Wednesday, September 27, 2017

Is a lot of spam our own damn faults?

I got an unsolicited sales inquiry from a major company the other day.  Each day, 10 to 20 junk emails make it through our spam filter.  Usually, I can delete them after only a second or two, but this one sounded like I might already have a business relationship with them.  I don't want to risk insulting a customer or vendor, so I responded, asking what it was about.  The salesman was honest; he said that he thought somebody with my title would be interested.  I wasn't.  Not even close.

I've been on the Internet since it became easy to get on it.  When did it become acceptable to send blind solicitations?  When did the word "spam" come to mean only Nigerian princes and phishing schemes?  It used to be only desperate, border-line ethical, fly-by-night companies that sent junk email.  Now it's Box, Oracle, Microsoft, hell, I'm pretty sure my own employer does it!  Why have mainstream companies sunk so low as to send solicitations based on title?

Think back (if you're old enough) 20 years.  There were trade magazines that you could get "for free".  All you had to do is fill out a sheet that indicated in fair detail what your interests were, what industry you worked in, and the kinds of products over which you have purchase influence.  Vendors got very precisely-targeted lists, and we all knew that we would be getting solicitations.  We valued the magazine, so we didn't resent the ads.  Heck, although I don't remember specifically, I suspect I responded positively to one or two solicitations; the advertiser got their money's worth and I got a product that I wanted.

Those magazines don't exist any more, or at least not in my field.  We've all stopped reading the paper versions and instead look to the web for the information we're interested in.  We subscribe to blogs,  podcasts, slash-dot, LinkedIn groups, and any number of other curated content providers.  But the Internet evolved from an early non-commercial birth.  Early adopters resented the commercialization of the Internet, and refused to give information about themselves.  We create throw-away email addresses to subscribe.  We want to remain anonymous.  So the information curators never established the model of "you tell me about yourself for marketing purposes, and I'll give you information you want."  Some companies tried to get that going, but the internet "culture" prevented it from catching on.

So guess what?  I and my fellow-junk-email-haters are suffering from the unintended consequences of our own behavior.  Vendors no longer have precisely-targeted lists available to them.  So they substitute quantity for quality; send a million emails, and you're sure to find some prospects.  It's the new normal.

Idealists like me want a total paradigm change.  We want unsolicited advertisements to go away completely.  Back in the day, if I knew I wanted a C compiler, what did I do?  Open the yellow pages?  Sorry, no entries in the Yellow Pages for C compilers.  No, I *depended* on those trade magazines' advertisers to give me access to vendors of C compilers.  But now that search engines exist, we can do away with outgoing advertisements.  Instead of push marketing, go with pull marketing.  If I want a C compiler, I won't open my "junk" folder to find an unsolicited ad, I'll do a web search.  And this model *does* work!  We put some useful information on our web site, and attracted more than one customer who came for that information and stayed for our product.

And yet, the realist in me knows that human nature is what it is.  Research has proven again and again that advertising works.  I suspect modern email campaigns generate a lot of "unsubscribe me" responses, some of which may be less than polite, but I also suspect that they generate at least some interest.  Cast a wide-enough net and you'll catch some fish.

So if I have an emotional response to junk mail that is out of proportion to it's actual cost to me, that's my problem, not the advertisers.  I guess I need to get over it.

Thursday, September 21, 2017

Solaris Multicast Deafness Bug

Once again, the mighty Dave Zabel (of two different fames) has found another Multicast-related bug, this time in Solaris.  I think that recent versions of Solaris fix it, and I don't have the energy to track down *when* they fixed it, but if you have Solaris servers that you haven't kept updated for a while, you might have this bug.


DEAFNESS DEMONSTRATED

You'll need Informatica's "mtools" package for Solaris.  These are great tools offered for free in both binary and source form at https://marketplace.informatica.com/solutions/informatica_mtools

And you'll need two hosts: A and B.  Host B should be Solaris 6.10 that hasn't been updated in a long time.  Host A can be anything.

1. On host A, run this:
    msend 239.0.3.13 12000 15

2. On host B, open two windows.  In the first, enter this:
    mdump 239.0.3.13 12000

Admire the printouts of the multicast packets for a while.  Isn't technology wonderful?  :-)

3. In a second host B window, enter:
    mdump 239.128.3.13 12000

Note that the first window continues to print, but the second window is silent.  No surprise; it is listening to a different and unused multicast group!  Of course it is silent.

