EC2 instances don't have a console, so it's pointless to start
emergency mode if a mount fails. (This happened to me with an
encrypted filesystem where the key wasn't sent on time using "charon
send-keys".) Better to cross fingers and continue booting.
Starting an authenticated root shell is a security hole, so don't do
it by default. The kernel command line parameter
‘initrd.shell_on_fail’ restores the original. (Of course, this only
improves security if you have a password on GRUB to prevent the kernel
command line from being edited by unauthorized users.)
The 'memtest86' package didn't work on any of my machines. 'memtest86plus', on
the other hand, seems to work just fine. Does anyone know why we keep the
seemingly older version around still?
The "S|s|single" option is handled by systemd (starting
rescue.target). And the rescue target basically removes the need for
a special debug shell. (Also, there is "systemd.crash_shell=1" for
starting a shell if systemd crashes.)
Also, symlink kbrequest.target to rescue.target as suggested by the
systemd.special manpage. This way, you can start a sulogin rescue
shell by pressing Alt+Up.
During a configuration switch, changed units are stopped in the old
configuration, then started in the new configuration (i.e. after
running the activation script and running "systemctl daemon-reload").
This ensures that services are stopped using the ExecStop/ExecStopPost
commands from the old configuration.
However, for some services it's undesirable to stop them; in
particular dhcpcd, which deconfigures its network interfaces when it
stops. This is dangerous when doing remote upgrades - usually things
go right (especially because the switch script ignores SIGHUP), but
not always (see 9aa69885f0). Likewise,
sshd should be kept running for as long as possible to prevent a
lock-out if the switch fails.
So the new option ‘stopIfChanged = false’ causes "systemctl restart"
to be used instead of "systemctl stop" followed by "systemctl start".
This is only proper for services that don't have stop commands. (And
it might not handle dependencies properly in some cases, but I'm not
sure.)
This is mainly useful for specifying mounts that depend on other
units. For example sshfs or davfs need network (and possibly
nameservices).
While systemd makes a distinction between local and remote
filesystems, this only works for in-kernel filesystems such as
nfs and cifs.
fuse-based filesystems (such as sshfs and davs) are classified as
local, so they fail without networking. By explicitly declaring these
mounts as full systemd units (as opposed to having systemd generate
them automatically from /etc/fstab), dependencies can be specified as
on every other unit.
In the future, we can probably port NixOS' filesystems handling to use
these native systemd.mount units and skip /etc/fstab altogether, but
this probably requires additional changes, such as starting systemd
even earlier during boot (stage 1).
I'm not any good at perl, and I only came up with this after many
slow attempts. Any review welcome.
But until this, memtest was broken, and extraPrepareConfig as well, in grub.
Having all services with DefaultDependencies=yes depend on
local-fs.target is annoying, because some of those services might be
necessary to mount local filesystems. For instance, Charon's
send-keys feature requires sshd to be running in order to receive LUKS
encryption keys, which in turn requires dhcpcd, and so on. So we drop
this dependency (and swap.target as well for consistency). If
services require a specific mount, they should use RequiresMountsFor
in any case.
This has rendered my system unbootable, because I forgot to enable AUTOFS4 in my
custom kernel. In addition to AUTOFS4, this includes (hopefully) all other
kernel features needed by systemd, as listed in the README:
REQUIREMENTS:
Linux kernel >= 2.6.39
with devtmpfs
with cgroups (but it's OK to disable all controllers)
optional but strongly recommended: autofs4, ipv6
Autofs4 is not a requirement here, but in our case it turns out that the system
is not able to boot properly with a LUKS-enabled system (or at least not on _my_
system).
Signed-off-by: aszlig <aszlig@redmoonstudios.org>
Running agetty on ttyS0 interferes with the backdoor, which uses ttyS0
as its standard error. After agetty starts, writes to the stderr file
descriptor will return EIO (though doing "exec 2>/proc/self/fd/2" will
miracuously fix this).
http://hydra.nixos.org/build/3252782
‘systemd-vconsole-setup’ by default operates on /dev/tty0, the
currently active tty. Since it puts /dev/tty0 in Unicode or ASCII
mode, if the X server is currently active when it runs, keys such as
Alt-F4 won't reach the X server anymore. So use /dev/tty1 instead.
The triggers are just arbitrary strings that are included in the unit
under X-Restart-Triggers. The idea is that if they change between
reconfigurations, switch-to-configuration will restart the unit
because its store path changed. This is mostly useful for services
that implicitly depend on generated files in /etc. Thus you can say
restartTriggers = [ confFile ];
where ‘confFile’ is the derivation that generated the /etc file in
question.
For each statically configured interface, we now create a unit
‘<interface>-cfg.service’ which gets started as soon as the network
device comes up. Similarly, each bridge defined in
‘networking.bridges’ and virtual interface in ‘networking.interfaces’
is created by a service ‘<interface>.service’.
So if we have
networking.bridges.br0.interfaces = [ "eth0" "eth1" ];
networking.interfaces =
[ { name = "br0";
ipAddress = "192.168.1.1";
}
];
then there will be a unit ‘br0.service’ that depends on
‘sys-subsystem-net-devices-eth0.device’ and
‘sys-subsystem-net-devices-eth1.device’, and a unit ‘br0-cfg.service’
that depends on ‘sys-subsystem-net-devices-br0.device’.
So instead of:
boot.systemd.services."foo".serviceConfig =
''
StartLimitInterval=10
CPUShare=500
'';
you can say:
boot.systemd.services."foo".serviceConfig.StartLimitInterval = 10;
boot.systemd.services."foo".serviceConfig.CPUShare = 500;
This way all unit options are available and users can set/override
options in configuration.nix.
Systemd sets locale variables like $LANG when running services, so
$LOCALE_ARCHIVE should also be set to prevent warnings like "perl:
warning: Setting locale failed.".
kernel 3.4+ needs cifs-utils to mount CIFS filesystems.
the kernel itself (and busybox's cifs mount code) are no longer able
to do this in some/most cases and will error out saying:
"CIFS VFS: connecting to DFS root not implemented yet"
Nixos' qemu-vm target is hurt by this, as it wants to mount /nix/store
via cifs very early in the boot process.
This commit makes sure the initrd for affected kernels is built with
cifs-utils if needed.
Subtle: dhcpcd.service would call resolvconf during shutdown, which in
turn would start invalidate-nscd.service, causing the shutdown to be
cancelled. Instead, give nscd.service a proper reload action, and do
"systemctl reload --no-block nscd.service". The --no-block is
necessary to prevent that command from waiting until a timeout occurs
(bug in systemd?).
As non-QWERTY keyboards don't feel so warm and cozy if they hug QWERTY LUKS
password prompts, it was on honor for me to serve King Dvorak XV to fight the
glorious keyboard war against... what?! Yes, I'm awake!
We're fighting with loadkeys to spit out busybox binary keymaps against loadkmap
(loadkeys does have a special target -b for that).
And yep, I'm somewhat abusing preLVMCommands, if someone got issues with that,
feel free to introduce a new substitute in stage-i-init.sh.
Sent from my iPhone
We had a "mount -o remount,rw none /" that was setting back 'relatime',
although we had set 'noatime' at initrd mount. Removing the word 'none' fixed
it.
Specifying a device (in this case 'none'), makes mount to forget previous
device options. According to manpage, it says not to read fstab or mtab. But the
effect is that of setting 'relatime', if it was mounted 'noatime.
