Hi Marco and Dmitry, any comments about the following replay, thanks.

On 2021/7/6 12:07, Kefeng Wang wrote:
>
> Hi Marco and Dmitry,
>
> On 2021/7/5 23:04, Marco Elver wrote:
>> On Mon, Jul 05, 2021 at 07:14PM +0800, Kefeng Wang wrote:
>> [...]
>>> +#ifdef CONFIG_KASAN_VMALLOC
>>> +void __init __weak kasan_populate_early_vm_area_shadow(void *start,
>>> +						       unsigned long size)
>> This should probably not be __weak, otherwise you now have 2 __weak
>> functions.
> Indeed, forget it.
>>> +{
>>> +	unsigned long shadow_start, shadow_end;
>>> +
>>> +	if (!is_vmalloc_or_module_addr(start))
>>> +		return;
>>> +
>>> +	shadow_start = (unsigned long)kasan_mem_to_shadow(start);
>>> +	shadow_start = ALIGN_DOWN(shadow_start, PAGE_SIZE);
>>> +	shadow_end = (unsigned long)kasan_mem_to_shadow(start + size);
>>> +	shadow_end = ALIGN(shadow_end, PAGE_SIZE);
>>> +	kasan_map_populate(shadow_start, shadow_end,
>>> +			   early_pfn_to_nid(virt_to_pfn(start)));
>>> +}
>>> +#endif
>> This function looks quite generic -- would any of this also apply to
>> other architectures? I see that ppc and sparc at least also define
>> CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK.
>
> I can't try ppc/sparc, but only ppc support KASAN_VMALLOC,
>
> I check the x86, it supports CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK,
>
> looks this issue is existing on x86 and ppc.
>
>>>   void __init kasan_init(void)
>>>   {
>>>   	kasan_init_shadow();
>>> diff --git a/include/linux/kasan.h b/include/linux/kasan.h
>>> index 5310e217bd74..79d3895b0240 100644
>>> --- a/include/linux/kasan.h
>>> +++ b/include/linux/kasan.h
>>> @@ -49,6 +49,8 @@ extern p4d_t kasan_early_shadow_p4d[MAX_PTRS_PER_P4D];
>>>   int kasan_populate_early_shadow(const void *shadow_start,
>>>   				const void *shadow_end);
>>>   
>>> +void kasan_populate_early_vm_area_shadow(void *start, unsigned long size);
>>> +
>>>   static inline void *kasan_mem_to_shadow(const void *addr)
>>>   {
>>>   	return (void *)((unsigned long)addr >> KASAN_SHADOW_SCALE_SHIFT)
>>> diff --git a/mm/kasan/init.c b/mm/kasan/init.c
>>> index cc64ed6858c6..d39577d088a1 100644
>>> --- a/mm/kasan/init.c
>>> +++ b/mm/kasan/init.c
>>> @@ -279,6 +279,11 @@ int __ref kasan_populate_early_shadow(const void *shadow_start,
>>>   	return 0;
>>>   }
>>>   
>>> +void __init __weak kasan_populate_early_vm_area_shadow(void *start,
>>> +						       unsigned long size)
>>> +{
>>> +}
>> I'm just wondering if this could be a generic function, perhaps with an
>> appropriate IS_ENABLED() check of a generic Kconfig option
>> (CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK ?) to short-circuit it, if it's
>> not only an arm64 problem.
>
> kasan_map_populate() is arm64 special function, and the x86 has kasan_shallow_populate_pgds(),
> ppc has kasan_init_shadow_page_tables(), so look those ARCHs should do the same way like ARM64,
>
> Here we can't use kasan_populate_early_shadow(), this functions will make the early shadow maps
> everything to a single page of zeroes（kasan_early_shadow_page), and set it pte_wrprotect, see
> zero_pte_populate(), right?
>
> Also I try this, it crashs on ARM64 when change kasan_map_populate() to kasan_populate_early_shadow(),
>
> Unable to handle kernel write to read-only memory at virtual address ffff700002938000
> ...
> Call trace:
>   __memset+0x16c/0x1c0
>   kasan_unpoison+0x34/0x6c
>   kasan_unpoison_vmalloc+0x2c/0x3c
>   __get_vm_area_node.constprop.0+0x13c/0x240
>   __vmalloc_node_range+0xf4/0x4f0
>   __vmalloc_node+0x80/0x9c
>   init_IRQ+0xe8/0x130
>   start_kernel+0x188/0x360
>   __primary_switched+0xc0/0xc8
>
>
>> But I haven't looked much further, so would appeal to you to either
>> confirm or reject this idea.
>>
>> Thanks,
>> -- Marco
>> .
>>