Re: Kernel panic at Ubuntu: IMA + Apparmor

[ebiederm@xmission.com (Eric W. Biederman)]

Dmitry Kasatkin <dmitry.kasatkin@gmail.com> writes:

> On 25 April 2014 23:01, Oleg Nesterov <oleg@redhat.com> wrote:
>> On 04/25, Eric W. Biederman wrote:
>>>
>>> Oleg Nesterov <oleg@redhat.com> writes:
>>>
>>> > Well. I _think_ that __fput() and ima_file_free() in particular should not
>>> > depend on current and/or current->nsproxy. If nothing else, fput() can be
>>> > called by the unrelated task which looks into /proc/pid/.
>>> >
>>> > But again, task_work_add() has more and more users, and it seems that even
>>> > __fput() paths can do "everything", so perhaps it would be safer to allow
>>> > to use ->nsproxy in task_work_run.
>>>
>>> Like I said, give me a clear motivating case.
>>
>> I agree, we need a reason. Currently I do not see one.
>>
>>> Right now not allowing
>>> nsproxy is turning up bugs in __fput.  Which seems like a good thing.
>>
>> This is what I certainly agree with ;)
>>
>
> Hi,
>
> IMA uses kernel_read API which does not know anything about caller.
> And of course security frameworks are at guard as usual.
>
> Exactly after reading first Eric's respons, I thought why to scratch
> the head when task work queues are indeed designed for tasks...

__fput has no guarantee of running in the task that close the file
descriptor.  If your code depends on that your code is broken.

> And if you to dig for the history, IMA-appraisal was stuck due to
> lockdep reporting even though it was on non-everlaping cases.
> IIRC files vs. directories...
>
> After that IIRC Al Viro discussed about delayed fput and IIRC Oleg
> (sorry if I am wrong) introduced task work queues.
>
> So IMA-appraisal was able to be upstreamed... That was ~3.4 time frame, IIRC
>
> Name space also dated around ~3.4??
> Apparmor namespace change was also around that time.
>
> 3.10 introduces this name space order change and broke IMA-appraisal.

IMA-appraisal is fundamentally broken because I can take a mandatory
file lock and prevent IMA-apprasial.

Using kernel_read is what allows this.

> Isn't it a clear motivating case???

kernel_read is not appropriate for IMA use.  The rest of this is just
the messenger.

IMA needs to use a cousin of kernel_read that operates at a lower level
than vfs_read.  A function that all of the permission checks and the
fsnotify work.

I am sorry to be the bearer of bad news.  But kernel_read is totally
inappropriate for IMA.

Eric