path: root/doc/src/sgml/runtime.sgml

<!-- 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>

 <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.) It is not advisable to install executables owned by this 
   user because compromised systems could then modify their own 
   binaries.
  </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>.
   (<acronym>SQL</acronym> 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 will be 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. To initialize a
   database cluster, use the command <xref
   linkend="app-initdb">,<indexterm><primary>initdb</></> which is
   installed with <productname>PostgreSQL</productname>. The desired
   file system location of your database cluster is indicated by the
   <option>-D</option> option, for example:
<screen>
<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</></> like so:
<screen>
<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. It is likely that it will not
   have the permission to do so (if you followed our advice and created
   an unprivileged account). In that case you should create the
   directory yourself (as root) and change the owner to be the
   <productname>PostgreSQL</productname> user. Here is how this might
   be done:
<screen>
root# <userinput>mkdir /usr/local/pgsql/data</userinput>
root# <userinput>chown postgres /usr/local/pgsql/data</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
   looks like it has already been initialized.</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.
  </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</><secondary>of the
   superuser</></indexterm>  Also, specify <option>-A md5</> or
   <option>-A password</> so that the default <literal>trust</> authentication
   mode is not used; or modify the generated <filename>pg_hba.conf</filename>
   file after running <command>initdb</command>, but
   <emphasis>before</> you start the server for the first time. (Other
   reasonable approaches include using <literal>ident</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</></> 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</> or <literal>POSIX</>. 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>

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

   <indexterm zone="creating-cluster-nfs">
    <primary>Network File Systems</primary>
   </indexterm>
   <indexterm><primary><acronym>NFS</></><see>Network File Systems</></>
   <indexterm><primary>Network Attached Storage (<acronym>NAS</>)</><see>Network File Systems</></>

   <para>
    Many installations create database clusters on network file systems.
    Sometimes this is done directly via <acronym>NFS</>, or by using a
    Network Attached Storage (<acronym>NAS</>) device that uses
    <acronym>NFS</> internally.  <productname>PostgreSQL</> does nothing
    special for <acronym>NFS</> file systems, meaning it assumes
    <acronym>NFS</> behaves exactly like locally-connected drives
    (<acronym>DAS</>, Direct Attached Storage).  If client and server
    <acronym>NFS</> implementations have non-standard semantics, this can
    cause reliability problems (see <ulink
    url="http://www.time-travellers.org/shane/papers/NFS_considered_harmful.html"></ulink>).
    Specifically, delayed (asynchronous) writes to the <acronym>NFS</>
    server can cause reliability problems;   if possible, mount
    <acronym>NFS</> file systems synchronously (without caching) to avoid
    this.  Also, soft-mounting <acronym>NFS</> is not recommended.
    (Storage Area Networks (<acronym>SAN</>) use a low-level
    communication protocol rather than <acronym>NFS</>.)
   </para>

  </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</></>
   The <command>postgres</command> program must know where to
   find the data it is supposed to use. This is done with the
   <option>-D</option> option. Thus, the simplest way to start the
   server is:
<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</> and
   <systemitem>stderr</> 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</><secondary>starting
   the server during</></> Autostart scripts are operating-system-specific.
   There are a few distributed with
   <productname>PostgreSQL</productname> in the
   <filename>contrib/start-scripts</> 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>rc.d</> 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 -c '...'
   postgres</literal>.  For example:
<programlisting>
su -c 'pg_ctl start -D /usr/local/pgsql/data -l serverlog' postgres
</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</><secondary>start script</secondary></>
     </para>
    </listitem>

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

    <listitem>
     <para>
      On <productname>Linux</productname> systems either add
      <indexterm><primary>Linux</><secondary>start script</secondary></>
<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 look at the file
      <filename>contrib/start-scripts/linux</filename> in the
      <productname>PostgreSQL</productname> source distribution.
     </para>
    </listitem>

    <listitem>
     <para>
      On <productname>NetBSD</productname>, either use the
      <productname>FreeBSD</productname> or
      <productname>Linux</productname> start scripts, depending on
      preference. <indexterm><primary>NetBSD</><secondary>start script</secondary></>
     </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</><secondary>start script</secondary></>
<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</> as
      <filename>S99postgresql</>.
     </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 socket: 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 TCP/IP listen socket
</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 socket: Permission denied
HINT:  Is another postmaster already running on port 666? If not, wait a few seconds and retry.
FATAL:  could not create TCP/IP listen socket
</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). Or it could
     mean that you do not have System-V-style shared memory support
     configured into your kernel at all. As a temporary workaround, you
     can try starting the server with a smaller-than-normal number of
     buffers (<xref linkend="guc-shared-buffers">). You will eventually want
     to 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</> 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>
     If you get an <quote>illegal system call</> error, it is likely that
     shared memory or semaphores are not supported in your kernel at
     all. In that case your only option is to reconfigure the kernel to
     enable these features.
    </para>

