Threat model
This page analyzes Duchess's use of the JNI APIs to explain how it guarantees memory safety. Sections:
- Assumptions: requirements for safe usage of Duchess which Duchess itself cannot enforce.
- Code invariants: invariants that Duchess maintains
- Threat vectors that cause UB: ways to create undefined behavior using JNI, and how Duchess prevents them (references code invariants and assumptions)
- Threat vectors that do not cause UB: suboptimal uses of the JNI that do not create UB; duchess prevents some of these but not all (references code invariants and assumptions)
Assumptions
We assume three things
- The user does not attempt to start the JVM via some other crate in parallel with using duchess methods
- The user does not use the JNI
PushLocalFrame
method to introduce "local variable frames" within the context ofJvm::with
call. - The Java
.class
files that are present at build time have the same type signatures and public interfaces as the class files that will be present at runtime. Although there are no known memory-safety vulnerabilities stemming from failure to maintain this invariant, failing to provide matching classfiles will result in failures at runtime.
Code invariants
This section introduces invariants maintained by Duchess using Rust's type system as well as careful API design.
Possessing a &mut Jvm<'jvm>
implies attached thread
Jvm
references are obtained with Jvm::with
. This codepath guarantees
- The JVM has been started (see the assumptions), using default settings if needed
- The current thread is attached
- We maintain a thread-local-variable tracking the current thread status.
- If the thread is recorded as attached, nothing happens.
- Otherwise the JNI method to attach the current thread
AttachCurrentThread
is invoked; in that case, the thread will be detached oncewith
returns. - Users can also invoke
Jvm::attach_thread_permanently
to avoid the overhead of attaching/detaching, which simply sets the thread-local variable to a permanent state and avoids detaching.
The 'jvm
lifetime &mut Jvm<'jvm>
is the innermost scope for local variables
References to Java objects of type J
are stored in a Local<'jvm, J>
holder. Local references can come from the arguments to native functions or from JvmOp::do_jni
calls. do_jni
calls use the 'jvm
' lifetime found on the Jvm<'jvm>
argument. This allows the Local
to be used freely within that scope. It is therefore important that 'jvm
be constrained to the innermost valid scope.
Inductive argument that this invariant is maintained:
- Base case: Users can only obtain a
Jvm<'jvm>
value viaJvm::with
, which takes a closure argument of typefor<'jvm> impl FnMut(&mut Jvm<'jvm>)
. Therefore, this closure cannot assume that'jvm
will outlive the closure call and all local values cannot escape the closure body. - Inductive case: All operations performed within a
Jvm::with
maintain the invariant. Violating the invariant would require introducing a new JNI local frame, which can happen in two ways:- invoking
PushLocalFrame
: Duchess does not expose this operation, and we assume users do not do this via some other crate. - calling into Java code which in turn calls back into Rust code via a
native
method: In this case, we would have a stack with Rust codeR1
, then Java codeJ
, then a Rust functionR2
that implements a Java native method.R1
must have invokedJvm::with
to obtain a&mut Jvm<'jvm>
. IfR1
could somehow give thisJvm<'jvm>
value toR2
,R2
could create locals that would outlive its dynamic extent, violating the invariant. However,R1
to invoke Java codeJ
,R1
had to invoke a duchess method with&mut Jvm<'jvm>
as argument, which means that it has given the Java code unique access to the (unique)Jvm<'jvm>
value, leant out its only reference, and the Java code does not give this value toR2
.
- invoking
Flaw:
It is theoretically possible to do something like this...
Jvm::with(|jvm1| ...)
- stash the
jvm1
somewhere in thread-local data using unsafe code Jvm::with(|jvm2| ...)
- invoke jvm code that calls back into Rust
- from inside that call, recover the
Jvm<'jvm1>
, allocate a newLocal
with it, and store the result back (unsafely)
- from inside that call, recover the
- invoke jvm code that calls back into Rust
- recover the pair of
jvm1
and the object that was created
- stash the
...it is difficult to write the code that would do this and it requires unsafe code, but that unsafe code doesn't seem to be doing anything that should not theoretically work. Avoiding this is difficult, but if we focus on execute
, we can make it so that users never directly get their hands on a Jvm
and make this safe.
All references to impl JavaObject
types are JNI local or global references
The JavaObject
trait is an unsafe
trait. When implemented on a struct S
, it means that every &S
reference must be a JNI local or global references. This trait is implemented for all the structs that duchess creates to represent Java types, e.g., duchess::java::lang::Object
. This invariant is enforced by the following pattern:
- Each such struct has a private field of type
Infallible
, ensuring it could never be constructed via safe code. - To "construct" an instance of this struct you would use a constructor like
Object::new
which returns an implJavaConstructor
; when evaluated it will yield aLocal
wrapper. Locals are only constructed for pointers we get from JNI. Global can be created from Locals (and hence come from JNI too).
