Conversation
Jarkko Sakkinen
jarkko
First performance fixes for TPM HMAC encryption:
https://lore.kernel.org/linux-integrity/D4727YOJY8KZ.L6RKMRBKRCSN@kernel.org/T/#m22ed621c65c75c75900fe77c6b963ac98eb1b624
I don't expect this to fully address the performance challenges. It is the first axis here. Other axis is the session creation but these are independent issues.
https://lore.kernel.org/linux-integrity/D4727YOJY8KZ.L6RKMRBKRCSN@kernel.org/T/#m22ed621c65c75c75900fe77c6b963ac98eb1b624
I don't expect this to fully address the performance challenges. It is the first axis here. Other axis is the session creation but these are independent issues.
1
0
2
I was a bit stressed out of this and did the first thing that I could think of and it seems to deliver promising results:
https://lore.kernel.org/linux-integrity/CALSz7m1WG7fZ9UuO0URgCZEDG7r_wB4Ev_4mOHJThH_d1Ed1nw@mail.gmail.com/
I was hoping to get exactly that 5 seconds drop based on calculating relative times of call stacks measured by bpftrace, but did not expect that the calculations I made in the notebook would be exact match of the empirical results (this never happens) 🤷
I make a similar fix now for auth session objects (this was for not eagerly swapping the public key used as the communications end point).
https://lore.kernel.org/linux-integrity/CALSz7m1WG7fZ9UuO0URgCZEDG7r_wB4Ev_4mOHJThH_d1Ed1nw@mail.gmail.com/
I was hoping to get exactly that 5 seconds drop based on calculating relative times of call stacks measured by bpftrace, but did not expect that the calculations I made in the notebook would be exact match of the empirical results (this never happens) 🤷
I make a similar fix now for auth session objects (this was for not eagerly swapping the public key used as the communications end point).
1
0
2
1
0
0
Post-poning hwrng registeration will not help and was a false conclusion .
I attached kprobes using boot-time config and these appear when init process is already running:
start_event: (tpm_transmit+0x4/0x4b0)
end_event: (tpm_transmit_cmd+0x33/0xc0 <- tpm_transmit)
start_event: (tpm_transmit+0x4/0x4b0)
end_event: (tpm_transmit_cmd+0x33/0xc0 <- tpm_transmit)
start_event: (tpm_transmit+0x4/0x4b0)
end_event: (tpm_transmit_cmd+0x33/0xc0 <- tpm_transmit)
These emit from tpm2_get_random() invoked by hwrng.
I guess 8.9 seconds is what we satisfy for the moment :-) I put out one more version where extras have been cut and performance fixes also have fixes tags as we most likely want these also to 6.10/6.11.
I attached kprobes using boot-time config and these appear when init process is already running:
start_event: (tpm_transmit+0x4/0x4b0)
end_event: (tpm_transmit_cmd+0x33/0xc0 <- tpm_transmit)
start_event: (tpm_transmit+0x4/0x4b0)
end_event: (tpm_transmit_cmd+0x33/0xc0 <- tpm_transmit)
start_event: (tpm_transmit+0x4/0x4b0)
end_event: (tpm_transmit_cmd+0x33/0xc0 <- tpm_transmit)
These emit from tpm2_get_random() invoked by hwrng.
I guess 8.9 seconds is what we satisfy for the moment :-) I put out one more version where extras have been cut and performance fixes also have fixes tags as we most likely want these also to 6.10/6.11.
0
1
0