[PATCH v10 05/14] mm: multi-gen LRU: groundwork

[Yu Zhao]

On Tue, Apr 26, 2022 at 5:42 PM Andrew Morton <akpm at linux-foundation.org> wrote:
>
> On Tue, 26 Apr 2022 16:39:07 -0600 Yu Zhao <yuzhao at google.com> wrote:
>
> > On Mon, Apr 11, 2022 at 8:16 PM Andrew Morton <akpm at linux-foundation.org> wrote:
> > >
> > > On Wed,  6 Apr 2022 21:15:17 -0600 Yu Zhao <yuzhao at google.com> wrote:
> > >
> > > > Evictable pages are divided into multiple generations for each lruvec.
> > > > The youngest generation number is stored in lrugen->max_seq for both
> > > > anon and file types as they are aged on an equal footing. The oldest
> > > > generation numbers are stored in lrugen->min_seq[] separately for anon
> > > > and file types as clean file pages can be evicted regardless of swap
> > > > constraints. These three variables are monotonically increasing.
> > > >
> > > > ...
> > > >
> > > > +static inline bool lru_gen_del_folio(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
> > >
> > > There's a lot of function inlining here.  Fortunately the compiler will
> > > ignore it all, because some of it looks wrong.  Please review (and
> > > remeasure!).  If inlining is reqlly justified, use __always_inline, and
> > > document the reasons for doing so.
> >
> > I totally expect modern compilers to make better decisions than I do.
> > And personally, I'd never use __always_inline; instead, I'd strongly
> > recommend FDO/LTO.
>
> My (badly expressed) point is that there's a lot of inlining of large
> functions here.
>
> For example, lru_gen_add_folio() is huge and has 4(?) call sites.  This
> may well produce slower code due to the icache footprint.
>
> Experiment: moving lru_gen_del_folio() into mm/vmscan.c shrinks that
> file's .text from 80612 bytes to 78956.
>
> I tend to think that out-of-line regular old C functions should be the
> default and that the code should be inlined only when a clear benefit
> is demonstrable, or has at least been seriously thought about.

I can move those functions to vmscan.c if you think it would improve
performance. I don't have a strong opinion here -- I was able to
measure the bloat but not the performance impact.

> > > > --- a/mm/Kconfig
> > > > +++ b/mm/Kconfig
> > > > @@ -909,6 +909,14 @@ config ANON_VMA_NAME
> > > >         area from being merged with adjacent virtual memory areas due to the
> > > >         difference in their name.
> > > >
> > > > +config LRU_GEN
> > > > +     bool "Multi-Gen LRU"
> > > > +     depends on MMU
> > > > +     # the following options can use up the spare bits in page flags
> > > > +     depends on !MAXSMP && (64BIT || !SPARSEMEM || SPARSEMEM_VMEMMAP)
> > > > +     help
> > > > +       A high performance LRU implementation to overcommit memory.
> > > > +
> > > >  source "mm/damon/Kconfig"
> > >
> > > This is a problem.  I had to jump through hoops just to be able to
> > > compile-test this.  Turns out I had to figure out how to disable
> > > MAXSMP.
> > >
> > > Can we please figure out a way to ensure that more testers are at least
> > > compile testing this?  Allnoconfig, defconfig, allyesconfig, allmodconfig.
> > >
> > > Also, I suggest that we actually make MGLRU the default while in linux-next.
> >
> > The !MAXSMP is to work around [1], which I haven't had the time to
> > fix. That BUILD_BUG_ON() shouldn't assert sizeof(struct page) == 64
> > since the true size depends on WANT_PAGE_VIRTUAL as well as
> > LAST_CPUPID_NOT_IN_PAGE_FLAGS. My plan is here [2].
> >
> > [1] https://lore.kernel.org/r/20190905154603.10349-4-aneesh.kumar@linux.ibm.com/
> > [2] https://lore.kernel.org/r/Ygl1Gf+ATBuI%2Fm2q@google.com/
>
> OK, thanks.  This is fairly urgent for -next and -rc inclusion.  If
> practically nobody is compiling the feature then practically nobody is
> testing it.  Let's come up with a way to improves the expected coverage
> by a lot.

Let me just remove !MAXSMP, since I wasn't able to reproduce this
build error [1] anymore.

[1] https://lore.kernel.org/r/1792f0b2e29.d72f70c9807100.8179330337708563324@xanmod.org/