Merge pull request #126 from razvandimescu/fix/self-resolver-loop

fix(bootstrap): route numa HTTPS via IP-literal bootstrap resolver (#122)

benches/recursive_compare.rs

@@ -383,7 +383,7 @@ fn run_default(rt: &tokio::runtime::Runtime) {
 
 /// Library-to-library: Numa forward_query_raw vs Hickory resolver.lookup.
 fn run_direct(rt: &tokio::runtime::Runtime) {
-    let upstream = numa::forward::parse_upstream(DOH_UPSTREAM, 443).expect("failed to parse");
+    let upstream = numa::forward::parse_upstream(DOH_UPSTREAM, 443, None).expect("failed to parse");
     let resolver = rt.block_on(build_hickory_resolver());
     let timeout = Duration::from_secs(10);
 
@@ -609,9 +609,11 @@ fn run_hedge_multi(rt: &tokio::runtime::Runtime, iterations: usize) {
         DOMAINS.len()
     );
 
-    let primary = numa::forward::parse_upstream(DOH_UPSTREAM, 443).expect("failed to parse");
+    let primary = numa::forward::parse_upstream(DOH_UPSTREAM, 443, None).expect("failed to parse");
-    let primary_dual = numa::forward::parse_upstream(DOH_UPSTREAM, 443).expect("failed to parse");
+    let primary_dual =
-    let secondary_dual = numa::forward::parse_upstream(DOH_UPSTREAM, 443).expect("failed to parse");
+        numa::forward::parse_upstream(DOH_UPSTREAM, 443, None).expect("failed to parse");
+    let secondary_dual =
+        numa::forward::parse_upstream(DOH_UPSTREAM, 443, None).expect("failed to parse");
     let resolver = rt.block_on(build_hickory_resolver());
 
     println!("Warming up...");
@@ -810,7 +812,7 @@ fn run_diag(rt: &tokio::runtime::Runtime) {
 fn run_diag_clients(rt: &tokio::runtime::Runtime) {
     println!("Client diagnostic: reqwest vs Hickory (20 queries to {DOH_UPSTREAM})\n");
 
-    let upstream = numa::forward::parse_upstream(DOH_UPSTREAM, 443).expect("failed to parse");
+    let upstream = numa::forward::parse_upstream(DOH_UPSTREAM, 443, None).expect("failed to parse");
     let resolver = rt.block_on(build_hickory_resolver());
     let timeout = Duration::from_secs(10);
 

src/blocklist.rs

@@ -357,12 +357,17 @@ mod tests {
 
 const RETRY_DELAYS_SECS: &[u64] = &[2, 10, 30];
 
-pub async fn download_blocklists(lists: &[String]) -> Vec<(String, String)> {
+pub async fn download_blocklists(
-    let client = reqwest::Client::builder()
+    lists: &[String],
+    resolver: Option<std::sync::Arc<crate::bootstrap_resolver::NumaResolver>>,
+) -> Vec<(String, String)> {
+    let mut builder = reqwest::Client::builder()
         .timeout(Duration::from_secs(30))
-        .gzip(true)
+        .gzip(true);
-        .build()
+    if let Some(r) = resolver {
-        .unwrap_or_default();
+        builder = builder.dns_resolver(r);
+    }
+    let client = builder.build().unwrap_or_default();
 
     let fetches = lists.iter().map(|url| {
         let client = &client;

