RFR: 8250989: Consolidate buffer allocation code for CDS static/dynamic dumping

Ioi Lam iklam at openjdk.java.net
Tue Feb 2 00:05:46 UTC 2021


On Mon, 1 Feb 2021 18:31:08 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> Before this PR, when a static CDS archive is dumped, we would unconditionally allocate a 4GB ReservedSpace and use that as the buffer for writing the archive content. We usually don't need such a big ReservedSpace. this also unnecessarily complicates Metaspace::global_initialize().
>> 
>> After this PR, both we first load all the classes and then allocate a buffer with an appropriate size.
>> 
>> I also simplified the pointer relocation code that prepares the archive to be mapped at the "requested location" (usually 0x800000000), and improved the comments.
>> 
>> We used to have lots of special cases for adjusting pointers during dynamic dump. All of those are removed.
>> 
>> Reviewers, please start with
>> - Header comments in archiveBuilder.hpp that describe the overall process of archive dumping
>> - `MetaspaceShared::initialize_for_static_dump()`
>> - `ArchiveBuilder::reserve_buffer()`
>> - Comments around `RelocateBufferToRequested` in archiveBuilder.cpp
>
> src/hotspot/share/memory/metaspace.cpp line 681:
> 
>> 679: 
>> 680:   // If UseCompressedClassPointers=1, we have two cases:
>> 681:   // a) if CDS archive has been mapped, it will create the ccs
> 
> This comment is confusing. So, at this point the ccs archive has already be mapped? Where?

I've changed it to:

  // a) if CDS archive(s) have been mapped, we would have already reserved the
  //    ccs (immediately above the mapped archives) and initialized
  //    the CompressedKlassPointers encoding.

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

PR: https://git.openjdk.java.net/jdk/pull/2296


More information about the hotspot-runtime-dev mailing list