["your highness" <zanash@xxxxxxxxx>]

On 2/19/07, your highness <zanash@xxxxxxxxx> wrote:
> On 2/16/07, Guy Harris <guy@xxxxxxxxxxxx> wrote:
> >
> > On Feb 16, 2007, at 7:06 AM, your highness wrote:
> >
> > > Hi everyone,
> > >
> > > I managed to write a dissector for a new protocol (Condor) and it
> > > works most of the time ; ) ... Thanks to Jaap Keuter and the README
> > > files. I am starting a new thread for a new problem.
> > >
> > > The problem is that I get errors for packets that are spread over
> > > multiple TCP packets.
> > >
> > > Condor runs on top of TCP. In the header of each condor packet the
> > > length of the whole packet is given by a 3-octets field, thus as
> > > suggested in the README.developer I used tcp_dissect_pdus() for
> > > reassembling condor packets that span across more than one TCP
> > > packets.
> >
> > It also handles multiple packets in *one* TCP segment.
> >
> > > The skeleton of my code that is relevant to this issue is as follows:
> > >
> > > static guint
> > > get_condor_pdu_len(tvbuff_t *tvb, int offset)
> > > {
> > > // total packet length in 2nd, 3rd and 4th octets of a Condor packet
> > > return tvb_get_ntoh24(tvb, offset+1);
> > > }
> >
> > So it's a 3-octet field at an offset of 1 octet into the packet, so
> > you have to have at least 4 bytes of the packet in order to get the
> > length field?
> >
> > If so, then:
> >
> > > static void
> > > dissect_condor_tcp(tvbuff_t *tvb, packet_info *pinfo, proto_tree
> > > *tree)
> > > {
> > > tcp_dissect_pdus(tvb, pinfo, tree, condor_desegment, 3,
> > > get_condor_pdu_len,
> > >    dissect_condor_tcp_pdu);
> > > }
> >
> > ...that should be
> >
> >         tcp_dissect_pdus(tvb, pinfo, tree, condor_desegment, 4,
> > get_condor_pdu_len,
> >            dissect_condor_tcp_pdu);
> >
> > as you need the first *4* bytes of the packet in order to get the
> > packet length.
>
> That is correct, the first 4 bytes are needed for the reason you
> mentioned. I have modified this and it resolves the "Malformed packet"
> error, but now it gives "TCP segment of a reassemebled PDU". Please
> read further in this email for more details about this message.
>
> >
> > > I first got the error message [Packet size limited during capture]
> > > when I sent a number of packets together (the length of the packets
> > > together was more than 1024 bytes).
> >
> > "Packet size limited during capture" usually means that the capture
> > was done with a "snapshot length" shorter than the maximum packet
> > length.  tcpdump, by default, has a "snapshot length" of 68 or 96.
> >
> > However, there are some cases where that error is incorrectly given
> > for malformed or un-reassembled packets.
> >
> > > But must be caused by the
> > > segmentation by the Data Link Layer.
> >
> > Probably not.
> >
> > Most data link layers either don't do segmentation (e.g., Ethernet and
> > PPP), or do the reassembly on the adapter or in the adapter driver, so
> > that the host doesn't see the segments (802.11, ATM AAL5).
> >
> > Some Wireshark dissectors for data link layers that do fragmentation
> > also support data link layer reassembly, for cases where the capture
> > doesn't have reassembled packets (because, for example, the adapter
> > was told not to do reassembly, or because the capture mechanism is
> > supplied with raw packets rather than reassembled packets if the
> > driver's doing reassembly); the 802.11 dissector can do that.  We
> > don't have that for ATM.
> >
> > However, data link layers also have a maximum packet size that they'll
> > *accept* from a higher layer - they might internally fragment, at the
> > data link layer, packets at or below that maximum packet size, or they
> > might not (Ethernet, for example, doesn't).  This means that if a
> > higher-layer protocol needs to send a packet larger than that,
> > segmentation needs to be done at *that* layer.
> >
> > That segmentation isn't done by the data link layer, even though it's
> > caused by a limitation of the data link layer.  It might be done at
> > the network layer (IP or CLNP fragmentation, for example), or it might
> > be done at the transport layer (TCP breaking the data stream into TCP
> > segments, COTP breaking a COTP PDU into segments).
> >
> > If, for example, you send a 16K Condor packet, that won't fit into
> > most link-layer packets.  The segmentation for that is done at the TCP
> > layer, *not* by the data link layer.
> >
> > > Using tcp_dissect_pdus() solved
> > > this problem. I don't understand how this happens because I supposed
> > > that when a protocol runs on top of TCP, the tcp dissector itself
