Re: [PATCH v3] mm/kasan: dump alloc and free stack for page allocator

[Andrey Ryabinin <aryabinin@virtuozzo.com>]

On 9/12/19 5:31 PM, Vlastimil Babka wrote:
> On 9/12/19 4:08 PM, Walter Wu wrote:
>>
>>>   extern void __reset_page_owner(struct page *page, unsigned int order);
>>> diff --git a/lib/Kconfig.kasan b/lib/Kconfig.kasan
>>> index 6c9682ce0254..dc560c7562e8 100644
>>> --- a/lib/Kconfig.kasan
>>> +++ b/lib/Kconfig.kasan
>>> @@ -41,6 +41,8 @@ config KASAN_GENERIC
>>>       select SLUB_DEBUG if SLUB
>>>       select CONSTRUCTORS
>>>       select STACKDEPOT
>>> +    select PAGE_OWNER
>>> +    select PAGE_OWNER_FREE_STACK
>>>       help
>>>         Enables generic KASAN mode.
>>>         Supported in both GCC and Clang. With GCC it requires version 4.9.2
>>> @@ -63,6 +65,8 @@ config KASAN_SW_TAGS
>>>       select SLUB_DEBUG if SLUB
>>>       select CONSTRUCTORS
>>>       select STACKDEPOT
>>> +    select PAGE_OWNER
>>> +    select PAGE_OWNER_FREE_STACK
>>>       help
>>
>> What is the difference between PAGE_OWNER+PAGE_OWNER_FREE_STACK and
>> DEBUG_PAGEALLOC?
> 
> Same memory usage, but debug_pagealloc means also extra checks and restricting memory access to freed pages to catch UAF.
> 
>> If you directly enable PAGE_OWNER+PAGE_OWNER_FREE_STACK
>> PAGE_OWNER_FREE_STACK,don't you think low-memory device to want to use
>> KASAN?
> 
> OK, so it should be optional? But I think it's enough to distinguish no PAGE_OWNER at all, and PAGE_OWNER+PAGE_OWNER_FREE_STACK together - I don't see much point in PAGE_OWNER only for this kind of debugging.
> 
> So how about this? KASAN wouldn't select PAGE_OWNER* but it would be recommended in the help+docs. When PAGE_OWNER and KASAN are selected by user, PAGE_OWNER_FREE_STACK gets also selected, and both will be also runtime enabled without explicit page_owner=on.
> I mostly want to avoid another boot-time option for enabling PAGE_OWNER_FREE_STACK.
> Would that be enough flexibility for low-memory devices vs full-fledged debugging?

Originally I thought that with you patch users still can disable page_owner via "page_owner=off" boot param.
But now I realized that this won't work. I think it should work, we should allow users to disable it.



Or another alternative option (and actually easier one to implement), leave PAGE_OWNER as is (no "select"s in Kconfigs)
Make PAGE_OWNER_FREE_STACK like this:

+config PAGE_OWNER_FREE_STACK
+	def_bool KASAN || DEBUG_PAGEALLOC
+	depends on PAGE_OWNER
+

So, users that want alloc/free stack will have to enable CONFIG_PAGE_OWNER=y and add page_owner=on to boot cmdline.


Basically the difference between these alternative is whether we enable page_owner by default or not. But there is always a possibility to disable it.