CVE-2021-0928, writeToParcel
/createFromParcel
serialization mismatch in android.hardware.camera2.params.OutputConfiguration
This is exploit using that vulnerability for privilege escalation from installed Android app into Android Settings app (or any other app installed app could send to <receiver>
declared in AndroidManifest.xml
, privilege escalation by sending to <activity>
was possible too, although not presented here)
I've found issue originally on Android 12 Developer Preview 3
Exploit version present in this repo works on Android 12 Beta 2 and 3
Vulnerability was fixed in first official Android 12 release
Writeup below was originally written for Google for consideration of this report as complete exploit chain
At time of writing Android 12 was not available in AOSP (Android Developer Preview/Beta releases are not open source)
Most of IPC on Android is done through class called Parcel
Basic usage of Parcel is as following:
Parcel p = Parcel.obtain();
p.writeInt(1);
p.writeString("Hello");
Then Parcel
is sent to another process through Binder
. Alternatively for testing one can call p.setDataPosition(0)
to rewind parcel to beginning position and start reading:
int a = p.readInt(); // a = 1
String b = p.readString(); // b = "Hello"
It should be noted that parcel internally holds position from which reads are performed. It it responsibility of Parcel
class user to ensure that read*
methods match previously used write*
methods, otherwise subsequent reads will be from wrong positions in buffer
Parcel also provides ability to write custom objects, preferred way to do so is by implementing Parcelable
interface
Here is example implementation of Parcelable
interface (irrelevant code removed, WindowContainerTransaction
class is used in exploit as part of gadget chain, however there isn't anything wrong with it)
package android.window;
public final class WindowContainerTransaction implements Parcelable {
private final ArrayMap<IBinder, Change> mChanges = new ArrayMap<>();
private final ArrayList<HierarchyOp> mHierarchyOps = new ArrayList<>();
private WindowContainerTransaction(Parcel in) {
in.readMap(mChanges, null /* loader */);
in.readList(mHierarchyOps, null /* loader */);
}
@Override
public void writeToParcel(@NonNull Parcel dest, int flags) {
dest.writeMap(mChanges);
dest.writeList(mHierarchyOps);
}
@NonNull
public static final Creator<WindowContainerTransaction> CREATOR =
new Creator<WindowContainerTransaction>() {
@Override
public WindowContainerTransaction createFromParcel(Parcel in) {
return new WindowContainerTransaction(in);
}
};
}
As can be seen above, writeToParcel()
method is used during writing. Then while reading CREATOR.createFromParcel()
factory method is called. It is responsibility of Parcelable
implementation to ensure that createFromParcel
reads same amount of data as was written by writeToParcel
, otherwise all subsequent reads from that Parcel
will read data from wrong offset
Such class can be written to/read from Parcel
through:
- Directly calling
obj.writeToParcel(parcel, 0)
/obj = WindowContainerTransaction.CREATOR.createFromParcel()
, this is often used when type of class is known, for example whenParcelable
has field with differentParcelable
or in code generated by AIDL when defined RPC method hasParcelable
as argument - Through
Parcel.writeParcelable
/readParcelable
.writeParcelable
first writes name of class and then callswriteToParcel
method fromParcelable
interface.readParcelable
reads written class name, finds class with that name in providedClassLoader
orBOOTCLASSPATH
if null was provided. Once class is found it's static fieldCREATOR
is used to obtainParcelable.Creator
instance which is factory used to read that class. It is important to note that whenreadParcelable
method is used it can read anyParcelable
available in class path as name of object to be created is read from sameParcel
readParcelable
is used by many otherParcel
methods, for examplereadList
seen in example above reads elements throughreadValue
, which is most generic method of transferring objects inParcel
and one of ways it uses is throughreadParcelable
. Also in above example due to Java's Type ErasureArrayList<HierarchyOp> mHierarchyOps
field can actually contain any objects supported by Parcel, not only those compatible with type specified in generic type declaration
As noted above it is responsibility of Parcelable
interface implementation to ensure that createFromParcel
reads same amount of data from Parcel
as matching writeToParcel
has previously written. Whenever there is in BOOTCLASSPATH
a Parcelable
which can violate that contract it creates a vulnerability as it allows for following scenario:
- An evil application sends to
system_server
