[PATCH v4 3/3] powerpc/32: Add KASAN support

[Andrey Ryabinin]

On 2/12/19 4:08 AM, Daniel Axtens wrote:
> Andrey Ryabinin <aryabinin at virtuozzo.com> writes:
> 
>> On 2/11/19 3:25 PM, Andrey Konovalov wrote:
>>> On Sat, Feb 9, 2019 at 12:55 PM christophe leroy
>>> <christophe.leroy at c-s.fr> wrote:
>>>>
>>>> Hi Andrey,
>>>>
>>>> Le 08/02/2019 à 18:40, Andrey Konovalov a écrit :
>>>>> On Fri, Feb 8, 2019 at 6:17 PM Christophe Leroy <christophe.leroy at c-s.fr> wrote:
>>>>>>
>>>>>> Hi Daniel,
>>>>>>
>>>>>> Le 08/02/2019 à 17:18, Daniel Axtens a écrit :
>>>>>>> Hi Christophe,
>>>>>>>
>>>>>>> I've been attempting to port this to 64-bit Book3e nohash (e6500),
>>>>>>> although I think I've ended up with an approach more similar to Aneesh's
>>>>>>> much earlier (2015) series for book3s.
>>>>>>>
>>>>>>> Part of this is just due to the changes between 32 and 64 bits - we need
>>>>>>> to hack around the discontiguous mappings - but one thing that I'm
>>>>>>> particularly puzzled by is what the kasan_early_init is supposed to do.
>>>>>>
>>>>>> It should be a problem as my patch uses a 'for_each_memblock(memory,
>>>>>> reg)' loop.
>>>>>>
>>>>>>>
>>>>>>>> +void __init kasan_early_init(void)
>>>>>>>> +{
>>>>>>>> +    unsigned long addr = KASAN_SHADOW_START;
>>>>>>>> +    unsigned long end = KASAN_SHADOW_END;
>>>>>>>> +    unsigned long next;
>>>>>>>> +    pmd_t *pmd = pmd_offset(pud_offset(pgd_offset_k(addr), addr), addr);
>>>>>>>> +    int i;
>>>>>>>> +    phys_addr_t pa = __pa(kasan_early_shadow_page);
>>>>>>>> +
>>>>>>>> +    BUILD_BUG_ON(KASAN_SHADOW_START & ~PGDIR_MASK);
>>>>>>>> +
>>>>>>>> +    if (early_mmu_has_feature(MMU_FTR_HPTE_TABLE))
>>>>>>>> +            panic("KASAN not supported with Hash MMU\n");
>>>>>>>> +
>>>>>>>> +    for (i = 0; i < PTRS_PER_PTE; i++)
>>>>>>>> +            __set_pte_at(&init_mm, (unsigned long)kasan_early_shadow_page,
>>>>>>>> +                         kasan_early_shadow_pte + i,
>>>>>>>> +                         pfn_pte(PHYS_PFN(pa), PAGE_KERNEL_RO), 0);
>>>>>>>> +
>>>>>>>> +    do {
>>>>>>>> +            next = pgd_addr_end(addr, end);
>>>>>>>> +            pmd_populate_kernel(&init_mm, pmd, kasan_early_shadow_pte);
>>>>>>>> +    } while (pmd++, addr = next, addr != end);
>>>>>>>> +}
>>>>>>>
>>>>>>> As far as I can tell it's mapping the early shadow page, read-only, over
>>>>>>> the KASAN_SHADOW_START->KASAN_SHADOW_END range, and it's using the early
>>>>>>> shadow PTE array from the generic code.
>>>>>>>
>>>>>>> I haven't been able to find an answer to why this is in the docs, so I
>>>>>>> was wondering if you or anyone else could explain the early part of
>>>>>>> kasan init a bit better.
>>>>>>
>>>>>> See https://www.kernel.org/doc/html/latest/dev-tools/kasan.html for an
>>>>>> explanation of the shadow.
>>>>>>
>>>>>> When shadow is 0, it means the memory area is entirely accessible.
>>>>>>
>>>>>> It is necessary to setup a shadow area as soon as possible because all
>>>>>> data accesses check the shadow area, from the begining (except for a few
>>>>>> files where sanitizing has been disabled in Makefiles).
>>>>>>
>>>>>> Until the real shadow area is set, all access are granted thanks to the
>>>>>> zero shadow area beeing for of zeros.
>>>>>
>>>>> Not entirely correct. kasan_early_init() indeed maps the whole shadow
>>>>> memory range to the same kasan_early_shadow_page. However as kernel
>>>>> loads and memory gets allocated this shadow page gets rewritten with
>>>>> non-zero values by different KASAN allocator hooks. Since these values
>>>>> come from completely different parts of the kernel, but all land on
>>>>> the same page, kasan_early_shadow_page's content can be considered
>>>>> garbage. When KASAN checks memory accesses for validity it detects
>>>>> these garbage shadow values, but doesn't print any reports, as the
>>>>> reporting routine bails out on the current->kasan_depth check (which
>>>>> has the value of 1 initially). Only after kasan_init() completes, when
>>>>> the proper shadow memory is mapped, current->kasan_depth gets set to 0
>>>>> and we start reporting bad accesses.
>>>>
>>>> That's surprising, because in the early phase I map the shadow area
>>>> read-only, so I do not expect it to get modified unless RO protection is
>>>> failing for some reason.
>>>
>>> Actually it might be that the allocator hooks don't modify shadow at
>>> this point, as the allocator is not yet initialized. However stack
>>> should be getting poisoned and unpoisoned from the very start. But the
>>> generic statement that early shadow gets dirtied should be correct.
>>> Might it be that you don't use stack instrumentation?
>>>
>>
>> Yes, stack instrumentation is not used here, because shadow offset which we pass to
>> the -fasan-shadow-offset= cflag is not specified here. So the logic in scrpits/Makefile.kasan
>> just fallbacks to CFLAGS_KASAN_MINIMAL, which is outline and without stack instrumentation.
>>
>> Christophe, you can specify KASAN_SHADOW_OFFSET either in Kconfig (e.g. x86_64) or
>> in Makefile (e.g. arm64). And make early mapping writable, because compiler generated code will write
>> to shadow memory in function prologue/epilogue.
> 
> Hmm. Is this limitation just that compilers have not implemented
> out-of-line support for stack instrumentation, or is there a deeper
> reason that stack/global instrumentation relies upon inline
> instrumentation?
> 

Yes, it's simply wasn't implemented in compilers. Stack [un]poisoning code is always inlined.

But globals is the opposite of that, they all poisoned out-of-line via __asan_register_globals() call.

> I ask because it's very common on ppc64 to have the virtual address
> space split up into discontiguous blocks. I know this means we lose
> inline instrumentation, but I didn't realise we'd also lose stack and
> global instrumentation...
> 
> I wonder if it would be worth, in the distant future, trying to
> implement a smarter scheme in compilers where we could insert more
> complex inline mapping schemes.
> 

I'd say it depends on performance boost that inline might give for those complex inline schemes.
The whole inline instrumentation thing exists only because it gives better performance.


> Regards,
> Daniel
>