Jamie Lokier wrote:
> Blue Swirl wrote:
> > On Mon, Nov 16, 2009 at 2:47 PM, Kevin Wolf <kw...@redhat.com> wrote:
> > > Am 13.11.2009 22:05, schrieb Blue Swirl:
> > >> On Fri, Nov 13, 2009 at 5:17 PM, Kevin Wolf <kw...@redhat.com> wrote:
> > >>> We're leaking file descriptors to child processes. Set FD_CLOEXEC on
> > >>> file
> > >>> descriptors that don't need to be passed to children to stop this
> > >>> misbehaviour.
> > >>
> > >>> - c = accept(s, (struct sockaddr *)&addr, &addrlen);
> > >>> + c = qemu_accept(s, (struct sockaddr *)&addr, &addrlen);
> > >>
> > >> Would it be possible to improve the interface so that no casts are
> > >> needed for the calling code?
> > >
> > > How exactly would you do that? The only way I see to do it would be
> > > using void*, but I'm not sure if this really is an improvement.
> >
> > Instead of sockaddr_in vs. sockaddr and the lame casts in between, we
> > could have QSockAddr which magically works. Or if we only ever use
> > sockaddr_in, just use that.
>
> int qemu_accept(int s, union __attribute__((__transparent_union__)) {
> struct sockaddr *sa;
> struct sockaddr_in *sin;
> struct sockaddr_in6 *sin6;
> } addr, socklen_t len);
>
> #define qemu_accept(s, addr) qemu_accept(s, addr, sizeof(*addr))
>
> Seems to work. :-)
The transparent_union is what Glibc uses for accept(), by the way,
when _GNU_SOURCE is defined. That's why the old code using
accept() compiled without warnings despite the type mismatch.
Grep for __SOCKADDR_ARG in Glibc's /usr/include/sys/socket.h.
-- Jamie
