  <para>
   Normally it is better to start <command>postgres</command> in the
   background.  For this, use the usual Unix shell syntax:
<screen>
$ <userinput>postgres -D /usr/local/pgsql/data &gt;logfile 2&gt;&amp;1 &amp;</userinput>
</screen>
   It is important to store the server's <systemitem>stdout</systemitem> and
   <systemitem>stderr</systemitem> output somewhere, as shown above. It will help
   for auditing purposes and to diagnose problems. (See <xref
   linkend="logfile-maintenance"/> for a more thorough discussion of log
   file handling.)
  </para>

  <para>
   The <command>postgres</command> program also takes a number of other
   command-line options. For more information, see the
   <xref linkend="app-postgres"/> reference page
   and <xref linkend="runtime-config"/> below.
  </para>

  <para>
   This shell syntax can get tedious quickly.  Therefore the wrapper
   program
   <xref linkend="app-pg-ctl"/><indexterm><primary>pg_ctl</primary></indexterm>
   is provided to simplify some tasks.  For example:
<programlisting>
pg_ctl start -l logfile
</programlisting>
   will start the server in the background and put the output into the
   named log file. The <option>-D</option> option has the same meaning
   here as for <command>postgres</command>. <command>pg_ctl</command>
   is also capable of stopping the server.
  </para>

  <para>
   Normally, you will want to start the database server when the
   computer boots.<indexterm>
     <primary>booting</primary>
     <secondary>starting the server during</secondary>
   </indexterm>
   Autostart scripts are operating-system-specific.
   There are a few example scripts distributed with
   <productname>PostgreSQL</productname> in the
   <filename>contrib/start-scripts</filename> directory. Installing one will require
   root privileges.
  </para>

  <para>
   Different systems have different conventions for starting up daemons
   at boot time. Many systems have a file
   <filename>/etc/rc.local</filename> or
   <filename>/etc/rc.d/rc.local</filename>. Others use <filename>init.d</filename> or
   <filename>rc.d</filename> directories. Whatever you do, the server must be
   run by the <productname>PostgreSQL</productname> user account
   <emphasis>and not by root</emphasis> or any other user. Therefore you
   probably should form your commands using
   <literal>su postgres -c '...'</literal>.  For example:
<programlisting>
su postgres -c 'pg_ctl start -D /usr/local/pgsql/data -l serverlog'
</programlisting>
  </para>

  <para>
   Here are a few more operating-system-specific suggestions. (In each
   case be sure to use the proper installation directory and user
   name where we show generic values.)

   <itemizedlist>
    <listitem>
     <para>
      For <productname>FreeBSD</productname>, look at the file
      <filename>contrib/start-scripts/freebsd</filename> in the
      <productname>PostgreSQL</productname> source distribution.
      <indexterm><primary>FreeBSD</primary><secondary>start script</secondary></indexterm>
     </para>
    </listitem>

    <listitem>
     <para>
      On <productname>OpenBSD</productname>, add the following lines
      to the file <filename>/etc/rc.local</filename>:
      <indexterm><primary>OpenBSD</primary><secondary>start script</secondary></indexterm>
<programlisting>
if [ -x /usr/local/pgsql/bin/pg_ctl -a -x /usr/local/pgsql/bin/postgres ]; then
    su -l postgres -c '/usr/local/pgsql/bin/pg_ctl start -s -l /var/postgresql/log -D /usr/local/pgsql/data'
    echo -n ' postgresql'
fi
</programlisting>
     </para>
    </listitem>

    <listitem>
     <para>
      On <productname>Linux</productname> systems either add
      <indexterm><primary>Linux</primary><secondary>start script</secondary></indexterm>
<programlisting>
/usr/local/pgsql/bin/pg_ctl start -l logfile -D /usr/local/pgsql/data
</programlisting>
      to <filename>/etc/rc.d/rc.local</filename>
      or <filename>/etc/rc.local</filename> or look at the file
      <filename>contrib/start-scripts/linux</filename> in the
      <productname>PostgreSQL</productname> source distribution.
     </para>

     <para>
      When using <application>systemd</application>, you can use the following
      service unit file (e.g.,
      at <filename>/etc/systemd/system/postgresql.service</filename>):<indexterm><primary>systemd</primary></indexterm>
<programlisting>
[Unit]
Description=PostgreSQL database server
Documentation=man:postgres(1)

[Service]
Type=notify
User=postgres
ExecStart=/usr/local/pgsql/bin/postgres -D /usr/local/pgsql/data
ExecReload=/bin/kill -HUP $MAINPID
KillMode=mixed
KillSignal=SIGINT
TimeoutSec=infinity

[Install]
WantedBy=multi-user.target
</programlisting>
      Using <literal>Type=notify</literal> requires that the server binary was
      built with <literal>configure --with-systemd</literal>.
     </para>

     <para>
      Consider carefully the timeout
      setting.  <application>systemd</application> has a default timeout of 90
      seconds as of this writing and will kill a process that does not report
      readiness within that time.  But a <productname>PostgreSQL</productname>
      server that might have to perform crash recovery at startup could take
      much longer to become ready.  The suggested value
      of <literal>infinity</literal> disables the timeout logic.
     </para>
    </listitem>

    <listitem>
     <para>
      On <productname>NetBSD</productname>, use either the
      <productname>FreeBSD</productname> or
      <productname>Linux</productname> start scripts, depending on
      preference.
      <indexterm><primary>NetBSD</primary><secondary>start script</secondary></indexterm>
     </para>
    </listitem>

    <listitem>
     <para>
      On <productname>Solaris</productname>, create a file called
      <filename>/etc/init.d/postgresql</filename> that contains
      the following line:
      <indexterm><primary>Solaris</primary><secondary>start script</secondary></indexterm>
<programlisting>
su - postgres -c "/usr/local/pgsql/bin/pg_ctl start -l logfile -D /usr/local/pgsql/data"
</programlisting>
      Then, create a symbolic link to it in <filename>/etc/rc3.d</filename> as
      <filename>S99postgresql</filename>.
     </para>
    </listitem>
   </itemizedlist>

  </para>

   <para>
    While the server is running, its
    <acronym>PID</acronym> is stored in the file
    <filename>postmaster.pid</filename> in the data directory. This is
    used to prevent multiple server instances from
    running in the same data directory and can also be used for
    shutting down the server.
   </para>

   <sect2 id="server-start-failures">
    <title>Server Start-up Failures</title>

    <para>
     There are several common reasons the server might fail to
     start. Check the server's log file, or start it by hand (without
     redirecting standard output or standard error) and see what error
     messages appear. Below we explain some of the most common error
     messages in more detail.
    </para>

    <para>
<screen>
LOG:  could not bind IPv4 address "127.0.0.1": Address already in use
HINT:  Is another postmaster already running on port 5432? If not, wait a few seconds and retry.
FATAL:  could not create any TCP/IP sockets
</screen>
     This usually means just what it suggests: you tried to start
     another server on the same port where one is already running.
     However, if the kernel error message is not <computeroutput>Address
     already in use</computeroutput> or some variant of that, there might
     be a different problem. For example, trying to start a server
     on a reserved port number might draw something like:
<screen>
$ <userinput>postgres -p 666</userinput>
LOG:  could not bind IPv4 address "127.0.0.1": Permission denied
HINT:  Is another postmaster already running on port 666? If not, wait a few seconds and retry.
FATAL:  could not create any TCP/IP sockets
</screen>
    </para>

    <para>
     A message like:
<screen>
FATAL:  could not create shared memory segment: Invalid argument
DETAIL:  Failed system call was shmget(key=5440001, size=4011376640, 03600).
</screen>
     probably means your kernel's limit on the size of shared memory is
     smaller than the work area <productname>PostgreSQL</productname>
     is trying to create (4011376640 bytes in this example).
     This is only likely to happen if you have set <literal>shared_memory_type</literal>
     to <literal>sysv</literal>.  In that case, you
     can try starting the server with a smaller-than-normal number of
     buffers (<xref linkend="guc-shared-buffers"/>), or
     reconfigure your kernel to increase the allowed shared memory
     size. You might also see this message when trying to start multiple
     servers on the same machine, if their total space requested
     exceeds the kernel limit.
    </para>

    <para>
     An error like:
<screen>
FATAL:  could not create semaphores: No space left on device
DETAIL:  Failed system call was semget(5440126, 17, 03600).
</screen>
     does <emphasis>not</emphasis> mean you've run out of disk
     space. It means your kernel's limit on the number of <systemitem
     class="osname">System V</systemitem> semaphores is smaller than the number
     <productname>PostgreSQL</productname> wants to create. As above,
     you might be able to work around the problem by starting the
     server with a reduced number of allowed connections
     (<xref linkend="guc-max-connections"/>), but you'll eventually want to
     increase the kernel limit.
    </para>

    <para>
     Details about configuring <systemitem class="osname">System V</systemitem>
     <acronym>IPC</acronym> facilities are given in <xref linkend="sysvipc"/>.
    </para>
   </sect2>

   <sect2 id="client-connection-problems">
    <title>Client Connection Problems</title>

    <para>
     Although the error conditions possible on the client side are quite
     varied and application-dependent, a few of them might be directly
     related to how the server was started. Conditions other than
     those shown below should be documented with the respective client
     application.
    </para>

    <para>
<screen>
psql: error: connection to server at "server.joe.com" (123.123.123.123), port 5432 failed: Connection refused
        Is the server running on that host and accepting TCP/IP connections?
</screen>
     This is the generic <quote>I couldn't find a server to talk
     to</quote> failure. It looks like the above when TCP/IP
     communication is attempted. A common mistake is to forget to
     configure the server to allow TCP/IP connections.
    </para>

    <para>
     Alternatively, you might get this when attempting Unix-domain socket
     communication to a local server:
<screen>
psql: error: connection to server on socket "/tmp/.s.PGSQL.5432" failed: No such file or directory
        Is the server running locally and accepting connections on that socket?
</screen>
     If the server is indeed running, check that the client's idea of the
     socket path (here <literal>/tmp</literal>) agrees with the server's
     <xref linkend="guc-unix-socket-directories"/> setting.
    </para>

    <para>
     A connection failure message always shows the server address or socket
     path name, which is useful in verifying that the client is trying to
     connect to the right place. If there is in fact no server
     listening there, the kernel error message will typically be either
     <computeroutput>Connection refused</computeroutput> or
     <computeroutput>No such file or directory</computeroutput>, as
     illustrated. (It is important to realize that
     <computeroutput>Connection refused</computeroutput> in this context
     does <emphasis>not</emphasis> mean that the server got your
     connection request and rejected it. That case will produce a
     different message, as shown in <xref
     linkend="client-authentication-problems"/>.) Other error messages
     such as <computeroutput>Connection timed out</computeroutput> might
     indicate more fundamental problems, like lack of network
     connectivity, or a firewall blocking the connection.
    </para>
   </sect2>
  </sect1>

