| Age | Commit message (Collapse) | Author |
|---|
|
Produce more debugging output for a failing test.
As part of that, we allocate additional finalizable objects, and
print how many of them were finalized. This should allow us to
tell if there were young GCs, but the PhantomReference was
promoted. Also print whether the reference was enqueued after
explicitly triggering a GC.
Mark various allocation functions noinline to make sure that
we don't accidentally keep the allocated objects live. I don't really
expect this to be the issue, but ...
Increase the delay after we fail until we recheck the reference.
Bug: 123921229
Test: Inspect message when PhantomReference is kept live.
Change-Id: Ic5cc9aba3ed383def3174ef0fc1c6b4ed61678a1
|
|
Define kCheckImmediatelyThreshold in terms of kNotifyNativeInterval.
We previously raised kNotifyNativeInterval without adjusting it,
making it possible to go for much longer periods with checking
whether we should GC. This effectively reduces
kCheckImmediatelyThreshold, by about a factor of 2 in production,
and more on host. This is unlikely to fix the failures.
Have 004-NativeAllocations produce much more detailed output on
failure.
MUST_SLEEP to keep Lint happy.
Bug: 123921229
Test: TreeHugger
Change-Id: Ibe61ea16be8a5f25ec241f682ef3b4958a50116f
|
|
Add -XX:FinalizerTimeoutMs commandline argument, and make it
available for libcore.
Add 1336-short-finalizer-timeout to test it. Blacklist it for
the cases in which 030-bad-finalizers was known not to work,
since it's largely a clone. Disable it for gcstress, since timeout
margins are tight.
Increase timeout in 030-bad-finalizers due to gcstress issues.
Use the result in 004-NativeAllocations and 030-bad-finalizers
to avoid timeouts. Make 004-NativeAllocations more proactive in
avoiding timeout.
This is intended as a replacement for aosp/836642 .
Bug: 129350738
Bug: 68792448
Test: Boot AOSP, TreeHugger
Change-Id: I9aef9c9bacb7e8d851bd2c879ed9fb5b596abc60
|
|
Under a GSS configuration, calls to registerNativeAllocation trigger
blocking GC instead of concurrent GC, which means the main thread in the
test can end up holding on to deadlockLock for longer than the finalizer
timeout.
The fix is to reduce the calls to registerNativeAlloation required to
trigger the deadlock scenario.
Bug: 68792448
Test: ./test/testrunner/testrunner.py -b -t 004-NativeAllocations
Test: Verify test catches deadlock case on buggy versions of art.
Change-Id: Ide417de7a5cad593afd2877054f88588b7a32be3
|
|
Simplifies test logic for testing that RegisterNativeAllocation triggers
GC and doesn't deadlock.
Removes the test that RegisterNativeAllocation bounds the maximum
allocated native bytes, because we no longer wish to provide a guarantee
of that kind.
Test: ./test/testrunner/testrunner.py -b -t 004-NativeAllocations --host
Test: Verify the test fails if RegisterNativeAllocation does nothing.
Test: Verify the test fails if kNativeAllocationFinalizeTimeout is 0.
Bug: 68792448
Bug: 70831911
Change-Id: I3f6625869a89976e22918e2bca79c87f54aa9596
|
|
Fixes 12 tests with javac/dx.
Previously, javac with the default bootclasspath (unspecified) was used,
meaning that it would usually pickup rt.jar from a developer's JDK system
install.
With javac-helper.sh, libcore is used for the javac -bootclasspath,
enabling art tests to load libcore-specific packages such as libcore or
dalvik (in particular, it uses the output of art/tools/bootjars.sh which
calculates the correct bootclasspath from the build artifacts).
To get libcore on the bootclasspath, source files must be placed in a
'art/test/$TESTNAME/src-art' directory. Otherwise the old behavior is
kept to encourage tests to be cross-platform compatible.
The minimal set of tests that relied on 'import dalvik/import libcore'
had their src directories renamed to src-art, and are now building
successfully.
Test: ANDROID_COMPILE_WITH_JACK=false art/test.py
Bug: 36902714
Change-Id: Iafd245de9e04c312c5ac107897e34d7b97191726
|