man/daemon.xml

   1 <?xml version='1.0'?> <!--*-nxml-*-->
   2 <!DOCTYPE refentry PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
   3         "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd">
   4
   5 <!--
   6   This file is part of systemd.
   7
   8   Copyright 2010 Lennart Poettering
   9
  10   systemd is free software; you can redistribute it and/or modify it
  11   under the terms of the GNU Lesser General Public License as published by
  12   the Free Software Foundation; either version 2.1 of the License, or
  13   (at your option) any later version.
  14
  15   systemd is distributed in the hope that it will be useful, but
  16   WITHOUT ANY WARRANTY; without even the implied warranty of
  17   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  18   Lesser General Public License for more details.
  19
  20   You should have received a copy of the GNU Lesser General Public License
  21   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  22 -->
  23
  24 <refentry id="daemon">
  25
  26         <refentryinfo>
  27                 <title>daemon</title>
  28                 <productname>systemd</productname>
  29
  30                 <authorgroup>
  31                         <author>
  32                                 <contrib>Developer</contrib>
  33                                 <firstname>Lennart</firstname>
  34                                 <surname>Poettering</surname>
  35                                 <email>lennart@poettering.net</email>
  36                         </author>
  37                 </authorgroup>
  38         </refentryinfo>
  39
  40         <refmeta>
  41                 <refentrytitle>daemon</refentrytitle>
  42                 <manvolnum>7</manvolnum>
  43         </refmeta>
  44
  45         <refnamediv>
  46                 <refname>daemon</refname>
  47                 <refpurpose>Writing and packaging system daemons</refpurpose>
  48         </refnamediv>
  49
  50         <refsect1>
  51                 <title>Description</title>
  52
  53                 <para>A daemon is a service process that runs in the
  54                 background and supervises the system or provides
  55                 functionality to other processes. Traditionally,
  56                 daemons are implemented following a scheme originating
  57                 in SysV Unix. Modern daemons should follow a simpler
  58                 yet more powerful scheme (here called "new-style"
  59                 daemons), as implemented by
  60                 <citerefentry><refentrytitle>systemd</refentrytitle><manvolnum>1</manvolnum></citerefentry>. This
  61                 manual page covers both schemes, and in
  62                 particular includes recommendations for daemons that
  63                 shall be included in the systemd init system.</para>
  64
  65                 <refsect2>
  66                         <title>SysV Daemons</title>
  67
  68                         <para>When a traditional SysV daemon
  69                         starts, it should execute the following steps
  70                         as part of the initialization. Note that these
  71                         steps are unnecessary for new-style daemons (see below),
  72                         and should only be implemented if compatibility
  73                         with SysV is essential.</para>
  74
  75                         <orderedlist>
  76                                 <listitem><para>Close all open file
  77                                 descriptors except standard input, output,
  78                                 and error (i.e. the first three file
  79                                 descriptors 0, 1, 2). This ensures
  80                                 that no accidentally passed file
  81                                 descriptor stays around in the daemon
  82                                 process. On Linux, this is best
  83                                 implemented by iterating through
  84                                 <filename>/proc/self/fd</filename>,
  85                                 with a fallback of iterating from file
  86                                 descriptor 3 to the value returned by
  87                                 <function>getrlimit()</function> for
  88                                 RLIMIT_NOFILE.</para></listitem>
  89
  90                                 <listitem><para>Reset all signal
  91                                 handlers to their default. This is
  92                                 best done by iterating through the
  93                                 available signals up to the limit of
  94                                 _NSIG and resetting them to
  95                                 <constant>SIG_DFL</constant>.</para></listitem>
  96
  97                                 <listitem><para>Reset the signal mask
  98                                 using
  99                                 <function>sigprocmask()</function>.</para></listitem>
 100
 101                                 <listitem><para>Sanitize the
 102                                 environment block, removing or
 103                                 resetting environment variables that
 104                                 might negatively impact daemon
 105                                 runtime.</para></listitem>
 106
 107                                 <listitem><para>Call <function>fork()</function>,
 108                                 to create a background
 109                                 process.</para></listitem>
 110
 111                                 <listitem><para>In the child, call
 112                                 <function>setsid()</function> to
 113                                 detach from any terminal and create an
 114                                 independent session.</para></listitem>
 115
 116                                 <listitem><para>In the child, call
 117                                 <function>fork()</function> again, to
 118                                 ensure that the daemon can never re-acquire
 119                                 a terminal again.</para></listitem>
 120
 121                                 <listitem><para>Call <function>exit()</function> in the
 122                                 first child, so that only the second
 123                                 child (the actual daemon process)
 124                                 stays around. This ensures that the
 125                                 daemon process is re-parented to
 126                                 init/PID 1, as all daemons should
 127                                 be.</para></listitem>
 128
 129                                 <listitem><para>In the daemon process,
 130                                 connect <filename>/dev/null</filename>
 131                                 to standard input, output, and error.
