ghcide hangs during typechecking on a single-core machine or setting `-j1`

[expipiplus1]

Copied from #573 as this happens with just the ghcide command.

I am bumping into this on NixOS with HLS 0.6.0 8.8.4, or ghcide 0.5.0 8.8.4

--verbose doesn't say much interesting aside from [DEBUG] Set files of interest to: [(NormalizedFilePath "/home/alice/Foo.hs". OnDisk)] after Type checking the files`

Here is a reproducer in the form of a nixos test:

import <nixpkgs/nixos/tests/make-test-python.nix> ({ pkgs, ... }: {
  name = "test";
  machine = { ... }: {
    imports = [ <nixpkgs/nixos/tests/common/user-account.nix> ];
    environment.systemPackages =
      [ pkgs.vim pkgs.ghc pkgs.haskellPackages.ghcide ];
  };
  testScript = ''
    # Login
    machine.wait_for_unit("multi-user.target")

    # Create a hie.yaml file and Foo.hs (this happens with and without the yaml file)
    machine.succeed('echo "cradle:\n  direct:\n    arguments: []" > hie.yaml')
    machine.succeed('echo "module Foo where\nfoo = ()" > Foo.hs')

    # Run ghcide
    machine.succeed("ghcide")

    # Record for posterity
    machine.screenshot("ghcide-output")
  '';
})

Either

• Run with nix-build repro.nix and observe the hang
• Start the VM to experiment with $(nb tests/vim.nix -A driverInteractive)/bin/nixos-run-vms
  • logging in with alice/foobar

I suspect that it could be something about the minimality of the environment which causes this?

strace tells me that it's hung in FUTEX_WAIT_PRIVATE, but that's not really surprising

[expipiplus1]

If I set virtualisation.cores = 2 then there is no hang, looks like this is a problem just for single core machines.

[pepeiborra]

I recently discovered this independently. It happens also with +RTS -N1.

It is caused by the Shake thread pool being bound by the number of capabilities. Ghcide runs a "daemon"-like thread to process Shake actions from a TQueue, and a main thread to do the typchecking. If the thread pool is of size 1, the daemon thread will be unable to progress and the typechecking thread will be unable to run.

In general a threadpool should never be of size 1, so I would say this needs to be fixed in Shake. /cc @ndmitchell

[ndmitchell]

shakeThreads controls the number of threads in the Shake thread pool, and also threads should never mutually depend on each other in a dead-locking way since "normal" Shake programs shouldn't be communicating via TQueue. I suggest that you increase shakeThreads by 1 more than the number of threads you really want, which will ensure its at least 2.

[pepeiborra]

Cool, so we need to set shakeThreads to if numCapabilities == 1 then 2 else 0 or simply numCapabilities + 1.
Joe, will you test and send a PR?

[ndmitchell]

I'd go for numCapabilities + 1 - if you're taking control of shakeThreads and want it at a precise value better to make sure there's no chance Shake infers a different number of threads to you. At one point Shake tried to infer the number of processors you had if it looked like you weren't on the threaded runtime, which would mess up any implicitness.

[expipiplus1]

Why the + 1, @ndmitchell? (instead of max 2 numCapabilities)

so

shakeThreads = max 2 (fromMaybe numCapabilities (optThreads options))

[ndmitchell]

If Ghcide runs a mostly-blocked thread then + 1 is appropriate. If it's a mostly-busy thread then max 2 ... is the right thing to do. The code has changed so much since I was last deeply involved in it I've no idea which it is!