On Thu, Jun 1, 2017 at 11:02 AM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: > This patch is based on an earlier POC by Miklos Szeredi. > > When iterating an "impure" upper directory which may contain copy up > entries, lookup the overlay dentries to see if we know their copy up > origin and if we do, call vfs_getattr() on the overlay dentry > to make sure that d_ino will be consistent with st_ino from stat(2). Getting back to this, because there are issues: - ".." can have different origin even if dir is pure (pure upper residing in merged dir, pure lower residing in dir which has another lower layer (which is the origin) above it) - we check impure at open time, but dir can become impure between two getdents*(2) calls, we can get wrong inode numbers Becoming impure needs to be handled differently from what this patch does, because we can't switch to cached readdir in the middle of the directory listing. To handle that we need to add our own actor to the is_real case to fix up the inode numbers. This would fix the ".." case as well. The problem is, we are inside underlying fs code when actor is called, so there's not much we can do without risking deadlocks. One thing we can do (hopefully) is call d_lookup() to check for a cached overlay dentry, and get the inode number out. We need to store the full inode number in the overlay inode, because we can't do vfs_getattr() either, but that's a good optimization anyway. If the overlay dentry is not in the cache we need to abort the iteration, lookup the dentry, and continue iterating. Thoughts? I'll have a go at this tomorrow. Thanks, Miklos -- 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
