Get rid of the getrandom and OsRng requirement

[stevefan1999-personal]

This PR solves the getrandom dependency problem, due to OsRng being added to the library code itself. Since OsRng assumed to be running under an operating system, this makes it impossible to run in freestanding environment such as bare-metal kernel and isolated raw WebAssembly.

It has now been removed and the dependency on RNG source in the provider itself has been inversed, meaning that the user has to supply a RNG provider from the start in the initial provider initialization call, and during execution, we take get take back the RNG implementation in rustls global code.

Assuming the code that needs randomness in this library would only be called by the library itself (e.g. RustCrypto Provider -> ECDHE key gen), we can safely assume there has to be a crypto provider and it has to be our own library.

no_std users can still enable getrandom but they have to make the implementation manually. This can simply be done by adding the getrandom crate in their end, and then set the default RNG using OsRng themselves, rather than pushing the responsibility to us to always assume OsRng to be available.

Now, it is really debatable whether CryptoRng is also needed or not for RNG, as it is intended for embedded system where getrandom is unlikely to be ported to the platform, so the user might just want to use a good initial entropy (like from CPU jitter or other micro-movements) then shove it to a PRNG that is fast and generates random numbers for the rest of the lifetime. This is not cryptographically safe at all, but so far this is a necessary evil if you want it to run in no_std environment where the existence of a good CSPRNG implementation is not guaranteed.

Thus, impl CryptoRng for CryptoProviderRng can be considered a phony implementation out of necessity for the code to compile. It is now user-controlled and does not guarantee the output to be of cryptographic quality.

[newpavlov]

I don't think I like this approach. It feels really hacky to me.

What exact problem are you trying to solve? Building/testing on no_std targets?

[stevefan1999-personal]

@newpavlov Agreed. I rethought about it and changed it like that without using OnceCell

[newpavlov]

I had a more fundamental question: why exactly do you want to make dependency on getrandom optional?

If you want a way to change entropy source (e.g. for deterministic testing), then I think its better to introduce ProviderWithRng generic over CryptoRng + Default and define alias type Provider = ProviderWithRng<OsRng>; gated on the getrandom feature. With this approach you can define a testing CSPRNG seeded with a fixed seed and state stored in a static.

[stevefan1999-personal]

I had a more fundamental question: why exactly do you want to make dependency on getrandom optional?

If you want a way to change entropy source (e.g. for deterministic testing), then I think its better to introduce ProviderWithRng generic over CryptoRng + Default and define alias type Provider = ProviderWithRng<OsRng>; gated on the getrandom feature. With this approach you can define a testing CSPRNG seeded with a fixed seed and state stored in a static.

For one, it broke no_std as the automatic target selection of getrandom looks very arcane (https://docs.rs/getrandom/latest/getrandom/), that means whenever you have a new triplet you are not likely to port getrandom on day one, and second it assumes the existence of an operating system, at least within the definition of rand_core, which is not always the case (as I raised the example of using it in MCUs).

But I do think your approach would also be alright, let me try that out as well.

[newpavlov]

it broke no_std as the automatic target selection of getrandom looks very arcane

Do you perhaps mean the custom backend? For most targets supported out-of-box automatic selection is a pretty straightforward cfg_if!.

As for the custom backend, I agree that it's not the most straightforward thing, but, unfortunately, Rust does not provide good tools for this problem. We tried to describe how to enable custom backends in the register_custom_getrandom! docs. I don't think it's the most complex thing required for configuring a no_std target. Note that there is also an additional escape hatch in the form of patching the getrandom dependency by a custom crate.

that means whenever you have a new triplet you are not likely to port getrandom on day one

It's no surprise that you need additional configuration on no_std targets. I think it's much better to do it in one place than have several crates to invent their own ad hoc way of selecting user-defined entropy sources.

second it assumes the existence of an operating system

No, it does not. OsRng is a somewhat misleading name, but it works just fine on no_std targets assuming getrandom was configured properly.