To make choice of focus window consistent for each focus method, a number of guidelines should be followed. (For purposes of discussion here, I'm excluding things like the panel and the desktop from "windows". It is technically incorrect to do this, but I'm lazy and "windows" is shorter than something like "normal windows". See the end of the discussion for how these special cases are handled.) The basics are easy: Focus method Behavior click Focus the window on top sloppy When an EnterNotify is received, focus the window mouse Same as sloppy, but also defocus when mouse enters DESKTOP window Note that these choices (along with the choice that clicking on a window raises it for the click focus method) introduces the following invariants for focus from mouse activity: Focus method Invariant click The window on top is focused sloppy If the mouse is in a window, then it is focused; if the mouse is not in a window, then the most recently used window is focused. mouse If the mouse is in a non-DESKTOP window, then it is focused; otherwise, the designated "no_focus_window" is focused However, there are a number of cases where the current focus window becomes invalid and another should be chosen. Some examples are when a focused window is closed or minimized, or when the user changes workspaces. In these cases, there needs to be a rule consistent with the above about the new window to choose. Focus method Behavior click Focus the most recently used window (same as the window on top) sloppy Focus the window containing the pointer if there is such a window, otherwise focus the most recently used window. mouse Focus the non-DESKTOP window containing the pointer if there is one, otherwise focus the designated "no_focus_window". Note that "most recently used window", as used here, has a slightly different connotation than "most recent to have keyboard focus". This is because when a user activates a window that is a transient, its ancestor(s) should be considered to be more recently used than other windows that have had the keyboard focus more recently. (See bug 157360; this may mean that the alt-tab order should also change simultaneously, although the current implementation does not do that.) Also, sometimes a new window will be mapped (e.g. unminimizing a window or launching a new application). Most users want to interact with new windows right away, so these should typically be focused. This does conflict with the invariants for sloppy and mouse focus modes, so this wouldn't be true for a strict-pointer-focus mode. For all other modes (non-strict-pointer-focus modes), there are only two cases in which a new window shouldn't be focused: 1) If the window takes a while to launch and the user starts interacting with a different application, the new window should not take focus. 2) If the window that will appear was not launched by the user (error dialogs, instant messaging windows, etc.), then the window should not take focus when it appears. To handle these cases, Marco compares timestamps of the event that caused the launch and the timestamp of the last interaction with the focused window. (Case 2 is handled by the application providing a special timestamp of 0 for the launch time, which ensures that the window that appears doesn't get focus) If the newly launched window isn't focused, some things should be done to alert the user that there is a window to work with: 1) The _NET_WM_DEMANDS_ATTENTION hint should be set 2) If the new window isn't modal for the focused window, it should appear below the focused window so that it doesn't obscure the focused window that the user is interacting with. 3) If the new window is modal to the focused window, the currently focused window should lose focus but the modal window should appear on top. Additionally, the user may decide to use the keyboard instead of the mouse to navigate between windows (referred to as "keynav"). This poses no problems for click-to-focus (because the same invariant can be maintained), but for sloppy and mouse focus it requires extra work to attempt to handle the INHERENTLY CONFLICTING CONSTRAINTS. Marco does this by having a mouse_mode boolean used to determine which of the two sets of invariants holds. This mode is set according to which method was most recently used to choose a focus window: 1) When receiving EnterNotify events from mouse movement, set mouse_mode to TRUE. 2) When using keynav to choose a focus window (e.g. alt-tab, alt-esc, alt-f2, move-window-to-workspace keybindings), set mouse_mode to FALSE. 3) When handling events that don't choose a focus window but rather need a focus_window chosen for them (e.g. switch-to-workspace keybindings), don't change the mouse_mode and just use the current value. Note that grabs present a special case since they can generate EnterNotify and LeaveNotify events without using the mouse, thus these events should be ignored when the crossing mode is NotifyGrab or NotifyUngrab. THIS MOUSENAV/KEYNAV MODERATION METHOD IS NOT PERFECT--there are corner cases when trying to mix-and-match between mousenav and keynav simultaneously that cause problems; but it appears to be the most reasonable tradeoff and works well in most cases, especially if the user sticks to just mousenav for a long time or just keynav for a long time. Finally, windows of type WM_DOCK or WM_DESKTOP (e.g. the desktop and the panel) present a special case, at least partially due to the lack of decorations. For WM_DESKTOP windows, we only focus them if the user explicitly requests it (e.g. clicks on the window, uses Ctrl-Alt-Tab to navigate to it, uses a keybinding to show the desktop, etc.). For WM_DOCK windows, we do not focus unless we receive a very explicit request (e.g. Ctrl-Alt-Tab or a _NET_ACTIVE_WINDOW message; not normal clicks). To read more about the bugs that inspired these choices: - When a focused window becomes invalid and another should be chosen http://bugzilla.gnome.org/show_bug.cgi?id=135810 - When a new window is mapped http://bugzilla.gnome.org/show_bug.cgi?id=118372 Also, the EWMH spec, especially the parts relating to _NET_WM_USER_TIME - Modal vs. non-modal dialogs that get denied focus when mapped http://bugzilla.gnome.org/show_bug.cgi?id=151996 - Mousenav vs. Keynav in mouse and sloppy focus modes http://bugzilla.gnome.org/show_bug.cgi?id=167545 http://bugzilla.gnome.org/show_bug.cgi?id=101190 http://bugzilla.gnome.org/show_bug.cgi?id=357695 - Not focusing panels http://bugzilla.gnome.org/show_bug.cgi?id=160470 http://bugzilla.gnome.org/show_bug.cgi?id=120100 There were many bugs which had to be fixed to get all the above working; they helped form these policies and/or show the difficulties in implementing this policy (my apologies