Looking at the security of Plasma/Wayland

Our beta release of Plasma 5.5 includes a Wayland session file which allows to start in a Plasma on Wayland session directly from a login manager supporting Wayland session. This is a good point in time to look at the security of our Plasma on Wayland session.

Situation on X11

X11 is not secure and has severe conceptual issues like

  • any client connected to the X server (either remote or local) can read all input events
  • any client can get information about when another window rendered and get the content of the window
  • any client can change any X attribute of any other window
  • any window can position itself
  • many more issues

This can be used to create very interesting attacks. It’s one of the reasons why I for example think it’s a very bad idea to start the file manager as root on the same X server. I’m quite certain that if I wanted to I could exploit this relatively easily just through what X provides.

The insecurity of X11 also influenced the security design of applications running on X11. It’s pointless to think about preventing potential attacks if you could get the same by just using core X11 functionality. For example KWin’s scripting functionality allows to interact with the X11 windows. In general one could say that’s dangerous as it allows untrusted code to change aspects of the managed windows, but it’s nothing you could not get with plain X11.

Improvements on Wayland

Wayland removes the threats from the X11 world. The protocols are designed in a secure way. A client cannot in any way interact with windows from other clients. This implies:

  • reading of key events for other windows is not possible
  • getting window content of other windows is not possible

In addition the protocols do not allow to e.g. position windows themselves, or raise themselves, change stacking order and so on.

Security removed in Plasma

But lots of these security restrictions do not work for us in Plasma. Our desktop shell need to be able to get information of other windows (e.g. Tasks applet), be able to mark a panel as a panel (above other windows) and need to be able to position windows itself.

Given that we removed some of the security aspects again and introduced a few custom protocols to provide window management facilities and also to allow Plasma windows to position themselves. At the moment we have no security restrictions on that yet which gives this functionality to all clients.

We will address this in a future release. There are multiple ways how this could be addressed, e.g. using the Wayland security module library or use systemd in some way to restrict access. Overall I think it will require rethinking security on a Linux user session in general, more on that later on.

Security added in Plasma compared to X11

The most important change on the security front of the desktop session is a secure screen locker. With Plasma 5.5 on Wayland we are able to provide that and address some long standing issues from X11. The screen locks even if a context menu is open or anything else grabbing input. The compositor knows that the screen is locked and knows which window is the screen locker. This is a huge change compared to X11: the XServer has no concept of a screen locker. Our compositor can now do the right thing when the screen is locked:

  • don’t render other windows
  • ensure input events are only handled in the lock screen
  • prevent access to screen grabbing functionality while screen is locked

As a matter of fact the Wayland protocol itself doesn’t know anything about screen locking either. This is now something we added directly to KWin and doesn’t need any additional custom Wayland interfaces.

How to break the security?

Now imagine you want to write a key logger in a Plasma/Wayland world. How would you do it? I asked myself this question recently, thought about it, found a possible solution and had a key logger in less than 10 minutes: ouch.

Of course there is no way to get a client to act as a key logger. The Wayland protocol is designed in a secure way and also our protocol additions do not weaken that. So the key to get a key logger is to attack KWin.

So what can an attacker do with KWin if he owns it? Well pretty much anything. KWin internally has a very straight forward trust model: everything is trusted and everything can access anything. There is not much to do about that, this is kind of how binaries work.

For example as a Qt application each loaded plugin has access to the QCoreApplication::instance. From there one could just use Qt’s meta object inspection to e.g. get to the InputRedirection model and connect to the internal signal on each key press:

<code>void ExamplePlugin::installKeyLogger()
    const auto children = QCoreApplication::instance()-&gt;children();
    for (auto it = children.begin(); it != children.end(); ++it) {
        const QMetaObject *meta = (*it)-&gt;metaObject();
        if (qstrcmp(meta-&gt;className(), "KWin::InputRedirection") != 0) {
        connect(*it, SIGNAL(keyStateChanged(quint32,InputRedirection::KeyboardKeyState)), this, SLOT(keyPressed(quint32)), Qt::DirectConnection);

void ExamplePlugin::keyPressed(quint32 key)
    qDebug() &lt;&lt; "!!!! Key: " &lt;&lt; key;

