Category Archives: Tutorials and Guides

tstats: afterburners for your Splunk threat hunting

Recently, @da_667 posted an excellent introduction to threat hunting in Splunk. The information in Sysmon EID 1 and Windows EID 4688 process execution events is invaluable for this task. Depending on your environment, however, you might find these searches frustratingly slow, especially if you are trying to look at a large time window. You may also have noticed that although these logs concern the same underlying event, you are using two different searches to find the same thing. Is there anything we can do to improve our searches? Spoiler: yes!

One of the biggest advantages Splunk grants is in the way it turns the traditional model of indexing SIEM events on its head. Instead of parsing all the fields from every event as they arrive for insertion into a behemoth of a SQL database, they decided it was far more efficient to just sort them by the originating host, source type, and time, and extract everything else on the fly when you search. It’s a superb model, but does come with some drawbacks. Some searches that might have been fast in a database are not so rapid here. Again, because there is no database, limitations on the number of fields you can manage are removed – but there is nothing to keep fields with similar data having the same name, so every type of data has its own naming conventions.

The answer to both of these problems is datamodels, and in particular, Splunk’s Common Information Model (CIM). Datamodels allow you to define a schema to address similar events across diverse sources. For example, instead of searching

index=wineventlog EventCode=4688 New_Process_Name="*powershell.exe" 

and

index=sysmon EventCode=1 Image=*powershell.exe 

separately, you can search the Endpoint.Processes datamodel for process_name=powershell.exe and get results for both. The CIM is a set of predefined datamodels for, as the name implies, types of information that are common. Once you have defined a datamodel and mapped a sourcetype to it, you can “accelerate” it, which generates indexes of the fields in the model. This process carries a storage, CPU and RAM cost and is not on by default, so you need to understand the implications before enabling it.

Turn it up to 11

Let’s take this example, based on the fourth search in @da_667’s blog. In my (very limited) data set, according to the Job Inspector, it took 10.232 seconds to search 30 days’ worth of data. That’s not so bad, but I only have a few thousand events here, and you might be searching millions, or tens of millions – or more!

Splunk Job Inspector showing search time and cost breakdown

What happens if we try searching an accelerated datamodel instead? Is there much of a difference?

Splunk Job Inspector information for accelerated datamodel search

Holy shitballs yes it can. This search returned in 0.038 seconds, that’s nearly 270x faster! What sorcery is this? Well, the command used was:

| tstats summariesonly=true count from datamodel=Endpoint.Processes where Processes.process_name=powershell.exe by Processes.process_path Processes.process

What’s going on in this command? First of all, instead of going to a Splunk index and running all events that match the time range through filters to find “*.powershell.exe“, my tstats command is telling it to search just the tsidx files – the accelerated indexes mentioned earlier – related to the Endpoint datamodel. Part of the indexing operation has broken out the process name in to a separate field, so we can search for an explicit name rather than wildcarding the path.

The statistics argument count and the by clause work similarly to the traditional stats command, but you will note that the search specifies Processes.process_name – a quirk of the structure of data models means that where you are searching a subset of a datamodel (a dataset in Splunk parlance), you need to specify your search in the form

datamodel=DatamodelName[.DatasetName] where [DatasetName.]field_name=somevalue by [DatasetName.]field2_name [DatasetName.]field3_name

The DatasetName components are not always needed – it depends whether you’re searching fields that are part of the root datamodel or not (it took me ages to get the hang of this so please don’t feel stupid if you’re struggling with it).

Filtered, like my coffee

Just as with the Hurricane Labs blog, options for filtering and manipulating tstats output can be managed with the same operations.

| tstats summariesonly=true count from datamodel=Endpoint.Processes where Processes.process_name=powershell.exe NOT Processes.parent_process_name IN ("code.exe", "officeclicktorun.exe") by Processes.process_path Processes.process | `drop_dm_object_name("Processes")`

You can filter on any of the fields present in the data model, and also by time, and the original index and sourcetype. The resulting data can be piped to whatever other manipulation/visualisation commands you want, which is particularly handy for charts and other dashboard features – your dashboards will be vastly sped up if you can base them on tstats searches.

