Remote target (for distributed actor) arguments are not properly deinitialized

[ser-0xff]

Remote target arguments reconstructed on the remote side of the distributed actor are not properly deinitialised.
If the remote target argument is a struct with some class instance references then those class instances are never released, which cause a memory leak.

We implemented minimised test
It perform a remote call on the distributed actor with a struct argument having a reference to the ClassValue instance.
Can find the ClassValue instance initialized twice (first time on the caller side, second time on the remote side), but reinitialised only once (on the caller side):

git clone https://github.com/ordo-one/external-reproducers.git
Cloning into 'external-reproducers'...
remote: Enumerating objects: 135, done.
remote: Counting objects: 100% (135/135), done.
remote: Compressing objects: 100% (88/88), done.
remote: Total 135 (delta 49), reused 91 (delta 23), pack-reused 0
Receiving objects: 100% (135/135), 27.13 KiB | 524.00 KiB/s, done.
Resolving deltas: 100% (49/49), done.
aaa % cd external-reproducers/swift/distributed-system-leak 
distributed-system-leak % swift test
Building for debugging...
[5/5] Linking DistributedSystemPackageTests
Build complete! (7.65s)
Test Suite 'All tests' started at 2023-05-11 12:45:03.890
Test Suite 'DistributedSystemPackageTests.xctest' started at 2023-05-11 12:45:03.891
Test Suite 'DistributedSystemTests' started at 2023-05-11 12:45:03.891
Test Case '-[DistributedSystemTests.DistributedSystemTests test1]' started.
Test Case '-[DistributedSystemTests.DistributedSystemTests test1]' passed (0.002 seconds).
Test Suite 'DistributedSystemTests' passed at 2023-05-11 12:45:03.893.
	 Executed 1 test, with 0 failures (0 unexpected) in 0.002 (0.002) seconds
Test Suite 'DistributedSystemPackageTests.xctest' passed at 2023-05-11 12:45:03.893.
	 Executed 1 test, with 0 failures (0 unexpected) in 0.002 (0.002) seconds
Test Suite 'All tests' passed at 2023-05-11 12:45:03.893.
	 Executed 1 test, with 0 failures (0 unexpected) in 0.002 (0.003) seconds
assign<TestActor>: 1
actorReady<TestActor>: 1
resolve<TestActor>
init: 0x0000600002a301d0
recordArgument
remoteCallVoid
init: 0x0000600002a301f0
value=100
deinit: 0x0000600002a301d0
distributed-system-leak % 

We would expect the second ClassValue instance constructed on the remote side would be deinitialised, thus the output would also had a line

deinit: 0x0000600002a301f0

Environment

• Swift compiler version info

distributed-system-leak % swift --version
swift-driver version: 1.62.15 Apple Swift version 5.7.2 (swiftlang-5.7.2.135.5 clang-1400.0.29.51)
Target: arm64-apple-macosx13.0

• Xcode version info
• Deployment target:

distributed-system-leak % uname -a
Darwin mbpro.local 22.3.0 Darwin Kernel Version 22.3.0: Mon Jan 30 20:38:37 PST 2023; root:xnu-8792.81.3~2/RELEASE_ARM64_T6000 arm64

[ktoso]

Thanks let me try to reproduce this

[ktoso]

Thanks, I was able to reproduce this -- we didn't really test with messages being classes as that's a bit unusual but of course it must be supported correctly. I'm investigating this more now.

[ktoso]

rdar://109207043

[ser-0xff]

Thank you for the update.

[ktoso]

We reproduced and have a fix incoming, thank you very much for noticing this.

[hassila]

Super, thanks! Just for our planning (whether we need to workaround this or not), is this going into 5.9 or is it later?

[ktoso]

There's no hard promises with these things but yes we've cherry picked it over to 5.9 as well, it should be able to land there AFAICS.

[ktoso]

More completet fix incoming in:

• [Distributed] Ensure to swift_release during destroying decoded distributed call arguments #65952
• 🍒[5.9][IRGen] Distributed: Destroy loaded arguments after the call #65941

[ser-0xff]

Hi!
We run our minimised test with the new swift toolchain and see there is no leak now.
Thank you for fixing that.