fix(kernel,audioout,relinker): thread/mutex correctness fixes and --windows-gui
### libkernel (threads, mutexes, condition variables)
- **Deadlock in `scePthreadCondTimedwait` (normal/error-check mutexes):** `_owner` was not restored after the wait. If another thread locked and unlocked the mutex during the wait, the waiter woke up owning the mutex while `_owner` was empty. Its next `scePthreadMutexUnlock` returned `EPERM`, leaving the mutex locked forever (`lock → timedwait → unlock` pattern).
- **Wrong timeout code:** `SCE_KERNEL_ERROR_ETIMEDOUT` in `Cond.cpp` was `0x80020062`; it is `0x8002003C` (FreeBSD `ETIMEDOUT` = 60, matching `Semaphore.hpp`).
- **`scePthreadSelf()`** returned `nullptr` on Linux, and on both platforms for threads not created by `scePthreadCreate` (including the main thread). Such threads now get a lazily adopted, non-joinable handle (`scePthreadJoin`/`Detach` on it return `EINVAL`).
- **Linux `scePthreadExit`** skipped `FinishThread`, so `scePthreadJoin` got a garbage return value. Cleanup now runs from an RAII guard, so it survives glibc forced unwinding.
- **Linux detached threads** were never freed. They now use the same reference count as on Windows.
- **Recursive mutex unlock** by a non-owner, or with a zero count, was UB (`std::recursive_timed_mutex::unlock`). It now returns `SCE_KERNEL_ERROR_EPERM`, and `_owner`/`_count` are also saved and restored around a recursive cond wait.
- **`sceKernelLseek`** returned `int` instead of `off_t`; offsets above `INT_MAX` threw.
- **`sceKernelDeleteSema`** implemented: waiters are woken with `SCE_KERNEL_ERROR_EACCES` and the object is freed once they leave.
- `sceAudioOutOpen` accepted `freq == 0` / `len == 0`; the frequency was later used as a divisor. It now returns `SCE_AUDIO_OUT_ERROR_INVALID_SAMPLE_FREQ` / `INVALID_SIZE`.
On Linux scePthreadExit called pthread_exit, whose forced unwind runs
the personality routine of every guest frame. Those resolve to
libc.prx's __gxx_personality_v0, which cannot read libgcc's unwind
context and aborts, so guest_thread_self crashed on Linux.
Record a jump target in the thread start routine and longjmp back to it
from scePthreadExit instead. Like _endthreadex on Windows, this skips
the guest frames without unwinding them; the start routine still
releases the thread normally. Threads not created by scePthreadCreate
keep the pthread_exit fallback.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018emyZ4Rk8yoSJvxTFBe4DL
Bring in the resolved#25 branch, which already carries upstream main
including boykopovar/AnyPS5#10, and re-apply the KernelErrors.hpp
refactor onto the rewritten files:
- Equeue: drop the local EQUEUE_ERROR_* constants and use the shared
SCE_KERNEL_ERROR_* names.
- Pthread Attr/Thread: drop local error constants and include
KernelErrors.hpp.
- Pthread Cond/Mutex: take upstream; they no longer define local
SCE_KERNEL_ERROR_* constants.
- Semaphore: drop KERNEL_SEMA_ERROR_* in favour of KernelErrors.hpp.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018emyZ4Rk8yoSJvxTFBe4DL
Resolve conflicts with boykopovar/AnyPS5#10 (sce module loading) and
later upstream changes.
- Pthread Cond/Mutex: take upstream. It already releases mutex
ownership across condition waits, returns ETIMEDOUT 0x8002003C, and
rejects unlocks from non-owners for every mutex type.
- Pthread Thread: keep upstream's lifecycle, Windows adoption and job
affinity; re-apply the POSIX fixes on top: reference-counted
release on join/detach/exit, self-join rejection on both platforms,
and scePthreadSelf adoption for threads not created by
scePthreadCreate.
- Semaphore: keep upstream's TimedWait::Condition and wait tracing;
re-apply sceKernelDeleteSema waking waiters with EACCES.
