Re: [PATCH] lkdtm: add bad USER_DS test

[Christian Borntraeger]

On 03/24/2017 04:17 PM, Thomas Garnier wrote:
> On Fri, Mar 24, 2017 at 1:14 AM, Heiko Carstens
> <heiko.carstens@xxxxxxxxxx> wrote:
>> On Thu, Mar 23, 2017 at 01:34:19PM -0700, Kees Cook wrote:
>>> This adds CORRUPT_USER_DS to check that the get_fs() test on syscall return
>>> still sees USER_DS during the new VERIFY_PRE_USERMODE_STATE checks.
>>>
>>> Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
>>
>> ...
>>
>>> +void lkdtm_CORRUPT_USER_DS(void)
>>> +{
>>> +     /*
>>> +      * Test that USER_DS has been set correctly on exiting a syscall.
>>> +      * Since setting this higher than USER_DS (TASK_SIZE) would introduce
>>> +      * an exploitable condition, we lower it instead, since that should
>>> +      * not create as large a problem on an unprotected system.
>>> +      */
>>> +     mm_segment_t lowfs;
>>> +#ifdef MAKE_MM_SEG
>>> +     lowfs = MAKE_MM_SEG(TASK_SIZE - PAGE_SIZE);
>>> +#else
>>> +     lowfs = TASK_SIZE - PAGE_SIZE;
>>> +#endif
>>> +
>>> +     pr_info("setting bad task size limit\n");
>>> +     set_fs(lowfs);
>>> +}
>>
>> This won't work on architectures where the set_fs() argument does not
>> contain an address but an address space identifier. This is true e.g. for
>> s390 and as far as I know also for sparc.
>> On s390 we have complete distinct address spaces for kernel and user space
>> that each start at address zero.
>>
> 
> The patch that enforce USER_DS is disabled on s390 anyway. I guess, we
> can just do a set_fs(KERNEL_DS) for the others.

that would enable the test, but it would also mean that lkdtm can be used by
a program to escalate its rights. I think that is the reason why Kees did this
lowfs things.