[ovs-dev] [PATCH v12 0/1] Balance-tcp bond mode optimization

[Matteo Croce]

On Wed, Mar 11, 2020 at 3:26 PM Vishal Deep Ajmera
<vishal.deep.ajmera at ericsson.com> wrote:
>
> v11->v12:
>  Addressed most of comments from Ilya and Eelco.
>  https://mail.openvswitch.org/pipermail/ovs-dev/2020-February/367832.html
>  https://mail.openvswitch.org/pipermail/ovs-dev/2020-February/367842.html
>  Rebased to OVS master.
>

Hi Vishal,

I tested your patch, but it seems there is a memory leak which depletes
all the buffers in a very short time.
I've found that reverting a chunk fixed it, this is the change I made:

--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -7354,15 +7354,17 @@ dp_execute_lb_output_action(struct dp_netdev_pmd_thread *pmd,
                             struct dp_packet_batch *packets_,
                             bool should_steal, uint32_t bond)
 {
-    struct dp_packet *packet;
-    uint32_t i;
-    const uint32_t cnt = dp_packet_batch_size(packets_);
     struct tx_bond *p_bond = tx_bond_lookup(&pmd->tx_bonds, bond);
 
     if (!p_bond) {
         return false;
     }
-    DP_PACKET_BATCH_REFILL_FOR_EACH (i, cnt, packet, packets_) {
+
+    struct dp_packet_batch del_pkts;
+    dp_packet_batch_init(&del_pkts);
+
+    struct dp_packet *packet;
+    DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) {
         /*
          * Lookup the bond-hash table using hash to get the slave.
          */
@@ -7379,11 +7381,16 @@ dp_execute_lb_output_action(struct dp_netdev_pmd_thread *pmd,
             non_atomic_ullong_add(&s_entry->n_packets, 1);
             non_atomic_ullong_add(&s_entry->n_bytes, size);
         } else {
-            dp_packet_batch_refill(packets_, packet, i);
+            dp_packet_batch_add(&del_pkts, packet);
         }
     }
 
-    return dp_packet_batch_is_empty(packets_);
+    /* Delete packets that failed OUTPUT action */
+    COVERAGE_ADD(datapath_drop_invalid_port,
+                 dp_packet_batch_size(&del_pkts));
+    dp_packet_delete_batch(&del_pkts, should_steal);
+
+    return true;
 }
 
 static void

Regards,
-- 
Matteo Croce
per aspera ad upstream