aBundle
ORParcelable
containing faultyParcelable
instance along with specifically constructed data that will be actually read in step 3 but passed verbatim during step 2 system_server
verifiesBundle
is safe and then forwards it ORsystem_server
passes providedParcelable
to AIDL method that also has critical data passed in next parameter (if data received in that parameter could be modified that would cause security issue)- Another app receives data from
system_server
and trusts it, however due to faulty serialization data that it actually sees differs from datasystem_server
intended to send
I've used "OR" in above steps as these steps describe both an old exploit variant which leads to starting arbitrary Activity which I've published in 2017 (on the left side of "OR") and a new variant which I'll describe here in next section
From the point of application developer using APIs available in Android SDK the way BroadcastReceiver
works is that one application calls sendBroadcast
(although often apps want to receive Broadcasts from system, not app) and then broadcasted Intent is matched to <receiver>
defined in AndroidManifest.xml
, when that happens system starts process of receiving application, instantiates BroadcastReceiver
subclass as defined in <receiver android:name>
attribute and then calls onReceive
method
Let's take a look at communication with system_server
happening in process receiving broadcast:
- When application process is initially started it calls
IActivityManager.attachApplication()
, by doing so it passesIApplicationThread
handle which is used by system to tell application process what to do - When system wants to execute manifest-registered
BroadcastReceiver
in application process, it callsscheduleReceiver
method usingIApplicationThread
described in previous point. This method has multiple arguments but here most important ones are first two:Intent intent
, which was previously passed to system whensendBroadcast()
was calledActivityInfo info
, which contains information about component that has to be executed. Value to this parameter is taken by system from Package Manager Service. Most importantly data passed in this parameter includes path to file from which Java class handling received broadcast will be loaded
At this point you probably can guess what this new exploit path is: call sendBroadcast()
passing an Intent
that will cause that when system tries to call scheduleReceiver
it'll cause that application in which scheduleReceiver
is invoked will see tampered ActivityInfo
It should be noted that this new exploit path became viable in Android 12 as previously there was no way to put arbitrary Parcelable
s in Intent
(Intent extras don't count as they are put into Bundle
which has its whole length written into Parcel and is read as single blob, so extras cannot cause misinterpretation of Intent
object containing them)
Most of the time writeToParcel
/createFromParcel
mismatches are cases where in one of these methods one of the fields is forgotten or written twice, in such case sending such object will always trigger mismatch. (Most of the time that happens when object while Parcelable
, isn't really used across processes, otherwise that would be quickly noticed during normal usage)
This time however that wasn't the case and triggering mismatch isn't obvious
Let's take a look at vulnerable class (original was here, lines marked // New in Android 12
were manually added as they weren't present in AOSP at time of writing) (Here's commit originally introducing vulnerability, however it was published after Android 12 was released)
package android.hardware.camera2.params;
public final class OutputConfiguration implements Parcelable {
private OutputConfiguration(@NonNull Parcel source) {
int rotation = source.readInt();
int surfaceSetId = source.readInt();
int surfaceType = source.readInt();
int width = source.readInt();
int height = source.readInt();
boolean isDeferred = source.readInt() == 1;
boolean isShared = source.readInt() == 1;
ArrayList<Surface> surfaces = new ArrayList<Surface>();
source.readTypedList(surfaces, Surface.CREATOR);
String physicalCameraId = source.readString();
boolean isMultiResolution = source.readInt() == 1; // New in Android 12
ArrayList<Integer> sensorPixelModesUsed = new ArrayList<Integer>(); // New in Android 12
source.readList(sensorPixelModesUsed, Integer.class.getClassLoader()); // New in Android 12
// SNIP: copy values from variables set above to fields of this class
}
public static final @android.annotation.NonNull Parcelable.Creator<OutputConfiguration> CREATOR =
new Parcelable.Creator<OutputConfiguration>() {
@Override
public OutputConfiguration createFromParcel(Parcel source) {
try {
