Consider the scenario where you run DHCPv4 on the primary bridge
interface.
You have no real interface to "wait upon", so it's OK. Nonetheless,
anything depending on successful completion of DHCPv4, e.g. adding a
default route, will block `s6-rc -v2 up change default`.
The way new interfaces are attached to the bridge is via `s6-rc -b -u
change $attach-oneshot-service`, this introduce in turn a deadlock.
At some point, DHCPv4 will timeout, unblocking the deadlock and
attaching the members to the primary bridge interface, making it ready
to send L2 broadcast packets for DHCP, unblocking DHCP in turn again.
This is not satisfying because we really want to have a no-hiccups
bring-up.
To fix this, we proceed to multiple changes:
- we remove `svc.ifwait.build` out of band `s6-rc -b -u $oneshot-attach`
call, which is, by design, wrong here.
- users can now depend on the members service to know when a bridge is
fully operational (we could make it more granular and let them depend
on the LAN member joining rather than WLAN, etc.)
- users can also depend on the primary service being brought up rather
than just being present, this is useful if you need to bring it up
when it has AT LEAST one member to get link local address or MAC
addresses (fixing DHCPv6 bring up as well because `ff02::1` is used
there).
One thing is not addressed yet, if you are running a WLAN service using
RADIUS attached to the bridge, at bring up time, it will try to reach
out the external RADIUS server and *fail*.
To solve this, granular dependency on the DHCPv4 once LAN is joined.
Then the hostapd can wait on defaultroute4 completion so that
connectivity is available to reach RADIUS server.
It can join the bridge later on without any hiccup as well.
This is left as a TODO as hostapd can survive RADIUS authentication
failure and retry later.
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
This avoids the OPERSTATE unknown when the bridge is brought up but the
members are not ready yet.
This will make OPERSTATE to down, enabling us to wait until we have
brought up completely all the members.
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
The way the parsing works is examining one character at a time.
First, if we had `rootfstype=... root=...`, the parsing would jump and
ignore `root=...`, which sucks.
To fix this, we scan multiple times a copy of the cmdline.
Now, we have a new problem: `root=... altroot=...` lead to opts.device
being equal to the altroot as we are looking one char at a time, so we
will arrive at a moment looking at `root=...` for `altroot=...`.
To avoid this, we rename `altroot` in `rootalt`, cheap, I know.
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
Normal command line and TFTP command line can be sometimes very
different.
e.g. We don't want to load UBI filesystems for a TFTP boot as it may
interfere with our root device loading.
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
It's not a proper overlay because it breaks the splicing for luaOnBuild
and therefore breaks the rest.
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
e.g. colmena can make use of it to declare an heterogeneous set of
systems, mixing Liminix and NixOS systems.
Signed-off-by: Raito Bezarius <masterancpp@gmail.com>
s6-supervise sends signals (e.g. SIGTERM) to the pid of the process
running "run", so how do we know if the ceanup commands are even
getting executed if the shell interpreter that is supposed to do that
got killed already?
