[RFC] Interface for injection of input events

[Jonas Ådahl]

On Wed, Mar 22, 2017 at 12:23:46PM +1000, Peter Hutterer wrote:
> Hi all,
> 
> This is an RFC for a new interface to generate input events from arbitrary
> clients. Note that this is only a few days old, so **do not** assume this is
> anything more a thought experiment right now. This email is supposed to start a
> discussion and collection of the various points that need to be addressed.
> 
> First: why? There are some commandline tools like xdotool that allow for some
> scripting of desktop interactions. xdotool supports two categories: input
> device emulation and window management.
> 
> This RFC primarily addresses the input device emulation bits but there is
> room for adding window management capabilities. I have a basic repo here:
> http://github.com/whot/woodotool but it doesn't contain much beyond what's
> in this email.
> 
> This will be a discussion of the interface only because the implementations
> are so specific that there is no real code-sharing beyond the interface
> itself. I've talked privately to some of you already, the general mood is
> somewhere around reluctant acceptance.
> 
> So here's a list of talking points:
> 
> == DBus ==
> What we need is basic IPC and not directly Wayland related, DBus provides a
> bunch of extras over the wayland protocol: introspection, ease of
> extensibility, bindings, etc. Also, Mir.
> 
> == Functionality-based interfaces ==
> We need a mix of capabilities and features, not all of which will/should be
> available to all clients. Right now, I have two for devices:
>  org.freedesktop.WoodoTool.Keyboard (Press, Release)
>  org.freedesktop.WoodoTool.Mouse (Press, Release, MoveRelative, MoveAbsolute)
> Compositors can implement one, both, either, etc. For future extensions,
> having a Touch interface, Joystick, or whatever is deemed useful is
> technically trivial.

I wouldn't say Joystick, at the current state, is trivial at all, as
it'd require a compositor to have a setuid uinput process doing this.
There hasn't been a consensus on whether to add a joystick/gamepad
protocol or introduce a lower (evdev) level protocol to Wayland, and
AFAIK reading the evdev node directly is how it's done on X11.

> 
> There's a manager interface too but that's a technical detail, see the repo
> for more details.
> 
> == Control of the event stream ==
> The events are coming in through a custom interface, so it's relatively
> trivial to ignore events based on context, e.g. ignore fake key events while
> the screen is locked. Any uinput-based solution would lack this context.
> 
> == Authentication/Identification ==
> The goal is to filter clients based on some white/blacklist, so that e.g.
> xdotool can access this interface but others cannot.
> 
> This is a big ¯\_(ツ)_/¯ for now, I don't now how to do this reliably.
> It's trivial to do per user, but per-process is difficult. DBus filters
> are largely limited to per-users. It's possible to get the process ID of a
> sender but going beyond that is unreliable (kernel doesn't guarantee comm
> being accurate).
> 
> Requiring applications to bind to a bus name merely restricts them to being
> a singleton, there is no guarantee the application that binds
> org.freedesktop.org.WoodoTool.auth.xdotool is actually xdotool.
> 
> The option that comes closest so far is some pre-shared key between
> compositor and application. That would need to be worked into the API, but
> it also relies on all participants to keep the key encrypted in memory and
> the various configuration files.
> 
> So it's not clear whether we can do anything beyond a basic on/off toggle on
> whether to allow events from fake input devices. Debatable if such a crude
> mechanism is useful.
> 
> 
> Either way, this is a problem that *must* be solved but not necessarily one
> that affects the API itself (beyond what is required to make it
> technically feasable, e.g. passing cookies around)

This could be left up to flatpak et.al, couldn't it? Coming up with a
authentication mechanism that likely can be worked around without proper
sandboxing doesn't sound relaible. CC:ing Alex regarding this.

> 
> == Isolation of devices ==
> Compositors should create separate virtual input devices for each client so
> one client can't mess with the state of another one or even detect if
> there's another one active. Plus we get to re-use all the existing code that
> merge state from different (physical) devices. This just makes the actual
> device handling implementation trivial.
> 
> It gets a bit more difficult if we want to have this per-seat though. Seat
> is a wayland concept so we could fold that in, but then you're leaking the
> seat configuration. So I'm not 100% sure we even want this, or whether we
> should just make sure it can be retrofitted in the future.

One option, that I think has been experimented with in terms of remote
desktoping, is to create new seats. I.e. a scripting client can
create its own seat, and create devices on that. This could be added to the
protocol already without requiring multiple seat support, by having a
GetDefaultSeat() thing that one creates devices from, and then maybe
add two optional methods: GetSeat(name) CreateNewSeat() for the other
cases.

> 
> == Use for testing ==
> Came up several times, I'm not yet convinced this is a good compositor testing
> interface beyond basic input. A good testing interface likely requires
> something more compositor-specific where more of the window state is made
> available. But it could work as an interim solution.

I suppose it could alreay be useful for fullscreen application testing,
where the position is guaranteed.

> 
> == Input coordinate handling ==
> Keys and mouse buttons are trivial unless we want custom focus (we don't).
> Relative coordinates are ok too, absolute ones are trickier because they
> rely on screen layout not available to the client.
> 
> So the specification needs to include a behaviour we can stick to forever,
> something like "in pixels as measured from the logical top-left of the
> top-left-most screen" etc. Not difficult per se, but this stuff is usually
> prone to corner-case unhappiness.

Does this mean we'd need to add introspection type of API to get the
output layout? Otherwise it'd be hard to script "click bottom right
corner" etc.

> 
> == Input coordinate filtering ==
> One use-case that was mentioned to me was effectively "intercept button X
> and send key events 'abc' instead". This is not possible with the current
> proposal and I'm not sure it can be without going overboard with
> specifications. It *may* be possible to provide some global hotkey hooks I
> have not come up with a good solution to that.
> 
> == Window management ==
> This is beyond the current suggestion but also where it gets both
> interesting and difficult.
> 
> I have a sample org.freedesktop.WoodoTool.Desktop interface that would send
> Edge signals to clients. The compositor decides when these are triggered,
> the client can react to these things with custom commmands.
> 
> But this is where we get into the proper scripting territory and that's also
> where the opinions will diverge quite quickly.
> 
> For example, xdotool supports things like "search for window with name 'foo'
> and activate it". Implementing this is ... tricky. Now we need some
> definition of what a window classifies as and how windows are sorted within
> a process - compare multiple GIMP windows with multiple gnome-terminal
> windows for example. Not specifying exactly what order we return leaves us
> open to behavioural dependencies which may break in the future.

It's especially tricky since this API would deal with multiple window
systems at the same time, while hopefully making that opaque to the
scripting client. There are some similarities between Wayland and X11
(e.g. WM_CLASS and xdg shell app id), and probalby Mir too, but X11
being X11, I suspect there might wierd scenarios with interesting window
combinations making up a real window.

Jonas

> 
> In other words, an interface to search for windows of a given application
> is technically feasable but extremly hard to get right.
> 
> Anyway, with the separation of interfaces this is not something we need in
> the first iterations. But one legitimate question is whether just an
> implementation for virtual input devices is sufficient or whether it's
> largely pointless without any additional window-management capabilities.
> 
> == Implementation ==
> Because of what it does, there is no real code-sharing between compositors -
> they would just use their own intrastructure to hook to the dbus interface
> and create virtual devices. On the client-side it's much the same thing,
> binding to dbus is trivial.
> 
> 
> So 99% of the work here would be to define the interface and get everyone to
> agree on it.
> 
> Any comments or suggestions?
> 
> Cheers,
>   Peter