Re: [RFC PATCH 1/5] ima: extend clone() with IMA namespace support

[Stefan Berger <stefanb@linux.vnet.ibm.com>]

On 07/25/2017 03:48 PM, Mimi Zohar wrote:
> On Tue, 2017-07-25 at 12:08 -0700, James Bottomley wrote:
>> On Tue, 2017-07-25 at 14:04 -0500, Serge E. Hallyn wrote:
>>> On Tue, Jul 25, 2017 at 11:49:14AM -0700, James Bottomley wrote:
>>>> On Tue, 2017-07-25 at 12:53 -0500, Serge E. Hallyn wrote:
>>>>> On Thu, Jul 20, 2017 at 06:50:29PM -0400, Mehmet Kayaalp wrote:
>>>>>>
>>>>>> From: Yuqiong Sun <suny@us.ibm.com>
>>>>>>
>>>>>> Add new CONFIG_IMA_NS config option.  Let clone() create a new
>>>>>> IMA namespace upon CLONE_NEWNS flag. Add ima_ns data structure
>>>>>> in nsproxy. ima_ns is allocated and freed upon IMA namespace
>>>>>> creation and exit. Currently, the ima_ns contains no useful IMA
>>>>>> data but only a dummy interface. This patch creates the
>>>>>> framework for namespacing the different aspects of IMA (eg.
>>>>>> IMA-audit, IMA-measurement, IMA-appraisal).
>>>>>>
>>>>>> Signed-off-by: Yuqiong Sun <suny@us.ibm.com>
>>>>>>
>>>>>> Changelog:
>>>>>> * Use CLONE_NEWNS instead of a new CLONE_NEWIMA flag
>>>>> Hi,
>>>>>
>>>>> So this means that every mount namespace clone will clone a new
>>>>> IMA namespace.  Is that really ok?
>>>> Based on what: space concerns (struct ima_ns is reasonably small)?
>>>> or whether tying it to the mount namespace is the correct thing to
>>>> do.  On
>>> Mostly the latter.  The other would be not so much space concerns as
>>> time concerns.  Many things use new mounts namespaces, and we
>>> wouldn't want multiple IMA calls on all file accesses by all of
>>> those.
>>>
>>>> the latter, it does seem that this should be a property of either
>>>> the mount or user ns rather than its own separate ns.  I could see
>>>> a use where even a container might want multiple ima keyrings
>>>> within the container (say containerised apache service with
>>>> multiple tenants), so instinct tells me that mount ns is the
>>>> correct granularity for this.
>>> I wonder whether we could use echo 1 > /sys/kernel/security/ima/newns
>>> as the trigger for requesting a new ima ns on the next
>>> clone(CLONE_NEWNS).
>> I could go with that, but what about the trigger being installing or
>> updating the keyring?  That's the only operation that needs namespace
>> separation, so on mount ns clone, you get a pointer to the old ima_ns
>> until you do something that requires a new key, which then triggers the
>> copy of the namespace and installing it?
> It isn't just the keyrings that need to be namespaced, but the
> measurement list and policy as well.
>
> IMA-measurement, IMA-appraisal and IMA-audit are all policy based.
>
> As soon as the namespace starts, measurements should be added to the
> namespace specific measurement list, not it's parent.

IMA is about measuring things, logging what was executed, and finally 
someone looking at the measurement log and detecting 'things'. So at 
least one attack that needs to be prevented is a malicious person 
opening an IMA namespace, executing something malicious, and not leaving 
any trace on the host because all the logs went into the measurement 
list of the IMA namespace, which disappeared. That said, I am wondering 
whether there has to be a minimum set of  namespaces (PID, UTS) 
providing enough 'isolation' that someone  may actually open an IMA 
namespace and run their code. To avoid leaving no traces one could argue 
to implement recursive logging, so something that is logged inside the 
namespace will be detected in all parent containers up to the 
init_ima_ns (host) because it's logged (and TPM extended) there as well. 
The challenge with that is that logging costs memory and that can be 
abused as well until the machine needs a reboot... I guess the solution 
could be requesting an IMA namespace in one way or another but requiring 
several other namespace flags in the clone() to actually 'get' it. 
Jumping namespaces with setns() may have to be restricted as well once 
there is an IMA namespace.

    Stefan

>
> Mimi
>