feat(dns): implement dns service, the half a reboot takes away

docs/hosting-a-moshpit-name.md

@@ -98,21 +98,61 @@ anyone scanning the box the app with the name stripped off the front.
 On every machine that should see the name:
 
 ```sh
-sudo moshcode dns enable
+moshcode dns enable
 ```
 
+Do not prefix it with `sudo`. The command escalates the one step that needs
+root and re-runs only that; `sudo moshcode` takes every path from the wrong
+`$HOME`.
+
 That does two things: writes a `systemd-resolved` drop-in routing Moshpit
 endings at the bridge, and starts the bridge. The drop-in is a file and survives
 a reboot. **The bridge process does not** — so after a restart the routing
 points at a port with nothing behind it, and every Moshpit name stops resolving
-with no obvious cause. The bundled `deploy/moshcode-dns.service` is the missing
-half:
+with no obvious cause. This is the whole reason Moshpit DNS holds on servers,
+which got a unit installed by hand, and quietly falls over on desktops, which
+never did.
+
+`dns service` is the missing half:
+
+```sh
+moshcode dns service            # print the unit, change nothing
+moshcode dns service --write    # install and start it, no root needed
+```
+
+That installs a **user** service, which is the right scope for a per-user
+install: it runs as you, so `$HOME` and the mise/nvm shims are reachable, and
+the pidfile lands where an unprivileged `dns status` looks for it. A user
+service stops at logout unless you allow it to linger:
+
+```sh
+loginctl enable-linger "$USER"
+```
+
+For a box with moshcode installed system-wide, a system unit is available and
+does need root to place:
 
 ```sh
-sudo cp deploy/moshcode-dns.service /etc/systemd/system/
+moshcode dns service --system | sudo tee /etc/systemd/system/moshcode-dns.service
 sudo systemctl enable --now moshcode-dns
 ```
 
+Either way the unit is **generated from the running process**, so `ExecStart`
+names the interpreter that is demonstrably working and the install that is
+really there. Do not hand-write one or copy another machine's: moshcode lives
+under `$HOME` and its wrapper execs whatever `node` is first on `PATH`, and
+systemd has neither on its own `PATH`. A unit that guesses fails at `203/EXEC`
+with nothing useful in the journal. Regenerate it after moving or reinstalling
+either program.
+
+Removing it again:
+
+```sh
+moshcode dns service --remove
+```
+
+That takes away the service only. The routing is `dns disable`.
+
 ## Verifying, one layer at a time
 
 Every layer fails identically in a browser, so do not debug from one.

examples/templates/bun-caddy-sqlite/deploy/moshcode-dns.service

@@ -7,8 +7,22 @@
 # and every Moshpit name stops resolving with no obvious cause. This unit is
 # the missing half.
 #
-#   sudo cp deploy/moshcode-dns.service /etc/systemd/system/
-#   sudo systemctl enable --now moshcode-dns
+# GENERATE IT, DO NOT COPY THIS ONE:
+#
+#   moshcode dns service --write     # this user's own service, needs no root
+#   moshcode dns service --system | sudo tee /etc/systemd/system/moshcode-dns.service
+#
+# `dns service` writes ExecStart from the running process, so it names the node
+# that demonstrably works and the install that is actually on the box. A static
+# file cannot: moshcode installs under $HOME, and its wrapper execs whatever
+# `node` is first on PATH — which on any mise, nvm or asdf machine is another
+# shim under $HOME. systemd has neither on its PATH and, with ProtectHome or
+# DynamicUser set, could not read them if it did. A unit that gets this wrong
+# fails at 203/EXEC with nothing useful in the journal.
+#
+# What is left below therefore assumes the one case it can safely assume: both
+# moshcode and node installed system-wide, service running as root. Check
+# `systemctl status moshcode-dns` before believing it.
 #
 # Install this on machines that need to REACH Moshpit names. A box that only
 # serves one does not need it — Caddy answers whatever Host header arrives and
@@ -22,18 +36,15 @@ Before=systemd-resolved.service
 
 [Service]
 Type=simple
-# Port 5354 is unprivileged, so this does not need root. The trade-off is that
-# the parking responder cannot take port 80 and falls back to the public
-# parking address — which only affects names that point nowhere yet.
-ExecStart=/usr/bin/env moshcode dns start --port 5354
+# Port 5354 is unprivileged, so this does not need root for the DNS side. The
+# trade-off is that the parking responder cannot take port 80 and falls back to
+# the public parking address — which only affects names that point nowhere yet.
+ExecStart=/usr/local/bin/moshcode dns start --port 5354
 Restart=always
 RestartSec=2
 
