On Thu, 2026-10-22 at 12:54 +0200, Jiri Pirko wrote:
Wed, Oct 21, 2026 at 09:05:09PM +0200, zohar@linux.ibm.com wrote:
quoted
On Wed, 2026-10-21 at 11:35 -0300, Jason Gunthorpe wrote:
quoted
On Wed, Oct 21, 2026 at 10:12:34AM -0400, Mimi Zohar wrote:
quoted
quoted
quoted
I think for now we should not include a new dynamic log here, the uAPI
around this Jiri drafted can remain a future step. I don't want to
open complicated discussions about new functionality at this
stage. The whole CTLV idea in general is a big enough thing to bite
off for this series.
Okay, dropping the log from this set.
Thank you. Please keep in mind going forward, that the scarce hardware register
limitation (like the 4 TDX RTMRs) should be resolved (e.g. by multiplexing the
extension path), not by combining distinct measurement lists.
I'm not sure what you mean?
Since the HW doesn't support sealing the only *technical* requirement
is the verifier can reproduce the measurement register hash. So long
as we can get there we can do whatever.
Multiplexing extension alone isn't enough. We need some scheme to
reconstruct interleaved logs too.
For example we could add an atomic sequence number for each
extend. Every log sharing a register can report the sequence
number. The verifier would have to globally sort all logs to confirm
the hash.
Hopefully we don't have to tackle this problem. I hope that 4
registers is enough for Linux. One for IMA, one for systemd, one for
user? Plus a spare!
Agreed, I hope you don't have to tackle it either.
quoted
Maybe we can build a multiplexed scheme on the spare? I don't know.
To avoid interleaving the logs, AI had suggested multiplexing or virtualization
of the RTMRs. Whether is has any merit, I'll leave to you. Here's the very
short AI explanation:
"Instead of allowing external drivers or subsystems to execute a direct, raw
extend command to a specific hardware register index, the new attestation
subsystem should act as a Hypervisor for Registers."
Isn't the hypervisor what we are doing now? I mean multiple extension
originators share the same register and eventually the same log. Or do
you suggest to go even further having a "generic_registerX" exposed for
every hw implementation allowing the user to have full abstraction?
The "generic_registerX" would be always alias to something hw specific,
like RTMR2 for example of TDX.
I definitely prefer the latter, as everything continues to work as is with
minimal, if any, change.
AI's wording:
• The Virtual Mapping: The [attestation] subsystem exposes "Virtual RTMRs"
(vRTMRs) internally to the kernel.
• IMA's Dedicated Lane: The kernel config routes IMA to vRTMR_IMA.
Mimi