Merge pull request #145 from razvandimescu/docs/recipes

docs: lift user-facing guides to recipes/, drop dangling docs/ refs
This commit was merged in pull request #145.
This commit is contained in:
Razvan Dimescu
2026-04-24 15:22:44 +03:00
committed by GitHub
11 changed files with 207 additions and 16 deletions

View File

@@ -51,8 +51,8 @@ api_port = 5380
# relay_ip = "178.104.229.30" # optional: pin IPs so numa doesn't leak the # relay_ip = "178.104.229.30" # optional: pin IPs so numa doesn't leak the
# target_ip = "104.16.249.249" # relay/target hostnames via the bootstrap # target_ip = "104.16.249.249" # relay/target hostnames via the bootstrap
# # resolver on cold boot when numa is its # # resolver on cold boot when numa is its
# # own system DNS. See docs/implementation/ # # own system DNS. See
# # bootstrap-resolver.md. # # recipes/odoh-upstream.md.
# root_hints = [ # only used in recursive mode # root_hints = [ # only used in recursive mode
# "198.41.0.4", # a.root-servers.net (Verisign) # "198.41.0.4", # a.root-servers.net (Verisign)
# "199.9.14.201", # b.root-servers.net (USC-ISI) # "199.9.14.201", # b.root-servers.net (USC-ISI)

View File

@@ -2,10 +2,10 @@
Single-container deploy that runs Numa as an ODoH (RFC 9230) client: every Single-container deploy that runs Numa as an ODoH (RFC 9230) client: every
DNS query routes through an independent relay + target so neither operator DNS query routes through an independent relay + target so neither operator
sees both your IP and your question. See the [ODoH integration doc][odoh] sees both your IP and your question. See the [ODoH upstream recipe][odoh]
for the full protocol and privacy trade-offs. for the protocol details and the bootstrap-pinning trade-offs.
[odoh]: ../../docs/implementation/odoh-integration.md [odoh]: ../../recipes/odoh-upstream.md
## Prerequisites ## Prerequisites

View File

@@ -1,7 +1,7 @@
# Numa — ODoH client mode (docker-compose starter). # Numa — ODoH client mode (docker-compose starter).
# Sends every DNS query through an independent relay + target pair so # Sends every DNS query through an independent relay + target pair so
# neither operator sees both your IP and your question. See # neither operator sees both your IP and your question. See
# docs/implementation/odoh-integration.md for the protocol details and # recipes/odoh-upstream.md for the protocol details and
# packaging/client/README.md for deploy notes. # packaging/client/README.md for deploy notes.
[server] [server]

11
recipes/README.md Normal file
View File

@@ -0,0 +1,11 @@
# Recipes
Scenario-driven configs for common Numa deployments. Each recipe is self-contained: copy the snippet, adjust the marked fields, reload.
## Transport / encryption
- [DoH on the LAN](doh-on-lan.md) — expose Numa's built-in DNS-over-HTTPS to local clients.
- [dnsdist in front of Numa](dnsdist-front.md) — terminate public TLS externally, keep Numa on loopback.
- [ODoH upstream with bootstrap pinning](odoh-upstream.md) — oblivious DNS client mode without leaking the relay/target hostnames.
Missing a scenario? Open an issue or PR — these are plain Markdown with no build step.

64
recipes/dnsdist-front.md Normal file
View File

@@ -0,0 +1,64 @@
# dnsdist in front of Numa
For public DoH with a real (ACME-signed) cert, terminate TLS outside Numa and forward plain DNS (or loopback-only DoH) to the resolver. Cert renewal, rate-limiting, and load-balancing live in the front-end; Numa stays focused on resolution.
## When to use this
- Public hostname (`dns.example.com`) with a Let's Encrypt or internal PKI cert.
- You want a dedicated front-end for DoH/DoT/DoQ while Numa stays loopback-bound.
- You plan to run multiple Numa instances behind one endpoint.
## Architecture
```
public 443/DoH ┐
public 853/DoT ├─► dnsdist ─► 127.0.0.1:53 (Numa UDP/TCP)
public 443/DoQ ┘
```
## dnsdist config
```lua
-- /etc/dnsdist/dnsdist.conf
newServer({address="127.0.0.1:53", name="numa", checkType="A", checkName="numa.rs."})
addDOHLocal(
"0.0.0.0:443",
"/etc/letsencrypt/live/dns.example.com/fullchain.pem",
"/etc/letsencrypt/live/dns.example.com/privkey.pem",
"/dns-query",
{doTCP=true, reusePort=true}
)
addTLSLocal(
"0.0.0.0:853",
"/etc/letsencrypt/live/dns.example.com/fullchain.pem",
"/etc/letsencrypt/live/dns.example.com/privkey.pem"
)
addAction(AllRule(), PoolAction("", false))
```
## Numa config
```toml
[proxy]
enabled = true # keep if you still use *.numa service routing
bind_addr = "127.0.0.1" # stays default
```
No changes to `[server]` — Numa keeps serving plain DNS on UDP/TCP 53, which dnsdist forwards.
## Caveat: client IPs
Without PROXY protocol support in Numa, the query log shows the front-end's IP on every query, not the real client. dnsdist can emit PROXY v2 (`useProxyProtocol=true` on `newServer`), but Numa doesn't yet parse it — tracked in the wish-list under #143. Until then, accept the blind spot or correlate against dnsdist's own logs.
## Verify
```bash
kdig +https @dns.example.com example.com
kdig +tls @dns.example.com example.com
```
Both should return clean answers. Numa's `/queries` API should show the request landing, sourced from the front-end IP.

