Race condition during module compilation engine about wazero

[mathetake]

please provide the reproducer, otherwise we can't verify

[mathetake]

closing until you give us the minimal repro. Note that executing a module instance concurrently is inherently unsafe and not possible unless you use the threading proposal enabled + your binary is compiled in that way.

[martinlakov]

As described in this comment - a compilation engine is getting re-used/shared between runtimes.

The wazevo engine comes with an internal map named compiledModules. When creating runtimes & compiling modules happens concurrently - a race condition occurs with accessing the said map. This happens in here

[martinlakov]

I'd be glad to provide example(s), propose a fix etc. Just want to see this one getting resolved 😊

[martinlakov]

A bit of background to the story - issue has been observed while working on a high-level library that steps on top of wazero and takes care to hide away complexity associated with working with the linear memory etc. This one - we plan to open-source as soon we have the APIs somewhat final.

[martinlakov]

@mathetake - update of my previous comment (was on my phone, on my way to work)

As described in this comment - a compilation engine is getting re-used/shared between runtimes.

The wazevo engine comes with an internal map named compiledModules. When creating runtimes & compiling modules happens concurrently - a race condition occurs with accessing the said map. This happens like so:

• in here
• invoked from
• invoked from
• invoked from

As the "callstack" suggests - there's no logic to govern access to the said map and the said map is not thread-safe. As a result - a race condition occurs as soon as you try and use a single cache instance among multiple runtime instances and you try to load/execute modules concurrently.

It is evident that we're talking about a bug in here. If you take a look at the implementation of CompiledModuleCount for wazevo (see: here) - read-only access to the said map is governed appropriately.

[martinlakov]

A potential fix for the problem is to create a "private" function to check if a module has been compiled already. Something like:

func (e *engine) isCompiled(m *wasm.Module) bool {
	e.mux.RLock()
	defer e.mux.RUnlock()

	_, ok := e.compiledModules[m.ID]
	return ok
}

and replace the direct access to the map in here with calling the said function.