-DynamicUser=yes
 NoNewPrivileges=yes
 PrivateTmp=yes
-ProtectSystem=strict
-ProtectHome=yes
 
 [Install]
 WantedBy=multi-user.target

examples/templates/caddy-proxy/deploy/moshcode-dns.service

@@ -7,8 +7,22 @@
 # and every Moshpit name stops resolving with no obvious cause. This unit is
 # the missing half.
 #
-#   sudo cp deploy/moshcode-dns.service /etc/systemd/system/
-#   sudo systemctl enable --now moshcode-dns
+# GENERATE IT, DO NOT COPY THIS ONE:
+#
+#   moshcode dns service --write     # this user's own service, needs no root
+#   moshcode dns service --system | sudo tee /etc/systemd/system/moshcode-dns.service
+#
+# `dns service` writes ExecStart from the running process, so it names the node
+# that demonstrably works and the install that is actually on the box. A static
+# file cannot: moshcode installs under $HOME, and its wrapper execs whatever
+# `node` is first on PATH — which on any mise, nvm or asdf machine is another
+# shim under $HOME. systemd has neither on its PATH and, with ProtectHome or
+# DynamicUser set, could not read them if it did. A unit that gets this wrong
+# fails at 203/EXEC with nothing useful in the journal.
+#
+# What is left below therefore assumes the one case it can safely assume: both
+# moshcode and node installed system-wide, service running as root. Check
+# `systemctl status moshcode-dns` before believing it.
 #
 # Install this on machines that need to REACH Moshpit names. A box that only
 # serves one does not need it — Caddy answers whatever Host header arrives and
@@ -22,18 +36,15 @@ Before=systemd-resolved.service
 
 [Service]
 Type=simple
-# Port 5354 is unprivileged, so this does not need root. The trade-off is that
-# the parking responder cannot take port 80 and falls back to the public
-# parking address — which only affects names that point nowhere yet.
-ExecStart=/usr/bin/env moshcode dns start --port 5354
+# Port 5354 is unprivileged, so this does not need root for the DNS side. The
+# trade-off is that the parking responder cannot take port 80 and falls back to
+# the public parking address — which only affects names that point nowhere yet.
+ExecStart=/usr/local/bin/moshcode dns start --port 5354
 Restart=always
 RestartSec=2
 
-DynamicUser=yes
 NoNewPrivileges=yes
 PrivateTmp=yes
-ProtectSystem=strict
-ProtectHome=yes
 
 [Install]
 WantedBy=multi-user.target

examples/templates/caddy-static/deploy/moshcode-dns.service

@@ -7,8 +7,22 @@
 # and every Moshpit name stops resolving with no obvious cause. This unit is
 # the missing half.
 #
-#   sudo cp deploy/moshcode-dns.service /etc/systemd/system/
-#   sudo systemctl enable --now moshcode-dns
+# GENERATE IT, DO NOT COPY THIS ONE:
+#
+#   moshcode dns service --write     # this user's own service, needs no root
+#   moshcode dns service --system | sudo tee /etc/systemd/system/moshcode-dns.service
+#
+# `dns service` writes ExecStart from the running process, so it names the node
+# that demonstrably works and the install that is actually on the box. A static
+# file cannot: moshcode installs under $HOME, and its wrapper execs whatever
+# `node` is first on PATH — which on any mise, nvm or asdf machine is another
+# shim under $HOME. systemd has neither on its PATH and, with ProtectHome or
+# DynamicUser set, could not read them if it did. A unit that gets this wrong
+# fails at 203/EXEC with nothing useful in the journal.
+#
+# What is left below therefore assumes the one case it can safely assume: both
+# moshcode and node installed system-wide, service running as root. Check
+# `systemctl status moshcode-dns` before believing it.
 #
 # Install this on machines that need to REACH Moshpit names. A box that only
 # serves one does not need it — Caddy answers whatever Host header arrives and
@@ -22,18 +36,15 @@ Before=systemd-resolved.service
 
 [Service]
 Type=simple
-# Port 5354 is unprivileged, so this does not need root. The trade-off is that
-# the parking responder cannot take port 80 and falls back to the public
-# parking address — which only affects names that point nowhere yet.
-ExecStart=/usr/bin/env moshcode dns start --port 5354
+# Port 5354 is unprivileged, so this does not need root for the DNS side. The
+# trade-off is that the parking responder cannot take port 80 and falls back to
+# the public parking address — which only affects names that point nowhere yet.
+ExecStart=/usr/local/bin/moshcode dns start --port 5354
 Restart=always
 RestartSec=2
 
