RFR: 8029186: regression-hotspot nightly failure: assert(FLAG_IS_DEFAULT(MaxNewSize) || MaxNewSize < MaxHeapSize) failed

[Stefan Johansson]

Thanks Jesper for looking at this.

I actually found an issue myself when going over the code. The problem 
is in initialize_flags for the TwoGenerationalPolicy, where I verify 
that the desired new_size is within the allowed bounds. I realized that 
I need to check against the flag MaxNewSize since the member might not 
have been initialized yet. Below is the diff for the change and an 
updated webrev can be found here:
http://cr.openjdk.java.net/~sjohanss/8029186/webrev.01/

---
@@ -422,7 +422,9 @@
    if (FLAG_IS_CMDLINE(OldSize) && !FLAG_IS_CMDLINE(NewSize)) {
      if (OldSize < _initial_heap_byte_size) {
        size_t new_size = _initial_heap_byte_size - OldSize;
-      if (new_size >= _min_gen0_size && new_size <= _max_gen0_size) {
+      // Need to compare against the flag value for max since 
_max_gen0_size
+      // might not have been set yet.
+      if (new_size >= _min_gen0_size && new_size <= MaxNewSize) {
          FLAG_SET_ERGO(uintx, NewSize, new_size);
          _initial_gen0_size = NewSize;
        }
---

Thanks,
Stefan

On 2014-03-25 17:33, Jesper Wilhelmsson wrote:
> Ship it!
> /Jesper
>
> Stefan Johansson skrev 25/3/14 16:29:
>> Hi,
>>
>> Please review this fix for:
>> https://bugs.openjdk.java.net/browse/JDK-8029186
>>
>> Webrev:
>> http://cr.openjdk.java.net/~sjohanss/8029186/webrev.00/
>>
>> Summary:
>> There have been a lot of bugs due to corner cases with heap sizing 
>> flags. This
>> change strives to fix this particular problem and at the same time 
>> somewhat
>> simplify the collector policy. I've also added some assertions to 
>> ensure that
>> the sizing of the heap is more correct.
>>
>> Testing:
>> * JPRT sanity and JTREG
>> * Adhoc aurora run for tmtools tests
>> * Script that verifies that we don't get assertions during startup 
>> for different
>> combinations of NewSize, MaxNewSize, OldSize, Xms and Xmx.
>>
>> Thanks,
>> StefanJ