[Felipebmaitan]

What

Add deferred cancellation for POSIX read, _read, readv, _readv, pread, preadv, recv, recvfrom, recvmsg, _recvfrom and _recvmsg. This includes read/readv on guest sockets and the existing accepted native-file, pipe, standard-descriptor, shared-memory, archive and random-device paths. Existing directory/driver/flag rejection and positioned socket/pipe errors remain. SCE read and libSceNet entry points acquire no new POSIX cancellation point.

Unix interrupts native calls with an otherwise unused host SIGUSR2. Windows cancels synchronous file/pipe calls or cancels and drains the exact overlapped socket request. Operation buffers, HostWrite and ownership guards end before cleanup/TSD or deferred handlers run. Disabled cancellation remains pending. Successful, zero and partial results preserve their result when cancellation arrives concurrently, following FreeBSD 11.4 libthr's failure-only leave check.

Per-descriptor admission prevents close/replacement/reuse races without hidden public aliases, additional descriptor slots, child inheritance or record-lock release. POSIX close, sceKernelClose and guest-socket close interrupt admitted POSIX operations before waiting for ownership to drain. Wake registrations are independent of cancellation state and thread adoption, and retry until native submission/completion races are resolved. Completed/partial results win; otherwise the read returns EINTR without signal restart. Descriptor replacement and stdio closure may still wait indefinitely. No blanket socket shutdown was added: surviving-alias coverage applies to pipes and GuestSockets-managed sockets; borrowed libSceNet retains its existing shutdown behavior. Windows append handle replacement joins this ownership. FILE locking precedes descriptor exclusion. Asynchronous caught signals wake the native wait and run after these guards release; restart requires unchanged descriptor generations.

Addressing Dean’s close/read comment: close now wakes the admitted native requests before waiting, and keeps descriptor ownership until native requests and HostWrite scopes have drained. Windows pre-call interruption also initializes errno, _doserrno and GetLastError, avoiding stale EAGAIN/EOF remapping. This interrupts/drains the read rather than reproducing the reported console behavior where the read stays blocked after close returns; no console verification is claimed.

Saved-context implementation metadata distinguishes suspended fibres from abandoned frames, including copied records and manually constructed stacks. It uses opaque jump-buffer tail/machine-context spare storage; ABI sizes/offsets, architectural registers/FPU/masks and uc_link/stack/flags remain. This internal format has no console verification and makes no foreign-libc saved-record compatibility claim. POSIX getcontext treats uc_mcontext as opaque.

Milestone 1 of four: reads/receives, writes/sends, accept/connect, multiplexed waits. No game/title unblock, console or performance claim.

Prerequisites: #10817 through e782144 and #10755 through 92f4276. All original commits/authorship are preserved. Main through 228cd64 is composed with them. Original implementation comparison excluding prerequisites (before this close fix): cf6dbf6...e93539c.

Overlap coordination: requests and confirmations. The author of both PRs has selected #11872 before #10856 and #11625; both later PRs now list Depends on: #11872 for merge ordering, with unchanged source heads. Specific order requests have been posted to the authors of #7090, #7135, #8916, #9938, #10552, #11341, #11831 and the new MSG_WAITALL PR #11896. #11896 now confirms #11872 first and declares that dependency; the other external requests remain pending. No combined runtime validation is claimed. The requested receive integration preserves descriptor/HostWrite ownership and cancellation/signal/close wakeups, including post-partial MSG_WAITALL waits. The FILE, caller-stack, Darwin and platform/debt seams are specified in the linked requests. Existing code prerequisites remain #10817 and #10755. Reverse-order requests will need an explicit agreed dependency before merge. Inconclusive bot comparisons remain unresolved. The recvfrom warning is the moved existing wrapper, not a duplicate export; the public-header audit passes.

