[dpdk-dev] [Bug 870] Fragmented packets incorrect RSS distribution with TAP PMD on Hyper-V platform

[bugzilla@dpdk.org]

https://bugs.dpdk.org/show_bug.cgi?id=870

            Bug ID: 870
           Summary: Fragmented packets incorrect RSS distribution with TAP
                    PMD on Hyper-V platform
           Product: DPDK
           Version: 20.11
          Hardware: x86
                OS: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: Normal
         Component: ethdev
          Assignee: dev@dpdk.org
          Reporter: madhuker.mythri@oracle.com
  Target Milestone: ---

On Hyper-V based platforms(Azure with Accelerated-mode disabled), the network
ports come up on Failsafe-PMD(and underneath TAP PMD) configured with 4-queues
RSS enabled.
The normal Rx/Tx packets were received and transmitted well. But, when the IP
fragmented packets received on these ports, all the fragments of single-packet
were distributed across different queues, instead of arriving to single-queue.

As per the TAP PMD RSS(based on BPF) algorithm mentioned in the 
"tap_bpf_program.c" file, the RSS algorithm written based on 4-tuple(L3 Src/Dst
and L4 Src-port/Dst-port) as mentioned in below code snippet, the RSS algorithm
does not check for the IP fragmented packets.
It assumes all the incoming IP based packets will have L4-header and trying to
access the L4 Src-port/Dst-port fields using the pointer of L3-header
end-address  without cross checking the L4 header presence.  since, in-case of
fragmented packets the L4 header will not be present expect for first fragment
and thus should not consider L4 Src-port/Dst-port under the 4-tuple hash
calculation.

So, this RSS hash calculation has a bug in-case of fragmented packets.

File-name: ./drivers/net/tap/tap_bpf_program.c
================ =========================
rss_l3_l4(struct __sk_buff *skb)
{
.....
.....

  __u8 *src_dst_addr = data + off + offsetof(struct iphdr, saddr);
        __u8 *src_dst_port = data + off + sizeof(struct iphdr);
        struct ipv4_l3_l4_tuple v4_tuple = {
            .src_addr = IPv4(*(src_dst_addr + 0),
                    *(src_dst_addr + 1),
                    *(src_dst_addr + 2),
                    *(src_dst_addr + 3)),
            .dst_addr = IPv4(*(src_dst_addr + 4),
                    *(src_dst_addr + 5),
                    *(src_dst_addr + 6),
                    *(src_dst_addr + 7)),
            .sport = PORT(*(src_dst_port + 0),
                    *(src_dst_port + 1)),
            .dport = PORT(*(src_dst_port + 2),
                    *(src_dst_port + 3)),
        };
....
....

}
=========================================

So, here we can see the L4 src/dst ports were fetched without cross-checking
the packet has L4 header and in-case of fragmented packets the L4 src/dst ports
will be some junk values.

Due to this all the fragments of same packet were landing on to different
queues and thus re-assembling the fragments got failed.

Tested with DPDK-version "20.11.2" and the Linux Kernel-version is "5.4.17".

-- 
You are receiving this mail because:
You are the assignee for the bug.