0.38 release planning

[sunfishcode]

As of this writing, Rustix 0.37 was released about a month ago, and many crates started adopting it a few days ago. We hopefully won't need another semver bump for quite a while.

This issue is for collecting changes that we should make in 0.38, when the time comes. So far, I'm aware of:

• chmodat_with introduced in FreeBSD RESOLVE_BENEATH support and more #541 should be renamed to just chmodat.
• The epoll EventVec iterator should return Events rather than tuples of their parts. If we publicly expose the Event struct, we can also add a EventVec::from_raw_parts, which is needed by mustang.
• Update to bitflags 2.x
  • Use the new - BitFlags trait to simplify conv.rs
• Harmonizing the types between libc and linux_raw as discussed in Better factoring for flags code #513
  • Some of this has been done; there's more to do, but we don't need to block 0.38 on it.
• Improve strict-provenance compatibility in the epoll API. #693
• Finer-grained API modules
  • Consider splitting some things out of rustix::io, such as the various polling APIs into a rustix::poll or so, and maybe moving the procfs API to rustix::procfs.
  • Split "system" and "membarrier" APIs out of process and into a new system module. Maybe also fs::sync?
• Rename getsockopt_nosigpipe and setsockopt_nosigpipe to get_socket_nosigpipe and set_socket_nosigpipe, for consistency with the rest of the getsockopt/setsockopt function names.
• Rename sendmsg_noaddr to sendmsg
• - AcceptFlags is now just an alias for - SocketFlags, so we should remove it.
• Move poll, epoll, kqueue, and port out of rustix::io, and put them behind a feature flag
• Overhaul the termios API. #615
• Pid::from_raw{,_nonzero} should not be unsafe #604
• Improve perf of all buffer consuming APIs #502
• Bitflags workarounds #695

[valpackett]

• Consider changing kqueue kevent(…) API a little
  • It's strange that changelist and eventlist aren't Option<>; normally when you manage events you pass NULL, 0 as eventlist, nevents and when you wait for events you pass NULL, 0 as changelist, nchanges. Dummy references to unused things are mehhh
  • I'm not a fan of the changelist vec capacity thing, the implicit requirement to allocate vec capacity isn't obvious, it's a footgun: if you try to wait for events but you pass a vec with 0 capacity, you'll spin endlessly instead of waiting. I've documented this in Add kqueue signal filter #582 but I'm not sure what the best solution would be
    • just implicitly reserving some small number sounds reasonable I guess?
    • or get rid of the vec and use a slice, making it a lot more obvious? (also this would allow using arena allocators etc. instead of vec, haha)

[notgull]

Something to consider might be making it so you can store epoll/kqueue/etc events on the stack instead of in a Vec. Would probably require a trait over some kind of ArrayVec

[valpackett]

kqueue only uses Vec for the length manipulation, we could really just extract that part and offer an fn kevent_raw or something (name subject to bikeshedding) without the semver bump

[sunfishcode]

The 0.38-planning branch is now landed on main. There are a few more changes listed above that I'd like to get in before the 0.38 release, as listed in the top comment here, but hopefully those won't take long.

[sunfishcode]

Drafting a 0.38 API changes guide (I'll add this to the github release notes too):

• Meta:

  • Rust 1.63
  • bitflags 2
  • significant compile-time improvements

• Some APIs were factored out into new toplevel API modules, with their
  corresponding cargo feature flags:

  • rustix::event, for epoll, eventfd, and poll
  • rustix::pipe, for pipe, tee, splice, and related functions
  • rustix::procfs, for procfs-related APIs
  • rustix::stdio, for *std{in,out,err}
  • rustix::system, for uname, sysinfo, and sethostname

• rustix::io changes:

  • SeekFrom was moved to rustix::fs.

• rustix::net changes:

  • socket now takes an Option<Protocol> and Protocol no longer implements Default; change Protocol::default() to None.
  • sendmsg_noaddr is renamed to sendmsg.
  • getsockopt_nosigpipe and setsockopt_nosigpipe are renamed to get_socket_nosigpipe and set_socket_nosigpipe.
  • AcceptFlags is renamed to SocketFlags.

• rustix::event changes:

  • epoll changes:
    • epoll::epoll_add, epoll::epoll_mod, and epoll::epoll_del were renamed to epoll::add, epoll::modify, and epoll::delete.
    • The EventVec iterator now returns Events rather than tuples.
    • epoll::Event::data is changed from a u64 to a custom EventData type which can hold either a u64 or a *mut c_void.

• rustix::fs changes:

  • rustix::fs::chmodat_with is renamed to rustix::fs::chmodat.
  • rustix::fs::cwd() is replaced with rustix::fs::CWD.
    • Non-*at versions of several functions that implicitly operate on CWD were also added.

• rustix::termios changes:

  • The API now fully supports arbitrary speeds. The cfsetispeed, cfsetospeed, cfsetspeed, tcsetattr2, and tcgetattr2 functions, as well as the CBAUD and CIBAUD constants, are replaced by set_input_speed, set_output_speed, and set_speed functions on Termios.
  • The speed macros such as B9600 are no longer required (B9600 now always has the value 9600, and so on for all the other speed macros), but they can still be used. They're now in the rustix::termios::speed module.
  • cfmakeraw is now called Termios::make_raw.

• rustix::process changes:

  • Pid::from_raw is no longer unsafe.
  • Pid::from_raw_nonzero is removed, and a new Pid::from_raw_unchecked is added.
  • ReaperStatus::pid is now an Option<Pid>, as the OS doesn't always provide it.

• rustix::io_uring changes:

  • io_uring_cqe now supports the big_cqe field, and no longer implements Copy and Clone.

[sunfishcode]

There's now a 0.38.0-alpha.2 prelrelease up with everything I'm planning to include in the 0.38.0 release, except for #695.

[sunfishcode]

rustix 0.38.0 is now released! See the release notes for the list of user-visible changes.