4. Kill that second mdump.

WHOA!  The first mdump stops printing!  It went deaf to 239.0.3.13.  When trying this same experiment on Linux, or on our Solaris 5.11 machines, it does not go deaf.  But we have several old, non-updated 5.10 machines where the first mdump does go deaf on this step.

5. Enter:
    netstat -g

The OS still thinks it is listening to the multicast group.

6. Enter:
    snoop -P host 239.0.3.13

The packets are still being received!  But they aren't being delivered to the first mdump.

7. Enter:
    mdump 239.128.3.13 12000

WHOA!  The first mdump starts printing again!  The second mdump is still silent since there still isn't any traffic on its multicast group.


MAYBE IT'S MY PROBLEM?

Maybe PEBKAC?  Or a bug in mtools?

Nope.  Let's start over and try it again with a small change in step 3:

1. On host A, run this:
    msend 239.0.3.13 12000 15

2. On host B, open two windows.  In the first, enter this:
    mdump 239.0.3.13 12000

3. In a second host B window, enter:
    mdump 239.64.3.13 12000

See what I did there?  I changed the 128 to 64.  As before, the first window continues to print, but the second window is silent.

4. Kill that second mdump.

Lookie there!  The first mdump continues to print the messages.  No deafness.


WHAT'S GOING ON?

Well, I'm not sure, but I think it's got to be related to multicast group aliasing.  Remember that there are 2**28 different IP multicast groups.  But what about Ethernet?  There are only 2**23 Ethernet multicast MAC addresses allocated for use by IP multicast.  It turns out that 239.0.3.13 and 239.128.3.13 map to the same Ethernet multicast MAC address: 01-00-5E-00-03-13.

The IGMP protocol doesn't care about that; host B still tells the switch which multicast groups are subscribed, and it treats 239.0.3.13 and 239.128.3.13 as different.  But when the IP layer interfaces with Ethernet, it needs to program the NIC with the same multicast MAC address for those two IP groups.  And apparently older versions of Solaris didn't do the book keeping right.

I've tried this experiment on other OSes and they all work as you would expect (no deafness).  Our Solaris 5.11 machine does it right.  And even a recently-installed 5.10 system works right.  But older systems that haven't been updated in a while all have this problem.



THE MORAL OF THE STORY

The obvious moral is to update your systems.

But even then, you should avoid using multicast groups that alias on top of each other.  The whole point of multicast is that you don't receive packets that you aren't interested in.  But if you have traffic published to both 239.0.3.13 and 239.128.3.13,  a host subscribing to only one of them will get data for both.  The IP layer will do the right thing (discard the undesired packets), but it still produces an unnecessary load.


ANY OTHER GOTCHAS?

Sure.  Watch out for well-known and ad-hoc multicast protocols in the range 224.0.0.0 - 224.4.255.255.  Are any of those in use anywhere on your network?  No?  Are you sure they never will be?

Look at the multicast group we tested with: 239.0.3.13.  That aliases on top of 224.0.3.13, which is in an ad-hoc1 range labeled "RFE Generic Service".  I don't know what that is (and Google doesn't seem to know either), but I'm thinking I want to avoid aliasing, even if low probability.

You should be fine if you use multicast groups between 239.0.5.0 - 239.127.255.255.

Oh, and update your systems too.  Good hygiene and all that.


UPDATE: UPGRADING FIXES IT

We've upgraded one of our "problem servers" to the latest Solaris 5.11 and it fixed the deafness problem.

I'm not interested in figuring out exactly which minor release they fixed it in.

Saturday, July 8, 2017

Most Random Password Generators are Bad

Good for you!  You're taking the advice of experts and clicking "Generate Password", resulting in 10 characters of gibberish.  There!  Now your password will take thousands of years to crack.

Um ... not necessarily.  Try a few days.


RAND() LIMITS ENTROPY TO 32 BITS

When using the pseudo-random number generator supplied by most language libraries, the entropy of the resulting password is limited to 32 bits!

Let's take XKCD's algorithm: ~2000 word dictionary, randomly select 4 words, produces 2000**4 different possible passwords, which is 16 trillion.  Log base 2 gives 43.9 bits of entropy.

