Re: [RFC] [PATCH] Avoid livelock for fsync

[Jan Kara <jack@suse.cz>]

On Tue 27-10-09 19:26:14, Nikanth Karthikesan wrote:
> On Monday 26 October 2009 23:43:14 Jan Kara wrote:
> >   Hi,
> > 
> >   on my way back from Kernel Summit, I've coded the attached patch which
> > implements livelock avoidance for write_cache_pages. We tag patches that
> > should be written in the beginning of write_cache_pages and then write
> > only tagged pages (see the patch for details). The patch is based on Nick's
> > idea.
> 
> As I understand, livelock can be caused only by dirtying new pages.
> 
> So theoretically, if a process can dirty pages faster than we can tag pages 
> for writeback, even now isn't there a chance for livelock? But if it is really 
  Yes, theoretically the livelock is still there but practically, I don't
think it's triggerable (the amount of work needed to do either write(2) or
page fault is much higher than just looking up a page in radix tree and
setting there one bit). If the file has lots of dirty pages, I belive user
can create a few more while we are tagging but not much...

> a very fast operation and livelock is not possible, why not hold the tree_lock 
> during the entire period of tagging the pages for writeback i.e., call 
> tag_pages_for_writeback() under mapping->tree_lock? Would it cause 
> deadlock/starvation or some other serious problems?
  I'm dropping tree_lock because I don't think I can hold it during
pagevec_lookup_tag. Even if that was worked-around, if the file has lots of
dirty pages, it could take us long enough to tag all of them that it would
matter latency-wise for other users of the lock. So I'd leave the code as
is.

									Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR