Two weeks ago (yeah, I am a slacker) FIRST technical colloquium was held in Prague and we (me and Sasha) were given the opportunity to attend (the fact the Derrick serves as FIRST chair in the steering comittee has of course something to do with it).

I only attended one day of the technical colloquium (Tuesday 29th). The day was filled with various talks and presentations. Most of them were performed by various CERT teams members from around the world. This was because this event was a joint meeting of FIRST and TF-CSIRT. It was definitely interesting to see very different approaches to the shared problem set (dealing with incidents, setting up honey pots, building forensic analysis labs, etc.). Not only these differences stemmed from sizes of the networks and organizations but also (and that was kind of funny) from nationalities.
In the morning I talked about the integration of Netcat into Solaris, describing the process, current features and planned enhancements and extensions.

The most anticipated talk was by Adam Laurie who is entertaining guy involved in many hacker-like activities (see e.g. A hacker games the hotel article by Wired) directed at proving insecurities in many publicly used systems.
Adam (brother of Ben Laurie, author of Apache-SSL and OpenSSL contributor) first started with intro about satellite scanning, insecure hotel safes (with backdoors installed by the manufacturers which can be overcome by a screwdriver). Then he proceeded to talk about RFID chips, mainly about cloning.

Also, at the "social event" in the evening I had the pleasure to share a table with Ken van Wyk who is overall cool fellow and the author of Secure coding and Incident response books from O'Reilly.

In overall, it was interesting to see so many security types in a room and get to know some of them.

It seems that many developers and dtrace users found themselves in a position where they wanted to add some SDT probes to a module to get more insight into what's going on but the had to pause and were thinking "okay, more probes. But where to put them ? Do I really need the additional probes when I already have the fbt ones ?". To do this, systematic approach is needed in order not to over-do or under-do. I will use KSSL (Solaris kernel SSL proxy [1]) for illustration.

With CR 6556447, tens of SDT probes were introduced into KSSL module and other modules which interface with it. Also, in addition to the new SDT probes, in KSSL we got rid of the KSSL_DEBUG macros compiled only in DEBUG kernels and substituted them with SDT probes. As a result, much better observability and error detection was achieved with both debug and non-debug kernels. The other option would be to create KSSL dtrace provider but that would be too big gun for what is needed to achieve.

Generically, the following interesting data points for data gathering/observation can be identified in code:

• data paths
  When there is a more than one path how data could flow into a subsystem. E.g. for TCP we have couple of cases how SSL data could reach KSSL input queue. To identify where from exactly was tcp_kssl_input() called we use SDT probes:

  	if (tcp->tcp_listener || tcp->tcp_hard_binding) {
  ...
  		if (tcp->tcp_kssl_pending) {
  			DTRACE_PROBE1(kssl_mblk__ksslinput_pending,
  			    mblk_t *, mp);
  			tcp_kssl_input(tcp, mp);
  		} else {
  			tcp_rcv_enqueue(tcp, mp, seg_len);
  		}
  	} else {
  ...
  		/* Does this need SSL processing first? */
  			if ((tcp->tcp_kssl_ctx != NULL) &&
  			    (DB_TYPE(mp) == M_DATA)) {
  				DTRACE_PROBE1(kssl_mblk__ksslinput_data1,
  				    mblk_t *, mp);
  				tcp_kssl_input(tcp, mp);
  			} else {
  				putnext(tcp->tcp_rq, mp);
  				if (!canputnext(tcp->tcp_rq))
  					tcp->tcp_rwnd -= seg_len;
  			}
  ...
  

• data processed in while/for cycles
  To observe what happens in each iteration of the cycle. Can be used in code like this:

  while (mp != NULL) {
  
    DTRACE_PROBE1(kssl_mblk__handle_record_cycle, mblk_t *, mp);
  
    /* process the data */
    ...
  
    mp = mp->b_cont;
  }
  

• switch statements
  If significant/non-trivial processing happens inside switch it may be useful to add SDT probes there too. E.g.:

    content_type = (SSL3ContentType)mp->b_rptr[0];
    switch(content_type) {
      /* select processing according to type */
      case content_alert:
         DTRACE_PROBE1(kssl_mblk__content_alert, mblk_t *, mp);
         ...
         break;
      case content_change_cipher_spec:
         DTRACE_PROBE1(kssl_mblk__change_cipher_spec, mblk_t *, mp);
         ...
         break;
      default:
         DTRACE_PROBE1(kssl_mblk__unexpected_msg, mblk_t *, mp);
         break;
    }
  