 132                                 </para></listitem>
 133
 134                                 <listitem><para>In the daemon process,
 135                                 reset the umask to 0, so that the file
 136                                 modes passed to <function>open()</function>, <function>mkdir()</function> and
 137                                 suchlike directly control the access
 138                                 mode of the created files and
 139                                 directories.</para></listitem>
 140
 141                                 <listitem><para>In the daemon process,
 142                                 change the current directory to the
 143                                 root directory (/), in order to avoid
 144                                 that the daemon involuntarily
 145                                 blocks mount points from being
 146                                 unmounted.</para></listitem>
 147
 148                                 <listitem><para>In the daemon process,
 149                                 write the daemon PID (as returned by
 150                                 <function>getpid()</function>) to a
 151                                 PID file, for example
 152                                 <filename>/run/foobar.pid</filename>
 153                                 (for a hypothetical daemon "foobar")
 154                                 to ensure that the daemon cannot be
 155                                 started more than once. This must be
 156                                 implemented in race-free fashion so
 157                                 that the PID file is only updated when
 158                                 it is verified at the same time that
 159                                 the PID previously stored in the PID
 160                                 file no longer exists or belongs to a
 161                                 foreign process.</para></listitem>
 162
 163                                 <listitem><para>In the daemon process,
 164                                 drop privileges, if possible and
 165                                 applicable.</para></listitem>
 166
 167                                 <listitem><para>From the daemon
 168                                 process, notify the original process
 169                                 started that initialization is
 170                                 complete. This can be implemented via
 171                                 an unnamed pipe or similar
 172                                 communication channel that is created
 173                                 before the first
 174                                 <function>fork()</function> and hence
 175                                 available in both the original and the
 176                                 daemon process.</para></listitem>
 177
 178                                 <listitem><para>Call
 179                                 <function>exit()</function> in the
 180                                 original process. The process that
 181                                 invoked the daemon must be able to
 182                                 rely on that this
 183                                 <function>exit()</function> happens
 184                                 after initialization is complete and
 185                                 all external communication channels
 186                                 are established and
 187                                 accessible.</para></listitem>
 188                         </orderedlist>
 189
 190                         <para>The BSD <function>daemon()</function> function should not be
 191                         used, as it implements only a subset of these steps.</para>
 192
 193                         <para>A daemon that needs to provide
 194                         compatibility with SysV systems should
 195                         implement the scheme pointed out
 196                         above. However, it is recommended to make this
 197                         behavior optional and configurable via a
 198                         command line argument to ease debugging as
 199                         well as to simplify integration into systems
 200                         using systemd.</para>
 201                 </refsect2>
 202
 203                 <refsect2>
 204                         <title>New-Style Daemons</title>
 205
 206                         <para>Modern services for Linux should be
 207                         implemented as new-style daemons. This makes it
 208                         easier to supervise and control them at
 209                         runtime and simplifies their
 210                         implementation.</para>
 211
 212                         <para>For developing a new-style daemon, none
 213                         of the initialization steps recommended for
 214                         SysV daemons need to be implemented. New-style
 215                         init systems such as systemd make all of them
 216                         redundant. Moreover, since some of these steps
 217                         interfere with process monitoring, file
 218                         descriptor passing and other functionality of
 219                         the init system, it is recommended not to
 220                         execute them when run as new-style
 221                         service.</para>
 222
 223                         <para>Note that new-style init systems
 224                         guarantee execution of daemon processes in a
 225                         clean process context: it is guaranteed that
 226                         the environment block is sanitized, that the
 227                         signal handlers and mask is reset and that no
 228                         left-over file descriptors are passed. Daemons
 229                         will be executed in their own session, with
 230                         standard input/output/error connected to
 231                         <filename>/dev/null</filename> unless
 232                         otherwise configured. The umask is reset.
