[Ondra97]

What

On Windows, a zero-length sceNetRecv, sceNetRecvfrom or sceNetRecvmsg on a healthy connected TCP socket can wait for peer data, or return EAGAIN with MSG_DONTWAIT/SO_NBIO. A caller asking to receive no bytes should complete with zero without waiting or consuming queued stream data. Fresh unconnected sockets also expose Winsock's EINVAL through recvfrom/recvmsg, instead of ENOTCONN.

The Windows receive-preparation helper previously bypassed every stream socket. The exports then reached the readiness check or native receive even when their requested size was zero. I extend that existing helper to complete an empty receive only when getpeername identifies a connection and a nonblocking WSAPoll check reports no error, hangup or invalid descriptor. It does not read or clear SO_ERROR. Existing export paths set the empty sender length and clear recvmsg output flags/control length while leaving payload, address and control storage untouched.

The stream-specific branch is guarded by _WIN32. The helper is also compiled on macOS in current main; macOS keeps its existing immediate stream return and never sees WSAPOLLFD or WSAPoll from this branch.

For a stream that has never attempted a connection, the helper reports ENOTCONN, including bound and listening sockets. Connection attempts and accepted connections are tracked so an unsuccessful getpeername does not invent that fresh-socket result for a pending or failed connection. Pending-error, disconnected and connecting paths continue through the existing host receive path; their exact console error/output precedence remains documented as debt. Nonempty receives and UDP keep their existing paths: in particular, a zero-length UDP receive still consumes a queued datagram.

The behavioral reference is FreeBSD 11.4 soreceive_generic, lines 1568-1583: an unconnected connection-required socket fails first, and zero residual length returns before the nonblocking wait decision. This is a BSD reference, not a new PS5 measurement.

I add 24 Windows runtime regressions: all three exports with an empty stream, MSG_DONTWAIT, MSG_PEEK | MSG_DONTWAIT, SO_NBIO, queued data, null/empty vectors, invalid descriptors/flags and fresh/bound/listening sockets, plus a UDP consumption control for each export. The tests check output canaries and bound the empty receive to 500 ms, with cleanup if the baseline waits.

Tested

Latest validation (2026-10-24)

Base 522e32e3b618154e0e221a9f4d28da87dd2cf187; submitted and tested head 3d0c15528847dc042ad50428c7c21c37bc7ad7f1. Windows 11 Pro build 26300, Ryzen 9 7900X, 31.66 GiB RAM, RTX 4070 Ti SUPER (no GPU tests); WinLibs GCC 15.2.0 UCRT POSIX SEH r7, Ninja Release, BUILD_TESTING=ON, bundled dependencies and the existing FFmpeg prebuilt tree.

• All production libraries, the relinker and the selected runtime fixtures built successfully on the final commit, without a source overlay.
• 62/62 selected runtime tests passed, zero failures and zero skips in 10.11 s, including all 24 empty-stream receive cases, datagram and guest-buffer controls, refused connections, nonblocking/blocking transfers, TCP and SSL. patched_libraries loaded 368 libraries, zero failed. The original regression source is unchanged.
• Preprocessing the exact helper region from the unguarded automatic merge of old head 58108841b1d351689df4f32091bbdb209ce6895d and the pinned base exposes WSAPOLLFD/WSAPoll with _WIN32 undefined and __APPLE__ defined. The final source excludes both and retains the stream return 0. Windows preprocessing is byte-identical before and after the guard. This checks the conditional source branch, not a native macOS build.
• Conventions, strict main-sync and whitespace checks passed. The overlap scan identifies feat(net): expose stream and datagram protocol #9930's already-agreed accepted-socket/debt insertion order. The shared-file patches in conflicting fix(libSceNet): stop Sackboy loading retries on peer sockets #5836, Macos port #7201, fix(libSceNet): reject the flags and names the module rejects #9727, fix(libSceNet): keep MSG_DONTWAIT and aborted stream sends short on macOS hosts #9733 and fix(libSceNet): print the interface and socket tables from sceNetShowIfconfig and sceNetShowNetstat #10502 were inspected manually: peer transports/options, resolver/platform handling, argument validation, macOS sends and diagnostic tables do not implement this empty-receive helper. This is source comparison, not joint runtime validation of those PRs.

