On Mon, Mar 7, 2016 at 9:28 AM, Arnaldo Carvalho de Melo
[off-list ref] wrote:

quoted

I have gone down that route, but I don't get user stack traces at all.
  I suspect this is because I'm on ARM 3.14.52, which does not support
CONFIG_USER_STACKTRACE_SUPPORT.  Nor does it support
CONFIG_ARCH_SUPPORTS_UPROBES.
All that said, I DO get stack traces when using straight perf record
with "--call-graph dwarf".  (Does that imply "-e cpu-clock"?).  So, I

Ok, good data point, as what 'perf record --callgraph dwarf' does is to
take register and userland stack samples so that later it can use it
together with CFI information from the binaries involved to do the stack
unwinding.

Right, that's my understanding as well.  Given that I do get stack
traces in some cases, this makes me wonder what the difference between
the events which yield user stack traces (seemingly "Software event"
in perf list) and events which do not yield user stack traces
(seemingly "Tracepoint" in perf list).  My next steps are to work
through these mechanisms and try to understand the difference.

quoted

was hoping that whatever is enabling that event to get stack traces
might be available to me.

So what you are not managing to do is to collect userland stacktraces
from tracepoints, right?

Correct.

Any ARM person here to try to help diagnose why this does work on x86_64
but not on ARM?

Any pointers appreciated!  Thanks!

Jonah