Stream: wasi

Topic: bindgen + resources

Andrew Brown (Jun 13 2024 at 18:14):

I ran into an issue where implementing impl<T: WasiNnView> gen::tensor::HostTensor for T returned the following error:

error[E0119]: conflicting implementations of trait `HostTensor` for type `&mut _`
   --> crates/wasi-nn/src/wit.rs:26:5
    |
26  | /     wasmtime::component::bindgen!({
27  | |         world: "ml",
28  | |         path: "wit/wasi-nn.wit",
29  | |         trappable_imports: true,
30  | |     });
    | |______^ conflicting implementation for `&mut _`
...
265 |   impl<T: WasiNnView> gen::tensor::HostTensor for T {
    |   ------------------------------------------------- first implementation here
    |
    = note: this error originates in the macro `wasmtime::component::bindgen` (in Nightly builds, run with -Z macro-backtrace for more info)

This is only a problem now that I started using resources in the WIT file. I noticed that wasmtime-wasi-http goes about this with a wrapper struct so I decided to add the following:

pub struct WasiNnImpl<T>(pub T);
impl<T: WasiNnView> gen::tensor::HostTensor for WasiNnImpl<T> { ... }

Is this the right approach? I don't seem to see the same kind of wrapping in wasmtime-wasi–how come?

Andrew Brown (Jun 13 2024 at 18:15):

cc: @Pat Hickey?

Alex Crichton (Jun 13 2024 at 19:14):

If you can avoid it I'd skip WasiNnView entirely

Alex Crichton (Jun 13 2024 at 19:14):

For resources, if you can get away with it, I might recommend:

pub struct WasiNnView<'a> {
    pub table: &'a mut ResourceTable,
}

impl gen::tensor::HostTensor for WasiNnView<'_> { /* ... */ }

Andrew Brown (Jun 13 2024 at 19:15):

hm...

Alex Crichton (Jun 13 2024 at 19:15):

(sorry editing the message above)

Alex Crichton (Jun 13 2024 at 19:15):

and then add_to_linker would look like:

fn add_to_linker<T>(linker: &mut Linker<T>, f: fn(&mut T) -> WasiNnView<'_>) { /* ... */ }

Alex Crichton (Jun 13 2024 at 19:16):

for wasmtime-wasi we can in theory do this

Andrew Brown (Jun 13 2024 at 19:16):

Why do I see views in wasmtime-wasi and wasmtime-wasi-http?

Alex Crichton (Jun 13 2024 at 19:16):

but wasmtime-wasi-http we can't do this naively because it has customization hooks via a trait

Alex Crichton (Jun 13 2024 at 19:16):

so if you want customization hooks for wasi-nn eventually on the embedder size you can go the route of a WasiNnView trait

Alex Crichton (Jun 13 2024 at 19:16):

but if you don't know of any right now it's much easier to stick with concrete types/structures

Andrew Brown (Jun 13 2024 at 19:18):

what type of customization hooks does wasi-http need? where can I look at that?

Andrew Brown (Jun 13 2024 at 19:19):

oh, like each embedder of Wasmtime may want to "do special stuff" for each HTTP request so they provide their own implementation of WasiHttpView?

Alex Crichton (Jun 13 2024 at 19:21):

for wasi:http it's primarily send_request

Alex Crichton (Jun 13 2024 at 19:22):

it's basically which layer of abstraction you want to work at

Alex Crichton (Jun 13 2024 at 19:22):

so for example embedders can implement all the wasi traits themselves whenver they want

Alex Crichton (Jun 13 2024 at 19:22):

that's a bit of a big task though so ideally you want to use wasmtime-wasi

Alex Crichton (Jun 13 2024 at 19:22):

but then often embedders want to customize just a few things here and there

Alex Crichton (Jun 13 2024 at 19:22):

for some requests we can do that via WasiCtx and various configuration variables in there

Alex Crichton (Jun 13 2024 at 19:23):

but for other cases you want a full-blown host-defined callback

Alex Crichton (Jun 13 2024 at 19:23):

but for callbacks it's a question of how do you close over the interesting state, and threading &mut T from Store<T> isn't viable here

Alex Crichton (Jun 13 2024 at 19:23):

so the next best thing we've "figured out" is trait FooView

Alex Crichton (Jun 13 2024 at 19:23):

where the implementation of FooView has whatever state it needs and the callbacks are trait methods

Alex Crichton (Jun 13 2024 at 19:24):

overall it's not a great design but IMO it's the best we've got so far and it's a reflection of a tricky problem space

Andrew Brown (Jun 13 2024 at 19:27):

ok, I see what you mean; it seems like the approach should be "start with the concrete type" and eventually--and only if needed--"refactor with a view trait"

Alex Crichton (Jun 13 2024 at 19:34):

Yeah if that works it's the path I'd recommend, dealing with blanket impls and extra customization generally leads to headaches

Andrew Brown (Jun 13 2024 at 23:44):

Oh, I realized there might be another reason for creating a WasiNnView other than customizability. For wasmtime run and wasmtime serve, we end up creating a struct Host with a single ResourceTable to be shared between all the WASI proposals that are enabled at the time. In that case, we _do_ end up needing a bridge to the shared table and the current approach is via *View traits from each proposal. I think I do need to implement WasiNnView then, right?

Alex Crichton (Jun 13 2024 at 23:47):

Not necessarily, that's the 'a in WasiNnView<'a> above, you wouldn't actually store WasiNnView anywhere but it'd only be created as a temporary in a closure

Alex Crichton (Jun 13 2024 at 23:47):

but it'd borrow a ResourceTable from elsewhere

Andrew Brown (Jun 13 2024 at 23:47):

ah, ok!

Last updated: Nov 22 2024 at 17:03 UTC