On 24 August 2013 06:30, Jesse Gross <je...@nicira.com> wrote:
> Putting a big lock around the majority of the packet processing
> doesn't seem like a particularly good idea for performance and you
> would need to make sure that you get all the entry points. You would
> probably be better off serializing the parts that actually need it,
> like updating stats counters.
Obviously.. I didn't knew earlier what exactly are we looking to protect
here :)
Does something like this makes sense?
diff --git a/datapath/actions.c b/datapath/actions.c
index fa4b904..37ab132 100644
--- a/datapath/actions.c
+++ b/datapath/actions.c
@@ -607,7 +607,12 @@ int ovs_execute_actions(struct datapath *dp,
struct sk_buff *skb)
int error;
/* Check whether we've looped too much. */
+#ifdef CONFIG_PREEMPT_RT_FULL
+ get_cpu_light();
loop = &__get_cpu_var(loop_counters);
+#else
+ loop = &get_cpu_var(loop_counters);
+#endif
if (unlikely(++loop->count > MAX_LOOPS))
loop->looping = true;
if (unlikely(loop->looping)) {
@@ -629,5 +634,10 @@ out_loop:
if (!--loop->count)
loop->looping = false;
+#ifdef CONFIG_PREEMPT_RT_FULL
+ put_cpu_light();
+#else
+ put_cpu_var(loop_counters);
+#endif
return error;
}
diff --git a/datapath/datapath.c b/datapath/datapath.c
index 27deec8..a0fa3de 100644
--- a/datapath/datapath.c
+++ b/datapath/datapath.c
@@ -940,9 +940,7 @@ static int ovs_packet_cmd_execute(struct sk_buff
*skb, struct genl_info *info)
if (!dp)
goto err_unlock;
- local_bh_disable();
err = ovs_execute_actions(dp, packet);
- local_bh_enable();
rcu_read_unlock();
ovs_flow_free(flow, false);
-------------x---------------x----------
Thanks a lot Jesse.
--
Viresh
_______________________________________________
discuss mailing list
discuss@openvswitch.org
http://openvswitch.org/mailman/listinfo/discuss