 <sect1 id="kernel-resources">
  <title>Managing Kernel Resources</title>

  <para>
   <productname>PostgreSQL</productname> can sometimes exhaust various operating system
   resource limits, especially when multiple copies of the server are running
   on the same system, or in very large installations.  This section explains
   the kernel resources used by <productname>PostgreSQL</productname> and the steps you
   can take to resolve problems related to kernel resource consumption.
  </para>

  <sect2 id="sysvipc">
   <title>Shared Memory and Semaphores</title>

   <indexterm zone="sysvipc">
    <primary>shared memory</primary>
   </indexterm>

   <indexterm zone="sysvipc">
    <primary>semaphores</primary>
   </indexterm>

   <para>
    <productname>PostgreSQL</productname> requires the operating system to provide
    inter-process communication (<acronym>IPC</acronym>) features, specifically
    shared memory and semaphores.  Unix-derived systems typically provide
    <quote><systemitem class="osname">System V</systemitem></quote> <acronym>IPC</acronym>,
    <quote><systemitem class="osname">POSIX</systemitem></quote> <acronym>IPC</acronym>, or both.
    <systemitem class="osname">Windows</systemitem> has its own implementation of
    these features and is not discussed here.
   </para>

   <para>
    By default, <productname>PostgreSQL</productname> allocates
    a very small amount of System V shared memory, as well as a much larger
    amount of anonymous <function>mmap</function> shared memory.
    Alternatively, a single large System V shared memory region can be used
    (see <xref linkend="guc-shared-memory-type"/>).

    In addition a significant number of semaphores, which can be either
    System V or POSIX style, are created at server startup.  Currently,
    POSIX semaphores are used on Linux and FreeBSD systems while other
    platforms use System V semaphores.
   </para>

   <para>
    System V <acronym>IPC</acronym> features are typically constrained by
    system-wide allocation limits.
    When <productname>PostgreSQL</productname> exceeds one of these limits,
    the server will refuse to start and
    should leave an instructive error message describing the problem
    and what to do about it. (See also <xref
    linkend="server-start-failures"/>.) The relevant kernel
    parameters are named consistently across different systems; <xref
    linkend="sysvipc-parameters"/> gives an overview. The methods to set
    them, however, vary. Suggestions for some platforms are given below.
   </para>

   <table id="sysvipc-parameters">
    <title><systemitem class="osname">System V</systemitem> <acronym>IPC</acronym> Parameters</title>

    <tgroup cols="3">
     <colspec colname="col1" colwidth="1*"/>
     <colspec colname="col2" colwidth="3*"/>
     <colspec colname="col3" colwidth="3*"/>
     <thead>
      <row>
       <entry>Name</entry>
       <entry>Description</entry>
       <entry>Values needed to run one <productname>PostgreSQL</productname> instance</entry>
      </row>
     </thead>

     <tbody>
      <row>
       <entry><varname>SHMMAX</varname></entry>
       <entry>Maximum size of shared memory segment (bytes)</entry>
       <entry>at least 1kB, but the default is usually much higher</entry>
      </row>

      <row>
       <entry><varname>SHMMIN</varname></entry>
       <entry>Minimum size of shared memory segment (bytes)</entry>
       <entry>1</entry>
      </row>

      <row>
       <entry><varname>SHMALL</varname></entry>
       <entry>Total amount of shared memory available (bytes or pages)</entry>
       <entry>same as <varname>SHMMAX</varname> if bytes,
        or <literal>ceil(SHMMAX/PAGE_SIZE)</literal> if pages,
        plus room for other applications</entry>
      </row>

      <row>
       <entry><varname>SHMSEG</varname></entry>
       <entry>Maximum number of shared memory segments per process</entry>
       <entry>only 1 segment is needed, but the default is much higher</entry>
      </row>

       <row>
        <entry><varname>SHMMNI</varname></entry>
        <entry>Maximum number of shared memory segments system-wide</entry>
        <entry>like <varname>SHMSEG</varname> plus room for other applications</entry>
       </row>

       <row>
        <entry><varname>SEMMNI</varname></entry>
        <entry>Maximum number of semaphore identifiers (i.e., sets)</entry>
        <entry>at least <literal>ceil((max_connections + autovacuum_max_workers + max_wal_senders + max_worker_processes + 5) / 16)</literal> plus room for other applications</entry>
       </row>

       <row>
        <entry><varname>SEMMNS</varname></entry>
        <entry>Maximum number of semaphores system-wide</entry>
        <entry><literal>ceil((max_connections + autovacuum_max_workers + max_wal_senders + max_worker_processes + 5) / 16) * 17</literal> plus room for other applications</entry>
       </row>

       <row>
        <entry><varname>SEMMSL</varname></entry>
        <entry>Maximum number of semaphores per set</entry>
        <entry>at least 17</entry>
       </row>

       <row>
        <entry><varname>SEMMAP</varname></entry>
        <entry>Number of entries in semaphore map</entry>
        <entry>see text</entry>
       </row>

       <row>
        <entry><varname>SEMVMX</varname></entry>
        <entry>Maximum value of semaphore</entry>
        <entry>at least 1000 (The default is often 32767; do not change unless necessary)</entry>
       </row>

     </tbody>
    </tgroup>
   </table>

   <para>
    <productname>PostgreSQL</productname> requires a few bytes of System V shared memory
    (typically 48 bytes, on 64-bit platforms) for each copy of the server.
    On most modern operating systems, this amount can easily be allocated.
    However, if you are running many copies of the server or you explicitly
    configure the server to use large amounts of System V shared memory (see
    <xref linkend="guc-shared-memory-type"/> and <xref
    linkend="guc-dynamic-shared-memory-type"/>), it may be necessary to
    increase <varname>SHMALL</varname>, which is the total amount of System V shared
    memory system-wide.  Note that <varname>SHMALL</varname> is measured in pages
    rather than bytes on many systems.
   </para>

   <para>
    Less likely to cause problems is the minimum size for shared
    memory segments (<varname>SHMMIN</varname>), which should be at most
    approximately 32 bytes for <productname>PostgreSQL</productname> (it is
    usually just 1). The maximum number of segments system-wide
    (<varname>SHMMNI</varname>) or per-process (<varname>SHMSEG</varname>) are unlikely
    to cause a problem unless your system has them set to zero.
   </para>

   <para>
    When using System V semaphores,
    <productname>PostgreSQL</productname> uses one semaphore per allowed connection
    (<xref linkend="guc-max-connections"/>), allowed autovacuum worker process
    (<xref linkend="guc-autovacuum-max-workers"/>) and allowed background
    process (<xref linkend="guc-max-worker-processes"/>), in sets of 16.
    Each such set will
    also contain a 17th semaphore which contains a <quote>magic
    number</quote>, to detect collision with semaphore sets used by
    other applications. The maximum number of semaphores in the system
    is set by <varname>SEMMNS</varname>, which consequently must be at least
    as high as <varname>max_connections</varname> plus
    <varname>autovacuum_max_workers</varname> plus <varname>max_wal_senders</varname>,
    plus <varname>max_worker_processes</varname>, plus one extra for each 16
    allowed connections plus workers (see the formula in <xref
    linkend="sysvipc-parameters"/>).  The parameter <varname>SEMMNI</varname>
    determines the limit on the number of semaphore sets that can
    exist on the system at one time.  Hence this parameter must be at
    least <literal>ceil((max_connections + autovacuum_max_workers + max_wal_senders + max_worker_processes + 5) / 16)</literal>.
    Lowering the number
    of allowed connections is a temporary workaround for failures,
    which are usually confusingly worded <quote>No space
    left on device</quote>, from the function <function>semget</function>.
   </para>

   <para>
    In some cases it might also be necessary to increase
    <varname>SEMMAP</varname> to be at least on the order of
    <varname>SEMMNS</varname>.  If the system has this parameter
    (many do not), it defines the size of the semaphore
    resource map, in which each contiguous block of available semaphores
    needs an entry. When a semaphore set is freed it is either added to
    an existing entry that is adjacent to the freed block or it is
    registered under a new map entry. If the map is full, the freed
    semaphores get lost (until reboot). Fragmentation of the semaphore
    space could over time lead to fewer available semaphores than there
    should be.
   </para>

   <para>
    Various other settings related to <quote>semaphore undo</quote>, such as
    <varname>SEMMNU</varname> and <varname>SEMUME</varname>, do not affect
    <productname>PostgreSQL</productname>.
   </para>

   <para>
    When using POSIX semaphores, the number of semaphores needed is the
    same as for System V, that is one semaphore per allowed connection
    (<xref linkend="guc-max-connections"/>), allowed autovacuum worker process
    (<xref linkend="guc-autovacuum-max-workers"/>) and allowed background
    process (<xref linkend="guc-max-worker-processes"/>).
    On the platforms where this option is preferred, there is no specific
    kernel limit on the number of POSIX semaphores.
   </para>


    <variablelist>
     <varlistentry>
      <term><systemitem class="osname">AIX</systemitem>
      <indexterm><primary>AIX</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        It should not be necessary to do
        any special configuration for such parameters as
        <varname>SHMMAX</varname>, as it appears this is configured to
        allow all memory to be used as shared memory.  That is the
        sort of configuration commonly used for other databases such
        as <application>DB/2</application>.</para>

       <para> It might, however, be necessary to modify the global
       <command>ulimit</command> information in
       <filename>/etc/security/limits</filename>, as the default hard
       limits for file sizes (<varname>fsize</varname>) and numbers of
       files (<varname>nofiles</varname>) might be too low.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">FreeBSD</systemitem>
      <indexterm><primary>FreeBSD</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        The default shared memory settings are usually good enough, unless
        you have set <literal>shared_memory_type</literal> to <literal>sysv</literal>.
        System V semaphores are not used on this platform.
       </para>

       <para>
        The default IPC settings can be changed using
        the <command>sysctl</command> or
        <command>loader</command> interfaces.  The following
        parameters can be set using <command>sysctl</command>:
<screen>
<prompt>#</prompt> <userinput>sysctl kern.ipc.shmall=32768</userinput>
<prompt>#</prompt> <userinput>sysctl kern.ipc.shmmax=134217728</userinput>
</screen>
        To make these settings persist over reboots, modify
        <filename>/etc/sysctl.conf</filename>.
       </para>

