Re: [PATCH] mm: Introduce GFP_PGTABLE

[Anshuman Khandual]

On 01/12/2019 09:19 PM, Matthew Wilcox wrote:
> On Sat, Jan 12, 2019 at 02:49:29PM +0100, Christophe Leroy wrote:
>> As far as I can see,
>>
>> #define GFP_KERNEL_ACCOUNT (GFP_KERNEL | __GFP_ACCOUNT)
>>
>> So what's the difference between:
>>
>> (GFP_KERNEL_ACCOUNT | __GFP_ZERO) & ~__GFP_ACCOUNT
>>
>> and
>>
>> (GFP_KERNEL | __GFP_ZERO) & ~__GFP_ACCOUNT
> 
> Nothing.  But there's a huge difference in the other parts of that same
> file where GFP_ACCOUNT is _not_ used.
> 
> I think this unification is too small to bother with.  Something I've
> had on my todo list for some time and have not done anything about
> is to actually unify all of the architecture pte/pmd/... allocations.
> There are tricks some architectures use that others would benefit from.

Sure. Could you please elaborate on this ?

Invariably all kernel pgtable page allocations should use GFP_PGTABLE and
all user page table allocation should use (GFP_PGTABLE | __GFP_ACCOUNT).
Ideally there should be default generic functions user_pgtable_gfp() or
kernel_pgtable_gfp() returning these values. Overrides can be provided if
an arch still wants some more control.