src/bootstrap_resolver.rs

@@ -0,0 +1,236 @@
+//! `reqwest` DNS resolver used by numa-originated HTTPS (DoH upstream, ODoH
+//! 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),
+//! the default `getaddrinfo` path loops back through numa before numa can
+//! answer — a chicken-and-egg that deadlocks cold boot. See issue #122 and
+//! `docs/implementation/bootstrap-resolver.md`.
+//!
+//! Resolution order per hostname:
+//! 1. Per-hostname overrides (e.g. ODoH `relay_ip` / `target_ip`) → return
+//!    immediately, no DNS query. Preserves ODoH's "zero plain-DNS leak"
+//!    property for configured endpoints.
+//! 2. Otherwise, query A + AAAA in parallel via UDP to IP-literal bootstrap
+//!    servers, with TCP fallback on UDP timeout (for networks that block
+//!    outbound UDP:53 — see memory: `project_network_udp_hostile.md`).
+
+use std::collections::HashMap;
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+use std::time::Duration;
+
+use log::{debug, info, warn};
+use reqwest::dns::{Addrs, Name, Resolve, Resolving};
+
+use crate::forward::{forward_tcp, forward_udp};
+use crate::packet::DnsPacket;
+use crate::question::QueryType;
+use crate::record::DnsRecord;
+
+const UDP_TIMEOUT: Duration = Duration::from_millis(800);
+const TCP_TIMEOUT: Duration = Duration::from_millis(1500);
+const DEFAULT_BOOTSTRAP: &[SocketAddr] = &[
+    SocketAddr::new(IpAddr::V4(Ipv4Addr::new(9, 9, 9, 9)), 53),
+    SocketAddr::new(IpAddr::V4(Ipv4Addr::new(1, 1, 1, 1)), 53),
+];
+
+pub struct NumaResolver {
+    bootstrap: Vec<SocketAddr>,
+    overrides: HashMap<String, Vec<IpAddr>>,
+}
+
+impl NumaResolver {
+    /// Build a resolver from the configured `upstream.fallback` list and any
+    /// per-hostname overrides (e.g. ODoH's `relay_ip`/`target_ip`).
+    ///
+    /// `fallback` entries are filtered to IP literals only — hostnames would
+    /// re-introduce the self-loop inside the resolver itself. Empty or
+    /// unusable fallback yields the hardcoded default (Quad9 + Cloudflare).
+    pub fn new(fallback: &[String], overrides: HashMap<String, Vec<IpAddr>>) -> Self {
+        let mut bootstrap: Vec<SocketAddr> = Vec::with_capacity(fallback.len());
+        for entry in fallback {
+            match crate::forward::parse_upstream_addr(entry, 53) {
+                Ok(addr) => bootstrap.push(addr),
+                Err(_) => {
+                    warn!(
+                        "bootstrap_resolver: skipping non-IP fallback '{}' \
+                         (hostnames would re-enter the self-loop)",
+                        entry
+                    );
+                }
+            }
+        }
+        let source = if bootstrap.is_empty() {
+            bootstrap = DEFAULT_BOOTSTRAP.to_vec();
+            "default (no IP-literal in upstream.fallback)"
+        } else {
+            "upstream.fallback"
+        };
+        let ips: Vec<String> = bootstrap.iter().map(|s| s.ip().to_string()).collect();
+        info!(
+            "bootstrap resolver: {} via {} — used for numa-originated HTTPS hostname resolution",
+            ips.join(", "),
+            source
+        );
+        if !overrides.is_empty() {
+            let mut pairs: Vec<String> = overrides
+                .iter()
+                .flat_map(|(host, ips)| ips.iter().map(move |ip| format!("{}={}", host, ip)))
+                .collect();
+            pairs.sort();
+            info!(
+                "bootstrap resolver: host overrides (skip DNS, connect direct): {}",
+                pairs.join(", ")
+            );
+        }
+        Self {
+            bootstrap,
+            overrides,
+        }
+    }
+
+    #[cfg(test)]
+    pub fn bootstrap(&self) -> &[SocketAddr] {
+        &self.bootstrap
+    }
+}
+
+impl Resolve for NumaResolver {
+    fn resolve(&self, name: Name) -> Resolving {
+        let hostname = name.as_str().to_string();
+
+        if let Some(ips) = self.overrides.get(&hostname) {
+            let addrs: Vec<SocketAddr> = ips.iter().map(|ip| SocketAddr::new(*ip, 0)).collect();
+            debug!(
+                "bootstrap_resolver: override hit for {} → {:?}",
+                hostname, ips
+            );
+            return Box::pin(async move { Ok(Box::new(addrs.into_iter()) as Addrs) });
+        }
+
+        let bootstrap = self.bootstrap.clone();
+        Box::pin(async move {
+            let addrs = resolve_via_bootstrap(&hostname, &bootstrap).await?;
+            debug!(
+                "bootstrap_resolver: resolved {} → {} addr(s)",
+                hostname,
+                addrs.len()
+            );
+            Ok(Box::new(addrs.into_iter()) as Addrs)
+        })
+    }
+}
+
+async fn resolve_via_bootstrap(
+    hostname: &str,
+    bootstrap: &[SocketAddr],
+) -> Result<Vec<SocketAddr>, Box<dyn std::error::Error + Send + Sync>> {
+    let mut last_err: Option<String> = None;
+    for &server in bootstrap {
+        let q_a = DnsPacket::query(0xBEEF, hostname, QueryType::A);
+        let q_aaaa = DnsPacket::query(0xBEF0, hostname, QueryType::AAAA);
+        let (a_res, aaaa_res) = tokio::join!(
+            query_with_tcp_fallback(&q_a, server),
+            query_with_tcp_fallback(&q_aaaa, server),
+        );
+
+        let mut out = Vec::new();
+        match a_res {
+            Ok(pkt) => extract_addrs(&pkt, &mut out),
+            Err(e) => last_err = Some(format!("{} A failed: {}", server, e)),
+        }
+        match aaaa_res {
+            Ok(pkt) => extract_addrs(&pkt, &mut out),
+            // AAAA is optional — many hosts return NXDOMAIN/empty. Don't
+            // treat as the primary error if A succeeded.
+            Err(e) => debug!("bootstrap {} AAAA for {} failed: {}", server, hostname, e),
+        }
+        if !out.is_empty() {
+            return Ok(out);
+        }
+    }
+    Err(last_err
+        .unwrap_or_else(|| "no bootstrap servers reachable".into())
+        .into())
+}
+
+async fn query_with_tcp_fallback(
+    query: &DnsPacket,
+    server: SocketAddr,
+) -> crate::Result<DnsPacket> {
+    match forward_udp(query, server, UDP_TIMEOUT).await {
+        Ok(pkt) => Ok(pkt),
+        Err(e) => {
+            debug!(
+                "bootstrap UDP {} failed ({}), falling back to TCP",
+                server, e
+            );
+            forward_tcp(query, server, TCP_TIMEOUT).await
+        }
+    }
+}
+
+fn extract_addrs(pkt: &DnsPacket, out: &mut Vec<SocketAddr>) {
+    for r in &pkt.answers {
+        match r {
+            DnsRecord::A { addr, .. } => out.push(SocketAddr::new(IpAddr::V4(*addr), 0)),
+            DnsRecord::AAAA { addr, .. } => out.push(SocketAddr::new(IpAddr::V6(*addr), 0)),
+            _ => {}
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::net::{Ipv4Addr, Ipv6Addr};
+
+    #[test]
+    fn empty_fallback_uses_defaults() {
+        let r = NumaResolver::new(&[], HashMap::new());
+        let got: Vec<String> = r.bootstrap().iter().map(|s| s.to_string()).collect();
+        assert_eq!(got, vec!["9.9.9.9:53", "1.1.1.1:53"]);
+    }
+
+    #[test]
+    fn fallback_accepts_ip_literals_only() {
+        let fallback = vec![
+            "9.9.9.9".to_string(),
+            "dns.quad9.net".to_string(),
+            "1.1.1.1:5353".to_string(),
+        ];
+        let r = NumaResolver::new(&fallback, HashMap::new());
+        let got: Vec<String> = r.bootstrap().iter().map(|s| s.to_string()).collect();
+        assert_eq!(got, vec!["9.9.9.9:53", "1.1.1.1:5353"]);
+    }
+
+    #[test]
+    fn override_returns_configured_ips_without_dns() {
+        let mut overrides = HashMap::new();
+        overrides.insert(
+            "odoh-relay.example".to_string(),
+            vec![IpAddr::V4(Ipv4Addr::new(178, 104, 229, 30))],
+        );
+        let r = NumaResolver::new(&[], overrides);
+        let name: Name = "odoh-relay.example".parse().unwrap();
+        let fut = r.resolve(name);
+        let res = futures::executor::block_on(fut).unwrap();
+        let addrs: Vec<_> = res.collect();
+        assert_eq!(addrs.len(), 1);
+        assert_eq!(addrs[0].ip(), IpAddr::V4(Ipv4Addr::new(178, 104, 229, 30)));
+    }
+
+    #[test]
+    fn override_supports_multiple_ips_including_ipv6() {
+        let mut overrides = HashMap::new();
+        overrides.insert(
+            "dual.example".to_string(),
+            vec![
+                IpAddr::V4(Ipv4Addr::new(1, 2, 3, 4)),
+                IpAddr::V6(Ipv6Addr::LOCALHOST),
+            ],
+        );
+        let r = NumaResolver::new(&[], overrides);
+        let res = futures::executor::block_on(r.resolve("dual.example".parse().unwrap())).unwrap();
+        let addrs: Vec<_> = res.collect();
+        assert_eq!(addrs.len(), 2);
+    }
+}

