[dpdk-dev] rte_eth_tx_burst() hangs because of nofreedescriptors

yingzhi kaitoy at qq.com
Tue Oct 25 09:10:23 CEST 2016


Hi Helin,


I'm doing comparison tests today, (with fragmented and small traffic).


On the other hand, I checked some example code from DPDK source: 
load_balancer:
app_lcore_io_tx:


			n_pkts = rte_eth_tx_burst(
				port,
				0,
				lp->tx.mbuf_out[port].array,
				(uint16_t) n_mbufs);



			if (unlikely(n_pkts < n_mbufs)) {
				uint32_t k;
				for (k = n_pkts; k < n_mbufs; k ++) {
					struct rte_mbuf *pkt_to_free = lp->tx.mbuf_out[port].array[k];
					rte_pktmbuf_free(pkt_to_free);
				}
			}

And fragmentation example
http://dpdk.org/doc/api-2.2/ip_fragmentation_2main_8c-example.html
        ret = rte_eth_tx_burst(port, queueid, m_table, n);
        if (unlikely(ret < n)) {
                do {
                        rte_pktmbuf_free(m_table[ret]);
                } while (++ret < n);
        }



They works basically the same as in my app, is there any other way to free segmented packets?
The question is:
if pkt_mbuf next pointer is not none, rte_pktmbuf_free() will free all the packet of that linklist, but due to packet re-order, I guess those packets may not sequentially stores in the mbuf array? So how can we correctly free those fragmented packets if  rte_eth_tx_burst() didn't sent out them so we need to manually free?


Thanks



------------------ Original ------------------
From:  "Zhang, Helin";<helin.zhang at intel.com>;
Date:  Tue, Oct 25, 2016 09:05 AM
To:  "yingzhi"<kaitoy at qq.com>; 
Cc:  "dev at dpdk.org"<dev at dpdk.org>; 
Subject:  RE: RE: [dpdk-dev] rte_kni_tx_burst() hangs because of nofreedescriptors



  
 
 
 
    
From: yingzhi [mailto:kaitoy at qq.com] 
 Sent: Monday, October 24, 2016 6:39 PM
 To: Zhang, Helin
 Cc: dev at dpdk.org
 Subject: Re: RE: [dpdk-dev] rte_kni_tx_burst() hangs because of no freedescriptors
 
 
 
 
  
Hi Helin,
 
  
 
 
  
Thanks for your response, to answer your questions:
 
  
1. we send only one packet each time calling rte_kni_tx_burst(), which means the last argument is 1.
 
  
2. it returns 0 because the free mbuf function inside tx_burst will not free any mbuf:
 
  
 
 
   
          if (txq->nb_tx_free < txq->tx_free_thresh)
 
  
                      ixgbe_tx_free_bufs(txq);
 
 
  
 
 
  
after this operation, the txq->nb_tx_free is not increased and keeps to be "0" eventually.
 
  
 
 
  
I did some tests today, I commented out this section of ixgbe_rxtx_vec_common.h -> ixgbe_tx_free_bufs
 
  
 
 
   
          status = txq->tx_ring[txq->tx_next_dd].wb.status;
 
  
          if (!(status & IXGBE_ADVTXD_STAT_DD))
 
  
                      return 0;
 
 
   
 
 
  
After ignoring DD bit check, our app runs for about 6 hours without issue. I suspect there is something wrong in my program set the DD bit somewhere. One of the possible cause currently I suspect is as far as I know, rte_pktmbuf_free(mbuf.array[k])  will free the mbuf of the packet and any fragmented packets following by it. But in our application such as below code snippet:
 
  
 
 
   
        auto nb_tx = rte_eth_tx_burst(port, queue, mbuf.array, (uint16_t) nb_rx);
 
  
        if (unlikely(nb_tx < nb_rx)) {
 
  
            for (unsigned k = nb_tx; k < nb_rx; k++) {
 
  
                rte_pktmbuf_free(mbuf.array[k]);
 
  
            }
 
  
        }
 
[Zhang, Helin] it seems above code piece has memory leak, if the buffer is chained. After all memory leaked, then the issue comes. Please try to check  if this is the root cause!
 
 
  
 
 
  
In this case if there are fragmented packets and failed transmission, may cause a mbuf be freed multiple times.
 