 233                         </para>
 234
 235                         <para>It is recommended for new-style daemons
 236                         to implement the following:</para>
 237
 238                         <orderedlist>
 239                                 <listitem><para>If <constant>SIGTERM</constant> is
 240                                 received, shut down the daemon and
 241                                 exit cleanly.</para></listitem>
 242
 243                                 <listitem><para>If <constant>SIGHUP</constant> is received,
 244                                 reload the configuration files, if
 245                                 this applies.</para></listitem>
 246
 247                                 <listitem><para>Provide a correct exit
 248                                 code from the main daemon process, as
 249                                 this is used by the init system to
 250                                 detect service errors and problems. It
 251                                 is recommended to follow the exit code
 252                                 scheme as defined in the <ulink
 253                                 url="http://refspecs.linuxbase.org/LSB_3.1.1/LSB-Core-generic/LSB-Core-generic/iniscrptact.html">LSB
 254                                 recommendations for SysV init
 255                                 scripts</ulink>.</para></listitem>
 256
 257                                 <listitem><para>If possible and
 258                                 applicable, expose the daemon's control
 259                                 interface via the D-Bus IPC system and
 260                                 grab a bus name as last step of
 261                                 initialization.</para></listitem>
 262
 263                                 <listitem><para>For integration in
 264                                 systemd, provide a
 265                                 <filename>.service</filename> unit
 266                                 file that carries information about
 267                                 starting, stopping and otherwise
 268                                 maintaining the daemon. See
 269                                 <citerefentry><refentrytitle>systemd.service</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 270                                 for details.</para></listitem>
 271
 272                                 <listitem><para>As much as possible,
 273                                 rely on the init system's
 274                                 functionality to limit the access of
 275                                 the daemon to files, services and
 276                                 other resources, i.e. in the case of
 277                                 systemd, rely on systemd's resource
 278                                 limit control instead of implementing
 279                                 your own, rely on systemd's privilege
 280                                 dropping code instead of implementing
 281                                 it in the daemon, and similar. See
 282                                 <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 283                                 for the available
 284                                 controls.</para></listitem>
 285
 286                                 <listitem><para>If D-Bus is used, make
 287                                 your daemon bus-activatable by
 288                                 supplying a D-Bus service activation
 289                                 configuration file. This has multiple
 290                                 advantages: your daemon may be started
 291                                 lazily on-demand; it may be started in
 292                                 parallel to other daemons requiring it
 293                                 -- which maximizes parallelization and
 294                                 boot-up speed; your daemon can be
 295                                 restarted on failure without losing
 296                                 any bus requests, as the bus queues
 297                                 requests for activatable services. See
 298                                 below for details.</para></listitem>
 299
 300                                 <listitem><para>If your daemon
 301                                 provides services to other local
 302                                 processes or remote clients via a
 303                                 socket, it should be made
 304                                 socket-activatable following the
 305                                 scheme pointed out below. Like D-Bus
 306                                 activation, this enables on-demand
 307                                 starting of services as well as it
 308                                 allows improved parallelization of
 309                                 service start-up. Also, for state-less
 310                                 protocols (such as syslog, DNS), a
 311                                 daemon implementing socket-based
 312                                 activation can be restarted without
 313                                 losing a single request. See below for
 314                                 details.</para></listitem>
 315
 316                                 <listitem><para>If applicable, a daemon
 317                                 should notify the init system about
 318                                 startup completion or status updates
 319                                 via the
 320                                 <citerefentry><refentrytitle>sd_notify</refentrytitle><manvolnum>3</manvolnum></citerefentry>
 321                                 interface.</para></listitem>
 322
 323                                 <listitem><para>Instead of using the
 324                                 <function>syslog()</function> call to
 325                                 log directly to the system syslog
 326                                 service, a new-style daemon may choose
 327                                 to simply log to standard error via
 328                                 <function>fprintf()</function>, which
 329                                 is then forwarded to syslog by the
 330                                 init system. If log priorities are
 331                                 necessary, these can be encoded by
 332                                 prefixing individual log lines with
 333                                 strings like "&lt;4&gt;" (for log
 334                                 priority 4 "WARNING" in the syslog
 335                                 priority scheme), following a similar
 336                                 style as the Linux kernel's
 337                                 <function>printk()</function> priority
 338                                 system. For details, see
 339                                 <citerefentry><refentrytitle>sd-daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>
 340                                 and
 341                                 <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para></listitem>
 342
 343                         </orderedlist>
 344
 345                         <para>These recommendations are similar but
 346                         not identical to the <ulink