• labels which cannot be (easily) identified in other way
  Useful if code which follows the label is generic (assignments, no function calls), e.g.:

                                  /*
                                   * Give this session a chance to fall back to
                                   * userland SSL
                                   */
                                  if (ctxmp == NULL)
                                          goto no_can_do;
  ...
  no_can_do:
                                  DTRACE_PROBE1(kssl_no_can_do, tcp_t *, tcp);
                                  listener = tcp->tcp_listener;
                                  ind_mp = tcp->tcp_conn.tcp_eager_conn_ind;
                                  ASSERT(ind_mp != NULL);
  

You've surely noticed that same of the probe definitions above have common prefix (kssl_mblk-). This is one of the things which make SDT probes so attractive. With prefixes it is possible to do the e.g. following:

sdt:::kssl_err-*
{
  trace(timestamp);
  printf("hit error in %s\n", probefunc);
  stack(); ustack();
}

The important part is that we do not specify module of function name. The implicit wildcard (funcname/probename left out) combined with explicit wildcard (prefix + asterisk) will lead to all KSSL error probes to be activated regardless of in which module or function there are defined. This is obviously very useful for technologies which span multiple Solaris subsystems or modules (such as KSSL).

The nice thing about the error probes is that they could be leveraged in test suites. For each test case we can first run dtrace script with the above probeset covering all KSSL errors in the background and after the test completes just check if it produced some data. If it did, then the test case can be considered as failed. No need to check kstat(1M) (and other counters), log files, etc.

Also, thanks to the way how dtrace probes are activated we can have both generic probeset (using this for lack of better term) as above with addition of probe specific action, e.g.:

/* probeset of all KSSL error probes */
sdt:::kssl_err-*
{
  trace(timestamp);
  printf("hit error in %s\n", probefunc);
}

/* 
  the probe definition is:
     DTRACE_PROBE2(kssl_err__bad_record_size,
         uint16_t, rec_sz, int, spec->cipher_bsize);
 */
sdt:kssl:kssl_handle_record:kssl_err-bad_record_size
{
  trace(timestamp);
  tracemem(arg0, 32);
  printf("rec_sz = %d , cipher_bsize = %d\n", arg1, arg2);
}

If probe kssl_err-bad_record_size gets activated the generic probe will be activated (and fires) too because the probeset contains the probe.

Similarly to the error prefix, we can have data specific prefix. For KSSL it is kssl_mblk- prefix which could be used for tracing all mblks (msgb(9S)) as they flow through TCP/IP, STREAMS and KSSL modules. With such probes it is possible to do e.g. the following:

/* how many bytes from a mblk to dump */
#define DUMP_SIZE       48

/* use macros from  */
#define MBLKL(mp)       ((mp)->b_wptr - (mp)->b_rptr)
#define DB_FLAGS(mp)    ((mp)->b_datap->db_flags)

#define PRINT_MBLK_INFO(mp)                                             \
        printf("mblk = 0x%p\n", mp);                                    \
        printf("mblk size = %d\n", MBLKL((mblk_t *)mp));                \
        PRINT_MBLK_PTRS(mp);

#define PRINT_MBLK(mp)                                                  \
                trace(timestamp);                                       \
                printf("\n");                                           \
                PRINT_MBLK_INFO(mp);                                    \
                printf("DB_FLAGS = 0x%x", DB_FLAGS((mblk_t *)mp));      \
                tracemem(((mblk_t *)mp)->b_rptr, DUMP_SIZE);            \
                tracemem(((mblk_t *)mp)->b_wptr - DUMP_SIZE,            \
                        DUMP_SIZE);

sdt:::kssl_mblk-*
{
  trace(timestamp);
  printf("\n");
  PRINT_MBLK(arg0)
}

This is actually an excerpt from my (currently internal) KSSL debugging suite.
An example of output from such probe can be seen in my Coloring dtrace output post.

