On Fri, Apr 26, 2019 at 09:36:00AM -0700, Linus Torvalds wrote: > On Fri, Apr 26, 2019 at 9:25 AM Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > > > Is it really ok to union the count and rcu_head there? > > It should be fine, because the rcu_count should only ever be used once > the count has gone to zero and the name cannot be found any more. > > And while RCU path walking may find and use the *name* after the > dentry has been killed off (but not free'd yet), all the actual > external_name() accesses should be serialized by the dentry lock, so > there's no access to those fields once the dentry is dead. It's not quite that; access to external_name contents is fine, ->d_lock or not. __d_lookup_rcu() does read it under rcu_read_lock alone. However: * we never free it without an RCU delay after the final drop of refcount. RCU delay might happen on dentry->d_rcu (if it's dentry_free()) or on name->p.rcu (if it's release_dentry_name_snapshot() or d_move() dropping the final reference). * it's never observed in ->d_name after the refcount reaches zero. * no lockless access ever looks at the refcount. It can look at ->name[], but that's it. What I don't understand is why would anyone want to mess with name snapshots for dentry_path() lookalikes...
