OpenAFS Cache Manager tracing using eBPF
Presenters:
- Yadavendra Yadav, IBM
- Todd DeSantis, IBM
For debugging Linux Kernel issues in OpenAFS we use dynamic tracers like systemTap, perf, kprobe/kretprobe etc. Dynamic tracer like systemTap requires certain packages to be installed on a system, and it creates a kernel module which gets loaded in kernel to make probes active, same is a case with kprobe & kretprobe. eBPF can be used to trace a custom kernel module (like OpenAFS) and is part of Linux kernel itself. Also eBPF creates a register based Virtual machine which runs JIT native compiled BPF (Berkeley Packet Filter) programs and this makes it more safe to use in production environment. eBPF stack also have verifier in place which will make sure BPF programs doesn’t creates issues, e.g. loops, null pointer deference etc. eBPF can also be used for tracing user space programs. In this talk we will focus on architecture of eBPF and how it can be used for profiling, dynamic tracing and instrumentation of OpenAFS kernel module. We will also touch on tracing userspace processes like AFS servers.