Home page logo

wireshark logo Wireshark mailing list archives

Re: Memory consumption in tshark
From: Evan Huus <eapache () gmail com>
Date: Tue, 27 Aug 2013 18:17:13 -0400

On Tue, Aug 27, 2013 at 6:02 PM, Anders Broman <a.broman () bredband net>wrote:

Joerg Mayer skrev 2013-08-27 23:24:

 On Tue, Aug 27, 2013 at 05:09:19PM -0400, Evan Huus wrote:

IIRC, two-pass allows for most/all of the reassembly/request-response
which we want to do sometimes. Any ideas why we have to keep some

 Two-pass requires us to keep *all* the state around through the first
so that it is available during the second pass (at which point it can be
discarded).  Even in single-pass mode, there is some state that we can't
always immediately discard. If I see a fragment of a TCP message then it
doesn't make sense to discard that until the other fragments have arrived
and been reassembled. If I see a request, I probably need to keep state
from that request until the response (which may never show up).

We already do reassembly and a lot of other stateful work in single-pass
mode. The only thing two-pass mode provides is the ability to "see the
future" (for example, saying: this request has a response 5 packets

So (assuming we really free everything we could already) could add a
possibly configurable foresight horizon of 10000 packets. If a packet
number is older than 10000 packets, forget it?

 I haven't really looked but we do keep the reassembled fragments around
even in tshark as there is no
mechanism to discard them selectively if run by tshark as opposed to
wireshark and that's the really big memory eater I would think.

This is the primary problem. Any state saved with p_add_proto_data, or the
reassembly API, or the conversation API, or (god-forbid) globals is not
being freed right now. Any memory allocated with seasonal/file scope is not
being freed right now. We do free frame data, column data and misc. other
bits and pieces, but the majority of the state we create is not freed until
the end of the file.

As Anders says, this is because we have no way right now to selectively
discard it: much of the data is stored in a way that we can only get rid of
all of it, or none. Part of that is because there was no se_free call (for
which wmem_free is the convenient solution) and part of that is just
because nobody's ever added it to, for example, the reassembly API.

I'm sure there are some significant improvements we could make if somebody
figures out how, but you also have to beware of bugs like 9027 [1] which is
occurring precisely because we were trying to free old reassembly data. It
turns out there are certain cases where that data, at least, could still be
in use. The logic to selectively free state is not simple, and is often


[1] https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=9027

P.S. Given the above it would actually be fairly easy to create a fully
state-less tshark: make file-scope memory redirect to packet-scope, and
make our stateful APIs (p_add_proto_data, reassembly, conversations)
no-ops. I'm sure this will break some dissectors that use globals, but they
should be fixed anyways :)
Sent via:    Wireshark-dev mailing list <wireshark-dev () wireshark org>
Archives:    http://www.wireshark.org/lists/wireshark-dev
Unsubscribe: https://wireshark.org/mailman/options/wireshark-dev
             mailto:wireshark-dev-request () wireshark org?subject=unsubscribe

  By Date           By Thread  

Current thread:
[ Nmap | Sec Tools | Mailing Lists | Site News | About/Contact | Advertising | Privacy ]