Add a Sync bounds to errors

[sfackler]

Box<Error + Sync + Send> is far more common than Box<Error + Send> - common conveniences like conversion from strings are not
implemented for Box<Error + Send>, Box<Error + Send> cannot be
stored in io::Errors, etc.

This is a breaking change, but non-Sync errors should be rare so the
fallout shouldn't be too bad.

r? @brson I've confirmed that rustup builds without modification with this change.

[Yamakaky]

Do you have an example of code that requires this change?

[sfackler]

Returning an error_chain generated error as a Box<Error + Sync + Send>.

[Yamakaky]

Most errors are Sync I think, since they are read-only?

[sfackler]

Yep!

[Yamakaky]

Please update the changelog.

[sfackler]

Done

[Yamakaky]

common conveniences like conversion from strings are not implemented for Box<Error + Send>

What do you mean?

[sfackler]

This works: let e: Box<Error> = "uh oh".into(), as does let e: Box<Error + Sync + Send> = "uh oh.into(), but not let e = Box<Error + Send> = "uh oh".into().

[Yamakaky]

Hum, strange, do you know why?

[sfackler]

We never bothered implementing it since Box<Error + Send> is exceedingly rare.

[Yamakaky]

Ok, let see if it breaks something important in another crate ;)

[diwic]

JFTR, my main D-Bus error struct won't implement Sync, because it is a wrapper around an FFI struct, and as such I cannot provide any guarantees that the "C" side of the struct is (or will remain) Sync.

[Yamakaky]

That's what I feared...

[durka]

What is the suggested workaround for errors that cannot be made Sync? Can I just not use error-chain anymore?

[Yamakaky]

I'll just remove it in 0.9.

[sfackler]

Things break without the bound just as they do with it. Can we make the error bounds configurable in the error_chain! {} macro?

[Yamakaky]

Yes, see #121.

[jsgf]

I ran into this because std::io::Error requires a nested error to be Sync. It makes it hard to incorporate error-chain errors into interfaces which return std::io::Result.

[Yamakaky]

See #134

[rnewman]

Reverting this means that error_chain can't easily be used with combine. In Mentat we've had to roll back to 0.8.1 until this is fixed: mozilla/mentat#418