Message-ID: <CAAmzW4MWwDwqtXukHZ48g9O40DPJeidJmKdG5pMV4k9TefMfkQ@mail.gmail.com>
Date:   Fri, 17 Apr 2020 12:57:11 +0900
From:   Joonsoo Kim <js1304@...il.com>
To:     Johannes Weiner <hannes@...xchg.org>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        Linux Memory Management List <linux-mm@...ck.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Michal Hocko <mhocko@...nel.org>,
        Hugh Dickins <hughd@...gle.com>,
        Minchan Kim <minchan@...nel.org>,
        Vlastimil Babka <vbabka@...e.cz>,
        Mel Gorman <mgorman@...hsingularity.net>, kernel-team@....com,
        Joonsoo Kim <iamjoonsoo.kim@....com>
Subject: Re: [PATCH v5 05/10] mm/swap: charge the page when adding to the swap cache

2020년 4월 17일 (금) 오후 12:31, Johannes Weiner <hannes@...xchg.org>님이 작성:
>
> On Fri, Apr 17, 2020 at 10:38:53AM +0900, Joonsoo Kim wrote:
> > 2020년 4월 17일 (금) 오전 1:11, Johannes Weiner <hannes@...xchg.org>님이 작성:
> > >
> > > Hello Joonsoo,
> > >
> > > On Fri, Apr 03, 2020 at 02:40:43PM +0900, js1304@...il.com wrote:
> > > > @@ -112,7 +112,7 @@ void show_swap_cache_info(void)
> > > >   * but sets SwapCache flag and private instead of mapping and index.
> > > >   */
> > > >  int add_to_swap_cache(struct page *page, swp_entry_t entry,
> > > > -                     gfp_t gfp, void **shadowp)
> > > > +                     struct vm_area_struct *vma, gfp_t gfp, void **shadowp)
> > > >  {
> > > >       struct address_space *address_space = swap_address_space(entry);
> > > >       pgoff_t idx = swp_offset(entry);
> > > > @@ -120,14 +120,26 @@ int add_to_swap_cache(struct page *page, swp_entry_t entry,
> > > >       unsigned long i, nr = compound_nr(page);
> > > >       unsigned long nrexceptional = 0;
> > > >       void *old;
> > > > +     bool compound = !!compound_order(page);
> > > > +     int error;
> > > > +     struct mm_struct *mm = vma ? vma->vm_mm : current->mm;
> > > > +     struct mem_cgroup *memcg;
> > > >
> > > >       VM_BUG_ON_PAGE(!PageLocked(page), page);
> > > >       VM_BUG_ON_PAGE(PageSwapCache(page), page);
> > > >       VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
> > > >
> > > >       page_ref_add(page, nr);
> > > > +     /* PageSwapCache() prevent the page from being re-charged */
> > > >       SetPageSwapCache(page);
> > > >
> > > > +     error = mem_cgroup_try_charge(page, mm, gfp, &memcg, compound);
> > > > +     if (error) {
> > > > +             ClearPageSwapCache(page);
> > > > +             page_ref_sub(page, nr);
> > > > +             return error;
> > > > +     }
> > > > +
> > > >       do {
> > > >               xas_lock_irq(&xas);
> > > >               xas_create_range(&xas);
> > > > @@ -153,11 +165,16 @@ int add_to_swap_cache(struct page *page, swp_entry_t entry,
> > > >               xas_unlock_irq(&xas);
> > > >       } while (xas_nomem(&xas, gfp));
> > > >
> > > > -     if (!xas_error(&xas))
> > > > +     if (!xas_error(&xas)) {
> > > > +             mem_cgroup_commit_charge(page, memcg, false, compound);
> > >
> > > Unfortunately you cannot commit here yet because the rmap isn't set up
> > > and that will cause memcg_charge_statistics() to account the page
> > > incorrectly as file. And rmap is only set up during a page fault.
> >
> > I also found this problem a few days ago. In my investigation, what we need for
> > anonymous page to make accounting correct is a way to find the type of the page,
> > file or anon, since there is no special code to use the rmap. And, I
> > think that it
> > could be done by checking NULL mapping or something else.
>
> page->mapping is NULL for truncated file pages, file pages before they
> are inserted into the page cache, and anon pages before the rmap. It's
> not straight-forward to robustly tell those apart inside memcg.

Okay.

> But fundamentally, it's a silly problem to fix. We only need to tell
> page types apart to maintain the MEMCG_CACHE and MEMCG_RSS
> counters. But these are unnecessary duplicates of the NR_FILE_PAGES
> and NR_ANON_MAPPED vmstat counters - counters for which we already
> have accounting sites in generic code, and that already exist in the
> per-cgroup vmstat arrays. We just need to link them.
>
> So that's what I'm fixing instead: I'm adjusting the charging sequence
> slightly so that page->mem_cgroup is stable when the core VM code
> accounts pages by type. And then I'm hooking into these places with
> mod_lruvec_page_state and friends, and ditching MEMCG_CACHE/MEMCG_RSS.
>
> After that, memcg doesn't have to know about the types of pages at
> all. It can focus on maintaining page_counters and page->mem_cgroup,
> and leave the vmstat breakdown to generic VM code.
>
> Then we can charge pages right after allocation, regardless of type.
>
> [ Eventually, the memcg accounting interface shouldn't be much more
>   than GFP_ACCOUNT (with memalloc_use_memcg() for faults, swap etc.),
>   and the vmstat hooks. My patches don't quite get there, but that's
>   the direction they're pushing. ]

Thanks for explanation! It will help me understand your future patches.

> > Is there anything I missed? And, I cannot find the function,
> > memcg_charge_statistics(). Please let me know the file name of this
> > function.
>
> The correct name is mem_cgroup_charge_statistics(), my apologies.

No problem.

> > > This needs a bit of a rework of the memcg charging code that appears
> > > out of scope for your patches. I'm preparing a series right now to do
> > > just that. It'll also fix the swap ownership tracking problem when the
> > > swap controller is disabled, so we don't have to choose between
> > > charging the wrong cgroup or hampering swap readahead efficiency.
> >
> > Sound good! I also think that these patches are out of scope of my series.
> > I will wait your patches. Could you let me know when your series is submitted?
> > I'd like to plan my work schedule based on your patch schedule.
>
> I just finished the first draft of the series a few minutes ago. I'll
> polish it a bit, make sure it compiles etc. ;-), and send it out for
> review, testing and rebasing, hopefully tomorrow or early next week.

Sound good! See you again next week. :)

Thanks.

Powered by blists - more mailing lists