Add opt_in crate feature

[newpavlov]

This feature resolves the lock file bloat issue at the cost of less ergonomic use of opt-in backends.

TODO: update docs

[dhardy]

I think it would make more sense to call this feature wasm_backends. In case another backend were added / made optional, we'd then add another feature for that backend.

I also don't see much point in opt_in affecting other backends like rdrand which aren't adding dependencies.

[josephlr]

I think the ideal situation would be to get @dhardy's idea in #433 (comment) to work somehow. That way an external crate (say getrandom-js) could have code like:

use getrandom::{register_backend,Error};register_backend!("js", web_crypto_fill);unsafeextern"C"fnweb_crypto_fill(ptr:*mutu8,len:size) -> Option<Error>{// ...}

And then passing --cfg getrandom_backend="js" would "just work". This would also remove the concept of a "custom" backend, as any custom backend could be declared and used in the same way. The "rdrand", "rndr", and "linux_getrandom" would stay as-is being "builtin" alternative backends.

[dhardy]

@josephlr: I don't think register_backend! can work like that: IIUC the only real mechanism we have is to define a #[no_mangle] fn with a precise signature exactly once in the binary.

[newpavlov]

I think it would make more sense to call this feature wasm_backends.

The opt_in feature is not WASM-specific. It allows users to enable optional backends, including the ability to overwrite the default backends. The idea here is to have a unified approach which would be the same for any potential future opt-in backend which may bring additional dependencies (e.g. like linux_rustix) without proliferation of opt-in-specific features.

@josephlr
I am not sure how your idea would work. Can you explain it in a more detail?

As I wrote here, one alternative approach is to move WASM JS implementation into a separate crate and rely on the custom backend.

[dhardy]

The opt_in feature is not WASM-specific. It allows users to enable optional backends, including the ability to overwrite the default backends. The idea here is to have a unified approach which would be the same for any potential future opt-in backend which may bring additional dependencies (e.g. like linux_rustix) without proliferation of opt-in-specific features.

Using a separate feature flag for each backend (or group of backends for some target platform) makes more sense IMO, so e.g. wasm_backends and rustix_backends (or just js and rustix).

Unless your intention is that --cfg getrandom_backend="X" only be usable when opt_in is enabled... but I'm not sure we can lint that properly (lints.rust.unexpected_cfgs can't be affected by feature flags I think).

[newpavlov]

Using a separate feature flag for each backend (or group of backends for some target platform) makes more sense IMO, so e.g. wasm_backends and rustix_backends (or just js and rustix).

What about the custom and rdrand/rndr backends? Note the potential security implications of using crate features for those.

Unless your intention is that --cfg getrandom_backend="X" only be usable when opt_in is enabled...

Yes, it's exactly my intention.

but I'm not sure we can lint that properly (lints.rust.unexpected_cfgs can't be affected by feature flags I think).

Hm, I don't see why it would be a problem. IIUC unexpected_cfgs does not require for listed cfgs to be used in code.

UPD: I updated the code and, as you can see, CI passes without any lint warnings.

[newpavlov]

Closing in favor of #574.