[RFC/PATCH RESEND -next 01/21] Add kernel address sanitizer infrastructure.

[Andrey Ryabinin]

On 07/10/14 17:31, Sasha Levin wrote:
> On 07/10/2014 09:01 AM, Andrey Ryabinin wrote:
>> On 07/10/14 15:55, Sasha Levin wrote:
>>>> On 07/09/2014 07:29 AM, Andrey Ryabinin wrote:
>>>>>> Address sanitizer for kernel (kasan) is a dynamic memory error detector.
>>>>>>
>>>>>> The main features of kasan is:
>>>>>>  - is based on compiler instrumentation (fast),
>>>>>>  - detects out of bounds for both writes and reads,
>>>>>>  - provides use after free detection,
>>>>>>
>>>>>> This patch only adds infrastructure for kernel address sanitizer. It's not
>>>>>> available for use yet. The idea and some code was borrowed from [1].
>>>>>>
>>>>>> This feature requires pretty fresh GCC (revision r211699 from 2014-06-16 or
>>>>>> latter).
>>>>>>
>>>>>> Implementation details:
>>>>>> The main idea of KASAN is to use shadow memory to record whether each byte of memory
>>>>>> is safe to access or not, and use compiler's instrumentation to check the shadow memory
>>>>>> on each memory access.
>>>>>>
>>>>>> Address sanitizer dedicates 1/8 of the low memory to the shadow memory and uses direct
>>>>>> mapping with a scale and offset to translate a memory address to its corresponding
>>>>>> shadow address.
>>>>>>
>>>>>> Here is function to translate address to corresponding shadow address:
>>>>>>
>>>>>>      unsigned long kasan_mem_to_shadow(unsigned long addr)
>>>>>>      {
>>>>>>                 return ((addr - PAGE_OFFSET) >> KASAN_SHADOW_SCALE_SHIFT)
>>>>>>                              + kasan_shadow_start;
>>>>>>      }
>>>>>>
>>>>>> where KASAN_SHADOW_SCALE_SHIFT = 3.
>>>>>>
>>>>>> So for every 8 bytes of lowmemory there is one corresponding byte of shadow memory.
>>>>>> The following encoding used for each shadow byte: 0 means that all 8 bytes of the
>>>>>> corresponding memory region are valid for access; k (1 <= k <= 7) means that
>>>>>> the first k bytes are valid for access, and other (8 - k) bytes are not;
>>>>>> Any negative value indicates that the entire 8-bytes are unaccessible.
>>>>>> Different negative values used to distinguish between different kinds of
>>>>>> unaccessible memory (redzones, freed memory) (see mm/kasan/kasan.h).
>>>>>>
>>>>>> To be able to detect accesses to bad memory we need a special compiler.
>>>>>> Such compiler inserts a specific function calls (__asan_load*(addr), __asan_store*(addr))
>>>>>> before each memory access of size 1, 2, 4, 8 or 16.
>>>>>>
>>>>>> These functions check whether memory region is valid to access or not by checking
>>>>>> corresponding shadow memory. If access is not valid an error printed.
>>>>>>
>>>>>> [1] https://code.google.com/p/address-sanitizer/wiki/AddressSanitizerForKernel
>>>>>>
>>>>>> Signed-off-by: Andrey Ryabinin <a.ryabinin at samsung.com>
>>>>
>>>> I gave it a spin, and it seems that it fails for what you might call a "regular"
>>>> memory size these days, in my case it was 18G:
>>>>
>>>> [    0.000000] Kernel panic - not syncing: ERROR: Failed to allocate 0xe0c00000 bytes below 0x0.
>>>> [    0.000000]
>>>> [    0.000000] CPU: 0 PID: 0 Comm: swapper Not tainted 3.16.0-rc4-next-20140710-sasha-00044-gb7b0579-dirty #784
>>>> [    0.000000]  ffffffffb9c2d3c8 cd9ce91adea4379a 0000000000000000 ffffffffb9c2d3c8
>>>> [    0.000000]  ffffffffb9c2d330 ffffffffb7fe89b7 ffffffffb93c8c28 ffffffffb9c2d3b8
>>>> [    0.000000]  ffffffffb7fcff1d 0000000000000018 ffffffffb9c2d3c8 ffffffffb9c2d360
>>>> [    0.000000] Call Trace:
>>>> [    0.000000] <UNK> dump_stack (lib/dump_stack.c:52)
>>>> [    0.000000] panic (kernel/panic.c:119)
>>>> [    0.000000] memblock_alloc_base (mm/memblock.c:1092)
>>>> [    0.000000] memblock_alloc (mm/memblock.c:1097)
>>>> [    0.000000] kasan_alloc_shadow (mm/kasan/kasan.c:151)
>>>> [    0.000000] zone_sizes_init (arch/x86/mm/init.c:684)
>>>> [    0.000000] paging_init (arch/x86/mm/init_64.c:677)
>>>> [    0.000000] setup_arch (arch/x86/kernel/setup.c:1168)
>>>> [    0.000000] ? printk (kernel/printk/printk.c:1839)
>>>> [    0.000000] start_kernel (include/linux/mm_types.h:462 init/main.c:533)
>>>> [    0.000000] ? early_idt_handlers (arch/x86/kernel/head_64.S:344)
>>>> [    0.000000] x86_64_start_reservations (arch/x86/kernel/head64.c:194)
>>>> [    0.000000] x86_64_start_kernel (arch/x86/kernel/head64.c:183)
>>>>
>>>> It got better when I reduced memory to 1GB, but then my system just failed to boot
>>>> at all because that's not enough to bring everything up.
>>>>
>> Thanks.
>> I think memory size is not a problem here. I tested on my desktop with 16G.
>> Seems it's a problem with memory holes cited by Dave.
>> kasan tries to allocate ~3.5G. It means that lowmemsize is 28G in your case.
> 
> That's correct (I've mistyped and got 18 instead of 28 above).
> 
> However, I'm a bit confused here, I thought highmem/lowmem split was a 32bit
> thing, so I'm not sure how it applies here.
> 
Right. By lowmemsize here I mean size of direct
mapping of all phys. memory (which usually called as lowmem on 32bit systems).



> Anyways, the machine won't boot with more than 1GB of RAM, is there a solution to
> get KASAN running on my machine?
> 
Could you share you .config? I'll try to boot it by myself. It could be that some options conflicting with kasan.
Also boot cmdline might help.

> 
> Thanks,
> Sasha
> 
>