in advance for producing a list heavily lopsided to what I've done; it's just that these bugs are the ones I'm the most familiar with): bug 72314 ignore LeaveNotify events from grabs bug 82921 focus windows on map bug 87531 only show focus for sticky windows on active workspace (pager) bug 94545 focus window on workspace switch is non-deterministic bug 95747 should ignore EnterNotify events with NotifyInferior detail set bug 97635 sticky windows always keep focus when switching workspaces bug 102665 a window unminimized from the tasklist should be focused bug 107347 focus windows that manually position themselves too bug 108643 focus in MRU order instead of stack order bug 110970 moving a window to another workspace loses focus bug 112031 closing a dialog can result in a strange focus window bug 115650 add _NET_WM_USER_TIME support to gtk+ (see also 150502) bug 120100 panel shouldn't be focused after workspace applet usage bug 123803 need final EnterNotify after workspace switch (see also 124798) bug 124981 focus clicked window in pager only if on current workspace bug 125492 catch the xserver unfocusing everything and fix its braindeadedness bug 128200 focus correct window on libwnck window minimize (see 107681 too) bug 131582 fix race condition on window minimize/close bug 133120 wrong window focused when changing workspaces bug 135024 _NET_ACTIVE_WINDOW messages need timestamps bug 135786 middle-clicking on focused window to lower it should defocus too bug 136581 window minimization vs. activation for mouse focus bug 144900 fix focus choice on "un-showing" the desktop bug 147475 don't lock keyboard on workspace change bug 148364 DEMANDS_ATTENTION support for marco & libwnck (and other stuff) bug 149028 focus-stealing-prevention for marco-dialog (and other stuff) bug 149366 windows denied focus on map occur in wrong order in alt-tab list bug 149543 consistent focus window when unshowing desktop bug 149589 race in focus choice from libwnck messages bug 150271 make sure "run application" dialog gets focused bug 150668 update gtk+ _NET_ACTIVE_WINDOW support bug 151245 application startup notification forwarding (partially rejected) bug 151984 Soeren's idea--backup timestamp when startup notification not used bug 151990 prevent focus inconsistencies by only providing one focus method bug 151996 modal dialogs denied focus should not be lowered bug 152000 fix race on window close followed by rapid mouse movement bug 152004 ways to handle new window versus mouse invariants bug 153220 catch the root window getting focus and reset to default window bug 157360 focus parents of dismissed transient windows in preference to the window that most recently had focus bug 159257 focus the desktop when showing it bug 160470 don't focus panels on click bug 163450 correct highlighting in workspace switcher popup bug 164716 refuse to focus a window with a modal transient, and focus the transient instead bug 166524 avoid new windows being obscured by the focus window bug 167545 mousenav vs. keynav in mouse and sloppy focus modes Addendum on sloppy and mouse focus You may occasionally hear people refer to sloppy or mouse focus modes as inherently buggy. This is what they mean by that: 1) Keynav doesn't maintain the same invariants as mouse navigation for these focus modes; switching back and forth between navigation methods, therefore, may have or appear to have inconsistencies. Examples: a) If the user uses Alt-Tab to change the window with focus, then starts to move the mouse, at that moment the window where the mouse is does not have focus. b) Users expect that a workspace they previously used will not change when the return to it. This means things like window position and stacking order, but also the focus window. Unfortunately, using the original focus window (which would be the most recently used window on that workspace) will sometimes conflict with the invariants for mouse and sloppy focus modes. Users are much more surprised by the invariant being broken than by having the focus window changed (see bug 94545 and probably others), so we maintain the invariant. This only matters when using Ctrl-Alt-Arrow to switch workspaces instead of clicking in the workspace switcher, so this really is a keynav vs mouse issue. Either that, or a windows-are-being-mapped exception. ;-) c) Opening a menu, then moving the mouse to a different window, and then pressing escape to dismiss the menu will result in the window containing the mouse not being focused. This is actually correct behavior (because pressing escape shows that the user is using key navigation to interact with the window containing the menu) but is one of those hard-to-get-right keynav and mouse focus mixture cases. (See bug 101190 for more details) d) Similar to (c), moving the mouse off the menu doesn't immediately focus the window that the mouse goes over, due to an application grab (we couldn't change this and wouldn't want to, but technically it does break the invariant). e) If mouse_mode is off and the user does something to cause focus to change (e.g. switch workspaces, close or minimize a window, etc.) and simultaneously tries to move the mouse, the choice of which window to focus is inherently race-y. (You probably can't satisfy both keynav and mousenav invariants simultaneously...) 2) The sloppy/mouse invariants are often not strictly maintained; for example, we provide an exception to the invariant for newly mapped windows. (Most find that not allowing this exception is confusing) 3) There are an awful lot of little cases to handle to get any focus mode right, even for click-to-focus. Since mouse and sloppy focus have sometimes been hard to even determine what correct behavior is, it is much harder to get them completely right. Plus mouse and sloppy focus users are a minority, decreasing the motivation of window manager implementors to get those focus modes right. 4) Because of -1-, -2-, and -3-, implementations are often buggy or inconsistent and people form their opinions from usage of these implementations. 5) Sloppy focus suffers from a bit of a discoverability problem (for example, I have seen a scientist sit down to a computer for which sloppy focus was in use and take a few minutes before figuring out how window activation worked; granted the layout of the windows in that situation was a bit unusual but it still illustrates that sloppy focus is harder than it should be to figure out). Mouse focus solves this problem; however, people that have experience with other computing environments are accustomed to being able to move their mouse outside the window they are working with and still continue interacting with that window, which conflicts with mouse focus.