[jdk11u-dev] RFR: 8303215: Make thread stacks not use huge pages

[Paul Hohensee]

On Tue, 21 Oct 2025 11:11:57 GMT, Takuya Kiriyama <tkiriyama at openjdk.org> wrote:

> Hi All,  
> 
> I would like to backport the fix for bug JDK-8303215 from JDK 17 to JDK 11. This bug addresses an issue with Transparent Huge Pages (THP) and thread stack allocation.
> Backport is unclean, 
> 
> **1. Backport:**
> The fix in os_linux.cpp cannot be cleanly backported.
> Since jdk11 does not include JDK-8244752 and JDK-8256155, it differs from the fix in jdk17. It uses setup_large_page_size() instead of scan_default_large_page_size().
> 
> 
> **2. Bug Reproduction:**
> The bug was successfully reproduced in Linux x86_64 with THP enabled (`/sys/kernel/mm/transparent_hugepage/enabled = always`).
> Reproduction was confirmed using the following dummy thread program: 
> 
> public class ThreadStackHugePageTest {
>         public static void main(String[] args) {
>                 final int NUM_THREADS = 1000;
>                 final long SLEEP_SECONDS = 3600;
>                 final long SLEEP_MILISECONDS = SLEEP_SECONDS * 1000;
> 
>                 List<Thread> threads = new ArrayList<>();
> 
>                 for (int i = 0; i < NUM_THREADS; i++) {
>                         Runnable task = () -> {
>                                 try {
>                                         Thread.sleep(SLEEP_MILISECONDS);
>                                 } catch (InterruptedException e) {
>                                         Thread.currentThread().interrupt();
>                                 }
>                         };
> 
>                         Thread thread = new Thread(task);
>                         threads.add(thread);
> 
>                         thread.start();
>                 }
> 
>                 System.out.println("all threads are started");
> 
>                 for (Thread thread : threads) {
>                         try {
>                                 thread.join();
>                         } catch (InterruptedException e) {
>                                 Thread.currentThread().interrupt();
>                         }
>                 }
>         }
> }
> 
> The fix significantly reduces the number of anonymous regions with fully occupied RSS.
> 
> **3. Regression Testing:**
> I ran the `test/hotspot` tests on Linux x64 as a regression test. No failures were found.
> 
> Thank you.

Yes, those two. But they in turn have related issues.

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

PR Comment: https://git.openjdk.org/jdk11u-dev/pull/3106#issuecomment-3435457787