Tested

Current head 8c68efe40774567be1ea5801f330c0351727f518, Windows 11/WinLibs GCC 15.2 and Linux/WSL imav-tao/GCC 13:

• Full production libs, relinker and registered focused test target builds pass on both platforms.
• Windows: 320/320 focused runtime tests pass. Linux: 380/380 pass. These execute production libraries.
• The bridge repair at 4527224 passes 68/68 Windows patched-library/header/native-thread checks and the separate Linux audit plus 64/64 compiler batches. Its mounted-source audit first exceeded the existing 90-second fixture bound. After composing main 228cd64, both header audits again verify 368 libraries/12,811 exports, and patched-library loading passes within the focused suites. Test timeout policy is unchanged.
• git diff --check and python3 tools/check_conventions.py --base origin/main pass with no convention warnings. The overlap report from publication remains below; subsequent coordination is tracked in the linked comment.

The standalone I/O signal ownership repair is eebaac5b24, within the read prerequisite. It acquires owning host TLS before publishing signal state, and teardown looks up existing I/O ownership without allocating an FLS slot/value. The empty-I/O and restart-only FLS-exhaustion controls fail before their fixes (0.45 s and 0.35 s) and pass afterward. New coverage includes 96 ordinary/raw native threads and guest return/exit cleanup, TSD and join ordering. This removes allocation introduced by I/O retirement; it makes no claim about unrelated existing libc-internal TLS allocation.

Launching Human Fall Flat PPSA02429 v01.017.000 found a second integration defect: patched libc renamed the descriptor class's C++ methods while consumers imported the original names. Existing patched_libraries fails on the preceding production (Windows 0.20 s), and title startup fails on both OSes. 4527224683 uses inline class wrappers over C-linkage _nid_no_patch helpers, retaining the descriptor registry and operation TLS inside libc. Patched-library loading and title startup now pass that boundary. The title reaches rendering/save setup and actual blocking UDP receives; the transport comparison belongs to #11969's follow-up evidence. No title unblock, console, TCP title or performance claim is made.

Fresh independent read-only MICRO and MACRO pairs inspected both exact repair candidates and their production callers. No confirmed defect remained in either final candidate. A further fresh MICRO/MACRO pair checked the 8c68efe merge of main's signal-pointer fixes and found no confirmed integration defect. Earlier Windows address-waking fixture flakiness remains documented in #11969; this head's full serial focused run passes.

Coverage: real blocked pipes and loopback sockets; ordinary/zero/partial completion, cancellation, disabled/pending states, 32-round completion races, scalar/vector/positioned APIs, WAITALL progress, timeout/nonblocking/DONTWAIT, zero-total recvmsg, exactly-once cleanup/TSD, dup2/close/reuse, descriptor capacity and alias visibility, child inheritance, cross-process record locks, borrowed libSceNet closure, FILE order, Windows positioned/vector serialization, append replacement, protected stdin, invalid-source dup2, registry bounds, native/raised signals, native masks, context copies/non-LIFO suspension/nonlocal exits/manual stacks and adopted threads. Exception scenarios are split into bounded cases while preserving their address-wakeup iteration coverage. New cases have 15-second CTest limits; split exception cases have 25-second limits.

Additional close coverage: pipe read/readv, socket read/readv/recv/recvmsg, partial MSG_WAITALL, disabled pending cancellation and exactly-once cleanup/TSD, unregistered callers, multiple readers, direct guest-socket close, mixed sceKernelClose and POSIX reads, surviving managed aliases and reuse, borrowed-socket POSIX close, and deterministic Windows pre-call native error slots.

Negative controls against actual production paths:

• Exact preceding production (main plus preserved prerequisites, 24c755f1c8) fails all three selected new pipe/read, socket/recv and borrowed-socket POSIX-close regressions; all pass after the fix. The original production sources were restored byte-for-byte after this control.

• Initial close candidate 0403250 fails the two mixed sceKernelClose regressions and Windows stale native-error regression; all pass after the review fixes.

• Removing native file/socket wakeups makes both blocked-cancellation cases fail because external I/O is required.

• Reinstating public-alias retention fails capacity and alias-visibility cases.

• The older signal implementation fails Windows/Linux handler-close cases; numeric/bounds retirement fails fibre/mask cases.

• Reinstating the POSIX stack registry fails context-churn RSS; removing early dup2 validation fails the invalid-source/blocked-target case in about two seconds.