For more complex projects it would be waste to stop here. Prefixes could be further structured. However, this has some drawbacks. In particular, I was thinking about having kssl_mblk- and kssl_err- prefixes. Now what to do for places where an error condition occurred _and_ we would like to see the associated mblk ? Using something like kssl_mblk_err-* comes to ones mind. However, there is a problem with that - what about the singleton cases (only mblk, only err). Sure, using multiple wildcards in dtrace is possible (e.g. syscall::*read*:) but this will make it ugly and complicated given the number of mblk+err cases (it's probably safe to assume that the number of such cases will be low). Simply, it's not worth the hassle. Rather, I went with 2 probes.
To conclude, using structured prefixes is highly beneficial only for set of probes where categories/sub-prefixes create non-intersecting sets (e.g. data type and debug level).

Of course, all of the above is not valid only for kernel but also for custom userland probes !

[1] High-level description of KSSL can be found in blueprint 819-5782.

When my friend from Usti nad Labem (9th biggest city in Czech republic) asked me to present about OpenSolaris at local Linux conference, I got hooked. First, Usti is interesting city (the city is surrounded by beautiful countryside, yet has chemical plants in the vicinity of city center) and I haven't been there for a long time and second, having the opportunity to present about OpenSolaris technologies to Linux folks is unique.

When we (Sasha agreed to go present with me) arrived to Usti we were greeted by slightly apocalyptic weather (see pictures below). The environment where all the presentations took place compensated that, fortunately.

40 minutes to present about something which most people in the room are not very aware of is challenging. The fact that OpenSolaris is open source and it is a home for several disruptive technologies makes that both easier and harder. We have greatly leveraged Bryan Cantrill's Dtrace review video taped at Google for doing second part of the presentation where we demo'ed dtrace. I have even borrowed some of his quotes. I am pretty sure he wouldn't object since his presentations were perused in the past. To make the list of attributions complete, we have substantial material in the first part from Lukas' past presentations about OpenSolaris project.

It's much better to demo the technology than just talk about how great it is (I remember a funny moment in Bryan's presentation where a dtrace probe didn't "want" to fire where Bryan jokingly said "bad demo!" to the screen. I nearly fell off of my chair at that moment.). "So, I have finally seen dtrace in action !" was one of the great things to hear after the presentation.

The "OpenSolaris technologies" presentation can be downloaded here.

As you might know, Netcat implementation is going to be part of OpenSolaris. The initial Netcat integration is based on a reimplementation from OpenBSD (here's why).

As Jeff Bonwick said, open sourced code is nothing compared to the fact that all design discussions and decisions suddenly happen in the public (loosely paraphrased). This is a great wave to ride and I have jumped on it when it was not really small so I have at least posted the webrev pointer for initial Netcat integration (CR 4664622) to the opensolaris-code mailing list (which is roughly the equivalent of freebsd-hackers, openbsd-tech or similar mailing lists) to get some code review comments.

Since then couple of things changed. Thanks to Dan Price and others it's now possible to upload webrevs to cr.opensolaris.org. I have started using the service so the new and official place for the Netcat webrev is

cr.opensolaris.org/~vkotal/netcat-webrev/

The webrev has moved location but what I said in the opensolaris-code post still holds true:

Any constructive notes regarding the code are welcome. (I am mainly looking for functional/logic errors, packaging flaws or parts of code which could mismatch the PSARC case)

The following things could help any potential code reviewer:

• Summary of the changes done to the original implementation in order to adapt it to Solaris environment..
• PSARC 2007/389 case covering interfaces delivered by this project. For more information about ARCs see Architecture Process and Tools community pages.
• SUNWnetcat package (x86) which contains /usr/bin/nc binary
• webrev of the differences between my version of Netcat and the one which is currently in OpenBSD.
  Only the *.[ch] files matter, of course. (This is very easy thing to do with distributed SCM since it only requires one to reparent and regenerate webrev against new parent workspace)
• Updated manual page
  This is slightly different from the man page in the PSARC materials because it contains new section about using nc with privileges and associated set of examples in the EXAMPLES section. The man page in the PSARC materials will not be updated because after a case is approved, the man page is updated only in case some architectural changes were needed. In the case of privileges, it is only addition describing specific usage, no architectural changes.

The conclusion for non code reviewers ? I hope it is clear the in (Open)Solaris land we value quality and transparency. Peer reviews and architectural reviews are just (albeit crucial) pieces which help to achieve that goal.