- Relinker: keep --windows-gui alongside upstream's --skip-sce-module,
trampoline-aware Patch signature and new Python tests.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_018emyZ4Rk8yoSJvxTFBe4DL
feat(fiber): implement fibers on linux
The state machine, the cross-thread resume handling and the entry trampoline are now shared between platforms. Only the context switch differs:
- `Aps5FiberSwitchStack` saves the SysV callee-saved set (`rbp rbx r12–r15`, MXCSR and the x87 control word) instead of the Win64 set.
- `Aps5FiberTrampoline` passes the fiber in `rdi`.
- There is no TIB to update, so `CurrentBounds`/`SetBounds` do nothing on Linux.
fix(agc): keep BGRA color targets resident between draws
BGRA color targets fail the exact guest-texture-format lookup and fall back to a synchronous framebuffer round-trip for every draw. This keeps them resident using compatible RGBA storage with the original BGRA attachment view. Presentation checks the actual display channel order, attachment views request only attachment usage, and host-write barriers include the host stage.
This follows the existing linked PRX/Vulkan path. It adds no interpreter, separate runtime process, or new shader compilation path.
feat(agc): ACB markers, index buffer unbinding and address in errors
- The DCB marker helpers move from an anonymous namespace in `DcbState/src/Marker.cpp` to `Agc::Marker::Push`/`Pop` (declared in `Marker.hpp`). The ACB entry points use the same helpers, so ACB and DCB markers produce the same custom NOP packets. `SetMarker` is a push followed by a pop, and returns the push packet, as on the DCB. The color is ignored on both, as before.
- `sceAgcDcbSetIndexBuffer` emits the packet with address 0 instead of rejecting it. A non-null address still has to be 2-byte aligned.
- `Agc::Command::CheckAddress` includes the rejected address in its message.
feat(agc): DMA_DATA cache policy hints and GDS source/destination
- **Cache policy:** the source (bits 13–14) and destination (bits 25–26) cache-policy fields of `DMA_DATA` dword 1 are accepted. They're cache hints with no effect on the copied data. The remaining reserved bits are still rejected.
- **GDS storage:** GDS is modelled as one 64 KiB, zero-initialised store shared by the whole device, behind a mutex.
- **GDS source/destination:** select `1` reads from or writes to that store:
- the validator requires the high address dword to be 0 and the range to fit inside the 64 KiB;
- `Execute` copies between memory, immediate data and GDS in any combination;
- `ResolveStore` returns nothing for a GDS destination, since there's no guest memory to write, and for a GDS source it resolves the bytes from the store.
- Register and prefetch destinations are still rejected.
- Nothing is faked for shaders: a shader that uses GDS still fails validation (unsupported descriptor role).
fix(agc): reject primitive restart only for indexed draws
- `DecodeState` still reads `0x24b` (so it stays in the recorded draw-key registers) but no longer rejects on it.
- `DrawRejection` takes an `indexed` flag, and reports restart only for indexed draws.
- The driver's draw path rejects an indexed draw with restart enabled before the precheck, so indexed draws with restart are still refused when the precheck is off. Restart itself is still not implemented.
fix(agc): keep a replaced Vulkan device alive
In the presenter path, after `PrepareForReplacement()` (which flushes cached textures, publishes shadows and waits for idle), the old device is kept in a function-local static list instead of being released. Its child objects stay valid for as long as anything still refers to them. The cost is one idle device for the life of the process; the replacement happens once, at the first flip.
feat(json2): implement sce::Json values, strings, arrays, objects, parser and null-access callback
- **Layouts:** `String`, `Array`, `Object` and `Value` are pointer-sized handles to host-side storage. The title stores `String`/`Array`/`Object` at an 8-byte stride and `Value` in 0x20-byte slots, so none of them can be bigger. The pair an `Object::iterator` dereferences to has its key at +0x00 and its `Value` at +0x10, which the title reads directly. Array and object iterators are 8 bytes.
- **Parser:** a strict RFC 8259 parser.
- Numbers: integers that fit `int64` become `Integer`, larger non-negative ones `UInteger`, and everything else `Real`.
- Strings: `\u` escapes, including surrogate pairs, are decoded to UTF-8.
- Errors: on any error `parse` returns a negative code (`0x80920101`) and leaves the target value untouched.
- **Serialisation:** compact output. Strings are escaped; reals use `%.17g`, and a non-finite real throws.
