On Thu, Oct 11, 2018 at 01:27:58PM +0000, Winkler, Tomas wrote:
> >
> > On Wed, Oct 10, 2018 at 08:06:38AM -0600, Jason Gunthorpe wrote:
> > > On Wed, Oct 10, 2018 at 03:38:17PM +0200, Gustavo A. R. Silva wrote:
> > > > Currently, there are some values assigned to variable *rc*, which
> > > > are never actually used in any computation, because such variable is
> > > > updated at line 550, before they can be used:
> > > >
> > > > 549out:
> > > > 550 rc = tpm_go_idle(chip, flags);
> > > > 551 if (rc)
> > > > 552 goto out;
> > > >
> > > > Fix this by removing such assignments.
> > >
> > > Should this be done by not quashing rc during the error unwind rather
> > > than dropping the errors?
> >
> > Yeah.`
> >
> > Wondering if tpm_go_idle() should simply be a void-function i.e. issue just a
> > warning inside (disclaimer: did not revisit its code when writing this).
>
> We did have rather a long discussion about it when it was merged.
> There are two flows that may crash
> rc = tpm2_commit_space()
>
> but you still can need to
>
> rc = go_idle()
>
> which also may crash which may override the previous value.
>
> Frankly the second one is fatal, the stack will go out of sync.
> We may do void here as the stack will crash in a subsequent command.
>
> The 'goto out' is quite a bug, probably caused by code movement.
I just looked at the code properly and noticed that there is a regression
caused by 627448e85c76 ("tpm: separate cmd_ready/go_idle from
runtime_pm") i.e. when tpm_go_idle() fails it loops back and retries
tpm_go_idle().
/Jarkko
