win32: inject a socket-aware version of CloseHandle() into the CRT

_close() on an fd calls CloseHandle(), which is illegal if the fd
contains a socket handle.

We previously worked around this by having our own close(), which called
closesocket() before calling _close()
(e601c439adce167078ac7b49550c0418ace86f94).

Amusingly, the author of that solution thought it's just a temporary
workaround:

  /*
   * close RTL fd while respecting sockets
   * added as temporary measure until PerlIO has real
   * Win32 native layer
   *   -- BKS, 11-11-2000
   */

To make it thread-safe, we had to manipulate the internals of file
descriptors, which kept changing
(b47a847f6284f6f98ad7509cf77a4aeb802d8fce).

Unfortunately, the C runtime has been rewritten and it no longer exposes
them at all. We had to disable the thread-safety fix in Visual C++ 2015
builds (1f664ef5314fb6e438137c44c95cf5ecdbdb5e9b). It also wouldn't work
with MinGW configured to use UCRT.

This commit introduces a new solution: we inject a socket-aware version
of CloseHandle() into the C runtime library. Hopefully, this will be
less fragile.

This also fixes a few issues that the original solution didn't:

  - Closing a busy pipe doesn't cause a deadlock (fixes #19963)

  - _dup2 properly closes an overwritten socket (fixes #20920)

1 files changed, 12 insertions, 2 deletions

diff --git a/pod/perldelta.pod b/pod/perldelta.pod
index 4183474c53..1953ca6e55 100644
--- a/pod/perldelta.pod
+++ b/pod/perldelta.pod
@@ -365,9 +365,19 @@ L</Modules and Pragmata> section.
 
 =over 4
 
-=item XXX-some-platform
+=item Windows
 
-XXX
+=over 4
+
+=item *
+
+C<POSIX::dup2> no longer creates broken sockets. [GH #20920]
+
+=item *
+
+Closing a busy pipe could cause Perl to hang. [GH #19963]
+
+=back
 
 =back