No native macOS build/runtime, title or console run, default all-target build or complete CTest suite was performed. The earlier Windows red/green results and host-specific observations are retained below with their original revision IDs.

Commands, from the source tree with the documented toolchain on PATH, FFMPEG_PREBUILT_DIR set and the sibling Release build configured:

cmake -S . -B ../build
cmake --build ../build --target libs relinker patched_libraries_tests net_empty_stream_tests guest_datagram_receive_tests guest_refused_connect_tests guest_socket_host_write_tests guest_ssl_tests guest_ssl_load_tests guest_ssl_ps5_readers_tests guest_ssl_transport_tests net_dontwait_tests net_blocking_send_tests guest_sce_net_tests guest_tcp_tests --parallel 8
ctest --test-dir ../build -R '^(net_empty_stream_.*|guest_sce_net|guest_datagram_.*|guest_refused_connect|guest_socket_host_write_.*|guest_ssl.*|patched_libraries|net_dontwait_.*|guest_tcp)$' -V --output-on-failure --no-tests=error --output-junit tests.xml
python -B tools/check_conventions.py --base origin/main
python -B tools/check_main_sync.py --base origin/main --strict
python -B tools/pr_overlap.py --repo boykopovar/AnyPS5 --branch codex/pr8686-rebase-20261022 --jobs 3
git diff --check origin/main...HEAD

Historical validation on earlier revisions

Environment: Windows 11 Pro 10.0.26300; Ryzen 9 7900X, 32 GB RAM, RTX 4070 Ti SUPER (no GPU tests); WinLibs GCC 15.2.0 UCRT POSIX SEH r7; Ninja Release with BUILD_TESTING=ON, Python 3.14 and the existing FFmpeg prebuilt dependency. These are synthetic loopback tests calling the implemented exports. No original PRX/SPRX modules, console or PRAGMATA run was available. This is a partial host-side correction related to #8005; title-level impact and completion of the overall audit remain unverified.

Historical October 22 validation: base a6dcea9279b53b3a39c13695eac66f7c8b5b503c, head 58108841b1d351689df4f32091bbdb209ce6895d. The original production patch and all 24 new test cases are preserved; the documentation resolution retains the current-main debt entries and adds only this PR's receive-state caveat.

With the unchanged production file from current main and the final regression fixture, 7/24 tests pass and 17/24 fail. After restoring the fixed production file and rebuilding, all 24 pass. The queued-data and UDP consumption controls pass in both runs.

The production build and selected fixtures pass without a source overlay. The expanded runtime selection passes 62/62 with zero skips in 10.14 s, covering all 24 empty-receive cases, datagram and guest-buffer controls, refused connections, nonblocking/blocking transfers, TCP, SSL and patched-library loading.

cmake --build build --target libs relinker patched_libraries_tests net_empty_stream_tests guest_datagram_receive_tests guest_refused_connect_tests guest_socket_host_write_tests guest_ssl_tests guest_ssl_load_tests guest_ssl_ps5_readers_tests guest_ssl_transport_tests net_dontwait_tests net_blocking_send_tests guest_sce_net_tests guest_tcp_tests --parallel 8
ctest --test-dir build -R '^(net_empty_stream_.*|guest_sce_net|guest_datagram_.*|guest_refused_connect|guest_socket_host_write_.*|guest_ssl.*|patched_libraries|net_dontwait_.*|guest_tcp)$' -V --output-on-failure --no-tests=error
python -B tools/check_conventions.py --base origin/main
python -B tools/check_main_sync.py --base origin/main --strict
python -B tools/pr_overlap.py --repo boykopovar/AnyPS5 --branch codex/pr8686-rebase-20261022 --jobs 3
git diff --check origin/main...HEAD

