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

[Casey Schaufler]

On 6/6/2019 10:16 AM, Stephen Smalley wrote:
> On 6/6/19 12:43 PM, Casey Schaufler wrote:
>> ...
>> I don't agree. That is, I don't believe it is sufficient.
>> There is no guarantee that being able to set a watch on an
>> object implies that every process that can trigger the event
>> can send it to you.
>>
>>     Watcher has Smack label W
>>     Triggerer has Smack label T
>>     Watched object has Smack label O
>>
>>     Relevant Smack rules are
>>
>>     W O rw
>>     T O rw
>>
>> The watcher will be able to set the watch,
>> the triggerer will be able to trigger the event,
>> but there is nothing that would allow the watcher
>> to receive the event. This is not a case of watcher
>> reading the watched object, as the event is delivered
>> without any action by watcher.
>
> You are allowing arbitrary information flow between T and W above.  Who cares about notifications?

I do. If Watched object is /dev/null no data flow is possible.
There are many objects on a modern Linux system for which this
is true. Even if it's "just a file" the existence of one path
for data to flow does not justify ignoring the rules for other
data paths.

>
> How is it different from W and T mapping the same file as a shared mapping and communicating by reading and writing the shared memory?  You aren't performing a permission check directly between W and T there.

In this case there is one object O, two subjects, W and T and two accesses.

	W open O
	T open O

They fiddle about with the data in O.

In the event case, there are two objects, O and W, two subjects W and T, and
three accesses.

	W watch O
	T trigger O
	T write-event W

You can't wave away the flow of data. Different objects are involved.

An analogy is that two processes with different UIDs can open a file,
but still can't signal each other. Different mechanisms have different
policies. I'm not saying that's good, but it's the context we're in.