But using a pseudo-random number generator with a 32-bit initial seed means that it will only generate 2**32 different sequences, or 4 billion.  That's .027% of the total!  In other words, 99.97% of the possible XKCD-style passwords CANNOT BE GENERATED by that program!

Normally, you can add more bits of entropy by either expanding the dictionary size (the number of words to choose from), or increasing the number of words in the password.  But because of the pseudo-random number generator, you are STUCK at 32 bits of entropy.  An attacker could even pre-generate the 4 billion possible XKCD-style passwords that a standard Linux rand() produces.

My point is not that 32 bits bits of entropy aren't enough, it's that you aren't necessarily going to get what you think you're getting if you use the stock pseudo-random number generator.

So if you're running somebody's application and you click "generate random password" and you see a string of gibberish that claims to have a crack time of thousands of years, it is probably wrong.  32 bits of entropy at 1000 guesses per second has a brute force crack time of under 50 days.  (And modern crackers go MUCH faster than 1000/sec.)


TRULY RANDOM NUMBERS ELIMINATE THE LIMIT

For my program, I offer the "-r" option, which reaches out to https://random.org to get random numbers.  It doesn't need very many -- you only need 4 random numbers to generate an XKCD-style password -- but the important feature is that random.org is truly random.  There is no seed.  Each number is uniformly random and independent from the previous number.  (Or at least, so claims the owner of random.org.)  I'm pretty sure this removes any artificial limit on entropy, so you can get as much entropy as you want by increasing the dictionary size and/or the number of words in the password.

Using random.org is not the only way to solve this problem.  Have you ever generated an SSL certificate?  It can take several seconds while the software "generates" enough entropy for long key lengths.  I'm not personally familiar with how that is done, but I've heard the the OS uses external physical events, like keystrokes, network interrupts, etc.  I think I've heard that it also uses disk interrupts, which makes me wonder if SSD drives make it harder for kernels to generate entropy.

If you're going to be demanding a lot of entropy for your application, you should not abuse random.org.  Instead learn how to use locally-generated entropy.

The vast majority of on-line password generators are written in Javascript.  I'm not sure how to get truly random numbers (i.e. entropy) in Javascript, but this might be a good starting point.

(By the way, a bit of reading on my part shows me that I have a lot to learn.  But my reading to date does reinforce my primary point: simply using rand() or a similar/derived function does not produce passwords that take thousands of years to crack.  At best they rely on "security through obscurity".)


PASSWORD MANAGERS' RANDOMIZER???

I'm a little worried about the "random password generators" included in password managers.  The idea is that you should have a different password for every on-line account, and you let the password manager deal with the hundreds of passwords you end up with.  Since you don't need to remember, or even type those passwords, you might as well make them be random character gibberish.  Only the password manager's master password needs to be memorized.

However, if your password manager just uses the normal pseudo-random number generator in the system, that sequence of random characters will not have as much entropy as you think.  I can tell you that LastPass's online password generator just uses Javascript's get_random() function, which only has 32-bits of entropy.  Now maybe their laptop application uses /dev/random, but also maybe the fact that their on-line generator uses built-in random indicates they didn't give the issue much thought.

I haven't done an exhaustive search, but I would wager that 90% of "generate password" functions just use the language's default random number generator, which has a 32-bit seed (or less!).

My suggestion is to use https://www.random.org/passwords/ to generate your gibberish passwords, or my program to generate XKCD-style passwords.


SEED V.S. PERIOD

The rand() man page says that Linux rand() uses the same algorithm as random().  And srandom()'s man page says that the seed is an unsigned int, which is 32 bits.  It also says:
The period of this random number generator is very large,
approximately 16 * ((2^31) - 1).
I.e. the period is approximately 34 billion, which is about 35 bits.  But the seed is 32 bits.  This means that you cannot start the random number generator at any arbitrary point in its period.  Even if you figure out a way to fully-leverage all 35 bits of random()'s period, that still gives you a crack time of 397 days, at 1000 guesses/sec.  And by the way, modern password crackers go much faster than 1000/sec.

XKCD-style Password Generator

I got to thinking about passwords again today.

I wrote my own program to produce XKCD-style passwords from a list of 2126 common words, and calculated some stats.  I reproduced XKCD's calculation of 44 bits of entropy for 4 randomly-selected words.  And I made a few mildly-interesting discoveries, and one more-interesting realization.