But Martin, why don’t you just remove the signal, why should any other aspect of KWin see the key events? Because this is just the example of the most trivial exploit. Of course it’s not the only one. If you have enough time and money you could write more sophisticated ones. For example look at this scenario:

KWin uses logind to open restricted files like the input event files or the DRM node. For this KWin registers as the session controller in logind. Now a binary plugin could just send a DBus call to logind to also open the input event files and read all events. Or open the DRM node and take over rendering from KWin. There is nothing logind could do about it: how should it be able to distinguish a valid from an invalid request coming from KWin?

How to secure again?

As we can see the threat is in loading plugins. So all we need to do is ensure that KWin doesn’t load any plugins from not trusted locations (that is not from any user owned locations). This is easy enough for QML plugins where we have the complete control. In fact it’s easy to ensure for any of KWin’s own plugins. We can restrict the location of all of them.

And even more: by default a system is setup in a way that no binary plugins are loaded from user’s home. So yeah, no problem after all? Well, unfortunately not. During session startup various scripts are sourced which can override the environment variables to influence the loading of plugins. And this allows to also use the well known LD_PRELOAD hack. My naive approach to circumvent this issue didn’t work out at all as I had to learn that already the session startup and the PAM interaction source scripts. So your session might be owned very early.

An approach to black list (unset) env variables is futile. There are too many libraries KWin relies on which in turn load plugins through custom env variables. Most obvious examples are Qt and Mesa. But there are probably many more. If we forget to unset one variable the protection is broken.

A different approach would be to white list some known secure env variables to be passed to KWin. But this also requires that at the point where we want to do the restriction the session is not already completely broken. This in turn means that neither PAM nor the session manager may load any variables into the session before starting the session startup. And that’s unfortunately outside what we can do in our session startup.

So for Plasma 5.5 I think there is nothing we can do to get this secure, which is fine given that the Wayland session is still under development. For Plasma 5.6 we need to rethink the approach completely and that might involve changing the startup overall. We need to have a secure and controlled startup process. Only once KWin is started we can think about sourcing env variables from user locations.

So how big is the threat? By default it’s of course secure. Only if there is already a malicious program running in the system there is a chance of installing a key logger in this way. If one is able to exploit e.g. a browser in a way that it can store an env variable script in one of the locations, you are owned. Or if someone is able to get physical access to your unencrypted hard disk, there is a threat. There are easy workarounds for a user: make all locations from where scripts are sourced during session startup non-writable and non-executable, best change ownership to root and encrypt your home location.

Overall it means that Plasma 5.5 on Wayland is not yet able to provide the security I would have liked to have, but it’s still a huge improvement over X11. And I’m quite certain that we will be able to solve this.

October Plasma on Wayland Update: all about geometry

Last month our Wayland efforts made a huge step forward. In KWin we are now at a state where I think the big underlying work is finished, we entered the finishing line of the KWin Wayland porting. The whole system though still needs a little bit more work.

The big remaining task which I worked on last month was geometry handling. That is simplified: moving and resizing windows. Sounds relatively easy, but isn’t. Moving and resizing windows or in general the geometry handling is one of the core aspects of a window manager. It’s where our expertise is, the code which makes KWin such a good window manager. Naturally we don’t want to throw that code out and want to reuse it in a Wayland world.

And that meant: refactoring. The big problem here was that the code in question is also highly X11 specific. Examples are: the sync protocol for synced resizing, window gravity handling, various X11 specific quirks for moving X11 window moving smooth (e.g. only update at end of move if compositing is active, direct updates if not active).

The task was now to separate the X11 specific geometry handling from generic geometry handling so that it can support both Wayland and X11 windows. With this now in place all the geometry handling works also for Wayland windows. This includes move/resize through the alt+f3 menu, but also triggered from the window itself. During move/resize the quick tiling areas are of course triggered correctly, show the outline and snap to the area if released. Resizing is automatically synced to the speed the client supports and that also in a much better way than on X11 (yay for Wayland and double-buffered state!). During move/resize the windows snap to screen borders and other windows.

