Re: [PATCH] erofs-utils: lib: follow symlinks on intermediate tar path components
Gao Xiang <[email protected]> Sun, 2 Aug 2026 21:00:56 +0800
| Newsgroups | org.ozlabs.lists.linux-erofs |
|---|---|
| Message-ID | <[email protected]> |
Hi Mikhail, On Sat, Jul 25, 2026 at 03:09:20PM +0000, [email protected] wrote: > From: Mikhail Malyshev <[email protected]> > > mkfs.erofs --tar fails to convert layers that store files underneath a > symlinked directory, a very common "usr-merge" layout in RHEL/UBI/Fedora > based OCI images. For example a base layer contains > > lib64 -> usr/lib64 (symlink) > usr/lib64/ (directory) > lib64/libc.so.6 (regular file, path goes THROUGH the symlink) > > erofs_rebuild_get_dentry() walks the path component by component and, when > it hits the "lib64" component, finds an existing dentry that is a symlink > rather than a directory. Since there are still components left to resolve > it bails out with -ENOTDIR. In practice this surfaces as a bogus ~2 TiB > image and "Could not format the device", making the whole layer > unconvertible even though nothing is wrong with the tar. > > A real filesystem (and containerd's walking differ, which applies the same > layers via fs.RootPath) resolves such a path by following the symlink to > its target directory. Do the same during tar rebuild: when an intermediate > path component resolves to a symlink, follow it to the target directory > (relative targets from the containing directory, absolute targets from the > tree root), chasing chained symlinks up to a bounded depth to avoid loops. > > Signed-off-by: Mikhail Malyshev <[email protected]> Thanks for the patch and sorry for late reply. Yes, I assume containerd will resolve the directories like this, but as a quick reference, could you give some pointer how `tar` utility works and `posix` documents this behavior? I plan to release erofs-utils 1.9.3 without this patch soon, and it can be addressed in erofs-utils 1.9.4 version. Thanks, Gao Xiang