61
recipes/doh-on-lan.md Normal file
View File

@@ -0,0 +1,61 @@
# DoH on the LAN
Numa ships an RFC 8484 DoH endpoint (`POST /dns-query`) on the `[proxy]` HTTPS listener. By default it binds `127.0.0.1:443` with a self-signed cert — invisible to anything off the box. Three changes make it reachable from the LAN.
## When to use this
- Your phone/laptop is on the same network as Numa and you want encrypted DNS without a cloud resolver.
- You're OK installing Numa's self-signed CA on every client (one-time, via `/ca.pem` + the mobileconfig flow).
For a publicly-trusted cert, see [dnsdist in front of Numa](dnsdist-front.md) instead.
## Minimal config
```toml
[proxy]
enabled = true # default
bind_addr = "0.0.0.0" # was 127.0.0.1 — expose to LAN
tls_port = 443 # default; DoH is served here
tld = "numa" # default — self-resolving, see below
```
`tld` is the DoH gate: Numa accepts the DoH request only when the `Host` header is loopback or equals (or is a subdomain of) `tld`. Clients therefore dial `https://numa/dns-query`.
With the default `tld = "numa"`, there's no DNS bootstrap to configure: Numa already resolves `numa` and `*.numa` to its own LAN IP for remote clients (that's how the `*.numa` service-proxy feature works). Any client that uses Numa as its resolver will resolve `numa` correctly on first try.
If you'd rather use a hostname that resolves via normal DNS (e.g. you want DoH-only clients that never talk plain DNS to Numa), set `tld = "dns.example.com"` and add a matching A record in whichever DNS your clients consult before reaching Numa.
## Trust the CA on each client
Numa generates a self-signed CA at startup. Fetch it once, import it wherever you'll run the DoH client:
```bash
curl -o numa-ca.pem http://<numa-ip>:5380/ca.pem
```
- **macOS** — `sudo security add-trusted-cert -d -r trustRoot -k /Library/Keychains/System.keychain numa-ca.pem`
- **iOS** — install the mobileconfig from the API (same CA, signed profile). Flip *Settings → General → About → Certificate Trust Settings* on after install.
- **Linux** — drop into `/usr/local/share/ca-certificates/` and run `sudo update-ca-certificates`.
- **Android** — requires the user-installed CA path; browsers may still refuse it for DoH. Consider the [dnsdist front](dnsdist-front.md) route instead.
## Verify
```bash
kdig +https @numa example.com
```
Without `+https` kdig uses plain DNS. With `+https` the same answers should flow over port 443.
Raw check:
```bash
curl -H 'accept: application/dns-message' \
--data-binary @query.bin \
https://numa/dns-query
```
## Gotchas
- Port 443 is privileged on Linux/macOS. Run Numa via the provided service units, or grant `CAP_NET_BIND_SERVICE` (`sudo setcap 'cap_net_bind_service=+ep' /path/to/numa`).
- Non-matching `Host` header → HTTP 404 from the proxy's fallback handler. Double-check `tld`.
- ChromeOS enrollment rejects user-installed CAs for some flows — known pain point, see issue #136.

