JDK-8179084 : HotSpot VM fails to start when AggressiveHeap is set
  • Type: Bug
  • Component: hotspot
  • Sub-Component: gc
  • Affected Version: 8u131,9
  • Priority: P2
  • Status: Closed
  • Resolution: Fixed
  • OS: generic
  • CPU: generic
  • Submitted: 2017-04-21
  • Updated: 2018-06-21
  • Resolved: 2017-04-27
The Version table provides details related to the release that this issue/RFE will be addressed.

Unresolved : Release in which this issue/RFE will be addressed.
Resolved: Release in which this issue/RFE has been resolved.
Fixed : Release in which this issue/RFE has been fixed. The release containing this fix may be available for download as an Early Access Release or a General Availability Release.

To download the current JDK release, click here.
JDK 10 JDK 8 JDK 9 Other
10Fixed 8u151Fixed 9 b168Fixed openjdk7uFixed
Related Reports
Duplicate :  
JDK-8187900 - JRE fails to start by only specifying -XX:+AggressiveHeap
Relates :  
JDK-8147910 - Cache initial active_processor_count
Description
FULL PRODUCT VERSION :
java version "1.8.0_131"
Java(TM) SE Runtime Environment (build 1.8.0_131-b11)
Java HotSpot(TM) 64-Bit Server VM (build 25.131-b11, mixed mode)

FULL OS VERSION :
Mac OS X Sierra 10.12.4
Darwin ID-PC14001.local 16.5.0 Darwin Kernel Version 16.5.0: Fri Mar  3 16:52:33 PST 2017; root:xnu-3789.51.2~3/RELEASE_X86_64 x86_64

A DESCRIPTION OF THE PROBLEM :
When specifying -XX:+AggressiveHeap as JVM argument the VM now aborts with an error message . This used to work perfectly fine in all previous versions of the JRE/JDK

THE PROBLEM WAS REPRODUCIBLE WITH -Xint FLAG: Yes

THE PROBLEM WAS REPRODUCIBLE WITH -server FLAG: Yes

REGRESSION.  Last worked in version 8u121

STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
In the simplest form even this command line aborts
"java -XX:+AggressiveHeap -version" with error message 
"The Parallel GC can not be combined with -XX:ParallelGCThreads=0"

In previous versions this simple cmd line would print the version. But any other combination doesn't work either. 

Additionally trying to use PrintFlagsFinal to check the actual value still causes the error.
Many products ship with the setting in some config file out of the box (as do our own when set to production level during install)

PrintFlagsFInal confirms that it was using 8 threads on my machine prior to u131.

EXPECTED VERSUS ACTUAL BEHAVIOR :
That the virtual machine starts up with AggressiveHeap enabled without error
ERROR MESSAGES/STACK TRACES THAT OCCUR :
"The Parallel GC can not be combined with -XX:ParallelGCThreads=0"

REPRODUCIBILITY :
This bug can be reproduced always.

---------- BEGIN SOURCE ----------
java -XX:+AggressiveHeap -version
---------- END SOURCE ----------

CUSTOMER SUBMITTED WORKAROUND :
You can work arround this by explicitly specifying a value for ParallelGCthreads only if that argument is earlier in the argument list than AggressiveHeap.

