[ovs-dev] [PATCH v8 11/13] netdev-dpdk: copy large packet to multi-seg. mbufs

[Lam, Tiago]

On 18/06/2018 14:15, Eelco Chaudron wrote:
> 
> 
> On 11 Jun 2018, at 18:21, Tiago Lam wrote:
> 
>> From: Mark Kavanagh <mark.b.kavanagh at intel.com>
>>
>> Currently, packets are only copied to a single segment in the function
>> dpdk_do_tx_copy(). This could be an issue in the case of jumbo frames,
>> particularly when multi-segment mbufs are involved.
>>
>> This patch calculates the number of segments needed by a packet and
>> copies the data to each segment.
>>
>> A new function, dpdk_buf_alloc(), has also been introduced as a 
>> wrapper
>> around the nonpmd_mp_mutex to serialise allocations from a non-pmd
>> context.
>>
>> Co-authored-by: Michael Qiu <qiudayu at chinac.com>
>> Co-authored-by: Tiago Lam <tiago.lam at intel.com>
>>
>> Signed-off-by: Mark Kavanagh <mark.b.kavanagh at intel.com>
>> Signed-off-by: Michael Qiu <qiudayu at chinac.com>
>> Signed-off-by: Tiago Lam <tiago.lam at intel.com>
>> ---
>>  lib/netdev-dpdk.c | 94 
>> +++++++++++++++++++++++++++++++++++++++++++++++++------
>>  1 file changed, 84 insertions(+), 10 deletions(-)
>>
>> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
>> index 9b1fb9a..0079e28 100644
>> --- a/lib/netdev-dpdk.c
>> +++ b/lib/netdev-dpdk.c
>> @@ -515,6 +515,22 @@ dpdk_rte_mzalloc(size_t sz)
>>      return rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE);
>>  }
>>
>> +static struct rte_mbuf *
>> +dpdk_buf_alloc(struct rte_mempool *mp)
>> +{
>> +    if (!dpdk_thread_is_pmd()) {
>> +        ovs_mutex_lock(&nonpmd_mp_mutex);
> Can you explain why the lock is needed here for non PMD threads?
>> +    }
>> +
>> +    struct rte_mbuf *mbuf = rte_pktmbuf_alloc(mp);
>> +
>> +    if (!dpdk_thread_is_pmd()) {
>> +        ovs_mutex_unlock(&nonpmd_mp_mutex);
>> +    }
>> +
>> +    return mbuf;
>> +}
>> +
>>  void
>>  free_dpdk_buf(struct dp_packet *p)
>>  {
>> @@ -2167,6 +2183,71 @@ out:
>>      }
>>  }
>>
>> +static int
> 
> a bool, true for success might be more usefull here.

I've updated this part to return better error codes. More specifically,
`ENOMEM` so the caller knows there's no memory.

> 
>> +dpdk_prep_tx_buf(struct dp_packet *packet, struct rte_mbuf **head,
>> +                 struct rte_mempool *mp)
> 
> Maybe the function name should be more what it's doing, i.e 
> dpdk_clone_dp_packet_to_mbuf().
> 

Sounds clearer to me, aside from the clone since it's different types of
packets. What about `dpdk_copy_dp_packet_to_mbuf()`?

>> +{
>> +    struct rte_mbuf *temp;
>> +    uint32_t size = dp_packet_size(packet);
>> +    uint16_t max_data_len, data_len;
>> +    uint32_t nb_segs = 0;
>> +    int i;
>> +
>> +    temp = *head = dpdk_buf_alloc(mp);
>> +    if (OVS_UNLIKELY(!temp)) {
>> +        return 1;
>> +    }
>> +
>> +    /* All new allocated mbuf's max data len is the same */
>> +    max_data_len = temp->buf_len - temp->data_off;
>> +
>> +    /* Calculate # of output mbufs. */
>> +    nb_segs = size / max_data_len;
>> +    if (size % max_data_len) {
>> +        nb_segs = nb_segs + 1;
>> +    }
>> +
>> +    /* Allocate additional mbufs when multiple output mbufs required. 
>> */
>> +    for (i = 1; i < nb_segs; i++) {
>> +        temp->next = dpdk_buf_alloc(mp);
>> +        if (!temp->next) {
>> +            free_dpdk_buf((struct dp_packet *) *head);
>> +            *head = NULL;
>> +            break;
>> +        }
>> +        temp = temp->next;
>> +    }
>> +    /* We have to do a copy for now */
>> +    rte_pktmbuf_pkt_len(*head) = size;
>> +    temp = *head;
>> +
>> +    data_len = size < max_data_len ? size: max_data_len;
> 
> Can we use max_data_len to copy? It's only valid if rte_pktmbuf_mtod() 
> returns the first byte, but after alloc it's 
> rte_pktmbuf_reset_headroom(). So we will overwrite invalid memory.
> 
>> +    if (packet->source == DPBUF_DPDK) {
>> +        *head = &(packet->mbuf);
>> +        while (temp && head && size > 0) {
>> +            rte_memcpy(rte_pktmbuf_mtod(temp, void *),
>> +                    dp_packet_data((struct dp_packet *)head), 
>> data_len);
> 
> Here you assume source and destination mbuf sizes are the same...
> Also use container_of for "(struct dp_packet *)head)"
> 
>> +            rte_pktmbuf_data_len(temp) = data_len;
>> +            *head = (*head)->next;
>> +            size = size - data_len;
>> +            data_len =  size < max_data_len ? size: max_data_len;
>> +            temp = temp->next;
>> +        }
>> +    } else {
> 
> Why not use dp_packet_mbuf_write() here?
> 

I missed this. I'll take this approach as well, thanks.

Tiago.