Sequence of Processor Instructions Leads to Unexpected Behavior

Description

Demonstrations

The following examples help to illustrate the nature of this weakness and describe methods or techniques which can be used to mitigate the risk.

Note that the examples here are by no means exhaustive and any given weakness may have many subtle varieties, each of which may require different detection methods or runtime controls.

Example One

The Pentium F00F bug is a real-world example of how a sequence of instructions can lock a processor. The "cmpxchg8b" instruction compares contents of registers with a memory location. The operand is expected to be a memory location, but in the bad code snippet it is the eax register. Because the specified operand is illegal, an exception is generated, which is the correct behavior and not a security issue in itself. However, when prefixed with the "lock" instruction, the processor deadlocks because locked memory transactions require a read and write pair of transactions to occur before the lock on the memory bus is released. The exception causes a read to occur but there is no corresponding write, as there would have been if a legal operand had been supplied to the cmpxchg8b instruction. [REF-1331]

lock cmpxchg8b eax

Example Two

The Cyrix Coma bug was capable of trapping a Cyrix 6x86, 6x86L, or 6x86MX processor in an infinite loop. An infinite loop on a processor is not necessarily an issue on its own, as interrupts could stop the loop. However, on select Cyrix processors, the x86 Assembly 'xchg' instruction was designed to prevent interrupts. On these processors, if the loop was such that a new 'xchg' instruction entered the instruction pipeline before the previous one exited, the processor would become deadlocked. [REF-1323]

Example Three

The Motorola MC6800 microprocessor contained the first documented instance of a Halt and Catch Fire instruction - an instruction that causes the normal function of a processor to stop. If the MC6800 was given the opcode 0x9D or 0xDD, the processor would begin to read all memory very quickly, in sequence, and without executing any other instructions. This will cause the processor to become unresponsive to anything but a hard reset. [REF-1324]

Example Four

The example code is taken from the commit stage inside the processor core of the HACK@DAC'19 buggy CVA6 SoC [REF-1342]. To ensure the correct execution of atomic instructions, the CPU must guarantee atomicity: no other device overwrites the memory location between the atomic read starts and the atomic write finishes. Another device may overwrite the memory location only before the read operation or after the write operation, but never between them, and finally, the content will still be consistent.

Atomicity is especially critical when the variable to be modified is a mutex, counting semaphore, or similar piece of data that controls access to shared resources. Failure to ensure atomicity may result in two processors accessing a shared resource simultaneously, permanent lock-up, or similar disastrous behavior.

if (csr_exception_i.valid && csr_exception_i.cause[63] && commit_instr_i[0].fu != CSR) begin
  exception_o = csr_exception_i;
  exception_o.tval = commit_instr_i[0].ex.tval;

end

The above vulnerable code checks for CSR interrupts and gives them precedence over any other exception. However, the interrupts should not occur when the processor runs a series of atomic instructions. In the above vulnerable code, the required check must be included to ensure the processor is not in the middle of a series of atomic instructions.

Refrain from interrupting if the intention is to commit an atomic instruction that should not be interrupted. This can be done by adding a condition to check whether the current committing instruction is atomic. [REF-1343]

if (csr_exception_i.valid && csr_exception_i.cause[63] && !amo_valid_commit_o && commit_instr_i[0].fu != CSR) begin
  exception_o = csr_exception_i;
  exception_o.tval = commit_instr_i[0].ex.tval;

end

See Also