Messages in this thread | | | From | Amir Goldstein <> | Date | Wed, 30 Nov 2016 17:05:25 +0200 | Subject | Re: [POC/RFC PATCH] overlayfs: constant inode numbers |
| |
On Tue, Nov 29, 2016 at 11:49 PM, Miklos Szeredi <miklos@szeredi.hu> wrote: > On Tue, Nov 29, 2016 at 1:03 PM, Amir Goldstein <amir73il@gmail.com> wrote: ... > I meant that we can unify OVL_XATTR_INO with "redirect/fh" > functionality and get something good out of it. > >> Perhaps you meant for non-dir: >> >> 5. If redirect_dir=fh, *propagate* lowest-handle on non-dir copy up >> 6. In ovl_lookup() of non-dir, decode lowest-handle to set oe->ino > > Yes. > > OVL_XATTR_FH would be safe to ignore, so this is back and forward > compatible.. And the cost is probably not prohitive, since copy ups > should be relatively rare. > > After a backup + restore it is not expected that we get back the old > inode numbers so it's fine to ignore the stale file handles. >
FYI, there are 2 interesting corner case of "semi stale" handles: - Copy of layers to same fs (without deleting old layers) - Old layers are deleted but an old deleted file is still open
I have handled both these cases in the last version of redirect_fh that I pushed yesterday, but not 100% sure that I handled them correctly.
Anyway, I will get to work on adjusting redirect_fh for use by stable inodes.
> The following issues are left: > > - performance of readdir;
Here is one very simple optimization for WIP: @@ -157,6 +157,8 @@ static int ovl_fill_lowest(struct ovl_readdir_data *rdd, list_move_tail(&p->l_node, &rdd->middle); } else { p = ovl_cache_entry_new(rdd, name, namelen, ino, d_type); + if (p) + p->ino = ino;
For non-lowets entry, we can provide mount option 'readdir_ino'. With readdir_ino, readdir pays a penalty of getxattr for any non-lowest entry (either OVL_XATTR_FH or OVL_XATTR_INO). Without readdir_ino, readdir will get d_ino = 0, in which case, at least `find <path> --inum=<n>` does the right thing (fallback to fstat for this dirent).
> - what to do if not all layers are on the same fs;
Same as what I did for redirect_fh - turn the feature off. We can also export this state in /proc/mounts options and maybe allow to explicitly turn off stable inodes, but I don't think that we should, because there shouldn't be a program which relies on inode numbers NOT being stable.
> - hard link copy ups. >
I'll start by setting up a TODO Wiki page and writing xfstests for all those issues. Maybe even track them on github..
Amir.
| |