    <para>
     Details about configuring <systemitem class="osname">System V</>
     <acronym>IPC</> 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: could not connect to server: Connection refused
        Is the server running on host "server.joe.com" and accepting
        TCP/IP connections on port 5432?
</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'll get this when attempting Unix-domain socket
     communication to a local server:
<screen>
psql: could not connect to server: No such file or directory
        Is the server running locally and accepting
        connections on Unix domain socket "/tmp/.s.PGSQL.5432"?
</screen>
    </para>

    <para>
     The last line is useful in verifying that the client is trying to
     connect to the right place. If there is in fact no server
     running 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.
    </para>
   </sect2>
  </sect1>

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

  <para>
   A large <productname>PostgreSQL</> installation can quickly exhaust
   various operating system resource limits. (On some systems, the
   factory defaults are so low that you don't even need a really
   <quote>large</> installation.) If you have encountered this kind of
   problem, keep reading.
  </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>
    Shared memory and semaphores are collectively referred to as
    <quote><systemitem class="osname">System V</>
    <acronym>IPC</></quote> (together with message queues, which are not
    relevant for <productname>PostgreSQL</>). Almost all modern
    operating systems provide these features, but many of them don't have
    them turned on or sufficiently sized by default, especially as
    available RAM and the demands of database applications grow.
    (On <systemitem class="osname">Windows</>,
    <productname>PostgreSQL</> provides its own replacement 
    implementation of these facilities, so most of this section
    can be disregarded.)
   </para>

   <para>
    The complete lack of these facilities is usually manifested by an
    <errorname>Illegal system call</> error upon server start. In
    that case there is no alternative but to reconfigure your
    kernel.  <productname>PostgreSQL</> won't work without them.
    This situation is rare, however, among modern operating systems.
   </para>

   <para>
    When <productname>PostgreSQL</> exceeds one of the various hard
    <acronym>IPC</> 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</> <acronym>IPC</> parameters</>

    <tgroup cols="3">
     <thead>
      <row>
       <entry>Name</>
       <entry>Description</>
       <entry>Reasonable values</>
      </row>
     </thead>

     <tbody>
      <row>
       <entry><varname>SHMMAX</></>
       <entry>Maximum size of shared memory segment (bytes)</>
       <entry>at least several megabytes (see text)</entry>
      </row>

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

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

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

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

       <row>
        <entry><varname>SEMMNI</></>
        <entry>Maximum number of semaphore identifiers (i.e., sets)</>
        <entry>at least <literal>ceil((max_connections + autovacuum_max_workers) / 16)</literal></>
       </row>

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

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

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

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

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


   <para>
    <indexterm><primary>SHMMAX</primary></indexterm> The most important
    shared memory parameter is <varname>SHMMAX</>, the maximum size, in
    bytes, of a shared memory segment. If you get an error message from
    <function>shmget</> like <quote>Invalid argument</>, it is
    likely that this limit has been exceeded.  The size of the required
    shared memory segment varies depending on several
    <productname>PostgreSQL</> configuration parameters, as shown in
    <xref linkend="shared-memory-parameters">.  (Any error message you might
    get will include the exact size of the failed allocation request.)
    You can, as a temporary solution, lower some of those settings to
    avoid the failure.  While it is possible to get
    <productname>PostgreSQL</> to run with <varname>SHMMAX</> as small as
    2 MB, you need considerably more for acceptable performance.  Desirable
    settings are in the hundreds of megabytes to a few gigabytes.
   </para>

   <para>
    Some systems also have a limit on the total amount of shared memory in
    the system (<varname>SHMALL</>).  Make sure this is large enough
    for <productname>PostgreSQL</> plus any other applications that
    are using shared memory segments.  Note that <varname>SHMALL</>
    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</>), which should be at most
    approximately 500 kB for <productname>PostgreSQL</> (it is
    usually just 1). The maximum number of segments system-wide
    (<varname>SHMMNI</>) or per-process (<varname>SHMSEG</>) are unlikely
    to cause a problem unless your system has them set to zero.
   </para>

