On Thu, 2024-09-12 at 09:26 -0400, James Bottomley wrote:
> On Thu, 2024-09-12 at 16:16 +0300, Jarkko Sakkinen wrote:
> > On Wed Sep 11, 2024 at 3:21 PM EEST, James Bottomley wrote:
> > > On Wed, 2024-09-11 at 10:53 +0200, Roberto Sassu wrote:
> [...]
> > > > I made few measurements. I have a Fedora 38 VM with TPM
> > > > passthrough.
> > > >
> > > > Kernels: 6.11-rc2+ (guest), 6.5.0-45-generic (host)
> > > >
> > > > QEMU:
> > > >
> > > > rc qemu-kvm 1:4.2-
> > > > 3ubuntu6.27
> > > > ii qemu-system-x86 1:6.2+dfsg-
> > > > 2ubuntu6.22
> > > >
> > > >
> > > > TPM2_PT_MANUFACTURER:
> > > > raw: 0x49465800
> > > > value: "IFX"
> > > > TPM2_PT_VENDOR_STRING_1:
> > > > raw: 0x534C4239
> > > > value: "SLB9"
> > > > TPM2_PT_VENDOR_STRING_2:
> > > > raw: 0x36373000
> > > > value: "670"
> > > >
> > > >
> > > > No HMAC:
> > > >
> > > > # tracer: function_graph
> > > > #
> > > > # CPU DURATION FUNCTION CALLS
> > > > # | | | | | | |
> > > > 0) | tpm2_pcr_extend() {
> > > > 0) 1.112 us | tpm_buf_append_hmac_session();
> > > > 0) # 6360.029 us | tpm_transmit_cmd();
> > > > 0) # 6415.012 us | }
> > > >
> > > >
> > > > HMAC:
> > > >
> > > > # tracer: function_graph
> > > > #
> > > > # CPU DURATION FUNCTION CALLS
> > > > # | | | | | | |
> > > > 1) | tpm2_pcr_extend() {
> > > > 1) | tpm2_start_auth_session() {
> > > > 1) * 36976.99 us | tpm_transmit_cmd();
> > > > 1) * 84746.51 us | tpm_transmit_cmd();
> > > > 1) # 3195.083 us | tpm_transmit_cmd();
> > > > 1) @ 126795.1 us | }
> > > > 1) 2.254 us | tpm_buf_append_hmac_session();
> > > > 1) 3.546 us | tpm_buf_fill_hmac_session();
> > > > 1) * 24356.46 us | tpm_transmit_cmd();
> > > > 1) 3.496 us | tpm_buf_check_hmac_response();
> > > > 1) @ 151171.0 us | }
> > >
> > > Well, unfortunately, that tells us that it's the TPM itself that's
> > > taking the time processing the security overhead. The ordering of
> > > the commands in tpm2_start_auth_session() shows
> > >
> > > 37ms for context restore of null key
> > > 85ms for start session with encrypted salt
> > > 3ms to flush null key
> > > -----
> > > 125ms
> > >
> > > If we context save the session, we'd likely only bear a single 37ms
> > > cost to restore it (replacing the total 125ms). However, there's
> > > nothing we can do about the extend execution going from 6ms to
> > > 24ms, so I could halve your current boot time with security enabled
> > > (it's currently 149ms, it would go to 61ms, but it's still 10x
> > > slower than the unsecured extend at 6ms)
> > >
> > > James
> >
> > I'll hold for better benchmarks.
>
> Well, yes, I'd like to see this for a variety of TPMs.
>
> This one clearly shows it's the real time wait for the TPM (since it
> dwarfs the CPU time calculation there's not much optimization we can do
> on the kernel end). The one thing that's missing in all of this is
> what was the TPM? but even if it's an outlier that's really bad at
> crypto what should we do? We could have a blacklist that turns off the
> extend hmac (or a whitelist that turns it on), but we can't simply say
> too bad you need a better TPM.
Ops, sorry. I pasted the TPM properties. Was not that clear:
Infineon Optiga SLB9670 (interpreting the properties).
Roberto
