RFR: 8342848: Shenandoah: Marking bitmap may not be completely cleared in generational mode

[William Kemper]

On Wed, 23 Oct 2024 16:06:25 GMT, Xiaolong Peng <xpeng at openjdk.org> wrote:

>> In the investigation of the crashe I saw in PR https://github.com/openjdk/shenandoah/pull/516, I happened to reproduce the crash GenShen TIP as well.  
>> 
>> The crash was reproduced multi times on both AWS r7g-4xlarge and r7i-4xlarge instances by running test below repeatedly:
>> 
>> 
>> CONF=linux-aarch64-server-fastdebug  make clean test TEST=gc/stress/gcold/TestGCOldWithShenandoah.java#generational JTREG="REPEAT_COUNT=1000" 
>> ``` 
>> 
>> Crash:
>> 
>> #
>> # A fatal error has been detected by the Java Runtime Environment:
>> #
>> #  Internal Error (/home/xlpeng/repos/jdk-xlpeng/src/hotspot/share/gc/shenandoah/shenandoahConcurrentGC.cpp:642), pid=24134, tid=24158
>> #  assert(_generation->is_bitmap_clear()) failed: need clear marking bitmap
>> #
>> # JRE version: OpenJDK Runtime Environment (24.0) (fastdebug build 24-internal-adhoc.xlpeng.jdk-xlpeng)
>> # Java VM: OpenJDK 64-Bit Server VM (fastdebug 24-internal-adhoc.xlpeng.jdk-xlpeng, mixed mode, sharing, tiered, compressed oops, compressed class ptrs, shenandoah gc, linux-aarch64)
>> # Problematic frame:
>> # V  [libjvm.so+0x15eadc4]  ShenandoahConcurrentGC::op_init_mark()+0x358
>> #
>> # No core dump will be written. Core dumps have been disabled. To enable core dumping, try "ulimit -c unlimited" before starting Java again
>> #
>> # An error report file with more information is saved as:
>> # /local/home/xlpeng/repos/jdk-xlpeng/build/linux-aarch64-server-fastdebug/test-support/jtreg_test_hotspot_jtreg_gc_stress_gcold_TestGCOldWithShenandoah_java_generational/scratch/0/hs_err_pid24134.log
>> #
>> # If you would like to submit a bug report, please visit:
>> #   https://bugreport.java.com/bugreport/crash.jsp
>> #
>> 
>> 
>> With logging/instrumentation,  it seems to be caused by the one line  code `bool needs_reset = _generation->contains(region) || !region->is_affiliated(); `, considering bitmap reset is a concurrent operation, if is possible mutator thread changes the affiliation from FREE to YOUNG when  bitmap reset is running, both `_generation->contains(region)` and `!region->is_affiliated()` can be false when affiliation is FREE and mutator is updating it at the same time.
>> 
>> Logs from instrumentation:
>> 
>> [32.793s][info][gc          ] GC(19) Not reseting bitmap for YOUNG region (0x0000ffff8c1a6100)(affiliation before test: FREE)
>> 
>> ...
>> 
>> [32.807s][info][gc,task     ] GC(20) Using 8 of 8 workers for init marking
>> [32.808s][info][gc          ] GC(20) Region (0x0000ffff8c1a6100) doesn't have clear bitmap, [1, 1, 1]
> ...
>
> src/hotspot/share/gc/shenandoah/shenandoahGeneration.cpp line 80:
> 
>> 78:     ShenandoahMarkingContext* const ctx = heap->marking_context();
>> 79:     while (region != nullptr) {
>> 80:       bool needs_reset = !region->is_affiliated() || _generation->contains(region);
> 
> We should really read the affiliation only once for the whole test, but will have to create a new method to achieve that to keep the code clean and encapsulated.

Hmm, you reckon the affiliation is changing between first testing for `_generation->contains` and then testing for `!region->is_affiliated()`? I suppose this _could_ happen, but the fix feels fragile (not sure we can guarantee the compiler won't reorder the tests). We do have `ShenandoahHeapRegion::affiliation`. What about:

  auto affiliation = region->affiliation();
  bool needs_reset = (affiliation == ShenandoahAffiliation::FREE || _generation->contains(affiliation))

Where `ShenandoahGeneration::contains(ShenandoaAffiliation affiliation)` tests if the affiliation and the generation type match?

-------------

PR Review Comment: https://git.openjdk.org/shenandoah/pull/523#discussion_r1813300281