std::io::Write::write_vectored() to io::stdout() does not produce all expected lines.

[rustyconover]

When trying to write an array of 8192 IoSlices consisting of "1" followed by "\n" only the first character is written on Mac OS X.

use std::io;
use std::io::prelude::*;
use std::io::IoSlice;

fn main() {
    let mut write_array: Vec<IoSlice> = Vec::new();
    for _ in 1..1024 * 4 {
        write_array.push(IoSlice::new(b"1"));
        write_array.push(IoSlice::new(b"\n"));
    }
    io::stdout().write_vectored(&write_array).unwrap();
    io::stdout().flush().unwrap();
}

Output:

1

Just one byte is written:

./target/debug/vector-bug | wc 
       0       1       1

I expected 4096 pairs of "1\n" to be written.

Meta

$ rustc --version --verbose
rustc 1.40.0 (73528e339 2019-12-16)
binary: rustc
commit-hash: 73528e339aae0f17a15ffa49a8ac608f50c6cf14
commit-date: 2019-12-16
host: x86_64-apple-darwin
release: 1.40.0
LLVM version: 9.0

[jonas-schievink]

On nightly this panics with an InvalidInput I/O error instead (at least on the playground).

[tmiasko]

The write_vectored was implemented for LineWriter in 2fee28e, which changed behaviour here.
Now all buffers will be written at once and most likely trigger OS specific limit as described in #68042.

[sfackler]

Like write, write_vectored does not necessarily write the entire set of data out. You need to check the value it returns.

[jonas-schievink]

This behavior does not match the documentation:

Data is copied from each buffer in order, with the final buffer read from possibly being only partially consumed.

[sfackler]

"final buffer" means the final buffer that was read from, not the last buffer in the input slice. A concrete example of short writes beyond the OS doing that is the default implementation:

The default implementation calls write with either the first nonempty buffer provided, or an empty one if none exists.

[rustyconover]

Would there be opposition to adding write_all_vectored() rather than having everybody implement the same logic of retrying until all data is written?

[sfackler]

I think that would definitely make sense, though the implementation would be pretty tricky. Even more reason that it should be done once though I guess!

[rustyconover]

I'll work on writing it.

…

On Thu, Jan 9, 2020 at 10:33 AM Steven Fackler ***@***.***> wrote: I think that would definitely make sense, though the implementation would be pretty tricky. Even more reason that it should be done once though I guess! — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#68041?email_source=notifications&email_token=AAFSWJOMZWO2OR3ASGQX523Q447THA5CNFSM4KESBSX2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEIQWNWQ#issuecomment-572614362>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAFSWJNS4IP2OXNVEMAQGDDQ447THANCNFSM4KESBSXQ> .

[rustyconover]

It seems others are working on it in #62726, but are blocked since they can't advance the pointer inside of the pending or partially written IoSlice.

[rustyconover]

Here is an example write_all_vectored that works with nightly:

fn write_all_vectored(file: &mut File, mut bufs: &mut [IoSlice<'_>]) -> std::io::Result<()> {
    while bufs.len() > 0 {
        match file.write_vectored(bufs) {
            Ok(0) => {
                return Err(std::io::Error::new(
                    std::io::ErrorKind::WriteZero,
                    "failed to write whole buffer",
                ));
            }
            Ok(n) => bufs = IoSlice::advance(bufs, n),
            Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
            Err(e) => return Err(e),
        }
    }
    Ok(())
}

[sfackler]

That modifies the input slice array, which may or may not be a thing we'd want in an implementation in Read itself.