/cc @ianlancetaylor

Option 3: do nothing. It's an internal API.

If you have examples where it affects user code, that's more interesting.

If we went with option 2, then

	for {
		err = syscall.Sendto(fd.Sysfd, p, 0, sa)
		if err == syscall.EAGAIN {
			if err = fd.pd.waitWrite(); err == nil {
				continue
			}
		}
		break
	}
	if err == nil {
		n = len(p)
	}
	return

Can be rewritten to

	for {
		err := syscall.Sendto(fd.Sysfd, p, 0, sa)
		if err == syscall.EAGAIN {
			if err = fd.pd.waitWrite(); err == nil {
				continue
			}
		}
		return len(p), err
	}

I think there is value in pursuing this.

As far as I can see, the internal/poll package is correctly reflecting the semantics that were implemented for Write, Pwrite, WriteTo, and WriteMsg before the internal poll package was created. That is, internal/poll is just a refactoring of existing functionality, and the functionality of these functions is, as far as I can tell, unchanged. We aren't simply talking about internal/poll; we are talking about the corresponding functions in os and net.

I think it is clear that we can not change the behavior of Write or Pwrite. They can return a number of bytes written and an error. That is permitted by the io.Writer interface. Changing them to not report the number of bytes written on error would be a bug.

WriteMsg, unlike Write and Pwrite, does not use a loop for writing out data. It has a loop, but only to look for EAGAIN. Once WriteMsg succeeds once, it returns. You say above that WriteMsg returns a non-zero count of bytes on error, but that is not accurate. On an error return from the sendmsg system call, we know that no data is wrtten. So WriteMsg is implemented correctly: on error, it reports that no data has been written, which is correct. On success, it reports the length of the data that was written, which must always include all the out-of-band data.

WriteTo, like WriteMsg, does not use a loop. It acts like WriteMsg.

The upshot is that all the functions correctly report the number of bytes written on error. It happens that for WriteTo and WriteMsg the number of bytes written on error is always zero, but for Write and Pwrite it may sometimes be non-zero.

I don't think there is anything to change.

Thanks Ian.