Java 25 File I/O (PPSA99866)

[NevesHenry]

Existing thread

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

Java 25 File I/O (PPSA99866)

Content version tested

01.000.004. It is the PPSA99866-file-io-pass-01.000.004.zip release of mateo-cogeanu/ps5-java25-homebrew. Its SHA-256 is a9a0159806034da9dc5f1164046c9c05e9d86340172fca9a071b8ee497304b15, which matches the release's IO_Probe-SHA256SUMS-01.000.004.txt. It is one of the project's OpenJDK 25 probe apps (HotSpot 25.0.4, interpreted mode) and reports ALL PASS on a PS5 (firmware 13.60).

AnyPS5 commit

main 6ef5a1c

System tested

Ubuntu 22.04 on WSL2, Ryzen 9 7900, lavapipe, GCC 13, Release

Furthest point

Menu

What is shown

The probe menu shows and the runtime constructors run (Runtime constructors: 359; page size: 16384 bytes). The core image opens. Picking the first row stops inside JNI_CreateJavaVM, while HotSpot protects the main thread's stack guard pages:

STAGE: Calling real JNI_CreateJavaVM
[0.404s][warning][os,thread] Attempt to protect stack guard pages failed (0x00007fffff7ff000-0x00007fffff80b000).

hs_err_java25_*.log:

# Native memory allocation (mprotect) failed to protect 49152 bytes. Error detail: memory to guard stack pages
#  Internal Error (stackOverflow.cpp:113), pid=114037, tid=100001

While writing that report, the VM reads kern.boottime, and the process ends with:

terminate called after throwing an instance of 'std::runtime_error'
  what():  sysctl_nid_postfix: unsupported MIB 1.21

Current blocker

On Linux, the thread stacks that scePthreadAttrGet / pthread_attr_get_np / sceKernelIsStack report do not start on a 16 KiB page boundary. HotSpot aligns its guard zone down to its 16 KiB page, so mprotect lands below the reported stack and is refused.

Open pull request: #9684 (aligns created and adopted thread stacks on Linux).

The same stop shows on Java 25 ZIP/JAR (PPSA99867, 01.000.000) and Java 25 Dependencies (PPSA99873, 01.000.001) from the same project. With the fix, all three titles pass every row, and the VM is created and destroyed.

How it was run

1. Unpacked the release zip and unwrapped eboot.bin from its fake-signed SELF.
2. Relinked with relinker --trap-sdk-syscall-stubs --registry.
3. Ran app.elf with the rows driven by tools/pad_input.py, using APS5_PAD_INPUT_FILE.

AI-assisted report (Claude Code).