On Fri, Jun 13, 2025 at 03:41:34AM +0100, Al Viro wrote: > On Fri, Jun 13, 2025 at 12:37:33PM +1000, NeilBrown wrote: > > > If two threads in the same namespace look up the same name at the same > > time (which previously didn't exist), they will both enter > > d_alloc_parallel() where neither will notice the other, so both will > > create and install d_in_lookup() dentries, and then both will call > > ->lookup, creating two identical inodes. > > > > I suspect that isn't fatal, but it does seem odd. > > > > Maybe proc_sys_compare should return 0 for d_in_lookup() (aka !inode) > > dentries, and then proc_sys_revalidate() can perform the is_seen test > > and return -EAGAIN if needed, and __lookup_slow() and others could > > interpret that as meaning to "goto again" without calling > > d_invalidate(). > > Umm... Not sure it's the best solution; let me think a bit. Just need > to finish going through the ported rpc_pipefs series for the final look > and posting it; should be about half an hour or so... FWIW, I think we need the following: mismatch in name/len => return 1 in_lookup => return 0, let the fucker get rechecked later when it ceases to be in_lookup; can only happen when we are called from d_alloc_parallel(). otherwise, NULL inode => return 1; we are seeing a dentry halfway through __dentry_kill(); caller is a lockless dcache lookup, under RCU otherwise, check ->sysctl and sysctl_is_seen(). And yes, you do need rcu_dereference() there. Caller must be holding rcu_read_lock or dentry->d_lock or have a counting reference to dentry.
