Re: [PATCH v3 -next 00/18] landlock: Add READ_METADATA and WRITE_METADATA access rights
From: Cai Xinchen
Date: Thu Nov 12 2026 - 23:31:33 EST
You are right, and thanks for catching it.
I have been working on these modifications over the past two weeks and have deployed sashiko locally. However, my current changes are still being flagged for some bugs by my local sashiko, and I am still in the process of fixing them. I expect to release v4 in a few days. Regarding these issues, my current fixes are as follows: - AT_EMPTY_PATH: v4 will use CLASS(fd_raw, f)(dfd), the same lookup path_init() does for an empty path, so O_PATH descriptors work again. do_utimes_fd() (futimens(2)) intentionally keeps CLASS(fd, f), matching fchmod(2)/fchown(2); AT_EMPTY_PATH returns -EINVAL there anyway. The same patch rejects a NULL pathname with -EFAULT before CLASS(filename_maybe_null) can turn it into an empty path relative to AT_FDCWD, and adds the missing audit_file() on the do_utimes_path() fd branch. - __nullable (patches 5-9, 11-13): v4 will extend kernel/bpf/bpf_lsm_proto.c (the existing strong-override file used for mmap_file) with strong definitions for all 11 hooks that gained a file argument, each declared file__nullable, so the verifier sees PTR_MAYBE_NULL and a program must check before dereferencing. - ntfs named stream setattr (patch 1): v3 forwarded with __notify_change(idmap, base_dentry, ...). v4 will use notify_change_target(idmap, attr->ia_path, base_dentry, ...), so the hook is checked against the path the syscall used while the attributes still land on the base inode. - overlayfs fileattr (patch 8): v3 removed the ovl_security_fileattr() workaround but then called the unchecked __vfs_fileattr_get()/_set() on the real dentry, which bypasses the hooks. v4 will call the checked vfs_fileattr_get()/_set() with the real upper/lower path, so the hooks also run for the underlying layer; ecryptfs is switched the same way.
Best regards.
Cai
On 11/11/2026 12:52 AM, Christian Brauner wrote:
This regresses AT_EMPTY_PATH. fchmodat2(), fchownat() and utimensat()
with an empty path now go through CLASS(fd) which is fdget() and that
refuses O_PATH file descriptors. IOW, this is very very broken.
Please also go through Sashiko's review of v3 (cf. [1]). The new struct
file arguments need __nullable or bpf programs can dereference NULL,
and the ntfs3 stream setattr and overlayfs fileattr paths that bypass
the hooks need treatment.