Subject: Re: Prototype kernel continuation-passing for NetBSD
To: Jonathan Stone <firstname.lastname@example.org>
From: Stephan Uphoff <email@example.com>
Date: 01/28/2004 20:15:59
> I have some immediate uses for a kcont-like interface; so any feedback
> on this would be appreciated.
I think you need a SA like behavior for your process context continuations.
When a continuation blocks - add a new worker thread.
Otherwise you might add hidden dependencies leading to deadlocks.
( Well you probably have to limit the # of threads .. but it is something to
keep in mind )
You can also add some parallelism by using one thread per CPU.
Adding priorities to process context continuations would also be desirable
- especially once NetBSD gains in kernel preemption.
The framework looks nice - with the exception of kcont_defer_malloc
where I fail to see how you could recover from a failed allocation.
I like it - I am just wondering if some additional infrastructure
would allow code using continuations to be less complex.
OK - here are some initial thoughts about this to start
some discussions - I reserve the right to change my opinions once
I had time to really think about it ;-)
I played with continuation once on a C++ project where strange requirements
forced me to a strange design. (Scaling from system with threads down to
systems requiring shared memory and I/O processes)
Error handling in continuations add a lot of coding complexity
and I am especially worried about memory allocation.
I am wondering if it would be beneficial to offer any continuation
function a way to "call me again when I can allocate some memory".
( Or allocate memory - and then continue me)
While the real benefit of continuation is the ability to do things
parallel and out of order - there are also a lot of instances where
you end up with hierarchical continuations.
( Yes you could flatten everything .. but this becomes unmaintainable real fast)
Having an argument stack associated with the continuation
( MS's IRPs come to mind ) can also eliminate some error checking and is probably
easier on the cache.
( Yes a parent continuation would have to check the stack space for non-checking