nixpkgs/nixos/modules/tasks/filesystems.nix
Florian Klink 3c74e48d9c nixos/filesystems: ensure keys gid on /run/keys mountpoint
boot.specialFileSystems is used to describe mount points to be set up in
stage 1 and 2.

We use it to create /run/keys already there, so sshd-in-initrd scenarios
can consume keys sent over through nixops send-keys.

However, it seems the kernel only supports the gid=… option for tmpfs,
not ramfs, causing /run/keys to be owned by the root group, not keys
group.

This was/is worked around in nixops by running a chown root:keys
/run/keys whenever pushing keys [1], and as machines had to have pushed keys
to be usable, this was pretty much always the case.

This is causing regressions in setups not provisioned via nixops, that
still use /run/keys for secrets (through cloud provider startup scripts
for example), as suddenly being an owner of the "keys" group isn't
enough to access the folder.

This PR removes the defunct gid=… option in the mount script called in
stage 1 and 2, and introduces a tmpfiles rule which takes care of fixing
up permissions as part of sysinit.target (very early in systemd bootup,
so before regular services are started).

In case of nixops deployments, this doesn't change anything.
nixops-based deployments receiving secrets from nixops send-keys in
initrd will simply have the permissions already set once tmpfiles is
started.

Fixes #42344

[1]: 884d6c3994/nixops/backends/__init__.py (L267-L269)
2020-02-05 01:53:26 +01:00

