Re: [RFC][PATCH 00/10] Mount, FS, Block and Keyrings notifications [ver #3]

From: Stephen Smalley
Date: Thu Jun 06 2019 - 10:09:46 EST

On 6/6/19 9:16 AM, David Howells wrote:
Stephen Smalley <sds@xxxxxxxxxxxxx> wrote:

This might be easier to discuss if you can reply to:

https://lore.kernel.org/lkml/5393.1559768763@xxxxxxxxxxxxxxxxxxxxxx/

which is on the ver #2 posting of this patchset.

Sorry for being late to the party. Not sure whether you're asking me to respond only there or both there and here to your comments below. I'll start here but can revisit if it's a problem.

LSM support is included, but controversial:

(1) The creds of the process that did the fput() that reduced the refcount
to zero are cached in the file struct.

(2) __fput() overrides the current creds with the creds from (1) whilst
doing the cleanup, thereby making sure that the creds seen by the
destruction notification generated by mntput() appears to come from
the last fputter.

(3) security_post_notification() is called for each queue that we might
want to post a notification into, thereby allowing the LSM to prevent
covert communications.

(?) Do I need to add security_set_watch(), say, to rule on whether a watch
may be set in the first place? I might need to add a variant per
watch-type.

(?) Do I really need to keep track of the process creds in which an
implicit object destruction happened? For example, imagine you create
an fd with fsopen()/fsmount(). It is marked to dissolve the mount it
refers to on close unless move_mount() clears that flag. Now, imagine
someone looking at that fd through procfs at the same time as you exit
due to an error. The LSM sees the destruction notification come from
the looker if they happen to do their fput() after yours.


I'm not in favor of this approach.

Which bit? The last point? Keeping track of the process creds after an
implicit object destruction.

(1), (2), (3), and the last point.

Can we check permission to the object being watched when a watch is set
(read-like access),

Yes, and I need to do that. I think it's likely to require an extra hook for
each entry point added because the objects are different:

int security_watch_key(struct watch *watch, struct key *key);
int security_watch_sb(struct watch *watch, struct path *path);
int security_watch_mount(struct watch *watch, struct path *path);
int security_watch_devices(struct watch *watch);

make sure every access that can trigger a notification requires a
(write-like) permission to the accessed object,

"write-like permssion" for whom? The triggerer or the watcher?

The former, i.e. the process that performed the operation that triggered the notification. Think of it as a write from the process to the accessed object, which triggers a notification (another write) on some related object (the watched object), which is then read by the watcher.

There are various 'classes' of events:

(1) System events (eg. hardware I/O errors, automount points expiring).

(2) Direct events (eg. automounts, manual mounts, EDQUOT, key linkage).

(3) Indirect events (eg. exit/close doing the last fput and causing an
unmount).

Class (1) are uncaused by a process, so I use init_cred for them. One could
argue that the automount point expiry should perhaps take place under the
creds of whoever triggered it in the first place, but we need to be careful
about long-term cred pinning.

This seems equivalent to just checking whether the watcher is allowed to get that kind of event, no other cred truly needed.

Class (2) the causing process must've had permission to cause them - otherwise
we wouldn't have got the event.

So we've already done a check on the causing process, and we're going to check whether the watcher can set the watch. We just need to establish the connection between the accessed object and the watched object in some manner.

Class (3) is interesting since it's currently entirely cleanup events and the
process may have the right to do them (close, dup2, exit, but also execve)
whether the LSM thinks it should be able to cause the object to be destroyed
or not.

It gets more complicated than that, though: multiple processes with different
security attributes can all have fds pointing to a common file object - and
the last one to close carries the can as far as the LSM is concerned.

Yes, I'd prefer to avoid that. You can't write policy that is stable and meaningful that way. This may fall under a similar situation as class (1) - all we can meaningfully do is check whether the watcher is allowed to see all such events.

And yet more complicated when you throw in unix sockets with partially passed
fds still in their queues. That's what patch 01 is designed to try and cope
with.

and make sure there is some sane way to control the relationship between the
accessed object and the watched object (write-like)?

This is the trick. Keys and superblocks have object labels of their own and
don't - for now - propagate their watches. With these, the watch is on the
object you initially assign it to and it goes no further than that.

mount_notify() is the interesting case since we want to be able to detect
mount topology change events from within the vfs subtree rooted at the watched
directory without having to manually put a watch on every directory in that
subtree - or even just every mount object. >
Or, maybe, that's what I'll have to do: make it mount_notify() can only apply
to the subtree within its superblock, and the caller must call mount_notify()
for every mount object it wants to monitor. That would at least ensure that
the caller can, at that point, reach all those mount points.

Would that at least make it consistent with fanotify (not that it provides a great example)?

For cases where we have no object per se or at least no security
structure/label associated with it, we may have to fall back to a
coarse-grained "Can the watcher get this kind of notification in general?".

Agreed - and we should probably have that anyway.

David