bug#65116: 29.1; query-replace-read-args fails reading second arg in detached minibuf

[Alan Mackenzie <acm@muc.de>]

Hello, Eli and Po.

On Sat, Jan 13, 2024 at 08:34:28 +0200, Eli Zaretskii wrote:
> > Date: Fri, 12 Jan 2024 21:44:11 +0000
> > Cc: 65116@debbugs.gnu.org, Eli Zaretskii <eliz@gnu.org>, acm@muc.de
> > From: Alan Mackenzie <acm@muc.de>

> > On Fri, Jan 12, 2024 at 12:57:42 -0600, Jim Rees wrote:
> > > Well that's a relief. I do have an unusual setup with detached minibuf and
> > > focus follows mouse. There has been a lot of churn in replace.el and frame.c
> > > lately and I keep hoping the bug will go away on its own. I don't really
> > > understand all the focus changes in the code but I do see why they are
> > > necessary.

> > > I have a workaround, I have bound this to a key and use it to re-focus to
> > > the minibuf so I can enter the 'to' text:

> > > (select-frame-set-input-focus (window-frame (minibuffer-window)))

> > > But that requires manual intervention so for now I'm sticking with 28.1.

> > I've been playing with the setup for an hour or two.  It seems that
> > performing some action in the minibuffer (say, M-x auto-revert-mode, but
> > anything will do) causes M-% to work properly.  But then, the moment the
> > mouse leaves the active frame or window (I'm not sure which), M-% no
> > longer works properly, until the next minibuffer action.

> > I know this isn't much help to you, but it should be a help to us,
> > tracking down what's going wrong.

> If this is WM-specific, maybe Po Lu (CC'ed) could help us understand
> what happens here?  Perhaps some message we expect from X is not being
> received in this scenario?

I've got some more info which might be useful.  There are two calls to
read_minibuf (the lowest level of minibuffer access in src/minibuf.c)
during the query-replace processing.

during the first access, the focus gets redirected to the minibuffer at
L812 of minibuf.c, with

  if (!EQ (mini_frame, selected_frame))
      Fredirect_frame_focus (selected_frame, mini_frame);

..  [I don't know what undoes this focus redirection later.]  After the
recursive edit, provided it wasn't aborted with C-g, the code moves the
focus back to the main frame with, at L963,

    call2 (Qselect_frame_set_input_focus, calling_frame, Qnil);

..  In the bug scenario, if this statement at L963 is commented out, the
bug symptoms are no longer evident.

So it seems something in the second read_minibuf call is preventing the
Fredirect_frame_focus at L812 from working.

We can hardly avoid needing to use GDB, here.

-- 
Alan Mackenzie (Nuremberg, Germany).