feat(libkernel): JIT shared memory on macOS hosts

core/libs/prx/libkernel/DirectMemory/DirectMemory.cpp

@@ -1216,6 +1216,9 @@ int DoMapShared(void** addr, size_t len, int prot, HostSharedMemory object) {
     try {
 #ifdef _WIN32
         GuestArena::GuestArenaMap_nid_postfix(mapped, len, object, 0, WinProtFromPosix(nativeProtection));
+#elif defined(__APPLE__)
+        if (::mmap(mapped, len, nativeProtection & ~PROT_EXEC, MAP_SHARED | MAP_FIXED, object, 0) == MAP_FAILED) throw std::system_error(errno, std::generic_category(), "map shared memory");
+        if ((nativeProtection & PROT_EXEC) != 0 && ::mprotect(mapped, len, nativeProtection) != 0) throw std::system_error(errno, std::generic_category(), "make shared memory executable");
 #else
         if (::mmap(mapped, len, nativeProtection, MAP_SHARED | MAP_FIXED, object, 0) == MAP_FAILED) throw std::system_error(errno, std::generic_category(), "map shared memory");
 #endif

core/libs/prx/libkernel/DirectMemory/Jit.cpp

@@ -26,6 +26,9 @@
 #include <fcntl.h>
 #include <sys/mman.h>
 #include <unistd.h>
+#ifdef __APPLE__
+#include <filesystem>
+#endif
 #endif
 
 namespace {
@@ -140,8 +143,19 @@ int DuplicateDescriptor(int descriptor) {
 }
 #else
 int CreateBacking(std::size_t bytes, int* descriptor, std::shared_ptr<JitObject>* object) {
+#ifdef __APPLE__
+    auto path = (std::filesystem::temp_directory_path() / "anyps5-jit-XXXXXX").string();
+    const int file = mkstemp(path.data());
+    if (file < 0) return HostFailure("create JIT shared memory", errno);
+    if (unlink(path.c_str()) != 0 || fcntl(file, F_SETFD, FD_CLOEXEC) != 0) {
+        const int error = errno;
+        ::close(file);
+        return HostFailure("create JIT shared memory", error);
+    }
+#else
     const int file = memfd_create("anyps5 jit shared memory", MFD_CLOEXEC);
     if (file < 0) return HostFailure("create JIT shared memory", errno);
+#endif
     const bool representable = bytes <= static_cast<std::uint64_t>(std::numeric_limits<off_t>::max());
     if (!representable || ftruncate(file, static_cast<off_t>(bytes)) != 0) {
         const int error = !representable || errno == EFBIG ? ENOMEM : errno;

core/libs/tests/GuestJitSharedMemory.cpp

@@ -187,6 +187,8 @@ int main() {
     WriteThroughAliasRunsThroughExecutableView();
     OneViewWritesAndRuns();
     Dup2KeepsTheDescriptorTableInStep();
+#ifndef __APPLE__
     MmapOfTheDescriptorSharesTheObject();
+#endif
     InvalidArgumentsAreRejected();
 }