 347                         url="https://developer.apple.com/library/mac/documentation/MacOSX/Conceptual/BPSystemStartup/Chapters/CreatingLaunchdJobs.html">Apple
 348                         MacOS X Daemon Requirements</ulink>.</para>
 349                 </refsect2>
 350
 351         </refsect1>
 352         <refsect1>
 353                 <title>Activation</title>
 354
 355                 <para>New-style init systems provide multiple
 356                 additional mechanisms to activate services, as
 357                 detailed below. It is common that services are
 358                 configured to be activated via more than one mechanism
 359                 at the same time. An example for systemd:
 360                 <filename>bluetoothd.service</filename> might get
 361                 activated either when Bluetooth hardware is plugged
 362                 in, or when an application accesses its programming
 363                 interfaces via D-Bus. Or, a print server daemon might
 364                 get activated when traffic arrives at an IPP port, or
 365                 when a printer is plugged in, or when a file is queued
 366                 in the printer spool directory. Even for services that
 367                 are intended to be started on system bootup
 368                 unconditionally, it is a good idea to implement some of
 369                 the various activation schemes outlined below, in
 370                 order to maximize parallelization. If a daemon
 371                 implements a D-Bus service or listening socket,
 372                 implementing the full bus and socket activation scheme
 373                 allows starting of the daemon with its clients in
 374                 parallel (which speeds up boot-up), since all its
 375                 communication channels are established already, and no
 376                 request is lost because client requests will be queued
 377                 by the bus system (in case of D-Bus) or the kernel (in
 378                 case of sockets) until the activation is
 379                 completed.</para>
 380
 381                 <refsect2>
 382                         <title>Activation on Boot</title>
 383
 384                         <para>Old-style daemons are usually activated
 385                         exclusively on boot (and manually by the
 386                         administrator) via SysV init scripts, as
 387                         detailed in the <ulink
 388                         url="http://refspecs.linuxbase.org/LSB_3.1.1/LSB-Core-generic/LSB-Core-generic/iniscrptact.html">LSB
 389                         Linux Standard Base Core
 390                         Specification</ulink>. This method of
 391                         activation is supported ubiquitously on Linux
 392                         init systems, both old-style and new-style
 393                         systems. Among other issues, SysV init scripts
 394                         have the disadvantage of involving shell
 395                         scripts in the boot process. New-style init
 396                         systems generally employ updated versions of
 397                         activation, both during boot-up and during
 398                         runtime and using more minimal service
 399                         description files.</para>
 400
 401                         <para>In systemd, if the developer or
 402                         administrator wants to make sure that a service or
 403                         other unit is activated automatically on boot,
 404                         it is recommended to place a symlink to the
 405                         unit file in the <filename>.wants/</filename>
 406                         directory of either
 407                         <filename>multi-user.target</filename> or
 408                         <filename>graphical.target</filename>, which
 409                         are normally used as boot targets at system
 410                         startup. See
 411                         <citerefentry><refentrytitle>systemd.unit</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 412                         for details about the
 413                         <filename>.wants/</filename> directories, and
 414                         <citerefentry><refentrytitle>systemd.special</refentrytitle><manvolnum>7</manvolnum></citerefentry>
 415                         for details about the two boot targets.</para>
 416
 417                 </refsect2>
 418
 419                 <refsect2>
 420                         <title>Socket-Based Activation</title>
 421
 422                         <para>In order to maximize the possible
 423                         parallelization and robustness and simplify
 424                         configuration and development, it is
 425                         recommended for all new-style daemons that
 426                         communicate via listening sockets to employ
 427                         socket-based activation. In a socket-based
 428                         activation scheme, the creation and binding of
 429                         the listening socket as primary communication
 430                         channel of daemons to local (and sometimes
 431                         remote) clients is moved out of the daemon
 432                         code and into the init system. Based on
 433                         per-daemon configuration, the init system
 434                         installs the sockets and then hands them off
 435                         to the spawned process as soon as the
 436                         respective daemon is to be started.
 437                         Optionally, activation of the service can be
 438                         delayed until the first inbound traffic
 439                         arrives at the socket to implement on-demand
 440                         activation of daemons. However, the primary
 441                         advantage of this scheme is that all providers
 442                         and all consumers of the sockets can be
 443                         started in parallel as soon as all sockets
 444                         are established. In addition to that, daemons
 445                         can be restarted with losing only a minimal
 446                         number of client transactions, or even any
 447                         client request at all (the latter is
 448                         particularly true for state-less protocols,
 449                         such as DNS or syslog), because the socket
 450                         stays bound and accessible during the restart,
 451                         and all requests are queued while the daemon
 452                         cannot process them.</para>
 453
 454                         <para>New-style daemons which support socket
 455                         activation must be able to receive their
 456                         sockets from the init system instead of
 457                         creating and binding them themselves. For
 458                         details about the programming interfaces for
 459                         this scheme provided by systemd, see
 460                         <citerefentry><refentrytitle>sd_listen_fds</refentrytitle><manvolnum>3</manvolnum></citerefentry>
 461                         and
 462                         <citerefentry><refentrytitle>sd-daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>. For