But having to figure out why certain services/java based apps no longer start and having to edit the config files in the right location and order is annoying at best.
Comments
Regression test exists.
26-07-2017
It is a critical pre approval to push the fix into the mainline repo.
28-04-2017
Fix Request Without this fix we have a regression of a failure to start java when using a not uncommon command line option. The fix is low risk. The removed defaulting of ParallelGCThreads is covered by a later defaulting of that option to the same value. The proposed change also includes a test to verify simple use of the option doesn't crash and has some of the expected effects. Webrev of the fix is here: http://cr.openjdk.java.net/~kbarrett/8179084/hotspot.00/
25-04-2017
There don't seem to be any tests at all for -XX:+AggressiveHeap. Also, the documentation for it seems pretty thin. I think [~dholmes] suggestion of removing the FLAG_SET_DEFAULT(ParallelGCThreads, ...) form from set_aggressive_heap_flags() is okay for JDK 9 and 8u. In the longer term, we might want to do something a little more principled. -XX:+AggressiveHeap is handled in a rather strange way. It is not handled as a normal flag option. (-XX:-AggressiveHeap isn't even recognized.) Instead, while running down the list of CLAs, we call set_aggressive_heap_flags as soon as that option is seen. So the problem at hand can't even be worked around by adding -XX:ParallelGCThreads=<n> to the end of the options; it must preceed the AggressiveHeap option. set_aggressive_heap_flags sets a number of options. Some of that looks plausible. But where the value is obtained from something not directly and obviously available at that point, there may be problems. The too early attempt to get the active processor count is one such. I think there might be some similar problems in the heap and new space sizing.
24-04-2017
Calling sequence: os::initial_active_processor_count() => Abstract_VM_Version::nof_parallel_worker_threads() ===> Abstract_VM_Version::calc_parallel_worker_threads =====> Abstract_VM_Version::parallel_worker_threads() =======> Arguments::set_parnew_gc_flags =========> Arguments::set_cms_and_parnew_gc_flags() ===========> Arguments::set_gc_specific_flags() =========> Arguments::set_gc_specific_flags() =======> Arguments::set_parallel_gc_flags() =========> Arguments::set_gc_specific_flags() =======> Arguments::set_g1_gc_flags() =========> Arguments::set_gc_specific_flags() =======> Arguments::parse_each_vm_init_arg() // JDK 9 factors code into set_aggressive_heap_flags() rather than directly inline All except the last end at set_gc_specific_flags(): => ArgumentsExt::set_gc_specific_flags() ===> Arguments::apply_ergo() So normally the initial_active_processor_count is only used in this context after Arguments::apply_ergo(), hence os::init_before_ergo() was a suitable location to do initialization. For Arguments::parse_each_vm_init_arg() => Arguments::parse_vm_init_args() ===> Arguments::parse() => Arguments::parse_options_environment_variable() ===> Arguments::parse_java_options_environment_variable() ===> Arguments::parse_java_tool_options_environment_variable() =====> Arguments::parse_vm_init_args() ======> Arguments::parse() Will now check initialization possibilities for os::initialize_initial_active_processor_count().
24-04-2017
The code actually sets ParallelGCThreads twice when AggressiveHeap is used: // Arguments::parse_each_vm_init_arg() // Enable parallel GC and adaptive generation sizing FLAG_SET_CMDLINE(bool, UseParallelGC, true); FLAG_SET_DEFAULT(ParallelGCThreads, Abstract_VM_Version::parallel_worker_threads()); <= will be zero at this point and is cached then later: // Arguments::set_gc_specific_flags() if (UseParallelGC || UseParallelOldGC) { set_parallel_gc_flags(); which in turn does: FLAG_SET_DEFAULT(ParallelGCThreads, Abstract_VM_Version::parallel_worker_threads()); <= Safe to call here but zero already cached. if (ParallelGCThreads == 0) { jio_fprintf(defaultStream::error_stream(), "The Parallel GC can not be combined with -XX:ParallelGCThreads=0\n"); vm_exit(1); } so it seems to me that the simplest fix here is to delete the first: FLAG_SET_DEFAULT(ParallelGCThreads, Abstract_VM_Version::parallel_worker_threads()); as we will set it later anyway.
24-04-2017
os::initialize_initial_active_processor_count() calls os::active_processor_count() to do the initialization. Analysis of the per-OS implementations show the following VM initialization dependencies: - AIX: None (uses sysconf) - BSD: None (already initialized during os::init() - Linux: None (uses sched_getaffinity) - Solaris: None (uses sysconf and pset_bind) - Windows: None (uses GetProcessAffinityMask) So it is perfectly safe to move os::initialize_initial_active_processor_count() into os::init() (specific placement is needed on BSD). The downside of this is that os::init() is platform specific so we have to change all 5 implementations. Alternatively we could call os::initialize_initial_active_processor_count() directly in create_vm immediately after os::init().
24-04-2017
Yes this is caused by JDK-8147910. The AggressiveHeap option is processed before we initialize the value: // Parse arguments jint parse_result = Arguments::parse(args); <= AggressiveHeap processed in here if (parse_result != JNI_OK) return parse_result; os::init_before_ergo(); <= initial processor count initilaized here. Need to re-examine the initialization paths to see where we can move this to. Not sure why it isn't done in os::init().
24-04-2017
This is a regression introduced in 8u131 b06/ 9ea b131 == jdk/8u131/fcs/b11/binaries/linux-x64/bin/java -XX:+AggressiveHeap -version The Parallel GC can not be combined with -XX:ParallelGCThreads=0 == 8u121 - Pass 8u131 - Fail 9 ea b130 - Pass 9 ea b131 - fail 9 ea b166 - Fail == Suspecting this could be introduced by JDK-8147910
21-04-2017