[ducloc080202]

What

Rebases #3861 (open but conflicting since Oct 12) onto current main. Serializes Windows file-position transfers with a mutex per descriptor, shared across duplicated descriptors: positioned transfers (NativePositioned_nid_no_patch), plain transfers (NativeTransfer, NativeRead, NativeWrite) and seeks (NativeLseek) of one descriptor can no longer interleave SetFilePointerEx positioning. Entries persist per descriptor number; a lock is only ever held alone, so no new lock order.

Supersedes #3861; that PR can be closed if this one is accepted.

Tested

• Windows (MinGW 15.2.0, Release): full build passes; new guest_positioned_race (concurrent pread against read+lseek on one descriptor, verifying both sides) fails on main on the first run and passes 10/10 with the fix; file_position, guest_posix_filesystem, guest_filesystem, guest_file_vector, file_streams, guest_aio all pass (7/7). Linux/macOS rely on CI.
• Cat Quest III (PPSA15065), Windows: loads all 454 asset bundles to the loading screen with no file-read errors. Linux/macOS rely on CI.

Checklist

• Based on current main; no other open PR implements the same functions (rebases conflicting fix(libkernel): serialize positioned file transfers on Windows #3861; supersedes it)
• 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 file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: Cat Quest III (PPSA15065) on Windows: after loading ~450 bundles the title stays on the loading screen; it polls scePadRead and receives buttons but never advances #4909 (needed to reach the loading screen; the title still stalls there)
• AI-assisted: yes

[oneandonlydean]

Taking this one over #3861. Three things before it can go in:

• The branch has a merge commit (e2e4ba2a0); please rebase the original commit onto main instead, keeping its author.
• Open.cpp:46-48: NativeLseek can still run between the save and the restore in NativePositioned_nid_no_patch (Stdio.cpp:186-197), and the restore then overwrites the seek. Take the same lock there.
• A Windows test that runs pread against read and lseek on one descriptor from two threads and fails on main.

A lock per descriptor instead of one for all files would be better, as asked on #3861.

[ducloc080202]

Done, pushed as 31bcf98..e69306f (branch rebased, no merge commit): 1) original commit cherry-picked with its author kept; 2) the same per-descriptor lock now covers NativeLseek, so a seek between the positioned save and restore can no longer clobber either side; 3) new guest_positioned_race test (pread vs read+lseek on one descriptor, both sides verified) fails on main first run, passes 10/10 with the fix; 4) one mutex per descriptor number, shared across dup, entries persist so close/reuse cannot strand a waiter. The debt entry for the Windows positioned race now describes the serialized behavior.

[nicolasfalesy]

Linux build (Arch, GCC 16.2.1, Release, BUILD_TESTING=ON), head f9881adaa merged alone onto main 4bb2c7f: 592 targets fail to link, everything that links libkernel.prx:

/usr/bin/ld: core/libs/libs/unpatched/libkernel.prx: undefined reference to `File::SharePositionedTransfer(int, int)'

File::SharePositionedTransfer is defined in File/src/Stdio.cpp inside the #ifdef _WIN32 block that starts at line 50, but ShareDescriptor in File/src/DescriptorControl.cpp now calls it on every platform (the call sits before that function's own #ifdef _WIN32). Either guarding the call with #ifdef _WIN32 like the GuestFileLockShare_nid_no_patch line below it, or giving Linux an empty definition, should fix it. The CI linux job is skipped for this PR, so it would not show there.

[xXPBSXx]

Windows counter-test (follow-up): Windows 10 (19045), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release, this PR merged onto main 2a2a3dc65. guest_positioned_race passes with the PR; with the PR's source files replaced by main's, same test, it aborts at line 74.

Not tested: Linux, a console, a title run.

AI-assisted (Claude Code).

[p3n1cillin]

Data point from a third Unity title, for scope: Cult of the Lamb (PPSA06465) also faults intermittently on Loading.PreloadManager at start-up on Windows. This PR merged onto main a0f5916a2 doesn't change its rate: in 8 alternated 40 s runs per side, there were 5 faults with the PR and 4 without. So its fault is a different race from the one fixed here; details in #5490. This says nothing against the PR for Hotline Miami or Cat Quest III.

[oneandonlydean]

Rebased onto main (bf8dc61) and force-pushed. The only conflict was the include list in File/src/Stdio.cpp, which now keeps both <unordered_map> and <utility>. Your three commits keep their authors.
Added one commit: DescriptorControl.cpp:568 now calls File::SharePositionedTransfer under _WIN32, next to GuestFileLockShare_nid_no_patch. The lock is only defined in the Windows part of Stdio.cpp, so libkernel didn't link on Linux. Windows behaviour is unchanged.