On Thu, Jan 12, 2012 at 05:03:11PM +0900, Minchan Kim wrote:
> On Thu, Jan 12, 2012 at 01:47:02PM +0800, Huang Shijie wrote:
> > In the real tests, there are maybe many times the cc->nr_migratepages is zero,
> > but isolate_migratepages() returns ISOLATE_SUCCESS.
> >
> > Memory in our mx6q board:
> > 2G memory, 8192 pages per page block
> >
> > We use the following command to test in two types system loads:
> > #echo 1 > /proc/sys/vm/compact_memory
> >
> > Test Result:
> > [1] little load(login in the ubuntu):
> > all the scanned pageblocks : 79
> > pageblocks which get no pages : 46
> >
> > The ratio of `get no pages` pageblock is 58.2%.
> >
> > [2] heavy load(start thunderbird, firefox, ..etc):
> > all the scanned pageblocks : 89
> > pageblocks which get no pages : 36
> >
> > The ratio of `get no pages` pageblock is 40.4%.
> >
> > In order to get better performance, we should check the number of the
> > really isolated pages. And do the optimazition for this case.
> >
> > Also fix the confused comments(from Mel Gorman).
> >
> > Tested this patch in MX6Q board.
> >
> > Signed-off-by: Huang Shijie <b32955@xxxxxxxxxxxxx>
> > Acked-by: Mel Gorman <mgorman@xxxxxxx>
> > ---
> > mm/compaction.c | 28 ++++++++++++++++------------
> > 1 files changed, 16 insertions(+), 12 deletions(-)
> >
> > diff --git a/mm/compaction.c b/mm/compaction.c
> > index f4f514d..41d1b72a 100644
> > --- a/mm/compaction.c
> > +++ b/mm/compaction.c
> > @@ -246,8 +246,8 @@ static bool too_many_isolated(struct zone *zone)
> > /* possible outcome of isolate_migratepages */
> > typedef enum {
> > ISOLATE_ABORT, /* Abort compaction now */
> > - ISOLATE_NONE, /* No pages isolated, continue scanning */
> > - ISOLATE_SUCCESS, /* Pages isolated, migrate */
> > + ISOLATE_NONE, /* No pages scanned, consider next pageblock*/
> > + ISOLATE_SUCCESS, /* Pages scanned and maybe isolated, migrate */
> > } isolate_migrate_t;
> >
>
> Hmm, I don't like this change.
> ISOLATE_NONE mean "we don't isolate any page at all"
> ISOLATE_SUCCESS mean "We isolaetssome pages"
> It's very clear but you are changing semantic slighly.
>
That is somewhat the point of his patch - isolate_migratepages()
can return ISOLATE_SUCCESS even though no pages were isolated. Note that
he does not change when ISOLATE_NONE or ISOLATE_SUCCESS gets returned,
he updates the comment to match what the code is actually doing. This
should be visible from the tracepoint. My machine has been up for days
and loaded when I started a process that mapped a large anonymous
region. THP would kick in and I see from the tracepoints excerpts like
this
malloc-13964 [007] 221636.457022: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457022: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457023: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457023: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457024: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457025: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457025: mm_compaction_isolate_migratepages: nr_scanned=1 nr_taken=0
malloc-13964 [007] 221636.457049: mm_compaction_isolate_migratepages: nr_scanned=512 nr_taken=16
malloc-13964 [007] 221636.457102: mm_compaction_isolate_migratepages: nr_scanned=512 nr_taken=16
malloc-13964 [007] 221636.457143: mm_compaction_isolate_migratepages: nr_scanned=512 nr_taken=17
malloc-13964 [007] 221636.457189: mm_compaction_isolate_migratepages: nr_scanned=433 nr_taken=32
malloc-13964 [007] 221636.457253: mm_compaction_isolate_migratepages: nr_scanned=205 nr_taken=32
malloc-13964 [007] 221636.457319: mm_compaction_isolate_migratepages: nr_scanned=389 nr_taken=7
These "nr_scanned=1 nr_taken=0" are during async compaction where the
scanner is skipping over pageblocks that are not MIGRATE_MOVABLE. As the
function only deals in pageblocks, it means the function returns after
only scanning 1 page expecting that compact_zone() will move to the next
block.
