Re: [PATCH 3/3] Provide control over unmapped pages

From: Balbir Singh
Date: Wed Dec 01 2010 - 00:22:46 EST


* Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> [2010-12-01 10:24:21]:

> * Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> [2010-11-30 14:25:09]:
>
> > On Tue, 30 Nov 2010 15:46:31 +0530
> > Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > > Provide control using zone_reclaim() and a boot parameter. The
> > > code reuses functionality from zone_reclaim() to isolate unmapped
> > > pages and reclaim them as a priority, ahead of other mapped pages.
> > >
> > > Signed-off-by: Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx>
> > > ---
> > > include/linux/swap.h | 5 ++-
> > > mm/page_alloc.c | 7 +++--
> > > mm/vmscan.c | 72 +++++++++++++++++++++++++++++++++++++++++++++++++-
> > > 3 files changed, 79 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/include/linux/swap.h b/include/linux/swap.h
> > > index eba53e7..78b0830 100644
> > > --- a/include/linux/swap.h
> > > +++ b/include/linux/swap.h
> > > @@ -252,11 +252,12 @@ extern int vm_swappiness;
> > > extern int remove_mapping(struct address_space *mapping, struct page *page);
> > > extern long vm_total_pages;
> > >
> > > -#ifdef CONFIG_NUMA
> > > -extern int zone_reclaim_mode;
> > > extern int sysctl_min_unmapped_ratio;
> > > extern int sysctl_min_slab_ratio;
> >
> > This change will need to be moved into the first patch.
> >
>
> OK, will do, thanks for pointing it out
>
> > > extern int zone_reclaim(struct zone *, gfp_t, unsigned int);
> > > +extern bool should_balance_unmapped_pages(struct zone *zone);
> > > +#ifdef CONFIG_NUMA
> > > +extern int zone_reclaim_mode;
> > > #else
> > > #define zone_reclaim_mode 0
> > > static inline int zone_reclaim(struct zone *z, gfp_t mask, unsigned int order)
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 62b7280..4228da3 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -1662,6 +1662,9 @@ zonelist_scan:
> > > unsigned long mark;
> > > int ret;
> > >
> > > + if (should_balance_unmapped_pages(zone))
> > > + wakeup_kswapd(zone, order);
> >
> > gack, this is on the page allocator fastpath, isn't it? So
> > 99.99999999% of the world's machines end up doing a pointless call to a
> > pointless function which pointlessly tests a pointless global and
> > pointlessly returns? All because of some whacky KSM thing?
> >
> > The speed and space overhead of this code should be *zero* if
> > !CONFIG_UNMAPPED_PAGECACHE_CONTROL and should be minimal if
> > CONFIG_UNMAPPED_PAGECACHE_CONTROL=y. The way to do the latter is to
> > inline the test of unmapped_page_control into callers and only if it is
> > true (and use unlikely(), please) do we call into the KSM gunk.
> >
>
> Will do, should_balance_unmapped_pages() will be a made a no-op in the
> absence of CONFIG_UNMAPPED_PAGECACHE_CONTROL
>
> > > --- a/mm/vmscan.c
> > > +++ b/mm/vmscan.c
> > > @@ -145,6 +145,21 @@ static DECLARE_RWSEM(shrinker_rwsem);
> > > #define scanning_global_lru(sc) (1)
> > > #endif
> > >
> > > +static unsigned long balance_unmapped_pages(int priority, struct zone *zone,
> > > + struct scan_control *sc);
> > > +static int unmapped_page_control __read_mostly;
> > > +
> > > +static int __init unmapped_page_control_parm(char *str)
> > > +{
> > > + unmapped_page_control = 1;
> > > + /*
> > > + * XXX: Should we tweak swappiness here?
> > > + */
> > > + return 1;
> > > +}
> > > +__setup("unmapped_page_control", unmapped_page_control_parm);
> >
> > aw c'mon guys, everybody knows that when you add a kernel parameter you
> > document it in Documentation/kernel-parameters.txt.
>
> Will do - feeling silly on missing it out, that is where reviews help.
>
> >
> > > static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
> > > struct scan_control *sc)
> > > {
> > > @@ -2223,6 +2238,12 @@ loop_again:
> > > shrink_active_list(SWAP_CLUSTER_MAX, zone,
> > > &sc, priority, 0);
> > >
> > > + /*
> > > + * We do unmapped page balancing once here and once
> > > + * below, so that we don't lose out
> > > + */
> > > + balance_unmapped_pages(priority, zone, &sc);
> > > +
> > > if (!zone_watermark_ok_safe(zone, order,
