Re: Endianness issue with selecting non-fragmented packets

[Guy Harris <gharris () sonic net>]

On Jul 27, 2018, at 11:21 AM, Richard Clayton <richard () highwayman com> wrote:

> I am running tcpdump under FreeBSD 11 on an AMD64.
>
> I have a file containing UDP packets and IP fragments.
>
> This command (the filter corresponds to the information on the man
> page):
>
>    tcpdump -r file.pcap "(ip[6:2] & 0x1FFF = 0)"
>
> unexpectedly prints all of the packets :-(
>
> The command:
>
>    tcpdump -r file.pcap "(ip[6:2] & 0xFF1F = 0)"
>
> skips all the fragments and only prints complete packets.
>
>
> This is clearly an endianness issue ... but shouldn't tcpdump/libpcap be
> hiding that from me ?

Yes.  The BPF code generated for that expression is

        # load 2-byte value at offset 12, i.e. Ethernet type
        (000) ldh      [12]
        # Compare for equality against 0x800, i.e. IPv4; if equal,
        # go to instruction 002, otherwise go to instruction 005.
        (001) jeq      #0x800           jt 2    jf 5
        # Load 2-byte value at offset 20, i.e. 14+6, so at an offset
        # of 6 from the beginning of the Ethernet payload, so at
        # an offset of 6 from the beginning of the IPv4 header,
        # so the flags and fragment offset.
        (002) ldh      [20]
        # Test bits 0x1fff; if any are set, go to instruction 0005,
        # and if they're all clear, go to instruction 0004.
        (003) jset     #0x1fff          jt 5    jf 4
        # Return the snapshot length, meaning "accept the packet
        # and return that many bytes of packet data".
        (004) ret      #65535
        # Return 0, meaning "reject the packet".
        (005) ret      #0

The load instructions are all loading the data as big-endian, i.e. they don't load a machine word in its native byte 
order, they load it as if it were in network byte order, so the same value should be loaded on a little-endian x86 
machine like yours or a big-endian machine such as a SPARC/{most}PowerPC/{S/390, z/Architecture} machine.

And I tested its on my x86-64 Mac, using the tip-of-the-1.9-branch libpcap and the tip-of-the-4.9-branch tcpdump, with 
a capture file with some fragmented IP packets, and the filter "(ip[6:2] & 0x1FFF = 0)" rejected the fragmented 
packets.  The filter "(ip[6:2] & 0xFF1F = 0)" rejected *all* the packets.

Nothing was done to libpcap or tcpdump since the 1.9.0/4.9.2 releases to affect that.

_______________________________________________
tcpdump-workers mailing list
tcpdump-workers () lists tcpdump org
https://lists.sandelman.ca/mailman/listinfo/tcpdump-workers