dns: fix RFC 8305 family interleaving of getaddrinfo results

[crishoj]

(Disclosure: Code and writeup by Claude Fable 5)

Fixes #33278.

A cold fetch() (or any usockets connect) on a network with advertised-but-unroutable IPv6 stalls for the full OS connect timeout (~70 s) when the host resolves to several AAAA records before any A record: all CONCURRENT_CONNECTIONS = 4 parallel attempts go to the first four addresses in resolver order — all dead IPv6 — and blackholed SYNs never fail, so the IPv4 addresses are never tried.

Root cause

process_results() in src/runtime/dns_jsc/dns.rs — the function that packs getaddrinfo results for the usockets/QUIC connect cache — has always had a // sort (interleave ipv4 and ipv6) pass, but it never worked:

• It is a line-for-line port of a Zig loop whose for … else { break } fired unconditionally (the port note said so explicitly), so the outer loop dies after a single index.
• want starts at AF_INET6 and only toggles when a swap happens, so any list starting with IPv6 — the exact failing shape — just continues and returns completely untouched.

Fix

Ordering now happens while the entries are copied into the packed allocation: alternate between the first entry's family and the rest, preserving resolver order within each family (RFC 8305 §4). Starting from the resolver's first family (rather than hardcoding IPv6) respects the OS's RFC 6724 policy, so v4-first results stay v4-first. The slot mapping is a closed-form bijection over 0..count, so the MaybeUninit/assume_init invariant is unchanged; the old post-hoc swap block is deleted.

With interleaving in place, the first batch of 4 attempts covers both families, so a run of dead same-family addresses can no longer monopolize it. This is deliberately scoped to the existing (broken) mechanism; making dns.setDefaultResultOrder visible to the native path is a separate concern tracked in #28817.

Test

The packing is exposed through bun:internal-for-testing (dnsInternals.getaddrinfoInterleave), following the socket_body.rs / linear_fifo_testing.rs pattern: the probe builds a synthetic getaddrinfo chain from a list of families, runs the real process_results, and walks the rebuilt ai_next chain (original index stashed in the port so stability is observable). 8 cases in test/js/bun/dns/dns-prefetch.test.ts cover the issue's exact shape, v4-first, uneven counts, single-family, and empty inputs.

Verified the test fails for the right reason: with the copy loop reverted to sequential order, the three mixed-family cases fail (the others are fixed points of interleaving by design). The slot formula was additionally checked exhaustively for every family sequence up to length 8 (bijection, alternating prefix, per-family stability).

Honest caveat on the end-to-end repro: the network I filed the issue from has since changed — IPv6 now fails in ~2 ms (unreachable) instead of blackholing, and fast-failing attempts always recovered via the refill-on-error path — so I could not re-measure the 70 s → sub-second improvement live. Deleting the fix and re-running the new tests is the deterministic evidence.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: adcd99f1-8a49-499a-9cd0-acd9c626d613

📥 Commits

Reviewing files that changed from the base of the PR and between 1498d7b and 8cf3e01.

📒 Files selected for processing (4)

• src/js/internal-for-testing.ts
• src/runtime/dns_jsc/dns.rs
• src/runtime/dns_jsc/mod.rs
• test/js/bun/dns/dns-prefetch.test.ts

---

Walkthrough

Adds RFC 8305 address-family interleaving to DNS getaddrinfo result processing in dns.rs, replacing a post-copy sort/interleave pass with in-place interleaved placement during the copy loop. Introduces a bun:internal-for-testing bridge (testing_apis.getaddrinfo_interleave) with a new dns re-export module, exposed via dnsInternals.getaddrinfoInterleave in internal-for-testing.ts, and covered by new interleaving tests.

Changes

RFC 8305 getaddrinfo interleaving

Layer / File(s)	Summary
Interleaving algorithm in process_results src/runtime/dns_jsc/dns.rs	internal::process_results now computes the first observed family and count, places each addrinfo node at an interleaved destination index during the copy loop, and finalizes the result buffer directly, removing the prior separate sort/interleave pass.
Testing bridge and JS exposure src/runtime/dns_jsc/dns.rs, src/runtime/dns_jsc/mod.rs, src/js/internal-for-testing.ts	A new testing_apis module with getaddrinfo_interleave() builds a synthetic AddrInfo chain from given families, runs process_results, and returns encoded order; mod.rs re-exports it via a new dns module; dnsInternals.getaddrinfoInterleave exposes it to JS via $newRustFunction.
Interleaving tests test/js/bun/dns/dns-prefetch.test.ts	Imports dnsInternals and adds a parameterized describe/test.each suite validating getaddrinfoInterleave output against expected interleaved, encoded family orders.

Sequence Diagram(s)

sequenceDiagram
  participant Test as dns-prefetch.test.ts
  participant JS as dnsInternals.getaddrinfoInterleave
  participant Bridge as testing_apis.getaddrinfo_interleave
  participant Core as internal::process_results

  Test->>JS: getaddrinfoInterleave(families)
  JS->>Bridge: invoke Rust function
  Bridge->>Bridge: build synthetic AddrInfo chain
  Bridge->>Core: process_results(chain)
  Core-->>Bridge: interleaved ResultEntry list
  Bridge-->>JS: encoded order array
  JS-->>Test: interleaved family order

Loading

Related issues: #33278 (fetch stalls on unroutable IPv6 due to lack of RFC 8305 family interleaving).

Suggested labels: dns, rust, testing

Suggested reviewers: (none identified from provided context)

---

🐇 A hare hops through addresses four and six,
Interleaved now, no more stuck-in-a-fix,
Rust hums beneath, JS calls with glee,
Tests confirm the order, as it ought to be.

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: RFC 8305 family interleaving for getaddrinfo results.
Description check	✅ Passed	The description covers the change and verification steps, though it doesn't use the template headings verbatim.
Linked Issues check	✅ Passed	The PR implements RFC 8305 interleaving for fetch's native DNS/connect path, which addresses the linked issue's primary bug.
Out of Scope Changes check	✅ Passed	The code changes stay focused on DNS interleaving, test exposure, and coverage, with no clear unrelated scope.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}