syzbot
KCSAN: data-race in generic_permission / kernfs_iop_permission
Status:
moderation: reported on 2026/11/20 00:03
Subsystems:
kernfs
Labels:
prio:low
race:benign
[Documentation on labels]
Reported-by: syzbot+5a5937e5f82f0e8fa8ec@syzkaller.appspotmail.com
First crash: 2d07h, last: 4h35m
✨ AI-generated root cause summary:
kernfs_iop_permission() in fs/kernfs/inode.c acquires only a shared read lock (down_read(&root->kernfs_iattr_rwsem)) before calling kernfs_refresh_inode(kn, inode) and generic_permission(). Because down_read() permits concurrent execution on the same inode, one task's plain 2-byte write to inode->i_mode in kernfs_refresh_inode() races with another task's plain 2-byte read of inode->i_mode in acl_permission_check() (via generic_permission()), as well as with concurrent kernfs_refresh_inode() calls in kernfs_iop_permission() and kernfs_iop_getattr().
▶
▼
✨ Full AI root cause analysis
Root Cause Analysis
The KCSAN data race occurs because kernfs_iop_permission() in fs/kernfs/inode.c mutates fields of the shared VFS struct inode while holding only a shared reader lock (down_read(&root->kernfs_iattr_rwsem)), allowing concurrent VFS lookups on the same kernfs node to read and write inode->i_mode without synchronization.
Technical Mechanism
-
Shared locking in kernfs_iop_permission() (fs/kernfs/inode.c:273-292):
- When VFS path traversal (
link_path_walk() -> may_lookup() -> inode_permission() -> do_inode_permission()) checks permissions on a kernfs inode (in non-RCU mode, i.e., without MAY_NOT_BLOCK), it invokes inode->i_op->permission, which points to kernfs_iop_permission() (registered in kernfs_iops, kernfs_dir_iops, and kernfs_symlink_iops).
kernfs_iop_permission() acquires root->kernfs_iattr_rwsem for reading:
down_read(&root->kernfs_iattr_rwsem);
kernfs_refresh_inode(kn, inode);
ret = generic_permission(&nop_mnt_idmap, inode, mask);
up_read(&root->kernfs_iattr_rwsem);
kernfs_iop_getattr() (fs/kernfs/inode.c:182-196) follows the same pattern, calling kernfs_refresh_inode() and generic_fillattr() under down_read(&root->kernfs_iattr_rwsem).
-
Unprotected writes in kernfs_refresh_inode() (fs/kernfs/inode.c:165-180):
- Because
root->kernfs_iattr_rwsem is only acquired for reading, multiple tasks can enter kernfs_iop_permission() (and kernfs_iop_getattr()) simultaneously for the same struct kernfs_node (kn) and struct inode (inode).
kernfs_refresh_inode() unconditionally performs plain (non-atomic) writes to inode fields:
inode->i_mode = kn->mode; (fs/kernfs/inode.c:169), a 2-byte umode_t write at offset 0 of struct inode.
set_inode_attr(inode, attrs); (fs/kernfs/inode.c:176), which writes inode->i_uid, inode->i_gid, and timestamps.
set_nlink(inode, kn->dir.subdirs + 2); (fs/kernfs/inode.c:179), which writes inode->__i_nlink for active directories.
-
Concurrent read in generic_permission() / acl_permission_check() (fs/namei.c:438-531):
- After
kernfs_refresh_inode(), kernfs_iop_permission() calls generic_permission(&nop_mnt_idmap, inode, mask) (fs/kernfs/inode.c:288), which calls acl_permission_check() (fs/namei.c:529).
acl_permission_check() performs a plain 2-byte read of inode->i_mode (unsigned int mode = inode->i_mode; at fs/namei.c:441) and may also read inode->i_uid and inode->i_gid.
- In the reported trace, Task 3731 (
openat -> kernfs_iop_permission() -> kernfs_refresh_inode() at fs/kernfs/inode.c:169) writes 2 bytes to inode->i_mode (0xffff888107516f88) concurrently with Task 3010 (readlink -> kernfs_iop_permission() -> generic_permission() -> acl_permission_check() at fs/namei.c:441) reading those 2 bytes.
Lifecycle and Invariants
- Modifications to
kn->mode (via __kernfs_setattr() in kernfs_setattr() / kernfs_iop_setattr()) and kn->iattr attributes are serialized under down_write(&root->kernfs_iattr_rwsem).
kernfs_iattr_rwsem protects the kernfs_node attributes against concurrent setattr, but it does not serialize the copy/refresh of those attributes from kn into the VFS struct inode, nor does kernfs_iop_permission() hold inode->i_lock or inode->i_rwsem exclusively. Consequently, any two concurrent readers executing kernfs_iop_permission() or kernfs_iop_getattr() on the same kernfs inode violate the C memory model invariant against unsynchronized concurrent read/write (and write/write) accesses to inode->i_mode (as well as i_uid, i_gid, and __i_nlink).
Sample crash report:
* Struck through repros no longer work on HEAD.