]> git.ipfire.org Git - thirdparty/systemd.git/blame - man/daemon.xml
TODO: Remove Elapsed Time DHCPv6 option as it is done
[thirdparty/systemd.git] / man / daemon.xml
CommitLineData
64aba792
LP
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
5430f7f2
LP
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
64aba792
LP
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
5430f7f2 18 Lesser General Public License for more details.
64aba792 19
5430f7f2 20 You should have received a copy of the GNU Lesser General Public License
64aba792
LP
21 along with systemd; If not, see <http://www.gnu.org/licenses/>.
22-->
23
62adf224 24<refentry id="daemon">
64aba792
LP
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>
34511ca7 47 <refpurpose>Writing and packaging system daemons</refpurpose>
64aba792
LP
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
62adf224
LP
58 yet more powerful scheme (here called "new-style"
59 daemons), as implemented by
99ffae46
LP
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>
64aba792
LP
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
62adf224 71 steps are unnecessary for new-style daemons (see below),
436c44a5 72 and should only be implemented if compatibility
64aba792
LP
73 with SysV is essential.</para>
74
75 <orderedlist>
76 <listitem><para>Close all open file
6667311d
JE
77 descriptors except standard input, output,
78 and error (i.e. the first three file
64aba792
LP
79 descriptors 0, 1, 2). This ensures
80 that no accidentally passed file
81 descriptor stays around in the daemon
409dee2e 82 process. On Linux, this is best
64aba792
LP
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
62adf224 87 <function>getrlimit()</function> for
64aba792
LP
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
05cc7267 95 <constant>SIG_DFL</constant>.</para></listitem>
64aba792
LP
96
97 <listitem><para>Reset the signal mask
62adf224
LP
98 using
99 <function>sigprocmask()</function>.</para></listitem>
64aba792 100
62adf224
LP
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>,
64aba792
LP
108 to create a background
109 process.</para></listitem>
110
111 <listitem><para>In the child, call
62adf224
LP
112 <function>setsid()</function> to
113 detach from any terminal and create an
114 independent session.</para></listitem>
64aba792
LP
115
116 <listitem><para>In the child, call
62adf224 117 <function>fork()</function> again, to
409dee2e 118 ensure that the daemon can never re-acquire
62adf224 119 a terminal again.</para></listitem>
64aba792 120
62adf224 121 <listitem><para>Call <function>exit()</function> in the
64aba792
LP
122 first child, so that only the second
123 child (the actual daemon process)
124 stays around. This ensures that the
2ab38e24 125 daemon process is re-parented to
64aba792
LP
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>
6db27428
ZJS
131 to standard input, output, and error.
132 </para></listitem>
64aba792
LP
133
134 <listitem><para>In the daemon process,
135 reset the umask to 0, so that the file
62adf224 136 modes passed to <function>open()</function>, <function>mkdir()</function> and
64aba792
LP
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
62adf224
LP
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
41a55c46 152 <filename>/run/foobar.pid</filename>
409dee2e 153 (for a hypothetical daemon "foobar")
62adf224
LP
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
409dee2e 158 it is verified at the same time that
62adf224
LP
159 the PID previously stored in the PID
160 file no longer exists or belongs to a
c4b834a4 161 foreign process.</para></listitem>
62adf224 162
64aba792
LP
163 <listitem><para>In the daemon process,
164 drop privileges, if possible and
165 applicable.</para></listitem>
166
167 <listitem><para>From the daemon
409dee2e 168 process, notify the original process
64aba792
LP
169 started that initialization is
170 complete. This can be implemented via
171 an unnamed pipe or similar
172 communication channel that is created
62adf224
LP
173 before the first
174 <function>fork()</function> and hence
175 available in both the original and the
176 daemon process.</para></listitem>
64aba792 177
62adf224
LP
178 <listitem><para>Call
179 <function>exit()</function> in the
64aba792
LP
180 original process. The process that
181 invoked the daemon must be able to
c5315881 182 rely on that this
62adf224
LP
183 <function>exit()</function> happens
184 after initialization is complete and
185 all external communication channels
c5315881 186 are established and
64aba792
LP
187 accessible.</para></listitem>
188 </orderedlist>
189
62adf224
LP
190 <para>The BSD <function>daemon()</function> function should not be
191 used, as it implements only a subset of these steps.</para>
64aba792
LP
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
c5315881 197 behavior optional and configurable via a
409dee2e 198 command line argument to ease debugging as
64aba792
LP
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
409dee2e 212 <para>For developing a new-style daemon, none
64aba792
LP
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
409dee2e 219 the init system, it is recommended not to
64aba792
LP
220 execute them when run as new-style
221 service.</para>
222
62adf224 223 <para>Note that new-style init systems
6db27428
ZJS
224 guarantee execution of daemon processes in a
225 clean process context: it is guaranteed that
62adf224
LP
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
6db27428
ZJS
229 will be executed in their own session, with
230 standard input/output/error connected to
62adf224 231 <filename>/dev/null</filename> unless
6db27428
ZJS
232 otherwise configured. The umask is reset.
