nixos/doc: Document the activation script

This may be helpful to new module developers, curious users, and people
who just need a reference without having to look at the implementation
This commit is contained in:
Janne Heß 2022-01-30 00:37:55 +01:00
parent 08cd8ab8b6
commit 8d925cc8db
No known key found for this signature in database
GPG Key ID: 69165158F05265DF
7 changed files with 574 additions and 0 deletions

View File

@ -0,0 +1,72 @@
# Activation script {#sec-activation-script}
The activation script is a bash script called to activate the new
configuration which resides in a NixOS system in `$out/activate`. Since its
contents depend on your system configuration, the contents may differ.
This chapter explains how the script works in general and some common NixOS
snippets. Please be aware that the script is executed on every boot and system
switch, so tasks that can be performed in other places should be performed
there (for example letting a directory of a service be created by systemd using
mechanisms like `StateDirectory`, `CacheDirectory`, ... or if that's not
possible using `preStart` of the service).
Activation scripts are defined as snippets using
[](#opt-system.activationScripts). They can either be a simple multiline string
or an attribute set that can depend on other snippets. The builder for the
activation script will take these dependencies into account and order the
snippets accordingly. As a simple example:
```nix
system.activationScripts.my-activation-script = {
deps = [ "etc" ];
# supportsDryActivation = true;
text = ''
echo "Hallo i bims"
'';
};
```
This example creates an activation script snippet that is run after the `etc`
snippet. The special variable `supportsDryActivation` can be set so the snippet
is also run when `nixos-rebuild dry-activate` is run. To differentiate between
real and dry activation, the `$NIXOS_ACTION` environment variable can be
read which is set to `dry-activate` when a dry activation is done.
An activation script can write to special files instructing
`switch-to-configuration` to restart/reload units. The script will take these
requests into account and will incorperate the unit configuration as described
above. This means that the activation script will "fake" a modified unit file
and `switch-to-configuration` will act accordingly. By doing so, configuration
like [systemd.services.\<name\>.restartIfChanged](#opt-systemd.services) is
respected. Since the activation script is run **after** services are already
stopped, [systemd.services.\<name\>.stopIfChanged](#opt-systemd.services)
cannot be taken into account anymore and the unit is always restarted instead
of being stopped and started afterwards.
The files that can be written to are `/run/nixos/activation-restart-list` and
`/run/nixos/activation-reload-list` with their respective counterparts for
dry activation being `/run/nixos/dry-activation-restart-list` and
`/run/nixos/dry-activation-reload-list`. Those files can contain
newline-separated lists of unit names where duplicates are being ignored. These
files are not create automatically and activation scripts must take the
possiblility into account that they have to create them first.
## NixOS snippets {#sec-activation-script-nixos-snippets}
There are some snippets NixOS enables by default because disabling them would
most likely break you system. This section lists a few of them and what they
do:
- `binsh` creates `/bin/sh` which points to the runtime shell
- `etc` sets up the contents of `/etc`, this includes systemd units and
excludes `/etc/passwd`, `/etc/group`, and `/etc/shadow` (which are managed by
the `users` snippet)
- `hostname` sets the system's hostname in the kernel (not in `/etc`)
- `modprobe` sets the path to the `modprobe` binary for module auto-loading
- `nix` prepares the nix store and adds a default initial channel
- `specialfs` is responsible for mounting filesystems like `/proc` and `sys`
- `users` creates and removes users and groups by managing `/etc/passwd`,
`/etc/group` and `/etc/shadow`. This also creates home directories
- `usrbinenv` creates `/usr/bin/env`
- `var` creates some directories in `/var` that are not service-specific
- `wrappers` creates setuid wrappers like `ping` and `sudo`

View File

@ -12,6 +12,7 @@
<xi:include href="../from_md/development/sources.chapter.xml" />
<xi:include href="../from_md/development/writing-modules.chapter.xml" />
<xi:include href="../from_md/development/building-parts.chapter.xml" />
<xi:include href="../from_md/development/what-happens-during-a-system-switch.chapter.xml" />
<xi:include href="../from_md/development/writing-documentation.chapter.xml" />
<xi:include href="../from_md/development/building-nixos.chapter.xml" />
<xi:include href="../from_md/development/nixos-tests.chapter.xml" />