  
 
 
  
Above is just my suspect, need to do some tests later today or tomorrow.
 
  
 
 
  
Thanks
 
  
------------------ Original ------------------
 
   
From:  "Zhang, Helin";<helin.zhang at intel.com>;
 
  
Date:  Mon, Oct 24, 2016 11:33 AM
 
  
To:  "yingzhi"<kaitoy at qq.com>; 
 
  
Cc:  "dev at dpdk.org"<dev at dpdk.org>; 
 
  
Subject:  RE: [dpdk-dev] rte_kni_tx_burst() hangs because of no freedescriptors
 
 
  
 
 
 
Hi Yingzhi
 
 Thank you for the reporting! The description is not so clear at least for me.
 Please help to narrown down the issue by youself.
 How many packets would it have for calling TX function?
 Why it would return 0 after calling TX function? No memory? Or return from else? Have you found anything?
 
 Regards,
 Helin
 
 > -----Original Message-----
 > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of yingzhi
 > Sent: Sunday, October 23, 2016 9:30 PM
 > To: users; dev at dpdk.org
 > Subject: [dpdk-dev] rte_kni_tx_burst() hangs because of no free descriptors
 > 
 > -
 > Hi Experts,
 > 
 > Background:
 > 
 > We are using DPDK to develop a LoadBalancer following below logic: When
 > a new packet is received:
 >  1. if the dst_addr is management IP, forward to KNI. 2. if the dst_addr is in
 > VIP list, select backend and forward(modify dst mac address). 3. otherwise
 > drop the packet.
 > 
 > At this stage, we use one single thread for KNI forwarding and another for
 > VIP forwarding(forward to eth).
 > 
 > DPDK version: 16.07
 >  NIC: 82599ES 10-Gigabit SFI/SFP+ Network Connection
 >  Linux: 14.04.1-Ubuntu x64
 > 
 > Promblem description:
 > 
 > The program runs correctly for sometime(around 2 hours for 400Mb traffic).
 > But it it will hang. When problem happens, rte_eth_tx_burst() will not able to
 > send out any packets(always returns 0). We tracked into that function and
 > noticed it is actually calling ixgbe driver's ixgbe_xmit_pkts_vec() function in
 > our environment, because we use default tx queue configuration, after
 > printing some info, we found if the free function works fine:
 >  tx_rs_thresh: 32, tx_free_thresh: 32, nb_tx_free: 31
 > 
 > it will trigger free and make 32 more free descriptors:
 >  tx_rs_thresh: 32, tx_free_thresh: 32, nb_tx_free: 62
 > 
 > but when something going wrong, it will no longer free anything:
 >  tx_rs_thresh: 32, tx_free_thresh: 32, nb_tx_free: 0 tx_rs_thresh: 32,
 > tx_free_thresh: 32, nb_tx_free: 0
 > 
 > It may related with the DD flag of the descriptor but we are not quite sure.
 > 
 > Our program logic:
 > 
 > create two mbuf pools on socket 0, one for rx_queue and one for kni. (all
 > lcore threads runs on socket0)
 > 
 > init kni interface with rte_kni_alloc()
 > 
 > 
 > init one NIC interface with
 >  rte_eth_dev_configure(); rte_eth_rx_queue_setup();
 > rte_eth_tx_queue_setup(); rte_eth_dev_start();
 > 
 > 
 > 
 > in the eth main loop: (code is simplified)
 >  while(1) {     n = rte_eth_rx_burst(packets);     for (i = 0; i < n; ++i)      {         if
 > (SEND_TO_KNI) {             m = rte_kni_tx_burst(packets[i]);             if (m != 1))
 > {                 rte_pktmbuf_free(packets[i]);             }         }         if (SEND_TO_ETH)
 > {             // after modify the packet             m = rte_eth_tx_burst(packets[i]);
 > if (m != 1)) {                 rte_pktmbuf_free(packets[i]);             }         }         //
 > otherwise drop the packet         rte_pktmbuf_free(packets[i]);     } }
 > 
 > 
 > Please advise if I'm using DPDK in a wrong way. Sorry if I missed something
 > basic, I'm new to DPDK.
 > 
 > Thanks in advance
 >  Best regards


More information about the dev mailing list