linux-kernel - Re: init_ima() adds 8 % to boot time

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:   Sun, 20 Feb 2022 10:19:37 -0500
From:   Mimi Zohar <zohar@...ux.ibm.com>
To:     Paul Menzel <pmenzel@...gen.mpg.de>,
        Dmitry Kasatkin <dmitry.kasatkin@...il.com>,
        Jarkko Sakkinen <jarkko@...nel.org>
Cc:     linux-integrity@...r.kernel.org,
        LKML <linux-kernel@...r.kernel.org>, Petr Vorel <pvorel@...e.cz>
Subject: Re: init_ima() adds 8 % to boot time

[Cc'ing Jarkko, Petr Vorel]

Hi Paul,

On Sat, 2022-02-19 at 10:44 +0100, Paul Menzel wrote:
> Dear Linux folks,
> 
> 
> Debian builds its Linux kernel image with `CONFIG_IMA=y` since version 
> 5.13.9 [1]. Unfortunately, on the Dell Latitude E7250 `init_ima` takes 
> around 33 ms, adding 8 % to the boot time up to loading the initrd.
> 
>      [    0.000000] Linux version 5.17.0-rc4-amd64 
> (debian-kernel@...ts.debian.org) (gcc-11 (Debian 11.2.0-16) 11.2.0, GNU 
> ld (GNU Binutils for Debian) 2.38) #1 SMP PREEMPT Debian 5.17~rc4-1~exp1 
> (2022-02-18)
>      […]
>      [    0.238520] calling  init_tis+0x0/0xde @ 1
>      [    0.254749] tpm_tis 00:08: 1.2 TPM (device-id 0x3205, rev-id 80)
>      [    0.285665] initcall init_tis+0x0/0xde returned 0 after 46038 usecs
>      […]
>      [    0.301327] calling  init_ima+0x0/0xb5 @ 1
>      [    0.301332] ima: Allocated hash algorithm: sha256
>      [    0.335502] ima: No architecture policies found
>      [    0.335520] initcall init_ima+0x0/0xb5 returned 0 after 33389 usecs
>      […]
>      [    0.447312] Run /init as init process
> 
> Tracing `init_ima` with a depth of 5 shows 
> `ima_calc_boot_aggregate_tfm()` takes 24 ms, and 
> `ima_add_template_entry()` takes 10 ms.
> 
>          1.282630 |   1)   swapper-1    |               | 
> ima_add_boot_aggregate() {
>          1.282631 |   1)   swapper-1    |               | 
> ima_calc_boot_agg:0regate() {
>          1.282631 |   1)   swapper-1    |   0.153 us    | 
> ima_alloc_tfm();
>          1.282631 |   1)   swapper-1    | * 24404.59 us | 
> ima_calc_boot_aggregate_tfm();
>          1.307037 |   1)   swapper-1    |   0.482 us    | 
> ima_free_tfm.part.0();
>          1.307038 |   1)   swapper-1    | * 24407.06 us |        } /* 
> ima_calc_boot_aggregate */
>          1.307038 |   1)   swapper-1    |               | 
> ima_alloc_init_template() {
>          1.307038 |   1)   swapper-1    |   0.173 us    | 
> ima_template_desc_current();
>          1.307039 |   1)   swapper-1    |   0.836 us    | 
> __kmalloc();
>          1.307040 |   1)   swapper-1    |   0.580 us    | 
> __kmalloc();
>          1.307041 |   1)   swapper-1    |   1.555 us    | 
> ima_eventdigest_ng_init();
>          1.307043 |   1)   swapper-1    |   1.275 us    | 
> ima_eventname_ng_init();
>          1.307044 |   1)   swapper-1    |   0.256 us    | 
> ima_eventsig_init();
>          1.307045 |   1)   swapper-1    |   6.618 us    |        } /* 
> ima_alloc_init_template */
>          1.307045 |   1)   swapper-1    |               | 
> ima_store_template() {
>          1.307045 |   1)   swapper-1    |   5.049 us    | 
> ima_calc_field_array_hash();
>          1.307051 |   1)   swapper-1    | # 9316.953 us | 
> ima_add_template_entry();
>          1.316369 |   1)   swapper-1    | # 9323.728 us |        } /* 
> ima_store_template */
>          1.316369 |   1)   swapper-1    | * 33738.54 us |      } /* 
> ima_add_boot_aggregate */
> 
> Tracing `ima_calc_boot_aggregate_tfm()` (attached) shows that the first 
> `tpm1_pcr_read()` takes 16 ms in `tpm_transmit()`. Is communicating with 
> the TPM supposed to be that slow?
> 
> In the last years, Linux decreased it’s boot time a lot, so do you see a 
> way to move things out of the hot path and get `init_ima` well below 10 
> ms? (As systems get faster and faster, having systems with standard 
> distributions to be up below two seconds after pressing the power button 
> should be a reasonable goal (500 ms firmware (like coreboot) + 500 ms 
> Linux kernel + 1 s user space).
> 
> 
> [1]: 
> https://salsa.debian.org/kernel-team/linux/-/commit/6e679322d7d98d30b4a8a3d1b659c899a6e9d4df

Thank you including the initial and other TPM delays.  The main reason
for the "boot_aggregate" is to tie the pre-OS measurements to the post
OS measurement list.  Without the TPM based 'boot_aggregate', any IMA
measurement list could be used to verify a TPM quote.  The
'boot_aggregate' is calculated, originally, based on PCRs 0 - 7 and
more recently may include PCRs 8 & 9 as well.  The 'boot_aggregate' is
the first record in the IMA measurement list and the first record after
a soft reboot (kexec).  It is the one and only IMA measurement record
not dependent on policy.

There are TPM 1.2 & 2.0 standards' requirements, but there are also
buggy TPMs which don't adhere to them to such an extent that IMA goes
into 'TPM-bypass' mode. Perhaps for those not interested in extending
the concepts of trusted boot to the running OS, defining a new boot
command line option to force IMA into this 'TPM-bypass' mode would be
an acceptable alternative to the delay.  The IMA measurement list would
still include a 'boot_aggregate' record, but one containing 0's.

thanks,

Mimi