pipe, pipe2: don't corrupt fd on error

pipe, pipe2: don't corrupt fd on error

I noticed a potential subtle double-close bug in libvirt.  There,
a common idiom is to initialize an int fd[2]={-1,-1}, then have
multiple error paths goto common cleanup code.  In the cleanup
code, the fds are closed if they are not already -1; this works
if the error label is reached before the pipe call, or after
pipe succeeds, but if it was the pipe call itself that jumped
to the error label, then it is relying on failed pipe() not
altering the values already in fd array prior to the failure.
Our pipe2 replacement violated this assumption, and could leave
a non-negative value in the array, which in turn would let
libvirt close an already-closed fd, possibly nuking an unrelated
fd opened by another thread that happened to get the same value.

As a result, I raised a POSIX issue regarding the behavior of
pipe on failure: http://austingroupbugs.net/view.php?id=467

Using that test program, I learned that most systems leave fd
unchanged on error, but that mingw always assigns -1 into the
array.  This fixes the mingw pipe() replacement, as well as
the gnulib pipe2() replacement.

I don't know of any race-free way to work around a cygwin crash:
http://cygwin.com/ml/cygwin/2011-06/msg00328.html - we could
always open() and then close() two fds to guess whether two
spare fd still remain before calling pipe(), but that is racy.

* lib/pipe.c (pipe): Leave fd unchanged on error.
* lib/pipe2.c (pipe2): Likewise.
* doc/posix-functions/pipe.texi (pipe): Document cygwin issue.
* doc/glibc-functions/pipe2.texi (pipe2): Likewise.

Signed-off-by: Eric Blake <eblake@redhat.com>