Re: [PATCH 3/7 v2] mm: rework __isolate_lru_page() file/anon filter

[Konstantin Khlebnikov]

Hugh Dickins wrote:
> On Fri, 9 Mar 2012, Konstantin Khlebnikov wrote:
> > Hugh Dickins wrote:
> > > I like very much the look of what he's come up with, but I'm still
> > > puzzling over why it barely makes any improvement to __isolate_lru_page():
> > > seems significantly inferior (in code size terms) to his original (which
> > > I imagine Glauber's compromise would be equivalent to).
> > >
> > > At some point I ought to give up on niggling about this,
> > > but I haven't quite got there yet.
> >
> > (with if())
> > $ ./scripts/bloat-o-meter built-in.o built-in.o-v1
> > add/remove: 0/0 grow/shrink: 2/1 up/down: 32/-20 (12)
> > function                                     old     new   delta
> > static.shrink_active_list                    837     853     +16
> > shrink_inactive_list                        1259    1275     +16
> > static.isolate_lru_pages                    1055    1035     -20
> >
> > (with switch())
> > $ ./scripts/bloat-o-meter built-in.o built-in.o-v2
> > add/remove: 0/0 grow/shrink: 4/2 up/down: 111/-23 (88)
> > function                                     old     new   delta
> > __isolate_lru_page                           301     377     +76
> > static.shrink_active_list                    837     853     +16
> > shrink_inactive_list                        1259    1275     +16
> > page_evictable                               170     173      +3
> > __remove_mapping                             322     319      -3
> > static.isolate_lru_pages                    1055    1035     -20
> >
> > (without __always_inline on page_lru())
> > $ ./scripts/bloat-o-meter built-in.o built-in.o-v5-noinline
> > add/remove: 0/0 grow/shrink: 5/2 up/down: 93/-23 (70)
> > function                                     old     new   delta
> > __isolate_lru_page                           301     333     +32
> > isolate_lru_page                             359     385     +26
> > static.shrink_active_list                    837     853     +16
> > putback_inactive_pages                       635     651     +16
> > page_evictable                               170     173      +3
> > __remove_mapping                             322     319      -3
> > static.isolate_lru_pages                    1055    1035     -20
> >
> > $ ./scripts/bloat-o-meter built-in.o built-in.o-v5
> > add/remove: 0/0 grow/shrink: 3/4 up/down: 35/-67 (-32)
> > function                                     old     new   delta
> > static.shrink_active_list                    837     853     +16
> > __isolate_lru_page                           301     317     +16
> > page_evictable                               170     173      +3
> > __remove_mapping                             322     319      -3
> > mem_cgroup_lru_del                            73      65      -8
> > static.isolate_lru_pages                    1055    1035     -20
> > __mem_cgroup_commit_charge                   676     640     -36
> >
> > Actually __isolate_lru_page() even little bit bigger
>
> I was coming to realize that it must be your page_lru()ing:
> although it's dressed up in one line, there's several branches there.

Yes, but I think we can optimize page_lru(): we can prepare ready-to-use
page lru index in lower bits of page->flags, if we swap page flags and split
LRU_UNEVICTABLE into FILE/ANON parts.

> I think you'll find you have a clear winner at last, if you just pass
> lru on down as third arg to __isolate_lru_page(), where file used to
> be passed, instead of re-evaluating it inside.
>
> shrink callers already have the lru, and compaction works it out
> immediately afterwards.

No, for non-lumpy isolation we don't need this check at all,
because all pages already picked from right lru list.

I'll send separate patch for this (on top v5 patchset), after meditation =)

> Though we do need to be careful: the lumpy case would then have to
> pass page_lru(cursor_page).  Oh, actually no (though it would deserve
> a comment): since the lumpy case selects LRU_ALL_EVICTABLE, it's
> irrelevant what it passes for lru, so might as well stick with
> the one passed down.  Though you may decide I'm being too tricky
> there, and prefer to calculate page_lru(cursor_page) anyway, it
> not being the hottest path.
>
> Whether you'd still want page_lru(page) __always_inline, I don't know.
>
> Hugh

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/