 463                         details about porting existing daemons to
 464                         socket-based activation, see below. With
 465                         minimal effort, it is possible to implement
 466                         socket-based activation in addition to
 467                         traditional internal socket creation in the
 468                         same codebase in order to support both
 469                         new-style and old-style init systems from the
 470                         same daemon binary.</para>
 471
 472                         <para>systemd implements socket-based
 473                         activation via <filename>.socket</filename>
 474                         units, which are described in
 475                         <citerefentry><refentrytitle>systemd.socket</refentrytitle><manvolnum>5</manvolnum></citerefentry>. When
 476                         configuring socket units for socket-based
 477                         activation, it is essential that all listening
 478                         sockets are pulled in by the special target
 479                         unit <filename>sockets.target</filename>. It
 480                         is recommended to place a
 481                         <varname>WantedBy=sockets.target</varname>
 482                         directive in the <literal>[Install]</literal>
 483                         section to automatically add such a
 484                         dependency on installation of a socket
 485                         unit. Unless
 486                         <varname>DefaultDependencies=no</varname> is
 487                         set, the necessary ordering dependencies are
 488                         implicitly created for all socket units. For
 489                         more information about
 490                         <filename>sockets.target</filename>, see
 491                         <citerefentry><refentrytitle>systemd.special</refentrytitle><manvolnum>7</manvolnum></citerefentry>. It
 492                         is not necessary or recommended to place any
 493                         additional dependencies on socket units (for
 494                         example from
 495                         <filename>multi-user.target</filename> or
 496                         suchlike) when one is installed in
 497                         <filename>sockets.target</filename>.</para>
 498                 </refsect2>
 499
 500                 <refsect2>
 501                         <title>Bus-Based Activation</title>
 502
 503                         <para>When the D-Bus IPC system is used for
 504                         communication with clients, new-style daemons
 505                         should employ bus activation so that they are
 506                         automatically activated when a client
 507                         application accesses their IPC
 508                         interfaces. This is configured in D-Bus
 509                         service files (not to be confused with systemd
 510                         service unit files!). To ensure that D-Bus
 511                         uses systemd to start-up and maintain the
 512                         daemon, use the
 513                         <varname>SystemdService=</varname> directive
 514                         in these service files to configure the
 515                         matching systemd service for a D-Bus
 516                         service. e.g.: For a D-Bus service whose D-Bus
 517                         activation file is named
 518                         <filename>org.freedesktop.RealtimeKit.service</filename>,
 519                         make sure to set
 520                         <varname>SystemdService=rtkit-daemon.service</varname>
 521                         in that file to bind it to the systemd
 522                         service
 523                         <filename>rtkit-daemon.service</filename>. This
 524                         is needed to make sure that the daemon is
 525                         started in a race-free fashion when activated
 526                         via multiple mechanisms simultaneously.</para>
 527                 </refsect2>
 528
 529                 <refsect2>
 530                         <title>Device-Based Activation</title>
 531
 532                         <para>Often, daemons that manage a particular
 533                         type of hardware should be activated only when
 534                         the hardware of the respective kind is plugged
 535                         in or otherwise becomes available. In a
 536                         new-style init system, it is possible to bind
 537                         activation to hardware plug/unplug events. In
 538                         systemd, kernel devices appearing in the
 539                         sysfs/udev device tree can be exposed as units
 540                         if they are tagged with the string
 541                         <literal>systemd</literal>. Like any other
 542                         kind of unit, they may then pull in other units
 543                         when activated (i.e. plugged in) and thus
 544                         implement device-based activation. systemd
 545                         dependencies may be encoded in the udev
 546                         database via the
 547                         <varname>SYSTEMD_WANTS=</varname>
 548                         property. See
 549                         <citerefentry><refentrytitle>systemd.device</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 550                         for details. Often, it is nicer to pull in
 551                         services from devices only indirectly via
 552                         dedicated targets. Example: Instead of pulling
 553                         in <filename>bluetoothd.service</filename>
 554                         from all the various bluetooth dongles and
 555                         other hardware available, pull in
 556                         bluetooth.target from them and
 557                         <filename>bluetoothd.service</filename> from
 558                         that target. This provides for nicer
 559                         abstraction and gives administrators the
 560                         option to enable
 561                         <filename>bluetoothd.service</filename> via
 562                         controlling a
 563                         <filename>bluetooth.target.wants/</filename>
 564                         symlink uniformly with a command like
 565                         <command>enable</command> of
 566                         <citerefentry><refentrytitle>systemctl</refentrytitle><manvolnum>1</manvolnum></citerefentry>
 567                         instead of manipulating the udev
 568                         ruleset.</para>
 569                 </refsect2>
 570
 571                 <refsect2>
 572                         <title>Path-Based Activation</title>
 573
 574                         <para>Often, runtime of daemons processing
 575                         spool files or directories (such as a printing
 576                         system) can be delayed until these file system
 577                         objects change state, or become
 578                         non-empty. New-style init systems provide a
 579                         way to bind service activation to file system
