Re: [PATCH v2 1/4] tpm: Implement tpm_chip_find() and tpm_chip_put() for other subsystems

[Stefan Berger]

On 06/21/2018 03:06 PM, Jason Gunthorpe wrote:
> On Thu, Jun 21, 2018 at 02:19:44PM -0400, Stefan Berger wrote:
> > On 06/21/2018 01:56 PM, Jason Gunthorpe wrote:
> > > On Thu, Jun 21, 2018 at 01:45:03PM -0400, Stefan Berger wrote:
> > > > On 06/21/2018 01:15 PM, Jarkko Sakkinen wrote:
> > > > > On Wed, Jun 20, 2018 at 04:42:33PM -0400, Stefan Berger wrote:
> > > > > > Implement tpm_chip_find() for other subsystems to find a TPM chip and
> > > > > > get a reference to that chip. Once done with using the chip, the reference
> > > > > > is released using tpm_chip_put().
> > > > > >
> > > > > > Signed-off-by: Stefan Berger <stefanb@xxxxxxxxxxxxxxxxxx>
> > > > > You should sort this out in a way that we don't end up with duplicate
> > > > > functions.
> > > > Do you want me to create a function *like* tpm_chip_find_get() that takes an
> > > > additional parameter whether to get the ops semaphore and have that function
> > > > called by the existing tpm_chip_find_get() and the new tpm_chip_find(). The
> > > > latter would then not get the ops semphore. I didn't want to do this since
> > > > one time the function returns with a lock held and the other time not.
> > > Another option, and I haven't looked, is to revise the callers of
> > > tpm_chip_find_get to not require it to hold the ops semaphore for
> > > them.
> > We have tpm_chip_unregister calling tpm_del_char_device to set the ops to
> > NULL once a chip is unregistered. All existing callers, if they pass in a
> > tpm_chip != NULL, currently fail if the ops are NULL. (If they pass in
> > tpm_chip = NULL, they shouldn't find a chip once ops are null and it has
> > been removed from the IDR). I wouldn't change that since IMA will call in
> > with a tpm_chip != NULL and we want to protect the ops. All existing code
> > within the tpm subsystem does seem to call tpm_chip_find_get() with a NULL
> > pointer, though. Also trusted keys seems to pass in a NULL pointer every
> > time.
> >
> > > Either by giving them an API to do it, or revising the TPM entry
> > > points to do it.
> > >
> > > I didn't look, but how did the ops semaphore get grabbed in your
> > > revised patches? They do grab it, right?
> > The revised patches do not touch the existing code much but will call
> > tpm_chip_find_get() and get that semaphore every time before the ops are
> > used. IMA is the only caller of tpm_chip_find() that now gets an additional
> > reference to the tpm_chip and these APIs get called like this from IMA:
> >
> > ima init: chip = tpm_chip_find()
> >
> > ima::tpm: tpm_chip_find_get(chip) ... tpm_put_ops(chip)
> >
> > ima::tpm: tpm_chip_find_get(chip) ... tpm_put_ops(chip)
> >
> > [repeat]
> >
> > ima shutdown: tpm_chip_put(chip)
> Maybe just change tpm_chip_find_get() into tpm_get_ops(chip) and
> convert all callers?

And then re-introduce tpm_chip_find_get() for IMA to call ?