fix compilation on uefi targets

[usamoi]

• Added a CHANGELOG.md entry

Summary

On uefi targets, the return type of std::io::Error::raw_os_error is Option<usize>, instead of Option<i32>. In rand_core, it is assumed as Option<i32>. This PR uses Option<usize> as the return type on the uefi targets.

getrandom has already handled this: https://github.com/rust-random/getrandom/blob/18d89843981b93032b2a2c6f1e33897075a8d727/src/error.rs#L6-L19

[dhardy]

This doesn't feel like the right way to do this: clearly #1537 failed to isolate rand from getrandom changes.

Do you need a patch release for rand_core?

[newpavlov]

Note that technically it's a breaking change. I think we can cfg the implementation for UEFI and replace it with panic and fix it properly in the next breaking release.

UPD: On the second thought, since the current code does not compile on the UEFI targets, then this change can be considered non-breaking since it only affects "broken" targets.

[usamoi]

Do you need a patch release for rand_core?

No. I can use the patch locally, so this is not necessary for me.

Note that technically it's a breaking change.

I'm fine with waiting to merge it into 0.10.

[dhardy]

@usamoi could you add a note to rand_core/CHANGELOG.md please?

@newpavlov I think RawOsError should be a pub type in getrandom and re-exported in rand_core to avoid this: it's a foot-gun that the docs just say raw_os_error returns Option<i32>. This is what e.g. libc does for platform-dependent types.

[newpavlov]

@dhardy
I don't think it makes sense to both re-export RawOsError and wrap getrandom::Error. I think we should either expose getrandom::Error in rand_core (in the next breaking release) or duplicate RawOsError.

[dhardy]

@newpavlov the argument not to export getrandom::Error was to minimise the chance that a breaking release of getrandom would be breaking in rand_core. So far that hasn't actually helped us though.

getrandom should make the RawOsError type explicit anyway IMO.

[dhardy]

So there is a fundamental difference between a pub-export of getrandom::Error and getrandom::RawOsError: the former is a type, the latter an alias. In other words, getrandom_v0_3::Error and getrandom_v0_3::Error will be distinct types; getrandom_v0_3::RawOsError and getrandom_v0_3::RawOsError will not be.

Implication: depending on getrandom::Error automatically makes a bump to the getrandom (non-patch) version breaking; depending on getrandom::RawOsError does not (so long as the aliases are compatible).

Once getrandom v1.0 is out I'm fine with depending on getrandom::Error. See also #1537.

[dhardy]

Marking as 👍 but if the next getrandom release is made soon this may be replaced.