View File

@ -0,0 +1,57 @@
# Unit handling {#sec-unit-handling}
To figure out what units need to be started/stopped/restarted/reloaded, the
script first checks the current state of the system, similar to what `systemctl
list-units` shows. For each of the units, the script goes through the following
checks:
- Is the unit file still in the new system? If not, **stop** the service unless
it sets `X-StopOnRemoval` in the `[Unit]` section to `false`.
- Is it a `.target` unit? If so, **start** it unless it sets
`RefuseManualStart` in the `[Unit]` section to `true` or `X-OnlyManualStart`
in the `[Unit]` section to `true`. Also **stop** the unit again unless it
sets `X-StopOnReconfiguration` to `false`.
- Are the contents of the unit files different? They are compared by parsing
them and comparing their contents. If they are different but only
`X-Reload-Triggers` in the `[Unit]` section is changed, **reload** the unit.
The NixOS module system allows setting these triggers with the option
[systemd.services.\<name\>.reloadTriggers](#opt-systemd.services). If the
unit files differ in any way, the following actions are performed:
- `.path` and `.slice` units are ignored. There is no need to restart them
since changes in their values are applied by systemd when systemd is
reloaded.
- `.mount` units are **reload**ed. These mostly come from the `/etc/fstab`
parser.
- `.socket` units are currently ignored. This is to be fixed at a later
point.
- The rest of the units (mostly `.service` units) are then **reload**ed if
`X-ReloadIfChanged` in the `[Service]` section is set to `true` (exposed
via [systemd.services.\<name\>.reloadIfChanged](#opt-systemd.services)).
- If the reload flag is not set, some more flags decide if the unit is
skipped. These flags are `X-RestartIfChanged` in the `[Service]` section
(exposed via
[systemd.services.\<name\>.restartIfChanged](#opt-systemd.services)),
`RefuseManualStop` in the `[Unit]` section, and `X-OnlyManualStart` in the
`[Unit]` section.
- The rest of the behavior is decided whether the unit has `X-StopIfChanged`
in the `[Service]` section set (exposed via
[systemd.services.\<name\>.stopIfChanged](#opt-systemd.services)). This is
set to `true` by default and must be explicitly turned off if not wanted.
If the flag is enabled, the unit is **stop**ped and then **start**ed. If
not, the unit is **restart**ed. The goal of the flag is to make sure that
the new unit never runs in the old environment which is still in place
before the activation script is run.
- The last thing that is taken into account is whether the unit is a service
and socket-activated. Due to a bug, this is currently only done when
`X-StopIfChanged` is set. If the unit is socket-activated, the socket is
stopped and started, and the service is stopped and to be started by socket
activation.

View File

@ -0,0 +1,53 @@
# What happens during a system switch? {#sec-switching-systems}
Running `nixos-rebuild switch` is one of the more common tasks under NixOS.
This chapter explains some of the internals of this command to make it simpler
for new module developers to configure their units correctly and to make it
easier to understand what is happening and why for curious administrators.
`nixos-rebuild`, like many deployment solutions, calls `switch-to-configuration`
which resides in a NixOS system at `$out/bin/switch-to-configuration`. The
script is called with the action that is to be performed like `switch`, `test`,
`boot`. There is also the `dry-activate` action which does not really perform
the actions but rather prints what it would do if you called it with `test`.
This feature can be used to check what service states would be changed if the
configuration was switched to.
If the action is `switch` or `boot`, the bootloader is updated first so the
configuration will be the next one to boot. Unless `NIXOS_NO_SYNC` is set to
`1`, `/nix/store` is synced to disk.
If the action is `switch` or `test`, the currently running system is inspected
and the actions to switch to the new system are calculated. This process takes
two data sources into account: `/etc/fstab` and the current systemd status.
Mounts and swaps are read from `/etc/fstab` and the corresponding actions are
generated. If a new mount is added, for example, the proper `.mount` unit is
marked to be started. The current systemd state is inspected, the difference
between the current system and the desired configuration is calculated and
actions are generated to get to this state. There are a lot of nuances that can
be controlled by the units which are explained here.
After calculating what should be done, the actions are carried out. The order
of actions is always the same:
- Stop units (`systemctl stop`)
- Run activation script (`$out/activate`)
- See if the activation script requested more units to restart
- Restart systemd if needed (`systemd daemon-reexec`)
- Forget about the failed state of units (`systemctl reset-failed`)
- Reload systemd (`systemctl daemon-reload`)
- Reload systemd user instances (`systemctl --user daemon-reload`)
- Set up tmpfiles (`systemd-tmpfiles --create`)
- Reload units (`systemctl reload`)
- Restart units (`systemctl restart`)
- Start units (`systemctl start`)
- Inspect what changed during these actions and print units that failed and
that were newly started
Most of these actions are either self-explaining but some of them have to do
with our units or the activation script. For this reason, these topics are
explained in the next sections.
```{=docbook}
<xi:include href="unit-handling.section.xml" />
<xi:include href="activation-script.section.xml" />
```