Conventions, strict main-sync and whitespace checks pass. The overlap tool identifies the existing remote version of this PR and requests manual review of shared files in #5836, #6044, #6879, #7201, #8204, #8225, #8232, #8234, #8235, #8254, #8362 and #9076. Their changes cover other transports/options, test paths, platform handling, address parsing, diagnostic/configuration exports, socket information and allocation flags; none supplies this empty-stream receive behavior. I also reviewed related connection, timeout and closed-stream-readiness changes; this is source comparison, not runtime validation of combined open PRs.

The default all-target build and complete CTest suite were not run on the rebased head. The original build, probe and host-specific guest_poll observations below are historical and pinned to their stated revisions.

Revision under test: 54d4adf7f58fe0d9b454b4d377a66e5c454020f9, based on aa1b1bfcea0e9fc2a457577ad8dce94010da7cad. The before/after comparison uses the same final test source, build configuration and dependencies, replacing only the production file changed by this PR with main's Export.cpp (blob 6a679a251fdd22b452127a24db2a0c50b838e268, also unchanged at a6a4101e190ec408c3ed471d20f06872cae1b4a3). Each table entry is one CTest invocation, not a performance benchmark.

Metric	Before (base)	After (fix)
New receive regressions	7/24 passed, 17 failed	24/24 passed, 0 failed
sceNetRecv cases	3/8 passed	8/8 passed
sceNetRecvfrom cases	2/8 passed	8/8 passed
sceNetRecvmsg cases	2/8 passed	8/8 passed

The failing base cases cover waiting/EAGAIN on empty streams and EINVAL on fresh recvfrom/recvmsg. The queued-data and UDP controls passed before and after.

Validation of the revision above:

• Full default Windows Release build completed successfully with the test prerequisite described below; the original prerequisite file was restored byte-for-byte afterwards and the worktree is clean.
• The focused runtime selection passed 50/50, with 0 failed and 0 skipped. This includes the 24 new cases, existing UDP/TCP receive and guest-buffer write checks, refused connection handling, SSL/Net/kernel transport checks and patched_libraries (354 loaded, 0 failed).
• check_conventions.py, git diff --check and the test-registration audit passed; the latter reported 1,127 test executables and no built-but-unregistered executable.
• The separately rerun existing guest_poll test failed with exit 0xc0000409. It links libkernel/libc, not libSceNet; neither its source nor its kernel socket implementation is changed here. A standalone native Winsock urgent-data probe on this host also reported no exceptional readiness and delivered the urgent byte through normal receive even with SO_OOBINLINE=0. This remains an unresolved host/test limitation; I do not claim a full repository CTest pass or a fix for it.

The full Windows test build requires the single-file test correction from #8310 at ca55e46f98ad85861dd7567abad0c028274daca1. Without it, this MinGW toolchain rejects the existing alignas(65536) static array in ImageArrayWindowPastLayerLimit.cpp because its supported maximum is 8192. I applied only that reviewed test-file change for validation and restored the original bytes afterwards; it is absent from this PR. The build result therefore applies to this head plus that test prerequisite, not an unqualified standalone-head build.

Reproduction after configuring the documented Windows toolchain and dependencies: put the GCC toolchain and Python on PATH, and set FFMPEG_PREBUILT_DIR in the environment to the existing FFmpeg prebuilt tree. Fetch #8310's exact revision before generating its test-only patch.