src/config.rs

@@ -56,7 +56,7 @@ impl ForwardingRuleConfig {
         }
         let mut primary = Vec::with_capacity(self.upstream.len());
         for s in &self.upstream {
-            let u = crate::forward::parse_upstream(s, 53)
+            let u = crate::forward::parse_upstream(s, 53, None)
                 .map_err(|e| format!("forwarding rule for upstream '{}': {}", s, e))?;
             primary.push(u);
         }
@@ -241,6 +241,26 @@ pub struct OdohUpstream {
     pub target_bootstrap: Option<SocketAddr>,
 }
 
+impl OdohUpstream {
+    /// Per-host IP overrides for the bootstrap resolver, lifted from
+    /// `relay_ip`/`target_ip`. Keeps the "zero plain-DNS leak for ODoH
+    /// endpoints" property when numa is its own system resolver.
+    pub fn host_ip_overrides(&self) -> std::collections::HashMap<String, Vec<std::net::IpAddr>> {
+        let mut out = std::collections::HashMap::new();
+        if let Some(addr) = self.relay_bootstrap {
+            out.entry(self.relay_host.clone())
+                .or_insert_with(Vec::new)
+                .push(addr.ip());
+        }
+        if let Some(addr) = self.target_bootstrap {
+            out.entry(self.target_host.clone())
+                .or_insert_with(Vec::new)
+                .push(addr.ip());
+        }
+        out
+    }
+}
+
 impl UpstreamConfig {
     /// Validate and extract ODoH-specific fields. Called during `load_config`
     /// so misconfigured ODoH fails fast at startup, the same care we take

