RFR: 8307567: Avoid relocating global roots to metaspaceObjs in CDS dump

[Matias Saavedra Silva]

On Sat, 6 May 2023 17:11:54 GMT, Ioi Lam <iklam at openjdk.org> wrote:

> The functions `ArchiveBuilder::relocate_roots()` and `ArchiveBuilder::relocate_vm_classes()` relocate global roots such as `Universe::_typeArrayKlassObjs[]` and `vmClasses::_klasses[]` to point to locations inside the CDS "buffer" space.
> 
> As a result, the VM is no longer usable after CDS dumping has finished.
> 
> Previously, we tried to save some of the VM states and restore them after CDS dump has finished. See [JDK-8264735](https://bugs.openjdk.org/browse/JDK-8264735), which added functions like `FileMapInfo::clone_shared_path_table()`.
> 
> However, this is too cumbersome and doesn't cover all cases. The intention of the global root relocation is to make it easy to write the roots using `WriteClosure::do_ptr()`, which could handle only pointers into the "buffer" space. If we call `get_buffered_addr()` inside `WriteClosure::do_ptr()`, then there's no need to modify the global roots.
> 
> Notes for reviewers:
> 
> - Start with changes in `WriteClosure::do_ptr()`
> - Some code in SystemDictionaryShared also needed to call `get_buffered_addr()`
> - I added new function `ArchiveBuilder::write_pointer_in_buffer()` to simplify writing pointers into the CDS buffer
> -  `ArchiveBuilder::relocate_roots()`, `ArchiveBuilder::relocate_vm_classes()` and `FileMapInfo::clone_shared_path_table()` are removed.
>     - I am leaving `CloneDumpTimeClassTable` and `CloneDumpTimeLambdaProxyClassTable` for now. I plan to remove them in a follow up RFE after more testing.

I just have a few questions but otherwise this looks good!

src/hotspot/share/cds/archiveBuilder.cpp line 458:

> 456: 
> 457: void ArchiveBuilder::remember_embedded_pointer_in_gathered_obj(address src_obj, address* field_addr) {
> 458:   // field_addr must be a valid field inside src_obj.

Is there a way to check this with an assert?

src/hotspot/share/classfile/systemDictionaryShared.cpp line 1169:

> 1167:   bool do_entry(LambdaProxyClassKey& key, DumpTimeLambdaProxyClassInfo& info) {
> 1168:     int len = info._proxy_klasses->length();
> 1169:     InstanceKlass* last_buff_k = nullptr;

Nice, this is a much cleaner approach!

src/hotspot/share/memory/iterator.hpp line 359:

> 357:   }
> 358: 
> 359:   // Useful aliases

This is a nice addition but it might be buried under the rest of this PR. Would it be worth making this a separate enhancement?

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

Marked as reviewed by matsaave (Committer).

PR Review: https://git.openjdk.org/jdk/pull/13854#pullrequestreview-1421158727
PR Review Comment: https://git.openjdk.org/jdk/pull/13854#discussion_r1190249712
PR Review Comment: https://git.openjdk.org/jdk/pull/13854#discussion_r1190256133
PR Review Comment: https://git.openjdk.org/jdk/pull/13854#discussion_r1190256871