Re: [RFC PATCH] fs: introduce mkdirat2 syscall for atomic mkdir

["J. Bruce Fields" <bfields@fieldses.org>]

On Mon, Mar 01, 2021 at 07:35:37PM +0000, Matthew Wilcox wrote:
> On Mon, Mar 01, 2021 at 02:09:03PM -0500, J. Bruce Fields wrote:
> > On Sun, Feb 28, 2021 at 08:57:20AM -0500, Drew DeVault wrote:
> > > On Sat Feb 27, 2021 at 11:03 PM EST, Matthew Wilcox wrote:
> > > > > 1. Program A creates a directory
> > > > > 2. Program A is pre-empted
> > > > > 3. Program B deletes the directory
> > > > > 4. Program A creates a file in that directory
> > > > > 5. RIP
> > > >
> > > > umm ... program B deletes the directory. program A opens it in order to
> > > > use openat(). program A gets ENOENT and exits, confused. that's the
> > > > race you're removing here -- and it seems fairly insignificant to me.
> > > 
> > > Yes, that is the race being eliminated here. Instead of this, program A
> > > has an fd which holds a reference to the directory, so it just works. A
> > > race is a race. It's an oversight in the API.
> > 
> > Step 4 still fails either way, because you can't create a file in an
> > unlinked directory, even if you hold a reference to that directory.
> > What's the behavior change at step 4 that you're hoping for?
> 
> If step 3 is 'mv foo bar', then the behaviour change will be that the
> files still get created, just as bar/quux, instead of foo/quux.  It's not
> clear to me this is necessarily an improvement in behaviour.

Oh, OK.

Yeah, it'd be useful to have some more detail on how this would be used.

--b.

> (as an aside, i think there's a missing feature in posix -- being able
> to atomically replace one directory with another.  you can atomically
> replace one file with another with hard links, but since you can't
> hardlink a directory, you can't do the same trick.  Maybe you should
> just always move files out of a directory instead of moving directories
> as a single operation)