[xXPBSXx]

What

fcntl on a guest socket answered F_GETFL and F_SETFL and failed every other command with EINVAL. libcurl calls fcntl(fd, F_SETFD, FD_CLOEXEC) on every new socket, and it failed:

• Title evidence: PPSA99001 (ProsperoRadio, bundled libcurl 8.18.0 and OpenSSL 3.5.2 over libkernel sockets), traced by screamm on main in this comment on #4796: fcntl(F_SETFD) on a socket returns EINVAL (the comment lists it among the "smaller mismatches curl tolerates, for later"); libcurl issues fcntl(fd, F_SETFD, FD_CLOEXEC) on every new socket. The same group of titles (PSXS5, SwanStation, ProsperoTV, ProsperoRadio) uses that libcurl. No title fails because of this answer: curl ignores the error, so the change replaces a wrong answer in the trace of these titles and does not unblock one.

F_GETFD and F_SETFD are now answered for a guest socket as FreeBSD 11.4's kern_fcntl does (FD_CLOEXEC, value 1):

• F_SETFD sets or clears the flag, F_GETFD reads it back (a new socket starts at 0);
• SOCK_CLOEXEC in the type of socket sets it at creation;
• dup and dup2 give the copy no flag (FreeBSD's kern_dup clears it) and leave the original's; close forgets it, so the next socket on the same number starts at 0.

The flag is only recorded: there is no exec in the guest. Not changed, and in the TechnicalDebt entry for libkernel: the other commands on a socket still fail with EINVAL (F_DUPFD, F_DUPFD_CLOEXEC, F_GETOWN, F_SETOWN), and an accepted socket starts without the flag.

Overlap: this branch was merged with main 00f4b3351 (conflict in docs/dev/debt/libkernel.md next to the ioctl entry, resolved by keeping both entries; git merge-tree against main is clean). #4954 (fcntl and O_APPEND), #5186 (FIOCLEX and FIONCLEX) and #5569 (ENOTTY for malformed ioctl requests) touch the same part of docs/dev/debt/libkernel.md; whichever lands second rebases its entry.

Tested

Windows 10 (19045), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release, head 863e506b2 (this branch merged with main 00f4b3351):

• The socket group passes, 25 of 25 (guest_socket*, including the twelve guest_socket_host_write_*, guest_sce_net, guest_udp, guest_tcp, guest_poll, guest_dup, guest_kernel_fcntl), ctest -V, serial, after building the test executables of exactly these tests (log run_5020_fix2.log). guest_socket_descriptor_flags (new) is one of them: F_GETFD on a new socket, F_SETFD and read-back, dup and dup2 give the copy no flag and keep the original's, F_SETFD 0, SOCK_CLOEXEC, and the next socket after a closed one starts at 0.
• Without the change to Udp.cpp (taken from main, test kept; measured on the previous head b09eccdc4, whose Udp.cpp is the same) it fails: a new socket must not be close-on-exec (F_GETFD 0) (errno 22) (log run_5020.log).
• A first run on that previous head showed 12 failures in the guest_socket_host_write_* tests; the cause was an outdated test executable (not rebuilt). After rebuilding it they pass, here and on main.
• python tools/check_conventions.py --base upstream/main passes.
• Not tested: Linux (one source for both hosts, only Windows was built), a console, a title run with this change (the evidence is screamm's trace).

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 (docs/dev/debt/libkernel.md)
• No notes, investigation .md files or images added to the repository
• New third-party code is a submodule built from source (none added)
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: PPSA99001 (ProsperoRadio) and the titles of its group: a clean fcntl trace on their sockets; no bring-up discussion exists
• AI-assisted: yes

AI-assisted (Claude Code).

🤖 Generated with Claude Code

[adriadam10]

Checked on current main (merged and rebased): build + ctest green (169 tests in core/libs and libkernel), and guest_socket_descriptor_flags fails without the Udp.cpp change. F_GETFD/F_SETFD, SOCK_CLOEXEC, dup/dup2 and close match FreeBSD's kern_fcntl/kern_dup.

Added on your branch: rebased onto main (conflict in docs/dev/debt/libkernel.md, kept both entries) and updated the fcntl entry, which still said F_GETFD/F_SETFD on a socket fail with EINVAL. Merging once the full Build is green.