IMO write() should write everything, otherwise it is just too surprising. Doesn't it do that currently?
For really specialistic use cases writing "not all" could be appropriate, but that should be very obvious from the name of the function (e.g. writePartial).

@piscisaureus Not really. For example, write on a Conn have the following layers of implications (assuming Unix only):
ConnImpl.write ->
tokio_write::write ->
<Resource as AsyncWrite>::poll_write ->
tokio::net::TcpStream::poll_write ->
tokio::net::TcpStream::write ->
PollEvented<mio::net::TcpStream>::write ->
mio::net::TcpStream::write ->
sys::TcpStream::write ->
std::net::TcpStream::write ->
std::sys_common::net::TcpStream::write ->
libc::send

As it eventually goes to send (2) and none of the above step does a write-all attempt, it is possible to write and return partially if something like signal interrupts happens when we have already written some data (from manpages on macOS, EINTR only triggers on no data written yet:)

[EINTR]            A signal interrupts the system call before any data is transmitted.

I've rebased on current master.