1:1 correspondence between JNI global/local references and Global
/Local
Every time we create a Global
value (resp. Local
), it is created with a new global or local reference on the JNI side as well. The Drop
for Global
releases the global (resp., local) reference.
Threat vectors that cause UB
What follows is a list of specific threat vectors identified by based on the documentation JNI documentation as well as a checklist of common JNI failures found on IBM documentation.
PushLocalFrame
invoked by a mechanism external to Duchess
Outcome of nonadherence: UB
Duchess not expose PushLocalFrame
, but it is possible to invoke this method via unsafe code or from other crates (e.g., the jni
crate's push_local_frame
method). This method will cause local variables created within its dynamic scope to be released when PopLocalFrame
is invoked. The 'jvm
lifetime mechanism used to ensure local variables do not escape their scope could be invalidated by these methods. See the section on the jvm lifetime for more details.
How Duchess avoids this: Duchess carefully controls use of this method internally. We explicitly assume that users do not invoke this method directly via alternative means.
Multiple JVMs started in the same process
Outcome of nonadherence: UB
There can only be one JVM per process. If multiple JVMs are started concurrently, crashes or UB can occur.
How Duchess avoids this: Documentation and synchronization: Within the Duchess library, all JVM accesses are internally synchronized. Duchess will lazily start the JVM if it has not been started already. However, Duchess cannot control the behavior of other libraries (or another major version of the package). Duchess mitigates this with documentation recommending users start the JVM explicitly in main. Future work may improve this with a centralized "start-jvm
" crate that is shared between jni
, duchess
and any other JNI based Rust libraries. Duchess may also add mitigations to prevent multiple major versions of Duchess from being used in the same dependency closure.
When you update a Java object in native code, ensure synchronization of access.
Outcome of nonadherence: Memory corruption
How Duchess avoids this: We do not support updating objects in native code.
Cached method and field IDs
From the JNI documentation:
A field or method ID does not prevent the VM from unloading the class from which the ID has been derived. After the class is unloaded, the method or field ID becomes invalid. The native code, therefore, must make sure to:
- keep a live reference to the underlying class, or
- recompute the method or field ID
if it intends to use a method or field ID for an extended period of time.
Duchess caches method and field IDs in various places. In all cases, the id is derived from a Class
reference obtained by invoking JavaObject::class
. The JavaObject::class
method is defined to permanently (for the lifetime of the process) cache a global reference to the class object, fulfilling the first criteria ("keep a live reference to the underlying class").
Local references are tied to the lifetime of a JNI method call
The JNI manual documents that local references are "valid for the duration of a native method call. Once the method returns, these references will be automatically out of scope." In Duchess, each newly created local reference is assigned to a Local<'jvm, T>
. This type carries a lifetime ('jvm
) that derives from the duchess::Jvm<'jvm>
argument provided to the JvmOp::do_jni
method. Therefore, the local cannot escape the 'jvm
lifetime on the Jvm<'jvm>
value; duchess maintains an invariant that 'jvm
is the innermost JNI local scope.
Local references cannot be saved in global variables.
Outcome of nonadherence: Random crashes
How Duchess avoids this: See discussion here and the jvm
invariant.
Always check for exceptions (or return codes) on return from a JNI function. Always handle a deferred exception immediately you detect it.
Outcome of nonadherence: Unexplained exceptions or undefined behavior, crashes
How Duchess avoids this: End-users do not directly invoke JNI functions. Within Duchess, virtually all calls to JNI functions use the EnvPtr::invoke
helper function which checks for exceptions. A small number use invoke_unchecked
:
array.rs
- invokes
invoke_unchecked
onGetArrayLength
, which is not documented as having failure conditions - invokes primitive setter with known-valid bounds
- invokes primitive getter with known-valid bounds
- invokes
cast.rs
- invokes infallible method
IsInstanceOf
- invokes infallible method
find.rs
- invokes
GetMethodID
andGetStaticMethodID
"unchecked" but checks the return value for null and handles exception that occurs
- invokes
raw.rs
- invokes
invoke_unchecked
in the implementation ofinvoke
:)
- invokes
ref_.rs
- invokes
NewLocalRef
with a known-non-null argument - invokes
NewLocalRef
with a known-non-null argument
- invokes
str.rs
- invokes
GetStringLength
— infallible - invokes
GetStringUTFLength
— infallible - invokes
GetStringUTFRegion
with known-valid bounds
- invokes
jvm.rs
Usage of Throw and ThrowNew
The native method can choose to return immediately, causing the exception to be thrown in the Java code that initiated the native method call.
Outcome of nonadherence: Undefined behavior.
How Duchess avoids this: Throw and ThrowNew are invoked from the java_function
macro as part of native_function_returning_object
and native_function_returning_scalar
. After duchess raises the exception, no more JNI calls are made and null or 0 is returned forcing the Java caller to handle the exception.
Clear exceptions before invoking other JNI calls
After an exception has been raised, the native code must first clear the exception before making other JNI calls.