docs/dev/debt/libkernel.md

@@ -120,6 +120,7 @@
 - [Guest thread scheduling](../../../core/libs/prx/libkernel/Pthread/src/Thread.cpp) (libkernel) - the policies and priorities that `sched_get_priority_max`, `sched_get_priority_min`, `pthread_setschedparam` and `scePthreadSetschedparam` accept, and those a created thread takes, follow a PS5 (firmware 13.60), but they are only recorded and reported: they are not applied to host threads, so a title's priority order has no effect on host scheduling
 - [sceKernelMprotect, sceKernelMtypeprotect](../../../core/libs/prx/libkernel/DirectMemory/DirectMemory.cpp) (libkernel) over a range with a gap that is not registered in the allocation tracker protect the registered parts and skip the gap, and return 0, as FreeBSD 11.4's [`vm_map_protect`](https://github.com/freebsd/freebsd-src/blob/release/11.4.0/sys/vm/vm_map.c) does (it has no check for holes); a range whose start is unmapped (also when later pages are mapped) returns `SCE_KERNEL_ERROR_EINVAL` (`mprotect`: `EINVAL`) because an existing test pins that for an unmapped start; FreeBSD answers 0 there too (`vm_map_protect` takes the next entry), so this errno is not taken from the source and was not changed. `sceKernelMprotect` threw for a gap before, which ended the title. Every part is checked for a GPU or image pin before the first one is protected; the parts are then protected one after the other, so a host protection failure in a later part leaves the earlier parts changed (FreeBSD's `vm_map_protect` cannot fail in its second pass). The behaviour on the console is not known. None of this was measured on a console.
 - [sceKernelMprotect protection bits](../../../core/libs/prx/libkernel/DirectMemory/Export.cpp) (libkernel) - bits outside `0x3F7` are dropped instead of refused, in the POSIX `mprotect` as well (FreeBSD 11.4 masks the argument, `sys/vm/vm_mmap.c:613`; the console was measured only through `sceKernelMprotect`), as a PS5 (firmware 13.60) drops them: Onimusha: Way of the Sword (PPSA27836 01.000.001) maps direct memory with `0xf2` and then protects it with `0xfffffff2`, which the console reports as `0xf3` and then `0x3f3`. In a payload process the console also drops the CPU execute bit (`0x4`) of flexible, direct and JIT memory, and POSIX `mprotect` with write and execute reports write only (`0x2`). A title process may keep execute, so it is kept here. The map calls and `sceKernelMtypeprotect` refuse undefined bits on the console with `EINVAL` (`0x80020016`), and the direct map calls refuse execute; here both throw for undefined bits, and the direct map calls map execute
+- [JIT shared memory on macOS](../../../core/libs/prx/libkernel/DirectMemory/Jit.cpp) (libkernel) - the object is an unlinked `mkstemp` file in the temporary directory, as `shm_open` makes its anonymous objects. macOS refuses an executable shared file mapping (`EPERM`), so an executable view is mapped without execute and then given it with `mprotect`. Under Rosetta 2 (Apple M2, macOS 26), code rewritten through one view after it ran from another keeps its old translation until the executable view's protection changes; first writes and a single read-write-execute view, as XPSemu PPSA97358 and VLC PPSA85300 map, are seen. `mmap` of a JIT descriptor is not supported, as file-backed guest `mmap` throws on macOS.
 - [JIT shared memory](../../../core/libs/prx/libkernel/DirectMemory/Jit.cpp) (libkernel) - `sceKernelJitCreateSharedMemory`, `sceKernelJitCreateAliasOfSharedMemory` and `sceKernelJitMapSharedMemory` take the signatures of [Orbital](https://github.com/AlexAltea/orbital)'s `tools/ida/db_types.json` (name, length, maximum protection, descriptor out; descriptor, maximum protection, descriptor out; descriptor, protection, address in and out), which ps4sdk, ps5-jar-loader and the call sites in RetroArch PPSA99169, XPSemu PPSA97358 and VLC PPSA85300 match; SharpProspero declares the alias as returning the descriptor. The object is a host anonymous file shared by its descriptors (a `memfd` on Linux, a delete-on-close temporary file mapped as an executable section on Windows), the alias is a duplicate descriptor, and every map maps the whole object, so writes through one view show through the others; a non-null address is a hint, as in `mmap` without `MAP_FIXED`. Not measured on the console: a length that is zero or not a multiple of 16 KiB