       <para>
        If you have set <literal>shared_memory_type</literal> to
        <literal>sysv</literal>, you might also want to configure your kernel
        to lock System V shared memory into RAM and prevent it from being paged
        out to swap.  This can be accomplished using the <command>sysctl</command>
        setting <literal>kern.ipc.shm_use_phys</literal>.
       </para>

       <para>
        If running in a FreeBSD jail, you should set its
        <literal>sysvshm</literal> parameter to <literal>new</literal>, so that
        it has its own separate System V shared memory namespace.
        (Before FreeBSD 11.0, it was necessary to enable shared access to
        the host's IPC namespace from jails, and take measures to avoid
        collisions.)
       </para>

      </listitem>
     </varlistentry>

     <varlistentry>
      <term><systemitem class="osname">NetBSD</systemitem>
      <indexterm><primary>NetBSD</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        The default shared memory settings are usually good enough, unless
        you have set <literal>shared_memory_type</literal> to <literal>sysv</literal>.
        You will usually want to increase <literal>kern.ipc.semmni</literal>
        and <literal>kern.ipc.semmns</literal>,
        as <systemitem class="osname">NetBSD</systemitem>'s default settings
        for these are uncomfortably small.
       </para>

       <para>
        IPC parameters can be adjusted using <command>sysctl</command>,
        for example:
<screen>
<prompt>#</prompt> <userinput>sysctl -w kern.ipc.semmni=100</userinput>
</screen>
        To make these settings persist over reboots, modify
        <filename>/etc/sysctl.conf</filename>.
       </para>

       <para>
        If you have set <literal>shared_memory_type</literal> to
        <literal>sysv</literal>, you might also want to configure your kernel
        to lock System V shared memory into RAM and prevent it from being paged
        out to swap.  This can be accomplished using the <command>sysctl</command>
        setting <literal>kern.ipc.shm_use_phys</literal>.
       </para>
      </listitem>
     </varlistentry>

     <varlistentry>
      <term><systemitem class="osname">OpenBSD</systemitem>
      <indexterm><primary>OpenBSD</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        The default shared memory settings are usually good enough, unless
        you have set <literal>shared_memory_type</literal> to <literal>sysv</literal>.
        You will usually want to
        increase <literal>kern.seminfo.semmni</literal>
        and <literal>kern.seminfo.semmns</literal>,
        as <systemitem class="osname">OpenBSD</systemitem>'s default settings
        for these are uncomfortably small.
       </para>

       <para>
        IPC parameters can be adjusted using <command>sysctl</command>,
        for example:
<screen>
<prompt>#</prompt> <userinput>sysctl kern.seminfo.semmni=100</userinput>
</screen>
        To make these settings persist over reboots, modify
        <filename>/etc/sysctl.conf</filename>.
       </para>

      </listitem>
     </varlistentry>

     <varlistentry>
      <term><systemitem class="osname">HP-UX</systemitem>
      <indexterm><primary>HP-UX</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        The default settings tend to suffice for normal installations.
       </para>
       <para>
        <acronym>IPC</acronym> parameters can be set in the <application>System
        Administration Manager</application> (<acronym>SAM</acronym>) under
        <menuchoice><guimenu>Kernel
        Configuration</guimenu><guimenuitem>Configurable Parameters</guimenuitem></menuchoice>. Choose
        <guibutton>Create A New Kernel</guibutton> when you're done.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">Linux</systemitem>
      <indexterm><primary>Linux</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        The default shared memory settings are usually good enough, unless
        you have set <literal>shared_memory_type</literal> to <literal>sysv</literal>,
        and even then only on older kernel versions that shipped with low defaults.
        System V semaphores are not used on this platform.
       </para>

       <para>
        The shared memory size settings can be changed via the
        <command>sysctl</command> interface.  For example, to allow 16 GB:
<screen>
<prompt>$</prompt> <userinput>sysctl -w kernel.shmmax=17179869184</userinput>
<prompt>$</prompt> <userinput>sysctl -w kernel.shmall=4194304</userinput>
</screen>
        To make these settings persist over reboots, see
        <filename>/etc/sysctl.conf</filename>.
       </para>

      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">macOS</systemitem>
      <indexterm><primary>macOS</primary><secondary>IPC configuration</secondary></indexterm>
      </term>
      <listitem>
       <para>
        The default shared memory and semaphore settings are usually good enough, unless
        you have set <literal>shared_memory_type</literal> to <literal>sysv</literal>.
       </para>
       <para>
        The recommended method for configuring shared memory in macOS
        is to create a file named <filename>/etc/sysctl.conf</filename>,
        containing variable assignments such as:
<programlisting>
kern.sysv.shmmax=4194304
kern.sysv.shmmin=1
kern.sysv.shmmni=32
kern.sysv.shmseg=8
kern.sysv.shmall=1024
</programlisting>
        Note that in some macOS versions,
        <emphasis>all five</emphasis> shared-memory parameters must be set in
        <filename>/etc/sysctl.conf</filename>, else the values will be ignored.
       </para>

       <para>
        <varname>SHMMAX</varname> can only be set to a multiple of 4096.
       </para>

       <para>
        <varname>SHMALL</varname> is measured in 4 kB pages on this platform.
       </para>

       <para>
        It is possible to change all but <varname>SHMMNI</varname> on the fly, using
        <application>sysctl</application>.  But it's still best to set up your preferred
        values via <filename>/etc/sysctl.conf</filename>, so that the values will be
        kept across reboots.
       </para>

      </listitem>
     </varlistentry>

     <varlistentry>
      <term><systemitem class="osname">Solaris</systemitem></term>
      <term><systemitem class="osname">illumos</systemitem></term>
      <listitem>
       <para>
        The default shared memory and semaphore settings are usually good enough for most
        <productname>PostgreSQL</productname> applications.  Solaris defaults
        to a <varname>SHMMAX</varname> of one-quarter of system <acronym>RAM</acronym>.
        To further adjust this setting, use a project setting associated
        with the <literal>postgres</literal> user.  For example, run the
        following as <literal>root</literal>:
<programlisting>
projadd -c "PostgreSQL DB User" -K "project.max-shm-memory=(privileged,8GB,deny)" -U postgres -G postgres user.postgres
</programlisting>
       </para>

       <para>
        This command adds the <literal>user.postgres</literal> project and
        sets the shared memory maximum for the <literal>postgres</literal>
        user to 8GB, and takes effect the next time that user logs
        in, or when you restart <productname>PostgreSQL</productname> (not reload).
        The above assumes that <productname>PostgreSQL</productname> is run by
        the <literal>postgres</literal> user in the <literal>postgres</literal>
        group.  No server reboot is required.
       </para>

       <para>
        Other recommended kernel setting changes for database servers which will
        have a large number of connections are:
<programlisting>
project.max-shm-ids=(priv,32768,deny)
project.max-sem-ids=(priv,4096,deny)
project.max-msg-ids=(priv,4096,deny)
</programlisting>
       </para>

       <para>
        Additionally, if you are running <productname>PostgreSQL</productname>
        inside a zone, you may need to raise the zone resource usage
        limits as well.  See "Chapter2:  Projects and Tasks" in the
        <citetitle>System Administrator's Guide</citetitle> for more
        information on <literal>projects</literal> and <command>prctl</command>.
       </para>
      </listitem>
     </varlistentry>

    </variablelist>

  </sect2>

  <sect2 id="systemd-removeipc">
   <title>systemd RemoveIPC</title>

   <indexterm>
    <primary>systemd</primary>
    <secondary>RemoveIPC</secondary>
   </indexterm>

   <para>
    If <productname>systemd</productname> is in use, some care must be taken
    that IPC resources (including shared memory) are not prematurely
    removed by the operating system.  This is especially of concern when
    installing PostgreSQL from source.  Users of distribution packages of
    PostgreSQL are less likely to be affected, as
    the <literal>postgres</literal> user is then normally created as a system
    user.
   </para>

   <para>
    The setting <literal>RemoveIPC</literal>
    in <filename>logind.conf</filename> controls whether IPC objects are
    removed when a user fully logs out.  System users are exempt.  This
    setting defaults to on in stock <productname>systemd</productname>, but
    some operating system distributions default it to off.
   </para>

   <para>
    A typical observed effect when this setting is on is that shared memory
    objects used for parallel query execution are removed at apparently random
    times, leading to errors and warnings while attempting to open and remove
    them, like
<screen>
WARNING:  could not remove shared memory segment "/PostgreSQL.1450751626": No such file or directory
</screen>
    Different types of IPC objects (shared memory vs. semaphores, System V
    vs. POSIX) are treated slightly differently
    by <productname>systemd</productname>, so one might observe that some IPC
    resources are not removed in the same way as others.  But it is not
    advisable to rely on these subtle differences.
   </para>

   <para>
    A <quote>user logging out</quote> might happen as part of a maintenance
    job or manually when an administrator logs in as
    the <literal>postgres</literal> user or something similar, so it is hard
    to prevent in general.
   </para>

   <para>
    What is a <quote>system user</quote> is determined
    at <productname>systemd</productname> compile time from
    the <symbol>SYS_UID_MAX</symbol> setting
    in <filename>/etc/login.defs</filename>.
   </para>

   <para>
    Packaging and deployment scripts should be careful to create
    the <literal>postgres</literal> user as a system user by
    using <literal>useradd -r</literal>, <literal>adduser --system</literal>,
    or equivalent.
   </para>

   <para>
    Alternatively, if the user account was created incorrectly or cannot be
    changed, it is recommended to set
<programlisting>
RemoveIPC=no
</programlisting>
    in <filename>/etc/systemd/logind.conf</filename> or another appropriate
    configuration file.
   </para>

   <caution>
    <para>
     At least one of these two things has to be ensured, or the PostgreSQL
     server will be very unreliable.
    </para>
   </caution>
  </sect2>

  <sect2>
   <title>Resource Limits</title>

   <para>
    Unix-like operating systems enforce various kinds of resource limits
    that might interfere with the operation of your
    <productname>PostgreSQL</productname> server. Of particular
    importance are limits on the number of processes per user, the
    number of open files per process, and the amount of memory available
    to each process. Each of these have a <quote>hard</quote> and a
    <quote>soft</quote> limit. The soft limit is what actually counts
    but it can be changed by the user up to the hard limit. The hard
    limit can only be changed by the root user. The system call
    <function>setrlimit</function> is responsible for setting these
    parameters. The shell's built-in command <command>ulimit</command>
    (Bourne shells) or <command>limit</command> (<application>csh</application>) is
    used to control the resource limits from the command line. On
    BSD-derived systems the file <filename>/etc/login.conf</filename>
    controls the various resource limits set during login. See the
    operating system documentation for details. The relevant
    parameters are <varname>maxproc</varname>,
    <varname>openfiles</varname>, and <varname>datasize</varname>. For
    example:
<programlisting>
default:\
...
        :datasize-cur=256M:\
        :maxproc-cur=256:\
        :openfiles-cur=256:\
...
</programlisting>
    (<literal>-cur</literal> is the soft limit.  Append
    <literal>-max</literal> to set the hard limit.)
   </para>

