Apologies if my long message derailed this discussion. I meant to support Mimi's concern here and project a future vision where TCG_TPM2_HMAC doesn't conflict with other features.
More concisely, I think that: > tpm2_get_random() is costly when TCG_TPM2_HMAC is enabled is not a compelling argument for removing TPM as an RNG source, because TCG_TPM2_HMAC is known to have poor performance already anyway. Thanks Chris On Thu, Mar 5, 2026 at 7:37 AM Mimi Zohar <[email protected]> wrote: > > On Tue, 2026-03-03 at 23:32 +0200, Jarkko Sakkinen wrote: > > On Fri, Feb 20, 2026 at 10:30:26AM -0800, Chris Fenner wrote: > > > My conclusion about TCG_TPM2_HMAC after [1] and [2] was that > > > TPM2_TCG_HMAC doesn't (or didn't at the time) actually solve the > > > threat model it claims to (active interposer adversaries), while > > > dramatically increasing the cost of many kernel TPM activities beyond > > > the amount that would have been required to just solve for > > > passive/bus-sniffer interposer adversaries. The added symmetric crypto > > > required to secure a TPM transaction is almost not noticeable; the big > > > performance problem is the re-bootstrapping of the session with ECDH > > > for every command. > > > > > > My primary concern at that time was, essentially, that TCG_TPM2_HMAC > > > punts on checking that the key that was used to secure the session was > > > actually resident in a real TPM and not just an interposer adversary. > > > I wrote up my understanding at > > > https://www.dlp.rip/decorative-cryptography, for anyone who wants a > > > long-form opinionated take :). > > > > > > Unless I'm wrong, or TCG_TPM2_HMAC has changed dramatically since > > > August, I don't think "TPM2_TCG_HMAC makes this too costly" is a > > > compelling reason to make a security decision. (There could be other > > > reasons to make choices about whether to use the TPM as a source of > > > randomness in the kernel! This just isn't one IMHO.) > > > > > > The version of TCG_TPM2_HMAC that I'd like to see someday would be one > > > that fully admits that its threat model is only passive interposers, > > > and sets up one session upon startup and ContextSaves/ContextLoads it > > > back into the TPM as needed in order to secure parameter encryption > > > for e.g., GetRandom() and Unseal() calls. > > > > Neither agreeing nor disagreeing but this patch set clearly does not > > move forward and I spent already enough energy for this. For better > > ideas the patches are available in queue branch. > > Jarkko, you totally ignored my comments below. I object to your removing the > TPM trusted-keys RNG support. > > Mimi > > > > > High-level takes don't move anything forward (or backward), sorry. > > > > > > > > [1]: > > > https://lore.kernel.org/linux-integrity/CAMigqh2nwuRRxaLyOJ+QaTJ+XGmkQj=rmj5k9gp1bccxp2o...@mail.gmail.com/ > > > [2]: > > > https://lore.kernel.org/linux-integrity/[email protected]/ > > > > > > Thanks > > > Chris > > > > > > On Fri, Feb 20, 2026 at 10:04 AM Mimi Zohar <[email protected]> wrote: > > > > > > > > [Cc: Chris Fenner, Jonathan McDowell, Roberto] > > > > > > > > On Sun, 2026-01-25 at 21:25 +0200, Jarkko Sakkinen wrote: > > > > > 1. tpm2_get_random() is costly when TCG_TPM2_HMAC is enabled and thus > > > > > its > > > > > use should be pooled rather than directly used. This both reduces > > > > > latency and improves its predictability. > > > > > > > > If the concern is the latency of encrypting the bus session, please > > > > remember > > > > that: > > > > > > > > - Not all environments expose the TPM bus to sniffing. > > > > - The current TPM trusted keys design is based on TPM RNG, but already > > > > allows it > > > > to be replaced with the kernel RNG via the "trusted_rng=kernel" boot > > > > command > > > > line option. > > > > - The proposed patch removes that possibility for no reason. > > > > > > > > Mimi & Elaine > > > > > > > > > > > > BR, Jarkko
