Dalvik Bytecode Verifier Notes

The bytecode verifier in the Dalvik VM attempts to provide the same sorts of checks and guarantees that other popular virtual machines do. We perform generally the same set of checks as are described in _The Java Virtual Machine Specification, Second Edition_, including the updates planned for the Third Edition.

Verification can be enabled for all classes, disabled for all, or enabled only for "remote" (non-bootstrap) classes. It should be performed for any class that will be processed with the DEX optimizer, and in fact the default VM behavior is to only optimize verified classes.

Why Verify?

The verification process adds additional time to the build and to the installation of new applications. It's fairly quick for app-sized DEX files, but rather slow for the big "core" and "framework" files. Why do it all, when our system relies on UNIX processes for security?

1. Optimizations. The interpreter can ignore a lot of potential error cases because the verifier guarantees that they are impossible. Also, we can optimize the DEX file more aggressively if we start with a stronger set of assumptions about the bytecode.
2. "Exact" GC. The work peformed during verification has significant overlap with the work required to compute register use maps for exact GC. Improper register use, caught by the verifier, could lead to subtle problems with an "exact" GC.
3. Intra-application security. If an app wants to download bits of interpreted code over the network and execute them, it can safely do so using well-established security mechanisms.
4. 3rd party app failure analysis. We have no way to control the tools and post-processing utilities that external developers employ, so when we get bug reports with a weird exception or native crash it's very helpful to start with the assumption that the bytecode is valid.

Verifier Differences

There are a few checks that the Dalvik bytecode verifier does not perform, because they're not relevant. For example:

• Type restrictions on constant pool references are not enforced, because Dalvik does not have a pool of typed constants. (Dalvik uses a simple index into type-specific pools.)
• Verification of the operand stack size is not performed, because Dalvik does not have an operand stack.
• Limitations on jsr and ret do not apply, because Dalvik doesn't support subroutines.

• In a conventional VM, backward branches and exceptions are forbidden when a local variable holds an uninitialized reference. The restriction was changed to mark registers as invalid when they hold references to the uninitialized result of a previous invocation of the same new-instance instruction. This solves the same problem -- trickery potentially allowing uninitialized objects to slip past the verifier -- without unduly limiting branches.

• The move-exception instruction can only appear as the first instruction in an exception handler.
• The move-result* instructions can only appear immediately after an appropriate invoke-* or filled-new-array instruction.

The Dalvik verifier is more restrictive than other VMs in one area: type safety on sub-32-bit integer widths. These additional restrictions should make it impossible to, say, pass a value outside the range [-128, 127] to a function that takes a byte as an argument.

Verification Failures

When the verifier rejects a class, it always throws a VerifyError. This is different in some cases from other implementations. For example, if a class attempts to perform an illegal access on a field, the expected behavior is to receive an IllegalAccessError at runtime the first time the field is actually accessed. The Dalvik verifier will reject the entire class immediately.

It's difficult to throw the error on first use in Dalvik. Possible ways to implement this behavior include:

1. We could replace the invalid field access instruction with a special instruction that generates an illegal access error, and allow class verification to complete successfully. This type of verification must often be deferred to first class load, rather than be performed ahead of time during DEX optimization, which means the bytecode instructions will be mapped read-only during verification. So this won't work.
2. We can perform the access checks when the field/method/class is resolved. In a typical VM implementation we would do the check when the entry is resolved in the context of the current classfile, but our DEX files combine multiple classfiles together, merging the field/method/class resolution results into a single large table. Once one class successfully resolves the field, every other class in the same DEX file would be able to access the field. This is bad.
3. Perform the access checks on every field/method/class access. This adds significant overhead. This is mitigated somewhat by the DEX optimizer, which will convert many field/method/class accesses into a simpler form after performing the access check. However, not all accesses can be optimized (e.g. accesses to classes unknown at dexopt time), and we don't currently have an optimized form of certain instructions (notably static field operations).

Other implementations are possible, but they all involve allocating some amount of additional memory or spending additional cycles on non-DEX-optimized instructions. We don't want to throw an IllegalAccessError at verification time, since that would indicate that access to the class being verified was illegal.

One approach that might be worth pursuing: for situations like illegal accesses, the verifier makes an in-RAM private copy of the method, and alters the instructions there. The class object is altered to point at the new copy of the instructions. This requires minimal memory overhead and provides a better experience for developers.

The VerifyError is accompanied by detailed, if somewhat cryptic, information in the log file. From this it's possible to determine the exact instruction that failed, and the reason for the failure. We can also constructor the VerifyError with an IllegalAccessError passed in as the cause.