On Tue, Feb 28, 2017 at 9:44 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > On Tue, Feb 28, 2017 at 11:08:31AM -0800, Josh England wrote: > >> The behavior I've noted so far from this patch seems to be very well >> defined and follows the rule of least surprise (at least for me). I >> can change files/directories in the shared root and the changes are >> instantly visible to all clients. If a file/directory has been copied >> up on the client that change takes precedence over changes in the >> lower. Deletes work as expected. I haven't found a downside yet. > > Try to do something that invalidates a directory in underlying layer > with stuff in overlay on its subdirectories. That's an interesting test to run, but why does ovl_dentry_revalidate() need to return 1/-ESTALE instead of ret? Why not propagate the return values 0 and -ECHILD from NFS? After all, NFS lower directory inode can be invalidated after lookup and then other overlayfs ops (e.g. readdir) could fail with -ESTALE just the same as those ops would fail directly over NFS. No? What am I missing?
