[PATCH v5 3/4] LSM: Define SELinux function to measure state and policy

[Stephen Smalley]

On 8/3/20 6:08 PM, Lakshmi Ramasubramanian wrote:

> On 8/3/20 2:07 PM, Stephen Smalley wrote:
>
>>>>> [   68.870715] irq event stamp: 23486085
>>>>> [   68.870715] hardirqs last  enabled at (23486085):
>>>>> [<ffffffffaa419406>] _raw_spin_unlock_irqrestore+0x46/0x60
>>>>> [   68.870715] hardirqs last disabled at (23486084):
>>>>> [<ffffffffaa419443>] _raw_spin_lock_irqsave+0x23/0x90
>>>>> [   68.870715] softirqs last  enabled at (23486074):
>>>>> [<ffffffffaa8004f3>] __do_softirq+0x4f3/0x662
>>>>> [   68.870715] softirqs last disabled at (23486067):
>>>>> [<ffffffffaa601072>] asm_call_on_stack+0x12/0x20
>>>>> [   68.870715] ---[ end trace fb02740ff6f4d0cd ]---
>>>>
>>>> I think one issue here is that systemd loads SELinux policy first, 
>>>> then IMA policy, so it doesn't know whether it needs to measure 
>>>> SELinux policy on first policy load, and another issue is that the 
>>>> policy is too large to just queue the policy data itself this way 
>>>> (or you need to use an allocator that can handle larger sizes).
>>>>
>>>
>>> The problem seems to be that a lock is held when the IMA hook to 
>>> measure the LSM state is called. So memory allocation is not 
>>> allowed, but the hook is doing an allocation. I'll address this - 
>>> thanks for catching it.
>>>
>>> I have the following CONFIGs enabled, but I still don't see the 
>>> above issue on my machine.
>>>
>> The warning has to do with the memory allocation order being above 
>> the max order supported for kmalloc.  I think the problem is that 
>> ima_alloc_data_entry() is using kmemdup() to duplicate a payload of 
>> arbitrary size.  Policies on e.g. Fedora can be quite large, so you 
>> can't assume they can be allocated with kmalloc and friends.
>>
>
> Thanks for clarifying. Yes ima_alloc_entry() does use kmemdup to save 
> the given buffer (to be measured) until IMA loads custom policy.
>
> On my machine the SELinux policy size is about 2MB.
>
> Perhaps vmalloc would be better than using kmalloc? If there are 
> better options for such large buffer allocation, please let me know.

kvmalloc() can be used to select whichever one is most appropriate.