Is there any demo or example that can run on Neoverse N2/V2?

[rejohn]

Is there any demo or example that can run on Neoverse N2/V2? I mean Alibaba Yitian710 or AWS graviton3?

[algrant-arm]

CSAL aims to support any CoreSight-compatible system, so that would include systems based on N2/V2, but to run the demo on any specific system, its CoreSight devices would need to be visible in the application core memory map. That relies on the vendor taking action to make the devices accessible and document them.

Generally, CSAL is designed for embedded systems, or for embedding into systems as part of system firmware, or by vendors during system bring-up and troubleshooting. If you're a regular server user and you don't have firmware-level access, then CSAL is unlikely to be helpful.

For instruction trace specifically, Armv9 replaces ETM with ETE/TRBE as a built-in feature on each individual core, and this is programmed via system registers. So the old-style memory-mapped interface to trace (which CSAL uses) is no longer required, and instruction trace no longer relies on CoreSight ATB components like funnels and ETRs - although it can be routed that way for low-level off-chip debug. The Linux kernel supports ETE/TRBE, it should appear as "perf record cs_etm..." just like ETM did. If you want instruction trace for profiling, and the ETE/TRBE driver isn't enabled, best check with the system vendor.

[rejohn]

Okay, thank you. I'm using cs_etm for instruction tracing. Because I want to control when the cs_etm trace starts and stops, I'm using perf_event_open + ioctl, and then writing the data from mmap memory to the file trace.bin. However, I don't know how to process this type of trace data—I know that perf records can be processed directly using perf + opencsd. Do you have any suggestions?

[algrant-arm]

It's not easy. Firstly, don't try writing your own ETE/ETM decoder - the ETE/ETM trace stream format is very complex, a decoder is 1000s of lines of code and it's not straightforward. Use OpenCSD directly, or via perf.

If the trace is from a single process, or from the kernel, you may be able to use OpenCSD stand-alone - you just need to arrange that it sees the correct binaries at the correct locations so it can follow the branches. The example code in OpenCSD sources should give some idea of how to do this. See also documentation in the OpenCSD repo: https://github.com/Linaro/OpenCSD

If the trace is from multiple processes, being scheduled in and out over time, or there are dynamic library loads and unloads, the problem becomes far more complicated. You have to arrange that OpenCSD sees the correct binary at the correct location at the correct time. This would likely mean collecting other perf records (PERF_RECORD_MMAP etc.) and establishing the state of the address space at each timestamp, correlating with the timestamps in the ETE/ETM trace stream. There is logic in perf to do this, but again, it's complex. If you have a raw trace.bin file and you have memory mappings from some other place, you could try synthesizing a perf.data file.