src/ctx.rs

@@ -209,45 +209,32 @@ pub async fn resolve_query(
             {
                 // Conditional forwarding takes priority over recursive mode
                 // (e.g. Tailscale .ts.net, VPC private zones)
-                upstream_transport = pool.preferred().map(|u| u.transport());
+                let key = (qname.clone(), qtype);
-                match forward_with_failover_raw(
+                let (resp, path, err) =
+                    resolve_coalesced(&ctx.inflight, key, &query, QueryPath::Forwarded, || async {
+                        let wire = forward_with_failover_raw(
                             raw_wire,
                             pool,
                             &ctx.srtt,
                             ctx.timeout,
                             ctx.hedge_delay,
                         )
-                .await
+                        .await?;
-                {
+                        cache_and_parse(ctx, &qname, qtype, &wire)
-                    Ok(resp_wire) => match cache_and_parse(ctx, &qname, qtype, &resp_wire) {
+                    })
-                        Ok(resp) => (resp, QueryPath::Forwarded, DnssecStatus::Indeterminate),
+                    .await;
-                        Err(e) => {
+                log_coalesced_outcome(src_addr, qtype, &qname, path, err.as_deref(), "FORWARD");
-                            error!("{} | {:?} {} | PARSE ERROR | {}", src_addr, qtype, qname, e);
+                if path == QueryPath::Forwarded {
-                            (
+                    upstream_transport = pool.preferred().map(|u| u.transport());
-                                DnsPacket::response_from(&query, ResultCode::SERVFAIL),
-                                QueryPath::UpstreamError,
-                                DnssecStatus::Indeterminate,
-                            )
-                        }
-                    },
-                    Err(e) => {
-                        error!(
-                            "{} | {:?} {} | FORWARD ERROR | {}",
-                            src_addr, qtype, qname, e
-                        );
-                        (
-                            DnsPacket::response_from(&query, ResultCode::SERVFAIL),
-                            QueryPath::UpstreamError,
-                            DnssecStatus::Indeterminate,
-                        )
-                    }
                 }
+                (resp, path, DnssecStatus::Indeterminate)
             } else if ctx.upstream_mode == UpstreamMode::Recursive {
                 // Recursive resolution makes UDP hops to roots/TLDs/auths;
                 // tag as Udp so the dashboard can aggregate plaintext-wire
                 // egress honestly. Only mark on success — errors stay None.
                 let key = (qname.clone(), qtype);
-                let (resp, path, err) = resolve_coalesced(&ctx.inflight, key, &query, || {
+                let (resp, path, err) =
+                    resolve_coalesced(&ctx.inflight, key, &query, QueryPath::Recursive, || {
                         crate::recursive::resolve_recursive(
                             &qname,
                             qtype,
@@ -258,57 +245,32 @@ pub async fn resolve_query(
                         )
                     })
                     .await;
-                if path == QueryPath::Coalesced {
+                log_coalesced_outcome(src_addr, qtype, &qname, path, err.as_deref(), "RECURSIVE");
-                    debug!("{} | {:?} {} | COALESCED", src_addr, qtype, qname);
+                if path == QueryPath::Recursive {
-                } else if path == QueryPath::UpstreamError {
-                    error!(
-                        "{} | {:?} {} | RECURSIVE ERROR | {}",
-                        src_addr,
-                        qtype,
-                        qname,
-                        err.as_deref().unwrap_or("leader failed")
-                    );
-                } else {
                     upstream_transport = Some(crate::stats::UpstreamTransport::Udp);
                 }
                 (resp, path, DnssecStatus::Indeterminate)
             } else {
                 let pool = ctx.upstream_pool.lock().unwrap().clone();
-                match forward_with_failover_raw(
+                let key = (qname.clone(), qtype);
+                let (resp, path, err) =
+                    resolve_coalesced(&ctx.inflight, key, &query, QueryPath::Upstream, || async {
+                        let wire = forward_with_failover_raw(
                             raw_wire,
                             &pool,
                             &ctx.srtt,
                             ctx.timeout,
                             ctx.hedge_delay,
                         )
-                .await
+                        .await?;
-                {
+                        cache_and_parse(ctx, &qname, qtype, &wire)
-                    Ok(resp_wire) => match cache_and_parse(ctx, &qname, qtype, &resp_wire) {
+                    })
-                        Ok(resp) => {
+                    .await;
+                log_coalesced_outcome(src_addr, qtype, &qname, path, err.as_deref(), "UPSTREAM");
+                if path == QueryPath::Upstream {
                     upstream_transport = pool.preferred().map(|u| u.transport());
-                            (resp, QueryPath::Upstream, DnssecStatus::Indeterminate)
-                        }
-                        Err(e) => {
-                            error!("{} | {:?} {} | PARSE ERROR | {}", src_addr, qtype, qname, e);
-                            (
-                                DnsPacket::response_from(&query, ResultCode::SERVFAIL),
-                                QueryPath::UpstreamError,
-                                DnssecStatus::Indeterminate,
-                            )
-                        }
-                    },
-                    Err(e) => {
-                        error!(
-                            "{} | {:?} {} | UPSTREAM ERROR | {}",
-                            src_addr, qtype, qname, e
-                        );
-                        (
-                            DnsPacket::response_from(&query, ResultCode::SERVFAIL),
-                            QueryPath::UpstreamError,
-                            DnssecStatus::Indeterminate,
-                        )
-                    }
                 }
+                (resp, path, DnssecStatus::Indeterminate)
             }
         }
     };
@@ -611,11 +573,15 @@ fn acquire_inflight(inflight: &Mutex<InflightMap>, key: (String, QueryType)) ->
 
 /// Run a resolve function with in-flight coalescing. Multiple concurrent calls
 /// for the same key share a single resolution — the first caller (leader)
-/// executes `resolve_fn`, and followers wait for the broadcast result.
+/// executes `resolve_fn`, and followers wait for the broadcast result. The
+/// leader's successful path is tagged with `leader_path` so callers that
+/// share this helper (recursive, forwarded-rule, forward-upstream) keep their
+/// own observability without duplicating the inflight map.
 async fn resolve_coalesced<F, Fut>(
     inflight: &Mutex<InflightMap>,
     key: (String, QueryType),
     query: &DnsPacket,
+    leader_path: QueryPath,
     resolve_fn: F,
 ) -> (DnsPacket, QueryPath, Option<String>)
 where
@@ -644,7 +610,7 @@ where
             match result {
                 Ok(resp) => {
                     let _ = tx.send(Some(resp.clone()));
-                    (resp, QueryPath::Recursive, None)
+                    (resp, leader_path, None)
                 }
                 Err(e) => {
                     let _ = tx.send(None);
@@ -671,6 +637,33 @@ impl Drop for InflightGuard<'_> {
     }
 }
 
+/// Emit the log lines shared by the three upstream branches (Forwarded,
+/// Recursive, Upstream) after `resolve_coalesced` returns. Leader-success
+/// and transport-tagging stay at the call site since they diverge per
+/// branch, but the Coalesced debug and UpstreamError error are identical
+/// except for the label.
+fn log_coalesced_outcome(
+    src_addr: SocketAddr,
+    qtype: QueryType,
+    qname: &str,
+    path: QueryPath,
+    err: Option<&str>,
+    label: &str,
+) {
+    match path {
+        QueryPath::Coalesced => debug!("{} | {:?} {} | COALESCED", src_addr, qtype, qname),
+        QueryPath::UpstreamError => error!(
+            "{} | {:?} {} | {} ERROR | {}",
+            src_addr,
+            qtype,
+            qname,
+            label,
+            err.unwrap_or("leader failed")
+        ),
+        _ => {}
+    }
+}
+
 fn special_use_response(query: &DnsPacket, qname: &str, qtype: QueryType) -> DnsPacket {
     use std::net::{Ipv4Addr, Ipv6Addr};
     if qname == "ipv4only.arpa" {
@@ -909,7 +902,7 @@ mod tests {
             let key = ("coalesce.test".to_string(), QueryType::A);
             let query = DnsPacket::query(100 + i, "coalesce.test", QueryType::A);
             handles.push(tokio::spawn(async move {
-                resolve_coalesced(&inf, key, &query, || async {
+                resolve_coalesced(&inf, key, &query, QueryPath::Recursive, || async {
                     count.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                     tokio::time::sleep(Duration::from_millis(200)).await;
                     Ok(mock_response("coalesce.test"))
@@ -953,6 +946,7 @@ mod tests {
                 &inf1,
                 ("same.domain".to_string(), QueryType::A),
                 &query_a,
+                QueryPath::Recursive,
                 || async {
                     count1.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                     tokio::time::sleep(Duration::from_millis(100)).await;
@@ -966,6 +960,7 @@ mod tests {
                 &inf2,
                 ("same.domain".to_string(), QueryType::AAAA),
                 &query_aaaa,
+                QueryPath::Recursive,
                 || async {
                     count2.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                     tokio::time::sleep(Duration::from_millis(100)).await;
@@ -995,6 +990,7 @@ mod tests {
             &inflight,
             ("will-fail.test".to_string(), QueryType::A),
             &query,
+            QueryPath::Recursive,
             || async { Err::<DnsPacket, _>("upstream timeout".into()) },
         )
         .await;
@@ -1016,6 +1012,7 @@ mod tests {
                     &inf,
                     ("fail.test".to_string(), QueryType::A),
                     &query,
+                    QueryPath::Recursive,
                     || async {
                         tokio::time::sleep(Duration::from_millis(200)).await;
                         Err::<DnsPacket, _>("upstream error".into())
@@ -1056,6 +1053,7 @@ mod tests {
             &inflight,
             ("question.test".to_string(), QueryType::A),
             &query,
+            QueryPath::Recursive,
             || async { Err::<DnsPacket, _>("fail".into()) },
         )
         .await;
@@ -1080,6 +1078,7 @@ mod tests {
             &inflight,
             ("err-msg.test".to_string(), QueryType::A),
             &query,
+            QueryPath::Recursive,
             || async { Err::<DnsPacket, _>("connection refused by upstream".into()) },
         )
         .await;

