Proposal to add 'raise' and 'lower' to xdg-shell in wayland-protocols

Mon Dec 12 14:11:16 UTC 2016

Hi,

I put together an initial proposal for this at
https://patchwork.freedesktop.org/series/16642/

Please take a look!

Thanks,

Adam

On Thu, Dec 8, 2016 at 11:08 AM, Adam Goode <agoode at google.com> wrote:

> On Thu, Dec 8, 2016 at 10:42 AM, Mike Blumenkrantz <
> michael.blumenkrantz at gmail.com> wrote:
>
>> Allowing clients to raise/lower has nothing to do with trust and
>> everything to do with control. If clients can affect stacking, they can
>> affect the compositor's window placement policy. Wayland is all about
>> having the compositor be in absolute control of this to the point of
>> clients having the perception of being sandboxed, hence why there is no
>> method for windows to set/request/know their position on screen.
>>
>> I have no plans at this time to r-b any wayland-protocols addition which
>> allows direct client manipulation of window stacking, regardless of how
>> cleverly written it may be.
>>
>> Specifying CSD titlebar regions is just going to lead to even more
>> complexity; if you're dead set on this, I'd suggest an extension which just
>> notifies the compositor when the titlebar has been clicked; the compositor
>> can then choose to take action based on its policies.
>>
>>
> This would be equivalent to renaming the current xdg-shell 'move' request
> to 'titlebar_grab' (and 'resize' to 'edge_grab'), and removing the
> description of the actual move and resize semantics from the protocol
> description. I believe it is essentially what Bill Spitzak initially
> proposed.
>
> If we want to have the semantics of server-side decorations (where the
> compositor is completely in control of windows), but allow clients complete
> control over the appearance of such decorations, then we should make that
> split clear and talk of "decorations" in the protocol, not "actions".
>
> This would allow a clear delegation of policy to the compositor without
> material change to the protocol. The use of "grab" makes it clearer that
> the compositor is in charge of deciding what to do (it is less specific
> than "move" or "resize").
>
> You could also fold these two requests together into just "grab", change
> the "resize_edge" enum to be called "grab_region", and add a new entry for
> "titlebar".
>
>
> Adam
>
>
>
>
>>
>> On Tue, Dec 6, 2016 at 3:51 PM Adam Goode <agoode at google.com> wrote:
>>
>>> On Mon, Dec 5, 2016 at 11:11 AM, Mike Blumenkrantz <
>>> michael.blumenkrantz at gmail.com> wrote:
>>>
>>> To echo Jonas's comments, I'm also strongly opposed to adding window
>>> stacking manipulation to the xdg-shell protocol. It's already a mess
>>> handling windows which try to raise/focus themselves in X11, this is not an
>>> issue I want to handle under Wayland.
>>>
>>> EFL also has functionality in the toolkit for this on client-side, but
>>> it does nothing on Wayland and we discourage its use under X11.
>>>
>>> On Fri, Dec 2, 2016 at 11:25 AM Adam Goode <agoode at google.com> wrote:
>>>
>>> On Fri, Dec 2, 2016 at 1:29 AM, Jonas Ådahl <jadahl at gmail.com> wrote:
>>>
>>> On Thu, Dec 01, 2016 at 09:28:11AM -0500, Adam Goode wrote:
>>> > On Thu, Dec 1, 2016 at 12:56 AM, Jonas Ådahl <jadahl at gmail.com> wrote:
>>> >
>>> > > On Wed, Nov 30, 2016 at 12:04:33PM -0500, Adam Goode wrote:
>>> > > > Hi,
>>> > > >
>>> > > > See https://bugzilla.redhat.com/show_bug.cgi?id=1349225 and
>>> > > > https://bugzilla.gnome.org/show_bug.cgi?id=767967
>>> > > >
>>> > > > When using Client Side Decorations, toolkits cannot bind raise or
>>> lower
>>> > > to
>>> > > > user actions. This binding is traditionally used in the "middle
>>> click
>>> > > > titlebar to lower" action, which no longer works with CSD on
>>> Wayland.
>>> > > > Additionally, when click-to-raise is disabled, a click on a CSD
>>> titlebar
>>> > > > will not raise the window.
>>> > > >
>>> > > > I would like to add 'raise' and 'lower' to xdg-shell. These
>>> requests will
>>> > > > take no arguments, similarly to 'set_maximized' (which is commonly
>>> bound
>>> > > to
>>> > > > double-click titlebar).
>>> > >
>>> > > A client should not be able to raise itself on demand like that.
>>> Usually
>>> > > when raising, what they actually wanted to do is get attention
>>> because
