The usage of languages like Go and Rust is growing, and while such languages come with observability features/libraries and powerful associated tools, using pre-existing investigative, data collection workflows is something desirable, to take advantage of familiarity and training.
This presentation will cover roadblocks found in trying to use uprobes and BPF to collect metrics in workloads written in golang using the prometheus metrics APIs. Issues include calling conventions, floating pointing register access, data structure layouts obtained from DWARF, libbpf skeletons, measuring instrumentation overhead. The results in the improvement of tools such as perf and pahole will also be discussed.
Being able to use the same observability tools to investigate/correlate issues in the kernel, BPF, and userspace in various languages is something of value, but requires addressing language-specific implementation details, some of which will be presented and hopefully new ideas on how to tackle those issues will be gathered from fellow developers.
Arnaldo Carvalho de Melo created pahole, a tool to help in optimizing data structures, used in Linux, glibc, KDE, xine & others. Maintainer of ‘perf’ (profiling, tracing, debugging, etc). Lately getting lost in BPF land.