Re: Cygwin port of Guile 2.2

[Andy Wingo <wingo@pobox.com>]

Greets,

On Fri 12 May 2017 16:13, Derek Upham <sand@blarg.net> writes:

> Andy Wingo <wingo@pobox.com> writes:
>
>> scm_join_thread isn't actually implemented in terms of
>> scm_i_pthread_join any more.  Probably that's what's going wrong here --
>> and probably that should be fixed to ensure that we actually join the
>> thread.  (Otherwise it would be a memory leak too AFAIU.)  Bcc'ing
>> bug-guile to create a bug for that.
>
> I noticed that scm_join_thread was calling back into Scheme-land.  Are these statements all correct?
>
> - We are using call-with-new-thread underneath the hood.

Underneath the hood of what? :)

> - call-with-new-thread is documented to return a Scheme object from a
> thunk/handler.  Any underlying pthreads should be implementation
> details.

Correct.  In practice call-with-new-thread will create a pthread but I
can imagine circumstances in which it might (in the future) spawn an
auxiliary pthread for some reason, and I wouldn't want to rule that out.

> - The spawned thread sends the Scheme object to the condition variable
> as soon as the user thunk exits.  Any number of operations can happen
> afterwards; the thread is still running in Scheme-land at this point,
> in call-with-new-thread’s wrapping thunk.
> - join-thread waits on the condition variable only.

These are implementation details.  They are correct but probably the
implementation should change to do the scm_i_pthread_join and we should
guarantee that after the join, the thread is really dead.  This is bug
26858.

> So at the end of join-thread we need to add a call to
> scm_i_pthread_join (which we implement in threads.c) to ensure that
> the pthread is completely gone before that join-thread returns.  Is
> that accurate?

Well... yes, but we have to ensure that we call scm_i_pthread_join at
most once.  I think calling pthread_join twice on a thread is
undefined.  So there are some gnarlies here.  Need to fix this.

> Unfortunately, I think the GC threads are going to end up being
> immovable objects in the path to full process-form support.

You can disable marker threads with the GC_MARKERS environment variable,
and the finalization thread should come and go as needed.  Probably this
is not a blocker from your POV.  Signal handling is probably the most
serious issue; perhaps we can avoid the thread somehow, since we handle
signals asynchronously anyway..

Andy