@hergertme Is there a performance penalty with this?
And if sframe support becomes default (after it is accepted), then we could teach the kernel to do user space stack tracing without frame pointers.
0
0
2
@hergertme sframe is similar to ORC. I guess it we should take some time to do the benchmarks with and without frame pointers, on critical loads.
ORC has solved the issue for the kernel, but currently profiling user space without frame pointers is “interesting”.
0
0
0
@hergertme as for getting that information in the kernel. If there’s an elf section for it, it shouldn’t be too hard. We are already looking into it. The idea is to flag that the task has the sframe section available, and when perf wants a trace, it will set a callback of some kind, and when the task is about to go back to user space, the section will be mapped in and call the perf callback to with the user space stacktrace.
It doesn’t matter when the user space stacktrace happens inside the kernel, it’s not going to change.
0
0
0