PASSWORD LENGTH: MORE = BETTER?

My average XKCD-style password length is 20.8 letters (over a large sample), which is a lot of typing.  So I decided to limit word size.  By filtering my list of 2129 words to nothing longer than 4 letters, I ended up with 709 words.  That's not many, and 4 of them together only gives 37 bits of entropy.  Not so hot.  But if you string 5 of 709 words together, you get 47 bits of entropy, which is better than XKCD!  And the average password drops to 18.3 characters.

I find that interesting: shorter passwords which produce more bits of entropy than longer passwords.  Seems counter-intuitive, until you realize that opening it up to the full 2129 words increases average word length more than it increases entropy.  (See below for the math.)


THE PASSWORDS

So, what do these passwords look like?  Here's 10 of the XKCD-style: 4 words from the 2126 word set:

password: MostlyRelativeSpinAdvanced
length: 26
password: ForBasicallyThinkingExplain
length: 27
password: CookieArmyMysteryConference
length: 27
password: ExpectConvertQuarterbackPresentation
length: 36
password: EverybodyProductHotDemonstrate
length: 30
password: RockIndexWellFloat
length: 18
password: BehaviorNearlyPromotePercentage
length: 31
password: PocketSurviveFourLab
length: 20
password: MuchWeekWillAnd
length: 15
password: DivideMorePeakSeveral
length: 21


So, how easy are those to remember?  Memorizing an XKCD-style password is about creating a mental picture or story around it.  Use some imagination.  It usually helps to make it amusing.  How about the first one: "MostlyRelativeSpinAdvanced"?  Well, I'm a bit of a science geek, so this one makes perfect sense.  You have a particle stream, and most of the particles are moving at relativistic speeds.  So measuring each particle's spin is a pretty advanced thing to do.  Hmm ... what's the amusing part of that?  Oh maybe that an actual physicist would roll his eyes at my explanation and say that I don't know the first thing about particle physics.  But basically, I was able to imagine a mini-story or mental picture for each of those passwords, so while I might not be able to memorize all 10, I could easily memorize one of them.

What about the shorter passwords consisting of 5 words from the 709 words of 4 letters or less?

password: BuyWingSadRideSeed
length: 18
password: SeedPairTankJailDo
length: 18
password: PanBuryDenyDataOld
length: 18
password: GeneRiceTeaYetSin
length: 17
password: WallJailLabNextTent
length: 19
password: HallSnapCashRichRead
length: 20
password: WarmUsKeepRoseLess
length: 18
password: PortMarkSirYouLeaf
length: 18
password: HiAgoHipAnyBe
length: 13
password: EaseSkyRealTossFate
length: 19


Even though those are shorter (and more secure) passwords, I guess I find them more difficult to remember them.  It's about creating a mental picture or story around those words.  Since the words are random, they don't come out in any conceptually correlated way.  So you stretch your imagination to encompass them.  The more words in the password, the more you have to stretch.

Take the last password up there, "EaseSkyRealTossFate", and drop that last word to make it 4 words: "Ease sky real toss".  My first thought is that "toss" is the children's game "ring toss".  Sky and ease kind of fit since the game is usually pretty easy and you toss things towards the sky.  The word "real" is kind of left out, but I imagine throwing something "real", like a laptop or a dinner plate, instead of a game piece.  So I imagine the ease of tossing a real laptop into the sky.  Yeah, that's stretching the imagination a bit, but maybe not too much.  I could pretty easily remember EaseSkyRealToss.

But now throw "fate" in there and my whole mental picture falls apart.  I guess I could say that when the laptop lands, its fate will be sealed, but ... not sure why ... but I would have much more trouble remembering it.

So I'll be sticking to 4 words and more typing.


THE MATH

Passwords are basically taking a set of N things, and taking L of them out with replacement.  For example, a 4-digit PIN consists of a set of 10 digits (N=10), and you take 4 digits out (L=4) with replacement.  The "with replacement" simply means that you might take the same digit out more than once (e.g. 2338).  So the entropy of a 4-digit PIN is 10**4 *(10 to the power of 4), which is 10,000.  To get that in terms of bits, take the log base 2 of it to get 13.  So 13 bits of entropy.

