Re: [PATCH 1/8] kasan,x86: Frob kasan_report() in an exception

[Andrey Ryabinin]

On 2/28/19 6:52 PM, Dmitry Vyukov wrote:
> On Thu, Feb 28, 2019 at 4:46 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>
>> On Thu, Feb 28, 2019 at 04:22:04PM +0100, Dmitry Vyukov wrote:
>>> On Thu, Feb 28, 2019 at 4:05 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>>>
>>>> Because __asan_{load,store}{N,1,2,4,8,16}_noabort() get called from
>>>> UACCESS context, and kasan_report() is most definitely _NOT_ safe to
>>>> be called from there, move it into an exception much like BUG/WARN.
>>>>
>>>> *compile tested only*
>>>
>>>
>>> Please test it by booting KASAN kernel and then loading module
>>> produced by CONFIG_TEST_KASAN=y. There are too many subtle aspects to
>>> rely on "compile tested only", reviewers can't catch all of them
>>> either.
>>
>> Sure, I'll do that. I just wanted to share the rest of the patches.
>>
>> A quick test shows it dies _REAAAAAAAALY_ early, as in:
>>
>> "Booting the kernel."
>>
>> is the first and very last thing it says... I wonder how I did that :-)
> 
> One thing is that during early boot kasan_report is called multiple
> times, but these are false positives related to the fact that we don't
> have a proper shadow yet (setup later). So during early boot we set
> kasan_disable=1 (or some global or per-task flag), and then
> kasan_report checks it and returns.
> Once we setup proper shadow, the flag is reset and from now on
> kasan_report actually reports bug.
> 

Yup, see report_enabled() function.