networkd fails to set transient hostname

[martinpitt]

Submission type

• Bug report
• Request for enhancement (RFE)

systemd version the issue has been seen with

232

Used distribution

Debian

While investigating issue #4504 and what would break when networkd starts before D-Bus, I noticed that the only reason why networkd currently runs after dbus.service is setting a transient hostname received through DHCP. But this doesn't actually seem to work: When DHCP does send a hostname, networkd says

systemd-networkd[516]: ens3: Could not set hostname: Interactive authentication required.

Presumably because networkd drops its privileges to user systemd-network and the polkit policy of hostnamed only allows root to change the host name.

FTR, I reproduced this in QEMU with -net nic,model=virtio -net user,hostname=fizzy which will send a hostname along with DHCP.

Fixing this should go along with actually testing this in test/networkd-test.py, as this is quite a subtle problem that isn't obvious when reviewing PRs for privilege reduction.

[poettering]

i figure adding the "systemd-network" user to the default PK file should be relatively easy. At least for folks who grok PK policy well, which I don't ;-)

[martinpitt]

Confirming that this works if you set <allow_any>yes</allow_any> for org.freedesktop.hostname1.set-hostname in /usr/share/polkit-1/actions/org.freedesktop.hostname1.policy (this is of course only for testing). *.policy files don't have a syntax for "also grant this policy to Unix users, that's where the pklocalauthority(8) files come in -- and unfortunately there's two formats around: the old ini-style declarative ones (Identity=systemd-network\nAction=org.freedesktop.hostname1.set-hostname\nResultAny=yes) and the "free-form JavaScript" ones from newer polkits. IMNSHO these have been a ridiculously bad idea for several reasons and Debian/Ubuntu never got those, but Fedora did. So we could auto-detect the available polkit version during package build, or ship both files and the distro picks, or we provide a ./configure switch for those.

[martinpitt]

martinpitt/systemd@a3ccebc fixes this with both the pkla-based and JS rules based polkit. I think we should just ship both, as (1) they don't really get into each other's way, (2) there is no sensible build-time detection for it, and (3) the user may upgrade polkit at runtime.

I will write a test case for this before I post a PR.

[keszybz]

Agreed.