View File

@ -0,0 +1,150 @@
<section xmlns="http://docbook.org/ns/docbook" xmlns:xlink="http://www.w3.org/1999/xlink" xml:id="sec-activation-script">
<title>Activation script</title>
<para>
The activation script is a bash script called to activate the new
configuration which resides in a NixOS system in
<literal>$out/activate</literal>. Since its contents depend on your
system configuration, the contents may differ. This chapter explains
how the script works in general and some common NixOS snippets.
Please be aware that the script is executed on every boot and system
switch, so tasks that can be performed in other places should be
performed there (for example letting a directory of a service be
created by systemd using mechanisms like
<literal>StateDirectory</literal>,
<literal>CacheDirectory</literal>, … or if thats not possible using
<literal>preStart</literal> of the service).
</para>
<para>
Activation scripts are defined as snippets using
<xref linkend="opt-system.activationScripts" />. They can either be
a simple multiline string or an attribute set that can depend on
other snippets. The builder for the activation script will take
these dependencies into account and order the snippets accordingly.
As a simple example:
</para>
<programlisting language="bash">
system.activationScripts.my-activation-script = {
deps = [ &quot;etc&quot; ];
# supportsDryActivation = true;
text = ''
echo &quot;Hallo i bims&quot;
'';
};
</programlisting>
<para>
This example creates an activation script snippet that is run after
the <literal>etc</literal> snippet. The special variable
<literal>supportsDryActivation</literal> can be set so the snippet
is also run when <literal>nixos-rebuild dry-activate</literal> is
run. To differentiate between real and dry activation, the
<literal>$NIXOS_ACTION</literal> environment variable can be read
which is set to <literal>dry-activate</literal> when a dry
activation is done.
</para>
<para>
An activation script can write to special files instructing
<literal>switch-to-configuration</literal> to restart/reload units.
The script will take these requests into account and will
incorperate the unit configuration as described above. This means
that the activation script will <quote>fake</quote> a modified unit
file and <literal>switch-to-configuration</literal> will act
accordingly. By doing so, configuration like
<link linkend="opt-systemd.services">systemd.services.&lt;name&gt;.restartIfChanged</link>
is respected. Since the activation script is run
<emphasis role="strong">after</emphasis> services are already
stopped,
<link linkend="opt-systemd.services">systemd.services.&lt;name&gt;.stopIfChanged</link>
cannot be taken into account anymore and the unit is always
restarted instead of being stopped and started afterwards.
</para>
<para>
The files that can be written to are
<literal>/run/nixos/activation-restart-list</literal> and
<literal>/run/nixos/activation-reload-list</literal> with their
respective counterparts for dry activation being
<literal>/run/nixos/dry-activation-restart-list</literal> and
<literal>/run/nixos/dry-activation-reload-list</literal>. Those
files can contain newline-separated lists of unit names where
duplicates are being ignored. These files are not create
automatically and activation scripts must take the possiblility into
account that they have to create them first.
</para>
<section xml:id="sec-activation-script-nixos-snippets">
<title>NixOS snippets</title>
<para>
There are some snippets NixOS enables by default because disabling
them would most likely break you system. This section lists a few
of them and what they do:
</para>
<itemizedlist spacing="compact">
<listitem>
<para>
<literal>binsh</literal> creates <literal>/bin/sh</literal>
which points to the runtime shell
</para>
</listitem>
<listitem>
<para>
<literal>etc</literal> sets up the contents of
<literal>/etc</literal>, this includes systemd units and
excludes <literal>/etc/passwd</literal>,
<literal>/etc/group</literal>, and
<literal>/etc/shadow</literal> (which are managed by the
<literal>users</literal> snippet)
</para>
</listitem>
<listitem>
<para>
<literal>hostname</literal> sets the systems hostname in the
kernel (not in <literal>/etc</literal>)
</para>
</listitem>
<listitem>
<para>
<literal>modprobe</literal> sets the path to the
<literal>modprobe</literal> binary for module auto-loading
</para>
</listitem>
<listitem>
<para>
<literal>nix</literal> prepares the nix store and adds a
default initial channel
</para>
</listitem>
<listitem>
<para>
<literal>specialfs</literal> is responsible for mounting
filesystems like <literal>/proc</literal> and
<literal>sys</literal>
</para>
</listitem>
<listitem>
<para>
<literal>users</literal> creates and removes users and groups
by managing <literal>/etc/passwd</literal>,
<literal>/etc/group</literal> and
<literal>/etc/shadow</literal>. This also creates home
directories
</para>
</listitem>
<listitem>
<para>
<literal>usrbinenv</literal> creates
<literal>/usr/bin/env</literal>
</para>
</listitem>
<listitem>
<para>
<literal>var</literal> creates some directories in
<literal>/var</literal> that are not service-specific
</para>
</listitem>
<listitem>
<para>
<literal>wrappers</literal> creates setuid wrappers like
<literal>ping</literal> and <literal>sudo</literal>
</para>
</listitem>
</itemizedlist>
</section>
</section>

