On Wed, Mar 17, 2021 at 1:42 PM Jan Kara [off-list ref] wrote:

On Wed 17-03-21 13:01:35, Amir Goldstein wrote:

quoted

On Tue, Mar 16, 2021 at 5:55 PM Jan Kara [off-list ref] wrote:

quoted

On Thu 04-03-21 13:29:19, Amir Goldstein wrote:

quoted

Jan,

These patches try to implement a minimal set and least controversial
functionality that we can allow for unprivileged users as a starting
point.

The patches were tested on top of v5.12-rc1 and the fanotify_merge
patches using the unprivileged listener LTP tests written by Matthew
and another LTP tests I wrote to test the sysfs tunable limits [1].

Thanks. I've added both patches to my tree.

Great!
I'll go post the LTP tests and work on the man page updates.

BTW, I noticed that you pushed the aggregating for_next branch,
but not the fsnotify topic branch.

Is this intentional?

Not really, pushed now. Thanks for reminder.

quoted

I am asking because I am usually basing my development branches
off of your fsnotify branch, but I can base them on the unpushed branch.

Heads up. I am playing with extra privileges we may be able to
allow an ns_capable user.
For example, watching a FS_USERNS_MOUNT filesystem that the user
itself has mounted inside userns.

Another feature I am investigating is how to utilize the new idmapped
mounts to get a subtree watch functionality. This requires attaching a
userns to the group on fanotify_init().

<hand waving>
If the group's userns are the same or below the idmapped mount userns,
then all the objects accessed via that idmapped mount are accessible
to the group's userns admin. We can use that fact to filter events very
early based on their mnt_userns and the group's userns, which should be
cheaper than any subtree permission checks.
<\hand waving>

Yeah, I agree this should work. Just it seems to me the userbase for this
functionality will be (at least currently) rather limited. While full

That may change when systemd home dirs feature starts to use
idmapped mounts.
Being able to watch the user's entire home directory is a big win
already.

subtree watches would be IMO interesting to much more users.

Agreed.

I was looking into that as well, using the example of nfsd_acceptable()
to implement the subtree permission check.

The problem here is that even if unprivileged users cannot compromise
security, they can still cause significant CPU overhead either queueing
events or filtering events and that is something I haven't been able to
figure out a way to escape from.

BUT, if you allow userns admin to setup subtree watches (a.k.a filtered
filesystem marks) on a userns filesystem/idmapped mount, now users
can watch subtrees in their home directories and other processes will
pay CPU penalty only for file access in the users home directories.

That might be acceptable.

Thanks,
Amir.