> > > high_wmark_pages(zone), 0, 0)) {
> > > end_zone = i;
> > > @@ -2258,6 +2279,11 @@ loop_again:
> > > continue;
> > >
> > > sc.nr_scanned = 0;
> > > + /*
> > > + * Balance unmapped pages upfront, this should be
> > > + * really cheap
> > > + */
> > > + balance_unmapped_pages(priority, zone, &sc);
> >
> > More unjustifiable overhead on a commonly-executed codepath.
> >
>
> Will refactor with a CONFIG suggested above.
>
> > > /*
> > > * Call soft limit reclaim before calling shrink_zone.
> > > @@ -2491,7 +2517,8 @@ void wakeup_kswapd(struct zone *zone, int order)
> > > pgdat->kswapd_max_order = order;
> > > if (!waitqueue_active(&pgdat->kswapd_wait))
> > > return;
> > > - if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
> > > + if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0) &&
> > > + !should_balance_unmapped_pages(zone))
> > > return;
> > >
> > > trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
> > > @@ -2740,6 +2767,49 @@ zone_reclaim_unmapped_pages(struct zone *zone, struct scan_control *sc,
> > > }
> > >
> > > /*
> > > + * Routine to balance unmapped pages, inspired from the code under
> > > + * CONFIG_NUMA that does unmapped page and slab page control by keeping
> > > + * min_unmapped_pages in the zone. We currently reclaim just unmapped
> > > + * pages, slab control will come in soon, at which point this routine
> > > + * should be called balance cached pages
> > > + */
> >
> > The problem I have with this comment is that it uses the term "balance"
> > without ever defining it. Plus "balance" is already a term which is used
> > in memory reclaim.
> >
> > So if you can think up a unique noun then that's good but whether or
> > not that is done, please describe with great care what that term
> > actually means in this context.
>
> I used balance as a not a 1:1 balance, but to balance the proportion
> of unmapped page cache based on a sysctl/tunable.
>
> >
> > > +static unsigned long balance_unmapped_pages(int priority, struct zone *zone,
> > > + struct scan_control *sc)
> > > +{
> > > + if (unmapped_page_control &&
> > > + (zone_unmapped_file_pages(zone) > zone->min_unmapped_pages)) {
> > > + struct scan_control nsc;
> > > + unsigned long nr_pages;
> > > +
> > > + nsc = *sc;
> > > +
> > > + nsc.swappiness = 0;
> > > + nsc.may_writepage = 0;
> > > + nsc.may_unmap = 0;
> > > + nsc.nr_reclaimed = 0;
> >
> > Doing a clone-and-own of a scan_control is novel. What's going on here?
>
> This code overwrites the swappiness, may_* and nr_reclaimed for
> correct stats. The idea is to vary the reclaim behaviour/bias it.
>
> >
> > > + nr_pages = zone_unmapped_file_pages(zone) -
> > > + zone->min_unmapped_pages;
> > > + /* Magically try to reclaim eighth the unmapped cache pages */
> > > + nr_pages >>= 3;
> > > +
> > > + zone_reclaim_unmapped_pages(zone, &nsc, nr_pages);
> > > + return nsc.nr_reclaimed;
> > > + }
> > > + return 0;
> > > +}
> > > +
> > > +#define UNMAPPED_PAGE_RATIO 16
> >
> > Well. Giving 16 a name didn't really clarify anything. Attentive
> > readers will want to know what this does, why 16 was chosen and what
> > the effects of changing it will be.
>
> Sorry, I documented that in the changelog of the first patchset. I'll
> document it here as well. The reason for choosing 16 is based on
> heuristics and test, the tradeoff being overenthusiastic reclaim
> versus size of cache/performance.
>
> >
> > > +bool should_balance_unmapped_pages(struct zone *zone)
> > > +{
> > > + if (unmapped_page_control &&
> > > + (zone_unmapped_file_pages(zone) >
> > > + UNMAPPED_PAGE_RATIO * zone->min_unmapped_pages))
> > > + return true;
> > > + return false;
> > > +}
> >
> >
> > > Reviewed-by: Christoph Lameter <cl@xxxxxxxxx>
> >
> > So you're OK with shoving all this flotsam into 100,000,000 cellphones?
> > This was a pretty outrageous patchset!
>
> I'll do a better one, BTW, a lot of embedded folks are interested in
> page cache control outside of cgroup behaviour.
>
> Thanks for the detailed review!
>

My local MTA failed to deliver the message, trying again.

--
Three Cheers,
Balbir
--
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/