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DFFMPEG_PREBUILT_DIR="$env:FFMPEG_PREBUILT_DIR"
cmake --build build --target net_empty_stream_tests --parallel 1
ctest --test-dir build -R '^net_empty_stream_' --output-on-failure
git fetch https://github.com/boykopovar/AnyPS5.git ca55e46f98ad85861dd7567abad0c028274daca1
git diff ca55e46f98ad85861dd7567abad0c028274daca1^ ca55e46f98ad85861dd7567abad0c028274daca1 --output=../array-window-test.patch -- core/libs/prx/libSceAgcDriver/tests/execution/ImageArrayWindowPastLayerLimit.cpp
git apply ../array-window-test.patch
cmake --build build --parallel 6
git apply -R ../array-window-test.patch
ctest --test-dir build -R '^(net_empty_stream_.*|guest_sce_net|guest_datagram_.*|guest_refused_connect|guest_socket_host_write_.*|guest_ssl.*|patched_libraries)$' --output-on-failure
ctest --test-dir build -R '^guest_poll$' --output-on-failure
python tools/check_test_registration.py build
python tools/check_conventions.py --base origin/main
python tools/pr_overlap.py --repo boykopovar/AnyPS5 --branch codex/fix-net-empty-stream-receive --jobs 4

For a failing baseline, retain this PR's test source and CMake registration, replace core/libs/prx/libSceNet/Export.cpp with the file from the base revision, rebuild the target and repeat the same CTest command. Restore the fixed file afterwards. This isolates the sole production-file change rather than attributing unrelated changes on moving main to the fix.

Overlap review: the scan found no conflicting changed lines with #8653 (per-call nonblocking transport) or #8598 (epoll registration lifetime). I also compared their shared implementations: neither handles empty stream receives, and their transport/state changes compose with this guard; static three-way merges of all three shared files were clean. The kernel-only empty-receive change in #8492 is separate from these libSceNet exports. I manually reviewed the files the checker could not compare in #5836, #6044, #6879, #7201 and #8204: peer transports/options, test organization, host includes/macOS resolver handling and Ethernet address parsing do not implement this behavior. This is source review, not runtime validation of combined PRs.

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 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: feat(net): expose stream and datagram protocol #9930 for the agreed merge order; no socket runtime dependency
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• 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 5836	core/libs/prx/libSceNet/CMakeLists.txt core/libs/prx/libSceNet/Export.cpp docs/dev/debt/libSceNet.md
PR 7201	core/libs/prx/libSceNet/Export.cpp
PR 9727	core/libs/prx/libSceNet/Export.cpp docs/dev/debt/libSceNet.md
PR 9733	core/libs/prx/libSceNet/Export.cpp docs/dev/debt/libSceNet.md
PR 9948	core/libs/prx/libSceNet/Export.cpp docs/dev/debt/libSceNet.md
PR 10502	core/libs/prx/libSceNet/Export.cpp docs/dev/debt/libSceNet.md

[cadenasynyster]

Independent Windows run of this PR, in case it helps review.

Environment: Windows 11 10.0.26300 (same build as the PR description), Ryzen 7 9800X3D, WinLibs GCC 15.2.0 r7 UCRT, Release, BUILD_TESTING=ON. I built only the net_empty_stream_tests target, so the #8310 alignas workaround was not needed.

Results:

• PR head: ctest -R net_empty_stream passes 24/24.
• Baseline: I kept the PR's tests and replaced only core/libs/prx/libSceNet/Export.cpp with main's version, then rebuilt. 17 of 24 fail and 7 pass, the same as the table in the description. The failing names are the recv, recvfrom and recvmsg empty, dontwait, peek, nonblocking and nullbuf cases, plus recvfrom_errors and recvmsg_errors.
• I restored the PR's Export.cpp afterwards.

guest_poll: the description lists it as failing with 0xc0000409 on this OS build. On my machine it passes 3/3 (0.33 s each) on main plus an unrelated commit, so that failure looks specific to the author's host rather than to Windows 11 10.0.26300 itself.

Not run: the standalone Winsock urgent-data probe, Linux/macOS, the full repository CTest, and any title or console run. These are synthetic loopback tests only.

AI-assisted: yes

[MarcZX10]

The independent Windows run materially strengthens the report: it reproduces the 24/24 head and 17/24 baseline split on the same OS build, using only the production-file swap for the baseline. It also shows guest_poll passes on another Windows 10.0.26300 host, so the reported 0xc0000409 remains a host-specific observation rather than an OS-build-wide limitation. The stated test scope and missing Linux/macOS/title coverage are clear.