- **Null access:** a typed getter on a value of another type, a missing key or an out-of-range index goes through the null-access callback. The callback's result is used when its type matches (any type for `operator[]`). Otherwise a static default value of the requested type is returned. `terminate` clears the callback.
- **Integer getters:** `getInteger`/`getUInteger`/`getReal` convert between the integer kinds and to real where the value fits.
- `setGlobalNullAccessCallBack` (NID `00oCq0RwSAY`) was identified by hashing its mangled name.
- `MemAllocator`'s base constructor and destructor install the abstract class's vtable. A derived allocator replaces it with its own, so its entries (which throw, like a pure virtual call) run only if a pure virtual is called. `Initializer::initialize` throws on a null parameter instead of silently accepting it.
- Mutating a value that isn't the expected type (`referArray` on a string, etc.) throws.
feat(libkernel): pread, pwrite, getdents and getdirentries on linux
- **pread/pwrite:** use the existing `NativePread`/`NativePwrite` and don't move the file offset. A null buffer with a non-zero size returns `EFAULT`, a negative offset `EINVAL`, and host failures go through `SceErrorFromErrno`.
- **getdirentries:**
- reads host `getdents64` records and converts them to the FreeBSD `struct dirent` layout the guest expects (`d_fileno` u32, `d_reclen` u16, `d_type` u8, `d_namlen` u8, NUL-terminated name, record padded to 4 bytes);
- `*basep` gets the directory position from before the read;
- when a record doesn't fit in the guest buffer, the descriptor is moved with `lseek` back to that entry's cookie, so the next call resumes there;
- if not even the first record fits, the call returns `EINVAL` without consuming anything, and a name longer than 255 bytes returns `ENAMETOOLONG`.
- **getdents** is `getdirentries` without `basep`.
- Adds the guest `EFAULT`, `EINVAL` and `ENAMETOOLONG` constants.
feat(relinker): lower register-form EXTRQ and relocate following instructions for short sites
- **Lowering:** register-form EXTRQ is emitted out of line with SSE2 only. Two scratch XMM registers (distinct from both operands) are spilled; one gets the index (`src >> 8 & 63`), the other `(64 - length) & 63` (`length` masked to 6 bits, so 0 means 64). The destination is shifted right by the index, then left and right by the complement, which keeps the low `length` bits. INSERTQ register form is still reported as unsupported.
- **Short sites:** when a site is shorter than a `jmp rel32`, the converter extends it over the instructions that follow it until it's 5 bytes long:
- following EXTRQ/INSERTQ instructions (before any other instruction) are lowered into the same stub, via a new `MatchSequence`;
- other instructions are copied after the lowered body (`LowerOutOfLine`'s new `trailing` bytes), but only if they're sequential, have no RIP-relative operand and no branch target;
- anything else (a return, a branch, RIP-relative code, the segment ending) still fails the relink, as does a branch target inside the extended site.
- `docs/TechnicalDebt.md` now lists only the INSERTQ register form, MONITORX/MWAITX/CLZERO/RDPRU/MCOMMIT and SHA-NI.
feat(userservice): user name, user number and accessibility settings of the initial user
They report the initial user, following `sceUserServiceGetGamePresets`:
- **Name:** `Player`. A buffer shorter than 17 bytes returns `SCE_USER_SERVICE_ERROR_BUFFER_TOO_SHORT` (`0x8096000A`).
- **User number:** 1.
- **Accessibility settings:** 0, the setting-off value on a fresh console.
fix(libc): don't walk guest frame pointers when reporting an unhandled exception
Only log __builtin_return_address(0), which doesn't depend on the guest's frames. After the fix, the log prints the exception type and the terminate handler prints what().
feat(libc): mspace memalign, malloc stats, thread-unsafe flag and nested mspaces
- Memalign uses the existing aligned allocation.
- Stats fill `SceLibcMallocManagedSize` (size checked against 0x28) from new `inUse`/`peakInUse` counters on the arena. The system size is the arena's extent.
- The thread-unsafe flag is accepted: the implementation always locks, so skipping locks is only an optimisation it doesn't need. Other flags are still rejected.
