On 03/30/2017 09:45 PM, Linus Torvalds wrote:

On Thu, Mar 30, 2017 at 11:26 AM, Christoph Hellwig [off-list ref] wrote:

quoted

That would be nice, but still won't work as we blindly copy f_flags
into F_GETFL, not even masking our internal FMODE_ bits.

Ok, *that* is just silly of us, and we could try to just fix, and even backport.

There's no possible valid use I could see where that should break
(famous last words - user code does some damn odd things at times).

Of course, that won't fix old kernels that are out there, but then
neither would your original patch...

Side note: I think you *can* detect the O_ATOMIC support by using
F_SETFL, because F_SETFL only allows you to change flags that we
recognize. So somebody who really wants to *guarantee* that O_ATOMIC
is there and honored even with old kernels could presumable do
something like

   fd = open(..); // *no* O_ATOMIC
   fcnt(fd, F_SETFL, O_ATOMIC);
   if (fcnt(fd, F_GETFL, NULL) & O_ATOMIC)
        // Yay! We actually got it
   else
        // I guess we need to fall back on old behavior

although I agree that that is ridiculously inconvenient and not a
great thing, and it's worth trying to aim for some better model.

Perhaps in that case it is time for an F_GETFL2 an F_GET_REAL_FL
that gives you the nice simple user code Linus wanted for new applications.
and solves forward and backwords for applications and Kernels?

Just my $0.017
Boaz

                    Linus