cross-posted from: https://programming.dev/post/39212874
I recently migrated my services from rootful
docker
to rootlesspodman quadlets
. It went smoothly, since nothing I use actually needs to be rootful. Well, except forcaddy
. It needs to be able to attach to privileged ports 80 and 443.My current way to bypass it is using
HAProxy
running as root and forwarding connections using proxy protocol. (Tried to usefirewalld
but that makes the client IP opaque tocaddy
.) But that adds an extra layer, which means extra latency. It’s perfectly usable, but I’d like to get rid of it, if possible.I’m willing to run
caddy
in rootfulpodman
if needed. But from what I understand, that means I can’t have it in the same rootless network as my other containers. I really don’t wanna open most of my containers’ ports, so that’s not an option.So, I’m asking whether any of these three things are possible.
- Use
firewalld
to forward ports tocaddy
without obscuring the client’s IP.- Make rootful
caddy
share a network with other rootless containers.- Assign privileged ports to caddy somehow, in rootless mode. (I know there’s a way to make all these ports unprivileged, but is it possible to only assign these 2 ports as unprivileged?)
Or maybe there’s a fourth way that I’m missing. I feel like this is a common enough setup, that there must be a way to do it. Any pointers are appreciated, thanks.
Hi,
The client IP problem is a longstanding issue in podman’s virtual bridge networks.
As a workaround I’d run HAProxy rootless, using the
pasta
networking mode as that one allows seeing native client IP. With pasta’s-T
flag (see docs) I’d forward traffic to another caddy container binding to127.0.0.1:8080
or something similar.This would coincide with your firewalld/HAProxy port-forwarding setup, but it has more rootlessness to it. It’s still not perfect and you’d still need to tweak sysctls, but I hope it may be useful
You can use rootless caddy via systemd socket activation, here’s a basic setup:
- rootless-caddy.service
[Unit] Description=rootless-caddy Requires=rootless-caddy.socket After=rootless-caddy.socket [Service] # a non root user here User=El_Quentinator ExecStart=podman run --name caddy --rm -v [...] docker.io/caddy:alpine [Install] WantedBy=default.target
- rootless-caddy.socket
[Socket] BindIPv6Only=both ### sockets for the HTTP reverse proxy # fd/3 ListenStream=[::]:443 # fdgram/4 ListenDatagram=[::]:443 [Install] WantedBy=sockets.target
- Caddyfile
{$SITE_ADDRESS} { # tcp/443 bind fd/3 { protocols h1 h2 } # udp/443 bind fdgram/4 { protocols h3 } [...] }
And that’s it really.
You can find a few more examples over here: https://github.com/eriksjolund/podman-caddy-socket-activation
Systemd socket activation has a few more interesting advantages on top of unlocking binding priviliged ports:
- allowing to not bind any port on the container, which itself allows to use
--network none
while still being able to connect to it via systemd socket, pretty neat to expose some web app while completely cutting its own external access. - getting rootful networking performance
- starting up your service only when connecting to it (and potentially shutting it back down after some time)
Drawbacks is that the file descriptor binding is a bit awkward and not always supported (caddy/nginx/haproxy do support it though). And that podman pods / kube do not support it (or at least not yet).
It seems that I’d still need to modify
net.ipv4.ip_unprivileged_port_start=80
in sysctl, which I don’t want to do. If I do it, the socket isn’t even strictly necessary.TBH I haven’t played with passing caddy’s podman network to other containers, mine is a simple reverse proxy to other standalone containers but not directly connected via
podman run --network
(or quadlet network). In my scenario I can at least confirm thatnet.ipv4.ip_unprivileged_port_start
doesn’t need to be modified, the only annoyance is that I cannot use a systemd user service, even though the end process doesn’t run as root.EDIT: Actually looking at the examples a bit more closely I think the primary difference with my setup is that the systemd socket is started with
systemd --user
which thus requires the sysctl change, whereas I’m not using a systemd user service, relying instead onUser=some-non-root-user
to use rootless podman, but requiring root privileges to manage the systemd service.
This is interesting. I need to figure out how it works for
podman
and it’ll be the perfect setup.
I’m not running your configuration so can’t tell you with the assurance that I have it working but Forwarding ports with firewalld appears to address port forwarding to rootless podman using firewalld. If that doesn’t work for you you might need to clarify what your firewalld configuration is that obscures the client IP. I wouldn’t expect a simple port mapping to affect IP address.
Be aware you might have to resort to nftables if firewalld doesn’t work. I use localhost a lot and the routing rules are different in that case.
I think it’s the masquerade that’s causing problems for me. I have to keep it enabled since I’m running a
tailscale
exit node. But maybe I can selectively disable it here.
Rootless podman caddy doesn’t need those priviliged ports, if you have your server behind a firewall device. You can map your ports on the firewall/router 80:8080 and then on the caddy container 8080:80. This way there is no need for priviliged ports and the traffic seems to go on ports 80 (and 443 the same way).
I mentioned in the post that it seems to make the client IP opaque to
caddy
.I’ve never used your exact setup, but I have had issues with a web server behind a WAF not getting the client IP (all user traffic was shown as the WAF IP). In my case, the WAF was appending the client IP in a header, and I just had to tell web app to use that header as the client IP instead of the actual IP. Again, not sure if this helps since I have never used podman or caddy (this setup was with Wordpress and an Azure Application Gateway) but the same principles might apply.
@SinTan1729 Using privileged ports can be activated with a sysctl setting:
https://access.redhat.com/solutions/7044059But that just makes most ports unprivileged. That is a solution, but less preferred than my current one.
@SinTan1729 How many user do you have on your machine, which could open and run a service on a privileged port?
And when there is no application, which is providing a service on a privileged port, then there is no security issue from my point of view.
And if you want to get absolutely secure, then you can restrict the access only to specific ports based on firewall rules.
https://www.digitalocean.com/community/tutorials/ufw-essentials-common-firewall-rules-and-commands#how-to-allow-all-incoming-http-and-httpsJust a couple of friends use it. But I’d like to use this as a learning opportunity and do it the proper way. It seems that if I turn of masquerade in general, and use
firewalld
fine-grained rules to enable it when I actually need it, I might be able to achieve what I want. I’ll post an update to the original post if I can get it to work.@SinTan1729 Thank you, now I can better understand why you want to avoid to open the privileged ports for non-root users which makes sense for your scenario.
I’m in the easy situation, that I don’t have to think about such a scenario, because my selfhosting system is exclusive for me.
I don’t know the exact agreement with your friends, but to avoid security issues I personally would use following way:
- deny usage of all ports by firewall
- allow only necessary ports by firewall
- enable privileged ports by sysctl
So it reduces additional layers and complexity.If one of your friends would provide a service on a specific port it has to be discussed with you.
And if this is a privileged port, it is also possible.Or you can handle e.g. a web request with a rule in caddy.
Why does it have to be a priviledged port?
Most people expect a domain to work without adding 8080 as a port number in the URL. Hell, I’d say a majority don’t even know that it’s possible.
You don’t have to add 8080.
You can just remove the restriction if you don’t want it. But that’s just a classical xy problem. OP does not want to use 80 just doesn’t know that he doesn’t want it because he’s looking for something else.
I’m confused. What do you even mean?