On Sun, 16 Nov 2003, Guy Harris wrote:
> On Sun, Nov 16, 2003 at 10:04:32AM -0800, Richard Sharpe wrote:
> > Yes, then the big problem is that we need to run an AWK script across all
> > the source changing all those protocol index def'ns.
>
> No - we continue to have "proto_register_protocol()" return an index
> (which is what "proto_tree_add_item()" and
> "proto_tree_add_protocol_format()" expect); we just change what's inside
> a dissector handle, and have "dissector_add()" and the like do the
> translation.
OK, I must be thick, because I am not understanding ...
The performance problem, at least according to gprof data seems to be that
compare_proto_id is called far too many time. It is called because
find_protocol_by_id is walking the list of protocols and passing
compare_proto_id as the equality function:
list_entry = g_list_find_custom(protocols, &proto_id, compare_proto_id);
Now, a large number of functions in proto.c do this:
proto_get_protocol_<attr>(int proto_id)
{
protocol_t *protocol;
protocol = find_protocol_by_id(proto_id);
return protocol-><attr>;
}
and proto_get_protocol_short_name is the biggest user of
find_protocol_by_id in my gprof trace, followed by
proto_is_protocol_enabled.
It seems to me that we cannot simply change a proto_id into a pointer but
cast it to an int, as that will cause problems on 64-bit platforms.
So, I want to change all those routines that currently deal with a
proto_id to deal with a protocol_t * ...
But that requires changes to lots of places, including all dissectors
(although the changes will be minor in each of them).
Regards
-----
Richard Sharpe, rsharpe[at]ns.aus.com, rsharpe[at]samba.org,
sharpe[at]ethereal.com, http://www.richardsharpe.com