Re: [PATCH] tpm: Create cleanup class for tpm_buf

[Jarkko Sakkinen <jarkko@kernel.org>]

On Thu, Jun 26, 2025 at 11:49:15AM -0300, Jason Gunthorpe wrote:
> On Thu, Jun 26, 2025 at 12:37:56AM +0300, Jarkko Sakkinen wrote:
> > @@ -323,7 +323,7 @@ unsigned long tpm1_calc_ordinal_duration(struct tpm_chip *chip, u32 ordinal)
> >   */
> >  static int tpm1_startup(struct tpm_chip *chip)
> >  {
> > -	struct tpm_buf buf;
> > +	CLASS(tpm_buf, buf)();
> >  	int rc;
> >  
> >  	dev_info(&chip->dev, "starting up the TPM manually\n");
> > @@ -335,7 +335,6 @@ static int tpm1_startup(struct tpm_chip *chip)
> >  	tpm_buf_append_u16(&buf, TPM_ST_CLEAR);
> >  
> >  	rc = tpm_transmit_cmd(chip, &buf, 0, "attempting to start the TPM");
> > -	tpm_buf_destroy(&buf);
> >  	return rc;
> >  }
> 
> So, Linus has spoken negatively about just converting existing code to
> use cleanup.h, fearful it would introduce more bugs.

I did not do this for the sake of conversion. It's just that tpm_buf is
a pretty good fit for such construct, as it is always in function scope
and always heap allocated.

> I would certainly split this into more patches, and it would be nice
> if something mechanical like coccinelle could do the change.

I took this a bit in further:

https://lore.kernel.org/linux-integrity/aF2NNHilFfZwBoxA@kernel.org/T/#t

I did that few dozen times while developing this, running always at
minimum:

1. https://codeberg.org/jarkko/linux-tpmdd-test/src/branch/main/board/pc_x86_64/test_tpm2_kselftest.exp.in
2. https://codeberg.org/jarkko/linux-tpmdd-test/src/branch/main/board/pc_x86_64/test_tpm2_trusted.exp.in

A few times I run some ad-hoc tests too.

And despite 89% is mechanical work there was at least a dozen code
blocks where you need to understand the context too. So actually with
this careful manual work was not that bad idea in the end.

> 
> At least I would add the class and drop the tpm_buf_destroy() as one
> patch, and another would be to cleanup any empty gotos.
> 
> Also, I think the style guide for cleanup.h is to not use the
> variable block, so it should be more like:
> 
> CLASS(tpm_buf, buf)();
> if (!tpm_buf)
>    return -ENOMEM;
> 
> AFAICT, but that seems to be some kind of tribal knowledge.

This was improved in v2 :-) If you have some proposal how you'd
liked that version to be splitted, please give feedback.
> 
> Jason


BR, Jarkko