View File

@ -0,0 +1,119 @@
<section xmlns="http://docbook.org/ns/docbook" xmlns:xlink="http://www.w3.org/1999/xlink" xml:id="sec-unit-handling">
<title>Unit handling</title>
<para>
To figure out what units need to be
started/stopped/restarted/reloaded, the script first checks the
current state of the system, similar to what
<literal>systemctl list-units</literal> shows. For each of the
units, the script goes through the following checks:
</para>
<itemizedlist>
<listitem>
<para>
Is the unit file still in the new system? If not,
<emphasis role="strong">stop</emphasis> the service unless it
sets <literal>X-StopOnRemoval</literal> in the
<literal>[Unit]</literal> section to <literal>false</literal>.
</para>
</listitem>
<listitem>
<para>
Is it a <literal>.target</literal> unit? If so,
<emphasis role="strong">start</emphasis> it unless it sets
<literal>RefuseManualStart</literal> in the
<literal>[Unit]</literal> section to <literal>true</literal> or
<literal>X-OnlyManualStart</literal> in the
<literal>[Unit]</literal> section to <literal>true</literal>.
Also <emphasis role="strong">stop</emphasis> the unit again
unless it sets <literal>X-StopOnReconfiguration</literal> to
<literal>false</literal>.
</para>
</listitem>
<listitem>
<para>
Are the contents of the unit files different? They are compared
by parsing them and comparing their contents. If they are
different but only <literal>X-Reload-Triggers</literal> in the
<literal>[Unit]</literal> section is changed,
<emphasis role="strong">reload</emphasis> the unit. The NixOS
module system allows setting these triggers with the option
<link linkend="opt-systemd.services">systemd.services.&lt;name&gt;.reloadTriggers</link>.
If the unit files differ in any way, the following actions are
performed:
</para>
<itemizedlist>
<listitem>
<para>
<literal>.path</literal> and <literal>.slice</literal> units
are ignored. There is no need to restart them since changes
in their values are applied by systemd when systemd is
reloaded.
</para>
</listitem>
<listitem>
<para>
<literal>.mount</literal> units are
<emphasis role="strong">reload</emphasis>ed. These mostly
come from the <literal>/etc/fstab</literal> parser.
</para>
</listitem>
<listitem>
<para>
<literal>.socket</literal> units are currently ignored. This
is to be fixed at a later point.
</para>
</listitem>
<listitem>
<para>
The rest of the units (mostly <literal>.service</literal>
units) are then <emphasis role="strong">reload</emphasis>ed
if <literal>X-ReloadIfChanged</literal> in the
<literal>[Service]</literal> section is set to
<literal>true</literal> (exposed via
<link linkend="opt-systemd.services">systemd.services.&lt;name&gt;.reloadIfChanged</link>).
</para>
</listitem>
<listitem>
<para>
If the reload flag is not set, some more flags decide if the
unit is skipped. These flags are
<literal>X-RestartIfChanged</literal> in the
<literal>[Service]</literal> section (exposed via
<link linkend="opt-systemd.services">systemd.services.&lt;name&gt;.restartIfChanged</link>),
<literal>RefuseManualStop</literal> in the
<literal>[Unit]</literal> section, and
<literal>X-OnlyManualStart</literal> in the
<literal>[Unit]</literal> section.
</para>
</listitem>
<listitem>
<para>
The rest of the behavior is decided whether the unit has
<literal>X-StopIfChanged</literal> in the
<literal>[Service]</literal> section set (exposed via
<link linkend="opt-systemd.services">systemd.services.&lt;name&gt;.stopIfChanged</link>).
This is set to <literal>true</literal> by default and must
be explicitly turned off if not wanted. If the flag is
enabled, the unit is
<emphasis role="strong">stop</emphasis>ped and then
<emphasis role="strong">start</emphasis>ed. If not, the unit
is <emphasis role="strong">restart</emphasis>ed. The goal of
the flag is to make sure that the new unit never runs in the
old environment which is still in place before the
activation script is run.
</para>
</listitem>
<listitem>
<para>
The last thing that is taken into account is whether the
unit is a service and socket-activated. Due to a bug, this
is currently only done when
<literal>X-StopIfChanged</literal> is set. If the unit is
socket-activated, the socket is stopped and started, and the
service is stopped and to be started by socket activation.
</para>
</listitem>
</itemizedlist>
</listitem>
</itemizedlist>
</section>