Unfortunately our checks to ensure that the window titlebar is not moved outside the visible area doesn’t work, because of client side decorations. Yay, everything is awesome! Have I mentioned that client side decorations are a stupid idea because it breaks useful stuff? No, cannot be. Related to that: QtWayland enters move mode if you click the title bar to activate the window. Of course there should be a delay which KWin implements for it’s decorations. After all we have more than 15 years of experience in doing window decorations. Yay, client side decos! Given control to the application! Let’s leave them in charge of window management, what could go wrong? Wohoo! Totally awesome! Broken window management! That’s the way to go! Of course that could be fixed in Qt, but well it doesn’t fix the problem. We have seen this in Chromium years ago, we even had to adjust our quick tiling/maximization behavior because of it. Another fun fact: QtWayland deco has a minimize button which does nothing because it’s not supported in wl_shell protocol. Have I mentioned that client side decos are awesome?

All right, all right, I’ll end the sarcasm mode now. Our geometry handling has a few more very handy features like window packing: you can configure a shortcut to move the window left (or other direction) and it will move to the next window or the screen border. Similar we support growing or shrinking the window to the next window. I was hardly aware of that feature before doing the refactoring, so I thought I should point it out. It can be used from scripting, so should be useful for ideas like poor man tiling or resizing multiple quick tiled windows at the same time. As a nice addition this code is now covered by auto tests.

So with geometry handling in place it’s possible to do real testing and one of my systems (my notebook) migrated from X11 to Wayland. Actually I have been using Wayland on that system already for watching videos since April as it gives tear-free rendering. A clear plus when watching videos.

And this gives us already the outlook for what we will see in the November update: I’ll focus on stabilizing the current state of Plasma/Wayland and fix bugs, fix bugs and fix bugs. My aim is to have a useable early-adopters version ready for the Plasma 5.5 release. It’s looking good, so I’m confident that we will reach that state.

Upgrading libhybris

One of the most important dependencies for our phone project is libhybris. Libhybris is a neat technology to allow interfacing with Android drivers allowing for example to bring Wayland to a device where all we have are Android drivers.

Given that KWin provides a hwcomposer backend which uses libhybris to create an OpenGL context. All other applications need libhybris indirectly to have the Wayland OpenGL buffer exchange work automatically.

When we started the work on the hwcomposer backend we based it on the libhybris version used in Ubuntu (0.1.0+git20131207) as we used Ubuntu as the reference platform. Soon enough we noticed that this version diverged a lot from the upstream version. Lots of recent changes are missing and there are API incompatible changes.

This made working with it difficult. How much time should we invest in investigating issues? Should we write code which we know might break once Ubuntu decides to upgrade libhybris? How well is Wayland integrated in the Ubuntu version given that they don’t need it? If we need help, who to talk to? Ubuntu who will tell us that they don’t know anything about Wayland, or the libhybris devs who might just tell us: use later version?

Furthermore we want other distributions to provide Plasma for the phone. This means they need to provide libhybris. Of course this is difficult if we need to tell them that we need exactly the version used by Ubuntu. And even more it might conflict with other uses. Considering distributions like Mer would have to chose between a libhybris for Plasma and a libhybris for lipstick.

With that in mind we wanted to invest some time on upgrading libhybris in our stack in this release cycle and then fix the issues we were seeing in the stack. Our awesome packagers did the job of creating packages so that I can port KWin against it. And in deed after some hacking I had KWin rendering again. A more difficult task was to get other applications to work as we run into a problem that libwayland-egl does not use the alternatives system. Thus our packagers needed to do some ld tricks to get this worked around. But with that we had a nice rendering system.

A surprise in this exercise was that our input handling code in the hwcomposer backend didn’t compile any more. The code was gone. While that was in the first moment an unpleasant surprise, it soon turned into something wonderful. If that code is not needed at all on an Android powered device it means that we must be able to get libinput to work with it. 400 lines of code deleted and it’s using the shared input stack through libinput. I’m very happy about that!