-DynamicUser=yes
 NoNewPrivileges=yes
 PrivateTmp=yes
-ProtectSystem=strict
-ProtectHome=yes
 
 [Install]
 WantedBy=multi-user.target

src/cli-schema.mjs

@@ -390,7 +390,7 @@ export const CORE_CLI_COMMANDS = [
       ["--no-proxy", "with enable: answer origins rather than the local proxy", ""],
     ],
     examples: [
-      ["sudo moshcode dns enable", "route Moshpit endings here"],
+      ["moshcode dns enable", "route Moshpit endings here (it escalates itself)"],
       ["moshcode dns resolve blue.eggs", "what a machine actually gets"],
     ],
     seeAlso: ["doh", "site"],
@@ -1009,7 +1009,16 @@ export const DNS_VERBS = [
   { name: "refresh", description: "re-apply routing for endings claimed since" },
   { name: "start", description: "run the bridge in the foreground" },
   { name: "install", description: "print the resolver config without applying it" },
-  { name: "service", description: "install or remove the background service" },
+  {
+    name: "service",
+    description: "keep the bridge running across reboots — the half `enable` does not survive",
+    synopsis: [
+      ["moshcode dns service", "print the unit for this install, change nothing"],
+      ["moshcode dns service --write", "install and start it as this user; needs no root"],
+      ["moshcode dns service --system", "a system unit instead — place it with sudo tee"],
+      ["moshcode dns service --remove", "stop it and take the unit away"],
+    ],
+  },
   { name: "tlds", description: "list the endings claimed in the Pit" },
   { name: "resolve", description: "what a name resolves to, and why" },
   { name: "trust", description: "trust one name's certificate, after checking it against the registry pin" },

