On Sun, Dec 8, 2019 at 6:54 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > > On Thu, Nov 28, 2019 at 04:48:58PM +0100, Miklos Szeredi wrote: > > > Turning rename_lock into a per-sb lock mitigates the above issues. > > ... and completely fucks d_lookup(), since false negative can be > caused by rename of *ANYTHING* you've encountered while walking > a hash chain. Drat. Could we use the bitlock in the hash chain head for verification of a negative? Seems like much finer grained than having to rely on a global seqlock for hash integrity. Thanks, Miklos
