Re: Concurrency, again (was: Emacs Lisp's future)

[Eli Zaretskii <eliz@gnu.org>]

> From: John Wiegley <jwiegley@gmail.com>
> Cc: Toon Claes <toon@iotcl.com>,  lars@nocrew.org,  emacs-devel@gnu.org
> Date: Sun, 09 Oct 2016 19:59:09 -0700
> 
> >>>>> Eli Zaretskii <eliz@gnu.org> writes:
> 
> > There's already a concurrency branch in the Emacs repository. It needs some
> > work before it could be merged to master, so if there are people here who
> > wants this badly enough, I suggest that they continue work almost done by
> > Tom Tromey, who developed that branch, instead of starting anew.
> 
> I think it needs more than just a little work. I spoke to Tom on the phone
> last year, and we both agreed it's not a foregone conclusion that that branch
> represents the right way of approaching concurrency for an application like
> Emacs.

We will never know if we don't try it.  We will also never collect the
experience we need to define "the right way" of approaching
concurrency in Emacs, if what's already on the branch is not good
enough.  Meanwhile, chances of someone coming aboard with the kind of
experience and intimate knowledge of Emacs and concurrency that even
approaches that of Tom, and motivated enough to do that kind of
development using some other model, are non-existent.

IOW, saying that the concurrency branch is not good for us without
trying to use it seriously is tantamount to saying that Emacs will
never get any concurrency features at all, and all this sub-thread is
yet another example of issues which never advance anywhere beyond mere
talking about them.  That's too bad, IMO, because I think something
like that could open opportunities for important new features.

> In general, users don't want programmatic concurrency.

AFAIK, the concurrency branch offers programmatic concurrency only as
an opt-in option.  Some concurrency happens automatically without any
code Lisp programmers need to write.  So I see no problem here.  What
we need is for creative Lisp programmers to take this model and
explore how it could be applied to solving real-life problems, with or
without programmatic control of the concurrency, see what are its
limitations and advantages, and generally collect the kind of
experience we sorely need to advance Emacs in that direction.

> I fear we'd be debugging subtle interaction issues for the rest of
> our lives if we just merged Tom's branch in today. It implements a Java-style
> form of concurrency, dependent on mutexes and locking to achieve coherency,
> which is fiendishly difficult to get right, even if you confine its use to
> just a few small tasks. The hardest bug I ever debugged in my life was a
> concurrency bug; the second hardest wasn't even close (and thankfully, I
> didn't try to solve it at the same time).

Did you look at the code, though?  If not, I suggest you do.  The
model is very simple, and its implementation is also simple.  Suffices
it to say that at the time I could port that to MS-Windows and debug
the few problems in a matter of few days work.  When I finished, the
only known problem I left unsolved was the one described here:

  http://lists.gnu.org/archive/html/emacs-devel/2013-08/msg00803.html

and in a followup:

  http://lists.gnu.org/archive/html/emacs-devel/2013-09/msg00009.html

For those who want some background, I pointed to long discussions of
the related issues here:

  http://lists.gnu.org/archive/html/emacs-devel/2015-11/msg00068.html

> I'd much rather we re-examine the goals we want to achieve with concurrency,
> an then ask if there are other ways to get there besides, well, concurrency.
> For example, if we had a lightweight forking mechanism with transparent
> communication between sub-ordinate processes (think async.el, but in C and
> highly efficient), I think we could achieve what users want, without the
> downsides developers don't want. Even the popular web browsers are moving in
> this direction, since it gives a similar effect to threading, but without as
> many of the downsides; or take the popularity of the Actor model, used to
> reduce the coherency problem down to just mailboxes.

Sorry for being blunt, but without someone actively working on this,
this is just talk.  Who do we have on board who can implement this in
C in a way that will work on more than a single platform?  On some
important platforms, forking is anything but "lightweight", btw.  And
if async.el is any evidence, it's fiendishly hard to debug and even
understand.  E.g., it took me some time to understand where to plug
the necessary coding stuff, back when we had problems with debbugs
clients related to non-ASCII characters.

In general, debugging processing that is distributed between several
processes is much harder than debugging a single multi-threaded
application.  Since Emacs is mostly in maintenance mode, this aspect
should IMO be significant in any development decision we make.

> The reason I love Haskell for its STM concurrency (software transactional
> memory) is that it makes certain classes of problems impossible to express. I
> believe we would need a mechanism like that for Emacs Lisp, so no one ever has
> to hunt down cyclic mutex locks, or reference counts, or why two operations
> that should be atomic aren't. I'd rather have a single-threaded Emacs for a
> quite a while longer before inviting these sorts problems into our lives.

But we don't even know whether these problems are relevant to what Tom
implemented in the concurrency branch.

We've gone a long way since this issue was first brought up.  We've
changed our sources in significant ways to support concurrency: all
those BVAR and KVAR macros that are all around the C sources were
introduced for that very purpose.  Likewise, the globals.h header
file, and the fact that each variable exposed to Lisp is a member of
some C struct -- all this was for supporting concurrency.  The code
for this is written, debugged, and is only a little ways from being
ready for prime time.  I'm sorry, but in these conditions discarding
all that work without even giving it a serious try makes very little
sense to me.  If we indeed don't intend to use the concurrency branch,
how about removing all that BVAR/KVAR ballast from our sources, to put
our money where our mouth is?