Support passing FDs (socket activation) about caddy

[flokli]

Sockets can have names attached (so the user can name them http and https for example, or api and metrics), and we could add a syntax to refer to them via these names in Caddyfile. I could say I want a http server on sd-listen:http, which would then expect a listener named http to be passed to caddy.

All these passed FDs also give you a net.Listener interface, so even without explicit config caddy could still check the properties of it and apply some heuristics too (detect port 80 and 443 if you got two unnamed TCP sockets), if we want to apply some out-of-the-box behaviour in these scenarios. But getting the basic support for it (using an externally-passed FD by its name/index) and defining the syntax for it would be a nice first step.

You can play around with this through systemd-socket-activate -l 8088 -l 8089 --fdname=foo:bar -- /path/to-caddy, which will give you two TCP sockets listening on the two ports, named foo and bar.

[WeidiDeng]

@flokli So that means you're fine with mixing passing FD and current binding behavior? And since you will use bind explicitly, it's an error to bind to an non existent FD.

The problem with names is that one name can map to many sockets with different addresses, how do you think caddy handle this situation?

[mohammed90]

for those that just care about binding to ports <1024 AND not running Caddy as root,

There was never a need to run Caddy as root on Linux. Our standard systemd unit file is shipped with CAP_NET_BIND_SERVICE which allows the service to run without root. The SocketBindAllow and SocketBindDeny allows further restriction to specific ports rather than any port below 1024.

[francislavoie]

Are you looking for the bind directive? https://caddyserver.com/docs/caddyfile/directives/bind

And see https://caddyserver.com/docs/conventions#network-addresses, you can use unix sockets in reverse_proxy upstreams.

I'm not sure what you're asking for if not that.

[mholt]

It sounds like what is being asked for is graceful upgrades/restarts.

Caddy 1 had this feature, and I quite liked how it worked: pass the socket directly to the next process. It worked on all Unix systems without relying on a separate system service, and it was smart enough to understand Caddy configuration: if the new config didn't use a socket, it wouldn't be kept; rather than blindly moving all the sockets over.

I'd probably rather bring the implementation from Caddy 1 into Caddy 2.

[flokli]

It sounds like what is being asked for is graceful upgrades/restarts.

No, getting this for free is only one side-effect of supporting socket-activation.

socket-activation will also cause caddy to get started lazily whenever the first connection to the (externally configured) socket address happens, which simplifies declaring service dependencies too.

The article linked from my link elaborates a bit more on this.

Caddy 1 had this feature, and I quite liked how it worked: pass the socket directly to the next process.

This still requires caddy to do manual coordination with its new process and pass it around explicitly. The point of simply taking the FDs passed by the service manager is that caddy does not have to be aware of whether it's the first process being started on the system, or you start a new version with another config. caddy simply gets an FD, where new connections will appear on.

[flokli]

Ah yes, and because caddy just takes FDs, it doesn't need to bind() on its own, which allows applying stronger sandboxing from the outside.

[mholt]

@flokli

This still requires caddy to do manual coordination with its new process and pass it around explicitly. The point of simply taking the FDs passed by the service manager is that caddy does not have to be aware of whether it's the first process being started on the system, or you start a new version with another config. caddy simply gets an FD, where new connections will appear on.

But what is Caddy supposed to do with that socket? How does it know the configuration associated with it? You can't just hand a server a socket and expect it to know what to do with it, without any configuration... maybe I am missing something about how it works.

[mholt]

Oh I see, so you'd still have your Caddy config, you'd just specify a different network name for the listener address, and Caddy will then get it from the service manager rather than binding a new socket.

[flokli]

Yes! Or well, I don't want caddy to do any bind on its own at all, but pass in every socket via this mechanism.

[mholt]

In that case you can use bind in your site blocks to get the socket from the service manager. We'd just need to implement a package that calls caddy.RegisterNetwork(). For example the caddy-tailscale package does this so that Tailscale can provide a listener.

Anyone is welcome to pick this up.

[WeidiDeng]

@mholt I did some experiments with registering custom network, it's too much trouble to be worth it. Every site block needs an explicit bind and that includes http port and http3 udp socket.

@flokli I'm thinking on unix, we can try preferring socket activation but fallback to the old behavior. What do you think of it? Or should caddy just exit unsuccessfully if socket activation environments variables are found but not sockets matching listening critertia are found? Or if some warning logs are emitted?

As mentioned above, you are responsible to pass every socket yourself, including 80 tcp and 443 udp if auto http->https and http3 are enabled respectively. And admin socket if enabled as well. Assuming you restart caddy instead of reload it.

[climba03003]

I would really see it happen and it can greatly reduce my network stack complexity.
Currently, I am have two caddy in front of server and I face a lot of instability because of podman networking.
I change to using socket to see if it works better (no more DNS resolution).

flowchart TD
    A[Caddy] -->|Reverse Proxy| B{Container Network}
    B -->|Serve Frontend| C[Caddy]
    C -->|Reverse Proxy| D[Server]

When socket activation becomes a thing, it can also reduce resources usage. Because the middle caddy can be terminated when no one connected for some time. If the outer one can be socket activated, it will directly pass the socket to inner one and benefit of direct network connection.

[flokli]

@mholt I did some experiments with registering custom network, it's too much trouble to be worth it. Every site block needs an explicit bind and that includes http port and http3 udp socket.

@flokli I'm thinking on unix, we can try preferring socket activation but fallback to the old behavior. What do you think of it? Or should caddy just exit unsuccessfully if socket activation environments variables are found but not sockets matching listening critertia are found? Or if some warning logs are emitted?

I think ti makes sense to first land the feature with explicit configuration, which might mean explicit bind statements, and once that's in, think about having more opinionated defaults in case we are in a socket-activated environment.

The good thing is, it's pretty safe to detect whether caddy is running in a socket-activated environment or not, so we are able to change defaults in this case, without breaking existing usecases.

[eliasp]

Until this is implemented: for those that just care about binding to ports <1024 AND not running Caddy as root, can use systemd's SocketBindAllow= (available since systemd 249)