233 </para>
62adf224 234
64aba792
LP
235 <para>It is recommended for new-style daemons
236 to implement the following:</para>
237
238 <orderedlist>
05cc7267 239 <listitem><para>If <constant>SIGTERM</constant> is
64aba792
LP
240 received, shut down the daemon and
241 exit cleanly.</para></listitem>
242
05cc7267 243 <listitem><para>If <constant>SIGHUP</constant> is received,
64aba792
LP
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
62adf224 252 scheme as defined in the <ulink
27d14fb3 253 url="http://refspecs.linuxbase.org/LSB_3.1.1/LSB-Core-generic/LSB-Core-generic/iniscrptact.html">LSB
62adf224
LP
254 recommendations for SysV init
255 scripts</ulink>.</para></listitem>
64aba792 256
64aba792 257 <listitem><para>If possible and
409dee2e 258 applicable, expose the daemon's control
64aba792
LP
259 interface via the D-Bus IPC system and
260 grab a bus name as last step of
261 initialization.</para></listitem>
262
99ffae46
LP
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,
c5315881 273 rely on the init system's
99ffae46
LP
274 functionality to limit the access of
275 the daemon to files, services and
409dee2e 276 other resources, i.e. in the case of
99ffae46
LP
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
64aba792 286 <listitem><para>If D-Bus is used, make
409dee2e 287 your daemon bus-activatable by
64aba792
LP
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
409dee2e 295 restarted on failure without losing
64aba792 296 any bus requests, as the bus queues
62adf224
LP
297 requests for activatable services. See
298 below for details.</para></listitem>
64aba792
LP
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
409dee2e 306 activation, this enables on-demand
64aba792 307 starting of services as well as it
436c44a5 308 allows improved parallelization of
64aba792 309 service start-up. Also, for state-less
409dee2e 310 protocols (such as syslog, DNS), a
64aba792
LP
311 daemon implementing socket-based
312 activation can be restarted without
62adf224
LP
313 losing a single request. See below for
314 details.</para></listitem>
64aba792 315
409dee2e 316 <listitem><para>If applicable, a daemon
64aba792 317 should notify the init system about
62adf224
LP
318 startup completion or status updates
319 via the
320 <citerefentry><refentrytitle>sd_notify</refentrytitle><manvolnum>3</manvolnum></citerefentry>
64aba792
LP
321 interface.</para></listitem>
322
323 <listitem><para>Instead of using the
8c63bf4a
LP
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
cb07866b 339 <citerefentry><refentrytitle>sd-daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>
62adf224
LP
340 and
341 <citerefentry><refentrytitle>systemd.exec</refentrytitle><manvolnum>5</manvolnum></citerefentry>.</para></listitem>
64aba792
LP
342
343 </orderedlist>
62adf224
LP
344
345 <para>These recommendations are similar but
346 not identical to the <ulink
0afedd30 347 url="https://developer.apple.com/library/mac/documentation/MacOSX/Conceptual/BPSystemStartup/Chapters/CreatingLaunchdJobs.html">Apple
62adf224 348 MacOS X Daemon Requirements</ulink>.</para>
64aba792
LP
349 </refsect2>
350
99ffae46
LP
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
409dee2e 368 unconditionally, it is a good idea to implement some of
99ffae46 369 the various activation schemes outlined below, in
409dee2e 370 order to maximize parallelization. If a daemon
99ffae46
LP
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
409dee2e 378 case of sockets) until the activation is
99ffae46
LP
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
27d14fb3 388 url="http://refspecs.linuxbase.org/LSB_3.1.1/LSB-Core-generic/LSB-Core-generic/iniscrptact.html">LSB
99ffae46
LP
389 Linux Standard Base Core
390 Specification</ulink>. This method of
8f28cbcd 391 activation is supported ubiquitously on Linux
99ffae46 392 init systems, both old-style and new-style
409dee2e 393 systems. Among other issues, SysV init scripts
99ffae46
LP
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
e0e009c0 402 administrator wants to make sure that a service or
409dee2e 403 other unit is activated automatically on boot,
99ffae46
LP
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
64aba792 419 <refsect2>
62adf224 420 <title>Socket-Based Activation</title>
99ffae46
LP
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
409dee2e 428 activation scheme, the creation and binding of
99ffae46
LP
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
409dee2e 433 per-daemon configuration, the init system
99ffae46
LP
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.
