On Thu, Dec 3, 2020 at 4:17 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Wed 02-12-20 00:23:29, Pavel Tatashin wrote:
> [...]
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 611799c72da5..7a6d86d0bc5f 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -3766,20 +3766,25 @@ alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
> > return alloc_flags;
> > }
> >
> > -static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
> > - unsigned int alloc_flags)
> > +static inline unsigned int cma_alloc_flags(gfp_t gfp_mask,
> > + unsigned int alloc_flags)
> > {
> > #ifdef CONFIG_CMA
> > - unsigned int pflags = current->flags;
> > -
> > - if (!(pflags & PF_MEMALLOC_NOMOVABLE) &&
> > - gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
> > + if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
> > alloc_flags |= ALLOC_CMA;
> > -
> > #endif
> > return alloc_flags;
> > }
> >
> > +static inline gfp_t current_gfp_checkmovable(gfp_t gfp_mask)
> > +{
> > + unsigned int pflags = current->flags;
> > +
> > + if ((pflags & PF_MEMALLOC_NOMOVABLE))
> > + return gfp_mask & ~__GFP_MOVABLE;
> > + return gfp_mask;
> > +}
> > +
>
> It sucks that we have to control both ALLOC and gfp flags. But wouldn't
> it be simpler and more straightforward to keep current_alloc_flags as is
> (module PF rename) and hook the gfp mask evaluation into current_gfp_context
> and move it up before the first allocation attempt?
We could do that, but perhaps as a separate patch? I am worried about
hidden implication of adding extra scope (GFP_NOIO|GFP_NOFS) to the
fast path. Also, current_gfp_context() is used elsewhere, and in some
places removing __GFP_MOVABLE from gfp_mask means that we will need to
also change other things. For example [1], in try_to_free_pages() we
call current_gfp_context(gfp_mask) which can reduce the maximum zone
idx, yet we simply set it to: reclaim_idx = gfp_zone(gfp_mask), not to
the newly determined gfp_mask.
[1] https://soleen.com/source/xref/linux/mm/vmscan.c?r=2da9f630#3239
All scope flags
> should be applicable to the hot path as well. It would add few cycles to
> there but the question is whether that would be noticeable over just
> handling PF_MEMALLOC_NOMOVABLE on its own. The cache line would be
> pulled in anyway.
Let's try it in a separate patch? I will add it in the next version of
this series.
Thank you,
Pasha
