[PATCH V2] mm: Introduce GFP_PGTABLE

[Christophe Leroy]

Le 16/01/2019 à 14:18, Matthew Wilcox a écrit :
> On Wed, Jan 16, 2019 at 06:42:22PM +0530, Anshuman Khandual wrote:
>> On 01/16/2019 06:00 PM, Matthew Wilcox wrote:
>>> On Wed, Jan 16, 2019 at 07:57:03AM +0100, Michal Hocko wrote:
>>>> On Wed 16-01-19 11:51:32, Anshuman Khandual wrote:
>>>>> All architectures have been defining their own PGALLOC_GFP as (GFP_KERNEL |
>>>>> __GFP_ZERO) and using it for allocating page table pages. This causes some
>>>>> code duplication which can be easily avoided. GFP_KERNEL allocated and
>>>>> cleared out pages (__GFP_ZERO) are required for page tables on any given
>>>>> architecture. This creates a new generic GFP flag flag which can be used
>>>>> for any page table page allocation. Does not cause any functional change.
>>>>>
>>>>> GFP_PGTABLE is being added into include/asm-generic/pgtable.h which is the
>>>>> generic page tabe header just to prevent it's potential misuse as a general
>>>>> allocation flag if included in include/linux/gfp.h.
>>>>
>>>> I haven't reviewed the patch yet but I am wondering whether this is
>>>> really worth it without going all the way down to unify the common code
>>>> and remove much more code duplication. Or is this not possible for some
>>>> reason?
>>>
>>> Exactly what I suggested doing in response to v1.
>>>
>>> Also, the approach taken here is crazy.  x86 has a feature that no other
>>> architecture has bothered to implement yet -- accounting page tables
>>> to the process.  Yet instead of spreading that goodness to all other
>>> architectures, Anshuman has gone to more effort to avoid doing that.
>>
>> The basic objective for this patch is to create a common minimum allocation
>> flag that can be used by architectures but that still allows archs to add
>> on additional constraints if they see fit. This patch does not intend to
>> change functionality for any arch.
> 
> I disagree with your objective.  Making more code common is a great idea,
> but this patch is too unambitious.  We should be heading towards one or
> two page table allocation functions instead of having every architecture do
> its own thing.
> 
> So start there.  Move the x86 function into common code and convert one
> other architecture to use it too.

Are we talking about pte_alloc_one_kernel() and pte_alloc_one() ?

I'm not sure x86 function is the best common one, as it seems to 
allocate a multiple of PAGE_SIZE only.

Some arches like powerpc use pagetables which are smaller than a page, 
for instance powerpc 8xx uses 4k pagetables even with 16k pages, which 
means a single page can be used by 4 pagetables.

Therefore, I would suggest to start with powerpc functions.

Christophe