409dee2e 437 Optionally, activation of the service can be
99ffae46 438 delayed until the first inbound traffic
409dee2e 439 arrives at the socket to implement on-demand
99ffae46
LP
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
20604ebc 443 started in parallel as soon as all sockets
409dee2e 444 are established. In addition to that, daemons
99ffae46 445 can be restarted with losing only a minimal
409dee2e 446 number of client transactions, or even any
99ffae46
LP
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
409dee2e 456 sockets from the init system instead of
99ffae46
LP
457 creating and binding them themselves. For
458 details about the programming interfaces for
409dee2e 459 this scheme provided by systemd, see
99ffae46
LP
460 <citerefentry><refentrytitle>sd_listen_fds</refentrytitle><manvolnum>3</manvolnum></citerefentry>
461 and
cb07866b 462 <citerefentry><refentrytitle>sd-daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>. For
99ffae46 463 details about porting existing daemons to
409dee2e
JE
464 socket-based activation, see below. With
465 minimal effort, it is possible to implement
99ffae46
LP
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
409dee2e 477 activation, it is essential that all listening
99ffae46
LP
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>
409dee2e 483 section to automatically add such a
99ffae46
LP
484 dependency on installation of a socket
485 unit. Unless
486 <varname>DefaultDependencies=no</varname> is
409dee2e 487 set, the necessary ordering dependencies are
99ffae46
LP
488 implicitly created for all socket units. For
489 more information about
409dee2e 490 <filename>sockets.target</filename>, see
99ffae46
LP
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>
64aba792
LP
498 </refsect2>
499
500 <refsect2>
62adf224 501 <title>Bus-Based Activation</title>
99ffae46
LP
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
409dee2e 512 daemon, use the
99ffae46 513 <varname>SystemdService=</varname> directive
409dee2e 514 in these service files to configure the
99ffae46 515 matching systemd service for a D-Bus
409dee2e 516 service. e.g.: For a D-Bus service whose D-Bus
99ffae46
LP
517 activation file is named
518 <filename>org.freedesktop.RealtimeKit.service</filename>,
519 make sure to set
520 <varname>SystemdService=rtkit-daemon.service</varname>
409dee2e 521 in that file to bind it to the systemd
99ffae46
LP
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
409dee2e 536 new-style init system, it is possible to bind
ee5762e3
LP
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
909f413d 541 <literal>systemd</literal>. Like any other
409dee2e
JE
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
99ffae46
LP
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>
409dee2e 550 for details. Often, it is nicer to pull in
99ffae46 551 services from devices only indirectly via
409dee2e 552 dedicated targets. Example: Instead of pulling
99ffae46
LP
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>
ee5762e3
LP
564 symlink uniformly with a command like
565 <command>enable</command> of
566 <citerefentry><refentrytitle>systemctl</refentrytitle><manvolnum>1</manvolnum></citerefentry>
99ffae46
LP
567 instead of manipulating the udev
568 ruleset.</para>
64aba792
LP
569 </refsect2>
570
571 <refsect2>
62adf224 572 <title>Path-Based Activation</title>
99ffae46
LP
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>
62adf224
LP
597 </refsect2>
598
99ffae46
LP
599 <refsect2>
600 <title>Other Forms of Activation</title>
601
602 <para>Other forms of activation have been
603 suggested and implemented in some
409dee2e 604 systems. However, there are often simpler or
99ffae46
LP
605 better alternatives, or they can be put
606 together of combinations of the schemes
409dee2e 607 above. Example: Sometimes, it appears useful to
99ffae46
LP
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