   <para>
    <productname>PostgreSQL</> uses one semaphore per allowed connection
    (<xref linkend="guc-max-connections">) and allowed autovacuum worker
    process (<xref linkend="guc-autovacuum-max-workers">), 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</>, which consequently must be at least
    as high as <varname>max_connections</> plus
    <varname>autovacuum_max_workers</>, plus one extra for each 16
    allowed connections plus workers (see the formula in <xref
    linkend="sysvipc-parameters">).  The parameter <varname>SEMMNI</>
    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) / 16)</>.
    Lowering the number
    of allowed connections is a temporary workaround for failures,
    which are usually confusingly worded <quote>No space
    left on device</>, from the function <function>semget</>.
   </para>

   <para>
    In some cases it might also be necessary to increase
    <varname>SEMMAP</> to be at least on the order of
    <varname>SEMMNS</>. This parameter 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>
    The <varname>SEMMSL</> parameter, which determines how many
    semaphores can be in a set, must be at least 17 for
    <productname>PostgreSQL</>.
   </para>

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


    <variablelist>
     <varlistentry>
      <term><systemitem class="osname">AIX</></term>
      <indexterm><primary>AIX</><secondary>IPC configuration</></>
      <listitem>
       <para>
        At least as of version 5.1, 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">BSD/OS</></term>
      <indexterm><primary>BSD/OS</><secondary>IPC configuration</></>
      <listitem>
       <formalpara>
        <title>Shared Memory</>
        <para>
         By default, only 4 MB of shared memory is supported. Keep in
         mind that shared memory is not pageable; it is locked in RAM.
         To increase the amount of shared memory supported by your
         system, add something like the following to your kernel configuration
         file:
<programlisting>
options "SHMALL=8192"
options "SHMMAX=\(SHMALL*PAGE_SIZE\)"
</programlisting>
         <varname>SHMALL</> is measured in 4 kB pages, so a value of
         1024 represents 4 MB of shared memory. Therefore the above increases
         the maximum shared memory area to 32 MB.
         For those running 4.3 or later, you will probably also need to increase
         <varname>KERNEL_VIRTUAL_MB</> above the default <literal>248</>.
         Once all changes have been made, recompile the kernel, and reboot.
        </para>
       </formalpara>

       <formalpara>
        <title>Semaphores</>
        <para>
         You will probably want to increase the number of semaphores
         as well; the default system total of 60 will only allow about
         50 <productname>PostgreSQL</productname> connections.  Set the
         values you want in your kernel configuration file, e.g.:
<programlisting>
options "SEMMNI=40"
options "SEMMNS=240"
</programlisting>
        </para>
       </formalpara>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">FreeBSD</></term>
      <indexterm><primary>FreeBSD</><secondary>IPC configuration</></>
      <listitem>
       <para>
        The default settings are only suitable for small installations
        (for example, default <varname>SHMMAX</varname> is 32
        MB). Changes can be made via the <command>sysctl</command> or
        <command>loader</command> interfaces.  The following
        parameters can be set using <command>sysctl</command>:
<screen>
<prompt>$</prompt> <userinput>sysctl -w kern.ipc.shmall=32768</userinput>
<prompt>$</prompt> <userinput>sysctl -w kern.ipc.shmmax=134217728</userinput>
<prompt>$</prompt> <userinput>sysctl -w kern.ipc.semmap=256</userinput>
</screen>
        To have these settings persist over reboots, modify
        <filename>/etc/sysctl.conf</filename>.
       </para>

       <para>
        The remaining semaphore settings are read-only as far as
        <command>sysctl</command> is concerned, but can be changed
        before boot using the <command>loader</command> prompt:
<screen>
<prompt>(loader)</prompt> <userinput>set kern.ipc.semmni=256</userinput>
<prompt>(loader)</prompt> <userinput>set kern.ipc.semmns=512</userinput>
<prompt>(loader)</prompt> <userinput>set kern.ipc.semmnu=256</userinput>
</screen>
        Similarly these can be saved between reboots in
        <filename>/boot/loader.conf</filename>.
       </para>

