RFR: 8350621: Code cache stops scheduling GC [v5]

[Kim Barrett]

On Mon, 21 Jul 2025 07:25:06 GMT, Thomas Schatzl <tschatzl at openjdk.org> wrote:

>> Hi all,
>> 
>>   please review this change to avoid CodeCache triggered GCs temporarily being ignored.
>> 
>> In particular, G1 does not make sure when its `collect()` method is called during a concurrent cycle, that a `Remark` pause that does code unloading etc. actually occurs after that request. This makes it so that some internal flag is not reset appropriately (via some callback to the caller). After this event, there will never be another attempt by the compiler threads to issue a garbage collection. The compiler threads are stuck until the next code unloading (caused by e.g. a regular concurrent cycle being triggered).
>> 
>> The [original PR](https://github.com/openjdk/jdk/pull/23656) also stated a similar with Parallel GC - however due to recent changes (https://bugs.openjdk.org/browse/JDK-8192647 and others) I do not think it is possible any more that the `collect()`call can silently be ignored. (I.e. gclocker could abort a full gc that is the only reason why code cache unloading and that callback will not occur as expected).
>> 
>> So for G1, the change makes the caller of `collect()` (i.e. the compiler thread) wait until that situation passes and retry. With that the compiler thread is sure that the callback the compiler threads are waiting for occurs. This does have the disadvantage that that compiler thread may not be available for compilation any more for a bit.
>> 
>> Testing: tier1-5, test case passing, failing before
>> 
>> Thanks,
>>   Thomas
>
> Thomas Schatzl has updated the pull request incrementally with one additional commit since the last revision:
> 
>   * kbarrett review, remove unsupported gc causes for that branch

Still not quite...

src/hotspot/share/gc/g1/g1CollectedHeap.cpp line 1892:

> 1890:       assert(cause == GCCause::_g1_humongous_allocation ||
> 1891:              cause == GCCause::_g1_periodic_collection ||
> 1892:              cause == GCCause::_wb_breakpoint, "Unsupported cause %s", GCCause::to_string(cause));

`_wb_breakpoint` is unexpected here too, since it's completely handled by an earlier clause in the if-ladder.
The first two are the entire residue that should be able to get here.

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

Changes requested by kbarrett (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/26189#pullrequestreview-3039578093
PR Review Comment: https://git.openjdk.org/jdk/pull/26189#discussion_r2220169890