Re: [PATCH] tpm: avoid accessing cleared ops during shutdown

[Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>]

On Thu, Jul 16, 2020 at 10:38:00AM -0700, Guenter Roeck wrote:
> On Thu, Jul 16, 2020 at 10:28 AM Jarkko Sakkinen
> <jarkko.sakkinen@linux.intel.com> wrote:
> >
> > On Tue, Jul 14, 2020 at 08:48:38AM -0700, Guenter Roeck wrote:
> > > On Tue, Jul 14, 2020 at 4:32 AM Jarkko Sakkinen
> > > <jarkko.sakkinen@linux.intel.com> wrote:
> > > >
> > > > On Fri, Jul 10, 2020 at 11:25:44AM -0700, Andrey Pronin wrote:
> > > > > > Why does not tpm_del_char_device need this?
> > > > >
> > > > > "Not" is a typo in the sentence above, right? tpm_del_char_device *does*
> > > > > need the fix. When tpm_class_shutdown is called it sets chip->ops to
> > > > > NULL. If tpm_del_char_device is called after that, it doesn't check if
> > > > > chip->ops is NULL (normal kernel API and char device API calls go
> > > > > through tpm_try_get_ops, but tpm_del_char_device doesn't) and proceeds to
> > > > > call tpm2_shutdown(), which tries sending the command and dereferences
> > > > > chip->ops.
> > > >
> > > > It's a typo, yes. Sorry about that.
> > > >
> > > > tpm_class_shutdown() is essentially tail of tpm_del_char_device().
> > > >
> > > > To clean things up, I'd suggest dropping tpm_del_char_device() and
> > > > call tpm_class_shutdown() in tpm_chip_unregisters() along, and open
> > > > coding things that prepend it in tpm_del_char_device().
> > > >
> > >
> > > Personally I would have preferred two separate patches, one to fix the
> > > immediate problem (with Cc: stable) and one for the cleanup, but I
> > > guess merging both into one is ok as long as it is marked for stable.
> > >
> > > Thanks,
> > > Guenter
> >
> > Not sure about stable as this issue does not afaik concern earlier
> > kernel versions?
> >
> 
> I just had a quick look into linux-5.4.y, and it seemed to me that it
> is affected. Maybe I am wrong. Either case, we already applied this
> patch to all affected ChromeOS kernel branches, so from our
> perspective it doesn't really matter.
> 
> Thanks,
> Guenter

I'm fine with cc'ing stable after consideration given the benefits.

Given that conclusion, it is better to break this down to two part
series as you proposed.

/Jarkko