SELinux stands for Security-Enhanced Linux. It’s a Linux Security Module (LSM) that enables permissions and restrictions with a much finer granularity than traditional Unix permissions.
In short, SELinux works by labeling resources with types and establishing
permissions for certain operations between types. For example, a process
running the program
/usr/sbin/varnishd has the type
varnishd_t and the
files inside the
/etc/varnish directory have the type
There exists a rule to allow
varnishd_t to read
varnishd_etc_t files, for
/etc/varnish/default.vcl. SELinux also has its own flavor of users
and roles and comes with a significant learning curve.
varnish service is started, you should be able to observe two
$ ps -C varnishd,cache-main -o comm,label COMMAND LABEL varnishd system_u:system_r:varnishd_t:s0 cache-main system_u:system_r:varnishd_t:s0
The label gives us the security context for
system_uis the built-in system user meant for system services
system_ris the built-in system role
varnishd_tis a type
s0is a range
The important part in this case is that
varnishd is labeled with the
varnishd_t type. Even though
varnishd is executed as the root user,
SELinux will deny any operation that a root user could normally perform
if that operation is not allowed for the
It used to be almost systematic to solve a permission problem by disabling SELinux and you should avoid that. What SELinux solves is the loophole of the root user: in theory the root user (UID 0) is the privileged user and can do anything. For instance, port numbers below 1024 are privileged ports, so in order to simply run a standard web server on ports 443 and 80 you need root privileges.
Varnish was designed with security in mind, and the workload is split in two processes:
With its jail facility Varnish will permanently drop privileges in the cache process, and temporarily drop them in the management process to only reacquire privileges with the smallest possible window to perform privileged operations.
While Varnish proactively tries to mitigate many attack vectors, what happens when a severe security vulnerability is discovered in any piece of software? If a remote code execution can be achieved on a process running as root then an attacker can virtually do anything. SELinux defends against that and will for example prevent Varnish from reading your mail server configuration if you happen to run both services on the same system. SELinux comes with a policy covering a large part of critical system software like Varnish and can greatly reduce the blast radius of certain classes of vulnerabilities.
You shouldn’t disable SELinux.
When SELinux is enabled it can essentially run in two modes:
In enforcing mode, anything denied by the policy will be blocked and recorded
in the system’s audit log. The application will likely see denied operations
fail with Operation not permitted (
EPERM), and how this error is handled is
up to the application.
In permissive mode, an unauthorized operation is only logged, not denied, which allows the application to keep making progress while informing system administrators that the policy would have denied certain operations.
Ideally, running a comprehensive workload in permissive mode should allow you to catch most if not all the things that would go wrong in enforcing mode.
You can then search the audit log with
ausearch and even generate SELinux
audit2allow. When working with Varnish Enterpise, the best way
to help us study SELinux denials is to send us a
varnishgather report after
running your workload in permissive mode.
See varnishgather on GitHub.
You don’t have to jeopardize the security of your entire system by putting it in permissive mode. Instead, you can put a specific type from the policy in permissive mode:
semanage permissive -a varnishd_t # add to the list of permissive types semanage permissive -d varnishd_t # delete permissive type semanage permissive -l # list permissive types
As mentioned earlier, SELinux will prevent Varnish from reading your mail server configuration. This works by abstracting resources such as paths, programs or port numbers with types, and then allowing certain types to perform specific operations on other types.
The SELinux policy contains several types related to Varnish, you may put in permissive mode the one that is denied operations according to the audit log.
It will very likely be one of:
varnishd_t is allowed to read configuration files labeled as
varnishd_etc_t. Some utilities like
ls can take a
-Z option to print the
security context of files and directories:
$ ls -Z /etc/varnish/*.vcl unconfined_u:object_r:user_home_t:s0 /etc/varnish/custom.vcl system_u:object_r:varnishd_etc_t:s0 /etc/varnish/default.vcl
If you observe files with the wrong security context, this can be corrected
$ restorecon -v -r /etc/varnish Relabeled /etc/varnish/custom.vcl from unconfined_u:object_r:user_home_t:s0 to unconfined_u:object_r:varnishd_etc_t:s0
If your SELinux troubleshooting reveals something more complicated than an
unexpected label on a file, remember to contact our support if you are a
Varnish Enterprise customer. For the quickest possible resolution, include a
varnishgather report after running your workload with the problematic type
in permissive mode.
Varnish Enterprise supports SELinux on Red Hat Enterprise Linux (RHEL) and compatible systems. The policy is developed and maintained by Red Hat, and we at Varnish Software contribute fixes and enhancements to the policy when we update our downstream policy.
SELinux is modular, so there is a base policy and a set of modules on top of that. There is a varnishd semodule covering several components from Varnish Cache.
In broad strokes, the
varnishd program is allowed to do the following:
It also allows
varnishncsa to access Varnish’s shared
memory log and persist logs in
On systems where we support SELinux we disable the
varnishd semodule and add
varnish-plus semodule instead. We do this because Varnish Enterprise
has more features and some of them would otherwise be denied by default.
We could have added a complimentary semodule and leave the
varnishd one alone,
but we support both RHEL7 and RHEL8 and their policies target respectively
6.0 between which there is a noticeable gap.
It was much simpler to maintain a single policy for both platforms. By
effectively forking the
varnishd semodule we also took care of isolating
Enterprise-only permissions, which gives us in return a better understanding
of whether a change we are making is relevant for the upstream policy and
should be contributed back.
The one noticeable departure from the upstream policy is the split of the
varnishd_port_t in two:
All VMODs are running inside
varnishd, and are subject to the same SELinux
policy. Even our
vmod_memcached would likely be denied a connection to a
Memcached server based on Memcached’s default port number.
There are several ways to deal with this. You could add the Memcached port to
varnishd_http_port_t directly from a command line with the
utility. Or you could create your own semodule to grant
varnishd_t access to
Unless you are proficient enough with SELinux to extend the policy, you could simply turn a built-in boolean on:
semanage boolean --modify --on varnishd_connect_any
When in doubt, contact our support.
One often reoccurring mistake is to copy files via SSH and move them to the
/etc/varnish directory. Whether it is done manually or through automation,
it should be noted that when a file is effectively moved, it will keep the
security context it inherited when it was created.
So it is common that VCL files in
/etc/varnish end up with the
type. You can fix this as part of your deployment by adding a
step or by updating the security context during the move:
$ ls -Z custom.vcl unconfined_u:object_r:user_home_t:s0 custom.vcl $ mv -v -Z custom.vcl /etc/varnish copied 'custom.vcl' -> '/etc/varnish/custom.vcl' removed 'custom.vcl' $ ls -Z /etc/varnish/custom.vcl unconfined_u:object_r:varnishd_etc_t:s0 /etc/varnish/custom.vcl
There are other utilities like
ps that can take a
-Z option, not just
Our packaging comes with a
logrotate configuration, and the idea is to always
append access logs and rotate them to avoid overwriting and losing logs on
disk as we are building them up. The lack of arbitrary write permission comes
from the upstream
varnishd semodule and it looks sensible enough that we
decided to keep it that way.