>>> > > something happened, and that is what an API is supposed to do. I
>>> think
>>> > > the last time this was discussed it was referred to as "present" or
>>> > > something. GTK+ have a private protocol for this until we have
>>> something
>>> > > else.
>>> > >
>>> > > Regarding 'lower', any reason why this cannot be made a compositor
>>> side
>>> > > modifier->middle-click kind of thing? It'd work on the whole window
>>> and
>>> > > it'd work on all clients without any need for any protocol. There has
>>> > > also been discussions about having a protocol for specifying a
>>> "window
>>> > > title area" kind of thing, which the compositor can handle with
>>> special
>>> > > care would so be needed.
>>> > >
>>> > >
>>> > The compositor side modifier->middle-click does work this way in mutter
>>> > today. Having support for bare clicks with special meaning in certain
>>> > regions is the more subtle case. What this means today is that if you
>>> > configure your compositor with no-raise-on-click (for the traditional X
>>> > behavior), then windows with CSD don't get raised when decorations are
>>> > clicked (unless you hold a modifier key).
>>> >
>>> > The idea of having a special title-bar area known to the compositor
>>> seems
>>> > worth pursuing, but that could be abused by clients. CSD blended the
>>> lines
>>> > between client and compositor roles, and X was happy to be permissive
>>> with
>>> > that situation. I'm glad Wayland is non-permissive by default, but it
>>> does
>>> > make things tricky.
>>>
>>> Right, and since the lines are now comparably blurry, I think it makes
>>> sense to move away from clients doing window management when we have the
>>> chance. Personally I'd vote for getting rid of the
>>> "middle-click-on-title-lowers" binding in mutter X11 SSD paths rather
>>> than adding it to GTK+.
>>>
>>>
>>> Unfortunately, it is in GTK+ for a while now. We have duplication
>>> between Mutter and GTK+, CSD required GTK+ to directly implement these
>>> features. Example: https://github.com/GNOME/gtk/blob/master/gtk/gtkwin
>>> dow.c#L1372
>>>
>>> This duplication is unfortunate. Hopefully it could be resolved someday.
>>>
>>>
>>>
>>> >
>>> > I would really like to avoid having direct ways for a client to
>>> > > interfere with the window stacking, and especially not ones that
>>> require
>>> > > round trips.
>>> > >
>>> > >
>>> > As an alternative, what do you think of a raise/lower protocol that
>>> > required evidence of a click to trigger? The client could decide on
>>> its own
>>> > which buttons did what on which region of its surface (necessary for
>>> CSD),
>>> > but prove to the compositor that a click of some kind actually
>>> occurred.
>>> >
>>>
>>> There is still no reason why a client needs to have the ability to raise
>>> itself when it can communicate what it actually wanted. IIRC the
>>> "present" proposal used input event serials (i.e. allowed the compositor
>>> to match against a click/touch/...) at its discretion.
>>>
>>>
>>>
>>> I will have to look at the "present" proposal. Thanks for your comments.
>>>
>>>
>>>
>>> Jonas
>>>
>>> >
>>> >
>>> > > >
>>> > > > Is there any objection to adding these to xdg-shell, or should I
>>> > > > investigate another solution?
>>> > > >
>>> > >
>>> > > xdg-shell is in a state where I don't think we should add anything
>>> that
>>> > > is not very crucial until we have managed to declare it stable. A
>>> thing
>>> > > like lower/raise, which has been discussed plenty of times and is on
>>> the
>>> > > more controversial side, should not delay any stabilization of
>>> > > xdg-shell. With that said, things can be added as separate
>>> extensions in
>>> > > the mean time.
>>> > >
>>> > >
>>> > > Jonas
>>> > >
>>> > > >
>>> > > >
>>> > > > Thanks,
>>> > > >
>>> > > > Adam
>>> > >
>>> > > > _______________________________________________
>>> > > > wayland-devel mailing list
>>> > > > wayland-devel at lists.freedesktop.org
>>> > > > https://lists.freedesktop.org/mailman/listinfo/wayland-devel
>>> > >
>>> > >
>>>
>>> _______________________________________________
>>> wayland-devel mailing list
>>> wayland-devel at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/wayland-devel
>>>
>>>
>>>
>>> So I looked a bit at the present proposal from some time ago. It is not
