From gem5
Jump to: navigation, search

The Checker allows for dynamic verification of any CPUs that use DynInst. Currently it does not support SMT or MP systems. Extending the Checker for those systems should not be difficult, but makes it impossible to verify memory value of loads.

What it verifies

The Checker can verify:

  • Correct instruction path
  • Correct fetched instruction
  • Correct execution results
  • Correct PC redirection due to branches, faults, or PC events

What it can't verify

The Checker cannot verify:

  • Correct interrupt handling
  • Store values being correct in memory
  • Certain instructions, such as RPCC (reads cycle counter), RC/RS (reads interrupt flags)

The Checker does not have access to the main CPU's interrupt signals, so it can't ensure that the interrupts are detected and handled correctly.

Stores, although they commit and are issued to memory in program order, may actually complete out of order. This is especially true with store conditionals, which don't actually complete until the store conditional result is written back. Because instructions are verified in program order and only after they complete, the value in memory may not reflect the store being verified but may instead be due to a younger (but also committed) store.

Instructions such as RPCC will have different results by the time the instruction has completed versus when it executed. Thus these instructions must be marked as IsUnverifiable in the decoder.isa file.

How it works

The Checker works by keeping a separate SimpleThread, which it updates with its own results of instruction fetching and execution. Once the main CPU completes an instruction, it sends that instruction to the Checker. The Checker verifies these instructions in program order. It ensures that each instruction matches the instruction that the Checker fetches. It does the same with the execution of the instruction, as well as any faults that the instruction generates. In the cases of loads it checks memory to verify the load's value matches what is in memory. This verification can not be done on anything other than UP systems because memory's value may have changed between the time the load executes and the time the load completes.

The Checker defines a ThreadContext, called CheckerThreadContext, which serves as a wrapper for the main CPU's ThreadContext and its own SimpleThread. It takes any calls made to CheckerThreadContext and forwards them to either: both the Checker and the main CPU (in the case of setting registers); or just the main CPU alone (in the case of reading registers). This allows the Checker to be able to keep its state updated properly even when the main CPU's state is updated externally (e.g. through a fault). It also allows CheckerThreadContext to be used transparently in place of the main CPU's ThreadContext because it will return the main CPU's state properly.

How to use

For a CPU that uses DynInst, the following steps must be taken to be able to use the Checker:

  • Add the Checker to its SimObject params.
  • Setup the Icache and Dcache ports of the Checker when they are created at Fetch and in the LSQ.
  • Store the result of any destination register write in the result variable inside DynInst.
  • Use the Checker's ThreadContext in place of its normal ThreadContext when the checker is enabled.
  • Add calls to tell the Checker when instructions have completed. Usually this is when the instruction is committed. However stores do not actually complete until their data is written back.
  • If the CPU modifies its own state anywhere inside its code (and not through its ThreadContext), be sure to call the right functions to keep the Checker up to date.

See the Detailed CPU for an example of using the Checker.