On Thu, Nov 2, 2017 at 6:33 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote: > On Wed, Nov 1, 2017 at 9:22 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: >> Miklos, >> >> My trigger for fixing up the non-samefs patches is that I found out >> while testing NFS export that reconnecting a dir dentry doesn't work >> for non-samefs case. >> I tracked the problem down to the default implementation of get_name() >> in expfs.c that looks up a child inode in parent dir by d_ino. > > That may be a good fix, but.. Not that good, it creates dir cache for pure lower without handling NULL upper properly. Pushed tested fix to ovl-nonsamefs-v7 with some re-arrange of the patches and more blurb in commit messages. > we shouldn't have to rely on correct > d_ino to get the name. > > In fact we can just recursively call exportfs_get_name() on underlying > fs, since we have the parent and the child handle in both the dir and > non-dir cases. > Yes, I have that implementation in the early pure upper decode patches, but threw it out later on. > Am I missing something? > No, in fact implementation of get_name() is even simpler than that. The trick with copy up of dir on encode is that we are guarantied to have a connected upper real dir dentry before we decode the overlay dentry (see simple implementation of get_parent() for example), so we can lookup the name in dcache. I just thought, hey, I'll apply Chandan's patch to fix get_name(), but it turned out to be a lot more. V7 should be fine now, although I still need to improve the xfstest to make sure I covered all cases properly. I did already torture V7 (+ovl-nfs-export-wip +ovl_snapshot_wip) with snapshot tests, so I am more confident about it than V6. Will post V7 patched to list after I have the test, but feel free to review. Thanks, Amir. -- To unsubscribe from this list: send the line "unsubscribe linux-unionfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
