Re: [PATCH] tpm: fix a race condition in tpm2_unseal_trusted()

[Jason Gunthorpe <jgunthorpe@obsidianresearch.com>]

On Mon, Aug 29, 2016 at 06:25:21PM +0300, Jarkko Sakkinen wrote:
> On Sun, Aug 28, 2016 at 12:51:49PM -0600, Jason Gunthorpe wrote:
> > On Sun, Aug 28, 2016 at 08:36:52AM +0200, Jarkko Sakkinen wrote:
> > >  
> > > @@ -576,7 +576,8 @@ static int tpm2_load(struct tpm_chip *chip,
> > >  		goto out;
> > >  	}
> > >  
> > > -	rc = tpm_transmit_cmd(chip, buf.data, PAGE_SIZE, "loading blob");
> > > +	rc = tpm_transmit_cmd(chip, buf.data, PAGE_SIZE, TPM_TRANSMIT_UNLOCKED,
> > > +			      "loading blob");
> > 
> > I still don't like this, required mutex's should not be split outside the
> > function that needs them without more a more obvious indication:
> > 
> > > +	mutex_lock(&chip->tpm_mutex);
> > >  	rc = tpm2_load(chip, payload, options, &blob_handle);
> > >  	if (rc)
> > > -		return rc;
> > 
> > I recommend you stick with the idiom and do this:
> > 
> >         mutex_lock(&chip->tpm_mutex);
> >   	rc = tpm2_load(chip, payload, options, &blob_handle, TPM_TRANSMIT_UNLOCKED);
> > 
> > Which makes it easy to see we are doing it right everywhere.
> 
> Why consume stack for unnecessary stuff? This is a static function. For
> me this sounds like cutting hairs really.

Well, tpm2_load looks like any other normal command that would grab
the mutex, so something has to be done to indicate to the reader it is
the unlocked version.

I wouldn't worry about the stack, the compiler will inline that away
anyhow.

> One thing that would improve readability would be to rename internal
> functions tpm2_load and tpm2_unseal to tpm2_load_cmd and tpm2_unseal_cmd
> in order to underline that they are command wrappers and not to mix with
> tpm2_unseal_trusted().

That seems reasonable as well, as long as all _cmd varients are unlocked.

Jason