lkml.org 
[lkml]   [2010]   [Jul]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 8/8] vmscan: Kick flusher threads to clean pages when reclaim is encountering dirty pages
    On Mon, Jul 26, 2010 at 03:28:32PM +0800, Wu Fengguang wrote:
    > On Mon, Jul 19, 2010 at 09:11:30PM +0800, Mel Gorman wrote:
    > > There are a number of cases where pages get cleaned but two of concern
    > > to this patch are;
    > > o When dirtying pages, processes may be throttled to clean pages if
    > > dirty_ratio is not met.
    > > o Pages belonging to inodes dirtied longer than
    > > dirty_writeback_centisecs get cleaned.
    > >
    > > The problem for reclaim is that dirty pages can reach the end of the LRU
    > > if pages are being dirtied slowly so that neither the throttling cleans
    > > them or a flusher thread waking periodically.
    > >
    > > Background flush is already cleaning old or expired inodes first but the
    > > expire time is too far in the future at the time of page reclaim. To mitigate
    > > future problems, this patch wakes flusher threads to clean 1.5 times the
    > > number of dirty pages encountered by reclaimers. The reasoning is that pages
    > > were being dirtied at a roughly constant rate recently so if N dirty pages
    > > were encountered in this scan block, we are likely to see roughly N dirty
    > > pages again soon so try keep the flusher threads ahead of reclaim.
    > >
    > > This is unfortunately very hand-wavy but there is not really a good way of
    > > quantifying how bad it is when reclaim encounters dirty pages other than
    > > "down with that sort of thing". Similarly, there is not an obvious way of
    > > figuring how what percentage of dirty pages are old in terms of LRU-age and
    > > should be cleaned. Ideally, the background flushers would only be cleaning
    > > pages belonging to the zone being scanned but it's not clear if this would
    > > be of benefit (less IO) or not (potentially less efficient IO if an inode
    > > is scattered across multiple zones).
    > >
    > > Signed-off-by: Mel Gorman <mel@csn.ul.ie>
    > > ---
    > > mm/vmscan.c | 18 +++++++++++-------
    > > 1 files changed, 11 insertions(+), 7 deletions(-)
    > >
    > > diff --git a/mm/vmscan.c b/mm/vmscan.c
    > > index bc50937..5763719 100644
    > > --- a/mm/vmscan.c
    > > +++ b/mm/vmscan.c
    > > @@ -806,6 +806,8 @@ restart_dirty:
    > > }
    > >
    > > if (PageDirty(page)) {
    > > + nr_dirty++;
    > > +
    > > /*
    > > * If the caller cannot writeback pages, dirty pages
    > > * are put on a separate list for cleaning by either
    > > @@ -814,7 +816,6 @@ restart_dirty:
    > > if (!reclaim_can_writeback(sc, page)) {
    > > list_add(&page->lru, &dirty_pages);
    > > unlock_page(page);
    > > - nr_dirty++;
    > > goto keep_dirty;
    > > }
    > >
    > > @@ -933,13 +934,16 @@ keep_dirty:
    > > VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
    > > }
    > >
    > > + /*
    > > + * If reclaim is encountering dirty pages, it may be because
    > > + * dirty pages are reaching the end of the LRU even though
    > > + * the dirty_ratio may be satisified. In this case, wake
    > > + * flusher threads to pro-actively clean some pages
    > > + */
    > > + wakeup_flusher_threads(laptop_mode ? 0 : nr_dirty + nr_dirty / 2);
    >
    > Ah it's very possible that nr_dirty==0 here! Then you are hitting the
    > number of dirty pages down to 0 whether or not pageout() is called.
    >

    True, this has been fixed to only wakeup flusher threads when this is
    the file LRU, dirty pages have been encountered and the caller has
    sc->may_writepage.

    > Another minor issue is, the passed (nr_dirty + nr_dirty / 2) is
    > normally a small number, much smaller than MAX_WRITEBACK_PAGES.
    > The flusher will sync at least MAX_WRITEBACK_PAGES pages, this is good
    > for efficiency.
    > And it seems good to let the flusher write much more
    > than nr_dirty pages to safeguard a reasonable large
    > vmscan-head-to-first-dirty-LRU-page margin. So it would be enough to
    > update the comments.
    >

    Ok, the reasoning had been to flush a number of pages that was related
    to the scanning rate but if that is inefficient for the flusher, I'll
    use MAX_WRITEBACK_PAGES.

    Thanks

    --
    Mel Gorman
    Part-time Phd Student Linux Technology Center
    University of Limerick IBM Dublin Software Lab


    \
     
     \ /
      Last update: 2010-07-26 11:29    [W:0.025 / U:0.084 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site