Re: [PATCH v2] fadvise: move active pages to inactive list with POSIX_FADV_DONTNEED

[Andrea Righi <andrea@betterlinux.com>]

On Fri, Jun 24, 2011 at 01:57:54PM +0800, Shaohua Li wrote:
> On Thu, 2011-06-23 at 14:36 +0800, Andrea Righi wrote:
> > There were some reported problems in the past about trashing page cache
> > when a backup software (i.e., rsync) touches a huge amount of pages (see
> > for example [1]).
> > 
> > This problem has been almost fixed by the Minchan Kim's patch [2] and a
> > proper use of fadvise() in the backup software. For example this patch
> > set [3] has been proposed for inclusion in rsync.
> > 
> > However, there can be still other similar trashing problems: when the
> > backup software reads all the source files, some of them may be part of
> > the actual working set of the system. When a
> > posix_fadvise(POSIX_FADV_DONTNEED) is performed _all_ pages are evicted
> > from pagecache, both the working set and the use-once pages touched only
> > by the backup software.
> > 
> > With the following solution when posix_fadvise(POSIX_FADV_DONTNEED) is
> > called for an active page instead of removing it from the page cache it
> > is added to the tail of the inactive list. Otherwise, if it's already in
> > the inactive list the page is removed from the page cache.
> > 
> > In this way if the backup was the only user of a page, that page will
> > be immediately removed from the page cache by calling
> > posix_fadvise(POSIX_FADV_DONTNEED). If the page was also touched by
> > other processes it'll be moved to the inactive list, having another
> > chance of being re-added to the working set, or simply reclaimed when
> > memory is needed.
> > 
> > Testcase:
> > 
> >   - create a 1GB file called "zero"
> >   - run md5sum zero to read all the pages in page cache (this is to
> >     simulate the user activity on this file)
> >   - run "rsync zero zero_copy" (rsync is patched with [3])
> >   - re-run md5sum zero (user activity on the working set) and measure
> >     the time to complete this command
> > 
> > The test has been performed using 3.0.0-rc4 vanilla and with this patch
> > applied (3.0.0-rc4-fadvise).
> > 
> > Results:
> >                   avg elapsed time      block:block_bio_queue
> >  3.0.0-rc4                  4.127s                      8,214
> >  3.0.0-rc4-fadvise          2.146s                          0
> > 
> > In the first case the file is evicted from page cache completely and we
> > must re-read it from the disk. In the second case the file is still in
> > page cache (in the inactive list) and we don't need any other additional
> > I/O operation.
> > 
> > [1] http://marc.info/?l=rsync&m=128885034930933&w=2
> > [2] https://lkml.org/lkml/2011/2/20/57
> > [3] http://lists.samba.org/archive/rsync/2010-November/025827.html
> > 
> > ChangeLog v1 -> v2:
> >  - fix comment in invalidate_mapping_pages()
> > 
> > Acked-by: Rik van Riel <riel@redhat.com>
> > Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
> > Signed-off-by: Andrea Righi <andrea@betterlinux.com>
> > ---
> >  mm/swap.c     |    9 +++++----
> >  mm/truncate.c |   10 +++++++---
> >  2 files changed, 12 insertions(+), 7 deletions(-)
> > 
> > diff --git a/mm/swap.c b/mm/swap.c
> > index 3a442f1..fc8bb76 100644
> > --- a/mm/swap.c
> > +++ b/mm/swap.c
> > @@ -411,10 +411,11 @@ void add_page_to_unevictable_list(struct page *page)
> >   *
> >   * 1. active, mapped page -> none
> >   * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
> > - * 3. inactive, mapped page -> none
> > - * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
> > - * 5. inactive, clean -> inactive, tail
> > - * 6. Others -> none
> > + * 3. active, clean -> inactive, tail
> > + * 4. inactive, mapped page -> none
> > + * 5. inactive, dirty/writeback page -> inactive, head, PG_reclaim
> > + * 6. inactive, clean -> inactive, tail
> > + * 7. Others -> none
> >   *
> >   * In 4, why it moves inactive's head, the VM expects the page would
> >   * be write it out by flusher threads as this is much more effective
> > diff --git a/mm/truncate.c b/mm/truncate.c
> > index 3a29a61..a36af48 100644
> > --- a/mm/truncate.c
> > +++ b/mm/truncate.c
> > @@ -357,11 +357,15 @@ unsigned long invalidate_mapping_pages(struct address_space *mapping,
> >  			if (lock_failed)
> >  				continue;
> >  
> > -			ret = invalidate_inode_page(page);
> > +			if (PageActive(page))
> > +				ret = 0;
> > +			else
> > +				ret = invalidate_inode_page(page);
> >  			unlock_page(page);
> >  			/*
> > -			 * Invalidation is a hint that the page is no longer
> > -			 * of interest and try to speed up its reclaim.
> > +			 * Invalidation of an inactive page is a hint that the
> > +			 * page is no longer of interest and try to speed up
> > +			 * its reclaim.
> >  			 */
> >  			if (!ret)
> >  				deactivate_page(page);
> this looks changed behavior, active page will not be invalidated.
> invalidate_mapping_pages is not just used by fadvise. for
> example, /proc/sys/vm/drop_cache can't drop active pages any more with
> the patch in the first invoke. Please audit other use cases too.

Yes, changing the invalidate_mapping_pages() behavior completely is not
good. With drop_cache we may want to actually drop the pages.

I was considering to implement the Pádraig's suggestion, so do not
change the current invalidate_mapping_pages() behavior and use
POSIX_FADV_NOREUSE to implement the "reduce cache eligibility" logic.

I'll post a new patch soon.

Thanks,
-Andrea