Formal specification and verification of a jvm and its bytecode verifier

How do we know that a bytecode-verified Java program will run safely? This dissertation addresses the question by building a precise model of the JVM and its bytecode verifier. We also built a machine and its bytecode verifier to demonstrate an approach to solving this problem. We proved that for any program on the small machine that has been vetted by the small bytecode verifier, then that program will run safely on the small machine. We created substantial libraries of ACL2 definitions and lemmas towards specifying and proving that the JVM safely executes verified programs. The fundamental problem is to connect the abstract execution of the bytecode verifier with the concrete execution of the JVM. These diverge in two ways: (1) the bytecode verifier executes on more abstract states and (2) its execution of INVOKE-family and BRANCH-family instructions differs from their execution by the JVM. Our contribution was identification of a critical on-track property that, despite these divergences between the bytecode verifier and the JVM, enables one to use the success of bytecode verification to predict the safety of concrete execution. The second difficulty is that the official specification describes many procedural aspects of the bytecode verification process. These aspects obscure the checks conducted by the bytecode verifier. We introduce an alternative bytecode verifier without such procedural aspects. We use the new bytecode verifier as a stepping stone for proving that the official bytecode verifier is effective. Following this methodology allowed us to prove, on our machine, that executions of bytecode-verified programs never overflow the operand stack. We note that significant effort is required in order to extend this result from our machine to the full JVM. We have formulated appropriate stronger notions of safe execution for programs on the full JVM. We introduced an alternative bytecode verifier. We proved the “reduction theorem” that relates the official bytecode verifier with the alternative bytecode verifier. We completed proofs of several thousand lemmas towards proving the safe execution of bytecode-verified programs on the full JVM. Our results are organized into supporting lemma libraries.