- An overlapping create is allowed only when the new range lies entirely inside a live block of the other mspace. Every other overlap is still rejected.
feat(libkernel): hinted and no-overwrite mappings, virtual range names and usleep
- **Non-fixed mapping with an address hint:** `MapAligned` threw `Non-fixed mapping address hints are not implemented`. On the PS5 the hint is where the search starts. On Linux the mapping is now placed at the first free, aligned range at or above the hint, found from `/proc/self/maps` and mapped with `MAP_FIXED_NOREPLACE` so nothing existing can be replaced. `RequireAvailable` is now only checked for fixed mappings, because a hint may point at used memory. Windows still throws.
- **`SCE_KERNEL_MAP_NO_OVERWRITE` (`0x80`):** it was rejected as an unknown flag. With `MAP_FIXED` on Linux it maps with `MAP_FIXED_NOREPLACE`, and it no longer takes the "remap fixed into registered" path, so it fails instead of replacing a registered range. The unsupported-flags error now includes the flag value.
- **Virtual range names:** `sceKernelSetVirtualRangeName` was `NotImplemented` and `sceKernelClearVirtualRangeName` wasn't exported. Names are now kept in a registry, split and trimmed on overlap. `sceKernelVirtualQuery` reports the name and splits its reported region at name boundaries. `sceKernelMapNamedDirectMemory` and `sceKernelMapNamedFlexibleMemory` now name the range instead of ignoring the name.
- **`usleep`** is exported (POSIX, beside `nanosleep`).
feat(input): support game controllers and configurable bindings
- Add optional per-game keyboard and mouse bindings through `anyps5-input.ini` or `ANYPS5_INPUT_CONFIG`. Unspecified actions keep their built-in bindings, and repeated action entries provide alternatives.
- Keep wheel presses and focus-loss state handling compatible with dynamically sized bindings. Report configuration errors through the existing video and pad failure path.
- Document the format and remove the resolved input-mapping debt item.
fix(relinker): report buffered output write failures
Small buffered writes can report success even when flushing fails. Both FileWriter::Write overloads now close the stream before checking for errors.
Added Linux regression tests using /dev/full. Both fail before the fix
fix(recompiler): widen constants of 64-bit operands instead of repeating them
TranslationContext::readU32Pair builds a 64-bit source from two 32-bit reads, taking the high half from offsetOperand(operand, 1). For registers that is the next register. For a literal or an inline constant, offsetOperand returns the same operand, so the constant ended up in both halves: the literal L was read as L:L instead of 0:L.
feat(libc): sscanf_s on linux
sscanf_s was a NotImplemented placeholder in libc/Export.cpp, and a retail title uses it to parse strings. It's now implemented on Linux. Windows still throws NotImplemented.
feat(agc): emit indexed indirect draws
This implements `sceAgcDcbDrawIndexIndirect` and `sceAgcDcbDrawIndexIndirectMulti`. They were stubs that aborted with "not implemented".
They follow the existing non-indexed `sceAgcDcbDrawIndirect` / `DrawIndirectMulti`:
- `DrawIndexIndirect` emits `DRAW_INDEX_INDIRECT` (0x25): data offset, base vertex / start instance SGPR locations, draw initiator (5 dwords, matching `GetSize` = 20).
- `DrawIndexIndirectMulti` emits `DRAW_INDEX_INDIRECT_MULTI` (0x38): the same fields, plus the draw-index location / count-indirect control word, count, count address and stride (10 dwords, matching `GetSize` = 40). The stride must be at least 20, the size of an indexed record.
fix: support SDK startup, linear framebuffers and indexed draws
Fix SDK startup with an empty allocator replacement table, add linear framebuffer presentation through the current presentation path, and implement explicit indexed draws with packet validation. Retain buffer-address regression coverage for the correction already upstream, and restrict the POSIX exception-runtime test to Unix so Windows test builds succeed.
Also fix two issues reproduced after rebasing: render-target format lookup probed unsupported formats by repeatedly throwing exceptions, costing about 78 ms per draw; the lookup now searches supported entries directly. Recorder buffers now use Windows FLS cleanup to avoid MinGW freeing emulated TLS storage before running their destructors. Cross-DLL tests check independent values and destruction on standard and native Windows threads.
SamG-Coder