src/forward.rs

@@ -113,10 +113,7 @@ impl fmt::Display for Upstream {
     }
 }
 
-pub(crate) fn parse_upstream_addr(
+pub fn parse_upstream_addr(s: &str, default_port: u16) -> std::result::Result<SocketAddr, String> {
-    s: &str,
-    default_port: u16,
-) -> std::result::Result<SocketAddr, String> {
     // Try full socket addr first: "1.2.3.4:5353" or "[::1]:5353"
     if let Ok(addr) = s.parse::<SocketAddr>() {
         return Ok(addr);
@@ -129,19 +126,28 @@ pub(crate) fn parse_upstream_addr(
 }
 
 /// Parse a slice of upstream address strings into `Upstream` values, failing
-/// on the first invalid entry.
+/// on the first invalid entry. DoH entries use `resolver` (when provided) as
-pub fn parse_upstream_list(addrs: &[String], default_port: u16) -> Result<Vec<Upstream>> {
+/// their hostname resolver.
+pub fn parse_upstream_list(
+    addrs: &[String],
+    default_port: u16,
+    resolver: Option<Arc<crate::bootstrap_resolver::NumaResolver>>,
+) -> Result<Vec<Upstream>> {
     addrs
         .iter()
-        .map(|s| parse_upstream(s, default_port))
+        .map(|s| parse_upstream(s, default_port, resolver.clone()))
         .collect()
 }
 
-pub fn parse_upstream(s: &str, default_port: u16) -> Result<Upstream> {
+pub fn parse_upstream(
+    s: &str,
+    default_port: u16,
+    resolver: Option<Arc<crate::bootstrap_resolver::NumaResolver>>,
+) -> Result<Upstream> {
     if s.starts_with("https://") {
         return Ok(Upstream::Doh {
             url: s.to_string(),
-            client: build_https_client(),
+            client: build_https_client_with_resolver(1, resolver),
         });
     }
     // tls://IP:PORT#hostname  or  tls://IP#hostname  (default port 853)
@@ -163,12 +169,16 @@ pub fn parse_upstream(s: &str, default_port: u16) -> Result<Upstream> {
 }
 
 /// HTTP/2 client tuned for DoH/ODoH: small windows for low latency, long-lived
-/// keep-alive. Shared by the DoH upstream and the ODoH config-fetcher +
+/// keep-alive. Pool defaults to one idle conn per host — good for resolvers
-/// seal/open path. Pool defaults to one idle conn per host — good for
+/// that talk to a single upstream; relays that fan out to many targets
-/// resolvers that talk to a single upstream; relays that fan out to many
+/// should use [`build_https_client_with_pool`].
-/// targets should use [`build_https_client_with_pool`].
+///
+/// Uses the system resolver. Callers running inside `serve::run` pass the
+/// shared [`crate::bootstrap_resolver::NumaResolver`] via
+/// [`build_https_client_with_resolver`] to avoid the self-loop documented
+/// in `docs/implementation/bootstrap-resolver.md`.
 pub fn build_https_client() -> reqwest::Client {
-    build_https_client_with_pool(1)
+    build_https_client_with_resolver(1, None)
 }
 
 /// Same shape as [`build_https_client`], but caller picks
@@ -176,20 +186,18 @@ pub fn build_https_client() -> reqwest::Client {
 /// and benefit from a larger pool so warm connections survive concurrent
 /// fan-out.
 pub fn build_https_client_with_pool(pool_max_idle_per_host: usize) -> reqwest::Client {
-    https_client_builder(pool_max_idle_per_host)
+    build_https_client_with_resolver(pool_max_idle_per_host, None)
-        .build()
-        .unwrap_or_default()
 }
 