Outcome of nonadherence: Undefined behavior.
How Duchess avoids this: When we detect an exception, we always clear the exception immediately before returning a Result
.
Illegal argument types
Reporting Programming Errors
The JNI does not check for programming errors such as passing in NULL pointers or illegal argument types.
The programmer must not pass illegal pointers or arguments of the wrong type to JNI functions. Doing so could result in arbitrary consequences, including a corrupted system state or VM crash.
How Duchess avoids this: We generate strongly typed interfaces based on the signatures found in the class files and we assume that the same class files are present at runtime.
Example tests:
type_mismatch_*.rs
in the test directory
Native references crossing threads
Local references are only valid in the thread in which they are created. The native code must not pass local references from one thread to another.
Outcome of nonadherence: Undefined Behavior
How Duchess avoids this: Duchess prevents this because Local
is !Sync
.
Example Tests:
- doctest in
ref_.rs
Threat vectors that do not cause UB
Invoke execution occurred regularly
Native methods should insert
ExceptionOccurred()
checks in necessary places (such as in a tight loop without other exception checks) to ensure that the current thread responds to asynchronous exceptions in a reasonable amount of time.
Outcome of nonadherence: Asynchronous exceptions won't be detected.
How Duchess avoids this: We check this flag at every interaction with the JVM but not other times; it is possible for Rust code to execute for arbitrary amounts of time without checking the flag. Asynchronous exceptions are not recommended in modern code and the outcome of not checking is not undefined behavior.
Local variable capacity
Each JNI frame has a guaranteed capacity which can be extended via EnsureLocalCapacity
. This limit is largely advisory, and exceeding it does not cause UB. The documentation states:
For backward compatibility, the VM allocates local references beyond the ensured capacity. (As a debugging support, the VM may give the user warnings that too many local references are being created. In the JDK, the programmer can supply the -verbose:jni command line option to turn on these messages.) The VM calls FatalError if no more local references can be created beyond the ensured capacity.
Outcome of nonadherence: Slower performance or, in extreme cases, aborting the process via reporting a Fatal Error.
How Duchess avoids this:
- Duchess is not aware of this limit and does not limit the number of local variables that will be created. If needed, we could support annotations or other means.
- However, if using
Duchess
in its recommended configuration (withexecute
calls), all local variables will be cleaned up in between operations, and operations always create a finite (and statically known) number of locals
Ensure that every global reference created has a path that deletes that global reference.
Outcome of nonadherence: Memory leak
How Duchess avoids this: Because there is a 1:1 correspondence between JNI global references
Every time we create a global reference, we store it in a Global
type. The destructor on this type will free the reference.
Memory exhaustion from too many local references
However, there are times when the programmer should explicitly free a local reference. Consider, for example, the following situations:
- A native method accesses a large Java object, thereby creating a local reference to the Java object. The native method then performs additional computation before returning to the caller. The local reference to the large Java object will prevent the object from being garbage collected, even if the object is no longer used in the remainder of the computation.
- A native method creates a large number of local references, although not all of them are used at the same time. Since the VM needs a certain amount of space to keep track of a local reference, creating too many local references may cause the system to run out of memory. For example, a native method loops through a large array of objects, retrieves the elements as local references, and operates on one element at each iteration. After each iteration, the programmer no longer needs the local reference to the array element.
The JNI allows the programmer to manually delete local references at any point within a native method.
Outcome of nonadherence: Memory exhaustion.
How Duchess avoids this: We do not expect users to do fine-grained interaction with Java objects in this fashion and we do not provide absolute protection from memory exhaustion. However, we do mitigate the likelihood, as the Local
type has a destructor that deletes local references. Therefore common usage patterns where a Local
is created and then dropped within a loop (but not live across loop iterations) would result in intermediate locals being deleted.
Ensure that you use the isCopy and mode flags correctly. See Copying and pinning.
Outcome of nonadherence: Memory leaks and/or heap fragmentation
How Duchess avoids this: Duchess does not currently make use of the methods to gain direct access to Java array contents, so this is not relevant.
Ensure that array and string elements are always freed.
Outcome of nonadherence: Memory leak
How Duchess avoids this: Unclear what this exactly means, to be honest, but we make no special effort to prevent it. However, memory leaks are largely unlikely in Duchess due to having a destructor on Global
.
Mismatch of .class
files
Duchess cannot guarantee that the .class
files used during compilation match the classfiles at runtime.
How Duchess avoids this Duchess will load JNI methods at runtime using their method descriptor. The method descriptor fully captures arguments and return type of a method, (with the exception of generics). If a correct classfile has not been provided, this will result in a safe runtime error. In the case of generics, we operate on pointers instead and do not directly transmute memory returned from the JVM. These values are only useful when interacting with the JVM. Since we are operating on pointers and interacting with the JVM, there is no avenue to create memory unsafety.
Outcome of nonadherence: Errors returned from .execute()