1600658 - junit harness for content process crash tests needs redesign

Status: RESOLVED FIXED

(GeckoView :: Sandboxing, defect, P1)

Description

[(No longer employed by Mozilla) Aaron Klotz]

TL;DR (emphasis mine):

If the process name begins with a ':' character, a new process private to that application will be created when needed to run that component (allowing you to spread your application across multiple processes).

Our current setup for the crashContent test assumes that any content processes started by the RemoteGeckoService will be charged to the RemoteGeckoService process, when in fact the content processes are charged to the androidTest application as a whole.

In other words, RemoteGeckoService's instance of Gecko thinks it is creating a new content process, while Android just hands it the existing content process from androidTest.

Prior to bug 1594820, this seemed to kind of mostly work due to fluky timing between the various processes. With the patches from bug 1594820 applied, the timing changed enough that, while the content process does successfully crash, that crash report is consumed by androidTest's crash handler instead of RemoteGeckoService's crash handler. Ergo, crashContent perma-failure.

Comment 1

[(No longer employed by Mozilla) Aaron Klotz]

Attachment: Bug 1600658: Part 1 - Remove setting of MOZ_CRASHREPORTER_NO_REPORT from runjunit.py; r=gbrown

We now need the junit crash handler to see crash reports.

Depends on D56852

Comment 2

[(No longer employed by Mozilla) Aaron Klotz]

Attachment: Bug 1600658: Part 2 - Move content crash test into a session test and modify TestCrashHandler to selectively evaluate crash report submission; r=snorp

• We add bidirectional messaging to TestCrashHandler:

  1. The test sends a message to the TestCrashHandler, notifying it that an
     upcoming crash is intentional and its intent should be checked.
  2. Upon receipt of the crash report, the service reviews the contents of the
     crash intent, and then sends a message back to the test with the test results.
  3. The service deletes any crash dump artifacts belonging to crash intents
     that have been evaluated so that the harness doesn't pick up any
     intentional crashes.

• We remove crashContent from CrashTest.kt and create ContentCrashTest.kt
  for that case. The crashParent test remains unchanged other than switching
  its crash handler over to TestCrashHandler.

• We remove the CrashTestHandler service, as both crashContent and
  crashParent tests now use TestCrashHandler.

Depends on D56853

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5bd66ef01543
Part 1 - Remove setting of MOZ_CRASHREPORTER_NO_REPORT from runjunit.py; r=gbrown
https://hg.mozilla.org/integration/autoland/rev/2a1e94a02eef
Part 2 - Move content crash test into a session test and modify TestCrashHandler to selectively evaluate crash report submission; r=snorp

Comment 5

[Cristina Coroiu [:ccoroiu]]

https://hg.mozilla.org/mozilla-central/rev/5bd66ef01543
https://hg.mozilla.org/mozilla-central/rev/2a1e94a02eef

Comment 6

[Chris Peterson [:cpeterson]]

Moving content process management bugs to the new GeckoView::Sandboxing component.