Another example: 8 randomly-selected letters for a password.  Let's assume lower-case only, and no digits or special characters.  The set of N things is the letters of the alphabet, so N=26.  By taking 8 characters, L=8.  26**8 = 208 billion.  Log base 2 of that is 37 bits of entropy.  Cool.  Now let's do random upper/lower case.  N=52, and 52**8 = 53 trillion, giving 45.6 bits of entropy.  Add in 0-9: N=62, 62**8 = 218 trillion, giving 47.6 bits of entropy.

So, back to my XKCD-style passwords.  My original set of 2126 words, taking 4 at a time, gives 2126**4 = 20 trillion, which is 44.2 bits.  My reduced set of 709 short words, taking 5 at a time gives 709**5 = 179 trillion, which is 47.3 bits.

However, see my next blog post for an observation about random password generators and entropy.


MORE WORDS?

My list of 2126 words actually comes from a list of 3000 words from Education First.  I filtered it to limit word length to 7 or fewer characters, resulting in my 2126 words.  Note to the rigorous: you'll find that I'm 2 words short; it made my code easier to ignore the first and last words.

So how about if I remove that filter and pick 4 words from the entire set of 2998?

2998 ** 4 = 80 trillion, which is 46 bits.  I.e. going from 2126 words to 2998 increases the entropy by 2 bits.  My average password length jumps to 25.6, which is 5 more characters.  I tried a few other word length limits and decided that 7 is best.


THE PROGRAM

See https://github.com/fordsfords/pgen Be sure to use "-r" if generating an actual password you want to use.

Monday, May 22, 2017

Some multicast programming tips

Never too old to learn.  :-)

There are lots of multicast example programs out there, so I won't try to compete with them.  But I did run across several things that weren't explained very well.


Single Socket, Multiple Groups

Yes, you can create a single socket and have it receive datagrams from multiple multicast groups.  Just include multiple calls to:
  setsockopt(recv_sock, IPPROTO_IP, IP_ADD_MEMBERSHIP, ...


Multiple Sockets, One Group per Socket

This is another common use case, where you create multiple sockets for receiving, with each socket joined to a different multicast group.


Binding the Receive Socket

Since a socket needs to be bound to a port to receive any kind of UDP datagram, multicast or unicast, you need to include a call to bind().  You pass in a sockaddr_in with the sin_port set as desired (remember to pass it in network order).  But what about the sin_addr?  What do you set that to?

Many people set it to INADDR_ANY, which is what I did in a recent program.  But in the multiple sockets, one group per socket case, it had an unexpected side effect.  All of my sockets were bound to the same destination port, but joined to different multicast groups.  With sin_addr set to INADDR_ANY, the kernel replicated the received datagrams and delivered a copy to *every* socket! I.e. simply doing the IP_ADD_MEMBERSHIP didn't do any filtering.  When a multicast datagram was received, the kernel just used the destination port and delivered a copy to every socket.

I had to do some extra searching to find out that you can set the sin_addr to the multicast group.  I have some reason to suspect that this is not portable across all operating systems, but at least it works on Linux.  Now I can have 10 sockets, each bound to the same port (don't forget SO_REUSEADDR) but different multicast groups.  When a multicast datagram is received, it is delivered *only* to the socket which is bound to the right port/multicast group pair.


Single Socket, Multiple Groups, reprise

So, what about the case where you have a single socket joined to multiple groups?  In that case, you *do* want to use INADDR_ANY in the bind.


Mix and Match?

I guess this poses a restriction.  You can't have, say, 2 sockets that you distribute 4 multicast groups across, with two groups each.  Why would you want to do that?  Maybe to load-balance across threads.  But assuming they all want to bind to the same port, you can't do it.  Setting the sin_addr to INADDR_ANY will mean that both sockets will receive a copy of each datagram sent. But you can't set sin_addr to multiple multicast groups.

So if you want to have multiple sockets, you need to have one group per socket, and bind that socket to the group.

Monday, May 15, 2017

WannaCrypt / WannaCry ransomeware

I'm not a security researcher, and I don't follow the subject very closely.  But here is an interesting read by the person who slowed the spread of the recent WannaCrypt / WannaCry ransomware outbreak.

https://www.malwaretech.com/2017/05/how-to-accidentally-stop-a-global-cyber-attacks.html

Sunday, April 30, 2017

Fraudulent spam email claiming to be Netflix

I got a phishing email.  So what?  I get lots of phishing emails.  Why blog about this one?

