[dpdk-dev] [PATCH v2 1/3] net/af_xdp: enable zero copy by extbuf

[Ye Xiaolong]

Hi, 

On 06/11, William Tu wrote:
[snip]
>> @@ -294,16 +326,26 @@ eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
>>
>>                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
>>                 mbuf = bufs[i];
>> -
>> -               desc->addr = (uint64_t)addrs[i];
>>                 desc->len = mbuf->pkt_len;
>> -               pkt = xsk_umem__get_data(umem->mz->addr,
>> -                                        desc->addr);
>> -               rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
>> -                          desc->len);
>> -               tx_bytes += mbuf->pkt_len;
>>
>> -               rte_pktmbuf_free(mbuf);
>> +               /*
>> +                * We need to make sure the external mbuf address is within
>> +                * current port's umem memzone range
>> +                */
>> +               if (pmd_zc && RTE_MBUF_HAS_EXTBUF(mbuf) &&
>> +                               in_umem_range(umem, (uint64_t)mbuf->buf_addr)) {
>> +                       desc->addr = (uint64_t)mbuf->buf_addr -
>> +                               umem->mz->addr_64;
>> +                       mbuf->buf_addr = xsk_umem__get_data(umem->mz->addr,
>> +                                       (uint64_t)addrs[i]);
>> +               } else {
>> +                       desc->addr = (uint64_t)addrs[i];
>> +                       pkt = xsk_umem__get_data(umem->mz->addr,
>> +                                       desc->addr);
>> +                       rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
>> +                                       desc->len);
>> +               }
>> +               tx_bytes += mbuf->pkt_len;
>>         }
>>
>>         xsk_ring_prod__submit(&txq->tx, nb_pkts);
>> @@ -313,6 +355,9 @@ eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
>>         txq->stats.tx_pkts += nb_pkts;
>>         txq->stats.tx_bytes += tx_bytes;
>>
>> +       for (i = 0; i < nb_pkts; i++)
>> +               rte_pktmbuf_free(bufs[i]);
>> +
>
>Is it ok to free the mbuf here?
>If the AF_XDP is running pmd_zc=true, the packet mbuf is still in the tx
>ring and might not be sent out yet.

For pmd_zc=ture case, here mbuf->buf_addr has been exchanged to available addr
dequeued from umem->buf_ring, rte_pktmbuf_free would just call the free callback
umem_buf_release_to_fq to enqueue the addr to the buf_ring.

Thanks,
Xiaolong

>
>Regards,
>William
>