Re: [PATCH RFC 0/2] skiplists for range indexes

[Dave Chinner <dchinner@redhat.com>]

On Sat, May 04, 2013 at 07:11:51AM -0400, Chris Mason wrote:
> Quoting Dave Chinner (2013-05-03 23:25:36)
> > 
> > I've got two cases I care about. The first is the buffer cache
> > indexes which have a 1000:1 read:modify ratio and I'd really like the
> > lookups to be lockless. The other case is the extent tree, where we
> > do lots of inserts when the extent tree is first read, and after
> > than it's typically 2 lookups for every insert/remove. Having one
> > tree that works for both would be handy...
> 
> Ok, we're in a similar boat then.  I'll finish off some of the API and
> test the pure RCU side harder.
> 
> For the extent tree, are you doing a lot of merging once things are in
> the tree?  I'm not planning on doing pure-rcu for items that get merged
> quiet yet.

Yes, we merge extents where ever possible. Almost all contiguous
allocations and unwritten extent conversions merge extents in some
manner...

> Also, I'm using unsigned longs right now.  My guess is we'll both want
> u64s, which means I have to do an i_size_read/write trick in a few
> spots.

Yup, definitely needs to be u64 for XFS...

Cheers,

Dave.
-- 
Dave Chinner
dchinner@redhat.com