331 lines
12 KiB
Nix
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

{ config, lib, pkgs, utils, ... }:
with lib;
with utils;
let
addCheckDesc = desc: elemType: check: types.addCheck elemType check
// { description = "${elemType.description} (with check: ${desc})"; };
nonEmptyStr = addCheckDesc "non-empty" types.str
(x: x != "" && ! (all (c: c == " " || c == "\t") (stringToCharacters x)));
fileSystems' = toposort fsBefore (attrValues config.fileSystems);
fileSystems = if fileSystems' ? result
then # use topologically sorted fileSystems everywhere
fileSystems'.result
else # the assertion below will catch this,
# but we fall back to the original order
# anyway so that other modules could check
# their assertions too
(attrValues config.fileSystems);
prioOption = prio: optionalString (prio != null) " pri=${toString prio}";
specialFSTypes = [ "proc" "sysfs" "tmpfs" "ramfs" "devtmpfs" "devpts" ];
coreFileSystemOpts = { name, config, ... }: {
options = {
mountPoint = mkOption {
example = "/mnt/usb";
type = nonEmptyStr;
description = "Location of the mounted the file system.";
};
device = mkOption {
default = null;
example = "/dev/sda";
type = types.nullOr nonEmptyStr;
description = "Location of the device.";
};
fsType = mkOption {
default = "auto";
example = "ext3";
type = nonEmptyStr;
description = "Type of the file system.";
};
options = mkOption {
default = [ "defaults" ];
example = [ "data=journal" ];
description = "Options used to mount the file system.";
type = types.listOf nonEmptyStr;
};
};
config = {
mountPoint = mkDefault name;
device = mkIf (elem config.fsType specialFSTypes) (mkDefault config.fsType);
};
};
fileSystemOpts = { config, ... }: {
options = {
label = mkOption {
default = null;
example = "root-partition";
type = types.nullOr nonEmptyStr;
description = "Label of the device (if any).";
};
autoFormat = mkOption {
default = false;
type = types.bool;
description = ''
If the device does not currently contain a filesystem (as
determined by <command>blkid</command>, then automatically
format it with the filesystem type specified in
<option>fsType</option>. Use with caution.
'';
};
formatOptions = mkOption {
default = "";
type = types.str;
description = ''
If <option>autoFormat</option> option is set specifies
extra options passed to mkfs.
'';
};
autoResize = mkOption {
default = false;
type = types.bool;
description = ''
If set, the filesystem is grown to its maximum size before
being mounted. (This is typically the size of the containing
partition.) This is currently only supported for ext2/3/4
filesystems that are mounted during early boot.
'';
};
noCheck = mkOption {
default = false;
type = types.bool;
description = "Disable running fsck on this filesystem.";
};
};
config = let
defaultFormatOptions =
# -F needed to allow bare block device without partitions
if (builtins.substring 0 3 config.fsType) == "ext" then "-F"
# -q needed for non-interactive operations
else if config.fsType == "jfs" then "-q"
# (same here)
else if config.fsType == "reiserfs" then "-q"
else null;
in {
options = mkIf config.autoResize [ "x-nixos.autoresize" ];
formatOptions = mkIf (defaultFormatOptions != null) (mkDefault defaultFormatOptions);
};
};
# Makes sequence of `specialMount device mountPoint options fsType` commands.
# `systemMount` should be defined in the sourcing script.
makeSpecialMounts = mounts:
pkgs.writeText "mounts.sh" (concatMapStringsSep "\n" (mount: ''
specialMount "${mount.device}" "${mount.mountPoint}" "${concatStringsSep "," mount.options}" "${mount.fsType}"
'') mounts);
in
{
###### interface
options = {
fileSystems = mkOption {
default = {};
example = literalExample ''
{
"/".device = "/dev/hda1";
"/data" = {
device = "/dev/hda2";
fsType = "ext3";
options = [ "data=journal" ];
};
"/bigdisk".label = "bigdisk";
}
'';
type = types.loaOf (types.submodule [coreFileSystemOpts fileSystemOpts]);
description = ''
The file systems to be mounted. It must include an entry for
the root directory (<literal>mountPoint = "/"</literal>). Each
entry in the list is an attribute set with the following fields:
<literal>mountPoint</literal>, <literal>device</literal>,
<literal>fsType</literal> (a file system type recognised by
<command>mount</command>; defaults to
<literal>"auto"</literal>), and <literal>options</literal>
(the mount options passed to <command>mount</command> using the
<option>-o</option> flag; defaults to <literal>[ "defaults" ]</literal>).
Instead of specifying <literal>device</literal>, you can also
specify a volume label (<literal>label</literal>) for file
systems that support it, such as ext2/ext3 (see <command>mke2fs
-L</command>).
'';
};
system.fsPackages = mkOption {
internal = true;
default = [ ];
description = "Packages supplying file system mounters and checkers.";
};
boot.supportedFilesystems = mkOption {
default = [ ];
example = [ "btrfs" ];
type = types.listOf types.str;
description = "Names of supported filesystem types.";
