The Express Data Path (XDP) is a way to circumvent the traditional Linux
networking stack and instead run an eBPF program on your NIC, that makes
the decision to provide Knot with certain packets. This is way faster
and more scalable but comes at the cost of reduced introspection.
Unfortunately the `knotc conf-check` command fails hard with missing
interfaces or IP addresses configured in `xdp.listen`, so we disable it
for now, once the `xdp` config section is set. We also promote the config
check condition to a proper option, so our conditions become public
documentation, and we allow users to deal with corner cases, that we have
not thought of yet.
We follow the pre-requisites documented in the Knot 3.3 manual, and set
up the required capabilities and allow the AF_XDP address family.
But on top of that, due to our strict hardening, we found two more
requirements, that were communicated upstream while debugging this.
- There is a requirement on AF_NETLINK, likely to query for and configure
the relevant network interface
- Running eBPF programs requires access to the `bpf` syscall, which we
deny through the `~@privileged` configuration.
In summary We now conditionally loosen the hardening of the unit once we
detect that an XDP configuration is wanted. And since we cannot
introspect arbitrary files from the `settingsFiles` option, we expose XDP
support through the `enableXDP` toggle option on the module.
Murmur provides an official systemd service file in their repo,
which contains various service hardening settings:
c4b5858d14/auxiliary_files/config_files/mumble-server.service.in (L7)
The service configuration in nixpkgs does not include these hardening settings.
This commit adds the hardening settings to the murmur service in nixpkgs.
This drops the `systemd-analyze security` score of murmur.service from 9.2 (UNSAFE) to 2.1 (OK).
With DefaultDependencies enabled, systemd adds "After=basic.target" to
service units. `basic.target` has a dependency on `sockets.target`, so
the `nftables` has (amongst others) the following order constraints:
* Before=network-pre.target
* After=sockets.target
Those constraints are often unsatisfiable. For example, `systemd-networkd`
has a dependency `After=network-pre.target`. When a socket unit now uses
`BindToDevice=` on a device managed by `networkd`, a timeout occurs
because `networkd` waits for `network-pre.target`, but
`network-pre.target` depends (through nftables) on `sockets.target`, but
the device to bind the socket to is never brought up, as this would
happen through `networkd`.
This is fixed by removing the implicit dependency on `basic.target`.