You’ll also note the macro drop_dm_object_name – this reformats the field names to exclude the Processes prefix, which is handy when you want to manipulate the data further as it makes the field names simpler to reference.

A need for speed

How do I get me some of this sweet, sweet acceleration I hear you ask? The first thing to understand is that it needs to be done carefully. You will see an increase in CPU and I/O on your indexers and search heads. This is because the method involves the search head running background searches that populate the index. There will be a noticeable increase in storage use, with the amount depending on the summary range (i.e. time period covered by detailed indexing) and how busy your data sources are.

With this in mind, you can start looking at the Common Information Model app and the documentation on accelerating data models. I highly recommend consulting Splunk’s Professional Services before forging ahead, unless your admins are particularly experienced. The basic process is as follows:

  • Ensure that your sourcetypes are CIM compliant. For most Splunk-supported apps, this is already done.
  • Ensure that you have sufficient resources to handle the increased load
  • Deploy the CIM app
  • Enable acceleration for the desired datamodels, and specify the indexes to be included (blank = all indexes. Inefficient – do not do this)
  • Wait for the summary indexes to build – you can view progress in Settings > Data models
  • Start your glorious tstats journey
Configuration for Endpoint datamodel in Splunk CIM app
Detail from Settings > Data models

Datamodels are hugely powerful and if you skim through the documentation you will see they can be applied to far more than just process execution. You can gather all of your IDS platforms under one roof, no matter the vendor. Get email logs from both Exchange and another platform? No problem! One search for all your email! One search for all your proxy logs, inbound and outbound! Endless possibilities are yours.

One search to rule them all, one search to find them… happy Splunking!

Collecting Netscaler web logs

A little while ago I wrote about collecting AppFlow output from a Citrix Netscaler and turning it into Apache-style access logs. Whilst that might technically work, there are a few drawbacks – first and foremost that Logstash gobbles CPU cycles like nobody’s business.

Furthermore, since the Netscaler outputs separate AppFlow records for request and response, if you want a normal reverse proxy log, you need to put them back together yourself. Although I have already described how to achieve that, as you can see above it is also not terribly efficient. So, is there a better way? There certainly is!

NetScaler Web Log Client

In order to deliver responses to requests correctly, the Netscaler must track the state of connections internally. Instead of creating our own Frankenstein’s Monster of a state machine to reassemble request and response from AppFlow, it would be much simpler if we could get everything from a place that already has the combined state. The good news is that Citrix have provided a client application to do just that. The bad news is that their documentation is a little on the shonky side, and it isn’t always clear what they mean. To fill in some of the gaps, I have written a brief guide to getting it running on CentOS 7. I will assume for this that you have installed CentOS 7 Minimal and updated it through yum.

Obtain the client

Citrix’s description of where to find the client on their site isn’t terribly helpful. Here’s how to get there at the time of writing:

    Citrix Portal > Downloads > Citrix Netscaler ADC > Firmware > [your version] > Weblog Clients

Prep the Netscaler

Ensure Web logging is turned on

    System > Settings > Configure Advanced Features > Web Logging

Ensure remote authentication is OFF for the nsroot user (not expecting many people to encounter this problem but it’s not easy to troubleshoot – the client just shows an authentication failure even if you entered the password correctly)

    System > User Administration > Users > nsroot > Enable External Authentication

Install and configure the NSWL client

Extract the .rpm from the zip downloaded from the Citrix portal and transfer it to your CentOS system. Run the following commands as root:

    $> yum install glibc.i686
    $> rpm -i nswl_linux-[citrix_version].rpm

You need to be able to connect from the system you are running the client on to your Netscaler reverse proxy on port 3011.