409dee2e 628 features of the operating system, in
99ffae46
LP
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
2dddca4a 638 bandwidth available to other processes, it
99ffae46
LP
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
62adf224
LP
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>
409dee2e 662 <listitem><para>If possible, do not use
62adf224
LP
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
828f33e8
LP
673 <varname>Type=dbus</varname> in the
674 service file if
62adf224
LP
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
828f33e8
LP
700 an <literal>[Install]</literal>
701 section including installation
702 information for the unit file. See
62adf224
LP
703 <citerefentry><refentrytitle>systemd.unit</refentrytitle><manvolnum>5</manvolnum></citerefentry>
704 for details. To activate your service
409dee2e 705 on boot, make sure to add a
62adf224
LP
706 <varname>WantedBy=multi-user.target</varname>
707 or
828f33e8
LP
708 <varname>WantedBy=graphical.target</varname>
709 directive. To activate your socket on
710 boot, make sure to add
409dee2e 711 <varname>WantedBy=sockets.target</varname>. Usually,
828f33e8 712 you also want to make sure that when
409dee2e 713 your service is installed, your socket
828f33e8
LP
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>
62adf224
LP
720
721 </orderedlist>
64aba792
LP
722 </refsect2>
723
724 <refsect2>
99ffae46 725 <title>Installing Systemd Service Files</title>
62adf224
LP
726
727 <para>At the build installation time
728 (e.g. <command>make install</command> during
409dee2e 729 package build), packages are recommended to
62adf224
LP
730 install their systemd unit files in the
731 directory returned by <command>pkg-config
732 systemd
ee5762e3 733 --variable=systemdsystemunitdir</command> (for
16dad32e 734 system services) or <command>pkg-config
ee5762e3 735 systemd
af2d49f7
LP
736 --variable=systemduserunitdir</command>
737 (for user services). This will make the
62adf224
LP
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>
409dee2e 742 by the administrator), symlinks should be
62adf224 743 created in the systemd configuration
ee5762e3
LP
744 directories via the <command>enable</command>
745 command of the
746 <citerefentry><refentrytitle>systemctl</refentrytitle><manvolnum>1</manvolnum></citerefentry>
409dee2e 747 tool to activate them automatically on
62adf224
LP
748 boot.</para>
749
750 <para>Packages using
5aded369 751 <citerefentry project='die-net'><refentrytitle>autoconf</refentrytitle><manvolnum>1</manvolnum></citerefentry>
62adf224
LP
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
758AC_ARG_WITH([systemdsystemunitdir],
fc9acf25 759 [AS_HELP_STRING([--with-systemdsystemunitdir=DIR], [Directory for systemd service files])],,
5486855f
MG
760 [with_systemdsystemunitdir=auto])
761AS_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],
fc9acf25 768 [with_systemdsystemunitdir="$def_systemdsystemunitdir"])])
5486855f
MG
769AS_IF([test "x$with_systemdsystemunitdir" != "xno"],
770 [AC_SUBST([systemdsystemunitdir], [$with_systemdsystemunitdir])])
fc9acf25 771AM_CONDITIONAL([HAVE_SYSTEMD], [test "x$with_systemdsystemunitdir" != "xno"])</programlisting>
62adf224
LP
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
9f7dad77 778 user unit directory is left as an exercise for the
62adf224
LP
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
5aded369 786 <citerefentry project='die-net'><refentrytitle>automake</refentrytitle><manvolnum>1</manvolnum></citerefentry>-based
62adf224
LP
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
795systemdsystemunit_DATA = \
796 foobar.socket \
797 foobar.service
798endif</programlisting>
799
800 <para>In the
5aded369 801 <citerefentry project='die-net'><refentrytitle>rpm</refentrytitle><manvolnum>8</manvolnum></citerefentry>
409dee2e 802 <filename>.spec</filename> file, use snippets
8a422bb2
LP
