Re: [PATCH 1/4] mm/page_alloc: fix non cma alloc context

[Joonsoo Kim]

2020ë 7ì 17ì (ê) ìí 5:15, Vlastimil Babka <vbabka@xxxxxxx>ëì ìì:
>
> On 7/17/20 10:10 AM, Vlastimil Babka wrote:
> > On 7/17/20 9:29 AM, Joonsoo Kim wrote:
> >> 2020ë 7ì 16ì (ë) ìí 4:45, Vlastimil Babka <vbabka@xxxxxxx>ëì ìì:
> >>>
> >>> On 7/16/20 9:27 AM, Joonsoo Kim wrote:
> >>> > 2020ë 7ì 15ì (ì) ìí 5:24, Vlastimil Babka <vbabka@xxxxxxx>ëì ìì:
> >>> >> >  /*
> >>> >> >   * get_page_from_freelist goes through the zonelist trying to allocate
> >>> >> >   * a page.
> >>> >> > @@ -3706,6 +3714,8 @@ get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
> >>> >> >       struct pglist_data *last_pgdat_dirty_limit = NULL;
> >>> >> >       bool no_fallback;
> >>> >> >
> >>> >> > +     current_alloc_flags(gfp_mask, &alloc_flags);
> >>> >>
> >>> >> I don't see why to move the test here? It will still be executed in the
> >>> >> fastpath, if that's what you wanted to avoid.
> >>> >
> >>> > I want to execute it on the fastpath, too. Reason that I moved it here
> >>> > is that alloc_flags could be reset on slowpath. See the code where
> >>> > __gfp_pfmemalloc_flags() is on. This is the only place that I can apply
> >>> > this option to all the allocation paths at once.
> >>>
> >>> But get_page_from_freelist() might be called multiple times in the slowpath, and
> >>> also anyone looking for gfp and alloc flags setup will likely not examine this
> >>> function. I don't see a problem in having it in two places that already deal
> >>> with alloc_flags setup, as it is now.
> >>
> >> I agree that anyone looking alloc flags will miss that function easily. Okay.
> >> I will place it on its original place, although we now need to add one
> >> more place.
> >> *Three places* are gfp_to_alloc_flags(), prepare_alloc_pages() and
> >> __gfp_pfmemalloc_flags().
> >
> > Hm the check below should also work for ALLOC_OOM|ALLOC_NOCMA then.
> >
> > /* Avoid allocations with no watermarks from looping endlessly */
> >    if (tsk_is_oom_victim(current) &&
> >         (alloc_flags == ALLOC_OOM ||
> >          (gfp_mask & __GFP_NOMEMALLOC)))
> >             goto nopage;
> >
> > Maybe it's simpler to change get_page_from_freelist() then. But document well.
>
> But then we have e.g. should_reclaim_retry() which calls __zone_watermark_ok()
> where ALLOC_CMA plays a role too, so that means we should have alloc_mask set up
> correctly wrt ALLOC_CMA at the __alloc_pages_slowpath() level...

Good catch! Hmm... Okay. It would be necessarily handled in three places.
I will fix it on the next version. Anyway, we need some clean-up about
alloc_flags
handling since it looks not good for maintenance.

Thanks.