C-g while exiting the minibuffer

[Stefan Monnier <monnier@iro.umontreal.ca>]

I've been trying to track down a performance problem I've been
experimenting for a while now.  Along the way I found that the problem
shows up when frame-focus is redirected (at least, it's a necessary
ingredient, tho not sufficient).

But I also found out that sometimes I have a redirected frame-focus for
no reason.  And I finally yesterday caught the bugger that causes this:
C-h right while exiting the minibuffer: my minibuffer is in a separate
frame, so while in the minibuffer, the frame-focus is redirected (this
is perfectly normal so far).  When exiting the minibuffer, this focus is
supposed to be canceled, but this cancellation can fail:

the cancellation is done implicitly by the restoration of the previous
window-configuration.  The problem being that Fset_window_configuration
does

      /* In the following call to `select-window', prevent "swapping out
	 point" in the old selected window using the buffer that has
	 been restored into it.  We already swapped out that point from
	 that window's old buffer.  */
      select_window (data->current_window, Qnil, 1);
      BVAR (XBUFFER (XWINDOW (selected_window)->contents), last_selected_window)
	= selected_window;

just before

      if (NILP (data->focus_frame)
	  || (FRAMEP (data->focus_frame)
	      && FRAME_LIVE_P (XFRAME (data->focus_frame))))
	Fredirect_frame_focus (frame, data->focus_frame);

IOW just before resetting the frame-focus, we select the
selected-window.  So far this looks innocuous, but select_window (with
a 1 as last argument) calls record_buffer which calls Fdelq which uses
QUIT, so if you happen to hit C-g at the right time, the frame-focus
redirection is not properly canceled.

Now I wonder what should be the best way to fix this.  I can just switch
the above two code blocks, which will fix the immediate problem (a C-g
will not prevent canceling the frame-focus redirected), but I think this
points to a larger problem of hitting C-g while processing unwind forms.
Of course binding inhibit-quit during processing of unwind forms could
be a source of problem in itself (if those forms fail to terminate).

Any comment/idea/suggestion?


        Stefan