[Ondra97]

What

Closing the native VLC v0.4 (PPSA85300) window aborts the process, as reported in the #5791 validation. SDL_QUIT calls LibcRequestExit_nid_postfix, whose detached thread currently runs the guest finalizers. VLC still has a joinable C++ thread object; its destructor calls terminate() and then abort().

The asynchronous host exit request now uses LibcExitWithoutHandlers_nid_no_patch from #5236. It runs the registered runtime shutdown callbacks, flushes the host streams and terminates without guest or host exit handlers. The new guest_requested_exit mode reuses the existing exit fixture: shutdown must complete, another guest thread must stay parked during shutdown, both kinds of exit handler must remain uncalled, and a second request with status 7 must leave the first request's status 0 intact. Its parent checks the child's actual exit code.

#5236 is merged into main. This PR now contains only the host-window-close change and its regression (three files). #5978 remains a separate dependent follow-up for explicit guest exit() and return from main.

Tested

Rebase validation on Windows 11 Pro build 26300, WinLibs GCC 15.2.0, Ninja Release: head eb8146b73, based on main 0ae1deb85. #5236 is merged into this base.

• Full default build, including the patched runtime libraries, passed.
• All 20 focused tests passed, with no failures or skips: eight exit modes, eleven related shutdown/lifecycle cases and patched_libraries.
• python tools/check_conventions.py --base origin/main, git diff --check and python tools/check_main_sync.py --base origin/main passed; clean merge, 0 behind at publication.
• git range-diff confirms that the host-window-close patch is unchanged. Production Process.cpp is byte-identical to the original title-tested PR head. The PR now has one commit and three changed files; the merged prerequisite is no longer carried as a separate commit.

No new title run or Linux run was performed for this rebase. The original title validation below was performed before the rebase:

Windows 11 Pro build 26300, WinLibs GCC 15.2.0, Ninja Release. Original head e0748f372, based on main 5a6ad0ed3 plus #5236:

• Full default build and patched runtime libraries passed. All 19 relevant tests passed in 2.37 s: asynchronous requested exit, SystemService exit/parameters, quick/immediate exit, return from main, relaunch/nested exit, KillApp/status/parameters, application heap ordering, thread atexit, user context/exit and four AGC shutdown tests.
• Negative control: the new child mode on fix(libSceSystemService): end LoadExec exit without running guest exit handlers #5236 alone still prints Guest shutdown completed followed by Guest exit handler ran; CTest fails in 0.17 s. That PR alone does not fix host window closure.
• Original official VLC v0.4, converted with --windows --registry --windows-diagnostics --windows-fixed-base 0x400000, including its own libc module. Executable ELF SHA-256: eec51934aa2debed99152c3996f39a41be25e17d9ca14de01cee2d17b62d8e3f. All 132 installed runtime files were checked against the build by SHA-256.
• On main 550764236, a visible 20 s run ends with 0xc0000409 after CloseMainWindow() returns true. GDB reaches LibcExit -> CxaFinalize; the guest destructor at 0x2261f30 calls termination when its first word is nonzero, and abort_nid_postfix is reached from 0x22c49af. The adjacent functions carry thread::join failed and thread::detach failed, and the destructor matches libc++'s thread destructor. Process.cpp is byte-identical on that baseline and final base 5a6ad0ed3.
• With this change, a visible 20 s run with the normal draw cache and a separate 60 s run with APS5_NO_DRAW_SRT_ENTRIES=1 both close normally with exit code 0, without a timeout or forced kill. The latter displays the complete welcome screen and 16 language buttons. Disabling that cache isolates shutdown from the rendering bug fixed by fix(agc): validate mesh index descriptors on cached draws #5791; this PR changes no rendering code. Media playback and a local Linux run were not tested.
• python tools/check_conventions.py --base origin/main, git diff --check and python tools/check_main_sync.py --base origin/main passed; clean merge, 0 behind.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt
• No notes, investigation .md files or images added to the repository (evidence recorded in this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: PPSA85300 VLC (PPSA85300) #5732 (abort on host window closure)
• GPUs tested: NVIDIA GeForce RTX 4070 Ti SUPER, driver 617.14 / 32.0.16.1714, Windows 11 Pro build 26300
• AI-assisted: yes — Codex

[MarcZX10]

I have a separate follow-up for #5931 covering explicit guest exit() and return-from-main: run guest finalizers while VideoOut/AGC are still available, then run the registered shutdown callbacks. It leaves your host-window-close path on LibcExitWithoutHandlers unchanged and declares Depends on: #5843. The regression checks the shutdown token is still live during a guest finalizer, including the normal nested-exit path. Please let me know if that stacking order conflicts with your intended change; I will keep the scope separate.

[Ondra97]

Thanks @MarcZX100, that ordering fits the intended scope of #5843. The dependency order would be #5236 → #5843 → your follow-up.

Please keep guest_requested_exit and guest_system_exit passing: both must still skip guest and host exit handlers. For explicit guest exit() and return-from-main, keeping VideoOut/AGC available through the remaining finalizers, including nested exit(), makes sense. The shutdown callbacks should then run once.

