Why Unbound kept retrying over TCP on large DNSSEC answers
Looking at unbound-control stats_noreset on gw01, roughly 4% of cold-cache queries
showed a 1.5-second latency spike paired with incrementing num.query.tcpout. All of them were
DNSSEC-signed zones during ZSK rollover over IPv6 transport.
Fragmented UDP over IPv6
An old config snippet I had copied years ago still set edns-buffer-size: 4096. When an
authoritative server replied with a 1780-byte signed UDP packet over IPv6, it attached an IPv6 Fragment
Header; somewhere upstream of my PPPoE session, fragmented UDP/53 packets were dropped on sight.
Unbound waited for UDP timeout before falling back to EDNS 512 / TCP.
Stick to the Flag Day 2020 default (1232)
Modern Unbound already defaults to 1232 bytes (1280-byte minimum IPv6 MTU minus 40-byte IPv6
header minus 8-byte UDP header) as long as you delete stale overrides:
# /etc/unbound/unbound.conf.d/local.conf
server:
edns-buffer-size: 1232
max-udp-size: 1232
harden-algo-downgrade: yes
qname-minimisation: yes
With 1232, the authoritative server sees the smaller buffer advertisement immediately, sets the
TC=1 truncated bit in a single unfragmented UDP reply, and Unbound retries over TCP in one RTT
instead of burning a full UDP retransmission timer first.
$ dig @127.0.0.1 +dnssec +bufsize=1232 org. DNSKEY | grep -E 'flags:|Query time:|MSG SIZE'
;; flags: qr rd ra ad; QUERY: 1, ANSWER: 5, AUTHORITY: 0, ADDITIONAL: 1
;; Query time: 38 msec
;; MSG SIZE rcvd: 1624