@Lukasa : I updated it with ConditionLock. Is this what you had in mind?

This is still blocking a thread from the concurrency pool and then depending on another thread to unblock it. Will that not be a problem? The only way I see around that is to add an async hook to NIOAsyncWriter.

For testing purposes, that blocking will be safe. We will eventually resolve it, and we don't rely on Dispatch to make that happen (it requires the NIO thread to do it), so it'll be ok.

For testing purposes, that blocking will be safe. We will eventually resolve it, and we don't rely on Dispatch to make that happen (it requires the NIO thread to do it), so it'll be ok.

I'm not sure I fully understand what you're saying here, but I assume it's fine then.

I do still have that 'FIXME' line in there, which would probably be good to remove. Well..., this got merged while I was typing, so I'll open a new pull request for that.

Edit: Opened #3059

It appears this pull gets the NIO tests to hang in the Android x86_64 emulator, but weirdly only when compiled and run on macOS, ie the linux CI runs shown there work fine. I don't know if that's a macOS cross-compilation issue or some incompatibility with the macOS emulator triggered by this pull, possibly both. Reverting this pull gets my Android CI runs on macOS to work again.

Let me know what you guys think.

That’s very interesting. Do you have a smaller repro? The OS-specificity makes me think this is probably a Swift bug, but if it’s easy enough to reproduce we may be able to figure out if the test that hangs is just flaky.

A bit tough for me to reproduce since I don't use macOS. @marcprux, any interest in looking into this on your Mac?

When I test against a local Android emulator (with skip android test) on my ARM macOS 15 machine with your released 6.0.3 Android SDK, the NIOAsyncWriterTests tests (which appears to be where your CI is hanging) do all pass:

Test Suite 'NIOAsyncWriterTests' passed at 2025-01-16 09:23:46.416
	 Executed 33 tests, with 0 failures (0 unexpected) in 0.022 (0.022) seconds

However, the test run does ultimately fail due to a leaked promise.

Edit: see next comment.

I also set up an Android CI in a fork of swift-nio and both macos-13 and ubuntu-24.04. In the run, macOS hangs like you are seeing and Ubuntu fails with the same leaked promise fatalError. I tried both with and without applying your patches in the workflow.

I'm not sure why the CI's macOS testing hangs but my local tests complete, but one difference is that the CI's Android emulator is API 29 and mine is API 34. I'll try with some different Android API levels and see if that changes anything.

Another possibility is architecture: my local emulator is running on aarch64, but the CI runs x86_64 (which is why we are running on macos-13: it is the last public macOS image to support x86). I won't have a macOS x86 machine handy, but I can get access to one later on, so I can try that out if you suspect that it could be architecture related.

Here's the leaked promise, although it is likely a red herring:

Test Case 'ChannelTests.testAskForLocalAndRemoteAddressesAfterChannelIsClosed' started at 2025-01-16 09:23:53.419
NIOPosixTests/TestUtils.swift:697: Fatal error: leaking promise created at (file: "NIOPosixTests/TestUtils.swift", line: 697)
Current stack trace:
0    libswiftCore.so                    0x00000075619ad878 _swift_stdlib_reportFatalErrorInFile + 128
1    libswiftCore.so                    0x00000075617cfb64 _assertionFailure(_:_:file:line:flags:) + 244
2    swift-nioPackageTests.xctest       0x00000061dced52f0 <unavailable> + 19673840
3    swift-nioPackageTests.xctest       0x00000061dcf77a40 <unavailable> + 20339264
4    swift-nioPackageTests.xctest       0x00000061dcf77e34 <unavailable> + 20340276
5    swift-nioPackageTests.xctest       0x00000061dcee49ec <unavailable> + 19737068
6    swift-nioPackageTests.xctest       0x00000061dcee4aac <unavailable> + 19737260
7    libswiftCore.so                    0x000000756191a79c <unavailable> + 4884380
8    libswiftCore.so                    0x000000756191b274 <unavailable> + 4887156
9    swift-nioPackageTests.xctest       0x00000061dda7ca48 <unavailable> + 31894088
10   swift-nioPackageTests.xctest       0x00000061dda705b0 <unavailable> + 31843760
11   swift-nioPackageTests.xctest       0x00000061dda7334c <unavailable> + 31855436
12   swift-nioPackageTests.xctest       0x00000061dda754fc <unavailable> + 31864060
13   swift-nioPackageTests.xctest       0x00000061dda7cf88 <unavailable> + 31895432
14   swift-nioPackageTests.xctest       0x00000061dda7bab4 <unavailable> + 31890100
15   swift-nioPackageTests.xctest       0x00000061dda7b4c8 <unavailable> + 31888584
16   swift-nioPackageTests.xctest       0x00000061dd7611d4 <unavailable> + 28635604
17   swift-nioPackageTests.xctest       0x00000061dd76e1b8 <unavailable> + 28688824
18   libXCTest.so                       0x000000755b6ccbcc <unavailable> + 236492
19   libXCTest.so                       0x000000755b6ce564 XCTAssertNoThrow<A>(_:_:file:line:) + 68
20   swift-nioPackageTests.xctest       0x00000061dd7907cc <unavailable> + 28829644
21   swift-nioPackageTests.xctest       0x00000061ddd54300 <unavailable> + 34874112
22   swift-nioPackageTests.xctest       0x00000061dce81c6c <unavailable> + 19332204
23   swift-nioPackageTests.xctest       0x00000061ddd829ac <unavailable> + 35064236
24   libXCTest.so                       0x000000755b6c5c44 <unavailable> + 207940
25   libXCTest.so                       0x000000755b6c50bc <unavailable> + 204988
26   libXCTest.so                       0x000000755b6c5bac <unavailable> + 207788
27   libXCTest.so                       0x000000755b6c5b8c <unavailable> + 207756
28   libXCTest.so                       0x000000755b6c5f00 <unavailable> + 208640
29   libXCTest.so                       0x000000755b6c405c <unavailable> + 200796
30   libXCTest.so                       0x000000755b6c6164 XCTestCase.invokeTest() + 356
31   libXCTest.so                       0x000000755b6c5f48 XCTestCase.perform(_:) + 128
32   libXCTest.so                       0x000000755b6c8a54 XCTest.run() + 164
33   libXCTest.so                       0x000000755b6c706c XCTestSuite.perform(_:) + 164
34   libXCTest.so                       0x000000755b6c8a54 XCTest.run() + 164
35   libXCTest.so                       0x000000755b6c706c XCTestSuite.perform(_:) + 164
36   libXCTest.so                       0x000000755b6c8a54 XCTest.run() + 164
37   libXCTest.so                       0x000000755b6c706c XCTestSuite.perform(_:) + 164
38   libXCTest.so                       0x000000755b6c8a54 XCTest.run() + 164
39   libXCTest.so                       0x000000755b6c40ec XCTMain(_:arguments:observers:) + 312
40   libXCTest.so                       0x000000755b6c42c4 XCTMain(_:) + 40
41   swift-nioPackageTests.xctest       0x00000061ddddf170 <unavailable> + 35443056
42   swift-nioPackageTests.xctest       0x00000061ddddf190 <unavailable> + 35443088
43   swift-nioPackageTests.xctest       0x00000061ddddf1a8 <unavailable> + 35443112
44   libc.so                            0x000000755d8f25a8 __libc_init + 108
Trap 

