Port to rustix 1.0.

[sunfishcode]

rustix 1.0 introduced a new Buffer API which is used in epoll::wait, kqueue, and polling, so this patch updates the code to use the new spare_capacity mechanism for writing to the spare capacity of a Vec.

rustix 1.0 also changes timeout parameters from i32 milliseconds to Timespec seconds plus nanoseconds, which is more general and aligns with newer Linux syscalls such as epoll_pwait2.

This also fixes a bug in the wait function; in rustix 0.38, calling epoll::wait would clear the events list before adding new events, however polling's documentation states that "New events will be appended to events." Using spare_capacity in rustix 1.0, the list is not cleared first, so events are appended.

Closes #234.

[sunfishcode]

I've now also tested this on illumos, to test the port backend. Everything passes except for the insert_twice test, however this doesn't appear to be a regression. The code is checking that inserting a fd twice gets an AlreadyExists error, however port_associate doesn't fail when the fd is already associated, and the port.rs code doesn't have any extra bookkeeping.

[taiki-e]

Thanks! This is generally LGTM.

[fogti]

For the mentioned problem, I think it would be a better idea to put rustix::event::Timespec into our public API, and re-export it from the crate root, instead of wrapping the fallible conversion.

Alternatively, putting a wrapper type for that into the crate root, and expose appropriate conversions.

[taiki-e]

I think the generally preferred behavior regarding time overflow is saturation or close to it.
smol-rs/async-io#87
smol-rs/async-io#86 (comment)

[sunfishcode]

Makes sense; I've now changed the code to handle Duration->Timespec overflow by saturating to an infinite timeout.

The i686-unknown-freebsd failure looks real. I'm investigating.

[sunfishcode]

The i686-unknown-freebsd failure is real; I'm working on a fix.

[taiki-e]

Thanks!

[fogti]

Sorry, I currently don't have time to properly review this (as I'm trying to finish a thesis); maybe after next week. I wish I'd knew how to test this for Hermit, but I currently don't.

[fogti]

Manual test suite run passes on x64_64-unknown-freebsd and x86_64-unknown-linux-gnu.

Please write a CHANGELOG.md entry for this, given that it changes how event vectors are handled: they aren't cleared anymore, this is a bugfix that brings the code back in line with what it should be, but that itself doesn't prevent downstream from accidentally ending up relying on the former, buggy behavior. (non-blocker)

[taiki-e]

Published in 3.8.0.