59
recipes/odoh-upstream.md Normal file
View File

@@ -0,0 +1,59 @@
# ODoH upstream with bootstrap pinning
Numa can run as an Oblivious DoH (RFC 9230) client: the relay sees your IP but not the question, the target sees the question but not your IP. Neither party alone can re-identify a query. This recipe covers the minimal config and the bootstrap leak that `relay_ip` / `target_ip` close.
## When to use this
- You want split-trust encrypted DNS without a single provider seeing both who you are and what you asked.
- Numa is your system resolver (so there's no "other" DNS to ask).
## Minimal config
```toml
[upstream]
mode = "odoh"
relay = "https://odoh-relay.numa.rs/relay"
target = "https://odoh.cloudflare-dns.com/dns-query"
strict = true # refuse to fall back to a non-oblivious path on relay failure
```
`strict = true` means a relay-level HTTPS failure returns SERVFAIL instead of silently downgrading. Set it to `false` and configure `[upstream].fallback` if you'd rather keep resolving (at the cost of the oblivious property).
## The bootstrap leak
When Numa is the system resolver and needs to reach the relay/target, *something* has to translate `odoh-relay.numa.rs` → IP. If Numa asks itself, you deadlock. If Numa asks a bootstrap resolver (1.1.1.1, 9.9.9.9), that resolver learns which ODoH endpoint you use in cleartext — it can't see your questions, but it sees the destination. That's the leak ODoH was supposed to close.
`relay_ip` and `target_ip` tell Numa the IPs directly, so it never asks anyone:
```toml
[upstream]
mode = "odoh"
relay = "https://odoh-relay.numa.rs/relay"
target = "https://odoh.cloudflare-dns.com/dns-query"
relay_ip = "178.104.229.30" # pin the relay — no hostname lookup
target_ip = "104.16.249.249" # pin the target — no hostname lookup
```
Numa still validates TLS against the hostnames in `relay` / `target`, so a hijacked IP can't masquerade — pinning skips only the DNS step.
## Finding current IPs
```bash
dig +short odoh-relay.numa.rs
dig +short odoh.cloudflare-dns.com
```
Re-pin when an operator rotates. The community-maintained list at <https://github.com/DNSCrypt/dnscrypt-resolvers/blob/master/v3/odoh-relays.md> is a useful cross-reference.
## Verify
```bash
kdig @127.0.0.1 example.com
```
Numa's `/queries` API and startup banner should label the upstream as `odoh://`. Look for `ODoH relay returned ...` errors in the logs if routing fails.
## Known gotchas
- **Same-operator refused.** Numa's eTLD+1 check blocks configs where the relay and target belong to the same operator (pointless — same party sees both sides). Override only when testing.
- **Single relay.** Current config accepts one relay and one target. Multi-entry rotation/failover is tracked in #140.

View File

@@ -2,8 +2,7 @@
//! relay/target, blocklist CDN). When numa is its own system resolver //! relay/target, blocklist CDN). When numa is its own system resolver
//! (`/etc/resolv.conf → 127.0.0.1`, HAOS add-on, Pi-hole-style container), //! (`/etc/resolv.conf → 127.0.0.1`, HAOS add-on, Pi-hole-style container),
//! the default `getaddrinfo` path loops back through numa before numa can //! the default `getaddrinfo` path loops back through numa before numa can
//! answer — a chicken-and-egg that deadlocks cold boot. See issue #122 and //! answer — a chicken-and-egg that deadlocks cold boot. See issue #122.
//! `docs/implementation/bootstrap-resolver.md`.
//! //!
//! Resolution order per hostname: //! Resolution order per hostname:
//! 1. Per-hostname overrides (e.g. ODoH `relay_ip` / `target_ip`) → return //! 1. Per-hostname overrides (e.g. ODoH `relay_ip` / `target_ip`) → return

View File

@@ -175,8 +175,7 @@ pub fn parse_upstream(
/// ///
/// Uses the system resolver. Callers running inside `serve::run` pass the /// Uses the system resolver. Callers running inside `serve::run` pass the
/// shared [`crate::bootstrap_resolver::NumaResolver`] via /// shared [`crate::bootstrap_resolver::NumaResolver`] via
/// [`build_https_client_with_resolver`] to avoid the self-loop documented /// [`build_https_client_with_resolver`] to avoid the self-loop (issue #122).
/// in `docs/implementation/bootstrap-resolver.md`.
pub fn build_https_client() -> reqwest::Client { pub fn build_https_client() -> reqwest::Client {
build_https_client_with_resolver(1, None) build_https_client_with_resolver(1, None)
} }

View File

@@ -7,11 +7,10 @@
//! Both handlers call [`HealthResponse::build`] to assemble the JSON //! Both handlers call [`HealthResponse::build`] to assemble the JSON
//! response from `HealthMeta` + live inputs. //! response from `HealthMeta` + live inputs.
//! //!
//! JSON schema is documented in `docs/implementation/ios-companion-app.md` //! The iOS companion app's `HealthInfo` struct is the canonical consumer;
//! §4.2. The iOS companion app's `HealthInfo` struct is the canonical //! any change to this response must keep that struct decoding cleanly (all
//! consumer; any change to this response must keep that struct decoding //! consumed fields are optional on the Swift side, but `lan_ip` is
//! cleanly (all consumed fields are optional on the Swift side, but //! load-bearing for the pipeline).
//! `lan_ip` is load-bearing for the pipeline).
use std::net::Ipv4Addr; use std::net::Ipv4Addr;
use std::path::Path; use std::path::Path;

View File

@@ -52,7 +52,6 @@ pub async fn run(config_path: String) -> crate::Result<()> {
// Routes numa-originated HTTPS (DoH upstream, ODoH relay/target, blocklist // Routes numa-originated HTTPS (DoH upstream, ODoH relay/target, blocklist
// CDN) away from the system resolver so lookups don't loop back through // CDN) away from the system resolver so lookups don't loop back through
// numa when it's its own system DNS. // numa when it's its own system DNS.
// See `docs/implementation/bootstrap-resolver.md`.
let resolver_overrides = match config.upstream.mode { let resolver_overrides = match config.upstream.mode {
crate::config::UpstreamMode::Odoh => config crate::config::UpstreamMode::Odoh => config
.upstream .upstream