[PATCH v8 01/14] kasan: sw_tags: Use arithmetic shift for shadow computation

[Maciej Wieczor-Retman]

Thanks for looking at the patches :)

On 2026-01-15 at 23:42:02 +0100, Andrey Ryabinin wrote:
>
>
>On 1/12/26 6:27 PM, Maciej Wieczor-Retman wrote:
>  
>> diff --git a/mm/kasan/report.c b/mm/kasan/report.c
>> index 62c01b4527eb..b5beb1b10bd2 100644
>> --- a/mm/kasan/report.c
>> +++ b/mm/kasan/report.c
>> @@ -642,11 +642,39 @@ void kasan_non_canonical_hook(unsigned long addr)
>>  	const char *bug_type;
>>  
>>  	/*
>> -	 * All addresses that came as a result of the memory-to-shadow mapping
>> -	 * (even for bogus pointers) must be >= KASAN_SHADOW_OFFSET.
>> +	 * For Generic KASAN, kasan_mem_to_shadow() uses the logical right shift
>> +	 * and never overflows with the chosen KASAN_SHADOW_OFFSET values (on
>> +	 * both x86 and arm64). Thus, the possible shadow addresses (even for
>> +	 * bogus pointers) belong to a single contiguous region that is the
>> +	 * result of kasan_mem_to_shadow() applied to the whole address space.
>>  	 */
>> -	if (addr < KASAN_SHADOW_OFFSET)
>> -		return;
>> +	if (IS_ENABLED(CONFIG_KASAN_GENERIC)) {
>> +		if (addr < (unsigned long)kasan_mem_to_shadow((void *)(0ULL)) ||
>> +		    addr > (unsigned long)kasan_mem_to_shadow((void *)(~0ULL)))
>> +			return;
>> +	}
>> +
>> +	/*
>> +	 * For Software Tag-Based KASAN, kasan_mem_to_shadow() uses the
>> +	 * arithmetic shift. Normally, this would make checking for a possible
>> +	 * shadow address complicated, as the shadow address computation
>> +	 * operation would overflow only for some memory addresses. However, due
>> +	 * to the chosen KASAN_SHADOW_OFFSET values and the fact the
>> +	 * kasan_mem_to_shadow() only operates on pointers with the tag reset,
>> +	 * the overflow always happens.
>> +	 *
>> +	 * For arm64, the top byte of the pointer gets reset to 0xFF. Thus, the
>> +	 * possible shadow addresses belong to a region that is the result of
>> +	 * kasan_mem_to_shadow() applied to the memory range
>> +	 * [0xFF000000000000, 0xFFFFFFFFFFFFFFFF]. Despite the overflow, the
>                  ^ Missing couple 00 here
>
>> +	 * resulting possible shadow region is contiguous, as the overflow
>> +	 * happens for both 0xFF000000000000 and 0xFFFFFFFFFFFFFFFF.
>                                  ^ same as above

Hah, right, thank you!

>
>> +	 */
>> +	if (IS_ENABLED(CONFIG_KASAN_SW_TAGS) && IS_ENABLED(CONFIG_ARM64)) {
>> +		if (addr < (unsigned long)kasan_mem_to_shadow((void *)(0xFFULL << 56)) ||
>
>This will not work for inline mode because compiler uses logical shift.
>Consider NULL-ptr derefernce. Compiler will calculate shadow address for 0 as:
>      (((0x0 | 0xffULL) << 56) >> 4)+0xffff800000000000ULL = 0x0fef8000....0
>Which is less than ((0xFF00...00LL) >> 4) +  0xffff800000000000ULL = 0xffff800...0
>So we will bail out here.
>Perhaps we could do addr |= 0xFFLL to fix this

I suppose it should work; tried it in a python script by shoving various
addresses into this check. Pushing addresses through a logical shift
memory_to_shadow normally would return early as you noticed, and after 'addr |=
0xFFLL' it seems to work as expected. And I didn't really catch any incorrect
address slipping by this scheme either. Thanks, I'll correct it.

>
>> +		    addr > (unsigned long)kasan_mem_to_shadow((void *)(~0ULL)))
>> +			return;
>> +	}
>>  
>>  	orig_addr = (unsigned long)kasan_shadow_to_mem((void *)addr);
>>  

-- 
Kind regards
Maciej Wieczór-Retman