   <para>
    Kernels can also have system-wide limits on some resources.
    <itemizedlist>
     <listitem>
      <para>
      On <productname>Linux</productname>
      <filename>/proc/sys/fs/file-max</filename> determines the
      maximum number of open files that the kernel will support.  It can
      be changed by writing a different number into the file or by
      adding an assignment in <filename>/etc/sysctl.conf</filename>.
      The maximum limit of files per process is fixed at the time the
      kernel is compiled; see
      <filename>/usr/src/linux/Documentation/proc.txt</filename> for
      more information.
      </para>
     </listitem>
    </itemizedlist>
   </para>

   <para>
    The <productname>PostgreSQL</productname> server uses one process
    per connection so you should provide for at least as many processes
    as allowed connections, in addition to what you need for the rest
    of your system.  This is usually not a problem but if you run
    several servers on one machine things might get tight.
   </para>

   <para>
    The factory default limit on open files is often set to
    <quote>socially friendly</quote> values that allow many users to
    coexist on a machine without using an inappropriate fraction of
    the system resources.  If you run many servers on a machine this
    is perhaps what you want, but on dedicated servers you might want to
    raise this limit.
   </para>

   <para>
    On the other side of the coin, some systems allow individual
    processes to open large numbers of files; if more than a few
    processes do so then the system-wide limit can easily be exceeded.
    If you find this happening, and you do not want to alter the
    system-wide limit, you can set <productname>PostgreSQL</productname>'s <xref
    linkend="guc-max-files-per-process"/> configuration parameter to
    limit the consumption of open files.
   </para>
  </sect2>

  <sect2 id="linux-memory-overcommit">
   <title>Linux Memory Overcommit</title>

   <indexterm>
    <primary>memory overcommit</primary>
   </indexterm>

   <indexterm>
    <primary>OOM</primary>
   </indexterm>

   <indexterm>
    <primary>overcommit</primary>
   </indexterm>

   <para>
    The default virtual memory behavior on Linux is not
    optimal for <productname>PostgreSQL</productname>. Because of the
    way that the kernel implements memory overcommit, the kernel might
    terminate the <productname>PostgreSQL</productname> postmaster (the
    supervisor server process) if the memory demands of either
    <productname>PostgreSQL</productname> or another process cause the
    system to run out of virtual memory.
   </para>

   <para>
    If this happens, you will see a kernel message that looks like
    this (consult your system documentation and configuration on where
    to look for such a message):
<programlisting>
Out of Memory: Killed process 12345 (postgres).
</programlisting>
    This indicates that the <filename>postgres</filename> process
    has been terminated due to memory pressure.
    Although existing database connections will continue to function
    normally, no new connections will be accepted.  To recover,
    <productname>PostgreSQL</productname> will need to be restarted.
   </para>

   <para>
    One way to avoid this problem is to run
    <productname>PostgreSQL</productname> on a machine where you can
    be sure that other processes will not run the machine out of
    memory.  If memory is tight, increasing the swap space of the
    operating system can help avoid the problem, because the
    out-of-memory (OOM) killer is invoked only when physical memory and
    swap space are exhausted.
   </para>

   <para>
    If <productname>PostgreSQL</productname> itself is the cause of the
    system running out of memory, you can avoid the problem by changing
    your configuration.  In some cases, it may help to lower memory-related
    configuration parameters, particularly
    <link linkend="guc-shared-buffers"><varname>shared_buffers</varname></link>,
    <link linkend="guc-work-mem"><varname>work_mem</varname></link>, and
    <link linkend="guc-hash-mem-multiplier"><varname>hash_mem_multiplier</varname></link>.
    In other cases, the problem may be caused by allowing too many
    connections to the database server itself.  In many cases, it may
    be better to reduce
    <link linkend="guc-max-connections"><varname>max_connections</varname></link>
    and instead make use of external connection-pooling software.
   </para>

   <para>
    It is possible to modify the
    kernel's behavior so that it will not <quote>overcommit</quote> memory.
    Although this setting will not prevent the <ulink
    url="https://lwn.net/Articles/104179/">OOM killer</ulink> from being invoked
    altogether, it will lower the chances significantly and will therefore
    lead to more robust system behavior.  This is done by selecting strict
    overcommit mode via <command>sysctl</command>:
<programlisting>
sysctl -w vm.overcommit_memory=2
</programlisting>
    or placing an equivalent entry in <filename>/etc/sysctl.conf</filename>.
    You might also wish to modify the related setting
    <varname>vm.overcommit_ratio</varname>.  For details see the kernel documentation
    file <ulink url="https://www.kernel.org/doc/Documentation/vm/overcommit-accounting"></ulink>.
   </para>

   <para>
    Another approach, which can be used with or without altering
    <varname>vm.overcommit_memory</varname>, is to set the process-specific
    <firstterm>OOM score adjustment</firstterm> value for the postmaster process to
    <literal>-1000</literal>, thereby guaranteeing it will not be targeted by the OOM
    killer.  The simplest way to do this is to execute
<programlisting>
echo -1000 > /proc/self/oom_score_adj
</programlisting>
    in the postmaster's startup script just before invoking the postmaster.
    Note that this action must be done as root, or it will have no effect;
    so a root-owned startup script is the easiest place to do it.  If you
    do this, you should also set these environment variables in the startup
    script before invoking the postmaster:
<programlisting>
export PG_OOM_ADJUST_FILE=/proc/self/oom_score_adj
export PG_OOM_ADJUST_VALUE=0
</programlisting>
    These settings will cause postmaster child processes to run with the
    normal OOM score adjustment of zero, so that the OOM killer can still
    target them at need.  You could use some other value for
    <envar>PG_OOM_ADJUST_VALUE</envar> if you want the child processes to run
    with some other OOM score adjustment.  (<envar>PG_OOM_ADJUST_VALUE</envar>
    can also be omitted, in which case it defaults to zero.)  If you do not
    set <envar>PG_OOM_ADJUST_FILE</envar>, the child processes will run with the
    same OOM score adjustment as the postmaster, which is unwise since the
    whole point is to ensure that the postmaster has a preferential setting.
   </para>

  </sect2>

  <sect2 id="linux-huge-pages">
   <title>Linux Huge Pages</title>

   <para>
    Using huge pages reduces overhead when using large contiguous chunks of
    memory, as <productname>PostgreSQL</productname> does, particularly when
    using large values of <xref linkend="guc-shared-buffers"/>.  To use this
    feature in <productname>PostgreSQL</productname> you need a kernel
    with <varname>CONFIG_HUGETLBFS=y</varname> and
    <varname>CONFIG_HUGETLB_PAGE=y</varname>. You will also have to configure
    the operating system to provide enough huge pages of the desired size.
    To estimate the number of huge pages needed, start
    <productname>PostgreSQL</productname> without huge pages enabled and check
    the postmaster's anonymous shared memory segment size, as well as the
    system's default and supported huge page sizes, using the
    <filename>/proc</filename> and <filename>/sys</filename> file systems.
    This might look like:
<programlisting>
$ <userinput>head -1 $PGDATA/postmaster.pid</userinput>
4170
$ <userinput>pmap 4170 | awk '/rw-s/ &amp;&amp; /zero/ {print $2}'</userinput>
6490428K
$ <userinput>grep ^Hugepagesize /proc/meminfo</userinput>
Hugepagesize:       2048 kB
$ <userinput>ls /sys/kernel/mm/hugepages</userinput>
hugepages-1048576kB  hugepages-2048kB
</programlisting>

     In this example the default is 2MB, but you can also explicitly request
     either 2MB or 1GB with <xref linkend="guc-huge-page-size"/>.

     Assuming <literal>2MB</literal> huge pages,
     <literal>6490428</literal> / <literal>2048</literal> gives approximately
     <literal>3169.154</literal>, so in this example we need at
     least <literal>3170</literal> huge pages.  A larger setting would be
     appropriate if other programs on the machine also need huge pages.
     We can set this with:
<programlisting>
# <userinput>sysctl -w vm.nr_hugepages=3170</userinput>
</programlisting>
     Don't forget to add this setting to <filename>/etc/sysctl.conf</filename>
     so that it is reapplied after reboots.  For non-default huge page sizes,
     we can instead use:
<programlisting>
# <userinput>echo 3170 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages</userinput>
</programlisting>
    It is also possible to provide these settings at boot time using
    kernel parameters such as <literal>hugepagesz=2M hugepages=3170</literal>.
   </para>

   <para>
    Sometimes the kernel is not able to allocate the desired number of huge
    pages immediately due to fragmentation, so it might be necessary
    to repeat the command or to reboot.  (Immediately after a reboot, most of
    the machine's memory should be available to convert into huge pages.)
    To verify the huge page allocation situation for a given size, use:
<programlisting>
$ <userinput>cat /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages</userinput>
</programlisting>
   </para>

   <para>
    It may also be necessary to give the database server's operating system
    user permission to use huge pages by setting
    <varname>vm.hugetlb_shm_group</varname> via <application>sysctl</application>, and/or
    give permission to lock memory with <command>ulimit -l</command>.
   </para>

   <para>
    The default behavior for huge pages in
    <productname>PostgreSQL</productname> is to use them when possible, with
    the system's default huge page size, and
    to fall back to normal pages on failure. To enforce the use of huge
    pages, you can set <xref linkend="guc-huge-pages"/>
    to <literal>on</literal> in <filename>postgresql.conf</filename>.
    Note that with this setting <productname>PostgreSQL</productname> will fail to
    start if not enough huge pages are available.
   </para>

   <para>
    For a detailed description of the <productname>Linux</productname> huge
    pages feature have a look
    at <ulink url="https://www.kernel.org/doc/Documentation/vm/hugetlbpage.txt"></ulink>.
   </para>

  </sect2>
 </sect1>


 <sect1 id="server-shutdown">
  <title>Shutting Down the Server</title>

  <indexterm zone="server-shutdown">
   <primary>shutdown</primary>
  </indexterm>

  <para>
   There are several ways to shut down the database server.
   Under the hood, they all reduce to sending a signal to the supervisor
   <command>postgres</command> process.
  </para>

