Re: dup2() vs dup3() inconsistency when

[Bodo Eggert <7eggert@gmx.de>]

On Fri, 10 Oct 2008, Michael Kerrisk wrote:
> On Fri, Oct 10, 2008 at 1:42 PM, Bodo Eggert <7eggert@gmx.de> wrote:
> > Michael Kerrisk <mtk.manpages@googlemail.com> wrote:
> >
> >> Well, as long as we are fixing the dup3() interface in the way that Al
> >> and Ulrich have suggested, what about another fix:
> >>
> >> give an error if newfd is already open, thus forcing the user to do an
> >> explicit close
> >>
> >> ?
> >>
> >> This silent close in dup2() is an implementation blemish.  Why not eliminate
> >> it?
> >
> > I think it might be usefull:
> > Thread B does some logging to fd 42
> > Thread A switches the logfile by creating a new file, writing a header and
> > then does dup3(fd, 42, O_WRONLY|O_APPEND|O_CLOEXEC); close(fd);
> 
> I don't know the details of the kernel locks here, so perhaps this is
> a naive question: but, as things stand is there not the potential for
> some nasty race if one thread is writing to fd 42 at the same time as
> another thread does a dup2(fd, 42)?

I strongly hope there would not be any ...

> > (Off cause this is not yet implemented, O_RDONLY would give some problems,
> >  O_CLOEXEC alone might be better done while open()ing the file, ... but you
> >  get the idea.)
> >
> >
> > BTW: I think dup3(fd, -1, flags) should use the file descriptor dup() would
> > return. Or should there be a dupf(fd, flags) syscall instead?
> 
> If one did this, maybe it would be better to have an extra flag that
> said: "use the first free file descriptor >= newfd", thus giving the
> more general functionality like fcntl(F_DUPFD).

I think you are right. I thought about something similar, too, but was 
distracted enough to not think about the connection.
-- 
Funny quotes:
10. Nothing is fool proof to a talented fool.