    $> nc -v [netscaler_ip] 3011

Add the target IP and nsroot account credentials to the config file as described in the Citrix docs (yes, some of their instructions are accurate – just not everything):

    $> /usr/local/netscaler/bin/nswl -addns -f /usr/local/netscaler/etc/log.conf

Edit the config file to set the format, log output directory, rotation settings etc.

----extract from /usr/local/netscaler/etc/log.conf----
logFormat    NCSA %h %v %l %u %p [%t] "%r" %s %j %J %{ms}T "%{referer}i" "%{user-agent}i"
logInterval			Daily
logFileSizeLimit		1024
logFilenameFormat		/var/log/netscaler/nswl-%{%y%m%d}t.log
------------------------------------------------------

Note: Citrix do not appear to provide a complete breakdown of what format strings are accepted, so I used the Apache documentation as a reference. However, not all of the variables are supported by the NSWL client, and some work in a different manner than expected. For example, %D does not output microseconds, but the %{UNIT}T style does work.

Configure a service to run the NSWL client

    $> vim /etc/systemd/system/nswl.service

[Unit]
Description=nswl

[Service]
Type=simple
User=nswl
Group=nswl
ExecStart=/usr/local/netscaler/bin/nswl -start -f /usr/local/netscaler/etc/log.conf	

[Install]
WantedBy=multi-user.target

    $> useradd -d <log directory> -s /sbin/nologin nswl
    $> chown -R nswl:nswl <log directory>
    $> systemctl daemon-reload
    $> service nswl start

SIEM configuration and log rotation

The logFormat directive shown above is similar to the standard Apache Combined format, but not identical. To parse the output, a slightly tweaked version of the regex is necessary:

^(?<src_ip>\S+) (?<site>\S+) (?:-|(?<ident>\S+)) (?:-|(?<user>\S+)) (?<dest_port>\d+) \[[^\]]*] "(?<request>[^"]+)" (?<status>\d+) (?<request_bytes>\d+) (?<response_bytes>\d+) (?<response_time>\d+) "(?:-|(?<http_referer>[^"]*))" "(?:-|(?<http_user_agent>.*))"

You should use a prefix pattern to match files to collect – do NOT use a suffix pattern like ‘*.<extension>‘ to track files. The NSWL client creates a new file with ‘.<number>‘ appended under many circumstances, including when the service is restarted, when the logFileSizeLimit is reached, and others. For example, if the service was restarted while writing to ‘nswl-20191001.log‘, it would begin writing ‘nswl-20191001.log.0‘.

Make sure to take this into account when configuring log rotation – e.g. move the files before compressing: ‘$> gzip nswl-20191001.log‘ results in ‘nswl-20191001.log.gz‘, which matches the pattern ‘nswl-*‘; SIEM agents may consider the latter file to be new and index it again, resulting in duplicate data.

Results

Using 1% CPU and a single process as opposed to the previous method of attempting to melt a CPU into the motherboard substrate is a definite improvement. Another plus is that it’s an officially supported tool, so in theory if something’s not working you can actually get some help with it.

I’m pretty proud of my eldritch horror of a python script, it ran for nearly two years in production with no significant problems (unlike Logstash which needed CPR every 6 weeks or so), but it’s high time my code was retired.

Via squid and second host running tor

Intercepting SSL with squid proxy and routing to tor

There was a time when practically all malware communicated with its command and control (C2) servers unencrypted. Those days are long gone, and now much of what we would wish to see is hidden under HTTPS.  What are we to do if we want to know what is going on within that traffic?

Introduction

(for those who are unfamiliar with the HTTPS protocol and public key encryption)

The foundation of HTTPS is the Public Key Infrastructure. When traffic is to be encrypted, the destination server provides a public key with which to encrypt a message. Only that server, which is in possession of the linked private key, can decrypt the message. Public key, or asymmetric encryption, is relatively slow so instead of all traffic being secured with this, the client and server use this stage only to negotiate a new key in secret for a symmetrically encrypted connection. If we wish to be able to read the traffic, we need to obtain the symmetric encryption key.

