Proposal: Memory grow event / handler
dcodeIO opened this issue · 12 comments
I have a use case here where a WebAssembly module is able to grow its memory from the inside, making it necessary to import a function from JS that is called to notify the outside that the underlying buffer is now in a detached state and that any potential views on the memory buffer must be recreated.
It would be easier / more interoperable / involve less boilerplate if the WebAssembly.Memory instance would emit an event / call a handler whenever growing of memory happens, for example a grow event / an .ongrow property that can be assigned a callback.
Is this something worth considering? :)
Maybe that's a more common js API:
const m = WebAssembly.Memory({ ... });
m.addEventListener("change", fn);change/onchange event.
I think that It would be quite confusing for people to have to handle that and doesn't seem like a common use-case in JS. Maybe we could have a live view of the Memory object somehow?
Sure, a "live view" of sorts that survives a buffer becoming detached would be ideal, if that's an option. Regarding an event: change is fine as well, of course.
malloc can enlarge the memory from within the Wasm code; so in my JS glue on every operation that accesses the memory views I have to check that the view hasn't become detached, and return a fresh view if so. I do this with a wrapper:
let _u8View = undefined;
function getU8View() {
if (_u8View === undefined || _u8View.buffer !== memory.buffer)
_u8View = new Uint8Array(memory.buffer);
return _u8View;
}
You get the idea... a way to avoid this nuisance would be nice!
Something like this seems useful. It would be useful to collect precedent in other Web APIs, and make this proposal more concrete.
@xtuc addEventListener seems odd to me; that's more of a web API than a JS API.
@NWilson Are you providing JS glue for arbitrary wasm modules? If you're generating the modules yourself it seems like it would be better to add an import hook instead. Even if not, you could create a new module with the hook by searching for all grow_memory instructions.
@dcodeIO WebAssembly.Memory never becomes detached, only the underlying ArrayBuffer. I believe having growable ArrayBuffers has been discussed before, but it would be a much larger change, IMO.
@binji That's right, technically this doesn't affect the Memory instance, but the views on its .buffer. Examples for such views are HEAP32 etc. in Emscripten. Without an event (or growable Arraybuffers), the JS-side doesn't know when the buffer becomes detached without either adding custom boilerplate (i.e. an imported env.onMemoryGrow function that is called from WebAssembly to notify the host) or explicitly checking for this condition whenever reading/writing to HEAP32, which adds quite some overhead.
Effectively it's still the memory that is being grown, with the arraybuffer becoming detached as a side-effect, hence the proposal for adding the event to the memory instance - but I'm of course fine with any more clever way to detect or even avoid the condition in question.
I think the Observables proposal (in stage 1) is intended to offer an alternative to addEventListener for standards that aren't part of the DOM (where EventTarget is standardized). An observable could be used for the event like this: memory.events.change.subscribe(event => { console.log(event.newBuffer); });
I don't think relying on Observable is a good idea; I don't expect them to be soon in JS (if at all) and that would be very uncommon.
The onchange/ongrow callback sounds great. Having to rewrite the binary to call an imported function after each memory.grow sounds very unpredictable (probably a workaround for now).
@NWilson Are you providing JS glue for arbitrary wasm modules? If you're generating the modules yourself it seems like it would be better to add an import hook instead. Even if not, you could create a new module with the hook by searching for all grow_memory instructions.
Yes, for arbitrary modules. The import hook would be awkward, since the grow_memory instructions are inside libc, which shouldn't really call external functions unless necessary... I guess I could maybe make it a Wasm syscall. There is precedent for Wasm-specific syscalls in Musl. Post-processing with an external tool would be another solution, but it would be a bit of a fiddle (doable I agree).
Would this event be fired synchronously? So any function that potentially allocates, and therefore potentially grows memory, could potentially run JS when it didn't before?
In #1271 (comment), it's pointed out that a synchronous event can't work anyway because a different thread can grow the memory.
An asynchronous event seems error-prone: a webassembly function that merely calls malloc or new before calling an imported javascript function that references the webassembly's memory buffer will occasionally fail because malloc/new will sometimes cause the memory to grow, and if the webassembly doesn't release control of the event loop between those steps, there's no opportunity for the asynchronous event to be dispatched and for the javascript to update its buffer references and views. (A grow event might be useful for other things, but I don't think it is for solving the buffer/view-updating problem.)