With that all in place we finally were able to investigate the rendering issues we were seeing. My hope was that just upgrading libhybris would fix the visual tearing, but unfortunately not. While I’m still surprised that it’s possible to get tearing in the first place on Android devices (hey ever heard of things like Atomic Mode Settings, Android?), it at least gives us a vsync event. Unfortunately the only tear-free solution I could find invokes blocking till we get the event. I don’t like that and I think that’s a bad architecture. One can have blocking free and tear free rendering. Our DRM (kernel mode setting) backend can do so with an easy to use API. Really disappointing that the Android stack is in that regard not better than the glx backend. But well at least it’s tear free 🙂

As we now use upstream libhybris I hope to see distributions to pick up the work and provide a Plasma phone spin. I’d love to see an openSUSE phone or a Fedora phone (or any other distribution). Distributions: you can of course ask us on how to integrate 🙂

The return of kwin_gles

Back in 4.x we provided two binaries for KWin: one compiled against OpenGL (kwin) and one compiled against OpenGL ES (kwin_gles). The reason for that is that one can only reasonably link either OpenGL or OpenGL ES and OpenGL ES is only a subset of OpenGL, so one needs to hide the OpenGL calls (especially the OpenGL 1 calls).

With 5.x we were no longer able to provide these two binaries. The reason for that is that OpenGL got “upgraded” in Qt and QtGui itself links either OpenGL or OpenGL ES. To keep the system’s sanity we decided to follow how Qt is compiled. If Qt is compiled with OpenGL support KWin gets compiled with OpenGL support, otherwise with OpenGL ES support.

That’s of course a reasonable design, but it means that it becomes difficult to test the OpenGL ES code paths. One needs a dedicated Qt and all other dependencies compiled against OpenGL ES. Or one needs a nice device like a Nexus 5 with Plasma mobile. This had resulted in breakage already as we were not able to test enough. Such times belong to the past as I have a nice Plasma mobile device to compile and test my KWin on.

But since we introduced OpenGL ES support through a compile time switch, many things have changed. KWin dropped the OpenGL 1 support which means that most of the code which wouldn’t compile with OpenGL ES is just gone. Furthermore we switched to use libepoxy, so we don’t link OpenGL or OpenGL ES at all, but libepoxy which does the right thing for us. With that we are able to remove all the compile time checks. Of course we need runtime checks to ensure that we don’t call functionality which isn’t available on OpenGL ES.

Now with the upcoming 5.5 release we are able to have one binary which serves both OpenGL and OpenGL ES. Note to distributions: the artifact kwinglesutils.so is no longer compiled, please adjust your packaging rules. KWin will use either OpenGL or OpenGL ES depending on what Qt uses.

Given that nowadays it’s also possible to create both an OpenGL and OpenGL ES context at runtime we can also make use of that and introduced a new value for our KWIN_COMPOSE environment variable: O2ES. If that’s specified KWin will use the EGL backend and create an OpenGL ES context. Although it in general is also possible to create an OpenGL ES context through glx we do not support that for simplification. As proof debug output (qdbus org.kde.KWin /KWin supportInformation) from my running KWin instance:

Compositing is active
Compositing Type: OpenGL ES 2.0
OpenGL vendor string: Intel Open Source Technology Center
OpenGL renderer string: Mesa DRI Intel(R) Ivybridge Desktop 
OpenGL version string: OpenGL ES 3.0 Mesa 10.6.8
OpenGL platform interface: EGL
OpenGL shading language version string: OpenGL ES GLSL ES 3.00
Driver: Intel
GPU class: IvyBridge
OpenGL version: 3.0
GLSL version: 3.0
Mesa version: 10.6.8
X server version: 1.17.3
Linux kernel version: 4.2
Direct rendering: Requires strict binding: no
GLSL shaders:  yes
Texture NPOT support:  yes
Virtual Machine:  no
OpenGL 2 Shaders are used
Painting blocks for vertical retrace:  yes

So in a way we have kwin_gles back, it’s different as it’s no longer a dedicated binary, but it’s runtime switchable. For the moment the only way will be the environment variable. I’m reluctant to add a config option as that sounds like quite some chance for breakage.