[Evan Huus <eapache@xxxxxxxxx>]

On Tue, Aug 27, 2013 at 6:02 PM, Anders Broman <a.broman@xxxxxxxxxxxx> 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
stuff,
which we want to do sometimes. Any ideas why we have to keep some
information
indefinitely?

Two-pass requires us to keep *all* the state around through the first pass
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 later).

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 dissector-dependent.

Cheers,
Evan

[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 :)