DragonFlyBSD threading model
dillon at apollo.backplane.com
Mon Aug 1 08:17:34 PDT 2005
:I have been reading up about how the different BSD systems are
:handling SMP and threading. I gather NetBSD uses scheduler
:activations, where userland programs are given virtual CPUs which then
:get mapped to kernel threads for execution (please correct me if my
:analysis is wrong). FreeBSD appears to use a variant of this idea
:called KSEs. I believe this is known as M:N.
:I am not however sure exactly how DragonFlyBSD threads work - each CPU
:has an LWKT scheduler, and that above this layer there is also a
:userland scheduler (partially equivalent to the functionality provided
:by sched_4bsd.c under FreeBSD?) - but my question is how that relates
:to userland programs - when they have threads..does each program
:thread map directly to one of the LWKT schedulers on one of the CPUs?
:Does that make it 1:1?
:I apologise if this seems like an extremly novide question, this stuff
:is all very new to me.
:Thanks in advance, Alex J Burke.
Basically everything gets mapped onto a LWKT (Light Weight Kernel Thread),
including userland processes.
The LWKT scheduler is a fixed priority scheduler. For example, if you
do a 'ps ax' you will see a ton of kernel threads in addition to userland,
including threads for handling interrupts. Priorities work something
Soft interrupt threads
Kernel threads (pageout demon, network protocol stacks, etc)
User Threads running in the kernel on behalf of userland
User Threads running in userland
LWKT scheduler helper thread (helps with cross-cpu scheduling)
The LWKT scheduler implements minimal preemption for interrupt threads
and will round-robin at the same priority level, but basically the
highest priority runnable thread is the one that gets run.
The userland scheduler is responsible for handling user threads only
while they are running in (or about to be return to) userland. The
userland scheduler will only schedule one thread at a time per cpu on
behalf of userland.
So, for example, if there are 100 user processes all ready to run, the
userland scheduler will only schedule one of those (per cpu) onto the
If there are 100 user processes all ready to run and 10 of them are
in the kernel, those 10 are directly scheduled via LWKT and the
userland scheduler only deals with the remaining 90, choosing one at
a time (per cpu) to schedule via LWKT. The userland scheduler is
preemptive, meaning that it can choose to deschedule one userland
process and reschedule another to LWKT at any time.
We don't have a priority inversion problem for e.g. idle vs normal vs
'realtime' (which isn't real realtime) userland processes because the
userland scheduler does not deal with threads once they have entered
the kernel. That's why there are two priorities associated with
This separation also allows us to theoretically implement multiple
userland schedulers, either exclusively or in tandem, or even locked
to a subset of available cpus. These are features I want to add.
<dillon at xxxxxxxxxxxxx>
More information about the Kernel