Re: [PATCH v6 1/3] workingset: refactor LRU refault to expose refault recency check

From: Brian Foster
Date: Fri Jan 20 2023 - 13:55:54 EST

On Fri, Jan 20, 2023 at 09:29:46AM -0800, Nhat Pham wrote:
> On Fri, Jan 20, 2023 at 6:33 AM Brian Foster <bfoster@xxxxxxxxxx> wrote:
> >
> > On Tue, Jan 17, 2023 at 11:59:57AM -0800, Nhat Pham wrote:
> > > In preparation for computing recently evicted pages in cachestat,
> > > refactor workingset_refault and lru_gen_refault to expose a helper
> > > function that would test if an evicted page is recently evicted.
> > >
> > > Signed-off-by: Nhat Pham <nphamcs@xxxxxxxxx>
> > > ---
> >
> > Hi Nhat,
> >
> > I'm not terribly familiar with the workingset management code, but a few
> > thoughts now that I've stared at it a bit...
> >
> > > include/linux/swap.h | 1 +
> > > mm/workingset.c | 129 ++++++++++++++++++++++++++++++-------------
> > > 2 files changed, 92 insertions(+), 38 deletions(-)
> > >
> > > diff --git a/include/linux/swap.h b/include/linux/swap.h
> > > index a18cf4b7c724..dae6f6f955eb 100644
> > > --- a/include/linux/swap.h
> > > +++ b/include/linux/swap.h
> > > @@ -361,6 +361,7 @@ static inline void folio_set_swap_entry(struct folio *folio, swp_entry_t entry)
> > > }
> > >
> > > /* linux/mm/workingset.c */
> > > +bool workingset_test_recent(void *shadow, bool file, bool *workingset);
> > > void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages);
> > > void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg);
> > > void workingset_refault(struct folio *folio, void *shadow);
> > > diff --git a/mm/workingset.c b/mm/workingset.c
> > > index 79585d55c45d..006482c4e0bd 100644
> > > --- a/mm/workingset.c
> > > +++ b/mm/workingset.c
> > > @@ -244,6 +244,33 @@ static void *lru_gen_eviction(struct folio *folio)
> > > return pack_shadow(mem_cgroup_id(memcg), pgdat, token, refs);
> > > }
> > >
> > > +/*
> > > + * Test if the folio is recently evicted.
> > > + *
> > > + * As a side effect, also populates the references with
> > > + * values unpacked from the shadow of the evicted folio.
> > > + */
> > > +static bool lru_gen_test_recent(void *shadow, bool file, bool *workingset)
> > > +{
> > > + struct mem_cgroup *eviction_memcg;
> > > + struct lruvec *lruvec;
> > > + struct lru_gen_struct *lrugen;
> > > + unsigned long min_seq;
> > > +
> >
> > Extra whitespace looks a bit funny here.
> >
> > > + int memcgid;
> > > + struct pglist_data *pgdat;
> > > + unsigned long token;
> > > +
> > > + unpack_shadow(shadow, &memcgid, &pgdat, &token, workingset);
> > > + eviction_memcg = mem_cgroup_from_id(memcgid);
> > > +
> > > + lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
> > > + lrugen = &lruvec->lrugen;
> > > +
> > > + min_seq = READ_ONCE(lrugen->min_seq[file]);
> > > + return !((token >> LRU_REFS_WIDTH) != (min_seq & (EVICTION_MASK >> LRU_REFS_WIDTH)));
> >
> > I think this might be more readable without the double negative.
>
> Hmm indeed. I was just making sure that I did not mess up Yu's
> original logic here (by just wrapping it in a parentheses and
> negate the whole thing), but if I understand it correctly it's just
> an equality check. I'll fix it in the next version to make it cleaner.
>
> >
> > Also it looks like this logic is pulled from lru_gen_refault(). Any
> > reason the caller isn't refactored to use this helper, similar to how
> > workingset_refault() is modified? It seems like a potential landmine to
> > duplicate the logic here for cachestat purposes and somewhere else for
> > actual workingset management.
>
> In V2, it is actually refactored analogously as well - but we had a discussion
> about it here:
>
> https://lkml.org/lkml/2022/12/5/1321
>

Yeah, sorry.. replied to Johannes.

