Java 25 LAN / DNS (PPSA99869)

[ahmad3a4]

Existing thread

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

Title

Java 25 LAN / DNS (PPSA99869)

Content version tested

01.000.000, release LAN_DNS_Probe-v01.000.000 of mateo-cogeanu/ps5-java25-homebrew (PPSA99869-lan-dns-pass-01.000.000.zip, SHA-256 3cfeed2f321334e742fa93070e6983bed201332f7837d907af2db52d547c11cc, matching the release's SHA256SUMS). It is one of the project's OpenJDK 25 probe apps and ships its own sce_module/libc.prx.

AnyPS5 commit

main 0aa4cb1 (first report), rechecked on main e34c479

System tested

Windows 11 Pro (26200), RTX 3060 Ti (NVIDIA 617.14), WinLibs GCC 15.2.0 r7, Release

Furthest point

On main e34c479: menu. Stages 1, 2 and 5 pass. The DNS stage passes where file I/O is fast and times out in the title's 8-second limit where it isn't (see the comments).

What is shown

Without --trap-sdk-syscall-stubs, relinking stops at the payload SDK's syscall stubs:

FAIL: Forbidden syscall instruction at code offset 0xd98027 (offset 0xd98027)

With --trap-sdk-syscall-stubs (#3804) it relinks. tools/import_audit.py finds 459 of 470 imports implemented, 7 that resolve to throwing stubs (__tls_get_addr from two libraries, posix_spawn, execve, execvp, cpuset_getaffinity, kevent) and 4 absent:

NID	Name	Library
UkEV4hvMo1E	sysctlnametomib	libkernel
psx0YiAKm7k	fegetenv	libSceLibcInternal
+hFjyohmOU0	mincore	libkernel
KZ-4qlqlpmo	sigsuspend	libkernel

The loader stops at the first one:

FAIL: unresolved ELF import UkEV4hvMo1E from libkernel.prx

Current blocker

None for start-up: the four absent imports are merged (#5502 sysctlnametomib, #5563 fegetenv, #5643 mincore, #5666 sigsuspend), and so is #5187 (getrlimit(RLIMIT_NOFILE)), which the VM needs next.

Open: #5831, because the title's trace log is damaged by concurrent O_APPEND writes on Windows. The DNS stage depends on file I/O speed, because the title opens, appends to and closes its trace file once per traced call.

How it was run

The release zip unpacked, eboot.bin and sce_module/libc.prx unwrapped from their fake-signed SELF containers. Relinked with --windows --to-intel --trap-sdk-syscall-stubs --registry, then run from the output folder with app0 mirrored and no data folder. The menu is driven with the default keyboard bindings (arrows and Enter).

AI-assisted: yes

[ahmad3a4]

Progress on Windows 11, main 9a3e2ac (#5187 merged) with #5502 (now merged), #5563, #5643 and #5666 stacked, relinked with --trap-sdk-syscall-stubs. tools/import_audit.py reports 0 absent imports, and the title runs its menu ("Java 25 - LAN / DNS 01.000.000"). The stages, driven with the default keyboard bindings:

Stage	Result
Initialize embedded core runtime	PASS: 359 constructors, page size 16384
Create Java 25 VM (interpreter mode)	PASS: real JNI_CreateJavaVM, -Xint, 128 MiB heap, code 0
LAN TCP / UDP echo	not run (needs the author's Mac helper)
DNS resolve example.com	FAIL: java.io.IOException: DNS lookup exceeded 8 seconds
Destroy Java VM	PASS: real DestroyJavaVM

The DNS stage times out in the title's own 8-second limit. During those 8 seconds its trace (/download0/java25-vm-trace.log) shows only metaspace allocation and zeroing, with no network call yet. The VM is still loading classes in -Xint, and it appends one trace line per traced memcpy/memcmp/allocation, about 280,000 sceKernelOpen calls in the first 45 seconds. This looks like speed, not a missing function. Not checked further yet.

The VM also prints [warning][os,thread] Attempt to allocate stack guard pages failed. during creation. It is not fatal, and not looked into yet.

While the DNS stage runs, two threads append to that trace through separate O_APPEND descriptors. On Windows, 86 lines lost their start and 99 lost their newline. #5831 makes those appends atomic; with it, no line is cut and no newline is lost.

AI-assisted.

[pushkarreddyy]

Taking the follow-up for setrlimit(RLIMIT_NOFILE) on Linux (clamping rlim_cur <= rlim_max matching FreeBSD sys_setrlimit semantics to unblock HotSpot VM startup when rlim_max is large); branch and PR in progress.

[pipe1os]

The DNS stage times out in the title's own 8-second limit.

It doesn't on my machine: i5-12400F, Windows 11 Pro (26300), WinLibs GCC 15.2.0 r7, Release, relinked with --trap-sdk-syscall-stubs, rows driven through APS5_PAD_INPUT_FILE.

• main e34c47921 (feat(libkernel): implement sysctlnametomib for the sysctl names AnyPS5 serves #5502, feat(libc): implement fegetenv #5563, feat(libkernel): implement mincore #5643, feat(libkernel): implement sigsuspend on the pending-signal set #5666 and feat(libkernel): report RLIMIT_NOFILE through getrlimit #5187 merged): STAGE: DNS: resolve example.com, then STAGE: LAN/DNS fixture returned normally, and the next row ends with DestroyJavaVM complete.
• The older main 107190d11 with those five stacked (feat(libkernel): report RLIMIT_NOFILE through getrlimit #5187 at a915c0723): same result.

The LAN row times out here too, since it points at the author's Mac. So I don't see a regression between those two commits. The probe's stage lines carry no timestamps, so I can't say how long the DNS stage took here.

[ahmad3a4]

Thanks for running it. On this machine the same main e34c479 still times out in the DNS stage, 3 of 3 runs: from a folder under %TEMP%, from C:\dev, and with AnyPS5's output sent to NUL. So it's this machine's file I/O speed, not AnyPS5.

The probe opens, appends to and closes /download0/java25-vm-trace.log once per traced call, about 92,000 times during the DNS stage. A standalone loop of raw CreateFileW + WriteFile + CloseHandle on the same drive takes about 97 µs per cycle here (Windows Defender real-time protection on). That is slightly more than the ~87 µs per trace line the title gets through AnyPS5, so AnyPS5 adds next to nothing on top of it.

I've updated the thread: on main the title starts and reaches its menu, and the DNS stage passes where file I/O is fast enough.

AI-assisted.