[milescoviello]

What

Fixes #4479. Carries @screamm's fix commit from that issue (screamm:feat/libkernel-record-locks-windows, b7cc2856), keeping the commit and its authorship, as the issue invites. The branch sits on top of #3649's two commits, which this commit does not need (#3649 is only how the titles reach the call); it cherry-picks onto current main without changes and builds there.

On Windows, fcntl record locks (F_GETLK, F_SETLK, F_SETLKW and the old F_OGETLK/F_OSETLK/F_OSETLKW forms) throw fcntl: record locks are not implemented on Windows, which #3840 recorded as a limit. With #3649 applied, ProsperoRadio (PPSA99001) stops there on F_SETLK, and ProsperoTV (PPSA99003) does too once #4478's fix is applied. This commit applies FreeBSD 11.4's argument checks in FreeBSD's order (kern_fcntl, lf_advlockasync) and then follows lf_blocks, where only locks of different owners conflict: every lock belongs to the one guest process, so F_SETLK and F_SETLKW succeed without waiting and F_GETLK reports F_UNLCK. Nothing is held on the host, because Windows byte-range locks are mandatory and would fail the guest's own reads and writes through its other descriptors; the fcntl TechnicalDebt entry records that other host processes neither see nor are seen by these locks. A SEEK_END lock on a directory and locks on character devices still throw. Linux is unchanged.

guest_kernel_fcntl replaces the old Windows "throws" expectation with the round trip and adds RecordLocksBelongToTheProcess (overlap merge and split, two descriptors on one range, another thread's F_SETLK/F_SETLKW/F_GETLK, SEEK_CUR/SEEK_END, negative lengths, the 24-byte old struct, EINVAL, EBADF on read-only and write-only descriptors, relock after close); the EOVERFLOW, error-order, F_GETLK write-back and directory cases are Windows-only.

Tested

• From the issue: Windows 11, WinLibs MinGW-w64 GCC 15.2.0, Release: guest_kernel_fcntl, guest_poll, guest_dup, guest_file_vector 3/3; 620/620 of ctest -j8 -E "^agc_driver_|^video_out_flip"; ProsperoRadio relinked against the build (with feat(libkernel): poll on file descriptors and Windows pipe read ends #3649) gets past the fcntl stop and reaches libkernel: poll throws for POLLPRI on guest sockets under WSAPoll (PSXS5, SwanStation, ProsperoTV after #3649); fix branch ready, PSXS5 renders #4478's.
• screamm's combined Windows run (main 33ae282e + feat(libkernel): poll on file descriptors and Windows pipe read ends #3649 + libkernel: poll throws for POLLPRI on guest sockets under WSAPoll (PSXS5, SwanStation, ProsperoTV after #3649); fix branch ready, PSXS5 renders #4478's fix + this commit): ProsperoRadio and ProsperoTV get past poll and the record locks and stop next on the graphics side (fix(recompiler): read P10 of a flat input as the provoking vertex #4030).
• Linux, added for this carry (the issue left the shared test part to Linux CI): Ubuntu 24.04, GCC 13.3 (CI's compiler), Release with CI's configure flags, on main 33ae282. guest_kernel_fcntl, guest_poll, guest_dup and guest_file_vector pass, and all 306 guest_* tests pass. The new shared RecordLocksBelongToTheProcess cases also pass with main's DescriptorControl.cpp, as expected since Linux forwards record locks to the host. I could not run the Windows path myself.
• Windows run of this head by @LeandroLP23 (comment): Windows 11, WinLibs GCC 15.2.0, Release: guest_kernel_fcntl 3/3 and fails with main's DescriptorControl.cpp (fcntl: record locks are not implemented on Windows, command=12, fd=3); ctest -E '^agc_' 554/554; ProsperoRadio with feat(libkernel): poll on file descriptors and Windows pipe read ends #3649 gets past the database's record lock to its HTTP code.
• Title run by @babyhuey (comment), Windows 11, WinLibs GCC 15.2: ProsperoTV (PPSA99003) on main d1fd937 + fix(libkernel): support POLLPRI on Windows sockets #4525 stops at what(): fcntl: record locks are not implemented on Windows, command=12, fd=4 (F_SETLK on a regular file); with this PR on top it gets past the lock and renders its menu for the full 60 s with no throw.
• tools/check_conventions.py --base upstream/main passes.

