[Bug]: TCPForwarderTest fails intermittently on macOS 27.2 after ConnectHandlerRaceTest: loopback connections are reset before accept

Aperta
#2,501 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
50/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
swift

Direzione di ricerca

Start with TCPForwarderTest.swift at the reported failure lines and compare testTCPForwarder with ConnectHandlerRaceTest.testRapidConnectDisconnect, the parallel test implicated by the reproduction. Run the supplied filtered SocketForwarderTests command repeatedly, then compare parallel and --no-parallel runs. Done means identifying and addressing the cause so the parallel suite passes reliably without dropping client connections.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

I have done the following
  • I have searched the existing issues
  • If possible, I've reproduced the issue using the 'main' branch of this project
Steps to reproduce

Run the SocketForwarderTests target repeatedly on main (16e8d48). The failure is intermittent.

swift test --skip-build -c debug -Xswiftc -warnings-as-errors -Xswiftc -enable-testing --filter SocketForwarderTests

Results on my machine, with the conditions interleaved in blocks:

Condition Failures
TCPForwarderTest alone 0 / 20
TCPForwarderTest|ConnectHandlerRaceTest 7 / 20
SocketForwarderTests 19 / 40
SocketForwarderTests, with the two suites nested under one @Suite(.serialized) parent 32 / 40
Pair, with the stress test pointed at a plain echo listener instead of a TCPForwarder 8 / 15
Pair, with the stress test opening 100 connections instead of 500 4 / 15

The samples are small and the machine was under load for some blocks, so the rates are rough. An earlier batch on c05a560 gave 0 / 60 for TCPForwarderTest alone and 0 / 60 with --no-parallel; I have not rerun --no-parallel and can't explain why it differs from the serialized result.

I have only seen this on my own machine. In the CI logs I looked at (macOS 27.0), testTCPForwarder passed while starting at the same moment as testRapidConnectDisconnect.

Problem description

TCPForwarderTest.testTCPForwarder fails intermittently when ConnectHandlerRaceTest.testRapidConnectDisconnect is part of the same test run. The two do not need to overlap: with the suites serialized, the stress test finished first in every run, and testTCPForwarder, started right after it, still failed. It fails in one of two ways.

Some of the 100 clients get no echo back:

✘ Test testTCPForwarder() recorded an issue at TCPForwarderTest.swift:108:13: Expectation failed: sortedResponses[i] == "\(i): success-tcp"

Or a client read fails:

✘ Test testTCPForwarder() recorded an issue at TCPForwarderTest.swift:25:6: Caught error: read(descriptor:pointer:size:): Connection reset by peer) (errno: 54)

With temporary tracing added to the forwarder and the tests, the failing connections die before either listener accepts them:

  • Clients that fail with Connection reset by peer had connected and written, but the proxy never accepted them; ConnectHandler.channelActive never ran for those connections.
  • Clients that get no echo were accepted by the proxy, and the backend connect succeeded, but the echo server never accepted that backend connection. The backend read then failed with ECONNRESET and GlueHandler.errorCaught closed the frontend, which is the intended behaviour. In one failing run, 100 backend connects succeeded, the echo server accepted 76, and the other 24 were reset.
  • During failing runs, netstat -Lan showed the test's own listeners holding 30-50 connections in the incomplete queue for hundreds of milliseconds. Neither listener has more than 100 pending connections, so the backlog (clamped to kern.ipc.somaxconn, 128) is not exceeded.

So the forwarder does not appear to be at fault: it reacts to a backend connection that was reset underneath it. The failure also occurs when the stress test targets a plain listener, with no TCPForwarder on its side at all.

What I can't explain is why loopback connections are reset before accept for a short time after a burst of connect/close on a different listener. I could not read TCP drop counters or capture packets without root, and I have only tested on macOS 27.2.

I don't have a fix to propose. Serializing the suites does not avoid it, and a delay or a retry in the test would only hide it. I'd expect the suite to pass reliably, so I'm leaving this open in case it shows up elsewhere. Feel free to close it if it does not reproduce outside my setup.

Environment
- OS: macOS 27.2 (26B123)
- Xcode: Version 27.2 (27B52)
- Container: main (16e8d48)
Code of Conduct
  • I agree to follow this project's Code of Conduct
Lingua principale
Swift
Stelle
51.9k
Fork
1.8k
Merge medio
2g 3h
PR unite (30g)
27

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di apple/container

Tutte le issue di apple/container

Issue simili

Altre issue su Swift

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.