Detecting who used the EC2 metadata server with BCC

Recently at work we had a minor incident involving exhaustion of the EC2 metadata server on some of our hosts. I was able to get enough detail to delegate the rest to a team to fix the issue.

AWS EC2 has this thing called the metadata server. As far as the user can tell it runs inside the hypervisor and is exposed directly to your host via an http server at You can use curl(1) to get basic info about your server, like what instance type it is, for example.

The metadata server also hosts per instance authentication data. Your hosts reach out to the metadata server, get some auth material, and use that auth material for each request that interacts with AWS.

This means that if you somehow exhaust the metadata server processes will not be able to authenticate with AWS. I haven’t dove in deeply to understand if it’s a rate limit or a concurrency limit, but I can say that in any case we ran into it.

🔗 Detecting the Bad Actor

Initially I was going to use tcpdump(1) to figure out what was happening, but as far as I know it does not expose process ids, and even if it did I suspect you’d have to do a dump per process.

My go to for “stuff lower level than strace(1) is BCC. BCC is a Linux-ish DTrace; and I say that in every sense. Just like Zones are a single, standalone thing in Solaris, and containers are a combination of two or more complex Linux features, BCC takes advantage of two or more compilers, kprobes, uprobes, and surely more. I don’t know everything there is to know about BCC, but generally speaking I don’t have to because there is a nice suite of tools to give you what you want.

I loaded up the tool listing, searched for tcp, and the second tool is tcpconnect. Here’s a basic example:

$ sudo /usr/share/bcc/tools/tcpconnect
PID    COMM         IP SADDR            DADDR            DPORT
15100  curl         4   80  
15110  curl         4    80  

In the actual incident though I only wanted, so I changed my command to:

$ sudo /usr/share/bcc/tools/tcpconnect | grep -F

I stopped getting any output at all, but from experience I know that’s because that tcpconnect is now buffering. Dominus recently had a blog post that discusses this, including solutions, so I tweaked the command to be:

$ sudo stdbuf -oL /usr/share/bcc/tools/tcpconnect |
       stdbuf -oL grep -F

I should figure out if I can just do something like exec stdbuf -oL $SHELL, but anyway the above works. So now the output will be something like this, printed as the connections are made:

15100  curl         4    80
15110  curl         4    80

🔗 Getting More Detail

This is great, but our processes set their name and the COMM field above truncates it. Side note: if you are running a fork based service, set your process name to something relevant. It’s really useful and basically free.

In Perl you can set it by simply doing $0 = "...".

My next step was to add a dash of Perl to grab the full process name. Here’s what I ended up with:

$ sudo stdbuf -oL /usr/share/bcc/tools/tcpconnect |
    stdbuf -oL grep -F |
    stdbuf -oL perl -pae'$F[1] = `cat /proc/$F[0]/cmdline`;
      $_ = join("\t", scalar(localtime), @F) . "\n"'

Perl’s -a flag makes it act a bit like awk(1), in that it tokenizes input on whitespace and populates @F with your data. So $F[0] is the pid, $F[1] becomes the untruncated name. I also added the timestamp. Here’s a (sanitized) example of the output:

Wed Jun 20 14:21:33 2018        perform-queued-tasks-manager send-welcome-email      4 80                          
Wed Jun 20 14:21:33 2018        perform-queued-tasks-manager send-welcome-email      4 80                          
Wed Jun 20 14:21:33 2018        perform-queued-tasks-manager send-welcome-email      4 80                                  
Wed Jun 20 14:21:33 2018        perform-queued-tasks-manager send-welcome-email      4 80                                  
Wed Jun 20 14:21:33 2018        perform-queued-tasks-manager send-welcome-email      4 80   

I’ve wanted to use low level Linux instrumentation in anger for years, and the fact that I did without thinking much about it is delightful. Thankfully I don’t need this kind of information very often, but having it available is great.

(The following includes affiliate links.)

I don’t think there is a book about BCC (yet.) I think the closest thing would be Brendan Gregg’s Systems Performance. It’s got a ton of detail and a good helping of methodology that will help with the kind of stuff that one tends to use BCC for.

BCC is very much implemented atop Linux, so it is worth knowing Linux and Unix if you ever need to do something more advanced than use an out-of-the-box tool. I suggest reading The Linux Programming Interface for that kind of information.

Posted Thu, Jun 21, 2018

If you're interested in being notified when new posts are published, you can subscribe here; you'll get an email once a week at the most.