How can this be achieved? If we are in a position to intercept the traffic, we could provide a public key that we are in control of to the client, and establish our own connection to the server. The traffic would be decrypted at our interception point with our key, and re-encrypted as we pass it to the server with the server’s key. However, because HTTPS must be able to keep information confidential, it has defences designed with this attack in mind. A key issued by a server is normally provided along with the means to verify that it is genuine, not falsified as we wish to do. The key is accompanied by a cryptographic signature from a Certificate Authority (CA), and computers and other devices using HTTPS to communicate hold a list of CAs which are considered trustworthy and authorised to verify that keys are valid. Comparing the signature against the client’s stored list enables the client to verify the authenticity of the public key.

If we wish to inspect encrypted communication, we must both intercept the secret key during the exchange, and convince the client that the certificate it receives is genuine. This post will walk through the process needed to achieve those two goals.

Design

Starting point

I have already been running a sandbox that routes traffic via tor. It is loosely based on Sean Whalen’s Cuckoo guide, and implements the tor routing without going via privoxy, as shown below.

Initial setup

Using this method allows me to run malware without revealing the public IP of my lab environment. It has certain drawbacks; some malware will recognise that it is being routed via tor and stop functioning, however the tradeoff is acceptable to me.

squid | tor

Using squid with tor comes with some caveats that make the eventual configuration a little complicated. The version of squid I am using (3.5.23) cannot directly connect to a tor process running on the local host. In order to route via tor locally you will need a parent cache peer to which the connection can be forwarded. Privoxy is capable of serving this purpose, so initially I attempted the setup shown below:

Via privoxy

This configuration will function just fine if all you want is to proxy via squid. Unfortunately, this version of squid does not support SSL/TLS interception when a parent cache is being used. So, since we cannot use privoxy, and squid cannot route to tor on the same host, what can we do? Run tor on a different host!

Via squid and second host running tor

Implementation

squid with ssl intercept/ssl-bump

In order to use squid with ssl-bump, you must have compiled squid with the –with-openssl and –enable-ssl-crtd options. The default package on Debian is not compiled this way, so to save you some time I have provided the commands I used to compile it:

apt-get source squid
cd squid3-3.5.23/
./configure --build=x86_64-linux-gnu --prefix=/usr --includedir=${prefix}/include --mandir=${prefix}/share/man --infodir=${prefix}/share/info --sysconfdir=/etc --localstatedir=/var --libexecdir=${prefix}/lib/squid3 --srcdir=. --disable-maintainer-mode --disable-dependency-tracking --disable-silent-rules 'BUILDCXXFLAGS=-g -O2 -fdebug-prefix-map=/build/squid3-4PillG/squid3-3.5.23=. -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -Wl,-z,relro -Wl,-z,now -Wl,--as-needed' --datadir=/usr/share/squid --sysconfdir=/etc/squid --libexecdir=/usr/lib/squid --mandir=/usr/share/man --enable-inline --disable-arch-native --enable-async-io=8 --enable-storeio=ufs,aufs,diskd,rock --enable-removal-policies=lru,heap --enable-delay-pools --enable-cache-digests --enable-icap-client --enable-follow-x-forwarded-for --enable-auth-basic=DB,fake,getpwnam,LDAP,NCSA,NIS,PAM,POP3,RADIUS,SASL,SMB --enable-auth-digest=file,LDAP --enable-auth-negotiate=kerberos,wrapper --enable-auth-ntlm=fake,smb_lm --enable-external-acl-helpers=file_userip,kerberos_ldap_group,LDAP_group,session,SQL_session,time_quota,unix_group,wbinfo_group --enable-url-rewrite-helpers=fake --enable-eui --enable-esi --enable-icmp --enable-zph-qos --enable-ecap --disable-translation --with-swapdir=/var/spool/squid --with-logdir=/var/log/squid --with-pidfile=/var/run/squid.pid --with-filedescriptors=65536 --with-large-files --with-default-user=proxy --enable-build-info='Debian linux' --enable-linux-netfilter build_alias=x86_64-linux-gnu 'CFLAGS=-g -O2 -fdebug-prefix-map=/build/squid3-4PillG/squid3-3.5.23=. -fstack-protector-strong -Wformat -Werror=format-security -Wall' 'LDFLAGS=-Wl,-z,relro -Wl,-z,now -Wl,--as-needed' 'CPPFLAGS=-Wdate-time -D_FORTIFY_SOURCE=2' 'CXXFLAGS=-g -O2 -fdebug-prefix-map=/build/squid3-4PillG/squid3-3.5.23=. -fstack-protector-strong -Wformat -Werror=format-security' --with-openssl --enable-ssl-crtd
make && make install