OutputConfiguration outputConfiguration = new OutputConfiguration(source);
return outputConfiguration;
} catch (Exception e) {
Log.e(TAG, "Exception creating OutputConfiguration from parcel", e);
return null;
}
}
@Override
public OutputConfiguration[] newArray(int size) {
return new OutputConfiguration[size];
}
};
@Override
public void writeToParcel(Parcel dest, int flags) {
if (dest == null) {
throw new IllegalArgumentException("dest must not be null");
}
dest.writeInt(mRotation);
dest.writeInt(mSurfaceGroupId);
dest.writeInt(mSurfaceType);
dest.writeInt(mConfiguredSize.getWidth());
dest.writeInt(mConfiguredSize.getHeight());
dest.writeInt(mIsDeferredConfig ? 1 : 0);
dest.writeInt(mIsShared ? 1 : 0);
dest.writeTypedList(mSurfaces);
dest.writeString(mPhysicalCameraId);
dest.writeInt(mIsMultiResolution ? 1 : 0); // New in Android 12
dest.writeList(mSensorPixelModesUsed); // New in Android 12
}
private ArrayList<Surface> mSurfaces;
private final int mRotation;
private final int mSurfaceGroupId;
private final int mSurfaceType;
private final Size mConfiguredSize;
private final int mConfiguredFormat;
private final int mConfiguredDataspace;
private final int mConfiguredGenerationId;
private final boolean mIsDeferredConfig;
private boolean mIsShared;
private String mPhysicalCameraId;
private boolean mIsMultiResolution; // New in Android 12
private ArrayList<Integer> mSensorPixelModesUsed; // New in Android 12
}
So whats wrong here and why this change introduces vulnerability? As I've said while discussing WindowContainerTransaction
example Parcelable
, readList
can actually fill list with any object supported by Parcel
, not only ones matching generic declaration (ArrayList<Integer>
). However since we're just using this class as part of serialization gadget chain and not actually using it that field won't be used for anything else than reading and writing to Parcel
(and attempts to use elements from ArrayList
containing elements mismatching its generic declaration would only lead to ClassCastException
anyway), that isn't problem in itself
In this class there's also a try-catch
within createFromParcel
, which means that if during read an Exception
is thrown, reading of OutputConfiguration
will be stopped and reading of object containing OutputConfiguration
will proceed. When that happens whole OutputConfiguration
will be written to Parcel but it'll be read only to point at which Exception
happened. This creates mismatch as unconsumed data written within OutputConfiguration.writeToParcel
will actually be read by object that was calling OutputConfiguration.CREATOR.createFromParcel
Now, combination of these two (allowing nesting arbitrary objects supported by Parcel
and wrapping that within try-catch
without rethrow) gives ability to construct Parcelable
that can be written by system_server
and later be read in a way that is controlled by app that initially constructed Parcelable
being forwarded by system_server
Ok, so in order to construct such Parcelable
now we need to find something to be put in mSensorPixelModesUsed
that will be successfully read in system_server
(as this object is being received from attacker app through Parcel
), successfully written by system_server
and then will fail to unparcel and throw an Exception
in victim app
One of ways to do so is to use class that is present within system_server
but not in apps, so that attempting to deserialize it would lead to ClassNotFoundException
. I cannot however pick a Parcelable
from system_server
as readParcelable
without explicitly specified ClassLoader
will only search BOOTCLASSPATH
which won't contain system_server
specific classes. Solution to that problem is to use one of Serializable
classes as ObjectInputStream
will pick ClassLoader
from first non-BOOTCLASSPATH
method in stack trace
I've picked PackageManagerException
, however before we use it there's one more thing we need to do. In OutputConfiguration
constructor when readList
is called loader
argument is explicitly set to Integer.class.getClassLoader()
. That loader
value is propagated to readValue()
, then to readSerializable()
and within readSerializable()
if loader parameter isn't null it is used instead of resolveClass
from ObjectInputStream
(that c != null
check does nothing because when Class.forName
doesn't find class it'll throw exception instead of returning null). The way around is quite simple though, we just need to wrap PackageManagerException
in some Parcelable
that does readList
without specifying ClassLoader
. This is where described above WindowContainerTransaction
class comes in
So, at this point we have following object:
OutputConfiguration
mSensorPixelModesUsed.get(0) = WindowContainerTransaction
mHierarchyOps.get(0) = PackageManagerException
