RFR: 8251158: Implementation of JEP 387: Elastic Metaspace [v11]

[Richard Reingruber]

On Mon, 5 Oct 2020 13:38:03 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> Hi all,
>> 
>> this is the continuation of the ongoing review for the JEP387 implementation (last rounds see [1] [2]). Sorry for the
>> delay, had vacation then the entrance of Skara delayed things a bit.
>> For the delta diff please see [3].
>> 
>> This is the first time I do a large PR after Skara, so if something is wrong please bear with me. I cannot answer all
>> feedback individually in this PR body, but I incorporated almost all into the new revision.
>> What changed since the last version:
>> 
>> - I renamed most metaspace files back to the original naming scheme or to something similar, hopefully capturing the
>>   group consent.
>> 
>> - I changed the way allocation guards are checked if MetaspaceGuardAllocations is enabled. Before, I would test for
>>   overwrites upon CLD destruction, but since that check was subject to VerifyMetaspaceInterval it only ran for every nth
>>   class loader which made it rather pointless. Now I run it always.
>> 
>> - I also improved the printout on block corruption, and log block corruption unconditionally before asserting.
>> 
>> - I also fixed up and commented the death test which tests for allocation overwriters (test_allocationGuard.cpp)
>> 
>> Side note, I find the corruption check very useful but if you guys think it is too much I still can remove the feature.
>> 
>> - In ChunkManager::purge() I improved the comments after discussions with Leo.
>> 
>> - I fixed a bug with VerifyMetaspaceInterval: if set to 1 the "SOMETIMES" sections were supposed to fire always, but due
>>   to a one-off error they only fired every second time. Now, if -XX:VerifyMetaspaceInterval=1, the checks really run
>>   every time.
>> 
>> - Fixed indentation issues as Leo requested
>> 
>> - Rewrote the condition and the assert in VirtualSpaceList::allocate_root_chunk() as Leo requested
>> 
>> - I removed the "can_purge" logic from VirtualSpaceList. The list does not need to know. It just should iterate all nodes
>>   and attempt purging, and if a node does not own its ReservedSpace, it refuses to be purged. That is simpler and more
>>   flexible since it allows us to have list with purge-able and non-purge-able nodes.
>> 
>> - and various smaller fixes, mainly on request of Leo.
>> 
>> @lkorinth:
>> 
>>> VirtualSpaceNode.hpp
>>>
>>>102   // Start pointer of the area.
>>>103   MetaWord* const _base;
>>>
>>>How does this differ from _rs._base? Really needed?
>>>
>>>105   // Size, in words, of the whole node
>>>106   const size_t _word_size;
>>>
>>>Can we not calculate this from _rs.size()?
>> 
>> You are right, _base and _word_size are directly related to the underlying space. But I'd prefer to leave it the way it
>> is. Mainly because ReservedSpace::_base and ::_size are nonconst and theoretically can change under me. It is highly
>> improbable but I'd like to know. Note that VirtualSpaceNode::verify checks that.  Should we clean up ReservedSpace at
>> some point and make those members const - as they should be - then I would rewrite this as you suggest.
>> Thanks, again, for all your review work!
>> 
>> ------
>> 
>> 
>> [1] https://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2020-August/041162.html
>> [2] https://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2020-September/041628.html
>> [3] https://github.com/openjdk/jdk/commit/731f795bc0c1c502dc6cac8f866ff45a15bdd02d
>
> Thomas Stuefe has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Review Feedback Richard 1

Hi Thomas,

this is batch 3 with another 3 points. I'm sending it now because I think I cannot reply to your comments otherwise.

Cheers, Richard.

src/hotspot/share/memory/metaspace/metaspaceSettings.hpp line 45:

> 43:   // The default size of a non-class VirtualSpaceNode (unless created differently).
> 44:   // Must be a multiple of the root chunk size.
> 45:   static const size_t VirtualSpaceNodeDefaultWordSize = chunklevel::MAX_CHUNK_WORD_SIZE * 2; // lets go with 8mb
> virt size. Seems a good compromise betw. virt and mapping fragmentation.

Member name do not match convention. See doc/hotspot-style.md:206:

`VirtualSpaceNodeDefaultWordSize`
`VirtualSpaceNodeReserveAlignmentWordSize`
`EnlargeChunksInPlace`

src/hotspot/share/memory/metaspace/metaspaceArena.cpp line 239:

> 237:       UL2(trace, "taken from fbl (now: %d, " SIZE_FORMAT ").",
> 238:           _fbl->count(), _fbl->total_size());
> 239:       // Note: Space in the freeblock dictionary counts as already used (see retire_current_chunk()) -

`retire_current_chunk()` does not exist.

src/hotspot/share/memory/metaspace/metaspaceArena.cpp line 465:

> 463:   assert(p != NULL && word_size > 0, "Sanity");
> 464:   bool found = false;
> 465:   if (!found) {

The if is redundant. The condition is always true.

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

Changes requested by rrich (Committer).

PR: https://git.openjdk.java.net/jdk/pull/336