Tracking Issue for `core_io_borrowed_buf`

[jmillikin]

View all comments

Feature gate: #![feature(core_io_borrowed_buf)]

This is a tracking issue for an MVP of core::io, which contains an OS-independent subset of std::io.

Public API

The initial API of this module consists of BorrowedBuf and BorrowedCursor, which were previously only available in std.

Steps / History

• std::io: migrate ReadBuf to BorrowedBuf/BorrowedCursor #97015
• Move BorrowedBuf and BorrowedCursor from std:io to core::io #117694
• Final comment period (FCP)¹
• Stabilization PR

See also: #78485

Unresolved Questions

• Now that it's generic (Make BorrowedBuf and BorrowedCursor generic over the data #149749) double-check any bounds we might want before stabilization
  • For example, it used to always be Debug (which was easy since u8: Debug), but for generic do we want always-Debug or only for T: Debug?

Naming:

• BorrowedCursor::written needs renaming to buf_written (implementation work only) Tracking Issue for core_io_borrowed_buf #117693 (comment)
• core::io or core::mem? Tracking Issue for core_io_borrowed_buf #117693 (comment), #t-libs > `io::BorrowedBuf` or `mem::BorrowedBuf`
• Should BorrowedCursor be renamed to BorrowedWriter? It's more similar to BufWriter than Cursor Tracking Issue for core_io_borrowed_buf #117693 (comment), #t-libs > Why is it called `io::BorrowedCursor`?
• General name confusion, specifically read_buf Tracking Issue for core_io_borrowed_buf #117693 (comment) (Concern 5)

Documentation:

• All API is missing examples, some is missing tests (implementation work only) Tracking Issue for core_io_borrowed_buf #117693 (comment) / Tracking Issue for core_io_borrowed_buf #117693 (comment)
• General documentation issues: needs to clarify why these two types exist rather than being combined Tracking Issue for core_io_borrowed_buf #117693 (comment), what exists is confusing and methods are not in useful order Tracking Issue for core_io_borrowed_buf #117693 (comment) (concern 1)

API confusion:

• BorrowedCursor::capacity indicates remaining capacity rather than total capacity, unlike Vec Tracking Issue for core_io_borrowed_buf #117693 (comment)
• Feature split with borrowed_buf_init is confusing: advance and advance_checked are split, affects as_mut Tracking Issue for core_io_borrowed_buf #117693 (comment), Tracking Issue for core_io_borrowed_buf #117693 (comment) (concern 3)
• Do we need both into_filled and into_filled_mut? Tracking Issue for core_io_borrowed_buf #117693 (comment)
• Should we add inherent constructors to avoid inference issues? (Not blocking) Tracking Issue for core_io_borrowed_buf #117693 (comment)
• Should we have is_empty? (Not blocking) Tracking Issue for core_io_borrowed_buf #117693 (comment)
• Unnecessary return of self Tracking Issue for core_io_borrowed_buf #117693 (comment)
• written confusion, position as an alternative Tracking Issue for core_io_borrowed_buf #117693 (comment)

Variance:

• We have BorrowedCursor<'data>, which would be invariant, but make it covariant with a lot of unsafe. This is fragile compared to BorrowedCursor<'buf, 'data> Tracking Issue for core_io_borrowed_buf #117693 (comment), Tracking Issue for core_io_borrowed_buf #117693 (comment), #t-libs > Why is it called `io::BorrowedCursor`? @ 💬

Trait and bounds:

• Should we print the filled portion during Debug? We don't have T: Debug in the current impl Tracking Issue for core_io_borrowed_buf #117693 (comment)
• Needless gates on T: Copy Tracking Issue for core_io_borrowed_buf #117693 (comment)

Other:

• This API has been contentious: confirm it is what we want over something like #t-libs > `buffer-trait` would overlap with `BorrowedBuf`

Footnotes

1. https://std-dev-guide.rust-lang.org/feature-lifecycle/stabilization.html ↩

[tgross35]

These types in core could use a core::fmt::Write impl, but I am not sure how this is possible without conflicting with the blanket impl for std::io::Write.

[taralx]

Is the intent to stabilize this at the same time as #78485 or independently of it?

[kornelski]

I'm surprised that this doesn't have any safe integration with Vec. I'd like to have vec.as_borrowed_buf() that can safely append to the uninitialized capacity, without having to manually perform unsafe vec.set_len correction afterwards.

[keepsimple1]

Could you please add a link in the summary to the PR that originally added BorrowedBuf and BorrowedCursor ? (As both this move and the original changes are unstable, it is somewhat difficult to track back the changes).

I believe the link is: #97015

[keepsimple1]

If I understand, BorrowedBuf (like MaybeUninit) essentially is safe to write, may I suggest adding a method to the API of BorrowedBuf:

/// Copies the bytes of `buf` to self starting at `offset`.
///  Returns the number of bytes actually written.
///
/// If `offset` is equal or larger than `self.buf.len()`, returns 0.
/// If `offset` is larger than `self.init`, the region `self.buf[self.init..offset]` is zeroed. 
pub fn write(&mut self, offset: usize, buf: &[u8]) -> usize

(this method updates filled and init accordingly)

The idea is to make the API even easier to use:

• Simpler: no need to use BorrowedCursor for such use cases.
• No need to use different APIs to write filled / init section and uninit sections.
• No panics (from this method) as it returns the written length.
• Compatible with existing API.

What do you think?