Now such object can successfully deserialized within system_server
: when WindowContainerTransaction
calls readList
it'll try finding PackageManagerException
class using system servers ClassLoader
(not BootClassLoader
), as it can find it in stack trace. That class loader happens to be present within stack trace because while all of following methods weren't from system server class path: Binder#execTransact()
, IActivityManager$Stub#onTransact()
generated by AIDL and methods from all used Parcelable
classes, there was method declared within system server in stack trace: an overridden onTransact
within ActivityManagerService
. Therefore system_server
can read and later write such object to Parcel
and when target app attempts reading it PackageManagerException
class won't be available and therefore ClassNotFoundException
will be thrown, wrapped into RuntimeException
and then caught by OutputConfiguration
CREATOR
So we triggered this mismatch. Well, not really yet at this point because exception was caught when no unread data was left by OutputConfiguration.writeToParcel
, but we can easily add another item to mSensorPixelModesUsed
List
and that item will be written through Parcel.writeValue
and left unread after reading OutputConfiguration
As noted above I'll want to trigger mismatch from Intent
object, as it will be passed by system_server
to an AIDL method which has Intent
in first parameter and execution information in second parameter, so that serialization/deserialization of Intent
passed in first parameter lead to modification of value in second parameter
In Intent.readFromParcel()
all values are read through dedicated typed methods so there we cannot specify custom Parcelable
class
Within Intent
though, there's nested ClipData
and since Android 12 in ClipData$Item
there's a new field ActivityInfo mActivityInfo
(was not present in AOSP at time of initial writing, here's commit introducing that field, this field is read through in.readTypedObject(ActivityInfo.CREATOR)
inside ClipData(Parcel in)
constructor)
Then within ActivityInfo(Parcel source)
constructor again there isn't way to put custom Parcelable
, but as ActivityInfo
extends from ComponentInfo
it has applicationInfo
field
Finally within ApplicationInfo
there's SparseArray<int[]> splitDependencies
field, which is read through readSparseArray
, which in turn uses readValue
to read SparseArray
items
At this point we could place OutputConfiguration
within splitDependencies
, however reading splitDependencies
is followed by few readString8()
calls and it'd be nice to have full control over unconsumed data after mismatch happens so we can directly place empty strings there and not worry about different interpretation of unconsumed data
To do so, first we need to put some raw data container within OutputConfiguration.mSensorPixelModesUsed
that will be written through writeValue
. I've chosen Bundle
. That way in unconsumed data we'll have left:
writeValue
VAL_BUNDLE
tag- Length of raw data (this link also applies to remaining items in this list)
BUNDLE_MAGIC
- Raw data passed verbatim through
Parcel.appendFrom
So we have three Parcel.writeInt
items we'd have unconsumed, we can get rid of them by wrapping OutputConfiguration
within some Parcelable
that while reading it reads arbitrary Parcelable
value followed by three ints. I've found that in ZenPolicy CREATOR
To sum up we've got following object hierarchy (that is present in system_server
and which it attempts passing to scheduleReceiver
)
Intent
mClipData = ClipData
mItems.get(0).mActivityInfo = ActivityInfo
applicationInfo = ApplicationInfo
splitDependencies.get(0) = ZenPolicy
mVisualEffects.get(0) = OutputConfiguration
mSensorPixelModesUsed.get(0) = WindowContainerTransaction
mHierarchyOps.get(0) = PackageManagerException
mSensorPixelModesUsed.get(1) = Bundle
That is written by system_server
. Then receiving application reads everything up to (and including readSerializable
data of) PackageManagerException
normally, however after Serializable
data for PackageManagerException
are read an exception is thrown and reading of everything below OutputConfiguration
is cancelled, leaving Bundle
unread. Reading proceeds to ZenPolicy
which consumes three ints that precede raw data within Bundle
. Then ApplicationInfo
reading proceeds with reading data that were previously raw data passed verbatim in Bundle. Reading of that raw data will continue with remaining objects in this stack (ApplicationInfo
, ActivityInfo
, ClipData
and Intent
) and then that raw data will be used for reading next handleReceiver
method parameter
As I've just said below now remaining scheduleReceiver
parameters are read from buffer controlled by attacker.
