os: support runtime poller with os.File on Windows

[ianlancetaylor]

In order to use the runtime poller with os.File on Windows, the file or pipe will have to be opened with FILE_FLAG_OVERLAPPED. We shouldn't do that unconditionally unless we can remove it if the program calls the Fd method. Programs currently expect the Fd method to return a handle that uses ordinary synchronous I/O.

[gopherbot]

CL https://golang.org/cl/36800 mentions this issue.

[rasky]

ReOpenFile() can do just that, adding/removing FILE_FLAG_OVERLAPPED (https://msdn.microsoft.com/en-us/library/aa365497(VS.85).aspx)

[polarina]

Overlapped reads on Windows may run synchronously under some conditions. Files using NTFS compression can not be read asynchronously, for example.

See https://support.microsoft.com/en-us/help/156932/asynchronous-disk-i-o-appears-as-synchronous-on-windows

[ianlancetaylor]

Thanks. For our purposes it doesn't matter if I/O is sometimes synchronous. The effect will be that the I/O operation will tie up a thread. That is not ideal, but it is no different than any other blocking syscall.

[jstarks]

There's no way to flip the bit on an existing file handle, and ReOpenFile will only work properly for real local files, only if the file does not have the delete on close bit set (I think), and only if they have been opened with compatible sharing flags. And of course Fd() uintptr does not provide a mechanism to return an error if something goes wrong.

I'm afraid this may be impossible to achieve with the current Windows API.

This is probably not such a shame for normal files opened for cached IO, although recent versions of Windows do perform cached reads asynchronously in some cases. But it certainly is limiting for named pipes, files opened with FILE_FLAG_NO_BUFFERING, and other Windows file handle types that reliably perform IO asynchronously.

Maybe this is something that we can improve in Windows. But even if we do, I think for Windows there would still be a lot of value in providing direct access to the runtime poller. There are several APIs that take OVERLAPPED structures that don't fit into os.File (ConnectNamedPipe, MergeVirtualDisk, DeviceIoControl, etc.), and it would be ideal to be able to use the existing poller to avoid blocking threads on these APIs.

That seems at odds with some of the discussion in #18507, unfortunately.