Re: [PATCH v6 27/39] kasan, mm: only define ___GFP_SKIP_KASAN_POISON with HW_TAGS

[Vlastimil Babka <vbabka@suse.cz>]

On 3/23/22 14:36, Andrey Konovalov wrote:
> On Wed, Mar 23, 2022 at 2:02 PM Sebastian Andrzej Siewior
> <bigeasy@linutronix.de> wrote:
>>
>> On 2022-03-23 12:48:29 [+0100], Vlastimil Babka wrote:
>>>> +#ifdef CONFIG_KASAN_HW_TAGS
>>>>  #define ___GFP_SKIP_KASAN_POISON   0x1000000u
>>>> +#else
>>>> +#define ___GFP_SKIP_KASAN_POISON   0
>>>> +#endif
>>>>  #ifdef CONFIG_LOCKDEP
>>>>  #define ___GFP_NOLOCKDEP   0x2000000u
>>>>  #else
>>>> @@ -251,7 +255,9 @@ struct vm_area_struct;
>>>>  #define __GFP_NOLOCKDEP ((__force gfp_t)___GFP_NOLOCKDEP)
>>>>
>>>>  /* Room for N __GFP_FOO bits */
>>>> -#define __GFP_BITS_SHIFT (25 + IS_ENABLED(CONFIG_LOCKDEP))
>>>> +#define __GFP_BITS_SHIFT (24 +                                     \
>>>> +                     IS_ENABLED(CONFIG_KASAN_HW_TAGS) +    \
>>>> +                     IS_ENABLED(CONFIG_LOCKDEP))
>>>
>>> This breaks __GFP_NOLOCKDEP, see:
>>> https://lore.kernel.org/all/YjoJ4CzB3yfWSV1F@linutronix.de/
>>
>> This could work because ___GFP_NOLOCKDEP is still 0x2000000u. In
>>         ("kasan, page_alloc: allow skipping memory init for HW_TAGS")
>>         https://lore.kernel.org/all/0d53efeff345de7d708e0baa0d8829167772521e.1643047180.git.andreyknvl@google.com/
>>
>> This is replaced with 0x8000000u which breaks lockdep.
>>
>> Sebastian
> 
> Hi Sebastian,
> 
> Indeed, sorry for breaking lockdep. Thank you for the report!
> 
> I wonder what's the proper fix for this. Perhaps, don't hide KASAN GFP
> bits under CONFIG_KASAN_HW_TAGS? And then do:
> 
> #define __GFP_BITS_SHIFT (27 + IS_ENABLED(CONFIG_LOCKDEP))
> 
> Vlastimil, Andrew do you have any preference?

I guess it's the simplest thing to do for now. For the future we can
still improve and handle all combinations of kasan/lockdep to occupy as
few bits as possible and set the shift/mask appropriately. Or consider
first if it's necessary anyway. I don't know if we really expect at any
point to start triggering the BUILD_BUG_ON() in radix_tree_init() and
then only some combination of configs will reduce the flags to a number
that works. Or is there anything else that depends on __GFP_BITS_SHIFT?
I mean if we don't expect to go this way, we can just define
__GFP_BITS_SHIFT as a constant that assumes all the config-dependent
flags to be defined (not zero).

> If my suggestion sounds good, Andrew, could you directly apply the
> changes? They are needed for these 3 patches:
> 
> kasan, page_alloc: allow skipping memory init for HW_TAGS
> kasan, page_alloc: allow skipping unpoisoning for HW_TAGS
> kasan, mm: only define ___GFP_SKIP_KASAN_POISON with HW_TAGS
> 
> As these depend on each other, I can't send separate patches that can
> be folded for all 3.
> 
> Thanks!