The configuration above is identical to the precompiled one in the Debian Stretch repository, apart from the addition of the SSL options. If you are using a different distro the above command may not work.

Most of my configuration is based on the guide in the official squid documentation. My squid configuration is as follows:

acl ftp proto FTP
acl SSL_ports port 443
acl SSL_ports port 1025-65535
acl Safe_ports port 80 # http
acl Safe_ports port 21 # ftp
acl Safe_ports port 443 # https
acl Safe_ports port 70 # gopher
acl Safe_ports port 210 # wais
acl Safe_ports port 1025-65535 # unregistered ports
acl Safe_ports port 280 # http-mgmt
acl Safe_ports port 488 # gss-http
acl Safe_ports port 591 # filemaker
acl Safe_ports port 777 # multiling http
acl CONNECT method CONNECT
acl LANnet src 192.168.80.0/24 # local network for virtual machines
acl step1 at_step SslBump1
http_access deny !Safe_ports
http_access deny CONNECT !SSL_ports
http_access allow localhost manager
http_access allow LANnet
http_access deny manager
http_access allow localhost
http_access deny all
http_port 3128 intercept # intercept required for transparent proxy
https_port 3129 intercept ssl-bump \
    cert=/etc/squid/antfarm.pem \
    generate-host-certificates=on dynamic_cert_mem_cache_size=4MB
ssl_bump peek step1
ssl_bump bump all
sslcrtd_program /usr/lib/squid/ssl_crtd -s /var/lib/ssl_db -M 4MB
sslcrtd_children 8 startup=1 idle=1
access_log daemon:/var/log/squid/access.log logformat=combined
pid_filename /var/run/squid/squid.pid
coredump_dir /var/spool/squid
refresh_pattern ^ftp: 1440 20% 10080
refresh_pattern ^gopher: 1440 0% 1440
refresh_pattern -i (/cgi-bin/|\?) 0 0% 0
refresh_pattern . 0 20% 4320
request_header_access X-Forwarded-For deny all
httpd_suppress_version_string on
always_direct allow all

Use the SSL certificate generation process shown in the linked guide. Once you have created the .pem file, copy the section from —–BEGIN CERTIFICATE—– to —–END CERTIFICATE—– into a new file with the extension .crt.

A few notes here:

  • The ‘intercept’ keyword is necessary if you are using iptables to redirect ports to squid as a transparent proxy. If you configure your client to explicitly use a proxy, you should not use it.
  • The always_direct clause is used because we are routing squid’s output to another host (running tor) as the default gateway. If you wanted to use the squid → privoxy → tor configuration locally, you would use ‘never_direct’ instead.
  • The path for the ssl_crtd tool in Debian is /usr/local/squid/ssl_crtd – no libexec.
  • When setting permissions for the cache directories in Debian, use “proxy:proxy” instead of “squid:squid” as this is the default user that Debian creates to run the squid service.

In order for the virtual machine to treat the falsified public keys as genuine, we must instruct it to trust the certificate as created above. For a Windows 7 host like mine, double click the .crt file and import the certificate in to the Trusted Root Certification Authorities store.

