On Fri, 04 Jul 2025, Al Viro wrote: > On Mon, Jun 16, 2025 at 12:49:51PM +1000, NeilBrown wrote: > > > The reality is that ->sysctl does not need rcu protection. There is no > > concurrent update except that it can be set to NULL which is pointless. > > I would rather *not* leave a dangling pointer there, and yes, it can > end up being dangling. kfree_rcu() from inside the ->evict_inode() > may very well happen earlier than (also RCU-delayed) freeing of struct > inode itself. In that case could we move the proc_sys_evict_inode() call from proc_evict_inode() to proc_free_inode(), and replace kfree_rcu() with kfree()? Or does the inode need to be deleted from ->sibling_inodes earlier than free_inode? > > What we can do is WRITE_ONCE() to set it to NULL on the evict_inode > side and READ_ONCE() in the proc_sys_compare(). That is likely the simplest change. Thanks, NeilBrown > > The reason why the latter is memory-safe is that ->d_compare() for > non-in-lookup dentries is called either under rcu_read_lock() (in which > case observing non-NULL means that kfree_rcu() couldn't have gotten to > freeing the sucker) *or* under ->d_lock, in which case the inode can't > reach ->evict_inode() until we are done. > > So this predicate is very much relevant. Have that fucker called with > neither rcu_read_lock() nor ->d_lock, and you might very well end up > with dereferencing an already freed ctl_table_header. >