> >
> > > +}
> > > +
> > > static void lru_gen_refault(struct folio *folio, void *shadow)
> > > {
> > > int hist, tier, refs;
> > > @@ -306,6 +333,11 @@ static void *lru_gen_eviction(struct folio *folio)
> > > return NULL;
> > > }
> > >
> > > +static bool lru_gen_test_recent(void *shadow, bool file, bool *workingset)
> > > +{
> > > + return true;
> > > +}
> >
> > I guess this is a no-op for !MGLRU but given the context (i.e. special
> > treatment for "recent" refaults), perhaps false is a more sane default?
>
> Hmm, fair point. Let me fix that in the next version.
>
> >
> > > +
> > > static void lru_gen_refault(struct folio *folio, void *shadow)
> > > {
> > > }
> > > @@ -373,40 +405,31 @@ void *workingset_eviction(struct folio *folio, struct mem_cgroup *target_memcg)
> > > folio_test_workingset(folio));
> > > }
> > >
> > > -/**
> > > - * workingset_refault - Evaluate the refault of a previously evicted folio.
> > > - * @folio: The freshly allocated replacement folio.
> > > - * @shadow: Shadow entry of the evicted folio.
> > > +/*
> > > + * Test if the folio is recently evicted by checking if
> > > + * refault distance of shadow exceeds workingset size.
> > > *
> > > - * Calculates and evaluates the refault distance of the previously
> > > - * evicted folio in the context of the node and the memcg whose memory
> > > - * pressure caused the eviction.
> > > + * As a side effect, populate workingset with the value
> > > + * unpacked from shadow.
> > > */
> > > -void workingset_refault(struct folio *folio, void *shadow)
> > > +bool workingset_test_recent(void *shadow, bool file, bool *workingset)
> > > {
> > > - bool file = folio_is_file_lru(folio);
> > > struct mem_cgroup *eviction_memcg;
> > > struct lruvec *eviction_lruvec;
> > > unsigned long refault_distance;
> > > unsigned long workingset_size;
> > > - struct pglist_data *pgdat;
> > > - struct mem_cgroup *memcg;
> > > - unsigned long eviction;
> > > - struct lruvec *lruvec;
> > > unsigned long refault;
> > > - bool workingset;
> > > +
> > > int memcgid;
> > > - long nr;
> > > + struct pglist_data *pgdat;
> > > + unsigned long eviction;
> > >
> > > - if (lru_gen_enabled()) {
> > > - lru_gen_refault(folio, shadow);
> > > - return;
> > > - }
> > > + if (lru_gen_enabled())
> > > + return lru_gen_test_recent(shadow, file, workingset);
> >
> > Hmm.. so this function is only called by workingset_refault() when
> > lru_gen_enabled() == false, otherwise it calls into lru_gen_refault(),
> > which as noted above duplicates some of the recency logic.
> >
> > I'm assuming this lru_gen_test_recent() call is so filemap_cachestat()
> > can just call workingset_test_recent(). That seems reasonable, but makes
> > me wonder...
>
> You're right. It's a bit clunky...
>
> >
> > >
> > > - unpack_shadow(shadow, &memcgid, &pgdat, &eviction, &workingset);
> > > + unpack_shadow(shadow, &memcgid, &pgdat, &eviction, workingset);
> > > eviction <<= bucket_order;
> > >
> > > - rcu_read_lock();
> > > /*
> > > * Look up the memcg associated with the stored ID. It might
> > > * have been deleted since the folio's eviction.
> > > @@ -425,7 +448,8 @@ void workingset_refault(struct folio *folio, void *shadow)
> > > */
> > > eviction_memcg = mem_cgroup_from_id(memcgid);
> > > if (!mem_cgroup_disabled() && !eviction_memcg)
> > > - goto out;
> > > + return false;
> > > +
> > > eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
> > > refault = atomic_long_read(&eviction_lruvec->nonresident_age);
> > >
> > > @@ -447,21 +471,6 @@ void workingset_refault(struct folio *folio, void *shadow)
> > > */
> > > refault_distance = (refault - eviction) & EVICTION_MASK;
> > >
> > > - /*
> > > - * The activation decision for this folio is made at the level
> > > - * where the eviction occurred, as that is where the LRU order
> > > - * during folio reclaim is being determined.
> > > - *
> > > - * However, the cgroup that will own the folio is the one that
> > > - * is actually experiencing the refault event.
> > > - */
> > > - nr = folio_nr_pages(folio);
> > > - memcg = folio_memcg(folio);
> > > - pgdat = folio_pgdat(folio);
> > > - lruvec = mem_cgroup_lruvec(memcg, pgdat);
> > > -
> > > - mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
> > > -
> > > mem_cgroup_flush_stats_delayed();
> > > /*
> > > * Compare the distance to the existing workingset size. We
> > > @@ -483,8 +492,51 @@ void workingset_refault(struct folio *folio, void *shadow)
> > > NR_INACTIVE_ANON);
> > > }
> > > }
> > > - if (refault_distance > workingset_size)
> > > +
> > > + return refault_distance <= workingset_size;
> > > +}
> > > +
> > > +/**
> > > + * workingset_refault - Evaluate the refault of a previously evicted folio.
> > > + * @folio: The freshly allocated replacement folio.
> > > + * @shadow: Shadow entry of the evicted folio.
> > > + *
> > > + * Calculates and evaluates the refault distance of the previously
> > > + * evicted folio in the context of the node and the memcg whose memory
> > > + * pressure caused the eviction.
> > > + */
> > > +void workingset_refault(struct folio *folio, void *shadow)
> > > +{
> > > + bool file = folio_is_file_lru(folio);
> > > + struct pglist_data *pgdat;
> > > + struct mem_cgroup *memcg;
> > > + struct lruvec *lruvec;
> > > + bool workingset;
> > > + long nr;
> > > +
> > > + if (lru_gen_enabled()) {
> > > + lru_gen_refault(folio, shadow);
> > > + return;
> > > + }
> >
> > ... if perhaps this should call workingset_test_recent() a bit earlier,
> > since it also covers the lru_gen_*() case..? That may or may not be
> > cleaner. It _seems like_ it might produce a bit more consistent logic,
> > but just a thought and I could easily be missing details.
>
> Hmm you mean before/in place of the lru_gen_refault call?
> workingset_test_recent only covers lru_gen_test_recent,
> not the rest of the logic of lru_gen_refault I believe.
>

