[dpdk-dev] Question regarding rte_eth_tx_burst()

[Jesus Velazquez <jvelazquez@optenet.com>]

Hi all,

Any ideas would be helpful, I have an application that uses a rte_ring 
for queuing packets to be sent (a single consumer - multiple producer).

The problem I am facing is with rte_eth_tx_burst(), once I dequeue N 
packets from tx ring, there is no guaranty that all packets will be 
sent, if not all packets
are sent, rte_ring does not allow pushing packets at front (it is a 
FIFO).  So, if I re-queue them in the ring, packets will be network 
reordered.

There is no mechanism to know the number of TX free descriptors in a 
particular queue. I'm working with ixgbe and I found that nb_tx_free is 
an approximation of that,
it is updated every time at ixgbe_xmit_cleanup() that is called during 
ixgbe_xmit_pkts().

I was wondering to add a function that returns nb_tx_free value in order 
to have an idea of how many packets can be safety sent.

Right now my code does this:

....
   nbulk = RTE_MIN(rte_ring_count(queue->tx_ring), (uint16_t)TX_DRAIN_PKTS);

   ret = rte_ring_sc_dequeue_bulk(queue->tx_ring, (void **)ptrs, nbulk);
   if(unlikely(ret < 0))
     return 0; /* Not enough objects?, I asked a moment ago */

   ret = rte_eth_tx_burst(queue->port, queue->id, ptrs, nbulk);
   if(unlikely(ret > nbulk)) {
     TRACE(RTE, ">>>Imposible sent more than existing\n");
     return -1;
   }

   while(ret < nbulk) {
     /* bad case: packet reordering */
     ret2 = rte_ring_mp_enqueue(queue->tx_ring, (void *)&ptrs[ret]);
     if(ret2 < 0) {
       /* TODO: may be an emergency tx queue? */
       /* wrost case: dropping */
       rte_pktmbuf_free(ptrs[ret]);
     }
     ret++;
   }

I would appreciate any  ideas of how to overcome this situation.

Thanks and regards,

Jesus