commit c0f481003b889aadb7b367e1546065626655e3b5
author Ryan Savitski <rsavitski@google.com> Tue Apr 14 17:20:56 2020 +0100
committer Ryan Savitski <rsavitski@google.com> Tue Apr 14 16:21:39 2020 +0000
tree 7f964c6c446bc617318ac36f2b71d7bae861cbec
parent fa3fa65e8d0dd0b2d6c3f78c9a9810fd60d437c7

traced_perf: optional periodic unwinder cache clearing

Preface:
* targeting muiltiple processes at once is normal for this profiler.
* unwindstack cache seems quite necessary in practice (for cpu%), but
  also doesn't actually have an inherent limit (so over time, footprint
  will grow to cover unwinding for all sampled processes).
* for longer traces, it's likely that the working set necessary at
  any given medium-term interval is much lower than the aforementioned
  worst case (so periodic clears could help).
* we want to try using ring buffer traces (with a finalization trigger)
  with stack sampling in the field.

I don't want to change the libunwindstack cache atm, so using it as-is.
Periodic clears are a bit crude, and I've considered alternatives (think
a memory guardrail that does this purging when hit, instead of shutting
down the data source / daemon). But stopping at a periodic option for
now.

Separate from incremental state clears since it's a heavy operation,
which should be performed infrequently. While incremental state clearing
needs to be frequent enough to make any given ring buffer snapshot have
most of the necessary incremental state for decoding.

Thoughts/suggestions welcomed. I'm kinda struggling with building a good
set of knobs for controlling the memory usage when running as an
unattended daemon.

Bug: 144281346
Change-Id: I6342fa636c9ee7eca214a7be4c74c0b187246b0f