[ovs-discuss] ovs netdev-dpdk priv descriptor.

[Daniele Di Proietto]

Hi,

As part of his MTU series, Mark sent a patch to do this:

http://openvswitch.org/pipermail/dev/2016-February/066406.html

On 19/02/2016 10:40, "David Evans" <davidjoshuaevans at gmail.com> wrote:

>Hey Pravin & OVS wizards.
>
>Just found something fun. If you're updating to dpdk 2.2 the support is
>better for tx handling when doing rte_mempool_create() is now ok to use
>the private data size
>
>
>You won't need the funky pointer arithmetic in  __rte_pktmbuf_init() any
>more as you can use the standard rte_pktmbuf_init() inside your
>ovs_rte_pktmbuf_init()
>
>
>And - the following to set the private data size to fit in your dp_packet
>
>
>dmp->mp = rte_mempool_create(mp_name, mp_size-1, MBUF_SIZE(mtu),
>MP_CACHE_SZ,
>sizeof(struct dp_packet) - sizeof(struct rte_mbuf),
>rte_pktmbuf_pool_init, NULL,
>ovs_rte_pktmbuf_init, NULL,
>socket_id, 0);
>
>
>
>This way, if you ever decide to extend to including stuff like
>ipv4_fragmentation (EG if you have  outgoing packets you want to match
>the MTU) the dpdk frag library uses indirect buffers you can be sure that
>all the pointers will line up right for
> attach / detatch / free.
>
>
>so you can make one of these
>
>
>dmp->mp_indirect = rte_mempool_create(mp_name_indirect, mp_size-1,
>sizeof(struct dp_packet), 32,
>sizeof(struct dp_packet) - sizeof(struct rte_mbuf),
>NULL, NULL,
>rte_pktmbuf_init, NULL,
>socket_id, 0);
>
>
>
>and then based on packet size do the same as the ip_fragmentation example
>in dpdk.
>
>
>Inside the rte_pktmbuf_free_seg routine is some pointer arithmetic that
>needs the private data size to be set correct  to find the original
>direct buffer from the indirect one.
>
>
>Hope this is useful. I don't want to submit a patch myself as my code has
>all manner of other crazy application specific tricks beyond this.
>
>
>Regards,
>Dave.
>
>
>
>