Re: [PATCH 4/5] writeback: avoid livelocking WB_SYNC_ALL writeback
From: Wu Fengguang
Date: Tue Nov 09 2010 - 21:26:33 EST
On Wed, Nov 10, 2010 at 07:18:40AM +0800, Jan Kara wrote:
> On Tue 09-11-10 14:43:46, Andrew Morton wrote:
> > I don't really see how this patch changes anything. For WB_SYNC_ALL
> > requests the code will still try to write out 2^63 pages, only it does
> > it all in a single writeback_inodes_wb() call. What prevents that call
Sorry sync() works on one super block after another, so it's some
__writeback_inodes_sb() call. I'll update the comment.
> > itself from getting livelocked?
__writeback_inodes_sb() livelock is prevented by
- working on a finite set of files by doing queue_io() once at the beginning
- working on a finite set of pages by PAGECACHE_TAG_TOWRITE page tagging
> I'm referring to the livelock avoidance using page tagging. Fengguang
> actually added a note about this into a comment in the code but it's not
> in the changelog. And you're right it should be here.
OK, I'll add the above to changelog.
> > Perhaps the unmentioned problem here is that each call to
> > writeback_inodes_wb(MAX_WRITEBACK_PAGES) will restart its walk across
> > the inode lists. So instead of giving up on a being-written-to-file,
> > we continuously revisit it again and again and again.
> >
> > Correct? If so, please add the description. If incorrect, please add
> > the description as well ;)
> Yes, that's the problem.
writeback_inodes_wb(MAX_WRITEBACK_PAGES) will put the not full written
inode to head of b_more_io, and pick up the next inode from tail of
b_io next time it is called. Here the tail of b_io serves as the
cursor.
b_io b_more_io
|----------------|-----------------|
^head ^cursor ^tail
> > Root cause time: it's those damn per-sb inode lists *again*. They're
> > just awful. We need some data structure there which is more amenable
> > to being iterated over. Something against which we can store cursors,
> > for a start.
> This would be definitely nice. But in this particular case, since we have
> that page tagging livelock avoidance, we can just do all we need in a one
> big sweep so we are OK.
The main problem of list_head is the awkward superblock walks in
move_expired_inodes(). It may take inode_lock for too long time.
It helps to break up b_dirty into a rb-tree. That will make
redirty_tail() more straightforward, too.
> Suggestion for the new changelog:
> When wb_writeback() is called in WB_SYNC_ALL mode, work->nr_to_write is
> usually set to LONG_MAX. The logic in wb_writeback() then calls
> __writeback_inodes_sb() with nr_to_write == MAX_WRITEBACK_PAGES and
> we easily end up with negative nr_to_write after the function returns.
> This is because write_cache_pages() does not stop writing when
> nr_to_write drops to zero in WB_SYNC_ALL mode.
It will return with (nr_to_write <=0) regardless of the
write_cache_pages() trick to ignore nr_to_write. So I changed the
above to:
we easily end up with non-positive nr_to_write after the function
returns, if the inode has more than MAX_WRITEBACK_PAGES dirty pages
at the moment.
Others look good. I'll repost the series with updated changelog.
Thanks,
Fengguang
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/