Re: slab_pre_alloc_hook() strips __GFP_NOLOCKDEP away.

[Vlastimil Babka <vbabka@suse.cz>]

On 3/23/22 12:30, Sebastian Andrzej Siewior wrote:
> On 2022-03-23 11:09:35 [+0100], Vlastimil Babka wrote:
>>> --- a/mm/slab.h
>>> +++ b/mm/slab.h
>>> @@ -717,7 +717,7 @@ static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
>>>  						     struct obj_cgroup **objcgp,
>>>  						     size_t size, gfp_t flags)
>>>  {
>>> -	flags &= gfp_allowed_mask;
>>> +	flags &= gfp_allowed_mask | __GFP_NOLOCKDEP;
>>
>> Hmm but gfp_allowed_mask already should contain __GFP_NOLOCKDEP after
>> kernel_init_freeable() is reached. I doubt we can reach fs or reclaim
>> code before that?
> 
> That is way past init. I have
>  gfp_allowed_mask = 0x1ffffff
>  __GFP_NOLOCKDEP = 0x8000000
> 
> Looking at the origin of gfp_allowed_mask
> 
> | /* Room for N __GFP_FOO bits */
> | #define __GFP_BITS_SHIFT (24 +                                          \
> |                           3 * IS_ENABLED(CONFIG_KASAN_HW_TAGS) +        \
> |                           IS_ENABLED(CONFIG_LOCKDEP))
> | #define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))

Mainline has this:
#define __GFP_BITS_SHIFT (25 + IS_ENABLED(CONFIG_LOCKDEP))

which looks OK.

The '24' comes from "kasan, mm: only define ___GFP_SKIP_KASAN_POISON
with HW_TAGS" in mmotm.
It makes ___GFP_SKIP_KASAN_POISON optional, that's why the decrease to
24. However this is not about the number of flags, but the highest bit,
which can be that of __GFP_NOLOCKDEP and it's then not covered by
__GFP_BITS_SHIFT if lockdep is enabled and kasan not...

> which in my case 24 + 1 but it would need to be something like
> 	27 * IS_ENABLED(CONFIG_LOCKDEP)
> 
> or a better way to come up with __GFP_BITS_MASK.
> 
> Sebastian