On Mon, Oct 08, 2012 at 11:39:18PM -0700, Kent Overstreet wrote:

The refcounting before wasn't very clear; there are two refcounts in
struct kioctx, with an unclear relationship between them (or between
them and ctx->dead).

Now, reqs_active holds a refcount on users (when reqs_active is
nonzero), and the initial refcount is taken on reqs_active - when
ctx->dead goes to 1, we drop that initial refcount.

I agree that it's a mess, but let's rethink this work on top of the
series I'm sending out that gets rid of the retry and cancel code.  It
makes the code a lot easier to follow.  (And Jens also has some patches
to take fewer locks in the submission path, we'll want to take them into
account too.)

- z