[PATCH 1/2 v3] tpm: add a function that performs selftest + startup

Simon Glass sjg at chromium.org
Fri Feb 17 03:55:01 CET 2023 

Hi Ilias,

On Thu, 16 Feb 2023 at 14:10, Ilias Apalodimas
<ilias.apalodimas at linaro.org> wrote:
>
> Hi Simon,
>
> Apologies for the late reply.
>
> On Tue, Feb 07, 2023 at 06:38:54AM -0700, Simon Glass wrote:
> > Hi Ilias,
> >
> > On Mon, 6 Feb 2023 at 06:25, Ilias Apalodimas
> > <ilias.apalodimas at linaro.org> wrote:
> > >
> > > Hi Simon,
> > > On Sat, Jan 28, 2023 at 03:01:22PM -0700, Simon Glass wrote:
> > > > Hi Ilias,
> > > >
> > > > On Thu, 26 Jan 2023 at 01:18, Ilias Apalodimas
> > > > <ilias.apalodimas at linaro.org> wrote:
> > > > >
> > > > > As described in [0] if a command requires use of an untested algorithm
> > > > > or functional module, the TPM performs the test and then completes the
> > > > > command actions.
> > > > >
> > > > > Since we don't check for TPM_RC_NEEDS_TEST (which is the return code of
> > > > > the TPM in that case) and even if we would, it would complicate our TPM
> > > > > code for no apparent reason,  add a wrapper function that performs both
> > > > > the selftest and the startup sequence of the TPM.
> > > > >
> > > > > It's worth noting that this is implemented on TPMv2.0.  The code for
> > > > > 1.2 would look similar,  but I don't have a device available to test.
> > > > >
> > > > > [0]
> > > > > https://trustedcomputinggroup.org/wp-content/uploads/TPM-Rev-2.0-Part-1-Architecture-01.07-2014-03-13.pdf
> > > > > §12.3 Self-test modes
> > > > >
> > > > > Signed-off-by: Ilias Apalodimas <ilias.apalodimas at linaro.org>
> > > > > ---
> > > > > Changes since v2:
> > > > > - add tpm_init() to auto start
> > > > >
> > > > > Changes since v1:
> > > > > - Remove a superfluous if statement
> > > > > - Move function comments to the header file
> > > > >  include/tpm-v2.h  | 19 +++++++++++++++++++
> > > > >  include/tpm_api.h |  8 ++++++++
> > > > >  lib/tpm-v2.c      | 24 ++++++++++++++++++++++++
> > > > >  lib/tpm_api.c     |  8 ++++++++
> > > > >  4 files changed, 59 insertions(+)
> > > >
> > > > I think this is a good idea, but it should be implemented at the API
> > > > level. Please see below.
> > >
> > > It is implemented at the API level.  I could try testing this in QEMU using
> > > swtpm, but the driver I added for mmio is TPMv2 only.  So I don't really feel
> > > comfortable adding support for a device I can't test.  If someone has a
> > > tpm1.2 and is willing to test it, I can modify the patch accordingly.
> >
> > swtpm is a bit of a pain, as we discussed and as you are showing by
> > this comment. IMO it is good enough to use the sandbox testing for
> > this function. See below.
> >
>
> The comment says the exact opposite.  It never hints on any swtpm
> limitations or problems.

Manual testing is fine but it doesn't last. It is better to invest in
automated tests.

>
> fwiw swtpm is fine.  It even offers a socket we could hook against instead of
> re-inventing the wheel with the sandbox tpm (which is missing 95% of the
> tpm functionality anyway).  The reason we can't use it is that the u-boot
> driver only implements 2.0.

I wish I could convince you that emulators are not reinventing the
wheel. They are designed for testing. They run quickly, are easy to
debug and can test failure conditions, not just the happy path.

In Chrome OS we recently took our Zephyr code base from ~35% to 90%
test coverage using emulators, without a QEMU in sight[1]. Emulators
are a long-established software technique, one that has taken U-Boot a
long way.

>
> [...]
>
> > > > > +       rc = tpm_init(dev);
> > > > > +       if (rc && rc != -EBUSY)
> > > >
> > > > Does that work, with rc being unsigned?
> > >
> > >
> > > Yes integer promotion is fine here.  As long as we don't try to do anything
> > > silly e.g start doing math on the result or compare it against bigger types
> > > this is fine.
> >
> > OK
> >
> > > There was a patch from Sughosh in the past that you rejected and it was
> > > converting enough of the TPM API return calls to int.  I think the reason
> > > the API works with u32, is that the spec return codes from the device
> > > itself are described as u32.  But that shouldn't affect the internal U-Boot
> > > API.  I can send a patch afterwards moving the U-Boot TPM API functions and
> > > return int.
> >
> > Well, the question is, what is the return value? If it is the actual
> > TPM return code, then u32 is (unfortunately) correct. If it is an
> > errno then it should be int. The challenge is that some code wants to
> > know about the TPM internals, special error codes, etc. so we cannot
> > easily move to errno. Is that right?
>
> Maybe, I'll have to take a closer look, but in principle I don't see why
> the internal API should return device error codes verbatim.  If we care
> about the tpm error that much we can pass a u32 * and still return an int.
> In any case as we discussed integer promotion is fine here.

