|
On Thu, Mar 28, 2013 at 10:57:12AM +0200, Julian Anastasov wrote:
>
> Hello,
>
> On Thu, 28 Mar 2013, Simon Horman wrote:
>
> > On Fri, Mar 22, 2013 at 11:46:35AM +0200, Julian Anastasov wrote:
> > > This is the second patchset with IPVS optimizations.
> > > Now we convert the schedulers, dests and services to RCU.
> > >
> > > All patches are for net-next based on the first
> > > patchset v3. The idea is after discussion and review Simon to
> > > apply the patchset after a week or so to ipvs-next tree.
> > >
> > > The changes in this patchset eliminate global locks
> > > from packet processing by using RCU. There are more details
> > > in the patches.
> > >
> > > After this patchset the situation is as follows:
> > >
> > > - dests:
> > > - lookups under RCU lock allow ip_vs_dest_hold, used
> > > for binding dest to conn or to select dest by scheduler
> > > - dests are freed by dest_trash code long after grace period
> > >
> > > - services:
> > > - no global read_lock
> > > - lookups under RCU lock allow scheduler to select
> > > dests under RCU lock
> > > - grace period implemented with IP_VS_WAIT_WHILE is
> > > gone allowing scheduler's dest selection and scheduler
> > > reconfiguration to run in parallel
> > >
> > > - schedulers:
> > > - schedule method runs under RCU lock, needs _rcu
> > > if using svc->destinations, needs _bh suffix to locks
> > > because it can be called in LOCAL_OUT hook
> > > - when dest is added, the add_dest method is called
> > > instead of update_service
> > > - when dest is deleted, the del_dest method is called
> > > instead of update_service
> > > - when dest is updated, the upd_dest method is called
> > > instead of update_service
> > > - scheduler can hold dests in its state long after they are
> > > unlinked from svc, even without providing del_dest handler.
> > > But such dests must not be returned by the
> > > schedule method (needs IP_VS_DEST_F_AVAILABLE check)
> > > - sched_data must be freed after grace period and
> > > module exit should be delayed with synchronize_rcu
> > > to wait all RCU read-side critical sections to
> > > complete
> > >
> > > - BH:
> > > - we do not disable BHs in LOCAL_OUT and sync code anymore
> > > - _bh suffixes are added to all places that need them,
> > > except timer handlers
> >
> > Hi Julian,
> >
> > what is the status of this series?
> >
> > N.B: THis series is different to "[PATCH 00/15 v3] IPVS optimizations
> > (repost)"
>
> Yes, I posted 2 parts with optimizations. Part 1 is
> at v3 while part 2 is at its first version. Both series
> are ready for applying. BTW, Hans plans more tests in
> the next days, may be we will have some numbers for
> comparison.
Thanks, I'll apply these changes.
--
To unsubscribe from this list: send the line "unsubscribe lvs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
|
