Add the ability to choose in-kernel frame pointer unwinding of the user stack

[mstange]

The ability to use frame pointer unwinding landed in ab3bf51. But as far as I can tell this still copies the user stack bytes and does the unwinding in user-land, which is slow (copies a lot of memory) and which leads to truncated stacks because it can only unwind the top 64KB of the stack.

perfetto/src/profiling/perf/event_config.cc

Lines 623 to 638 in 0ce5e90

	if (IsUserFramesEnabled(unwind_mode)) {
	pe.sample_type |= PERF_SAMPLE_STACK_USER | PERF_SAMPLE_REGS_USER;
	// PERF_SAMPLE_STACK_USER:
	// Needs to be < ((u16)(~0u)), and have bottom 8 bits clear.
	// Note that the kernel still needs to make space for the other parts of the
	// sample (up to the max record size of 64k), so the effective maximum
	// can be lower than this.
	pe.sample_stack_user = (1u << 16) - 256;
	// PERF_SAMPLE_REGS_USER:
	pe.sample_regs_user =
	PerfUserRegsMaskForArch(unwindstack::Regs::CurrentArch());
	}
	if (kernel_frames) {
	pe.sample_type |= PERF_SAMPLE_CALLCHAIN;
	pe.exclude_callchain_user = true;
	}

I would like to have a path which asks the kernel to do the unwinding of the user stack, i.e. which uses PERF_SAMPLE_CALLCHAIN, and which does not drop the user frames (exclude_callchain_user = false). This path would not set PERF_SAMPLE_STACK_USER, i.e. it would skip the copying of the stack bytes.

Could such a path be added? cc @LalitMaganti

[LalitMaganti]

Thanks for filing! It's a lot more clear now what you're asking for, sorry for the confusion on the bug :)

I would like to have a path which asks the kernel to do the unwinding of the user stack, i.e. which uses PERF_SAMPLE_CALLCHAIN, and which does not drop the user frames (exclude_callchain_user = false). This path would not set PERF_SAMPLE_STACK_USER, i.e. it would skip the copying of the stack bytes.

I would say that this makes sense as a feature for us to have in Perfetto. However, it's also not something we would use ourselves (as Android does not do frame pointer unwinding and I don't have any hope of that changing). So I think this is one of those PRs we might get to our of personal curiousity but not otherwise.

To anyone reading: if you're interested in this, by far the most reliable way to make this happen is sending us a PR. We're happy to guide on how this should be designed and review the code.

However, I'm by no means the authority on the recording code of Perfetto so will defer to @KirillTim and @rsavitski to be have the final say.

[rsavitski]

+1 that the request makes sense, but would need to be a contribution since the profiler primarily targets whole-system profiling on Android, where the lowest common denominator requires userspace stacks + DWARF unwinding.

Fwiw I've seen this option in a pending contribution from ARM (#1984), so we'll probably address this request as part of that PR.