Re: [PATCH] selftest/trustedkeys: TPM 1.2 trusted keys test

[Jerry Snitselaar <jsnitsel@redhat.com>]

On Mon Sep 16 19, Mimi Zohar wrote:
>On Mon, 2019-09-16 at 00:42 -0700, Jerry Snitselaar wrote:
>> On Sun Sep 15 19, Mimi Zohar wrote:
>> >On Sun, 2019-09-15 at 16:52 -0400, Mimi Zohar wrote:
>> >> On Fri, 2019-09-13 at 15:08 +0100, Jarkko Sakkinen wrote:
>> >> > On Wed, Sep 11, 2019 at 08:00:40AM -0400, Mimi Zohar wrote:
>> >> > > On Tue, 2019-09-10 at 19:24 -0400, Mimi Zohar wrote:
>> >> > > > On Tue, 2019-09-10 at 19:18 -0400, Mimi Zohar wrote:
>> >> > > > > Create, save and load trusted keys test
>> >> > > >
>> >> > > > Creating trusted keys is failing with the following messages.  Any idea why?
>> >> > > >
>> >> > > > [  147.014653] tpm tpm0: A TPM error (34) occurred attempting to a send a command
>> >> > > > [  147.014678] trusted_key: srkseal failed (-1)
>> >> > > > [  147.014687] trusted_key: key_seal failed (-1)
>> >> > >
>> >> > > This is a regression, that needs to be resolved.  The test works on
>> >> > > kernels prior to 5.1.
>> >> >
>> >> > It breaks on 5.2?
>> >>
>> >> No, the regression is in 5.1.
>> >>
>> >> >
>> >> > Can you bisect the failing commit?
>> >>
>> >> git bisect start -- drivers/char/tpm/
>> >> git bisect bad
>> >> git bisect good v5.0 
>> >>
>> >> # first bad commit: [412eb585587a1dc43c9622db79de9663b6c4c238] tpm:
>> >> use tpm_buf in tpm_transmit_cmd() as the IO parameter
>> >
>> >In tpm_send(), setting buf.data directly to cmd, instead of calling
>> >tpm_buf_init() fixes the problem.
>> >
>> >Mimi
>> >
>> >
>>
>> The problem is that the command buffer is copied into the tpm_buf with
>> the memcpy, but after the tpm_transmit_cmd, nothing gets copied back
>> to be used by the code path that called tpm_send. There is code that
>> looks at that buffer after trusted_tpm_send returns. Both
>> security/keys/trusted.c and crypto/asymmetric_keys/asym_tpm.c run into
>> this.
>>
>> While playing around with it, adding a memcpy() after the transmit
>> call worked for me as well as setting buf.data directly instead of the
>> tpm_buf_init/memcpy/tpm_buf_destroy calls.
>
>Thanks!  I really appreciate your looking at the problem.  I haven't
>yet tried the fix with a TPM 2.0.
>>

I don't know the security/keys/trusted.c code well at all, but it looks
like parts of it are wired up to differentiate between tpm1.2 and tpm2.0,
and others are not. There is a tpm_seal_trusted and tpm_unseal_trusted that
gets used for tpm2, and are called in trusted_instantiate, which does a
tpm_is_tpm2 check, but trusted_update just calls the tpm1.2 code and
has no tpm2 checks from what I can tell.

>> I'm wondering if it would be worthwhile to convert the
>> security/keys/trusted.c and crypto/asymmertic_keys/aym_tpm.c code to
>> use the same tpm_buf and tpm_buf manipulation code as gets used in
>> drivers/char/tpm.
>
>For now, let's keep the regression fix, that is backported, simple and
>straight forward.
>

Agreed, I was just talking about in general. but it looks like Sumit has
already taken that task on.

>Trusted keys is evolving to be more than just TPM based keys.  I'm
>sure there will be more changes coming.  At least, now, there is a TPM
>1.2 regression test.
>
>thanks,
>
>Mimi
>