diff options
| author | 2022-11-13 20:14:10 +0000 | |
|---|---|---|
| committer | 2022-12-01 23:32:14 +0000 | |
| commit | 7dde6edda2bc0091f54a2d07ff80d124090fa964 (patch) | |
| tree | c49e7cd2677c3acb28f9ea59df3503db27bda62d /runtime/native/java_lang_VMClassLoader.cc | |
| parent | dd161502900d8d28aca56c17eab9045770917a91 (diff) | |
Grab perfetto heap dump on OOME
In order to make OutOfMemoryErrors actionable, enable collecting heap
dumps at the time of the error when there is a present tracing session.
For now userdebug-only
This change introduces a new hook which is used by the existing
perfetto plugin to collect the data.
Usage: Setup a perfetto tracing session using data source
android.java_hprof.oom, with a START_TRACING trigger activated by
"com.android.telemetry.art-outofmemory"
Important points:
- Unlike the current impl, we need to block the execution while we
collect the data. This is because propagating the OOME will
result in the process exiting (and cleanup will eliminate the entire
cgroup, including the heap dump children)
- Processes do not know whether a START_TRACING session exists but
nevertheless need to wait for the producer start signal (in case one
exists). We wait up to 500ms which will delay all OOMing processes.
For more details see go/art-oom-heap-dump
Test: manual
system_server OOME (collected heap dump, verified execution
resumes without a tracing session)
Bug: 211178923
Change-Id: I90e260fbc6f113e77d00f0c958551fb561babb4e
Diffstat (limited to 'runtime/native/java_lang_VMClassLoader.cc')
0 files changed, 0 insertions, 0 deletions