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
62adf224
LP
807 the packaging guidelines of your distribution
808 for details and the equivalent for other
8a422bb2
LP
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>
62adf224
LP
817
818 <programlisting>%post
8a422bb2 819%systemd_post foobar.service foobar.socket
62adf224
LP
820
821%preun
8a422bb2 822%systemd_preun foobar.service foobar.socket
ee5762e3
LP
823
824%postun
8a422bb2
LP
825%systemd_postun</programlisting>
826
827 <para>If the service shall be restarted during
409dee2e 828 upgrades, replace the
8a422bb2
LP
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>
828f33e8 845
6908d384
LP
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
67ff6134 852 <programlisting>%triggerun -- foobar &lt; 0.47.11-1
63415a2d 853if /sbin/chkconfig --level 5 foobar ; then
d12d0e64 854 /bin/systemctl --no-reload enable foobar.service foobar.socket >/dev/null 2>&amp;1 || :
6908d384
LP
855fi</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
409dee2e 860 time the unit file is installed, it will be
6908d384 861 enabled if and only if the SysV init script is
0a9962a1 862 enabled, thus making sure that the enable
6908d384
LP
863 status is not changed. Note that
864 <command>chkconfig</command> is a command
485ccf9a 865 specific to Fedora which can be used to check
6908d384
LP
866 whether a SysV init script is enabled. Other
867 operating systems will have to use different
868 commands here.</para>
62adf224 869 </refsect2>
64aba792
LP
870 </refsect1>
871
99ffae46
LP
872 <refsect1>
873 <title>Porting Existing Daemons</title>
874
875 <para>Since new-style init systems such as systemd are
409dee2e 876 compatible with traditional SysV init systems, it is
99ffae46 877 not strictly necessary to port existing daemons to the
409dee2e 878 new style. However, doing so offers additional
ad678a06 879 functionality to the daemons as well as simplifying
99ffae46
LP
880 integration into new-style init systems.</para>
881
409dee2e 882 <para>To port an existing SysV compatible daemon, the
99ffae46
LP
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
74d00578 895 local system via local <constant>AF_UNIX</constant> sockets,
99ffae46 896 consider implementing socket-based activation
409dee2e 897 (see above). Usually, a minimal patch is
99ffae46
LP
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
12f25b6e 906 that the file system socket nodes for local
74d00578 907 <constant>AF_UNIX</constant> sockets used in the socket-based
99ffae46
LP
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
ad678a06 918 open file descriptors if sockets are
99ffae46
LP
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>
64aba792 934
9546c6ed
LP
935 <refsect1>
936 <title>Placing Daemon Data</title>
937
908c32e0 938 <para>It is recommended to follow the general
9546c6ed
LP
939 guidelines for placing package files, as discussed in
940 <citerefentry><refentrytitle>file-hierarchy</refentrytitle><manvolnum>7</manvolnum></citerefentry>.</para>
941 </refsect1>
942
64aba792 943 <refsect1>
160cd5c9
LP
944 <title>See Also</title>
945 <para>
946 <citerefentry><refentrytitle>systemd</refentrytitle><manvolnum>1</manvolnum></citerefentry>,
cb07866b 947 <citerefentry><refentrytitle>sd-daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
62adf224
LP
948 <citerefentry><refentrytitle>sd_listen_fds</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
949 <citerefentry><refentrytitle>sd_notify</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
99ffae46 950 <citerefentry><refentrytitle>daemon</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
9546c6ed
LP
951 <citerefentry><refentrytitle>systemd.service</refentrytitle><manvolnum>5</manvolnum></citerefentry>,
952 <citerefentry><refentrytitle>file-hierarchy</refentrytitle><manvolnum>7</manvolnum></citerefentry>
160cd5c9 953 </para>
64aba792
LP
954 </refsect1>
955
956</refentry>