  <para>
   If you are using a pre-packaged version
   of <productname>PostgreSQL</productname>, and you used its provisions
   for starting the server, then you should also use its provisions for
   stopping the server.  Consult the package-level documentation for
   details.
  </para>

  <para>
   When managing the server directly, you can control the type of shutdown
   by sending different signals to the <command>postgres</command>
   process:

   <variablelist>
    <varlistentry>
     <term><systemitem>SIGTERM</systemitem><indexterm><primary>SIGTERM</primary></indexterm></term>
     <listitem>
      <para>
       This is the <firstterm>Smart Shutdown</firstterm> mode.
       After receiving <systemitem>SIGTERM</systemitem>, the server
       disallows new connections, but lets existing sessions end their
       work normally. It shuts down only after all of the sessions terminate.
       If the server is in online backup mode, it additionally waits
       until online backup mode is no longer active.  While backup mode is
       active, new connections will still be allowed, but only to superusers
       (this exception allows a superuser to connect to terminate
       online backup mode).  If the server is in recovery when a smart
       shutdown is requested, recovery and streaming replication will be
       stopped only after all regular sessions have terminated.
      </para>
     </listitem>
    </varlistentry>

    <varlistentry>
     <term><systemitem>SIGINT</systemitem><indexterm><primary>SIGINT</primary></indexterm></term>
     <listitem>
      <para>
       This is the <firstterm>Fast Shutdown</firstterm> mode.
       The server disallows new connections and sends all existing
       server processes <systemitem>SIGTERM</systemitem>, which will cause them
       to abort their current transactions and exit promptly. It then
       waits for all server processes to exit and finally shuts down.
       If the server is in online backup mode, backup mode will be
       terminated, rendering the backup useless.
      </para>
     </listitem>
    </varlistentry>

    <varlistentry>
     <term><systemitem>SIGQUIT</systemitem><indexterm><primary>SIGQUIT</primary></indexterm></term>
     <listitem>
      <para>
      This is the <firstterm>Immediate Shutdown</firstterm> mode.
      The server will send <systemitem>SIGQUIT</systemitem> to all child
      processes and wait for them to terminate.  If any do not terminate
      within 5 seconds, they will be sent <systemitem>SIGKILL</systemitem>.
      The supervisor server process exits as soon as all child processes have
      exited, without doing normal database shutdown processing.
      This will lead to recovery (by
      replaying the WAL log) upon next start-up. This is recommended
      only in emergencies.
      </para>
     </listitem>
    </varlistentry>
   </variablelist>
  </para>

  <para>
   The <xref linkend="app-pg-ctl"/> program provides a convenient
   interface for sending these signals to shut down the server.
   Alternatively, you can send the signal directly using <command>kill</command>
   on non-Windows systems.
   The <acronym>PID</acronym> of the <command>postgres</command> process can be
   found using the <command>ps</command> program, or from the file
   <filename>postmaster.pid</filename> in the data directory. For
   example, to do a fast shutdown:
<screen>
$ <userinput>kill -INT `head -1 /usr/local/pgsql/data/postmaster.pid`</userinput>
</screen>
  </para>

  <important>
   <para>
    It is best not to use <systemitem>SIGKILL</systemitem> to shut down the
    server.  Doing so will prevent the server from releasing shared memory and
    semaphores.  Furthermore, <systemitem>SIGKILL</systemitem> kills
    the <command>postgres</command> process without letting it relay the
    signal to its subprocesses, so it might be necessary to kill the
    individual subprocesses by hand as well.
   </para>
  </important>

  <para>
   To terminate an individual session while allowing other sessions to
   continue, use <function>pg_terminate_backend()</function> (see <xref
   linkend="functions-admin-signal-table"/>) or send a
   <systemitem>SIGTERM</systemitem> signal to the child process associated with
   the session.
  </para>
 </sect1>

 <sect1 id="upgrading">
  <title>Upgrading a <productname>PostgreSQL</productname> Cluster</title>

  <indexterm zone="upgrading">
   <primary>upgrading</primary>
  </indexterm>

  <indexterm zone="upgrading">
   <primary>version</primary>
   <secondary>compatibility</secondary>
  </indexterm>

  <para>
   This section discusses how to upgrade your database data from one
   <productname>PostgreSQL</productname> release to a newer one.
  </para>

  <para>
   Current <productname>PostgreSQL</productname> version numbers consist of a
   major and a minor version number.  For example, in the version number 10.1,
   the 10 is the major version number and the 1 is the minor version number,
   meaning this would be the first minor release of the major release 10.  For
   releases before <productname>PostgreSQL</productname> version 10.0, version
   numbers consist of three numbers, for example, 9.5.3.  In those cases, the
   major version consists of the first two digit groups of the version number,
   e.g., 9.5, and the minor version is the third number, e.g., 3, meaning this
   would be the third minor release of the major release 9.5.
  </para>

  <para>
   Minor releases never change the internal storage format and are always
   compatible with earlier and later minor releases of the same major version
   number.  For example, version 10.1 is compatible with version 10.0 and
   version 10.6.  Similarly, for example, 9.5.3 is compatible with 9.5.0,
   9.5.1, and 9.5.6.  To update between compatible versions, you simply
   replace the executables while the server is down and restart the server.
   The data directory remains unchanged &mdash; minor upgrades are that
   simple.
  </para>

  <para>
   For <emphasis>major</emphasis> releases of <productname>PostgreSQL</productname>, the
   internal data storage format is subject to change, thus complicating
   upgrades.  The traditional method for moving data to a new major version
   is to dump and restore the database, though this can be slow.  A
   faster method is <xref linkend="pgupgrade"/>.  Replication methods are
   also available, as discussed below.
   (If you are using a pre-packaged version
   of <productname>PostgreSQL</productname>, it may provide scripts to
   assist with major version upgrades.  Consult the package-level
   documentation for details.)
  </para>

  <para>
   New major versions also typically introduce some user-visible
   incompatibilities, so application programming changes might be required.
   All user-visible changes are listed in the release notes (<xref
   linkend="release"/>);  pay particular attention to the section
   labeled "Migration".  Though you can upgrade from one major version
   to another without upgrading to intervening versions, you should read
   the major release notes of all intervening versions.
  </para>

  <para>
   Cautious users will want to test their client applications on the new
   version before switching over fully; therefore, it's often a good idea to
   set up concurrent installations of old and new versions.  When
   testing a <productname>PostgreSQL</productname> major upgrade, consider the
   following categories of possible changes:
  </para>

  <variablelist>

   <varlistentry>
    <term>Administration</term>
    <listitem>
     <para>
      The capabilities available for administrators to monitor and control
      the server often change and improve in each major release.
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>SQL</term>
    <listitem>
     <para>
      Typically this includes new SQL command capabilities and not changes
      in behavior, unless specifically mentioned in the release notes.
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>Library API</term>
    <listitem>
     <para>
      Typically libraries like <application>libpq</application> only add new
      functionality, again unless mentioned in the release notes.
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>System Catalogs</term>
    <listitem>
     <para>
      System catalog changes usually only affect database management tools.
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>Server C-language API</term>
    <listitem>
     <para>
      This involves changes in the backend function API, which is written
      in the C programming language.  Such changes affect code that
      references backend functions deep inside the server.
     </para>
    </listitem>
   </varlistentry>

  </variablelist>

  <sect2 id="upgrading-via-pgdumpall">
   <title>Upgrading Data via <application>pg_dumpall</application></title>

   <para>
    One upgrade method is to dump data from one major version of
    <productname>PostgreSQL</productname> and restore it in another &mdash;  to do
    this, you must use a <emphasis>logical</emphasis> backup tool like
    <application>pg_dumpall</application>; file system
    level backup methods will not work. (There are checks in place that prevent
    you from using a data directory with an incompatible version of
    <productname>PostgreSQL</productname>, so no great harm can be done by
    trying to start the wrong server version on a data directory.)
   </para>

   <para>
    It is recommended that you use the <application>pg_dump</application> and
    <application>pg_dumpall</application> programs from the <emphasis>newer</emphasis>
    version of
    <productname>PostgreSQL</productname>, to take advantage of enhancements
    that might have been made in these programs.  Current releases of the
    dump programs can read data from any server version back to 8.0.
   </para>

   <para>
    These instructions assume that your existing installation is under the
    <filename>/usr/local/pgsql</filename> directory, and that the data area is in
    <filename>/usr/local/pgsql/data</filename>.  Substitute your paths
    appropriately.
   </para>

   <procedure>
    <step>
     <para>
      If making a backup, make sure that your database is not being updated.
      This does not affect the integrity of the backup, but the changed
      data would of course not be included. If necessary, edit the
      permissions in the file <filename>/usr/local/pgsql/data/pg_hba.conf</filename>
      (or equivalent) to disallow access from everyone except you.
      See <xref linkend="client-authentication"/> for additional information on
      access control.
     </para>

     <para>
      <indexterm>
       <primary>pg_dumpall</primary>
       <secondary>use during upgrade</secondary>
      </indexterm>

      To back up your database installation, type:
<screen>
<userinput>pg_dumpall &gt; <replaceable>outputfile</replaceable></userinput>
</screen>
     </para>

     <para>
      To make the backup, you can use the <application>pg_dumpall</application>
      command from the version you are currently running;  see <xref
      linkend="backup-dump-all"/> for more details.  For best
      results, however, try to use the <application>pg_dumpall</application>
      command from <productname>PostgreSQL</productname> &version;,
      since this version contains bug fixes and improvements over older
      versions.  While this advice might seem idiosyncratic since you
      haven't installed the new version yet, it is advisable to follow
      it if you plan to install the new version in parallel with the
      old version.  In that case you can complete the installation
      normally and transfer the data later.  This will also decrease
      the downtime.
     </para>
    </step>

    <step>
     <para>
      Shut down the old server:
<screen>
<userinput>pg_ctl stop</userinput>
</screen>
      On systems that have <productname>PostgreSQL</productname> started at boot time,
      there is probably a start-up file that will accomplish the same thing. For
      example, on a <systemitem class="osname">Red Hat Linux</systemitem> system one
      might find that this works:
<screen>
<userinput>/etc/rc.d/init.d/postgresql stop</userinput>
</screen>
      See <xref linkend="runtime"/> for details about starting and
      stopping the server.
     </para>
    </step>