       <para>
        You might also want to configure your kernel to lock 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 FreeBSD jails by enabling <application>sysctl</>'s
        <literal>security.jail.sysvipc_allowed</>, <application>postmaster</>s
        running in different jails should be run by different operating system
        users.  This improves security because it prevents non-root users
        from interfering with shared memory or semaphores in different jails,
        and it allows the PostgreSQL IPC cleanup code to function properly.
        (In FreeBSD 6.0 and later the IPC cleanup code does not properly detect
        processes in other jails, preventing the running of postmasters on the
        same port in different jails.)
       </para>

       <para>
        <systemitem class="osname">FreeBSD</> versions before 4.0 work like 
        <systemitem class="osname">NetBSD</> and <systemitem class="osname">
        OpenBSD</> (see below).
       </para>
      </listitem>
     </varlistentry>

     <varlistentry>
      <term><systemitem class="osname">NetBSD</></term>
      <term><systemitem class="osname">OpenBSD</></term>
      <indexterm><primary>NetBSD</><secondary>IPC configuration</></>
      <indexterm><primary>OpenBSD</><secondary>IPC configuration</></>
      <listitem>
       <para>
        The options <varname>SYSVSHM</> and <varname>SYSVSEM</> need
        to be enabled when the kernel is compiled. (They are by
        default.) The maximum size of shared memory is determined by
        the option <varname>SHMMAXPGS</> (in pages). The following
        shows an example of how to set the various parameters on
        <systemitem class="osname">NetBSD</>
        (<systemitem class="osname">OpenBSD</> uses <literal>option</> instead):
<programlisting>
options        SYSVSHM
options        SHMMAXPGS=4096
options        SHMSEG=256

options        SYSVSEM
options        SEMMNI=256
options        SEMMNS=512
options        SEMMNU=256
options        SEMMAP=256
</programlisting>
       </para>

       <para>
        You might also want to configure your kernel to lock 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">HP-UX</></term>
      <indexterm><primary>HP-UX</><secondary>IPC configuration</></>
      <listitem>
       <para>
        The default settings tend to suffice for normal installations.
        On <productname>HP-UX</> 10, the factory default for
        <varname>SEMMNS</> is 128, which might be too low for larger
        database sites.
       </para>
       <para>
        <acronym>IPC</> parameters can be set in the <application>System
        Administration Manager</> (<acronym>SAM</>) under
        <menuchoice><guimenu>Kernel
        Configuration</><guimenuitem>Configurable Parameters</></>. Choose
        <guibutton>Create A New Kernel</> when you're done.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">Linux</></term>
      <indexterm><primary>Linux</><secondary>IPC configuration</></>
      <listitem>
       <para>
        The default maximum segment size is 32 MB, which is only adequate
        for very small <productname>PostgreSQL</productname>
        installations.  The default maximum total size is 2097152
        pages.  A page is almost always 4096 bytes except in unusual
        kernel configurations with <quote>huge pages</quote>
        (use <literal>getconf PAGE_SIZE</literal> to verify).  That
        makes a default limit of 8 GB, which is often enough, but not
        always.
       </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>
        In addition these settings can be preserved between reboots in
        the file <filename>/etc/sysctl.conf</filename>.  Doing that is
        highly recommended.
       </para>

       <para>
        Ancient distributions might not have the <command>sysctl</command> program,
        but equivalent changes can be made by manipulating the 
        <filename>/proc</filename> file system:
<screen>
<prompt>$</prompt> <userinput>echo 17179869184 &gt;/proc/sys/kernel/shmmax</userinput>
<prompt>$</prompt> <userinput>echo 4194304 &gt;/proc/sys/kernel/shmall</userinput>
</screen>
       </para>

       <para>
        The remaining defaults are quite generously sized, and usually
        do not require changes.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">MacOS X</></term>
      <indexterm><primary>MacOS X</><secondary>IPC configuration</></>
      <listitem>
       <para>
        The recommended method for configuring shared memory in OS X
        is to create a file named <filename>/etc/sysctl.conf</>,
        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 OS X versions,
        <emphasis>all five</> shared-memory parameters must be set in
        <filename>/etc/sysctl.conf</>, else the values will be ignored.
       </para>

       <para>
        Beware that recent releases of OS X ignore attempts to set
        <varname>SHMMAX</> to a value that isn't an exact multiple of 4096.
       </para>

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

       <para>
        In older OS X versions, you will need to reboot to have changes in the
        shared memory parameters take effect.  As of 10.5 it is possible to
        change all but <varname>SHMMNI</> on the fly, using
        <application>sysctl</>.  But it's still best to set up your preferred
        values via <filename>/etc/sysctl.conf</>, so that the values will be
        kept across reboots.
       </para>

