2026.9.0 segfaults in quic-go QueueProbePacket (nil receiver) when running in a Docker bridge network — 2026.8.2 unaffected #1737

Describe the bug

cloudflared 2026.9.0 crashes with SIGSEGV roughly every 75 seconds when running
in a Docker container on a bridge network. The panic is a nil-receiver dereference
inside the bundled quic-go fork, on the probe-timeout path.

The same container on 2026.8.2 is completely stable. Notably, a second tunnel
running the same 2026.9.0 binary on the same machine, as a host systemd service,
has zero crashes — the only meaningful difference is that it uses the host
network directly instead of the Docker bridge/NAT.

Since QueueProbePacket is only reached on a PTO (probe timeout), the NAT path
seems to make packet loss / timeouts frequent enough to hit a latent race that the
direct path rarely triggers.

To Reproduce

1. Run cloudflare/cloudflared:2026.9.0 in a Docker container on a user-defined
   bridge network (container IP 172.20.0.2, NAT via iptables MASQUERADE).

2. Command: tunnel --no-autoupdate --loglevel info run --token <token>
   (remotely-managed tunnel, 3 ingress rules pointing at an internal service).

3. Let it run. It starts cleanly every time — all connectivity pre-checks PASS,
   4 connections register (tpe01 / khh01, quic) — then panics after roughly a
   minute and is restarted by restart: unless-stopped.
   Observed: 145 crashes in ~6 hours.

4. Tunnel ID: <填入或留空>

5. cloudflared config: token-based (remotely managed), no local config file.

Expected behavior

The tunnel stays connected without crashing, as it does on 2026.8.2.

Environment and versions

• OS: Ubuntu 24.04.5 LTS, kernel 6.8.0-139-generic (Docker container)
• Architecture: AMD64 — Intel Core 2 Duo P8800, 2 cores (old hardware, no AES-NI;
  noting it in case goroutine timing is relevant to the race)
• Version: 2026.9.0 (built 2026-09-09-18:43 UTC), GoVersion go1.26.8
• Last known good: 2026.8.2 (built 2026-08-14-12:28 UTC)

Logs and errors

[signal SIGSEGV: segmentation violation code=0x1 addr=0x4c pc=0xdae18d]

goroutine 128 [running]:
github.com/quic-go/quic-go/internal/ackhandler.(*sentPacketHandler).getPacketNumberSpace(...)
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/internal/ackhandler/sent_packet_handler.go:374
github.com/quic-go/quic-go/internal/ackhandler.(*sentPacketHandler).QueueProbePacket(0x0?, 0xfc?)
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/internal/ackhandler/sent_packet_handler.go:1042 +0x12d
github.com/quic-go/quic-go.(*Conn).sendProbePacket(0x1524c553b208, 0xd8?, 0x35a126441a6)
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/connection.go:2722 +0x104
github.com/quic-go/quic-go.(*Conn).triggerSending(0x1524c553b208, 0x35a126441a6)
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/connection.go:2486 +0x16a
github.com/quic-go/quic-go.(*Conn).run(0x1524c553b208)
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/connection.go:741 +0x705
github.com/quic-go/quic-go.(*Transport).doDial.func1()
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/transport.go:326 +0x3a
created by github.com/quic-go/quic-go.(*Transport).doDial in goroutine 107
/go/pkg/mod/github.com/chungthuang/quic-go@v0.45.1-0.20260529212404-a9fddf436fc4/transport.go:325 +0x737

The pc is identical on every crash (pc=0xdae18d), with addr varying only
between 0x4b–0x4d — a deterministic nil dereference, not memory corruption.

Additional context

• Not OOM: OOMKilled=false, memory usage 17.4 MiB against a 128 MiB limit.
• Same binary, same host, no crash: the host-side systemd tunnel runs the same
  2026.9.0 with 0 SIGSEGV over 8+ hours.
• Differences between the two processes on this machine:
  • container: Docker bridge NAT (172.20.0.2), ICMP proxy enabled
  • host: direct on the physical NIC, ICMP proxy disabled (ping_group_range)
• net.core.rmem_max / wmem_max are raised to 7500000 on this host, so both
  processes get the full 7 MiB UDP receive buffer they ask for. The stable host
  process runs with the same setting, so this does not appear to be the trigger.
• Rolling the container back to 2026.8.2 resolves it completely: 0 crashes in
  over an hour, where the observed rate would have predicted ~48.