1. export the comm.opencon function. Does all the connect/request
timeout, recv_before, and first data payload stuff that tryssl does, but
without trying SSL. Ought to save some boilerplate in some scripts
2. Make opencon use setup_connect instead of duplicating code.
3. Move a debug message about tryssl to tryssl from opencon
4. Transparently handle UDP in tryssl, in case someone does that. Debug
message about DTLS not being supported is printed, but otherwise just
connects with opencon. Previously, doing this would result in a
connection to the TCP version of the port, even if you passed in a port
table with port.protocol=="udp".
As reported by nnposter (http://seclists.org/nmap-dev/2014/q3/472) using
the rtt-based timeouts for read timeouts is not a good idea, since host
processing time can be considerably longer, especially for SSL
connections. comm.lua already allowed for different connect_ and
request_timeout values to reflect this truth, so this commit switches to
using the rtt-based timeout for the connect timeout and adding 6 seconds
to get the request timeout. This value is based on the totalwaitms value
in nmap-service-probes, and is still well short of the default 30s nsock
timeout.
There's no reason we can't use other verbs besides GET and POST. Other
verbs are handled like GET requests (parameters in the URI string). Any
redirect responses will be followed with GET requests, though.
std::list::size() was hanging because list traversal was broken for this
particular static list (timedout_hops). Could not reproduce with a small
test program, but another AIX user confirmed. Exhausted other options
before switching it to dynamic allocation (of the list itself, not the
elements, which are always dynamically allocated), which somehow
bypasses the problem.
Doesn't matter on most platforms because our own object (.o) files
require the symbols from libnbase, so all of them have been linked by
the time libnsock is encountered, but vasprintf and asprintf were
causing problems on platforms that need the versions defined in
libnbase.
This chunk of code in read_replies was generating valgrind errors
because reply.ttl was uninitialized:
if (host->state == HostState::COUNTING_DOWN) {
/* If this probe was past the target, skip ahead to what we think the
actual distance is. */
int distance = get_initial_ttl_guess(reply.ttl) - reply.ttl + 1;
if (distance > 0 && distance < host->current_ttl)
host->current_ttl = MIN(distance, MAX_TTL);
}
My compiler was setting the initial value to 0, which meant that
distance was always 33, clearly wrong. Setting reply.ttl in decode_reply
silences the error and causes distance to be more accurate.
This was a hassle to support on AIX (several broken implmentations) and
the move to a more portable solution (using autoconfig macros) was also
troublesome. Upon review, nse_fs.cc only uses directory and link
functions, so large files support shouldn't matter. Stripping it out.
nse_fs.cc was forcing large files support, and it was broken on AIX due
to this GCC bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=20366
Now we let autoconf handle setting the appropriate defines, and also
check for this particular bug before defining.
Changes fall into these categories:
1. Avoid pathological string building. Loops over x = x .. "foo" can
become very slow. Instead, use strbuf.lua, table.concat, or just one
continuous concatenation; a = x .. y .. z is one operation, better than
a = x .. y; a = a .. z
2. Use hex-escaped strings instead of string.char. I find this more
readable in many cases, and it avoids a table lookup and function call.
3. Don't duplicate code. A few libraries and scripts had re-implemented
stdnse.generate_random_string or openssl.rand_bytes.
