On Mon, 21 May 2012, Rusty Russell wrote:

On Fri, 20 Apr 2012 16:18:04 +0200 (CEST), Thomas Gleixner [off-list ref] wrote:

quoted

On Fri, 20 Apr 2012, Srivatsa S. Bhat wrote:
The whole notifier business needs a redesign as well, because we don't
have a way to express proper dependencies, we add random notifier
points and the teardown path is ass backwards. The whole thing wants
to be a tree which can be walked in either direction and from any
point. Right now we cut the trunk first and keep the single limb up
with a helicopter and start dismantling it.

But there are two ways to do it.  One is to eliminate the need for
callbacks.  The other is to make a full dependency-based callback

What do you mean with fully eliminating the need for callbacks. Do you
want to put the necessary bringup/shutdown function calls just in the
core code so we get rid of the notifiers or do you have something
different in mind ?

system.

Let's try the first before the second.  I implemented a full dep-based
system for initcalls once, and it was not as nice as I'd hoped.

Yeah, a full dependency thing might be overkill.

quoted

Flat notifiers are not working for this as they do not allow a tree
structure and prevent us to do things in parallel.

Not sure whether calling notifiers in parallel is going to be a big win:
they'll end up fighting over the cpu we're taking down.  But I could be
wrong.

I'm not going to aim for parallel in the first place. That was just an
idea and if we chose the right implementation then parallelism can be
added later.
 

quoted

That really needs to be completely reworked. There is also a lot of
stuff which wants to be moved into the starting/dying CPU
context. Right now we kinda do that by trampling on the CPU with a
high prio stomper thread, but that's really just a bandaid and steady
cause of trouble.

If you look at facilities which use kthreads, then there is lots other
setup which does not need a notifier at all, as it can be done in the
context of the thread when we have a way to start/park those threads
at the right time in the up/down process.

That's a very nice idea.  Should be simpler for kthreads park/unpark
themselves, rather than having some notifier to kill them.

Well, something needs to tell them to park/unpark, but what should
move into the thread is the setup/teardown of context.
 

The original concept of stopping the machine for cpu hotplug and trying
not to effect any other kernel code has jumped the shark: I think we
need to look seriously at a complete rewrite where we don't use
stop_machine.

Yep. Working on it. :)

Thanks,

	tglx