On Wed, Apr 19, 2017 at 4:52 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote: > On Wed, Apr 19, 2017 at 12:37 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: >> On Wed, Apr 19, 2017 at 12:16 PM, Miklos Szeredi <miklos@xxxxxxxxxx> wrote: >>> On Tue, Apr 18, 2017 at 8:37 PM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: >>>> >>>> On Mon, Apr 17, 2017 at 2:59 AM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: >>>> > Overlayfs inodes are considered unstable in several aspects, >>>> > because on a copy up event: >>>> > 1. st_ino can change >>>> > 2. st_dev can change >>>> > 3. hardlinks are broken >>>> > 4. NFS handle would become stale >>>> > 5. content of read-only file descriptor would become stale >>>> > >>>> > This patch set 'stabilizes' overlayfs inodes w.r.t. st_ino/st_dev >>>> > and takes some big steps in the direction of stabilizing hardlinks >>>> > and NFS handles. >>>> > >>>> >>>> I realized I forgot to mention in the cover letter that stable inodes >>>> are only available for the overlay configuration where all layers >>>> are on the same underlying fs and that underlying fs support >>>> NFS export (I think all eligible upper fs support NFS export anyway). >>> >>> Hmm, we could keep inode numbers stable across copy up even if layers >>> are on different filesystems: just need to use a separate st_dev for >>> lower layers and keep st_dev and st_inode constant. The only extra >>> thing needed compared to the samefs case is the allocation of dummy >>> device numbers for lower layers. Of course "find -xdev" and the like >>> still won't work properly, and we wouldn't be able to provide sane >>> d_ino values in readdir. >> >> Not sure that is going to be worth the effort, but we'll see. >> Anyway, not sure if you already read far enough into the series, >> but the fact that overlay inodes are hashed by stable inode ino >> helps solving a lot of the problems with minimal code changes, >> so in the grand scheme of things, I think it would be easier to >> say: same_fs can give you POSIX. non same_fs cannot. > > The effort should be small, and the reward is substantially less weird behavior. > > In fact I looked and SUSv4 > (http://pubs.opengroup.org/onlinepubs/9699919799/) only talks about > "mount point" in the context of directories. It does *not* require > st_dev to be the as the st_dev of the parent directory for > non-directories. The only requirement is that st_ino and st_dev > together uniquely identify a file in the system, which is why we need > to generate a dummy st_dev for lower files in this case. It also > explicitly only talks about directories in the context of "-xdev" and > the like. > I did see that 'find' does list files from overlay which have differnt st_dev than parent, but 'du -x' does not count the files, which should be very annoying to users. I'm surprised I haven't heard about this. > So even in the non-samefs case we could stamp it with "POSIX > compliant" because strictly speaking it is. > > If that's not enough, I think we *can* do unified ino space even in > most non-samefs cases. And here's why: look at the inode numbers of > any filesystem; they will always be "small" so we can just partition > the 64 bit ino space between layers and map inode numbers into its own > partition. This does not work in the general case, and it is a hack. > But it's a very simple hack and it probably works fine. Similar thing > is assumed by the 32bit compat code, which just returns EOVERFLOW if > the ino happens to be too large, which I guess doesn't happen too > often for most filesystems... > Well, if you are lucky you can run into a filesystem that exports a file handle of type FILEID_INO32_GEN, then you *know* you're good to go. ext* will do that and xfs that was forever mounted with -o inode32. Even with xfs -o inode64, it will not use the MSB ino bits unless you are in the exabytes fs sizes. Anyway, I will keep that in the back on my mind when working on stable inode to keep the implementation open for such improvements in the future. Amir.
