O_TMPFILE detection

O_TMPFILE detection

[Christoph Hellwig]

I've started implementing support for O_TMPFILE in xfs_io in preparation
for proper testcases and found something rather annoying:

Given that we implicitly pass O_DIRECTORY as part of the full O_TMPFILE
flag an open using O_TMPFILE will simply succeed on a kernel old enough
to not have O_TMPFILE support, and give you a file descriptor for the
directory passed in.  Of course this only works if you open the fd
read-only which isn't all that useful for real temporary files, but it
would be the natural choice for a feature test.

I guess we can't really fix this up any more, but does anyone have a
good idea how to document this properly in the man page?

Re: O_TMPFILE detection

[Linus Torvalds]

On Tue, Feb 11, 2014 at 8:03 AM, Christoph Hellwig <hch@infradead.org> wrote:
> I've started implementing support for O_TMPFILE in xfs_io in preparation
> for proper testcases and found something rather annoying:
>
> Given that we implicitly pass O_DIRECTORY as part of the full O_TMPFILE
> flag an open using O_TMPFILE will simply succeed on a kernel old enough
> to not have O_TMPFILE support, and give you a file descriptor for the
> directory passed in.  Of course this only works if you open the fd
> read-only which isn't all that useful for real temporary files, but it
> would be the natural choice for a feature test.

But O_TMPFILE without write access fails on modern kernels, so what
you test doesn't make sense.

In fact, that "O_TMPFILE fails with O_RDONLY" was done exactly so that
you could not have a program that works on kernels with O_TMPFILE
support _and_ opens a directory on old kernels without it. Basically
"O_TMPFILE | O_RDONLY" is an insane thing to do, and will never create
a temporary file.

              Linus