mDNS receiver actor spams "dropping message due to closed mailbox" after DropGuard is discarded

[jgeluk]

Summary

After an iroh::Endpoint has been running for a while, the swarm-discovery receiver actor enters a tight loop logging DEBUG "dropping message due to closed mailbox" every ~12 ms for every incoming mDNS packet. The messages come from acto::tokio::TokioSender::send and never stop until the process exits.

Root cause

Two issues combine:

1. iroh discards the DropGuard (iroh side)

In MdnsAddressLookup::spawn_discoverer() (src/address_lookup/mdns.rs), the DropGuard returned by Discoverer::spawn() is not stored — only the iroh-side actor AbortOnDropHandle is kept:

// spawn_discoverer() — simplified
let guard = discoverer.spawn(rt)?;   // DropGuard returned here
// … guard is dropped at the end of this scope

When the DropGuard drops, it aborts the guardian actor, which eventually tears down the sender actor. But the receiver actors are left running because their UDP sockets are still alive.

2. Receiver ignores failed sends (swarm-discovery side)

In swarm-discovery/src/receiver.rs:

loop {
    let (len, addr) = socket.recv_from(&mut buf).await?;
    if let Some(msg) = handle_msg(msg, &service_name, addr.ip()) {
        target.send(msg);  // returns false when target is dead — ignored
    }
}

The receiver never checks the return value of target.send(). Once the sender actor is dead, every mDNS packet produces a failed send + a DEBUG log line from acto, indefinitely.

Reproduction

On any LAN with mDNS traffic (typical macOS/Linux desktop), create an endpoint with mDNS discovery enabled and observe logs at DEBUG level. The spam starts within seconds after the DropGuard is dropped.

Observed log output

22:52:56.955    DEBUG    dropping message due to closed mailbox
22:52:56.967    DEBUG    dropping message due to closed mailbox
22:52:56.978    DEBUG    dropping message due to closed mailbox
… (repeats every ~12 ms indefinitely)

Suggested fix

iroh: Store the DropGuard as a field in MdnsAddressLookup so the swarm-discovery runtime stays alive as long as the endpoint does.

swarm-discovery (secondary): The receiver should check target.send() and exit the loop when it returns false, so a dead target doesn't cause unbounded log spam.

Versions

• iroh 0.97.0
• swarm-discovery 0.5.0
• acto 0.8.0

[matheus23]

1. iroh discards the DropGuard (iroh side)

In MdnsAddressLookup::spawn_discoverer() (src/address_lookup/mdns.rs), the DropGuard returned by Discoverer::spawn() is not stored — only the iroh-side actor AbortOnDropHandle is kept:

// spawn_discoverer() — simplified
let guard = discoverer.spawn(rt)?; // DropGuard returned here
// … guard is dropped at the end of this scope

When the DropGuard drops, it aborts the guardian actor, which eventually tears down the sender actor. But the receiver actors are left running because their UDP sockets are still alive.

That's incorrect. spawn_discoverer returns Result<DropGuard, _>, which is moved into the iroh-side actor, and should be kept for as long as MdnsAddressLookup is running.
We don't drop the guard at the end of the scope.

iroh: Store the DropGuard as a field in MdnsAddressLookup so the swarm-discovery runtime stays alive as long as the endpoint does.

This does not work, because we need access to the DropGuard inside the iroh-side actor: We call DropGuard::remove_all, add, set_txt_attribute, etc in that actor and DropGuard is not clonable.

---

On any LAN with mDNS traffic (typical macOS/Linux desktop), create an endpoint with mDNS discovery enabled and observe logs at DEBUG level. The spam starts within seconds after the DropGuard is dropped.

I can reproduce this. For anyone else, run RUST_LOG=acto=debug cargo run -p iroh --example mdns_address_lookup --features=mdns-address-lookup from the iroh repository.

That's because the MdnsAddressLookup is dropped, which drops the AbortOnDropHandle, which closes the mailbox that the acto actor on the swarm-discovery side uses to inform the iroh side about things.

• We should consider making the swam-discovery tasks abort on drop at the same time that MdnsAddressLookup is dropped.
• swarm-discovery should probably close its actor when it realizes it can't send messages out anymore (although there might be a technical reason why it doesn't do that yet and instead chooses to ignore the messages).

---

Aside: A bunch of "information" in your issue is likely AI-generated slop. Please don't create issues like this. It hides the legitimate bug report behind plausible-sounding misinformation.
I can see this being well-intentioned, trying to provide a bug report that's as good as possible, but unfortunately ends up costing me more time than a bug report that only contains the most relevant information possible with high signal-to-noise ratio.

[matheus23]

It's also worth clarifying that this behavior can only be seen when either the MdnsAddressLookup is dropped (e.g. when the iroh::Endpoint is closed), or when the iroh-side mdns address lookup actor panics.
It's unclear if the latter happened, although I suspect that's not what happens. In case it is, that'd be a valueable bug-report.