On Thu, 20 Apr 2023 02:30:41 +0000 "Hu, Jiayu" <jiayu...@intel.com> wrote:
> Hi Cheng, > > > -----Original Message----- > > From: jiangheng (G) <jianghen...@huawei.com> > > Sent: Saturday, April 15, 2023 10:46 PM > > To: us...@dpdk.org; Hu, Jiayu <jiayu...@intel.com>; dev@dpdk.org > > Subject: [GRO] check whether ip_id continuity needs to be checked when > > two TCP packets are merged. > > > > Hi jiayu.hu > > > > It cannot be guaranteed that 16bit identification field of ip packets in the > > same tcp stream will be continuous. > > Please help check whether ip_id continuity needs to be checked when two > > TCP packets are merged? > > Seems to modify the following code, gro will aggregate better, and work > > better: > > > > diff --git a/lib/gro/gro_tcp4.h b/lib/gro/gro_tcp4.h index > > 212f97a042..06faead7b5 100644 > > --- a/lib/gro/gro_tcp4.h > > +++ b/lib/gro/gro_tcp4.h > > @@ -291,12 +291,10 @@ check_seq_option(struct gro_tcp4_item *item, > > /* check if the two packets are neighbors */ > > len = pkt_orig->pkt_len - l2_offset - pkt_orig->l2_len - > > pkt_orig->l3_len - tcp_hl_orig; > > - if ((sent_seq == item->sent_seq + len) && (is_atomic || > > - (ip_id == item->ip_id + 1))) > > + if (sent_seq == item->sent_seq + len) > > For atomic packets, the IP ID field is ignored, as it can be set in various > ways. > For non-atomic packets, it follows Linux kernel tcp_gro_receive(). > > Is this change specific to your case? Can you give more details on why it > helps? > > Thanks, > Jiayu Agreed, DPDK GRO should follow Linux to avoid bugs.