    <step>
     <para>
      If restoring from backup, rename or delete the old installation
      directory if it is not version-specific.  It is a good idea to
      rename the directory, rather than
      delete it, in case you have trouble and need to revert to it.  Keep
      in mind the directory might consume significant disk space.  To rename
      the directory, use a command like this:
<screen>
<userinput>mv /usr/local/pgsql /usr/local/pgsql.old</userinput>
</screen>
     (Be sure to move the directory as a single unit so relative paths
     remain unchanged.)
     </para>
    </step>

    <step>
     <para>
      Install the new version of <productname>PostgreSQL</productname> as
      outlined in <xref linkend="install-procedure"/>.
     </para>
    </step>

    <step>
     <para>
      Create a new database cluster if needed.  Remember that you must
      execute these commands while logged in to the special database user
      account (which you already have if you are upgrading).
<programlisting>
<userinput>/usr/local/pgsql/bin/initdb -D /usr/local/pgsql/data</userinput>
</programlisting>
     </para>
    </step>

    <step>
     <para>
      Restore your previous <filename>pg_hba.conf</filename> and any
      <filename>postgresql.conf</filename> modifications.
     </para>
    </step>

    <step>
     <para>
      Start the database server, again using the special database user
      account:
<programlisting>
<userinput>/usr/local/pgsql/bin/postgres -D /usr/local/pgsql/data</userinput>
</programlisting>
     </para>
    </step>

    <step>
     <para>
      Finally, restore your data from backup with:
<screen>
<userinput>/usr/local/pgsql/bin/psql -d postgres -f <replaceable>outputfile</replaceable></userinput>
</screen>
      using the <emphasis>new</emphasis> <application>psql</application>.
     </para>
    </step>
   </procedure>

   <para>
    The least downtime can be achieved by installing the new server in
    a different directory and running both the old and the new servers
    in parallel, on different ports. Then you can use something like:

<programlisting>
pg_dumpall -p 5432 | psql -d postgres -p 5433
</programlisting>
    to transfer your data.
   </para>

  </sect2>

  <sect2 id="upgrading-via-pg-upgrade">
   <title>Upgrading Data via <application>pg_upgrade</application></title>

   <para>
    The <xref linkend="pgupgrade"/> module allows an installation to
    be migrated in-place from one major <productname>PostgreSQL</productname>
    version to another.  Upgrades can be performed in minutes,
    particularly with <option>--link</option> mode.  It requires steps similar to
    <application>pg_dumpall</application> above, e.g.,  starting/stopping the server,
    running <application>initdb</application>.  The <application>pg_upgrade</application> <link
    linkend="pgupgrade">documentation</link> outlines the necessary steps.
   </para>

  </sect2>

  <sect2 id="upgrading-via-replication">
   <title>Upgrading Data via Replication</title>

   <para>
    It is also possible to use logical replication methods to create a standby
    server with the updated version of <productname>PostgreSQL</productname>.
    This is possible because logical replication supports
    replication between different major versions of
    <productname>PostgreSQL</productname>.  The standby can be on the same computer or
    a different computer.  Once it has synced up with the primary server
    (running the older version of <productname>PostgreSQL</productname>), you can
    switch primaries and make the standby the primary and shut down the older
    database instance.  Such a switch-over results in only several seconds
    of downtime for an upgrade.
   </para>

   <para>
    This method of upgrading can be performed using the built-in logical
    replication facilities as well as using external logical replication
    systems such as <productname>pglogical</productname>,
    <productname>Slony</productname>, <productname>Londiste</productname>, and
    <productname>Bucardo</productname>.
   </para>
  </sect2>
 </sect1>

 <sect1 id="preventing-server-spoofing">
  <title>Preventing Server Spoofing</title>

  <indexterm zone="preventing-server-spoofing">
   <primary>server spoofing</primary>
  </indexterm>

  <para>
   While the server is running, it is not possible for a malicious user
   to take the place of the normal database server.  However, when the
   server is down, it is possible for a local user to spoof the normal
   server by starting their own server.  The spoof server could read
   passwords and queries sent by clients, but could not return any data
   because the <varname>PGDATA</varname> directory would still be secure because
   of directory permissions. Spoofing is possible because any user can
   start a database server; a client cannot identify an invalid server
   unless it is specially configured.
  </para>

  <para>
   One way to prevent spoofing of <literal>local</literal>
   connections is to use a Unix domain socket directory (<xref
   linkend="guc-unix-socket-directories"/>) that has write permission only
   for a trusted local user.  This prevents a malicious user from creating
   their own socket file in that directory.  If you are concerned that
   some applications might still reference <filename>/tmp</filename> for the
   socket file and hence be vulnerable to spoofing, during operating system
   startup create a symbolic link <filename>/tmp/.s.PGSQL.5432</filename> that points
   to the relocated socket file.  You also might need to modify your
   <filename>/tmp</filename> cleanup script to prevent removal of the symbolic link.
  </para>

  <para>
   Another option for <literal>local</literal> connections is for clients to use
   <link linkend="libpq-connect-requirepeer"><literal>requirepeer</literal></link>
   to specify the required owner of the server process connected to
   the socket.
  </para>

  <para>
   To prevent spoofing on TCP connections, either use
   SSL certificates and make sure that clients check the server's certificate,
   or use GSSAPI encryption (or both, if they're on separate connections).
  </para>

  <para>
   To prevent spoofing with SSL, the server
   must be configured to accept only <literal>hostssl</literal> connections (<xref
   linkend="auth-pg-hba-conf"/>) and have SSL key and certificate files
   (<xref linkend="ssl-tcp"/>). The TCP client must connect using
   <literal>sslmode=verify-ca</literal> or
   <literal>verify-full</literal> and have the appropriate root certificate
   file installed (<xref linkend="libq-ssl-certificates"/>).
  </para>

  <para>
    To prevent spoofing with GSSAPI, the server must be configured to accept
    only <literal>hostgssenc</literal> connections
    (<xref linkend="auth-pg-hba-conf"/>) and use <literal>gss</literal>
    authentication with them.  The TCP client must connect
    using <literal>gssencmode=require</literal>.
  </para>
 </sect1>

 <sect1 id="encryption-options">
  <title>Encryption Options</title>

  <indexterm zone="encryption-options">
   <primary>encryption</primary>
  </indexterm>

  <para>
   <productname>PostgreSQL</productname> offers encryption at several
   levels, and provides flexibility in protecting data from disclosure
   due to database server theft, unscrupulous administrators, and
   insecure networks. Encryption might also be required to secure
   sensitive data such as medical records or financial transactions.
  </para>

  <variablelist>

  <varlistentry>
   <term>Password Encryption</term>
   <listitem>

    <para>
     Database user passwords are stored as hashes (determined by the setting
     <xref linkend="guc-password-encryption"/>), so the administrator cannot
     determine the actual password assigned to the user. If SCRAM or MD5
     encryption is used for client authentication, the unencrypted password is
     never even temporarily present on the server because the client encrypts
     it before being sent across the network. SCRAM is preferred, because it
     is an Internet standard and is more secure than the PostgreSQL-specific
     MD5 authentication protocol.
    </para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term>Encryption For Specific Columns</term>

   <listitem>
    <para>
     The <xref linkend="pgcrypto"/> module allows certain fields to be
     stored encrypted.
     This is useful if only some of the data is sensitive.
     The client supplies the decryption key and the data is decrypted
     on the server and then sent to the client.
    </para>

    <para>
     The decrypted data and the decryption key are present on the
     server for a brief time while it is being decrypted and
     communicated between the client and server. This presents a brief
     moment where the data and keys can be intercepted by someone with
     complete access to the database server, such as the system
     administrator.
    </para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term>Data Partition Encryption</term>

   <listitem>
    <para>
     Storage encryption can be performed at the file system level or the
     block level.  Linux file system encryption options include eCryptfs
     and EncFS, while FreeBSD uses PEFS.  Block level or full disk
     encryption options include dm-crypt + LUKS on Linux and GEOM
     modules geli and gbde on FreeBSD.  Many other operating systems
     support this functionality, including Windows.
    </para>

    <para>
     This mechanism prevents unencrypted data from being read from the
     drives if the drives or the entire computer is stolen. This does
     not protect against attacks while the file system is mounted,
     because when mounted, the operating system provides an unencrypted
     view of the data. However, to mount the file system, you need some
     way for the encryption key to be passed to the operating system,
     and sometimes the key is stored somewhere on the host that mounts
     the disk.
    </para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term>Encrypting Data Across A Network</term>

    <listitem>
     <para>
      SSL connections encrypt all data sent across the network: the
      password, the queries, and the data returned. The
      <filename>pg_hba.conf</filename> file allows administrators to specify
      which hosts can use non-encrypted connections (<literal>host</literal>)
      and which require SSL-encrypted connections
      (<literal>hostssl</literal>). Also, clients can specify that they
      connect to servers only via SSL.
     </para>

     <para>
      GSSAPI-encrypted connections encrypt all data sent across the network,
      including queries and data returned.  (No password is sent across the
      network.)  The <filename>pg_hba.conf</filename> file allows
      administrators to specify which hosts can use non-encrypted connections
      (<literal>host</literal>) and which require GSSAPI-encrypted connections
      (<literal>hostgssenc</literal>).  Also, clients can specify that they
      connect to servers only on GSSAPI-encrypted connections
      (<literal>gssencmode=require</literal>).
     </para>

     <para>
      <application>Stunnel</application> or
      <application>SSH</application> can also be used to encrypt
      transmissions.
     </para>
    </listitem>
  </varlistentry>

  <varlistentry>
   <term>SSL Host Authentication</term>

   <listitem>
    <para>
     It is possible for both the client and server to provide SSL
     certificates to each other. It takes some extra configuration
     on each side, but this provides stronger verification of identity
     than the mere use of passwords. It prevents a computer from
     pretending to be the server just long enough to read the password
     sent by the client. It also helps prevent <quote>man in the middle</quote>
     attacks where a computer between the client and server pretends to
     be the server and reads and passes all data between the client and
     server.
    </para>
   </listitem>
  </varlistentry>

  <varlistentry>
   <term>Client-Side Encryption</term>

   <listitem>
    <para>
     If the system administrator for the server's machine cannot be trusted,
     it is necessary
     for the client to encrypt the data; this way, unencrypted data
     never appears on the database server. Data is encrypted on the
     client before being sent to the server, and database results have
     to be decrypted on the client before being used.
    </para>
   </listitem>
  </varlistentry>

