On Mon May 27, 2024 at 5:51 PM EEST, Jarkko Sakkinen wrote: > On Thu May 23, 2024 at 10:59 AM EEST, Vitor Soares wrote: > > On Wed, 2024-05-22 at 19:11 +0300, Jarkko Sakkinen wrote: > > > On Wed May 22, 2024 at 5:58 PM EEST, Vitor Soares wrote: > > > > I did run with ftrace, but need some more time to go through it. > > > > > > > > Here the step I did: > > > > kernel config: > > > > CONFIG_FUNCTION_TRACER > > > > CONFIG_FUNCTION_GRAPH_TRACER > > > > > > > > ftrace: > > > > # set filters > > > > echo tpm* > set_ftrace_filter > > > > > > > > # set tracer > > > > echo function_graph > current_tracer > > > > > > > > # take the sample > > > > echo 1 > tracing_on; time modprobe tpm_tis_spi; echo 0 > tracing_on > > > > > > > > regards, > > > > Vitor Soares > > > > > > I'm now compiling distro kernel (OpenSUSE) for NUC7 with v6.10 contents. > > > > > > After I have that setup, I'll develop a perf test either with perf or > > > bpftrace. I'll come back with the possible CONFIG_* that should be in > > > place in your kernel. Might take up until next week as I have some > > > conference stuff to prepare but I try to have stuff ready early next > > > week. > > > > > > No need to rush with this as long as possible patches go to rc2 or rc3. > > > Let's do a proper analysis instead. > > > > > > In the meantime you could check if you get perf and/or bpftrace to > > > your image that use to boot up your device. Preferably both but > > > please inform about this. > > > > > > > I already have perf running, for the bpftrace I might not be able to help. > > The interesting function to look at with/without hmac is probably > tpm2_get_random(). > > I attached a patch that removes hmac shenigans out of tpm2_get_random() > for the sake of proper comparative testing. Other thing that we need to measure is to split the cost into two parts: 1. Handshake, i.e. setting up and shutdowning a session. 2. Transaction, payload TPM command. This could be done by setting up couple of kprobes_events: payload_event: tpm2_get_random() etc. hmac_event: tpm2_start_auth_session(), tpm2_end_auth_session() etc. And just summing up the time for a boot to get a cost for hmac. I'd use bootconfig for this: https://www.kernel.org/doc/html/v6.9/trace/boottime-trace.html So I've made up plans how measure the incident but not sure when I have time to pro-actively work on a benchmark (thus sharing details). So I think with just proper bootconfig wtih no other tools uses this can be measured. BR, Jarkko