When reading through the code I got the impression that if
workingset_test_recent() -> lru_gen_test_recent() returned false, then
lru_gen_refault() didn't really do anything. I.e., the token/min_seq
check fails and it skips out to the end of the function. That had me
wondering whether workingset_refault() could just skip out of either
mode if workingset_test_recent() returns false (since it calls
lru_gen_test_recent()), but that may not work. Specifically I'm not sure
about that mod_lruvec_state()) call (discussed below) for the MGLRU
case. If that call is correct as is for !MGLRU, maybe it could be made
conditional based on lru_gen_enabled()..?

I guess what I was trying to get at here is that you've created this
nice workingset_test_recent() helper to cover both modes for
filemap_cachestat(), so it would be nice if it could be used in that way
in the core workingset code as well. :)

I suppose yet another option could be to skip the lru_gen_test_recent()
check in workingset_test_recent() and instead call it from
lru_gen_refault(), but then I guess you'd have to open code both checks
in the filemap_cachestat() call, which sounds kind of ugly..

>
> >
> > > +
> > > + rcu_read_lock();
> > > +
> > > + nr = folio_nr_pages(folio);
> > > + memcg = folio_memcg(folio);
> > > + pgdat = folio_pgdat(folio);
> > > + lruvec = mem_cgroup_lruvec(memcg, pgdat);
> > > +
> > > + if (!workingset_test_recent(shadow, file, &workingset)) {
> > > + /*
> > > + * The activation decision for this folio is made at the level
> > > + * where the eviction occurred, as that is where the LRU order
> > > + * during folio reclaim is being determined.
> > > + *
> > > + * However, the cgroup that will own the folio is the one that
> > > + * is actually experiencing the refault event.
> > > + */
> >
> > IIUC, this comment is explaining the difference between using the
> > eviction lru (based on the shadow entry) to calculate recency vs. the
> > lru for the current folio to process the refault. If so, perhaps it
> > should go right above the workingset_test_recent() call? (Then the if
> > braces could go away as well..).
>
> You're right! I think it should go above `nr = folio_nr_pages(folio);` call.
>

Sounds good.

> >
> > > goto out;
> > > + }
> > >
> > > folio_set_active(folio);
> > > workingset_age_nonresident(lruvec, nr);
> > > @@ -498,6 +550,7 @@ void workingset_refault(struct folio *folio, void *shadow)
> > > mod_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file, nr);
> > > }
> > > out:
> > > + mod_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file, nr);
> >
> > Why not just leave this up earlier in the function (i.e. before the
> > recency check) as it was originally?
>
> Let me double check, but I think this is a relic from the old (and incorrect)
> version of workingset code.
>
> Originally, mod_lruvec_state uses the lruvec computed from a variable
> (pgdat) that was unpacked from the shadow. So this mod_lruvec_state
> has to go after the unpack_shadow call (which has been moved inside
> of workingset_test_recent).
>
> This is actually wrong - we actually want the pgdat from the folio. It
> has been fixed in a separate patch:
>
> https://lore.kernel.org/all/20230104222944.2380117-1-nphamcs@xxxxxxxxx/T/#u
>

Yep, I had applied this series on top of that one..

> But I didn't update it here. Let me stare at it a bit more to make sure,
> and then fix it in the next version. It should not change the behavior,
> but it should be cleaner.
>

Sounds good. FWIW it looked like the logic hadn't changed with this
series so I just assumed it was correct, just possibly moved around
unnecessarily. I'll try to make more sense of it in the next version.
Thanks.

Brian

> >
> > Brian
> >
> > > rcu_read_unlock();
> > > }
> > >
> > > --
> > > 2.30.2
> > >
> >
>