  </variablelist>

 </sect1>

 <sect1 id="ssl-tcp">
  <title>Secure TCP/IP Connections with SSL</title>

  <indexterm zone="ssl-tcp">
   <primary>SSL</primary>
  </indexterm>

  <para>
   <productname>PostgreSQL</productname> has native support for using
   <acronym>SSL</acronym> connections to encrypt client/server communications
   for increased security. This requires that
   <productname>OpenSSL</productname> is installed on both client and
   server systems and that support in <productname>PostgreSQL</productname> is
   enabled at build time (see <xref linkend="installation"/>).
  </para>

  <sect2 id="ssl-setup">
   <title>Basic Setup</title>

  <para>
   With <acronym>SSL</acronym> support compiled in, the
   <productname>PostgreSQL</productname> server can be started with
   <acronym>SSL</acronym> enabled by setting the parameter
   <xref linkend="guc-ssl"/> to <literal>on</literal> in
   <filename>postgresql.conf</filename>.  The server will listen for both normal
   and <acronym>SSL</acronym> connections on the same TCP port, and will negotiate
   with any connecting client on whether to use <acronym>SSL</acronym>.  By
   default, this is at the client's option; see <xref
   linkend="auth-pg-hba-conf"/> about how to set up the server to require
   use of <acronym>SSL</acronym> for some or all connections.
  </para>

  <para>
   To start in <acronym>SSL</acronym> mode, files containing the server certificate
   and private key must exist.  By default, these files are expected to be
   named <filename>server.crt</filename> and <filename>server.key</filename>, respectively, in
   the server's data directory, but other names and locations can be specified
   using the configuration parameters <xref linkend="guc-ssl-cert-file"/>
   and <xref linkend="guc-ssl-key-file"/>.
  </para>

  <para>
   On Unix systems, the permissions on <filename>server.key</filename> must
   disallow any access to world or group; achieve this by the command
   <command>chmod 0600 server.key</command>.  Alternatively, the file can be
   owned by root and have group read access (that is, <literal>0640</literal>
   permissions).  That setup is intended for installations where certificate
   and key files are managed by the operating system.  The user under which
   the <productname>PostgreSQL</productname> server runs should then be made a
   member of the group that has access to those certificate and key files.
  </para>

  <para>
    If the data directory allows group read access then certificate files may
    need to be located outside of the data directory in order to conform to the
    security requirements outlined above.  Generally, group access is enabled
    to allow an unprivileged user to backup the database, and in that case the
    backup software will not be able to read the certificate files and will
    likely error.
  </para>

  <para>
   If the private key is protected with a passphrase, the
   server will prompt for the passphrase and will not start until it has
   been entered.
   Using a passphrase by default disables the ability to change the server's
   SSL configuration without a server restart, but see <xref
   linkend="guc-ssl-passphrase-command-supports-reload"/>.
   Furthermore, passphrase-protected private keys cannot be used at all
   on Windows.
  </para>

  <para>
   The first certificate in <filename>server.crt</filename> must be the
   server's certificate because it must match the server's private key.
   The certificates of <quote>intermediate</quote> certificate authorities
   can also be appended to the file.  Doing this avoids the necessity of
   storing intermediate certificates on clients, assuming the root and
   intermediate certificates were created with <literal>v3_ca</literal>
   extensions.  (This sets the certificate's basic constraint of
   <literal>CA</literal> to <literal>true</literal>.)
   This allows easier expiration of intermediate certificates.
  </para>

  <para>
   It is not necessary to add the root certificate to
   <filename>server.crt</filename>.  Instead, clients must have the root
   certificate of the server's certificate chain.
  </para>
  </sect2>

  <sect2 id="ssl-openssl-config">
   <title>OpenSSL Configuration</title>

  <para>
   <productname>PostgreSQL</productname> reads the system-wide
   <productname>OpenSSL</productname> configuration file. By default, this
   file is named <filename>openssl.cnf</filename> and is located in the
   directory reported by <literal>openssl version -d</literal>.
   This default can be overridden by setting environment variable
   <envar>OPENSSL_CONF</envar> to the name of the desired configuration file.
  </para>

  <para>
   <productname>OpenSSL</productname> supports a wide range of ciphers
   and authentication algorithms, of varying strength.  While a list of
   ciphers can be specified in the <productname>OpenSSL</productname>
   configuration file, you can specify ciphers specifically for use by
   the database server by modifying <xref linkend="guc-ssl-ciphers"/> in
   <filename>postgresql.conf</filename>.
  </para>

  <note>
   <para>
    It is possible to have authentication without encryption overhead by
    using <literal>NULL-SHA</literal> or <literal>NULL-MD5</literal> ciphers.  However,
    a man-in-the-middle could read and pass communications between client
    and server.  Also, encryption overhead is minimal compared to the
    overhead of authentication.  For these reasons NULL ciphers are not
    recommended.
   </para>
  </note>
  </sect2>

  <sect2 id="ssl-client-certificates">
   <title>Using Client Certificates</title>

  <para>
   To require the client to supply a trusted certificate,
   place certificates of the root certificate authorities
   (<acronym>CA</acronym>s) you trust in a file in the data
   directory, set the parameter <xref linkend="guc-ssl-ca-file"/> in
   <filename>postgresql.conf</filename> to the new file name, and add the
   authentication option <literal>clientcert=verify-ca</literal> or
   <literal>clientcert=verify-full</literal> to the appropriate
   <literal>hostssl</literal> line(s) in <filename>pg_hba.conf</filename>.
   A certificate will then be requested from the client during SSL
   connection startup.  (See <xref linkend="libpq-ssl"/> for a description
   of how to set up certificates on the client.)
  </para>

  <para>
   For a <literal>hostssl</literal> entry with
   <literal>clientcert=verify-ca</literal>, the server will verify
   that the client's certificate is signed by one of the trusted
   certificate authorities. If <literal>clientcert=verify-full</literal>
   is specified, the server will not only verify the certificate
   chain, but it will also check whether the username or its mapping
   matches the <literal>cn</literal> (Common Name) of the provided certificate.
   Note that certificate chain validation is always ensured when the
   <literal>cert</literal> authentication method is used
   (see <xref linkend="auth-cert"/>).
  </para>

  <para>
   Intermediate certificates that chain up to existing root certificates
   can also appear in the <xref linkend="guc-ssl-ca-file"/> file if
   you wish to avoid storing them on clients (assuming the root and
   intermediate certificates were created with <literal>v3_ca</literal>
   extensions).  Certificate Revocation List (CRL) entries are also
   checked if the parameter <xref linkend="guc-ssl-crl-file"/> or
   <xref linkend="guc-ssl-crl-dir"/> is set.
  </para>

  <para>
   The <literal>clientcert</literal> authentication option is available for
   all authentication methods, but only in <filename>pg_hba.conf</filename> lines
   specified as <literal>hostssl</literal>.  When <literal>clientcert</literal> is
   not specified, the server verifies the client certificate against its CA
   file only if a client certificate is presented and the CA is configured.
  </para>

  <para>
   There are two approaches to enforce that users provide a certificate during login.
  </para>

  <para>
   The first approach makes use of the <literal>cert</literal> authentication
   method for <literal>hostssl</literal> entries in <filename>pg_hba.conf</filename>,
   such that the certificate itself is used for authentication while also
   providing ssl connection security. See <xref linkend="auth-cert"/> for details.
   (It is not necessary to specify any <literal>clientcert</literal> options
   explicitly when using the <literal>cert</literal> authentication method.)
   In this case, the <literal>cn</literal> (Common Name) provided in
   the certificate is checked against the user name or an applicable mapping.
  </para>

  <para>
   The second approach combines any authentication method for <literal>hostssl</literal>
   entries with the verification of client certificates by setting the
   <literal>clientcert</literal> authentication option to <literal>verify-ca</literal>
   or <literal>verify-full</literal>. The former option only enforces that
   the certificate is valid, while the latter also ensures that the
   <literal>cn</literal> (Common Name) in the certificate matches
   the user name or an applicable mapping.
  </para>
  </sect2>

  <sect2 id="ssl-server-files">
   <title>SSL Server File Usage</title>

   <para>
    <xref linkend="ssl-file-usage"/> summarizes the files that are
    relevant to the SSL setup on the server.  (The shown file names are default
    names.  The locally configured names could be different.)
   </para>

  <table id="ssl-file-usage">
   <title>SSL Server File Usage</title>
   <tgroup cols="3">
    <thead>
     <row>
      <entry>File</entry>
      <entry>Contents</entry>
      <entry>Effect</entry>
     </row>
    </thead>

    <tbody>

     <row>
      <entry><xref linkend="guc-ssl-cert-file"/> (<filename>$PGDATA/server.crt</filename>)</entry>
      <entry>server certificate</entry>
      <entry>sent to client to indicate server's identity</entry>
     </row>

     <row>
      <entry><xref linkend="guc-ssl-key-file"/> (<filename>$PGDATA/server.key</filename>)</entry>
      <entry>server private key</entry>
      <entry>proves server certificate was sent by the owner; does not indicate
      certificate owner is trustworthy</entry>
     </row>

     <row>
      <entry><xref linkend="guc-ssl-ca-file"/></entry>
      <entry>trusted certificate authorities</entry>
      <entry>checks that client certificate is
      signed by a trusted certificate authority</entry>
     </row>

     <row>
      <entry><xref linkend="guc-ssl-crl-file"/></entry>
      <entry>certificates revoked by certificate authorities</entry>
      <entry>client certificate must not be on this list</entry>
     </row>

    </tbody>
   </tgroup>
  </table>

   <para>
    The server reads these files at server start and whenever the server
    configuration is reloaded.  On <systemitem class="osname">Windows</systemitem>
    systems, they are also re-read whenever a new backend process is spawned
    for a new client connection.
   </para>

   <para>
    If an error in these files is detected at server start, the server will
    refuse to start.  But if an error is detected during a configuration
    reload, the files are ignored and the old SSL configuration continues to
    be used.  On <systemitem class="osname">Windows</systemitem> systems, if an error in
    these files is detected at backend start, that backend will be unable to
    establish an SSL connection.  In all these cases, the error condition is
    reported in the server log.
   </para>
  </sect2>

  <sect2 id="ssl-certificate-creation">
   <title>Creating Certificates</title>

   <para>
     To create a simple self-signed certificate for the server, valid for 365
     days, use the following <productname>OpenSSL</productname> command,
     replacing <replaceable>dbhost.yourdomain.com</replaceable> with the
     server's host name:
<programlisting>
openssl req -new -x509 -days 365 -nodes -text -out server.crt \
  -keyout server.key -subj "/CN=<replaceable>dbhost.yourdomain.com</replaceable>"
</programlisting>
    Then do:
<programlisting>
chmod og-rwx server.key
</programlisting>
    because the server will reject the file if its permissions are more
    liberal than this.
    For more details on how to create your server private key and
    certificate, refer to the <productname>OpenSSL</productname> documentation.
   </para>

