Re: [OKS] Module removal

[Werner Almesberger <wa@almesberger.net>]

Keith Owens wrote:
> For netfilter, the use count reflects the number of packets being
> processed.  Complex and potentially high overhead.

Good example - netfilter may access a huge number of tiny
modules when working on a packet. While this by itself is a
performance issue, the need to keep reference counts right
doesn't necessarily help.

> All of this requires that the module information be passed in multiple
> structures and assumes that all code is careful about reference
> counting the code it is about to execute.

It's not really just the module information. If I can, say, get
callbacks from something even after I unregister, I may well
have destroyed the data I need to process the callbacks, and
oops or worse.

> There has to be a better way!

Well yes, there are other approaches that make sure something is
not used, e.g.

If you can a) make sure no new references are generated, and b)
send a sequential marker through the subsystem, you can be sure
that all references are gone, when the marker re-emerges. (Or use
multiple markers, e.g. one per CPU.)

Likewise, if you can disable restarting of a subsystem, and then
wait until the subsystem is idle, you're safe. E.g. tasklet_disable
works like this.

- Werner

-- 
  _________________________________________________________________________
 / Werner Almesberger, Buenos Aires, Argentina         wa@almesberger.net /
/_http://icapeople.epfl.ch/almesber/_____________________________________/