Repository navigation
libkernel/libc: rmdir, sceKernelRmdir and remove delete the directory for "dir/." instead of failing with EINVAL; fix branch ready #3116
Description
Guest paths are normalized lexically (Resolve in libc/src/General.cpp), so dir/. reaches the host as dir/ and the directory is removed. FreeBSD's kern_rmdirat rejects a final . with EINVAL ("No rmdir "." please."), and POSIX requires rmdir to fail with EINVAL when the last component is dot. remove calls rmdir for a directory (FreeBSD lib/libc/stdio/remove.c).
On main (Linux), each of these returns 0 and the directory is gone: rmdir("probe_dir/."), sceKernelRmdir("probe_dir/."), remove("probe_dir/.").
Fix on my fork: milescoviello/AnyPS5:fix/rmdir-dot-component (one commit, 9aaf6240, rebased onto main ec77a61 after #3395): sceKernelRmdir (and therefore rmdir) and remove return EINVAL (0x80020016) and keep the directory. A missing path or a file before the dot keeps its current error. guest_filesystem gains checks that fail on main and pass with the fix. Linux only; Windows was not built or tested. The check runs on the guest path before any host call, so the code is portable.
Setup: Gentoo, Linux 7.2.8, GCC 15.3.0, Release with CI's configure flags, main 4d27834.
Tested: the new checks fail on main and pass with the fix. Full ctest with the fix on Linux (RTX 5090 Laptop GPU): no new failures; the only failures are the known agc_driver_dcc_keys_first_write (#2767) and two guest_video_out_* tests that time out with the displays powered off. tools/check_conventions.py passes. All five fixes I am filing today also pass their tests merged together on the current main.
Not opened as a PR because intake is paused for new contributors; a maintainer is welcome to carry the commit.
AI-assisted: yes
Activity
[-]libkernel/libc: rmdir, sceKernelRmdir and remove delete the directory for "dir/." instead of failing with EINVAL[/-][+]libkernel/libc: rmdir, sceKernelRmdir and remove delete the directory for "dir/." instead of failing with EINVAL; fix branch ready[/+]Silverash-i8 commentedon Oct 11, 2026
Windows run, since the issue says it wasn't built or tested there: Windows 11, WinLibs GCC 15.2.0 r7, Release,
fix/rmdir-dot-component(a43ccf41) merged ontomain06da117 together with #3117, #3181 and #3169 (#3169 is needed here becauseguest_filesystemotherwise stops at the symlink checks on a machine without the symlink privilege).RemoveDirectoryWcollapses the dot itself, so a plainrmdirofdir\.removesdir(checked directly with Python'sos.rmdir).guest_filesystemsit outside the non-Windows block and pass:rmdir,sceKernelRmdirandremoveofdir/.anddir/./return EINVAL (0x80020016 forsceKernelRmdir) and the directory stays. The backslash paths the test builds on Windows work sinceFinalPathComponentIsDot_nid_no_patchsplits on both separators.ctest -E "^agc_") pass with the four branches merged.AI-assisted (Claude Code); I ran this on my machine.
milescoviello commentedon Oct 11, 2026
Thanks @Silverash-i8 for the Windows runs on this and #3117, and for checking
RemoveDirectoryWdirectly. Good to know the dot case is fixed on Windows too. The\\?\idea for..on Windows is useful as well.milescoviello commentedon Oct 11, 2026
Rebased onto
mainec77a61: #3395 changed the same lines ofsceKernelRmdir. The dot check now runs right after its ENOTDIR/ENOENT block, which keeps FreeBSD's order (lookup errors first, then EINVAL, then ENOTEMPTY). New commit9aaf6240. The 33 file, path and save-data tests pass, including #3395's new checks, andguest_filesystemfails with the check reverted.AI-assisted: yes
Silverash-i8 commentedon Oct 11, 2026
Windows recheck of the rebased
9aaf6240onmainec77a61 (Windows 11, WinLibs GCC 15.2.0 r7, Release):guest_filesystempasses, now all the way through since #3169 is in, and the non-GPU suite (ctest -E "^agc_") is 453/454. The one failure iswindows_exceptions, which fails the same way on plainmain(#3411), so nothing from this branch.AI-assisted (Claude Code); I ran this on my machine.