If we don't return the 'real' error codes then I worry that it will be
impossible to implement certain things. If we plan to map TPM errors
to errno, and can document all the cases, perhaps that would work, but
it seems like a lot of work.

>
> >
> > >
> > > >
> > > > > +               return rc;
> > > > > +       rc = tpm2_self_test(dev, TPMI_YES);
> > > >
> > > > Can we call tpm_self_test_full() ? If not, please update the API as needed.
> > > >
> > > > > +
> > > > > +       if (rc == TPM2_RC_INITIALIZE) {
> > > > > +               rc = tpm2_startup(dev, TPM2_SU_CLEAR);
> > > >
> > > > Should call tpm_startup()
> > >
> > >
> > > Same logic applies to both of these.  Yes that makes sense to use the top
> > > level API but only if we add support for 1.2 as well.  Otherwise the only
> > > thing this is going to create is circular dependencies between the v2 code
> > > and the API library.
> >
> > Then just add 1.2 support.
>
> That doesn't make any sense and I personally don't agree with the comment.
> The patch improves and fixes problems we have in TPM 2.0 which is what
> (hopefully) all devices use the last couple of years.
>
> Unfortunately I don't have the time to fix 1.2 and I don't see why this
> should hinder our efforts for decent TPM2.0 support.

The patch would actually be easier to write if you didn't limit it to
TPM2. To repeat myself, you are going around the TPM API we created,
because you don't want to use the sandbox emulator, which probably has
the functionality you need, or could do with 10 lines of code.

It really isn't saving time. Whose time is it saving?

>
> >
> > >
> > > >
> > > > > +               if (rc)
> > > > > +                       return rc;
> > > > > +
> > > > > +               rc = tpm2_self_test(dev, TPMI_YES);
> > > >
> > > > Again, tpm_self_test_full(). We are trying to provide a TPM API that
> > > > covers v1 and v2, to the extent possible.
> > >
> > > See above.
> > >
> > > >
> > > > > +       }
> > > > > +
> > > > > +       return rc;
> > > > > +}
> > > > > +
> > > >
> > > > Also please add a test for this to test/dm/tpm.c
> > >
> > > sure
> > >
> > > >
> > > > >  u32 tpm2_clear(struct udevice *dev, u32 handle, const char *pw,
> > > > >                const ssize_t pw_sz)
> > > > >  {
> > > > > diff --git a/lib/tpm_api.c b/lib/tpm_api.c
> > > > > index 7e8df8795ef3..5b2c11a277cc 100644
> > > > > --- a/lib/tpm_api.c
> > > > > +++ b/lib/tpm_api.c
> > > > > @@ -35,6 +35,14 @@ u32 tpm_startup(struct udevice *dev, enum tpm_startup_type mode)
> > > > >         }
> > > > >  }
> > > > >
> > > > > +u32 tpm_auto_start(struct udevice *dev)
> > > > > +{
> > > > > +       if (tpm_is_v2(dev))
> > > > > +               return tpm2_auto_start(dev);
> > > >
> > > > Hopefully all the code from above can move here.
> > >
> > > Repeating myself here, but I don't want to add untested code.  So I really
> > > prefer the patch as is, until someone verifies the 1.2 init sequence for
> > > me.
> >
> > I don't see any tests in this patch anyway. We do have some very basic
> > tests for the TPM in test/dm/tpm.c and I'm sure you can add your
> > function there. That should be enough for now. It really doesn't make
> > sense to work around the existing API just as a way of not
> > implementing it for something you cannot manually test. It just adds
> > confusion when some functions use the API and some don't.
>
> I've already added the tests and having selftests in test/dm/tpm.c makes
> sense.  What doesn't is requesting to fix 1.2 as well.  I'll send a v2 with
> the tests included, but I can't change the API side of things.
> I can send an RFC that fixes tpm1.2 if someone can test that on
> real hardware.  But I don't have time to spend adding 1.2 support in sandbox.

OK, well I will take the time to do this properly in the API. I will
keep track of how long I spend on it and let you know.

Regards,
Simon

[1] https://www.youtube.com/watch?v=WmAgXYTj4Fg

More information about the U-Boot mailing list