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 in TechnicalDebt
• 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: PPSA99003 (ProsperoTV) on Windows together with fix(libkernel): support POLLPRI on Windows sockets #4525, and PPSA99001 (ProsperoRadio) together with feat(libkernel): poll on file descriptors and Windows pipe read ends #3649; I found no Game bring-up discussion for them
• AI-assisted: yes

🤖 Generated with Claude Code

[LeandroLP23]

The PR body says the Windows path wasn't run for this carry, so I ran it. Windows 11 (26300), WinLibs GCC 15.2.0, Ninja, Release, BUILD_TESTING=ON, head 43ffba87e merged onto main 2f4cc8983.

Tests

• guest_kernel_fcntl: 3/3.
• ctest -E '^agc_' -j 4: 554/554.
• With main's DescriptorControl.cpp restored, guest_kernel_fcntl fails with fcntl: record locks are not implemented on Windows, command=12, fd=3. That is the same message as in libkernel: fcntl record locks throw on Windows (ProsperoRadio, ProsperoTV after #3649); fix branch ready #4479.

ProsperoRadio (PPSA99001, catalog 01.000.010), 60 s runs with --windows:

Build	Result
main + #3649	exit 0xc0000409 after 40 s. download0/prosperoradio.log ends at the sceKernelOpen of /download0/radio-browser.sqlite3 (125 lines)
main + #3649 + this PR	exit 0xc0000409 after 36 s. The log goes past the database to [ProsperoRadio][http] non-blocking sockets: NOT set (143 lines)

The title's log redirect doesn't keep the what() line in my runs, so the table shows how far the log gets, not the stop message. That fits the description: the title gets past the record lock on the database and stops later in its HTTP code, where #4479 expects the POLLPRI stop (#4525).

ProsperoTV (PPSA99003) exits after 10 s both with and without this PR. It needs #4525 first, and #4525 doesn't merge together with #3649: both change Socket/src/Udp.cpp and tests/GuestPoll.cpp. So I couldn't build the combination for it.

AI-assisted: yes (Claude Code)

[Dextroid17]

Review of #4480 (fcntl record locks on Windows), head 43ffba8

Verdict: approve, relying on LeandroLP23's Windows run for the host path. One non-blocking note.

Scope first: I ran Linux only. On Linux the #else branch uses the host's POSIX locks, so the Windows ProcessRecordLock is not compiled there. What I could check is the range/ordering/errno logic, by copying it into a stub harness.

What I ran (PR head merged onto upstream/main ee49621, clean env):

• guest_kernel_fcntl_tests (Linux build): 10/10 runs exit 0.
• Differential check of the range/ordering logic against my own transcription of FreeBSD fcntl() + lf_advlockasync() (freebsd-src 8c639945a): 2,352,000 cases (F_GETLK/F_SETLK, lock types incl. unknown, whence incl. invalid, start/length at int64 extremes, open mode, file size, file offset >= 0 only), built with UBSan. 0 mismatches. LockRangeError and the ProcessRecordLock checks are copied from the PR with the host calls stubbed. The reference is only as independent as my reading of FreeBSD, and it covers range and error order only, not F_SETLKW blocking, F_UNLCKSYS or the HostWrite write-back.
• Mutation controls: dropping the length - 1 overflow term gives 6,912 mismatches; dropping the F_GETLK unknown-type check gives 241,704; dropping the start == 0 guard gives 0 (when length < 0, start + length < 0 already catches it, so the guard is redundant but harmless).
• git merge-tree of this PR with #4555 (also edits DescriptorControl.cpp InitialEntry) is clean.

From LeandroLP23's comment, not run by me: Windows 554/554, guest_kernel_fcntl 3/3, and the failure message with main's DescriptorControl.cpp restored. In their ProsperoRadio runs the title exits 0xc0000409 both before and after this PR; with the PR the log gets past the database open to the HTTP code.

Not checked: the "SEEK_END on a directory still throws" case (the test asserts it under _WIN32 only).

Note: the start == 0 guard in LockRangeError is redundant when length < 0. Fine to leave.

[babyhuey]

A title that needs this: ProsperoTV (PPSA99003) on Windows 11 (WinLibs GCC 15.2, RTX 3080). On main d1fd937 + #4525 (POLLPRI under WSAPoll) it gets into its HUI and then stops at:

what():  fcntl: record locks are not implemented on Windows, command=12, fd=4

That's F_SETLK on a regular file. With this PR stacked on top (fork test/verify-tvlock, c28bb5e; merges clean), it gets past the lock and renders its menu for the full 60 s with no throw: [TV] splash hidden=1, then [TV] frames=600 mean=23.50ms max=25.38ms ... draws=2 shapes=407. guest_kernel_fcntl passes 5/5 on that head, with no warnings in DescriptorControl.cpp or GuestKernelFcntl.cpp.