Architecture

Inspektor Gadget is deployed to each node as a privileged DaemonSet. It uses in-kernel BPF helper programs to monitor events mainly related to syscalls from userspace programs in a pod. The BPF programs are run by the kernel and gather the log data. Inspektor Gadget’s userspace utilities fetch the log data from ring buffers and display it. What BPF programs are and how Inspektor Gadget uses them is briefly explained here.

k8s integration

The Linux kernel has an inbuilt virtual machine for BPF bytecode, allowing userspace to run small scripts in kernel space with limited impact (actually it is called eBPF to distinguish it from the historical BPF). The BPF programs are supplied by userspace in a binary format. The kernel then verifies the program through static analysis, so that no memory corruption can happen and no out of bounds access can leak sensitive data. In BPF no loops are allowed and the maximum number of instructions is limited, so that a BPF program with logical bugs can not hang up the kernel. Read more on BPF here and here .

To trace pods, Inspektor Gadget attaches BPF programs to kernel functions and the kernel will run them always when the functions are executed. Therefore, the BPF programs need to detect if the syscall that triggered the function comes from a pod that Inspektor Gadget should trace. To do that the program looks up the current cgroup id in a BPF map containing the list of pods to trace, if it’s not found the program exits early. Finally, the BPF program gathers the information to trace, e.g., syscall parameters, and writes them to a ring buffer or BPF map. Inspektor Gadget’s userspace utility listens or reads on this ring buffer or BPF map and fetches new events. If the tracing ends, the BPF program is removed again.

The Gadget Tracer Manager keeps a list of running gadgets and containers. Each running gadget has an associated BPF map that is filled with the cgroup ids of the containers to be traced according to the namespace, labels, pod name, etc. parameters passed to the gadget. The Gadget Tracer Manager also exposes a gRPC interface that is called each time a container is created or destroyed by the OCI PreStart and PostStop hooks. It updates the corresponding BPF maps of each gadget if the container satisfies the matching criteria.

Gadget Tracer Manager

The execsnoop, opensnoop, tcptop and tcpconnect subcommands use programs from bcc with additional filtering modifications . They are directly started on the nodes and their output is forwarded to Inspektor Gadget.

Sometimes it is useful to run a BPF program always in the background. It can trace everything and save it into different ringbuffers per pod. The userspace utility can then accesses a ring buffer retrospectively only if needed for introspection or errors. This is currently done for traceloop (not present in bcc), which uses an overwriteable ring buffer to only log a small amount of recent events. It uses a service per each node which provides a Unix Domain Socket accepting HTTP requests to list the available traces and to dump them (this stops tracing even if the pod did not already crash).

Debugging with traceloop on Kubernetes

Previous talks