On Wed May 22, 2024 at 4:35 PM EEST, James Bottomley wrote: > On Wed, 2024-05-22 at 09:18 +0100, Vitor Soares wrote: > > On Tue, 2024-05-21 at 08:33 -0400, James Bottomley wrote: > > > On Tue, 2024-05-21 at 10:10 +0300, Jarkko Sakkinen wrote: > > > > This benchmark could be done in user space using /dev/tpm0. > > > > > > Let's actually try that. If you have the ibmtss installed, the > > > command to time primary key generation from userspace on your tpm > > > is > > > > > > time tsscreateprimary -hi n -ecc nistp256 > > > > > > > > > And just for chuckles and grins, try it in the owner hierarchy as > > > well (sometimes slow TPMs cache this) > > > > > > time tsscreateprimary -hi o -ecc nistp256 > > > > > > And if you have tpm2 tools, the above commands should be: > > > > > > time tpm2_createprimary -C n -G ecc256 > > > time tpm2_createprimary -C o -G ecc256 > > > > > > James > > > > > > > > > > Testing on an arm64 platform I get the following results. > > > > hmac disabled: > > time modprobe tpm_tis_spi > > real 0m2.776s > > user 0m0.006s > > sys 0m0.015s > > > > time tpm2_createprimary -C n -G ecc256 > > real 0m0.686s > > user 0m0.044s > > sys 0m0.025s > > > > time tpm2_createprimary -C o -G ecc256 > > real 0m0.638s > > user 0m0.048s > > sys 0m0.009s > > > > > > hmac enabled: > > time modprobe tpm_tis_spi > > real 8m5.840s > > user 0m0.005s > > sys 0m0.018s > > > > > > time tpm2_createprimary -C n -G ecc256 > > real 5m27.678s > > user 0m0.059s > > sys 0m0.009s > > > > (after first command) > > real 0m0.395s > > user 0m0.040s > > sys 0m0.015s > > > > time tpm2_createprimary -C o -G ecc256 > > real 0m0.418s > > user 0m0.049s > > sys 0m0.009s > > That's interesting: it suggests the create primary is fast (as > expected) but that the TPM is blocked for some reason. Is there > anything else in dmesg if you do > > dmesg|grep -i tpm > > ? > > Unfortunately we don't really do timeouts on our end (we have the TPM > do it instead), but we could instrument your kernel with command and > time sent and returned. That may tell us where the problem lies. If there was possibility to use bpftrace it is trivial to get histogram of time used where. I can bake a script but I need to know first if it is available in the first place before going through that trouble. BR, Jarkko