AI-assisted: yes. — Codex

[ahmad3a4]

Windows check with the title from #5931. With this PR (e0748f3) on main ed835aa, closing Syobon Action (PPSA47402) with WM_CLOSE ends with exit code 0 in 3 of 3 runs. On main e34c479 it aborted 4 of 4 times with libc: exit thread cannot wait for itself. guest_requested_exit and the other exit tests pass here too.

The run was stacked with #5978, so the merge resolution is in #5978 (comment). Keep LibcRunShutdown_nid_postfix() in LibcExitWithoutHandlers_nid_no_patch once StartExit no longer runs it.

AI-assisted.

[Ondra97]

Thanks @ahmad3a4 for validating the combined tree. The merge resolution is correct: once StartExit() no longer runs shutdown, LibcExitWithoutHandlers_nid_no_patch() must call LibcRunShutdown_nid_postfix() before flushing and terminating.

@MarcZX100, one remaining coverage gap in #5978: guest_nested_exit currently checks the remaining handler and exit code, but does not register a shutdown callback or check the token. Please extend it to verify that the remaining finalizer sees a live token and that shutdown runs exactly once afterward.

AI-assisted: yes. — Codex

[screamm]

Windows/NVIDIA check with VLC (PPSA85300 v0.4, catalog release).

• Head eb8146b73 merged onto main 2559e9623: clean, no conflicts (merge 87c4dc4a6). Windows 11 Pro 26220, WinLibs MinGW-w64 GCC 15.2.0 r7, Ninja, Release, NVIDIA GeForce RTX 4070 (driver 596.21).
• Build: full default build and libs pass, no new warnings.
• --repeat until-fail:3: guest_requested_exit, guest_system_exit, guest_system_exit_parameters, guest_quick_exit, guest_immediate_exit, guest_return_from_main, guest_system_relaunch, guest_nested_exit: 8/8, 3 of 3.
• Negative control: the same guest_system_exit_tests.exe --requested-exit-driver with main's libc.prx first on PATH prints Guest shutdown completed, Guest exit handler ran, FATAL: abort() (SIGABRT) (3 of 3).
• ctest -j8 --timeout 120 -E '^agc_driver_|^video_out_flip': 813/813 passed (2 skipped, the same as on main: agc_host_import_sparse, guest_video_out_wqhd_detection_no_vulkan).
• VLC relinked with relinker --windows --registry unused-filter=0 --windows-fixed-base 0x400000, identical procedure for both builds, visible window, WM_CLOSE posted to it at 45 s:

Build	Runs	After WM_CLOSE	Window title at close
main 2559e96	3	exit 0xC0000409 after 0.8 / 3.2 / 1.4 s	VLC | FPS: 46.23 (2458), 48.98 (2263), 41.19 (1647)
main + #5843	2	exit 0 after 0.8 / 1.4 s	VLC | FPS: 57.05 (3020), 75.20 (2692)

Neither build prints a stop line. Every log ends at [src/Thread.cpp:527 RunThread] RunThread entry=0x22d3f60 arg=00000020000228c0 self=…, and on main the abort leaves only the exit code. Both builds show the same screen (gradient and page dots; the text needs #5791).

Verdict: on NVIDIA, closing the VLC window now exits 0 instead of aborting, with no test regressions.

AI-assisted: yes (Claude Code)

[milescoviello]

A Linux run with the title from #5931: Syobon Action (PPSA47402, catalog 01.001.000), Ubuntu 24.04, GCC 13.3, Release, main ab9a476. The window was closed with wmctrl -c 20 s in, while the title was presenting:

• main: exits 134 with libc: exit thread cannot wait for itself.
• main + this PR alone: exits 0.
• main + this PR + fix(libc): run guest finalizers before driver shutdown #5978: exits 0.

AI-assisted: yes

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 6044	core/libs/prx/libSceSystemService/CMakeLists.txt core/libs/tests/GuestSystemExit.cpp
PR 6447	core/libs/prx/libc/src/Process.cpp

[adriadam10]

Checked on the PR merged onto current main: build and scoped ctest green (119/119, no skips). guest_requested_exit fails when the Process.cpp change is reverted and passes with it; LibcExitWithoutHandlers_nid_no_patch still runs the shutdown callbacks through StartExit. Windows and Linux title runs by @ahmad3a4, @screamm and @milescoviello cover what we can't run here; merging once the full Build is green.

[MarcZX10]

I added the missing coverage to #5978 and rebased that PR onto current main (de974eb83). The nested-exit child now registers a shutdown callback and a remaining finalizer that asserts the shutdown token is still live; the exact marker sequence also verifies the finalizer runs before shutdown and the callback runs once. The existing exit-code and remaining-handler checks stay in place.

The test source syntax check and conventions check pass. I have not rerun the runtime CTest on this host; the earlier attempt stopped at the host CPU-feature guard before entering the test body. Hosted checks are starting on the rebased head.