};
boot.specialFileSystems = mkOption {
default = {};
type = types.loaOf (types.submodule coreFileSystemOpts);
internal = true;
description = ''
Special filesystems that are mounted very early during boot.
'';
};
};
###### implementation
config = {
assertions = let
ls = sep: concatMapStringsSep sep (x: x.mountPoint);
notAutoResizable = fs: fs.autoResize && !(hasPrefix "ext" fs.fsType || fs.fsType == "f2fs");
in [
{ assertion = ! (fileSystems' ? cycle);
message = "The fileSystems option can't be topologically sorted: mountpoint dependency path ${ls " -> " fileSystems'.cycle} loops to ${ls ", " fileSystems'.loops}";
}
{ assertion = ! (any notAutoResizable fileSystems);
message = let
fs = head (filter notAutoResizable fileSystems);
in
"Mountpoint '${fs.mountPoint}': 'autoResize = true' is not supported for 'fsType = \"${fs.fsType}\"':${if fs.fsType == "auto" then " fsType has to be explicitly set and" else ""} only the ext filesystems and f2fs support it.";
}
];
# Export for use in other modules
system.build.fileSystems = fileSystems;
system.build.earlyMountScript = makeSpecialMounts (toposort fsBefore (attrValues config.boot.specialFileSystems)).result;
boot.supportedFilesystems = map (fs: fs.fsType) fileSystems;
# Add the mount helpers to the system path so that `mount' can find them.
system.fsPackages = [ pkgs.dosfstools ];
environment.systemPackages = with pkgs; [ fuse3 fuse ] ++ config.system.fsPackages;
environment.etc.fstab.text =
let
fsToSkipCheck = [ "none" "bindfs" "btrfs" "zfs" "tmpfs" "nfs" "vboxsf" "glusterfs" ];
skipCheck = fs: fs.noCheck || fs.device == "none" || builtins.elem fs.fsType fsToSkipCheck;
# https://wiki.archlinux.org/index.php/fstab#Filepath_spaces
escape = string: builtins.replaceStrings [ " " "\t" ] [ "\\040" "\\011" ] string;
in ''
# This is a generated file. Do not edit!
#
# To make changes, edit the fileSystems and swapDevices NixOS options
# in your /etc/nixos/configuration.nix file.
# Filesystems.
${concatMapStrings (fs:
(if fs.device != null then escape fs.device
else if fs.label != null then "/dev/disk/by-label/${escape fs.label}"
else throw "No device specified for mount point ${fs.mountPoint}.")
+ " " + escape fs.mountPoint
+ " " + fs.fsType
+ " " + builtins.concatStringsSep "," fs.options
+ " 0"
+ " " + (if skipCheck fs then "0" else
if fs.mountPoint == "/" then "1" else "2")
+ "\n"
) fileSystems}
# Swap devices.
${flip concatMapStrings config.swapDevices (sw:
"${sw.realDevice} none swap${prioOption sw.priority}\n"
)}
'';
# Provide a target that pulls in all filesystems.
systemd.targets.fs =
{ description = "All File Systems";
wants = [ "local-fs.target" "remote-fs.target" ];
};
# Emit systemd services to format requested filesystems.
systemd.services =
let
formatDevice = fs:
let
mountPoint' = "${escapeSystemdPath fs.mountPoint}.mount";
device' = escapeSystemdPath fs.device;
device'' = "${device'}.device";
in nameValuePair "mkfs-${device'}"
{ description = "Initialisation of Filesystem ${fs.device}";
wantedBy = [ mountPoint' ];
before = [ mountPoint' "systemd-fsck@${device'}.service" ];
requires = [ device'' ];
after = [ device'' ];
path = [ pkgs.utillinux ] ++ config.system.fsPackages;
script =
''
if ! [ -e "${fs.device}" ]; then exit 1; fi
# FIXME: this is scary. The test could be more robust.
type=$(blkid -p -s TYPE -o value "${fs.device}" || true)
if [ -z "$type" ]; then
echo "creating ${fs.fsType} filesystem on ${fs.device}..."
mkfs.${fs.fsType} ${fs.formatOptions} "${fs.device}"
fi
'';
unitConfig.RequiresMountsFor = [ "${dirOf fs.device}" ];
unitConfig.DefaultDependencies = false; # needed to prevent a cycle
serviceConfig.Type = "oneshot";
};
in listToAttrs (map formatDevice (filter (fs: fs.autoFormat) fileSystems));
systemd.tmpfiles.rules = [
"Z /run/keys 0750 root ${toString config.ids.gids.keys}"
];
# Sync mount options with systemd's src/core/mount-setup.c: mount_table.
boot.specialFileSystems = {
"/proc" = { fsType = "proc"; options = [ "nosuid" "noexec" "nodev" ]; };
"/run" = { fsType = "tmpfs"; options = [ "nosuid" "nodev" "strictatime" "mode=755" "size=${config.boot.runSize}" ]; };
"/dev" = { fsType = "devtmpfs"; options = [ "nosuid" "strictatime" "mode=755" "size=${config.boot.devSize}" ]; };
"/dev/shm" = { fsType = "tmpfs"; options = [ "nosuid" "nodev" "strictatime" "mode=1777" "size=${config.boot.devShmSize}" ]; };
"/dev/pts" = { fsType = "devpts"; options = [ "nosuid" "noexec" "mode=620" "ptmxmode=0666" "gid=${toString config.ids.gids.tty}" ]; };
# To hold secrets that shouldn't be written to disk
"/run/keys" = { fsType = "ramfs"; options = [ "nosuid" "nodev" "mode=750" ]; };
} // optionalAttrs (!config.boot.isContainer) {
# systemd-nspawn populates /sys by itself, and remounting it causes all
# kinds of weird issues (most noticeably, waiting for host disk device
# nodes).
"/sys" = { fsType = "sysfs"; options = [ "nosuid" "noexec" "nodev" ]; };
};
};
}