Let's take a look at what happens once that method is invoked.
First scheduleReceiver
packs values from all arguments and uses sendMessage()
to pass execution to main thread
Next, on main thread handleReceiver
is called
handleReceiver
calls getPackageInfoNoCheck
, passing it ApplicationInfo
which it received as part of ActivityInfo
which was passed to scheduleReceiver
argument
getPackageInfo
checks if package with given name is already present in cache and if not it constructs new LoadedApk
instance, passing it ApplicationInfo
object received earlier (Since attacker wants to cause new LoadedApk
to be constructed a packageName
of package that wasn't earlier seen in this process is used)
Then ContextImpl.getClassLoader()
method is used, which at first run delegates to mPackageInfo.getClassLoader()
, with mPackageInfo
being a LoadedApk
constructed in previous paragraph
Then there's createOrUpdateClassLoaderLocked
, which calls makePaths
to populate zipPaths
with paths to be used in ClassLoader
, then they are joined and assigned to zip
variable and that is passed to createClassLoader
makePaths
fills zipPaths
using information from ApplicationInfo
, most importantly this includes sourceDir
. Attacker application makes injected ApplicationInfo
with sourceDir
set to path to own apk, therefore receiver class will be actually loaded from attacker apk. This directly leads to execution of attacker-controlled code within application receiving broadcast
Note on hidden API checks
There was one more thing that needed to be bypassed: hidden API checks. These were never meant to be security boundary (as application can always use NDK and call underlying syscall directly), but in this case they were bypassed by constructing crafted ClipData
by manually writing data to Parcel
and then using readParcelable
. Such ClipData
could be then normally attached to Intent
and then passed to sendBroadcast()
so sending broadcast itself was done using only public APIs
The above writeup was originally sent to Google and looks like they've made use of it as there are multiple fixes resulting from it (I think, I have no proof about direct causality)
Released with Android 12:
- Exception swallowing from
OutputConfiguration
and related classes was removed OutputConfiguration#mSensorPixelModesUsed
is no longer written throughwriteValue
ClipData#mActivityInfo
is no longer written toParcel
unless explicitly requested during write (soIntent
can no longer contain arbitraryParcelable
s fromBOOTCLASSPATH
, eliminating this exploitation technique)
Present only on master
branch at time of writing, not in released versions, probably will appear in Android 13 (not in 12L):
- There are new
List
reading methods on Parcel that check type of items and untyped versions have been marked as deprecated - There's new method
Parcel#enforceNoDataAvail()
that checks that there are no unread data left in Parcel, that apparently will be used by AIDL after reading RPC call arguments. Usually my exploits usually relied on fact that after all data was read from Parcel everything else was ignored, that would no longer will be the case, although I think in many cases one could construct data that will cause at end seek to final position so this one isn't strong mitigation. Anyway sometimes such issues occur naturally undetected so that would catch those. Further discussion about that is in issue #3 - Every item in
Bundle
will have its length saved separately. This pretty much kills whole bug class from which I've been reporting bugs privately to Google since 2014, published description in 2017 and code itself about year later. If I'm counting correctly that will be 8 years of bug class life (I honestly have no idea if that is a lot, although there still might be non-Bundle
exploit variants, like exactly this one (although exactly this one was already fixed))