fix(libkernel): change O_APPEND on a Windows file through fcntl

core/libs/prx/libkernel/File/src/DescriptorControl.cpp

@@ -22,8 +22,10 @@
 #include <string>
 
 #ifdef _WIN32
+#include <fcntl.h>
 #include <io.h>
 #include <windows.h>
+#include "prx/libc/include/General.hpp"
 #include "prx/libc/include/WindowsDescriptor.hpp"
 #else
 #include <fcntl.h>
@@ -180,8 +182,33 @@ Entry InitialEntry(int fd) {
     return {false, std::make_shared<int>(status)};
 }
 
+void SetHostAppend(int fd, bool append, const char* function) {
+    const int translation = ::_setmode(fd, _O_BINARY);
+    if (translation == -1 || ::_setmode(fd, translation) == -1)
+        throw std::runtime_error(std::string(function) + ": cannot read the translation mode of descriptor " + std::to_string(fd) + ", errno=" + std::to_string(errno));
+    HANDLE copy = nullptr;
+    if (!::DuplicateHandle(::GetCurrentProcess(), WindowsDescriptorHandle_nid_no_patch(fd), ::GetCurrentProcess(), &copy, 0, TRUE, DUPLICATE_SAME_ACCESS))
+        throw std::runtime_error(std::string(function) + ": cannot duplicate the handle of descriptor " + std::to_string(fd) + ", error=" + std::to_string(::GetLastError()));
+    const int replacement = ::_open_osfhandle(reinterpret_cast<std::intptr_t>(copy), (append ? _O_APPEND : 0) | translation);
+    if (replacement < 0) {
+        ::CloseHandle(copy);
+        throw std::runtime_error(std::string(function) + ": cannot open a replacement for descriptor " + std::to_string(fd) + ", errno=" + std::to_string(errno));
+    }
+    auto* const standard = HostStandardHandle_nid_no_patch(fd);
+    const int moved = ::_dup2(replacement, fd);
+    const int error = errno;
+    ::_close(replacement);
+    KeepStandardHandle_nid_no_patch(fd, standard);
+    if (moved != 0)
+        throw std::runtime_error(std::string(function) + ": cannot replace descriptor " + std::to_string(fd) + ", errno=" + std::to_string(error));
+}
+
 void ApplyHostStatus(int fd, Kind kind, int changed, int next, const char* function) {
-    if ((changed & Append) && kind == Kind::File) Unsupported(function, "changing O_APPEND", kind, fd);
+    if ((changed & Append) && kind == Kind::File) {
+        const auto status = g_entries.at(fd).status;
+        for (const auto& [descriptor, entry] : g_entries)
+            if (entry.status == status && DescriptorIsOpen_nid_no_patch(descriptor)) SetHostAppend(descriptor, (next & Append) != 0, function);
+    }
     if ((changed & Nonblocking) && kind == Kind::Device) Unsupported(function, "changing O_NONBLOCK", kind, fd);
     if ((changed & Nonblocking) && kind == Kind::Pipe) SetPipeNonblocking(fd, (next & Nonblocking) != 0, function);
 }

