GNU bug report logs - #79367
31.0.50; magit-commit sometimes doesn't work if diff-hl-update-async is t

Previous Next

Full log

Message #56 received at 79367 <at> debbugs.gnu.org (full text, mbox):

From: Eli Zaretskii <eliz <at> gnu.org>
To: Dmitry Gutov <dmitry <at> gutov.dev>
Cc: i <at> fuzy.me, sbaugh <at> janestreet.com, 79367 <at> debbugs.gnu.org
Subject: Re: bug#79367: 31.0.50; magit-commit sometimes doesn't work if
 diff-hl-update-async is t
Date: Wed, 03 Sep 2025 14:53:42 +0300

> Date: Tue, 2 Sep 2025 22:45:54 +0300
> Cc: sbaugh <at> janestreet.com, i <at> fuzy.me, 79367 <at> debbugs.gnu.org
> From: Dmitry Gutov <dmitry <at> gutov.dev>
> 
> On 02/09/2025 22:18, Eli Zaretskii wrote:
> 
> >> I do believe that when a change in a long-standing implementation causes
> >> an immediate issue we first consider reverting, especially when there is
> >> no consensus on the approach.
> > 
> > No, we don't.  Reverting is actually the last resort, reserved for
> > plain and clear mistakes, and for changes that cause grave regressions
> > and cannot be fixed soon enough.
> 
> A "mistake" is often in the eyes of the beholder. The lack of consensus 
> is something more objective.

I said "clear mistakes".  Those aren't in the eyes of the beholder,
they are clear to everyone.

> >    . if the process on behalf of which we called
> >      server_accept_connection was not locked to some thread, undo what
> >      make_process did to the new process when it called pset_thread
> >    . otherwise, call set_proc_thread to lock the new process to the
> >      same thread as the one which caused this call
> 
> Sounds about right.

Thanks, will do.

> >> I don't want us to spend a lot of time arguing revert-or-no-revert now,
> >> but could you, Eli, try to decide on a full example of a buggy scenario
> >> which is really solved by thread locking being?
> > 
> > I already described that in so many words.  That is all I can afford.
> 
> We need a fuller scenario. If not with code, then a full use case 
> described in English.

I described some of them, and just now posted a list of my messages in
bug#79201 where I did that.  That's the best I can afford doing at
this time, sorry.

> I'm pretty sure more of your time is being wasted on answering the 
> emails with objections. Regrettably.

Yes.  And I still consider this a worthwhile investment, if as result
we will become on the same page regarding how the relevant code works
and what should be our expectations from it.

> > And please keep in mind that processes were being locked to threads
> > since day one, just incompletely so.  I didn't invent it just
> > yesterday out of the blue: it's the original design of processes with
> > threads.
> 
> IIUC the existing protection did very little in practice, all of these 
> years: it only stopped some thread calling 'accept-process-output' with 
> an explicit reference of a process belonging to another thread.

That's not "very little" at all.  And if some Lisp program called
set-process-thread, it would do exactly what is now done by default.

> Any kind 
> of buggy code would have to try hard to obtain that reference in the 
> first place - so that would almost never happen anyway. The new 
> protections seem a lot more strict.

Yes, it is stricter.  Because that's the intent, as is clear both from
the code and from the documentation, and also because it makes Lisp
programs much easier to reason about.

I also don't quite see a disaster this change is causing to code which
disregarded the documented restriction.  All we had till now is a
single bug report, which clearly indicates that my change was
incomplete, and is easy to fix.  I fail to understand the amount of
excitement due to a change that turns out not to be different from
many other changes we make in development.  Here's one recent example:

  https://lists.gnu.org/archive/html/emacs-devel/2025-09/msg00002.html

This even broke the build, which is clearly a serious regression.  And
the solution was that Mattias quickly installed two followup changes;
case closed.  Good job, Mattias!

> TBC I think we should continue to support locking, and your current fix 
> seems productive in this regard, but whether locking should be by 
> default is not obvious to me - and at least two smart devs who worked 
> with Emacs Lisp threads say no. That seems like grounds to re-evaluate. 
> Even if we reach the current default we'll have documented our 
> conclusions better rather than just saying "this is so".

I'm not saying we cannot revisit the decision to lock processes to
threads.  But given that the original design and implementation
clearly tell us that was the intent, I think we should start from what
we have, and only change it if we decide that the current default is
wrong in too many important cases.  For now all we have is 2 cases:
one was solved yesterday by unlocking the process, the other is the
one discussed here, which uncovered a flaw in the change I installed,
and should be easy to fix.  That's not enough to declare the idea of
locking wrong as the default, let alone bankrupt.  For such a
far-reaching conclusion we need much more data points, of both kinds.
My knowledge of this code and my experience, based on 12 years of
developing, maintaining, and fixing bugs in it, is that letting random
threads read output from subprocesses causes many unexpected behaviors
and thus triggers subtle bugs and non-determinism in seemingly valid
Lisp programs.  You think differently, but please don't dismiss my
opinions on this too quickly, given my familiarity with the code in
this area and the amount of time I spent on it.

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.