 580                         changes. systemd implements this scheme via
 581                         path-based activation configured in
 582                         <filename>.path</filename> units, as outlined
 583                         in
 584                         <citerefentry><refentrytitle>systemd.path</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para>
 585                 </refsect2>
 586
 587                 <refsect2>
 588                         <title>Timer-Based Activation</title>
 589
 590                         <para>Some daemons that implement clean-up
 591                         jobs that are intended to be executed in
 592                         regular intervals benefit from timer-based
 593                         activation. In systemd, this is implemented
 594                         via <filename>.timer</filename> units, as
 595                         described in
 596                         <citerefentry><refentrytitle>systemd.timer</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para>
 597                 </refsect2>
 598
 599                 <refsect2>
 600                         <title>Other Forms of Activation</title>
 601
 602                         <para>Other forms of activation have been
 603                         suggested and implemented in some
 604                         systems. However, there are often simpler or
 605                         better alternatives, or they can be put
 606                         together of combinations of the schemes
 607                         above. Example: Sometimes, it appears useful to
 608                         start daemons or <filename>.socket</filename>
 609                         units when a specific IP address is configured
 610                         on a network interface, because network
 611                         sockets shall be bound to the
 612                         address. However, an alternative to implement
 613                         this is by utilizing the Linux IP_FREEBIND
 614                         socket option, as accessible via
 615                         <varname>FreeBind=yes</varname> in systemd
 616                         socket files (see
 617                         <citerefentry><refentrytitle>systemd.socket</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 618                         for details). This option, when enabled,
 619                         allows sockets to be bound to a non-local, not
 620                         configured IP address, and hence allows
 621                         bindings to a particular IP address before it
 622                         actually becomes available, making such an
 623                         explicit dependency to the configured address
 624                         redundant. Another often suggested trigger for
 625                         service activation is low system
 626                         load. However, here too, a more convincing
 627                         approach might be to make proper use of
 628                         features of the operating system, in
 629                         particular, the CPU or IO scheduler of
 630                         Linux. Instead of scheduling jobs from
 631                         userspace based on monitoring the OS
 632                         scheduler, it is advisable to leave the
 633                         scheduling of processes to the OS scheduler
 634                         itself. systemd provides fine-grained access
 635                         to the CPU and IO schedulers. If a process
 636                         executed by the init system shall not
 637                         negatively impact the amount of CPU or IO
 638                         bandwidth available to other processes, it
 639                         should be configured with
 640                         <varname>CPUSchedulingPolicy=idle</varname>
 641                         and/or
 642                         <varname>IOSchedulingClass=idle</varname>. Optionally,
 643                         this may be combined with timer-based
 644                         activation to schedule background jobs during
 645                         runtime and with minimal impact on the system,
 646                         and remove it from the boot phase
 647                         itself.</para>
 648                 </refsect2>
 649
 650         </refsect1>
 651         <refsect1>
 652                 <title>Integration with Systemd</title>
 653
 654                 <refsect2>
 655                         <title>Writing Systemd Unit Files</title>
 656
 657                         <para>When writing systemd unit files, it is
 658                         recommended to consider the following
 659                         suggestions:</para>
 660
 661                         <orderedlist>
 662                                 <listitem><para>If possible, do not use
 663                                 the <varname>Type=forking</varname>
 664                                 setting in service files. But if you
 665                                 do, make sure to set the PID file path
 666                                 using <varname>PIDFile=</varname>. See
 667                                 <citerefentry><refentrytitle>systemd.service</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 668                                 for details.</para></listitem>
 669
 670                                 <listitem><para>If your daemon
 671                                 registers a D-Bus name on the bus,
 672                                 make sure to use
 673                                 <varname>Type=dbus</varname> in the
 674                                 service file if
 675                                 possible.</para></listitem>
 676
 677                                 <listitem><para>Make sure to set a
 678                                 good human-readable description string
 679                                 with
 680                                 <varname>Description=</varname>.</para></listitem>
 681
 682                                 <listitem><para>Do not disable
 683                                 <varname>DefaultDependencies=</varname>,
 684                                 unless you really know what you do and
 685                                 your unit is involved in early boot or
 686                                 late system shutdown.</para></listitem>
 687
 688                                 <listitem><para>Normally, little if
 689                                 any dependencies should need to
 690                                 be defined explicitly. However, if you
 691                                 do configure explicit dependencies, only refer to
 692                                 unit names listed on
 693                                 <citerefentry><refentrytitle>systemd.special</refentrytitle><manvolnum>7</manvolnum></citerefentry>
 694                                 or names introduced by your own
 695                                 package to keep the unit file
 696                                 operating
 697                                 system-independent.</para></listitem>
 698
 699                                 <listitem><para>Make sure to include
 700                                 an <literal>[Install]</literal>
 701                                 section including installation
 702                                 information for the unit file. See