       <para>
        The file <filename>/etc/sysctl.conf</> is only honored in OS X
        10.3.9 and later.  If you are running a previous 10.3.x release,
        you must edit the file <filename>/etc/rc</>
        and change the values in the following commands:
<programlisting>
sysctl -w kern.sysv.shmmax
sysctl -w kern.sysv.shmmin
sysctl -w kern.sysv.shmmni
sysctl -w kern.sysv.shmseg
sysctl -w kern.sysv.shmall
</programlisting>
        Note that
        <filename>/etc/rc</> is usually overwritten by OS X system updates,
        so you should expect to have to redo these edits after each update.
       </para>

       <para>
        In OS X 10.2 and earlier, instead edit these commands in the file
        <filename>/System/Library/StartupItems/SystemTuning/SystemTuning</>.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">SCO OpenServer</></term>
      <indexterm><primary>SCO OpenServer</><secondary>IPC configuration</></>
      <listitem>
       <para>
        In the default configuration, only 512 kB of shared memory per
        segment is allowed. To increase the setting, first change to the
        directory <filename>/etc/conf/cf.d</>. To display the current value of
        <varname>SHMMAX</>, run:
<programlisting>
./configure -y SHMMAX
</programlisting>
        To set a new value for <varname>SHMMAX</>, run:
<programlisting>
./configure SHMMAX=<replaceable>value</>
</programlisting>
        where <replaceable>value</> is the new value you want to use
        (in bytes). After setting <varname>SHMMAX</>, rebuild the kernel:
<programlisting>
./link_unix
</programlisting>
        and reboot.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">Solaris</></term>
      <indexterm><primary>Solaris</><secondary>IPC configuration</></>
      <listitem>
       <para>
        At least in version 2.6, the default maximum size of a shared
        memory segment is too low for <productname>PostgreSQL</>. The
        relevant settings can be changed in <filename>/etc/system</>,
        for example:
<programlisting>
set shmsys:shminfo_shmmax=0x2000000
set shmsys:shminfo_shmmin=1
set shmsys:shminfo_shmmni=256
set shmsys:shminfo_shmseg=256

set semsys:seminfo_semmap=256
set semsys:seminfo_semmni=512
set semsys:seminfo_semmns=512
set semsys:seminfo_semmsl=32
</programlisting>
        You need to reboot for the changes to take effect.
       </para>

       <para>
        See also <ulink
        url="http://sunsite.uakom.sk/sunworldonline/swol-09-1997/swol-09-insidesolaris.html"></>
        for information on shared memory under
        <productname>Solaris</>.
       </para>
      </listitem>
     </varlistentry>


     <varlistentry>
      <term><systemitem class="osname">UnixWare</></term>
      <indexterm><primary>UnixWare</><secondary>IPC configuration</></>
      <listitem>
       <para>
        On <productname>UnixWare</> 7, the maximum size for shared
        memory segments is only 512 kB in the default configuration.
        To display the current value of <varname>SHMMAX</>, run:
<programlisting>
/etc/conf/bin/idtune -g SHMMAX
</programlisting>
        which displays the current, default, minimum, and maximum
        values. To set a new value for <varname>SHMMAX</>,
        run:
<programlisting>
/etc/conf/bin/idtune SHMMAX <replaceable>value</>
</programlisting>
        where <replaceable>value</> is the new value you want to use
        (in bytes). After setting <varname>SHMMAX</>, rebuild the
        kernel:
<programlisting>
/etc/conf/bin/idbuild -B
</programlisting>
        and reboot.
       </para>
      </listitem>
     </varlistentry>

    </variablelist>


   <table id="shared-memory-parameters">
    <title><productname>PostgreSQL</productname> shared memory usage</>

    <tgroup cols="2">
     <thead>
      <row>
       <entry>Usage</>
       <entry>Approximate shared memory bytes required (as of 8.3)</>
      </row>
     </thead>

     <tbody>
      <row>
       <entry>Connections</>
       <entry>(1800 + 270 * <xref
       linkend="guc-max-locks-per-transaction">) * <xref
       linkend="guc-max-connections"></entry>
      </row>

      <row>
       <entry>Autovacuum workers</>
       <entry>(1800 + 270 * <xref
       linkend="guc-max-locks-per-transaction">) * <xref
       linkend="guc-autovacuum-max-workers"></entry>
      </row>

