Re: [PATCH] memcg: page_alloc: skip bulk allocator for __GFP_ACCOUNT

[Roman Gushchin <guro@fb.com>]

On Wed, Oct 13, 2021 at 03:26:11PM -0700, Shakeel Butt wrote:
> On Wed, Oct 13, 2021 at 3:03 PM Roman Gushchin <guro@fb.com> wrote:
> >
> > On Wed, Oct 13, 2021 at 12:43:38PM -0700, Shakeel Butt wrote:
> > > The commit 5c1f4e690eec ("mm/vmalloc: switch to bulk allocator in
> > > __vmalloc_area_node()") switched to bulk page allocator for order 0
> > > allocation backing vmalloc. However bulk page allocator does not support
> > > __GFP_ACCOUNT allocations and there are several users of
> > > kvmalloc(__GFP_ACCOUNT).
> > >
> > > For now make __GFP_ACCOUNT allocations bypass bulk page allocator. In
> > > future if there is workload that can be significantly improved with the
> > > bulk page allocator with __GFP_ACCCOUNT support, we can revisit the
> > > decision.
> > >
> > > Fixes: 5c1f4e690eec ("mm/vmalloc: switch to bulk allocator in __vmalloc_area_node()")
> > > Signed-off-by: Shakeel Butt <shakeelb@google.com>
> > > ---
> > >  mm/page_alloc.c | 4 ++++
> > >  1 file changed, 4 insertions(+)
> > >
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 668edb16446a..b3acad4615d3 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -5215,6 +5215,10 @@ unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
> > >       unsigned int alloc_flags = ALLOC_WMARK_LOW;
> > >       int nr_populated = 0, nr_account = 0;
> > >
> > > +     /* Bulk allocator does not support memcg accounting. */
> > > +     if (unlikely(gfp & __GFP_ACCOUNT))
> > > +             goto out;
> > > +
> >
> > Isn't it a bit too aggressive?
> >
> > How about
> >     if (WARN_ON_ONCE(gfp & __GFP_ACCOUNT))
> 
> We actually know that kvmalloc(__GFP_ACCOUNT) users exist and can
> trigger bulk page allocator through vmalloc, so I don't think the
> warning would be any helpful.
> 
> >        gfp &= ~__GFP_ACCOUNT;
> 
> Bulk allocator is best effort, so callers have adequate fallbacks.
> Transparently disabling accounting would be unexpected.

I see...

Shouldn't we then move this check to an upper level?

E.g.:

if (!(gfp & __GFP_ACCOUNT))
   call_into_bulk_allocator();
else
   call_into_per_page_allocator();

Not a big deal, I'm just worried about potential silent memory allocation
failures.

Thanks!