Well, it's at least a *little* different.

Most of them direct the victim to an existing web site which has been compromised.  I.e. the web site's real owner has no idea that his own site is being used for fraudulent purposes.

In this one, the victim is directed to the domain name "netflix-myaccount.com", which the scammer obtained properly.  Unfortunately, the scammer wasn't stupid enough to include his own contact information in the registry, instead choosing to hide behind privacyprotect.org.

Now there's nothing wrong with using privacyprotect.org to hide one's identity.  If anything, it removed any doubt in my mind (as if there were any) that the page isn't owned by Netflix.  So it reinforced that it is a phishing site.  I sent a complaint email to privacyprotect.org anyway.

Next up, the domain the registry: ilovewww.com.  Never heard of them.  Malaysian.  Sent them a complaint email too to suspend the registration.

Next, the IP address that netflix-myaccount.com resolves to: 80.82.67.155.  A whois lookup shows the block is owned by Quasi Networks LTD.  Abuse email to it as well.

Now to another nice site: phishcheck.me, a site that evaluates how likely a site is to be fraudulent.  It actually goes to the site and analyzes it.  So I went there and plugged in "http://netflix-myaccount.com", and sure enough, it says that it is probably a phishing site (no surprise there).  But on that phishcheck.me page is a tab named "resources", which shows details of the access to the site ... and well lookie there, "netflix-myaccount.com" redirects to "netflix-secureserver.com".  Which resolves to the same IP as "netflix-myaccount.com", and is registered in the same ways (ilovewww.com and privacyprotect.org).  So what the point in that?  Oh well, another set of complaint emails for the new domain name.

Finally, let's see if it is a compromised web site.  I would like to see what other domain names resolve to the same IP address.  Unfortunately, this appears not to be an exact science.  The few sites there are that claim to do this find *no* domains resolving to that IP.  However, a simple google search for "80.82.67.155" (*with* the double quotes) does find the names "netflix-myaccount.com" and a new one: "www.useraccountvalidation-apple.com".

Yep.  Another phishing site, leveraging Apple instead of Netflix.  Let's do the drill, starting with whois.  WHOA!!!  Did we hit paydirt?

Registrant Contact
Name: Jamie Wilson
Organization:
Mailing Address: 22 Madisson Road, London London SE12 8DH GB
Phone: +44.07873394485
Ext:
Fax:
Fax Ext:
Email:uktradergb@gmail.com

Now, don't be too hasty.  The *real* registrant is a scammer.  What are the chances he would list his own real contact info?  The only thing that might be valid is the email address, since I think he needs that to fully set up the domain, and even then it might have been a single-use throwaway.

Hmm ... not totally throw-away.  A google of "uktradergb@gmail.com" has 6 hits, including "netflix-iduser1.com" and "netflix-iduser2.com", both of which have Jamie as the registrant, but neither of which resolve to valid IP addresses.  So not sure there's anything actionable (i.e. complainable) there.

But just in case, I googled the phone number, and found this additional hit: "AppleId1-Cgi.com", which doesn't appear to resolve to a valid IP.

Well, much as I hate to, let's skate over to "domaintools.com", which wants my money in a bad way.  It tells me that uktradergb@gmail.com is associated with ~38 domains, but of course won't tell me what any of them are without paying them $99.  And even though I would love to send complaints regarding all 38, I wouldn't love it $99 worth.

Ok, one more thing.  http://domainbigdata.com/gmail.com/mj/LX7iN6iKwKFIRfkD7CsKXQ says that the owner of that email address is Adam Stormont, and that the email is associated with a few other sites (but not 37), including "hmrc-refundvalidation.com", which doesn't resolve to an IP.  And by the way, a whois of another uktradergb@gmail.com domain, "hni-4.com", says that the registrant is David Hassleman.  So yeah, ignore the Jamie Wilson contact.  He wasn't that stupid.  :-)

And now I've run out of gas.  Maybe those domain names will be disabled in the next few days.  Or maybe I've just wasted a half hour of my life.  (Well, I've learned a few things, so not totally wasted.)

Friday, March 31, 2017

Cisco Eating Multicast Fragments???


UPDATE: after upgrading the IOS our "MDF" switch, this problem went away.  None of my readers (all 2 of them?) have reported seeing this problem with their switches.  So I think this issue is closed.