      <row>
       <entry>Prepared transactions</>
       <entry>(770 + 270 * <xref
       linkend="guc-max-locks-per-transaction">) * <xref linkend="guc-max-prepared-transactions"></entry>
      </row>

      <row>
       <entry>Shared disk buffers</>
       <entry>(<xref linkend="guc-block-size"> + 208) * <xref linkend="guc-shared-buffers"></entry>
      </row>

      <row>
       <entry>WAL buffers</>
       <entry>(<xref linkend="guc-wal-block-size"> + 8) * <xref linkend="guc-wal-buffers"></entry>
      </row>

      <row>
       <entry>Fixed space requirements</>
       <entry>770 kB</entry>
      </row>
     </tbody>
    </tgroup>
   </table>

  </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</>) 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</>'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>

   <para>
    In Linux 2.4 and later, the default virtual memory behavior 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> server (the
    master server process) if the memory demands of
    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>
    On Linux 2.6 and later, it is possible to modify the
    kernel's behavior so that it will not <quote>overcommit</> memory.
    Although this setting will not prevent the <ulink
    url="http://lwn.net/Articles/104179/">OOM killer</> 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</>.
    You might also wish to modify the related setting
    <varname>vm.overcommit_ratio</>.  For details see the kernel documentation
    file <filename>Documentation/vm/overcommit-accounting</>.
   </para>

   <para>
    Another approach, which can be used with or without altering
    <varname>vm.overcommit_memory</>, is to set the process-specific
    <varname>oom_adj</> value for the postmaster process to <literal>-17</>,
    thereby guaranteeing it will not be targeted by the OOM killer.  The
    simplest way to do this is to execute
<programlisting>
echo -17 > /proc/self/oom_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 may also wish to build <productname>PostgreSQL</>
    with <literal>-DLINUX_OOM_ADJ=0</> added to <varname>CFLAGS</>.
    That will cause postmaster child processes to run with the normal
    <varname>oom_adj</> value of zero, so that the OOM killer can still
    target them at need.
   </para>

   <note>
   <para>
    Some vendors' Linux 2.4 kernels are reported to have early versions
    of the 2.6 overcommit <command>sysctl</command> parameter.  However, setting
    <literal>vm.overcommit_memory</> to 2
    on a 2.4 kernel that does not have the relevant code will make
    things worse, not better.  It is recommended that you inspect
    the actual kernel source code (see the function
    <function>vm_enough_memory</> in the file <filename>mm/mmap.c</>)
    to verify what is supported in your kernel before you try this in a 2.4
    installation.  The presence of the <filename>overcommit-accounting</>
    documentation file should <emphasis>not</> be taken as evidence that the
    feature is there.  If in any doubt, consult a kernel expert or your
    kernel vendor.
   </para>
   </note>
  </sect2>
 </sect1>


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

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

  <para>
   There are several ways to shut down the database server. You control
   the type of shutdown by sending different signals to the master
   <command>postgres</command> process.

   <variablelist>
    <varlistentry>
     <term><systemitem>SIGTERM</systemitem><indexterm><primary>SIGTERM</></></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</></></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</></></term>
     <listitem>
      <para>
      This is the <firstterm>Immediate Shutdown</firstterm> mode.
      The master <command>postgres</command> process will send a
      <systemitem>SIGQUIT</systemitem> to all child processes and exit
      immediately, without properly shutting itself down. The child processes
      likewise exit immediately upon receiving
      <systemitem>SIGQUIT</systemitem>. 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</>
   on non-Windows systems.
   The <acronym>PID</> 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, which might then have to be done
    manually before a new server can be started.  Furthermore,
    <systemitem>SIGKILL</systemitem> kills the <command>postgres</command>
    process without letting it relay the signal to its subprocesses,
    so it will 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()</> (see <xref
   linkend="functions-admin-signal-table">) or send a
   <systemitem>SIGTERM</> signal to the child process associated with
   the session.
  </para>
 </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</> 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>
   The simplest way to prevent spoofing for <literal>local</>
   connections is to use a Unix domain socket directory (<xref
   linkend="guc-unix-socket-directory">) 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</> for the
   socket file and hence be vulnerable to spoofing, during operating system
   startup create a symbolic link <filename>/tmp/.s.PGSQL.5432</> that points
   to the relocated socket file.  You also might need to modify your
   <filename>/tmp</> cleanup script to prevent removal of the symbolic link.
  </para>

