On some system, certain interfaces are not usable by an end-user. The
canonical example of this is the ethernet MAC connected to a switch,
i.e. the DSA port.
Avoid having these be picked up in the generated factory-config, which
would be confusing to a user, by classifying them as "internal" ports
to be ignored by confd and friends.
The kernel's mounting of the rootfs may race against the underlying
device being probed - especially for media like MMC cards and USB
sticks.
Therefore, tell the kernel to wait for the root device to show up.
In the future we want to build disk images from builds where we
haven't locally built Infix, we've just downloaded an existing
release. In those cases we also have to create rauc.status, even
though we're not creating a bundle.
Make the traditional FITs smaller by:
- Filtering out the kernel image from the rootfs, since the FIT way is
to ship it separately.
- Gzipping the kernel in the FIT, since aarch64 uses the uncompressed
"Image" format, and squashfs won't take care of it for us.
- Always respect the requested model to emulate. Before this change, a
device configured to run in TAP mode would ignore a user's requested
model and always emulate a e1000.
This worked on x86_64 where that driver is part of the kernel, but
that is not true for aarch64.
- Always supply a mactab. It was very confusing that the interface
names would change, in Infix, depending on the selected networking
mode.
- Remove old decoy variables that were assigned but never used
anywhere.
- Separately setup the host and target side of a nic (i.e. use a
-netdev/-device pair instead of the -nic shortcut). This lets us
share the host side setup across all modes.
This reverts an earlier change where nano was made the default editor in
a misguided attemt by yours truly to accomodate beginner users.
With the recent improvements of the CLI and the fact that the NETCONF
support has matured greatly, the need for a user-friendly editor have
diminished drastically. Editing system files is now limited to devs
and the core team prefers Mg over nano and vi.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
This is not just for online viewing, these markdown documents are also
made available in the CLI itself from admin-exec as:
help [introduction | keybindings | tutorial]
With this commit we also add lowdown and the real less application to
be the Infix help system.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
This commit enables a basic sanity check of usernames in BusyBox that by
default is disabled (!). The rules enabled by this are:
- the user or group name should consist only of letters, digits,
underscores, periods, and dashes¹, and not start with a dash (as
defined by IEEE Std 1003.1-2001), or periods²
- for compatibility with Samba machine accounts, $ is supported at the
end of names
- max length of user or group name is 256 chars
Changes in this commit also includes a sync with latest BusyBox 1.36,
which adds: tree, tsort, SHA hw acceleration, and a separate script
for udhcpc6.
_____
¹) The documentation actually mentions "and at-signs" as well, but that
has been explicitly removed from the source code.
²) The documentation does not mention that leading periods are not
allowed, though this is also explicitly mentioned in the source code.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
- Add /bin/clish (symlink) to /etc/shells, as well as the traditional
/bin/{true,false} for completeness
- Validate shell string, with fallback to /bin/bash and /bin/sh
- Step infix-system revision, keeping old model in tree
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
This allows an application to set a larger socket receive buffer size.
The default size is left unchanged (212992). 1MB is arbitrary chosen.
The main reason for this is to allow netlink kernel sockets to
hold more interface data while user-space is processing it. Typically
useful during main configuration changes where lots of kernel side
changes occur in short period of time, such as VLAN creation.
Signed-off-by: Richard Alpe <richard@bit42.se>
This let's us report the installed software versions, before any
updates have been made. Since the info should _not_ be erased in a
factory reset (since the installed software versions will not change),
move the status file to aux, which is the proper place for this class
of information.
- When running with full disk emulation (via GRUB on x86, or U-Boot on
aarch64), there's no way for us to influence the commandline passed
to the kernel, so the console is fixed to ttyS0/AMA0. Therefore,
reintroduce an option to choose the console.
- Use a fixed `nr`s for virtconsoles, so that the gdbserver always
listens on `hvc1`. Ideally, we should use udevs `/dev/virtio-ports/*`
symlinks, but for some reason these point to `/dev/vportXpY` and not
to `/dev/hvcY`. TODO: Figure out what the difference is.
- While we're here, remove the duplicated RTC setup which was
introduced with the virtio-console work a while back.
/dev/ram is accepted by the kernel during boot, but there's no actual
file with this name in /dev, so RAUC will get confused about which
slot is actually one we booted from.
This patch affects the various artiact file names in release tarballs.
The files /etc/os-release and /etc/version in the image use the actual
GIT commit (git describe), as per post-build.sh
Guiding principle: artifacts without release suffix are development or
"nightly" builds. Artifacts *with* release suffix are released ones.