core/libs/tests/GuestKernelFcntl.cpp

@@ -199,6 +199,7 @@ static void SocketsKeepTheirPath() {
 }
 
 static void FileFlagsRoundTrip() {
+    std::remove("kernel_fcntl_probe.bin");
     const int file = sceKernelOpen("kernel_fcntl_probe.bin", CreateReadWrite, 0644);
     Require(file >= 0);
     Require(sceKernelFcntl(file, GetFlags) == ReadWrite);
@@ -275,12 +276,26 @@ static void FileFlagsRoundTrip() {
     Require(fcntl_nid_postfix(appending, GetFlags) == (WriteOnly | Append));
     Require(fcntl_nid_postfix(appending, SetFlags, Append | NonBlocking) == 0);
     Require(fcntl_nid_postfix(appending, GetFlags) == (WriteOnly | Append | NonBlocking));
-#ifdef _WIN32
-    Require(ThrowsWith([appending] { fcntl_nid_postfix(appending, SetFlags, 0); }, "fcntl: changing O_APPEND on a file is not implemented"));
-#else
+    Require(write_nid_postfix(appending, "ab", 2) == 2);
     Require(fcntl_nid_postfix(appending, SetFlags, 0) == 0);
     Require(fcntl_nid_postfix(appending, GetFlags) == WriteOnly);
-#endif
+    Require(sceKernelLseek(appending, 0, FromStart) == 0);
+    Require(write_nid_postfix(appending, "X", 1) == 1);
+    const int sharing = fcntl_nid_postfix(appending, DuplicateCommand, 0);
+    Require(sharing >= 0);
+    Require(fcntl_nid_postfix(sharing, SetFlags, Append) == 0);
+    Require(fcntl_nid_postfix(appending, GetFlags) == (WriteOnly | Append));
+    Require(sceKernelLseek(appending, 0, FromStart) == 0);
+    Require(write_nid_postfix(appending, "c", 1) == 1);
+    Require(fcntl_nid_postfix(sharing, SetFlags, 0) == 0);
+    Require(sceKernelLseek(appending, 0, FromStart) == 0);
+    Require(write_nid_postfix(appending, "Y", 1) == 1);
+    Require(close_nid_postfix(sharing) == 0);
+    const int reading = sceKernelOpen("kernel_fcntl_probe.bin", ReadOnly, 0);
+    Require(reading >= 0);
+    char written[16]{};
+    Require(sceKernelRead(reading, written, sizeof(written)) == 8 && std::memcmp(written, "Ybcdeabc", 8) == 0);
+    Require(sceKernelClose(reading) == 0);
     Require(ThrowsWith([appending] { fcntl_nid_postfix(appending, SetFlags, Append | 0x80); }, "fcntl: changing O_FSYNC on a file is not implemented"));
     Require(sceKernelClose(appending) == 0);
     std::remove("kernel_fcntl_probe.bin");

