On Thu, May 16, 2019 at 01:26:29PM +0300, Amir Goldstein wrote:
> There is a common pattern among pseudo filesystems for removing a dentry
> from code paths that are NOT coming from vfs_{unlink,rmdir}, using a
> combination of simple_{unlink,rmdir} and d_delete().
>
> Create an helper to perform this common operation. This helper is going
> to be used as a place holder for the new fsnotify_{unlink,rmdir} hooks.
This is the wrong approach. What we have is a bunch of places trying
to implement recursive removal of a subtree. They are broken, each in
its own way, and I'm not talking about fsnotify crap - there are
much more unpleasant issues there.
Trying to untangle that mess is not going to be made easier by mandating
the calls of that helper of yours - if anything, it makes the whole
thing harder to massage.
It needs to be dealt with, no arguments here, but that's not a good
starting point for that... I've taken several stabs at that, never
got anywhere satisfactory with those ;-/ I'll try to dig out the
notes/existing attempts at patch series; if you are willing to participate
in discussing those and sorting the whole thing out, you are very welcome;
just ping me in a couple of days.