  <para>
   To prevent spoofing on TCP connections, the best solution is to use
   SSL certificates and make sure that clients check the server's certificate.
   To do that, the server
   must be configured to accept only <literal>hostssl</> connections (<xref
   linkend="auth-pg-hba-conf">) and have SSL
   <filename>server.key</filename> (key) and
   <filename>server.crt</filename> (certificate) files (<xref
   linkend="ssl-tcp">). The TCP client must connect using
   <literal>sslmode=verify-ca</> or
   <literal>verify-full</> and have the appropriate root certificate
   file installed (<xref linkend="libpq-connect">).
  </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 Storage Encryption</term>
   <listitem>

    <para>
     By default, database user passwords are stored as MD5 hashes, so
     the administrator cannot determine the actual password assigned
     to the user. If MD5 encryption is used for client authentication,
     the unencrypted password is never even temporarily present on the
     server because the client MD5-encrypts it before being sent
     across the network.
    </para>
   </listitem>
  </varlistentry>

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

   <listitem>
    <para>
     The <filename>contrib</> function library
     <link linkend="pgcrypto"><function>pgcrypto</function></link>
     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>
     On Linux, encryption can be layered on top of a file system
     using a <quote>loopback device</quote>. This allows an entire
     file system partition to be encrypted on disk, and decrypted by the
     operating system. On FreeBSD, the equivalent facility is called
     GEOM Based Disk Encryption (<acronym>gbde</acronym>), and 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 Passwords Across A Network</term>

   <listitem>
     <para>
      The <literal>MD5</> authentication method double-encrypts the
      password on the client before sending it to the server. It first
      MD5-encrypts it based on the user name, and then encrypts it
      based on a random salt sent by the server when the database
      connection was made. It is this double-encrypted value that is
      sent over the network to the server. Double-encryption not only
      prevents the password from being discovered, it also prevents
      another connection from using the same encrypted password to
      connect to the database server at a later time.
     </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</> file allows administrators to specify
      which hosts can use non-encrypted connections (<literal>host</>)
      and which require SSL-encrypted connections
      (<literal>hostssl</>). Also, clients can specify that they
      connect to servers only via SSL. <application>Stunnel</> or
      <application>SSH</> 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</>
     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</> has native support for using
   <acronym>SSL</> 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</> is
   enabled at build time (see <xref linkend="installation">).
  </para>

  <para>
   With <acronym>SSL</> support compiled in, the
   <productname>PostgreSQL</> server can be started with
   <acronym>SSL</> enabled by setting the parameter
   <xref linkend="guc-ssl"> to <literal>on</> in
   <filename>postgresql.conf</>.  The server will listen for both normal
   and <acronym>SSL</> connections on the same TCP port, and will negotiate
   with any connecting client on whether to use <acronym>SSL</>.  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</> for some or all connections.
  </para>

  <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</>.
   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</>.
  </para>

  <note>
   <para>
    It is possible to have authentication without encryption overhead by
    using <literal>NULL-SHA</> or <literal>NULL-MD5</> 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>

  <para>
   To start in <acronym>SSL</> mode, the files <filename>server.crt</>
   and <filename>server.key</> must exist in the server's data directory.
   These files should contain the server certificate and private key,
   respectively.
   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>.
   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.
  </para>

  <para>
   In some cases, the server certificate might be signed by an
   <quote>intermediate</> certificate authority, rather than one that is
   directly trusted by clients.  To use such a certificate, append the
   certificate of the signing authority to the <filename>server.crt</> file,
   then its parent authority's certificate, and so on up to a <quote>root</>
   authority that is trusted by the clients.  The root certificate should
   be included in every case where <filename>server.crt</> contains more than
   one certificate.
  </para>

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

   <para>
   To require the client to supply a trusted certificate, place
   certificates of the certificate authorities (<acronym>CA</acronym>s)
   you trust in the file <filename>root.crt</filename> in the data
   directory, and set the <literal>clientcert</literal> parameter
   to <literal>1</literal> on the appropriate <literal>hostssl</> line(s) in
   <filename>pg_hba.conf</>.
   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.)  The server will
   verify that the client's certificate is signed by one of the trusted
   certificate authorities.  Certificate Revocation List (CRL) entries
   are also checked if the file <filename>root.crl</filename> exists.
   <!-- If this URL changes replace it with a URL to www.archive.org. -->
   (See <ulink
   url="http://h71000.www7.hp.com/DOC/83final/BA554_90007/ch04s02.html"></>
   for diagrams showing SSL certificate usage.)
  </para>