   <para>
    While a self-signed certificate can be used for testing, a certificate
    signed by a certificate authority (<acronym>CA</acronym>) (usually an
    enterprise-wide root <acronym>CA</acronym>) should be used in production.
   </para>

   <para>
    To create a server certificate whose identity can be validated
    by clients, first create a certificate signing request
    (<acronym>CSR</acronym>) and a public/private key file:
<programlisting>
openssl req -new -nodes -text -out root.csr \
  -keyout root.key -subj "/CN=<replaceable>root.yourdomain.com</replaceable>"
chmod og-rwx root.key
</programlisting>
    Then, sign the request with the key to create a root certificate
    authority (using the default <productname>OpenSSL</productname>
    configuration file location on <productname>Linux</productname>):
<programlisting>
openssl x509 -req -in root.csr -text -days 3650 \
  -extfile /etc/ssl/openssl.cnf -extensions v3_ca \
  -signkey root.key -out root.crt
</programlisting>
    Finally, create a server certificate signed by the new root certificate
    authority:
<programlisting>
openssl req -new -nodes -text -out server.csr \
  -keyout server.key -subj "/CN=<replaceable>dbhost.yourdomain.com</replaceable>"
chmod og-rwx server.key

openssl x509 -req -in server.csr -text -days 365 \
  -CA root.crt -CAkey root.key -CAcreateserial \
  -out server.crt
</programlisting>
    <filename>server.crt</filename> and <filename>server.key</filename>
    should be stored on the server, and <filename>root.crt</filename> should
    be stored on the client so the client can verify that the server's leaf
    certificate was signed by its trusted root certificate.
    <filename>root.key</filename> should be stored offline for use in
    creating future certificates.
   </para>

   <para>
    It is also possible to create a chain of trust that includes
    intermediate certificates:
<programlisting>
# root
openssl req -new -nodes -text -out root.csr \
  -keyout root.key -subj "/CN=<replaceable>root.yourdomain.com</replaceable>"
chmod og-rwx root.key
openssl x509 -req -in root.csr -text -days 3650 \
  -extfile /etc/ssl/openssl.cnf -extensions v3_ca \
  -signkey root.key -out root.crt

# intermediate
openssl req -new -nodes -text -out intermediate.csr \
  -keyout intermediate.key -subj "/CN=<replaceable>intermediate.yourdomain.com</replaceable>"
chmod og-rwx intermediate.key
openssl x509 -req -in intermediate.csr -text -days 1825 \
  -extfile /etc/ssl/openssl.cnf -extensions v3_ca \
  -CA root.crt -CAkey root.key -CAcreateserial \
  -out intermediate.crt

# leaf
openssl req -new -nodes -text -out server.csr \
  -keyout server.key -subj "/CN=<replaceable>dbhost.yourdomain.com</replaceable>"
chmod og-rwx server.key
openssl x509 -req -in server.csr -text -days 365 \
  -CA intermediate.crt -CAkey intermediate.key -CAcreateserial \
  -out server.crt
</programlisting>
    <filename>server.crt</filename> and
    <filename>intermediate.crt</filename> should be concatenated
    into a certificate file bundle and stored on the server.
    <filename>server.key</filename> should also be stored on the server.
    <filename>root.crt</filename> should be stored on the client so
    the client can verify that the server's leaf certificate was signed
    by a chain of certificates linked to its trusted root certificate.
    <filename>root.key</filename> and <filename>intermediate.key</filename>
    should be stored offline for use in creating future certificates.
   </para>
  </sect2>

 </sect1>

 <sect1 id="gssapi-enc">
  <title>Secure TCP/IP Connections with GSSAPI Encryption</title>

  <indexterm zone="gssapi-enc">
   <primary>gssapi</primary>
  </indexterm>

  <para>
   <productname>PostgreSQL</productname> also has native support for
   using <acronym>GSSAPI</acronym> to encrypt client/server communications for
   increased security.  Support requires that a <acronym>GSSAPI</acronym>
   implementation (such as MIT Kerberos) is installed on both client and server
   systems, and that support in <productname>PostgreSQL</productname> is
   enabled at build time (see <xref linkend="installation"/>).
  </para>

  <sect2 id="gssapi-setup">
   <title>Basic Setup</title>

   <para>
    The <productname>PostgreSQL</productname> server will listen for both
    normal and <acronym>GSSAPI</acronym>-encrypted connections on the same TCP
    port, and will negotiate with any connecting client whether to
    use <acronym>GSSAPI</acronym> for encryption (and for authentication).  By
    default, this decision is up to the client (which means it can be
    downgraded by an attacker); see <xref linkend="auth-pg-hba-conf"/> about
    setting up the server to require the use of <acronym>GSSAPI</acronym> for
    some or all connections.
   </para>

   <para>
    When using <acronym>GSSAPI</acronym> for encryption, it is common to
    use <acronym>GSSAPI</acronym> for authentication as well, since the
    underlying mechanism will determine both client and server identities
    (according to the <acronym>GSSAPI</acronym> implementation) in any
    case.  But this is not required;
    another <productname>PostgreSQL</productname> authentication method
    can be chosen to perform additional verification.
   </para>

   <para>
    Other than configuration of the negotiation
    behavior, <acronym>GSSAPI</acronym> encryption requires no setup beyond
    that which is necessary for GSSAPI authentication.  (For more information
    on configuring that, see <xref linkend="gssapi-auth"/>.)
   </para>
  </sect2>
 </sect1>

 <sect1 id="ssh-tunnels">
  <title>Secure TCP/IP Connections with <application>SSH</application> Tunnels</title>

  <indexterm zone="ssh-tunnels">
   <primary>ssh</primary>
  </indexterm>

  <para>
   It is possible to use <application>SSH</application> to encrypt the network
   connection between clients and a
   <productname>PostgreSQL</productname> server. Done properly, this
   provides an adequately secure network connection, even for non-SSL-capable
   clients.
  </para>

  <para>
   First make sure that an <application>SSH</application> server is
   running properly on the same machine as the
   <productname>PostgreSQL</productname> server and that you can log in using
   <command>ssh</command> as some user;  you then can establish a
   secure tunnel to the remote server.  A secure tunnel listens on a
   local port and forwards all traffic to a port on the remote machine.
   Traffic sent to the remote port can arrive on its
   <literal>localhost</literal> address, or different bind
   address if desired;  it does not appear as coming from your
   local machine.  This command creates a secure tunnel from the client
   machine to the remote machine <literal>foo.com</literal>:
<programlisting>
ssh -L 63333:localhost:5432 joe@foo.com
</programlisting>
   The first number in the <option>-L</option> argument, 63333, is the
   local port number of the tunnel; it can be any unused port.  (IANA
   reserves ports 49152 through 65535 for private use.)  The name or IP
   address after this is the remote bind address you are connecting to,
   i.e., <literal>localhost</literal>, which is the default.  The second
   number, 5432, is the remote end of the tunnel, e.g., the port number
   your database server is using.  In order to connect to the database
   server using this tunnel, you connect to port 63333 on the local
   machine:
<programlisting>
psql -h localhost -p 63333 postgres
</programlisting>
   To the database server it will then look as though you are
   user <literal>joe</literal> on host <literal>foo.com</literal>
   connecting to the <literal>localhost</literal> bind address, and it
   will use whatever authentication procedure was configured for
   connections by that user to that bind address.  Note that the server will not
   think the connection is SSL-encrypted, since in fact it is not
   encrypted between the
   <application>SSH</application> server and the
   <productname>PostgreSQL</productname> server.  This should not pose any
   extra security risk because they are on the same machine.
  </para>

  <para>
   In order for the
   tunnel setup to succeed you must be allowed to connect via
   <command>ssh</command> as <literal>joe@foo.com</literal>, just
   as if you had attempted to use <command>ssh</command> to create a
   terminal session.
  </para>

  <para>
   You could also have set up port forwarding as
<programlisting>
ssh -L 63333:foo.com:5432 joe@foo.com
</programlisting>
   but then the database server will see the connection as coming in
   on its <literal>foo.com</literal> bind address, which is not opened by
   the default setting <literal>listen_addresses =
   'localhost'</literal>.  This is usually not what you want.
  </para>

  <para>
   If you have to <quote>hop</quote> to the database server via some
   login host, one possible setup could look like this:
<programlisting>
ssh -L 63333:db.foo.com:5432 joe@shell.foo.com
</programlisting>
   Note that this way the connection
   from <literal>shell.foo.com</literal>
   to <literal>db.foo.com</literal> will not be encrypted by the SSH
   tunnel.
   SSH offers quite a few configuration possibilities when the network
   is restricted in various ways.  Please refer to the SSH
   documentation for details.
  </para>

  <tip>
   <para>
    Several other applications exist that can provide secure tunnels using
    a procedure similar in concept to the one just described.
   </para>
  </tip>

 </sect1>

 <sect1 id="event-log-registration">
  <title>Registering <application>Event Log</application> on <systemitem
  class="osname">Windows</systemitem></title>

  <indexterm zone="event-log-registration">
   <primary>event log</primary>
   <secondary>event log</secondary>
  </indexterm>

  <para>
   To register a <systemitem class="osname">Windows</systemitem>
   <application>event log</application> library with the operating system,
   issue this command:
<screen>
<userinput>regsvr32 <replaceable>pgsql_library_directory</replaceable>/pgevent.dll</userinput>
</screen>
   This creates registry entries used by the event viewer, under the default
   event source named <literal>PostgreSQL</literal>.
  </para>

  <para>
   To specify a different event source name (see
   <xref linkend="guc-event-source"/>), use the <literal>/n</literal>
   and <literal>/i</literal> options:
<screen>
<userinput>regsvr32 /n /i:<replaceable>event_source_name</replaceable> <replaceable>pgsql_library_directory</replaceable>/pgevent.dll</userinput>
</screen>
  </para>

  <para>
   To unregister the <application>event log</application> library from
   the operating system, issue this command:
<screen>
<userinput>regsvr32 /u [/i:<replaceable>event_source_name</replaceable>] <replaceable>pgsql_library_directory</replaceable>/pgevent.dll</userinput>
</screen>
  </para>

  <note>
   <para>
    To enable event logging in the database server, modify
    <xref linkend="guc-log-destination"/> to include
    <literal>eventlog</literal> in <filename>postgresql.conf</filename>.
   </para>
  </note>
 </sect1>

</chapter>