Use process jobs when spawning subprocesses

[bgamari]

Many toolchain tools written for POSIX systems rely on the exec system
call. Unfortunately, it is not possible to implement exec in a
POSIX-compliant manner on Windows. In particular, the semantics of the
exec implementation provided by the widely-used msvcrt C runtime
will cause process's waiting on the exec'ing process to incorrectly
conclude that the process has successfully terminated when in fact it is
still running in another process.

For this reason, the process library exposes the use_process_jobs
flag to use a more strict (although still not POSIX-compliant) mechanism
for tracking process completion. This is explained in this comment [2].

Unfortunately, job support in the process library is currently quite
broken and was only recently fixed [1]. Consequently, we only enable job
object support for process releases >= 1.6.8.

[1] haskell/process#168
[2] https://github.com/haskell/process/blob/master/System/Process.hs#L399

[phadej]

Distribution/Simple/Utils.hs:689:22: error:
    Not in scope: type constructor or class ‘CreateProces’
    Perhaps you meant ‘Process.CreateProcess’ (imported from System.Process)
    |
689 | enableProcessJobs :: CreateProces -> CreateProcess

[phadej]

Patch looks a lot nicer now.

@bgamari do you need this for Cabal-3.2 / GHC-8.10, i.e. should this patch be backported?

[phadej]

There's small failure in

Distribution/Simple/Utils.hs:711:56:
    `Process.delegate_ctlc' is not a (visible) constructor field name

but I can take care of it, if the (restarted) OSX job turns out happy.

[phadej]

I won't merge this until process-1.6.8.0 is on Hackage, I have no means to verify this patch works when CPP conditionals are satisfied.

[phadej]

Also, is there a way to see difference between use_process_jobs = True and False, i.e. can we write a test that createProcess behaves correctly?

[bgamari]

What action is required here? I believe the process release is out.

Also, is there a way to see difference between use_process_jobs = True and False, i.e. can we write a test that createProcess behaves correctly?

You could spawn a process which execs another image which sleeps for a while and then performs some observable side effect (e.g. writing to a file). If use_process_jobs works correctly then waitForProcess won't return until the process finishes. This isn't perfectly race-free but I'm afraid it's the best we can do.

[phadej]

Merged as part of #6536