> How about this?
>
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -376,7 +376,7 @@ static isolate_migrate_t isolate_migratepages(struct zone *zone,
>
> trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
>
> - return ISOLATE_SUCCESS;
> + return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
> }
>
> /*
> @@ -542,6 +542,8 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
> unsigned long nr_migrate, nr_remaining;
> int err;
>
> + count_vm_event(COMPACTBLOCKS);
> +
> switch (isolate_migratepages(zone, cc)) {
> case ISOLATE_ABORT:
> ret = COMPACT_PARTIAL;
> @@ -559,7 +561,6 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
> update_nr_listpages(cc);
> nr_remaining = cc->nr_migratepages;
>
> - count_vm_event(COMPACTBLOCKS);
> count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
> if (nr_remaining)
> count_vm_events(COMPACTPAGEFAILED, nr_remaining);
>
> This patch's side effect is that it accounts COMPACTBLOCK although isolation is cancel by signal
> but I think it's very rare and doesn't give big effect for statistics of compaciton.
>
This came up during discussion the last time. My opinion was that
COMPACTBLOCK not being updated was a problem. In the existing code
ISOLATE_NONE returning also means the scan did not take place and
this does not need to be accounted for. However, if we scan the block
and isolate no pages, we still want to account for that. A rapidly
increasing COMPACTBLOCKS while COMPACTPAGES changes very little could
indicate that compaction is doing a lot of busy work without making
any useful progress for example.
It could easily be argued that if we skip over !MIGRATE_MOVABLE
pageblocks then we should not account for that in COMPACTBLOCKS either
because the scanning was minimal. In that case we would change this
/*
* For async migration, also only scan in MOVABLE blocks. Async
* migration is optimistic to see if the minimum amount of work
* satisfies the allocation
*/
pageblock_nr = low_pfn >> pageblock_order;
if (!cc->sync && last_pageblock_nr != pageblock_nr &&
get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
low_pfn += pageblock_nr_pages;
low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
last_pageblock_nr = pageblock_nr;
continue;
}
to return ISOLATE_NONE there instead of continue. I would be ok making
that part of this patch to clarify the difference between ISOLATE_NONE
and ISOLATE_SUCCESS and what it means for accounting.
>
> > /*
> > @@ -542,7 +542,7 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
> >
> > while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
> > unsigned long nr_migrate, nr_remaining;
> > - int err;
> > + int err = 0;
> >
> > switch (isolate_migratepages(zone, cc)) {
> > case ISOLATE_ABORT:
> > @@ -554,17 +554,21 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
> > ;
> > }
> >
> > - nr_migrate = cc->nr_migratepages;
> > - err = migrate_pages(&cc->migratepages, compaction_alloc,
> > - (unsigned long)cc, false,
> > - cc->sync);
> > - update_nr_listpages(cc);
> > - nr_remaining = cc->nr_migratepages;
> > + nr_migrate = nr_remaining = cc->nr_migratepages;
> > + if (nr_migrate) {
> > + err = migrate_pages(&cc->migratepages, compaction_alloc,
> > + (unsigned long)cc, false,
> > + cc->sync);
> > + update_nr_listpages(cc);
> > + nr_remaining = cc->nr_migratepages;
> > + count_vm_events(COMPACTPAGES,
> > + nr_migrate - nr_remaining);
> > + if (nr_remaining)
> > + count_vm_events(COMPACTPAGEFAILED,
> > + nr_remaining);
> > + }
> >
> > count_vm_event(COMPACTBLOCKS);
> > - count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
> > - if (nr_remaining)
> > - count_vm_events(COMPACTPAGEFAILED, nr_remaining);
> > trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
> > nr_remaining);
> >
--
Mel Gorman
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>