On 21.11.2012 09:53, Luigi Rizzo wrote:
On Wed, Nov 21, 2012 at 12:36 AM, Andre Oppermann <an...@freebsd.org> wrote:
On 21.11.2012 09:04, Luigi Rizzo wrote:
The only adjustment i'd suggest to your scheme, if possible, is to add
some control (as it existed in the old polling architecture) to make sure
that userspace is not starved by the ithreads and other kernel tasks
(otherwise you can have livelock, as it happens with NAPI).
I am afraid that simple priorities do not work, you either need some
kind of fair scheduler, or put some hard limit on the cpu fraction used
by the kernel tasks when there are userspace processes around.
That's the quantum stuff I talked about. When the ithread has used up
its CPU quantum it gives up the CPU but schedules itself again at the
same time. Also contrary to its name the ithread does not run in
interrupt context but as a normal kernel thread with elevated priority.
it is the elevated priority that worries me, because it has potential
to preempt userspace and cause livelock.
Again, lacking a proper fair process scheduler, i think the only reliable
way is
to try and track the (approximate) cpu cycles consumed overall by the
various ithreads over, say, the current tick, and once you go above
the threshold drop the priority of those threads just above IDLEPRI.
Then when the next tick arrives you raise the priorities again.
That is what happens. The first time the ithread is scheduled it is
run at high priority. When it yields after consuming its quantum it
drops much lower but above heavily nice'd processes. Have to verify
the exact level though.
Compared to when i did polling in 2000 (when i wanted to support the i486
soekris),
now we can probably assume that a cheap timecounter is available on
all architectures (even ARM and MIPS should have some tsc-like thing, right
?)
and the cycles used vs cycles per tick can be accounted with a relatively
fine grain.
I hope to use as much existing kernel infrastructure as possible and
to keep it simple. I certainly don't want to do cross-core load analysis.
--
Andre
_______________________________________________
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to "freebsd-net-unsubscr...@freebsd.org"