src/dns-service.mjs

@@ -0,0 +1,157 @@
+// Keeping the bridge running across reboots.
+//
+// `dns enable` sets up two halves with different lifetimes: a systemd-resolved
+// drop-in, which is a file and survives a reboot on its own, and the bridge
+// process, which does not. After a restart the routing still points at a port
+// with nothing behind it and every Moshpit name stops resolving with no
+// obvious cause. This is the missing half, and it is why Moshpit DNS has held
+// on servers — which got a unit installed by hand — and quietly fallen over on
+// desktops, which never did.
+//
+// The unit is *generated* rather than shipped as a file, because no static
+// unit can be correct for this tool. moshcode installs under the invoking
+// user's $HOME, and its wrapper execs whatever `node` is first on PATH — which
+// on any mise, nvm or asdf box is another shim under $HOME. The unit that
+// shipped in examples/templates said:
+//
+//     ExecStart=/usr/bin/env moshcode dns start --port 5354
+//     DynamicUser=yes
+//     ProtectHome=yes
+//
+// and could not start on an ordinary install three times over: `moshcode` is
+// not on systemd's PATH, the `node` its wrapper needs is not on it either, and
+// ProtectHome hides the install from the service even if both had been found.
+// It starts only where moshcode and node are both installed system-wide, which
+// is the server case — the one that was already working.
+//
+// So the unit is written from the running process instead. `process.execPath`
+// is an interpreter demonstrably able to run this code, because it is running
+// it, and the entry is the script this very command was invoked from. Nothing
+// is guessed and nothing depends on PATH.
+import { spawn } from "node:child_process";
+import { mkdir, rm, writeFile } from "node:fs/promises";
+import { homedir } from "node:os";
+import { dirname, join } from "node:path";
+
+export const UNIT_NAME = "moshcode-dns.service";
+
+/**
+ * Where the unit goes, and which systemctl reaches it.
+ *
+ * The user scope is the default because it is the one that fits how moshcode
+ * is installed: a per-user tool, run by the user who owns the install, with
+ * $HOME and the mise shims reachable exactly as they are in a shell. It also
+ * puts the pidfile where the unprivileged `dns status` looks — systemd sets
+ * XDG_RUNTIME_DIR for user units, so `pidfilePath()` resolves to
+ * /run/user/<uid>/moshpit-dns.pid for the daemon and for the person asking
+ * after it. Under a system unit those are two different paths.
+ */
+export function servicePaths({ system = false, home = homedir() } = {}) {
+  return system
+    ? { path: join("/etc/systemd/system", UNIT_NAME), systemctl: ["systemctl"], scope: "system" }
+    : { path: join(home, ".config/systemd/user", UNIT_NAME), systemctl: ["systemctl", "--user"], scope: "user" };
+}
+
+/**
+ * The unit text, pinned to this install.
+ *
+ * `ProtectHome` is deliberately absent rather than set to a weaker value: the
+ * whole program lives under $HOME, so there is no setting of it that both
+ * protects anything and lets the service start. The hardening that survives is
+ * the hardening that does not contradict where the code is.
+ */
+export function serviceUnit({
+  system = false,
+  execPath = process.execPath,
+  entry,
+  port,
+  registryBase = null,
+  user = process.env.USER || process.env.LOGNAME,
+} = {}) {
+  if (!entry) throw new Error("serviceUnit needs the entry script to run");
+
+  const args = [entry, "dns", "start", "--port", String(port)];
+  if (registryBase) args.push("--registry", registryBase);
+  const exec = [execPath, ...args].map((part) => (/\s/.test(part) ? JSON.stringify(part) : part)).join(" ");
+
+  const lines = [
+    "# Generated by `moshcode dns service`. Regenerate it rather than editing:",
+    "# the paths below are this install's, and a moshcode or node that moves",
+    "# leaves a unit that fails at 203/EXEC with nothing else to say.",
+    "[Unit]",
+    "Description=Moshpit DNS bridge",
+    "Documentation=https://github.com/moshcoder/moshcode",
+    "After=network-online.target",
+    "Wants=network-online.target",
+    "",
+    "[Service]",
+    "Type=simple",
+    `ExecStart=${exec}`,
+    "Restart=always",
+    "RestartSec=2",
+    // 5354 is unprivileged, so nothing here needs root or a capability.
+    "NoNewPrivileges=yes",
+    "PrivateTmp=yes",
+  ];
+
+  if (system) {
+    // A system unit has no user of its own, so it has to be told which install
+    // to run — and it is the only scope where that question has a wrong answer.
+    lines.push(`User=${user}`);
+  }
+
+  lines.push(
+    "",
+    "[Install]",
+    system ? "WantedBy=multi-user.target" : "WantedBy=default.target",
+    "",
+  );
+  return lines.join("\n");
+}
+
+function run(command, args) {
+  return new Promise((resolve) => {
+    const child = spawn(command, args, { stdio: ["ignore", "pipe", "pipe"] });
+    let err = "";
+    child.stderr.on("data", (d) => (err += d));
+    child.on("error", (error) => resolve({ ok: false, error: error.message }));
+    child.on("exit", (code) => resolve({ ok: code === 0, error: err.trim() }));
+  });
+}
+
+/** Write the unit and start it. Returns the steps taken, in order, for printing. */
+export async function installService(unit, { system = false, home = homedir(), exec = run } = {}) {
+  const { path, systemctl, scope } = servicePaths({ system, home });
+  const steps = [];
+  try {
+    await mkdir(dirname(path), { recursive: true });
+    await writeFile(path, unit);
+    steps.push({ step: `wrote ${path}`, ok: true });
+  } catch (error) {
+    return { ok: false, path, scope, steps: [{ step: `could not write ${path}: ${error.message}`, ok: false }] };
+  }
+
+  const [cmd, ...flags] = systemctl;
+  for (const args of [[...flags, "daemon-reload"], [...flags, "enable", "--now", UNIT_NAME]]) {
+    const result = await exec(cmd, args);
+    steps.push({ step: `${cmd} ${args.join(" ")}`, ok: result.ok, error: result.error });
+    if (!result.ok) return { ok: false, path, scope, steps };
+  }
+  return { ok: true, path, scope, steps };
+}
+
+/** Stop it and take the unit away. Missing is not a failure — removal is idempotent. */
+export async function removeService({ system = false, home = homedir(), exec = run } = {}) {
+  const { path, systemctl, scope } = servicePaths({ system, home });
+  const [cmd, ...flags] = systemctl;
+  const steps = [];
+  for (const args of [[...flags, "disable", "--now", UNIT_NAME]]) {
+    const result = await exec(cmd, args);
+    steps.push({ step: `${cmd} ${args.join(" ")}`, ok: result.ok, error: result.error });
+  }
+  await rm(path, { force: true });
+  steps.push({ step: `removed ${path}`, ok: true });
+  const reload = await exec(cmd, [...flags, "daemon-reload"]);
+  steps.push({ step: `${cmd} ${[...flags, "daemon-reload"].join(" ")}`, ok: reload.ok, error: reload.error });
+  return { ok: true, path, scope, steps };
+}