On Wed, Aug 23, 2023 at 5:52 PM Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>
> On Wed, 23 Aug 2023 at 16:43, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
>
> > If we do this, then both overlay.whiteout and overlay.xattr_whiteouts
> > xattrs will be xattrs that the overlayfs driver never sets.
> > It's a precedent, but as long as it is properly documented and encoded
> > in fstests, I will be fine with it. Not sure about Miklos.
>
> Firstly I need to properly understand the proposal. At this point I'm
> not sure what overlay.whiteout is supposed to mean. Does it mean the
> same as a whiteout (chrdev(0,0))? Or does it mean that overlayfs
> should not treat it as a whiteout, but instead transform that into a
> chrdev(0,0) for the top overlay to interpret as a whiteout? Or
> something else?
>
My proposal does not involve any transformation.
It is "just" to support another format for a whiteout.
Transforming a REG or FIFO real object to CHR ovl object
will be a pain IMO and I don't see why it is needed.
Let me try again from the top:
1. ovl_path_is_whiteout() checks if either ovl_is_whiteout() (chardev(0,0))
or regular file with "overlay.whiteout" xattr, so ovl_lookup()
will result in
a negative dentry if either whiteout format is found at topmost layer
2. To optimize unneeded getxattr, "overlay.whiteout" xattr could be checked
only in case the parent dir has xattr "overlay.xattr_whiteouts"
3. mkfs.composefs is responsible of creating the non-chardev whiteouts
as well as the marking the dirs that contains them with
"overlay.xattr_whiteouts" - overlayfs itself never does that
4. ovl_calc_d_ino() (from readdir on a merge dir) returns true if the
the iterated dir has "overlay.xattr_whiteouts" xattr
5. That will cause ovl_cache_update_ino() to lookup the
*overlay dentry* that will be negative (as per rule 1 above)
if either whiteout format is found at topmost layer and that
will cause the readdir dirent to be marked is_whiteout and
filtered out of readdir results
* The trick for readdir is that the the per dirent DT_CHR optimization
is traded off for a per parent dir optimization, but even the worst case
where all directories have xattr_whiteouts, readdir is not more
expensive than readdir with xino enabled, which is the default for
several Linux distros
Hope this is more clear?
Thanks,
Amir.
