[PATCH] arm: Use kernel mm when updating section permissions

[Kees Cook]

On Fri, Nov 6, 2015 at 1:06 PM, Kevin Hilman <khilman at kernel.org> wrote:
> Kees Cook <keescook at chromium.org> writes:
>
>> On Fri, Nov 6, 2015 at 12:11 PM, Kevin Hilman <khilman at kernel.org> wrote:
>>> On Fri, Nov 6, 2015 at 11:12 AM, Kees Cook <keescook at chromium.org> wrote:
>>>
>>> [...]
>>>
>>>> Hi Kevin and Kernel CI folks,
>>>>
>>>> Could lkdtm get added to the kernel-CI workflows? Extracting and
>>>> validating Oops details when poking lkdtm would be extremely valuable
>>>> for these cases. :)
>>>
>>> Yeah, we can add that.
>>>
>>> What arches should we expect this to be working on?  For starters
>>
>> This is a great question. ;) They're a mix of CONFIG and hardware
>> feature specific, so probably they should be run on all architectures
>> and we can figure out what's missing in each case.
>>
>> Everything built with CONFIG_DEBUG_RODATA should pass these:
>>
>> WRITE_RO
>> WRITE_KERN
>> EXEC_DATA
>> EXEC_STACK
>> EXEC_KMALLOC
>> EXEC_VMALLOC
>>
>> But architectures without CONFIG_DEBUG_RODATA should be shamed. ;)
>>
>> Passing EXEC_USERSPACE requires SMEP on x86, and PXN on arm64.
>> Passing ACCESS_USERSPACE rquires SMAP on x86, and PAN on arm64.
>>
>> The recent PAN emulation CONFIG_CPU_SW_DOMAIN_PAN on non-LPAE arm
>> should cover ACCESS_USERSPACE too, and maybe EXEC_USERSPACE, but I
>> haven't taken a close look.
>
> A quick test on arm32 and both ACCESS_ and EXEC_USERSPACE tests pass
> (meaning they trigger the WARNs).

I'd expect a full Oops, not a WARN, but maybe CONFIG_CPU_SW_DOMAIN_PAN
needs to use a bigger hammer.

Russell, what sort of trap is DOMAIN_PAN expected to be triggering?

>> It might be useful, frankly, to test everything in lkdtm.
>
> So I gave this a quick spin on an ARM board (qcom-apq8064-ifc6410)
> using a dumb script[1] (for now avoiding the tests that cause a lockup
> so I can test multiple features without a reboot.)  Seems like most of
> them are producing a failure.
>
> However, this got me to thinking that one should probably write a
> kselftest for this feature, and catch quite a few issues with the ones
> that don't cause a hard lockup.  One would just need to be a bit smarter
> than my script and do something to trap SIG* (or the parent catching
> SIGCHLD) in order to be able to help determine failure, then grab the
> dmesg and log it.
>
> Having these test integrated into kselftest, and maintained along with
> the the kernel features would be *way* better than trying to maintain a
> set of tests in kernel CI for this feature, since right now we're
> working just building/running all the selftests automatically.
>
> What do you think about coming up with a kselftest for this stuff?  At
> least the non-lockup stuff?

Well, all the stuff I wrote tests for in lkdtm expect the kernel to
entirely Oops, and examining the Oops from outside is needed to verify
it was the correct type of Oops. I don't think testing via lkdtm can
be done from kselftest sensibly.

-Kees

>
> I'm not volunteering to write up the kselftest, but I will guarantee
> that it get run on a broad range of boards once it exists. :)
>
> Kevin
>
> [1]
> #!/bin/sh
>
> crash_test_dummy() {
>   echo $1> /sys/kernel/debug/provoke-crash/DIRECT
> }
>
> # Find all the tests that don't lockup
> TESTS=$(cat /sys/kernel/debug/provoke-crash/DIRECT |grep -v types| grep -v LOCK |grep -v PANIC)
>
> for test in $TESTS; do
>   echo "Performing test: $test"
>   crash_test_dummy $test &
>   sleep 1
> done



-- 
Kees Cook
Chrome OS Security