This drops the $GITVER from .img files for consistency with the other
artifacts, e.g. .gns3a and .pkg.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
To set up proper factory defaults in Infix we need to drop the default
/etc/sysrepo datastores, initialized at compile-time by Buildroot when
installing netopeer2. The pre-initialized datastores cannot be set up
with new {factory} data in any other way.
The system factory-config is constructed from two parts: static data and
generated data. The latter is:
- hostname (product name + the three last octets of the base MAC)
- interface defaults (depends on number of ports on the device) that
enable IPv6 link-local on all interfaces, and
- per-device unique SSH hostkey for remote NETCONF access, requires
openssl binary on target. Can maybe use openssh for this (TODO).
The hostkeys are saved in /cfg/factory.d/ to prevent them from being
regenerated on every boot.
The script gen-hostkeys and the update of 10-netconf-server.json are
inspired by the netopeer2 scripts merge_hostkey.sh and merge_config.sh,
which do just about the same thing for regular systems.
We bootstrap sysrepo early with all the required models, netopeer2 as
well as our own, by calling sysrepoctl with initial data for {factory}
and {startup} datastores. A local patch to sysrepo is used to allow it
to look for an /etc/factory-default.json instead of compile-time default
data, the patch allows us to have variable amount of objects in the init
data, e.g. number of interface objects.
A system startup-config is generated, if one is mising, from {running}.
We then drop the initial {running} datastore before starting the daemons
sysrepo-plugind and netopeer2-server so that they, and our plugins, are
ready when we load the system startup-config.
Some sanity checks and debug possibilities:
- Ensure /etc/sysrepo exists before copying factory-config
- Keep factory-default.json for debug
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
Primarily intended for minimal/testing images, but likely also useful
for other use-cases as well.
Defaults remain to sign all images.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
Since qemu.sh is now tied to the directory it's ran from, it makes
sense to change into that directory before starting to resolve any
file paths. This let's us start the script from any directory without
the need for any (cd && qemu) subshell.
This change drops Qemu menuconfig from the regular 'make menuconfig' in
favor of full release tarball support. Meaning "make update-defconfig"
will not longer cause unwanted diffs for Qemu defaults checked in when
the defconfigs are updated.
Details:
- For "make run, run-menuconfig" and bootstrapping qemu.cfg we want to
use the Buildroot menuconfig system. However, we must set the proper
environment variables for it to not overwrite our product's .config,
so the calls to conf and mconf are not pretty.
- The qemu.sh script can probably be cleaned up further, I've just set
a CONFIG_ prefix to all variables and dropped any images/ prefix vs.
the original. The CONFIG_ prefix is to make the qemu.cfg compatible
with kconfig-mconf from the kconfig-frontends package.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
First basic step, we also want a menuconfig-like way of changing the
settings, e.g., networking options.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
When evolving to a new generation fails, users will typically abandon
it (so that a new attempt can be made). However, the generated
scripts and the logged output is very valuable from a debugging
perspective.
Therefore, keep the data around by just renaming the generation
directory.
Later, we can include these directories in the garbage collecting
process, if we want to.
Initial analysis of system default users gives that there is no real
need for a default 'factory' user to trigger a factory reset. On an
actual physical device we will have a reset button or a menu option in
the bootloader to trigger factory reset, and in both physical and
virtual (GNS3) devices a logged in admin user can call the factory
command.
A virtual device that cannot be logged in to will have to be removed and
have its persistent backing store deleted, but that is the only real
downside to removing this user. The idea is to set a kernel command
line flag from the bootloader, and that could also be done from, e.g.,
GNS3.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
The /var, and particularly /var/lib, directories are used for persistent
storage for various system daemons. While adding support for containers
it became clear we had to change how we make these directories writable.
Both LXC and Podman use OverlayFS for the writable layer, and you cannot
set up an overlay on top of another overlay.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
Podman (POD manager) is an open source tool for developing, managing,
and running containers on your Linux systems.
Based on work by Christian Stewart and Nathaniel Husted.
Signed-off-by: Joachim Wiberg <troglobit@gmail.com>
A generic replacement for src/net with some advantages:
- Shell script is well-suited to these kinds of tasks, making this
implementation much simpler.
- Allow arbitrary number of actions (not just init/exit), this can be
used to implement things like `dagger exec up/down`.
- Allow arbitrary number of scripts for each action/node tuple. This
will let us handle more complex scenarios in confd without having to
fiddle about with src/net.
