On Fri, Jun 25, 2021 at 11:36:12AM +0800, user wrote:
> Hi Darrell,
> I'm currently doing some tests on OVS-DPDK and trying to optimize it,
> in the process I found your patch(28274f7 conntrack: Fix missed 'conn' lookup
> checks.) made OVS check for the presence of the 'conn' entry before inserting
> it:
>
> @@ -1158,8 +1165,11 @@ process_one(struct conntrack *ct, struct dp_packet
> *pkt,
> ovs_rwlock_unlock(&ct->resources_lock);
>
> ovs_mutex_lock(&ct->ct_lock);
> - conn = conn_not_found(ct, pkt, ctx, commit, now, nat_action_info,
> - helper, alg_exp, ct_alg_ctl);
> + hash = conn_key_hash(&ctx->key, ct->hash_basis);
> + if (!conn_key_lookup(ct, &ctx->key, hash, now, NULL, NULL)) {
> + conn = conn_not_found(ct, pkt, ctx, commit, now, nat_action_info,
> + helper, alg_exp, ct_alg_ctl);
> + }
> ovs_mutex_unlock(&ct->ct_lock);
> }
>
> I'm very confused about the reason we should do this, you mentioned
> that there were some cases that were missed, so fix them, could you please
> explain what scenarios we need to do this？
Without looking for the details, I would guess that there is a race
between one thread creating a given connection entry and another thread
doing it, and that checking for the entry with the lock held fixes the
race.
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