-/// HTTPS client for the ODoH upstream, with bootstrap-IP overrides applied
+/// [`build_https_client`] with an optional custom DNS resolver. Numa wires
-/// so relay/target hostname resolution can bypass system DNS.
+/// [`crate::bootstrap_resolver::NumaResolver`] here.
-pub fn build_odoh_client(odoh: &crate::config::OdohUpstream) -> reqwest::Client {
+pub fn build_https_client_with_resolver(
-    let mut builder = https_client_builder(1);
+    pool_max_idle_per_host: usize,
-    if let Some(addr) = odoh.relay_bootstrap {
+    resolver: Option<Arc<crate::bootstrap_resolver::NumaResolver>>,
-        builder = builder.resolve(&odoh.relay_host, addr);
+) -> reqwest::Client {
-    }
+    let mut builder = https_client_builder(pool_max_idle_per_host);
-    if let Some(addr) = odoh.target_bootstrap {
+    if let Some(r) = resolver {
-        builder = builder.resolve(&odoh.target_host, addr);
+        builder = builder.dns_resolver(r);
     }
     builder.build().unwrap_or_default()
 }
@@ -553,6 +561,9 @@ async fn forward_doh_raw(
 
 /// Send a lightweight keepalive query to a DoH upstream to prevent
 /// the HTTP/2 + TLS connection from going idle and being torn down.
+/// The first call doubles as a startup warm-up: bootstrap-resolver failures
+/// (unreachable Quad9/Cloudflare defaults, misconfigured hostname upstream)
+/// surface here rather than on the first client query.
 pub async fn keepalive_doh(upstream: &Upstream) {
     if let Upstream::Doh { url, client } = upstream {
         // Query for . NS — minimal, always succeeds, response is small
@@ -565,7 +576,9 @@ pub async fn keepalive_doh(upstream: &Upstream) {
             0x00, 0x02, // type NS
             0x00, 0x01, // class IN
         ];
-        let _ = forward_doh_raw(wire, url, client, Duration::from_secs(5)).await;
+        if let Err(e) = forward_doh_raw(wire, url, client, Duration::from_secs(5)).await {
+            log::warn!("DoH keepalive to {} failed: {}", url, e);
+        }
     }
 }
 

src/lib.rs

@@ -1,5 +1,6 @@
 pub mod api;
 pub mod blocklist;
+pub mod bootstrap_resolver;
 pub mod buffer;
 pub mod cache;
 pub mod config;

src/serve.rs

@@ -13,12 +13,13 @@ use log::{error, info};
 use tokio::net::UdpSocket;
 
 use crate::blocklist::{download_blocklists, parse_blocklist, BlocklistStore};
+use crate::bootstrap_resolver::NumaResolver;
 use crate::buffer::BytePacketBuffer;
 use crate::cache::DnsCache;
 use crate::config::{build_zone_map, load_config, ConfigLoad};
 use crate::ctx::{handle_query, ServerCtx};
 use crate::forward::{
-    build_https_client, build_odoh_client, parse_upstream_list, Upstream, UpstreamPool,
+    build_https_client_with_resolver, parse_upstream_list, Upstream, UpstreamPool,
 };
 use crate::odoh::OdohConfigCache;
 use crate::override_store::OverrideStore;
@@ -48,6 +49,23 @@ pub async fn run(config_path: String) -> crate::Result<()> {
         (dummy, "recursive (root hints)".to_string())
     };
 
+    // Routes numa-originated HTTPS (DoH upstream, ODoH relay/target, blocklist
+    // CDN) away from the system resolver so lookups don't loop back through
+    // numa when it's its own system DNS.
+    // See `docs/implementation/bootstrap-resolver.md`.
+    let resolver_overrides = match config.upstream.mode {
+        crate::config::UpstreamMode::Odoh => config
+            .upstream
+            .odoh_upstream()
+            .map(|o| o.host_ip_overrides())
+            .unwrap_or_default(),
+        _ => std::collections::HashMap::new(),
+    };
+    let bootstrap_resolver: Arc<NumaResolver> = Arc::new(NumaResolver::new(
+        &config.upstream.fallback,
+        resolver_overrides,
+    ));
+
     let (resolved_mode, upstream_auto, pool, upstream_label) = match config.upstream.mode {
         crate::config::UpstreamMode::Auto => {
             info!("auto mode: probing recursive resolution...");
@@ -57,7 +75,7 @@ pub async fn run(config_path: String) -> crate::Result<()> {
                 (crate::config::UpstreamMode::Recursive, false, pool, label)
             } else {
                 log::warn!("recursive probe failed — falling back to Quad9 DoH");
-                let client = build_https_client();
+                let client = build_https_client_with_resolver(1, Some(bootstrap_resolver.clone()));
                 let url = DOH_FALLBACK.to_string();
                 let label = url.clone();
                 let pool = UpstreamPool::new(vec![Upstream::Doh { url, client }], vec![]);
@@ -82,8 +100,16 @@ pub async fn run(config_path: String) -> crate::Result<()> {
                 config.upstream.address.clone()
             };
 
-            let primary = parse_upstream_list(&addrs, config.upstream.port)?;
+            let primary = parse_upstream_list(
-            let fallback = parse_upstream_list(&config.upstream.fallback, config.upstream.port)?;
+                &addrs,
+                config.upstream.port,
+                Some(bootstrap_resolver.clone()),
+            )?;
+            let fallback = parse_upstream_list(
+                &config.upstream.fallback,
+                config.upstream.port,
+                Some(bootstrap_resolver.clone()),
+            )?;
 
             let pool = UpstreamPool::new(primary, fallback);
             let label = pool.label();
@@ -96,7 +122,7 @@ pub async fn run(config_path: String) -> crate::Result<()> {
         }
         crate::config::UpstreamMode::Odoh => {
             let odoh = config.upstream.odoh_upstream()?;
-            let client = build_odoh_client(&odoh);
+            let client = build_https_client_with_resolver(1, Some(bootstrap_resolver.clone()));
             let target_config = Arc::new(OdohConfigCache::new(
                 odoh.target_host.clone(),
                 client.clone(),
@@ -110,7 +136,11 @@ pub async fn run(config_path: String) -> crate::Result<()> {
             let fallback = if odoh.strict {
                 Vec::new()
             } else {
-                parse_upstream_list(&config.upstream.fallback, config.upstream.port)?
+                parse_upstream_list(
+                    &config.upstream.fallback,
+                    config.upstream.port,
+                    Some(bootstrap_resolver.clone()),
+                )?
             };
             let pool = UpstreamPool::new(primary, fallback);
             let label = pool.label();
@@ -405,8 +435,9 @@ pub async fn run(config_path: String) -> crate::Result<()> {
     if config.blocking.enabled && !blocklist_lists.is_empty() {
         let bl_ctx = Arc::clone(&ctx);
         let bl_lists = blocklist_lists.clone();
+        let bl_resolver = bootstrap_resolver.clone();
         tokio::spawn(async move {
-            load_blocklists(&bl_ctx, &bl_lists).await;
+            load_blocklists(&bl_ctx, &bl_lists, Some(bl_resolver.clone())).await;
 