• An exact prerequisite production build also reproduces the old Linux socket fixture's assumption that reset become visible within 100 immediate sends. That fixture now waits for socket error readiness with a finite deadline; production send behavior is unchanged.

Independent fresh MICRO and MACRO reviews inspected each immutable candidate and production integration. Confirmed fixes include alias close/reuse and record-lock loss, EMFILE/inheritance, global/FILE lock inversions, pipe-vector interleaving, registry growth, protected-close rejection, zero-total recvmsg, append handle replacement, handler self-close, suspended-stack retirement, non-LIFO delivery and invalid dup2 source admission. The adopted-thread deletion concern prompted an idempotent retirement guard and exit coverage, although the proposed baseline crash was not reproduced. A later errno/restart finding was dismissed after source and runtime investigation: __error aliases host errno, and delivery already saves/restores it. All 4 Windows/13 Linux errno-changing-handler cases pass without a production change.

Fresh MICRO and MACRO reviews of fdd4ba317c3b62dedbcf5b941dae8914030932b7 against 24c755f1c8 report no confirmed correctness findings. Both independently inspected immutable production code. Their preceding review found the mixed sceKernelClose/POSIX-read deadlock and stale Windows pre-call native error slots; both were reproduced, fixed and re-reviewed by two new tasks. The only later change is the concise debt wording in bcf5df1c7756b848fce3c3baf16fe1eb07d36e96. Focused close-fix comparison: 24c755f...bcf5df1.

Limits: no PS5, title, macOS, Emscripten or ARM64 run; no general asynchronous cancellation; Unix SIGUSR2 must be free; native drivers without interruption support can delay cancellation and close indefinitely; stdio/replacement can still wait indefinitely; Windows blocking MSG_PEEK|MSG_WAITALL remains unsupported; existing recvmsg output-copy/SCM_RIGHTS gaps remain in debt. WSL used the existing memory-comparison fallback because userfaultfd write protection is unavailable.

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: fix(libkernel): return ESRCH from join, detach and cancel of a released thread handle #10817, fix(libc): drop the kernel state of a descriptor that fclose or freopen closes #10755
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes

Validation log excerpts

Final production candidate fdd4ba317c3b62dedbcf5b941dae8914030932b7
Windows full libs + relinker: PASS
Windows focused runtime: 313/313 PASS, 22.31 seconds total
Windows public headers: 65/65 PASS, 50.62 seconds total
Linux full libs + relinker: PASS
Linux focused runtime: 377/377 PASS, 9.49 seconds total
Linux public headers: 368 libraries / 12811 exports audited;
64/64 native compiler batches PASS, 71.93 seconds total

Overlap report

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
PR 7090	core/libs/prx/libkernel/Socket/src/Udp.cpp L1717-1719
PR 7135	docs/dev/debt/libkernel.md L206-208
PR 8916	docs/dev/debt/libkernel.md L206
PR 9938	core/libs/prx/libkernel/Pthread/src/Thread.cpp L762-763
PR 10552	core/libs/prx/libkernel/File/src/Stdio.cpp L1561-1562, L1593-1594
PR 11341	docs/dev/debt/libkernel.md L187
PR 11831	core/libs/prx/libc/include/FileStream.hpp L7, L63-65, L150, L163-165, L205-206, L337-338, L340-341, L343
PR 11896	core/libs/prx/libkernel/Socket/src/Udp.cpp L772-773, L790-798, L804-806, L1395-1401, L1551-1557, L1717-1727 docs/dev/debt/libkernel.md L233-234

Exports introduced here that are already present in main:

