Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Extended Berkeley Packet Filter (eBPF) supports in-kernel observability, networking, and security tools.
Default eBPF restrictions
The image disables unprivileged loading with
kernel.unprivileged_bpf_disabled=2. Just-in-time (JIT) compilation is enabled
(net.core.bpf_jit_enable=1), and JIT constant blinding is off
(net.core.bpf_jit_harden=0). To turn on constant blinding, see
Decide whether to harden the JIT.
bpftool, kernel development files, and eBPF build libraries aren't installed
on the image.
Prerequisites
- An Azure Linux virtual machine (VM).
- A user account with
sudoprivileges.
The program inventory and build checks on this page require packages that aren't installed by default. Install them before you run those commands:
sudo dnf install -y bpftool kernel-headers kernel-devel libbpf-devel
Verify eBPF restrictions and loaded programs
Runtime restrictions, kernel build options, and loaded programs provide separate views of eBPF use.
Read the current restrictions
sysctl \
kernel.unprivileged_bpf_disabled \
net.core.bpf_jit_enable \
net.core.bpf_jit_harden
Expected result for the default configuration:
kernel.unprivileged_bpf_disabled = 2
net.core.bpf_jit_enable = 1
net.core.bpf_jit_harden = 0
kernel.unprivileged_bpf_disabled has three states, and the difference between
two of them matters for auditing:
| Value | Behavior |
|---|---|
0 |
Any process can call bpf(). |
1 |
Unprivileged calls are rejected, and the value can't be lowered again until the system restarts. |
2 |
Unprivileged calls are rejected, and a privileged process can still set the value back to 0. |
Verify the build-time configuration
grep -E \
'^CONFIG_(BPF|BPF_SYSCALL|BPF_JIT|BPF_JIT_ALWAYS_ON|BPF_UNPRIV_DEFAULT_OFF)=' \
/boot/config-$(uname -r)
Expected result for the image kernel:
CONFIG_BPF=y
CONFIG_BPF_SYSCALL=y
CONFIG_BPF_JIT=y
CONFIG_BPF_JIT_ALWAYS_ON=y
CONFIG_BPF_UNPRIV_DEFAULT_OFF=y
CONFIG_BPF_JIT_ALWAYS_ONcompiles out the eBPF interpreter, so a program either compiles to native code or doesn't run.CONFIG_BPF_UNPRIV_DEFAULT_OFFis whykernel.unprivileged_bpf_disabledstarts at2rather than0.
Inventory what's loaded
List loaded programs and maps. These commands require the bpftool package
from Prerequisites:
bpftool version
sudo bpftool prog list
sudo bpftool map show
The output is similar to the following example. Program IDs, tags, load times, and process IDs vary:
bpftool v7.6.0
using libbpf v1.6
features: llvm, skeletons
41: lsm name restrict_filesystems tag 31efb7c56239148b gpl
loaded_at 2026-10-30T10:07:35+0000 uid 0
xlated 520B jited 297B memlock 4096B map_ids 22
btf_id 45
pids systemd(1)
94: cgroup_skb name sd_fw_egress tag 772db7720b2728e9 gpl
loaded_at 2026-10-30T10:13:07+0000 uid 0
xlated 64B jited 67B memlock 4096B
pids systemd(1)
...
Systemd and container runtimes can load programs, so a nonempty list isn't
necessarily unexpected. On the default image, systemd loads programs such as
restrict_filesystems, sd_fw_egress, sd_fw_ingress, and sd_devices.
Investigate entries you can't attribute to a deployed component:
sudo bpftool prog show id <program-id>
Use the program type, name, load time, and available process information to correlate it with an agent. A loaded program isn't necessarily attached to a hook, and its owner ID alone doesn't identify the application.
Harden eBPF and prepare a build environment
Changes to loading restrictions or JIT settings can affect agents that use eBPF. Test compatibility and performance before rollout.
Make the loading restriction one-way
If your policy requires that the restriction can't be lifted while the system
runs, set it to 1:
Caution
You can't return to 0 or 2 without restarting the VM. Confirm first that
nothing on the system loads eBPF programs without privileges. Agents
typically run with CAP_BPF or CAP_SYS_ADMIN and aren't affected.
echo 'kernel.unprivileged_bpf_disabled = 1' |
sudo tee /etc/sysctl.d/90-bpf-hardening.conf
sudo sysctl --load /etc/sysctl.d/90-bpf-hardening.conf
sysctl kernel.unprivileged_bpf_disabled
Decide whether to harden the JIT
net.core.bpf_jit_enable = 1 means programs are compiled to native code
instead of interpreted. net.core.bpf_jit_harden controls constant blinding,
which stops an attacker from smuggling chosen values into executable memory
through program constants:
| Value | Effect |
|---|---|
0 |
No blinding. |
1 |
Blinds constants in programs loaded by unprivileged users. |
2 |
Blinds constants in all programs, including privileged ones. |
With unprivileged loading disabled, use 2 if your requirement also covers
privileged loaders. Measure its effect on your agents and network performance:
echo 'net.core.bpf_jit_harden = 2' |
sudo tee /etc/sysctl.d/90-bpf-jit-harden.conf
sudo sysctl --load /etc/sysctl.d/90-bpf-jit-harden.conf
sysctl net.core.bpf_jit_harden
To undo the JIT change, remove the file and restore the prior runtime value
with sudo sysctl -w net.core.bpf_jit_harden=<previous-value>. Unlike the
one-way unprivileged-loading restriction, this setting can be changed without
a restart.
Install build dependencies
Kernel development files and eBPF build libraries aren't installed on the image. Install them before you build eBPF programs:
sudo dnf install -y kernel-headers kernel-devel libbpf-devel
Follow your loader's build requirements. Check the kernel development package and whether the kernel exposes BPF Type Format (BTF) information for compile-once-run-everywhere (CO-RE) relocations:
uname -r
rpm -q kernel-devel
ls -l /sys/kernel/btf/vmlinux
After you install the build dependencies, the result is similar to the
following example. The kernel-devel version matches the running kernel:
6.18.50-1.5.azl4.x86_64
kernel-devel-6.18.50-1.5.azl4.x86_64
-r--r--r--. 1 root root 6642883 Oct 30 10:22 /sys/kernel/btf/vmlinux
If rpm -q kernel-devel returns package kernel-devel is not installed,
install the build dependencies first.
For builds that require matching kernel development files, compare the package
version with the running kernel. If /sys/kernel/btf/vmlinux is absent, check
the tool's BTF requirements before attempting to load it.
Troubleshoot eBPF loading and build failures
Loading a program fails with "Operation not permitted"
That's the expected result for an unprivileged process. Run the loader with
root privileges, or grant the service CAP_BPF and the capabilities its
program type needs.
Setting kernel.unprivileged_bpf_disabled back to 0 fails
The value was previously set to 1, which the kernel treats as one-way. Remove
the setting from /etc/sysctl.d/ and restart the VM.
bpftool prog list returns nothing
Confirm the command succeeded with sudo. An empty successful result means no
loaded programs were listed; it isn't an attachment-specific check.
A build fails after a kernel update
The kernel-devel package now matches the newly installed kernel, not the one
you're running. This check requires the kernel-devel package from
Prerequisites. Restart into the updated kernel, then rebuild:
uname -r
rpm -q kernel kernel-devel
The running kernel version and the kernel-devel version must match, as in
the following example:
6.18.50-1.5.azl4.x86_64
kernel-6.18.50-1.5.azl4.x86_64
kernel-devel-6.18.50-1.5.azl4.x86_64