Re: Cygwin port of Guile 2.2

[Derek Upham <sand@blarg.net>]

Andy Wingo <wingo@pobox.com> writes:

> I guess the new thing being the finalizer thread?  In theory before the
> fork, Guile will stop the finalizer thread.  If that's not working, that
> could be it.
>
> Another option would be to disable fork() but only if Guile is built
> with threads.  Actually this is probably better -- I bet there are Guile
> people that expect to be able to fork on Cygwin and would be OK if their
> Guile had no threads.
>
> Please feel free to push to master once there are adequate comments.
> Someone will come later and want to enable fork() on Cygwin with threads
> and will need to know why it's disabled and what they could do to fix it
> :)

There’s the finalizer thread, but there’s also the signal delivery
thread.  The “sigaction” code ensures that signal delivery thread is
running.  And note that “primitive-fork” is supposed to display a
warning if you are forking with outstanding threads, but it explicitly
ignores the signal delivery thread during its check.

I have found that what actually hangs after a fork are the mutexes
supporting the threads: they are kernel-level resources, referenced by
ID, and end up being shared between parent and child.

I’m baking a fix that wipes out user signals before calling
“primitve-fork”, and then restores them separately in the parent and
child processes.  Doing so gives them different mutexes.  It required
some support in scmsigs.c to:

- Have “restore-signals” return a description of the signal handlers,
  that we can use to restore those handlers.
- Have “restore-signals” close the signal pipe (which kills the signal
  delivery thread).
- When closing the signal pipe, wait for the signal delivery thread to
  exit.

I don’t think there’s any safe way to restore the finalizer thread and
support SCSH-style (begin ...) process forms.  Shutting down the
finalizer thread is the best we can do.

All of this is on GNU/Linux, of course, not Cygwin.

Derek

-- 
Derek Upham
sand@blarg.net