Re: [PATCH 3/3] net: xilinx: axienet: Relax partial rx checksum checks
From: Sean Anderson
Date: Thu Sep 05 2024 - 12:33:22 EST
On 9/5/24 10:59, Eric Dumazet wrote:
> On Thu, Sep 5, 2024 at 4:24 PM Sean Anderson <sean.anderson@xxxxxxxxx> wrote:
>>
>> On 9/4/24 12:30, Eric Dumazet wrote:
>> > On Tue, Sep 3, 2024 at 8:43 PM Sean Anderson <sean.anderson@xxxxxxxxx> wrote:
>> >>
>> >> The partial rx checksum feature computes a checksum over the entire
>> >> packet, regardless of the L3 protocol. Remove the check for IPv4.
>> >> Additionally, packets under 64 bytes should have been dropped by the
>> >> MAC, so we can remove the length check as well.
>> >
>> > Some packets have a smaller len (than 64).
>> >
>> > For instance, TCP pure ACK and no options over IPv4 would be 54 bytes long.
>> >
>> > Presumably they are not dropped by the MAC ?
>>
>> Ethernet frames have a minimum size on the wire of 64 bytes. From 802.3
>> section 4.2.4.2.2:
>>
>> | The shortest valid transmission in full duplex mode must be at least
>> | minFrameSize in length. While collisions do not occur in full duplex
>> | mode MACs, a full duplex MAC nevertheless discards received frames
>> | containing less than minFrameSize bits. The discarding of such a frame
>> | by a MAC is not reported as an error.
>>
>> where minFrameSize is 512 bits (64 bytes).
>>
>> On the transmit side, undersize frames are padded. From 802.3 section
>> 4.2.3.3:
>>
>> | The CSMA/CD Media Access mechanism requires that a minimum frame
>> | length of minFrameSize bits be transmitted. If frameSize is less than
>> | minFrameSize, then the CSMA/CD MAC sublayer shall append extra bits in
>> | units of octets (Pad), after the end of the MAC Client Data field but
>> | prior to calculating and appending the FCS (if not provided by the MAC
>> | client).
>>
>> That said, I could not find any mention of a minimum frame size
>> limitation for partial checksums in the AXI Ethernet documentation.
>> RX_CSRAW is calculated over the whole packet, so it's possible that this
>> check is trying to avoid passing it to the net subsystem when the frame
>> has been padded. However, skb->len is the length of the Ethernet packet,
>> so we can't tell how long the original packet was at this point. That
>> can only be determined from the L3 header, which isn't parsed yet. I
>> assume this is handled by the net subsystem.
>>
>
> The fact there was a check in the driver hints about something.
>
> It is possible the csum is incorrect if a 'padding' is added at the
> receiver, if the padding has non zero bytes, and is not included in
> the csum.
>
> Look at this relevant patch :
>
> Author: Saeed Mahameed <saeedm@xxxxxxxxxxxx>
> Date: Mon Feb 11 18:04:17 2019 +0200
>
> net/mlx4_en: Force CHECKSUM_NONE for short ethernet frames
Well, I tested UDP and it appears to be working fine. First I ran
# nc -lu
on the DUT. On the other host I used scapy to send a packet with some
non-zero padding:
>>> port = RandShort()
>>> send(IP(dst="10.0.0.2")/UDP(sport=port, dport=4444)/Raw(b'data\r\n')/Padding(load=b'padding'))
I verified that the packet was received correctly, both in netcat and
with tcpdump:
# tcpdump -i net4 -xXn
tcpdump: verbose output suppressed, use -v[v]... for full protocol decode
listening on net4, link-type EN10MB (Ethernet), snapshot length 262144 bytes
16:07:45.083795 IP 10.0.0.1.27365 > 10.0.0.2.4444: UDP, length 6
0x0000: 4500 0022 0001 0000 4011 66c8 0a00 0001 E.."....@.f.....
0x0010: 0a00 0002 6ae5 115c 000e 0005 6461 7461 ....j..\....data
0x0020: 0d0a 7061 6464 696e 6700 0000 0000 ..padding.....
and also checked for checksum errors:
# netstat -s | grep InCsumErrors
InCsumErrors: 0
to verify that checksums were being checked properly, I also sent a
packet with an invalid checksum:
>>> send(IP(dst="10.0.0.2")/UDP(sport=port, dport=4444, chksum=5)/Raw(b'data\r\n')/Padding(load=b'padding'))
and confirmed that there was no output on netcat, and that I had gotten
a UDP checksum error:
# netstat -s | grep InCsumErrors
InCsumErrors: 1
I can try to test TCP as well, but it is a bit trickier due to the 3-way
handshake. From the documentation, partial checksums should be agnostic
to the L3 protocol, so I don't think there should be any difference.
--Sean