Description

This PR refactors the Metering middleware to address the issues described in #999. In short, the current metering does not prevent execution of code segments that exceed the limit, but instead only checks if limits are exceeded after execution of the code segment. Additionally the current metering does not catch some situations where the limit exceeded in the last few instructions of a function.

This metering injects limit checks prior to each code segment, ensure that execution does not proceed if there is not enough gas to reach the next branching instruction. Although after an ExecutionLimitExceededError the reported gas used will be greater than the limit, the execution will not have actually proceeded through that code segment.

Additionally, this metering will check further ahead than eWasm metering in certain circumstances.

For example if we have multiple nested loops in a block without any branching instructions,

METER
Block
    ... // no branching ops
    Loop
        METER
        ... // no branching ops
        Loop
            METER
            ...
            IF // first branching op since first METER
                ...
            END
        End
    End
End

Each METER represents injected metering code that adds the cost of the next segment to the used points and checks it against the limit, trapping if exceeded.

Each METER that I listed will actually ensure that there is enough gas to reach the first branching op, which in my above example is IF.

I'm happy to answer more questions about the Metering design. I believe it would benefit from additional tests that use more complex wasm code.

Implementation details

In order to implement this metering as such, a refactor of parse::read_module() and codegen::MiddlewareChain.run() and codegen::EventSink was required.

This metering algorithm requires the ability to modify previously seen Events that have already been pushed onto the EventSink inside of feed_event. This means that this middleware chain needs to see all of the Events within a function prior to any subsequent chains seeing any of the Events. Additionally this means that all ParserState(Operator) events for a given function need to all be loaded into memory at the same time.

In parse::read_module, I clone each Operator and store them in an events Vec<Event>. I load a complete function and then I pass the whole events Ven to MiddlewareChain.run(), which passes all events through each middleware chain.

Unfortunately wasmparser::Operator does not implement clone, so I had to fork it and add the trait. I have opened an issue on the latest wasmparser but for compatability I applied my patch to the version that is currently being used.

Issues

Currently my refactor of parse::read_module and codegen::MiddlewareChain.run breaks the singlepass backend. This is not due to the metering code but due to my refactor of the runtime-core, as I have isolated that patch in testing. Alternatively you can just comment out the line that adds the metering middleware entirely from the middleware-common-tests to show that the issue lies deeper.

When running the middleware-common-tests compilation fails with singlepass I get the following failure:

Errors were encountered when committing before finalization: UnknownLabel(Dynamic(DynamicLabel(3)))

https://gist.github.com/AdamSLevy/455d2a50ec1211d20efaebcd9c06eb1c#file-gistfile1-txt-L42-L102

I have compared my code very closely with the master branch and I am 99% certain that all mcg and fcg functions are being called in the proper order in my refactor. The only difference I can see is that the Events are now all WasmOwned.

However my metering code works perfectly with the LLVM backend. I have been working with @MarkMcCaskey to come up with this refactor and I have run into this issue and can't work past it as I believe it would require me to learn the singlepass backend and the dynasmrt Assembler.

Review

[ ] Add a short description of the the change to the CHANGELOG.md file

+333 -111

3 commit in AdamSLevy/wasmer/tree/MeteringImproved wasmerio/wasmer/tree/master

24 comments

8 changed files

AdamSLevy

pr closed time in 3 minutes

pull request commentwasmerio/wasmer

Improved Metering

Yeah, this PR is outdated now as is based on the Wasmer 0.x codebase which doesn't include the new structure brought by the refactor for 1.0.

I believe the essence is in included in the new metering schema. Here's the new code in case is useful for review: https://github.com/wasmerio/wasmer/blob/master/lib/middlewares/src/metering.rs#L182-L216

New PRs would be more than welcome for improving the current metering middleware. Closing this PR now :)

AdamSLevy

comment created time in 3 minutes

push eventwasmerio/wasmer

Ivan Enderlin

commit sha 5c9d0230232cd6683e915f3c12747d501ee985ec

chore(vm) Extract `InstanceRef` into its own module.