On 2025-04-17 10:01:17 [+0200], Paolo Abeni wrote:
> @Sebastian: I think the 'owner' assignment could be optimized out at
> compile time for non RT build - will likely not matter for performances,
> but I think it will be 'nicer', could you please update the patches to
> do that?
If we don't assign the `owner' then we can't use the lock even on !RT
because lockdep should complain. What about this then:
diff --git a/net/openvswitch/datapath.c b/net/openvswitch/datapath.c
index a3989d450a67f..b8f766978466d 100644
--- a/net/openvswitch/datapath.c
+++ b/net/openvswitch/datapath.c
@@ -294,8 +294,11 @@ void ovs_dp_process_packet(struct sk_buff *skb, struct
sw_flow_key *key)
sf_acts = rcu_dereference(flow->sf_acts);
/* This path can be invoked recursively: Use the current task to
* identify recursive invocation - the lock must be acquired only once.
+ * Even with disabled bottom halves this can be preempted on PREEMPT_RT.
+ * Limit the provecc to RT to avoid assigning `owner' if it can be
+ * avoided.
*/
- if (ovs_pcpu->owner != current) {
+ if (IS_ENABLED(CONFIG_PREEMPT_RT) && ovs_pcpu->owner != current) {
local_lock_nested_bh(&ovs_pcpu_storage.bh_lock);
ovs_pcpu->owner = current;
ovs_pcpu_locked = true;
@@ -687,9 +690,11 @@ static int ovs_packet_cmd_execute(struct sk_buff *skb,
struct genl_info *info)
local_bh_disable();
local_lock_nested_bh(&ovs_pcpu_storage.bh_lock);
- this_cpu_write(ovs_pcpu_storage.owner, current);
+ if (IS_ENABLED(CONFIG_PREEMPT_RT))
+ this_cpu_write(ovs_pcpu_storage.owner, current);
err = ovs_execute_actions(dp, packet, sf_acts, &flow->key);
- this_cpu_write(ovs_pcpu_storage.owner, NULL);
+ if (IS_ENABLED(CONFIG_PREEMPT_RT))
+ this_cpu_write(ovs_pcpu_storage.owner, NULL);
local_unlock_nested_bh(&ovs_pcpu_storage.bh_lock);
local_bh_enable();
rcu_read_unlock();
> Thanks!
>
> Paolo
Sebastian
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