Importing a cert

With squid set up and certificate imported, you must then configure iptables on the hypervisor host to redirect traffic through squid.

iptables -t nat -A PREROUTING -i virbr0 -p tcp --dport 80 -j REDIRECT --to-port 3128
iptables -t nat -A PREROUTING -i virbr0 -p tcp --dport 443 -j REDIRECT --to-port 3129

where virbr0 is the name of the virtual interface in QEMU. You should adjust interface name and destination ports as required for your setup.

tor service

On the second host I have installed tor (version 0.2.5.16 from Debian Stretch repo). This is configured with ports to listen for TCP and DNS connections in /etc/tor/torrc:

TransPort 192.168.42.2:8081
DNSPort 192.168.42.2:53

Then with iptables, inbound traffic from the hypervisor host is redirected to tor:

-A PREROUTING -s 192.168.42.4/32 -i eth0 -p tcp -j REDIRECT --to-ports 8081
routing

Since the objective is to keep my real IP hidden, care must be taken to ensure the host’s routing does not leak information. In /etc/network/interfaces, instead of specifying a gateway, I added two routes:

up route add -net 192.168.0.0 netmask 255.255.0.0 gw 192.168.40.1
up route add -net 0.0.0.0 netmask 0.0.0.0 gw 192.168.40.2

This causes all traffic not intended for my internal network to be routed to the host running the tor service (on 192.168.40.2). I have then configured my firewall so that it only allows connections reaching in to this VLAN, or from the tor host, not from the malware VM hypervisor.  When updates are required, connectivity can be enabled temporarily, with the VMs paused or shut off. Alternative techniques include allowing the hypervisor host to update via tor (if I didn’t mind it being slow), or routing the traffic from the VMs without NAT and denying anything outbound from the VM network on my core router, but that’s something to look at another day.

With the gateways set up, the routing for the VM interface can then be applied on the hypervisor host:

iptables -t nat -A POSTROUTING -o eth0 -j MASQUERADE
iptables -A FORWARD -i virbr0 -j ACCEPT

After applying these rules you should have a fully functioning TLS/SSL intercept routed via tor. To test, start by attempting to resolve a few hostnames from the VM and verify that the traffic is hitting your tor service host BEFORE giving any web pages a spin. Move on to HTTP/HTTPS traffic once you are sure DNS is working correctly.

Usage

Once you have a functioning setup you should expect to see both HTTP and HTTPS URLs appearing in your squid access log. In addition, if you perform a packet capture on the hypervisor virtual interface (virbr0 in my case), you can use the key generated earlier to view the decrypted traffic in Wireshark. You will need to copy the private key section of the .pem file to a new file to use in Wireshark. When entering the protocol as described in the link above, use ‘http’ in lowercase – uppercase will not work.

importing an SSL key in wireshark

decrypted output of call to https://ipapi.co

BSides Workshop “My log obeys commands – Parse!”

Better late than never, as they say. Last week I went to BSides London, which was pretty awesome. In between hanging out with all sorts of awesome people and downing mojitos, I had the opportunity to present a workshop. It seemed to go pretty well – though I have definitely learned enough to improve it for next time.

The short version is that it was an introduction to the basic principles and techniques of log parsing, for people at the level of a junior SOC analyst. Minimal regex knowledge required.

Although I don’t have a recording of the workshop, I’m putting the slides up here in case they’re of use to anyone. Enjoy! If you have any questions, please tweet @http_error_418 😊

My log obeys commands – Parse!

Undocumented function “RegRenameKey”

Whilst learning some bits and pieces about the Windows API, I found a function that I wanted to use which does not appear to be documented on MSDN, “RegRenameKey”. Although Visual Studio will helpfully inform you of the datatypes required, it does not tell you what is expected within the datatypes. Since the answer doesn’t seem to be out there just yet I thought I’d do a quick writeup. The definition for the function is as follows:

LSTATUS WINAPI RegRenameKey( 
  _In_         HKEY    hKey, 
  _In_         LPCTSTR lpSubKeyName,
  _In_         LPCTSTR lpNewKeyName 
);

