On 5/15/2019 2:52 AM, Arvind Sankar wrote:
On Tue, May 14, 2019 at 04:54:12PM -0700, James Bottomley wrote:
On Tue, 2019-05-14 at 18:39 -0500, Rob Landley wrote:
On 5/14/19 2:18 PM, James Bottomley wrote:
I think Rob is right here. If /init was statically built into
the kernel image, it has no more ability to compromise the kernel
than anything else in the kernel. What's the problem here?
The specific problem is that unless you own the kernel signing key,
which is really untrue for most distribution consumers because the
distro owns the key, you cannot build the initrd statically into
the kernel. You can take the distro signed kernel, link it with
the initrd then resign the combination with your key, provided you
insert your key into the MoK variables as a trusted secure boot
key, but the distros have been unhappy recommending this as
standard practice.
If our model for security is going to be to link the kernel and the
initrd statically to give signature protection over the aggregate
then we need to figure out how to execute this via the distros. If
we accept that the split model, where the distro owns and signs the
kernel but the machine owner builds and is responsible for the
initrd, then we need to explore split security models like this
proposal.
You can have a built-in and an external initrd? The second extracts
over the first? (I know because once upon a time conflicting files
would append. It sounds like the desired behavior here is O_EXCL fail
and move on.)
Technically yes, because the first initrd could find the second by some
predefined means, extract it to a temporary directory and do a
pivot_root() and then the second would do some stuff, find the real
root and do a pivot_root() again. However, while possible, wouldn't it
just add to the rendezvous complexity without adding any benefits? even
if the first initrd is built and signed by the distro and the second is
built by you, the first has to verify the second somehow. I suppose
the second could be tar extracted, which would add xattrs, if that's
the goal?
James
You can specify multiple initrd's to the boot loader, and they get
loaded in sequence into memory and parsed by the kernel before /init is
launched. Currently I believe later ones will overwrite the earlier
ones, which is why we've been talking about adding an option to prevent
that. You don't have to mess with manually finding/parsing initramfs's
which wouldn't even be feasible since you may not have the drivers
loaded yet to access the device/filesystem on which they live.
Once that's done, the embedded /init is just going to do in userspace
wht the current patch does in the kernel. So all the files in the
external initramfs(es) would need to have IMA signatures via the special
xattr file.
So, the scheme you are proposing is not equivalent: using the distro key
to verify signatures, compared to adding a new user key to verify the
initramfs he builds. Why would it be necessary for the user to share
responsibility with the distro, if the only files he uses come from the
distro?
Note that if you want the flexibility to be able to load one or both of
two external initramfs's, the current in-kernel proposal wouldn't be
enough -- the xattr specification would have to be more flexible (eg
reading .xattr-list* to allow each initramfs to specifiy its own
xattrs. This sort of enhancement would be much easier to handle with the
userspace variant.
Yes, the alternative solution is to parse .xattr-list at the time it is
extracted. The .xattr-list of each initramfs will be processed. Also,
the CPIO parser doesn't have to reopen the file after all other files
have been extracted.
Roberto
--
HUAWEI TECHNOLOGIES Duesseldorf GmbH, HRB 56063
Managing Director: Bo PENG, Jian LI, Yanli SHI
