lkml.org 
[lkml]   [2019]   [Feb]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v4 3/3] powerpc/32: Add KASAN support
From
Date


On 2/11/19 3:25 PM, Andrey Konovalov wrote:
> On Sat, Feb 9, 2019 at 12:55 PM christophe leroy
> <christophe.leroy@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@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.

\
 
 \ /
  Last update: 2019-02-11 17:28    [W:0.072 / U:3.876 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site