while caching the last element from the conversation hash chain lists speeds-up
the operation when the hash/chain lists are actually built, it
does NOT help a lot when a certain random conversation which is in the hash
table is looked-up.
I did some profiling and tracing and I saw that a lot of cpu time is spent in
the function conversation_lookup_hashtable() when wireshark
is asked to show the "Flow Graph", "TCP Conversations", "Voip Calls". I used
two types of captures with over 500k packets:
- tcp packets having the _same_ src ip addr, src tcp port, dst ip addr, dst tcp
port
- (mostly) sip packets containing sdp payloads which advertise the _same_ ip
addr, udp port for media
these types of captures lead to _huge_ chain lists behind the same hash bucket
(to which the conversation is actually mapped)
the solution would be to cache the last found conversation into the head of the
chain list and to use it whenever it is possible; most of the time the look-up
will be in O(1) instead of O(n) (n - number
of elements in the list).
https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=7149
svn path=/trunk/; revision=42141
Do the right thing with conversation hash chains.
Adds two new functions: conversation_insert_into_hashtable() and
conversation_remove_from_hashtable() that do the right thing with conversation
hash table chains and ordering and all that. Converts conversation_new(),
conversation_set_addr2() and conversation_set_port2() to use the new functions.
https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=7085
svn path=/trunk/; revision=42104
has been called.
In the conversation cleanup routine, free the GSlist for any proto_data which
may have been hanging off the (se_allocated) conversation.
svn path=/trunk/; revision=39484
https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=4422
From me: Fix a number of instances where the function prototype or
the function definition wasn't changed so there was a mismatch
thus causing Windows (but not gcc) compilation errors.
svn path=/trunk/; revision=32365
does (i.e., it will add the address bytes to the value that's already
there - it will not initialize the value, so you have to clear it before
doing any hashing).
svn path=/trunk/; revision=21578
the supplied patch fixes a problem where the options value should really be used from the conversation found (using
conversation_lookup_hashtable(...) to create a new conversation based on the already stored conversation template (the CONVERSATION_TEMPLATE bit is set in the stored conversation) rather from the options argument passed to the function(s).
This solves a problem that otherwise shows itself where "DISSECTOR_ASSERT(!(conv->options & CONVERSATION_TEMPLATE) && "Use the conversation_create_from_template function when the CONVERSATION_TEMPLATE bit is set in the options mask");" fails sometimes.
svn path=/trunk/; revision=18825
fix bug in conversation_delete_proto_data
second argument to g_slist_remove() is a pointer to the data, not a GSlist containing a list of such pointers.
svn path=/trunk/; revision=14755
This is very naughty and will cause problems when we have assigned a dissector to a dynamic port using conversation_set_dissector().
To make ethereal handle this case I have changed the try_conversation_dissector() to allow it to fail and return 0, meaning yes there is indeed a protocol registered for this conversation but that protocol rejected this packet.
(which only happens for "new" style dissectors, "old" style dissectors will never reject a packet that way)
When this happens the decode_udp_port() helper will still allow other dissectors to be tried, in the hope that the conversation is now used for some other protocol and thus someone else might be able to decode the packet.
Update SNMP and TFTP dissectors to check that even if there already is a conversation but that conversation does NOT have snmp/tftp registered as the dissector for it, then create a new conversation anyway and attach the proper dissector.
Since ethereal keeps track of which frame number a conversation started in, this actually works really well.
svn path=/trunk/; revision=14345
they have LF at the end of the line on UN*X and CR/LF on Windows;
hopefully this means that if a CR/LF version is checked in on Windows,
the CRs will be stripped so that they show up only when checked out on
Windows, not on UN*X.
svn path=/trunk/; revision=11400
- conversation.[ch] - To support not setting port2 on matching a
conversation. This is used by protocols such as iSNS in which the client
registers a TCP/UDP port with the server for notifications and the server
sends notifications to this port from different source ports.
- packet-isns.c - Added support for handling zero-length TLVs and ESI & SCN
frames (when registering an SCN/ESI port, a conversation dissector is
setup).
svn path=/trunk/; revision=11320
"try_circuit_dissector()" and "try_conversation_dissector()", as both fo
them call "call_dissector()" and "call_dissector()" now does that stuff
itself.
svn path=/trunk/; revision=6520
equivalents for the epan/ directory but leave winsock2.h in inet_pton.c
and inet_ntop.c for now (can't estimate the consequences).
svn path=/trunk/; revision=5928
When we see PRTOMAP GETPORT calls for UDP, make sure all further UDP packets to or from
this port goes to the ONC-RPC dissector regardless of the port on the other side.
We need this because if there is ONC-RPC traffic going between the ONC-RPC Program port to a port which has a normal ethereal dissector, ethereal would dissect the traffic as the protocol associated with the other port instead.
svn path=/trunk/; revision=5430
1. Changes how can_desegment works so that can_desegment is
only != 0 for whichever dissector is running immediately on
top of whoever offers the can_desegment service.
Thus DCERPC needs no special handling to see if it can trust
can_desegment (which is currently only available ontop of TCP
and not ontop of tcp->nbss->smb).
2. Changes fragment reassembly of transaction smb to only show
the defragmented packet for the transaction smb holding the
first fragment.
To see why, test it with a transaction SMB containing a ~60kb
PDU or larger. The old behaviour had approximately quadratic
behaviour regarding runtime for dissecting such PDUs.
(example: NetShareEnum is a command which can grow really really
large if the number of shares and comments are large)
svn path=/trunk/; revision=4296
than a pointer to a dissector function, as an argument.
This means that the conversation dissector is called through
"call_dissector()", so the dissector itself doesn't have to worry about
checking whether the protocol is enabled or setting
"pinfo->current_proto", so get rid of the code that does that in
conversation dissectors. Also, make the conversation dissectors static.
Get rid of some direct calls to dissectors; replace them with calls
through handles, and, again, get rid of code to check whether a protocol
is enabled and set "pinfo->current_proto" where that code isn't needed.
Make those dissectors static if they aren't already static.
Add a routine "create_dissector_handle()" to create a dissector handle
without registering it by name, if the dissector isn't used outside the
module in which it's defined.
svn path=/trunk/; revision=4281
of protocol-id-plus-datum pairs, so that multiple protocols can attach
information to the same conversation.
Dissectors that attach information to a conversation should not assume
that if they find a conversation it has one of its data attached to it;
the conversation might've been created by another dissector.
svn path=/trunk/; revision=3901
that look up conversations in hash tables, unless they are arguments
that will be ignored; if they're not being ignored, then if the argument
is a null pointer you may get a crash if it's dereferenced, and if it's
not a null pointer you'll only get a match if the conversation has
whatever stuff the arguments points to as its first address or port.
If you match a conversation with a wildcarded address and/or port, and
the address and/or port matched a non-wildcarded search argument, and
the conversation is for a connection-oriented transport protocol, set
the wildcarded address and/or port for the conversation to the value
that matched it.
svn path=/trunk/; revision=3897
"try_conversation_dissector()" does - start with as exact matches as
possible, and then start doing wildcarding - so that it can find
conversations with wildcard addresses or ports even if both address and
port arguments are supplied to it.
svn path=/trunk/; revision=3893
Bill Meier