Re: [PATCH] Revert "libfs: Use d_children list to iterate simple_offset directories"

[Christoph Hellwig <hch@infradead.org>]

On Tue, Mar 11, 2025 at 05:47:18PM -0400, Kent Overstreet wrote:
> > guarantee that d_off will always increase in value as an application
> > walks the directory. That's an impossible thing to guarantee given
> > the way d_off values are chosen (at entry creation time, not at
> > directory iteration time).
> 
> Not sure why you're trying to cite POSIX when it's an actual application
> regression under discussion.

Because from what Darrick report they look very related.

> Sane d_offset behaviour is one of those "of _course_ things will break
> you screw that up and I don't want to be on the hook for debugging it"
> things to filesystem developers - you don't do it.

Yet the semantics are surprisingly underdocument, where the issues show
up is rather surprising and non-obvious and people do tend to get it
wrong surprisingly often (probably this thread, a whole bunch of btrfs
things showing up, the old v1 xfs format, etc).

I've actually started collection various issues in the past, the test
cases that triggered it and how it relates to writtent standards.  It's
a bit of a mess.  I hope to eventually have a coherent writeup on all
that.