On Wed 04-01-17 13:52:24, Vlastimil Babka wrote:
> On 01/04/2017 11:19 AM, Michal Hocko wrote:
> > From: Michal Hocko <mho...@suse.com>
> >
> > Our reclaim process has several tracepoints to tell us more about how
> > things are progressing. We are, however, missing a tracepoint to track
> > active list aging. Introduce mm_vmscan_lru_shrink_active which reports
> > the number of
> > - nr_scanned, nr_taken pages to tell us the LRU isolation
> > effectiveness.
>
> Well, this point is no longer true, is it...
ups, leftover
- nr_take - the number of isolated pages
> > - nr_referenced pages which tells us that we are hitting referenced
> > pages which are deactivated. If this is a large part of the
> > reported nr_deactivated pages then we might be hitting into
> > the active list too early because they might be still part of
> > the working set. This might help to debug performance issues.
> > - nr_activated pages which tells us how many pages are kept on the
>
> "nr_activated" is slightly misleading? They remain active, they are not
> being activated (that's why the pgactivate vmstat is also not increased
> on them, right?). I guess rename to "nr_active" ? Or something like
> "nr_remain_active" although that's longer.
will go with nr_active
>
> [...]
>
> > @@ -1857,6 +1859,7 @@ static void move_active_pages_to_lru(struct lruvec
> > *lruvec,
> > unsigned long pgmoved = 0;
> > struct page *page;
> > int nr_pages;
> > + int nr_moved = 0;
> >
> > while (!list_empty(list)) {
> > page = lru_to_page(list);
> > @@ -1882,11 +1885,15 @@ static void move_active_pages_to_lru(struct lruvec
> > *lruvec,
> > spin_lock_irq(&pgdat->lru_lock);
> > } else
> > list_add(&page->lru, pages_to_free);
> > + } else {
> > + nr_moved += nr_pages;
> > }
> > }
> >
> > if (!is_active_lru(lru))
> > __count_vm_events(PGDEACTIVATE, pgmoved);
>
> So we now have pgmoved and nr_moved. One is used for vmstat, other for
> tracepoint, and the only difference is that vmstat includes pages where
> we raced with page being unmapped from all pte's (IIUC?) and thus
> removed from lru, which should be rather rare? I guess those are being
> counted into vmstat only due to how the code evolved from using pagevec.
> If we don't consider them in the tracepoint, then I'd suggest we don't
> count them into vmstat either, and simplify this.
OK, but I would prefer to have this in a separate patch, OK?
--
Michal Hocko
SUSE Labs