[Queuse]

@Ondra97 I'm adding SO_PROTOCOL readback in #9930. Both branches add an assignment to the accepted socket next to connection.type: yours sets connectAttempted, mine copies the listener's protocol. Both need to stay.

Would landing #9930 before this PR work for you? Mine already follows #9543/#9525. If yours needs to go first, I can rebase and keep both assignments instead. I've left mine as a draft until we settle the order.

[Ondra97]

@Queuse I confirm that #9930 landing before #8686 works for my plans. I added #9930 to #8686's Depends on line for this merge order; your existing #9525 -> #9543 -> #9930 sequence stays intact.

I checked both patches. When #8686 is rebased after #9930 lands, the accepted socket must retain both connection.protocol = listener.protocol and connection.connectAttempted = true. The separate protocol and empty-stream CTest registrations and debt entries must also remain.

This is merge-order coordination only. No source changes or combined runtime test are claimed by this reply.

[Queuse]

Thanks, @Ondra97. I've updated #9930 with the agreed order: #9525 → #9543 → #9930 → #8686. Both accepted-socket assignments, test registrations and debt entries need to stay when yours rebases. The order with #8686 is settled; I'm keeping #9930 as a draft while the separate #6044 test-reorganization order is still pending.

[harisibrar1]

Checked against main 522e32e and FreeBSD 11.4 sys/kern/uipc_socket.c (soreceive_generic, lines 1568-1583):

• The ordering for the established, error-free case matches FreeBSD 11.4: the ENOTCONN check for a never-connected stream, then the uio_resid == 0 return, then the nonblocking EWOULDBLOCK check. The remaining states are listed in the debt file, which seems right to me.
• Possible macOS build problem after rebase: on main, wait_for_local_address is now compiled under #if defined(_WIN32) || defined(__APPLE__) (Export.cpp:690-759, widened in 78983da after your base). This PR's new stream branch uses WSAPOLLFD/WSAPoll without a _WIN32 guard, while every other WSAPoll use in the file has one (e.g. wait_for_datagram around :723-732), and nothing in the tree defines WSAPOLLFD for other hosts. The branch still merges textually (3-way git merge-tree shows no conflict markers), so this would only show up when building on macOS; I couldn't build there. Would you consider guarding the new stream branch with #ifdef _WIN32, so macOS keeps main's current early return for streams?
• Main doesn't already contain this change: its wait_for_local_address still returns 0 immediately for every stream socket (:737).

Reviewed with AI assistance (Claude Code); the checks above were run locally on my machine.

[Ondra97]

@harisibrar1 Fixed in 3d0c15528847dc042ad50428c7c21c37bc7ad7f1, with main 522e32e3b618154e0e221a9f4d28da87dd2cf187 merged in.

The stream branch now guards its WinSock logic with #ifdef _WIN32; the other branch returns 0, preserving main's macOS stream behavior. Preprocessing the exact helper region confirms no WSAPOLLFD/WSAPoll on the __APPLE__ branch and byte-identical Windows output compared with the unguarded merged control.

All production libraries, the relinker and selected fixtures built on the final commit. The focused Windows selection passed 62/62, including all 24 empty-stream cases, with zero skips; 368 patched libraries loaded. Conventions, strict main-sync and whitespace checks passed. I have not run a native macOS build or runtime test. The agreed #9930 merge order remains in Depends on.

AI-assisted: yes.

[adriadam10]

Re-checked against current main: the new libSceNet/Export.cpp changes (net thread entry, NetSockInfo6) are away from wait_for_local_address, accept and connect, and the merge is clean. The _WIN32 guard keeps macOS returning 0, connectAttempted is set in connect/accept, the 24 net_empty_stream_* tests fail on the old Export.cpp (17/24, also reproduced by @cadenasynyster), and the debt entry is present. UDP and non-empty receives are unchanged; merged.