• recvfrom

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 3986	core/libs/prx/libkernel/Pthread/src/Thread.cpp core/libs/prx/libkernel/System/src/Signals.cpp
PR 4563	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/Pthread/include/Pthread.hpp core/libs/prx/libkernel/Pthread/src/Thread.cpp
PR 4630	core/libs/prx/libkernel/System/src/Exception.cpp core/libs/tests/GuestRaiseException.cpp docs/dev/debt/libkernel.md
PR 5021	core/libs/prx/libc/CMakeLists.txt core/libs/prx/libc/include/FileStream.hpp core/libs/prx/libc/src/JumpBuffer.cpp core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Open.cpp core/libs/prx/libkernel/File/src/Stdio.cpp core/libs/prx/libkernel/Pthread/include/Pthread.hpp core/libs/prx/libkernel/Pthread/src/Thread.cpp core/libs/prx/libkernel/Socket/src/Udp.cpp core/libs/prx/libkernel/System/src/Exception.cpp core/libs/prx/libkernel/System/src/SignalSets.cpp docs/dev/debt/libc.md docs/dev/debt/libkernel.md
PR 5836	core/libs/prx/libSceNet/Export.cpp
PR 6085	core/libs/prx/libc/CMakeLists.txt
PR 6149	docs/dev/debt/libkernel.md
PR 6713	core/libs/prx/libc/CMakeLists.txt docs/dev/debt/libc.md
PR 6936	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Open.cpp core/libs/prx/libkernel/File/src/Stdio.cpp core/libs/tests/GuestKernelAppArchive.cpp docs/dev/debt/libkernel.md
PR 7081	core/libs/prx/libc/CMakeLists.txt core/libs/prx/libc/src/FileAndHeap.cpp docs/dev/debt/libc.md
PR 7088	core/libs/prx/libkernel/File/include/File.hpp core/libs/prx/libkernel/Socket/src/Udp.cpp docs/dev/debt/libkernel.md
PR 7105	docs/dev/debt/libkernel.md
PR 7121	core/libs/prx/libkernel/File/src/DescriptorControl.cpp core/libs/prx/libkernel/File/src/Open.cpp core/libs/prx/libkernel/File/src/Stdio.cpp docs/dev/debt/libkernel.md
PR 7149	core/libs/prx/libkernel/Pthread/include/Pthread.hpp core/libs/prx/libkernel/Pthread/src/Thread.cpp docs/dev/debt/libkernel.md
PR 7192	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/System/src/Exception.cpp docs/dev/debt/libkernel.md
PR 7386	docs/dev/debt/libc.md
PR 7447	core/libs/prx/libc/CMakeLists.txt core/libs/prx/libc/include/FileStream.hpp core/libs/prx/libc/src/FileAndHeap.cpp core/libs/prx/libkernel/File/src/Stdio.cpp docs/dev/debt/libc.md docs/dev/debt/libkernel.md
PR 7509	docs/dev/debt/libkernel.md
PR 7648	docs/dev/debt/libc.md
PR 7715	core/libs/prx/libc/CMakeLists.txt docs/dev/debt/libc.md
PR 7804	core/libs/prx/libkernel/Pthread/include/Pthread.hpp core/libs/prx/libkernel/Pthread/src/Thread.cpp docs/dev/debt/libkernel.md
PR 8150	core/libs/prx/libc/CMakeLists.txt docs/dev/debt/libc.md
PR 8417	core/libs/prx/libc/include/Api.hpp docs/dev/debt/libc.md
PR 8592	core/libs/prx/libkernel/Pthread/include/Pthread.hpp core/libs/prx/libkernel/Pthread/src/Thread.cpp core/libs/prx/libkernel/System/src/Exception.cpp
PR 8605	core/libs/prx/libkernel/Socket/src/Udp.cpp docs/dev/debt/libkernel.md
PR 8937	core/libs/prx/libc/CMakeLists.txt core/libs/prx/libc/include/Api.hpp docs/dev/debt/libc.md
