On Thu 14-04-11 08:30:45, Wu Fengguang wrote:
> On Thu, Apr 14, 2011 at 05:53:07AM +0800, Jan Kara wrote:
> > On Wed 13-04-11 16:59:39, Wu Fengguang wrote:
> > > When dd in 512bytes, balance_dirty_pages_ratelimited() could be called 8
> > > times for the same page, but obviously the page is only dirtied once.
> > >
> > > Fix it with a (slightly racy) PageDirty() test.
> > >
> > > Signed-off-by: Wu Fengguang <fengguang.wu@xxxxxxxxx>
> > > ---
> > > mm/filemap.c | 5 ++++-
> > > 1 file changed, 4 insertions(+), 1 deletion(-)
> > >
> > > --- linux-next.orig/mm/filemap.c 2011-04-13 16:46:01.000000000 +0800
> > > +++ linux-next/mm/filemap.c 2011-04-13 16:47:26.000000000 +0800
> > > @@ -2313,6 +2313,7 @@ static ssize_t generic_perform_write(str
> > > long status = 0;
> > > ssize_t written = 0;
> > > unsigned int flags = 0;
> > > + unsigned int dirty;
> > >
> > > /*
> > > * Copies from kernel address space cannot fail (NFSD is a big user).
> > > @@ -2361,6 +2362,7 @@ again:
> > > pagefault_enable();
> > > flush_dcache_page(page);
> > >
> > > + dirty = PageDirty(page);
> > This isn't completely right as we sometimes dirty the page in
> > ->write_begin() (see e.g. block_write_begin() when we allocate blocks under
> > an already uptodate page) and in such cases we would not call
> > balance_dirty_pages(). So I'm not sure we can really do this
> > optimization (although it's sad)...
>
> Good catch, thanks! I evaluated three possible options, the last one
> looks most promising (however is a radical change).
>
> - do radix_tree_tag_get() before calling ->write_begin()
> simple but heavy weight
Yes, moreover you cannot really do the check until you have the page
locked for write because otherwise someone could come and write the page
before ->write_begin starts working with it.
> - add balance_dirty_pages_ratelimited() in __block_write_begin()
> seems not easy, too
Yes, you would call balance_dirty_pages_ratelimited() with page lock held
which is not a good thing to do.
> - accurately account the dirtied pages in account_page_dirtied() rather than
> in balance_dirty_pages_ratelimited_nr(). This diff on top of my patchset
> illustrates the idea, but will need to sort out cases like direct IO ...
>
> --- linux-next.orig/mm/page-writeback.c 2011-04-14 07:50:09.000000000 +0800
> +++ linux-next/mm/page-writeback.c 2011-04-14 07:52:35.000000000 +0800
> @@ -1295,8 +1295,6 @@ void balance_dirty_pages_ratelimited_nr(
> if (!bdi_cap_account_dirty(bdi))
> return;
>
> - current->nr_dirtied += nr_pages_dirtied;
> -
> if (dirty_exceeded_recently(bdi, MAX_PAUSE)) {
> unsigned long max = current->nr_dirtied +
> (128 >> (PAGE_SHIFT - 10));
> @@ -1752,6 +1750,7 @@ void account_page_dirtied(struct page *p
> __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
> task_dirty_inc(current);
> task_io_account_write(PAGE_CACHE_SIZE);
> + current->nr_dirtied++;
> }
> }
I see. We could do ratelimit accounting in account_page_dirtied() and
only check limits in balance_dirty_pages(). The only downside of this I can
see is that we would do one-by-one increment instead of a simple addition
when several pages are dirtied (ocfs2, btrfs, and splice interface take
advantage of this). But that should not be a huge issue and it's probably
worth the better ratelimit accounting.
Honza
--
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx 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>