Re: fsnotify path hooks

[Amir Goldstein <amir73il@gmail.com>]

On Fri, Apr 9, 2021 at 1:08 PM Jan Kara <jack@suse.cz> wrote:
>
> On Thu 08-04-21 18:11:31, Amir Goldstein wrote:
> > > > FYI, I tried your suggested approach above for fsnotify_xattr(),
> > > > but I think I prefer to use an explicit flavor fsnotify_xattr_mnt()
> > > > and a wrapper fsnotify_xattr().
> > > > Pushed WIP to fsnotify_path_hooks branch. It also contains
> > > > some unstashed "fix" patches to tidy up the previous hooks.
> > >
> > > What's in fsnotify_path_hooks branch looks good to me wrt xattr hooks.
> > > I somewhat dislike about e.g. the fsnotify_create() approach you took is
> > > that there are separate hooks fsnotify_create() and fsnotify_create_path()
> > > which expose what is IMO an internal fsnotify detail of what are different
> > > event types. I'd say it is more natural (from VFS POV) to have just a
> > > single hook and fill in as much information as available... Also from
> >
> > So to be clear, you do NOT want additional wrappers like this and
> > you prefer to have the NULL mnt argument explicit in all callers?
> >
> > static inline void fsnotify_xattr(struct dentry *dentry)
> > {
> >         fsnotify_xattr_mnt(NULL, dentry);
> > }
> >
> > For fsnotify_xattr() it does not matter so much, but fsnotify_create/mkdir()
> > have quite a few callers in special filesystems.
>
> Yes, I prefer explicit NULL mnt argument to make it obvious we are going to
> miss something in this case. I agree it's going to be somewhat bigger churn
> but it isn't that bad (10 + 6 callers).
>
> > > outside view, it is unclear that e.g. vfs_create() will generate some types
> > > of fsnotify events but not all while e.g. do_mknodat() will generate all
> > > fsnotify events. That's why I'm not sure whether a helper like vfs_create()
> > > in your tree is the right abstraction since generating one type of fsnotify
> > > event while not generating another type should be a very conscious decision
> > > of the implementor - basically if you have no other option.
> >
> > I lost you here.
>
> Sorry, I was probably too philosophical here ;)
>
> > Are you ok with vfs_create() vs. vfs_create_nonotify()?
>
> I'm OK with vfs_create_nonotify(). I have a problem with vfs_create()
> because it generates inode + fs events but does not generate mount events
> which is just strange (although I appreciate the technical reason behind
> it :).
>
> > How do you propose to change fsnotify hooks in vfs_create()?
>
> So either pass 'mnt' to vfs_create() - as we discussed, this may be
> actually acceptable these days due to idmapped mounts work - and generate
> all events there, or make vfs_create() not generate any fsnotify events and
> create new vfs_create_notify() which will take the 'mnt' and generate
> events. Either is fine with me and more consistent than what you currently
> propose. Thoughts?
>

Jan,

I started to go down the vfs_create_notify() path and I guess it's looking
not too bad (?). Pushed WIP to branch fsnotify_path_hooks-wip.

I hit another bump though.  By getting fsnotify_{unlink,rmdir}() outside of the
vfs helpers, we break the rule:

        /* Expected to be called before d_delete() */
        WARN_ON_ONCE(d_is_negative(dentry));

I'm not sure how to solve this without passing mnt into the vfs helpers.

One solution is not adding support for delete events to mount mark.
I was trying to aim for maximum flexibility, but for the use case that
Christian mentioned (injecting bind mounts into container) it is only
really necessary to support FAN_CREATE and FAN_MOVED_TO
(or FAN_MOVE_SELF) on a mount mark for observing when a path
becomes positive.

For observing when a path becomes negative, it is sufficient to watch
all the ancestor directories.

Thoughts?

Thanks,
Amir.