             // Periodic refresh
             let mut interval = tokio::time::interval(Duration::from_secs(refresh_hours * 3600));
@@ -414,7 +445,7 @@ pub async fn run(config_path: String) -> crate::Result<()> {
             loop {
                 interval.tick().await;
                 info!("refreshing blocklists...");
-                load_blocklists(&bl_ctx, &bl_lists).await;
+                load_blocklists(&bl_ctx, &bl_lists, Some(bl_resolver.clone())).await;
             }
         });
     }
@@ -596,8 +627,8 @@ async fn network_watch_loop(ctx: Arc<ServerCtx>) {
     }
 }
 
-async fn load_blocklists(ctx: &ServerCtx, lists: &[String]) {
+async fn load_blocklists(ctx: &ServerCtx, lists: &[String], resolver: Option<Arc<NumaResolver>>) {
-    let downloaded = download_blocklists(lists).await;
+    let downloaded = download_blocklists(lists, resolver).await;
 
     // Parse outside the lock to avoid blocking DNS queries during parse (~100ms)
     let mut all_domains = std::collections::HashSet::new();
@@ -632,8 +663,10 @@ async fn warm_domain(ctx: &ServerCtx, domain: &str) {
 }
 
 async fn doh_keepalive_loop(ctx: Arc<ServerCtx>) {
+    // First tick fires immediately so we surface bootstrap-resolver failures
+    // (unreachable Quad9/Cloudflare, blocked :53, bad upstream hostname) in
+    // the startup logs instead of on the first client query.
     let mut interval = tokio::time::interval(Duration::from_secs(25));
-    interval.tick().await; // skip first immediate tick
     loop {
         interval.tick().await;
         let pool = ctx.upstream_pool.lock().unwrap().clone();

