Re: [PATCH 0/2] radix-tree: fix writeback livelock avoidance code

[Dave Chinner <david@fromorbit.com>]

On Fri, Aug 20, 2010 at 03:51:41PM +0200, Jan Kara wrote:
> On Fri 20-08-10 15:22:05, Dave Chinner wrote:
> > The following two patches fix bugs in the new radix tree functionality used to
> > implement the writeback livelock avoidance code. Both bugs manifest themselves
> > as stray PAGECACHE_TAG_TOWRITE tags in the mapping->page_tree radix tree
> > resulting in livelocks during tag lookups. More subtly, they also appear to
> > result in writeback tree walks occasionally terminating early and so not
> > actually writing all the pages they are supposed to.
>   Really, how that early termination could happen? I'm just wondering
> because I don't see that.. The code just mindlessly copies tags regardless
> of how target flags are set so that's why I'd think that any stale copied
> flags just don't matter...

With the debug I had in pace, I saw a couple of find_get_pages_tag
loops stop (nr_found == 0) rather than livelock when they
encountered a stray tag, which appears to result in writeback not
writing all the pages. I also saw invalidation removing pages from
the page cache that had the PAGECACHE_TAG_TOWRITE tag set, which
indicated that sometimes they weren't getting written back as they
should have been. These were quite rare - they maybe occurred once
for every 1000 livelock occurrences I saw....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com