Actually, scratch all that. It seems to be hanging randomly, on both macOS and Ubuntu runners, irrespective of API level. See the runs at https://github.com/marcprux/swift-nio/actions/runs/12812276673

I'm still not able to get it to hang on my local machine, but I'll keep trying.

Does it always hang on the same test?

No, it seems to vary. Sometimes in NIOAsyncSequenceProducerTests.testYield_whenStreaming_andNotSuspended_andDemandMore , sometimes in NIOAsyncWriterTests.testSetWritability_whenInitial. It's hard to tell exactly because I think there's some output buffering going on (because sometimes the log will just stop at an incomplete line like:

Test Case 'NIOAsyncSequenceProducerTests.testYield_whenS
Error: The operation was canceled.

Yeah, that seems fairly likely. Can you run the underlying test binary directly? Potentially running it under stdbuf -o0 to try to disable output buffering entirely.

On my Android CI, it is entirely consistent: linux CI runs pass every time while macOS runs always hang. See the run yesterday: across two attempts with this pull, I had four macOS hangs and six linux runs pass (ignore the devel macOS failures, that's a separate issue). So much so that today I only reverted this pull for macOS, finagolfin/swift-android-sdk@7476f2be5, and all the runs passed.

@finagolfin What happens if you only revert the test?

I don't think this test is the issue, as it appears to be failing elsewhere in the log, but I will check that and let you know.

Nope, you were right to suspect the test, as simply reverting this test on macOS got the NIO tests to pass again. 😄

Marc, can you confirm?

Marc, can you confirm?

Confirmed: if I remove the test, then the hangs never occurs in the emulator (running on macOS-13 or Ubuntu).

@marcprux, are you disabling all NIOAsyncWriterTests though, as shown in the linked commit above, marcprux/swift-nio@87d02310d? If so, it would be better to just disable this new test, as I did above, to see if this code change affects the older tests too, which you are no longer running.

I'm going to try running the linux-compiled tests on the macOS emulator and vice versa next, to see if that helps narrow down where this strange hang is coming from, as it could be that we're hitting a bug in the Android emulator for macOS.

This could be an emulator bug, but it could also be that we're seeing an ordering issue flushed out by the performance overhead of the emulator. It'd be useful to try to get a "sequence of events" log from your hanging test as well, as concurrent tests are horrible to debug.

Alright, I ran the NIO tests alone in the Android emulator, once swapping the mac-compiled tests into the linux-hosted emulator and vice versa, once keeping the mac-built tests in the mac-hosted emulator and so on, where only the mac version hung.

Looking at the results from the swapped run, the macos-hosted emulator again hung each time, with the linux-compiled tests not finishing for 20 minutes before I killed it (I have a 5 macOS job limit on the OSS free tier of github Actions, so I killed it once it seemed to hang, so other jobs could proceed). I just kicked off a swapped run again that I will let run for longer to make sure it does hang after longer (just killed it again after the macos-hosted emulators all hung for more than an hour and a half).

This means the problem is in the interaction between this test and the macOS-hosted Android emulator, since the linux-compiled test runs fine in the linux-hosted emulator but hangs in the mac-hosted emulator. The mac cross-compile and mac-hosted Android emulator are consistently slower on github Actions, because of greater contention for fewer macOS CI machines or whatever, so either that lack of speed screws up this test or there is a bug in the macOS-hosted Android emulator.

Regarding "a 'sequence of events' log," you want me to spray this test with print statements to figure out what's running when?

Yeah, that was what I was asking for. Presumably this is the result of a specific ordering sequence caused by the slowdown in the emulator.