I think we've discovered a bug in our Cisco switch related to UDP multicast and IP fragmentation.  Dave Zabel (of Windows corrupting UDP fame) did the initial detective work, and I did most of the analysis.  And I'm not quite ready to declare victory yet, but I'm pretty sure we know roughly what is going on.


BOTTOM LINE:

It appears that Cisco is not paying proper attention to whether a packet is fragmented when checking the UDP destination port for the BFD protocol.  The result is that it eats user packets that it misidentifies as being part of that protocol.


THE SETUP:

We have 4 Catalyst 3560 "LAB" switches (48 port) trunked to a Catalyst 4507 "MDF" switch.  Our lab test machines are distributed across the LAB switches.

Our messaging software multicasts UDP datagrams.  One of our regression tests involves sending messages of varying sizes with randomized data.  We saw that occasionally, one of the messages would be lost.  Doing packet captures showed that the missing datagram is NAKed and retransmitted multiple times, but the subscribing host never saw the datagram, even though it saw all the previous and subsequent datagrams.  (This particular test does not send at a particularly stressful rate.)

Further investigation showed that some hosts always got the message in question, while others never got the message.  Turns out that the hosts that got the message were on the same LAB switch as the sender.  The hosts that didn't get the message were on a different switch.

I narrowed it down to a minimal test datagram of 1476 bytes.  The first 1474 bytes can be any arbitrary values, but the last two bytes had to be either "0e c8" or "0e c9".  Any datagram with either of those two problematic byte pairs at that offset will be lost.  Note that the datagram will be split into 2 packets (IP fragments) by the sending host's IP stack.  Strategically placed tcpdumps indicated that the first IP fragment always makes it to the receiver, but the second one seems to be eaten by our "MDF" switch.

There's nothing magic about the size 1476 - it can be larger and the problem still happens.  1476 is just the smallest datagram which demonstrates the problem.


IP FRAGMENTATION:

IP fragmentation happens when UDP hands to IP a datagram that doesn't fit into a single MTU-sized Ethernet packet (1500 bytes).  A UDP datagram consists of an 8-byte header, followed by up to 65,527 bytes of UDP payload.  IP splits a large datagram up into fragments of 1480 bytes each and prepends its own 20-byte IP header to each fragment.  But note that only the first fragment will contain the UDP header.  So IP fragment #1 will hold the 8-byte UDP header and the first 1472 bytes of my datagram.

Since my test datagram is 1476 bytes long, IP fragment #2 will contain a 20-byte IP header followed by the last 4 bytes of my datagram.

I won't show you the first fragment of my test datagram because it's long and boring.  And it is successfully handled by Cisco, so it's also not relevant.

Here's a tcpdump of the second fragment of my test datagram (test datagram bytes highlighted).  Note that tcpdump includes a 14-byte Ethernet header in front of the 20-byte IP header, then the last 4 bytes of my test datagram, and finally 22 padding nulls to make up a minimum-size packet (those nulls are not counted as part of the IP payload).

07:56:38.518614 00:1e:c9:4e:a1:92 (oui Unknown) > 01:00:5e:65:03:01 (oui Unknown), ethertype IPv4 (0x0800), length 60: (tos 0x0, ttl   2, id 2132, offset 1480, flags [none], proto: UDP (17), length: 24) 10.29.3.88 > 239.101.3.1: udp
        0x0000:  0100 5e65 0301 001e c94e a192 0800 4500  ..^e.....N....E.
        0x0010:  0018 0854 00b9 0211 afed 0a1d 0358 ef65  ...T.........X.e
        0x0020:  0301 0000 0ec8 0000 0000 0000 0000 0000  ................
        0x0030:  0000 0000 0000 0000 0000 0000            ............

This is the packet which is successfully received by hosts on the same switch as the sender, but is never received by hosts on a different switch.  Change the "0e c8" byte pair to, for example, "1e c8" or "0e c7" and everything works fine - the packet is properly forwarded.


A CASE OF MISTAKEN IDENTITY?

In my problematic datagram, the last 4 bytes occupy the same packet position in fragment #2 as the UDP header in a non-fragmented packet.  In particular, the byte pair "0e c8" occupies the same packet position as the UDP destination port in a non-fragmented packet.  Those byte values correspond to port 3784, which is used by the BFD protocol.  BFD is used to quickly detect failures in the path between adjacent forwarding switches and routers, so it is of special interest to our switches.  (The other problematic byte pair "0e c9" corresponds to port 3785, which is also used by BFD.)

