Should we bump the MSRV for OwnedFd?

[notgull]

Rust 1.63.0 was just released, and the OwnedFd struct is now stable. Given that it essentially replaces the RawFdContainer type that we currently use, is it worth it to bump the MSRV to 1.63.0 in order to use it?

Pros:

• As the remainder of the Rust ecosystem (e.g. nix) evolves, it will likely used BorrowedFd and OwnedFd instead of RawFd. We don't want an ecosystem split
• We will no longer be responsible for the invariants of RawFdContainer, since it will be on the standard library's back.
• OwnedFd is niched, meaning that the layout of OwnedFd is optimized.
• Bumping the MSRV to 1.63.0 would mean we could use some of the features mentioned in Tracking of possibly useful post-1.56 features #538

Cons:

• Converting a RawFd to an OwnedFd requires unsafe, meaning that, unless nix also returns OwnedFd from its ancillary message function, we'd have to introduce unsafe code.
• Bumping the MSRV past 1.57.0 would mean that x11rb would not build with the version of rustc currently in the Ubuntu package repository.

[psychon]

I updated #538 for this. It now lists OwnedFd as a useful feature for Rust 1.63.

I have another "Con" to add: I would no longer be able to build x11rb locally.

As the remainder of the Rust ecosystem (e.g. nix) evolves, it will likely used BorrowedFd and OwnedFd instead of RawFd. We don't want an ecosystem split

Sorry, but I do not buy this argument at the present time. Yes, eventually we'd want to stop using RawFd, but for now it is Rust 1.63 that causes a split. All of the ecosystem right now uses RawFd.

For nix specifically: They are at 1.46 currently. I doubt that they would do such a large jump immediately. I did not find any issue/PR about RawFd there. Seems like they are in the process of raising MSRV to 1.51 due to an accident at serde: nix-rust/nix#1792

Edit: Now that issue talks about bumping nix to 1.56. Also, in that issue, I found rust-lang/libs-team#72 which talks about an MSRV policy for libc. That seems to be "bleeding edge" enough to also influence us.

[ids1024]

Looks like https://packages.debian.org/bookworm/rustc has 1.63 now. But it's also possible to use the io-lifetimes crate to support older versions, which some libraries that don't want to require 1.63 are doing.

winit (and softbuffer) is targetting 1.60.0 currently, so if it will be using x11rb (rust-windowing/winit#2614), it's best for x11rb to not require a newer version, unless that is bumped soon.

[psychon]

But it's also possible to use the io-lifetimes crate to support older versions

Oh, wow. I like that idea and went ahead and implemented it. However.... thanks to FromRawFd::from_raw_fd being unsafe, this breaks the crate (#![cfg_attr(not(feature = "allow-unsafe-code"), forbid(unsafe_code))]). x11rb::rust_connection::Stream currently implements FD-passing and abstract socket support thanks to nix making all necessary ingredients safe. That doesn't seem possible anymore with io-lifetimes.

I pushed my work to https://github.com/psychon/x11rb/tree/owned-fd.

[notgull]

I've been trying to get sendmsg/recvmsg into rustix, see bytecodealliance/rustix#505. rustix uses io-lifetimes by default, which would circumvent this issue.

[psychon]

Closed by #815