Vaishali Thakkar

Implementing profilers and tracers in BPF offers a high degree of flexibility and allows for tools that have lower overhead and can make them suitable for production usage. BPF has a helper to unwind native stacks with frame pointers that works great. Unfortunately, most Linux distros and compilation pipelines omit frame pointers. We've built a BPF program that uses DWARF-unwind information to walk native stacks without frame pointers in BPF. We've integrated it into our continuous profiler project, Parca Agent.

Other people: Javier Honduvilla Coto

Sampling CPU profilers periodically fetch the stacks of the profiled processes that are running on the CPU at a given time. Walking the stacks of native processes with a little work is easily possible when frame pointers(FPs) are present. But most binaries in the real world are not compiled with FPs. So it can get quite complicated if profilers have to walk the stacks when frame pointers are omitted. In this talk, we will talk about how we can walk the stacks using the DWARF CFI (mainly .eh_frame). We will also discuss how eBPF is helping us with that and how extending the current stack walking facilities can be useful especially in interpreted languages, such as Ruby, as well as runtimes with JITs, like the JVM.