wasi-nn: backend context leak

[yamt]

it seems that backend context is overwritten and leak if load is called multiple times by an instance.
i guess the support of multiple models is completely broken.
seemingly a regression made by #3692

[lum1n0us]

it seems that backend context is overwritten and leak if load is called multiple times by an instance.

Could you be more specific about it? I'm not quite following.

[yamt]

the first load call initializes wasi_nn_ctx->backend_ctx.
the second load call overwrites it. (and make the previous value leak)
at least that's what i understand by reading the code.

[lum1n0us]

Like continuously calling wasi_ephemeral_nn.load() several times? You are right, there is no protection to handle that. However, it seems the current wasi-nn design expects a fixed calling sequence. I'm not sure if we should implement such protection.

[yamt]

However, it seems the current wasi-nn design expects a fixed calling sequence.

where did you get the impression from?

I'm not sure if we should implement such protection.

our tensorflowlite backend seems to have some logic to allow an instance to load multiple graphs.
i suspect it used to work. (ie. this is a regression)

[yamt]

anyway, even if we don't want to support the operation, it's better to reject it gracefully than leaking.

[yamt]

fixed by #4366