Re: [PATCH] kasan: Add explicit preconditions to kasan_report()

[Vincenzo Frascino]

On 1/20/21 4:04 PM, Catalin Marinas wrote:
> On Tue, Jan 19, 2021 at 08:35:49PM +0000, Vincenzo Frascino wrote:
>> On 1/19/21 6:52 PM, Catalin Marinas wrote:
>>> On Tue, Jan 19, 2021 at 07:27:43PM +0100, Andrey Konovalov wrote:
>>>> On Tue, Jan 19, 2021 at 6:26 PM Vincenzo Frascino
>>>> <vincenzo.frascino@xxxxxxx> wrote:
>>>>>
>>>>> With the introduction of KASAN_HW_TAGS, kasan_report() dereferences
>>>>> the address passed as a parameter.
>>>>>
>>>>> Add a comment to make sure that the preconditions to the function are
>>>>> explicitly clarified.
>>>>>
>>>>> Note: An invalid address (e.g. NULL pointer address) passed to the
>>>>> function when, KASAN_HW_TAGS is enabled, leads to a kernel panic.
>>>>>
>>>>> Cc: Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx>
>>>>> Cc: Alexander Potapenko <glider@xxxxxxxxxx>
>>>>> Cc: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
>>>>> Cc: Leon Romanovsky <leonro@xxxxxxxxxxxx>
>>>>> Cc: Andrey Konovalov <andreyknvl@xxxxxxxxxx>
>>>>> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@xxxxxxx>
>>>>> ---
>>>>>  mm/kasan/report.c | 11 +++++++++++
>>>>>  1 file changed, 11 insertions(+)
>>>>>
>>>>> diff --git a/mm/kasan/report.c b/mm/kasan/report.c
>>>>> index c0fb21797550..2485b585004d 100644
>>>>> --- a/mm/kasan/report.c
>>>>> +++ b/mm/kasan/report.c
>>>>> @@ -403,6 +403,17 @@ static void __kasan_report(unsigned long addr, size_t size, bool is_write,
>>>>>         end_report(&flags);
>>>>>  }
>>>>>
>>>>> +/**
>>>>> + * kasan_report - report kasan fault details
>>>>> + * @addr: valid address of the allocation where the tag fault was detected
>>>>> + * @size: size of the allocation where the tag fault was detected
>>>>> + * @is_write: the instruction that caused the fault was a read or write?
>>>>> + * @ip: pointer to the instruction that cause the fault
>>>>> + *
>>>>> + * Note: When CONFIG_KASAN_HW_TAGS is enabled kasan_report() dereferences
>>>>> + * the address to access the tags, hence it must be valid at this point in
>>>>> + * order to not cause a kernel panic.
>>>>> + */
>>>>
>>>> It doesn't dereference the address, it just checks the tags, right?
>>>>
>>>> Ideally, kasan_report() should survive that with HW_TAGS like with the
>>>> other modes. The reason it doesn't is probably because of a blank
>>>> addr_has_metadata() definition for HW_TAGS in mm/kasan/kasan.h. I
>>>> guess we should somehow check that the memory comes from page_alloc or
>>>> kmalloc. Or otherwise make sure that it has tags. Maybe there's an arm
>>>> instruction to check whether the memory has tags?
>>>
>>> There isn't an architected way to probe whether a memory location has a
>>> VA->PA mapping. The tags are addressed by PA but you can't reach them if
>>> you get a page fault on the VA. So we either document the kasan_report()
>>> preconditions or, as you suggest, update addr_has_metadata() for the
>>> HW_TAGS case. Something like:
>>>
>>>         return is_vmalloc_addr(virt) || virt_addr_valid(virt));
>>>
>>
>> This seems not working on arm64 because according to virt_addr_valid 0 is a
>> valid virtual address, in fact:
>>
>> __is_lm_address(0) == true && pfn_valid(virt_to_pfn(0)) == true.
> 
> Ah, so __is_lm_address(0) is true. Maybe we should improve this since
> virt_to_pfn(0) doesn't make much sense.
> 

How do you propose to improve it?

>> An option could be to make an exception for virtual address 0 in
>> addr_has_metadata() something like:
>>
>> static inline bool addr_has_metadata(const void *addr)
>> {
>> 	if ((u64)addr == 0)
>> 		return false;
>>
>> 	return (is_vmalloc_addr(addr) || virt_addr_valid(addr));
>> }
> 
> As Andrey replied, passing a non-zero small value would still be
> incorrectly detected as valid.
> 

I would like to remove the check completely and have virt_addr_valid(addr) to
return the right thing if possible.

I admit, yesterday evening I did not thing it through completely before posting
this code that had the sole purpose to open the discussion. I agree in principle
on what Andrey said as well (addr < PAGE_SIZE).

-- 
Regards,
Vincenzo