1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
|
<html>
<head>
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />
<meta http-equiv="Content-Language" content="en" />
<title>s6: why another supervision suite</title>
<meta name="Description" content="s6: why another supervision suite" />
<meta name="Keywords" content="s6 supervision daemontools runit perp service svscan supervise" />
<!-- <link rel="stylesheet" type="text/css" href="//skarnet.org/default.css" /> -->
</head>
<body>
<p>
<a href="index.html">s6</a><br />
<a href="//skarnet.org/software/">Software</a><br />
<a href="//skarnet.org/">skarnet.org</a>
</p>
<h1> Why another supervision suite ? </h1>
<p>
Supervision suites are becoming quite common. Today, we already have:
</p>
<ul>
<li> Good (?) old System V init, which can be made to supervise services if you perform <tt>/etc/inittab</tt> voodoo.
BSD init can also be used the same way with the <tt>/etc/ttys</tt> file, but for some reason, nobody among BSD
developers is using <tt>/etc/ttys</tt> to this purpose, so I won't consider BSD init here. </li>
<li> <a href="http://cr.yp.to/daemontools.html">daemontools</a>, the pioneer </li>
<li> <a href="http://untroubled.org/daemontools-encore/">daemontools-encore</a>, Bruce Guenter's upgrade to daemontools </li>
<li> <a href="http://smarden.org/runit/">runit</a>, Gerrit Pape's suite, well-integrated with Debian </li>
<li> <a href="http://b0llix.net/perp/">perp</a>, Wayne Marshall's take on supervision </li>
<li> Integrated init systems providing a lot of features, process supervision being one of them.
For instance, <a href="http://upstart.ubuntu.com/">Upstart</a>, MacOS X's
<a href="http://developer.apple.com/library/mac/#documentation/Darwin/Reference/ManPages/man8/launchd.8.html">launchd</a>,
and Fedora's <a href="http://freedesktop.org/wiki/Software/systemd">systemd</a>. </li>
</ul>
<p>
Why is s6 needed ? What does it do differently ? Here are the criteria I used.
</p>
<h2> Supervision suites should not wake up unless notified. </h2>
<ul>
<li> System V init fails the test: it wakes up every 5 seconds, for the reason that
<tt>/dev/initctl</tt> might have changed.
<a href="http://demotivate.me/mediafiles/full/4162010103911AM_picard-no-facepalm.jpg"><tt>m(</tt></a> </li>
<li> daemontools fails the test: it wakes up every 5 seconds to check for new services. </li>
<li> daemontools-encore does the same. </li>
<li> the current version of runit fails the test: it wakes up every 14 seconds. But this is a workaround for a bug in some Linux kernels;
there is no design flaw in runit that prevents it from passing the test. </li>
<li> perp works. </li>
<li> Upstart works. I have no idea what other integrated init systems do: it's much too difficult to strace them
to see exactly where they're spending their time, and when it is possible, the trace output is so big that it's
hard to extract any valuable information from it. </li>
<li> s6 works. The <tt>-t</tt> option to s6-svscan makes it check its services
with a configurable timeout; by default, this timeout is infinite, i.e. it
never wakes up unless it receives a command via s6-svscanctl. </li>
</ul>
<h2> Supervision suites should provide a program that can run as process 1. </h2>
<ul>
<li> System V init <em>is</em> process 1, so no problem here. </li>
<li> Integrated init systems, by definition, provide a process 1. </li>
<li> daemontools was not designed to take over init, although
<a href="http://code.dogmap.org./svscan-1/">it can be made to work</a> with
enough hacking skills. Same thing with daemontools-encore. </li>
<li> runit provides an <em>init</em> functionality, but the mechanism is
separate from the supervision itself; the <tt>runit</tt> process, not the
<tt>runsvdir</tt> process, runs as process 1. This lengthens the supervision
chain. </li>
<li> perp was not designed to run as process 1. It probably could be made to work too
without too much trouble. </li>
<li> s6-svscan was designed from the start to be run as process 1, although it
does not have to. </li>
</ul>
<h2> Supervision suites should be bug-free, lightweight and easy to understand. </h2>
<ul>
<li> daemontools, daemontools-encore, runit and perp all qualify. All of this is excellent quality
code, <a href="//skarnet.org/software/skalibs/djblegacy.html">unsurprisingly</a>. </li>
<li> System V init is understandable, and reasonably lightweight; but it is still
too big for what it does - poorly. The <tt>/etc/inittab</tt> file needs to be parsed;
that parser has to be in process 1. There is support in process 1 for the whole
"runlevel" concept, which is a primitive form of service management. The same
executable handles all 3 stages of the machine's lifetime and does not separate
them properly. All in all, System V init does its job, but is showing its age
and nowadays we know much better designs. </li>
<li> This is where integrated init systems fail, hard. By wanting to organize
the way a the machine is operated - so, machine state management - in the
<em>same package</em> as the init and process supervision system, they add
incredible complexity where it does not belong.
<ul>
<li> Upstart uses <tt>ptrace</tt> to watch its children fork(), and links
process 1 against libdbus. This is insane.
