On Mon, Feb 6, 2017 at 5:25 AM, J. R. Okajima <hooanon05g@xxxxxxxxx> wrote: > James Bottomley: >> This allows any subtree to be uid/gid shifted and bound elsewhere. It > ::: > > Interesting. > But I am afraid that the inconsistency problem of the inode numbers will > happen. > Yet another example that overlayfs already is in the process of solving (it is fixed for stat of merged directory inode). In fact, fir the case of single layer overlay (as well as shiftfs) the solution is trivial - preserve underlying inode st_ino/d_ino and use the overlayed fs st_dev. > shiftfs_new_inode() uses get_next_ino() which means > - 1st time: inodeA is created and cached, inumA is assigned > - after using inodeA, it will be discarded from the cache > - 2nd time: inodeA is looked-up again, and another inode number (inumB) > is assgined. > > This inconsistency will not be a problem for the "pure virtual" fs such > as procfs and sysfs. But your shiftfs is not pure as them. Shiftfs will > be used as a wrapper (or "binder" which means bind-mount) of an orginary > filesystem. > The symptom of this problem from users perspective will be > - find -inum doesn't work > - git-status doesn't work, which keeps st_dev and st_ino and compares > the current files. > Of course they will be limited to when the target dir is huge and/or > system memory is low. As long as the inode cache is large enough to hold > all necessary inodes, the problem won't happen. > > If shiftfs will supports exporting via NFS in the future, the > consistency of inum will be important too. > > > J. R. Okajima
