RFR (XS): JDK-8058967 metaspace/shrink_grow/CompressedClassSpaceSize fails with OOM: Compressed class space

[Joseph Provino]

Here is the updated webrev:

http://cr.openjdk.java.net/~jprovino/8058967/webrev.01

thanks.

joe

On 5/15/2015 4:35 PM, Kim Barrett wrote:
> On May 15, 2015, at 3:16 PM, Joseph Provino <joseph.provino at oracle.com> wrote:
>> On 5/15/2015 2:29 PM, Jon Masamitsu wrote:
>>> I think the statement of the policy should be updated.
>> How does this sound?  Should the comment be with the "if" statement or
>> where _small_chunk_limit is declared?
>>
>>   /*
>>    * The policy is to allocate up to _small_chunk_limit small chunks
>>    * after which only medium chunks are allocated.  This is done to
>>    * reduce fragmentation.  In some cases, this can result in a lot
>>    * of small chunks being allocated to the point where it's not
>>    * possible to expand.  If this happens, there may be no medium chunks
>>    * available and  OOME would be thrown.  Instead of doing that,
>>    * if the allocation request size fits in a small chunk, an attempt
>>    * is made to allocate a small chunk.
>>    */
>>   if (next == NULL &&
>>       word_size + Metachunk::overhead() <= small_chunk_size() &&
>>       grow_chunks_by_words == medium_chunk_size()) {
>>     /*
>>      * There are no medium chunks available but a small chunk is big enough.
>>      * See if a small chunk is available.
>>      */
>>     next = get_new_chunk(word_size, small_chunk_size());
>>   }
>>
> After further consideration and offline discussion, I'm ok with this
> approach.  There are several open CRs to improve metaspace allocation,
> some of which might render this change moot.  However, as a short term
> fix for an intermittently failing test, this seems ok.
>
> So the above version looks good to me.
>
> https://bugs.openjdk.java.net/browse/JDK-8076476
> Coalesce Metachunks in the Metaspaces
>
> https://bugs.openjdk.java.net/browse/JDK-8035945
> Small Metaspace allocation requests can fail even though SmallChunks are available
> This might be a duplicate of 8058967.
>