PR 9156	core/libs/prx/libkernel/CMakeLists.txt docs/dev/debt/libkernel.md
PR 9199	docs/dev/debt/libkernel.md
PR 9242	docs/dev/debt/libkernel.md
PR 9284	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Stdio.cpp docs/dev/debt/libkernel.md
PR 9293	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Stdio.cpp docs/dev/debt/libkernel.md
PR 9297	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Stdio.cpp docs/dev/debt/libkernel.md
PR 9302	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Stdio.cpp docs/dev/debt/libkernel.md
PR 9371	docs/dev/debt/libc.md
PR 9383	core/libs/prx/libkernel/CMakeLists.txt
PR 9389	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/System/src/Signals.cpp docs/dev/debt/libkernel.md
PR 9447	docs/dev/debt/libc.md
PR 9473	core/libs/prx/libkernel/File/src/Open.cpp core/libs/prx/libkernel/File/src/Stdio.cpp
PR 9491	core/libs/prx/libkernel/CMakeLists.txt docs/dev/debt/libkernel.md
PR 9526	core/libs/prx/libc/src/FileAndHeap.cpp docs/dev/debt/libc.md
PR 9557	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/DescriptorControl.cpp docs/dev/debt/libkernel.md
PR 9566	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/DescriptorControl.cpp docs/dev/debt/libkernel.md
PR 9575	core/libs/prx/libkernel/CMakeLists.txt docs/dev/debt/libkernel.md
PR 9578	core/libs/prx/libSceNet/Export.cpp core/libs/prx/libkernel/Socket/src/Udp.cpp docs/dev/debt/libkernel.md
PR 9713	core/libs/prx/libkernel/Socket/src/Udp.cpp
PR 9775	core/libs/prx/libc/CMakeLists.txt
PR 9929	core/libs/prx/libc/CMakeLists.txt docs/dev/debt/libc.md
PR 9932	core/libs/prx/libSceNet/Export.cpp
PR 9947	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/File/src/Open.cpp docs/dev/debt/libkernel.md
PR 10192	docs/dev/debt/libkernel.md
PR 10319	core/libs/prx/libkernel/File/src/Open.cpp docs/dev/debt/libkernel.md
PR 10324	core/libs/prx/libc/CMakeLists.txt core/libs/prx/libc/include/Api.hpp core/libs/prx/libc/include/FileStream.hpp core/libs/prx/libc/src/FileAndHeap.cpp docs/dev/debt/libc.md
PR 10433	core/libs/prx/libc/src/FileAndHeap.cpp docs/dev/debt/libc.md
PR 10442	docs/dev/debt/libkernel.md
PR 10502	core/libs/prx/libSceNet/Export.cpp
PR 10558	docs/dev/debt/libkernel.md
PR 10620	core/libs/prx/libc/src/FileAndHeap.cpp docs/dev/debt/libc.md
PR 10725	docs/dev/debt/libkernel.md
PR 10749	core/libs/prx/libc/include/Api.hpp core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/System/src/Signals.cpp docs/dev/debt/libc.md docs/dev/debt/libkernel.md
PR 11002	core/libs/prx/libkernel/CMakeLists.txt core/libs/prx/libkernel/System/src/Signals.cpp docs/dev/debt/libkernel.md
PR 11005	docs/dev/debt/libkernel.md
PR 11042	docs/dev/debt/libkernel.md
PR 11085	docs/dev/debt/libkernel.md
PR 11134	core/libs/prx/libSceNet/Export.cpp
PR 11215	docs/dev/debt/libkernel.md
PR 11223	core/libs/prx/libkernel/Socket/src/Udp.cpp docs/dev/debt/libkernel.md
PR 11267	docs/dev/debt/libkernel.md
PR 11332	docs/dev/debt/libc.md