tests/docker/self-resolver-loop.sh

@@ -0,0 +1,155 @@
+#!/usr/bin/env bash
+#
+# Reproducer for issue #122 — chicken-and-egg when numa is its own system
+# resolver (HAOS add-on, Pi-hole-style container, laptop with
+# resolv.conf → 127.0.0.1).
+#
+# Topology:
+#   container /etc/resolv.conf  →  nameserver 127.0.0.1
+#   numa bound on :53           →  upstream DoH by hostname (quad9)
+#   numa boots → spawns blocklist download
+#   reqwest::get → getaddrinfo("cdn.jsdelivr.net")
+#     → loopback UDP :53 → numa → cache miss → DoH upstream
+#     → getaddrinfo("dns.quad9.net") → same loop → glibc EAI_AGAIN
+#
+# Expected on master: both assertions FAIL (bug reproduced).
+# Expected after bootstrap-IP fix: both assertions PASS.
+#
+# Requirements: docker (with internet access for external lists/DoH)
+# Usage:        ./tests/docker/self-resolver-loop.sh
+
+set -euo pipefail
+
+cd "$(dirname "$0")/../.."
+
+GREEN="\033[32m"; RED="\033[31m"; RESET="\033[0m"
+
+pass() { printf "  ${GREEN}✓${RESET} %s\n" "$1"; }
+fail() { printf "  ${RED}✗${RESET} %s\n" "$1"; printf "    %s\n" "$2"; FAILED=$((FAILED+1)); }
+FAILED=0
+
+OUT=/tmp/numa-self-resolver.out
+
+echo "── self-resolver-loop: building + reproducing on debian:bookworm ──"
+echo "  (first run is slow: image pull + cold cargo build, ~5-8 min)"
+echo
+
+docker run --rm \
+    -v "$PWD:/src:ro" \
+    -v numa-self-resolver-cargo:/root/.cargo \
+    -v numa-self-resolver-target:/work/target \
+    debian:bookworm bash -c '
+set -e
+
+# Phase 1: install deps + build with the container DNS as given by Docker
+# (resolves deb.debian.org, static.rust-lang.org, crates.io).
+apt-get update -qq && apt-get install -y -qq curl build-essential dnsutils 2>&1 | tail -3
+
+if ! command -v cargo &>/dev/null; then
+    curl -sSf https://sh.rustup.rs | sh -s -- -y --profile minimal --quiet
+fi
+. "$HOME/.cargo/env"
+
+mkdir -p /work
+tar -C /src --exclude=./target --exclude=./.git -cf - . | tar -C /work -xf -
+cd /work
+
+echo "── cargo build --release --locked ──"
+cargo build --release --locked 2>&1 | tail -5
+echo
+
+# Phase 2: flip system DNS to numa itself — this is the pathological
+# topology from issue #122 (HAOS add-on, resolv.conf → 127.0.0.1).
+# Everything after this point, any getaddrinfo call inside numa loops
+# back through :53.
+echo "nameserver 127.0.0.1" > /etc/resolv.conf
+echo "── /etc/resolv.conf inside container (post-flip) ──"
+cat /etc/resolv.conf
+echo
+
+cat > /tmp/numa.toml <<CONF
+[server]
+bind_addr = "0.0.0.0:53"
+api_port = 5380
+api_bind_addr = "127.0.0.1"
+data_dir = "/tmp/numa-data"
+
+[upstream]
+mode = "forward"
+address = ["https://dns.quad9.net/dns-query"]
+timeout_ms = 3000
+
+[blocking]
+enabled = true
+lists = ["https://cdn.jsdelivr.net/gh/hagezi/dns-blocklists@latest/hosts/pro.txt"]
+CONF
+
+mkdir -p /tmp/numa-data
+
+echo "── starting numa ──"
+RUST_LOG=info ./target/release/numa /tmp/numa.toml > /tmp/numa.log 2>&1 &
+NUMA_PID=$!
+
+# Wait up to 120s for blocklist to populate.
+# Retry delays 2+10+30s = 42s, plus ~4 × ~10s getaddrinfo timeouts under
+# self-loop = ~82s worst case. 120s leaves headroom.
+LOADED=0
+for i in $(seq 1 120); do
+    LOADED=$(curl -sf http://127.0.0.1:5380/blocking/stats 2>/dev/null \
+        | grep -o "\"domains_loaded\":[0-9]*" | cut -d: -f2 || echo 0)
+    [ "${LOADED:-0}" -gt 100 ] && break
+    sleep 1
+done
+
+# First cold DoH query — time it.
+START=$(date +%s%N)
+dig @127.0.0.1 example.com A +time=15 +tries=1 > /tmp/dig.out 2>&1 || true
+END=$(date +%s%N)
+LATENCY_MS=$(( (END - START) / 1000000 ))
+STATUS=$(grep -oE "status: [A-Z]+" /tmp/dig.out | head -1 || echo "status: TIMEOUT")
+
+kill $NUMA_PID 2>/dev/null || true
+wait $NUMA_PID 2>/dev/null || true
+
+echo
+echo "=== RESULT ==="
+echo "domains_loaded=$LOADED"
+echo "first_query_latency_ms=$LATENCY_MS"
+echo "first_query_${STATUS// /_}"
+echo
+echo "=== numa.log (tail 40) ==="
+tail -40 /tmp/numa.log
+echo
+echo "=== dig.out ==="
+cat /tmp/dig.out
+' 2>&1 | tee "$OUT"
+
+echo
+echo "── assertions ──"
+
+LOADED=$(grep '^domains_loaded=' "$OUT" | tail -1 | cut -d= -f2 || echo 0)
+LATENCY=$(grep '^first_query_latency_ms=' "$OUT" | tail -1 | cut -d= -f2 || echo 999999)
+STATUS_LINE=$(grep '^first_query_status_' "$OUT" | tail -1 || echo "first_query_status_TIMEOUT")
+
+if [ "${LOADED:-0}" -gt 100 ]; then
+    pass "blocklist downloaded (domains_loaded=$LOADED)"
+else
+    fail "blocklist downloaded (got domains_loaded=${LOADED:-0}, expected >100)" \
+         "chicken-and-egg: blocklist HTTPS client has no DNS bootstrap; getaddrinfo loops through numa"
+fi
+
+if [ "${LATENCY:-999999}" -lt 2000 ]; then
+    pass "first DoH query under 2s (latency=${LATENCY}ms, $STATUS_LINE)"
+else
+    fail "first DoH query under 2s (got ${LATENCY}ms, $STATUS_LINE)" \
+         "self-loop on getaddrinfo(upstream_host); plain DoH needs bootstrap-IP symmetry with ODoH"
+fi
+
+echo
+if [ "$FAILED" -eq 0 ]; then
+    printf "${GREEN}── self-resolver-loop passed (fix is in place) ──${RESET}\n"
+    exit 0
+else
+    printf "${RED}── self-resolver-loop failed ($FAILED assertion(s)) — bug #122 reproduced ──${RESET}\n"
+    exit 1
+fi

tests/integration.sh

@@ -975,6 +975,52 @@ check "Same-host relay+target rejected at startup" \
     "same host" \
     "$STARTUP_OUT"
 
+# relay_ip / target_ip must land in the bootstrap resolver's override map,
+# so reqwest connects direct to the configured IPs instead of resolving the
+# hostnames via plain DNS (ODoH's zero-plain-DNS-leak property). Using
+# RFC 5737 TEST-NET-1 IPs — never routable, so the OdohConfigCache won't
+# actually connect, but the override-map wiring is visible in the startup log.
+cat > "$CONFIG" << 'CONF'
+[server]
+bind_addr = "127.0.0.1:5354"
+api_port = 5381
+
+[upstream]
+mode = "odoh"
+relay = "https://odoh-relay.example.com/proxy"
+target = "https://odoh-target.example.org/dns-query"
+relay_ip = "192.0.2.1"
+target_ip = "192.0.2.2"
+
+[cache]
+max_entries = 10000
+
+[blocking]
+enabled = false
+
+[proxy]
+enabled = false
+CONF
+
+RUST_LOG=info "$BINARY" "$CONFIG" > "$LOG" 2>&1 &
+NUMA_PID=$!
+for _ in $(seq 1 30); do
+    curl -sf "http://127.0.0.1:$API_PORT/health" >/dev/null 2>&1 && break
+    sleep 0.1
+done
+
+OVERRIDE_LOG=$(grep 'bootstrap resolver: host overrides' "$LOG" || true)
+check "relay_ip wired into bootstrap override map" \
+    "odoh-relay.example.com=192.0.2.1" \
+    "$OVERRIDE_LOG"
+check "target_ip wired into bootstrap override map" \
+    "odoh-target.example.org=192.0.2.2" \
+    "$OVERRIDE_LOG"
+
+kill "$NUMA_PID" 2>/dev/null || true
+wait "$NUMA_PID" 2>/dev/null || true
+sleep 1
+
 fi  # end Suite 8
 
 # ---- Suite 9: Numa's own ODoH relay (--relay-mode) ----