-----邮件原件-----
发件人: David Marchand <david.march...@redhat.com>
发送时间: 2023年9月27日 23:43
收件人: Dexia Li <dexia...@jaguarmicro.com>
抄送: ovs-dev@openvswitch.org; i.maxim...@ovn.org; Qingmin Liu
<qingmin....@jaguarmicro.com>; Joey Xing <joey.x...@jaguarmicro.com>; Mike
Pattrick <m...@redhat.com>
主题: Re: [PATCH v3] netdev-dpdk: Add vxlan and geneve tunnel tso.
On Wed, Sep 27, 2023 at 11:29 AM Dexia Li <dexia...@jaguarmicro.com> wrote:
> > > static inline uint32_t *
> > > dp_packet_ol_flags_ptr(const struct dp_packet *b) { @@ -615,9
> > > +673,10 @@ dp_packet_set_size(struct dp_packet *b, uint32_t v)
> > > * (and thus 'v') will always be <= UINT16_MAX; this means that
> > > there is no
> > > * loss of accuracy in assigning 'v' to 'data_len'.
> > > */
> > > - b->mbuf.data_len = (uint16_t)v; /* Current seg length. */
> > > - b->mbuf.pkt_len = v; /* Total length of all segments
> > > linked to
> > > - * this segment. */
> > > + /* Current seg length. */
> > > + b->mbuf.data_len += (uint16_t)(v - b->mbuf.pkt_len);
> >
> > What is this change for?
>
> For multi-seg packets, pkt_len and data_len of the first seg are resized
> because tunnel encap.
OVS userspace (datapath) code does not support multi segments packets.
Besides, the default headroom of a dp-packet is enough for most tunnel headers.
"Non DPDK" dp-packet can be adjusted dynamically with dp_packet_resize(), only
DPDK dp-packet has some limitation (for which I have a pending patch).
> The 128 Bytes headroom are reserved between buff_addr and data_off of
> mbuf. Then data_off move forward for the len of tunnel encap. So the first
> seg of the multi-seg packet should resize for data_len.
> At this time, the data_len of the first seg is
> 2048(RTE_MBUF_DEFAULT_DATAROOM) plus tunnel encap len.
> Before the resize, the data_len of the first seg is 2048. The dataroom is
> full of the first seg.
>
> It will result in tunnel packets tso error if the data_len of the
> first seg equals pkt_len. In my test, This will result in the last packet of
> seg packets can not appear.
I am not following.
There is no first segment to consider because there is only one segment/buffer
in OVS dp-packets.
When encapsulating, OVS checks if current headroom is enough. If not, it
allocates a bigger buffer and copies data into it.
In vxlan tunnel case,
netdev_push_header
- netdev_tnl_push_udp_header
- netdev_tnl_push_ip_header
- dp_packet_push_uninit
- dp_packet_prealloc_headroom(b, size);
- dp_packet_resize(b, MAX(size, 64), dp_packet_tailroom(b));
- https://github.com/openvswitch/ovs/blob/master/lib/dp-packet.c#L253
- dp_packet_set_data(b, (char*)dp_packet_data(b) - size);
- dp_packet_set_size(b, dp_packet_size(b) + size);
For DPDK-source dp-packets, as long as the headroom is big enough, the current
code should handle this situation fine.
If the headroom is not big enough, well, you may have a look at below link, but
that required at least 3 levels of encapsulations iirc.
https://patchwork.ozlabs.org/project/openvswitch/patch/20220318153339.31083-1-david.march...@redhat.com/
Did I miss something?
No, I did not make it clear. we can look at the graph in the attachment
When using dpdk dp packet, dp_packet_set_size will set wrong data_len of first
seg of rte mbuf.
It will result in tso error.
--
David Marchand
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
