On 2/16/25 11:17 PM, Sumit Garg wrote:
On Thu, Feb 13, 2025 at 09:19:58AM -0600, Stuart Yoder wrote:
On 2/12/25 11:31 PM, Sumit Garg wrote:
+ Rob
On Thu, 13 Feb 2025 at 03:25, Stuart Yoder <stuart.yoder@xxxxxxx> wrote:
On 2/12/25 1:39 AM, Sumit Garg wrote:
On Tue, 11 Feb 2025 at 21:39, Stuart Yoder <stuart.yoder@xxxxxxx> wrote:
Hi Sumit,
On 2/11/25 12:45 AM, Sumit Garg wrote:
+ Jens
Hi Stuart,
On Tue, 11 Feb 2025 at 04:52, Stuart Yoder <stuart.yoder@xxxxxxx> wrote:
These patches add support for the CRB FF-A start method defined
in the TCG ACPI specification v1.4 and the FF-A ABI defined
in the Arm TPM Service CRB over FF-A (DEN0138) specification.
(https://developer.arm.com/documentation/den0138/latest/)
Nice to have a specification standardizing interface to TPM
managed/implemented by the firmware. Care to add corresponding kernel
documentation under Documentation/security/tpm/.
Yes, I can add some documentation there.
BTW, we already have drivers/char/tpm/tpm_ftpm_tee.c, so do you see
possibilities for an abstraction layer on top of communication channel
based on either FF-A or TEE or platform bus?
I think the CRB and OP-TEE based messaging approaches for interacting
with a TZ-based TPM are fundamentally different and I don't see how
to harmonize them through some abstraction.
The OP-TEE TPM protocol copies the TPM command into a temp shared memory
buffer and sends a message to the TPM referencing that buffer.
The CRB uses a permanently shared memory carve-out that in addition
to the command/response data has other fields for locality control,
command control, status, TPM idle, etc. The only 'message' needed is
something to signal 'start'. Any OS that is FF-A aware and has a
CRB driver can simply add a new start method, which is what this
patch series does.
Okay, I see how the CRB driver is closely tied to the ACPI based
systems.
The CRB driver is currently probed based on ACPI, but it fundamentally
doesn't have to be. If there was a DT binding for CRB-based
TPMs the different start methods would be defined there and the
CRB driver could support that.
Can't we rather enable the CRB driver itself probed based on FF-A bus
and rather dynamically discover the shared memory buffer via FF-A
instead? AFAIU, FF-A provides you with a discovery framework for
firmware bits.
Yes, you could do this. But, then the TPM CRB drivers in all the
ACPI-based OSes (Linux, Windows) and hypervisors need to be
taught this new method of discovery. Adding new start methods is
reasonably straightforward, but changing the basic discovery
mechanism is a much bigger change.
We will be teaching every other OS or hypervisor about FF-A
communication regardless. So it's rather about if we want to do it
properly leveraging auto discovery mechanisms supported by FF-A or not.
But if we still want to overload ACPI or DT with the
discoverable firmware bits then it seems like an overkill here.
I think it would make sense to do ACPI based discovery or FF-A
based discovery, but doing both I think would be overkill. For
ease of OS integration ACPI is the way to go. And, potentially
device tree in the future.
Encoding firmware bits in ACPI/DT can be seen as an easy upstream path
in the shorter run. But when the ACPI/DT becomes overloaded with
information that has to be passed from firmware to the OS rather than
purely describing hardware to the OS, it's ABI maintainability becomes
complex. We are already dealing with DT ABI compatibility challenges
especially the forward compatibility, so let's not make it even worse
with firmware information that can be discovered automatically.
The TCG defined ACPI table has the following:
-Physical address of the TPM
-Start method
-Start method specific parameters
-event log address
This has been in place 8+ years and this is what OSes expect.
The start method advertises the mechanism a driver uses to
signal the TPM that something has changed in the CRB, and
this allows different types of TPM implementations:
-memory mapped
-signal via ACPI
-signal via ARM SMC (legacy)
-signal via Pluton mailbox
-signal via FF-A
I don't see this as overloading the ACPI table, it's just what
the OS needs to know.
The TPM does not know (and should not know) the address of
the event log. An FF-A based TPM has no way to know this.
I don't see how changing TPM discovery to be via FF-A directly
would improve maintainability.
The other benefit of auto discovery is that platform enablement becomes
really smooth. Once the firmware starts supporting a particular feature
like TPM over FF-A then the OS can discover and support it.
If we added new CRB/FF-A ABIs to get the CRB physical address,
start method specific parameters, event log, it would mean that
all OSes and hypervisors need to re-architect their CRB drivers
or create new FF-A specific CRB drivers. That will not smooth
enablement for TPMs. And I don't see advantages for
maintainability.
Thanks,
Stuart
