Re: [PATCH 0/1] SBI: Introduce Physical Memory Protection Extension

Jonathan Behrens <behrensj@...>

There are different sorts of security threats to keep in mind. I'm generally not a fan of threat models where the "person who legally purchased the computer" is the threat, but that is admittedly a case hardware vendors often care about. Another security concern for many organizations is basically the reverse: somebody has attached an extra device or some DRAM with unauthorized M-mode code that the operating system software doesn't know about, and now it is being used to exfiltrate data from the system. This isn't necessarily the place to try to resolve this sort of threat, but it is worth keeping in mind that decisions we make don't make it even harder to address.


(sorry for the duplicate... my other address isn't subscribed to the mailing list)

On Wed, Apr 1, 2020 at 9:48 PM Nick Kossifidis <mick@...> wrote:
Hello all,

> Let's continue the discussion here as it has a wider audience than
> github PR. I have also cc'd all the possible stakeholders.
> The other alternatives propsed so far
> 1. Just parse the deveice tree node in S-mode software.
> Pros: No additional SBI extensions are required.
> Cons: U-Boot is responsible for copying the reserved-memory node from
> the DT passed by OpenSBI and set it to the final DT if it is a
> different one. OpenSBI also need to provide the DT where the previous
> stage (FSBL/U-Boot SPL) doesn't provide a DT.
> IMHO, this is not a very difficult problem to solve. The U-Boot
> implementation is available here (just ~40 lines of code).
> 2. Trap-n-Emulate PMP CSR reads from S-mode. The details are available
> here.
> Pros: No SBI extension or prior DTB fixup required in U-Boot.
> Cons: As PMP extensions proposals are already in place, this may need
> to be extended. Privilege spec may need to be modified to explicitly
> say that M-mode can provide PMP csr emulation.
> Any thoughts ?

The assumption that there'll always be a PMP rule for marking the
firmware's memory isn't always true. M-mode can access any region
without a matching PMP rule so basically the firmware can be anywhere
and work fine, without any PMP rule marking its region. With the current
PMP spec this is probably the best approach since it will also prevent
S-mode from ever accessing this region (S-mode will fail if there is no
matching PMP rule). Why waste a rule for the firmware's region to deny
any access from S-mode, when you can have the same result by not putting
that rule on PMP ? Even if that assumption was always true, what about
systems without PMP, systems with a different number of PMP regions, or
systems with custom PMP-equivalent solutions ? Trap-n-emulate is not
that generic (e.g. in the case of different number of PMP regions). Any
interface from S-mode to PMP or a PMP-equivalent mechanism, should be
more abstract than this IMHO.

We should also consider the security implications of exposing the PMP
configuration to a less privileged mode. It's one thing to let S-mode
know about the DRAM regions that it can't touch because the firmware is
there, and another to reveal the full PMP configuration. PMP may contain
settings that S-mode doesn't even know about, devices that are not on
the device tree for example or DRAM regions that are outside the device
tree's /memory node range (so S-mode won't touch them anyway). I'm not a
fan of security through obscurity but there are good reasons to keep PMP
configuration accessible only to M-mode (in addition to having a more
abstract API with S-mode to be able to handle the various system
configurations / implementations). We can have S-mode request M-mode to
protect specific physical regions that S-mode manages through PMP (or a
vendor-specific PMP-equivalent mechanism) but anything more than that
doesn't make much sense to me.

If we just want to prevent S-mode from touching firmware's memory we can
either exclude its region from the /memory node on the device tree, or
let the firmware modify the device tree it passes on to the next-stage
boot loader and add a reserved-memory node. I think that's the best
approach since it's the most generic one and it's standards-compliant,
plus the device-tree is platform-specific anyway, even if it's included
on the kernel blob directly without passing through OpenSBI or whatever
boot loader is there, we can always add there a reserved region, or
modify the /memory node manually.


Join { to automatically receive all group messages.