and protection bits other than CPU read, write and execute are `EINVAL`; an alias or a map asking for more than the descriptor's maximum protection (write and execute counting as read) is `EACCES`, as FreeBSD 11's `vm_map_protect` treats a protection past `max_protection`; a descriptor that is not a JIT object is `EBADF`. The name is ignored, the objects do not count against the flexible memory budget, `mprotect` of a view is not limited to the maximum protection, and `mmap` of the descriptors maps the object as a [regular file](../../../core/libs/prx/libkernel/DirectMemory/PosixMemory.cpp), so its views and the JIT views see each other's writes. On Linux the `memfd` is created without `MFD_EXEC`, which Linux 6.3 added and older kernels reject with `EINVAL`, and relies on the protection of the mapping: the flag only sets the file's execute permission bits, and the direct memory `memfd` is mapped executable the same way; a host with `vm.memfd_noexec` set to 2 refuses the creation and the call throws
 - [kqueue and kevent](../../../core/libs/prx/libkernel/Equeue/src/Kqueue.cpp) (libkernel) follow FreeBSD 11's [`kern_event.c`](https://github.com/freebsd/freebsd-src/blob/34bd34750187bae02a0086402c78c031c59dd8b3/sys/kern/kern_event.c) (`kqueue_kevent`, `kqueue_register`, `kqueue_scan`, the timer and user filters) over the event queue; nothing is measured on a console. The descriptor is a host null device registered as the kqueue: `close` and `sceKernelClose` close the queue and `dup`/`dup2` of it throw, but `read`, `write` and `fstat` act on the null device instead of failing with `EOPNOTSUPP` or reporting `S_IFIFO`. Read and write events are supported on guest sockets only, plus `EVFILT_READ` on the read end and `EVFILT_WRITE` on the write end of a pipe (see the kevent on a pipe entry) and the timer, user and [vnode](../../../core/libs/prx/libkernel/Equeue/src/FileEvent.cpp) filters and a read event on the host stream on descriptor 0, which never triggers, as `select` never reports descriptor 0 readable; PPSA34015 registers descriptor 0 for reading in a new kqueue at start-up and does not check the result; other filters, `EV_CLEAR` on a socket or a pipe off Linux (see the kevent EV_CLEAR entry), a `NOTE_NSECONDS` period above one second with a fraction (11.0 converts the fraction with the microsecond macro) and any change to an existing timer (11.0 keeps the armed expiry and applies the new period after it, 11.4's `filt_timertouch` restarts it) throw. `EV_DISABLE`, `EV_DISPATCH` and `EV_FORCEONESHOT` follow kern_event.c: a disabled knote stays registered, stops polling or firing and is not reported, a trigger raised while it is disabled is kept and reported after `EV_ENABLE`, where `knote` would drop it, `EV_DISPATCH` reports once and then disables the knote, and `EV_FORCEONESHOT` is ignored on a new knote while on an existing one it sets `EV_ONESHOT` and fires the knote; a negative `nevents` fails with `EINVAL`. A new user event keeps its `fflags` unmasked and is not triggered by `NOTE_TRIGGER`, because `filt_userattach` reads the already cleared `kn_fflags`. The timer limit `kq_calloutmax` (`ENOMEM`) is not enforced, and closing a kqueue wakes a thread waiting on it in `kevent` with `EBADF`, where `kqueue_drain` makes the close wait for that thread. A level-triggered event is reported once per call and moved behind the other events, as in `kqueue_scan`; the `sceKernel*Equeue` calls keep their own collection loop. `NOTE_LOWAT` on a socket sets the low-water mark as FreeBSD 11's [`filt_soread` and `filt_sowrite`](https://github.com/freebsd/freebsd-src/blob/release/11.0.0/sys/kern/uipc_socket.c) do, and a later change of the knote replaces it as `kqueue_register` replaces `kn_sdata`: a read event fires once the buffered bytes reach the mark, or at end of file or on a listening socket. A read mark above 1 is taken on stream sockets only: on a datagram socket Linux's `FIONREAD` reports the next datagram where FreeBSD's `sbavail` counts every queued byte, so it throws. A write event fires once the send buffer's free space reaches the mark, as `filt_sowrite` does with `sbspace`; the host reports that space as `SO_SNDBUF` less `SIOCOUTQ` on Linux and as `SO_SNDBUF` alone elsewhere, so off Linux it does not shrink while data is queued. Marks below 1, which make FreeBSD's read filter always fire, throw
