Fixed crash on gtk3-client unit selection dialog close with some gtk+
versions.

See bug #19846

(Browse SVN revision 22822)

> (I haven't tried to track down the actual commit and what
> releases it went into).

commit 5e2c23214564f7dcc687fa8467020eeb6b9407a9
Author: John Lindgren <john.lindgren@aol.com>
Date: Wed Mar 13 02:03:45 2013 +0100

treeview: Don't emit cursor-changed in destruction

This is a workaround until we have a good idea on how we want to handle
signal disconnection before/during/after dispose and destroy vfunc
calls.

https://bugzilla.gnome.org/show_bug.cgi?id=671939

So yes, some other change had already fixed the crash problem for our crosser builds before.

Ubuntu and Debian historically diverge in libgtk update timing and history. Current revisions in releases:

Debian Wheezy: 3.4.2
Debian Experimental: 3.8.0
Ubuntu Oneiric: 3.2.0
Ubuntu Precise:3.4.2
Ubuntu Quantal: 3.6.0
Ubuntu Raring: 3.6.4

I don't have tooling to get release history for other distributions installed. I don't know what to do about Wheezy (would need to meet a fairly high set of criteria to get applied), but the patch in bugzilla looks small enough, focused enough, and safe enough that it could probably be released as part of Precise updates, if that meets the local need. As with the tooling, I'm less certain how to arrange this for other distributions.

Gnome bug mentioned in comment #4 is still open, but the related gtk+ code has changed in 3.6, maybe unintentionally fixing this.

Obvious solution for us would be to make gtk+-3.6 the minimum requirement. Problem with that solution is that it would make gtk3-client unavailable to almost everyone (even not-yet-released Debian Wheezy has only 3.4, so it will be another release cycle before Debian has at least 3.6) That's not catastrophic from end-user point of view as we still have equally fully functional gtk2-client, and it will be a long time before freeciv-2.5 (with gtk3-client as default) is to be released anyway. However, it would make development work harder for now.

See also bug #20224 - presumably because of newer gtk+ this is no longer mere gtk+ assert, but crashes the client.

Well, right now I've got the client nearly in pieces (mostly due to gridding - nearly done now, just restoring the spacing and minor tweaks left), but it seems my first comment was on the spot - due to the order of signals being undefined, instead of
fc_assert_ret(pdialog);
it should just be
if (!pdialog) return;

Though I might have broken something else (unrelated to the above fix), as "Select"/"Deselect" seems to cease to work altogether.

There's quite a bit more:
- the original problem: AFAICT, it comes down to cursor-changed callback getting run after row-activated callback, so that usdlg_get(FALSE) returns NULL, as the dialog has already been destroyed
- the second part: "Select" button of this dialog seems to work only while in first tab (I suspect other buttons to be affected too, though row-activated callback works regardless of the tab)

Assertions don't seem to happen in gtk2 client (even if the buttons still fail in the same way) - it seems that either it's a minor problem in gtk3 or treeview is simply less strict there (i.e. I've noticed (IIRC in a different project) that treeselection returned by a treeview seems to be initially NULL in gtk3, while I haven't seen checks for such case in gtk2-based code).