  <para>
   The <literal>clientcert</literal> option in <filename>pg_hba.conf</> is
   available for all authentication methods, but only for rows specified as
   <literal>hostssl</>.  When <literal>clientcert</literal> is not specified
   or is set to <literal>0</>, the server will still verify presented client
   certificates against <filename>root.crt</filename> if that file exists
   &mdash; but it will not insist that a client certificate be presented.
  </para>

  <para>
   Note that <filename>root.crt</filename> lists the top-level CAs that are
   considered trusted for signing client certificates.  In principle it need
   not list the CA that signed the server's certificate, though in most cases
   that CA would also be trusted for client certificates.
  </para>

  <para>
   If you are setting up client certificates, you may wish to use
   the <literal>cert</> authentication method, so that the certificates
   control user authentication as well as providing connection security.
   See <xref linkend="auth-cert"> for details.
  </para>
  </sect2>

  <sect2 id="ssl-server-files">
   <title>SSL Server File Usage</title>
  <para>
   The files <filename>server.key</>, <filename>server.crt</>,
   <filename>root.crt</filename>, and <filename>root.crl</filename>
   are only examined during server start; so you must restart
   the server for changes in them to take effect.
  </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><filename>server.crt</></entry>
      <entry>server certificate</entry>
      <entry>sent to client to indicate server's identity</entry>
     </row>

     <row>
      <entry><filename>server.key</></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><filename>root.crt</></entry>
      <entry>trusted certificate authorities</entry>
      <entry>checks that client certificate is
      signed by a trusted certificate authority</entry>
     </row>

     <row>
      <entry><filename>root.crl</></entry>
      <entry>certificates revoked by certificate authorities</entry>
      <entry>client certificate must not be on this list</entry>
     </row>

    </tbody>
   </tgroup>
  </table>
  </sect2>

  <sect2 id="ssl-certificate-creation">
   <title>Creating a Self-Signed Certificate</title>

   <para>
    To create a quick self-signed certificate for the server, use the
    following <productname>OpenSSL</productname> command:
<programlisting>
openssl req -new -text -out server.req
</programlisting>
    Fill out the information that <application>openssl</> asks for. Make sure
    you enter the local host name as <quote>Common Name</>; the challenge
    password can be left blank. The program will generate a key that is
    passphrase protected; it will not accept a passphrase that is less
    than four characters long. To remove the passphrase (as you must if
    you want automatic start-up of the server), run the commands:
<programlisting>
openssl rsa -in privkey.pem -out server.key
rm privkey.pem
</programlisting>
    Enter the old passphrase to unlock the existing key. Now do:
<programlisting>
openssl req -x509 -in server.req -text -key server.key -out server.crt
</programlisting>
    to turn the certificate into a self-signed certificate and to copy
    the key and certificate to where the server will look for them.
    Finally 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</> documentation.
   </para>

   <para>
    A self-signed certificate can be used for testing, but a certificate
    signed by a certificate authority (<acronym>CA</>) (either one of the
    global <acronym>CAs</> or a local one) should be used in production
    so that clients can verify the server's identity. If all the clients
    are local to the organization, using a local <acronym>CA</> is
    recommended.
   </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. Then you can establish a secure
   tunnel with a command like this from the client machine:
<programlisting>
ssh -L 63333:localhost:5432 [email protected]
</programlisting>
   The first number in the <option>-L</option> argument, 63333, is the
   port number of your end of the tunnel; it can be any unused port.
   (IANA reserves ports 49152 through 65535 for private use.)  The
   second number, 5432, is the remote end of the tunnel: the port
   number your server is using. The name or IP address between the
   port numbers is the host with the database server you are going to
   connect to, as seen from the host you are logging in to, which
   is <literal>foo.com</literal> in this example. 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 really
   user <literal>joe</literal> on host <literal>foo.com</literal>
   connecting to <literal>localhost</literal> in that context, and it
   will use whatever authentication procedure was configured for
   connections from this user and host.  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 as long as 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>[email protected]</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 the port forwarding as
<programlisting>
ssh -L 63333:foo.com:5432 [email protected]
</programlisting>
   but then the database server will see the connection as coming in
   on its <literal>foo.com</literal> interface, 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 [email protected]
</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>

</chapter>