I assumed that lpSubKeyName and lpNewKeyName would accept the same form of input as the registry path expected for RegKeyCreate, e.g. “Software\\MyProduct\\MyKey”. However, attempting to use this returns error code 0x57 “The parameter is incorrect”. This is because lpNewKeyName seems to expect just a name without the path. A valid call looks like this:

 TCHAR keyname[] = L"Software\\testkey";
 TCHAR newkeyname[] = L"testkey2";
 LSTATUS renameerr = RegRenameKey(HKEY_CURRENT_USER, keyname, newkeyname);

Not a particularly difficult one, but hopefully this will save people some time!

Bonding rituals

I may have been quiet on here but not because I haven’t been doing lots of fun nerdy stuff. Unfortunately, there’s a fair amount of it that can’t be blogged about, hence the lack of new material here, but a problem came up the other day that was a royal pain in the ass pretty fun and interesting, and maybe some folks out there might be scratching their heads over it and appreciate there being something in the depths of t’interwebs to explain it.

Bonding is a pretty damn useful thing, especially to us NSM folks. Take a 1×1 tap and run the output cables up to a nice bit of tin running $distro_of_choice, a few minutes of tweaking interface config files, and hey presto! a bonded interface with both directions of traffic for Snort/Suricata/Bro/whatever to listen to, and your kit is safely out of line where the sysadmins can’t blame you when something breaks and takes out the internet (they’ll probably still try though).

So far, so standard. The other day I needed to do this in a VM – no problem, I thought. VMWare will let you pass traffic through to the guest; you need to put the switch into promiscuous mode because the interface in your guest/sniffer won’t have an IP assigned, which you can do in the vSwitch Security Policy.

With each output of the tap assigned its own vSwitch which was attached to an individual interface on the guest, I created a bond interface to combine the two. In the very best tradition of here’s one I made earlier let someone else make and plagiarised shamelessly, you can read a good guide here.  One notable exception – use mode 0 (round robin) and not  active/passive – we want to combine the outputs, instead of having the second only work if the first fails.

So, having done that, I brought up the bond0 interface and… weirdness happened. I was only seeing one side of the traffic. tcpdump on the bond0 interface was only showing the responses, not the requests. The slaved interfaces told a similar story, one had traffic (inbound), and the other was silent. Odd. Next check, was the ESXi host seeing the traffic but not passing it through? Checking this requires the use of pktcap-uw rather than VMWare’s implementation of tcpdump, which will not let you look at traffic on individual vSwitches. This showed the traffic was indeed present.

Proper head-scratching time now. The interface settings were all correct, the problem persisted through restarts of the interfaces, the networking service, even the OS. Next step was bringing up each interface manually one at a time; now it got even weirder. eth1 showed responses as expected. eth2 showed requests – awesome! bond0 showed… just the responses. Checked eth2 and it was now silent as the grave. Curses! This didn’t change when bond0 was shut down again; outbound traffic would only reappear when eth2 was brought up without bond0. Enabling bond0 killed it again until it was started without bond0 running. What the hell?

Having pretty much run out of ideas, a bit of experimentation was on the cards, starting with the ESXi config settings. This was clearly a stroke of genius, because upon setting MAC address changes to ‘accept’, it instantly started working. Why would this be?

One of the things that enabling bonding does is that the bond0 interface defaults to starting with the MAC of the first interface to join the bond. In round-robin mode, it then shuttles its MAC address around each interface to receive frames; VMWare’s (sensible) default is to ignore changes like this, and as a result, will stop transmitting traffic to the interface it sees as having violated the restriction until the interface is bounced. Thus, the first slave to join will receive traffic because its MAC stays the same, and the second stops being sent data because the vSwitch has seen its MAC change. Permitting changes on the vSwitch means the MAC can be assigned as necessary.

TLDR: If you want to use a bonded interface in an ESXi guest like this, you must set ‘Allow MAC address changes’ to accept on the vSwitches the slave interfaces connect to.