View File

@ -0,0 +1,122 @@
<chapter xmlns="http://docbook.org/ns/docbook" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xi="http://www.w3.org/2001/XInclude" xml:id="sec-switching-systems">
<title>What happens during a system switch?</title>
<para>
Running <literal>nixos-rebuild switch</literal> is one of the more
common tasks under NixOS. This chapter explains some of the
internals of this command to make it simpler for new module
developers to configure their units correctly and to make it easier
to understand what is happening and why for curious administrators.
</para>
<para>
<literal>nixos-rebuild</literal>, like many deployment solutions,
calls <literal>switch-to-configuration</literal> which resides in a
NixOS system at <literal>$out/bin/switch-to-configuration</literal>.
The script is called with the action that is to be performed like
<literal>switch</literal>, <literal>test</literal>,
<literal>boot</literal>. There is also the
<literal>dry-activate</literal> action which does not really perform
the actions but rather prints what it would do if you called it with
<literal>test</literal>. This feature can be used to check what
service states would be changed if the configuration was switched
to.
</para>
<para>
If the action is <literal>switch</literal> or
<literal>boot</literal>, the bootloader is updated first so the
configuration will be the next one to boot. Unless
<literal>NIXOS_NO_SYNC</literal> is set to <literal>1</literal>,
<literal>/nix/store</literal> is synced to disk.
</para>
<para>
If the action is <literal>switch</literal> or
<literal>test</literal>, the currently running system is inspected
and the actions to switch to the new system are calculated. This
process takes two data sources into account:
<literal>/etc/fstab</literal> and the current systemd status. Mounts
and swaps are read from <literal>/etc/fstab</literal> and the
corresponding actions are generated. If a new mount is added, for
example, the proper <literal>.mount</literal> unit is marked to be
started. The current systemd state is inspected, the difference
between the current system and the desired configuration is
calculated and actions are generated to get to this state. There are
a lot of nuances that can be controlled by the units which are
explained here.
</para>
<para>
After calculating what should be done, the actions are carried out.
The order of actions is always the same:
</para>
<itemizedlist spacing="compact">
<listitem>
<para>
Stop units (<literal>systemctl stop</literal>)
</para>
</listitem>
<listitem>
<para>
Run activation script (<literal>$out/activate</literal>)
</para>
</listitem>
<listitem>
<para>
See if the activation script requested more units to restart
</para>
</listitem>
<listitem>
<para>
Restart systemd if needed
(<literal>systemd daemon-reexec</literal>)
</para>
</listitem>
<listitem>
<para>
Forget about the failed state of units
(<literal>systemctl reset-failed</literal>)
</para>
</listitem>
<listitem>
<para>
Reload systemd (<literal>systemctl daemon-reload</literal>)
</para>
</listitem>
<listitem>
<para>
Reload systemd user instances
(<literal>systemctl --user daemon-reload</literal>)
</para>
</listitem>
<listitem>
<para>
Set up tmpfiles (<literal>systemd-tmpfiles --create</literal>)
</para>
</listitem>
<listitem>
<para>
Reload units (<literal>systemctl reload</literal>)
</para>
</listitem>
<listitem>
<para>
Restart units (<literal>systemctl restart</literal>)
</para>
</listitem>
<listitem>
<para>
Start units (<literal>systemctl start</literal>)
</para>
</listitem>
<listitem>
<para>
Inspect what changed during these actions and print units that
failed and that were newly started
</para>
</listitem>
</itemizedlist>
<para>
Most of these actions are either self-explaining but some of them
have to do with our units or the activation script. For this reason,
these topics are explained in the next sections.
</para>
<xi:include href="unit-handling.section.xml" />
<xi:include href="activation-script.section.xml" />
</chapter>