 703                                 <citerefentry><refentrytitle>systemd.unit</refentrytitle><manvolnum>5</manvolnum></citerefentry>
 704                                 for details. To activate your service
 705                                 on boot, make sure to add a
 706                                 <varname>WantedBy=multi-user.target</varname>
 707                                 or
 708                                 <varname>WantedBy=graphical.target</varname>
 709                                 directive. To activate your socket on
 710                                 boot, make sure to add
 711                                 <varname>WantedBy=sockets.target</varname>. Usually,
 712                                 you also want to make sure that when
 713                                 your service is installed, your socket
 714                                 is installed too, hence add
 715                                 <varname>Also=foo.socket</varname> in
 716                                 your service file
 717                                 <filename>foo.service</filename>, for
 718                                 a hypothetical program
 719                                 <filename>foo</filename>.</para></listitem>
 720
 721                         </orderedlist>
 722                 </refsect2>
 723
 724                 <refsect2>
 725                         <title>Installing Systemd Service Files</title>
 726
 727                         <para>At the build installation time
 728                         (e.g. <command>make install</command> during
 729                         package build), packages are recommended to
 730                         install their systemd unit files in the
 731                         directory returned by <command>pkg-config
 732                         systemd
 733                         --variable=systemdsystemunitdir</command> (for
 734                         system services) or <command>pkg-config
 735                         systemd
 736                         --variable=systemduserunitdir</command>
 737                         (for user services). This will make the
 738                         services available in the system on explicit
 739                         request but not activate them automatically
 740                         during boot. Optionally, during package
 741                         installation (e.g. <command>rpm -i</command>
 742                         by the administrator), symlinks should be
 743                         created in the systemd configuration
 744                         directories via the <command>enable</command>
 745                         command of the
 746                         <citerefentry><refentrytitle>systemctl</refentrytitle><manvolnum>1</manvolnum></citerefentry>
 747                         tool to activate them automatically on
 748                         boot.</para>
 749
 750                         <para>Packages using
 751                         <citerefentry project='die-net'><refentrytitle>autoconf</refentrytitle><manvolnum>1</manvolnum></citerefentry>
 752                         are recommended to use a configure script
 753                         excerpt like the following to determine the
 754                         unit installation path during source
 755                         configuration:</para>
 756
 757                         <programlisting>PKG_PROG_PKG_CONFIG
 758 AC_ARG_WITH([systemdsystemunitdir],
 759      [AS_HELP_STRING([--with-systemdsystemunitdir=DIR], [Directory for systemd service files])],,
 760      [with_systemdsystemunitdir=auto])
 761 AS_IF([test "x$with_systemdsystemunitdir" = "xyes" -o "x$with_systemdsystemunitdir" = "xauto"], [
 762      def_systemdsystemunitdir=$($PKG_CONFIG --variable=systemdsystemunitdir systemd)
 763
 764      AS_IF([test "x$def_systemdsystemunitdir" = "x"],
 765          [AS_IF([test "x$with_systemdsystemunitdir" = "xyes"],
 766                 [AC_MSG_ERROR([systemd support requested but pkg-config unable to query systemd package])])
 767           with_systemdsystemunitdir=no],
 768          [with_systemdsystemunitdir="$def_systemdsystemunitdir"])])
 769 AS_IF([test "x$with_systemdsystemunitdir" != "xno"],
 770       [AC_SUBST([systemdsystemunitdir], [$with_systemdsystemunitdir])])
 771 AM_CONDITIONAL([HAVE_SYSTEMD], [test "x$with_systemdsystemunitdir" != "xno"])</programlisting>
 772
 773                         <para>This snippet allows automatic
 774                         installation of the unit files on systemd
 775                         machines, and optionally allows their
 776                         installation even on machines lacking
 777                         systemd. (Modification of this snippet for the
 778                         user unit directory is left as an exercise for the
 779                         reader.)</para>
 780
 781                         <para>Additionally, to ensure that
 782                         <command>make distcheck</command> continues to
 783                         work, it is recommended to add the following
 784                         to the top-level <filename>Makefile.am</filename>
 785                         file in
 786                         <citerefentry project='die-net'><refentrytitle>automake</refentrytitle><manvolnum>1</manvolnum></citerefentry>-based
 787                         projects:</para>
 788
 789                         <programlisting>DISTCHECK_CONFIGURE_FLAGS = \
 790         --with-systemdsystemunitdir=$$dc_install_base/$(systemdsystemunitdir)</programlisting>
 791
 792                         <para>Finally, unit files should be installed in the system with an automake excerpt like the following:</para>
 793
 794                         <programlisting>if HAVE_SYSTEMD
 795 systemdsystemunit_DATA = \
 796         foobar.socket \
 797         foobar.service
 798 endif</programlisting>
 799
 800                         <para>In the
 801                         <citerefentry project='die-net'><refentrytitle>rpm</refentrytitle><manvolnum>8</manvolnum></citerefentry>
 802                         <filename>.spec</filename> file, use snippets
 803                         like the following to enable/disable the
 804                         service during
 805                         installation/deinstallation. This makes use of
 806                         the RPM macros shipped along systemd. Consult
 807                         the packaging guidelines of your distribution
 808                         for details and the equivalent for other
 809                         package managers.</para>
 810
 811                         <para>At the top of the file:</para>
 812
 813                         <programlisting>BuildRequires: systemd
 814 %{?systemd_requires}</programlisting>
 815
 816                         <para>And as scriptlets, further down:</para>
 817
 818                         <programlisting>%post
 819 %systemd_post foobar.service foobar.socket
 820
 821 %preun
 822 %systemd_preun foobar.service foobar.socket
 823
 824 %postun
 825 %systemd_postun</programlisting>
 826
 827                         <para>If the service shall be restarted during
 828                         upgrades, replace the
 829                         <literal>%postun</literal> scriptlet above
 830                         with the following:</para>
 831
 832                         <programlisting>%postun