Human Fall Flat read/transport/write startup socket evidence on Windows and Linux, including exact heads, call counts and limitations: #11969 (comment). No full-playthrough or title-unblock claim.

[oneandonlydean]

@Felipebmaitan thanks for the close fix. Close now wakes admitted readers and waits for them to drain, which resolves the earlier hang. Five points from two reads of the current head (21d5a97), none blocking:

1. Windows nonblocking receives. In ReceiveNative's cancellable branch, a nonblocking or MSG_DONTWAIT receive still creates an event, submits an overlapped WSARecv, waits 0 ms, then runs CancelIoEx and a blocking WSAGetOverlappedResult, then closes the event. An empty poll used to cost one recv returning WSAEWOULDBLOCK. A title that polls a UDP socket every frame now pays several kernel transitions each time. A nonblocking receive can't block, so could it use plain recv/recvfrom and skip the overlapped path?
2. Windows contended positioned reads. File::IoPositionedTransfer polls try_lock with a 1 µs WaitUntil, and nothing notifies the condition when the holder unlocks; only an interrupt does. With two threads reading the same descriptor positionally, the waiter either spins or sleeps a timer quantum on every acquisition. Could the unlock notify waiters, or could the wait use the existing lock with the interrupt as a wake?
3. SIGUSR2 disposition. NativeIo::State throws unless the disposition is SIG_DFL. SIG_IGN is inherited across exec, so a launcher that leaves SIGUSR2 ignored makes the first cancellable POSIX read abort the process. Could SIG_IGN be treated as free and replaced, keeping the throw for a real installed handler?
4. Per-call cost on Linux. A POSIX read of a regular file now goes through IoRestart, IoPoint and EndIoSignals (about six SetupBlocks, so roughly twelve rt_sigprocmask calls), a heap NativeIo::State with its own mask changes and sigpending, and about three GuestDescriptorAccess::Get calls, each taking a process-wide mutex plus an erase_if over the whole map. Could you post a microbenchmark of main against this head for read on a regular file and recvfrom on a non-blocking UDP socket, from several threads? If the overhead is measurable, could regular files skip the signal bracketing?
5. SignalManager::Run (IoSignals.cpp:180-195) keeps looping lock, wake and yield with no backoff while any state is pending, even when that state has no registered wait. With a driver that ignores SIGUSR2 this spins a core until the read returns. Would a short sleep or backoff after the first few rounds be acceptable?

[oneandonlydean]

@Felipebmaitan thanks for the close fix. Close now wakes admitted readers and waits for them to drain, which resolves the earlier hang. Five points from two reads of the current head (21d5a97), none blocking:

1. Windows nonblocking receives. In ReceiveNative's cancellable branch, a nonblocking or MSG_DONTWAIT receive still creates an event, submits an overlapped WSARecv, waits 0 ms, then runs CancelIoEx and a blocking WSAGetOverlappedResult, then closes the event. An empty poll used to cost one recv returning WSAEWOULDBLOCK. A title that polls a UDP socket every frame now pays several kernel transitions each time. A nonblocking receive can't block, so could it use plain recv/recvfrom and skip the overlapped path?
2. Windows contended positioned reads. File::IoPositionedTransfer polls try_lock with a 1 µs WaitUntil, and nothing notifies the condition when the holder unlocks; only an interrupt does. With two threads reading the same descriptor positionally, the waiter either spins or sleeps a timer quantum on every acquisition. Could the unlock notify waiters, or could the wait use the existing lock with the interrupt as a wake?
3. SIGUSR2 disposition. NativeIo::State throws unless the disposition is SIG_DFL. SIG_IGN is inherited across exec, so a launcher that leaves SIGUSR2 ignored makes the first cancellable POSIX read abort the process. Could SIG_IGN be treated as free and replaced, keeping the throw for a real installed handler?
4. Per-call cost on Linux. A POSIX read of a regular file now goes through IoRestart, IoPoint and EndIoSignals (about six SetupBlocks, so roughly twelve rt_sigprocmask calls), a heap NativeIo::State with its own mask changes and sigpending, and about three GuestDescriptorAccess::Get calls, each taking a process-wide mutex plus an erase_if over the whole map. Could you post a microbenchmark of main against this head for read on a regular file and recvfrom on a non-blocking UDP socket, from several threads? If the overhead is measurable, could regular files skip the signal bracketing?
5. SignalManager::Run (IoSignals.cpp:180-195) keeps looping lock, wake and yield with no backoff while any state is pending, even when that state has no registered wait. With a driver that ignores SIGUSR2 this spins a core until the read returns. Would a short sleep or backoff after the first few rounds be acceptable?

[Felipebmaitan]

Superseded by #12682 (read cancellation layer, head 3b0d2c6). Original branch, commits, review history and order tracker are preserved. The replacement composes current main and passes fresh MICRO/MACRO review, full production builds, 345 Windows / 405 Linux focused tests, both header checks and Human Fall Flat startup smoke runs. Transport and writes remain separate following layers; nothing is merged.