Re: [RFC][PATCH 0/8] Mount, FS, Block and Keyrings notifications [ver #2]

From: Stephen Smalley
Date: Wed Jun 05 2019 - 17:05:56 EST

On 6/5/19 3:28 PM, Greg KH wrote:
On Wed, Jun 05, 2019 at 02:25:33PM -0400, Stephen Smalley wrote:
On 6/5/19 1:47 PM, Andy Lutomirski wrote:

On Jun 5, 2019, at 10:01 AM, Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote:

On 6/5/2019 9:04 AM, Andy Lutomirski wrote:
On Wed, Jun 5, 2019 at 7:51 AM Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote:
On 6/5/2019 1:41 AM, David Howells wrote:
Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote:

I will try to explain the problem once again. If process A
sends a signal (writes information) to process B the kernel
checks that either process A has the same UID as process B
or that process A has privilege to override that policy.
Process B is passive in this access control decision, while
process A is active. In the event delivery case, process A
does something (e.g. modifies a keyring) that generates an
event, which is then sent to process B's event buffer.
I think this might be the core sticking point here. It looks like two
different situations:

(1) A explicitly sends event to B (eg. signalling, sendmsg, etc.)

(2) A implicitly and unknowingly sends event to B as a side effect of some
other action (eg. B has a watch for the event A did).

The LSM treats them as the same: that is B must have MAC authorisation to send
a message to A.

Threat is about what you can do, not what you intend to do.

And it would be really great if you put some thought into what
a rational model would be for UID based controls, too.

But there are problems with not sending the event:

(1) B's internal state is then corrupt (or, at least, unknowingly invalid).
Then B is a badly written program.
Either I'm misunderstanding you or I strongly disagree.

A program needs to be aware of the conditions under
which it gets event, *including the possibility that
it may not get an event that it's not allowed*. Do you
regularly write programs that go into corrupt states
if an open() fails? Or where read() returns less than
the amount of data you ask for?

I do not regularly write programs that handle read() omitting data in the middle of a TCP stream. I also donât write programs that wait for processes to die and need to handle the case where a child is dead, waitid() can see it, but SIGCHLD wasnât sent because âsecurityâ.

If B has
authority to detect a certain action, and A has authority to perform
that action, then refusing to notify B because B is somehow missing
some special authorization to be notified by A is nuts.

You are hand-waving the notion of authority. You are assuming
that if A can read X and B can read X that A can write B.

No, read it again please. Iâm assuming that if A can *write* X and B can read X then A can send information to B.

I guess the questions here are:

1) How do we handle recursive notification support, since we can't check
that B can read everything below a given directory easily? Perhaps we can
argue that if I have watch permission to / then that implies visibility to
everything below it but that is rather broad.

How do you handle fanotify today which I think can do this?

Doesn't appear to have been given much thought; looks like fanotify_init() checks capable(CAP_SYS_ADMIN) and fanotify_mark() checks inode_permission(MAY_READ) on the mount/directory/file. File descriptors for monitored files returned upon events at least get vetted through security_file_open() so that can prevent the monitoring process from receiving arbitrary descriptors. Would be preferable if fanotify_mark() did some kind of security_path_watch() or similar check, and distinguished mounts versus directories since monitoring of directories is not recursive.