 833 %systemd_postun_with_restart foobar.service</programlisting>
 834
 835                         <para>Note that
 836                         <literal>%systemd_post</literal> and
 837                         <literal>%systemd_preun</literal> expect the
 838                         names of all units that are installed/removed
 839                         as arguments, separated by
 840                         spaces. <literal>%systemd_postun</literal>
 841                         expects no
 842                         arguments. <literal>%systemd_postun_with_restart</literal>
 843                         expects the units to restart as
 844                         arguments.</para>
 845
 846                         <para>To facilitate upgrades from a package
 847                         version that shipped only SysV init scripts to
 848                         a package version that ships both a SysV init
 849                         script and a native systemd service file, use
 850                         a fragment like the following:</para>
 851
 852                         <programlisting>%triggerun -- foobar &lt; 0.47.11-1
 853 if /sbin/chkconfig --level 5 foobar ; then
 854         /bin/systemctl --no-reload enable foobar.service foobar.socket >/dev/null 2>&amp;1 || :
 855 fi</programlisting>
 856
 857                         <para>Where 0.47.11-1 is the first package
 858                         version that includes the native unit
 859                         file. This fragment will ensure that the first
 860                         time the unit file is installed, it will be
 861                         enabled if and only if the SysV init script is
 862                         enabled, thus making sure that the enable
 863                         status is not changed. Note that
 864                         <command>chkconfig</command> is a command
 865                         specific to Fedora which can be used to check
 866                         whether a SysV init script is enabled. Other
 867                         operating systems will have to use different
 868                         commands here.</para>
 869                 </refsect2>
 870         </refsect1>
 871
 872         <refsect1>
 873                 <title>Porting Existing Daemons</title>
 874
 875                 <para>Since new-style init systems such as systemd are
 876                 compatible with traditional SysV init systems, it is
 877                 not strictly necessary to port existing daemons to the
 878                 new style. However, doing so offers additional
 879                 functionality to the daemons as well as simplifying
 880                 integration into new-style init systems.</para>
 881
 882                 <para>To port an existing SysV compatible daemon, the
 883                 following steps are recommended:</para>
 884
 885                 <orderedlist>
 886                         <listitem><para>If not already implemented,
 887                         add an optional command line switch to the
 888                         daemon to disable daemonization. This is
 889                         useful not only for using the daemon in
 890                         new-style init systems, but also to ease
 891                         debugging.</para></listitem>
 892
 893                         <listitem><para>If the daemon offers
 894                         interfaces to other software running on the
 895                         local system via local <constant>AF_UNIX</constant> sockets,
 896                         consider implementing socket-based activation
 897                         (see above). Usually, a minimal patch is
 898                         sufficient to implement this: Extend the
 899                         socket creation in the daemon code so that
 900                         <citerefentry><refentrytitle>sd_listen_fds</refentrytitle><manvolnum>3</manvolnum></citerefentry>
 901                         is checked for already passed sockets
 902                         first. If sockets are passed (i.e. when
 903                         <function>sd_listen_fds()</function> returns a
 904                         positive value), skip the socket creation step
 905                         and use the passed sockets. Secondly, ensure
 906                         that the file system socket nodes for local
 907                         <constant>AF_UNIX</constant> sockets used in the socket-based
 908                         activation are not removed when the daemon
 909                         shuts down, if sockets have been
 910                         passed. Third, if the daemon normally closes
 911                         all remaining open file descriptors as part of
 912                         its initialization, the sockets passed from
 913                         the init system must be spared. Since
 914                         new-style init systems guarantee that no
 915                         left-over file descriptors are passed to
 916                         executed processes, it might be a good choice
 917                         to simply skip the closing of all remaining
 918                         open file descriptors if sockets are
 919                         passed.</para></listitem>
 920
 921                         <listitem><para>Write and install a systemd
 922                         unit file for the service (and the sockets if
 923                         socket-based activation is used, as well as a
 924                         path unit file, if the daemon processes a
 925                         spool directory), see above for
 926                         details.</para></listitem>
 927
 928                         <listitem><para>If the daemon exposes
 929                         interfaces via D-Bus, write and install a
 930                         D-Bus activation file for the service, see
 931                         above for details.</para></listitem>
 932                 </orderedlist>
 933         </refsect1>
 934
 935         <refsect1>
 936                 <title>Placing Daemon Data</title>
 937
 938                 <para>It is recommended to follow the general
 939                 guidelines for placing package files, as discussed in
 940                 <citerefentry><refentrytitle>file-hierarchy</refentrytitle><manvolnum>7</manvolnum></citerefentry>.</para>
 941         </refsect1>
 942
 943         <refsect1>
 944                 <title>See Also</title>
 945                 <para>
 946                         <citerefentry><refentrytitle>systemd</refentrytitle><manvolnum>1</manvolnum></citerefentry>,
 947                         <citerefentry><refentrytitle>sd-daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
 948                         <citerefentry><refentrytitle>sd_listen_fds</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
 949                         <citerefentry><refentrytitle>sd_notify</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
 950                         <citerefentry><refentrytitle>daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
 951                         <citerefentry><refentrytitle>systemd.service</refentrytitle><manvolnum>5</manvolnum></citerefentry>,
 952                         <citerefentry><refentrytitle>file-hierarchy</refentrytitle><manvolnum>7</manvolnum></citerefentry>
 953                 </para>
 954         </refsect1>
 955
 956 </refentry>