>>> exactly what I had in mind, but not too far off.
>>>
>>> Let me try to summarize what the state is here and possible future
>>> outcomes. Please correct me if I get any details wrong.
>>>
>>> Here are my current assumptions/observations:
>>> A1. Client-side decoration (CSD) is here to stay. At least GTK+ and
>>> Chrome use this to put widgets into the titlebar area. I assume Qt and EFL
>>> (and others) do this as well, but I haven't looked in detail.
>>> A2. CSD is used to create non-rectangular areas within titlebars that
>>> respond to user action. (Chrome tabs are not rectangular, and the titlebar
>>> sits behind them.)
>>> A3. Users should be able to bind 'raise' and 'lower' to mouse events in
>>> window titlebars.
>>> A4. Wayland developers would prefer to have all window manipulation
>>> events (maximize, move, resize, raise, lower) be exclusively in control of
>>> the compositor, preferably with no round trip to clients.
>>> A5. Wayland developers are very concerned about clients abusing raise
>>> and lower.
>>>
>>> One way to fulfill A4/A5 would be to do away with CSD, but A1 doesn't
>>> let us. So, as a compromise, we get these requests in xdg-shell:
>>> - show_window_menu
>>> - move
>>> - resize
>>> - set_maximized
>>> - set_minimized
>>>
>>> These requests allow the compositor to trust clients in requesting
>>> certain actions. But because of A5, raise and lower are not a part of
>>> xdg-shell. This is in conflict with A3.
>>>
>>>
>>> Some quick (but not great) ways forward:
>>>
>>> - Eliminate A1. Unlikely to happen. But then we could switch to giving
>>> full control back to the compositor and allow raise/lower to just work as
>>> it used to before CSD.
>>> - Eliminate A3. This is the simplest option, but the least satisfying
>>> from my perspective. I'll assume we're going to keep this requirement,
>>> since GNOME has supported this for a very long time.
>>> - (I am not going to propose relaxing A4 or A5 here.)
>>>
>>>
>>> Assuming a non-trivial solution, the way forward is to continue to build
>>> a protocol to simulate more of non-CSD behavior in a CSD world. xdg-shell
>>> covers a lot of this. It mostly comes down to how much the compositor is
>>> willing to trust clients. Here is my understanding of this currently:
>>>
>>> examples where the compositor mostly trusts clients:
>>> - set_maximized
>>> - set_minimized
>>> - close
>>> - set_fullscreen
>>> - configure
>>>
>>> These can be initiated by clients without any user interaction. The
>>> compositor has no way to know if a user triggered these actions, nor does
>>> it care. It can always apply some policy before acting.
>>>
>>>
>>> examples where the compositor asks for some verification:
>>> - show_window_menu
>>> - move
>>> - resize
>>>
>>> In these cases, the compositor requires a serial and other info to
>>> enforce that a user actually clicked or touched, to avoid abuse of grabbing.
>>>
>>>
>>> Now we are at raise and lower. By A5 above, these are sensitive requests
>>> and should have scrutiny. How sensitive? If they are no more sensitive than
>>> a grab, we could use the same precautions as move and resize and require
>>> the compositor to supply a valid serial and other information to the
>>> compositor to verify a user's intent. If this is acceptable, I can prepare
>>> some proposals that include serial and seat.
>>>
>>>
>>> If raise and lower are even more sensitive than move and resize (or if
>>> it is not desired to do it this way), then clients could tell the
>>> compositor the set of pixels that comprise each part of the CSD, and let
>>> the compositor have full control. This would allow us to avoid round trips
>>> for even move and resize, but (because of A2 above) might require a
>>> something like another surface to be sent to the compositor whenever the
>>> client changes in such a way that changes which pixels belong to the title
>>> bar. There is probably some way to do this more cleverly, but something
>>> like a bitmask would be needed.
>>>
>>>
>>>
>>> That summarizes my thoughts here. Hopefully I've captured most of the
>>> issues involved. Thanks for reading this far, comments welcome.
>>>
>>>
>>> Thanks,
>>>
>>> Adam
>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/wayland-devel/attachments/20161212/f9e5d336/attachment-0001.html>