So, when a LAB switch sends fragment #2 to the MDF, it looks like MDF is checking the UDP port WITHOUT looking at the IP header's "Fragment Offset" field.  It should only look for UDP port if the fragment offset is zero.  Here's that packet again with the fragment offset highlighted:

07:56:38.518614 00:1e:c9:4e:a1:92 (oui Unknown) > 01:00:5e:65:03:01 (oui Unknown), ethertype IPv4 (0x0800), length 60: (tos 0x0, ttl   2, id 2132, offset 1480, flags [none], proto: UDP (17), length: 24) 10.29.3.88 > 239.101.3.1: udp
        0x0000:  0100 5e65 0301 001e c94e a192 0800 4500  ..^e.....N....E.
        0x0010:  0018 0854 00b9 0211 afed 0a1d 0358 ef65  ...T.........X.e
        0x0020:  0301 0000 0ec8 0000 0000 0000 0000 0000  ................
        0x0030:  0000 0000 0000 0000 0000 0000            ............

For most (non-fragmented) packets, that byte will be zero, and the UDP header will be present, in which case the 0ec8 would be the port number.  The highlighted fragment offset of b9 hex is 185 decimal, and IP fragment offset is measured in units of 8-byte blocks, so the actual offset is 8*185=1480, which is tcpdump has for "offset".

It also seems strange to me that the switch ignores which multicast group I'm sending to.  I can send to any valid multicast group, and the problematic packet will be eaten by the "MDF" switch.  Shouldn't there be a specific multicast group for BFD?  Maybe I found 2 bugs?

My employer has a support contract with Cisco, and I'm working with the internal network group to get a Cisco ticket opened.  I'll update as I learn more, but it's slow climbing through the various levels of internal and external tech support, each one of whom starts out with, "are you sure it's plugged in?"  It may take weeks to find somebody who even knows what IP fragmentation is.


TRY IT YOURSELF

I would love to hear from others who can try this out on their own networks.  Grab the source files:


To build on Linux do:
gcc -o msend msend.c
gcc -o mdump mdump.c

Note that I've tried other operating systems (Widows and Solaris), with the same test results.  This is not an OS issue.

For this test, the main purpose of mdump is to get the host to join the multicast group.

Choose three hosts: A, B, and C.  Make sure A and B are on the same switch, and C is on a different switch.  In my case, all three hosts are on the same VLAN; I don't know if that is significant.  For this example, let's assume that the three hosts' IP addresses are 10.29.1.1, 10.29.1.2, and 10.29.1.3 respectively, and that all NICs are named "eth0".

Choose a multicast group and UDP port that aren't being used in your network.  I chose 239.101.3.1 and 12000.  I've tried others as well, with the same test results.

Note that the msend and mdump commands require you to put the hosts's primary IP address as the 3rd command-line parameter.  This is because multicast needs to be told explicitly which interface to use (normal IP routing doesn't know the "right" interface to use).

Open a window to A, and two windows each for B and C.  Enter the following commands:

B1: ./mdump  239.101.3.1 12000 10.29.1.2

B2: tcpdump -i eth0 -s2000 -vvv -XX -e host 239.101.3.1

C1: ./mdump  239.101.3.1 12000 10.29.1.3

C2: tcpdump -i eth0 -s2000 -vvv -XX -e host 239.101.3.1

A: ./msend 239.101.3.1 12000 10.29.1.1

The "msend" command sends two datagrams.  The first one is small and gives the sending host's name.  The second one is the 1476-byte datagram, whose second fragment gets eaten by the Cisco "MDF" switch.

Window B1 should show both datagrams fully received.

B2 should show 3 packets:
1. The short packet with the host name.
2. Fragment #1 of the long packet
3. Fragment #2 of the long packet

C1 should only show the first datagram.

C2 should show 2 packets:
1. The short packet with the host name.
2. Fragment #1 of the long packet.

Fragment #2 is missing from C2, presumably eaten by the "MDF" switch.

Note that the two "tcpdump" windows might show additional packets, which are for the "igmp" protocol, and are unrelated to the test.  If I had more time, I would figure out how to get "tcpdump" to ignore them.