 - [EVFILT_VNODE, sceKernelAddFileEvent and sceKernelDeleteFileEvent](../../../core/libs/prx/libkernel/Equeue/src/FileEvent.cpp) (libkernel) watch the host file or directory behind the descriptor and turn host changes into FreeBSD 11 vnode hints, which [`filt_vfsvnode`](https://github.com/freebsd/freebsd-src/blob/34bd34750187bae02a0086402c78c031c59dd8b3/sys/kern/vfs_subr.c#L5032) adds whole when they share a bit with the requested `fflags`: a write that grows a file is `NOTE_WRITE | NOTE_EXTEND`, another write `NOTE_WRITE`, a shrinking truncate or an attribute change `NOTE_ATTRIB`, a higher link count `NOTE_LINK`, a lower one or a pending delete `NOTE_DELETE`, a new path `NOTE_RENAME`; in a directory, an entry created, removed or renamed is `NOTE_WRITE`, with `NOTE_LINK` for a subdirectory, and on Linux an entry moved from or to another directory adds `NOTE_EXTEND`; inotify queues the two halves of a rename one after the other, so a watcher woken by the first can read it alone, and a half read without the other is classified only after a `getdents64` on the watched directory, which takes the directory lock that `rename` holds until both halves are queued ([`do_renameat2`](https://github.com/torvalds/linux/blob/v6.8/fs/namei.c#L4975) around `vfs_rename` and its `fsnotify_move`, [`iterate_dir`](https://github.com/torvalds/linux/blob/v6.8/fs/readdir.c#L103)), an ordering taken from the kernel source, not from inotify(7), which promises none; a directory registered with `NOTE_EXTEND` on Windows is accepted, as `vfs_kqfilter` does not check it, but never receives it, because the directory handle cannot tell a move between directories from a removal plus an addition. FreeBSD's hints depend on the filesystem and the console's filesystems are not known: a growing `ftruncate` reports `NOTE_WRITE | NOTE_EXTEND` where `vop_setattr_post` reports `NOTE_ATTRIB`, and on Windows `utimes` on a file reports `NOTE_WRITE`. The three Windows notifications are told apart in the error, which names the `parent`, `entries` or `subdirectories` watch, the path it was given and `GetLastError`, but a `FindFirstChangeNotification` that fails still throws rather than returning an error to the guest: the EVFILT_VNODE attach path is [`vfs_kqfilter`](https://github.com/freebsd/freebsd-src/blob/34bd34750187bae02a0086402c78c031c59dd8b3/sys/kern/vfs_subr.c#L4936), which returns `EINVAL`, `ENOMEM` or 0 after `fget` has turned a descriptor that is not open into `EBADF`, and consults no directory permission, so no FreeBSD errno stands for a host that refuses to notify, and what the console does on such a host is not measured. Only the `parent` watch can fail this way in practice: `entries` and `subdirectories` need the `FILE_LIST_DIRECTORY` that opening the watched directory already took. The parent handle produces no hint of its own, as `RunHostWatch` only calls `FindNextChangeNotification` on it, so a later change may leave the watch on its 50 ms poll instead of throwing; that is not done here. Windows uses `FindFirstChangeNotification` on the parent directory (and on a watched directory itself) and compares `GetFileInformationByHandleEx` snapshots at least every 50 ms, because a write through an open handle is reported only when the handle is flushed or closed, so file hints can arrive 50 ms late, and a file deleted while watched, which the POSIX delete of Windows 10 1903 moves into `$Extend\$Deleted` until its last handle closes, reports `NOTE_DELETE` without `NOTE_RENAME` and keeps its handle without a parent directory watch, as that directory cannot be watched; Linux uses inotify on the file and is not run. `NOTE_REVOKE`, `NOTE_OPEN`, `NOTE_CLOSE`, `NOTE_CLOSE_WRITE`, `NOTE_READ`, unknown bits and descriptors other than files, directories, pipes and sockets throw. `sceKernelAddFileEvent(eq, fd, watch, udata)` and `sceKernelDeleteFileEvent(eq, fd)` take the [OpenOrbis](https://github.com/OpenOrbis/OpenOrbis-PS4-Toolchain/blob/0a1aaf9d/include/orbis/libkernel.h#L35) and [SharpProspero](https://github.com/SvenGDK/SharpProspero/blob/5dfdb4b0/src/SharpProspero/Interop/Kernel/KernelEqueue.cs#L164) signatures; `watch` is taken as the `fflags` and the event is added with `EV_CLEAR`, neither confirmed on a title (feat(libkernel): implement kevent EV_DISABLE, EV_DISPATCH and EV_FORCEONESHOT) (fix(libkernel): follow kqueue_register for EV_FORCEONESHOT and modify flags)