base: Make it less likely for fdReady() to fail on Windows sockets.

See the added comment for details.

It's "less likely" because it can still fail if the socket happens to
have an FD larger than 1023, which can happen if many files are opened.

Until now, basic socket programs that use `hWaitForInput` were broken on
Windows.

That is because on Windows `FD_SETSIZE` defaults to 64, but pretty much
all GHC programs seem to have > 64 FDs open, so you can't actually
create a socket on which you can `select()`.

It errors with `fdReady: fd is too big` even with an example as simple
as the following (in this case, on my machine the `fd` is `284`):

  {-# LANGUAGE OverloadedStrings #-}

  import Control.Monad (forever)
  import Network.Socket
  import System.IO

  -- Simple echo server: Reads up to 10 chars from network, echoes them back.
  -- Uses the Handle API so that `hWaitForInput` can be used.
  main :: IO ()
  main = do
    sock <- socket AF_INET Stream 0
    setSocketOption sock ReuseAddr 1
    bind sock (SockAddrInet 1234 0x0100007f)
      -- 0x0100007f == 127.0.0.1 localhost
    listen sock 2
    forever $ do
      (connSock, _connAddr) <- accept sock
      putStrLn "Got connection"

      h <- socketToHandle connSock ReadWriteMode
      hSetBuffering h NoBuffering

      ready <- hWaitForInput h (5 * 1000) -- 5 seconds
      putStrLn $ "Ready: " ++ show ready

      line <- hGetLine h
      putStrLn "Got line"
      hPutStrLn h ("Got: " ++ line)
      hClose h

I'm not sure how this was not discovered earlier; for #13525 (where
`fdReady()` breaking completely was also discovered late) at least it
failed only when the timeout was non-zero, which is not used in ghc
beyond in `hWaitForInput`, but in this Windows socket case it breaks
even on the 0-timeout.

Maybe there is not actually anybody who uses sockets as handles on
Windows?

The workaround for now is to increase `FD_SETSIZE` on Windows;
increasing it is possible on Windows and BSD, see

https://stackoverflow.com/questions/7976388/increasing-limit-of-fd-setsi
ze-and-select

A real fix would be to move to IO Completion Ports on Windows, and thus
get rid of the last uses of `select()` (the other platforms already use
`poll()` but Windows doesn't have that).

Reviewers: bgamari, austin, hvr, erikd, simonmar

Reviewed By: bgamari

Subscribers: rwbarton, thomie

Differential Revision: https://phabricator.haskell.org/D3959

1 files changed, 16 insertions, 0 deletions

diff --git a/libraries/base/cbits/inputReady.c b/libraries/base/cbits/inputReady.c
index ddeee664d6..916633c9c0 100644
--- a/libraries/base/cbits/inputReady.c
+++ b/libraries/base/cbits/inputReady.c
@@ -4,6 +4,22 @@
  * hWaitForInput Runtime Support
  */
 
+/* FD_SETSIZE defaults to 64 on Windows, which makes even the most basic
+ * programs break that use select() on a socket FD.
+ * Thus we raise it here (before any #include of network-related headers)
+ * to 1024 so that at least those programs would work that would work on
+ * Linux if that used select() (luckily it uses poll() by now).
+ * See https://ghc.haskell.org/trac/ghc/ticket/13497#comment:23
+ * The real solution would be to remove all uses of select()
+ * on Windows, too, and use IO Completion Ports instead.
+ * Note that on Windows, one can simply define FD_SETSIZE to the desired
+ * size before including Winsock2.h, as described here:
+ *   https://msdn.microsoft.com/en-us/library/windows/desktop/ms740141(v=vs.85).aspx
+ */
+#if defined(_WIN32)
+#define FD_SETSIZE 1024
+#endif
+
 /* select and supporting types is not Posix */
 /* #include "PosixSource.h" */
 #include "HsBase.h"