Repository navigation
libc rename reports wrong errno values on Windows and fails to replace an empty directory #2403
Description
On Windows, the guest rename (rename_nid_postfix in core/libs/prx/libc/src/Filesystem.cpp) passes everything to std::filesystem::rename and maps the resulting std::error_code. MinGW's libstdc++ reports several cases differently from FreeBSD, so the guest sees wrong errno values, and one valid rename fails.
Probe calling rename_nid_postfix directly. Windows 11, WinLibs GCC 15.2.0 r7, current main:
| case | Windows result | FreeBSD |
|---|---|---|
| file → existing file | 0 | 0 |
| dir → new name | 0 | 0 |
| dir → existing empty dir | -1, errno 13 (EACCES) | 0 (replaces it) |
| dir → non-empty dir | -1, errno 13 | -1, ENOTEMPTY (66) |
| file → existing dir | -1, errno 21 | -1, EISDIR (21) |
| dir → existing file | -1, errno 20 | -1, ENOTDIR (20) |
| missing source | -1, errno 5 (EIO) | -1, ENOENT (2) |
| dir into its own subdirectory | -1, errno 5 | -1, EINVAL (22) |
| destination parent missing | -1, errno 5 | -1, ENOENT (2) |
On Linux, std::filesystem::rename is rename(2), so these already match.
The libkernel side (sceKernelRename) has the same problem; #2260 fixes it there. It checks source and destination with symlink_status, maps not-found and not-a-directory before calling rename, and removes an empty target directory first on Windows. The same approach would work here.
I haven't opened a PR, because #1192 moves rename_nid_postfix into a new Rename.cpp, and a fix in Filesystem.cpp would conflict with it. This is the data for whoever lands after it. I'm happy to do the follow-up once #1192 is merged or closed.
AI-assisted: yes (the probe was written and run with Claude Code).
Activity
fix(libc): report POSIX rename errors like the console
fix(libc): report POSIX rename errors like the console
MarcZX10 commentedon Oct 9, 2026
Validation update on current upstream
mainI rebased the prepared two-file candidate onto
cf8b129e2771d26eaffc300248c46145063f1794. The merge since the previous update (#2670) does not touch either changed file.guest_filesystem_tests(Linux x86_64, Release)cf8b129eguest_filesystem(--repeat until-fail:3)cf8b129emaingit diff --cached --checkThe Windows evidence from Achraf is a native Windows 11 probe on the original implementation branch (
480dd351, based on6cc33583): all nine reported rename behaviors matched FreeBSD, including replacing an empty directory. I confirmedRename.cppis unchanged between that branch and this candidate. The newly expandedguest_filesystemtest binary has not been rerun on Windows atcf8b129e; its added assertions have been run on Linux as shown above.The candidate remains local and staged. I have not opened a replacement PR while intake requires 20 merged PRs and my count is 15. AI-assisted change.
reyani-dev commentedon Oct 10, 2026
Confirmed on Windows, @MarcZX100: your branch (480dd35) merged onto current
main6cc3358, Windows 11, WinLibs GCC 15.2.0 r7, Release.The probe from this report (
rename_nid_postfixcalled directly) now matches FreeBSD in all nine cases, including the five that differed:The other four (file → existing file, dir → new name, file → existing dir 21, dir → existing file 20) are unchanged and correct.
guest_filesystemwith your regression passes 3/3 (--repeat until-fail:3), andctest -E "^agc_|video_out"passes 372/372.MarcZX10 commentedon Oct 10, 2026
Follow-up on current
mainThe ready change is now refreshed onto upstream
mainba2a0a31f15e9df224b2df904341fe00c0c26858incodex/fix-libc-rename-win-errors-2403-latest, commit7a41ca5e. It applies cleanly after #1192 moved the function toRename.cpp.Current Linux x86_64 / GCC 12.2 / Release validation:
guest_filesystem_testsbuilt.guest_filesystempassed (1/1).mainandgit diff --checkpassed.The test covers the reported errno cases and replacement of an empty destination directory. The native Windows table remains Achraf's independent run on the same implementation at
480dd351(based on earliermain); I have not run Windows on this refreshed commit. The repo still automatically closes PRs below 20 merged PRs; my live count is 15. The branch is pushed for a collaborator with intake access to carry, not an open PR. AI-assisted: yes.reyani-dev commentedon Oct 10, 2026
Windows rerun on the refreshed commit, @MarcZX100:
7a41ca5emerged ontomainba2a0a3, Windows 11, WinLibs GCC 15.2.0 r7, Release.main: empty-dir replace gives 0, non-empty dir 66, missing source 2, move into own subdirectory 22, missing destination parent 2.guest_filesystempasses 3/3 (--repeat until-fail:3).ctest -E "^agc_|video_out"passes 380/380.fix(libc): report POSIX rename errors like the console
oneandonlydean commentedon Oct 10, 2026
@MarcZX100 the intake check needs 10 merged PRs now, so 7a41ca5 can go up as a PR.