> > > There is another problem, when delayed copy is used. The content comes
> > > from one source, metadata from another.
> > >
> > > I initially created test-file-lower on the lower directory
> > > (overlayfs/data), before mounting overlayfs. After mount on
> > > overlayfs/mnt:
> > >
> > > # getfattr -m - -e hex -d overlayfs/mnt/test-file-lower
> > > # file: overlayfs/mnt/test-file-lower
> > > security.evm=0x02c86ec91a4c0cf024537fd24347b780b90973402e
> > > security.ima=0x0404f2ca1bb6c7e907d06dafe4687e579fce76b37e4e93b7605022da52e6ccc26fd2
> > > security.selinux=0x73797374656d5f753a6f626a6563745f723a756e6c6162656c65645f743a733000
> > >
> > > # chcon -t unconfined_t overlayfs/mnt/test-file-lower
> > >
> > > After this, IMA creates an empty file in the upper directory
> > > (overlayfs/root/data), and writes security.ima at file close.
> > > Unfortunately, this is what is presented from overlayfs, which is not
> > > in sync with the content.
> > >
> > > # getfattr -m - -e hex -d overlayfs/mnt/test-file-lower
> > > # file: overlayfs/mnt/test-file-lower
> > > security.evm=0x021d71e7df78c36745e3b651ce29cb9f47dc301248
> > > security.ima=0x04048855508aade16ec573d21e6a485dfd0a7624085c1a14b5ecdd6485de0c6839a4
> > > security.selinux=0x73797374656d5f753a6f626a6563745f723a756e636f6e66696e65645f743a733000
> > >
> > > # sha256sum overlayfs/mnt/test-file-lower
> > > f2ca1bb6c7e907d06dafe4687e579fce76b37e4e93b7605022da52e6ccc26fd2 overlayfs/mnt/test-file-lower
> > >
> > > # sha256sum overlayfs/root/data/test-file-lower
> > > 8855508aade16ec573d21e6a485dfd0a7624085c1a14b5ecdd6485de0c6839a4 overlayfs/root/data/test-file-lower (upperdir)
> > >
> > > We would need to use the lower security.ima until the copy is made, but
> > > at the same time we need to keep the upper valid (with all xattrs) so
> > > that IMA can update the next time overlayfs requests that.
> > >
> >
> > Yap.
> >
> > As Seth wrote, overlayfs is a combination of upper and lower.
> > The information that IMA needs should be accessible from either lower
> > or upper, but sometimes we will need to make the right choice.
> >
> > The case of security.ima is similar to that of st_blocks -
> > it is a data-related metadata, so it needs to be taken from the lowerdata inode
> > (not even the lower inode). See example of getting STATX_BLOCKS
> > in ovl_getattr().
> >
> > I would accept a patch that special cases security.ima in ovl_xattr_get()
> > and gets it from ovl_i_path_lowerdata(), which would need to be
> > factored out of ovl_path_lowerdata().
> >
> > I would also accept filtering out security.{ima,evm} from
> >
> > But I would only accept it if I know that IMA is not trying to write the
> > security.ima xattr when closing an overlayfs file, only when closing the
> > real underlying upper file.
>
> I don't see how that would be possible. As far as I'm aware, the
> correlation is between the overlay and the underlying lower/uppper
> file, not the other way around. How could a close on the underlying
> file trigger IMA on an overlay file?
>
Well, you are right. it cannot.
What I meant is that close of overlayfs file should NOT open and read
the overlayfs file and recalculate security.ima to store in overlayfs inode
because close of overlayfs file will follow a close of the upper file that
should recalculate and store security.ima in the upper inode.
It is possible that a close of an overlayfs file will update the security
state of the overlayfs inode by copying the security state from the
upper inode.
But then again, I could be misunderstanding the IMA workflows
and it could be more complicated than I try to present it.
This is the reason that I requested the documentation of how
IMA+overlayfs is *expected* to work.
Thanks,
Amir.
