Make sure to sync on file-io.rs tokio test

[Noratrieb]

Tokio AsyncWriteExt::write doesn't actually ensure that the contents have written, it just starts the write operation. To ensure that the file has actually been written, we need to sync_all first.

actually fixes #4367

[tiif]

To add more context, me and nora tried

async fn test_create_and_write() -> io::Result<()> {
    let path = utils::prepare("foo.txt");
    let mut file = File::create(&path).await?;

    // Write 10 bytes to the file.
    let a = file.write(b"some bytes").await?;
    assert_eq!(a, 10);
    assert_eq!(file.metadata().await.unwrap().len(), 10);
    
    remove_file(&path).unwrap();
    Ok(())
}

and the test only fails on second assert in certain seeds. That means the file metadata is not completely synced before we access it.

[RalfJung]

Isn't that a tokio bug? When a single task creates a file and then tries to read it, surely that should just work?

[Noratrieb]

The way the tokio code is written makes me think that this is deliberate behavior. I haven't found any docs about this though, so I will report that to tokio to at least ask them to document this behavior.

[RalfJung]

Not having program order establish happens-before is at least rather unconventional...

Cc @Darksonn

[Noratrieb]

filed tokio-rs/tokio#7378

[RalfJung]

Is this PR required to un-break CI or should the other one be enough?

[Darksonn]

Happy to improve docs, but this is intended behavior. Tokio's file behaves like BufWriter<File> in std.

[Noratrieb]

Is this PR required to un-break CI or should the other one be enough?

this one is required. i was able to produce failures with write_all locally (though it might be possible that just write_all reduces the amount of failures)