[PATCH] mm: Introduce GFP_PGTABLE

[Michal Hocko]

On Mon 14-01-19 09:30:55, Anshuman Khandual wrote:
> 
> 
> On 01/13/2019 11:05 PM, Michal Hocko wrote:
> > On Sat 12-01-19 15:56:38, 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.
> > 
> > I agree that some unification is due but GFP_PGTABLE is not something to
> > expose in generic gfp.h IMHO. It just risks an abuse. I would be looking
> 
> Why would you think that it risks an abuse ? It does not create new semantics
> of allocation in the buddy. Its just uses existing GFP_KERNEL allocation which
> is then getting zeroed out. The risks (if any) is exactly same as GFP_KERNEL.

Beucase my experience just tells me that people tend to use whatever
they find and name fits what they think they need.

> > at providing asm-generic implementation and reuse it to remove the code
> 
> Does that mean GFP_PGTABLE can be created but not in gfp.h but in some other
> memory related header file ?

I would just keep it close to its users. If that is a single arch
generic place then only better. But I suspect some arches have special
requirements.

> > duplication. But I haven't tried that to know that it will work out due
> > to small/subtle differences between arches.
> 
> IIUC from the allocation perspective GFP_ACCOUNT is the only thing which gets
> added with GFP_PGTABLE for user page table for memcg accounting purpose. There
> does not seem to be any other differences unless I am missing something.

It's been some time since I've checked the last time. Some arches were
using GPF_REPEAT (__GFP_RETRY_MAYFAIL) back then. I have removed most of
those but some were doing a higher order allocations so they probably
have stayed.
-- 
Michal Hocko
SUSE Labs