RFR: 8356716: ZGC: Cleanup Uncommit Logic [v4]

[Joel Sikström]

On Tue, 27 May 2025 06:08:10 GMT, Axel Boldt-Christmas <aboldtch at openjdk.org> wrote:

>> [JDK-8350441](https://bugs.openjdk.org/browse/JDK-8350441) required changing the way ZGC handle memory uncommitting (returning physical memory to the OS). Previously ZGC tracked how recently used memory was on a ZPage level. [JDK-8350441](https://bugs.openjdk.org/browse/JDK-8350441) did away with the ZPage abstraction for unused memory. But because of this ZGC does not have a convenient way of tracking the usage of a specific memory range. Instead [JDK-8350441](https://bugs.openjdk.org/browse/JDK-8350441) opted to keep a watermark in the cache unused mapped memory, to keep track of the amount of memory that was not used within the last ZUncommitDelay, and use this when deciding how much to uncommit.
>> 
>> Because this measurement is not as granular as previously, and because uncommitting memory is something we want to do conservatively, as a response to low memory utilization, [JDK-8350441](https://bugs.openjdk.org/browse/JDK-8350441) was written with the intent to spread out the uncommitting over some time interval.
>> 
>> The actual implementation in [JDK-8350441](https://bugs.openjdk.org/browse/JDK-8350441) has a few issues which this RFE tries to address:
>>   * Missing wait, the uncommitting is not actually spread out, but happens all at once.
>>   * Reactivity, if the process starts using memory that was below the previous watermark, uncommitting should stop.
>>   * Structure, the current implementation has a lot of different dependencies and has state spread out over multiple classes. Refactor to keep the logic contained to the ZUncommitter, and provide better named facilitating functions on the ZPartition and ZMappedCache. And make the lifecycle of ZUncommitter more explicit.
>>   * Events, overhaul the JFR uncommit events to be sent (and track the time for) a chunk of uncommits without any waits.
>> 
>> An alternative discussed has been to do uncommitting based on GC triggers rather than a periodically. So rather than using ZUncommitDelay, we could have our proactive GCs actually trigger and track uncommitting. This might be a future RFE, but it was not attempted here as it would change user facing APIs. [JDK-8329758](https://bugs.openjdk.org/browse/JDK-8329758) will more than likely overhaul the uncommit triggers as well, and the whole concept of ZUncommitDelay and having to tune how to uncommit will go away.
>
> Axel Boldt-Christmas has updated the pull request incrementally with six additional commits since the last revision:
> 
>  - Avoid excessive logging if ZUncommitDelay == 0
>  - Move uncommit logic from MappedCache to Uncommitter + cleanup and comment
>  - Better cycle activate / deactivate scoping
>  - Remove newline
>  - Cleanup time logging
>  - Comment describing `uncommitted == 0`

I really like how the Mapped Cache is now decoupled from the uncommitter, except for the notion of the watermark, which is perfectly fine IMO.

Some more thoughts:

src/hotspot/share/gc/z/zUncommitter.cpp line 212:

> 210:     // We are stopping
> 211:     return false;
> 212:   }

I'm unsure if we need this check. From what I can see, `activate_uncommit_cycle` is called right after `while(wait(...)) {` in `ZUncommitter::run_thread`, where the `should_continue()` check is essentially done in `wait(...)` by returning `!_stop`.

src/hotspot/share/gc/z/zUncommitter.cpp line 247:

> 245:   _uncommitted = 0;
> 246: 
> 247:   // Reset cache for next uncommit cycle

Suggestion:

  // Reset watermark for next uncommit cycle

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

PR Review: https://git.openjdk.org/jdk/pull/25198#pullrequestreview-2870706463
PR Review Comment: https://git.openjdk.org/jdk/pull/25198#discussion_r2108958713
PR Review Comment: https://git.openjdk.org/jdk/pull/25198#discussion_r2108945795