Hello, Thank you for your reply.
On Mon, 8 Sept 2025 at 10:43, chenyongchang via discuss
<[email protected]> wrote:
>>
>> Hi all,
>> I noticed an issue when recalculating UDP checksums in
>> dp_packet_ol_send_prepare -> packet_udp_complete_csum -> csum_finish.
> Avoid sending the same mail multiple times.
Sorry about the mistake, I’ll make sure to avoid it in the future.
>>
>> If the packet is the first fragment of a UDP fragmented packet, the
>> checksum can be wrong.
>> Example:
>> UDP payload 2000 bytes, fragmented with MTU 1500.
>> First fragment payload 1480 bytes, but udp->udp_len = 2008.
>> Checksum calculation cannot access later fragments, so it’s incorrect.
>>
>> Has anyone seen this before or have a recommended solution for handling UDP
>> checksum on fragmented packets?
> This report is a bit low on details.
> I'll assume you are talking about the main branch.
> There was a fix recently, you can have a look at: 3ffb1c4012b7 ("ipf:
> Complete l4 checksum before frag out.").
> If this does not help, does reverting e36793e11fe8 ("dp-packet:
> Resolve unknown checksums.") fix your issue?
> And we will need a reproducer.
Yes, we are debugging based on the main branch.
These two fixes are not related to this issue.
We are using OVS-DPDK with Intel 82599 or X710 NICs (neither of which supports
NETDEV_TX_OFFLOAD_OUTER_UDP_CKSUM).
The packet flow is as follows: a packet enters through a physical NIC managed
by OVS-DPDK, then OVS-DPDK performs NAT (only modifying the destination
address), encapsulates it into a Geneve tunnel, and finally sends it out
through another physical NIC managed by OVS-DPDK.
If OVS-DPDK receives a fragmented UDP packet, after performing NAT it needs to
recalculate the UDP checksum. The issue described above occurs during this
checksum recalculation.
The detailed debugging steps are as follows:
gdb) bt
#0 packet_udp_complete_csum (p=0x7fa8db453900, inner=true) at
lib/packets.c:2092
#1 0x00000000016ce823 in dp_packet_ol_send_prepare (p=0x7fa8db453900,
flags=128) at lib/dp-packet.c:618
#2 0x0000000001734933 in netdev_send (netdev=0x17f3e7840, qid=23,
batch=0x7fa8db43b030, concurrent_txq=false) at lib/netdev.c:940
#3 0x00000000016dcbb8 in dp_netdev_pmd_flush_output_on_port
(pmd=0x7fa8dc3ff680, p=0x7fa8db43b000) at lib/dpif-netdev.c:5578
#4 0x00000000016dcdb2 in dp_netdev_pmd_flush_output_packets
(pmd=0x7fa8dc3ff680, force=false) at lib/dpif-netdev.c:5619
#5 0x00000000016dd044 in dp_netdev_process_rxq_port (pmd=0x7fa8dc3ff680,
rxq=0x7fa911b15790, port_no=5) at lib/dpif-netdev.c:5678
#6 0x00000000016e16ca in pmd_thread_main (f_=0x7fa8dc3ff680) at
lib/dpif-netdev.c:7306
#7 0x00000000017ba1e1 in ovsthread_wrapper (aux_=0x7fa911a30fa0) at
lib/ovs-thread.c:429
#8 0x00007fa91288e802 in start_thread () from /lib64/libc.so.6
#9 0x00007fa91282e450 in clone3 () from /lib64/libc.so.6
(gdb) p udp_sz
$7 = 1480
(gdb) p /x *udp
$8 = {udp_src = 0xe803, udp_dst = 0xe803, udp_len = 0xd807, udp_csum = 0xd1f6}
(gdb) p /x udp->udp_csum
$9 = 0xd1f6
(gdb) c
Continuing.
Best regards
From: David Marchand
Date: 2025-09-08 17:03
To: chenyongchang
CC: ovs-discuss
Subject: Re: [ovs-discuss] UDP checksum issue on first fragment of fragmented
packets
_______________________________________________
discuss mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-discuss
