The tracefs and eventfs permissions are created dynamically based on what the mount point inode has or the instances directory inode has. But the way it worked had some inconsistencies that could lead to security issues as the file system is not behaving like admins would expect.
The files and directories could ignore the remount option that changes the gid or uid ownerships, leaving files susceptable to access that is not expected. This happens if a file had its value changed previously and then a remount changed all the files permissions. The one that was changed previously would not be affected. This change set resolves these inconsistencies. This also fixes the test_ownership.tc test as it would pass on the first time it is run, but fail on the second time, because of the inconsistant state of the permissions. Now you can run that test multiple times and it will always pass. Changes since v2: https://lore.kernel.org/linux-trace-kernel/20240502151547.973653...@goodmis.org/ - The eventfs_inode freeing was incorrect. The kref_put() would call release_ei() that freed the contents of the eventfs_inode then call kfree_rcu() on the eventfs_inode itself. The contents of the eventfs_inode needs to be freed after the RCU synchronization as well. The patches here add even more cases where that's a requirement. - Add a iput callback for the tracefs_inode to clear the TRACEFS_EVENT_INODE flag. This will prevent the clearing of flags in remount to go into the eventfs_remount() function. A RCU grace cycle happens between the clearing of this flag and where the eventfs_inode is freed, so it is OK if the iteration is happening at the same time, as it is done under rcu_read_lock(). Changes since v1: https://lore.kernel.org/linux-trace-kernel/20240502030024.062275...@goodmis.org/ - Testing showed that taking a mutex when freeing the tracefs_inode caused a lockdep splat as it can happen in the RCU softirq context. Convert the mutex to a spinlock for adding and removing the node from the link list, and free the node via call_rcu() so that the iteration of the list only needs to be protected by rcu_read_lock(). Steven Rostedt (Google) (6): eventfs: Free all of the eventfs_inode after RCU tracefs: Reset permissions on remount if permissions are options tracefs: Still use mount point as default permissions for instances eventfs: Do not differentiate the toplevel events directory eventfs: Do not treat events directory different than other directories eventfs: Have "events" directory get permissions from its parent ---- fs/tracefs/event_inode.c | 127 ++++++++++++++++++++++++++++------------------- fs/tracefs/inode.c | 92 ++++++++++++++++++++++++++++++++-- fs/tracefs/internal.h | 14 ++++-- 3 files changed, 175 insertions(+), 58 deletions(-)
