On Mon, 23 Mar 2015 13:37:38 +1100 NeilBrown <neilb@xxxxxxx> wrote: > Hi Al, > thanks for all your review help - particularly the fact that > dentry->d_inode is not stable in RCU-walk. That has lead to > a number of changes. > > I think this set addresses all of your review comments, improves > some documentation, and has a go at providing a solution for lustre. > > I hope to organize some proper testing soon, so I can confirm that it > makes certain loads a lot faster. Just FYI - I've now tested this series on a 64-CPU machine running load that has lots of threads doing lots of stats on lots of non-existent files. Without the patches, the test takes twice as long if there is a symlink early in the path for the stats, than if there is not. With the patches, there is no significant difference in the time the test takes between the symlink and no-symlink cases. So it appears to be achieving the goal. Thanks, NeilBrown
Attachment:
pgpLB1fo4cJfy.pgp
Description: OpenPGP digital signature