docs/dev/debt/libkernel.md

@@ -230,7 +230,7 @@
 - [kill, killpg](../../../core/libs/prx/libkernel/System/src/Signals.cpp) (libkernel) follow FreeBSD's `kern_kill` and `killpg1` for the only guest process: `EINVAL` for signals outside 0 to 128 and for negative `killpg` process groups, signal 0 only checks that the target exists, and its own pid, 0 and minus its own pid (the guest is taken to lead its own process group) signal it through `raise`, so they share `raise`'s limits above. Any other pid, and -1, which skips the caller, report `ESRCH`, so `EPERM` is never reported
 - [sceKernelFcntl](../../../core/libs/prx/libkernel/File/src/Open.cpp) (libkernel) runs the POSIX `fcntl` below and returns its guest errno as `SCE_KERNEL_ERROR_*`, the form the other `sceKernel` file exports use. The return values and error codes are unverified on a console
 - [Creating through a trailing slash](../../../core/libs/prx/libkernel/File/src/Open.cpp) (libkernel) - `open` and `sceKernelOpen` with `O_CREAT` on a missing name followed by a slash (`new/`) fail with `EINVAL` on Windows and with the host's `EISDIR` on Linux, where FreeBSD 11.4's `lookup()` (sys/kern/vfs_lookup.c) fails with `ENOENT` because only a directory may be created through a trailing slash
-- [fcntl](../../../core/libs/prx/libkernel/File/src/DescriptorControl.cpp) (libkernel) follows FreeBSD 11.4's [`kern_fcntl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_descrip.c#L479-L799) and [`kern_dup`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_descrip.c#L812) on file, directory, pipe and character device descriptors; a socket keeps the libSceNet path (`F_DUPFD`, `F_DUP2FD`, `F_DUPFD_CLOEXEC`, `F_DUP2FD_CLOEXEC`, `F_GETFL`, `F_SETFL` with `O_NONBLOCK`, `F_GETFD` and `F_SETFD`, `EINVAL` for every other command, see the next entry). On Linux and macOS a host socket that is not a guest socket, such as a standard descriptor a service manager connected to its journal, takes the flag handling of a pipe end and `FIONREAD` asks the host for its queued bytes (for a datagram socket Linux counts only the next datagram, where FreeBSD's [`soo_ioctl`](https://github.com/freebsd/freebsd-src/blob/release/11.4.0/sys/kern/sys_socket.c#L202-L205) counts every buffered byte); its record locks and read-ahead hints fail with `EBADF`, as FreeBSD fails them on a descriptor without a vnode, and `O_ASYNC` on it throws. The close-on-exec flag and the status flags live in a guest table, the status flags shared by `dup`, `dup2` and `F_DUPFD` as in an open file description; a descriptor the table has not seen, such as 0 to 2, starts from the host's flags, since the console's are not known. `F_DUPFD` and `F_DUPFD_CLOEXEC` return the lowest free host descriptor at or above the start, so the bound is the host table (`EMFILE` on Windows, `EINVAL` past `RLIMIT_NOFILE` on Linux), not the console's unknown `RLIMIT_NOFILE` and `kern.maxfilesperproc`. `O_NONBLOCK` switches a host pipe (`PIPE_NOWAIT` on Windows, where an empty read and a full write then report `EAGAIN`) and is only recorded on files, directories and the random device, as their reads never block; on a Windows console or other character device it throws. Changing `O_APPEND` on a Windows file, changing `O_FSYNC` on a file, `O_ASYNC` on a pipe or character device (`SIGIO` is not delivered), `F_GETOWN`, `F_SETOWN` and `F_SETLK_REMOTE` throw; `O_DIRECT`, `F_RDAHEAD` and `F_READAHEAD` are recorded as the hints they are. Record locks (`F_GETLK`, `F_SETLK`, `F_SETLKW` and their forms without `l_sysid`) use the host's POSIX locks on Linux and macOS. `F_GETLK` with `F_UNLCK` asks the host with `F_RDLCK`: as in [`lf_blocks`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_lockf.c#L866) only another owner's overlapping write lock conflicts with either request. On Windows they follow [`kern_fcntl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_descrip.c#L607-L752) and [`lf_advlockasync`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_lockf.c#L433-L487) for the one guest process, which owns every lock: `EOVERFLOW` for an `SEEK_CUR` start past `OFF_MAX`, then `EBADF` for `F_RDLCK` without read access or `F_WRLCK` without write access and `EINVAL` for another type, then `EINVAL` and `EOVERFLOW` for the range; as [`lf_blocks`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_lockf.c#L866) never lets locks of one owner conflict, whichever descriptor or thread asks, `F_SETLK` and `F_SETLKW` then succeed without waiting and `F_GETLK` reports `F_UNLCK`, so no lock list is kept and merging, splitting and the release on close have nothing to change. Nothing is held on the host, since Windows byte-range locks are mandatory and would fail the guest's own reads and writes through other descriptors, so other host processes neither see these locks nor are seen; a lock from the end of a directory throws. The return values and error codes are unverified on a console