> > > would take care of reassembly of the Data Link Layer packets. Is my
> > > reasoning correct?
> >
> > No.
> >
> > When a protocol runs on top of TCP, the TCP implementation on the
> > machine receiving packets will not reassemble all the data for a
> > packet for that protocol and only hand a reassembled packet to the
> > implementation of that higher-level protocol.  That's because TCP has
> > no idea when a packet for that protocol starts or ends; the service
> > TCP provides to protocols running on top of it is a stream of bytes,
> > with *NO* packet boundaries.
> >
> > That's why Condor - and many other protocols running on top of TCP -
> > need to put a length at the beginning of their packets.  That way, the
> > implementation of that protocol on the receiving side knows how big
> > the packet is, and keeps reading from TCP until it has all the bytes
> > of that packet.
> >
> > The TCP dissector in Wireshark has the same problem, so it has to be
> > told by the dissector for the protocol how big the packet is.  That
> > can be done by hand, but tcp_dissect_pdus() can do it for you in many
> > cases.
> >
> >
> > > So the problem that I get now appears when I send more than 8192 bytes
> > > (8K). In that case some Condor packets will span accross different TCP
> > > packets, right?
> >
> > If you're running atop Ethernet, the number is more like 1460 bytes,
> > assuming IPv4 with no options and TCP with no options (1514 bytes of
> > Ethernet header plus data = 14 bytes of header + 1500 bytes of data;
> > 1500 bytes of IPv4 header with no options + data = 20 bytes of IPv4
> > header with no options plus 1480 bytes of data; 1480 bytes of TCP
> > header with no options + data = 20 bytes of TCP header with no options
> > plus 1460 bytes of data).
>
> Right. My assumption was thus wrong and I was interpreting things
> wrong. I am running atop Ethernet and it seems that the maximum on my
> network is 1024 bytes of TCP data under the mentioned assumptions,
> except that 12 bytes of TCP options are used.
>
> >
> > Unless you're dealing with a link layer that handles packets with 8232
> > bytes of data (8192 bytes of TCP payload + 20 bytes of TCP header + 20
> > bytes of IP header; that 8232 does *NOT* count the link-layer header,
> > so it's 8232 + link-layer header size), or whatever the appropriate
> > value would be for IPv6 (or with IP or TCP options), you're going to
> > see segmentation before you get to 8192-byte packets.
>
> What I still don't understand is that I get problems *only* if I send
> more than 8192 bytes. So when I send 195 packets of 42 bytes (a total
> of 8190 bytes) dissection is ok and I see 8 TCP packets:
>  - Each of the TCP packets 1 through 7 contain 1024 bytes of Condor
> packets, which is 24 whole packets +  parts of 1 or 2 segmented
> packets
>  - The last TCP packet contains 1022 bytes of Condor packets
>
> But when I add one packet, so I send 196 packets (total of 8232 bytes)
> things go wrong. As I mentioned above, after changing the 5th
> parameter of tcp_dissect_pdus() from 3 to 4 I now get "TCP segment of
> a reassemebled PDU" for the 9th TCP packet and the last Condor packet
> is not dissected.
>
> >
> > > Only condor packets in the first TCP packet are
> > > recognized as valid. The other ones are tagged "Malfored packet".
> > >
> > > I am using netcat to send files that contain Condor packets for my
> > > tests. Attached is a captured file that contains condor packets of 42
> > > bytes. The first packet that is tagged Malformed has 2 bytes in the
> > > first TCP packets (0x20 and 0x00) and 40 bytes in the second TCP
> > > packet.
>
> I forgot to mention that the files are not sent over the network; I
> send and receive the file on the same machine using the loop back
> interface. I had to disable the checksum validation of TCP as proposed
> in section "Viewing traffic" of the FAQ, because of "TCP checksum
> offloading".
>
> I am going to test whether using the same interface of a machine for
> sending and receiving has any effect on this issue.

And it *has*! Sending the same file of more than 8192 bytes with
netcat from one machine to the one where wireshark runs results in
correct dissection.

Any explanation for the following?
- why things might go wrong when more than 8192 bytes are sent
- why only when the receiver and the sender use the same machine (the
same interface)

> >
> > I'll take a look at that later.
> > _______________________________________________
> > Wireshark-dev mailing list
> > Wireshark-dev@xxxxxxxxxxxxx
> > http://www.wireshark.org/mailman/listinfo/wireshark-dev
> >