Process 1 should be <em>absolutely stable</em>, it should be guaranteed
to never crash, so the whole of its source code should be under control. At
Upstart's level of complexity, those goals are outright impossible to achieve,
so this approach is flawed by design. </li>
<li> launchd suffers from the same kind of problems. Example: Services
running under launchd must be configured
<a href="https://developer.apple.com/library/mac/documentation/MacOSX/Conceptual/BPSystemStartup/Chapters/CreatingLaunchdJobs.html">using
XML</a>; the launchctl process interprets the XML, converts it into a
key-value store (which is strictly
less powerful than XML, so why do they even use XML in the first place?)
and sends it to launchd via a Mach-specific IPC. Process 1 needs to be
linked against the library that handles the Mach IPC, it needs to decode
the key-value store, and use it to run and supervise a daemon. And it
needs to keep everything in memory. This is
a lot more complex and resource-consuming than it needs to be. </li>
<li> systemd is much, much worse than the other ones, and a real danger
for the future of GNU/Linux. I have a <a href="systemd.html">special page</a>
dedicated to it. </li>
</ul>
What those systems fail to recognize is that process supervision, rooted in
process 1, is a good thing, and machine management is also a good thing, but
<strong>those are two different functions</strong>, and a good init system
needs, and <strong>should</strong>, only provide process supervision, in
order to keep such a crucial piece of code as easy to maintain as possible.
Machine management can be added <em>on top of</em> a process supervision
suite, in a different package, and it has nothing to do with process 1. </li>
<li> s6, which has been designed with embedded environments in mind, tries
harder than anyone to pass this. It tries so hard that <tt>s6-svscan</tt>
and <tt>s6-supervise</tt>, the two long-running programs that make the
supervision chain, <em>do not even allocate heap memory</em>, and their main
program source files are less than 500 lines long. </li>
</ul>
<h2> Supervision suites should provide a basis for high-level service management. </h2>
<ul>
<li> Neither System V init, daemontools, runit or perp
provides any hooks to wait for a service to go up or down. runit provides a
waiting mechanism, but it's based on polling, and the <tt>./check</tt> script
has to be manually written for every service. </li>
<li> daemontools-encore qualifies: the <em>notify script</em> can be used for
inter-service communication. But it's just a hook: all the real notification
work has to be done by the notify script itself, no notification framework is
provided. </li>
<li> Integrated init systems provide high-level service management
themselves. Again, this is not good design: service management has nothing
to do with init or process supervision, and should be implemented on top
of it, not as a part of it. </li>
<li> s6 comes with <a href="ftrig.html">an event notification
library</a>, and command-line tools based on this library, thus providing a simple
API for future service management tools to build upon. </li>
</ul>
<h2> Artistic considerations </h2>
<ul>
<li> <tt>s6-svscan</tt> and <tt>s6-supervise</tt> are <em>entirely asynchronous</em>.
Even during trouble (full process table, for instance), they'll remain reactive
and instantly respond to commands they may receive. <tt>s6-supervise</tt> has
even been implemented as a full deterministic finite automaton, to ensure it
always does the right thing under any circumstance. Other supervision suites
do not achieve that for now. </li>
<li> daemontools' <a href="http://cr.yp.to/daemontools/svscan.html">svscan</a>
maintains an open pipe between a daemon and its logger, so even if the daemon,
the logger, <em>and</em> both
<a href="http://cr.yp.to/daemontools/supervise.html">supervise</a> processes
die, the pipe is still the same so <em>no logs are lost, ever</em>, unless
svscan itself dies. </li>
<li> runit has only one supervisor, <a href="http://smarden.org/runit/runsv.8.html">runsv</a>,
for both a daemon and its logger. The pipe is maintained by <tt>runsv</tt>.
If the <tt>runsv</tt> process dies, the pipe disappears and logs are lost.
So, runit does not offer as strong a guarantee as daemontools. </li>
<li> perp has only one process, <a href="http://b0llix.net/perp/site.cgi?page=perpd.8">perpd</a>,
acting both as a "daemon and logger supervisor" (like <tt>runsv</tt>) and as a
"service directory scanner" (like <tt>runsvdir</tt>). It maintains the pipes
between the daemons and their respective loggers. If perpd dies, everything
is lost. Since perpd cannot be run as process 1, this is a possible SPOF for
a perp installation; however, perpd is well-written and has virtually no risk of
dying, especially compared to process 1 behemoths provided by integrated
init systems. </li>
<li> Besides, the <tt>runsv</tt> model, which has to handle both a daemon
and its logger, is more complex than the <tt>supervise</tt> model (which
only has to handle a daemon). Consequently, the <tt>runsvdir</tt> model is
simpler than the <tt>svscan</tt> model, but there is only one <tt>svscan</tt>
instance when there are several <tt>runsv</tt>s and <tt>supervise</tt>s.
The <tt>perpd</tt> model is obviously the most complex; while very understandable,
<tt>perpd</tt> is unarguably harder to maintain than the other two. </li>
<li> So, to achieve maximum simplicity and code reuse, and minimal memory
footprint, s6's design is close to daemontools' one.
And when <a href="s6-svscan-1.html">s6-svscan is run as process 1</a>,
pipes between daemons and loggers are never lost. </li>
</ul>
<h2> Conclusion </h2>
<p>
All in all, I believe that s6 offers the best overall implementation of a
supervision suite <em>as it should be designed</em>. At worst, it's just another
take on daemontools with a <a href="//skarnet.org/software/skalibs/">reliable
base library</a> and a few nifty features.
</p>
</body>
</html>
|