+- [fcntl](../../../core/libs/prx/libkernel/File/src/DescriptorControl.cpp) (libkernel) follows FreeBSD 11.4's [`kern_fcntl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_descrip.c#L479-L799) and [`kern_dup`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_descrip.c#L812) on file, directory, pipe and character device descriptors; a socket keeps the libSceNet path (`F_DUPFD`, `F_DUP2FD`, `F_DUPFD_CLOEXEC`, `F_DUP2FD_CLOEXEC`, `F_GETFL`, `F_SETFL` with `O_NONBLOCK`, `F_GETFD` and `F_SETFD`, `EINVAL` for every other command, see the next entry). On Linux and macOS a host socket that is not a guest socket, such as a standard descriptor a service manager connected to its journal, takes the flag handling of a pipe end and `FIONREAD` asks the host for its queued bytes (for a datagram socket Linux counts only the next datagram, where FreeBSD's [`soo_ioctl`](https://github.com/freebsd/freebsd-src/blob/release/11.4.0/sys/kern/sys_socket.c#L202-L205) counts every buffered byte); its record locks and read-ahead hints fail with `EBADF`, as FreeBSD fails them on a descriptor without a vnode, and `O_ASYNC` on it throws. The close-on-exec flag and the status flags live in a guest table, the status flags shared by `dup`, `dup2` and `F_DUPFD` as in an open file description; a descriptor the table has not seen, such as 0 to 2, starts from the host's flags, since the console's are not known. `F_DUPFD` and `F_DUPFD_CLOEXEC` return the lowest free host descriptor at or above the start, so the bound is the host table (`EMFILE` on Windows, `EINVAL` past `RLIMIT_NOFILE` on Linux), not the console's unknown `RLIMIT_NOFILE` and `kern.maxfilesperproc`. `O_NONBLOCK` switches a host pipe (`PIPE_NOWAIT` on Windows, where an empty read and a full write then report `EAGAIN`) and is only recorded on files, directories and the random device, as their reads never block; on a Windows console or other character device it throws. Changing `O_FSYNC` on a file, `O_ASYNC` on a pipe or character device (`SIGIO` is not delivered), `F_GETOWN`, `F_SETOWN` and `F_SETLK_REMOTE` throw; `O_DIRECT`, `F_RDAHEAD` and `F_READAHEAD` are recorded as the hints they are. Record locks (`F_GETLK`, `F_SETLK`, `F_SETLKW` and their forms without `l_sysid`) use the host's POSIX locks on Linux and macOS. `F_GETLK` with `F_UNLCK` asks the host with `F_RDLCK`: as in [`lf_blocks`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_lockf.c#L866) only another owner's overlapping write lock conflicts with either request. On Windows they follow [`kern_fcntl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_descrip.c#L607-L752) and [`lf_advlockasync`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_lockf.c#L433-L487) for the one guest process, which owns every lock: `EOVERFLOW` for an `SEEK_CUR` start past `OFF_MAX`, then `EBADF` for `F_RDLCK` without read access or `F_WRLCK` without write access and `EINVAL` for another type, then `EINVAL` and `EOVERFLOW` for the range; as [`lf_blocks`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/kern_lockf.c#L866) never lets locks of one owner conflict, whichever descriptor or thread asks, `F_SETLK` and `F_SETLKW` then succeed without waiting and `F_GETLK` reports `F_UNLCK`, so no lock list is kept and merging, splitting and the release on close have nothing to change. Nothing is held on the host, since Windows byte-range locks are mandatory and would fail the guest's own reads and writes through other descriptors, so other host processes neither see these locks nor are seen; a lock from the end of a directory throws. The return values and error codes are unverified on a console
 - [fcntl on a guest socket](../../../core/libs/prx/libkernel/Socket/src/Udp.cpp) (libkernel) answers `F_DUPFD`, `F_DUP2FD`, `F_DUPFD_CLOEXEC`, `F_DUP2FD_CLOEXEC`, `F_GETFD`, `F_SETFD`, `F_GETFL` and `F_SETFL` and fails every other command with `EINVAL`; FreeBSD 11.4's `kern_fcntl` also handles `F_GETOWN` and `F_SETOWN` on a socket. `FD_CLOEXEC` is only recorded (there is no `exec`): `dup`, `dup2`, `F_DUPFD`, `F_DUP2FD` onto another descriptor and an accepted socket start without it, `F_DUP2FD` onto the same descriptor keeps it, and `SOCK_CLOEXEC`, `F_DUPFD_CLOEXEC`, `F_DUP2FD_CLOEXEC` and `F_SETFD` set it. None of this was measured on a console.
 - [ioctl, _ioctl](../../../core/libs/prx/libkernel/Socket/src/Udp.cpp) (libkernel) handle only `FIONBIO` and `FIONREAD` on a socket, and any other request there fails with `ENOTTY`. On the other descriptors they follow FreeBSD's [`kern_ioctl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/sys_generic.c#L758), [`vn_ioctl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/vfs_vnops.c#L1488) and [`pipe_ioctl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/sys_pipe.c#L1346): `FIONBIO` sets `O_NONBLOCK` as `fcntl` does, `FIOCLEX` and `FIONCLEX` set and clear the close-on-exec flag `F_GETFD` reports, and `FIONREAD` reports the size past the offset of a file, the queued bytes of a pipe's read end, 0 on its write end and `ENOTTY` on the random device, while it throws on a directory or another character device. Every termios request (group `t`, `TIOCGWINSZ` included) fails with `ENOTTY`, since those descriptors are taken as not a terminal, as in `tcgetattr`, and any other well-formed request throws. An input request with length 0 is not rejected as malformed: [`sys_ioctl`](https://github.com/freebsd/freebsd-src/blob/8c639945a44588f6cfdddfb057bb5aad88735da0/sys/kern/sys_generic.c#L703-L718) rejects it only in a kernel built without `COMPAT_FREEBSD4`, `COMPAT_FREEBSD5` and `COMPAT_43`, and which of these the PS5 kernel has is unknown. Stremio Plus (PPSA74126) imports `_ioctl` for its statically linked `tc*` functions, which send `TIOCGETA`, `TIOCSETA`, `TIOCSPGRP` and the other termios requests
 - [ioctl, _ioctl](../../../core/libs/prx/libkernel/Socket/src/Udp.cpp) (libkernel) do not pass a `sceNetSocket` socket to libSceNet as `fcntl` does. On Linux its number duplicates the host socket, so `FIONBIO` switches the host `O_NONBLOCK` but not the library's `SO_NBIO` state; on Windows it is a `NUL` placeholder, so `FIONREAD` and enabling `FIONBIO` throw as on a character device. `sceNetIoctl` handles both requests on such a socket.